drivers/md/raid5.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * raid5.c : Multiple Devices driver for Linux
   4  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   5  *         Copyright (C) 1999, 2000 Ingo Molnar
   6  *         Copyright (C) 2002, 2003 H. Peter Anvin
   7  *
   8  * RAID-4/5/6 management functions.
   9  * Thanks to Penguin Computing for making the RAID-6 development possible
  10  * by donating a test server!
  11  */
  12
  13 /*
  14  * BITMAP UNPLUGGING:
  15  *
  16  * The sequencing for updating the bitmap reliably is a little
  17  * subtle (and I got it wrong the first time) so it deserves some
  18  * explanation.
  19  *
  20  * We group bitmap updates into batches.  Each batch has a number.
  21  * We may write out several batches at once, but that isn't very important.
  22  * conf->seq_write is the number of the last batch successfully written.
  23  * conf->seq_flush is the number of the last batch that was closed to
  24  *    new additions.
  25  * When we discover that we will need to write to any block in a stripe
  26  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  27  * the number of the batch it will be in. This is seq_flush+1.
  28  * When we are ready to do a write, if that batch hasn't been written yet,
  29  *   we plug the array and queue the stripe for later.
  30  * When an unplug happens, we increment bm_flush, thus closing the current
  31  *   batch.
  32  * When we notice that bm_flush > bm_write, we write out all pending updates
  33  * to the bitmap, and advance bm_write to where bm_flush was.
  34  * This may occasionally write a bit out twice, but is sure never to
  35  * miss any bits.
  36  */
  37
  38 #include <linux/blkdev.h>
  39 #include <linux/kthread.h>
  40 #include <linux/raid/pq.h>
  41 #include <linux/async_tx.h>
  42 #include <linux/module.h>
  43 #include <linux/async.h>
  44 #include <linux/seq_file.h>
  45 #include <linux/cpu.h>
  46 #include <linux/slab.h>
  47 #include <linux/ratelimit.h>
  48 #include <linux/nodemask.h>
  49
  50 #include <trace/events/block.h>
  51 #include <linux/list_sort.h>
  52
  53 #include "md.h"
  54 #include "raid5.h"
  55 #include "raid0.h"
  56 #include "md-bitmap.h"
  57 #include "raid5-log.h"
  58
  59 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
  60
  61 #define cpu_to_group(cpu) cpu_to_node(cpu)
  62 #define ANY_GROUP NUMA_NO_NODE
  63
  64 #define RAID5_MAX_REQ_STRIPES 256
  65
  66 static bool devices_handle_discard_safely = false;
  67 module_param(devices_handle_discard_safely, bool, 0644);
  68 MODULE_PARM_DESC(devices_handle_discard_safely,
  69                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  70 static struct workqueue_struct *raid5_wq;
  71
  72 static void raid5_quiesce(struct mddev *mddev, int quiesce);
  73
  74 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  75 {
  76         int hash = (sect >> RAID5_STRIPE_SHIFT(conf)) & HASH_MASK;
  77         return &conf->stripe_hashtbl[hash];
  78 }
  79
  80 static inline int stripe_hash_locks_hash(struct r5conf *conf, sector_t sect)
  81 {
  82         return (sect >> RAID5_STRIPE_SHIFT(conf)) & STRIPE_HASH_LOCKS_MASK;
  83 }
  84
  85 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  86         __acquires(&conf->device_lock)
  87 {
  88         spin_lock_irq(conf->hash_locks + hash);
  89         spin_lock(&conf->device_lock);
  90 }
  91
  92 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  93         __releases(&conf->device_lock)
  94 {
  95         spin_unlock(&conf->device_lock);
  96         spin_unlock_irq(conf->hash_locks + hash);
  97 }
  98
  99 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 100         __acquires(&conf->device_lock)
 101 {
 102         int i;
 103         spin_lock_irq(conf->hash_locks);
 104         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 105                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 106         spin_lock(&conf->device_lock);
 107 }
 108
 109 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 110         __releases(&conf->device_lock)
 111 {
 112         int i;
 113         spin_unlock(&conf->device_lock);
 114         for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 115                 spin_unlock(conf->hash_locks + i);
 116         spin_unlock_irq(conf->hash_locks);
 117 }
 118
 119 /* Find first data disk in a raid6 stripe */
 120 static inline int raid6_d0(struct stripe_head *sh)
 121 {
 122         if (sh->ddf_layout)
 123                 /* ddf always start from first device */
 124                 return 0;
 125         /* md starts just after Q block */
 126         if (sh->qd_idx == sh->disks - 1)
 127                 return 0;
 128         else
 129                 return sh->qd_idx + 1;
 130 }
 131 static inline int raid6_next_disk(int disk, int raid_disks)
 132 {
 133         disk++;
 134         return (disk < raid_disks) ? disk : 0;
 135 }
 136
 137 /* When walking through the disks in a raid5, starting at raid6_d0,
 138  * We need to map each disk to a 'slot', where the data disks are slot
 139  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 140  * is raid_disks-1.  This help does that mapping.
 141  */
 142 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 143                              int *count, int syndrome_disks)
 144 {
 145         int slot = *count;
 146
 147         if (sh->ddf_layout)
 148                 (*count)++;
 149         if (idx == sh->pd_idx)
 150                 return syndrome_disks;
 151         if (idx == sh->qd_idx)
 152                 return syndrome_disks + 1;
 153         if (!sh->ddf_layout)
 154                 (*count)++;
 155         return slot;
 156 }
 157
 158 static void print_raid5_conf(struct r5conf *conf);
 159
 160 static int stripe_operations_active(struct stripe_head *sh)
 161 {
 162         return sh->check_state || sh->reconstruct_state ||
 163                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 164                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 165 }
 166
 167 static bool stripe_is_lowprio(struct stripe_head *sh)
 168 {
 169         return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 170                 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 171                !test_bit(STRIPE_R5C_CACHING, &sh->state);
 172 }
 173
 174 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 175         __must_hold(&sh->raid_conf->device_lock)
 176 {
 177         struct r5conf *conf = sh->raid_conf;
 178         struct r5worker_group *group;
 179         int thread_cnt;
 180         int i, cpu = sh->cpu;
 181
 182         if (!cpu_online(cpu)) {
 183                 cpu = cpumask_any(cpu_online_mask);
 184                 sh->cpu = cpu;
 185         }
 186
 187         if (list_empty(&sh->lru)) {
 188                 struct r5worker_group *group;
 189                 group = conf->worker_groups + cpu_to_group(cpu);
 190                 if (stripe_is_lowprio(sh))
 191                         list_add_tail(&sh->lru, &group->loprio_list);
 192                 else
 193                         list_add_tail(&sh->lru, &group->handle_list);
 194                 group->stripes_cnt++;
 195                 sh->group = group;
 196         }
 197
 198         if (conf->worker_cnt_per_group == 0) {
 199                 md_wakeup_thread(conf->mddev->thread);
 200                 return;
 201         }
 202
 203         group = conf->worker_groups + cpu_to_group(sh->cpu);
 204
 205         group->workers[0].working = true;
 206         /* at least one worker should run to avoid race */
 207         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 208
 209         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 210         /* wakeup more workers */
 211         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 212                 if (group->workers[i].working == false) {
 213                         group->workers[i].working = true;
 214                         queue_work_on(sh->cpu, raid5_wq,
 215                                       &group->workers[i].work);
 216                         thread_cnt--;
 217                 }
 218         }
 219 }
 220
 221 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 222                               struct list_head *temp_inactive_list)
 223         __must_hold(&conf->device_lock)
 224 {
 225         int i;
 226         int injournal = 0;      /* number of date pages with R5_InJournal */
 227
 228         BUG_ON(!list_empty(&sh->lru));
 229         BUG_ON(atomic_read(&conf->active_stripes)==0);
 230
 231         if (r5c_is_writeback(conf->log))
 232                 for (i = sh->disks; i--; )
 233                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 234                                 injournal++;
 235         /*
 236          * In the following cases, the stripe cannot be released to cached
 237          * lists. Therefore, we make the stripe write out and set
 238          * STRIPE_HANDLE:
 239          *   1. when quiesce in r5c write back;
 240          *   2. when resync is requested fot the stripe.
 241          */
 242         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 243             (conf->quiesce && r5c_is_writeback(conf->log) &&
 244              !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 245                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 246                         r5c_make_stripe_write_out(sh);
 247                 set_bit(STRIPE_HANDLE, &sh->state);
 248         }
 249
 250         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 251                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 252                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 253                         list_add_tail(&sh->lru, &conf->delayed_list);
 254                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 255                            sh->bm_seq - conf->seq_write > 0)
 256                         list_add_tail(&sh->lru, &conf->bitmap_list);
 257                 else {
 258                         clear_bit(STRIPE_DELAYED, &sh->state);
 259                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 260                         if (conf->worker_cnt_per_group == 0) {
 261                                 if (stripe_is_lowprio(sh))
 262                                         list_add_tail(&sh->lru,
 263                                                         &conf->loprio_list);
 264                                 else
 265                                         list_add_tail(&sh->lru,
 266                                                         &conf->handle_list);
 267                         } else {
 268                                 raid5_wakeup_stripe_thread(sh);
 269                                 return;
 270                         }
 271                 }
 272                 md_wakeup_thread(conf->mddev->thread);
 273         } else {
 274                 BUG_ON(stripe_operations_active(sh));
 275                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 276                         if (atomic_dec_return(&conf->preread_active_stripes)
 277                             < IO_THRESHOLD)
 278                                 md_wakeup_thread(conf->mddev->thread);
 279                 atomic_dec(&conf->active_stripes);
 280                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 281                         if (!r5c_is_writeback(conf->log))
 282                                 list_add_tail(&sh->lru, temp_inactive_list);
 283                         else {
 284                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 285                                 if (injournal == 0)
 286                                         list_add_tail(&sh->lru, temp_inactive_list);
 287                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 288                                         /* full stripe */
 289                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 290                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 291                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 292                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 293                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 294                                         r5c_check_cached_full_stripe(conf);
 295                                 } else
 296                                         /*
 297                                          * STRIPE_R5C_PARTIAL_STRIPE is set in
 298                                          * r5c_try_caching_write(). No need to
 299                                          * set it again.
 300                                          */
 301                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 302                         }
 303                 }
 304         }
 305 }
 306
 307 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 308                              struct list_head *temp_inactive_list)
 309         __must_hold(&conf->device_lock)
 310 {
 311         if (atomic_dec_and_test(&sh->count))
 312                 do_release_stripe(conf, sh, temp_inactive_list);
 313 }
 314
 315 /*
 316  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 317  *
 318  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 319  * given time. Adding stripes only takes device lock, while deleting stripes
 320  * only takes hash lock.
 321  */
 322 static void release_inactive_stripe_list(struct r5conf *conf,
 323                                          struct list_head *temp_inactive_list,
 324                                          int hash)
 325 {
 326         int size;
 327         bool do_wakeup = false;
 328         unsigned long flags;
 329
 330         if (hash == NR_STRIPE_HASH_LOCKS) {
 331                 size = NR_STRIPE_HASH_LOCKS;
 332                 hash = NR_STRIPE_HASH_LOCKS - 1;
 333         } else
 334                 size = 1;
 335         while (size) {
 336                 struct list_head *list = &temp_inactive_list[size - 1];
 337
 338                 /*
 339                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 340                  * remove stripes from the list
 341                  */
 342                 if (!list_empty_careful(list)) {
 343                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 344                         if (list_empty(conf->inactive_list + hash) &&
 345                             !list_empty(list))
 346                                 atomic_dec(&conf->empty_inactive_list_nr);
 347                         list_splice_tail_init(list, conf->inactive_list + hash);
 348                         do_wakeup = true;
 349                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 350                 }
 351                 size--;
 352                 hash--;
 353         }
 354
 355         if (do_wakeup) {
 356                 wake_up(&conf->wait_for_stripe);
 357                 if (atomic_read(&conf->active_stripes) == 0)
 358                         wake_up(&conf->wait_for_quiescent);
 359                 if (conf->retry_read_aligned)
 360                         md_wakeup_thread(conf->mddev->thread);
 361         }
 362 }
 363
 364 static int release_stripe_list(struct r5conf *conf,
 365                                struct list_head *temp_inactive_list)
 366         __must_hold(&conf->device_lock)
 367 {
 368         struct stripe_head *sh, *t;
 369         int count = 0;
 370         struct llist_node *head;
 371
 372         head = llist_del_all(&conf->released_stripes);
 373         head = llist_reverse_order(head);
 374         llist_for_each_entry_safe(sh, t, head, release_list) {
 375                 int hash;
 376
 377                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 378                 smp_mb();
 379                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 380                 /*
 381                  * Don't worry the bit is set here, because if the bit is set
 382                  * again, the count is always > 1. This is true for
 383                  * STRIPE_ON_UNPLUG_LIST bit too.
 384                  */
 385                 hash = sh->hash_lock_index;
 386                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 387                 count++;
 388         }
 389
 390         return count;
 391 }
 392
 393 void raid5_release_stripe(struct stripe_head *sh)
 394 {
 395         struct r5conf *conf = sh->raid_conf;
 396         unsigned long flags;
 397         struct list_head list;
 398         int hash;
 399         bool wakeup;
 400
 401         /* Avoid release_list until the last reference.
 402          */
 403         if (atomic_add_unless(&sh->count, -1, 1))
 404                 return;
 405
 406         if (unlikely(!conf->mddev->thread) ||
 407                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 408                 goto slow_path;
 409         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 410         if (wakeup)
 411                 md_wakeup_thread(conf->mddev->thread);
 412         return;
 413 slow_path:
 414         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 415         if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
 416                 INIT_LIST_HEAD(&list);
 417                 hash = sh->hash_lock_index;
 418                 do_release_stripe(conf, sh, &list);
 419                 spin_unlock_irqrestore(&conf->device_lock, flags);
 420                 release_inactive_stripe_list(conf, &list, hash);
 421         }
 422 }
 423
 424 static inline void remove_hash(struct stripe_head *sh)
 425 {
 426         pr_debug("remove_hash(), stripe %llu\n",
 427                 (unsigned long long)sh->sector);
 428
 429         hlist_del_init(&sh->hash);
 430 }
 431
 432 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 433 {
 434         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 435
 436         pr_debug("insert_hash(), stripe %llu\n",
 437                 (unsigned long long)sh->sector);
 438
 439         hlist_add_head(&sh->hash, hp);
 440 }
 441
 442 /* find an idle stripe, make sure it is unhashed, and return it. */
 443 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 444 {
 445         struct stripe_head *sh = NULL;
 446         struct list_head *first;
 447
 448         if (list_empty(conf->inactive_list + hash))
 449                 goto out;
 450         first = (conf->inactive_list + hash)->next;
 451         sh = list_entry(first, struct stripe_head, lru);
 452         list_del_init(first);
 453         remove_hash(sh);
 454         atomic_inc(&conf->active_stripes);
 455         BUG_ON(hash != sh->hash_lock_index);
 456         if (list_empty(conf->inactive_list + hash))
 457                 atomic_inc(&conf->empty_inactive_list_nr);
 458 out:
 459         return sh;
 460 }
 461
 462 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
 463 static void free_stripe_pages(struct stripe_head *sh)
 464 {
 465         int i;
 466         struct page *p;
 467
 468         /* Have not allocate page pool */
 469         if (!sh->pages)
 470                 return;
 471
 472         for (i = 0; i < sh->nr_pages; i++) {
 473                 p = sh->pages[i];
 474                 if (p)
 475                         put_page(p);
 476                 sh->pages[i] = NULL;
 477         }
 478 }
 479
 480 static int alloc_stripe_pages(struct stripe_head *sh, gfp_t gfp)
 481 {
 482         int i;
 483         struct page *p;
 484
 485         for (i = 0; i < sh->nr_pages; i++) {
 486                 /* The page have allocated. */
 487                 if (sh->pages[i])
 488                         continue;
 489
 490                 p = alloc_page(gfp);
 491                 if (!p) {
 492                         free_stripe_pages(sh);
 493                         return -ENOMEM;
 494                 }
 495                 sh->pages[i] = p;
 496         }
 497         return 0;
 498 }
 499
 500 static int
 501 init_stripe_shared_pages(struct stripe_head *sh, struct r5conf *conf, int disks)
 502 {
 503         int nr_pages, cnt;
 504
 505         if (sh->pages)
 506                 return 0;
 507
 508         /* Each of the sh->dev[i] need one conf->stripe_size */
 509         cnt = PAGE_SIZE / conf->stripe_size;
 510         nr_pages = (disks + cnt - 1) / cnt;
 511
 512         sh->pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 513         if (!sh->pages)
 514                 return -ENOMEM;
 515         sh->nr_pages = nr_pages;
 516         sh->stripes_per_page = cnt;
 517         return 0;
 518 }
 519 #endif
 520
 521 static void shrink_buffers(struct stripe_head *sh)
 522 {
 523         int i;
 524         int num = sh->raid_conf->pool_size;
 525
 526 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 527         for (i = 0; i < num ; i++) {
 528                 struct page *p;
 529
 530                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 531                 p = sh->dev[i].page;
 532                 if (!p)
 533                         continue;
 534                 sh->dev[i].page = NULL;
 535                 put_page(p);
 536         }
 537 #else
 538         for (i = 0; i < num; i++)
 539                 sh->dev[i].page = NULL;
 540         free_stripe_pages(sh); /* Free pages */
 541 #endif
 542 }
 543
 544 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 545 {
 546         int i;
 547         int num = sh->raid_conf->pool_size;
 548
 549 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 550         for (i = 0; i < num; i++) {
 551                 struct page *page;
 552
 553                 if (!(page = alloc_page(gfp))) {
 554                         return 1;
 555                 }
 556                 sh->dev[i].page = page;
 557                 sh->dev[i].orig_page = page;
 558                 sh->dev[i].offset = 0;
 559         }
 560 #else
 561         if (alloc_stripe_pages(sh, gfp))
 562                 return -ENOMEM;
 563
 564         for (i = 0; i < num; i++) {
 565                 sh->dev[i].page = raid5_get_dev_page(sh, i);
 566                 sh->dev[i].orig_page = sh->dev[i].page;
 567                 sh->dev[i].offset = raid5_get_page_offset(sh, i);
 568         }
 569 #endif
 570         return 0;
 571 }
 572
 573 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 574                             struct stripe_head *sh);
 575
 576 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 577 {
 578         struct r5conf *conf = sh->raid_conf;
 579         int i, seq;
 580
 581         BUG_ON(atomic_read(&sh->count) != 0);
 582         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 583         BUG_ON(stripe_operations_active(sh));
 584         BUG_ON(sh->batch_head);
 585
 586         pr_debug("init_stripe called, stripe %llu\n",
 587                 (unsigned long long)sector);
 588 retry:
 589         seq = read_seqcount_begin(&conf->gen_lock);
 590         sh->generation = conf->generation - previous;
 591         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 592         sh->sector = sector;
 593         stripe_set_idx(sector, conf, previous, sh);
 594         sh->state = 0;
 595
 596         for (i = sh->disks; i--; ) {
 597                 struct r5dev *dev = &sh->dev[i];
 598
 599                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 600                     test_bit(R5_LOCKED, &dev->flags)) {
 601                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 602                                (unsigned long long)sh->sector, i, dev->toread,
 603                                dev->read, dev->towrite, dev->written,
 604                                test_bit(R5_LOCKED, &dev->flags));
 605                         WARN_ON(1);
 606                 }
 607                 dev->flags = 0;
 608                 dev->sector = raid5_compute_blocknr(sh, i, previous);
 609         }
 610         if (read_seqcount_retry(&conf->gen_lock, seq))
 611                 goto retry;
 612         sh->overwrite_disks = 0;
 613         insert_hash(conf, sh);
 614         sh->cpu = smp_processor_id();
 615         set_bit(STRIPE_BATCH_READY, &sh->state);
 616 }
 617
 618 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 619                                          short generation)
 620 {
 621         struct stripe_head *sh;
 622
 623         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 624         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 625                 if (sh->sector == sector && sh->generation == generation)
 626                         return sh;
 627         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 628         return NULL;
 629 }
 630
 631 static struct stripe_head *find_get_stripe(struct r5conf *conf,
 632                 sector_t sector, short generation, int hash)
 633 {
 634         int inc_empty_inactive_list_flag;
 635         struct stripe_head *sh;
 636
 637         sh = __find_stripe(conf, sector, generation);
 638         if (!sh)
 639                 return NULL;
 640
 641         if (atomic_inc_not_zero(&sh->count))
 642                 return sh;
 643
 644         /*
 645          * Slow path. The reference count is zero which means the stripe must
 646          * be on a list (sh->lru). Must remove the stripe from the list that
 647          * references it with the device_lock held.
 648          */
 649
 650         spin_lock(&conf->device_lock);
 651         if (!atomic_read(&sh->count)) {
 652                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 653                         atomic_inc(&conf->active_stripes);
 654                 BUG_ON(list_empty(&sh->lru) &&
 655                        !test_bit(STRIPE_EXPANDING, &sh->state));
 656                 inc_empty_inactive_list_flag = 0;
 657                 if (!list_empty(conf->inactive_list + hash))
 658                         inc_empty_inactive_list_flag = 1;
 659                 list_del_init(&sh->lru);
 660                 if (list_empty(conf->inactive_list + hash) &&
 661                     inc_empty_inactive_list_flag)
 662                         atomic_inc(&conf->empty_inactive_list_nr);
 663                 if (sh->group) {
 664                         sh->group->stripes_cnt--;
 665                         sh->group = NULL;
 666                 }
 667         }
 668         atomic_inc(&sh->count);
 669         spin_unlock(&conf->device_lock);
 670
 671         return sh;
 672 }
 673
 674 /*
 675  * Need to check if array has failed when deciding whether to:
 676  *  - start an array
 677  *  - remove non-faulty devices
 678  *  - add a spare
 679  *  - allow a reshape
 680  * This determination is simple when no reshape is happening.
 681  * However if there is a reshape, we need to carefully check
 682  * both the before and after sections.
 683  * This is because some failed devices may only affect one
 684  * of the two sections, and some non-in_sync devices may
 685  * be insync in the section most affected by failed devices.
 686  *
 687  * Most calls to this function hold &conf->device_lock. Calls
 688  * in raid5_run() do not require the lock as no other threads
 689  * have been started yet.
 690  */
 691 int raid5_calc_degraded(struct r5conf *conf)
 692 {
 693         int degraded, degraded2;
 694         int i;
 695
 696         degraded = 0;
 697         for (i = 0; i < conf->previous_raid_disks; i++) {
 698                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
 699
 700                 if (rdev && test_bit(Faulty, &rdev->flags))
 701                         rdev = READ_ONCE(conf->disks[i].replacement);
 702                 if (!rdev || test_bit(Faulty, &rdev->flags))
 703                         degraded++;
 704                 else if (test_bit(In_sync, &rdev->flags))
 705                         ;
 706                 else
 707                         /* not in-sync or faulty.
 708                          * If the reshape increases the number of devices,
 709                          * this is being recovered by the reshape, so
 710                          * this 'previous' section is not in_sync.
 711                          * If the number of devices is being reduced however,
 712                          * the device can only be part of the array if
 713                          * we are reverting a reshape, so this section will
 714                          * be in-sync.
 715                          */
 716                         if (conf->raid_disks >= conf->previous_raid_disks)
 717                                 degraded++;
 718         }
 719         if (conf->raid_disks == conf->previous_raid_disks)
 720                 return degraded;
 721         degraded2 = 0;
 722         for (i = 0; i < conf->raid_disks; i++) {
 723                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
 724
 725                 if (rdev && test_bit(Faulty, &rdev->flags))
 726                         rdev = READ_ONCE(conf->disks[i].replacement);
 727                 if (!rdev || test_bit(Faulty, &rdev->flags))
 728                         degraded2++;
 729                 else if (test_bit(In_sync, &rdev->flags))
 730                         ;
 731                 else
 732                         /* not in-sync or faulty.
 733                          * If reshape increases the number of devices, this
 734                          * section has already been recovered, else it
 735                          * almost certainly hasn't.
 736                          */
 737                         if (conf->raid_disks <= conf->previous_raid_disks)
 738                                 degraded2++;
 739         }
 740         if (degraded2 > degraded)
 741                 return degraded2;
 742         return degraded;
 743 }
 744
 745 static bool has_failed(struct r5conf *conf)
 746 {
 747         int degraded = conf->mddev->degraded;
 748
 749         if (test_bit(MD_BROKEN, &conf->mddev->flags))
 750                 return true;
 751
 752         if (conf->mddev->reshape_position != MaxSector)
 753                 degraded = raid5_calc_degraded(conf);
 754
 755         return degraded > conf->max_degraded;
 756 }
 757
 758 enum stripe_result {
 759         STRIPE_SUCCESS = 0,
 760         STRIPE_RETRY,
 761         STRIPE_SCHEDULE_AND_RETRY,
 762         STRIPE_FAIL,
 763         STRIPE_WAIT_RESHAPE,
 764 };
 765
 766 struct stripe_request_ctx {
 767         /* a reference to the last stripe_head for batching */
 768         struct stripe_head *batch_last;
 769
 770         /* first sector in the request */
 771         sector_t first_sector;
 772
 773         /* last sector in the request */
 774         sector_t last_sector;
 775
 776         /*
 777          * bitmap to track stripe sectors that have been added to stripes
 778          * add one to account for unaligned requests
 779          */
 780         DECLARE_BITMAP(sectors_to_do, RAID5_MAX_REQ_STRIPES + 1);
 781
 782         /* the request had REQ_PREFLUSH, cleared after the first stripe_head */
 783         bool do_flush;
 784 };
 785
 786 /*
 787  * Block until another thread clears R5_INACTIVE_BLOCKED or
 788  * there are fewer than 3/4 the maximum number of active stripes
 789  * and there is an inactive stripe available.
 790  */
 791 static bool is_inactive_blocked(struct r5conf *conf, int hash)
 792 {
 793         if (list_empty(conf->inactive_list + hash))
 794                 return false;
 795
 796         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
 797                 return true;
 798
 799         return (atomic_read(&conf->active_stripes) <
 800                 (conf->max_nr_stripes * 3 / 4));
 801 }
 802
 803 struct stripe_head *raid5_get_active_stripe(struct r5conf *conf,
 804                 struct stripe_request_ctx *ctx, sector_t sector,
 805                 unsigned int flags)
 806 {
 807         struct stripe_head *sh;
 808         int hash = stripe_hash_locks_hash(conf, sector);
 809         int previous = !!(flags & R5_GAS_PREVIOUS);
 810
 811         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 812
 813         spin_lock_irq(conf->hash_locks + hash);
 814
 815         for (;;) {
 816                 if (!(flags & R5_GAS_NOQUIESCE) && conf->quiesce) {
 817                         /*
 818                          * Must release the reference to batch_last before
 819                          * waiting, on quiesce, otherwise the batch_last will
 820                          * hold a reference to a stripe and raid5_quiesce()
 821                          * will deadlock waiting for active_stripes to go to
 822                          * zero.
 823                          */
 824                         if (ctx && ctx->batch_last) {
 825                                 raid5_release_stripe(ctx->batch_last);
 826                                 ctx->batch_last = NULL;
 827                         }
 828
 829                         wait_event_lock_irq(conf->wait_for_quiescent,
 830                                             !conf->quiesce,
 831                                             *(conf->hash_locks + hash));
 832                 }
 833
 834                 sh = find_get_stripe(conf, sector, conf->generation - previous,
 835                                      hash);
 836                 if (sh)
 837                         break;
 838
 839                 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 840                         sh = get_free_stripe(conf, hash);
 841                         if (sh) {
 842                                 r5c_check_stripe_cache_usage(conf);
 843                                 init_stripe(sh, sector, previous);
 844                                 atomic_inc(&sh->count);
 845                                 break;
 846                         }
 847
 848                         if (!test_bit(R5_DID_ALLOC, &conf->cache_state))
 849                                 set_bit(R5_ALLOC_MORE, &conf->cache_state);
 850                 }
 851
 852                 if (flags & R5_GAS_NOBLOCK)
 853                         break;
 854
 855                 set_bit(R5_INACTIVE_BLOCKED, &conf->cache_state);
 856                 r5l_wake_reclaim(conf->log, 0);
 857
 858                 /* release batch_last before wait to avoid risk of deadlock */
 859                 if (ctx && ctx->batch_last) {
 860                         raid5_release_stripe(ctx->batch_last);
 861                         ctx->batch_last = NULL;
 862                 }
 863
 864                 wait_event_lock_irq(conf->wait_for_stripe,
 865                                     is_inactive_blocked(conf, hash),
 866                                     *(conf->hash_locks + hash));
 867                 clear_bit(R5_INACTIVE_BLOCKED, &conf->cache_state);
 868         }
 869
 870         spin_unlock_irq(conf->hash_locks + hash);
 871         return sh;
 872 }
 873
 874 static bool is_full_stripe_write(struct stripe_head *sh)
 875 {
 876         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 877         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 878 }
 879
 880 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 881                 __acquires(&sh1->stripe_lock)
 882                 __acquires(&sh2->stripe_lock)
 883 {
 884         if (sh1 > sh2) {
 885                 spin_lock_irq(&sh2->stripe_lock);
 886                 spin_lock_nested(&sh1->stripe_lock, 1);
 887         } else {
 888                 spin_lock_irq(&sh1->stripe_lock);
 889                 spin_lock_nested(&sh2->stripe_lock, 1);
 890         }
 891 }
 892
 893 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 894                 __releases(&sh1->stripe_lock)
 895                 __releases(&sh2->stripe_lock)
 896 {
 897         spin_unlock(&sh1->stripe_lock);
 898         spin_unlock_irq(&sh2->stripe_lock);
 899 }
 900
 901 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 902 static bool stripe_can_batch(struct stripe_head *sh)
 903 {
 904         struct r5conf *conf = sh->raid_conf;
 905
 906         if (raid5_has_log(conf) || raid5_has_ppl(conf))
 907                 return false;
 908         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 909                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 910                 is_full_stripe_write(sh);
 911 }
 912
 913 /* we only do back search */
 914 static void stripe_add_to_batch_list(struct r5conf *conf,
 915                 struct stripe_head *sh, struct stripe_head *last_sh)
 916 {
 917         struct stripe_head *head;
 918         sector_t head_sector, tmp_sec;
 919         int hash;
 920         int dd_idx;
 921
 922         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 923         tmp_sec = sh->sector;
 924         if (!sector_div(tmp_sec, conf->chunk_sectors))
 925                 return;
 926         head_sector = sh->sector - RAID5_STRIPE_SECTORS(conf);
 927
 928         if (last_sh && head_sector == last_sh->sector) {
 929                 head = last_sh;
 930                 atomic_inc(&head->count);
 931         } else {
 932                 hash = stripe_hash_locks_hash(conf, head_sector);
 933                 spin_lock_irq(conf->hash_locks + hash);
 934                 head = find_get_stripe(conf, head_sector, conf->generation,
 935                                        hash);
 936                 spin_unlock_irq(conf->hash_locks + hash);
 937                 if (!head)
 938                         return;
 939                 if (!stripe_can_batch(head))
 940                         goto out;
 941         }
 942
 943         lock_two_stripes(head, sh);
 944         /* clear_batch_ready clear the flag */
 945         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 946                 goto unlock_out;
 947
 948         if (sh->batch_head)
 949                 goto unlock_out;
 950
 951         dd_idx = 0;
 952         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 953                 dd_idx++;
 954         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 955             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 956                 goto unlock_out;
 957
 958         if (head->batch_head) {
 959                 spin_lock(&head->batch_head->batch_lock);
 960                 /* This batch list is already running */
 961                 if (!stripe_can_batch(head)) {
 962                         spin_unlock(&head->batch_head->batch_lock);
 963                         goto unlock_out;
 964                 }
 965                 /*
 966                  * We must assign batch_head of this stripe within the
 967                  * batch_lock, otherwise clear_batch_ready of batch head
 968                  * stripe could clear BATCH_READY bit of this stripe and
 969                  * this stripe->batch_head doesn't get assigned, which
 970                  * could confuse clear_batch_ready for this stripe
 971                  */
 972                 sh->batch_head = head->batch_head;
 973
 974                 /*
 975                  * at this point, head's BATCH_READY could be cleared, but we
 976                  * can still add the stripe to batch list
 977                  */
 978                 list_add(&sh->batch_list, &head->batch_list);
 979                 spin_unlock(&head->batch_head->batch_lock);
 980         } else {
 981                 head->batch_head = head;
 982                 sh->batch_head = head->batch_head;
 983                 spin_lock(&head->batch_lock);
 984                 list_add_tail(&sh->batch_list, &head->batch_list);
 985                 spin_unlock(&head->batch_lock);
 986         }
 987
 988         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 989                 if (atomic_dec_return(&conf->preread_active_stripes)
 990                     < IO_THRESHOLD)
 991                         md_wakeup_thread(conf->mddev->thread);
 992
 993         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 994                 int seq = sh->bm_seq;
 995                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 996                     sh->batch_head->bm_seq > seq)
 997                         seq = sh->batch_head->bm_seq;
 998                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 999                 sh->batch_head->bm_seq = seq;
1000         }
1001
1002         atomic_inc(&sh->count);
1003 unlock_out:
1004         unlock_two_stripes(head, sh);
1005 out:
1006         raid5_release_stripe(head);
1007 }
1008
1009 /* Determine if 'data_offset' or 'new_data_offset' should be used
1010  * in this stripe_head.
1011  */
1012 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
1013 {
1014         sector_t progress = conf->reshape_progress;
1015         /* Need a memory barrier to make sure we see the value
1016          * of conf->generation, or ->data_offset that was set before
1017          * reshape_progress was updated.
1018          */
1019         smp_rmb();
1020         if (progress == MaxSector)
1021                 return 0;
1022         if (sh->generation == conf->generation - 1)
1023                 return 0;
1024         /* We are in a reshape, and this is a new-generation stripe,
1025          * so use new_data_offset.
1026          */
1027         return 1;
1028 }
1029
1030 static void dispatch_bio_list(struct bio_list *tmp)
1031 {
1032         struct bio *bio;
1033
1034         while ((bio = bio_list_pop(tmp)))
1035                 submit_bio_noacct(bio);
1036 }
1037
1038 static int cmp_stripe(void *priv, const struct list_head *a,
1039                       const struct list_head *b)
1040 {
1041         const struct r5pending_data *da = list_entry(a,
1042                                 struct r5pending_data, sibling);
1043         const struct r5pending_data *db = list_entry(b,
1044                                 struct r5pending_data, sibling);
1045         if (da->sector > db->sector)
1046                 return 1;
1047         if (da->sector < db->sector)
1048                 return -1;
1049         return 0;
1050 }
1051
1052 static void dispatch_defer_bios(struct r5conf *conf, int target,
1053                                 struct bio_list *list)
1054 {
1055         struct r5pending_data *data;
1056         struct list_head *first, *next = NULL;
1057         int cnt = 0;
1058
1059         if (conf->pending_data_cnt == 0)
1060                 return;
1061
1062         list_sort(NULL, &conf->pending_list, cmp_stripe);
1063
1064         first = conf->pending_list.next;
1065
1066         /* temporarily move the head */
1067         if (conf->next_pending_data)
1068                 list_move_tail(&conf->pending_list,
1069                                 &conf->next_pending_data->sibling);
1070
1071         while (!list_empty(&conf->pending_list)) {
1072                 data = list_first_entry(&conf->pending_list,
1073                         struct r5pending_data, sibling);
1074                 if (&data->sibling == first)
1075                         first = data->sibling.next;
1076                 next = data->sibling.next;
1077
1078                 bio_list_merge(list, &data->bios);
1079                 list_move(&data->sibling, &conf->free_list);
1080                 cnt++;
1081                 if (cnt >= target)
1082                         break;
1083         }
1084         conf->pending_data_cnt -= cnt;
1085         BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
1086
1087         if (next != &conf->pending_list)
1088                 conf->next_pending_data = list_entry(next,
1089                                 struct r5pending_data, sibling);
1090         else
1091                 conf->next_pending_data = NULL;
1092         /* list isn't empty */
1093         if (first != &conf->pending_list)
1094                 list_move_tail(&conf->pending_list, first);
1095 }
1096
1097 static void flush_deferred_bios(struct r5conf *conf)
1098 {
1099         struct bio_list tmp = BIO_EMPTY_LIST;
1100
1101         if (conf->pending_data_cnt == 0)
1102                 return;
1103
1104         spin_lock(&conf->pending_bios_lock);
1105         dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
1106         BUG_ON(conf->pending_data_cnt != 0);
1107         spin_unlock(&conf->pending_bios_lock);
1108
1109         dispatch_bio_list(&tmp);
1110 }
1111
1112 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
1113                                 struct bio_list *bios)
1114 {
1115         struct bio_list tmp = BIO_EMPTY_LIST;
1116         struct r5pending_data *ent;
1117
1118         spin_lock(&conf->pending_bios_lock);
1119         ent = list_first_entry(&conf->free_list, struct r5pending_data,
1120                                                         sibling);
1121         list_move_tail(&ent->sibling, &conf->pending_list);
1122         ent->sector = sector;
1123         bio_list_init(&ent->bios);
1124         bio_list_merge(&ent->bios, bios);
1125         conf->pending_data_cnt++;
1126         if (conf->pending_data_cnt >= PENDING_IO_MAX)
1127                 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
1128
1129         spin_unlock(&conf->pending_bios_lock);
1130
1131         dispatch_bio_list(&tmp);
1132 }
1133
1134 static void
1135 raid5_end_read_request(struct bio *bi);
1136 static void
1137 raid5_end_write_request(struct bio *bi);
1138
1139 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
1140 {
1141         struct r5conf *conf = sh->raid_conf;
1142         int i, disks = sh->disks;
1143         struct stripe_head *head_sh = sh;
1144         struct bio_list pending_bios = BIO_EMPTY_LIST;
1145         struct r5dev *dev;
1146         bool should_defer;
1147
1148         might_sleep();
1149
1150         if (log_stripe(sh, s) == 0)
1151                 return;
1152
1153         should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1154
1155         for (i = disks; i--; ) {
1156                 enum req_op op;
1157                 blk_opf_t op_flags = 0;
1158                 int replace_only = 0;
1159                 struct bio *bi, *rbi;
1160                 struct md_rdev *rdev, *rrdev = NULL;
1161
1162                 sh = head_sh;
1163                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1164                         op = REQ_OP_WRITE;
1165                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1166                                 op_flags = REQ_FUA;
1167                         if (test_bit(R5_Discard, &sh->dev[i].flags))
1168                                 op = REQ_OP_DISCARD;
1169                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1170                         op = REQ_OP_READ;
1171                 else if (test_and_clear_bit(R5_WantReplace,
1172                                             &sh->dev[i].flags)) {
1173                         op = REQ_OP_WRITE;
1174                         replace_only = 1;
1175                 } else
1176                         continue;
1177                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1178                         op_flags |= REQ_SYNC;
1179
1180 again:
1181                 dev = &sh->dev[i];
1182                 bi = &dev->req;
1183                 rbi = &dev->rreq; /* For writing to replacement */
1184
1185                 rdev = conf->disks[i].rdev;
1186                 rrdev = conf->disks[i].replacement;
1187                 if (op_is_write(op)) {
1188                         if (replace_only)
1189                                 rdev = NULL;
1190                         if (rdev == rrdev)
1191                                 /* We raced and saw duplicates */
1192                                 rrdev = NULL;
1193                 } else {
1194                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1195                                 rdev = rrdev;
1196                         rrdev = NULL;
1197                 }
1198
1199                 if (rdev && test_bit(Faulty, &rdev->flags))
1200                         rdev = NULL;
1201                 if (rdev)
1202                         atomic_inc(&rdev->nr_pending);
1203                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1204                         rrdev = NULL;
1205                 if (rrdev)
1206                         atomic_inc(&rrdev->nr_pending);
1207
1208                 /* We have already checked bad blocks for reads.  Now
1209                  * need to check for writes.  We never accept write errors
1210                  * on the replacement, so we don't to check rrdev.
