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 static bool devices_handle_discard_safely = false;
  65 module_param(devices_handle_discard_safely, bool, 0644);
  66 MODULE_PARM_DESC(devices_handle_discard_safely,
  67                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  68 static struct workqueue_struct *raid5_wq;
  69
  70 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  71 {
  72         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  73         return &conf->stripe_hashtbl[hash];
  74 }
  75
  76 static inline int stripe_hash_locks_hash(sector_t sect)
  77 {
  78         return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
  79 }
  80
  81 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  82 {
  83         spin_lock_irq(conf->hash_locks + hash);
  84         spin_lock(&conf->device_lock);
  85 }
  86
  87 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  88 {
  89         spin_unlock(&conf->device_lock);
  90         spin_unlock_irq(conf->hash_locks + hash);
  91 }
  92
  93 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
  94 {
  95         int i;
  96         spin_lock_irq(conf->hash_locks);
  97         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
  98                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
  99         spin_lock(&conf->device_lock);
 100 }
 101
 102 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 103 {
 104         int i;
 105         spin_unlock(&conf->device_lock);
 106         for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 107                 spin_unlock(conf->hash_locks + i);
 108         spin_unlock_irq(conf->hash_locks);
 109 }
 110
 111 /* Find first data disk in a raid6 stripe */
 112 static inline int raid6_d0(struct stripe_head *sh)
 113 {
 114         if (sh->ddf_layout)
 115                 /* ddf always start from first device */
 116                 return 0;
 117         /* md starts just after Q block */
 118         if (sh->qd_idx == sh->disks - 1)
 119                 return 0;
 120         else
 121                 return sh->qd_idx + 1;
 122 }
 123 static inline int raid6_next_disk(int disk, int raid_disks)
 124 {
 125         disk++;
 126         return (disk < raid_disks) ? disk : 0;
 127 }
 128
 129 /* When walking through the disks in a raid5, starting at raid6_d0,
 130  * We need to map each disk to a 'slot', where the data disks are slot
 131  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 132  * is raid_disks-1.  This help does that mapping.
 133  */
 134 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 135                              int *count, int syndrome_disks)
 136 {
 137         int slot = *count;
 138
 139         if (sh->ddf_layout)
 140                 (*count)++;
 141         if (idx == sh->pd_idx)
 142                 return syndrome_disks;
 143         if (idx == sh->qd_idx)
 144                 return syndrome_disks + 1;
 145         if (!sh->ddf_layout)
 146                 (*count)++;
 147         return slot;
 148 }
 149
 150 static void print_raid5_conf (struct r5conf *conf);
 151
 152 static int stripe_operations_active(struct stripe_head *sh)
 153 {
 154         return sh->check_state || sh->reconstruct_state ||
 155                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 156                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 157 }
 158
 159 static bool stripe_is_lowprio(struct stripe_head *sh)
 160 {
 161         return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 162                 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 163                !test_bit(STRIPE_R5C_CACHING, &sh->state);
 164 }
 165
 166 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 167 {
 168         struct r5conf *conf = sh->raid_conf;
 169         struct r5worker_group *group;
 170         int thread_cnt;
 171         int i, cpu = sh->cpu;
 172
 173         if (!cpu_online(cpu)) {
 174                 cpu = cpumask_any(cpu_online_mask);
 175                 sh->cpu = cpu;
 176         }
 177
 178         if (list_empty(&sh->lru)) {
 179                 struct r5worker_group *group;
 180                 group = conf->worker_groups + cpu_to_group(cpu);
 181                 if (stripe_is_lowprio(sh))
 182                         list_add_tail(&sh->lru, &group->loprio_list);
 183                 else
 184                         list_add_tail(&sh->lru, &group->handle_list);
 185                 group->stripes_cnt++;
 186                 sh->group = group;
 187         }
 188
 189         if (conf->worker_cnt_per_group == 0) {
 190                 md_wakeup_thread(conf->mddev->thread);
 191                 return;
 192         }
 193
 194         group = conf->worker_groups + cpu_to_group(sh->cpu);
 195
 196         group->workers[0].working = true;
 197         /* at least one worker should run to avoid race */
 198         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 199
 200         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 201         /* wakeup more workers */
 202         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 203                 if (group->workers[i].working == false) {
 204                         group->workers[i].working = true;
 205                         queue_work_on(sh->cpu, raid5_wq,
 206                                       &group->workers[i].work);
 207                         thread_cnt--;
 208                 }
 209         }
 210 }
 211
 212 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 213                               struct list_head *temp_inactive_list)
 214 {
 215         int i;
 216         int injournal = 0;      /* number of date pages with R5_InJournal */
 217
 218         BUG_ON(!list_empty(&sh->lru));
 219         BUG_ON(atomic_read(&conf->active_stripes)==0);
 220
 221         if (r5c_is_writeback(conf->log))
 222                 for (i = sh->disks; i--; )
 223                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 224                                 injournal++;
 225         /*
 226          * In the following cases, the stripe cannot be released to cached
 227          * lists. Therefore, we make the stripe write out and set
 228          * STRIPE_HANDLE:
 229          *   1. when quiesce in r5c write back;
 230          *   2. when resync is requested fot the stripe.
 231          */
 232         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 233             (conf->quiesce && r5c_is_writeback(conf->log) &&
 234              !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 235                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 236                         r5c_make_stripe_write_out(sh);
 237                 set_bit(STRIPE_HANDLE, &sh->state);
 238         }
 239
 240         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 241                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 242                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 243                         list_add_tail(&sh->lru, &conf->delayed_list);
 244                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 245                            sh->bm_seq - conf->seq_write > 0)
 246                         list_add_tail(&sh->lru, &conf->bitmap_list);
 247                 else {
 248                         clear_bit(STRIPE_DELAYED, &sh->state);
 249                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 250                         if (conf->worker_cnt_per_group == 0) {
 251                                 if (stripe_is_lowprio(sh))
 252                                         list_add_tail(&sh->lru,
 253                                                         &conf->loprio_list);
 254                                 else
 255                                         list_add_tail(&sh->lru,
 256                                                         &conf->handle_list);
 257                         } else {
 258                                 raid5_wakeup_stripe_thread(sh);
 259                                 return;
 260                         }
 261                 }
 262                 md_wakeup_thread(conf->mddev->thread);
 263         } else {
 264                 BUG_ON(stripe_operations_active(sh));
 265                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 266                         if (atomic_dec_return(&conf->preread_active_stripes)
 267                             < IO_THRESHOLD)
 268                                 md_wakeup_thread(conf->mddev->thread);
 269                 atomic_dec(&conf->active_stripes);
 270                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 271                         if (!r5c_is_writeback(conf->log))
 272                                 list_add_tail(&sh->lru, temp_inactive_list);
 273                         else {
 274                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 275                                 if (injournal == 0)
 276                                         list_add_tail(&sh->lru, temp_inactive_list);
 277                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 278                                         /* full stripe */
 279                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 280                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 281                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 282                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 283                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 284                                         r5c_check_cached_full_stripe(conf);
 285                                 } else
 286                                         /*
 287                                          * STRIPE_R5C_PARTIAL_STRIPE is set in
 288                                          * r5c_try_caching_write(). No need to
 289                                          * set it again.
 290                                          */
 291                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 292                         }
 293                 }
 294         }
 295 }
 296
 297 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 298                              struct list_head *temp_inactive_list)
 299 {
 300         if (atomic_dec_and_test(&sh->count))
 301                 do_release_stripe(conf, sh, temp_inactive_list);
 302 }
 303
 304 /*
 305  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 306  *
 307  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 308  * given time. Adding stripes only takes device lock, while deleting stripes
 309  * only takes hash lock.
 310  */
 311 static void release_inactive_stripe_list(struct r5conf *conf,
 312                                          struct list_head *temp_inactive_list,
 313                                          int hash)
 314 {
 315         int size;
 316         bool do_wakeup = false;
 317         unsigned long flags;
 318
 319         if (hash == NR_STRIPE_HASH_LOCKS) {
 320                 size = NR_STRIPE_HASH_LOCKS;
 321                 hash = NR_STRIPE_HASH_LOCKS - 1;
 322         } else
 323                 size = 1;
 324         while (size) {
 325                 struct list_head *list = &temp_inactive_list[size - 1];
 326
 327                 /*
 328                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 329                  * remove stripes from the list
 330                  */
 331                 if (!list_empty_careful(list)) {
 332                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 333                         if (list_empty(conf->inactive_list + hash) &&
 334                             !list_empty(list))
 335                                 atomic_dec(&conf->empty_inactive_list_nr);
 336                         list_splice_tail_init(list, conf->inactive_list + hash);
 337                         do_wakeup = true;
 338                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 339                 }
 340                 size--;
 341                 hash--;
 342         }
 343
 344         if (do_wakeup) {
 345                 wake_up(&conf->wait_for_stripe);
 346                 if (atomic_read(&conf->active_stripes) == 0)
 347                         wake_up(&conf->wait_for_quiescent);
 348                 if (conf->retry_read_aligned)
 349                         md_wakeup_thread(conf->mddev->thread);
 350         }
 351 }
 352
 353 /* should hold conf->device_lock already */
 354 static int release_stripe_list(struct r5conf *conf,
 355                                struct list_head *temp_inactive_list)
 356 {
 357         struct stripe_head *sh, *t;
 358         int count = 0;
 359         struct llist_node *head;
 360
 361         head = llist_del_all(&conf->released_stripes);
 362         head = llist_reverse_order(head);
 363         llist_for_each_entry_safe(sh, t, head, release_list) {
 364                 int hash;
 365
 366                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 367                 smp_mb();
 368                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 369                 /*
 370                  * Don't worry the bit is set here, because if the bit is set
 371                  * again, the count is always > 1. This is true for
 372                  * STRIPE_ON_UNPLUG_LIST bit too.
 373                  */
 374                 hash = sh->hash_lock_index;
 375                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 376                 count++;
 377         }
 378
 379         return count;
 380 }
 381
 382 void raid5_release_stripe(struct stripe_head *sh)
 383 {
 384         struct r5conf *conf = sh->raid_conf;
 385         unsigned long flags;
 386         struct list_head list;
 387         int hash;
 388         bool wakeup;
 389
 390         /* Avoid release_list until the last reference.
 391          */
 392         if (atomic_add_unless(&sh->count, -1, 1))
 393                 return;
 394
 395         if (unlikely(!conf->mddev->thread) ||
 396                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 397                 goto slow_path;
 398         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 399         if (wakeup)
 400                 md_wakeup_thread(conf->mddev->thread);
 401         return;
 402 slow_path:
 403         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 404         if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
 405                 INIT_LIST_HEAD(&list);
 406                 hash = sh->hash_lock_index;
 407                 do_release_stripe(conf, sh, &list);
 408                 spin_unlock_irqrestore(&conf->device_lock, flags);
 409                 release_inactive_stripe_list(conf, &list, hash);
 410         }
 411 }
 412
 413 static inline void remove_hash(struct stripe_head *sh)
 414 {
 415         pr_debug("remove_hash(), stripe %llu\n",
 416                 (unsigned long long)sh->sector);
 417
 418         hlist_del_init(&sh->hash);
 419 }
 420
 421 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 422 {
 423         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 424
 425         pr_debug("insert_hash(), stripe %llu\n",
 426                 (unsigned long long)sh->sector);
 427
 428         hlist_add_head(&sh->hash, hp);
 429 }
 430
 431 /* find an idle stripe, make sure it is unhashed, and return it. */
 432 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 433 {
 434         struct stripe_head *sh = NULL;
 435         struct list_head *first;
 436
 437         if (list_empty(conf->inactive_list + hash))
 438                 goto out;
 439         first = (conf->inactive_list + hash)->next;
 440         sh = list_entry(first, struct stripe_head, lru);
 441         list_del_init(first);
 442         remove_hash(sh);
 443         atomic_inc(&conf->active_stripes);
 444         BUG_ON(hash != sh->hash_lock_index);
 445         if (list_empty(conf->inactive_list + hash))
 446                 atomic_inc(&conf->empty_inactive_list_nr);
 447 out:
 448         return sh;
 449 }
 450
 451 static void shrink_buffers(struct stripe_head *sh)
 452 {
 453         struct page *p;
 454         int i;
 455         int num = sh->raid_conf->pool_size;
 456
 457         for (i = 0; i < num ; i++) {
 458                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 459                 p = sh->dev[i].page;
 460                 if (!p)
 461                         continue;
 462                 sh->dev[i].page = NULL;
 463                 put_page(p);
 464         }
 465 }
 466
 467 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 468 {
 469         int i;
 470         int num = sh->raid_conf->pool_size;
 471
 472         for (i = 0; i < num; i++) {
 473                 struct page *page;
 474
 475                 if (!(page = alloc_page(gfp))) {
 476                         return 1;
 477                 }
 478                 sh->dev[i].page = page;
 479                 sh->dev[i].orig_page = page;
 480         }
 481
 482         return 0;
 483 }
 484
 485 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 486                             struct stripe_head *sh);
 487
 488 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 489 {
 490         struct r5conf *conf = sh->raid_conf;
 491         int i, seq;
 492
 493         BUG_ON(atomic_read(&sh->count) != 0);
 494         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 495         BUG_ON(stripe_operations_active(sh));
 496         BUG_ON(sh->batch_head);
 497
 498         pr_debug("init_stripe called, stripe %llu\n",
 499                 (unsigned long long)sector);
 500 retry:
 501         seq = read_seqcount_begin(&conf->gen_lock);
 502         sh->generation = conf->generation - previous;
 503         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 504         sh->sector = sector;
 505         stripe_set_idx(sector, conf, previous, sh);
 506         sh->state = 0;
 507
 508         for (i = sh->disks; i--; ) {
 509                 struct r5dev *dev = &sh->dev[i];
 510
 511                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 512                     test_bit(R5_LOCKED, &dev->flags)) {
 513                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 514                                (unsigned long long)sh->sector, i, dev->toread,
 515                                dev->read, dev->towrite, dev->written,
 516                                test_bit(R5_LOCKED, &dev->flags));
 517                         WARN_ON(1);
 518                 }
 519                 dev->flags = 0;
 520                 dev->sector = raid5_compute_blocknr(sh, i, previous);
 521         }
 522         if (read_seqcount_retry(&conf->gen_lock, seq))
 523                 goto retry;
 524         sh->overwrite_disks = 0;
 525         insert_hash(conf, sh);
 526         sh->cpu = smp_processor_id();
 527         set_bit(STRIPE_BATCH_READY, &sh->state);
 528 }
 529
 530 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 531                                          short generation)
 532 {
 533         struct stripe_head *sh;
 534
 535         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 536         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 537                 if (sh->sector == sector && sh->generation == generation)
 538                         return sh;
 539         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 540         return NULL;
 541 }
 542
 543 /*
 544  * Need to check if array has failed when deciding whether to:
 545  *  - start an array
 546  *  - remove non-faulty devices
 547  *  - add a spare
 548  *  - allow a reshape
 549  * This determination is simple when no reshape is happening.
 550  * However if there is a reshape, we need to carefully check
 551  * both the before and after sections.
 552  * This is because some failed devices may only affect one
 553  * of the two sections, and some non-in_sync devices may
 554  * be insync in the section most affected by failed devices.
 555  */
 556 int raid5_calc_degraded(struct r5conf *conf)
 557 {
 558         int degraded, degraded2;
 559         int i;
 560
 561         rcu_read_lock();
 562         degraded = 0;
 563         for (i = 0; i < conf->previous_raid_disks; i++) {
 564                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 565                 if (rdev && test_bit(Faulty, &rdev->flags))
 566                         rdev = rcu_dereference(conf->disks[i].replacement);
 567                 if (!rdev || test_bit(Faulty, &rdev->flags))
 568                         degraded++;
 569                 else if (test_bit(In_sync, &rdev->flags))
 570                         ;
 571                 else
 572                         /* not in-sync or faulty.
 573                          * If the reshape increases the number of devices,
 574                          * this is being recovered by the reshape, so
 575                          * this 'previous' section is not in_sync.
 576                          * If the number of devices is being reduced however,
 577                          * the device can only be part of the array if
 578                          * we are reverting a reshape, so this section will
 579                          * be in-sync.
 580                          */
 581                         if (conf->raid_disks >= conf->previous_raid_disks)
 582                                 degraded++;
 583         }
 584         rcu_read_unlock();
 585         if (conf->raid_disks == conf->previous_raid_disks)
 586                 return degraded;
 587         rcu_read_lock();
 588         degraded2 = 0;
 589         for (i = 0; i < conf->raid_disks; i++) {
 590                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 591                 if (rdev && test_bit(Faulty, &rdev->flags))
 592                         rdev = rcu_dereference(conf->disks[i].replacement);
 593                 if (!rdev || test_bit(Faulty, &rdev->flags))
 594                         degraded2++;
 595                 else if (test_bit(In_sync, &rdev->flags))
 596                         ;
 597                 else
 598                         /* not in-sync or faulty.
 599                          * If reshape increases the number of devices, this
 600                          * section has already been recovered, else it
 601                          * almost certainly hasn't.
 602                          */
 603                         if (conf->raid_disks <= conf->previous_raid_disks)
 604                                 degraded2++;
 605         }
 606         rcu_read_unlock();
 607         if (degraded2 > degraded)
 608                 return degraded2;
 609         return degraded;
 610 }
 611
 612 static int has_failed(struct r5conf *conf)
 613 {
 614         int degraded;
 615
 616         if (conf->mddev->reshape_position == MaxSector)
 617                 return conf->mddev->degraded > conf->max_degraded;
 618
 619         degraded = raid5_calc_degraded(conf);
 620         if (degraded > conf->max_degraded)
 621                 return 1;
 622         return 0;
 623 }
 624
 625 struct stripe_head *
 626 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 627                         int previous, int noblock, int noquiesce)
 628 {
 629         struct stripe_head *sh;
 630         int hash = stripe_hash_locks_hash(sector);
 631         int inc_empty_inactive_list_flag;
 632
 633         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 634
 635         spin_lock_irq(conf->hash_locks + hash);
 636
 637         do {
 638                 wait_event_lock_irq(conf->wait_for_quiescent,
 639                                     conf->quiesce == 0 || noquiesce,
 640                                     *(conf->hash_locks + hash));
 641                 sh = __find_stripe(conf, sector, conf->generation - previous);
 642                 if (!sh) {
 643                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 644                                 sh = get_free_stripe(conf, hash);
 645                                 if (!sh && !test_bit(R5_DID_ALLOC,
 646                                                      &conf->cache_state))
 647                                         set_bit(R5_ALLOC_MORE,
 648                                                 &conf->cache_state);
 649                         }
 650                         if (noblock && sh == NULL)
 651                                 break;
 652
 653                         r5c_check_stripe_cache_usage(conf);
 654                         if (!sh) {
 655                                 set_bit(R5_INACTIVE_BLOCKED,
 656                                         &conf->cache_state);
 657                                 r5l_wake_reclaim(conf->log, 0);
 658                                 wait_event_lock_irq(
 659                                         conf->wait_for_stripe,
 660                                         !list_empty(conf->inactive_list + hash) &&
 661                                         (atomic_read(&conf->active_stripes)
 662                                          < (conf->max_nr_stripes * 3 / 4)
 663                                          || !test_bit(R5_INACTIVE_BLOCKED,
 664                                                       &conf->cache_state)),
 665                                         *(conf->hash_locks + hash));
 666                                 clear_bit(R5_INACTIVE_BLOCKED,
 667                                           &conf->cache_state);
 668                         } else {
 669                                 init_stripe(sh, sector, previous);
 670                                 atomic_inc(&sh->count);
 671                         }
 672                 } else if (!atomic_inc_not_zero(&sh->count)) {
 673                         spin_lock(&conf->device_lock);
 674                         if (!atomic_read(&sh->count)) {
 675                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 676                                         atomic_inc(&conf->active_stripes);
 677                                 BUG_ON(list_empty(&sh->lru) &&
 678                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 679                                 inc_empty_inactive_list_flag = 0;
 680                                 if (!list_empty(conf->inactive_list + hash))
 681                                         inc_empty_inactive_list_flag = 1;
 682                                 list_del_init(&sh->lru);
 683                                 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 684                                         atomic_inc(&conf->empty_inactive_list_nr);
 685                                 if (sh->group) {
 686                                         sh->group->stripes_cnt--;
 687                                         sh->group = NULL;
 688                                 }
 689                         }
 690                         atomic_inc(&sh->count);
 691                         spin_unlock(&conf->device_lock);
 692                 }
 693         } while (sh == NULL);
 694
 695         spin_unlock_irq(conf->hash_locks + hash);
 696         return sh;
 697 }
 698
 699 static bool is_full_stripe_write(struct stripe_head *sh)
 700 {
 701         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 702         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 703 }
 704
 705 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 706                 __acquires(&sh1->stripe_lock)
 707                 __acquires(&sh2->stripe_lock)
 708 {
 709         if (sh1 > sh2) {
 710                 spin_lock_irq(&sh2->stripe_lock);
 711                 spin_lock_nested(&sh1->stripe_lock, 1);
 712         } else {
 713                 spin_lock_irq(&sh1->stripe_lock);
 714                 spin_lock_nested(&sh2->stripe_lock, 1);
 715         }
 716 }
 717
 718 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 719                 __releases(&sh1->stripe_lock)
 720                 __releases(&sh2->stripe_lock)
 721 {
 722         spin_unlock(&sh1->stripe_lock);
 723         spin_unlock_irq(&sh2->stripe_lock);
 724 }
 725
 726 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 727 static bool stripe_can_batch(struct stripe_head *sh)
 728 {
 729         struct r5conf *conf = sh->raid_conf;
 730
 731         if (raid5_has_log(conf) || raid5_has_ppl(conf))
 732                 return false;
 733         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 734                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 735                 is_full_stripe_write(sh);
 736 }
 737
 738 /* we only do back search */
 739 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 740 {
 741         struct stripe_head *head;
 742         sector_t head_sector, tmp_sec;
 743         int hash;
 744         int dd_idx;
 745         int inc_empty_inactive_list_flag;
 746
 747         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 748         tmp_sec = sh->sector;
 749         if (!sector_div(tmp_sec, conf->chunk_sectors))
 750                 return;
 751         head_sector = sh->sector - STRIPE_SECTORS;
 752
 753         hash = stripe_hash_locks_hash(head_sector);
 754         spin_lock_irq(conf->hash_locks + hash);
 755         head = __find_stripe(conf, head_sector, conf->generation);
 756         if (head && !atomic_inc_not_zero(&head->count)) {
 757                 spin_lock(&conf->device_lock);
 758                 if (!atomic_read(&head->count)) {
 759                         if (!test_bit(STRIPE_HANDLE, &head->state))
 760                                 atomic_inc(&conf->active_stripes);
 761                         BUG_ON(list_empty(&head->lru) &&
 762                                !test_bit(STRIPE_EXPANDING, &head->state));
 763                         inc_empty_inactive_list_flag = 0;
 764                         if (!list_empty(conf->inactive_list + hash))
 765                                 inc_empty_inactive_list_flag = 1;
 766                         list_del_init(&head->lru);
 767                         if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 768                                 atomic_inc(&conf->empty_inactive_list_nr);
 769                         if (head->group) {
 770                                 head->group->stripes_cnt--;
 771                                 head->group = NULL;
 772                         }
 773                 }
 774                 atomic_inc(&head->count);
 775                 spin_unlock(&conf->device_lock);
 776         }
 777         spin_unlock_irq(conf->hash_locks + hash);
 778
 779         if (!head)
 780                 return;
 781         if (!stripe_can_batch(head))
 782                 goto out;
 783
 784         lock_two_stripes(head, sh);
 785         /* clear_batch_ready clear the flag */
 786         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 787                 goto unlock_out;
 788
 789         if (sh->batch_head)
 790                 goto unlock_out;
 791
 792         dd_idx = 0;
 793         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 794                 dd_idx++;
 795         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 796             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 797                 goto unlock_out;
 798
 799         if (head->batch_head) {
 800                 spin_lock(&head->batch_head->batch_lock);
 801                 /* This batch list is already running */
 802                 if (!stripe_can_batch(head)) {
 803                         spin_unlock(&head->batch_head->batch_lock);
 804                         goto unlock_out;
 805                 }
 806                 /*
 807                  * We must assign batch_head of this stripe within the
 808                  * batch_lock, otherwise clear_batch_ready of batch head
 809                  * stripe could clear BATCH_READY bit of this stripe and
 810                  * this stripe->batch_head doesn't get assigned, which
 811                  * could confuse clear_batch_ready for this stripe
 812                  */
 813                 sh->batch_head = head->batch_head;
 814
 815                 /*
 816                  * at this point, head's BATCH_READY could be cleared, but we
 817                  * can still add the stripe to batch list
 818                  */
 819                 list_add(&sh->batch_list, &head->batch_list);
 820                 spin_unlock(&head->batch_head->batch_lock);
 821         } else {
 822                 head->batch_head = head;
 823                 sh->batch_head = head->batch_head;
 824                 spin_lock(&head->batch_lock);
 825                 list_add_tail(&sh->batch_list, &head->batch_list);
 826                 spin_unlock(&head->batch_lock);
 827         }
 828
 829         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 830                 if (atomic_dec_return(&conf->preread_active_stripes)
 831                     < IO_THRESHOLD)
 832                         md_wakeup_thread(conf->mddev->thread);
 833
 834         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 835                 int seq = sh->bm_seq;
 836                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 837                     sh->batch_head->bm_seq > seq)
 838                         seq = sh->batch_head->bm_seq;
 839                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 840                 sh->batch_head->bm_seq = seq;
 841         }
 842
 843         atomic_inc(&sh->count);
 844 unlock_out:
 845         unlock_two_stripes(head, sh);
 846 out:
 847         raid5_release_stripe(head);
 848 }
 849
 850 /* Determine if 'data_offset' or 'new_data_offset' should be used
 851  * in this stripe_head.