1211                  */
1212                 while (op_is_write(op) && rdev &&
1213                        test_bit(WriteErrorSeen, &rdev->flags)) {
1214                         int bad = rdev_has_badblock(rdev, sh->sector,
1215                                                     RAID5_STRIPE_SECTORS(conf));
1216                         if (!bad)
1217                                 break;
1218
1219                         if (bad < 0) {
1220                                 set_bit(BlockedBadBlocks, &rdev->flags);
1221                                 if (!conf->mddev->external &&
1222                                     conf->mddev->sb_flags) {
1223                                         /* It is very unlikely, but we might
1224                                          * still need to write out the
1225                                          * bad block log - better give it
1226                                          * a chance*/
1227                                         md_check_recovery(conf->mddev);
1228                                 }
1229                                 /*
1230                                  * Because md_wait_for_blocked_rdev
1231                                  * will dec nr_pending, we must
1232                                  * increment it first.
1233                                  */
1234                                 atomic_inc(&rdev->nr_pending);
1235                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
1236                         } else {
1237                                 /* Acknowledged bad block - skip the write */
1238                                 rdev_dec_pending(rdev, conf->mddev);
1239                                 rdev = NULL;
1240                         }
1241                 }
1242
1243                 if (rdev) {
1244                         if (s->syncing || s->expanding || s->expanded
1245                             || s->replacing)
1246                                 md_sync_acct(rdev->bdev, RAID5_STRIPE_SECTORS(conf));
1247
1248                         set_bit(STRIPE_IO_STARTED, &sh->state);
1249
1250                         bio_init(bi, rdev->bdev, &dev->vec, 1, op | op_flags);
1251                         bi->bi_end_io = op_is_write(op)
1252                                 ? raid5_end_write_request
1253                                 : raid5_end_read_request;
1254                         bi->bi_private = sh;
1255
1256                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1257                                 __func__, (unsigned long long)sh->sector,
1258                                 bi->bi_opf, i);
1259                         atomic_inc(&sh->count);
1260                         if (sh != head_sh)
1261                                 atomic_inc(&head_sh->count);
1262                         if (use_new_offset(conf, sh))
1263                                 bi->bi_iter.bi_sector = (sh->sector
1264                                                  + rdev->new_data_offset);
1265                         else
1266                                 bi->bi_iter.bi_sector = (sh->sector
1267                                                  + rdev->data_offset);
1268                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1269                                 bi->bi_opf |= REQ_NOMERGE;
1270
1271                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1272                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1273
1274                         if (!op_is_write(op) &&
1275                             test_bit(R5_InJournal, &sh->dev[i].flags))
1276                                 /*
1277                                  * issuing read for a page in journal, this
1278                                  * must be preparing for prexor in rmw; read
1279                                  * the data into orig_page
1280                                  */
1281                                 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1282                         else
1283                                 sh->dev[i].vec.bv_page = sh->dev[i].page;
1284                         bi->bi_vcnt = 1;
1285                         bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1286                         bi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1287                         bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1288                         /*
1289                          * If this is discard request, set bi_vcnt 0. We don't
1290                          * want to confuse SCSI because SCSI will replace payload
1291                          */
1292                         if (op == REQ_OP_DISCARD)
1293                                 bi->bi_vcnt = 0;
1294                         if (rrdev)
1295                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1296
1297                         mddev_trace_remap(conf->mddev, bi, sh->dev[i].sector);
1298                         if (should_defer && op_is_write(op))
1299                                 bio_list_add(&pending_bios, bi);
1300                         else
1301                                 submit_bio_noacct(bi);
1302                 }
1303                 if (rrdev) {
1304                         if (s->syncing || s->expanding || s->expanded
1305                             || s->replacing)
1306                                 md_sync_acct(rrdev->bdev, RAID5_STRIPE_SECTORS(conf));
1307
1308                         set_bit(STRIPE_IO_STARTED, &sh->state);
1309
1310                         bio_init(rbi, rrdev->bdev, &dev->rvec, 1, op | op_flags);
1311                         BUG_ON(!op_is_write(op));
1312                         rbi->bi_end_io = raid5_end_write_request;
1313                         rbi->bi_private = sh;
1314
1315                         pr_debug("%s: for %llu schedule op %d on "
1316                                  "replacement disc %d\n",
1317                                 __func__, (unsigned long long)sh->sector,
1318                                 rbi->bi_opf, i);
1319                         atomic_inc(&sh->count);
1320                         if (sh != head_sh)
1321                                 atomic_inc(&head_sh->count);
1322                         if (use_new_offset(conf, sh))
1323                                 rbi->bi_iter.bi_sector = (sh->sector
1324                                                   + rrdev->new_data_offset);
1325                         else
1326                                 rbi->bi_iter.bi_sector = (sh->sector
1327                                                   + rrdev->data_offset);
1328                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1329                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1330                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1331                         rbi->bi_vcnt = 1;
1332                         rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1333                         rbi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1334                         rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1335                         /*
1336                          * If this is discard request, set bi_vcnt 0. We don't
1337                          * want to confuse SCSI because SCSI will replace payload
1338                          */
1339                         if (op == REQ_OP_DISCARD)
1340                                 rbi->bi_vcnt = 0;
1341                         mddev_trace_remap(conf->mddev, rbi, sh->dev[i].sector);
1342                         if (should_defer && op_is_write(op))
1343                                 bio_list_add(&pending_bios, rbi);
1344                         else
1345                                 submit_bio_noacct(rbi);
1346                 }
1347                 if (!rdev && !rrdev) {
1348                         if (op_is_write(op))
1349                                 set_bit(STRIPE_DEGRADED, &sh->state);
1350                         pr_debug("skip op %d on disc %d for sector %llu\n",
1351                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1352                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1353                         set_bit(STRIPE_HANDLE, &sh->state);
1354                 }
1355
1356                 if (!head_sh->batch_head)
1357                         continue;
1358                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1359                                       batch_list);
1360                 if (sh != head_sh)
1361                         goto again;
1362         }
1363
1364         if (should_defer && !bio_list_empty(&pending_bios))
1365                 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1366 }
1367
1368 static struct dma_async_tx_descriptor *
1369 async_copy_data(int frombio, struct bio *bio, struct page **page,
1370         unsigned int poff, sector_t sector, struct dma_async_tx_descriptor *tx,
1371         struct stripe_head *sh, int no_skipcopy)
1372 {
1373         struct bio_vec bvl;
1374         struct bvec_iter iter;
1375         struct page *bio_page;
1376         int page_offset;
1377         struct async_submit_ctl submit;
1378         enum async_tx_flags flags = 0;
1379         struct r5conf *conf = sh->raid_conf;
1380
1381         if (bio->bi_iter.bi_sector >= sector)
1382                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1383         else
1384                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1385
1386         if (frombio)
1387                 flags |= ASYNC_TX_FENCE;
1388         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1389
1390         bio_for_each_segment(bvl, bio, iter) {
1391                 int len = bvl.bv_len;
1392                 int clen;
1393                 int b_offset = 0;
1394
1395                 if (page_offset < 0) {
1396                         b_offset = -page_offset;
1397                         page_offset += b_offset;
1398                         len -= b_offset;
1399                 }
1400
1401                 if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
1402                         clen = RAID5_STRIPE_SIZE(conf) - page_offset;
1403                 else
1404                         clen = len;
1405
1406                 if (clen > 0) {
1407                         b_offset += bvl.bv_offset;
1408                         bio_page = bvl.bv_page;
1409                         if (frombio) {
1410                                 if (conf->skip_copy &&
1411                                     b_offset == 0 && page_offset == 0 &&
1412                                     clen == RAID5_STRIPE_SIZE(conf) &&
1413                                     !no_skipcopy)
1414                                         *page = bio_page;
1415                                 else
1416                                         tx = async_memcpy(*page, bio_page, page_offset + poff,
1417                                                   b_offset, clen, &submit);
1418                         } else
1419                                 tx = async_memcpy(bio_page, *page, b_offset,
1420                                                   page_offset + poff, clen, &submit);
1421                 }
1422                 /* chain the operations */
1423                 submit.depend_tx = tx;
1424
1425                 if (clen < len) /* hit end of page */
1426                         break;
1427                 page_offset +=  len;
1428         }
1429
1430         return tx;
1431 }
1432
1433 static void ops_complete_biofill(void *stripe_head_ref)
1434 {
1435         struct stripe_head *sh = stripe_head_ref;
1436         int i;
1437         struct r5conf *conf = sh->raid_conf;
1438
1439         pr_debug("%s: stripe %llu\n", __func__,
1440                 (unsigned long long)sh->sector);
1441
1442         /* clear completed biofills */
1443         for (i = sh->disks; i--; ) {
1444                 struct r5dev *dev = &sh->dev[i];
1445
1446                 /* acknowledge completion of a biofill operation */
1447                 /* and check if we need to reply to a read request,
1448                  * new R5_Wantfill requests are held off until
1449                  * !STRIPE_BIOFILL_RUN
1450                  */
1451                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1452                         struct bio *rbi, *rbi2;
1453
1454                         BUG_ON(!dev->read);
1455                         rbi = dev->read;
1456                         dev->read = NULL;
1457                         while (rbi && rbi->bi_iter.bi_sector <
1458                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1459                                 rbi2 = r5_next_bio(conf, rbi, dev->sector);
1460                                 bio_endio(rbi);
1461                                 rbi = rbi2;
1462                         }
1463                 }
1464         }
1465         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1466
1467         set_bit(STRIPE_HANDLE, &sh->state);
1468         raid5_release_stripe(sh);
1469 }
1470
1471 static void ops_run_biofill(struct stripe_head *sh)
1472 {
1473         struct dma_async_tx_descriptor *tx = NULL;
1474         struct async_submit_ctl submit;
1475         int i;
1476         struct r5conf *conf = sh->raid_conf;
1477
1478         BUG_ON(sh->batch_head);
1479         pr_debug("%s: stripe %llu\n", __func__,
1480                 (unsigned long long)sh->sector);
1481
1482         for (i = sh->disks; i--; ) {
1483                 struct r5dev *dev = &sh->dev[i];
1484                 if (test_bit(R5_Wantfill, &dev->flags)) {
1485                         struct bio *rbi;
1486                         spin_lock_irq(&sh->stripe_lock);
1487                         dev->read = rbi = dev->toread;
1488                         dev->toread = NULL;
1489                         spin_unlock_irq(&sh->stripe_lock);
1490                         while (rbi && rbi->bi_iter.bi_sector <
1491                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1492                                 tx = async_copy_data(0, rbi, &dev->page,
1493                                                      dev->offset,
1494                                                      dev->sector, tx, sh, 0);
1495                                 rbi = r5_next_bio(conf, rbi, dev->sector);
1496                         }
1497                 }
1498         }
1499
1500         atomic_inc(&sh->count);
1501         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1502         async_trigger_callback(&submit);
1503 }
1504
1505 static void mark_target_uptodate(struct stripe_head *sh, int target)
1506 {
1507         struct r5dev *tgt;
1508
1509         if (target < 0)
1510                 return;
1511
1512         tgt = &sh->dev[target];
1513         set_bit(R5_UPTODATE, &tgt->flags);
1514         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1515         clear_bit(R5_Wantcompute, &tgt->flags);
1516 }
1517
1518 static void ops_complete_compute(void *stripe_head_ref)
1519 {
1520         struct stripe_head *sh = stripe_head_ref;
1521
1522         pr_debug("%s: stripe %llu\n", __func__,
1523                 (unsigned long long)sh->sector);
1524
1525         /* mark the computed target(s) as uptodate */
1526         mark_target_uptodate(sh, sh->ops.target);
1527         mark_target_uptodate(sh, sh->ops.target2);
1528
1529         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1530         if (sh->check_state == check_state_compute_run)
1531                 sh->check_state = check_state_compute_result;
1532         set_bit(STRIPE_HANDLE, &sh->state);
1533         raid5_release_stripe(sh);
1534 }
1535
1536 /* return a pointer to the address conversion region of the scribble buffer */
1537 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1538 {
1539         return percpu->scribble + i * percpu->scribble_obj_size;
1540 }
1541
1542 /* return a pointer to the address conversion region of the scribble buffer */
1543 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1544                                  struct raid5_percpu *percpu, int i)
1545 {
1546         return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1547 }
1548
1549 /*
1550  * Return a pointer to record offset address.
1551  */
1552 static unsigned int *
1553 to_addr_offs(struct stripe_head *sh, struct raid5_percpu *percpu)
1554 {
1555         return (unsigned int *) (to_addr_conv(sh, percpu, 0) + sh->disks + 2);
1556 }
1557
1558 static struct dma_async_tx_descriptor *
1559 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1560 {
1561         int disks = sh->disks;
1562         struct page **xor_srcs = to_addr_page(percpu, 0);
1563         unsigned int *off_srcs = to_addr_offs(sh, percpu);
1564         int target = sh->ops.target;
1565         struct r5dev *tgt = &sh->dev[target];
1566         struct page *xor_dest = tgt->page;
1567         unsigned int off_dest = tgt->offset;
1568         int count = 0;
1569         struct dma_async_tx_descriptor *tx;
1570         struct async_submit_ctl submit;
1571         int i;
1572
1573         BUG_ON(sh->batch_head);
1574
1575         pr_debug("%s: stripe %llu block: %d\n",
1576                 __func__, (unsigned long long)sh->sector, target);
1577         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1578
1579         for (i = disks; i--; ) {
1580                 if (i != target) {
1581                         off_srcs[count] = sh->dev[i].offset;
1582                         xor_srcs[count++] = sh->dev[i].page;
1583                 }
1584         }
1585
1586         atomic_inc(&sh->count);
1587
1588         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1589                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1590         if (unlikely(count == 1))
1591                 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
1592                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1593         else
1594                 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1595                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1596
1597         return tx;
1598 }
1599
1600 /* set_syndrome_sources - populate source buffers for gen_syndrome
1601  * @srcs - (struct page *) array of size sh->disks
1602  * @offs - (unsigned int) array of offset for each page
1603  * @sh - stripe_head to parse
1604  *
1605  * Populates srcs in proper layout order for the stripe and returns the
1606  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1607  * destination buffer is recorded in srcs[count] and the Q destination
1608  * is recorded in srcs[count+1]].
1609  */
1610 static int set_syndrome_sources(struct page **srcs,
1611                                 unsigned int *offs,
1612                                 struct stripe_head *sh,
1613                                 int srctype)
1614 {
1615         int disks = sh->disks;
1616         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1617         int d0_idx = raid6_d0(sh);
1618         int count;
1619         int i;
1620
1621         for (i = 0; i < disks; i++)
1622                 srcs[i] = NULL;
1623
1624         count = 0;
1625         i = d0_idx;
1626         do {
1627                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1628                 struct r5dev *dev = &sh->dev[i];
1629
1630                 if (i == sh->qd_idx || i == sh->pd_idx ||
1631                     (srctype == SYNDROME_SRC_ALL) ||
1632                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1633                      (test_bit(R5_Wantdrain, &dev->flags) ||
1634                       test_bit(R5_InJournal, &dev->flags))) ||
1635                     (srctype == SYNDROME_SRC_WRITTEN &&
1636                      (dev->written ||
1637                       test_bit(R5_InJournal, &dev->flags)))) {
1638                         if (test_bit(R5_InJournal, &dev->flags))
1639                                 srcs[slot] = sh->dev[i].orig_page;
1640                         else
1641                                 srcs[slot] = sh->dev[i].page;
1642                         /*
1643                          * For R5_InJournal, PAGE_SIZE must be 4KB and will
1644                          * not shared page. In that case, dev[i].offset
1645                          * is 0.
1646                          */
1647                         offs[slot] = sh->dev[i].offset;
1648                 }
1649                 i = raid6_next_disk(i, disks);
1650         } while (i != d0_idx);
1651
1652         return syndrome_disks;
1653 }
1654
1655 static struct dma_async_tx_descriptor *
1656 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1657 {
1658         int disks = sh->disks;
1659         struct page **blocks = to_addr_page(percpu, 0);
1660         unsigned int *offs = to_addr_offs(sh, percpu);
1661         int target;
1662         int qd_idx = sh->qd_idx;
1663         struct dma_async_tx_descriptor *tx;
1664         struct async_submit_ctl submit;
1665         struct r5dev *tgt;
1666         struct page *dest;
1667         unsigned int dest_off;
1668         int i;
1669         int count;
1670
1671         BUG_ON(sh->batch_head);
1672         if (sh->ops.target < 0)
1673                 target = sh->ops.target2;
1674         else if (sh->ops.target2 < 0)
1675                 target = sh->ops.target;
1676         else
1677                 /* we should only have one valid target */
1678                 BUG();
1679         BUG_ON(target < 0);
1680         pr_debug("%s: stripe %llu block: %d\n",
1681                 __func__, (unsigned long long)sh->sector, target);
1682
1683         tgt = &sh->dev[target];
1684         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1685         dest = tgt->page;
1686         dest_off = tgt->offset;
1687
1688         atomic_inc(&sh->count);
1689
1690         if (target == qd_idx) {
1691                 count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1692                 blocks[count] = NULL; /* regenerating p is not necessary */
1693                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1694                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1695                                   ops_complete_compute, sh,
1696                                   to_addr_conv(sh, percpu, 0));
1697                 tx = async_gen_syndrome(blocks, offs, count+2,
1698                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1699         } else {
1700                 /* Compute any data- or p-drive using XOR */
1701                 count = 0;
1702                 for (i = disks; i-- ; ) {
1703                         if (i == target || i == qd_idx)
1704                                 continue;
1705                         offs[count] = sh->dev[i].offset;
1706                         blocks[count++] = sh->dev[i].page;
1707                 }
1708
1709                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1710                                   NULL, ops_complete_compute, sh,
1711                                   to_addr_conv(sh, percpu, 0));
1712                 tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1713                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1714         }
1715
1716         return tx;
1717 }
1718
1719 static struct dma_async_tx_descriptor *
1720 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1721 {
1722         int i, count, disks = sh->disks;
1723         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1724         int d0_idx = raid6_d0(sh);
1725         int faila = -1, failb = -1;
1726         int target = sh->ops.target;
1727         int target2 = sh->ops.target2;
1728         struct r5dev *tgt = &sh->dev[target];
1729         struct r5dev *tgt2 = &sh->dev[target2];
1730         struct dma_async_tx_descriptor *tx;
1731         struct page **blocks = to_addr_page(percpu, 0);
1732         unsigned int *offs = to_addr_offs(sh, percpu);
1733         struct async_submit_ctl submit;
1734
1735         BUG_ON(sh->batch_head);
1736         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1737                  __func__, (unsigned long long)sh->sector, target, target2);
1738         BUG_ON(target < 0 || target2 < 0);
1739         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1740         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1741
1742         /* we need to open-code set_syndrome_sources to handle the
1743          * slot number conversion for 'faila' and 'failb'
1744          */
1745         for (i = 0; i < disks ; i++) {
1746                 offs[i] = 0;
1747                 blocks[i] = NULL;
1748         }
1749         count = 0;
1750         i = d0_idx;
1751         do {
1752                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1753
1754                 offs[slot] = sh->dev[i].offset;
1755                 blocks[slot] = sh->dev[i].page;
1756
1757                 if (i == target)
1758                         faila = slot;
1759                 if (i == target2)
1760                         failb = slot;
1761                 i = raid6_next_disk(i, disks);
1762         } while (i != d0_idx);
1763
1764         BUG_ON(faila == failb);
1765         if (failb < faila)
1766                 swap(faila, failb);
1767         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1768                  __func__, (unsigned long long)sh->sector, faila, failb);
1769
1770         atomic_inc(&sh->count);
1771
1772         if (failb == syndrome_disks+1) {
1773                 /* Q disk is one of the missing disks */
1774                 if (faila == syndrome_disks) {
1775                         /* Missing P+Q, just recompute */
1776                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1777                                           ops_complete_compute, sh,
1778                                           to_addr_conv(sh, percpu, 0));
1779                         return async_gen_syndrome(blocks, offs, syndrome_disks+2,
1780                                                   RAID5_STRIPE_SIZE(sh->raid_conf),
1781                                                   &submit);
1782                 } else {
1783                         struct page *dest;
1784                         unsigned int dest_off;
1785                         int data_target;
1786                         int qd_idx = sh->qd_idx;
1787
1788                         /* Missing D+Q: recompute D from P, then recompute Q */
1789                         if (target == qd_idx)
1790                                 data_target = target2;
1791                         else
1792                                 data_target = target;
1793
1794                         count = 0;
1795                         for (i = disks; i-- ; ) {
1796                                 if (i == data_target || i == qd_idx)
1797                                         continue;
1798                                 offs[count] = sh->dev[i].offset;
1799                                 blocks[count++] = sh->dev[i].page;
1800                         }
1801                         dest = sh->dev[data_target].page;
1802                         dest_off = sh->dev[data_target].offset;
1803                         init_async_submit(&submit,
1804                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1805                                           NULL, NULL, NULL,
1806                                           to_addr_conv(sh, percpu, 0));
1807                         tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1808                                        RAID5_STRIPE_SIZE(sh->raid_conf),
1809                                        &submit);
1810
1811                         count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1812                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1813                                           ops_complete_compute, sh,
1814                                           to_addr_conv(sh, percpu, 0));
1815                         return async_gen_syndrome(blocks, offs, count+2,
1816                                                   RAID5_STRIPE_SIZE(sh->raid_conf),
1817                                                   &submit);
1818                 }
1819         } else {
1820                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1821                                   ops_complete_compute, sh,
1822                                   to_addr_conv(sh, percpu, 0));
1823                 if (failb == syndrome_disks) {
1824                         /* We're missing D+P. */
1825                         return async_raid6_datap_recov(syndrome_disks+2,
1826                                                 RAID5_STRIPE_SIZE(sh->raid_conf),
1827                                                 faila,
1828                                                 blocks, offs, &submit);
1829                 } else {
1830                         /* We're missing D+D. */
1831                         return async_raid6_2data_recov(syndrome_disks+2,
1832                                                 RAID5_STRIPE_SIZE(sh->raid_conf),
1833                                                 faila, failb,
1834                                                 blocks, offs, &submit);
1835                 }
1836         }
1837 }
1838
1839 static void ops_complete_prexor(void *stripe_head_ref)
1840 {
1841         struct stripe_head *sh = stripe_head_ref;
1842
1843         pr_debug("%s: stripe %llu\n", __func__,
1844                 (unsigned long long)sh->sector);
1845
1846         if (r5c_is_writeback(sh->raid_conf->log))
1847                 /*
1848                  * raid5-cache write back uses orig_page during prexor.
1849                  * After prexor, it is time to free orig_page
1850                  */
1851                 r5c_release_extra_page(sh);
1852 }
1853
1854 static struct dma_async_tx_descriptor *
1855 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1856                 struct dma_async_tx_descriptor *tx)
1857 {
1858         int disks = sh->disks;
1859         struct page **xor_srcs = to_addr_page(percpu, 0);
1860         unsigned int *off_srcs = to_addr_offs(sh, percpu);
1861         int count = 0, pd_idx = sh->pd_idx, i;
1862         struct async_submit_ctl submit;
1863
1864         /* existing parity data subtracted */
1865         unsigned int off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
1866         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1867
1868         BUG_ON(sh->batch_head);
1869         pr_debug("%s: stripe %llu\n", __func__,
1870                 (unsigned long long)sh->sector);
1871
1872         for (i = disks; i--; ) {
1873                 struct r5dev *dev = &sh->dev[i];
1874                 /* Only process blocks that are known to be uptodate */
1875                 if (test_bit(R5_InJournal, &dev->flags)) {
1876                         /*
1877                          * For this case, PAGE_SIZE must be equal to 4KB and
1878                          * page offset is zero.
1879                          */
1880                         off_srcs[count] = dev->offset;
1881                         xor_srcs[count++] = dev->orig_page;
1882                 } else if (test_bit(R5_Wantdrain, &dev->flags)) {
1883                         off_srcs[count] = dev->offset;
1884                         xor_srcs[count++] = dev->page;
1885                 }
1886         }
1887
1888         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1889                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1890         tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1891                         RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1892
1893         return tx;
1894 }
1895
1896 static struct dma_async_tx_descriptor *
1897 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1898                 struct dma_async_tx_descriptor *tx)
1899 {
1900         struct page **blocks = to_addr_page(percpu, 0);
1901         unsigned int *offs = to_addr_offs(sh, percpu);
1902         int count;
1903         struct async_submit_ctl submit;
1904
1905         pr_debug("%s: stripe %llu\n", __func__,
1906                 (unsigned long long)sh->sector);
1907
1908         count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_WANT_DRAIN);
1909
1910         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1911                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1912         tx = async_gen_syndrome(blocks, offs, count+2,
1913                         RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1914
1915         return tx;
1916 }
1917
1918 static struct dma_async_tx_descriptor *
1919 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1920 {
1921         struct r5conf *conf = sh->raid_conf;
1922         int disks = sh->disks;
1923         int i;
1924         struct stripe_head *head_sh = sh;
1925
1926         pr_debug("%s: stripe %llu\n", __func__,
1927                 (unsigned long long)sh->sector);
1928
1929         for (i = disks; i--; ) {
1930                 struct r5dev *dev;
1931                 struct bio *chosen;
1932
1933                 sh = head_sh;
1934                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1935                         struct bio *wbi;
1936
1937 again:
1938                         dev = &sh->dev[i];
1939                         /*
1940                          * clear R5_InJournal, so when rewriting a page in
1941                          * journal, it is not skipped by r5l_log_stripe()
1942                          */
1943                         clear_bit(R5_InJournal, &dev->flags);
1944                         spin_lock_irq(&sh->stripe_lock);
1945                         chosen = dev->towrite;
1946                         dev->towrite = NULL;
1947                         sh->overwrite_disks = 0;
1948                         BUG_ON(dev->written);
1949                         wbi = dev->written = chosen;
1950                         spin_unlock_irq(&sh->stripe_lock);
1951                         WARN_ON(dev->page != dev->orig_page);
1952
1953                         while (wbi && wbi->bi_iter.bi_sector <
1954                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1955                                 if (wbi->bi_opf & REQ_FUA)
1956                                         set_bit(R5_WantFUA, &dev->flags);
1957                                 if (wbi->bi_opf & REQ_SYNC)
1958                                         set_bit(R5_SyncIO, &dev->flags);
1959                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1960                                         set_bit(R5_Discard, &dev->flags);
1961                                 else {
1962                                         tx = async_copy_data(1, wbi, &dev->page,
1963                                                              dev->offset,
1964                                                              dev->sector, tx, sh,
1965                                                              r5c_is_writeback(conf->log));
1966                                         if (dev->page != dev->orig_page &&
1967                                             !r5c_is_writeback(conf->log)) {
1968                                                 set_bit(R5_SkipCopy, &dev->flags);
1969                                                 clear_bit(R5_UPTODATE, &dev->flags);
1970                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1971                                         }
1972                                 }
1973                                 wbi = r5_next_bio(conf, wbi, dev->sector);
1974                         }
1975
1976                         if (head_sh->batch_head) {
1977                                 sh = list_first_entry(&sh->batch_list,
1978                                                       struct stripe_head,
1979                                                       batch_list);
1980                                 if (sh == head_sh)
1981                                         continue;
1982                                 goto again;
1983                         }
1984                 }
1985         }
1986
1987         return tx;
1988 }
1989
1990 static void ops_complete_reconstruct(void *stripe_head_ref)
1991 {
1992         struct stripe_head *sh = stripe_head_ref;
1993         int disks = sh->disks;
1994         int pd_idx = sh->pd_idx;
1995         int qd_idx = sh->qd_idx;
1996         int i;
1997         bool fua = false, sync = false, discard = false;
1998
1999         pr_debug("%s: stripe %llu\n", __func__,
2000                 (unsigned long long)sh->sector);
2001
2002         for (i = disks; i--; ) {
2003                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
2004                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
2005                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
2006         }
2007
2008         for (i = disks; i--; ) {
2009                 struct r5dev *dev = &sh->dev[i];
2010
2011                 if (dev->written || i == pd_idx || i == qd_idx) {
2012                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
2013                                 set_bit(R5_UPTODATE, &dev->flags);
2014                                 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
2015                                         set_bit(R5_Expanded, &dev->flags);
2016                         }
2017                         if (fua)
2018                                 set_bit(R5_WantFUA, &dev->flags);
2019                         if (sync)
2020                                 set_bit(R5_SyncIO, &dev->flags);
2021                 }
2022         }
2023
2024         if (sh->reconstruct_state == reconstruct_state_drain_run)
2025                 sh->reconstruct_state = reconstruct_state_drain_result;
2026         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
2027                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
2028         else {
2029                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
2030                 sh->reconstruct_state = reconstruct_state_result;
2031         }
2032
2033         set_bit(STRIPE_HANDLE, &sh->state);
2034         raid5_release_stripe(sh);
2035 }
2036
2037 static void
2038 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
2039                      struct dma_async_tx_descriptor *tx)
2040 {
2041         int disks = sh->disks;
2042         struct page **xor_srcs;
2043         unsigned int *off_srcs;
2044         struct async_submit_ctl submit;
2045         int count, pd_idx = sh->pd_idx, i;
2046         struct page *xor_dest;
2047         unsigned int off_dest;
2048         int prexor = 0;
2049         unsigned long flags;
2050         int j = 0;
2051         struct stripe_head *head_sh = sh;
2052         int last_stripe;
2053
2054         pr_debug("%s: stripe %llu\n", __func__,
2055                 (unsigned long long)sh->sector);
2056
2057         for (i = 0; i < sh->disks; i++) {
2058                 if (pd_idx == i)
2059                         continue;
2060                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2061                         break;
2062         }
2063         if (i >= sh->disks) {
2064                 atomic_inc(&sh->count);
2065                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
2066                 ops_complete_reconstruct(sh);
2067                 return;
2068         }
2069 again:
2070         count = 0;
2071         xor_srcs = to_addr_page(percpu, j);
2072         off_srcs = to_addr_offs(sh, percpu);
2073         /* check if prexor is active which means only process blocks
2074          * that are part of a read-modify-write (written)
2075          */
2076         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2077                 prexor = 1;
2078                 off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
2079                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
2080                 for (i = disks; i--; ) {
2081                         struct r5dev *dev = &sh->dev[i];
2082                         if (head_sh->dev[i].written ||
2083                             test_bit(R5_InJournal, &head_sh->dev[i].flags)) {
2084                                 off_srcs[count] = dev->offset;
2085                                 xor_srcs[count++] = dev->page;
2086                         }
2087                 }
2088         } else {
2089                 xor_dest = sh->dev[pd_idx].page;
2090                 off_dest = sh->dev[pd_idx].offset;
2091                 for (i = disks; i--; ) {
2092                         struct r5dev *dev = &sh->dev[i];
2093                         if (i != pd_idx) {
2094                                 off_srcs[count] = dev->offset;
2095                                 xor_srcs[count++] = dev->page;
2096                         }
2097                 }
2098         }
2099
2100         /* 1/ if we prexor'd then the dest is reused as a source
2101          * 2/ if we did not prexor then we are redoing the parity
2102          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2103          * for the synchronous xor case
2104          */
2105         last_stripe = !head_sh->batch_head ||
2106                 list_first_entry(&sh->batch_list,
2107                                  struct stripe_head, batch_list) == head_sh;
2108         if (last_stripe) {
2109                 flags = ASYNC_TX_ACK |
2110                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
2111
2112                 atomic_inc(&head_sh->count);
2113                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
2114                                   to_addr_conv(sh, percpu, j));
2115         } else {
2116                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
2117                 init_async_submit(&submit, flags, tx, NULL, NULL,
2118                                   to_addr_conv(sh, percpu, j));
2119         }
2120
2121         if (unlikely(count == 1))
2122                 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
2123                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2124         else
2125                 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2126                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2127         if (!last_stripe) {
2128                 j++;
2129                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2130                                       batch_list);
2131                 goto again;
2132         }
2133 }
2134
2135 static void
2136 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
2137                      struct dma_async_tx_descriptor *tx)
2138 {
2139         struct async_submit_ctl submit;
2140         struct page **blocks;
2141         unsigned int *offs;
2142         int count, i, j = 0;
2143         struct stripe_head *head_sh = sh;
2144         int last_stripe;
2145         int synflags;
2146         unsigned long txflags;
2147
2148         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
2149
2150         for (i = 0; i < sh->disks; i++) {
2151                 if (sh->pd_idx == i || sh->qd_idx == i)
2152                         continue;
2153                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2154                         break;
2155         }
2156         if (i >= sh->disks) {
2157                 atomic_inc(&sh->count);
2158                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2159                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2160                 ops_complete_reconstruct(sh);
2161                 return;
2162         }
2163
2164 again:
2165         blocks = to_addr_page(percpu, j);
2166         offs = to_addr_offs(sh, percpu);
2167
2168         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2169                 synflags = SYNDROME_SRC_WRITTEN;
2170                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
2171         } else {
2172                 synflags = SYNDROME_SRC_ALL;
2173                 txflags = ASYNC_TX_ACK;
2174         }
2175
2176         count = set_syndrome_sources(blocks, offs, sh, synflags);
2177         last_stripe = !head_sh->batch_head ||
2178                 list_first_entry(&sh->batch_list,
2179                                  struct stripe_head, batch_list) == head_sh;
2180
2181         if (last_stripe) {
2182                 atomic_inc(&head_sh->count);
2183                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
2184                                   head_sh, to_addr_conv(sh, percpu, j));
2185         } else
2186                 init_async_submit(&submit, 0, tx, NULL, NULL,
2187                                   to_addr_conv(sh, percpu, j));
2188         tx = async_gen_syndrome(blocks, offs, count+2,
2189                         RAID5_STRIPE_SIZE(sh->raid_conf),  &submit);
2190         if (!last_stripe) {
2191                 j++;
2192                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2193                                       batch_list);
2194                 goto again;
2195         }
2196 }
2197
2198 static void ops_complete_check(void *stripe_head_ref)
2199 {
2200         struct stripe_head *sh = stripe_head_ref;
2201
2202         pr_debug("%s: stripe %llu\n", __func__,
2203                 (unsigned long long)sh->sector);
2204
2205         sh->check_state = check_state_check_result;
2206         set_bit(STRIPE_HANDLE, &sh->state);
2207         raid5_release_stripe(sh);
2208 }
2209
2210 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2211 {
2212         int disks = sh->disks;
2213         int pd_idx = sh->pd_idx;
2214         int qd_idx = sh->qd_idx;
2215         struct page *xor_dest;
2216         unsigned int off_dest;
2217         struct page **xor_srcs = to_addr_page(percpu, 0);
2218         unsigned int *off_srcs = to_addr_offs(sh, percpu);
2219         struct dma_async_tx_descriptor *tx;
2220         struct async_submit_ctl submit;
2221         int count;
2222         int i;
2223
2224         pr_debug("%s: stripe %llu\n", __func__,
2225                 (unsigned long long)sh->sector);
2226
2227         BUG_ON(sh->batch_head);
2228         count = 0;
2229         xor_dest = sh->dev[pd_idx].page;
2230         off_dest = sh->dev[pd_idx].offset;
2231         off_srcs[count] = off_dest;
2232         xor_srcs[count++] = xor_dest;
2233         for (i = disks; i--; ) {
2234                 if (i == pd_idx || i == qd_idx)
2235                         continue;
2236                 off_srcs[count] = sh->dev[i].offset;
2237                 xor_srcs[count++] = sh->dev[i].page;
2238         }
2239
2240         init_async_submit(&submit, 0, NULL, NULL, NULL,
2241                           to_addr_conv(sh, percpu, 0));
2242         tx = async_xor_val_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2243                            RAID5_STRIPE_SIZE(sh->raid_conf),
2244                            &sh->ops.zero_sum_result, &submit);
2245
2246         atomic_inc(&sh->count);
2247         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2248         tx = async_trigger_callback(&submit);
2249 }
2250
2251 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2252 {
2253         struct page **srcs = to_addr_page(percpu, 0);
2254         unsigned int *offs = to_addr_offs(sh, percpu);
2255         struct async_submit_ctl submit;
2256         int count;
2257
2258         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2259                 (unsigned long long)sh->sector, checkp);
2260
2261         BUG_ON(sh->batch_head);
2262         count = set_syndrome_sources(srcs, offs, sh, SYNDROME_SRC_ALL);
2263         if (!checkp)
2264                 srcs[count] = NULL;
2265
2266         atomic_inc(&sh->count);
2267         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2268                           sh, to_addr_conv(sh, percpu, 0));
2269         async_syndrome_val(srcs, offs, count+2,
2270                            RAID5_STRIPE_SIZE(sh->raid_conf),
2271                            &sh->ops.zero_sum_result, percpu->spare_page, 0, &submit);
2272 }
2273
2274 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2275 {
2276         int overlap_clear = 0, i, disks = sh->disks;
2277         struct dma_async_tx_descriptor *tx = NULL;
2278         struct r5conf *conf = sh->raid_conf;
2279         int level = conf->level;
2280         struct raid5_percpu *percpu;
2281
2282         local_lock(&conf->percpu->lock);
2283         percpu = this_cpu_ptr(conf->percpu);
2284         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2285                 ops_run_biofill(sh);
2286                 overlap_clear++;
2287         }
2288
2289         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2290                 if (level < 6)
2291                         tx = ops_run_compute5(sh, percpu);
2292                 else {
2293                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
2294                                 tx = ops_run_compute6_1(sh, percpu);
2295                         else
2296                                 tx = ops_run_compute6_2(sh, percpu);
2297                 }
2298                 /* terminate the chain if reconstruct is not set to be run */
2299                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2300                         async_tx_ack(tx);
2301         }
2302
2303         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2304                 if (level < 6)
2305                         tx = ops_run_prexor5(sh, percpu, tx);
2306                 else
2307                         tx = ops_run_prexor6(sh, percpu, tx);
2308         }
2309
2310         if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2311                 tx = ops_run_partial_parity(sh, percpu, tx);
2312
2313         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2314                 tx = ops_run_biodrain(sh, tx);
2315                 overlap_clear++;
2316         }
2317
2318         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2319                 if (level < 6)
2320                         ops_run_reconstruct5(sh, percpu, tx);
2321                 else
2322                         ops_run_reconstruct6(sh, percpu, tx);
2323         }
2324
2325         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2326                 if (sh->check_state == check_state_run)
2327                         ops_run_check_p(sh, percpu);
2328                 else if (sh->check_state == check_state_run_q)
2329                         ops_run_check_pq(sh, percpu, 0);
2330                 else if (sh->check_state == check_state_run_pq)
2331                         ops_run_check_pq(sh, percpu, 1);
2332                 else
2333                         BUG();
2334         }
2335
2336         if (overlap_clear && !sh->batch_head) {
2337                 for (i = disks; i--; ) {
2338                         struct r5dev *dev = &sh->dev[i];
2339                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2340                                 wake_up(&sh->raid_conf->wait_for_overlap);
2341                 }
2342         }
2343         local_unlock(&conf->percpu->lock);
2344 }
2345
2346 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2347 {
2348 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2349         kfree(sh->pages);
2350 #endif
2351         if (sh->ppl_page)
2352                 __free_page(sh->ppl_page);
2353         kmem_cache_free(sc, sh);
2354 }
2355
2356 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2357         int disks, struct r5conf *conf)
2358 {
2359         struct stripe_head *sh;
2360
2361         sh = kmem_cache_zalloc(sc, gfp);
2362         if (sh) {
2363                 spin_lock_init(&sh->stripe_lock);
2364                 spin_lock_init(&sh->batch_lock);
2365                 INIT_LIST_HEAD(&sh->batch_list);
2366                 INIT_LIST_HEAD(&sh->lru);
2367                 INIT_LIST_HEAD(&sh->r5c);
2368                 INIT_LIST_HEAD(&sh->log_list);
2369                 atomic_set(&sh->count, 1);
2370                 sh->raid_conf = conf;
2371                 sh->log_start = MaxSector;
2372
2373                 if (raid5_has_ppl(conf)) {
2374                         sh->ppl_page = alloc_page(gfp);
2375                         if (!sh->ppl_page) {
2376                                 free_stripe(sc, sh);
2377                                 return NULL;
2378                         }
2379                 }
2380 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2381                 if (init_stripe_shared_pages(sh, conf, disks)) {
2382                         free_stripe(sc, sh);
2383                         return NULL;
2384                 }
2385 #endif
2386         }
2387         return sh;
2388 }
2389 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2390 {
2391         struct stripe_head *sh;
2392
2393         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2394         if (!sh)
2395                 return 0;
2396
2397         if (grow_buffers(sh, gfp)) {
2398                 shrink_buffers(sh);
2399                 free_stripe(conf->slab_cache, sh);
2400                 return 0;
2401         }
2402         sh->hash_lock_index =
2403                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2404         /* we just created an active stripe so... */
2405         atomic_inc(&conf->active_stripes);
2406
2407         raid5_release_stripe(sh);
2408         WRITE_ONCE(conf->max_nr_stripes, conf->max_nr_stripes + 1);
2409         return 1;
2410 }
2411
2412 static int grow_stripes(struct r5conf *conf, int num)
2413 {
2414         struct kmem_cache *sc;
2415         size_t namelen = sizeof(conf->cache_name[0]);
2416         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2417
2418         if (mddev_is_dm(conf->mddev))
2419                 snprintf(conf->cache_name[0], namelen,
2420                         "raid%d-%p", conf->level, conf->mddev);
2421         else
2422                 snprintf(conf->cache_name[0], namelen,
2423                         "raid%d-%s", conf->level, mdname(conf->mddev));
2424         snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2425
2426         conf->active_name = 0;
2427         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2428                                struct_size_t(struct stripe_head, dev, devs),
2429                                0, 0, NULL);
2430         if (!sc)
2431                 return 1;
2432         conf->slab_cache = sc;
2433         conf->pool_size = devs;
2434         while (num--)
2435                 if (!grow_one_stripe(conf, GFP_KERNEL))
2436                         return 1;
2437
2438         return 0;
2439 }
2440
2441 /**
2442  * scribble_alloc - allocate percpu scribble buffer for required size
2443  *                  of the scribble region
2444  * @percpu: from for_each_present_cpu() of the caller
2445  * @num: total number of disks in the array
2446  * @cnt: scribble objs count for required size of the scribble region
2447  *
2448  * The scribble buffer size must be enough to contain:
2449  * 1/ a struct page pointer for each device in the array +2
2450  * 2/ room to convert each entry in (1) to its corresponding dma
2451  *    (dma_map_page()) or page (page_address()) address.
2452  *
2453  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2454  * calculate over all devices (not just the data blocks), using zeros in place
2455  * of the P and Q blocks.
2456  */
2457 static int scribble_alloc(struct raid5_percpu *percpu,
2458                           int num, int cnt)
2459 {
2460         size_t obj_size =
2461                 sizeof(struct page *) * (num + 2) +
2462                 sizeof(addr_conv_t) * (num + 2) +
2463                 sizeof(unsigned int) * (num + 2);
2464         void *scribble;
2465
2466         /*
2467          * If here is in raid array suspend context, it is in memalloc noio
2468          * context as well, there is no potential recursive memory reclaim
2469          * I/Os with the GFP_KERNEL flag.
2470          */
2471         scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2472         if (!scribble)
2473                 return -ENOMEM;
2474
2475         kvfree(percpu->scribble);
2476
2477         percpu->scribble = scribble;
2478         percpu->scribble_obj_size = obj_size;
2479         return 0;
2480 }
2481
2482 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2483 {
2484         unsigned long cpu;
2485         int err = 0;
2486
2487         /* Never shrink. */
2488         if (conf->scribble_disks >= new_disks &&
2489             conf->scribble_sectors >= new_sectors)
2490                 return 0;
2491
2492         raid5_quiesce(conf->mddev, true);
2493         cpus_read_lock();
2494
2495         for_each_present_cpu(cpu) {
2496                 struct raid5_percpu *percpu;
2497
2498                 percpu = per_cpu_ptr(conf->percpu, cpu);
2499                 err = scribble_alloc(percpu, new_disks,
2500                                      new_sectors / RAID5_STRIPE_SECTORS(conf));
2501                 if (err)
2502                         break;
2503         }
2504
2505         cpus_read_unlock();
2506         raid5_quiesce(conf->mddev, false);
2507
2508         if (!err) {
2509                 conf->scribble_disks = new_disks;
2510                 conf->scribble_sectors = new_sectors;
2511         }
2512         return err;
2513 }
2514
2515 static int resize_stripes(struct r5conf *conf, int newsize)
2516 {
2517         /* Make all the stripes able to hold 'newsize' devices.
2518          * New slots in each stripe get 'page' set to a new page.
2519          *
2520          * This happens in stages:
2521          * 1/ create a new kmem_cache and allocate the required number of
2522          *    stripe_heads.
2523          * 2/ gather all the old stripe_heads and transfer the pages across
2524          *    to the new stripe_heads.  This will have the side effect of
2525          *    freezing the array as once all stripe_heads have been collected,
2526          *    no IO will be possible.  Old stripe heads are freed once their
2527          *    pages have been transferred over, and the old kmem_cache is
2528          *    freed when all stripes are done.
2529          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2530          *    we simple return a failure status - no need to clean anything up.
2531          * 4/ allocate new pages for the new slots in the new stripe_heads.
2532          *    If this fails, we don't bother trying the shrink the
2533          *    stripe_heads down again, we just leave them as they are.
2534          *    As each stripe_head is processed the new one is released into
2535          *    active service.
2536          *
2537          * Once step2 is started, we cannot afford to wait for a write,
2538          * so we use GFP_NOIO allocations.
2539          */
2540         struct stripe_head *osh, *nsh;
2541         LIST_HEAD(newstripes);
2542         struct disk_info *ndisks;
2543         int err = 0;
2544         struct kmem_cache *sc;
2545         int i;
2546         int hash, cnt;
2547
2548         md_allow_write(conf->mddev);
2549
2550         /* Step 1 */
2551         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2552                                struct_size_t(struct stripe_head, dev, newsize),
2553                                0, 0, NULL);
2554         if (!sc)
2555                 return -ENOMEM;
2556
2557         /* Need to ensure auto-resizing doesn't interfere */
2558         mutex_lock(&conf->cache_size_mutex);
2559
2560         for (i = conf->max_nr_stripes; i; i--) {
2561                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2562                 if (!nsh)
2563                         break;
2564
2565                 list_add(&nsh->lru, &newstripes);
2566         }
2567         if (i) {
2568                 /* didn't get enough, give up */
2569                 while (!list_empty(&newstripes)) {
2570                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2571                         list_del(&nsh->lru);
2572                         free_stripe(sc, nsh);
2573                 }
2574                 kmem_cache_destroy(sc);
2575                 mutex_unlock(&conf->cache_size_mutex);
2576                 return -ENOMEM;
2577         }
2578         /* Step 2 - Must use GFP_NOIO now.
2579          * OK, we have enough stripes, start collecting inactive
2580          * stripes and copying them over
2581          */
2582         hash = 0;
2583         cnt = 0;
2584         list_for_each_entry(nsh, &newstripes, lru) {
2585                 lock_device_hash_lock(conf, hash);
2586                 wait_event_cmd(conf->wait_for_stripe,
2587                                     !list_empty(conf->inactive_list + hash),
2588                                     unlock_device_hash_lock(conf, hash),
2589                                     lock_device_hash_lock(conf, hash));
2590                 osh = get_free_stripe(conf, hash);
2591                 unlock_device_hash_lock(conf, hash);
2592
2593 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2594         for (i = 0; i < osh->nr_pages; i++) {
2595                 nsh->pages[i] = osh->pages[i];
2596                 osh->pages[i] = NULL;
2597         }
2598 #endif
2599                 for(i=0; i<conf->pool_size; i++) {
2600                         nsh->dev[i].page = osh->dev[i].page;
2601                         nsh->dev[i].orig_page = osh->dev[i].page;
2602                         nsh->dev[i].offset = osh->dev[i].offset;
2603                 }
2604                 nsh->hash_lock_index = hash;
2605                 free_stripe(conf->slab_cache, osh);
2606                 cnt++;
2607                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2608                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2609                         hash++;
2610                         cnt = 0;
2611                 }
2612         }
2613         kmem_cache_destroy(conf->slab_cache);
2614
2615         /* Step 3.
2616          * At this point, we are holding all the stripes so the array
2617          * is completely stalled, so now is a good time to resize
2618          * conf->disks and the scribble region
2619          */
2620         ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2621         if (ndisks) {
2622                 for (i = 0; i < conf->pool_size; i++)
2623                         ndisks[i] = conf->disks[i];
2624
2625                 for (i = conf->pool_size; i < newsize; i++) {
2626                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2627                         if (!ndisks[i].extra_page)
2628                                 err = -ENOMEM;
2629                 }
2630
2631                 if (err) {
2632                         for (i = conf->pool_size; i < newsize; i++)
2633                                 if (ndisks[i].extra_page)
2634                                         put_page(ndisks[i].extra_page);
2635                         kfree(ndisks);
2636                 } else {
2637                         kfree(conf->disks);
2638                         conf->disks = ndisks;
2639                 }
2640         } else
2641                 err = -ENOMEM;
2642
2643         conf->slab_cache = sc;
2644         conf->active_name = 1-conf->active_name;
2645
2646         /* Step 4, return new stripes to service */
2647         while(!list_empty(&newstripes)) {
2648                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2649                 list_del_init(&nsh->lru);
2650
2651 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2652                 for (i = 0; i < nsh->nr_pages; i++) {
2653                         if (nsh->pages[i])
2654                                 continue;
2655                         nsh->pages[i] = alloc_page(GFP_NOIO);
2656                         if (!nsh->pages[i])
2657                                 err = -ENOMEM;
2658                 }
2659
2660                 for (i = conf->raid_disks; i < newsize; i++) {
2661                         if (nsh->dev[i].page)
2662                                 continue;
2663                         nsh->dev[i].page = raid5_get_dev_page(nsh, i);
2664                         nsh->dev[i].orig_page = nsh->dev[i].page;
2665                         nsh->dev[i].offset = raid5_get_page_offset(nsh, i);
2666                 }
2667 #else
2668                 for (i=conf->raid_disks; i < newsize; i++)
2669                         if (nsh->dev[i].page == NULL) {
2670                                 struct page *p = alloc_page(GFP_NOIO);
2671                                 nsh->dev[i].page = p;
2672                                 nsh->dev[i].orig_page = p;
2673                                 nsh->dev[i].offset = 0;
2674                                 if (!p)
2675                                         err = -ENOMEM;
2676                         }
2677 #endif
2678                 raid5_release_stripe(nsh);
2679         }
2680         /* critical section pass, GFP_NOIO no longer needed */
2681
2682         if (!err)
2683                 conf->pool_size = newsize;
2684         mutex_unlock(&conf->cache_size_mutex);
2685
2686         return err;
2687 }
2688
2689 static int drop_one_stripe(struct r5conf *conf)
2690 {
2691         struct stripe_head *sh;
2692         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2693
2694         spin_lock_irq(conf->hash_locks + hash);
2695         sh = get_free_stripe(conf, hash);
2696         spin_unlock_irq(conf->hash_locks + hash);
2697         if (!sh)
2698                 return 0;
2699         BUG_ON(atomic_read(&sh->count));
2700         shrink_buffers(sh);
2701         free_stripe(conf->slab_cache, sh);
2702         atomic_dec(&conf->active_stripes);
2703         WRITE_ONCE(conf->max_nr_stripes, conf->max_nr_stripes - 1);
2704         return 1;
2705 }
2706
2707 static void shrink_stripes(struct r5conf *conf)
2708 {
2709         while (conf->max_nr_stripes &&
2710                drop_one_stripe(conf))
2711                 ;
2712
2713         kmem_cache_destroy(conf->slab_cache);
2714         conf->slab_cache = NULL;
2715 }
2716
2717 static void raid5_end_read_request(struct bio * bi)
2718 {
2719         struct stripe_head *sh = bi->bi_private;
2720         struct r5conf *conf = sh->raid_conf;
2721         int disks = sh->disks, i;
2722         struct md_rdev *rdev = NULL;
2723         sector_t s;
2724
2725         for (i=0 ; i<disks; i++)
2726                 if (bi == &sh->dev[i].req)
2727                         break;
2728
2729         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2730                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2731                 bi->bi_status);
2732         if (i == disks) {
2733                 BUG();
2734                 return;
2735         }
2736         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2737                 /* If replacement finished while this request was outstanding,
2738                  * 'replacement' might be NULL already.
2739                  * In that case it moved down to 'rdev'.
2740                  * rdev is not removed until all requests are finished.
2741                  */
2742                 rdev = conf->disks[i].replacement;
2743         if (!rdev)
2744                 rdev = conf->disks[i].rdev;
2745
2746         if (use_new_offset(conf, sh))
2747                 s = sh->sector + rdev->new_data_offset;
2748         else
2749                 s = sh->sector + rdev->data_offset;
2750         if (!bi->bi_status) {
2751                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2752                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2753                         /* Note that this cannot happen on a
2754                          * replacement device.  We just fail those on
2755                          * any error
2756                          */
2757                         pr_info_ratelimited(
2758                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %pg)\n",
2759                                 mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
2760                                 (unsigned long long)s,
2761                                 rdev->bdev);
2762                         atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
2763                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2764                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2765                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2766                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2767
2768                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2769                         /*
2770                          * end read for a page in journal, this
2771                          * must be preparing for prexor in rmw
2772                          */
2773                         set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2774
2775                 if (atomic_read(&rdev->read_errors))
2776                         atomic_set(&rdev->read_errors, 0);
2777         } else {
2778                 int retry = 0;
2779                 int set_bad = 0;
2780
2781                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2782                 if (!(bi->bi_status == BLK_STS_PROTECTION))
2783                         atomic_inc(&rdev->read_errors);
2784                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2785                         pr_warn_ratelimited(
2786                                 "md/raid:%s: read error on replacement device (sector %llu on %pg).\n",
2787                                 mdname(conf->mddev),
2788                                 (unsigned long long)s,
2789                                 rdev->bdev);
2790                 else if (conf->mddev->degraded >= conf->max_degraded) {
2791                         set_bad = 1;
2792                         pr_warn_ratelimited(
2793                                 "md/raid:%s: read error not correctable (sector %llu on %pg).\n",
2794                                 mdname(conf->mddev),
2795                                 (unsigned long long)s,
2796                                 rdev->bdev);
2797                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2798                         /* Oh, no!!! */
2799                         set_bad = 1;
2800                         pr_warn_ratelimited(
2801                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %pg).\n",
2802                                 mdname(conf->mddev),
2803                                 (unsigned long long)s,
2804                                 rdev->bdev);
2805                 } else if (atomic_read(&rdev->read_errors)
2806                          > conf->max_nr_stripes) {
2807                         if (!test_bit(Faulty, &rdev->flags)) {
2808                                 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2809                                     mdname(conf->mddev),
2810                                     atomic_read(&rdev->read_errors),
2811                                     conf->max_nr_stripes);
2812                                 pr_warn("md/raid:%s: Too many read errors, failing device %pg.\n",
2813                                     mdname(conf->mddev), rdev->bdev);
2814                         }
2815                 } else
2816                         retry = 1;
2817                 if (set_bad && test_bit(In_sync, &rdev->flags)
2818                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2819                         retry = 1;
2820                 if (retry)
2821                         if (sh->qd_idx >= 0 && sh->pd_idx == i)
2822                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2823                         else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2824                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2825                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2826                         } else
2827                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2828                 else {
2829                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2830                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2831                         if (!(set_bad
2832                               && test_bit(In_sync, &rdev->flags)
2833                               && rdev_set_badblocks(
2834                                       rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
2835                                 md_error(conf->mddev, rdev);
2836                 }
2837         }
2838         rdev_dec_pending(rdev, conf->mddev);
2839         bio_uninit(bi);
2840         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2841         set_bit(STRIPE_HANDLE, &sh->state);
2842         raid5_release_stripe(sh);
2843 }
2844
2845 static void raid5_end_write_request(struct bio *bi)
2846 {
2847         struct stripe_head *sh = bi->bi_private;
2848         struct r5conf *conf = sh->raid_conf;
2849         int disks = sh->disks, i;
2850         struct md_rdev *rdev;
2851         int replacement = 0;
2852
2853         for (i = 0 ; i < disks; i++) {
2854                 if (bi == &sh->dev[i].req) {
2855                         rdev = conf->disks[i].rdev;
2856                         break;
2857                 }
2858                 if (bi == &sh->dev[i].rreq) {
2859                         rdev = conf->disks[i].replacement;
2860                         if (rdev)
2861                                 replacement = 1;
2862                         else
2863                                 /* rdev was removed and 'replacement'
2864                                  * replaced it.  rdev is not removed
2865                                  * until all requests are finished.
2866                                  */
2867                                 rdev = conf->disks[i].rdev;
2868                         break;
2869                 }
2870         }
2871         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2872                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2873                 bi->bi_status);
2874         if (i == disks) {
2875                 BUG();
2876                 return;
2877         }
2878
2879         if (replacement) {
2880                 if (bi->bi_status)
2881                         md_error(conf->mddev, rdev);
2882                 else if (rdev_has_badblock(rdev, sh->sector,
2883                                            RAID5_STRIPE_SECTORS(conf)))
2884                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2885         } else {
2886                 if (bi->bi_status) {
2887                         set_bit(STRIPE_DEGRADED, &sh->state);
2888                         set_bit(WriteErrorSeen, &rdev->flags);
2889                         set_bit(R5_WriteError, &sh->dev[i].flags);
2890                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2891                                 set_bit(MD_RECOVERY_NEEDED,
2892                                         &rdev->mddev->recovery);
2893                 } else if (rdev_has_badblock(rdev, sh->sector,
2894                                              RAID5_STRIPE_SECTORS(conf))) {
2895                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2896                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2897                                 /* That was a successful write so make
2898                                  * sure it looks like we already did
2899                                  * a re-write.
2900                                  */
2901                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2902                 }
2903         }
2904         rdev_dec_pending(rdev, conf->mddev);
2905
2906         if (sh->batch_head && bi->bi_status && !replacement)
2907                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2908
2909         bio_uninit(bi);
2910         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2911                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2912         set_bit(STRIPE_HANDLE, &sh->state);
2913
2914         if (sh->batch_head && sh != sh->batch_head)
2915                 raid5_release_stripe(sh->batch_head);
2916         raid5_release_stripe(sh);
2917 }
2918
2919 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2920 {
2921         struct r5conf *conf = mddev->private;
2922         unsigned long flags;
2923         pr_debug("raid456: error called\n");
2924
2925         pr_crit("md/raid:%s: Disk failure on %pg, disabling device.\n",
2926                 mdname(mddev), rdev->bdev);
2927
2928         spin_lock_irqsave(&conf->device_lock, flags);
2929         set_bit(Faulty, &rdev->flags);
2930         clear_bit(In_sync, &rdev->flags);
2931         mddev->degraded = raid5_calc_degraded(conf);
2932
2933         if (has_failed(conf)) {
2934                 set_bit(MD_BROKEN, &conf->mddev->flags);
2935                 conf->recovery_disabled = mddev->recovery_disabled;
2936
2937                 pr_crit("md/raid:%s: Cannot continue operation (%d/%d failed).\n",
2938                         mdname(mddev), mddev->degraded, conf->raid_disks);
2939         } else {
2940                 pr_crit("md/raid:%s: Operation continuing on %d devices.\n",
2941                         mdname(mddev), conf->raid_disks - mddev->degraded);
2942         }
2943
2944         spin_unlock_irqrestore(&conf->device_lock, flags);
2945         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2946
2947         set_bit(Blocked, &rdev->flags);
2948         set_mask_bits(&mddev->sb_flags, 0,
2949                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2950         r5c_update_on_rdev_error(mddev, rdev);
2951 }
2952
2953 /*
2954  * Input: a 'big' sector number,
2955  * Output: index of the data and parity disk, and the sector # in them.
2956  */
2957 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2958                               int previous, int *dd_idx,
2959                               struct stripe_head *sh)
2960 {
2961         sector_t stripe, stripe2;
2962         sector_t chunk_number;
2963         unsigned int chunk_offset;
2964         int pd_idx, qd_idx;
2965         int ddf_layout = 0;
2966         sector_t new_sector;
2967         int algorithm = previous ? conf->prev_algo
2968                                  : conf->algorithm;
2969         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2970                                          : conf->chunk_sectors;
2971         int raid_disks = previous ? conf->previous_raid_disks
2972                                   : conf->raid_disks;
2973         int data_disks = raid_disks - conf->max_degraded;
2974
2975         /* First compute the information on this sector */
2976
2977         /*
2978          * Compute the chunk number and the sector offset inside the chunk
2979          */
2980         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2981         chunk_number = r_sector;
2982
2983         /*
2984          * Compute the stripe number
2985          */
2986         stripe = chunk_number;
2987         *dd_idx = sector_div(stripe, data_disks);
2988         stripe2 = stripe;
2989         /*
2990          * Select the parity disk based on the user selected algorithm.
2991          */
2992         pd_idx = qd_idx = -1;
2993         switch(conf->level) {
2994         case 4:
2995                 pd_idx = data_disks;
2996                 break;
2997         case 5:
2998                 switch (algorithm) {
2999                 case ALGORITHM_LEFT_ASYMMETRIC:
3000                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
3001                         if (*dd_idx >= pd_idx)
3002                                 (*dd_idx)++;
3003                         break;
3004                 case ALGORITHM_RIGHT_ASYMMETRIC:
3005                         pd_idx = sector_div(stripe2, raid_disks);
3006                         if (*dd_idx >= pd_idx)
3007                                 (*dd_idx)++;
3008                         break;
3009                 case ALGORITHM_LEFT_SYMMETRIC:
3010                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
3011                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3012                         break;
3013                 case ALGORITHM_RIGHT_SYMMETRIC:
3014                         pd_idx = sector_div(stripe2, raid_disks);
3015                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3016                         break;
3017                 case ALGORITHM_PARITY_0:
3018                         pd_idx = 0;
3019                         (*dd_idx)++;
3020                         break;
3021                 case ALGORITHM_PARITY_N:
3022                         pd_idx = data_disks;
3023                         break;
3024                 default:
3025                         BUG();
3026                 }
3027                 break;
3028         case 6:
3029
3030                 switch (algorithm) {
3031                 case ALGORITHM_LEFT_ASYMMETRIC:
3032                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3033                         qd_idx = pd_idx + 1;
3034                         if (pd_idx == raid_disks-1) {
3035                                 (*dd_idx)++;    /* Q D D D P */
3036                                 qd_idx = 0;
3037                         } else if (*dd_idx >= pd_idx)
3038                                 (*dd_idx) += 2; /* D D P Q D */
3039                         break;
3040                 case ALGORITHM_RIGHT_ASYMMETRIC:
3041                         pd_idx = sector_div(stripe2, raid_disks);
3042                         qd_idx = pd_idx + 1;
3043                         if (pd_idx == raid_disks-1) {
3044                                 (*dd_idx)++;    /* Q D D D P */
3045                                 qd_idx = 0;
3046                         } else if (*dd_idx >= pd_idx)
3047                                 (*dd_idx) += 2; /* D D P Q D */
3048                         break;
3049                 case ALGORITHM_LEFT_SYMMETRIC:
3050                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3051                         qd_idx = (pd_idx + 1) % raid_disks;
3052                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3053                         break;
3054                 case ALGORITHM_RIGHT_SYMMETRIC:
3055                         pd_idx = sector_div(stripe2, raid_disks);
3056                         qd_idx = (pd_idx + 1) % raid_disks;
3057                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3058                         break;
3059
3060                 case ALGORITHM_PARITY_0:
3061                         pd_idx = 0;
3062                         qd_idx = 1;
3063                         (*dd_idx) += 2;
3064                         break;
3065                 case ALGORITHM_PARITY_N:
3066                         pd_idx = data_disks;
3067                         qd_idx = data_disks + 1;
3068                         break;
3069
3070                 case ALGORITHM_ROTATING_ZERO_RESTART:
3071                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
3072                          * of blocks for computing Q is different.
3073                          */
3074                         pd_idx = sector_div(stripe2, raid_disks);
3075                         qd_idx = pd_idx + 1;
3076                         if (pd_idx == raid_disks-1) {
3077                                 (*dd_idx)++;    /* Q D D D P */
3078                                 qd_idx = 0;
3079                         } else if (*dd_idx >= pd_idx)
3080                                 (*dd_idx) += 2; /* D D P Q D */
3081                         ddf_layout = 1;
3082                         break;
3083
3084                 case ALGORITHM_ROTATING_N_RESTART:
3085                         /* Same a left_asymmetric, by first stripe is
3086                          * D D D P Q  rather than
3087                          * Q D D D P
3088                          */
3089                         stripe2 += 1;
3090                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3091                         qd_idx = pd_idx + 1;
3092                         if (pd_idx == raid_disks-1) {
3093                                 (*dd_idx)++;    /* Q D D D P */
3094                                 qd_idx = 0;
3095                         } else if (*dd_idx >= pd_idx)
3096                                 (*dd_idx) += 2; /* D D P Q D */
3097                         ddf_layout = 1;
3098                         break;
3099
3100                 case ALGORITHM_ROTATING_N_CONTINUE:
3101                         /* Same as left_symmetric but Q is before P */
3102                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3103                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
3104                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3105                         ddf_layout = 1;
3106                         break;
3107
3108                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3109                         /* RAID5 left_asymmetric, with Q on last device */
3110                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3111                         if (*dd_idx >= pd_idx)
3112                                 (*dd_idx)++;
3113                         qd_idx = raid_disks - 1;
3114                         break;
3115
3116                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3117                         pd_idx = sector_div(stripe2, raid_disks-1);
3118                         if (*dd_idx >= pd_idx)
3119                                 (*dd_idx)++;
3120                         qd_idx = raid_disks - 1;
3121                         break;
3122
3123                 case ALGORITHM_LEFT_SYMMETRIC_6:
3124                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3125                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3126                         qd_idx = raid_disks - 1;
3127                         break;
3128
3129                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3130                         pd_idx = sector_div(stripe2, raid_disks-1);
3131                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3132                         qd_idx = raid_disks - 1;
3133                         break;
3134
3135                 case ALGORITHM_PARITY_0_6:
3136                         pd_idx = 0;
3137                         (*dd_idx)++;
3138                         qd_idx = raid_disks - 1;
3139                         break;
3140
3141                 default:
3142                         BUG();
3143                 }
3144                 break;
3145         }
3146
3147         if (sh) {
3148                 sh->pd_idx = pd_idx;
3149                 sh->qd_idx = qd_idx;
3150                 sh->ddf_layout = ddf_layout;
3151         }
3152         /*
3153          * Finally, compute the new sector number
3154          */
3155         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
3156         return new_sector;
3157 }
3158
3159 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
3160 {
3161         struct r5conf *conf = sh->raid_conf;
3162         int raid_disks = sh->disks;
3163         int data_disks = raid_disks - conf->max_degraded;
3164         sector_t new_sector = sh->sector, check;
3165         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
3166                                          : conf->chunk_sectors;
3167         int algorithm = previous ? conf->prev_algo
3168                                  : conf->algorithm;
3169         sector_t stripe;
3170         int chunk_offset;
3171         sector_t chunk_number;
3172         int dummy1, dd_idx = i;
3173         sector_t r_sector;
3174         struct stripe_head sh2;
3175
3176         chunk_offset = sector_div(new_sector, sectors_per_chunk);
3177         stripe = new_sector;
3178
3179         if (i == sh->pd_idx)
3180                 return 0;
3181         switch(conf->level) {
3182         case 4: break;
3183         case 5:
3184                 switch (algorithm) {
3185                 case ALGORITHM_LEFT_ASYMMETRIC:
3186                 case ALGORITHM_RIGHT_ASYMMETRIC:
3187                         if (i > sh->pd_idx)
3188                                 i--;
3189                         break;
3190                 case ALGORITHM_LEFT_SYMMETRIC:
3191                 case ALGORITHM_RIGHT_SYMMETRIC:
3192                         if (i < sh->pd_idx)
3193                                 i += raid_disks;
3194                         i -= (sh->pd_idx + 1);
3195                         break;
3196                 case ALGORITHM_PARITY_0:
3197                         i -= 1;
3198                         break;
3199                 case ALGORITHM_PARITY_N:
3200                         break;
3201                 default:
3202                         BUG();
3203                 }
3204                 break;
3205         case 6:
3206                 if (i == sh->qd_idx)
3207                         return 0; /* It is the Q disk */
3208                 switch (algorithm) {
3209                 case ALGORITHM_LEFT_ASYMMETRIC:
3210                 case ALGORITHM_RIGHT_ASYMMETRIC:
3211                 case ALGORITHM_ROTATING_ZERO_RESTART:
3212                 case ALGORITHM_ROTATING_N_RESTART:
3213                         if (sh->pd_idx == raid_disks-1)
3214                                 i--;    /* Q D D D P */
3215                         else if (i > sh->pd_idx)
3216                                 i -= 2; /* D D P Q D */
3217                         break;
3218                 case ALGORITHM_LEFT_SYMMETRIC:
3219                 case ALGORITHM_RIGHT_SYMMETRIC:
3220                         if (sh->pd_idx == raid_disks-1)
3221                                 i--; /* Q D D D P */
3222                         else {
3223                                 /* D D P Q D */
3224                                 if (i < sh->pd_idx)
3225                                         i += raid_disks;
3226                                 i -= (sh->pd_idx + 2);
3227                         }
3228                         break;
3229                 case ALGORITHM_PARITY_0:
3230                         i -= 2;
3231                         break;
3232                 case ALGORITHM_PARITY_N:
3233                         break;
3234                 case ALGORITHM_ROTATING_N_CONTINUE:
3235                         /* Like left_symmetric, but P is before Q */
3236                         if (sh->pd_idx == 0)
3237                                 i--;    /* P D D D Q */
3238                         else {
3239                                 /* D D Q P D */
3240                                 if (i < sh->pd_idx)
3241                                         i += raid_disks;
3242                                 i -= (sh->pd_idx + 1);
3243                         }
3244                         break;
3245                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3246                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3247                         if (i > sh->pd_idx)
3248                                 i--;
3249                         break;
3250                 case ALGORITHM_LEFT_SYMMETRIC_6:
3251                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3252                         if (i < sh->pd_idx)
3253                                 i += data_disks + 1;
3254                         i -= (sh->pd_idx + 1);
3255                         break;
3256                 case ALGORITHM_PARITY_0_6:
3257                         i -= 1;
3258                         break;
3259                 default:
3260                         BUG();
3261                 }
3262                 break;
3263         }
3264
3265         chunk_number = stripe * data_disks + i;
3266         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3267
3268         check = raid5_compute_sector(conf, r_sector,
3269                                      previous, &dummy1, &sh2);
3270         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3271                 || sh2.qd_idx != sh->qd_idx) {
3272                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3273                         mdname(conf->mddev));
3274                 return 0;
3275         }
3276         return r_sector;
3277 }
3278
3279 /*
3280  * There are cases where we want handle_stripe_dirtying() and
3281  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3282  *
3283  * This function checks whether we want to delay the towrite. Specifically,
3284  * we delay the towrite when:
3285  *
3286  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3287  *      stripe has data in journal (for other devices).
3288  *
3289  *      In this case, when reading data for the non-overwrite dev, it is
3290  *      necessary to handle complex rmw of write back cache (prexor with
3291  *      orig_page, and xor with page). To keep read path simple, we would
3292  *      like to flush data in journal to RAID disks first, so complex rmw
3293  *      is handled in the write patch (handle_stripe_dirtying).
3294  *
3295  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3296  *
3297  *      It is important to be able to flush all stripes in raid5-cache.
3298  *      Therefore, we need reserve some space on the journal device for
3299  *      these flushes. If flush operation includes pending writes to the
3300  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3301  *      for the flush out. If we exclude these pending writes from flush
3302  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3303  *      Therefore, excluding pending writes in these cases enables more
3304  *      efficient use of the journal device.
3305  *
3306  *      Note: To make sure the stripe makes progress, we only delay
3307  *      towrite for stripes with data already in journal (injournal > 0).
3308  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3309  *      no_space_stripes list.
3310  *
3311  *   3. during journal failure
3312  *      In journal failure, we try to flush all cached data to raid disks
3313  *      based on data in stripe cache. The array is read-only to upper
3314  *      layers, so we would skip all pending writes.
3315  *
3316  */
3317 static inline bool delay_towrite(struct r5conf *conf,
3318                                  struct r5dev *dev,
3319                                  struct stripe_head_state *s)
3320 {
3321         /* case 1 above */
3322         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3323             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3324                 return true;
3325         /* case 2 above */
3326         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3327             s->injournal > 0)
3328                 return true;
3329         /* case 3 above */
3330         if (s->log_failed && s->injournal)
3331                 return true;
3332         return false;
3333 }
3334
3335 static void
3336 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3337                          int rcw, int expand)
3338 {
3339         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3340         struct r5conf *conf = sh->raid_conf;
3341         int level = conf->level;
3342
3343         if (rcw) {
3344                 /*
3345                  * In some cases, handle_stripe_dirtying initially decided to
3346                  * run rmw and allocates extra page for prexor. However, rcw is
3347                  * cheaper later on. We need to free the extra page now,
3348                  * because we won't be able to do that in ops_complete_prexor().
3349                  */
3350                 r5c_release_extra_page(sh);
3351
3352                 for (i = disks; i--; ) {
3353                         struct r5dev *dev = &sh->dev[i];
3354
3355                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3356                                 set_bit(R5_LOCKED, &dev->flags);
3357                                 set_bit(R5_Wantdrain, &dev->flags);
3358                                 if (!expand)
3359                                         clear_bit(R5_UPTODATE, &dev->flags);
3360                                 s->locked++;
3361                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3362                                 set_bit(R5_LOCKED, &dev->flags);
3363                                 s->locked++;
3364                         }
3365                 }
3366                 /* if we are not expanding this is a proper write request, and
3367                  * there will be bios with new data to be drained into the
3368                  * stripe cache
3369                  */
3370                 if (!expand) {
3371                         if (!s->locked)
3372                                 /* False alarm, nothing to do */
3373                                 return;
3374                         sh->reconstruct_state = reconstruct_state_drain_run;
3375                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3376                 } else
3377                         sh->reconstruct_state = reconstruct_state_run;
3378
3379                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3380
3381                 if (s->locked + conf->max_degraded == disks)
3382                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3383                                 atomic_inc(&conf->pending_full_writes);
3384         } else {
3385                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3386                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3387                 BUG_ON(level == 6 &&
3388                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3389                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3390
3391                 for (i = disks; i--; ) {
3392                         struct r5dev *dev = &sh->dev[i];
3393                         if (i == pd_idx || i == qd_idx)
3394                                 continue;
3395
3396                         if (dev->towrite &&
3397                             (test_bit(R5_UPTODATE, &dev->flags) ||
3398                              test_bit(R5_Wantcompute, &dev->flags))) {
3399                                 set_bit(R5_Wantdrain, &dev->flags);
3400                                 set_bit(R5_LOCKED, &dev->flags);
3401                                 clear_bit(R5_UPTODATE, &dev->flags);
3402                                 s->locked++;
3403                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3404                                 set_bit(R5_LOCKED, &dev->flags);
3405                                 s->locked++;
3406                         }
3407                 }
3408                 if (!s->locked)
3409                         /* False alarm - nothing to do */
3410                         return;
3411                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3412                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3413                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3414                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3415         }
3416
3417         /* keep the parity disk(s) locked while asynchronous operations
3418          * are in flight
3419          */
3420         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3421         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3422         s->locked++;
3423
3424         if (level == 6) {
3425                 int qd_idx = sh->qd_idx;
3426                 struct r5dev *dev = &sh->dev[qd_idx];
3427
3428                 set_bit(R5_LOCKED, &dev->flags);
3429                 clear_bit(R5_UPTODATE, &dev->flags);
3430                 s->locked++;
3431         }
3432
3433         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3434             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3435             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3436             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3437                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3438
3439         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3440                 __func__, (unsigned long long)sh->sector,
3441                 s->locked, s->ops_request);
3442 }
3443
3444 static bool stripe_bio_overlaps(struct stripe_head *sh, struct bio *bi,
3445                                 int dd_idx, int forwrite)
3446 {
3447         struct r5conf *conf = sh->raid_conf;
3448         struct bio **bip;
3449
3450         pr_debug("checking bi b#%llu to stripe s#%llu\n",
3451                  bi->bi_iter.bi_sector, sh->sector);
3452
3453         /* Don't allow new IO added to stripes in batch list */
3454         if (sh->batch_head)
3455                 return true;
3456
3457         if (forwrite)
3458                 bip = &sh->dev[dd_idx].towrite;
3459         else
3460                 bip = &sh->dev[dd_idx].toread;
3461
3462         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3463                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3464                         return true;
3465                 bip = &(*bip)->bi_next;
3466         }
3467
3468         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3469                 return true;
3470
3471         if (forwrite && raid5_has_ppl(conf)) {
3472                 /*
3473                  * With PPL only writes to consecutive data chunks within a
3474                  * stripe are allowed because for a single stripe_head we can
3475                  * only have one PPL entry at a time, which describes one data
3476                  * range. Not really an overlap, but wait_for_overlap can be
3477                  * used to handle this.