 852  */
 853 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 854 {
 855         sector_t progress = conf->reshape_progress;
 856         /* Need a memory barrier to make sure we see the value
 857          * of conf->generation, or ->data_offset that was set before
 858          * reshape_progress was updated.
 859          */
 860         smp_rmb();
 861         if (progress == MaxSector)
 862                 return 0;
 863         if (sh->generation == conf->generation - 1)
 864                 return 0;
 865         /* We are in a reshape, and this is a new-generation stripe,
 866          * so use new_data_offset.
 867          */
 868         return 1;
 869 }
 870
 871 static void dispatch_bio_list(struct bio_list *tmp)
 872 {
 873         struct bio *bio;
 874
 875         while ((bio = bio_list_pop(tmp)))
 876                 generic_make_request(bio);
 877 }
 878
 879 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
 880 {
 881         const struct r5pending_data *da = list_entry(a,
 882                                 struct r5pending_data, sibling);
 883         const struct r5pending_data *db = list_entry(b,
 884                                 struct r5pending_data, sibling);
 885         if (da->sector > db->sector)
 886                 return 1;
 887         if (da->sector < db->sector)
 888                 return -1;
 889         return 0;
 890 }
 891
 892 static void dispatch_defer_bios(struct r5conf *conf, int target,
 893                                 struct bio_list *list)
 894 {
 895         struct r5pending_data *data;
 896         struct list_head *first, *next = NULL;
 897         int cnt = 0;
 898
 899         if (conf->pending_data_cnt == 0)
 900                 return;
 901
 902         list_sort(NULL, &conf->pending_list, cmp_stripe);
 903
 904         first = conf->pending_list.next;
 905
 906         /* temporarily move the head */
 907         if (conf->next_pending_data)
 908                 list_move_tail(&conf->pending_list,
 909                                 &conf->next_pending_data->sibling);
 910
 911         while (!list_empty(&conf->pending_list)) {
 912                 data = list_first_entry(&conf->pending_list,
 913                         struct r5pending_data, sibling);
 914                 if (&data->sibling == first)
 915                         first = data->sibling.next;
 916                 next = data->sibling.next;
 917
 918                 bio_list_merge(list, &data->bios);
 919                 list_move(&data->sibling, &conf->free_list);
 920                 cnt++;
 921                 if (cnt >= target)
 922                         break;
 923         }
 924         conf->pending_data_cnt -= cnt;
 925         BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
 926
 927         if (next != &conf->pending_list)
 928                 conf->next_pending_data = list_entry(next,
 929                                 struct r5pending_data, sibling);
 930         else
 931                 conf->next_pending_data = NULL;
 932         /* list isn't empty */
 933         if (first != &conf->pending_list)
 934                 list_move_tail(&conf->pending_list, first);
 935 }
 936
 937 static void flush_deferred_bios(struct r5conf *conf)
 938 {
 939         struct bio_list tmp = BIO_EMPTY_LIST;
 940
 941         if (conf->pending_data_cnt == 0)
 942                 return;
 943
 944         spin_lock(&conf->pending_bios_lock);
 945         dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
 946         BUG_ON(conf->pending_data_cnt != 0);
 947         spin_unlock(&conf->pending_bios_lock);
 948
 949         dispatch_bio_list(&tmp);
 950 }
 951
 952 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
 953                                 struct bio_list *bios)
 954 {
 955         struct bio_list tmp = BIO_EMPTY_LIST;
 956         struct r5pending_data *ent;
 957
 958         spin_lock(&conf->pending_bios_lock);
 959         ent = list_first_entry(&conf->free_list, struct r5pending_data,
 960                                                         sibling);
 961         list_move_tail(&ent->sibling, &conf->pending_list);
 962         ent->sector = sector;
 963         bio_list_init(&ent->bios);
 964         bio_list_merge(&ent->bios, bios);
 965         conf->pending_data_cnt++;
 966         if (conf->pending_data_cnt >= PENDING_IO_MAX)
 967                 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
 968
 969         spin_unlock(&conf->pending_bios_lock);
 970
 971         dispatch_bio_list(&tmp);
 972 }
 973
 974 static void
 975 raid5_end_read_request(struct bio *bi);
 976 static void
 977 raid5_end_write_request(struct bio *bi);
 978
 979 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 980 {
 981         struct r5conf *conf = sh->raid_conf;
 982         int i, disks = sh->disks;
 983         struct stripe_head *head_sh = sh;
 984         struct bio_list pending_bios = BIO_EMPTY_LIST;
 985         bool should_defer;
 986
 987         might_sleep();
 988
 989         if (log_stripe(sh, s) == 0)
 990                 return;
 991
 992         should_defer = conf->batch_bio_dispatch && conf->group_cnt;
 993
 994         for (i = disks; i--; ) {
 995                 int op, op_flags = 0;
 996                 int replace_only = 0;
 997                 struct bio *bi, *rbi;
 998                 struct md_rdev *rdev, *rrdev = NULL;
 999
1000                 sh = head_sh;
1001                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1002                         op = REQ_OP_WRITE;
1003                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1004                                 op_flags = REQ_FUA;
1005                         if (test_bit(R5_Discard, &sh->dev[i].flags))
1006                                 op = REQ_OP_DISCARD;
1007                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1008                         op = REQ_OP_READ;
1009                 else if (test_and_clear_bit(R5_WantReplace,
1010                                             &sh->dev[i].flags)) {
1011                         op = REQ_OP_WRITE;
1012                         replace_only = 1;
1013                 } else
1014                         continue;
1015                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1016                         op_flags |= REQ_SYNC;
1017
1018 again:
1019                 bi = &sh->dev[i].req;
1020                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1021
1022                 rcu_read_lock();
1023                 rrdev = rcu_dereference(conf->disks[i].replacement);
1024                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1025                 rdev = rcu_dereference(conf->disks[i].rdev);
1026                 if (!rdev) {
1027                         rdev = rrdev;
1028                         rrdev = NULL;
1029                 }
1030                 if (op_is_write(op)) {
1031                         if (replace_only)
1032                                 rdev = NULL;
1033                         if (rdev == rrdev)
1034                                 /* We raced and saw duplicates */
1035                                 rrdev = NULL;
1036                 } else {
1037                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1038                                 rdev = rrdev;
1039                         rrdev = NULL;
1040                 }
1041
1042                 if (rdev && test_bit(Faulty, &rdev->flags))
1043                         rdev = NULL;
1044                 if (rdev)
1045                         atomic_inc(&rdev->nr_pending);
1046                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1047                         rrdev = NULL;
1048                 if (rrdev)
1049                         atomic_inc(&rrdev->nr_pending);
1050                 rcu_read_unlock();
1051
1052                 /* We have already checked bad blocks for reads.  Now
1053                  * need to check for writes.  We never accept write errors
1054                  * on the replacement, so we don't to check rrdev.
1055                  */
1056                 while (op_is_write(op) && rdev &&
1057                        test_bit(WriteErrorSeen, &rdev->flags)) {
1058                         sector_t first_bad;
1059                         int bad_sectors;
1060                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1061                                               &first_bad, &bad_sectors);
1062                         if (!bad)
1063                                 break;
1064
1065                         if (bad < 0) {
1066                                 set_bit(BlockedBadBlocks, &rdev->flags);
1067                                 if (!conf->mddev->external &&
1068                                     conf->mddev->sb_flags) {
1069                                         /* It is very unlikely, but we might
1070                                          * still need to write out the
1071                                          * bad block log - better give it
1072                                          * a chance*/
1073                                         md_check_recovery(conf->mddev);
1074                                 }
1075                                 /*
1076                                  * Because md_wait_for_blocked_rdev
1077                                  * will dec nr_pending, we must
1078                                  * increment it first.
1079                                  */
1080                                 atomic_inc(&rdev->nr_pending);
1081                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
1082                         } else {
1083                                 /* Acknowledged bad block - skip the write */
1084                                 rdev_dec_pending(rdev, conf->mddev);
1085                                 rdev = NULL;
1086                         }
1087                 }
1088
1089                 if (rdev) {
1090                         if (s->syncing || s->expanding || s->expanded
1091                             || s->replacing)
1092                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1093
1094                         set_bit(STRIPE_IO_STARTED, &sh->state);
1095
1096                         bio_set_dev(bi, rdev->bdev);
1097                         bio_set_op_attrs(bi, op, op_flags);
1098                         bi->bi_end_io = op_is_write(op)
1099                                 ? raid5_end_write_request
1100                                 : raid5_end_read_request;
1101                         bi->bi_private = sh;
1102
1103                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1104                                 __func__, (unsigned long long)sh->sector,
1105                                 bi->bi_opf, i);
1106                         atomic_inc(&sh->count);
1107                         if (sh != head_sh)
1108                                 atomic_inc(&head_sh->count);
1109                         if (use_new_offset(conf, sh))
1110                                 bi->bi_iter.bi_sector = (sh->sector
1111                                                  + rdev->new_data_offset);
1112                         else
1113                                 bi->bi_iter.bi_sector = (sh->sector
1114                                                  + rdev->data_offset);
1115                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1116                                 bi->bi_opf |= REQ_NOMERGE;
1117
1118                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1119                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1120
1121                         if (!op_is_write(op) &&
1122                             test_bit(R5_InJournal, &sh->dev[i].flags))
1123                                 /*
1124                                  * issuing read for a page in journal, this
1125                                  * must be preparing for prexor in rmw; read
1126                                  * the data into orig_page
1127                                  */
1128                                 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1129                         else
1130                                 sh->dev[i].vec.bv_page = sh->dev[i].page;
1131                         bi->bi_vcnt = 1;
1132                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1133                         bi->bi_io_vec[0].bv_offset = 0;
1134                         bi->bi_iter.bi_size = STRIPE_SIZE;
1135                         bi->bi_write_hint = sh->dev[i].write_hint;
1136                         if (!rrdev)
1137                                 sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1138                         /*
1139                          * If this is discard request, set bi_vcnt 0. We don't
1140                          * want to confuse SCSI because SCSI will replace payload
1141                          */
1142                         if (op == REQ_OP_DISCARD)
1143                                 bi->bi_vcnt = 0;
1144                         if (rrdev)
1145                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1146
1147                         if (conf->mddev->gendisk)
1148                                 trace_block_bio_remap(bi->bi_disk->queue,
1149                                                       bi, disk_devt(conf->mddev->gendisk),
1150                                                       sh->dev[i].sector);
1151                         if (should_defer && op_is_write(op))
1152                                 bio_list_add(&pending_bios, bi);
1153                         else
1154                                 generic_make_request(bi);
1155                 }
1156                 if (rrdev) {
1157                         if (s->syncing || s->expanding || s->expanded
1158                             || s->replacing)
1159                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1160
1161                         set_bit(STRIPE_IO_STARTED, &sh->state);
1162
1163                         bio_set_dev(rbi, rrdev->bdev);
1164                         bio_set_op_attrs(rbi, op, op_flags);
1165                         BUG_ON(!op_is_write(op));
1166                         rbi->bi_end_io = raid5_end_write_request;
1167                         rbi->bi_private = sh;
1168
1169                         pr_debug("%s: for %llu schedule op %d on "
1170                                  "replacement disc %d\n",
1171                                 __func__, (unsigned long long)sh->sector,
1172                                 rbi->bi_opf, i);
1173                         atomic_inc(&sh->count);
1174                         if (sh != head_sh)
1175                                 atomic_inc(&head_sh->count);
1176                         if (use_new_offset(conf, sh))
1177                                 rbi->bi_iter.bi_sector = (sh->sector
1178                                                   + rrdev->new_data_offset);
1179                         else
1180                                 rbi->bi_iter.bi_sector = (sh->sector
1181                                                   + rrdev->data_offset);
1182                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1183                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1184                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1185                         rbi->bi_vcnt = 1;
1186                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1187                         rbi->bi_io_vec[0].bv_offset = 0;
1188                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1189                         rbi->bi_write_hint = sh->dev[i].write_hint;
1190                         sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1191                         /*
1192                          * If this is discard request, set bi_vcnt 0. We don't
1193                          * want to confuse SCSI because SCSI will replace payload
1194                          */
1195                         if (op == REQ_OP_DISCARD)
1196                                 rbi->bi_vcnt = 0;
1197                         if (conf->mddev->gendisk)
1198                                 trace_block_bio_remap(rbi->bi_disk->queue,
1199                                                       rbi, disk_devt(conf->mddev->gendisk),
1200                                                       sh->dev[i].sector);
1201                         if (should_defer && op_is_write(op))
1202                                 bio_list_add(&pending_bios, rbi);
1203                         else
1204                                 generic_make_request(rbi);
1205                 }
1206                 if (!rdev && !rrdev) {
1207                         if (op_is_write(op))
1208                                 set_bit(STRIPE_DEGRADED, &sh->state);
1209                         pr_debug("skip op %d on disc %d for sector %llu\n",
1210                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1211                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1212                         set_bit(STRIPE_HANDLE, &sh->state);
1213                 }
1214
1215                 if (!head_sh->batch_head)
1216                         continue;
1217                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1218                                       batch_list);
1219                 if (sh != head_sh)
1220                         goto again;
1221         }
1222
1223         if (should_defer && !bio_list_empty(&pending_bios))
1224                 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1225 }
1226
1227 static struct dma_async_tx_descriptor *
1228 async_copy_data(int frombio, struct bio *bio, struct page **page,
1229         sector_t sector, struct dma_async_tx_descriptor *tx,
1230         struct stripe_head *sh, int no_skipcopy)
1231 {
1232         struct bio_vec bvl;
1233         struct bvec_iter iter;
1234         struct page *bio_page;
1235         int page_offset;
1236         struct async_submit_ctl submit;
1237         enum async_tx_flags flags = 0;
1238
1239         if (bio->bi_iter.bi_sector >= sector)
1240                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1241         else
1242                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1243
1244         if (frombio)
1245                 flags |= ASYNC_TX_FENCE;
1246         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1247
1248         bio_for_each_segment(bvl, bio, iter) {
1249                 int len = bvl.bv_len;
1250                 int clen;
1251                 int b_offset = 0;
1252
1253                 if (page_offset < 0) {
1254                         b_offset = -page_offset;
1255                         page_offset += b_offset;
1256                         len -= b_offset;
1257                 }
1258
1259                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1260                         clen = STRIPE_SIZE - page_offset;
1261                 else
1262                         clen = len;
1263
1264                 if (clen > 0) {
1265                         b_offset += bvl.bv_offset;
1266                         bio_page = bvl.bv_page;
1267                         if (frombio) {
1268                                 if (sh->raid_conf->skip_copy &&
1269                                     b_offset == 0 && page_offset == 0 &&
1270                                     clen == STRIPE_SIZE &&
1271                                     !no_skipcopy)
1272                                         *page = bio_page;
1273                                 else
1274                                         tx = async_memcpy(*page, bio_page, page_offset,
1275                                                   b_offset, clen, &submit);
1276                         } else
1277                                 tx = async_memcpy(bio_page, *page, b_offset,
1278                                                   page_offset, clen, &submit);
1279                 }
1280                 /* chain the operations */
1281                 submit.depend_tx = tx;
1282
1283                 if (clen < len) /* hit end of page */
1284                         break;
1285                 page_offset +=  len;
1286         }
1287
1288         return tx;
1289 }
1290
1291 static void ops_complete_biofill(void *stripe_head_ref)
1292 {
1293         struct stripe_head *sh = stripe_head_ref;
1294         int i;
1295
1296         pr_debug("%s: stripe %llu\n", __func__,
1297                 (unsigned long long)sh->sector);
1298
1299         /* clear completed biofills */
1300         for (i = sh->disks; i--; ) {
1301                 struct r5dev *dev = &sh->dev[i];
1302
1303                 /* acknowledge completion of a biofill operation */
1304                 /* and check if we need to reply to a read request,
1305                  * new R5_Wantfill requests are held off until
1306                  * !STRIPE_BIOFILL_RUN
1307                  */
1308                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1309                         struct bio *rbi, *rbi2;
1310
1311                         BUG_ON(!dev->read);
1312                         rbi = dev->read;
1313                         dev->read = NULL;
1314                         while (rbi && rbi->bi_iter.bi_sector <
1315                                 dev->sector + STRIPE_SECTORS) {
1316                                 rbi2 = r5_next_bio(rbi, dev->sector);
1317                                 bio_endio(rbi);
1318                                 rbi = rbi2;
1319                         }
1320                 }
1321         }
1322         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1323
1324         set_bit(STRIPE_HANDLE, &sh->state);
1325         raid5_release_stripe(sh);
1326 }
1327
1328 static void ops_run_biofill(struct stripe_head *sh)
1329 {
1330         struct dma_async_tx_descriptor *tx = NULL;
1331         struct async_submit_ctl submit;
1332         int i;
1333
1334         BUG_ON(sh->batch_head);
1335         pr_debug("%s: stripe %llu\n", __func__,
1336                 (unsigned long long)sh->sector);
1337
1338         for (i = sh->disks; i--; ) {
1339                 struct r5dev *dev = &sh->dev[i];
1340                 if (test_bit(R5_Wantfill, &dev->flags)) {
1341                         struct bio *rbi;
1342                         spin_lock_irq(&sh->stripe_lock);
1343                         dev->read = rbi = dev->toread;
1344                         dev->toread = NULL;
1345                         spin_unlock_irq(&sh->stripe_lock);
1346                         while (rbi && rbi->bi_iter.bi_sector <
1347                                 dev->sector + STRIPE_SECTORS) {
1348                                 tx = async_copy_data(0, rbi, &dev->page,
1349                                                      dev->sector, tx, sh, 0);
1350                                 rbi = r5_next_bio(rbi, dev->sector);
1351                         }
1352                 }
1353         }
1354
1355         atomic_inc(&sh->count);
1356         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1357         async_trigger_callback(&submit);
1358 }
1359
1360 static void mark_target_uptodate(struct stripe_head *sh, int target)
1361 {
1362         struct r5dev *tgt;
1363
1364         if (target < 0)
1365                 return;
1366
1367         tgt = &sh->dev[target];
1368         set_bit(R5_UPTODATE, &tgt->flags);
1369         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1370         clear_bit(R5_Wantcompute, &tgt->flags);
1371 }
1372
1373 static void ops_complete_compute(void *stripe_head_ref)
1374 {
1375         struct stripe_head *sh = stripe_head_ref;
1376
1377         pr_debug("%s: stripe %llu\n", __func__,
1378                 (unsigned long long)sh->sector);
1379
1380         /* mark the computed target(s) as uptodate */
1381         mark_target_uptodate(sh, sh->ops.target);
1382         mark_target_uptodate(sh, sh->ops.target2);
1383
1384         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1385         if (sh->check_state == check_state_compute_run)
1386                 sh->check_state = check_state_compute_result;
1387         set_bit(STRIPE_HANDLE, &sh->state);
1388         raid5_release_stripe(sh);
1389 }
1390
1391 /* return a pointer to the address conversion region of the scribble buffer */
1392 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1393 {
1394         return percpu->scribble + i * percpu->scribble_obj_size;
1395 }
1396
1397 /* return a pointer to the address conversion region of the scribble buffer */
1398 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1399                                  struct raid5_percpu *percpu, int i)
1400 {
1401         return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1402 }
1403
1404 static struct dma_async_tx_descriptor *
1405 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1406 {
1407         int disks = sh->disks;
1408         struct page **xor_srcs = to_addr_page(percpu, 0);
1409         int target = sh->ops.target;
1410         struct r5dev *tgt = &sh->dev[target];
1411         struct page *xor_dest = tgt->page;
1412         int count = 0;
1413         struct dma_async_tx_descriptor *tx;
1414         struct async_submit_ctl submit;
1415         int i;
1416
1417         BUG_ON(sh->batch_head);
1418
1419         pr_debug("%s: stripe %llu block: %d\n",
1420                 __func__, (unsigned long long)sh->sector, target);
1421         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1422
1423         for (i = disks; i--; )
1424                 if (i != target)
1425                         xor_srcs[count++] = sh->dev[i].page;
1426
1427         atomic_inc(&sh->count);
1428
1429         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1430                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1431         if (unlikely(count == 1))
1432                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1433         else
1434                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1435
1436         return tx;
1437 }
1438
1439 /* set_syndrome_sources - populate source buffers for gen_syndrome
1440  * @srcs - (struct page *) array of size sh->disks
1441  * @sh - stripe_head to parse
1442  *
1443  * Populates srcs in proper layout order for the stripe and returns the
1444  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1445  * destination buffer is recorded in srcs[count] and the Q destination
1446  * is recorded in srcs[count+1]].