3478                  */
3479                 sector_t sector;
3480                 sector_t first = 0;
3481                 sector_t last = 0;
3482                 int count = 0;
3483                 int i;
3484
3485                 for (i = 0; i < sh->disks; i++) {
3486                         if (i != sh->pd_idx &&
3487                             (i == dd_idx || sh->dev[i].towrite)) {
3488                                 sector = sh->dev[i].sector;
3489                                 if (count == 0 || sector < first)
3490                                         first = sector;
3491                                 if (sector > last)
3492                                         last = sector;
3493                                 count++;
3494                         }
3495                 }
3496
3497                 if (first + conf->chunk_sectors * (count - 1) != last)
3498                         return true;
3499         }
3500
3501         return false;
3502 }
3503
3504 static void __add_stripe_bio(struct stripe_head *sh, struct bio *bi,
3505                              int dd_idx, int forwrite, int previous)
3506 {
3507         struct r5conf *conf = sh->raid_conf;
3508         struct bio **bip;
3509         int firstwrite = 0;
3510
3511         if (forwrite) {
3512                 bip = &sh->dev[dd_idx].towrite;
3513                 if (!*bip)
3514                         firstwrite = 1;
3515         } else {
3516                 bip = &sh->dev[dd_idx].toread;
3517         }
3518
3519         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector)
3520                 bip = &(*bip)->bi_next;
3521
3522         if (!forwrite || previous)
3523                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3524
3525         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3526         if (*bip)
3527                 bi->bi_next = *bip;
3528         *bip = bi;
3529         bio_inc_remaining(bi);
3530         md_write_inc(conf->mddev, bi);
3531
3532         if (forwrite) {
3533                 /* check if page is covered */
3534                 sector_t sector = sh->dev[dd_idx].sector;
3535                 for (bi=sh->dev[dd_idx].towrite;
3536                      sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
3537                              bi && bi->bi_iter.bi_sector <= sector;
3538                      bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
3539                         if (bio_end_sector(bi) >= sector)
3540                                 sector = bio_end_sector(bi);
3541                 }
3542                 if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
3543                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3544                                 sh->overwrite_disks++;
3545         }
3546
3547         pr_debug("added bi b#%llu to stripe s#%llu, disk %d, logical %llu\n",
3548                  (*bip)->bi_iter.bi_sector, sh->sector, dd_idx,
3549                  sh->dev[dd_idx].sector);
3550
3551         if (conf->mddev->bitmap && firstwrite) {
3552                 /* Cannot hold spinlock over bitmap_startwrite,
3553                  * but must ensure this isn't added to a batch until
3554                  * we have added to the bitmap and set bm_seq.
3555                  * So set STRIPE_BITMAP_PENDING to prevent
3556                  * batching.
3557                  * If multiple __add_stripe_bio() calls race here they
3558                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3559                  * to complete "bitmap_startwrite" gets to set
3560                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3561                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3562                  * any more.
3563                  */
3564                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3565                 spin_unlock_irq(&sh->stripe_lock);
3566                 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3567                                      RAID5_STRIPE_SECTORS(conf), 0);
3568                 spin_lock_irq(&sh->stripe_lock);
3569                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3570                 if (!sh->batch_head) {
3571                         sh->bm_seq = conf->seq_flush+1;
3572                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3573                 }
3574         }
3575 }
3576
3577 /*
3578  * Each stripe/dev can have one or more bios attached.
3579  * toread/towrite point to the first in a chain.
3580  * The bi_next chain must be in order.
3581  */
3582 static bool add_stripe_bio(struct stripe_head *sh, struct bio *bi,
3583                            int dd_idx, int forwrite, int previous)
3584 {
3585         spin_lock_irq(&sh->stripe_lock);
3586
3587         if (stripe_bio_overlaps(sh, bi, dd_idx, forwrite)) {
3588                 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3589                 spin_unlock_irq(&sh->stripe_lock);
3590                 return false;
3591         }
3592
3593         __add_stripe_bio(sh, bi, dd_idx, forwrite, previous);
3594         spin_unlock_irq(&sh->stripe_lock);
3595         return true;
3596 }
3597
3598 static void end_reshape(struct r5conf *conf);
3599
3600 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3601                             struct stripe_head *sh)
3602 {
3603         int sectors_per_chunk =
3604                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3605         int dd_idx;
3606         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3607         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3608
3609         raid5_compute_sector(conf,
3610                              stripe * (disks - conf->max_degraded)
3611                              *sectors_per_chunk + chunk_offset,
3612                              previous,
3613                              &dd_idx, sh);
3614 }
3615
3616 static void
3617 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3618                      struct stripe_head_state *s, int disks)
3619 {
3620         int i;
3621         BUG_ON(sh->batch_head);
3622         for (i = disks; i--; ) {
3623                 struct bio *bi;
3624                 int bitmap_end = 0;
3625
3626                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3627                         struct md_rdev *rdev = conf->disks[i].rdev;
3628
3629                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3630                             !test_bit(Faulty, &rdev->flags))
3631                                 atomic_inc(&rdev->nr_pending);
3632                         else
3633                                 rdev = NULL;
3634                         if (rdev) {
3635                                 if (!rdev_set_badblocks(
3636                                             rdev,
3637                                             sh->sector,
3638                                             RAID5_STRIPE_SECTORS(conf), 0))
3639                                         md_error(conf->mddev, rdev);
3640                                 rdev_dec_pending(rdev, conf->mddev);
3641                         }
3642                 }
3643                 spin_lock_irq(&sh->stripe_lock);
3644                 /* fail all writes first */
3645                 bi = sh->dev[i].towrite;
3646                 sh->dev[i].towrite = NULL;
3647                 sh->overwrite_disks = 0;
3648                 spin_unlock_irq(&sh->stripe_lock);
3649                 if (bi)
3650                         bitmap_end = 1;
3651
3652                 log_stripe_write_finished(sh);
3653
3654                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3655                         wake_up(&conf->wait_for_overlap);
3656
3657                 while (bi && bi->bi_iter.bi_sector <
3658                         sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3659                         struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
3660
3661                         md_write_end(conf->mddev);
3662                         bio_io_error(bi);
3663                         bi = nextbi;
3664                 }
3665                 if (bitmap_end)
3666                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3667                                            RAID5_STRIPE_SECTORS(conf), 0, 0);
3668                 bitmap_end = 0;
3669                 /* and fail all 'written' */
3670                 bi = sh->dev[i].written;
3671                 sh->dev[i].written = NULL;
3672                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3673                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3674                         sh->dev[i].page = sh->dev[i].orig_page;
3675                 }
3676
3677                 if (bi) bitmap_end = 1;
3678                 while (bi && bi->bi_iter.bi_sector <
3679                        sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3680                         struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
3681
3682                         md_write_end(conf->mddev);
3683                         bio_io_error(bi);
3684                         bi = bi2;
3685                 }
3686
3687                 /* fail any reads if this device is non-operational and
3688                  * the data has not reached the cache yet.
3689                  */
3690                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3691                     s->failed > conf->max_degraded &&
3692                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3693                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3694                         spin_lock_irq(&sh->stripe_lock);
3695                         bi = sh->dev[i].toread;
3696                         sh->dev[i].toread = NULL;
3697                         spin_unlock_irq(&sh->stripe_lock);
3698                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3699                                 wake_up(&conf->wait_for_overlap);
3700                         if (bi)
3701                                 s->to_read--;
3702                         while (bi && bi->bi_iter.bi_sector <
3703                                sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3704                                 struct bio *nextbi =
3705                                         r5_next_bio(conf, bi, sh->dev[i].sector);
3706
3707                                 bio_io_error(bi);
3708                                 bi = nextbi;
3709                         }
3710                 }
3711                 if (bitmap_end)
3712                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3713                                            RAID5_STRIPE_SECTORS(conf), 0, 0);
3714                 /* If we were in the middle of a write the parity block might
3715                  * still be locked - so just clear all R5_LOCKED flags
3716                  */
3717                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3718         }
3719         s->to_write = 0;
3720         s->written = 0;
3721
3722         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3723                 if (atomic_dec_and_test(&conf->pending_full_writes))
3724                         md_wakeup_thread(conf->mddev->thread);
3725 }
3726
3727 static void
3728 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3729                    struct stripe_head_state *s)
3730 {
3731         int abort = 0;
3732         int i;
3733
3734         BUG_ON(sh->batch_head);
3735         clear_bit(STRIPE_SYNCING, &sh->state);
3736         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3737                 wake_up(&conf->wait_for_overlap);
3738         s->syncing = 0;
3739         s->replacing = 0;
3740         /* There is nothing more to do for sync/check/repair.
3741          * Don't even need to abort as that is handled elsewhere
3742          * if needed, and not always wanted e.g. if there is a known
3743          * bad block here.
3744          * For recover/replace we need to record a bad block on all
3745          * non-sync devices, or abort the recovery
3746          */
3747         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3748                 /* During recovery devices cannot be removed, so
3749                  * locking and refcounting of rdevs is not needed
3750                  */
3751                 for (i = 0; i < conf->raid_disks; i++) {
3752                         struct md_rdev *rdev = conf->disks[i].rdev;
3753
3754                         if (rdev
3755                             && !test_bit(Faulty, &rdev->flags)
3756                             && !test_bit(In_sync, &rdev->flags)
3757                             && !rdev_set_badblocks(rdev, sh->sector,
3758                                                    RAID5_STRIPE_SECTORS(conf), 0))
3759                                 abort = 1;
3760                         rdev = conf->disks[i].replacement;
3761
3762                         if (rdev
3763                             && !test_bit(Faulty, &rdev->flags)
3764                             && !test_bit(In_sync, &rdev->flags)
3765                             && !rdev_set_badblocks(rdev, sh->sector,
3766                                                    RAID5_STRIPE_SECTORS(conf), 0))
3767                                 abort = 1;
3768                 }
3769                 if (abort)
3770                         conf->recovery_disabled =
3771                                 conf->mddev->recovery_disabled;
3772         }
3773         md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
3774 }
3775
3776 static int want_replace(struct stripe_head *sh, int disk_idx)
3777 {
3778         struct md_rdev *rdev;
3779         int rv = 0;
3780
3781         rdev = sh->raid_conf->disks[disk_idx].replacement;
3782         if (rdev
3783             && !test_bit(Faulty, &rdev->flags)
3784             && !test_bit(In_sync, &rdev->flags)
3785             && (rdev->recovery_offset <= sh->sector
3786                 || rdev->mddev->recovery_cp <= sh->sector))
3787                 rv = 1;
3788         return rv;
3789 }
3790
3791 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3792                            int disk_idx, int disks)
3793 {
3794         struct r5dev *dev = &sh->dev[disk_idx];
3795         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3796                                   &sh->dev[s->failed_num[1]] };
3797         int i;
3798         bool force_rcw = (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW);
3799
3800
3801         if (test_bit(R5_LOCKED, &dev->flags) ||
3802             test_bit(R5_UPTODATE, &dev->flags))
3803                 /* No point reading this as we already have it or have
3804                  * decided to get it.
3805                  */
3806                 return 0;
3807
3808         if (dev->toread ||
3809             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3810                 /* We need this block to directly satisfy a request */
3811                 return 1;
3812
3813         if (s->syncing || s->expanding ||
3814             (s->replacing && want_replace(sh, disk_idx)))
3815                 /* When syncing, or expanding we read everything.
3816                  * When replacing, we need the replaced block.
3817                  */
3818                 return 1;
3819
3820         if ((s->failed >= 1 && fdev[0]->toread) ||
3821             (s->failed >= 2 && fdev[1]->toread))
3822                 /* If we want to read from a failed device, then
3823                  * we need to actually read every other device.
3824                  */
3825                 return 1;
3826
3827         /* Sometimes neither read-modify-write nor reconstruct-write
3828          * cycles can work.  In those cases we read every block we
3829          * can.  Then the parity-update is certain to have enough to
3830          * work with.
3831          * This can only be a problem when we need to write something,
3832          * and some device has failed.  If either of those tests
3833          * fail we need look no further.
3834          */
3835         if (!s->failed || !s->to_write)
3836                 return 0;
3837
3838         if (test_bit(R5_Insync, &dev->flags) &&
3839             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3840                 /* Pre-reads at not permitted until after short delay
3841                  * to gather multiple requests.  However if this
3842                  * device is no Insync, the block could only be computed
3843                  * and there is no need to delay that.
3844                  */
3845                 return 0;
3846
3847         for (i = 0; i < s->failed && i < 2; i++) {
3848                 if (fdev[i]->towrite &&
3849                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3850                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3851                         /* If we have a partial write to a failed
3852                          * device, then we will need to reconstruct
3853                          * the content of that device, so all other
3854                          * devices must be read.
3855                          */
3856                         return 1;
3857
3858                 if (s->failed >= 2 &&
3859                     (fdev[i]->towrite ||
3860                      s->failed_num[i] == sh->pd_idx ||
3861                      s->failed_num[i] == sh->qd_idx) &&
3862                     !test_bit(R5_UPTODATE, &fdev[i]->flags))
3863                         /* In max degraded raid6, If the failed disk is P, Q,
3864                          * or we want to read the failed disk, we need to do
3865                          * reconstruct-write.
3866                          */
3867                         force_rcw = true;
3868         }
3869
3870         /* If we are forced to do a reconstruct-write, because parity
3871          * cannot be trusted and we are currently recovering it, there
3872          * is extra need to be careful.
3873          * If one of the devices that we would need to read, because
3874          * it is not being overwritten (and maybe not written at all)
3875          * is missing/faulty, then we need to read everything we can.
3876          */
3877         if (!force_rcw &&
3878             sh->sector < sh->raid_conf->mddev->recovery_cp)
3879                 /* reconstruct-write isn't being forced */
3880                 return 0;
3881         for (i = 0; i < s->failed && i < 2; i++) {
3882                 if (s->failed_num[i] != sh->pd_idx &&
3883                     s->failed_num[i] != sh->qd_idx &&
3884                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3885                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3886                         return 1;
3887         }
3888
3889         return 0;
3890 }
3891
3892 /* fetch_block - checks the given member device to see if its data needs
3893  * to be read or computed to satisfy a request.
3894  *
3895  * Returns 1 when no more member devices need to be checked, otherwise returns
3896  * 0 to tell the loop in handle_stripe_fill to continue
3897  */
3898 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3899                        int disk_idx, int disks)
3900 {
3901         struct r5dev *dev = &sh->dev[disk_idx];
3902
3903         /* is the data in this block needed, and can we get it? */
3904         if (need_this_block(sh, s, disk_idx, disks)) {
3905                 /* we would like to get this block, possibly by computing it,
3906                  * otherwise read it if the backing disk is insync
3907                  */
3908                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3909                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3910                 BUG_ON(sh->batch_head);
3911
3912                 /*
3913                  * In the raid6 case if the only non-uptodate disk is P
3914                  * then we already trusted P to compute the other failed
3915                  * drives. It is safe to compute rather than re-read P.
3916                  * In other cases we only compute blocks from failed
3917                  * devices, otherwise check/repair might fail to detect
3918                  * a real inconsistency.
3919                  */
3920
3921                 if ((s->uptodate == disks - 1) &&
3922                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3923                     (s->failed && (disk_idx == s->failed_num[0] ||
3924                                    disk_idx == s->failed_num[1])))) {
3925                         /* have disk failed, and we're requested to fetch it;
3926                          * do compute it
3927                          */
3928                         pr_debug("Computing stripe %llu block %d\n",
3929                                (unsigned long long)sh->sector, disk_idx);
3930                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3931                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3932                         set_bit(R5_Wantcompute, &dev->flags);
3933                         sh->ops.target = disk_idx;
3934                         sh->ops.target2 = -1; /* no 2nd target */
3935                         s->req_compute = 1;
3936                         /* Careful: from this point on 'uptodate' is in the eye
3937                          * of raid_run_ops which services 'compute' operations
3938                          * before writes. R5_Wantcompute flags a block that will
3939                          * be R5_UPTODATE by the time it is needed for a
3940                          * subsequent operation.
3941                          */
3942                         s->uptodate++;
3943                         return 1;
3944                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3945                         /* Computing 2-failure is *very* expensive; only
3946                          * do it if failed >= 2
3947                          */
3948                         int other;
3949                         for (other = disks; other--; ) {
3950                                 if (other == disk_idx)
3951                                         continue;
3952                                 if (!test_bit(R5_UPTODATE,
3953                                       &sh->dev[other].flags))
3954                                         break;
3955                         }
3956                         BUG_ON(other < 0);
3957                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3958                                (unsigned long long)sh->sector,
3959                                disk_idx, other);
3960                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3961                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3962                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3963                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3964                         sh->ops.target = disk_idx;
3965                         sh->ops.target2 = other;
3966                         s->uptodate += 2;
3967                         s->req_compute = 1;
3968                         return 1;
3969                 } else if (test_bit(R5_Insync, &dev->flags)) {
3970                         set_bit(R5_LOCKED, &dev->flags);
3971                         set_bit(R5_Wantread, &dev->flags);
3972                         s->locked++;
3973                         pr_debug("Reading block %d (sync=%d)\n",
3974                                 disk_idx, s->syncing);
3975                 }
3976         }
3977
3978         return 0;
3979 }
3980
3981 /*
3982  * handle_stripe_fill - read or compute data to satisfy pending requests.
3983  */
3984 static void handle_stripe_fill(struct stripe_head *sh,
3985                                struct stripe_head_state *s,
3986                                int disks)
3987 {
3988         int i;
3989
3990         /* look for blocks to read/compute, skip this if a compute
3991          * is already in flight, or if the stripe contents are in the
3992          * midst of changing due to a write
3993          */
3994         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3995             !sh->reconstruct_state) {
3996
3997                 /*
3998                  * For degraded stripe with data in journal, do not handle
3999                  * read requests yet, instead, flush the stripe to raid
4000                  * disks first, this avoids handling complex rmw of write
4001                  * back cache (prexor with orig_page, and then xor with
4002                  * page) in the read path
4003                  */
4004                 if (s->to_read && s->injournal && s->failed) {
4005                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
4006                                 r5c_make_stripe_write_out(sh);
4007                         goto out;
4008                 }
4009
4010                 for (i = disks; i--; )
4011                         if (fetch_block(sh, s, i, disks))
4012                                 break;
4013         }
4014 out:
4015         set_bit(STRIPE_HANDLE, &sh->state);
4016 }
4017
4018 static void break_stripe_batch_list(struct stripe_head *head_sh,
4019                                     unsigned long handle_flags);
4020 /* handle_stripe_clean_event
4021  * any written block on an uptodate or failed drive can be returned.
4022  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
4023  * never LOCKED, so we don't need to test 'failed' directly.
4024  */
4025 static void handle_stripe_clean_event(struct r5conf *conf,
4026         struct stripe_head *sh, int disks)
4027 {
4028         int i;
4029         struct r5dev *dev;
4030         int discard_pending = 0;
4031         struct stripe_head *head_sh = sh;
4032         bool do_endio = false;
4033
4034         for (i = disks; i--; )
4035                 if (sh->dev[i].written) {
4036                         dev = &sh->dev[i];
4037                         if (!test_bit(R5_LOCKED, &dev->flags) &&
4038                             (test_bit(R5_UPTODATE, &dev->flags) ||
4039                              test_bit(R5_Discard, &dev->flags) ||
4040                              test_bit(R5_SkipCopy, &dev->flags))) {
4041                                 /* We can return any write requests */
4042                                 struct bio *wbi, *wbi2;
4043                                 pr_debug("Return write for disc %d\n", i);
4044                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
4045                                         clear_bit(R5_UPTODATE, &dev->flags);
4046                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
4047                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
4048                                 }
4049                                 do_endio = true;
4050
4051 returnbi:
4052                                 dev->page = dev->orig_page;
4053                                 wbi = dev->written;
4054                                 dev->written = NULL;
4055                                 while (wbi && wbi->bi_iter.bi_sector <
4056                                         dev->sector + RAID5_STRIPE_SECTORS(conf)) {
4057                                         wbi2 = r5_next_bio(conf, wbi, dev->sector);
4058                                         md_write_end(conf->mddev);
4059                                         bio_endio(wbi);
4060                                         wbi = wbi2;
4061                                 }
4062                                 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
4063                                                    RAID5_STRIPE_SECTORS(conf),
4064                                                    !test_bit(STRIPE_DEGRADED, &sh->state),
4065                                                    0);
4066                                 if (head_sh->batch_head) {
4067                                         sh = list_first_entry(&sh->batch_list,
4068                                                               struct stripe_head,
4069                                                               batch_list);
4070                                         if (sh != head_sh) {
4071                                                 dev = &sh->dev[i];
4072                                                 goto returnbi;
4073                                         }
4074                                 }
4075                                 sh = head_sh;
4076                                 dev = &sh->dev[i];
4077                         } else if (test_bit(R5_Discard, &dev->flags))
4078                                 discard_pending = 1;
4079                 }
4080
4081         log_stripe_write_finished(sh);
4082
4083         if (!discard_pending &&
4084             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
4085                 int hash;
4086                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
4087                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4088                 if (sh->qd_idx >= 0) {
4089                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
4090                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
4091                 }
4092                 /* now that discard is done we can proceed with any sync */
4093                 clear_bit(STRIPE_DISCARD, &sh->state);
4094                 /*
4095                  * SCSI discard will change some bio fields and the stripe has
4096                  * no updated data, so remove it from hash list and the stripe
4097                  * will be reinitialized
4098                  */
4099 unhash:
4100                 hash = sh->hash_lock_index;
4101                 spin_lock_irq(conf->hash_locks + hash);
4102                 remove_hash(sh);
4103                 spin_unlock_irq(conf->hash_locks + hash);
4104                 if (head_sh->batch_head) {
4105                         sh = list_first_entry(&sh->batch_list,
4106                                               struct stripe_head, batch_list);
4107                         if (sh != head_sh)
4108                                         goto unhash;
4109                 }
4110                 sh = head_sh;
4111
4112                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
4113                         set_bit(STRIPE_HANDLE, &sh->state);
4114
4115         }
4116
4117         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
4118                 if (atomic_dec_and_test(&conf->pending_full_writes))
4119                         md_wakeup_thread(conf->mddev->thread);
4120
4121         if (head_sh->batch_head && do_endio)
4122                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
4123 }
4124
4125 /*
4126  * For RMW in write back cache, we need extra page in prexor to store the
4127  * old data. This page is stored in dev->orig_page.
4128  *
4129  * This function checks whether we have data for prexor. The exact logic
4130  * is:
4131  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
4132  */
4133 static inline bool uptodate_for_rmw(struct r5dev *dev)
4134 {
4135         return (test_bit(R5_UPTODATE, &dev->flags)) &&
4136                 (!test_bit(R5_InJournal, &dev->flags) ||
4137                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
4138 }
4139
4140 static int handle_stripe_dirtying(struct r5conf *conf,
4141                                   struct stripe_head *sh,
4142                                   struct stripe_head_state *s,
4143                                   int disks)
4144 {
4145         int rmw = 0, rcw = 0, i;
4146         sector_t recovery_cp = conf->mddev->recovery_cp;
4147
4148         /* Check whether resync is now happening or should start.
4149          * If yes, then the array is dirty (after unclean shutdown or
4150          * initial creation), so parity in some stripes might be inconsistent.
4151          * In this case, we need to always do reconstruct-write, to ensure
4152          * that in case of drive failure or read-error correction, we
4153          * generate correct data from the parity.
4154          */
4155         if (conf->rmw_level == PARITY_DISABLE_RMW ||
4156             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
4157              s->failed == 0)) {
4158                 /* Calculate the real rcw later - for now make it
4159                  * look like rcw is cheaper
4160                  */
4161                 rcw = 1; rmw = 2;
4162                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
4163                          conf->rmw_level, (unsigned long long)recovery_cp,
4164                          (unsigned long long)sh->sector);
4165         } else for (i = disks; i--; ) {
4166                 /* would I have to read this buffer for read_modify_write */
4167                 struct r5dev *dev = &sh->dev[i];
4168                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4169                      i == sh->pd_idx || i == sh->qd_idx ||
4170                      test_bit(R5_InJournal, &dev->flags)) &&
4171                     !test_bit(R5_LOCKED, &dev->flags) &&
4172                     !(uptodate_for_rmw(dev) ||
4173                       test_bit(R5_Wantcompute, &dev->flags))) {
4174                         if (test_bit(R5_Insync, &dev->flags))
4175                                 rmw++;
4176                         else
4177                                 rmw += 2*disks;  /* cannot read it */
4178                 }
4179                 /* Would I have to read this buffer for reconstruct_write */
4180                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4181                     i != sh->pd_idx && i != sh->qd_idx &&
4182                     !test_bit(R5_LOCKED, &dev->flags) &&
4183                     !(test_bit(R5_UPTODATE, &dev->flags) ||
4184                       test_bit(R5_Wantcompute, &dev->flags))) {
4185                         if (test_bit(R5_Insync, &dev->flags))
4186                                 rcw++;
4187                         else
4188                                 rcw += 2*disks;
4189                 }
4190         }
4191
4192         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
4193                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
4194         set_bit(STRIPE_HANDLE, &sh->state);
4195         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
4196                 /* prefer read-modify-write, but need to get some data */
4197                 mddev_add_trace_msg(conf->mddev, "raid5 rmw %llu %d",
4198                                 sh->sector, rmw);
4199
4200                 for (i = disks; i--; ) {
4201                         struct r5dev *dev = &sh->dev[i];
4202                         if (test_bit(R5_InJournal, &dev->flags) &&
4203                             dev->page == dev->orig_page &&
4204                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
4205                                 /* alloc page for prexor */
4206                                 struct page *p = alloc_page(GFP_NOIO);
4207
4208                                 if (p) {
4209                                         dev->orig_page = p;
4210                                         continue;
4211                                 }
4212
4213                                 /*
4214                                  * alloc_page() failed, try use
4215                                  * disk_info->extra_page
4216                                  */
4217                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
4218                                                       &conf->cache_state)) {
4219                                         r5c_use_extra_page(sh);
4220                                         break;
4221                                 }
4222
4223                                 /* extra_page in use, add to delayed_list */
4224                                 set_bit(STRIPE_DELAYED, &sh->state);
4225                                 s->waiting_extra_page = 1;
4226                                 return -EAGAIN;
4227                         }
4228                 }
4229
4230                 for (i = disks; i--; ) {
4231                         struct r5dev *dev = &sh->dev[i];
4232                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4233                              i == sh->pd_idx || i == sh->qd_idx ||
4234                              test_bit(R5_InJournal, &dev->flags)) &&
4235                             !test_bit(R5_LOCKED, &dev->flags) &&
4236                             !(uptodate_for_rmw(dev) ||
4237                               test_bit(R5_Wantcompute, &dev->flags)) &&
4238                             test_bit(R5_Insync, &dev->flags)) {
4239                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
4240                                              &sh->state)) {
4241                                         pr_debug("Read_old block %d for r-m-w\n",
4242                                                  i);
4243                                         set_bit(R5_LOCKED, &dev->flags);
4244                                         set_bit(R5_Wantread, &dev->flags);
4245                                         s->locked++;
4246                                 } else
4247                                         set_bit(STRIPE_DELAYED, &sh->state);
4248                         }
4249                 }
4250         }
4251         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
4252                 /* want reconstruct write, but need to get some data */
4253                 int qread =0;
4254                 rcw = 0;
4255                 for (i = disks; i--; ) {
4256                         struct r5dev *dev = &sh->dev[i];
4257                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4258                             i != sh->pd_idx && i != sh->qd_idx &&
4259                             !test_bit(R5_LOCKED, &dev->flags) &&
4260                             !(test_bit(R5_UPTODATE, &dev->flags) ||
4261                               test_bit(R5_Wantcompute, &dev->flags))) {
4262                                 rcw++;
4263                                 if (test_bit(R5_Insync, &dev->flags) &&
4264                                     test_bit(STRIPE_PREREAD_ACTIVE,
4265                                              &sh->state)) {
4266                                         pr_debug("Read_old block "
4267                                                 "%d for Reconstruct\n", i);
4268                                         set_bit(R5_LOCKED, &dev->flags);
4269                                         set_bit(R5_Wantread, &dev->flags);
4270                                         s->locked++;
4271                                         qread++;
4272                                 } else
4273                                         set_bit(STRIPE_DELAYED, &sh->state);
4274                         }
4275                 }
4276                 if (rcw && !mddev_is_dm(conf->mddev))
4277                         blk_add_trace_msg(conf->mddev->gendisk->queue,
4278                                 "raid5 rcw %llu %d %d %d",
4279                                 (unsigned long long)sh->sector, rcw, qread,
4280                                 test_bit(STRIPE_DELAYED, &sh->state));
4281         }
4282
4283         if (rcw > disks && rmw > disks &&
4284             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4285                 set_bit(STRIPE_DELAYED, &sh->state);
4286
4287         /* now if nothing is locked, and if we have enough data,
4288          * we can start a write request
4289          */
4290         /* since handle_stripe can be called at any time we need to handle the
4291          * case where a compute block operation has been submitted and then a
4292          * subsequent call wants to start a write request.  raid_run_ops only
4293          * handles the case where compute block and reconstruct are requested
4294          * simultaneously.  If this is not the case then new writes need to be
4295          * held off until the compute completes.
4296          */
4297         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4298             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4299              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4300                 schedule_reconstruction(sh, s, rcw == 0, 0);
4301         return 0;
4302 }
4303
4304 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4305                                 struct stripe_head_state *s, int disks)
4306 {
4307         struct r5dev *dev = NULL;
4308
4309         BUG_ON(sh->batch_head);
4310         set_bit(STRIPE_HANDLE, &sh->state);
4311
4312         switch (sh->check_state) {
4313         case check_state_idle:
4314                 /* start a new check operation if there are no failures */
4315                 if (s->failed == 0) {
4316                         BUG_ON(s->uptodate != disks);
4317                         sh->check_state = check_state_run;
4318                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4319                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4320                         s->uptodate--;
4321                         break;
4322                 }
4323                 dev = &sh->dev[s->failed_num[0]];
4324                 fallthrough;
4325         case check_state_compute_result:
4326                 sh->check_state = check_state_idle;
4327                 if (!dev)
4328                         dev = &sh->dev[sh->pd_idx];
4329
4330                 /* check that a write has not made the stripe insync */
4331                 if (test_bit(STRIPE_INSYNC, &sh->state))
4332                         break;
4333
4334                 /* either failed parity check, or recovery is happening */
4335                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4336                 BUG_ON(s->uptodate != disks);
4337
4338                 set_bit(R5_LOCKED, &dev->flags);
4339                 s->locked++;
4340                 set_bit(R5_Wantwrite, &dev->flags);
4341
4342                 clear_bit(STRIPE_DEGRADED, &sh->state);
4343                 set_bit(STRIPE_INSYNC, &sh->state);
4344                 break;
4345         case check_state_run:
4346                 break; /* we will be called again upon completion */
4347         case check_state_check_result:
4348                 sh->check_state = check_state_idle;
4349
4350                 /* if a failure occurred during the check operation, leave
4351                  * STRIPE_INSYNC not set and let the stripe be handled again
4352                  */
4353                 if (s->failed)
4354                         break;
4355
4356                 /* handle a successful check operation, if parity is correct
4357                  * we are done.  Otherwise update the mismatch count and repair
4358                  * parity if !MD_RECOVERY_CHECK
4359                  */
4360                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4361                         /* parity is correct (on disc,
4362                          * not in buffer any more)
4363                          */
4364                         set_bit(STRIPE_INSYNC, &sh->state);
4365                 else {
4366                         atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4367                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4368                                 /* don't try to repair!! */
4369                                 set_bit(STRIPE_INSYNC, &sh->state);
4370                                 pr_warn_ratelimited("%s: mismatch sector in range "
4371                                                     "%llu-%llu\n", mdname(conf->mddev),
4372                                                     (unsigned long long) sh->sector,
4373                                                     (unsigned long long) sh->sector +
4374                                                     RAID5_STRIPE_SECTORS(conf));
4375                         } else {
4376                                 sh->check_state = check_state_compute_run;
4377                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4378                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4379                                 set_bit(R5_Wantcompute,
4380                                         &sh->dev[sh->pd_idx].flags);
4381                                 sh->ops.target = sh->pd_idx;
4382                                 sh->ops.target2 = -1;
4383                                 s->uptodate++;
4384                         }
4385                 }
4386                 break;
4387         case check_state_compute_run:
4388                 break;
4389         default:
4390                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4391                        __func__, sh->check_state,
4392                        (unsigned long long) sh->sector);
4393                 BUG();
4394         }
4395 }
4396
4397 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4398                                   struct stripe_head_state *s,
4399                                   int disks)
4400 {
4401         int pd_idx = sh->pd_idx;
4402         int qd_idx = sh->qd_idx;
4403         struct r5dev *dev;
4404
4405         BUG_ON(sh->batch_head);
4406         set_bit(STRIPE_HANDLE, &sh->state);
4407
4408         BUG_ON(s->failed > 2);
4409
4410         /* Want to check and possibly repair P and Q.
4411          * However there could be one 'failed' device, in which
4412          * case we can only check one of them, possibly using the
4413          * other to generate missing data
4414          */
4415
4416         switch (sh->check_state) {
4417         case check_state_idle:
4418                 /* start a new check operation if there are < 2 failures */
4419                 if (s->failed == s->q_failed) {
4420                         /* The only possible failed device holds Q, so it
4421                          * makes sense to check P (If anything else were failed,
4422                          * we would have used P to recreate it).
4423                          */
4424                         sh->check_state = check_state_run;
4425                 }
4426                 if (!s->q_failed && s->failed < 2) {
4427                         /* Q is not failed, and we didn't use it to generate
4428                          * anything, so it makes sense to check it
4429                          */
4430                         if (sh->check_state == check_state_run)
4431                                 sh->check_state = check_state_run_pq;
4432                         else
4433                                 sh->check_state = check_state_run_q;
4434                 }
4435
4436                 /* discard potentially stale zero_sum_result */
4437                 sh->ops.zero_sum_result = 0;
4438
4439                 if (sh->check_state == check_state_run) {
4440                         /* async_xor_zero_sum destroys the contents of P */
4441                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4442                         s->uptodate--;
4443                 }
4444                 if (sh->check_state >= check_state_run &&
4445                     sh->check_state <= check_state_run_pq) {
4446                         /* async_syndrome_zero_sum preserves P and Q, so
4447                          * no need to mark them !uptodate here
4448                          */
4449                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4450                         break;
4451                 }
4452
4453                 /* we have 2-disk failure */
4454                 BUG_ON(s->failed != 2);
4455                 fallthrough;
4456         case check_state_compute_result:
4457                 sh->check_state = check_state_idle;
4458
4459                 /* check that a write has not made the stripe insync */
4460                 if (test_bit(STRIPE_INSYNC, &sh->state))
4461                         break;
4462
4463                 /* now write out any block on a failed drive,
4464                  * or P or Q if they were recomputed
4465                  */
4466                 dev = NULL;
4467                 if (s->failed == 2) {
4468                         dev = &sh->dev[s->failed_num[1]];
4469                         s->locked++;
4470                         set_bit(R5_LOCKED, &dev->flags);
4471                         set_bit(R5_Wantwrite, &dev->flags);
4472                 }
4473                 if (s->failed >= 1) {
4474                         dev = &sh->dev[s->failed_num[0]];
4475                         s->locked++;
4476                         set_bit(R5_LOCKED, &dev->flags);
4477                         set_bit(R5_Wantwrite, &dev->flags);
4478                 }
4479                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4480                         dev = &sh->dev[pd_idx];
4481                         s->locked++;
4482                         set_bit(R5_LOCKED, &dev->flags);
4483                         set_bit(R5_Wantwrite, &dev->flags);
4484                 }
4485                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4486                         dev = &sh->dev[qd_idx];
4487                         s->locked++;
4488                         set_bit(R5_LOCKED, &dev->flags);
4489                         set_bit(R5_Wantwrite, &dev->flags);
4490                 }
4491                 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4492                               "%s: disk%td not up to date\n",
4493                               mdname(conf->mddev),
4494                               dev - (struct r5dev *) &sh->dev)) {
4495                         clear_bit(R5_LOCKED, &dev->flags);
4496                         clear_bit(R5_Wantwrite, &dev->flags);
4497                         s->locked--;
4498                 }
4499                 clear_bit(STRIPE_DEGRADED, &sh->state);
4500
4501                 set_bit(STRIPE_INSYNC, &sh->state);
4502                 break;
4503         case check_state_run:
4504         case check_state_run_q:
4505         case check_state_run_pq:
4506                 break; /* we will be called again upon completion */
4507         case check_state_check_result:
4508                 sh->check_state = check_state_idle;
4509
4510                 /* handle a successful check operation, if parity is correct
4511                  * we are done.  Otherwise update the mismatch count and repair
4512                  * parity if !MD_RECOVERY_CHECK
4513                  */
4514                 if (sh->ops.zero_sum_result == 0) {
4515                         /* both parities are correct */
4516                         if (!s->failed)
4517                                 set_bit(STRIPE_INSYNC, &sh->state);
4518                         else {
4519                                 /* in contrast to the raid5 case we can validate
4520                                  * parity, but still have a failure to write
4521                                  * back
4522                                  */
4523                                 sh->check_state = check_state_compute_result;
4524                                 /* Returning at this point means that we may go
4525                                  * off and bring p and/or q uptodate again so
4526                                  * we make sure to check zero_sum_result again
4527                                  * to verify if p or q need writeback
4528                                  */
4529                         }
4530                 } else {
4531                         atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4532                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4533                                 /* don't try to repair!! */
4534                                 set_bit(STRIPE_INSYNC, &sh->state);
4535                                 pr_warn_ratelimited("%s: mismatch sector in range "
4536                                                     "%llu-%llu\n", mdname(conf->mddev),
4537                                                     (unsigned long long) sh->sector,
4538                                                     (unsigned long long) sh->sector +
4539                                                     RAID5_STRIPE_SECTORS(conf));
4540                         } else {
4541                                 int *target = &sh->ops.target;
4542
4543                                 sh->ops.target = -1;
4544                                 sh->ops.target2 = -1;
4545                                 sh->check_state = check_state_compute_run;
4546                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4547                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4548                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4549                                         set_bit(R5_Wantcompute,
4550                                                 &sh->dev[pd_idx].flags);
4551                                         *target = pd_idx;
4552                                         target = &sh->ops.target2;
4553                                         s->uptodate++;
4554                                 }
4555                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4556                                         set_bit(R5_Wantcompute,
4557                                                 &sh->dev[qd_idx].flags);
4558                                         *target = qd_idx;
4559                                         s->uptodate++;
4560                                 }
4561                         }
4562                 }
4563                 break;
4564         case check_state_compute_run:
4565                 break;
4566         default:
4567                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4568                         __func__, sh->check_state,
4569                         (unsigned long long) sh->sector);
4570                 BUG();
4571         }
4572 }
4573
4574 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4575 {
4576         int i;
4577
4578         /* We have read all the blocks in this stripe and now we need to
4579          * copy some of them into a target stripe for expand.