1447  */
1448 static int set_syndrome_sources(struct page **srcs,
1449                                 struct stripe_head *sh,
1450                                 int srctype)
1451 {
1452         int disks = sh->disks;
1453         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1454         int d0_idx = raid6_d0(sh);
1455         int count;
1456         int i;
1457
1458         for (i = 0; i < disks; i++)
1459                 srcs[i] = NULL;
1460
1461         count = 0;
1462         i = d0_idx;
1463         do {
1464                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1465                 struct r5dev *dev = &sh->dev[i];
1466
1467                 if (i == sh->qd_idx || i == sh->pd_idx ||
1468                     (srctype == SYNDROME_SRC_ALL) ||
1469                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1470                      (test_bit(R5_Wantdrain, &dev->flags) ||
1471                       test_bit(R5_InJournal, &dev->flags))) ||
1472                     (srctype == SYNDROME_SRC_WRITTEN &&
1473                      (dev->written ||
1474                       test_bit(R5_InJournal, &dev->flags)))) {
1475                         if (test_bit(R5_InJournal, &dev->flags))
1476                                 srcs[slot] = sh->dev[i].orig_page;
1477                         else
1478                                 srcs[slot] = sh->dev[i].page;
1479                 }
1480                 i = raid6_next_disk(i, disks);
1481         } while (i != d0_idx);
1482
1483         return syndrome_disks;
1484 }
1485
1486 static struct dma_async_tx_descriptor *
1487 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1488 {
1489         int disks = sh->disks;
1490         struct page **blocks = to_addr_page(percpu, 0);
1491         int target;
1492         int qd_idx = sh->qd_idx;
1493         struct dma_async_tx_descriptor *tx;
1494         struct async_submit_ctl submit;
1495         struct r5dev *tgt;
1496         struct page *dest;
1497         int i;
1498         int count;
1499
1500         BUG_ON(sh->batch_head);
1501         if (sh->ops.target < 0)
1502                 target = sh->ops.target2;
1503         else if (sh->ops.target2 < 0)
1504                 target = sh->ops.target;
1505         else
1506                 /* we should only have one valid target */
1507                 BUG();
1508         BUG_ON(target < 0);
1509         pr_debug("%s: stripe %llu block: %d\n",
1510                 __func__, (unsigned long long)sh->sector, target);
1511
1512         tgt = &sh->dev[target];
1513         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1514         dest = tgt->page;
1515
1516         atomic_inc(&sh->count);
1517
1518         if (target == qd_idx) {
1519                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1520                 blocks[count] = NULL; /* regenerating p is not necessary */
1521                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1522                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1523                                   ops_complete_compute, sh,
1524                                   to_addr_conv(sh, percpu, 0));
1525                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1526         } else {
1527                 /* Compute any data- or p-drive using XOR */
1528                 count = 0;
1529                 for (i = disks; i-- ; ) {
1530                         if (i == target || i == qd_idx)
1531                                 continue;
1532                         blocks[count++] = sh->dev[i].page;
1533                 }
1534
1535                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1536                                   NULL, ops_complete_compute, sh,
1537                                   to_addr_conv(sh, percpu, 0));
1538                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1539         }
1540
1541         return tx;
1542 }
1543
1544 static struct dma_async_tx_descriptor *
1545 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1546 {
1547         int i, count, disks = sh->disks;
1548         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1549         int d0_idx = raid6_d0(sh);
1550         int faila = -1, failb = -1;
1551         int target = sh->ops.target;
1552         int target2 = sh->ops.target2;
1553         struct r5dev *tgt = &sh->dev[target];
1554         struct r5dev *tgt2 = &sh->dev[target2];
1555         struct dma_async_tx_descriptor *tx;
1556         struct page **blocks = to_addr_page(percpu, 0);
1557         struct async_submit_ctl submit;
1558
1559         BUG_ON(sh->batch_head);
1560         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1561                  __func__, (unsigned long long)sh->sector, target, target2);
1562         BUG_ON(target < 0 || target2 < 0);
1563         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1564         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1565
1566         /* we need to open-code set_syndrome_sources to handle the
1567          * slot number conversion for 'faila' and 'failb'
1568          */
1569         for (i = 0; i < disks ; i++)
1570                 blocks[i] = NULL;
1571         count = 0;
1572         i = d0_idx;
1573         do {
1574                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1575
1576                 blocks[slot] = sh->dev[i].page;
1577
1578                 if (i == target)
1579                         faila = slot;
1580                 if (i == target2)
1581                         failb = slot;
1582                 i = raid6_next_disk(i, disks);
1583         } while (i != d0_idx);
1584
1585         BUG_ON(faila == failb);
1586         if (failb < faila)
1587                 swap(faila, failb);
1588         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1589                  __func__, (unsigned long long)sh->sector, faila, failb);
1590
1591         atomic_inc(&sh->count);
1592
1593         if (failb == syndrome_disks+1) {
1594                 /* Q disk is one of the missing disks */
1595                 if (faila == syndrome_disks) {
1596                         /* Missing P+Q, just recompute */
1597                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1598                                           ops_complete_compute, sh,
1599                                           to_addr_conv(sh, percpu, 0));
1600                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1601                                                   STRIPE_SIZE, &submit);
1602                 } else {
1603                         struct page *dest;
1604                         int data_target;
1605                         int qd_idx = sh->qd_idx;
1606
1607                         /* Missing D+Q: recompute D from P, then recompute Q */
1608                         if (target == qd_idx)
1609                                 data_target = target2;
1610                         else
1611                                 data_target = target;
1612
1613                         count = 0;
1614                         for (i = disks; i-- ; ) {
1615                                 if (i == data_target || i == qd_idx)
1616                                         continue;
1617                                 blocks[count++] = sh->dev[i].page;
1618                         }
1619                         dest = sh->dev[data_target].page;
1620                         init_async_submit(&submit,
1621                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1622                                           NULL, NULL, NULL,
1623                                           to_addr_conv(sh, percpu, 0));
1624                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1625                                        &submit);
1626
1627                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1628                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1629                                           ops_complete_compute, sh,
1630                                           to_addr_conv(sh, percpu, 0));
1631                         return async_gen_syndrome(blocks, 0, count+2,
1632                                                   STRIPE_SIZE, &submit);
1633                 }
1634         } else {
1635                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1636                                   ops_complete_compute, sh,
1637                                   to_addr_conv(sh, percpu, 0));
1638                 if (failb == syndrome_disks) {
1639                         /* We're missing D+P. */
1640                         return async_raid6_datap_recov(syndrome_disks+2,
1641                                                        STRIPE_SIZE, faila,
1642                                                        blocks, &submit);
1643                 } else {
1644                         /* We're missing D+D. */
1645                         return async_raid6_2data_recov(syndrome_disks+2,
1646                                                        STRIPE_SIZE, faila, failb,
1647                                                        blocks, &submit);
1648                 }
1649         }
1650 }
1651
1652 static void ops_complete_prexor(void *stripe_head_ref)
1653 {
1654         struct stripe_head *sh = stripe_head_ref;
1655
1656         pr_debug("%s: stripe %llu\n", __func__,
1657                 (unsigned long long)sh->sector);
1658
1659         if (r5c_is_writeback(sh->raid_conf->log))
1660                 /*
1661                  * raid5-cache write back uses orig_page during prexor.
1662                  * After prexor, it is time to free orig_page
1663                  */
1664                 r5c_release_extra_page(sh);
1665 }
1666
1667 static struct dma_async_tx_descriptor *
1668 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1669                 struct dma_async_tx_descriptor *tx)
1670 {
1671         int disks = sh->disks;
1672         struct page **xor_srcs = to_addr_page(percpu, 0);
1673         int count = 0, pd_idx = sh->pd_idx, i;
1674         struct async_submit_ctl submit;
1675
1676         /* existing parity data subtracted */
1677         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1678
1679         BUG_ON(sh->batch_head);
1680         pr_debug("%s: stripe %llu\n", __func__,
1681                 (unsigned long long)sh->sector);
1682
1683         for (i = disks; i--; ) {
1684                 struct r5dev *dev = &sh->dev[i];
1685                 /* Only process blocks that are known to be uptodate */
1686                 if (test_bit(R5_InJournal, &dev->flags))
1687                         xor_srcs[count++] = dev->orig_page;
1688                 else if (test_bit(R5_Wantdrain, &dev->flags))
1689                         xor_srcs[count++] = dev->page;
1690         }
1691
1692         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1693                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1694         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1695
1696         return tx;
1697 }
1698
1699 static struct dma_async_tx_descriptor *
1700 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1701                 struct dma_async_tx_descriptor *tx)
1702 {
1703         struct page **blocks = to_addr_page(percpu, 0);
1704         int count;
1705         struct async_submit_ctl submit;
1706
1707         pr_debug("%s: stripe %llu\n", __func__,
1708                 (unsigned long long)sh->sector);
1709
1710         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1711
1712         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1713                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1714         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1715
1716         return tx;
1717 }
1718
1719 static struct dma_async_tx_descriptor *
1720 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1721 {
1722         struct r5conf *conf = sh->raid_conf;
1723         int disks = sh->disks;
1724         int i;
1725         struct stripe_head *head_sh = sh;
1726
1727         pr_debug("%s: stripe %llu\n", __func__,
1728                 (unsigned long long)sh->sector);
1729
1730         for (i = disks; i--; ) {
1731                 struct r5dev *dev;
1732                 struct bio *chosen;
1733
1734                 sh = head_sh;
1735                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1736                         struct bio *wbi;
1737
1738 again:
1739                         dev = &sh->dev[i];
1740                         /*
1741                          * clear R5_InJournal, so when rewriting a page in
1742                          * journal, it is not skipped by r5l_log_stripe()
1743                          */
1744                         clear_bit(R5_InJournal, &dev->flags);
1745                         spin_lock_irq(&sh->stripe_lock);
1746                         chosen = dev->towrite;
1747                         dev->towrite = NULL;
1748                         sh->overwrite_disks = 0;
1749                         BUG_ON(dev->written);
1750                         wbi = dev->written = chosen;
1751                         spin_unlock_irq(&sh->stripe_lock);
1752                         WARN_ON(dev->page != dev->orig_page);
1753
1754                         while (wbi && wbi->bi_iter.bi_sector <
1755                                 dev->sector + STRIPE_SECTORS) {
1756                                 if (wbi->bi_opf & REQ_FUA)
1757                                         set_bit(R5_WantFUA, &dev->flags);
1758                                 if (wbi->bi_opf & REQ_SYNC)
1759                                         set_bit(R5_SyncIO, &dev->flags);
1760                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1761                                         set_bit(R5_Discard, &dev->flags);
1762                                 else {
1763                                         tx = async_copy_data(1, wbi, &dev->page,
1764                                                              dev->sector, tx, sh,
1765                                                              r5c_is_writeback(conf->log));
1766                                         if (dev->page != dev->orig_page &&
1767                                             !r5c_is_writeback(conf->log)) {
1768                                                 set_bit(R5_SkipCopy, &dev->flags);
1769                                                 clear_bit(R5_UPTODATE, &dev->flags);
1770                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1771                                         }
1772                                 }
1773                                 wbi = r5_next_bio(wbi, dev->sector);
1774                         }
1775
1776                         if (head_sh->batch_head) {
1777                                 sh = list_first_entry(&sh->batch_list,
1778                                                       struct stripe_head,
1779                                                       batch_list);
1780                                 if (sh == head_sh)
1781                                         continue;
1782                                 goto again;
1783                         }
1784                 }
1785         }
1786
1787         return tx;
1788 }
1789
1790 static void ops_complete_reconstruct(void *stripe_head_ref)
1791 {
1792         struct stripe_head *sh = stripe_head_ref;
1793         int disks = sh->disks;
1794         int pd_idx = sh->pd_idx;
1795         int qd_idx = sh->qd_idx;
1796         int i;
1797         bool fua = false, sync = false, discard = false;
1798
1799         pr_debug("%s: stripe %llu\n", __func__,
1800                 (unsigned long long)sh->sector);
1801
1802         for (i = disks; i--; ) {
1803                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1804                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1805                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1806         }
1807
1808         for (i = disks; i--; ) {
1809                 struct r5dev *dev = &sh->dev[i];
1810
1811                 if (dev->written || i == pd_idx || i == qd_idx) {
1812                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1813                                 set_bit(R5_UPTODATE, &dev->flags);
1814                                 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1815                                         set_bit(R5_Expanded, &dev->flags);
1816                         }
1817                         if (fua)
1818                                 set_bit(R5_WantFUA, &dev->flags);
1819                         if (sync)
1820                                 set_bit(R5_SyncIO, &dev->flags);
1821                 }
1822         }
1823
1824         if (sh->reconstruct_state == reconstruct_state_drain_run)
1825                 sh->reconstruct_state = reconstruct_state_drain_result;
1826         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1827                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1828         else {
1829                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1830                 sh->reconstruct_state = reconstruct_state_result;
1831         }
1832
1833         set_bit(STRIPE_HANDLE, &sh->state);
1834         raid5_release_stripe(sh);
1835 }
1836
1837 static void
1838 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1839                      struct dma_async_tx_descriptor *tx)
1840 {
1841         int disks = sh->disks;
1842         struct page **xor_srcs;
1843         struct async_submit_ctl submit;
1844         int count, pd_idx = sh->pd_idx, i;
1845         struct page *xor_dest;
1846         int prexor = 0;
1847         unsigned long flags;
1848         int j = 0;
1849         struct stripe_head *head_sh = sh;
1850         int last_stripe;
1851
1852         pr_debug("%s: stripe %llu\n", __func__,
1853                 (unsigned long long)sh->sector);
1854
1855         for (i = 0; i < sh->disks; i++) {
1856                 if (pd_idx == i)
1857                         continue;
1858                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1859                         break;
1860         }
1861         if (i >= sh->disks) {
1862                 atomic_inc(&sh->count);
1863                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1864                 ops_complete_reconstruct(sh);
1865                 return;
1866         }
1867 again:
1868         count = 0;
1869         xor_srcs = to_addr_page(percpu, j);
1870         /* check if prexor is active which means only process blocks
1871          * that are part of a read-modify-write (written)
1872          */
1873         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1874                 prexor = 1;
1875                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1876                 for (i = disks; i--; ) {
1877                         struct r5dev *dev = &sh->dev[i];
1878                         if (head_sh->dev[i].written ||
1879                             test_bit(R5_InJournal, &head_sh->dev[i].flags))
1880                                 xor_srcs[count++] = dev->page;
1881                 }
1882         } else {
1883                 xor_dest = sh->dev[pd_idx].page;
1884                 for (i = disks; i--; ) {
1885                         struct r5dev *dev = &sh->dev[i];
1886                         if (i != pd_idx)
1887                                 xor_srcs[count++] = dev->page;
1888                 }
1889         }
1890
1891         /* 1/ if we prexor'd then the dest is reused as a source
1892          * 2/ if we did not prexor then we are redoing the parity
1893          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1894          * for the synchronous xor case
1895          */
1896         last_stripe = !head_sh->batch_head ||
1897                 list_first_entry(&sh->batch_list,
1898                                  struct stripe_head, batch_list) == head_sh;
1899         if (last_stripe) {
1900                 flags = ASYNC_TX_ACK |
1901                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1902
1903                 atomic_inc(&head_sh->count);
1904                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1905                                   to_addr_conv(sh, percpu, j));
1906         } else {
1907                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1908                 init_async_submit(&submit, flags, tx, NULL, NULL,
1909                                   to_addr_conv(sh, percpu, j));
1910         }
1911
1912         if (unlikely(count == 1))
1913                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1914         else
1915                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1916         if (!last_stripe) {
1917                 j++;
1918                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1919                                       batch_list);
1920                 goto again;
1921         }
1922 }
1923
1924 static void
1925 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1926                      struct dma_async_tx_descriptor *tx)
1927 {
1928         struct async_submit_ctl submit;
1929         struct page **blocks;
1930         int count, i, j = 0;
1931         struct stripe_head *head_sh = sh;
1932         int last_stripe;
1933         int synflags;
1934         unsigned long txflags;
1935
1936         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1937
1938         for (i = 0; i < sh->disks; i++) {
1939                 if (sh->pd_idx == i || sh->qd_idx == i)
1940                         continue;
1941                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1942                         break;
1943         }
1944         if (i >= sh->disks) {
1945                 atomic_inc(&sh->count);
1946                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1947                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1948                 ops_complete_reconstruct(sh);
1949                 return;
1950         }
1951
1952 again:
1953         blocks = to_addr_page(percpu, j);
1954
1955         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1956                 synflags = SYNDROME_SRC_WRITTEN;
1957                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1958         } else {
1959                 synflags = SYNDROME_SRC_ALL;
1960                 txflags = ASYNC_TX_ACK;
1961         }
1962
1963         count = set_syndrome_sources(blocks, sh, synflags);
1964         last_stripe = !head_sh->batch_head ||
1965                 list_first_entry(&sh->batch_list,
1966                                  struct stripe_head, batch_list) == head_sh;
1967
1968         if (last_stripe) {
1969                 atomic_inc(&head_sh->count);
1970                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1971                                   head_sh, to_addr_conv(sh, percpu, j));
1972         } else
1973                 init_async_submit(&submit, 0, tx, NULL, NULL,
1974                                   to_addr_conv(sh, percpu, j));
1975         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1976         if (!last_stripe) {
1977                 j++;
1978                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1979                                       batch_list);
1980                 goto again;
1981         }
1982 }
1983
1984 static void ops_complete_check(void *stripe_head_ref)
1985 {
1986         struct stripe_head *sh = stripe_head_ref;
1987
1988         pr_debug("%s: stripe %llu\n", __func__,
1989                 (unsigned long long)sh->sector);
1990
1991         sh->check_state = check_state_check_result;
1992         set_bit(STRIPE_HANDLE, &sh->state);
1993         raid5_release_stripe(sh);
1994 }
1995
1996 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1997 {
1998         int disks = sh->disks;
1999         int pd_idx = sh->pd_idx;
2000         int qd_idx = sh->qd_idx;
2001         struct page *xor_dest;
2002         struct page **xor_srcs = to_addr_page(percpu, 0);
2003         struct dma_async_tx_descriptor *tx;
2004         struct async_submit_ctl submit;
2005         int count;
2006         int i;
2007
2008         pr_debug("%s: stripe %llu\n", __func__,
2009                 (unsigned long long)sh->sector);
2010
2011         BUG_ON(sh->batch_head);
2012         count = 0;
2013         xor_dest = sh->dev[pd_idx].page;
2014         xor_srcs[count++] = xor_dest;
2015         for (i = disks; i--; ) {
2016                 if (i == pd_idx || i == qd_idx)
2017                         continue;
2018                 xor_srcs[count++] = sh->dev[i].page;
2019         }
2020
2021         init_async_submit(&submit, 0, NULL, NULL, NULL,
2022                           to_addr_conv(sh, percpu, 0));
2023         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2024                            &sh->ops.zero_sum_result, &submit);
2025
2026         atomic_inc(&sh->count);
2027         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2028         tx = async_trigger_callback(&submit);
2029 }
2030
2031 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2032 {
2033         struct page **srcs = to_addr_page(percpu, 0);
2034         struct async_submit_ctl submit;
2035         int count;
2036
2037         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2038                 (unsigned long long)sh->sector, checkp);
2039
2040         BUG_ON(sh->batch_head);
2041         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2042         if (!checkp)
2043                 srcs[count] = NULL;
2044
2045         atomic_inc(&sh->count);
2046         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2047                           sh, to_addr_conv(sh, percpu, 0));
2048         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2049                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2050 }
2051
2052 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2053 {
2054         int overlap_clear = 0, i, disks = sh->disks;
2055         struct dma_async_tx_descriptor *tx = NULL;
2056         struct r5conf *conf = sh->raid_conf;
2057         int level = conf->level;