4580          */
4581         struct dma_async_tx_descriptor *tx = NULL;
4582         BUG_ON(sh->batch_head);
4583         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4584         for (i = 0; i < sh->disks; i++)
4585                 if (i != sh->pd_idx && i != sh->qd_idx) {
4586                         int dd_idx, j;
4587                         struct stripe_head *sh2;
4588                         struct async_submit_ctl submit;
4589
4590                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4591                         sector_t s = raid5_compute_sector(conf, bn, 0,
4592                                                           &dd_idx, NULL);
4593                         sh2 = raid5_get_active_stripe(conf, NULL, s,
4594                                 R5_GAS_NOBLOCK | R5_GAS_NOQUIESCE);
4595                         if (sh2 == NULL)
4596                                 /* so far only the early blocks of this stripe
4597                                  * have been requested.  When later blocks
4598                                  * get requested, we will try again
4599                                  */
4600                                 continue;
4601                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4602                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4603                                 /* must have already done this block */
4604                                 raid5_release_stripe(sh2);
4605                                 continue;
4606                         }
4607
4608                         /* place all the copies on one channel */
4609                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4610                         tx = async_memcpy(sh2->dev[dd_idx].page,
4611                                           sh->dev[i].page, sh2->dev[dd_idx].offset,
4612                                           sh->dev[i].offset, RAID5_STRIPE_SIZE(conf),
4613                                           &submit);
4614
4615                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4616                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4617                         for (j = 0; j < conf->raid_disks; j++)
4618                                 if (j != sh2->pd_idx &&
4619                                     j != sh2->qd_idx &&
4620                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4621                                         break;
4622                         if (j == conf->raid_disks) {
4623                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4624                                 set_bit(STRIPE_HANDLE, &sh2->state);
4625                         }
4626                         raid5_release_stripe(sh2);
4627
4628                 }
4629         /* done submitting copies, wait for them to complete */
4630         async_tx_quiesce(&tx);
4631 }
4632
4633 /*
4634  * handle_stripe - do things to a stripe.
4635  *
4636  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4637  * state of various bits to see what needs to be done.
4638  * Possible results:
4639  *    return some read requests which now have data
4640  *    return some write requests which are safely on storage
4641  *    schedule a read on some buffers
4642  *    schedule a write of some buffers
4643  *    return confirmation of parity correctness
4644  *
4645  */
4646
4647 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4648 {
4649         struct r5conf *conf = sh->raid_conf;
4650         int disks = sh->disks;
4651         struct r5dev *dev;
4652         int i;
4653         int do_recovery = 0;
4654
4655         memset(s, 0, sizeof(*s));
4656
4657         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4658         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4659         s->failed_num[0] = -1;
4660         s->failed_num[1] = -1;
4661         s->log_failed = r5l_log_disk_error(conf);
4662
4663         /* Now to look around and see what can be done */
4664         for (i=disks; i--; ) {
4665                 struct md_rdev *rdev;
4666                 int is_bad = 0;
4667
4668                 dev = &sh->dev[i];
4669
4670                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4671                          i, dev->flags,
4672                          dev->toread, dev->towrite, dev->written);
4673                 /* maybe we can reply to a read
4674                  *
4675                  * new wantfill requests are only permitted while
4676                  * ops_complete_biofill is guaranteed to be inactive
4677                  */
4678                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4679                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4680                         set_bit(R5_Wantfill, &dev->flags);
4681
4682                 /* now count some things */
4683                 if (test_bit(R5_LOCKED, &dev->flags))
4684                         s->locked++;
4685                 if (test_bit(R5_UPTODATE, &dev->flags))
4686                         s->uptodate++;
4687                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4688                         s->compute++;
4689                         BUG_ON(s->compute > 2);
4690                 }
4691
4692                 if (test_bit(R5_Wantfill, &dev->flags))
4693                         s->to_fill++;
4694                 else if (dev->toread)
4695                         s->to_read++;
4696                 if (dev->towrite) {
4697                         s->to_write++;
4698                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4699                                 s->non_overwrite++;
4700                 }
4701                 if (dev->written)
4702                         s->written++;
4703                 /* Prefer to use the replacement for reads, but only
4704                  * if it is recovered enough and has no bad blocks.
4705                  */
4706                 rdev = conf->disks[i].replacement;
4707                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4708                     rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
4709                     !rdev_has_badblock(rdev, sh->sector,
4710                                        RAID5_STRIPE_SECTORS(conf)))
4711                         set_bit(R5_ReadRepl, &dev->flags);
4712                 else {
4713                         if (rdev && !test_bit(Faulty, &rdev->flags))
4714                                 set_bit(R5_NeedReplace, &dev->flags);
4715                         else
4716                                 clear_bit(R5_NeedReplace, &dev->flags);
4717                         rdev = conf->disks[i].rdev;
4718                         clear_bit(R5_ReadRepl, &dev->flags);
4719                 }
4720                 if (rdev && test_bit(Faulty, &rdev->flags))
4721                         rdev = NULL;
4722                 if (rdev) {
4723                         is_bad = rdev_has_badblock(rdev, sh->sector,
4724                                                    RAID5_STRIPE_SECTORS(conf));
4725                         if (s->blocked_rdev == NULL
4726                             && (test_bit(Blocked, &rdev->flags)
4727                                 || is_bad < 0)) {
4728                                 if (is_bad < 0)
4729                                         set_bit(BlockedBadBlocks,
4730                                                 &rdev->flags);
4731                                 s->blocked_rdev = rdev;
4732                                 atomic_inc(&rdev->nr_pending);
4733                         }
4734                 }
4735                 clear_bit(R5_Insync, &dev->flags);
4736                 if (!rdev)
4737                         /* Not in-sync */;
4738                 else if (is_bad) {
4739                         /* also not in-sync */
4740                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4741                             test_bit(R5_UPTODATE, &dev->flags)) {
4742                                 /* treat as in-sync, but with a read error
4743                                  * which we can now try to correct
4744                                  */
4745                                 set_bit(R5_Insync, &dev->flags);
4746                                 set_bit(R5_ReadError, &dev->flags);
4747                         }
4748                 } else if (test_bit(In_sync, &rdev->flags))
4749                         set_bit(R5_Insync, &dev->flags);
4750                 else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <= rdev->recovery_offset)
4751                         /* in sync if before recovery_offset */
4752                         set_bit(R5_Insync, &dev->flags);
4753                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4754                          test_bit(R5_Expanded, &dev->flags))
4755                         /* If we've reshaped into here, we assume it is Insync.
4756                          * We will shortly update recovery_offset to make
4757                          * it official.
4758                          */
4759                         set_bit(R5_Insync, &dev->flags);
4760
4761                 if (test_bit(R5_WriteError, &dev->flags)) {
4762                         /* This flag does not apply to '.replacement'
4763                          * only to .rdev, so make sure to check that*/
4764                         struct md_rdev *rdev2 = conf->disks[i].rdev;
4765
4766                         if (rdev2 == rdev)
4767                                 clear_bit(R5_Insync, &dev->flags);
4768                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4769                                 s->handle_bad_blocks = 1;
4770                                 atomic_inc(&rdev2->nr_pending);
4771                         } else
4772                                 clear_bit(R5_WriteError, &dev->flags);
4773                 }
4774                 if (test_bit(R5_MadeGood, &dev->flags)) {
4775                         /* This flag does not apply to '.replacement'
4776                          * only to .rdev, so make sure to check that*/
4777                         struct md_rdev *rdev2 = conf->disks[i].rdev;
4778
4779                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4780                                 s->handle_bad_blocks = 1;
4781                                 atomic_inc(&rdev2->nr_pending);
4782                         } else
4783                                 clear_bit(R5_MadeGood, &dev->flags);
4784                 }
4785                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4786                         struct md_rdev *rdev2 = conf->disks[i].replacement;
4787
4788                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4789                                 s->handle_bad_blocks = 1;
4790                                 atomic_inc(&rdev2->nr_pending);
4791                         } else
4792                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4793                 }
4794                 if (!test_bit(R5_Insync, &dev->flags)) {
4795                         /* The ReadError flag will just be confusing now */
4796                         clear_bit(R5_ReadError, &dev->flags);
4797                         clear_bit(R5_ReWrite, &dev->flags);
4798                 }
4799                 if (test_bit(R5_ReadError, &dev->flags))
4800                         clear_bit(R5_Insync, &dev->flags);
4801                 if (!test_bit(R5_Insync, &dev->flags)) {
4802                         if (s->failed < 2)
4803                                 s->failed_num[s->failed] = i;
4804                         s->failed++;
4805                         if (rdev && !test_bit(Faulty, &rdev->flags))
4806                                 do_recovery = 1;
4807                         else if (!rdev) {
4808                                 rdev = conf->disks[i].replacement;
4809                                 if (rdev && !test_bit(Faulty, &rdev->flags))
4810                                         do_recovery = 1;
4811                         }
4812                 }
4813
4814                 if (test_bit(R5_InJournal, &dev->flags))
4815                         s->injournal++;
4816                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4817                         s->just_cached++;
4818         }
4819         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4820                 /* If there is a failed device being replaced,
4821                  *     we must be recovering.
4822                  * else if we are after recovery_cp, we must be syncing
4823                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4824                  * else we can only be replacing
4825                  * sync and recovery both need to read all devices, and so
4826                  * use the same flag.
4827                  */
4828                 if (do_recovery ||
4829                     sh->sector >= conf->mddev->recovery_cp ||
4830                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4831                         s->syncing = 1;
4832                 else
4833                         s->replacing = 1;
4834         }
4835 }
4836
4837 /*
4838  * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4839  * a head which can now be handled.
4840  */
4841 static int clear_batch_ready(struct stripe_head *sh)
4842 {
4843         struct stripe_head *tmp;
4844         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4845                 return (sh->batch_head && sh->batch_head != sh);
4846         spin_lock(&sh->stripe_lock);
4847         if (!sh->batch_head) {
4848                 spin_unlock(&sh->stripe_lock);
4849                 return 0;
4850         }
4851
4852         /*
4853          * this stripe could be added to a batch list before we check
4854          * BATCH_READY, skips it
4855          */
4856         if (sh->batch_head != sh) {
4857                 spin_unlock(&sh->stripe_lock);
4858                 return 1;
4859         }
4860         spin_lock(&sh->batch_lock);
4861         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4862                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4863         spin_unlock(&sh->batch_lock);
4864         spin_unlock(&sh->stripe_lock);
4865
4866         /*
4867          * BATCH_READY is cleared, no new stripes can be added.
4868          * batch_list can be accessed without lock
4869          */
4870         return 0;
4871 }
4872
4873 static void break_stripe_batch_list(struct stripe_head *head_sh,
4874                                     unsigned long handle_flags)
4875 {
4876         struct stripe_head *sh, *next;
4877         int i;
4878         int do_wakeup = 0;
4879
4880         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4881
4882                 list_del_init(&sh->batch_list);
4883
4884                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4885                                           (1 << STRIPE_SYNCING) |
4886                                           (1 << STRIPE_REPLACED) |
4887                                           (1 << STRIPE_DELAYED) |
4888                                           (1 << STRIPE_BIT_DELAY) |
4889                                           (1 << STRIPE_FULL_WRITE) |
4890                                           (1 << STRIPE_BIOFILL_RUN) |
4891                                           (1 << STRIPE_COMPUTE_RUN)  |
4892                                           (1 << STRIPE_DISCARD) |
4893                                           (1 << STRIPE_BATCH_READY) |
4894                                           (1 << STRIPE_BATCH_ERR) |
4895                                           (1 << STRIPE_BITMAP_PENDING)),
4896                         "stripe state: %lx\n", sh->state);
4897                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4898                                               (1 << STRIPE_REPLACED)),
4899                         "head stripe state: %lx\n", head_sh->state);
4900
4901                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4902                                             (1 << STRIPE_PREREAD_ACTIVE) |
4903                                             (1 << STRIPE_DEGRADED) |
4904                                             (1 << STRIPE_ON_UNPLUG_LIST)),
4905                               head_sh->state & (1 << STRIPE_INSYNC));
4906
4907                 sh->check_state = head_sh->check_state;
4908                 sh->reconstruct_state = head_sh->reconstruct_state;
4909                 spin_lock_irq(&sh->stripe_lock);
4910                 sh->batch_head = NULL;
4911                 spin_unlock_irq(&sh->stripe_lock);
4912                 for (i = 0; i < sh->disks; i++) {
4913                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4914                                 do_wakeup = 1;
4915                         sh->dev[i].flags = head_sh->dev[i].flags &
4916                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4917                 }
4918                 if (handle_flags == 0 ||
4919                     sh->state & handle_flags)
4920                         set_bit(STRIPE_HANDLE, &sh->state);
4921                 raid5_release_stripe(sh);
4922         }
4923         spin_lock_irq(&head_sh->stripe_lock);
4924         head_sh->batch_head = NULL;
4925         spin_unlock_irq(&head_sh->stripe_lock);
4926         for (i = 0; i < head_sh->disks; i++)
4927                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4928                         do_wakeup = 1;
4929         if (head_sh->state & handle_flags)
4930                 set_bit(STRIPE_HANDLE, &head_sh->state);
4931
4932         if (do_wakeup)
4933                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4934 }
4935
4936 static void handle_stripe(struct stripe_head *sh)
4937 {
4938         struct stripe_head_state s;
4939         struct r5conf *conf = sh->raid_conf;
4940         int i;
4941         int prexor;
4942         int disks = sh->disks;
4943         struct r5dev *pdev, *qdev;
4944
4945         clear_bit(STRIPE_HANDLE, &sh->state);
4946
4947         /*
4948          * handle_stripe should not continue handle the batched stripe, only
4949          * the head of batch list or lone stripe can continue. Otherwise we
4950          * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
4951          * is set for the batched stripe.
4952          */
4953         if (clear_batch_ready(sh))
4954                 return;
4955
4956         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4957                 /* already being handled, ensure it gets handled
4958                  * again when current action finishes */
4959                 set_bit(STRIPE_HANDLE, &sh->state);
4960                 return;
4961         }
4962
4963         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4964                 break_stripe_batch_list(sh, 0);
4965
4966         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4967                 spin_lock(&sh->stripe_lock);
4968                 /*
4969                  * Cannot process 'sync' concurrently with 'discard'.
4970                  * Flush data in r5cache before 'sync'.
4971                  */
4972                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4973                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4974                     !test_bit(STRIPE_DISCARD, &sh->state) &&
4975                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4976                         set_bit(STRIPE_SYNCING, &sh->state);
4977                         clear_bit(STRIPE_INSYNC, &sh->state);
4978                         clear_bit(STRIPE_REPLACED, &sh->state);
4979                 }
4980                 spin_unlock(&sh->stripe_lock);
4981         }
4982         clear_bit(STRIPE_DELAYED, &sh->state);
4983
4984         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4985                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4986                (unsigned long long)sh->sector, sh->state,
4987                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4988                sh->check_state, sh->reconstruct_state);
4989
4990         analyse_stripe(sh, &s);
4991
4992         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4993                 goto finish;
4994
4995         if (s.handle_bad_blocks ||
4996             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4997                 set_bit(STRIPE_HANDLE, &sh->state);
4998                 goto finish;
4999         }
5000
5001         if (unlikely(s.blocked_rdev)) {
5002                 if (s.syncing || s.expanding || s.expanded ||
5003                     s.replacing || s.to_write || s.written) {
5004                         set_bit(STRIPE_HANDLE, &sh->state);
5005                         goto finish;
5006                 }
5007                 /* There is nothing for the blocked_rdev to block */
5008                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
5009                 s.blocked_rdev = NULL;
5010         }
5011
5012         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
5013                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
5014                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
5015         }
5016
5017         pr_debug("locked=%d uptodate=%d to_read=%d"
5018                " to_write=%d failed=%d failed_num=%d,%d\n",
5019                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
5020                s.failed_num[0], s.failed_num[1]);
5021         /*
5022          * check if the array has lost more than max_degraded devices and,
5023          * if so, some requests might need to be failed.
5024          *
5025          * When journal device failed (log_failed), we will only process
5026          * the stripe if there is data need write to raid disks
5027          */
5028         if (s.failed > conf->max_degraded ||
5029             (s.log_failed && s.injournal == 0)) {
5030                 sh->check_state = 0;
5031                 sh->reconstruct_state = 0;
5032                 break_stripe_batch_list(sh, 0);
5033                 if (s.to_read+s.to_write+s.written)
5034                         handle_failed_stripe(conf, sh, &s, disks);
5035                 if (s.syncing + s.replacing)
5036                         handle_failed_sync(conf, sh, &s);
5037         }
5038
5039         /* Now we check to see if any write operations have recently
5040          * completed
5041          */
5042         prexor = 0;
5043         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
5044                 prexor = 1;
5045         if (sh->reconstruct_state == reconstruct_state_drain_result ||
5046             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
5047                 sh->reconstruct_state = reconstruct_state_idle;
5048
5049                 /* All the 'written' buffers and the parity block are ready to
5050                  * be written back to disk
5051                  */
5052                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
5053                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
5054                 BUG_ON(sh->qd_idx >= 0 &&
5055                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
5056                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
5057                 for (i = disks; i--; ) {
5058                         struct r5dev *dev = &sh->dev[i];
5059                         if (test_bit(R5_LOCKED, &dev->flags) &&
5060                                 (i == sh->pd_idx || i == sh->qd_idx ||
5061                                  dev->written || test_bit(R5_InJournal,
5062                                                           &dev->flags))) {
5063                                 pr_debug("Writing block %d\n", i);
5064                                 set_bit(R5_Wantwrite, &dev->flags);
5065                                 if (prexor)
5066                                         continue;
5067                                 if (s.failed > 1)
5068                                         continue;
5069                                 if (!test_bit(R5_Insync, &dev->flags) ||
5070                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
5071                                      s.failed == 0))
5072                                         set_bit(STRIPE_INSYNC, &sh->state);
5073                         }
5074                 }
5075                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5076                         s.dec_preread_active = 1;
5077         }
5078
5079         /*
5080          * might be able to return some write requests if the parity blocks
5081          * are safe, or on a failed drive
5082          */
5083         pdev = &sh->dev[sh->pd_idx];
5084         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
5085                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
5086         qdev = &sh->dev[sh->qd_idx];
5087         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
5088                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
5089                 || conf->level < 6;
5090
5091         if (s.written &&
5092             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
5093                              && !test_bit(R5_LOCKED, &pdev->flags)
5094                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
5095                                  test_bit(R5_Discard, &pdev->flags))))) &&
5096             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
5097                              && !test_bit(R5_LOCKED, &qdev->flags)
5098                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
5099                                  test_bit(R5_Discard, &qdev->flags))))))
5100                 handle_stripe_clean_event(conf, sh, disks);
5101
5102         if (s.just_cached)
5103                 r5c_handle_cached_data_endio(conf, sh, disks);
5104         log_stripe_write_finished(sh);
5105
5106         /* Now we might consider reading some blocks, either to check/generate
5107          * parity, or to satisfy requests
5108          * or to load a block that is being partially written.
5109          */
5110         if (s.to_read || s.non_overwrite
5111             || (s.to_write && s.failed)
5112             || (s.syncing && (s.uptodate + s.compute < disks))
5113             || s.replacing
5114             || s.expanding)
5115                 handle_stripe_fill(sh, &s, disks);
5116
5117         /*
5118          * When the stripe finishes full journal write cycle (write to journal
5119          * and raid disk), this is the clean up procedure so it is ready for
5120          * next operation.
5121          */
5122         r5c_finish_stripe_write_out(conf, sh, &s);
5123
5124         /*
5125          * Now to consider new write requests, cache write back and what else,
5126          * if anything should be read.  We do not handle new writes when:
5127          * 1/ A 'write' operation (copy+xor) is already in flight.
5128          * 2/ A 'check' operation is in flight, as it may clobber the parity
5129          *    block.
5130          * 3/ A r5c cache log write is in flight.
5131          */
5132
5133         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
5134                 if (!r5c_is_writeback(conf->log)) {
5135                         if (s.to_write)
5136                                 handle_stripe_dirtying(conf, sh, &s, disks);
5137                 } else { /* write back cache */
5138                         int ret = 0;
5139
5140                         /* First, try handle writes in caching phase */
5141                         if (s.to_write)
5142                                 ret = r5c_try_caching_write(conf, sh, &s,
5143                                                             disks);
5144                         /*
5145                          * If caching phase failed: ret == -EAGAIN
5146                          *    OR
5147                          * stripe under reclaim: !caching && injournal
5148                          *
5149                          * fall back to handle_stripe_dirtying()
5150                          */
5151                         if (ret == -EAGAIN ||
5152                             /* stripe under reclaim: !caching && injournal */
5153                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
5154                              s.injournal > 0)) {
5155                                 ret = handle_stripe_dirtying(conf, sh, &s,
5156                                                              disks);
5157                                 if (ret == -EAGAIN)
5158                                         goto finish;
5159                         }
5160                 }
5161         }
5162
5163         /* maybe we need to check and possibly fix the parity for this stripe
5164          * Any reads will already have been scheduled, so we just see if enough
5165          * data is available.  The parity check is held off while parity
5166          * dependent operations are in flight.
5167          */
5168         if (sh->check_state ||
5169             (s.syncing && s.locked == 0 &&
5170              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5171              !test_bit(STRIPE_INSYNC, &sh->state))) {
5172                 if (conf->level == 6)
5173                         handle_parity_checks6(conf, sh, &s, disks);
5174                 else
5175                         handle_parity_checks5(conf, sh, &s, disks);
5176         }
5177
5178         if ((s.replacing || s.syncing) && s.locked == 0
5179             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
5180             && !test_bit(STRIPE_REPLACED, &sh->state)) {
5181                 /* Write out to replacement devices where possible */
5182                 for (i = 0; i < conf->raid_disks; i++)
5183                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
5184                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
5185                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
5186                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
5187                                 s.locked++;
5188                         }
5189                 if (s.replacing)
5190                         set_bit(STRIPE_INSYNC, &sh->state);
5191                 set_bit(STRIPE_REPLACED, &sh->state);
5192         }
5193         if ((s.syncing || s.replacing) && s.locked == 0 &&
5194             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5195             test_bit(STRIPE_INSYNC, &sh->state)) {
5196                 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5197                 clear_bit(STRIPE_SYNCING, &sh->state);
5198                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
5199                         wake_up(&conf->wait_for_overlap);
5200         }
5201
5202         /* If the failed drives are just a ReadError, then we might need
5203          * to progress the repair/check process
5204          */
5205         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
5206                 for (i = 0; i < s.failed; i++) {
5207                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
5208                         if (test_bit(R5_ReadError, &dev->flags)
5209                             && !test_bit(R5_LOCKED, &dev->flags)
5210                             && test_bit(R5_UPTODATE, &dev->flags)
5211                                 ) {
5212                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
5213                                         set_bit(R5_Wantwrite, &dev->flags);
5214                                         set_bit(R5_ReWrite, &dev->flags);
5215                                 } else
5216                                         /* let's read it back */
5217                                         set_bit(R5_Wantread, &dev->flags);
5218                                 set_bit(R5_LOCKED, &dev->flags);
5219                                 s.locked++;
5220                         }
5221                 }
5222
5223         /* Finish reconstruct operations initiated by the expansion process */
5224         if (sh->reconstruct_state == reconstruct_state_result) {
5225                 struct stripe_head *sh_src
5226                         = raid5_get_active_stripe(conf, NULL, sh->sector,
5227                                         R5_GAS_PREVIOUS | R5_GAS_NOBLOCK |
5228                                         R5_GAS_NOQUIESCE);
5229                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
5230                         /* sh cannot be written until sh_src has been read.
5231                          * so arrange for sh to be delayed a little
5232                          */
5233                         set_bit(STRIPE_DELAYED, &sh->state);
5234                         set_bit(STRIPE_HANDLE, &sh->state);
5235                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
5236                                               &sh_src->state))
5237                                 atomic_inc(&conf->preread_active_stripes);
5238                         raid5_release_stripe(sh_src);
5239                         goto finish;
5240                 }
5241                 if (sh_src)
5242                         raid5_release_stripe(sh_src);
5243
5244                 sh->reconstruct_state = reconstruct_state_idle;
5245                 clear_bit(STRIPE_EXPANDING, &sh->state);
5246                 for (i = conf->raid_disks; i--; ) {
5247                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
5248                         set_bit(R5_LOCKED, &sh->dev[i].flags);
5249                         s.locked++;
5250                 }
5251         }
5252
5253         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
5254             !sh->reconstruct_state) {
5255                 /* Need to write out all blocks after computing parity */
5256                 sh->disks = conf->raid_disks;
5257                 stripe_set_idx(sh->sector, conf, 0, sh);
5258                 schedule_reconstruction(sh, &s, 1, 1);
5259         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
5260                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
5261                 atomic_dec(&conf->reshape_stripes);
5262                 wake_up(&conf->wait_for_overlap);
5263                 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5264         }
5265
5266         if (s.expanding && s.locked == 0 &&
5267             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
5268                 handle_stripe_expansion(conf, sh);
5269
5270 finish:
5271         /* wait for this device to become unblocked */
5272         if (unlikely(s.blocked_rdev)) {
5273                 if (conf->mddev->external)
5274                         md_wait_for_blocked_rdev(s.blocked_rdev,
5275                                                  conf->mddev);
5276                 else
5277                         /* Internal metadata will immediately
5278                          * be written by raid5d, so we don't
5279                          * need to wait here.
5280                          */
5281                         rdev_dec_pending(s.blocked_rdev,
5282                                          conf->mddev);
5283         }
5284
5285         if (s.handle_bad_blocks)
5286                 for (i = disks; i--; ) {
5287                         struct md_rdev *rdev;
5288                         struct r5dev *dev = &sh->dev[i];
5289                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5290                                 /* We own a safe reference to the rdev */
5291                                 rdev = conf->disks[i].rdev;
5292                                 if (!rdev_set_badblocks(rdev, sh->sector,
5293                                                         RAID5_STRIPE_SECTORS(conf), 0))
5294                                         md_error(conf->mddev, rdev);
5295                                 rdev_dec_pending(rdev, conf->mddev);
5296                         }
5297                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5298                                 rdev = conf->disks[i].rdev;
5299                                 rdev_clear_badblocks(rdev, sh->sector,
5300                                                      RAID5_STRIPE_SECTORS(conf), 0);
5301                                 rdev_dec_pending(rdev, conf->mddev);
5302                         }
5303                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5304                                 rdev = conf->disks[i].replacement;
5305                                 if (!rdev)
5306                                         /* rdev have been moved down */
5307                                         rdev = conf->disks[i].rdev;
5308                                 rdev_clear_badblocks(rdev, sh->sector,
5309                                                      RAID5_STRIPE_SECTORS(conf), 0);
5310                                 rdev_dec_pending(rdev, conf->mddev);
5311                         }
5312                 }
5313
5314         if (s.ops_request)
5315                 raid_run_ops(sh, s.ops_request);
5316
5317         ops_run_io(sh, &s);
5318
5319         if (s.dec_preread_active) {
5320                 /* We delay this until after ops_run_io so that if make_request
5321                  * is waiting on a flush, it won't continue until the writes
5322                  * have actually been submitted.
5323                  */
5324                 atomic_dec(&conf->preread_active_stripes);
5325                 if (atomic_read(&conf->preread_active_stripes) <
5326                     IO_THRESHOLD)
5327                         md_wakeup_thread(conf->mddev->thread);
5328         }
5329
5330         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5331 }
5332
5333 static void raid5_activate_delayed(struct r5conf *conf)
5334         __must_hold(&conf->device_lock)
5335 {
5336         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5337                 while (!list_empty(&conf->delayed_list)) {
5338                         struct list_head *l = conf->delayed_list.next;
5339                         struct stripe_head *sh;
5340                         sh = list_entry(l, struct stripe_head, lru);
5341                         list_del_init(l);
5342                         clear_bit(STRIPE_DELAYED, &sh->state);
5343                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5344                                 atomic_inc(&conf->preread_active_stripes);
5345                         list_add_tail(&sh->lru, &conf->hold_list);
5346                         raid5_wakeup_stripe_thread(sh);
5347                 }
5348         }
5349 }
5350
5351 static void activate_bit_delay(struct r5conf *conf,
5352                 struct list_head *temp_inactive_list)
5353         __must_hold(&conf->device_lock)
5354 {
5355         struct list_head head;
5356         list_add(&head, &conf->bitmap_list);
5357         list_del_init(&conf->bitmap_list);
5358         while (!list_empty(&head)) {
5359                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5360                 int hash;
5361                 list_del_init(&sh->lru);
5362                 atomic_inc(&sh->count);
5363                 hash = sh->hash_lock_index;
5364                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5365         }
5366 }
5367
5368 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5369 {
5370         struct r5conf *conf = mddev->private;
5371         sector_t sector = bio->bi_iter.bi_sector;
5372         unsigned int chunk_sectors;
5373         unsigned int bio_sectors = bio_sectors(bio);
5374
5375         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5376         return  chunk_sectors >=
5377                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5378 }
5379
5380 /*
5381  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5382  *  later sampled by raid5d.
5383  */
5384 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5385 {
5386         unsigned long flags;
5387
5388         spin_lock_irqsave(&conf->device_lock, flags);
5389
5390         bi->bi_next = conf->retry_read_aligned_list;
5391         conf->retry_read_aligned_list = bi;
5392
5393         spin_unlock_irqrestore(&conf->device_lock, flags);
5394         md_wakeup_thread(conf->mddev->thread);
5395 }
5396
5397 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5398                                          unsigned int *offset)
5399 {
5400         struct bio *bi;
5401
5402         bi = conf->retry_read_aligned;
5403         if (bi) {
5404                 *offset = conf->retry_read_offset;
5405                 conf->retry_read_aligned = NULL;
5406                 return bi;
5407         }
5408         bi = conf->retry_read_aligned_list;
5409         if(bi) {
5410                 conf->retry_read_aligned_list = bi->bi_next;
5411                 bi->bi_next = NULL;
5412                 *offset = 0;
5413         }
5414
5415         return bi;
5416 }
5417
5418 /*
5419  *  The "raid5_align_endio" should check if the read succeeded and if it
5420  *  did, call bio_endio on the original bio (having bio_put the new bio
5421  *  first).
5422  *  If the read failed..
5423  */
5424 static void raid5_align_endio(struct bio *bi)
5425 {
5426         struct bio *raid_bi = bi->bi_private;
5427         struct md_rdev *rdev = (void *)raid_bi->bi_next;
5428         struct mddev *mddev = rdev->mddev;
5429         struct r5conf *conf = mddev->private;
5430         blk_status_t error = bi->bi_status;
5431
5432         bio_put(bi);
5433         raid_bi->bi_next = NULL;
5434         rdev_dec_pending(rdev, conf->mddev);
5435
5436         if (!error) {
5437                 bio_endio(raid_bi);
5438                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5439                         wake_up(&conf->wait_for_quiescent);
5440                 return;
5441         }
5442
5443         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5444
5445         add_bio_to_retry(raid_bi, conf);
5446 }
5447
5448 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5449 {
5450         struct r5conf *conf = mddev->private;
5451         struct bio *align_bio;
5452         struct md_rdev *rdev;
5453         sector_t sector, end_sector;
5454         int dd_idx;
5455         bool did_inc;
5456
5457         if (!in_chunk_boundary(mddev, raid_bio)) {
5458                 pr_debug("%s: non aligned\n", __func__);
5459                 return 0;
5460         }
5461
5462         sector = raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector, 0,
5463                                       &dd_idx, NULL);
5464         end_sector = sector + bio_sectors(raid_bio);
5465
5466         if (r5c_big_stripe_cached(conf, sector))
5467                 return 0;
5468
5469         rdev = conf->disks[dd_idx].replacement;
5470         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5471             rdev->recovery_offset < end_sector) {
5472                 rdev = conf->disks[dd_idx].rdev;
5473                 if (!rdev)
5474                         return 0;
5475                 if (test_bit(Faulty, &rdev->flags) ||
5476                     !(test_bit(In_sync, &rdev->flags) ||
5477                       rdev->recovery_offset >= end_sector))
5478                         return 0;
5479         }
5480
5481         atomic_inc(&rdev->nr_pending);
5482
5483         if (rdev_has_badblock(rdev, sector, bio_sectors(raid_bio))) {
5484                 rdev_dec_pending(rdev, mddev);
5485                 return 0;
5486         }
5487
5488         md_account_bio(mddev, &raid_bio);
5489         raid_bio->bi_next = (void *)rdev;
5490
5491         align_bio = bio_alloc_clone(rdev->bdev, raid_bio, GFP_NOIO,
5492                                     &mddev->bio_set);
5493         align_bio->bi_end_io = raid5_align_endio;
5494         align_bio->bi_private = raid_bio;
5495         align_bio->bi_iter.bi_sector = sector;
5496
5497         /* No reshape active, so we can trust rdev->data_offset */
5498         align_bio->bi_iter.bi_sector += rdev->data_offset;
5499
5500         did_inc = false;
5501         if (conf->quiesce == 0) {
5502                 atomic_inc(&conf->active_aligned_reads);
5503                 did_inc = true;
5504         }
5505         /* need a memory barrier to detect the race with raid5_quiesce() */
5506         if (!did_inc || smp_load_acquire(&conf->quiesce) != 0) {
5507                 /* quiesce is in progress, so we need to undo io activation and wait
5508                  * for it to finish
5509                  */
5510                 if (did_inc && atomic_dec_and_test(&conf->active_aligned_reads))
5511                         wake_up(&conf->wait_for_quiescent);
5512                 spin_lock_irq(&conf->device_lock);
5513                 wait_event_lock_irq(conf->wait_for_quiescent, conf->quiesce == 0,
5514                                     conf->device_lock);
5515                 atomic_inc(&conf->active_aligned_reads);
5516                 spin_unlock_irq(&conf->device_lock);
5517         }
5518
5519         mddev_trace_remap(mddev, align_bio, raid_bio->bi_iter.bi_sector);
5520         submit_bio_noacct(align_bio);
5521         return 1;
5522 }
5523
5524 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5525 {
5526         struct bio *split;
5527         sector_t sector = raid_bio->bi_iter.bi_sector;
5528         unsigned chunk_sects = mddev->chunk_sectors;
5529         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5530
5531         if (sectors < bio_sectors(raid_bio)) {
5532                 struct r5conf *conf = mddev->private;
5533                 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5534                 bio_chain(split, raid_bio);
5535                 submit_bio_noacct(raid_bio);
5536                 raid_bio = split;
5537         }
5538
5539         if (!raid5_read_one_chunk(mddev, raid_bio))
5540                 return raid_bio;
5541
5542         return NULL;
5543 }
5544
5545 /* __get_priority_stripe - get the next stripe to process
5546  *
5547  * Full stripe writes are allowed to pass preread active stripes up until
5548  * the bypass_threshold is exceeded.  In general the bypass_count
5549  * increments when the handle_list is handled before the hold_list; however, it
5550  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5551  * stripe with in flight i/o.  The bypass_count will be reset when the
5552  * head of the hold_list has changed, i.e. the head was promoted to the
5553  * handle_list.