2058         struct raid5_percpu *percpu;
2059         unsigned long cpu;
2060
2061         cpu = get_cpu();
2062         percpu = per_cpu_ptr(conf->percpu, cpu);
2063         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2064                 ops_run_biofill(sh);
2065                 overlap_clear++;
2066         }
2067
2068         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2069                 if (level < 6)
2070                         tx = ops_run_compute5(sh, percpu);
2071                 else {
2072                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
2073                                 tx = ops_run_compute6_1(sh, percpu);
2074                         else
2075                                 tx = ops_run_compute6_2(sh, percpu);
2076                 }
2077                 /* terminate the chain if reconstruct is not set to be run */
2078                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2079                         async_tx_ack(tx);
2080         }
2081
2082         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2083                 if (level < 6)
2084                         tx = ops_run_prexor5(sh, percpu, tx);
2085                 else
2086                         tx = ops_run_prexor6(sh, percpu, tx);
2087         }
2088
2089         if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2090                 tx = ops_run_partial_parity(sh, percpu, tx);
2091
2092         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2093                 tx = ops_run_biodrain(sh, tx);
2094                 overlap_clear++;
2095         }
2096
2097         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2098                 if (level < 6)
2099                         ops_run_reconstruct5(sh, percpu, tx);
2100                 else
2101                         ops_run_reconstruct6(sh, percpu, tx);
2102         }
2103
2104         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2105                 if (sh->check_state == check_state_run)
2106                         ops_run_check_p(sh, percpu);
2107                 else if (sh->check_state == check_state_run_q)
2108                         ops_run_check_pq(sh, percpu, 0);
2109                 else if (sh->check_state == check_state_run_pq)
2110                         ops_run_check_pq(sh, percpu, 1);
2111                 else
2112                         BUG();
2113         }
2114
2115         if (overlap_clear && !sh->batch_head)
2116                 for (i = disks; i--; ) {
2117                         struct r5dev *dev = &sh->dev[i];
2118                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2119                                 wake_up(&sh->raid_conf->wait_for_overlap);
2120                 }
2121         put_cpu();
2122 }
2123
2124 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2125 {
2126         if (sh->ppl_page)
2127                 __free_page(sh->ppl_page);
2128         kmem_cache_free(sc, sh);
2129 }
2130
2131 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2132         int disks, struct r5conf *conf)
2133 {
2134         struct stripe_head *sh;
2135         int i;
2136
2137         sh = kmem_cache_zalloc(sc, gfp);
2138         if (sh) {
2139                 spin_lock_init(&sh->stripe_lock);
2140                 spin_lock_init(&sh->batch_lock);
2141                 INIT_LIST_HEAD(&sh->batch_list);
2142                 INIT_LIST_HEAD(&sh->lru);
2143                 INIT_LIST_HEAD(&sh->r5c);
2144                 INIT_LIST_HEAD(&sh->log_list);
2145                 atomic_set(&sh->count, 1);
2146                 sh->raid_conf = conf;
2147                 sh->log_start = MaxSector;
2148                 for (i = 0; i < disks; i++) {
2149                         struct r5dev *dev = &sh->dev[i];
2150
2151                         bio_init(&dev->req, &dev->vec, 1);
2152                         bio_init(&dev->rreq, &dev->rvec, 1);
2153                 }
2154
2155                 if (raid5_has_ppl(conf)) {
2156                         sh->ppl_page = alloc_page(gfp);
2157                         if (!sh->ppl_page) {
2158                                 free_stripe(sc, sh);
2159                                 sh = NULL;
2160                         }
2161                 }
2162         }
2163         return sh;
2164 }
2165 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2166 {
2167         struct stripe_head *sh;
2168
2169         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2170         if (!sh)
2171                 return 0;
2172
2173         if (grow_buffers(sh, gfp)) {
2174                 shrink_buffers(sh);
2175                 free_stripe(conf->slab_cache, sh);
2176                 return 0;
2177         }
2178         sh->hash_lock_index =
2179                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2180         /* we just created an active stripe so... */
2181         atomic_inc(&conf->active_stripes);
2182
2183         raid5_release_stripe(sh);
2184         conf->max_nr_stripes++;
2185         return 1;
2186 }
2187
2188 static int grow_stripes(struct r5conf *conf, int num)
2189 {
2190         struct kmem_cache *sc;
2191         size_t namelen = sizeof(conf->cache_name[0]);
2192         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2193
2194         if (conf->mddev->gendisk)
2195                 snprintf(conf->cache_name[0], namelen,
2196                         "raid%d-%s", conf->level, mdname(conf->mddev));
2197         else
2198                 snprintf(conf->cache_name[0], namelen,
2199                         "raid%d-%p", conf->level, conf->mddev);
2200         snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2201
2202         conf->active_name = 0;
2203         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2204                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2205                                0, 0, NULL);
2206         if (!sc)
2207                 return 1;
2208         conf->slab_cache = sc;
2209         conf->pool_size = devs;
2210         while (num--)
2211                 if (!grow_one_stripe(conf, GFP_KERNEL))
2212                         return 1;
2213
2214         return 0;
2215 }
2216
2217 /**
2218  * scribble_alloc - allocate percpu scribble buffer for required size
2219  *                  of the scribble region
2220  * @percpu - from for_each_present_cpu() of the caller
2221  * @num - total number of disks in the array
2222  * @cnt - scribble objs count for required size of the scribble region
2223  *
2224  * The scribble buffer size must be enough to contain:
2225  * 1/ a struct page pointer for each device in the array +2
2226  * 2/ room to convert each entry in (1) to its corresponding dma
2227  *    (dma_map_page()) or page (page_address()) address.
2228  *
2229  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2230  * calculate over all devices (not just the data blocks), using zeros in place
2231  * of the P and Q blocks.
2232  */
2233 static int scribble_alloc(struct raid5_percpu *percpu,
2234                           int num, int cnt)
2235 {
2236         size_t obj_size =
2237                 sizeof(struct page *) * (num+2) +
2238                 sizeof(addr_conv_t) * (num+2);
2239         void *scribble;
2240
2241         /*
2242          * If here is in raid array suspend context, it is in memalloc noio
2243          * context as well, there is no potential recursive memory reclaim
2244          * I/Os with the GFP_KERNEL flag.
2245          */
2246         scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2247         if (!scribble)
2248                 return -ENOMEM;
2249
2250         kvfree(percpu->scribble);
2251
2252         percpu->scribble = scribble;
2253         percpu->scribble_obj_size = obj_size;
2254         return 0;
2255 }
2256
2257 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2258 {
2259         unsigned long cpu;
2260         int err = 0;
2261
2262         /*
2263          * Never shrink. And mddev_suspend() could deadlock if this is called
2264          * from raid5d. In that case, scribble_disks and scribble_sectors
2265          * should equal to new_disks and new_sectors
2266          */
2267         if (conf->scribble_disks >= new_disks &&
2268             conf->scribble_sectors >= new_sectors)
2269                 return 0;
2270         mddev_suspend(conf->mddev);
2271         get_online_cpus();
2272
2273         for_each_present_cpu(cpu) {
2274                 struct raid5_percpu *percpu;
2275
2276                 percpu = per_cpu_ptr(conf->percpu, cpu);
2277                 err = scribble_alloc(percpu, new_disks,
2278                                      new_sectors / STRIPE_SECTORS);
2279                 if (err)
2280                         break;
2281         }
2282
2283         put_online_cpus();
2284         mddev_resume(conf->mddev);
2285         if (!err) {
2286                 conf->scribble_disks = new_disks;
2287                 conf->scribble_sectors = new_sectors;
2288         }
2289         return err;
2290 }
2291
2292 static int resize_stripes(struct r5conf *conf, int newsize)
2293 {
2294         /* Make all the stripes able to hold 'newsize' devices.
2295          * New slots in each stripe get 'page' set to a new page.
2296          *
2297          * This happens in stages:
2298          * 1/ create a new kmem_cache and allocate the required number of
2299          *    stripe_heads.
2300          * 2/ gather all the old stripe_heads and transfer the pages across
2301          *    to the new stripe_heads.  This will have the side effect of
2302          *    freezing the array as once all stripe_heads have been collected,
2303          *    no IO will be possible.  Old stripe heads are freed once their
2304          *    pages have been transferred over, and the old kmem_cache is
2305          *    freed when all stripes are done.
2306          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2307          *    we simple return a failure status - no need to clean anything up.
2308          * 4/ allocate new pages for the new slots in the new stripe_heads.
2309          *    If this fails, we don't bother trying the shrink the
2310          *    stripe_heads down again, we just leave them as they are.
2311          *    As each stripe_head is processed the new one is released into
2312          *    active service.
2313          *
2314          * Once step2 is started, we cannot afford to wait for a write,
2315          * so we use GFP_NOIO allocations.
2316          */
2317         struct stripe_head *osh, *nsh;
2318         LIST_HEAD(newstripes);
2319         struct disk_info *ndisks;
2320         int err = 0;
2321         struct kmem_cache *sc;
2322         int i;
2323         int hash, cnt;
2324
2325         md_allow_write(conf->mddev);
2326
2327         /* Step 1 */
2328         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2329                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2330                                0, 0, NULL);
2331         if (!sc)
2332                 return -ENOMEM;
2333
2334         /* Need to ensure auto-resizing doesn't interfere */
2335         mutex_lock(&conf->cache_size_mutex);
2336
2337         for (i = conf->max_nr_stripes; i; i--) {
2338                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2339                 if (!nsh)
2340                         break;
2341
2342                 list_add(&nsh->lru, &newstripes);
2343         }
2344         if (i) {
2345                 /* didn't get enough, give up */
2346                 while (!list_empty(&newstripes)) {
2347                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2348                         list_del(&nsh->lru);
2349                         free_stripe(sc, nsh);
2350                 }
2351                 kmem_cache_destroy(sc);
2352                 mutex_unlock(&conf->cache_size_mutex);
2353                 return -ENOMEM;
2354         }
2355         /* Step 2 - Must use GFP_NOIO now.
2356          * OK, we have enough stripes, start collecting inactive
2357          * stripes and copying them over
2358          */
2359         hash = 0;
2360         cnt = 0;
2361         list_for_each_entry(nsh, &newstripes, lru) {
2362                 lock_device_hash_lock(conf, hash);
2363                 wait_event_cmd(conf->wait_for_stripe,
2364                                     !list_empty(conf->inactive_list + hash),
2365                                     unlock_device_hash_lock(conf, hash),
2366                                     lock_device_hash_lock(conf, hash));
2367                 osh = get_free_stripe(conf, hash);
2368                 unlock_device_hash_lock(conf, hash);
2369
2370                 for(i=0; i<conf->pool_size; i++) {
2371                         nsh->dev[i].page = osh->dev[i].page;
2372                         nsh->dev[i].orig_page = osh->dev[i].page;
2373                 }
2374                 nsh->hash_lock_index = hash;
2375                 free_stripe(conf->slab_cache, osh);
2376                 cnt++;
2377                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2378                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2379                         hash++;
2380                         cnt = 0;
2381                 }
2382         }
2383         kmem_cache_destroy(conf->slab_cache);
2384
2385         /* Step 3.
2386          * At this point, we are holding all the stripes so the array
2387          * is completely stalled, so now is a good time to resize
2388          * conf->disks and the scribble region
2389          */
2390         ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2391         if (ndisks) {
2392                 for (i = 0; i < conf->pool_size; i++)
2393                         ndisks[i] = conf->disks[i];
2394
2395                 for (i = conf->pool_size; i < newsize; i++) {
2396                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2397                         if (!ndisks[i].extra_page)
2398                                 err = -ENOMEM;
2399                 }
2400
2401                 if (err) {
2402                         for (i = conf->pool_size; i < newsize; i++)
2403                                 if (ndisks[i].extra_page)
2404                                         put_page(ndisks[i].extra_page);
2405                         kfree(ndisks);
2406                 } else {
2407                         kfree(conf->disks);
2408                         conf->disks = ndisks;
2409                 }
2410         } else
2411                 err = -ENOMEM;
2412
2413         mutex_unlock(&conf->cache_size_mutex);
2414
2415         conf->slab_cache = sc;
2416         conf->active_name = 1-conf->active_name;
2417
2418         /* Step 4, return new stripes to service */
2419         while(!list_empty(&newstripes)) {
2420                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2421                 list_del_init(&nsh->lru);
2422
2423                 for (i=conf->raid_disks; i < newsize; i++)
2424                         if (nsh->dev[i].page == NULL) {
2425                                 struct page *p = alloc_page(GFP_NOIO);
2426                                 nsh->dev[i].page = p;
2427                                 nsh->dev[i].orig_page = p;
2428                                 if (!p)
2429                                         err = -ENOMEM;
2430                         }
2431                 raid5_release_stripe(nsh);
2432         }
2433         /* critical section pass, GFP_NOIO no longer needed */
2434
2435         if (!err)
2436                 conf->pool_size = newsize;
2437         return err;
2438 }
2439
2440 static int drop_one_stripe(struct r5conf *conf)
2441 {
2442         struct stripe_head *sh;
2443         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2444
2445         spin_lock_irq(conf->hash_locks + hash);
2446         sh = get_free_stripe(conf, hash);
2447         spin_unlock_irq(conf->hash_locks + hash);
2448         if (!sh)
2449                 return 0;
2450         BUG_ON(atomic_read(&sh->count));
2451         shrink_buffers(sh);
2452         free_stripe(conf->slab_cache, sh);
2453         atomic_dec(&conf->active_stripes);
2454         conf->max_nr_stripes--;
2455         return 1;
2456 }
2457
2458 static void shrink_stripes(struct r5conf *conf)
2459 {
2460         while (conf->max_nr_stripes &&
2461                drop_one_stripe(conf))
2462                 ;
2463
2464         kmem_cache_destroy(conf->slab_cache);
2465         conf->slab_cache = NULL;
2466 }
2467
2468 static void raid5_end_read_request(struct bio * bi)
2469 {
2470         struct stripe_head *sh = bi->bi_private;
2471         struct r5conf *conf = sh->raid_conf;
2472         int disks = sh->disks, i;
2473         char b[BDEVNAME_SIZE];
2474         struct md_rdev *rdev = NULL;
2475         sector_t s;
2476
2477         for (i=0 ; i<disks; i++)
2478                 if (bi == &sh->dev[i].req)
2479                         break;
2480
2481         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2482                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2483                 bi->bi_status);
2484         if (i == disks) {
2485                 bio_reset(bi);
2486                 BUG();
2487                 return;
2488         }
2489         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2490                 /* If replacement finished while this request was outstanding,
2491                  * 'replacement' might be NULL already.
2492                  * In that case it moved down to 'rdev'.
2493                  * rdev is not removed until all requests are finished.
2494                  */
2495                 rdev = conf->disks[i].replacement;
2496         if (!rdev)
2497                 rdev = conf->disks[i].rdev;
2498
2499         if (use_new_offset(conf, sh))
2500                 s = sh->sector + rdev->new_data_offset;
2501         else
2502                 s = sh->sector + rdev->data_offset;
2503         if (!bi->bi_status) {
2504                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2505                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2506                         /* Note that this cannot happen on a
2507                          * replacement device.  We just fail those on
2508                          * any error
2509                          */
2510                         pr_info_ratelimited(
2511                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2512                                 mdname(conf->mddev), STRIPE_SECTORS,
2513                                 (unsigned long long)s,
2514                                 bdevname(rdev->bdev, b));
2515                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2516                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2517                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2518                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2519                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2520
2521                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2522                         /*
2523                          * end read for a page in journal, this
2524                          * must be preparing for prexor in rmw
2525                          */
2526                         set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2527
2528                 if (atomic_read(&rdev->read_errors))
2529                         atomic_set(&rdev->read_errors, 0);
2530         } else {
2531                 const char *bdn = bdevname(rdev->bdev, b);
2532                 int retry = 0;
2533                 int set_bad = 0;
2534
2535                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2536                 if (!(bi->bi_status == BLK_STS_PROTECTION))
2537                         atomic_inc(&rdev->read_errors);
2538                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2539                         pr_warn_ratelimited(
2540                                 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2541                                 mdname(conf->mddev),
2542                                 (unsigned long long)s,
2543                                 bdn);
2544                 else if (conf->mddev->degraded >= conf->max_degraded) {
2545                         set_bad = 1;
2546                         pr_warn_ratelimited(
2547                                 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2548                                 mdname(conf->mddev),
2549                                 (unsigned long long)s,
2550                                 bdn);
2551                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2552                         /* Oh, no!!! */
2553                         set_bad = 1;
2554                         pr_warn_ratelimited(
2555                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2556                                 mdname(conf->mddev),
2557                                 (unsigned long long)s,
2558                                 bdn);
2559                 } else if (atomic_read(&rdev->read_errors)
2560                          > conf->max_nr_stripes) {
2561                         if (!test_bit(Faulty, &rdev->flags)) {
2562                                 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2563                                     mdname(conf->mddev),
2564                                     atomic_read(&rdev->read_errors),
2565                                     conf->max_nr_stripes);
2566                                 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2567                                     mdname(conf->mddev), bdn);
2568                         }
2569                 } else
2570                         retry = 1;
2571                 if (set_bad && test_bit(In_sync, &rdev->flags)
2572                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2573                         retry = 1;
2574                 if (retry)
2575                         if (sh->qd_idx >= 0 && sh->pd_idx == i)
2576                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2577                         else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2578                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2579                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2580                         } else
2581                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2582                 else {
2583                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2584                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2585                         if (!(set_bad
2586                               && test_bit(In_sync, &rdev->flags)
2587                               && rdev_set_badblocks(
2588                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2589                                 md_error(conf->mddev, rdev);
2590                 }
2591         }
2592         rdev_dec_pending(rdev, conf->mddev);
2593         bio_reset(bi);
2594         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2595         set_bit(STRIPE_HANDLE, &sh->state);
2596         raid5_release_stripe(sh);
2597 }
2598
2599 static void raid5_end_write_request(struct bio *bi)
2600 {
2601         struct stripe_head *sh = bi->bi_private;
2602         struct r5conf *conf = sh->raid_conf;
2603         int disks = sh->disks, i;
2604         struct md_rdev *rdev;
2605         sector_t first_bad;
2606         int bad_sectors;
2607         int replacement = 0;
2608
2609         for (i = 0 ; i < disks; i++) {
2610                 if (bi == &sh->dev[i].req) {
2611                         rdev = conf->disks[i].rdev;
2612                         break;
2613                 }
2614                 if (bi == &sh->dev[i].rreq) {
2615                         rdev = conf->disks[i].replacement;
2616                         if (rdev)
2617                                 replacement = 1;
2618                         else
2619                                 /* rdev was removed and 'replacement'
2620                                  * replaced it.  rdev is not removed
2621                                  * until all requests are finished.
2622                                  */
2623                                 rdev = conf->disks[i].rdev;
2624                         break;
2625                 }
2626         }
2627         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2628                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2629                 bi->bi_status);
2630         if (i == disks) {
2631                 bio_reset(bi);
2632                 BUG();
2633                 return;
2634         }
2635
2636         if (replacement) {
2637                 if (bi->bi_status)
2638                         md_error(conf->mddev, rdev);
2639                 else if (is_badblock(rdev, sh->sector,
2640                                      STRIPE_SECTORS,
2641                                      &first_bad, &bad_sectors))
2642                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2643         } else {
2644                 if (bi->bi_status) {
2645                         set_bit(STRIPE_DEGRADED, &sh->state);
2646                         set_bit(WriteErrorSeen, &rdev->flags);
2647                         set_bit(R5_WriteError, &sh->dev[i].flags);
2648                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2649                                 set_bit(MD_RECOVERY_NEEDED,
2650                                         &rdev->mddev->recovery);
2651                 } else if (is_badblock(rdev, sh->sector,
2652                                        STRIPE_SECTORS,
2653                                        &first_bad, &bad_sectors)) {
2654                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2655                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2656                                 /* That was a successful write so make
2657                                  * sure it looks like we already did
2658                                  * a re-write.
2659                                  */
2660                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2661                 }
2662         }
2663         rdev_dec_pending(rdev, conf->mddev);
2664
2665         if (sh->batch_head && bi->bi_status && !replacement)
2666                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2667
2668         bio_reset(bi);
2669         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2670                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2671         set_bit(STRIPE_HANDLE, &sh->state);
2672         raid5_release_stripe(sh);
2673
2674         if (sh->batch_head && sh != sh->batch_head)
2675                 raid5_release_stripe(sh->batch_head);
2676 }
2677
2678 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2679 {
2680         char b[BDEVNAME_SIZE];
2681         struct r5conf *conf = mddev->private;
2682         unsigned long flags;
2683         pr_debug("raid456: error called\n");
2684
2685         spin_lock_irqsave(&conf->device_lock, flags);
2686
2687         if (test_bit(In_sync, &rdev->flags) &&
2688             mddev->degraded == conf->max_degraded) {
2689                 /*
2690                  * Don't allow to achieve failed state
2691                  * Don't try to recover this device
2692                  */
2693                 conf->recovery_disabled = mddev->recovery_disabled;
2694                 spin_unlock_irqrestore(&conf->device_lock, flags);
2695                 return;
2696         }
2697
2698         set_bit(Faulty, &rdev->flags);
2699         clear_bit(In_sync, &rdev->flags);
2700         mddev->degraded = raid5_calc_degraded(conf);
2701         spin_unlock_irqrestore(&conf->device_lock, flags);
2702         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2703
2704         set_bit(Blocked, &rdev->flags);
2705         set_mask_bits(&mddev->sb_flags, 0,
2706                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2707         pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2708                 "md/raid:%s: Operation continuing on %d devices.\n",
2709                 mdname(mddev),
2710                 bdevname(rdev->bdev, b),
2711                 mdname(mddev),
2712                 conf->raid_disks - mddev->degraded);
2713         r5c_update_on_rdev_error(mddev, rdev);
2714 }
2715
2716 /*
2717  * Input: a 'big' sector number,
2718  * Output: index of the data and parity disk, and the sector # in them.
2719  */
2720 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2721                               int previous, int *dd_idx,
2722                               struct stripe_head *sh)
2723 {
2724         sector_t stripe, stripe2;
2725         sector_t chunk_number;
2726         unsigned int chunk_offset;
2727         int pd_idx, qd_idx;
2728         int ddf_layout = 0;
2729         sector_t new_sector;
2730         int algorithm = previous ? conf->prev_algo
2731                                  : conf->algorithm;
2732         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2733                                          : conf->chunk_sectors;
2734         int raid_disks = previous ? conf->previous_raid_disks
2735                                   : conf->raid_disks;
2736         int data_disks = raid_disks - conf->max_degraded;
2737
2738         /* First compute the information on this sector */
2739
2740         /*
2741          * Compute the chunk number and the sector offset inside the chunk
2742          */
2743         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2744         chunk_number = r_sector;
2745
2746         /*
2747          * Compute the stripe number
2748          */
2749         stripe = chunk_number;
2750         *dd_idx = sector_div(stripe, data_disks);
2751         stripe2 = stripe;
2752         /*
2753          * Select the parity disk based on the user selected algorithm.
2754          */
2755         pd_idx = qd_idx = -1;
2756         switch(conf->level) {
2757         case 4:
2758                 pd_idx = data_disks;
2759                 break;
2760         case 5:
2761                 switch (algorithm) {
2762                 case ALGORITHM_LEFT_ASYMMETRIC:
2763                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2764                         if (*dd_idx >= pd_idx)
2765                                 (*dd_idx)++;
2766                         break;
2767                 case ALGORITHM_RIGHT_ASYMMETRIC:
2768                         pd_idx = sector_div(stripe2, raid_disks);
2769                         if (*dd_idx >= pd_idx)
2770                                 (*dd_idx)++;
2771                         break;
2772                 case ALGORITHM_LEFT_SYMMETRIC:
2773                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2774                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2775                         break;
2776                 case ALGORITHM_RIGHT_SYMMETRIC:
2777                         pd_idx = sector_div(stripe2, raid_disks);
2778                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2779                         break;
2780                 case ALGORITHM_PARITY_0:
2781                         pd_idx = 0;
2782                         (*dd_idx)++;
2783                         break;
2784                 case ALGORITHM_PARITY_N:
2785                         pd_idx = data_disks;
2786                         break;
2787                 default:
2788                         BUG();
2789                 }
2790                 break;
2791         case 6:
2792
2793                 switch (algorithm) {
2794                 case ALGORITHM_LEFT_ASYMMETRIC:
2795                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2796                         qd_idx = pd_idx + 1;
2797                         if (pd_idx == raid_disks-1) {
2798                                 (*dd_idx)++;    /* Q D D D P */
2799                                 qd_idx = 0;
2800                         } else if (*dd_idx >= pd_idx)
2801                                 (*dd_idx) += 2; /* D D P Q D */
2802                         break;
2803                 case ALGORITHM_RIGHT_ASYMMETRIC:
2804                         pd_idx = sector_div(stripe2, raid_disks);
2805                         qd_idx = pd_idx + 1;
2806                         if (pd_idx == raid_disks-1) {
2807                                 (*dd_idx)++;    /* Q D D D P */
2808                                 qd_idx = 0;
2809                         } else if (*dd_idx >= pd_idx)
2810                                 (*dd_idx) += 2; /* D D P Q D */
2811                         break;
2812                 case ALGORITHM_LEFT_SYMMETRIC:
2813                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2814                         qd_idx = (pd_idx + 1) % raid_disks;
2815                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2816                         break;
2817                 case ALGORITHM_RIGHT_SYMMETRIC:
2818                         pd_idx = sector_div(stripe2, raid_disks);
2819                         qd_idx = (pd_idx + 1) % raid_disks;
2820                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2821                         break;
2822
2823                 case ALGORITHM_PARITY_0:
2824                         pd_idx = 0;
2825                         qd_idx = 1;
2826                         (*dd_idx) += 2;
2827                         break;
2828                 case ALGORITHM_PARITY_N:
2829                         pd_idx = data_disks;
2830                         qd_idx = data_disks + 1;
2831                         break;
2832
2833                 case ALGORITHM_ROTATING_ZERO_RESTART:
2834                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2835                          * of blocks for computing Q is different.
2836                          */
2837                         pd_idx = sector_div(stripe2, raid_disks);
2838                         qd_idx = pd_idx + 1;
2839                         if (pd_idx == raid_disks-1) {
2840                                 (*dd_idx)++;    /* Q D D D P */
2841                                 qd_idx = 0;
2842                         } else if (*dd_idx >= pd_idx)
2843                                 (*dd_idx) += 2; /* D D P Q D */
2844                         ddf_layout = 1;
2845                         break;
2846
2847                 case ALGORITHM_ROTATING_N_RESTART:
2848                         /* Same a left_asymmetric, by first stripe is
2849                          * D D D P Q  rather than
2850                          * Q D D D P
2851                          */
2852                         stripe2 += 1;
2853                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2854                         qd_idx = pd_idx + 1;
2855                         if (pd_idx == raid_disks-1) {
2856                                 (*dd_idx)++;    /* Q D D D P */
2857                                 qd_idx = 0;
2858                         } else if (*dd_idx >= pd_idx)
2859                                 (*dd_idx) += 2; /* D D P Q D */
2860                         ddf_layout = 1;
2861                         break;
2862
2863                 case ALGORITHM_ROTATING_N_CONTINUE:
2864                         /* Same as left_symmetric but Q is before P */
2865                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2866                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2867                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2868                         ddf_layout = 1;
2869                         break;
2870
2871                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2872                         /* RAID5 left_asymmetric, with Q on last device */
2873                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2874                         if (*dd_idx >= pd_idx)
2875                                 (*dd_idx)++;
2876                         qd_idx = raid_disks - 1;
2877                         break;
2878
2879                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2880                         pd_idx = sector_div(stripe2, raid_disks-1);
2881                         if (*dd_idx >= pd_idx)
2882                                 (*dd_idx)++;
2883                         qd_idx = raid_disks - 1;
2884                         break;
2885
2886                 case ALGORITHM_LEFT_SYMMETRIC_6:
2887                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2888                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2889                         qd_idx = raid_disks - 1;
2890                         break;
2891
2892                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2893                         pd_idx = sector_div(stripe2, raid_disks-1);
2894                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2895                         qd_idx = raid_disks - 1;
2896                         break;
2897
2898                 case ALGORITHM_PARITY_0_6:
2899                         pd_idx = 0;
2900                         (*dd_idx)++;
2901                         qd_idx = raid_disks - 1;
2902                         break;
2903
2904                 default:
2905                         BUG();
2906                 }
2907                 break;
2908         }
2909
2910         if (sh) {
2911                 sh->pd_idx = pd_idx;
2912                 sh->qd_idx = qd_idx;
2913                 sh->ddf_layout = ddf_layout;
2914         }
2915         /*
2916          * Finally, compute the new sector number
2917          */
2918         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2919         return new_sector;
2920 }
2921
2922 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2923 {
2924         struct r5conf *conf = sh->raid_conf;
2925         int raid_disks = sh->disks;
2926         int data_disks = raid_disks - conf->max_degraded;
2927         sector_t new_sector = sh->sector, check;
2928         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2929                                          : conf->chunk_sectors;
2930         int algorithm = previous ? conf->prev_algo
2931                                  : conf->algorithm;
2932         sector_t stripe;
2933         int chunk_offset;
2934         sector_t chunk_number;
2935         int dummy1, dd_idx = i;
2936         sector_t r_sector;
2937         struct stripe_head sh2;
2938
2939         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2940         stripe = new_sector;
2941
2942         if (i == sh->pd_idx)
2943                 return 0;
2944         switch(conf->level) {
2945         case 4: break;
2946         case 5:
2947                 switch (algorithm) {
2948                 case ALGORITHM_LEFT_ASYMMETRIC:
2949                 case ALGORITHM_RIGHT_ASYMMETRIC:
2950                         if (i > sh->pd_idx)
2951                                 i--;
2952                         break;
2953                 case ALGORITHM_LEFT_SYMMETRIC:
2954                 case ALGORITHM_RIGHT_SYMMETRIC:
2955                         if (i < sh->pd_idx)
2956                                 i += raid_disks;
2957                         i -= (sh->pd_idx + 1);
2958                         break;
2959                 case ALGORITHM_PARITY_0:
2960                         i -= 1;
2961                         break;
2962                 case ALGORITHM_PARITY_N:
2963                         break;
2964                 default:
2965                         BUG();
2966                 }
2967                 break;
2968         case 6:
2969                 if (i == sh->qd_idx)
2970                         return 0; /* It is the Q disk */
2971                 switch (algorithm) {
2972                 case ALGORITHM_LEFT_ASYMMETRIC:
2973                 case ALGORITHM_RIGHT_ASYMMETRIC:
2974                 case ALGORITHM_ROTATING_ZERO_RESTART:
2975                 case ALGORITHM_ROTATING_N_RESTART:
2976                         if (sh->pd_idx == raid_disks-1)
2977                                 i--;    /* Q D D D P */
2978                         else if (i > sh->pd_idx)
2979                                 i -= 2; /* D D P Q D */
2980                         break;
2981                 case ALGORITHM_LEFT_SYMMETRIC:
2982                 case ALGORITHM_RIGHT_SYMMETRIC:
2983                         if (sh->pd_idx == raid_disks-1)
2984                                 i--; /* Q D D D P */
2985                         else {
2986                                 /* D D P Q D */
2987                                 if (i < sh->pd_idx)
2988                                         i += raid_disks;
2989                                 i -= (sh->pd_idx + 2);
2990                         }
2991                         break;
2992                 case ALGORITHM_PARITY_0:
2993                         i -= 2;
2994                         break;
2995                 case ALGORITHM_PARITY_N:
2996                         break;
2997                 case ALGORITHM_ROTATING_N_CONTINUE:
2998                         /* Like left_symmetric, but P is before Q */
2999                         if (sh->pd_idx == 0)
3000                                 i--;    /* P D D D Q */
3001                         else {
3002                                 /* D D Q P D */
3003                                 if (i < sh->pd_idx)
3004                                         i += raid_disks;
3005                                 i -= (sh->pd_idx + 1);
3006                         }
3007                         break;
3008                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3009                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3010                         if (i > sh->pd_idx)
3011                                 i--;
3012                         break;
3013                 case ALGORITHM_LEFT_SYMMETRIC_6:
3014                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3015                         if (i < sh->pd_idx)
3016                                 i += data_disks + 1;
3017                         i -= (sh->pd_idx + 1);
3018                         break;
3019                 case ALGORITHM_PARITY_0_6:
3020                         i -= 1;
3021                         break;
3022                 default:
3023                         BUG();
3024                 }
3025                 break;
3026         }
3027
3028         chunk_number = stripe * data_disks + i;
3029         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3030
3031         check = raid5_compute_sector(conf, r_sector,
3032                                      previous, &dummy1, &sh2);
3033         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3034                 || sh2.qd_idx != sh->qd_idx) {
3035                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3036                         mdname(conf->mddev));
3037                 return 0;
3038         }
3039         return r_sector;
3040 }
3041
3042 /*
3043  * There are cases where we want handle_stripe_dirtying() and
3044  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3045  *
3046  * This function checks whether we want to delay the towrite. Specifically,
3047  * we delay the towrite when:
3048  *
3049  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3050  *      stripe has data in journal (for other devices).
3051  *
3052  *      In this case, when reading data for the non-overwrite dev, it is
3053  *      necessary to handle complex rmw of write back cache (prexor with
3054  *      orig_page, and xor with page). To keep read path simple, we would
3055  *      like to flush data in journal to RAID disks first, so complex rmw
3056  *      is handled in the write patch (handle_stripe_dirtying).
3057  *
3058  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3059  *
3060  *      It is important to be able to flush all stripes in raid5-cache.
3061  *      Therefore, we need reserve some space on the journal device for
3062  *      these flushes. If flush operation includes pending writes to the
3063  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3064  *      for the flush out. If we exclude these pending writes from flush
3065  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3066  *      Therefore, excluding pending writes in these cases enables more
3067  *      efficient use of the journal device.
3068  *
3069  *      Note: To make sure the stripe makes progress, we only delay
3070  *      towrite for stripes with data already in journal (injournal > 0).
3071  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3072  *      no_space_stripes list.
3073  *
3074  *   3. during journal failure
3075  *      In journal failure, we try to flush all cached data to raid disks
3076  *      based on data in stripe cache. The array is read-only to upper
3077  *      layers, so we would skip all pending writes.
3078  *
3079  */
3080 static inline bool delay_towrite(struct r5conf *conf,
3081                                  struct r5dev *dev,
3082                                  struct stripe_head_state *s)
3083 {
3084         /* case 1 above */
3085         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3086             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3087                 return true;
3088         /* case 2 above */
3089         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3090             s->injournal > 0)
3091                 return true;
3092         /* case 3 above */
3093         if (s->log_failed && s->injournal)
3094                 return true;
3095         return false;
3096 }
3097
3098 static void
3099 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3100                          int rcw, int expand)
3101 {
3102         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3103         struct r5conf *conf = sh->raid_conf;
3104         int level = conf->level;
3105
3106         if (rcw) {
3107                 /*
3108                  * In some cases, handle_stripe_dirtying initially decided to
3109                  * run rmw and allocates extra page for prexor. However, rcw is
3110                  * cheaper later on. We need to free the extra page now,
3111                  * because we won't be able to do that in ops_complete_prexor().
3112                  */
3113                 r5c_release_extra_page(sh);
3114
3115                 for (i = disks; i--; ) {
3116                         struct r5dev *dev = &sh->dev[i];
3117
3118                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3119                                 set_bit(R5_LOCKED, &dev->flags);
3120                                 set_bit(R5_Wantdrain, &dev->flags);
3121                                 if (!expand)
3122                                         clear_bit(R5_UPTODATE, &dev->flags);
3123                                 s->locked++;
3124                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3125                                 set_bit(R5_LOCKED, &dev->flags);
3126                                 s->locked++;
3127                         }
3128                 }
3129                 /* if we are not expanding this is a proper write request, and
3130                  * there will be bios with new data to be drained into the
3131                  * stripe cache
3132                  */
3133                 if (!expand) {
3134                         if (!s->locked)
3135                                 /* False alarm, nothing to do */
3136                                 return;
3137                         sh->reconstruct_state = reconstruct_state_drain_run;
3138                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3139                 } else
3140                         sh->reconstruct_state = reconstruct_state_run;
3141
3142                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3143
3144                 if (s->locked + conf->max_degraded == disks)
3145                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3146                                 atomic_inc(&conf->pending_full_writes);
3147         } else {
3148                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3149                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3150                 BUG_ON(level == 6 &&
3151                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3152                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3153
3154                 for (i = disks; i--; ) {
3155                         struct r5dev *dev = &sh->dev[i];
3156                         if (i == pd_idx || i == qd_idx)
3157                                 continue;
3158
3159                         if (dev->towrite &&
3160                             (test_bit(R5_UPTODATE, &dev->flags) ||
3161                              test_bit(R5_Wantcompute, &dev->flags))) {
3162                                 set_bit(R5_Wantdrain, &dev->flags);
3163                                 set_bit(R5_LOCKED, &dev->flags);
3164                                 clear_bit(R5_UPTODATE, &dev->flags);
3165                                 s->locked++;
3166                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3167                                 set_bit(R5_LOCKED, &dev->flags);
3168                                 s->locked++;
3169                         }
3170                 }
3171                 if (!s->locked)
3172                         /* False alarm - nothing to do */
3173                         return;
3174                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3175                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3176                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3177                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3178         }
3179
3180         /* keep the parity disk(s) locked while asynchronous operations
3181          * are in flight
3182          */
3183         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3184         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3185         s->locked++;
3186
3187         if (level == 6) {
3188                 int qd_idx = sh->qd_idx;
3189                 struct r5dev *dev = &sh->dev[qd_idx];
3190
3191                 set_bit(R5_LOCKED, &dev->flags);
3192                 clear_bit(R5_UPTODATE, &dev->flags);
3193                 s->locked++;
3194         }
3195
3196         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3197             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3198             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3199             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3200                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3201
3202         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3203                 __func__, (unsigned long long)sh->sector,
3204                 s->locked, s->ops_request);
3205 }
3206
3207 /*
3208  * Each stripe/dev can have one or more bion attached.
3209  * toread/towrite point to the first in a chain.
3210  * The bi_next chain must be in order.
3211  */
3212 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3213                           int forwrite, int previous)
3214 {
3215         struct bio **bip;
3216         struct r5conf *conf = sh->raid_conf;
3217         int firstwrite=0;
3218
3219         pr_debug("adding bi b#%llu to stripe s#%llu\n",
3220                 (unsigned long long)bi->bi_iter.bi_sector,
3221                 (unsigned long long)sh->sector);
3222
3223         spin_lock_irq(&sh->stripe_lock);
3224         sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3225         /* Don't allow new IO added to stripes in batch list */
3226         if (sh->batch_head)
3227                 goto overlap;
3228         if (forwrite) {
3229                 bip = &sh->dev[dd_idx].towrite;
3230                 if (*bip == NULL)
3231                         firstwrite = 1;
3232         } else
3233                 bip = &sh->dev[dd_idx].toread;
3234         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3235                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3236                         goto overlap;
3237                 bip = & (*bip)->bi_next;
3238         }
3239         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3240                 goto overlap;
3241
3242         if (forwrite && raid5_has_ppl(conf)) {
3243                 /*
3244                  * With PPL only writes to consecutive data chunks within a
3245                  * stripe are allowed because for a single stripe_head we can
3246                  * only have one PPL entry at a time, which describes one data
3247                  * range. Not really an overlap, but wait_for_overlap can be
3248                  * used to handle this.
3249                  */
3250                 sector_t sector;
3251                 sector_t first = 0;
3252                 sector_t last = 0;
3253                 int count = 0;
3254                 int i;
3255
3256                 for (i = 0; i < sh->disks; i++) {
3257                         if (i != sh->pd_idx &&
3258                             (i == dd_idx || sh->dev[i].towrite)) {
3259                                 sector = sh->dev[i].sector;
3260                                 if (count == 0 || sector < first)
3261                                         first = sector;
3262                                 if (sector > last)
3263                                         last = sector;
3264                                 count++;
3265                         }
3266                 }
3267
3268                 if (first + conf->chunk_sectors * (count - 1) != last)
3269                         goto overlap;
3270         }
3271
3272         if (!forwrite || previous)
3273                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3274
3275         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3276         if (*bip)
3277                 bi->bi_next = *bip;
3278         *bip = bi;
3279         bio_inc_remaining(bi);
3280         md_write_inc(conf->mddev, bi);
3281
3282         if (forwrite) {
3283                 /* check if page is covered */
3284                 sector_t sector = sh->dev[dd_idx].sector;
3285                 for (bi=sh->dev[dd_idx].towrite;
3286                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3287                              bi && bi->bi_iter.bi_sector <= sector;
3288                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3289                         if (bio_end_sector(bi) >= sector)
3290                                 sector = bio_end_sector(bi);
3291                 }
3292                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3293                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3294                                 sh->overwrite_disks++;
3295         }
3296
3297         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3298                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3299                 (unsigned long long)sh->sector, dd_idx);
3300
3301         if (conf->mddev->bitmap && firstwrite) {
3302                 /* Cannot hold spinlock over bitmap_startwrite,
3303                  * but must ensure this isn't added to a batch until
3304                  * we have added to the bitmap and set bm_seq.