5554  */
5555 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5556         __must_hold(&conf->device_lock)
5557 {
5558         struct stripe_head *sh, *tmp;
5559         struct list_head *handle_list = NULL;
5560         struct r5worker_group *wg;
5561         bool second_try = !r5c_is_writeback(conf->log) &&
5562                 !r5l_log_disk_error(conf);
5563         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5564                 r5l_log_disk_error(conf);
5565
5566 again:
5567         wg = NULL;
5568         sh = NULL;
5569         if (conf->worker_cnt_per_group == 0) {
5570                 handle_list = try_loprio ? &conf->loprio_list :
5571                                         &conf->handle_list;
5572         } else if (group != ANY_GROUP) {
5573                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5574                                 &conf->worker_groups[group].handle_list;
5575                 wg = &conf->worker_groups[group];
5576         } else {
5577                 int i;
5578                 for (i = 0; i < conf->group_cnt; i++) {
5579                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5580                                 &conf->worker_groups[i].handle_list;
5581                         wg = &conf->worker_groups[i];
5582                         if (!list_empty(handle_list))
5583                                 break;
5584                 }
5585         }
5586
5587         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5588                   __func__,
5589                   list_empty(handle_list) ? "empty" : "busy",
5590                   list_empty(&conf->hold_list) ? "empty" : "busy",
5591                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5592
5593         if (!list_empty(handle_list)) {
5594                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5595
5596                 if (list_empty(&conf->hold_list))
5597                         conf->bypass_count = 0;
5598                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5599                         if (conf->hold_list.next == conf->last_hold)
5600                                 conf->bypass_count++;
5601                         else {
5602                                 conf->last_hold = conf->hold_list.next;
5603                                 conf->bypass_count -= conf->bypass_threshold;
5604                                 if (conf->bypass_count < 0)
5605                                         conf->bypass_count = 0;
5606                         }
5607                 }
5608         } else if (!list_empty(&conf->hold_list) &&
5609                    ((conf->bypass_threshold &&
5610                      conf->bypass_count > conf->bypass_threshold) ||
5611                     atomic_read(&conf->pending_full_writes) == 0)) {
5612
5613                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5614                         if (conf->worker_cnt_per_group == 0 ||
5615                             group == ANY_GROUP ||
5616                             !cpu_online(tmp->cpu) ||
5617                             cpu_to_group(tmp->cpu) == group) {
5618                                 sh = tmp;
5619                                 break;
5620                         }
5621                 }
5622
5623                 if (sh) {
5624                         conf->bypass_count -= conf->bypass_threshold;
5625                         if (conf->bypass_count < 0)
5626                                 conf->bypass_count = 0;
5627                 }
5628                 wg = NULL;
5629         }
5630
5631         if (!sh) {
5632                 if (second_try)
5633                         return NULL;
5634                 second_try = true;
5635                 try_loprio = !try_loprio;
5636                 goto again;
5637         }
5638
5639         if (wg) {
5640                 wg->stripes_cnt--;
5641                 sh->group = NULL;
5642         }
5643         list_del_init(&sh->lru);
5644         BUG_ON(atomic_inc_return(&sh->count) != 1);
5645         return sh;
5646 }
5647
5648 struct raid5_plug_cb {
5649         struct blk_plug_cb      cb;
5650         struct list_head        list;
5651         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5652 };
5653
5654 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5655 {
5656         struct raid5_plug_cb *cb = container_of(
5657                 blk_cb, struct raid5_plug_cb, cb);
5658         struct stripe_head *sh;
5659         struct mddev *mddev = cb->cb.data;
5660         struct r5conf *conf = mddev->private;
5661         int cnt = 0;
5662         int hash;
5663
5664         if (cb->list.next && !list_empty(&cb->list)) {
5665                 spin_lock_irq(&conf->device_lock);
5666                 while (!list_empty(&cb->list)) {
5667                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5668                         list_del_init(&sh->lru);
5669                         /*
5670                          * avoid race release_stripe_plug() sees
5671                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5672                          * is still in our list
5673                          */
5674                         smp_mb__before_atomic();
5675                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5676                         /*
5677                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5678                          * case, the count is always > 1 here
5679                          */
5680                         hash = sh->hash_lock_index;
5681                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5682                         cnt++;
5683                 }
5684                 spin_unlock_irq(&conf->device_lock);
5685         }
5686         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5687                                      NR_STRIPE_HASH_LOCKS);
5688         if (!mddev_is_dm(mddev))
5689                 trace_block_unplug(mddev->gendisk->queue, cnt, !from_schedule);
5690         kfree(cb);
5691 }
5692
5693 static void release_stripe_plug(struct mddev *mddev,
5694                                 struct stripe_head *sh)
5695 {
5696         struct blk_plug_cb *blk_cb = blk_check_plugged(
5697                 raid5_unplug, mddev,
5698                 sizeof(struct raid5_plug_cb));
5699         struct raid5_plug_cb *cb;
5700
5701         if (!blk_cb) {
5702                 raid5_release_stripe(sh);
5703                 return;
5704         }
5705
5706         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5707
5708         if (cb->list.next == NULL) {
5709                 int i;
5710                 INIT_LIST_HEAD(&cb->list);
5711                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5712                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5713         }
5714
5715         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5716                 list_add_tail(&sh->lru, &cb->list);
5717         else
5718                 raid5_release_stripe(sh);
5719 }
5720
5721 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5722 {
5723         struct r5conf *conf = mddev->private;
5724         sector_t logical_sector, last_sector;
5725         struct stripe_head *sh;
5726         int stripe_sectors;
5727
5728         /* We need to handle this when io_uring supports discard/trim */
5729         if (WARN_ON_ONCE(bi->bi_opf & REQ_NOWAIT))
5730                 return;
5731
5732         if (mddev->reshape_position != MaxSector)
5733                 /* Skip discard while reshape is happening */
5734                 return;
5735
5736         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5737         last_sector = bio_end_sector(bi);
5738
5739         bi->bi_next = NULL;
5740
5741         stripe_sectors = conf->chunk_sectors *
5742                 (conf->raid_disks - conf->max_degraded);
5743         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5744                                                stripe_sectors);
5745         sector_div(last_sector, stripe_sectors);
5746
5747         logical_sector *= conf->chunk_sectors;
5748         last_sector *= conf->chunk_sectors;
5749
5750         for (; logical_sector < last_sector;
5751              logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5752                 DEFINE_WAIT(w);
5753                 int d;
5754         again:
5755                 sh = raid5_get_active_stripe(conf, NULL, logical_sector, 0);
5756                 prepare_to_wait(&conf->wait_for_overlap, &w,
5757                                 TASK_UNINTERRUPTIBLE);
5758                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5759                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5760                         raid5_release_stripe(sh);
5761                         schedule();
5762                         goto again;
5763                 }
5764                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5765                 spin_lock_irq(&sh->stripe_lock);
5766                 for (d = 0; d < conf->raid_disks; d++) {
5767                         if (d == sh->pd_idx || d == sh->qd_idx)
5768                                 continue;
5769                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5770                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5771                                 spin_unlock_irq(&sh->stripe_lock);
5772                                 raid5_release_stripe(sh);
5773                                 schedule();
5774                                 goto again;
5775                         }
5776                 }
5777                 set_bit(STRIPE_DISCARD, &sh->state);
5778                 finish_wait(&conf->wait_for_overlap, &w);
5779                 sh->overwrite_disks = 0;
5780                 for (d = 0; d < conf->raid_disks; d++) {
5781                         if (d == sh->pd_idx || d == sh->qd_idx)
5782                                 continue;
5783                         sh->dev[d].towrite = bi;
5784                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5785                         bio_inc_remaining(bi);
5786                         md_write_inc(mddev, bi);
5787                         sh->overwrite_disks++;
5788                 }
5789                 spin_unlock_irq(&sh->stripe_lock);
5790                 if (conf->mddev->bitmap) {
5791                         for (d = 0;
5792                              d < conf->raid_disks - conf->max_degraded;
5793                              d++)
5794                                 md_bitmap_startwrite(mddev->bitmap,
5795                                                      sh->sector,
5796                                                      RAID5_STRIPE_SECTORS(conf),
5797                                                      0);
5798                         sh->bm_seq = conf->seq_flush + 1;
5799                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5800                 }
5801
5802                 set_bit(STRIPE_HANDLE, &sh->state);
5803                 clear_bit(STRIPE_DELAYED, &sh->state);
5804                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5805                         atomic_inc(&conf->preread_active_stripes);
5806                 release_stripe_plug(mddev, sh);
5807         }
5808
5809         bio_endio(bi);
5810 }
5811
5812 static bool ahead_of_reshape(struct mddev *mddev, sector_t sector,
5813                              sector_t reshape_sector)
5814 {
5815         return mddev->reshape_backwards ? sector < reshape_sector :
5816                                           sector >= reshape_sector;
5817 }
5818
5819 static bool range_ahead_of_reshape(struct mddev *mddev, sector_t min,
5820                                    sector_t max, sector_t reshape_sector)
5821 {
5822         return mddev->reshape_backwards ? max < reshape_sector :
5823                                           min >= reshape_sector;
5824 }
5825
5826 static bool stripe_ahead_of_reshape(struct mddev *mddev, struct r5conf *conf,
5827                                     struct stripe_head *sh)
5828 {
5829         sector_t max_sector = 0, min_sector = MaxSector;
5830         bool ret = false;
5831         int dd_idx;
5832
5833         for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5834                 if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5835                         continue;
5836
5837                 min_sector = min(min_sector, sh->dev[dd_idx].sector);
5838                 max_sector = max(max_sector, sh->dev[dd_idx].sector);
5839         }
5840
5841         spin_lock_irq(&conf->device_lock);
5842
5843         if (!range_ahead_of_reshape(mddev, min_sector, max_sector,
5844                                      conf->reshape_progress))
5845                 /* mismatch, need to try again */
5846                 ret = true;
5847
5848         spin_unlock_irq(&conf->device_lock);
5849
5850         return ret;
5851 }
5852
5853 static int add_all_stripe_bios(struct r5conf *conf,
5854                 struct stripe_request_ctx *ctx, struct stripe_head *sh,
5855                 struct bio *bi, int forwrite, int previous)
5856 {
5857         int dd_idx;
5858         int ret = 1;
5859
5860         spin_lock_irq(&sh->stripe_lock);
5861
5862         for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5863                 struct r5dev *dev = &sh->dev[dd_idx];
5864
5865                 if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5866                         continue;
5867
5868                 if (dev->sector < ctx->first_sector ||
5869                     dev->sector >= ctx->last_sector)
5870                         continue;
5871
5872                 if (stripe_bio_overlaps(sh, bi, dd_idx, forwrite)) {
5873                         set_bit(R5_Overlap, &dev->flags);
5874                         ret = 0;
5875                         continue;
5876                 }
5877         }
5878
5879         if (!ret)
5880                 goto out;
5881
5882         for (dd_idx = 0; dd_idx < sh->disks; dd_idx++) {
5883                 struct r5dev *dev = &sh->dev[dd_idx];
5884
5885                 if (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
5886                         continue;
5887
5888                 if (dev->sector < ctx->first_sector ||
5889                     dev->sector >= ctx->last_sector)
5890                         continue;
5891
5892                 __add_stripe_bio(sh, bi, dd_idx, forwrite, previous);
5893                 clear_bit((dev->sector - ctx->first_sector) >>
5894                           RAID5_STRIPE_SHIFT(conf), ctx->sectors_to_do);
5895         }
5896
5897 out:
5898         spin_unlock_irq(&sh->stripe_lock);
5899         return ret;
5900 }
5901
5902 enum reshape_loc {
5903         LOC_NO_RESHAPE,
5904         LOC_AHEAD_OF_RESHAPE,
5905         LOC_INSIDE_RESHAPE,
5906         LOC_BEHIND_RESHAPE,
5907 };
5908
5909 static enum reshape_loc get_reshape_loc(struct mddev *mddev,
5910                 struct r5conf *conf, sector_t logical_sector)
5911 {
5912         sector_t reshape_progress, reshape_safe;
5913         /*
5914          * Spinlock is needed as reshape_progress may be
5915          * 64bit on a 32bit platform, and so it might be
5916          * possible to see a half-updated value
5917          * Of course reshape_progress could change after
5918          * the lock is dropped, so once we get a reference
5919          * to the stripe that we think it is, we will have
5920          * to check again.
5921          */
5922         spin_lock_irq(&conf->device_lock);
5923         reshape_progress = conf->reshape_progress;
5924         reshape_safe = conf->reshape_safe;
5925         spin_unlock_irq(&conf->device_lock);
5926         if (reshape_progress == MaxSector)
5927                 return LOC_NO_RESHAPE;
5928         if (ahead_of_reshape(mddev, logical_sector, reshape_progress))
5929                 return LOC_AHEAD_OF_RESHAPE;
5930         if (ahead_of_reshape(mddev, logical_sector, reshape_safe))
5931                 return LOC_INSIDE_RESHAPE;
5932         return LOC_BEHIND_RESHAPE;
5933 }
5934
5935 static enum stripe_result make_stripe_request(struct mddev *mddev,
5936                 struct r5conf *conf, struct stripe_request_ctx *ctx,
5937                 sector_t logical_sector, struct bio *bi)
5938 {
5939         const int rw = bio_data_dir(bi);
5940         enum stripe_result ret;
5941         struct stripe_head *sh;
5942         sector_t new_sector;
5943         int previous = 0, flags = 0;
5944         int seq, dd_idx;
5945
5946         seq = read_seqcount_begin(&conf->gen_lock);
5947
5948         if (unlikely(conf->reshape_progress != MaxSector)) {
5949                 enum reshape_loc loc = get_reshape_loc(mddev, conf,
5950                                                        logical_sector);
5951                 if (loc == LOC_INSIDE_RESHAPE) {
5952                         ret = STRIPE_SCHEDULE_AND_RETRY;
5953                         goto out;
5954                 }
5955                 if (loc == LOC_AHEAD_OF_RESHAPE)
5956                         previous = 1;
5957         }
5958
5959         new_sector = raid5_compute_sector(conf, logical_sector, previous,
5960                                           &dd_idx, NULL);
5961         pr_debug("raid456: %s, sector %llu logical %llu\n", __func__,
5962                  new_sector, logical_sector);
5963
5964         if (previous)
5965                 flags |= R5_GAS_PREVIOUS;
5966         if (bi->bi_opf & REQ_RAHEAD)
5967                 flags |= R5_GAS_NOBLOCK;
5968         sh = raid5_get_active_stripe(conf, ctx, new_sector, flags);
5969         if (unlikely(!sh)) {
5970                 /* cannot get stripe, just give-up */
5971                 bi->bi_status = BLK_STS_IOERR;
5972                 return STRIPE_FAIL;
5973         }
5974
5975         if (unlikely(previous) &&
5976             stripe_ahead_of_reshape(mddev, conf, sh)) {
5977                 /*
5978                  * Expansion moved on while waiting for a stripe.
5979                  * Expansion could still move past after this
5980                  * test, but as we are holding a reference to
5981                  * 'sh', we know that if that happens,
5982                  *  STRIPE_EXPANDING will get set and the expansion
5983                  * won't proceed until we finish with the stripe.
5984                  */
5985                 ret = STRIPE_SCHEDULE_AND_RETRY;
5986                 goto out_release;
5987         }
5988
5989         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5990                 /* Might have got the wrong stripe_head by accident */
5991                 ret = STRIPE_RETRY;
5992                 goto out_release;
5993         }
5994
5995         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5996             !add_all_stripe_bios(conf, ctx, sh, bi, rw, previous)) {
5997                 /*
5998                  * Stripe is busy expanding or add failed due to
5999                  * overlap. Flush everything and wait a while.
6000                  */
6001                 md_wakeup_thread(mddev->thread);
6002                 ret = STRIPE_SCHEDULE_AND_RETRY;
6003                 goto out_release;
6004         }
6005
6006         if (stripe_can_batch(sh)) {
6007                 stripe_add_to_batch_list(conf, sh, ctx->batch_last);
6008                 if (ctx->batch_last)
6009                         raid5_release_stripe(ctx->batch_last);
6010                 atomic_inc(&sh->count);
6011                 ctx->batch_last = sh;
6012         }
6013
6014         if (ctx->do_flush) {
6015                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
6016                 /* we only need flush for one stripe */
6017                 ctx->do_flush = false;
6018         }
6019
6020         set_bit(STRIPE_HANDLE, &sh->state);
6021         clear_bit(STRIPE_DELAYED, &sh->state);
6022         if ((!sh->batch_head || sh == sh->batch_head) &&
6023             (bi->bi_opf & REQ_SYNC) &&
6024             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
6025                 atomic_inc(&conf->preread_active_stripes);
6026
6027         release_stripe_plug(mddev, sh);
6028         return STRIPE_SUCCESS;
6029
6030 out_release:
6031         raid5_release_stripe(sh);
6032 out:
6033         if (ret == STRIPE_SCHEDULE_AND_RETRY && reshape_interrupted(mddev)) {
6034                 bi->bi_status = BLK_STS_RESOURCE;
6035                 ret = STRIPE_WAIT_RESHAPE;
6036                 pr_err_ratelimited("dm-raid456: io across reshape position while reshape can't make progress");
6037         }
6038         return ret;
6039 }
6040
6041 /*
6042  * If the bio covers multiple data disks, find sector within the bio that has
6043  * the lowest chunk offset in the first chunk.
6044  */
6045 static sector_t raid5_bio_lowest_chunk_sector(struct r5conf *conf,
6046                                               struct bio *bi)
6047 {
6048         int sectors_per_chunk = conf->chunk_sectors;
6049         int raid_disks = conf->raid_disks;
6050         int dd_idx;
6051         struct stripe_head sh;
6052         unsigned int chunk_offset;
6053         sector_t r_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6054         sector_t sector;
6055
6056         /* We pass in fake stripe_head to get back parity disk numbers */
6057         sector = raid5_compute_sector(conf, r_sector, 0, &dd_idx, &sh);
6058         chunk_offset = sector_div(sector, sectors_per_chunk);
6059         if (sectors_per_chunk - chunk_offset >= bio_sectors(bi))
6060                 return r_sector;
6061         /*
6062          * Bio crosses to the next data disk. Check whether it's in the same
6063          * chunk.
6064          */
6065         dd_idx++;
6066         while (dd_idx == sh.pd_idx || dd_idx == sh.qd_idx)
6067                 dd_idx++;
6068         if (dd_idx >= raid_disks)
6069                 return r_sector;
6070         return r_sector + sectors_per_chunk - chunk_offset;
6071 }
6072
6073 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
6074 {
6075         DEFINE_WAIT_FUNC(wait, woken_wake_function);
6076         struct r5conf *conf = mddev->private;
6077         sector_t logical_sector;
6078         struct stripe_request_ctx ctx = {};
6079         const int rw = bio_data_dir(bi);
6080         enum stripe_result res;
6081         int s, stripe_cnt;
6082
6083         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
6084                 int ret = log_handle_flush_request(conf, bi);
6085
6086                 if (ret == 0)
6087                         return true;
6088                 if (ret == -ENODEV) {
6089                         if (md_flush_request(mddev, bi))
6090                                 return true;
6091                 }
6092                 /* ret == -EAGAIN, fallback */
6093                 /*
6094                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
6095                  * we need to flush journal device
6096                  */
6097                 ctx.do_flush = bi->bi_opf & REQ_PREFLUSH;
6098         }
6099
6100         md_write_start(mddev, bi);
6101         /*
6102          * If array is degraded, better not do chunk aligned read because
6103          * later we might have to read it again in order to reconstruct
6104          * data on failed drives.
6105          */
6106         if (rw == READ && mddev->degraded == 0 &&
6107             mddev->reshape_position == MaxSector) {
6108                 bi = chunk_aligned_read(mddev, bi);
6109                 if (!bi)
6110                         return true;
6111         }
6112
6113         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
6114                 make_discard_request(mddev, bi);
6115                 md_write_end(mddev);
6116                 return true;
6117         }
6118
6119         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6120         ctx.first_sector = logical_sector;
6121         ctx.last_sector = bio_end_sector(bi);
6122         bi->bi_next = NULL;
6123
6124         stripe_cnt = DIV_ROUND_UP_SECTOR_T(ctx.last_sector - logical_sector,
6125                                            RAID5_STRIPE_SECTORS(conf));
6126         bitmap_set(ctx.sectors_to_do, 0, stripe_cnt);
6127
6128         pr_debug("raid456: %s, logical %llu to %llu\n", __func__,
6129                  bi->bi_iter.bi_sector, ctx.last_sector);
6130
6131         /* Bail out if conflicts with reshape and REQ_NOWAIT is set */
6132         if ((bi->bi_opf & REQ_NOWAIT) &&
6133             (conf->reshape_progress != MaxSector) &&
6134             get_reshape_loc(mddev, conf, logical_sector) == LOC_INSIDE_RESHAPE) {
6135                 bio_wouldblock_error(bi);
6136                 if (rw == WRITE)
6137                         md_write_end(mddev);
6138                 return true;
6139         }
6140         md_account_bio(mddev, &bi);
6141
6142         /*
6143          * Lets start with the stripe with the lowest chunk offset in the first
6144          * chunk. That has the best chances of creating IOs adjacent to
6145          * previous IOs in case of sequential IO and thus creates the most
6146          * sequential IO pattern. We don't bother with the optimization when
6147          * reshaping as the performance benefit is not worth the complexity.
6148          */
6149         if (likely(conf->reshape_progress == MaxSector))
6150                 logical_sector = raid5_bio_lowest_chunk_sector(conf, bi);
6151         s = (logical_sector - ctx.first_sector) >> RAID5_STRIPE_SHIFT(conf);
6152
6153         add_wait_queue(&conf->wait_for_overlap, &wait);
6154         while (1) {
6155                 res = make_stripe_request(mddev, conf, &ctx, logical_sector,
6156                                           bi);
6157                 if (res == STRIPE_FAIL || res == STRIPE_WAIT_RESHAPE)
6158                         break;
6159
6160                 if (res == STRIPE_RETRY)
6161                         continue;
6162
6163                 if (res == STRIPE_SCHEDULE_AND_RETRY) {
6164                         /*
6165                          * Must release the reference to batch_last before
6166                          * scheduling and waiting for work to be done,
6167                          * otherwise the batch_last stripe head could prevent
6168                          * raid5_activate_delayed() from making progress
6169                          * and thus deadlocking.
6170                          */
6171                         if (ctx.batch_last) {
6172                                 raid5_release_stripe(ctx.batch_last);
6173                                 ctx.batch_last = NULL;
6174                         }
6175
6176                         wait_woken(&wait, TASK_UNINTERRUPTIBLE,
6177                                    MAX_SCHEDULE_TIMEOUT);
6178                         continue;
6179                 }
6180
6181                 s = find_next_bit_wrap(ctx.sectors_to_do, stripe_cnt, s);
6182                 if (s == stripe_cnt)
6183                         break;
6184
6185                 logical_sector = ctx.first_sector +
6186                         (s << RAID5_STRIPE_SHIFT(conf));
6187         }
6188         remove_wait_queue(&conf->wait_for_overlap, &wait);
6189
6190         if (ctx.batch_last)
6191                 raid5_release_stripe(ctx.batch_last);
6192
6193         if (rw == WRITE)
6194                 md_write_end(mddev);
6195         if (res == STRIPE_WAIT_RESHAPE) {
6196                 md_free_cloned_bio(bi);
6197                 return false;
6198         }
6199
6200         bio_endio(bi);
6201         return true;
6202 }
6203
6204 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
6205
6206 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
6207 {
6208         /* reshaping is quite different to recovery/resync so it is
6209          * handled quite separately ... here.
6210          *
6211          * On each call to sync_request, we gather one chunk worth of
6212          * destination stripes and flag them as expanding.
6213          * Then we find all the source stripes and request reads.
6214          * As the reads complete, handle_stripe will copy the data
6215          * into the destination stripe and release that stripe.
6216          */
6217         struct r5conf *conf = mddev->private;
6218         struct stripe_head *sh;
6219         struct md_rdev *rdev;
6220         sector_t first_sector, last_sector;
6221         int raid_disks = conf->previous_raid_disks;
6222         int data_disks = raid_disks - conf->max_degraded;
6223         int new_data_disks = conf->raid_disks - conf->max_degraded;
6224         int i;
6225         int dd_idx;
6226         sector_t writepos, readpos, safepos;
6227         sector_t stripe_addr;
6228         int reshape_sectors;
6229         struct list_head stripes;
6230         sector_t retn;
6231
6232         if (sector_nr == 0) {
6233                 /* If restarting in the middle, skip the initial sectors */
6234                 if (mddev->reshape_backwards &&
6235                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
6236                         sector_nr = raid5_size(mddev, 0, 0)
6237                                 - conf->reshape_progress;
6238                 } else if (mddev->reshape_backwards &&
6239                            conf->reshape_progress == MaxSector) {
6240                         /* shouldn't happen, but just in case, finish up.*/
6241                         sector_nr = MaxSector;
6242                 } else if (!mddev->reshape_backwards &&
6243                            conf->reshape_progress > 0)
6244                         sector_nr = conf->reshape_progress;
6245                 sector_div(sector_nr, new_data_disks);
6246                 if (sector_nr) {
6247                         mddev->curr_resync_completed = sector_nr;
6248                         sysfs_notify_dirent_safe(mddev->sysfs_completed);
6249                         *skipped = 1;
6250                         retn = sector_nr;
6251                         goto finish;
6252                 }
6253         }
6254
6255         /* We need to process a full chunk at a time.
6256          * If old and new chunk sizes differ, we need to process the
6257          * largest of these
6258          */
6259
6260         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
6261
6262         /* We update the metadata at least every 10 seconds, or when
6263          * the data about to be copied would over-write the source of
6264          * the data at the front of the range.  i.e. one new_stripe
6265          * along from reshape_progress new_maps to after where
6266          * reshape_safe old_maps to
6267          */
6268         writepos = conf->reshape_progress;
6269         sector_div(writepos, new_data_disks);
6270         readpos = conf->reshape_progress;
6271         sector_div(readpos, data_disks);
6272         safepos = conf->reshape_safe;
6273         sector_div(safepos, data_disks);
6274         if (mddev->reshape_backwards) {
6275                 if (WARN_ON(writepos < reshape_sectors))
6276                         return MaxSector;
6277
6278                 writepos -= reshape_sectors;
6279                 readpos += reshape_sectors;
6280                 safepos += reshape_sectors;
6281         } else {
6282                 writepos += reshape_sectors;
6283                 /* readpos and safepos are worst-case calculations.
6284                  * A negative number is overly pessimistic, and causes
6285                  * obvious problems for unsigned storage.  So clip to 0.
6286                  */
6287                 readpos -= min_t(sector_t, reshape_sectors, readpos);
6288                 safepos -= min_t(sector_t, reshape_sectors, safepos);
6289         }
6290
6291         /* Having calculated the 'writepos' possibly use it
6292          * to set 'stripe_addr' which is where we will write to.
6293          */
6294         if (mddev->reshape_backwards) {
6295                 if (WARN_ON(conf->reshape_progress == 0))
6296                         return MaxSector;
6297
6298                 stripe_addr = writepos;
6299                 if (WARN_ON((mddev->dev_sectors &
6300                     ~((sector_t)reshape_sectors - 1)) -
6301                     reshape_sectors - stripe_addr != sector_nr))
6302                         return MaxSector;
6303         } else {
6304                 if (WARN_ON(writepos != sector_nr + reshape_sectors))
6305                         return MaxSector;
6306
6307                 stripe_addr = sector_nr;
6308         }
6309
6310         /* 'writepos' is the most advanced device address we might write.
6311          * 'readpos' is the least advanced device address we might read.
6312          * 'safepos' is the least address recorded in the metadata as having
6313          *     been reshaped.
6314          * If there is a min_offset_diff, these are adjusted either by
6315          * increasing the safepos/readpos if diff is negative, or
6316          * increasing writepos if diff is positive.
6317          * If 'readpos' is then behind 'writepos', there is no way that we can
6318          * ensure safety in the face of a crash - that must be done by userspace
6319          * making a backup of the data.  So in that case there is no particular
6320          * rush to update metadata.
6321          * Otherwise if 'safepos' is behind 'writepos', then we really need to
6322          * update the metadata to advance 'safepos' to match 'readpos' so that
6323          * we can be safe in the event of a crash.
6324          * So we insist on updating metadata if safepos is behind writepos and
6325          * readpos is beyond writepos.
6326          * In any case, update the metadata every 10 seconds.
6327          * Maybe that number should be configurable, but I'm not sure it is
6328          * worth it.... maybe it could be a multiple of safemode_delay???
6329          */
6330         if (conf->min_offset_diff < 0) {
6331                 safepos += -conf->min_offset_diff;
6332                 readpos += -conf->min_offset_diff;
6333         } else
6334                 writepos += conf->min_offset_diff;
6335
6336         if ((mddev->reshape_backwards
6337              ? (safepos > writepos && readpos < writepos)
6338              : (safepos < writepos && readpos > writepos)) ||
6339             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
6340                 /* Cannot proceed until we've updated the superblock... */
6341                 wait_event(conf->wait_for_overlap,
6342                            atomic_read(&conf->reshape_stripes)==0
6343                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6344                 if (atomic_read(&conf->reshape_stripes) != 0)
6345                         return 0;
6346                 mddev->reshape_position = conf->reshape_progress;
6347                 mddev->curr_resync_completed = sector_nr;
6348                 if (!mddev->reshape_backwards)
6349                         /* Can update recovery_offset */
6350                         rdev_for_each(rdev, mddev)
6351                                 if (rdev->raid_disk >= 0 &&
6352                                     !test_bit(Journal, &rdev->flags) &&
6353                                     !test_bit(In_sync, &rdev->flags) &&
6354                                     rdev->recovery_offset < sector_nr)
6355                                         rdev->recovery_offset = sector_nr;
6356
6357                 conf->reshape_checkpoint = jiffies;
6358                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6359                 md_wakeup_thread(mddev->thread);
6360                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
6361                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6362                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6363                         return 0;
6364                 spin_lock_irq(&conf->device_lock);
6365                 conf->reshape_safe = mddev->reshape_position;
6366                 spin_unlock_irq(&conf->device_lock);
6367                 wake_up(&conf->wait_for_overlap);
6368                 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6369         }
6370
6371         INIT_LIST_HEAD(&stripes);
6372         for (i = 0; i < reshape_sectors; i += RAID5_STRIPE_SECTORS(conf)) {
6373                 int j;
6374                 int skipped_disk = 0;
6375                 sh = raid5_get_active_stripe(conf, NULL, stripe_addr+i,
6376                                              R5_GAS_NOQUIESCE);
6377                 set_bit(STRIPE_EXPANDING, &sh->state);
6378                 atomic_inc(&conf->reshape_stripes);
6379                 /* If any of this stripe is beyond the end of the old
6380                  * array, then we need to zero those blocks
6381                  */
6382                 for (j=sh->disks; j--;) {
6383                         sector_t s;
6384                         if (j == sh->pd_idx)
6385                                 continue;
6386                         if (conf->level == 6 &&
6387                             j == sh->qd_idx)
6388                                 continue;
6389                         s = raid5_compute_blocknr(sh, j, 0);
6390                         if (s < raid5_size(mddev, 0, 0)) {
6391                                 skipped_disk = 1;
6392                                 continue;
6393                         }
6394                         memset(page_address(sh->dev[j].page), 0, RAID5_STRIPE_SIZE(conf));
6395                         set_bit(R5_Expanded, &sh->dev[j].flags);
6396                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
6397                 }
6398                 if (!skipped_disk) {
6399                         set_bit(STRIPE_EXPAND_READY, &sh->state);
6400                         set_bit(STRIPE_HANDLE, &sh->state);
6401                 }
6402                 list_add(&sh->lru, &stripes);
6403         }
6404         spin_lock_irq(&conf->device_lock);
6405         if (mddev->reshape_backwards)
6406                 conf->reshape_progress -= reshape_sectors * new_data_disks;
6407         else
6408                 conf->reshape_progress += reshape_sectors * new_data_disks;
6409         spin_unlock_irq(&conf->device_lock);
6410         /* Ok, those stripe are ready. We can start scheduling
6411          * reads on the source stripes.
6412          * The source stripes are determined by mapping the first and last
6413          * block on the destination stripes.
6414          */
6415         first_sector =
6416                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
6417                                      1, &dd_idx, NULL);
6418         last_sector =
6419                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
6420                                             * new_data_disks - 1),
6421                                      1, &dd_idx, NULL);
6422         if (last_sector >= mddev->dev_sectors)
6423                 last_sector = mddev->dev_sectors - 1;
6424         while (first_sector <= last_sector) {
6425                 sh = raid5_get_active_stripe(conf, NULL, first_sector,
6426                                 R5_GAS_PREVIOUS | R5_GAS_NOQUIESCE);
6427                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
6428                 set_bit(STRIPE_HANDLE, &sh->state);
6429                 raid5_release_stripe(sh);
6430                 first_sector += RAID5_STRIPE_SECTORS(conf);
6431         }
6432         /* Now that the sources are clearly marked, we can release
6433          * the destination stripes
6434          */
6435         while (!list_empty(&stripes)) {
6436                 sh = list_entry(stripes.next, struct stripe_head, lru);
6437                 list_del_init(&sh->lru);
6438                 raid5_release_stripe(sh);
6439         }
6440         /* If this takes us to the resync_max point where we have to pause,
6441          * then we need to write out the superblock.
6442          */
6443         sector_nr += reshape_sectors;
6444         retn = reshape_sectors;
6445 finish:
6446         if (mddev->curr_resync_completed > mddev->resync_max ||
6447             (sector_nr - mddev->curr_resync_completed) * 2
6448             >= mddev->resync_max - mddev->curr_resync_completed) {
6449                 /* Cannot proceed until we've updated the superblock... */
6450                 wait_event(conf->wait_for_overlap,
6451                            atomic_read(&conf->reshape_stripes) == 0
6452                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6453                 if (atomic_read(&conf->reshape_stripes) != 0)
6454                         goto ret;
6455                 mddev->reshape_position = conf->reshape_progress;
6456                 mddev->curr_resync_completed = sector_nr;
6457                 if (!mddev->reshape_backwards)
6458                         /* Can update recovery_offset */
6459                         rdev_for_each(rdev, mddev)
6460                                 if (rdev->raid_disk >= 0 &&
6461                                     !test_bit(Journal, &rdev->flags) &&
6462                                     !test_bit(In_sync, &rdev->flags) &&
6463                                     rdev->recovery_offset < sector_nr)
6464                                         rdev->recovery_offset = sector_nr;
6465                 conf->reshape_checkpoint = jiffies;
6466                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6467                 md_wakeup_thread(mddev->thread);
6468                 wait_event(mddev->sb_wait,
6469                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6470                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6471                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6472                         goto ret;
6473                 spin_lock_irq(&conf->device_lock);
6474                 conf->reshape_safe = mddev->reshape_position;
6475                 spin_unlock_irq(&conf->device_lock);
6476                 wake_up(&conf->wait_for_overlap);
6477                 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6478         }
6479 ret:
6480         return retn;
6481 }
6482
6483 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6484                                           sector_t max_sector, int *skipped)
6485 {
6486         struct r5conf *conf = mddev->private;
6487         struct stripe_head *sh;
6488         sector_t sync_blocks;
6489         int still_degraded = 0;
6490         int i;
6491
6492         if (sector_nr >= max_sector) {
6493                 /* just being told to finish up .. nothing much to do */
6494
6495                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6496                         end_reshape(conf);
6497                         return 0;
6498                 }
6499
6500                 if (mddev->curr_resync < max_sector) /* aborted */
6501                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6502                                            &sync_blocks, 1);
6503                 else /* completed sync */
6504                         conf->fullsync = 0;
6505                 md_bitmap_close_sync(mddev->bitmap);
6506
6507                 return 0;
6508         }
6509
6510         /* Allow raid5_quiesce to complete */
6511         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6512
6513         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6514                 return reshape_request(mddev, sector_nr, skipped);
6515
6516         /* No need to check resync_max as we never do more than one
6517          * stripe, and as resync_max will always be on a chunk boundary,
6518          * if the check in md_do_sync didn't fire, there is no chance
6519          * of overstepping resync_max here
6520          */
6521
6522         /* if there is too many failed drives and we are trying
6523          * to resync, then assert that we are finished, because there is
6524          * nothing we can do.
6525          */
6526         if (mddev->degraded >= conf->max_degraded &&
6527             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6528                 sector_t rv = mddev->dev_sectors - sector_nr;
6529                 *skipped = 1;
6530                 return rv;
6531         }
6532         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6533             !conf->fullsync &&
6534             !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6535             sync_blocks >= RAID5_STRIPE_SECTORS(conf)) {
6536                 /* we can skip this block, and probably more */
6537                 do_div(sync_blocks, RAID5_STRIPE_SECTORS(conf));
6538                 *skipped = 1;
6539                 /* keep things rounded to whole stripes */
6540                 return sync_blocks * RAID5_STRIPE_SECTORS(conf);
6541         }
6542
6543         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6544
6545         sh = raid5_get_active_stripe(conf, NULL, sector_nr,
6546                                      R5_GAS_NOBLOCK);
6547         if (sh == NULL) {
6548                 sh = raid5_get_active_stripe(conf, NULL, sector_nr, 0);
6549                 /* make sure we don't swamp the stripe cache if someone else
6550                  * is trying to get access
6551                  */
6552                 schedule_timeout_uninterruptible(1);
6553         }
6554         /* Need to check if array will still be degraded after recovery/resync
6555          * Note in case of > 1 drive failures it's possible we're rebuilding
6556          * one drive while leaving another faulty drive in array.
6557          */
6558         for (i = 0; i < conf->raid_disks; i++) {
6559                 struct md_rdev *rdev = conf->disks[i].rdev;
6560
6561                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6562                         still_degraded = 1;
6563         }
6564
6565         md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6566
6567         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6568         set_bit(STRIPE_HANDLE, &sh->state);
6569
6570         raid5_release_stripe(sh);
6571
6572         return RAID5_STRIPE_SECTORS(conf);
6573 }
6574
6575 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6576                                unsigned int offset)
6577 {
6578         /* We may not be able to submit a whole bio at once as there
6579          * may not be enough stripe_heads available.
6580          * We cannot pre-allocate enough stripe_heads as we may need
6581          * more than exist in the cache (if we allow ever large chunks).
6582          * So we do one stripe head at a time and record in
6583          * ->bi_hw_segments how many have been done.
6584          *
6585          * We *know* that this entire raid_bio is in one chunk, so
6586          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6587          */
6588         struct stripe_head *sh;
6589         int dd_idx;
6590         sector_t sector, logical_sector, last_sector;
6591         int scnt = 0;
6592         int handled = 0;
6593
6594         logical_sector = raid_bio->bi_iter.bi_sector &
6595                 ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6596         sector = raid5_compute_sector(conf, logical_sector,
6597                                       0, &dd_idx, NULL);
6598         last_sector = bio_end_sector(raid_bio);
6599
6600         for (; logical_sector < last_sector;
6601              logical_sector += RAID5_STRIPE_SECTORS(conf),
6602                      sector += RAID5_STRIPE_SECTORS(conf),
6603                      scnt++) {
6604
6605                 if (scnt < offset)
6606                         /* already done this stripe */
6607                         continue;
6608
6609                 sh = raid5_get_active_stripe(conf, NULL, sector,
6610                                 R5_GAS_NOBLOCK | R5_GAS_NOQUIESCE);
6611                 if (!sh) {
6612                         /* failed to get a stripe - must wait */
6613                         conf->retry_read_aligned = raid_bio;
6614                         conf->retry_read_offset = scnt;
6615                         return handled;
6616                 }
6617
6618                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6619                         raid5_release_stripe(sh);
6620                         conf->retry_read_aligned = raid_bio;
6621                         conf->retry_read_offset = scnt;
6622                         return handled;
6623                 }
6624
6625                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6626                 handle_stripe(sh);
6627                 raid5_release_stripe(sh);
6628                 handled++;
6629         }
6630
6631         bio_endio(raid_bio);
6632
6633         if (atomic_dec_and_test(&conf->active_aligned_reads))
6634                 wake_up(&conf->wait_for_quiescent);
6635         return handled;
6636 }
6637
6638 static int handle_active_stripes(struct r5conf *conf, int group,
6639                                  struct r5worker *worker,
6640                                  struct list_head *temp_inactive_list)
6641                 __must_hold(&conf->device_lock)
6642 {
6643         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6644         int i, batch_size = 0, hash;
6645         bool release_inactive = false;
6646
6647         while (batch_size < MAX_STRIPE_BATCH &&
6648                         (sh = __get_priority_stripe(conf, group)) != NULL)
6649                 batch[batch_size++] = sh;
6650
6651         if (batch_size == 0) {
6652                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6653                         if (!list_empty(temp_inactive_list + i))
6654                                 break;
6655                 if (i == NR_STRIPE_HASH_LOCKS) {
6656                         spin_unlock_irq(&conf->device_lock);
6657                         log_flush_stripe_to_raid(conf);
6658                         spin_lock_irq(&conf->device_lock);
6659                         return batch_size;
6660                 }
6661                 release_inactive = true;
6662         }
6663         spin_unlock_irq(&conf->device_lock);
6664
6665         release_inactive_stripe_list(conf, temp_inactive_list,
6666                                      NR_STRIPE_HASH_LOCKS);
6667
6668         r5l_flush_stripe_to_raid(conf->log);
6669         if (release_inactive) {
6670                 spin_lock_irq(&conf->device_lock);
6671                 return 0;
6672         }
6673
6674         for (i = 0; i < batch_size; i++)
6675                 handle_stripe(batch[i]);
6676         log_write_stripe_run(conf);
6677
6678         cond_resched();
6679
6680         spin_lock_irq(&conf->device_lock);
6681         for (i = 0; i < batch_size; i++) {
6682                 hash = batch[i]->hash_lock_index;
6683                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6684         }
6685         return batch_size;
6686 }
6687
6688 static void raid5_do_work(struct work_struct *work)
6689 {
6690         struct r5worker *worker = container_of(work, struct r5worker, work);
6691         struct r5worker_group *group = worker->group;
6692         struct r5conf *conf = group->conf;
6693         struct mddev *mddev = conf->mddev;
6694         int group_id = group - conf->worker_groups;
6695         int handled;
6696         struct blk_plug plug;
6697
6698         pr_debug("+++ raid5worker active\n");
6699
6700         blk_start_plug(&plug);
6701         handled = 0;
6702         spin_lock_irq(&conf->device_lock);
6703         while (1) {
6704                 int batch_size, released;
6705
6706                 released = release_stripe_list(conf, worker->temp_inactive_list);
6707
6708                 batch_size = handle_active_stripes(conf, group_id, worker,
6709                                                    worker->temp_inactive_list);
6710                 worker->working = false;
6711                 if (!batch_size && !released)
6712                         break;
6713                 handled += batch_size;
6714                 wait_event_lock_irq(mddev->sb_wait,
6715                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6716                         conf->device_lock);
6717         }
6718         pr_debug("%d stripes handled\n", handled);
6719
6720         spin_unlock_irq(&conf->device_lock);
6721
6722         flush_deferred_bios(conf);
6723
6724         r5l_flush_stripe_to_raid(conf->log);
6725
6726         async_tx_issue_pending_all();
6727         blk_finish_plug(&plug);
6728
6729         pr_debug("--- raid5worker inactive\n");
6730 }
6731
6732 /*
6733  * This is our raid5 kernel thread.
6734  *
6735  * We scan the hash table for stripes which can be handled now.
6736  * During the scan, completed stripes are saved for us by the interrupt
6737  * handler, so that they will not have to wait for our next wakeup.