3305                  * So set STRIPE_BITMAP_PENDING to prevent
3306                  * batching.
3307                  * If multiple add_stripe_bio() calls race here they
3308                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3309                  * to complete "bitmap_startwrite" gets to set
3310                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3311                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3312                  * any more.
3313                  */
3314                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3315                 spin_unlock_irq(&sh->stripe_lock);
3316                 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3317                                      STRIPE_SECTORS, 0);
3318                 spin_lock_irq(&sh->stripe_lock);
3319                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3320                 if (!sh->batch_head) {
3321                         sh->bm_seq = conf->seq_flush+1;
3322                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3323                 }
3324         }
3325         spin_unlock_irq(&sh->stripe_lock);
3326
3327         if (stripe_can_batch(sh))
3328                 stripe_add_to_batch_list(conf, sh);
3329         return 1;
3330
3331  overlap:
3332         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3333         spin_unlock_irq(&sh->stripe_lock);
3334         return 0;
3335 }
3336
3337 static void end_reshape(struct r5conf *conf);
3338
3339 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3340                             struct stripe_head *sh)
3341 {
3342         int sectors_per_chunk =
3343                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3344         int dd_idx;
3345         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3346         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3347
3348         raid5_compute_sector(conf,
3349                              stripe * (disks - conf->max_degraded)
3350                              *sectors_per_chunk + chunk_offset,
3351                              previous,
3352                              &dd_idx, sh);
3353 }
3354
3355 static void
3356 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3357                      struct stripe_head_state *s, int disks)
3358 {
3359         int i;
3360         BUG_ON(sh->batch_head);
3361         for (i = disks; i--; ) {
3362                 struct bio *bi;
3363                 int bitmap_end = 0;
3364
3365                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3366                         struct md_rdev *rdev;
3367                         rcu_read_lock();
3368                         rdev = rcu_dereference(conf->disks[i].rdev);
3369                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3370                             !test_bit(Faulty, &rdev->flags))
3371                                 atomic_inc(&rdev->nr_pending);
3372                         else
3373                                 rdev = NULL;
3374                         rcu_read_unlock();
3375                         if (rdev) {
3376                                 if (!rdev_set_badblocks(
3377                                             rdev,
3378                                             sh->sector,
3379                                             STRIPE_SECTORS, 0))
3380                                         md_error(conf->mddev, rdev);
3381                                 rdev_dec_pending(rdev, conf->mddev);
3382                         }
3383                 }
3384                 spin_lock_irq(&sh->stripe_lock);
3385                 /* fail all writes first */
3386                 bi = sh->dev[i].towrite;
3387                 sh->dev[i].towrite = NULL;
3388                 sh->overwrite_disks = 0;
3389                 spin_unlock_irq(&sh->stripe_lock);
3390                 if (bi)
3391                         bitmap_end = 1;
3392
3393                 log_stripe_write_finished(sh);
3394
3395                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3396                         wake_up(&conf->wait_for_overlap);
3397
3398                 while (bi && bi->bi_iter.bi_sector <
3399                         sh->dev[i].sector + STRIPE_SECTORS) {
3400                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3401
3402                         md_write_end(conf->mddev);
3403                         bio_io_error(bi);
3404                         bi = nextbi;
3405                 }
3406                 if (bitmap_end)
3407                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3408                                            STRIPE_SECTORS, 0, 0);
3409                 bitmap_end = 0;
3410                 /* and fail all 'written' */
3411                 bi = sh->dev[i].written;
3412                 sh->dev[i].written = NULL;
3413                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3414                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3415                         sh->dev[i].page = sh->dev[i].orig_page;
3416                 }
3417
3418                 if (bi) bitmap_end = 1;
3419                 while (bi && bi->bi_iter.bi_sector <
3420                        sh->dev[i].sector + STRIPE_SECTORS) {
3421                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3422
3423                         md_write_end(conf->mddev);
3424                         bio_io_error(bi);
3425                         bi = bi2;
3426                 }
3427
3428                 /* fail any reads if this device is non-operational and
3429                  * the data has not reached the cache yet.
3430                  */
3431                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3432                     s->failed > conf->max_degraded &&
3433                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3434                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3435                         spin_lock_irq(&sh->stripe_lock);
3436                         bi = sh->dev[i].toread;
3437                         sh->dev[i].toread = NULL;
3438                         spin_unlock_irq(&sh->stripe_lock);
3439                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3440                                 wake_up(&conf->wait_for_overlap);
3441                         if (bi)
3442                                 s->to_read--;
3443                         while (bi && bi->bi_iter.bi_sector <
3444                                sh->dev[i].sector + STRIPE_SECTORS) {
3445                                 struct bio *nextbi =
3446                                         r5_next_bio(bi, sh->dev[i].sector);
3447
3448                                 bio_io_error(bi);
3449                                 bi = nextbi;
3450                         }
3451                 }
3452                 if (bitmap_end)
3453                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3454                                            STRIPE_SECTORS, 0, 0);
3455                 /* If we were in the middle of a write the parity block might
3456                  * still be locked - so just clear all R5_LOCKED flags
3457                  */
3458                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3459         }
3460         s->to_write = 0;
3461         s->written = 0;
3462
3463         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3464                 if (atomic_dec_and_test(&conf->pending_full_writes))
3465                         md_wakeup_thread(conf->mddev->thread);
3466 }
3467
3468 static void
3469 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3470                    struct stripe_head_state *s)
3471 {
3472         int abort = 0;
3473         int i;
3474
3475         BUG_ON(sh->batch_head);
3476         clear_bit(STRIPE_SYNCING, &sh->state);
3477         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3478                 wake_up(&conf->wait_for_overlap);
3479         s->syncing = 0;
3480         s->replacing = 0;
3481         /* There is nothing more to do for sync/check/repair.
3482          * Don't even need to abort as that is handled elsewhere
3483          * if needed, and not always wanted e.g. if there is a known
3484          * bad block here.
3485          * For recover/replace we need to record a bad block on all
3486          * non-sync devices, or abort the recovery
3487          */
3488         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3489                 /* During recovery devices cannot be removed, so
3490                  * locking and refcounting of rdevs is not needed
3491                  */
3492                 rcu_read_lock();
3493                 for (i = 0; i < conf->raid_disks; i++) {
3494                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3495                         if (rdev
3496                             && !test_bit(Faulty, &rdev->flags)
3497                             && !test_bit(In_sync, &rdev->flags)
3498                             && !rdev_set_badblocks(rdev, sh->sector,
3499                                                    STRIPE_SECTORS, 0))
3500                                 abort = 1;
3501                         rdev = rcu_dereference(conf->disks[i].replacement);
3502                         if (rdev
3503                             && !test_bit(Faulty, &rdev->flags)
3504                             && !test_bit(In_sync, &rdev->flags)
3505                             && !rdev_set_badblocks(rdev, sh->sector,
3506                                                    STRIPE_SECTORS, 0))
3507                                 abort = 1;
3508                 }
3509                 rcu_read_unlock();
3510                 if (abort)
3511                         conf->recovery_disabled =
3512                                 conf->mddev->recovery_disabled;
3513         }
3514         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3515 }
3516
3517 static int want_replace(struct stripe_head *sh, int disk_idx)
3518 {
3519         struct md_rdev *rdev;
3520         int rv = 0;
3521
3522         rcu_read_lock();
3523         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3524         if (rdev
3525             && !test_bit(Faulty, &rdev->flags)
3526             && !test_bit(In_sync, &rdev->flags)
3527             && (rdev->recovery_offset <= sh->sector
3528                 || rdev->mddev->recovery_cp <= sh->sector))
3529                 rv = 1;
3530         rcu_read_unlock();
3531         return rv;
3532 }
3533
3534 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3535                            int disk_idx, int disks)
3536 {
3537         struct r5dev *dev = &sh->dev[disk_idx];
3538         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3539                                   &sh->dev[s->failed_num[1]] };
3540         int i;
3541
3542
3543         if (test_bit(R5_LOCKED, &dev->flags) ||
3544             test_bit(R5_UPTODATE, &dev->flags))
3545                 /* No point reading this as we already have it or have
3546                  * decided to get it.
3547                  */
3548                 return 0;
3549
3550         if (dev->toread ||
3551             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3552                 /* We need this block to directly satisfy a request */
3553                 return 1;
3554
3555         if (s->syncing || s->expanding ||
3556             (s->replacing && want_replace(sh, disk_idx)))
3557                 /* When syncing, or expanding we read everything.
3558                  * When replacing, we need the replaced block.
3559                  */
3560                 return 1;
3561
3562         if ((s->failed >= 1 && fdev[0]->toread) ||
3563             (s->failed >= 2 && fdev[1]->toread))
3564                 /* If we want to read from a failed device, then
3565                  * we need to actually read every other device.
3566                  */
3567                 return 1;
3568
3569         /* Sometimes neither read-modify-write nor reconstruct-write
3570          * cycles can work.  In those cases we read every block we
3571          * can.  Then the parity-update is certain to have enough to
3572          * work with.
3573          * This can only be a problem when we need to write something,
3574          * and some device has failed.  If either of those tests
3575          * fail we need look no further.
3576          */
3577         if (!s->failed || !s->to_write)
3578                 return 0;
3579
3580         if (test_bit(R5_Insync, &dev->flags) &&
3581             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3582                 /* Pre-reads at not permitted until after short delay
3583                  * to gather multiple requests.  However if this
3584                  * device is no Insync, the block could only be computed
3585                  * and there is no need to delay that.
3586                  */
3587                 return 0;
3588
3589         for (i = 0; i < s->failed && i < 2; i++) {
3590                 if (fdev[i]->towrite &&
3591                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3592                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3593                         /* If we have a partial write to a failed
3594                          * device, then we will need to reconstruct
3595                          * the content of that device, so all other
3596                          * devices must be read.
3597                          */
3598                         return 1;
3599         }
3600
3601         /* If we are forced to do a reconstruct-write, either because
3602          * the current RAID6 implementation only supports that, or
3603          * because parity cannot be trusted and we are currently
3604          * recovering it, there is extra need to be careful.
3605          * If one of the devices that we would need to read, because
3606          * it is not being overwritten (and maybe not written at all)
3607          * is missing/faulty, then we need to read everything we can.
3608          */
3609         if (sh->raid_conf->level != 6 &&
3610             sh->sector < sh->raid_conf->mddev->recovery_cp)
3611                 /* reconstruct-write isn't being forced */
3612                 return 0;
3613         for (i = 0; i < s->failed && i < 2; i++) {
3614                 if (s->failed_num[i] != sh->pd_idx &&
3615                     s->failed_num[i] != sh->qd_idx &&
3616                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3617                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3618                         return 1;
3619         }
3620
3621         return 0;
3622 }
3623
3624 /* fetch_block - checks the given member device to see if its data needs
3625  * to be read or computed to satisfy a request.
3626  *
3627  * Returns 1 when no more member devices need to be checked, otherwise returns
3628  * 0 to tell the loop in handle_stripe_fill to continue
3629  */
3630 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3631                        int disk_idx, int disks)
3632 {
3633         struct r5dev *dev = &sh->dev[disk_idx];
3634
3635         /* is the data in this block needed, and can we get it? */
3636         if (need_this_block(sh, s, disk_idx, disks)) {
3637                 /* we would like to get this block, possibly by computing it,
3638                  * otherwise read it if the backing disk is insync
3639                  */
3640                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3641                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3642                 BUG_ON(sh->batch_head);
3643
3644                 /*
3645                  * In the raid6 case if the only non-uptodate disk is P
3646                  * then we already trusted P to compute the other failed
3647                  * drives. It is safe to compute rather than re-read P.
3648                  * In other cases we only compute blocks from failed
3649                  * devices, otherwise check/repair might fail to detect
3650                  * a real inconsistency.
3651                  */
3652
3653                 if ((s->uptodate == disks - 1) &&
3654                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3655                     (s->failed && (disk_idx == s->failed_num[0] ||
3656                                    disk_idx == s->failed_num[1])))) {
3657                         /* have disk failed, and we're requested to fetch it;
3658                          * do compute it
3659                          */
3660                         pr_debug("Computing stripe %llu block %d\n",
3661                                (unsigned long long)sh->sector, disk_idx);
3662                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3663                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3664                         set_bit(R5_Wantcompute, &dev->flags);
3665                         sh->ops.target = disk_idx;
3666                         sh->ops.target2 = -1; /* no 2nd target */
3667                         s->req_compute = 1;
3668                         /* Careful: from this point on 'uptodate' is in the eye
3669                          * of raid_run_ops which services 'compute' operations
3670                          * before writes. R5_Wantcompute flags a block that will
3671                          * be R5_UPTODATE by the time it is needed for a
3672                          * subsequent operation.
3673                          */
3674                         s->uptodate++;
3675                         return 1;
3676                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3677                         /* Computing 2-failure is *very* expensive; only
3678                          * do it if failed >= 2
3679                          */
3680                         int other;
3681                         for (other = disks; other--; ) {
3682                                 if (other == disk_idx)
3683                                         continue;
3684                                 if (!test_bit(R5_UPTODATE,
3685                                       &sh->dev[other].flags))
3686                                         break;
3687                         }
3688                         BUG_ON(other < 0);
3689                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3690                                (unsigned long long)sh->sector,
3691                                disk_idx, other);
3692                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3693                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3694                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3695                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3696                         sh->ops.target = disk_idx;
3697                         sh->ops.target2 = other;
3698                         s->uptodate += 2;
3699                         s->req_compute = 1;
3700                         return 1;
3701                 } else if (test_bit(R5_Insync, &dev->flags)) {
3702                         set_bit(R5_LOCKED, &dev->flags);
3703                         set_bit(R5_Wantread, &dev->flags);
3704                         s->locked++;
3705                         pr_debug("Reading block %d (sync=%d)\n",
3706                                 disk_idx, s->syncing);
3707                 }
3708         }
3709
3710         return 0;
3711 }
3712
3713 /**
3714  * handle_stripe_fill - read or compute data to satisfy pending requests.
3715  */
3716 static void handle_stripe_fill(struct stripe_head *sh,
3717                                struct stripe_head_state *s,
3718                                int disks)
3719 {
3720         int i;
3721
3722         /* look for blocks to read/compute, skip this if a compute
3723          * is already in flight, or if the stripe contents are in the
3724          * midst of changing due to a write
3725          */
3726         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3727             !sh->reconstruct_state) {
3728
3729                 /*
3730                  * For degraded stripe with data in journal, do not handle
3731                  * read requests yet, instead, flush the stripe to raid
3732                  * disks first, this avoids handling complex rmw of write
3733                  * back cache (prexor with orig_page, and then xor with
3734                  * page) in the read path
3735                  */
3736                 if (s->injournal && s->failed) {
3737                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3738                                 r5c_make_stripe_write_out(sh);
3739                         goto out;
3740                 }
3741
3742                 for (i = disks; i--; )
3743                         if (fetch_block(sh, s, i, disks))
3744                                 break;
3745         }
3746 out:
3747         set_bit(STRIPE_HANDLE, &sh->state);
3748 }
3749
3750 static void break_stripe_batch_list(struct stripe_head *head_sh,
3751                                     unsigned long handle_flags);
3752 /* handle_stripe_clean_event
3753  * any written block on an uptodate or failed drive can be returned.
3754  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3755  * never LOCKED, so we don't need to test 'failed' directly.
3756  */
3757 static void handle_stripe_clean_event(struct r5conf *conf,
3758         struct stripe_head *sh, int disks)
3759 {
3760         int i;
3761         struct r5dev *dev;
3762         int discard_pending = 0;
3763         struct stripe_head *head_sh = sh;
3764         bool do_endio = false;
3765
3766         for (i = disks; i--; )
3767                 if (sh->dev[i].written) {
3768                         dev = &sh->dev[i];
3769                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3770                             (test_bit(R5_UPTODATE, &dev->flags) ||
3771                              test_bit(R5_Discard, &dev->flags) ||
3772                              test_bit(R5_SkipCopy, &dev->flags))) {
3773                                 /* We can return any write requests */
3774                                 struct bio *wbi, *wbi2;
3775                                 pr_debug("Return write for disc %d\n", i);
3776                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3777                                         clear_bit(R5_UPTODATE, &dev->flags);
3778                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3779                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3780                                 }
3781                                 do_endio = true;
3782
3783 returnbi:
3784                                 dev->page = dev->orig_page;
3785                                 wbi = dev->written;
3786                                 dev->written = NULL;
3787                                 while (wbi && wbi->bi_iter.bi_sector <
3788                                         dev->sector + STRIPE_SECTORS) {
3789                                         wbi2 = r5_next_bio(wbi, dev->sector);
3790                                         md_write_end(conf->mddev);
3791                                         bio_endio(wbi);
3792                                         wbi = wbi2;
3793                                 }
3794                                 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3795                                                    STRIPE_SECTORS,
3796                                                    !test_bit(STRIPE_DEGRADED, &sh->state),
3797                                                    0);
3798                                 if (head_sh->batch_head) {
3799                                         sh = list_first_entry(&sh->batch_list,
3800                                                               struct stripe_head,
3801                                                               batch_list);
3802                                         if (sh != head_sh) {
3803                                                 dev = &sh->dev[i];
3804                                                 goto returnbi;
3805                                         }
3806                                 }
3807                                 sh = head_sh;
3808                                 dev = &sh->dev[i];
3809                         } else if (test_bit(R5_Discard, &dev->flags))
3810                                 discard_pending = 1;
3811                 }
3812
3813         log_stripe_write_finished(sh);
3814
3815         if (!discard_pending &&
3816             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3817                 int hash;
3818                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3819                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3820                 if (sh->qd_idx >= 0) {
3821                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3822                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3823                 }
3824                 /* now that discard is done we can proceed with any sync */
3825                 clear_bit(STRIPE_DISCARD, &sh->state);
3826                 /*
3827                  * SCSI discard will change some bio fields and the stripe has
3828                  * no updated data, so remove it from hash list and the stripe
3829                  * will be reinitialized
3830                  */
3831 unhash:
3832                 hash = sh->hash_lock_index;
3833                 spin_lock_irq(conf->hash_locks + hash);
3834                 remove_hash(sh);
3835                 spin_unlock_irq(conf->hash_locks + hash);
3836                 if (head_sh->batch_head) {
3837                         sh = list_first_entry(&sh->batch_list,
3838                                               struct stripe_head, batch_list);
3839                         if (sh != head_sh)
3840                                         goto unhash;
3841                 }
3842                 sh = head_sh;
3843
3844                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3845                         set_bit(STRIPE_HANDLE, &sh->state);
3846
3847         }
3848
3849         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3850                 if (atomic_dec_and_test(&conf->pending_full_writes))
3851                         md_wakeup_thread(conf->mddev->thread);
3852
3853         if (head_sh->batch_head && do_endio)
3854                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3855 }
3856
3857 /*
3858  * For RMW in write back cache, we need extra page in prexor to store the
3859  * old data. This page is stored in dev->orig_page.
3860  *
3861  * This function checks whether we have data for prexor. The exact logic
3862  * is:
3863  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3864  */
3865 static inline bool uptodate_for_rmw(struct r5dev *dev)
3866 {
3867         return (test_bit(R5_UPTODATE, &dev->flags)) &&
3868                 (!test_bit(R5_InJournal, &dev->flags) ||
3869                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3870 }
3871
3872 static int handle_stripe_dirtying(struct r5conf *conf,
3873                                   struct stripe_head *sh,
3874                                   struct stripe_head_state *s,
3875                                   int disks)
3876 {
3877         int rmw = 0, rcw = 0, i;
3878         sector_t recovery_cp = conf->mddev->recovery_cp;
3879
3880         /* Check whether resync is now happening or should start.
3881          * If yes, then the array is dirty (after unclean shutdown or
3882          * initial creation), so parity in some stripes might be inconsistent.
3883          * In this case, we need to always do reconstruct-write, to ensure
3884          * that in case of drive failure or read-error correction, we
3885          * generate correct data from the parity.