6738  */
6739 static void raid5d(struct md_thread *thread)
6740 {
6741         struct mddev *mddev = thread->mddev;
6742         struct r5conf *conf = mddev->private;
6743         int handled;
6744         struct blk_plug plug;
6745
6746         pr_debug("+++ raid5d active\n");
6747
6748         md_check_recovery(mddev);
6749
6750         blk_start_plug(&plug);
6751         handled = 0;
6752         spin_lock_irq(&conf->device_lock);
6753         while (1) {
6754                 struct bio *bio;
6755                 int batch_size, released;
6756                 unsigned int offset;
6757
6758                 if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
6759                         break;
6760
6761                 released = release_stripe_list(conf, conf->temp_inactive_list);
6762                 if (released)
6763                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6764
6765                 if (
6766                     !list_empty(&conf->bitmap_list)) {
6767                         /* Now is a good time to flush some bitmap updates */
6768                         conf->seq_flush++;
6769                         spin_unlock_irq(&conf->device_lock);
6770                         md_bitmap_unplug(mddev->bitmap);
6771                         spin_lock_irq(&conf->device_lock);
6772                         conf->seq_write = conf->seq_flush;
6773                         activate_bit_delay(conf, conf->temp_inactive_list);
6774                 }
6775                 raid5_activate_delayed(conf);
6776
6777                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6778                         int ok;
6779                         spin_unlock_irq(&conf->device_lock);
6780                         ok = retry_aligned_read(conf, bio, offset);
6781                         spin_lock_irq(&conf->device_lock);
6782                         if (!ok)
6783                                 break;
6784                         handled++;
6785                 }
6786
6787                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6788                                                    conf->temp_inactive_list);
6789                 if (!batch_size && !released)
6790                         break;
6791                 handled += batch_size;
6792
6793                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6794                         spin_unlock_irq(&conf->device_lock);
6795                         md_check_recovery(mddev);
6796                         spin_lock_irq(&conf->device_lock);
6797                 }
6798         }
6799         pr_debug("%d stripes handled\n", handled);
6800
6801         spin_unlock_irq(&conf->device_lock);
6802         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6803             mutex_trylock(&conf->cache_size_mutex)) {
6804                 grow_one_stripe(conf, __GFP_NOWARN);
6805                 /* Set flag even if allocation failed.  This helps
6806                  * slow down allocation requests when mem is short
6807                  */
6808                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6809                 mutex_unlock(&conf->cache_size_mutex);
6810         }
6811
6812         flush_deferred_bios(conf);
6813
6814         r5l_flush_stripe_to_raid(conf->log);
6815
6816         async_tx_issue_pending_all();
6817         blk_finish_plug(&plug);
6818
6819         pr_debug("--- raid5d inactive\n");
6820 }
6821
6822 static ssize_t
6823 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6824 {
6825         struct r5conf *conf;
6826         int ret = 0;
6827         spin_lock(&mddev->lock);
6828         conf = mddev->private;
6829         if (conf)
6830                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6831         spin_unlock(&mddev->lock);
6832         return ret;
6833 }
6834
6835 int
6836 raid5_set_cache_size(struct mddev *mddev, int size)
6837 {
6838         int result = 0;
6839         struct r5conf *conf = mddev->private;
6840
6841         if (size <= 16 || size > 32768)
6842                 return -EINVAL;
6843
6844         WRITE_ONCE(conf->min_nr_stripes, size);
6845         mutex_lock(&conf->cache_size_mutex);
6846         while (size < conf->max_nr_stripes &&
6847                drop_one_stripe(conf))
6848                 ;
6849         mutex_unlock(&conf->cache_size_mutex);
6850
6851         md_allow_write(mddev);
6852
6853         mutex_lock(&conf->cache_size_mutex);
6854         while (size > conf->max_nr_stripes)
6855                 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6856                         WRITE_ONCE(conf->min_nr_stripes, conf->max_nr_stripes);
6857                         result = -ENOMEM;
6858                         break;
6859                 }
6860         mutex_unlock(&conf->cache_size_mutex);
6861
6862         return result;
6863 }
6864 EXPORT_SYMBOL(raid5_set_cache_size);
6865
6866 static ssize_t
6867 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6868 {
6869         struct r5conf *conf;
6870         unsigned long new;
6871         int err;
6872
6873         if (len >= PAGE_SIZE)
6874                 return -EINVAL;
6875         if (kstrtoul(page, 10, &new))
6876                 return -EINVAL;
6877         err = mddev_lock(mddev);
6878         if (err)
6879                 return err;
6880         conf = mddev->private;
6881         if (!conf)
6882                 err = -ENODEV;
6883         else
6884                 err = raid5_set_cache_size(mddev, new);
6885         mddev_unlock(mddev);
6886
6887         return err ?: len;
6888 }
6889
6890 static struct md_sysfs_entry
6891 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6892                                 raid5_show_stripe_cache_size,
6893                                 raid5_store_stripe_cache_size);
6894
6895 static ssize_t
6896 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6897 {
6898         struct r5conf *conf = mddev->private;
6899         if (conf)
6900                 return sprintf(page, "%d\n", conf->rmw_level);
6901         else
6902                 return 0;
6903 }
6904
6905 static ssize_t
6906 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6907 {
6908         struct r5conf *conf = mddev->private;
6909         unsigned long new;
6910
6911         if (!conf)
6912                 return -ENODEV;
6913
6914         if (len >= PAGE_SIZE)
6915                 return -EINVAL;
6916
6917         if (kstrtoul(page, 10, &new))
6918                 return -EINVAL;
6919
6920         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6921                 return -EINVAL;
6922
6923         if (new != PARITY_DISABLE_RMW &&
6924             new != PARITY_ENABLE_RMW &&
6925             new != PARITY_PREFER_RMW)
6926                 return -EINVAL;
6927
6928         conf->rmw_level = new;
6929         return len;
6930 }
6931
6932 static struct md_sysfs_entry
6933 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6934                          raid5_show_rmw_level,
6935                          raid5_store_rmw_level);
6936
6937 static ssize_t
6938 raid5_show_stripe_size(struct mddev  *mddev, char *page)
6939 {
6940         struct r5conf *conf;
6941         int ret = 0;
6942
6943         spin_lock(&mddev->lock);
6944         conf = mddev->private;
6945         if (conf)
6946                 ret = sprintf(page, "%lu\n", RAID5_STRIPE_SIZE(conf));
6947         spin_unlock(&mddev->lock);
6948         return ret;
6949 }
6950
6951 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
6952 static ssize_t
6953 raid5_store_stripe_size(struct mddev  *mddev, const char *page, size_t len)
6954 {
6955         struct r5conf *conf;
6956         unsigned long new;
6957         int err;
6958         int size;
6959
6960         if (len >= PAGE_SIZE)
6961                 return -EINVAL;
6962         if (kstrtoul(page, 10, &new))
6963                 return -EINVAL;
6964
6965         /*
6966          * The value should not be bigger than PAGE_SIZE. It requires to
6967          * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
6968          * of two.
6969          */
6970         if (new % DEFAULT_STRIPE_SIZE != 0 ||
6971                         new > PAGE_SIZE || new == 0 ||
6972                         new != roundup_pow_of_two(new))
6973                 return -EINVAL;
6974
6975         err = mddev_suspend_and_lock(mddev);
6976         if (err)
6977                 return err;
6978
6979         conf = mddev->private;
6980         if (!conf) {
6981                 err = -ENODEV;
6982                 goto out_unlock;
6983         }
6984
6985         if (new == conf->stripe_size)
6986                 goto out_unlock;
6987
6988         pr_debug("md/raid: change stripe_size from %lu to %lu\n",
6989                         conf->stripe_size, new);
6990
6991         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
6992             mddev->reshape_position != MaxSector || mddev->sysfs_active) {
6993                 err = -EBUSY;
6994                 goto out_unlock;
6995         }
6996
6997         mutex_lock(&conf->cache_size_mutex);
6998         size = conf->max_nr_stripes;
6999
7000         shrink_stripes(conf);
7001
7002         conf->stripe_size = new;
7003         conf->stripe_shift = ilog2(new) - 9;
7004         conf->stripe_sectors = new >> 9;
7005         if (grow_stripes(conf, size)) {
7006                 pr_warn("md/raid:%s: couldn't allocate buffers\n",
7007                                 mdname(mddev));
7008                 err = -ENOMEM;
7009         }
7010         mutex_unlock(&conf->cache_size_mutex);
7011
7012 out_unlock:
7013         mddev_unlock_and_resume(mddev);
7014         return err ?: len;
7015 }
7016
7017 static struct md_sysfs_entry
7018 raid5_stripe_size = __ATTR(stripe_size, 0644,
7019                          raid5_show_stripe_size,
7020                          raid5_store_stripe_size);
7021 #else
7022 static struct md_sysfs_entry
7023 raid5_stripe_size = __ATTR(stripe_size, 0444,
7024                          raid5_show_stripe_size,
7025                          NULL);
7026 #endif
7027
7028 static ssize_t
7029 raid5_show_preread_threshold(struct mddev *mddev, char *page)
7030 {
7031         struct r5conf *conf;
7032         int ret = 0;
7033         spin_lock(&mddev->lock);
7034         conf = mddev->private;
7035         if (conf)
7036                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
7037         spin_unlock(&mddev->lock);
7038         return ret;
7039 }
7040
7041 static ssize_t
7042 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
7043 {
7044         struct r5conf *conf;
7045         unsigned long new;
7046         int err;
7047
7048         if (len >= PAGE_SIZE)
7049                 return -EINVAL;
7050         if (kstrtoul(page, 10, &new))
7051                 return -EINVAL;
7052
7053         err = mddev_lock(mddev);
7054         if (err)
7055                 return err;
7056         conf = mddev->private;
7057         if (!conf)
7058                 err = -ENODEV;
7059         else if (new > conf->min_nr_stripes)
7060                 err = -EINVAL;
7061         else
7062                 conf->bypass_threshold = new;
7063         mddev_unlock(mddev);
7064         return err ?: len;
7065 }
7066
7067 static struct md_sysfs_entry
7068 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
7069                                         S_IRUGO | S_IWUSR,
7070                                         raid5_show_preread_threshold,
7071                                         raid5_store_preread_threshold);
7072
7073 static ssize_t
7074 raid5_show_skip_copy(struct mddev *mddev, char *page)
7075 {
7076         struct r5conf *conf;
7077         int ret = 0;
7078         spin_lock(&mddev->lock);
7079         conf = mddev->private;
7080         if (conf)
7081                 ret = sprintf(page, "%d\n", conf->skip_copy);
7082         spin_unlock(&mddev->lock);
7083         return ret;
7084 }
7085
7086 static ssize_t
7087 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
7088 {
7089         struct r5conf *conf;
7090         unsigned long new;
7091         int err;
7092
7093         if (len >= PAGE_SIZE)
7094                 return -EINVAL;
7095         if (kstrtoul(page, 10, &new))
7096                 return -EINVAL;
7097         new = !!new;
7098
7099         err = mddev_suspend_and_lock(mddev);
7100         if (err)
7101                 return err;
7102         conf = mddev->private;
7103         if (!conf)
7104                 err = -ENODEV;
7105         else if (new != conf->skip_copy) {
7106                 struct request_queue *q = mddev->gendisk->queue;
7107                 struct queue_limits lim = queue_limits_start_update(q);
7108
7109                 conf->skip_copy = new;
7110                 if (new)
7111                         lim.features |= BLK_FEAT_STABLE_WRITES;
7112                 else
7113                         lim.features &= ~BLK_FEAT_STABLE_WRITES;
7114                 err = queue_limits_commit_update(q, &lim);
7115         }
7116         mddev_unlock_and_resume(mddev);
7117         return err ?: len;
7118 }
7119
7120 static struct md_sysfs_entry
7121 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
7122                                         raid5_show_skip_copy,
7123                                         raid5_store_skip_copy);
7124
7125 static ssize_t
7126 stripe_cache_active_show(struct mddev *mddev, char *page)
7127 {
7128         struct r5conf *conf = mddev->private;
7129         if (conf)
7130                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
7131         else
7132                 return 0;
7133 }
7134
7135 static struct md_sysfs_entry
7136 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
7137
7138 static ssize_t
7139 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
7140 {
7141         struct r5conf *conf;
7142         int ret = 0;
7143         spin_lock(&mddev->lock);
7144         conf = mddev->private;
7145         if (conf)
7146                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
7147         spin_unlock(&mddev->lock);
7148         return ret;
7149 }
7150
7151 static int alloc_thread_groups(struct r5conf *conf, int cnt,
7152                                int *group_cnt,
7153                                struct r5worker_group **worker_groups);
7154 static ssize_t
7155 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
7156 {
7157         struct r5conf *conf;
7158         unsigned int new;
7159         int err;
7160         struct r5worker_group *new_groups, *old_groups;
7161         int group_cnt;
7162
7163         if (len >= PAGE_SIZE)
7164                 return -EINVAL;
7165         if (kstrtouint(page, 10, &new))
7166                 return -EINVAL;
7167         /* 8192 should be big enough */
7168         if (new > 8192)
7169                 return -EINVAL;
7170
7171         err = mddev_suspend_and_lock(mddev);
7172         if (err)
7173                 return err;
7174         conf = mddev->private;
7175         if (!conf)
7176                 err = -ENODEV;
7177         else if (new != conf->worker_cnt_per_group) {
7178                 old_groups = conf->worker_groups;
7179                 if (old_groups)
7180                         flush_workqueue(raid5_wq);
7181
7182                 err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
7183                 if (!err) {
7184                         spin_lock_irq(&conf->device_lock);
7185                         conf->group_cnt = group_cnt;
7186                         conf->worker_cnt_per_group = new;
7187                         conf->worker_groups = new_groups;
7188                         spin_unlock_irq(&conf->device_lock);
7189
7190                         if (old_groups)
7191                                 kfree(old_groups[0].workers);
7192                         kfree(old_groups);
7193                 }
7194         }
7195         mddev_unlock_and_resume(mddev);
7196
7197         return err ?: len;
7198 }
7199
7200 static struct md_sysfs_entry
7201 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
7202                                 raid5_show_group_thread_cnt,
7203                                 raid5_store_group_thread_cnt);
7204
7205 static struct attribute *raid5_attrs[] =  {
7206         &raid5_stripecache_size.attr,
7207         &raid5_stripecache_active.attr,
7208         &raid5_preread_bypass_threshold.attr,
7209         &raid5_group_thread_cnt.attr,
7210         &raid5_skip_copy.attr,
7211         &raid5_rmw_level.attr,
7212         &raid5_stripe_size.attr,
7213         &r5c_journal_mode.attr,
7214         &ppl_write_hint.attr,
7215         NULL,
7216 };
7217 static const struct attribute_group raid5_attrs_group = {
7218         .name = NULL,
7219         .attrs = raid5_attrs,
7220 };
7221
7222 static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
7223                                struct r5worker_group **worker_groups)
7224 {
7225         int i, j, k;
7226         ssize_t size;
7227         struct r5worker *workers;
7228
7229         if (cnt == 0) {
7230                 *group_cnt = 0;
7231                 *worker_groups = NULL;
7232                 return 0;
7233         }
7234         *group_cnt = num_possible_nodes();
7235         size = sizeof(struct r5worker) * cnt;
7236         workers = kcalloc(size, *group_cnt, GFP_NOIO);
7237         *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
7238                                  GFP_NOIO);
7239         if (!*worker_groups || !workers) {
7240                 kfree(workers);
7241                 kfree(*worker_groups);
7242                 return -ENOMEM;
7243         }
7244
7245         for (i = 0; i < *group_cnt; i++) {
7246                 struct r5worker_group *group;
7247
7248                 group = &(*worker_groups)[i];
7249                 INIT_LIST_HEAD(&group->handle_list);
7250                 INIT_LIST_HEAD(&group->loprio_list);
7251                 group->conf = conf;
7252                 group->workers = workers + i * cnt;
7253
7254                 for (j = 0; j < cnt; j++) {
7255                         struct r5worker *worker = group->workers + j;
7256                         worker->group = group;
7257                         INIT_WORK(&worker->work, raid5_do_work);
7258
7259                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
7260                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
7261                 }
7262         }
7263
7264         return 0;
7265 }
7266
7267 static void free_thread_groups(struct r5conf *conf)
7268 {
7269         if (conf->worker_groups)
7270                 kfree(conf->worker_groups[0].workers);
7271         kfree(conf->worker_groups);
7272         conf->worker_groups = NULL;
7273 }
7274
7275 static sector_t
7276 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
7277 {
7278         struct r5conf *conf = mddev->private;
7279
7280         if (!sectors)
7281                 sectors = mddev->dev_sectors;
7282         if (!raid_disks)
7283                 /* size is defined by the smallest of previous and new size */
7284                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
7285
7286         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7287         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
7288         return sectors * (raid_disks - conf->max_degraded);
7289 }
7290
7291 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7292 {
7293         safe_put_page(percpu->spare_page);
7294         percpu->spare_page = NULL;
7295         kvfree(percpu->scribble);
7296         percpu->scribble = NULL;
7297 }
7298
7299 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7300 {
7301         if (conf->level == 6 && !percpu->spare_page) {
7302                 percpu->spare_page = alloc_page(GFP_KERNEL);
7303                 if (!percpu->spare_page)
7304                         return -ENOMEM;
7305         }
7306
7307         if (scribble_alloc(percpu,
7308                            max(conf->raid_disks,
7309                                conf->previous_raid_disks),
7310                            max(conf->chunk_sectors,
7311                                conf->prev_chunk_sectors)
7312                            / RAID5_STRIPE_SECTORS(conf))) {
7313                 free_scratch_buffer(conf, percpu);
7314                 return -ENOMEM;
7315         }
7316
7317         local_lock_init(&percpu->lock);
7318         return 0;
7319 }
7320
7321 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
7322 {
7323         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7324
7325         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
7326         return 0;
7327 }
7328
7329 static void raid5_free_percpu(struct r5conf *conf)
7330 {
7331         if (!conf->percpu)
7332                 return;
7333
7334         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7335         free_percpu(conf->percpu);
7336 }
7337
7338 static void free_conf(struct r5conf *conf)
7339 {
7340         int i;
7341
7342         log_exit(conf);
7343
7344         shrinker_free(conf->shrinker);
7345         free_thread_groups(conf);
7346         shrink_stripes(conf);
7347         raid5_free_percpu(conf);
7348         for (i = 0; i < conf->pool_size; i++)
7349                 if (conf->disks[i].extra_page)
7350                         put_page(conf->disks[i].extra_page);
7351         kfree(conf->disks);
7352         bioset_exit(&conf->bio_split);
7353         kfree(conf->stripe_hashtbl);
7354         kfree(conf->pending_data);
7355         kfree(conf);
7356 }
7357
7358 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
7359 {
7360         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7361         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
7362
7363         if (alloc_scratch_buffer(conf, percpu)) {
7364                 pr_warn("%s: failed memory allocation for cpu%u\n",
7365                         __func__, cpu);
7366                 return -ENOMEM;
7367         }
7368         return 0;
7369 }
7370
7371 static int raid5_alloc_percpu(struct r5conf *conf)
7372 {
7373         int err = 0;
7374
7375         conf->percpu = alloc_percpu(struct raid5_percpu);
7376         if (!conf->percpu)
7377                 return -ENOMEM;
7378
7379         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7380         if (!err) {
7381                 conf->scribble_disks = max(conf->raid_disks,
7382                         conf->previous_raid_disks);
7383                 conf->scribble_sectors = max(conf->chunk_sectors,
7384                         conf->prev_chunk_sectors);
7385         }
7386         return err;
7387 }
7388
7389 static unsigned long raid5_cache_scan(struct shrinker *shrink,
7390                                       struct shrink_control *sc)
7391 {
7392         struct r5conf *conf = shrink->private_data;
7393         unsigned long ret = SHRINK_STOP;
7394
7395         if (mutex_trylock(&conf->cache_size_mutex)) {
7396                 ret= 0;
7397                 while (ret < sc->nr_to_scan &&
7398                        conf->max_nr_stripes > conf->min_nr_stripes) {
7399                         if (drop_one_stripe(conf) == 0) {
7400                                 ret = SHRINK_STOP;
7401                                 break;
7402                         }
7403                         ret++;
7404                 }
7405                 mutex_unlock(&conf->cache_size_mutex);
7406         }
7407         return ret;
7408 }
7409
7410 static unsigned long raid5_cache_count(struct shrinker *shrink,
7411                                        struct shrink_control *sc)
7412 {
7413         struct r5conf *conf = shrink->private_data;
7414         int max_stripes = READ_ONCE(conf->max_nr_stripes);
7415         int min_stripes = READ_ONCE(conf->min_nr_stripes);
7416
7417         if (max_stripes < min_stripes)
7418                 /* unlikely, but not impossible */
7419                 return 0;
7420         return max_stripes - min_stripes;
7421 }
7422
7423 static struct r5conf *setup_conf(struct mddev *mddev)
7424 {
7425         struct r5conf *conf;
7426         int raid_disk, memory, max_disks;
7427         struct md_rdev *rdev;
7428         struct disk_info *disk;
7429         char pers_name[6];
7430         int i;
7431         int group_cnt;
7432         struct r5worker_group *new_group;
7433         int ret = -ENOMEM;
7434
7435         if (mddev->new_level != 5
7436             && mddev->new_level != 4
7437             && mddev->new_level != 6) {
7438                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
7439                         mdname(mddev), mddev->new_level);
7440                 return ERR_PTR(-EIO);
7441         }
7442         if ((mddev->new_level == 5
7443              && !algorithm_valid_raid5(mddev->new_layout)) ||
7444             (mddev->new_level == 6
7445              && !algorithm_valid_raid6(mddev->new_layout))) {
7446                 pr_warn("md/raid:%s: layout %d not supported\n",
7447                         mdname(mddev), mddev->new_layout);
7448                 return ERR_PTR(-EIO);
7449         }
7450         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
7451                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
7452                         mdname(mddev), mddev->raid_disks);
7453                 return ERR_PTR(-EINVAL);
7454         }
7455
7456         if (!mddev->new_chunk_sectors ||
7457             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
7458             !is_power_of_2(mddev->new_chunk_sectors)) {
7459                 pr_warn("md/raid:%s: invalid chunk size %d\n",
7460                         mdname(mddev), mddev->new_chunk_sectors << 9);
7461                 return ERR_PTR(-EINVAL);
7462         }
7463
7464         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
7465         if (conf == NULL)
7466                 goto abort;
7467
7468 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7469         conf->stripe_size = DEFAULT_STRIPE_SIZE;
7470         conf->stripe_shift = ilog2(DEFAULT_STRIPE_SIZE) - 9;
7471         conf->stripe_sectors = DEFAULT_STRIPE_SIZE >> 9;
7472 #endif
7473         INIT_LIST_HEAD(&conf->free_list);
7474         INIT_LIST_HEAD(&conf->pending_list);
7475         conf->pending_data = kcalloc(PENDING_IO_MAX,
7476                                      sizeof(struct r5pending_data),
7477                                      GFP_KERNEL);
7478         if (!conf->pending_data)
7479                 goto abort;
7480         for (i = 0; i < PENDING_IO_MAX; i++)
7481                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
7482         /* Don't enable multi-threading by default*/
7483         if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
7484                 conf->group_cnt = group_cnt;
7485                 conf->worker_cnt_per_group = 0;
7486                 conf->worker_groups = new_group;
7487         } else
7488                 goto abort;
7489         spin_lock_init(&conf->device_lock);
7490         seqcount_spinlock_init(&conf->gen_lock, &conf->device_lock);
7491         mutex_init(&conf->cache_size_mutex);
7492
7493         init_waitqueue_head(&conf->wait_for_quiescent);
7494         init_waitqueue_head(&conf->wait_for_stripe);
7495         init_waitqueue_head(&conf->wait_for_overlap);
7496         INIT_LIST_HEAD(&conf->handle_list);
7497         INIT_LIST_HEAD(&conf->loprio_list);
7498         INIT_LIST_HEAD(&conf->hold_list);
7499         INIT_LIST_HEAD(&conf->delayed_list);
7500         INIT_LIST_HEAD(&conf->bitmap_list);
7501         init_llist_head(&conf->released_stripes);
7502         atomic_set(&conf->active_stripes, 0);
7503         atomic_set(&conf->preread_active_stripes, 0);
7504         atomic_set(&conf->active_aligned_reads, 0);
7505         spin_lock_init(&conf->pending_bios_lock);
7506         conf->batch_bio_dispatch = true;
7507         rdev_for_each(rdev, mddev) {
7508                 if (test_bit(Journal, &rdev->flags))
7509                         continue;
7510                 if (bdev_nonrot(rdev->bdev)) {
7511                         conf->batch_bio_dispatch = false;
7512                         break;
7513                 }
7514         }
7515
7516         conf->bypass_threshold = BYPASS_THRESHOLD;
7517         conf->recovery_disabled = mddev->recovery_disabled - 1;
7518
7519         conf->raid_disks = mddev->raid_disks;
7520         if (mddev->reshape_position == MaxSector)
7521                 conf->previous_raid_disks = mddev->raid_disks;
7522         else
7523                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
7524         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
7525
7526         conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
7527                               GFP_KERNEL);
7528
7529         if (!conf->disks)
7530                 goto abort;
7531
7532         for (i = 0; i < max_disks; i++) {
7533                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
7534                 if (!conf->disks[i].extra_page)
7535                         goto abort;
7536         }
7537
7538         ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
7539         if (ret)
7540                 goto abort;
7541         conf->mddev = mddev;
7542
7543         ret = -ENOMEM;
7544         conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL);
7545         if (!conf->stripe_hashtbl)
7546                 goto abort;
7547
7548         /* We init hash_locks[0] separately to that it can be used
7549          * as the reference lock in the spin_lock_nest_lock() call
7550          * in lock_all_device_hash_locks_irq in order to convince
7551          * lockdep that we know what we are doing.
7552          */
7553         spin_lock_init(conf->hash_locks);
7554         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7555                 spin_lock_init(conf->hash_locks + i);
7556
7557         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7558                 INIT_LIST_HEAD(conf->inactive_list + i);
7559
7560         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7561                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7562
7563         atomic_set(&conf->r5c_cached_full_stripes, 0);
7564         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7565         atomic_set(&conf->r5c_cached_partial_stripes, 0);
7566         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7567         atomic_set(&conf->r5c_flushing_full_stripes, 0);
7568         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7569
7570         conf->level = mddev->new_level;
7571         conf->chunk_sectors = mddev->new_chunk_sectors;
7572         ret = raid5_alloc_percpu(conf);
7573         if (ret)
7574                 goto abort;
7575
7576         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7577
7578         ret = -EIO;
7579         rdev_for_each(rdev, mddev) {
7580                 raid_disk = rdev->raid_disk;
7581                 if (raid_disk >= max_disks
7582                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7583                         continue;
7584                 disk = conf->disks + raid_disk;
7585
7586                 if (test_bit(Replacement, &rdev->flags)) {
7587                         if (disk->replacement)
7588                                 goto abort;
7589                         disk->replacement = rdev;
7590                 } else {
7591                         if (disk->rdev)
7592                                 goto abort;
7593                         disk->rdev = rdev;
7594                 }
7595
7596                 if (test_bit(In_sync, &rdev->flags)) {
7597                         pr_info("md/raid:%s: device %pg operational as raid disk %d\n",
7598                                 mdname(mddev), rdev->bdev, raid_disk);
7599                 } else if (rdev->saved_raid_disk != raid_disk)
7600                         /* Cannot rely on bitmap to complete recovery */
7601                         conf->fullsync = 1;
7602         }
7603
7604         conf->level = mddev->new_level;
7605         if (conf->level == 6) {
7606                 conf->max_degraded = 2;
7607                 if (raid6_call.xor_syndrome)
7608                         conf->rmw_level = PARITY_ENABLE_RMW;
7609                 else
7610                         conf->rmw_level = PARITY_DISABLE_RMW;
7611         } else {
7612                 conf->max_degraded = 1;
7613                 conf->rmw_level = PARITY_ENABLE_RMW;
7614         }
7615         conf->algorithm = mddev->new_layout;
7616         conf->reshape_progress = mddev->reshape_position;
7617         if (conf->reshape_progress != MaxSector) {
7618                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7619                 conf->prev_algo = mddev->layout;
7620         } else {
7621                 conf->prev_chunk_sectors = conf->chunk_sectors;
7622                 conf->prev_algo = conf->algorithm;
7623         }
7624
7625         conf->min_nr_stripes = NR_STRIPES;
7626         if (mddev->reshape_position != MaxSector) {
7627                 int stripes = max_t(int,
7628                         ((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4,
7629                         ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4);
7630                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7631                 if (conf->min_nr_stripes != NR_STRIPES)
7632                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7633                                 mdname(mddev), conf->min_nr_stripes);
7634         }
7635         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7636                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7637         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7638         if (grow_stripes(conf, conf->min_nr_stripes)) {
7639                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7640                         mdname(mddev), memory);
7641                 ret = -ENOMEM;
7642                 goto abort;
7643         } else
7644                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7645         /*
7646          * Losing a stripe head costs more than the time to refill it,
7647          * it reduces the queue depth and so can hurt throughput.
7648          * So set it rather large, scaled by number of devices.
7649          */
7650         conf->shrinker = shrinker_alloc(0, "md-raid5:%s", mdname(mddev));
7651         if (!conf->shrinker) {
7652                 ret = -ENOMEM;
7653                 pr_warn("md/raid:%s: couldn't allocate shrinker.\n",
7654                         mdname(mddev));
7655                 goto abort;
7656         }
7657
7658         conf->shrinker->seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7659         conf->shrinker->scan_objects = raid5_cache_scan;
7660         conf->shrinker->count_objects = raid5_cache_count;
7661         conf->shrinker->batch = 128;
7662         conf->shrinker->private_data = conf;
7663
7664         shrinker_register(conf->shrinker);
7665
7666         sprintf(pers_name, "raid%d", mddev->new_level);
7667         rcu_assign_pointer(conf->thread,
7668                            md_register_thread(raid5d, mddev, pers_name));
7669         if (!conf->thread) {
7670                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7671                         mdname(mddev));
7672                 ret = -ENOMEM;
7673                 goto abort;
7674         }
7675
7676         return conf;
7677
7678  abort:
7679         if (conf)
7680                 free_conf(conf);
7681         return ERR_PTR(ret);
7682 }
7683
7684 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7685 {
7686         switch (algo) {
7687         case ALGORITHM_PARITY_0:
7688                 if (raid_disk < max_degraded)
7689                         return 1;
7690                 break;
7691         case ALGORITHM_PARITY_N:
7692                 if (raid_disk >= raid_disks - max_degraded)
7693                         return 1;
7694                 break;
7695         case ALGORITHM_PARITY_0_6:
7696                 if (raid_disk == 0 ||
7697                     raid_disk == raid_disks - 1)
7698                         return 1;
7699                 break;
7700         case ALGORITHM_LEFT_ASYMMETRIC_6:
7701         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7702         case ALGORITHM_LEFT_SYMMETRIC_6:
7703         case ALGORITHM_RIGHT_SYMMETRIC_6:
7704                 if (raid_disk == raid_disks - 1)
7705                         return 1;
7706         }
7707         return 0;
7708 }
7709
7710 static int raid5_set_limits(struct mddev *mddev)
7711 {
7712         struct r5conf *conf = mddev->private;
7713         struct queue_limits lim;
7714         int data_disks, stripe;
7715         struct md_rdev *rdev;
7716
7717         /*
7718          * The read-ahead size must cover two whole stripes, which is
7719          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices.
7720          */
7721         data_disks = conf->previous_raid_disks - conf->max_degraded;
7722
7723         /*
7724          * We can only discard a whole stripe. It doesn't make sense to
7725          * discard data disk but write parity disk
7726          */
7727         stripe = roundup_pow_of_two(data_disks * (mddev->chunk_sectors << 9));
7728
7729         md_init_stacking_limits(&lim);
7730         lim.io_min = mddev->chunk_sectors << 9;
7731         lim.io_opt = lim.io_min * (conf->raid_disks - conf->max_degraded);
7732         lim.features |= BLK_FEAT_RAID_PARTIAL_STRIPES_EXPENSIVE;
7733         lim.discard_granularity = stripe;
7734         lim.max_write_zeroes_sectors = 0;
7735         mddev_stack_rdev_limits(mddev, &lim, 0);
7736         rdev_for_each(rdev, mddev)
7737                 queue_limits_stack_bdev(&lim, rdev->bdev, rdev->new_data_offset,
7738                                 mddev->gendisk->disk_name);
7739
7740         /*
7741          * Zeroing is required for discard, otherwise data could be lost.
7742          *
7743          * Consider a scenario: discard a stripe (the stripe could be
7744          * inconsistent if discard_zeroes_data is 0); write one disk of the
7745          * stripe (the stripe could be inconsistent again depending on which
7746          * disks are used to calculate parity); the disk is broken; The stripe
7747          * data of this disk is lost.
7748          *
7749          * We only allow DISCARD if the sysadmin has confirmed that only safe
7750          * devices are in use by setting a module parameter.  A better idea
7751          * might be to turn DISCARD into WRITE_ZEROES requests, as that is
7752          * required to be safe.
7753          */
7754         if (!devices_handle_discard_safely ||
7755             lim.max_discard_sectors < (stripe >> 9) ||
7756             lim.discard_granularity < stripe)
7757                 lim.max_hw_discard_sectors = 0;
7758
7759         /*
7760          * Requests require having a bitmap for each stripe.
7761          * Limit the max sectors based on this.
7762          */
7763         lim.max_hw_sectors = RAID5_MAX_REQ_STRIPES << RAID5_STRIPE_SHIFT(conf);
7764
7765         /* No restrictions on the number of segments in the request */
7766         lim.max_segments = USHRT_MAX;
7767
7768         return queue_limits_set(mddev->gendisk->queue, &lim);
7769 }
7770
7771 static int raid5_run(struct mddev *mddev)
7772 {
7773         struct r5conf *conf;
7774         int dirty_parity_disks = 0;
7775         struct md_rdev *rdev;
7776         struct md_rdev *journal_dev = NULL;
7777         sector_t reshape_offset = 0;
7778         int i;
7779         long long min_offset_diff = 0;
7780         int first = 1;
7781         int ret = -EIO;
7782
7783         if (mddev->recovery_cp != MaxSector)
7784                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7785                           mdname(mddev));
7786
7787         rdev_for_each(rdev, mddev) {
7788                 long long diff;
7789
7790                 if (test_bit(Journal, &rdev->flags)) {
7791                         journal_dev = rdev;
7792                         continue;
7793                 }
7794                 if (rdev->raid_disk < 0)
7795                         continue;
7796                 diff = (rdev->new_data_offset - rdev->data_offset);
7797                 if (first) {
7798                         min_offset_diff = diff;
7799                         first = 0;
7800                 } else if (mddev->reshape_backwards &&
7801                          diff < min_offset_diff)
7802                         min_offset_diff = diff;
7803                 else if (!mddev->reshape_backwards &&
7804                          diff > min_offset_diff)
7805                         min_offset_diff = diff;
7806         }
7807
7808         if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7809             (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7810                 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7811                           mdname(mddev));
7812                 return -EINVAL;
7813         }
7814
7815         if (mddev->reshape_position != MaxSector) {
7816                 /* Check that we can continue the reshape.
7817                  * Difficulties arise if the stripe we would write to
7818                  * next is at or after the stripe we would read from next.
7819                  * For a reshape that changes the number of devices, this
7820                  * is only possible for a very short time, and mdadm makes
7821                  * sure that time appears to have past before assembling
7822                  * the array.  So we fail if that time hasn't passed.
7823                  * For a reshape that keeps the number of devices the same
7824                  * mdadm must be monitoring the reshape can keeping the
7825                  * critical areas read-only and backed up.  It will start
7826                  * the array in read-only mode, so we check for that.
7827                  */
7828                 sector_t here_new, here_old;
7829                 int old_disks;
7830                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7831                 int chunk_sectors;
7832                 int new_data_disks;
7833
7834                 if (journal_dev) {
7835                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7836                                 mdname(mddev));
7837                         return -EINVAL;
7838                 }
7839
7840                 if (mddev->new_level != mddev->level) {
7841                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7842                                 mdname(mddev));
7843                         return -EINVAL;
7844                 }
7845                 old_disks = mddev->raid_disks - mddev->delta_disks;
7846                 /* reshape_position must be on a new-stripe boundary, and one
7847                  * further up in new geometry must map after here in old
7848                  * geometry.
7849                  * If the chunk sizes are different, then as we perform reshape
7850                  * in units of the largest of the two, reshape_position needs
7851                  * be a multiple of the largest chunk size times new data disks.
7852                  */
7853                 here_new = mddev->reshape_position;
7854                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7855                 new_data_disks = mddev->raid_disks - max_degraded;
7856                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7857                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7858                                 mdname(mddev));
7859                         return -EINVAL;
7860                 }
7861                 reshape_offset = here_new * chunk_sectors;
7862                 /* here_new is the stripe we will write to */
7863                 here_old = mddev->reshape_position;
7864                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7865                 /* here_old is the first stripe that we might need to read
7866                  * from */
7867                 if (mddev->delta_disks == 0) {
7868                         /* We cannot be sure it is safe to start an in-place
7869                          * reshape.  It is only safe if user-space is monitoring
7870                          * and taking constant backups.
7871                          * mdadm always starts a situation like this in
7872                          * readonly mode so it can take control before
7873                          * allowing any writes.  So just check for that.
7874                          */
7875                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7876                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7877                                 /* not really in-place - so OK */;
7878                         else if (mddev->ro == 0) {
7879                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7880                                         mdname(mddev));
7881                                 return -EINVAL;
7882                         }
7883                 } else if (mddev->reshape_backwards
7884                     ? (here_new * chunk_sectors + min_offset_diff <=
7885                        here_old * chunk_sectors)
7886                     : (here_new * chunk_sectors >=
7887                        here_old * chunk_sectors + (-min_offset_diff))) {
7888                         /* Reading from the same stripe as writing to - bad */
7889                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7890                                 mdname(mddev));
7891                         return -EINVAL;
7892                 }
7893                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7894                 /* OK, we should be able to continue; */
7895         } else {
7896                 BUG_ON(mddev->level != mddev->new_level);
7897                 BUG_ON(mddev->layout != mddev->new_layout);
7898                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7899                 BUG_ON(mddev->delta_disks != 0);
7900         }
7901
7902         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7903             test_bit(MD_HAS_PPL, &mddev->flags)) {
7904                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7905                         mdname(mddev));
7906                 clear_bit(MD_HAS_PPL, &mddev->flags);
7907                 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7908         }
7909
7910         if (mddev->private == NULL)
7911                 conf = setup_conf(mddev);
7912         else
7913                 conf = mddev->private;
7914
7915         if (IS_ERR(conf))
7916                 return PTR_ERR(conf);
7917
7918         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7919                 if (!journal_dev) {
7920                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7921                                 mdname(mddev));
7922                         mddev->ro = 1;
7923                         set_disk_ro(mddev->gendisk, 1);
7924                 } else if (mddev->recovery_cp == MaxSector)
7925                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7926         }
7927
7928         conf->min_offset_diff = min_offset_diff;
7929         rcu_assign_pointer(mddev->thread, conf->thread);
7930         rcu_assign_pointer(conf->thread, NULL);
7931         mddev->private = conf;
7932
7933         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7934              i++) {
7935                 rdev = conf->disks[i].rdev;
7936                 if (!rdev)
7937                         continue;
7938                 if (conf->disks[i].replacement &&
7939                     conf->reshape_progress != MaxSector) {
7940                         /* replacements and reshape simply do not mix. */
7941                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7942                         goto abort;
7943                 }
7944                 if (test_bit(In_sync, &rdev->flags))
7945                         continue;
7946                 /* This disc is not fully in-sync.  However if it
7947                  * just stored parity (beyond the recovery_offset),
7948                  * when we don't need to be concerned about the
7949                  * array being dirty.
7950                  * When reshape goes 'backwards', we never have
7951                  * partially completed devices, so we only need
7952                  * to worry about reshape going forwards.
7953                  */
7954                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7955                 if (mddev->major_version == 0 &&
7956                     mddev->minor_version > 90)
7957                         rdev->recovery_offset = reshape_offset;
7958
7959                 if (rdev->recovery_offset < reshape_offset) {
7960                         /* We need to check old and new layout */
7961                         if (!only_parity(rdev->raid_disk,
7962                                          conf->algorithm,
7963                                          conf->raid_disks,
7964                                          conf->max_degraded))
7965                                 continue;
7966                 }
7967                 if (!only_parity(rdev->raid_disk,
7968                                  conf->prev_algo,
7969                                  conf->previous_raid_disks,
7970                                  conf->max_degraded))
7971                         continue;
7972                 dirty_parity_disks++;
7973         }
7974
7975         /*
7976          * 0 for a fully functional array, 1 or 2 for a degraded array.
7977          */
7978         mddev->degraded = raid5_calc_degraded(conf);
7979
7980         if (has_failed(conf)) {
7981                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7982                         mdname(mddev), mddev->degraded, conf->raid_disks);
7983                 goto abort;
7984         }
7985
7986         /* device size must be a multiple of chunk size */
7987         mddev->dev_sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7988         mddev->resync_max_sectors = mddev->dev_sectors;
7989
7990         if (mddev->degraded > dirty_parity_disks &&
7991             mddev->recovery_cp != MaxSector) {
7992                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7993                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7994                                 mdname(mddev));
7995                 else if (mddev->ok_start_degraded)
7996                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7997                                 mdname(mddev));
7998                 else {
7999                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
8000                                 mdname(mddev));
8001                         goto abort;
8002                 }
8003         }
8004
8005         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
8006                 mdname(mddev), conf->level,
8007                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
8008                 mddev->new_layout);
8009
8010         print_raid5_conf(conf);
8011
8012         if (conf->reshape_progress != MaxSector) {
8013                 conf->reshape_safe = conf->reshape_progress;
8014                 atomic_set(&conf->reshape_stripes, 0);
8015                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8016                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8017                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8018                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8019         }
8020
8021         /* Ok, everything is just fine now */
8022         if (mddev->to_remove == &raid5_attrs_group)
8023                 mddev->to_remove = NULL;
8024         else if (mddev->kobj.sd &&
8025             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
8026                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
8027                         mdname(mddev));
8028         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
8029
8030         if (!mddev_is_dm(mddev)) {
8031                 ret = raid5_set_limits(mddev);
8032                 if (ret)
8033                         goto abort;
8034         }
8035
8036         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
8037                 goto abort;
8038
8039         return 0;
8040 abort:
8041         md_unregister_thread(mddev, &mddev->thread);
8042         print_raid5_conf(conf);
8043         free_conf(conf);
8044         mddev->private = NULL;
8045         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
8046         return ret;
8047 }
8048
8049 static void raid5_free(struct mddev *mddev, void *priv)
8050 {
8051         struct r5conf *conf = priv;
8052
8053         free_conf(conf);
8054         mddev->to_remove = &raid5_attrs_group;
8055 }
8056
8057 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
8058 {
8059         struct r5conf *conf = mddev->private;
8060         int i;
8061
8062         lockdep_assert_held(&mddev->lock);
8063
8064         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
8065                 conf->chunk_sectors / 2, mddev->layout);
8066         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
8067         for (i = 0; i < conf->raid_disks; i++) {
8068                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
8069
8070                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
8071         }
8072         seq_printf (seq, "]");
8073 }
8074
8075 static void print_raid5_conf(struct r5conf *conf)
8076 {
8077         struct md_rdev *rdev;
8078         int i;
8079
8080         pr_debug("RAID conf printout:\n");
8081         if (!conf) {
8082                 pr_debug("(conf==NULL)\n");
8083                 return;
8084         }
8085         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
8086                conf->raid_disks,
8087                conf->raid_disks - conf->mddev->degraded);
8088
8089         for (i = 0; i < conf->raid_disks; i++) {
8090                 rdev = conf->disks[i].rdev;
8091                 if (rdev)
8092                         pr_debug(" disk %d, o:%d, dev:%pg\n",
8093                                i, !test_bit(Faulty, &rdev->flags),
8094                                rdev->bdev);
8095         }
8096 }
8097
8098 static int raid5_spare_active(struct mddev *mddev)
8099 {
8100         int i;
8101         struct r5conf *conf = mddev->private;
8102         struct md_rdev *rdev, *replacement;
8103         int count = 0;
8104         unsigned long flags;
8105
8106         for (i = 0; i < conf->raid_disks; i++) {
8107                 rdev = conf->disks[i].rdev;
8108                 replacement = conf->disks[i].replacement;
8109                 if (replacement
8110                     && replacement->recovery_offset == MaxSector
8111                     && !test_bit(Faulty, &replacement->flags)
8112                     && !test_and_set_bit(In_sync, &replacement->flags)) {
8113                         /* Replacement has just become active. */
8114                         if (!rdev
8115                             || !test_and_clear_bit(In_sync, &rdev->flags))
8116                                 count++;
8117                         if (rdev) {
8118                                 /* Replaced device not technically faulty,
8119                                  * but we need to be sure it gets removed
8120                                  * and never re-added.
8121                                  */
8122                                 set_bit(Faulty, &rdev->flags);
8123                                 sysfs_notify_dirent_safe(
8124                                         rdev->sysfs_state);
8125                         }
8126                         sysfs_notify_dirent_safe(replacement->sysfs_state);
8127                 } else if (rdev
8128                     && rdev->recovery_offset == MaxSector
8129                     && !test_bit(Faulty, &rdev->flags)
8130                     && !test_and_set_bit(In_sync, &rdev->flags)) {
8131                         count++;
8132                         sysfs_notify_dirent_safe(rdev->sysfs_state);
8133                 }
8134         }
8135         spin_lock_irqsave(&conf->device_lock, flags);
8136         mddev->degraded = raid5_calc_degraded(conf);
8137         spin_unlock_irqrestore(&conf->device_lock, flags);
8138         print_raid5_conf(conf);
8139         return count;
8140 }
8141
8142 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
8143 {
8144         struct r5conf *conf = mddev->private;
8145         int err = 0;
8146         int number = rdev->raid_disk;
8147         struct md_rdev **rdevp;
8148         struct disk_info *p;
8149         struct md_rdev *tmp;
8150
8151         print_raid5_conf(conf);
8152         if (test_bit(Journal, &rdev->flags) && conf->log) {
8153                 /*
8154                  * we can't wait pending write here, as this is called in
8155                  * raid5d, wait will deadlock.
8156                  * neilb: there is no locking about new writes here,
8157                  * so this cannot be safe.
8158                  */
8159                 if (atomic_read(&conf->active_stripes) ||
8160                     atomic_read(&conf->r5c_cached_full_stripes) ||
8161                     atomic_read(&conf->r5c_cached_partial_stripes)) {
8162                         return -EBUSY;
8163                 }
8164                 log_exit(conf);
8165                 return 0;
8166         }
8167         if (unlikely(number >= conf->pool_size))
8168                 return 0;
8169         p = conf->disks + number;
8170         if (rdev == p->rdev)
8171                 rdevp = &p->rdev;
8172         else if (rdev == p->replacement)
8173                 rdevp = &p->replacement;
8174         else
8175                 return 0;
8176
8177         if (number >= conf->raid_disks &&
8178             conf->reshape_progress == MaxSector)
8179                 clear_bit(In_sync, &rdev->flags);
8180
8181         if (test_bit(In_sync, &rdev->flags) ||
8182             atomic_read(&rdev->nr_pending)) {
8183                 err = -EBUSY;
8184                 goto abort;
8185         }
8186         /* Only remove non-faulty devices if recovery
8187          * isn't possible.
8188          */
8189         if (!test_bit(Faulty, &rdev->flags) &&
8190             mddev->recovery_disabled != conf->recovery_disabled &&
8191             !has_failed(conf) &&
8192             (!p->replacement || p->replacement == rdev) &&
8193             number < conf->raid_disks) {
8194                 err = -EBUSY;
8195                 goto abort;
8196         }
8197         WRITE_ONCE(*rdevp, NULL);
8198         if (!err) {
8199                 err = log_modify(conf, rdev, false);
8200                 if (err)
8201                         goto abort;
8202         }
8203
8204         tmp = p->replacement;
8205         if (tmp) {
8206                 /* We must have just cleared 'rdev' */
8207                 WRITE_ONCE(p->rdev, tmp);
8208                 clear_bit(Replacement, &tmp->flags);
8209                 WRITE_ONCE(p->replacement, NULL);
8210
8211                 if (!err)
8212                         err = log_modify(conf, tmp, true);
8213         }
8214
8215         clear_bit(WantReplacement, &rdev->flags);
8216 abort:
8217
8218         print_raid5_conf(conf);
8219         return err;
8220 }
8221
8222 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
8223 {
8224         struct r5conf *conf = mddev->private;
8225         int ret, err = -EEXIST;
8226         int disk;
8227         struct disk_info *p;
8228         struct md_rdev *tmp;
8229         int first = 0;
8230         int last = conf->raid_disks - 1;
8231
8232         if (test_bit(Journal, &rdev->flags)) {
8233                 if (conf->log)
8234                         return -EBUSY;
8235
8236                 rdev->raid_disk = 0;
8237                 /*
8238                  * The array is in readonly mode if journal is missing, so no
8239                  * write requests running. We should be safe
8240                  */
8241                 ret = log_init(conf, rdev, false);
8242                 if (ret)
8243                         return ret;
8244
8245                 ret = r5l_start(conf->log);
8246                 if (ret)
8247                         return ret;
8248
8249                 return 0;
8250         }
8251         if (mddev->recovery_disabled == conf->recovery_disabled)
8252                 return -EBUSY;
8253
8254         if (rdev->saved_raid_disk < 0 && has_failed(conf))
8255                 /* no point adding a device */
8256                 return -EINVAL;
8257
8258         if (rdev->raid_disk >= 0)
8259                 first = last = rdev->raid_disk;
8260
8261         /*
8262          * find the disk ... but prefer rdev->saved_raid_disk
8263          * if possible.
8264          */
8265         if (rdev->saved_raid_disk >= first &&
8266             rdev->saved_raid_disk <= last &&
8267             conf->disks[rdev->saved_raid_disk].rdev == NULL)
8268                 first = rdev->saved_raid_disk;
8269
8270         for (disk = first; disk <= last; disk++) {
8271                 p = conf->disks + disk;
8272                 if (p->rdev == NULL) {
8273                         clear_bit(In_sync, &rdev->flags);
8274                         rdev->raid_disk = disk;
8275                         if (rdev->saved_raid_disk != disk)
8276                                 conf->fullsync = 1;
8277                         WRITE_ONCE(p->rdev, rdev);
8278
8279                         err = log_modify(conf, rdev, true);
8280
8281                         goto out;
8282                 }
8283         }
8284         for (disk = first; disk <= last; disk++) {
8285                 p = conf->disks + disk;
8286                 tmp = p->rdev;
8287                 if (test_bit(WantReplacement, &tmp->flags) &&
8288                     mddev->reshape_position == MaxSector &&
8289                     p->replacement == NULL) {
8290                         clear_bit(In_sync, &rdev->flags);
8291                         set_bit(Replacement, &rdev->flags);
8292                         rdev->raid_disk = disk;
8293                         err = 0;
8294                         conf->fullsync = 1;
8295                         WRITE_ONCE(p->replacement, rdev);
8296                         break;
8297                 }
8298         }
8299 out:
8300         print_raid5_conf(conf);
8301         return err;
8302 }
8303
8304 static int raid5_resize(struct mddev *mddev, sector_t sectors)
8305 {
8306         /* no resync is happening, and there is enough space
8307          * on all devices, so we can resize.
8308          * We need to make sure resync covers any new space.
8309          * If the array is shrinking we should possibly wait until
8310          * any io in the removed space completes, but it hardly seems
8311          * worth it.
8312          */
8313         sector_t newsize;
8314         struct r5conf *conf = mddev->private;
8315
8316         if (raid5_has_log(conf) || raid5_has_ppl(conf))
8317                 return -EINVAL;
8318         sectors &= ~((sector_t)conf->chunk_sectors - 1);
8319         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
8320         if (mddev->external_size &&
8321             mddev->array_sectors > newsize)
8322                 return -EINVAL;
8323         if (mddev->bitmap) {
8324                 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
8325                 if (ret)
8326                         return ret;
8327         }
8328         md_set_array_sectors(mddev, newsize);
8329         if (sectors > mddev->dev_sectors &&
8330             mddev->recovery_cp > mddev->dev_sectors) {
8331                 mddev->recovery_cp = mddev->dev_sectors;
8332                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8333         }
8334         mddev->dev_sectors = sectors;
8335         mddev->resync_max_sectors = sectors;
8336         return 0;
8337 }
8338
8339 static int check_stripe_cache(struct mddev *mddev)
8340 {
8341         /* Can only proceed if there are plenty of stripe_heads.
8342          * We need a minimum of one full stripe,, and for sensible progress
8343          * it is best to have about 4 times that.
8344          * If we require 4 times, then the default 256 4K stripe_heads will
8345          * allow for chunk sizes up to 256K, which is probably OK.
8346          * If the chunk size is greater, user-space should request more
8347          * stripe_heads first.
8348          */
8349         struct r5conf *conf = mddev->private;
8350         if (((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8351             > conf->min_nr_stripes ||
8352             ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8353             > conf->min_nr_stripes) {
8354                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
8355                         mdname(mddev),
8356                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
8357                          / RAID5_STRIPE_SIZE(conf))*4);
8358                 return 0;
8359         }
8360         return 1;
8361 }
8362
8363 static int check_reshape(struct mddev *mddev)
8364 {
8365         struct r5conf *conf = mddev->private;
8366
8367         if (raid5_has_log(conf) || raid5_has_ppl(conf))
8368                 return -EINVAL;
8369         if (mddev->delta_disks == 0 &&
8370             mddev->new_layout == mddev->layout &&
8371             mddev->new_chunk_sectors == mddev->chunk_sectors)
8372                 return 0; /* nothing to do */
8373         if (has_failed(conf))
8374                 return -EINVAL;
8375         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
8376                 /* We might be able to shrink, but the devices must
8377                  * be made bigger first.
8378                  * For raid6, 4 is the minimum size.
8379                  * Otherwise 2 is the minimum
8380                  */
8381                 int min = 2;
8382                 if (mddev->level == 6)
8383                         min = 4;
8384                 if (mddev->raid_disks + mddev->delta_disks < min)
8385                         return -EINVAL;
8386         }
8387
8388         if (!check_stripe_cache(mddev))
8389                 return -ENOSPC;
8390
8391         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
8392             mddev->delta_disks > 0)
8393                 if (resize_chunks(conf,
8394                                   conf->previous_raid_disks
8395                                   + max(0, mddev->delta_disks),
8396                                   max(mddev->new_chunk_sectors,
8397                                       mddev->chunk_sectors)
8398                             ) < 0)
8399                         return -ENOMEM;
8400
8401         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
8402                 return 0; /* never bother to shrink */
8403         return resize_stripes(conf, (conf->previous_raid_disks
8404                                      + mddev->delta_disks));
8405 }
8406
8407 static int raid5_start_reshape(struct mddev *mddev)
8408 {
8409         struct r5conf *conf = mddev->private;
8410         struct md_rdev *rdev;
8411         int spares = 0;
8412         int i;
8413         unsigned long flags;
8414
8415         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
8416                 return -EBUSY;
8417
8418         if (!check_stripe_cache(mddev))
8419                 return -ENOSPC;
8420
8421         if (has_failed(conf))
8422                 return -EINVAL;
8423
8424         /* raid5 can't handle concurrent reshape and recovery */
8425         if (mddev->recovery_cp < MaxSector)
8426                 return -EBUSY;
8427         for (i = 0; i < conf->raid_disks; i++)
8428                 if (conf->disks[i].replacement)
8429                         return -EBUSY;
8430
8431         rdev_for_each(rdev, mddev) {
8432                 if (!test_bit(In_sync, &rdev->flags)
8433                     && !test_bit(Faulty, &rdev->flags))
8434                         spares++;
8435         }
8436
8437         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
8438                 /* Not enough devices even to make a degraded array
8439                  * of that size
8440                  */
8441                 return -EINVAL;
8442
8443         /* Refuse to reduce size of the array.  Any reductions in
8444          * array size must be through explicit setting of array_size
8445          * attribute.
8446          */
8447         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
8448             < mddev->array_sectors) {
8449                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
8450                         mdname(mddev));
8451                 return -EINVAL;
8452         }
8453
8454         atomic_set(&conf->reshape_stripes, 0);
8455         spin_lock_irq(&conf->device_lock);
8456         write_seqcount_begin(&conf->gen_lock);
8457         conf->previous_raid_disks = conf->raid_disks;
8458         conf->raid_disks += mddev->delta_disks;
8459         conf->prev_chunk_sectors = conf->chunk_sectors;
8460         conf->chunk_sectors = mddev->new_chunk_sectors;
8461         conf->prev_algo = conf->algorithm;
8462         conf->algorithm = mddev->new_layout;
8463         conf->generation++;
8464         /* Code that selects data_offset needs to see the generation update
8465          * if reshape_progress has been set - so a memory barrier needed.
8466          */
8467         smp_mb();
8468         if (mddev->reshape_backwards)
8469                 conf->reshape_progress = raid5_size(mddev, 0, 0);
8470         else
8471                 conf->reshape_progress = 0;
8472         conf->reshape_safe = conf->reshape_progress;
8473         write_seqcount_end(&conf->gen_lock);
8474         spin_unlock_irq(&conf->device_lock);
8475
8476         /* Now make sure any requests that proceeded on the assumption
8477          * the reshape wasn't running - like Discard or Read - have
8478          * completed.
8479          */
8480         raid5_quiesce(mddev, true);
8481         raid5_quiesce(mddev, false);
8482
8483         /* Add some new drives, as many as will fit.
8484          * We know there are enough to make the newly sized array work.
8485          * Don't add devices if we are reducing the number of
8486          * devices in the array.  This is because it is not possible
8487          * to correctly record the "partially reconstructed" state of
8488          * such devices during the reshape and confusion could result.
8489          */
8490         if (mddev->delta_disks >= 0) {
8491                 rdev_for_each(rdev, mddev)
8492                         if (rdev->raid_disk < 0 &&
8493                             !test_bit(Faulty, &rdev->flags)) {
8494                                 if (raid5_add_disk(mddev, rdev) == 0) {
8495                                         if (rdev->raid_disk
8496                                             >= conf->previous_raid_disks)
8497                                                 set_bit(In_sync, &rdev->flags);
8498                                         else
8499                                                 rdev->recovery_offset = 0;
8500
8501                                         /* Failure here is OK */
8502                                         sysfs_link_rdev(mddev, rdev);
8503                                 }
8504                         } else if (rdev->raid_disk >= conf->previous_raid_disks
8505                                    && !test_bit(Faulty, &rdev->flags)) {
8506                                 /* This is a spare that was manually added */
8507                                 set_bit(In_sync, &rdev->flags);
8508                         }
8509
8510                 /* When a reshape changes the number of devices,
8511                  * ->degraded is measured against the larger of the
8512                  * pre and post number of devices.
8513                  */
8514                 spin_lock_irqsave(&conf->device_lock, flags);
8515                 mddev->degraded = raid5_calc_degraded(conf);
8516                 spin_unlock_irqrestore(&conf->device_lock, flags);
8517         }
8518         mddev->raid_disks = conf->raid_disks;
8519         mddev->reshape_position = conf->reshape_progress;
8520         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8521
8522         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8523         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8524         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
8525         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8526         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8527         conf->reshape_checkpoint = jiffies;
8528         md_new_event();
8529         return 0;
8530 }
8531
8532 /* This is called from the reshape thread and should make any
8533  * changes needed in 'conf'
8534  */
8535 static void end_reshape(struct r5conf *conf)
8536 {
8537
8538         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8539                 struct md_rdev *rdev;
8540
8541                 spin_lock_irq(&conf->device_lock);
8542                 conf->previous_raid_disks = conf->raid_disks;
8543                 md_finish_reshape(conf->mddev);
8544                 smp_wmb();
8545                 conf->reshape_progress = MaxSector;
8546                 conf->mddev->reshape_position = MaxSector;
8547                 rdev_for_each(rdev, conf->mddev)
8548                         if (rdev->raid_disk >= 0 &&
8549                             !test_bit(Journal, &rdev->flags) &&
8550                             !test_bit(In_sync, &rdev->flags))
8551                                 rdev->recovery_offset = MaxSector;
8552                 spin_unlock_irq(&conf->device_lock);
8553                 wake_up(&conf->wait_for_overlap);
8554
8555                 mddev_update_io_opt(conf->mddev,
8556                         conf->raid_disks - conf->max_degraded);
8557         }
8558 }
8559
8560 /* This is called from the raid5d thread with mddev_lock held.
8561  * It makes config changes to the device.
8562  */
8563 static void raid5_finish_reshape(struct mddev *mddev)
8564 {
8565         struct r5conf *conf = mddev->private;
8566         struct md_rdev *rdev;
8567
8568         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8569
8570                 if (mddev->delta_disks <= 0) {
8571                         int d;
8572                         spin_lock_irq(&conf->device_lock);
8573                         mddev->degraded = raid5_calc_degraded(conf);
8574                         spin_unlock_irq(&conf->device_lock);
8575                         for (d = conf->raid_disks ;
8576                              d < conf->raid_disks - mddev->delta_disks;
8577                              d++) {
8578                                 rdev = conf->disks[d].rdev;
8579                                 if (rdev)
8580                                         clear_bit(In_sync, &rdev->flags);
8581                                 rdev = conf->disks[d].replacement;
8582                                 if (rdev)
8583                                         clear_bit(In_sync, &rdev->flags);
8584                         }
8585                 }
8586                 mddev->layout = conf->algorithm;
8587                 mddev->chunk_sectors = conf->chunk_sectors;
8588                 mddev->reshape_position = MaxSector;
8589                 mddev->delta_disks = 0;
8590                 mddev->reshape_backwards = 0;
8591         }
8592 }
8593
8594 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8595 {
8596         struct r5conf *conf = mddev->private;
8597
8598         if (quiesce) {
8599                 /* stop all writes */
8600                 lock_all_device_hash_locks_irq(conf);
8601                 /* '2' tells resync/reshape to pause so that all
8602                  * active stripes can drain
8603                  */
8604                 r5c_flush_cache(conf, INT_MAX);
8605                 /* need a memory barrier to make sure read_one_chunk() sees
8606                  * quiesce started and reverts to slow (locked) path.
8607                  */
8608                 smp_store_release(&conf->quiesce, 2);
8609                 wait_event_cmd(conf->wait_for_quiescent,
8610                                     atomic_read(&conf->active_stripes) == 0 &&
8611                                     atomic_read(&conf->active_aligned_reads) == 0,
8612                                     unlock_all_device_hash_locks_irq(conf),
8613                                     lock_all_device_hash_locks_irq(conf));
8614                 conf->quiesce = 1;
8615                 unlock_all_device_hash_locks_irq(conf);
8616                 /* allow reshape to continue */
8617                 wake_up(&conf->wait_for_overlap);
8618         } else {
8619                 /* re-enable writes */
8620                 lock_all_device_hash_locks_irq(conf);
8621                 conf->quiesce = 0;
8622                 wake_up(&conf->wait_for_quiescent);
8623                 wake_up(&conf->wait_for_overlap);
8624                 unlock_all_device_hash_locks_irq(conf);
8625         }
8626         log_quiesce(conf, quiesce);
8627 }
8628
8629 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8630 {
8631         struct r0conf *raid0_conf = mddev->private;
8632         sector_t sectors;
8633
8634         /* for raid0 takeover only one zone is supported */
8635         if (raid0_conf->nr_strip_zones > 1) {
8636                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8637                         mdname(mddev));
8638                 return ERR_PTR(-EINVAL);
8639         }
8640
8641         sectors = raid0_conf->strip_zone[0].zone_end;
8642         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8643         mddev->dev_sectors = sectors;
8644         mddev->new_level = level;
8645         mddev->new_layout = ALGORITHM_PARITY_N;
8646         mddev->new_chunk_sectors = mddev->chunk_sectors;
8647         mddev->raid_disks += 1;
8648         mddev->delta_disks = 1;
8649         /* make sure it will be not marked as dirty */
8650         mddev->recovery_cp = MaxSector;
8651
8652         return setup_conf(mddev);
8653 }
8654
8655 static void *raid5_takeover_raid1(struct mddev *mddev)
8656 {
8657         int chunksect;
8658         void *ret;
8659
8660         if (mddev->raid_disks != 2 ||
8661             mddev->degraded > 1)
8662                 return ERR_PTR(-EINVAL);
8663
8664         /* Should check if there are write-behind devices? */
8665
8666         chunksect = 64*2; /* 64K by default */
8667
8668         /* The array must be an exact multiple of chunksize */
8669         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8670                 chunksect >>= 1;
8671
8672         if ((chunksect<<9) < RAID5_STRIPE_SIZE((struct r5conf *)mddev->private))
8673                 /* array size does not allow a suitable chunk size */
8674                 return ERR_PTR(-EINVAL);
8675
8676         mddev->new_level = 5;
8677         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8678         mddev->new_chunk_sectors = chunksect;
8679
8680         ret = setup_conf(mddev);
8681         if (!IS_ERR(ret))
8682                 mddev_clear_unsupported_flags(mddev,
8683                         UNSUPPORTED_MDDEV_FLAGS);
8684         return ret;
8685 }
8686
8687 static void *raid5_takeover_raid6(struct mddev *mddev)
8688 {
8689         int new_layout;
8690
8691         switch (mddev->layout) {
8692         case ALGORITHM_LEFT_ASYMMETRIC_6:
8693                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8694                 break;
8695         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8696                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8697                 break;
8698         case ALGORITHM_LEFT_SYMMETRIC_6:
8699                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8700                 break;
8701         case ALGORITHM_RIGHT_SYMMETRIC_6:
8702                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8703                 break;
8704         case ALGORITHM_PARITY_0_6:
8705                 new_layout = ALGORITHM_PARITY_0;
8706                 break;
8707         case ALGORITHM_PARITY_N:
8708                 new_layout = ALGORITHM_PARITY_N;
8709                 break;
8710         default:
8711                 return ERR_PTR(-EINVAL);
8712         }
8713         mddev->new_level = 5;
8714         mddev->new_layout = new_layout;
8715         mddev->delta_disks = -1;
8716         mddev->raid_disks -= 1;
8717         return setup_conf(mddev);
8718 }
8719
8720 static int raid5_check_reshape(struct mddev *mddev)
8721 {
8722         /* For a 2-drive array, the layout and chunk size can be changed
8723          * immediately as not restriping is needed.
8724          * For larger arrays we record the new value - after validation
8725          * to be used by a reshape pass.
8726          */
8727         struct r5conf *conf = mddev->private;
8728         int new_chunk = mddev->new_chunk_sectors;
8729
8730         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8731                 return -EINVAL;
8732         if (new_chunk > 0) {
8733                 if (!is_power_of_2(new_chunk))
8734                         return -EINVAL;
8735                 if (new_chunk < (PAGE_SIZE>>9))
8736                         return -EINVAL;
8737                 if (mddev->array_sectors & (new_chunk-1))
8738                         /* not factor of array size */
8739                         return -EINVAL;
8740         }
8741
8742         /* They look valid */
8743
8744         if (mddev->raid_disks == 2) {
8745                 /* can make the change immediately */
8746                 if (mddev->new_layout >= 0) {
8747                         conf->algorithm = mddev->new_layout;
8748                         mddev->layout = mddev->new_layout;
8749                 }
8750                 if (new_chunk > 0) {
8751                         conf->chunk_sectors = new_chunk ;
8752                         mddev->chunk_sectors = new_chunk;
8753                 }
8754                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8755                 md_wakeup_thread(mddev->thread);
8756         }
8757         return check_reshape(mddev);
8758 }
8759
8760 static int raid6_check_reshape(struct mddev *mddev)
8761 {
8762         int new_chunk = mddev->new_chunk_sectors;
8763
8764         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8765                 return -EINVAL;
8766         if (new_chunk > 0) {
8767                 if (!is_power_of_2(new_chunk))
8768                         return -EINVAL;
8769                 if (new_chunk < (PAGE_SIZE >> 9))
8770                         return -EINVAL;
8771                 if (mddev->array_sectors & (new_chunk-1))
8772                         /* not factor of array size */
8773                         return -EINVAL;
8774         }
8775
8776         /* They look valid */
8777         return check_reshape(mddev);
8778 }
8779
8780 static void *raid5_takeover(struct mddev *mddev)
8781 {
8782         /* raid5 can take over:
8783          *  raid0 - if there is only one strip zone - make it a raid4 layout
8784          *  raid1 - if there are two drives.  We need to know the chunk size
8785          *  raid4 - trivial - just use a raid4 layout.
8786          *  raid6 - Providing it is a *_6 layout
8787          */
8788         if (mddev->level == 0)
8789                 return raid45_takeover_raid0(mddev, 5);
8790         if (mddev->level == 1)
8791                 return raid5_takeover_raid1(mddev);
8792         if (mddev->level == 4) {
8793                 mddev->new_layout = ALGORITHM_PARITY_N;
8794                 mddev->new_level = 5;
8795                 return setup_conf(mddev);
8796         }
8797         if (mddev->level == 6)
8798                 return raid5_takeover_raid6(mddev);
8799
8800         return ERR_PTR(-EINVAL);
8801 }
8802
8803 static void *raid4_takeover(struct mddev *mddev)
8804 {
8805         /* raid4 can take over:
8806          *  raid0 - if there is only one strip zone
8807          *  raid5 - if layout is right
8808          */
8809         if (mddev->level == 0)
8810                 return raid45_takeover_raid0(mddev, 4);
8811         if (mddev->level == 5 &&
8812             mddev->layout == ALGORITHM_PARITY_N) {
8813                 mddev->new_layout = 0;
8814                 mddev->new_level = 4;
8815                 return setup_conf(mddev);
8816         }
8817         return ERR_PTR(-EINVAL);
8818 }
8819
8820 static struct md_personality raid5_personality;
8821
8822 static void *raid6_takeover(struct mddev *mddev)
8823 {
8824         /* Currently can only take over a raid5.  We map the
8825          * personality to an equivalent raid6 personality
8826          * with the Q block at the end.
8827          */
8828         int new_layout;
8829
8830         if (mddev->pers != &raid5_personality)
8831                 return ERR_PTR(-EINVAL);
8832         if (mddev->degraded > 1)
8833                 return ERR_PTR(-EINVAL);
8834         if (mddev->raid_disks > 253)
8835                 return ERR_PTR(-EINVAL);
8836         if (mddev->raid_disks < 3)
8837                 return ERR_PTR(-EINVAL);
8838
8839         switch (mddev->layout) {
8840         case ALGORITHM_LEFT_ASYMMETRIC:
8841                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8842                 break;
8843         case ALGORITHM_RIGHT_ASYMMETRIC:
8844                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8845                 break;
8846         case ALGORITHM_LEFT_SYMMETRIC:
8847                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8848                 break;
8849         case ALGORITHM_RIGHT_SYMMETRIC:
8850                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8851                 break;
8852         case ALGORITHM_PARITY_0:
8853                 new_layout = ALGORITHM_PARITY_0_6;
8854                 break;
8855         case ALGORITHM_PARITY_N:
8856                 new_layout = ALGORITHM_PARITY_N;
8857                 break;
8858         default:
8859                 return ERR_PTR(-EINVAL);
8860         }
8861         mddev->new_level = 6;
8862         mddev->new_layout = new_layout;
8863         mddev->delta_disks = 1;
8864         mddev->raid_disks += 1;
8865         return setup_conf(mddev);
8866 }
8867
8868 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8869 {
8870         struct r5conf *conf;
8871         int err;
8872
8873         err = mddev_suspend_and_lock(mddev);
8874         if (err)
8875                 return err;
8876         conf = mddev->private;
8877         if (!conf) {
8878                 mddev_unlock_and_resume(mddev);
8879                 return -ENODEV;
8880         }
8881
8882         if (strncmp(buf, "ppl", 3) == 0) {
8883                 /* ppl only works with RAID 5 */
8884                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8885                         err = log_init(conf, NULL, true);
8886                         if (!err) {
8887                                 err = resize_stripes(conf, conf->pool_size);
8888                                 if (err)
8889                                         log_exit(conf);
8890                         }
8891                 } else
8892                         err = -EINVAL;
8893         } else if (strncmp(buf, "resync", 6) == 0) {
8894                 if (raid5_has_ppl(conf)) {
8895                         log_exit(conf);
8896                         err = resize_stripes(conf, conf->pool_size);
8897                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8898                            r5l_log_disk_error(conf)) {
8899                         bool journal_dev_exists = false;
8900                         struct md_rdev *rdev;
8901
8902                         rdev_for_each(rdev, mddev)
8903                                 if (test_bit(Journal, &rdev->flags)) {
8904                                         journal_dev_exists = true;
8905                                         break;
8906                                 }
8907
8908                         if (!journal_dev_exists)
8909                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8910                         else  /* need remove journal device first */
8911                                 err = -EBUSY;
8912                 } else
8913                         err = -EINVAL;
8914         } else {
8915                 err = -EINVAL;
8916         }
8917
8918         if (!err)
8919                 md_update_sb(mddev, 1);
8920
8921         mddev_unlock_and_resume(mddev);
8922
8923         return err;
8924 }
8925
8926 static int raid5_start(struct mddev *mddev)
8927 {
8928         struct r5conf *conf = mddev->private;
8929
8930         return r5l_start(conf->log);
8931 }
8932
8933 /*
8934  * This is only used for dm-raid456, caller already frozen sync_thread, hence
8935  * if rehsape is still in progress, io that is waiting for reshape can never be
8936  * done now, hence wake up and handle those IO.
8937  */
8938 static void raid5_prepare_suspend(struct mddev *mddev)
8939 {
8940         struct r5conf *conf = mddev->private;
8941
8942         wake_up(&conf->wait_for_overlap);
8943 }
8944
8945 static struct md_personality raid6_personality =
8946 {
8947         .name           = "raid6",
8948         .level          = 6,
8949         .owner          = THIS_MODULE,
8950         .make_request   = raid5_make_request,
8951         .run            = raid5_run,
8952         .start          = raid5_start,
8953         .free           = raid5_free,
8954         .status         = raid5_status,
8955         .error_handler  = raid5_error,
8956         .hot_add_disk   = raid5_add_disk,
8957         .hot_remove_disk= raid5_remove_disk,
8958         .spare_active   = raid5_spare_active,
8959         .sync_request   = raid5_sync_request,
8960         .resize         = raid5_resize,
8961         .size           = raid5_size,
8962         .check_reshape  = raid6_check_reshape,
8963         .start_reshape  = raid5_start_reshape,
8964         .finish_reshape = raid5_finish_reshape,
8965         .quiesce        = raid5_quiesce,
8966         .takeover       = raid6_takeover,
8967         .change_consistency_policy = raid5_change_consistency_policy,
8968         .prepare_suspend = raid5_prepare_suspend,
8969 };
8970 static struct md_personality raid5_personality =
8971 {
8972         .name           = "raid5",
8973         .level          = 5,
8974         .owner          = THIS_MODULE,
8975         .make_request   = raid5_make_request,
8976         .run            = raid5_run,
8977         .start          = raid5_start,
8978         .free           = raid5_free,
8979         .status         = raid5_status,
8980         .error_handler  = raid5_error,
8981         .hot_add_disk   = raid5_add_disk,
8982         .hot_remove_disk= raid5_remove_disk,
8983         .spare_active   = raid5_spare_active,
8984         .sync_request   = raid5_sync_request,
8985         .resize         = raid5_resize,
8986         .size           = raid5_size,
8987         .check_reshape  = raid5_check_reshape,
8988         .start_reshape  = raid5_start_reshape,
8989         .finish_reshape = raid5_finish_reshape,
8990         .quiesce        = raid5_quiesce,
8991         .takeover       = raid5_takeover,
8992         .change_consistency_policy = raid5_change_consistency_policy,
8993         .prepare_suspend = raid5_prepare_suspend,
8994 };
8995
8996 static struct md_personality raid4_personality =
8997 {
8998         .name           = "raid4",
8999         .level          = 4,
9000         .owner          = THIS_MODULE,
9001         .make_request   = raid5_make_request,
9002         .run            = raid5_run,
9003         .start          = raid5_start,
9004         .free           = raid5_free,
9005         .status         = raid5_status,
9006         .error_handler  = raid5_error,
9007         .hot_add_disk   = raid5_add_disk,
9008         .hot_remove_disk= raid5_remove_disk,
9009         .spare_active   = raid5_spare_active,
9010         .sync_request   = raid5_sync_request,
9011         .resize         = raid5_resize,
9012         .size           = raid5_size,
9013         .check_reshape  = raid5_check_reshape,
9014         .start_reshape  = raid5_start_reshape,
9015         .finish_reshape = raid5_finish_reshape,
9016         .quiesce        = raid5_quiesce,
9017         .takeover       = raid4_takeover,
9018         .change_consistency_policy = raid5_change_consistency_policy,
9019         .prepare_suspend = raid5_prepare_suspend,
9020 };
9021
9022 static int __init raid5_init(void)
9023 {
9024         int ret;
9025
9026         raid5_wq = alloc_workqueue("raid5wq",
9027                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
9028         if (!raid5_wq)
9029                 return -ENOMEM;
9030
9031         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
9032                                       "md/raid5:prepare",
9033                                       raid456_cpu_up_prepare,
9034                                       raid456_cpu_dead);
9035         if (ret) {
9036                 destroy_workqueue(raid5_wq);
9037                 return ret;
9038         }
9039         register_md_personality(&raid6_personality);
9040         register_md_personality(&raid5_personality);
9041         register_md_personality(&raid4_personality);
9042         return 0;
9043 }
9044
9045 static void raid5_exit(void)
9046 {
9047         unregister_md_personality(&raid6_personality);
9048         unregister_md_personality(&raid5_personality);
9049         unregister_md_personality(&raid4_personality);
9050         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
9051         destroy_workqueue(raid5_wq);
9052 }
9053
9054 module_init(raid5_init);
9055 module_exit(raid5_exit);
9056 MODULE_LICENSE("GPL");
9057 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
9058 MODULE_ALIAS("md-personality-4"); /* RAID5 */
9059 MODULE_ALIAS("md-raid5");
9060 MODULE_ALIAS("md-raid4");
9061 MODULE_ALIAS("md-level-5");
9062 MODULE_ALIAS("md-level-4");
9063 MODULE_ALIAS("md-personality-8"); /* RAID6 */
9064 MODULE_ALIAS("md-raid6");
9065 MODULE_ALIAS("md-level-6");
9066
9067 /* This used to be two separate modules, they were: */
9068 MODULE_ALIAS("raid5");
9069 MODULE_ALIAS("raid6");