3886          */
3887         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3888             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3889              s->failed == 0)) {
3890                 /* Calculate the real rcw later - for now make it
3891                  * look like rcw is cheaper
3892                  */
3893                 rcw = 1; rmw = 2;
3894                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3895                          conf->rmw_level, (unsigned long long)recovery_cp,
3896                          (unsigned long long)sh->sector);
3897         } else for (i = disks; i--; ) {
3898                 /* would I have to read this buffer for read_modify_write */
3899                 struct r5dev *dev = &sh->dev[i];
3900                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3901                      i == sh->pd_idx || i == sh->qd_idx ||
3902                      test_bit(R5_InJournal, &dev->flags)) &&
3903                     !test_bit(R5_LOCKED, &dev->flags) &&
3904                     !(uptodate_for_rmw(dev) ||
3905                       test_bit(R5_Wantcompute, &dev->flags))) {
3906                         if (test_bit(R5_Insync, &dev->flags))
3907                                 rmw++;
3908                         else
3909                                 rmw += 2*disks;  /* cannot read it */
3910                 }
3911                 /* Would I have to read this buffer for reconstruct_write */
3912                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3913                     i != sh->pd_idx && i != sh->qd_idx &&
3914                     !test_bit(R5_LOCKED, &dev->flags) &&
3915                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3916                       test_bit(R5_Wantcompute, &dev->flags))) {
3917                         if (test_bit(R5_Insync, &dev->flags))
3918                                 rcw++;
3919                         else
3920                                 rcw += 2*disks;
3921                 }
3922         }
3923
3924         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3925                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
3926         set_bit(STRIPE_HANDLE, &sh->state);
3927         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3928                 /* prefer read-modify-write, but need to get some data */
3929                 if (conf->mddev->queue)
3930                         blk_add_trace_msg(conf->mddev->queue,
3931                                           "raid5 rmw %llu %d",
3932                                           (unsigned long long)sh->sector, rmw);
3933                 for (i = disks; i--; ) {
3934                         struct r5dev *dev = &sh->dev[i];
3935                         if (test_bit(R5_InJournal, &dev->flags) &&
3936                             dev->page == dev->orig_page &&
3937                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3938                                 /* alloc page for prexor */
3939                                 struct page *p = alloc_page(GFP_NOIO);
3940
3941                                 if (p) {
3942                                         dev->orig_page = p;
3943                                         continue;
3944                                 }
3945
3946                                 /*
3947                                  * alloc_page() failed, try use
3948                                  * disk_info->extra_page
3949                                  */
3950                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3951                                                       &conf->cache_state)) {
3952                                         r5c_use_extra_page(sh);
3953                                         break;
3954                                 }
3955
3956                                 /* extra_page in use, add to delayed_list */
3957                                 set_bit(STRIPE_DELAYED, &sh->state);
3958                                 s->waiting_extra_page = 1;
3959                                 return -EAGAIN;
3960                         }
3961                 }
3962
3963                 for (i = disks; i--; ) {
3964                         struct r5dev *dev = &sh->dev[i];
3965                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3966                              i == sh->pd_idx || i == sh->qd_idx ||
3967                              test_bit(R5_InJournal, &dev->flags)) &&
3968                             !test_bit(R5_LOCKED, &dev->flags) &&
3969                             !(uptodate_for_rmw(dev) ||
3970                               test_bit(R5_Wantcompute, &dev->flags)) &&
3971                             test_bit(R5_Insync, &dev->flags)) {
3972                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3973                                              &sh->state)) {
3974                                         pr_debug("Read_old block %d for r-m-w\n",
3975                                                  i);
3976                                         set_bit(R5_LOCKED, &dev->flags);
3977                                         set_bit(R5_Wantread, &dev->flags);
3978                                         s->locked++;
3979                                 } else {
3980                                         set_bit(STRIPE_DELAYED, &sh->state);
3981                                         set_bit(STRIPE_HANDLE, &sh->state);
3982                                 }
3983                         }
3984                 }
3985         }
3986         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3987                 /* want reconstruct write, but need to get some data */
3988                 int qread =0;
3989                 rcw = 0;
3990                 for (i = disks; i--; ) {
3991                         struct r5dev *dev = &sh->dev[i];
3992                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3993                             i != sh->pd_idx && i != sh->qd_idx &&
3994                             !test_bit(R5_LOCKED, &dev->flags) &&
3995                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3996                               test_bit(R5_Wantcompute, &dev->flags))) {
3997                                 rcw++;
3998                                 if (test_bit(R5_Insync, &dev->flags) &&
3999                                     test_bit(STRIPE_PREREAD_ACTIVE,
4000                                              &sh->state)) {
4001                                         pr_debug("Read_old block "
4002                                                 "%d for Reconstruct\n", i);
4003                                         set_bit(R5_LOCKED, &dev->flags);
4004                                         set_bit(R5_Wantread, &dev->flags);
4005                                         s->locked++;
4006                                         qread++;
4007                                 } else {
4008                                         set_bit(STRIPE_DELAYED, &sh->state);
4009                                         set_bit(STRIPE_HANDLE, &sh->state);
4010                                 }
4011                         }
4012                 }
4013                 if (rcw && conf->mddev->queue)
4014                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4015                                           (unsigned long long)sh->sector,
4016                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4017         }
4018
4019         if (rcw > disks && rmw > disks &&
4020             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4021                 set_bit(STRIPE_DELAYED, &sh->state);
4022
4023         /* now if nothing is locked, and if we have enough data,
4024          * we can start a write request
4025          */
4026         /* since handle_stripe can be called at any time we need to handle the
4027          * case where a compute block operation has been submitted and then a
4028          * subsequent call wants to start a write request.  raid_run_ops only
4029          * handles the case where compute block and reconstruct are requested
4030          * simultaneously.  If this is not the case then new writes need to be
4031          * held off until the compute completes.
4032          */
4033         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4034             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4035              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4036                 schedule_reconstruction(sh, s, rcw == 0, 0);
4037         return 0;
4038 }
4039
4040 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4041                                 struct stripe_head_state *s, int disks)
4042 {
4043         struct r5dev *dev = NULL;
4044
4045         BUG_ON(sh->batch_head);
4046         set_bit(STRIPE_HANDLE, &sh->state);
4047
4048         switch (sh->check_state) {
4049         case check_state_idle:
4050                 /* start a new check operation if there are no failures */
4051                 if (s->failed == 0) {
4052                         BUG_ON(s->uptodate != disks);
4053                         sh->check_state = check_state_run;
4054                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4055                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4056                         s->uptodate--;
4057                         break;
4058                 }
4059                 dev = &sh->dev[s->failed_num[0]];
4060                 /* fall through */
4061         case check_state_compute_result:
4062                 sh->check_state = check_state_idle;
4063                 if (!dev)
4064                         dev = &sh->dev[sh->pd_idx];
4065
4066                 /* check that a write has not made the stripe insync */
4067                 if (test_bit(STRIPE_INSYNC, &sh->state))
4068                         break;
4069
4070                 /* either failed parity check, or recovery is happening */
4071                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4072                 BUG_ON(s->uptodate != disks);
4073
4074                 set_bit(R5_LOCKED, &dev->flags);
4075                 s->locked++;
4076                 set_bit(R5_Wantwrite, &dev->flags);
4077
4078                 clear_bit(STRIPE_DEGRADED, &sh->state);
4079                 set_bit(STRIPE_INSYNC, &sh->state);
4080                 break;
4081         case check_state_run:
4082                 break; /* we will be called again upon completion */
4083         case check_state_check_result:
4084                 sh->check_state = check_state_idle;
4085
4086                 /* if a failure occurred during the check operation, leave
4087                  * STRIPE_INSYNC not set and let the stripe be handled again
4088                  */
4089                 if (s->failed)
4090                         break;
4091
4092                 /* handle a successful check operation, if parity is correct
4093                  * we are done.  Otherwise update the mismatch count and repair
4094                  * parity if !MD_RECOVERY_CHECK
4095                  */
4096                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4097                         /* parity is correct (on disc,
4098                          * not in buffer any more)
4099                          */
4100                         set_bit(STRIPE_INSYNC, &sh->state);
4101                 else {
4102                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4103                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4104                                 /* don't try to repair!! */
4105                                 set_bit(STRIPE_INSYNC, &sh->state);
4106                                 pr_warn_ratelimited("%s: mismatch sector in range "
4107                                                     "%llu-%llu\n", mdname(conf->mddev),
4108                                                     (unsigned long long) sh->sector,
4109                                                     (unsigned long long) sh->sector +
4110                                                     STRIPE_SECTORS);
4111                         } else {
4112                                 sh->check_state = check_state_compute_run;
4113                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4114                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4115                                 set_bit(R5_Wantcompute,
4116                                         &sh->dev[sh->pd_idx].flags);
4117                                 sh->ops.target = sh->pd_idx;
4118                                 sh->ops.target2 = -1;
4119                                 s->uptodate++;
4120                         }
4121                 }
4122                 break;
4123         case check_state_compute_run:
4124                 break;
4125         default:
4126                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4127                        __func__, sh->check_state,
4128                        (unsigned long long) sh->sector);
4129                 BUG();
4130         }
4131 }
4132
4133 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4134                                   struct stripe_head_state *s,
4135                                   int disks)
4136 {
4137         int pd_idx = sh->pd_idx;
4138         int qd_idx = sh->qd_idx;
4139         struct r5dev *dev;
4140
4141         BUG_ON(sh->batch_head);
4142         set_bit(STRIPE_HANDLE, &sh->state);
4143
4144         BUG_ON(s->failed > 2);
4145
4146         /* Want to check and possibly repair P and Q.
4147          * However there could be one 'failed' device, in which
4148          * case we can only check one of them, possibly using the
4149          * other to generate missing data
4150          */
4151
4152         switch (sh->check_state) {
4153         case check_state_idle:
4154                 /* start a new check operation if there are < 2 failures */
4155                 if (s->failed == s->q_failed) {
4156                         /* The only possible failed device holds Q, so it
4157                          * makes sense to check P (If anything else were failed,
4158                          * we would have used P to recreate it).
4159                          */
4160                         sh->check_state = check_state_run;
4161                 }
4162                 if (!s->q_failed && s->failed < 2) {
4163                         /* Q is not failed, and we didn't use it to generate
4164                          * anything, so it makes sense to check it
4165                          */
4166                         if (sh->check_state == check_state_run)
4167                                 sh->check_state = check_state_run_pq;
4168                         else
4169                                 sh->check_state = check_state_run_q;
4170                 }
4171
4172                 /* discard potentially stale zero_sum_result */
4173                 sh->ops.zero_sum_result = 0;
4174
4175                 if (sh->check_state == check_state_run) {
4176                         /* async_xor_zero_sum destroys the contents of P */
4177                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4178                         s->uptodate--;
4179                 }
4180                 if (sh->check_state >= check_state_run &&
4181                     sh->check_state <= check_state_run_pq) {
4182                         /* async_syndrome_zero_sum preserves P and Q, so
4183                          * no need to mark them !uptodate here
4184                          */
4185                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4186                         break;
4187                 }
4188
4189                 /* we have 2-disk failure */
4190                 BUG_ON(s->failed != 2);
4191                 /* fall through */
4192         case check_state_compute_result:
4193                 sh->check_state = check_state_idle;
4194
4195                 /* check that a write has not made the stripe insync */
4196                 if (test_bit(STRIPE_INSYNC, &sh->state))
4197                         break;
4198
4199                 /* now write out any block on a failed drive,
4200                  * or P or Q if they were recomputed
4201                  */
4202                 dev = NULL;
4203                 if (s->failed == 2) {
4204                         dev = &sh->dev[s->failed_num[1]];
4205                         s->locked++;
4206                         set_bit(R5_LOCKED, &dev->flags);
4207                         set_bit(R5_Wantwrite, &dev->flags);
4208                 }
4209                 if (s->failed >= 1) {
4210                         dev = &sh->dev[s->failed_num[0]];
4211                         s->locked++;
4212                         set_bit(R5_LOCKED, &dev->flags);
4213                         set_bit(R5_Wantwrite, &dev->flags);
4214                 }
4215                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4216                         dev = &sh->dev[pd_idx];
4217                         s->locked++;
4218                         set_bit(R5_LOCKED, &dev->flags);
4219                         set_bit(R5_Wantwrite, &dev->flags);
4220                 }
4221                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4222                         dev = &sh->dev[qd_idx];
4223                         s->locked++;
4224                         set_bit(R5_LOCKED, &dev->flags);
4225                         set_bit(R5_Wantwrite, &dev->flags);
4226                 }
4227                 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4228                               "%s: disk%td not up to date\n",
4229                               mdname(conf->mddev),
4230                               dev - (struct r5dev *) &sh->dev)) {
4231                         clear_bit(R5_LOCKED, &dev->flags);
4232                         clear_bit(R5_Wantwrite, &dev->flags);
4233                         s->locked--;
4234                 }
4235                 clear_bit(STRIPE_DEGRADED, &sh->state);
4236
4237                 set_bit(STRIPE_INSYNC, &sh->state);
4238                 break;
4239         case check_state_run:
4240         case check_state_run_q:
4241         case check_state_run_pq:
4242                 break; /* we will be called again upon completion */
4243         case check_state_check_result:
4244                 sh->check_state = check_state_idle;
4245
4246                 /* handle a successful check operation, if parity is correct
4247                  * we are done.  Otherwise update the mismatch count and repair
4248                  * parity if !MD_RECOVERY_CHECK
4249                  */
4250                 if (sh->ops.zero_sum_result == 0) {
4251                         /* both parities are correct */
4252                         if (!s->failed)
4253                                 set_bit(STRIPE_INSYNC, &sh->state);
4254                         else {
4255                                 /* in contrast to the raid5 case we can validate
4256                                  * parity, but still have a failure to write
4257                                  * back
4258                                  */
4259                                 sh->check_state = check_state_compute_result;
4260                                 /* Returning at this point means that we may go
4261                                  * off and bring p and/or q uptodate again so
4262                                  * we make sure to check zero_sum_result again
4263                                  * to verify if p or q need writeback
4264                                  */
4265                         }
4266                 } else {
4267                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4268                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4269                                 /* don't try to repair!! */
4270                                 set_bit(STRIPE_INSYNC, &sh->state);
4271                                 pr_warn_ratelimited("%s: mismatch sector in range "
4272                                                     "%llu-%llu\n", mdname(conf->mddev),
4273                                                     (unsigned long long) sh->sector,
4274                                                     (unsigned long long) sh->sector +
4275                                                     STRIPE_SECTORS);
4276                         } else {
4277                                 int *target = &sh->ops.target;
4278
4279                                 sh->ops.target = -1;
4280                                 sh->ops.target2 = -1;
4281                                 sh->check_state = check_state_compute_run;
4282                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4283                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4284                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4285                                         set_bit(R5_Wantcompute,
4286                                                 &sh->dev[pd_idx].flags);
4287                                         *target = pd_idx;
4288                                         target = &sh->ops.target2;
4289                                         s->uptodate++;
4290                                 }
4291                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4292                                         set_bit(R5_Wantcompute,
4293                                                 &sh->dev[qd_idx].flags);
4294                                         *target = qd_idx;
4295                                         s->uptodate++;
4296                                 }
4297                         }
4298                 }
4299                 break;
4300         case check_state_compute_run:
4301                 break;
4302         default:
4303                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4304                         __func__, sh->check_state,
4305                         (unsigned long long) sh->sector);
4306                 BUG();
4307         }
4308 }
4309
4310 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4311 {
4312         int i;
4313
4314         /* We have read all the blocks in this stripe and now we need to
4315          * copy some of them into a target stripe for expand.
4316          */
4317         struct dma_async_tx_descriptor *tx = NULL;
4318         BUG_ON(sh->batch_head);
4319         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4320         for (i = 0; i < sh->disks; i++)
4321                 if (i != sh->pd_idx && i != sh->qd_idx) {
4322                         int dd_idx, j;
4323                         struct stripe_head *sh2;
4324                         struct async_submit_ctl submit;
4325
4326                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4327                         sector_t s = raid5_compute_sector(conf, bn, 0,
4328                                                           &dd_idx, NULL);
4329                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4330                         if (sh2 == NULL)
4331                                 /* so far only the early blocks of this stripe
4332                                  * have been requested.  When later blocks
4333                                  * get requested, we will try again
4334                                  */
4335                                 continue;
4336                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4337                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4338                                 /* must have already done this block */
4339                                 raid5_release_stripe(sh2);
4340                                 continue;
4341                         }
4342
4343                         /* place all the copies on one channel */
4344                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4345                         tx = async_memcpy(sh2->dev[dd_idx].page,
4346                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
4347                                           &submit);
4348
4349                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4350                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4351                         for (j = 0; j < conf->raid_disks; j++)
4352                                 if (j != sh2->pd_idx &&
4353                                     j != sh2->qd_idx &&
4354                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4355                                         break;
4356                         if (j == conf->raid_disks) {
4357                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4358                                 set_bit(STRIPE_HANDLE, &sh2->state);
4359                         }
4360                         raid5_release_stripe(sh2);
4361
4362                 }
4363         /* done submitting copies, wait for them to complete */
4364         async_tx_quiesce(&tx);
4365 }
4366
4367 /*
4368  * handle_stripe - do things to a stripe.
4369  *
4370  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4371  * state of various bits to see what needs to be done.
4372  * Possible results:
4373  *    return some read requests which now have data
4374  *    return some write requests which are safely on storage
4375  *    schedule a read on some buffers
4376  *    schedule a write of some buffers
4377  *    return confirmation of parity correctness
4378  *
4379  */
4380
4381 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4382 {
4383         struct r5conf *conf = sh->raid_conf;
4384         int disks = sh->disks;
4385         struct r5dev *dev;
4386         int i;
4387         int do_recovery = 0;
4388
4389         memset(s, 0, sizeof(*s));
4390
4391         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4392         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4393         s->failed_num[0] = -1;
4394         s->failed_num[1] = -1;
4395         s->log_failed = r5l_log_disk_error(conf);
4396
4397         /* Now to look around and see what can be done */
4398         rcu_read_lock();
4399         for (i=disks; i--; ) {
4400                 struct md_rdev *rdev;
4401                 sector_t first_bad;
4402                 int bad_sectors;
4403                 int is_bad = 0;
4404
4405                 dev = &sh->dev[i];
4406
4407                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4408                          i, dev->flags,
4409                          dev->toread, dev->towrite, dev->written);
4410                 /* maybe we can reply to a read
4411                  *
4412                  * new wantfill requests are only permitted while
4413                  * ops_complete_biofill is guaranteed to be inactive
4414                  */
4415                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4416                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4417                         set_bit(R5_Wantfill, &dev->flags);
4418
4419                 /* now count some things */
4420                 if (test_bit(R5_LOCKED, &dev->flags))
4421                         s->locked++;
4422                 if (test_bit(R5_UPTODATE, &dev->flags))
4423                         s->uptodate++;
4424                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4425                         s->compute++;
4426                         BUG_ON(s->compute > 2);
4427                 }
4428
4429                 if (test_bit(R5_Wantfill, &dev->flags))
4430                         s->to_fill++;
4431                 else if (dev->toread)
4432                         s->to_read++;
4433                 if (dev->towrite) {
4434                         s->to_write++;
4435                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4436                                 s->non_overwrite++;
4437                 }
4438                 if (dev->written)
4439                         s->written++;
4440                 /* Prefer to use the replacement for reads, but only
4441                  * if it is recovered enough and has no bad blocks.
4442                  */
4443                 rdev = rcu_dereference(conf->disks[i].replacement);
4444                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4445                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4446                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4447                                  &first_bad, &bad_sectors))
4448                         set_bit(R5_ReadRepl, &dev->flags);
4449                 else {
4450                         if (rdev && !test_bit(Faulty, &rdev->flags))
4451                                 set_bit(R5_NeedReplace, &dev->flags);
4452                         else
4453                                 clear_bit(R5_NeedReplace, &dev->flags);
4454                         rdev = rcu_dereference(conf->disks[i].rdev);
4455                         clear_bit(R5_ReadRepl, &dev->flags);
4456                 }
4457                 if (rdev && test_bit(Faulty, &rdev->flags))
4458                         rdev = NULL;
4459                 if (rdev) {
4460                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4461                                              &first_bad, &bad_sectors);
4462                         if (s->blocked_rdev == NULL
4463                             && (test_bit(Blocked, &rdev->flags)
4464                                 || is_bad < 0)) {
4465                                 if (is_bad < 0)
4466                                         set_bit(BlockedBadBlocks,
4467                                                 &rdev->flags);
4468                                 s->blocked_rdev = rdev;
4469                                 atomic_inc(&rdev->nr_pending);
4470                         }
4471                 }
4472                 clear_bit(R5_Insync, &dev->flags);
4473                 if (!rdev)
4474                         /* Not in-sync */;
4475                 else if (is_bad) {
4476                         /* also not in-sync */
4477                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4478                             test_bit(R5_UPTODATE, &dev->flags)) {
4479                                 /* treat as in-sync, but with a read error
4480                                  * which we can now try to correct
4481                                  */
4482                                 set_bit(R5_Insync, &dev->flags);
4483                                 set_bit(R5_ReadError, &dev->flags);
4484                         }
4485                 } else if (test_bit(In_sync, &rdev->flags))
4486                         set_bit(R5_Insync, &dev->flags);
4487                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4488                         /* in sync if before recovery_offset */
4489                         set_bit(R5_Insync, &dev->flags);
4490                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4491                          test_bit(R5_Expanded, &dev->flags))
4492                         /* If we've reshaped into here, we assume it is Insync.
4493                          * We will shortly update recovery_offset to make
4494                          * it official.
4495                          */
4496                         set_bit(R5_Insync, &dev->flags);
4497
4498                 if (test_bit(R5_WriteError, &dev->flags)) {
4499                         /* This flag does not apply to '.replacement'
4500                          * only to .rdev, so make sure to check that*/
4501                         struct md_rdev *rdev2 = rcu_dereference(
4502                                 conf->disks[i].rdev);
4503                         if (rdev2 == rdev)
4504                                 clear_bit(R5_Insync, &dev->flags);
4505                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4506                                 s->handle_bad_blocks = 1;
4507                                 atomic_inc(&rdev2->nr_pending);
4508                         } else
4509                                 clear_bit(R5_WriteError, &dev->flags);
4510                 }
4511                 if (test_bit(R5_MadeGood, &dev->flags)) {
4512                         /* This flag does not apply to '.replacement'
4513                          * only to .rdev, so make sure to check that*/
4514                         struct md_rdev *rdev2 = rcu_dereference(
4515                                 conf->disks[i].rdev);
4516                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4517                                 s->handle_bad_blocks = 1;
4518                                 atomic_inc(&rdev2->nr_pending);
4519                         } else
4520                                 clear_bit(R5_MadeGood, &dev->flags);
4521                 }
4522                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4523                         struct md_rdev *rdev2 = rcu_dereference(
4524                                 conf->disks[i].replacement);
4525                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4526                                 s->handle_bad_blocks = 1;
4527                                 atomic_inc(&rdev2->nr_pending);
4528                         } else
4529                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4530                 }
4531                 if (!test_bit(R5_Insync, &dev->flags)) {
4532                         /* The ReadError flag will just be confusing now */
4533                         clear_bit(R5_ReadError, &dev->flags);
4534                         clear_bit(R5_ReWrite, &dev->flags);
4535                 }
4536                 if (test_bit(R5_ReadError, &dev->flags))
4537                         clear_bit(R5_Insync, &dev->flags);
4538                 if (!test_bit(R5_Insync, &dev->flags)) {
4539                         if (s->failed < 2)
4540                                 s->failed_num[s->failed] = i;
4541                         s->failed++;
4542                         if (rdev && !test_bit(Faulty, &rdev->flags))
4543                                 do_recovery = 1;
4544                         else if (!rdev) {
4545                                 rdev = rcu_dereference(
4546                                     conf->disks[i].replacement);
4547                                 if (rdev && !test_bit(Faulty, &rdev->flags))
4548                                         do_recovery = 1;
4549                         }
4550                 }
4551
4552                 if (test_bit(R5_InJournal, &dev->flags))
4553                         s->injournal++;
4554                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4555                         s->just_cached++;
4556         }
4557         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4558                 /* If there is a failed device being replaced,
4559                  *     we must be recovering.
4560                  * else if we are after recovery_cp, we must be syncing
4561                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4562                  * else we can only be replacing
4563                  * sync and recovery both need to read all devices, and so
4564                  * use the same flag.
4565                  */
4566                 if (do_recovery ||
4567                     sh->sector >= conf->mddev->recovery_cp ||
4568                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4569                         s->syncing = 1;
4570                 else
4571                         s->replacing = 1;
4572         }
4573         rcu_read_unlock();
4574 }
4575
4576 static int clear_batch_ready(struct stripe_head *sh)
4577 {
4578         /* Return '1' if this is a member of batch, or
4579          * '0' if it is a lone stripe or a head which can now be
4580          * handled.
4581          */
4582         struct stripe_head *tmp;
4583         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4584                 return (sh->batch_head && sh->batch_head != sh);
4585         spin_lock(&sh->stripe_lock);
4586         if (!sh->batch_head) {
4587                 spin_unlock(&sh->stripe_lock);
4588                 return 0;
4589         }
4590
4591         /*
4592          * this stripe could be added to a batch list before we check
4593          * BATCH_READY, skips it
4594          */
4595         if (sh->batch_head != sh) {
4596                 spin_unlock(&sh->stripe_lock);
4597                 return 1;
4598         }
4599         spin_lock(&sh->batch_lock);
4600         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4601                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4602         spin_unlock(&sh->batch_lock);
4603         spin_unlock(&sh->stripe_lock);
4604
4605         /*
4606          * BATCH_READY is cleared, no new stripes can be added.
4607          * batch_list can be accessed without lock
4608          */
4609         return 0;
4610 }
4611
4612 static void break_stripe_batch_list(struct stripe_head *head_sh,
4613                                     unsigned long handle_flags)
4614 {
4615         struct stripe_head *sh, *next;
4616         int i;
4617         int do_wakeup = 0;
4618
4619         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4620
4621                 list_del_init(&sh->batch_list);
4622
4623                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4624                                           (1 << STRIPE_SYNCING) |
4625                                           (1 << STRIPE_REPLACED) |
4626                                           (1 << STRIPE_DELAYED) |
4627                                           (1 << STRIPE_BIT_DELAY) |
4628                                           (1 << STRIPE_FULL_WRITE) |
4629                                           (1 << STRIPE_BIOFILL_RUN) |
4630                                           (1 << STRIPE_COMPUTE_RUN)  |
4631                                           (1 << STRIPE_DISCARD) |
4632                                           (1 << STRIPE_BATCH_READY) |
4633                                           (1 << STRIPE_BATCH_ERR) |
4634                                           (1 << STRIPE_BITMAP_PENDING)),
4635                         "stripe state: %lx\n", sh->state);
4636                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4637                                               (1 << STRIPE_REPLACED)),
4638                         "head stripe state: %lx\n", head_sh->state);
4639
4640                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4641                                             (1 << STRIPE_PREREAD_ACTIVE) |
4642                                             (1 << STRIPE_DEGRADED) |
4643                                             (1 << STRIPE_ON_UNPLUG_LIST)),
4644                               head_sh->state & (1 << STRIPE_INSYNC));
4645
4646                 sh->check_state = head_sh->check_state;
4647                 sh->reconstruct_state = head_sh->reconstruct_state;
4648                 spin_lock_irq(&sh->stripe_lock);
4649                 sh->batch_head = NULL;
4650                 spin_unlock_irq(&sh->stripe_lock);
4651                 for (i = 0; i < sh->disks; i++) {
4652                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4653                                 do_wakeup = 1;
4654                         sh->dev[i].flags = head_sh->dev[i].flags &
4655                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4656                 }
4657                 if (handle_flags == 0 ||
4658                     sh->state & handle_flags)
4659                         set_bit(STRIPE_HANDLE, &sh->state);
4660                 raid5_release_stripe(sh);
4661         }
4662         spin_lock_irq(&head_sh->stripe_lock);
4663         head_sh->batch_head = NULL;
4664         spin_unlock_irq(&head_sh->stripe_lock);
4665         for (i = 0; i < head_sh->disks; i++)
4666                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4667                         do_wakeup = 1;
4668         if (head_sh->state & handle_flags)
4669                 set_bit(STRIPE_HANDLE, &head_sh->state);
4670
4671         if (do_wakeup)
4672                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4673 }
4674
4675 static void handle_stripe(struct stripe_head *sh)
4676 {
4677         struct stripe_head_state s;
4678         struct r5conf *conf = sh->raid_conf;
4679         int i;
4680         int prexor;
4681         int disks = sh->disks;
4682         struct r5dev *pdev, *qdev;
4683
4684         clear_bit(STRIPE_HANDLE, &sh->state);
4685         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4686                 /* already being handled, ensure it gets handled
4687                  * again when current action finishes */
4688                 set_bit(STRIPE_HANDLE, &sh->state);
4689                 return;
4690         }
4691
4692         if (clear_batch_ready(sh) ) {
4693                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4694                 return;
4695         }
4696
4697         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4698                 break_stripe_batch_list(sh, 0);
4699
4700         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4701                 spin_lock(&sh->stripe_lock);
4702                 /*
4703                  * Cannot process 'sync' concurrently with 'discard'.
4704                  * Flush data in r5cache before 'sync'.
4705                  */
4706                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4707                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4708                     !test_bit(STRIPE_DISCARD, &sh->state) &&
4709                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4710                         set_bit(STRIPE_SYNCING, &sh->state);
4711                         clear_bit(STRIPE_INSYNC, &sh->state);
4712                         clear_bit(STRIPE_REPLACED, &sh->state);
4713                 }
4714                 spin_unlock(&sh->stripe_lock);
4715         }
4716         clear_bit(STRIPE_DELAYED, &sh->state);
4717
4718         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4719                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4720                (unsigned long long)sh->sector, sh->state,
4721                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4722                sh->check_state, sh->reconstruct_state);
4723
4724         analyse_stripe(sh, &s);
4725
4726         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4727                 goto finish;
4728
4729         if (s.handle_bad_blocks ||
4730             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4731                 set_bit(STRIPE_HANDLE, &sh->state);
4732                 goto finish;
4733         }
4734
4735         if (unlikely(s.blocked_rdev)) {
4736                 if (s.syncing || s.expanding || s.expanded ||
4737                     s.replacing || s.to_write || s.written) {
4738                         set_bit(STRIPE_HANDLE, &sh->state);
4739                         goto finish;
4740                 }
4741                 /* There is nothing for the blocked_rdev to block */
4742                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4743                 s.blocked_rdev = NULL;
4744         }
4745
4746         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4747                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4748                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4749         }
4750
4751         pr_debug("locked=%d uptodate=%d to_read=%d"
4752                " to_write=%d failed=%d failed_num=%d,%d\n",
4753                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4754                s.failed_num[0], s.failed_num[1]);
4755         /*
4756          * check if the array has lost more than max_degraded devices and,
4757          * if so, some requests might need to be failed.
4758          *
4759          * When journal device failed (log_failed), we will only process
4760          * the stripe if there is data need write to raid disks
4761          */
4762         if (s.failed > conf->max_degraded ||
4763             (s.log_failed && s.injournal == 0)) {
4764                 sh->check_state = 0;
4765                 sh->reconstruct_state = 0;
4766                 break_stripe_batch_list(sh, 0);
4767                 if (s.to_read+s.to_write+s.written)
4768                         handle_failed_stripe(conf, sh, &s, disks);
4769                 if (s.syncing + s.replacing)
4770                         handle_failed_sync(conf, sh, &s);
4771         }
4772
4773         /* Now we check to see if any write operations have recently
4774          * completed
4775          */
4776         prexor = 0;
4777         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4778                 prexor = 1;
4779         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4780             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4781                 sh->reconstruct_state = reconstruct_state_idle;
4782
4783                 /* All the 'written' buffers and the parity block are ready to
4784                  * be written back to disk
4785                  */
4786                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4787                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4788                 BUG_ON(sh->qd_idx >= 0 &&
4789                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4790                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4791                 for (i = disks; i--; ) {
4792                         struct r5dev *dev = &sh->dev[i];
4793                         if (test_bit(R5_LOCKED, &dev->flags) &&
4794                                 (i == sh->pd_idx || i == sh->qd_idx ||
4795                                  dev->written || test_bit(R5_InJournal,
4796                                                           &dev->flags))) {
4797                                 pr_debug("Writing block %d\n", i);
4798                                 set_bit(R5_Wantwrite, &dev->flags);
4799                                 if (prexor)
4800                                         continue;
4801                                 if (s.failed > 1)
4802                                         continue;
4803                                 if (!test_bit(R5_Insync, &dev->flags) ||
4804                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4805                                      s.failed == 0))
4806                                         set_bit(STRIPE_INSYNC, &sh->state);
4807                         }
4808                 }
4809                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4810                         s.dec_preread_active = 1;
4811         }
4812
4813         /*
4814          * might be able to return some write requests if the parity blocks
4815          * are safe, or on a failed drive
4816          */
4817         pdev = &sh->dev[sh->pd_idx];
4818         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4819                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4820         qdev = &sh->dev[sh->qd_idx];
4821         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4822                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4823                 || conf->level < 6;
4824
4825         if (s.written &&
4826             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4827                              && !test_bit(R5_LOCKED, &pdev->flags)
4828                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4829                                  test_bit(R5_Discard, &pdev->flags))))) &&
4830             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4831                              && !test_bit(R5_LOCKED, &qdev->flags)
4832                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4833                                  test_bit(R5_Discard, &qdev->flags))))))
4834                 handle_stripe_clean_event(conf, sh, disks);
4835
4836         if (s.just_cached)
4837                 r5c_handle_cached_data_endio(conf, sh, disks);
4838         log_stripe_write_finished(sh);
4839
4840         /* Now we might consider reading some blocks, either to check/generate
4841          * parity, or to satisfy requests
4842          * or to load a block that is being partially written.
4843          */
4844         if (s.to_read || s.non_overwrite
4845             || (conf->level == 6 && s.to_write && s.failed)
4846             || (s.syncing && (s.uptodate + s.compute < disks))
4847             || s.replacing
4848             || s.expanding)
4849                 handle_stripe_fill(sh, &s, disks);
4850
4851         /*
4852          * When the stripe finishes full journal write cycle (write to journal
4853          * and raid disk), this is the clean up procedure so it is ready for
4854          * next operation.
4855          */
4856         r5c_finish_stripe_write_out(conf, sh, &s);
4857
4858         /*
4859          * Now to consider new write requests, cache write back and what else,
4860          * if anything should be read.  We do not handle new writes when:
4861          * 1/ A 'write' operation (copy+xor) is already in flight.
4862          * 2/ A 'check' operation is in flight, as it may clobber the parity
4863          *    block.
4864          * 3/ A r5c cache log write is in flight.
4865          */
4866
4867         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4868                 if (!r5c_is_writeback(conf->log)) {
4869                         if (s.to_write)
4870                                 handle_stripe_dirtying(conf, sh, &s, disks);
4871                 } else { /* write back cache */
4872                         int ret = 0;
4873
4874                         /* First, try handle writes in caching phase */
4875                         if (s.to_write)
4876                                 ret = r5c_try_caching_write(conf, sh, &s,
4877                                                             disks);
4878                         /*
4879                          * If caching phase failed: ret == -EAGAIN
4880                          *    OR
4881                          * stripe under reclaim: !caching && injournal
4882                          *
4883                          * fall back to handle_stripe_dirtying()
4884                          */
4885                         if (ret == -EAGAIN ||
4886                             /* stripe under reclaim: !caching && injournal */
4887                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4888                              s.injournal > 0)) {
4889                                 ret = handle_stripe_dirtying(conf, sh, &s,
4890                                                              disks);
4891                                 if (ret == -EAGAIN)
4892                                         goto finish;
4893                         }
4894                 }
4895         }
4896
4897         /* maybe we need to check and possibly fix the parity for this stripe
4898          * Any reads will already have been scheduled, so we just see if enough
4899          * data is available.  The parity check is held off while parity
4900          * dependent operations are in flight.
4901          */
4902         if (sh->check_state ||
4903             (s.syncing && s.locked == 0 &&
4904              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4905              !test_bit(STRIPE_INSYNC, &sh->state))) {
4906                 if (conf->level == 6)
4907                         handle_parity_checks6(conf, sh, &s, disks);
4908                 else
4909                         handle_parity_checks5(conf, sh, &s, disks);
4910         }
4911
4912         if ((s.replacing || s.syncing) && s.locked == 0
4913             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4914             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4915                 /* Write out to replacement devices where possible */
4916                 for (i = 0; i < conf->raid_disks; i++)
4917                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4918                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4919                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4920                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4921                                 s.locked++;
4922                         }
4923                 if (s.replacing)
4924                         set_bit(STRIPE_INSYNC, &sh->state);
4925                 set_bit(STRIPE_REPLACED, &sh->state);
4926         }
4927         if ((s.syncing || s.replacing) && s.locked == 0 &&
4928             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4929             test_bit(STRIPE_INSYNC, &sh->state)) {
4930                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4931                 clear_bit(STRIPE_SYNCING, &sh->state);
4932                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4933                         wake_up(&conf->wait_for_overlap);
4934         }
4935
4936         /* If the failed drives are just a ReadError, then we might need
4937          * to progress the repair/check process
4938          */
4939         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4940                 for (i = 0; i < s.failed; i++) {
4941                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4942                         if (test_bit(R5_ReadError, &dev->flags)
4943                             && !test_bit(R5_LOCKED, &dev->flags)
4944                             && test_bit(R5_UPTODATE, &dev->flags)
4945                                 ) {
4946                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4947                                         set_bit(R5_Wantwrite, &dev->flags);
4948                                         set_bit(R5_ReWrite, &dev->flags);
4949                                         set_bit(R5_LOCKED, &dev->flags);
4950                                         s.locked++;
4951                                 } else {
4952                                         /* let's read it back */
4953                                         set_bit(R5_Wantread, &dev->flags);
4954                                         set_bit(R5_LOCKED, &dev->flags);
4955                                         s.locked++;
4956                                 }
4957                         }
4958                 }
4959
4960         /* Finish reconstruct operations initiated by the expansion process */
4961         if (sh->reconstruct_state == reconstruct_state_result) {
4962                 struct stripe_head *sh_src
4963                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4964                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4965                         /* sh cannot be written until sh_src has been read.
4966                          * so arrange for sh to be delayed a little
4967                          */
4968                         set_bit(STRIPE_DELAYED, &sh->state);
4969                         set_bit(STRIPE_HANDLE, &sh->state);
4970                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4971                                               &sh_src->state))
4972                                 atomic_inc(&conf->preread_active_stripes);
4973                         raid5_release_stripe(sh_src);
4974                         goto finish;
4975                 }
4976                 if (sh_src)
4977                         raid5_release_stripe(sh_src);
4978
4979                 sh->reconstruct_state = reconstruct_state_idle;
4980                 clear_bit(STRIPE_EXPANDING, &sh->state);
4981                 for (i = conf->raid_disks; i--; ) {
4982                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4983                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4984                         s.locked++;
4985                 }
4986         }
4987
4988         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4989             !sh->reconstruct_state) {
4990                 /* Need to write out all blocks after computing parity */
4991                 sh->disks = conf->raid_disks;
4992                 stripe_set_idx(sh->sector, conf, 0, sh);
4993                 schedule_reconstruction(sh, &s, 1, 1);
4994         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4995                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4996                 atomic_dec(&conf->reshape_stripes);
4997                 wake_up(&conf->wait_for_overlap);
4998                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4999         }
5000
5001         if (s.expanding && s.locked == 0 &&
5002             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
5003                 handle_stripe_expansion(conf, sh);
5004
5005 finish:
5006         /* wait for this device to become unblocked */
5007         if (unlikely(s.blocked_rdev)) {
5008                 if (conf->mddev->external)
5009                         md_wait_for_blocked_rdev(s.blocked_rdev,
5010                                                  conf->mddev);
5011                 else
5012                         /* Internal metadata will immediately
5013                          * be written by raid5d, so we don't
5014                          * need to wait here.
5015                          */
5016                         rdev_dec_pending(s.blocked_rdev,
5017                                          conf->mddev);
5018         }
5019
5020         if (s.handle_bad_blocks)
5021                 for (i = disks; i--; ) {
5022                         struct md_rdev *rdev;
5023                         struct r5dev *dev = &sh->dev[i];
5024                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5025                                 /* We own a safe reference to the rdev */
5026                                 rdev = conf->disks[i].rdev;
5027                                 if (!rdev_set_badblocks(rdev, sh->sector,
5028                                                         STRIPE_SECTORS, 0))
5029                                         md_error(conf->mddev, rdev);
5030                                 rdev_dec_pending(rdev, conf->mddev);
5031                         }
5032                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5033                                 rdev = conf->disks[i].rdev;
5034                                 rdev_clear_badblocks(rdev, sh->sector,
5035                                                      STRIPE_SECTORS, 0);
5036                                 rdev_dec_pending(rdev, conf->mddev);
5037                         }
5038                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5039                                 rdev = conf->disks[i].replacement;
5040                                 if (!rdev)
5041                                         /* rdev have been moved down */
5042                                         rdev = conf->disks[i].rdev;
5043                                 rdev_clear_badblocks(rdev, sh->sector,
5044                                                      STRIPE_SECTORS, 0);
5045                                 rdev_dec_pending(rdev, conf->mddev);
5046                         }
5047                 }
5048
5049         if (s.ops_request)
5050                 raid_run_ops(sh, s.ops_request);
5051
5052         ops_run_io(sh, &s);
5053
5054         if (s.dec_preread_active) {
5055                 /* We delay this until after ops_run_io so that if make_request
5056                  * is waiting on a flush, it won't continue until the writes
5057                  * have actually been submitted.
5058                  */
5059                 atomic_dec(&conf->preread_active_stripes);
5060                 if (atomic_read(&conf->preread_active_stripes) <
5061                     IO_THRESHOLD)
5062                         md_wakeup_thread(conf->mddev->thread);
5063         }
5064
5065         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5066 }
5067
5068 static void raid5_activate_delayed(struct r5conf *conf)
5069 {
5070         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5071                 while (!list_empty(&conf->delayed_list)) {
5072                         struct list_head *l = conf->delayed_list.next;
5073                         struct stripe_head *sh;
5074                         sh = list_entry(l, struct stripe_head, lru);
5075                         list_del_init(l);
5076                         clear_bit(STRIPE_DELAYED, &sh->state);
5077                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5078                                 atomic_inc(&conf->preread_active_stripes);
5079                         list_add_tail(&sh->lru, &conf->hold_list);
5080                         raid5_wakeup_stripe_thread(sh);
5081                 }
5082         }
5083 }
5084
5085 static void activate_bit_delay(struct r5conf *conf,
5086         struct list_head *temp_inactive_list)
5087 {
5088         /* device_lock is held */
5089         struct list_head head;
5090         list_add(&head, &conf->bitmap_list);
5091         list_del_init(&conf->bitmap_list);
5092         while (!list_empty(&head)) {
5093                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5094                 int hash;
5095                 list_del_init(&sh->lru);
5096                 atomic_inc(&sh->count);
5097                 hash = sh->hash_lock_index;
5098                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5099         }
5100 }
5101
5102 static int raid5_congested(struct mddev *mddev, int bits)
5103 {
5104         struct r5conf *conf = mddev->private;
5105
5106         /* No difference between reads and writes.  Just check
5107          * how busy the stripe_cache is
5108          */
5109
5110         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5111                 return 1;
5112
5113         /* Also checks whether there is pressure on r5cache log space */
5114         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5115                 return 1;
5116         if (conf->quiesce)
5117                 return 1;
5118         if (atomic_read(&conf->empty_inactive_list_nr))
5119                 return 1;
5120
5121         return 0;
5122 }
5123
5124 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5125 {
5126         struct r5conf *conf = mddev->private;
5127         sector_t sector = bio->bi_iter.bi_sector;
5128         unsigned int chunk_sectors;
5129         unsigned int bio_sectors = bio_sectors(bio);
5130
5131         WARN_ON_ONCE(bio->bi_partno);
5132
5133         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5134         return  chunk_sectors >=
5135                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5136 }
5137
5138 /*
5139  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5140  *  later sampled by raid5d.
5141  */
5142 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5143 {
5144         unsigned long flags;
5145
5146         spin_lock_irqsave(&conf->device_lock, flags);
5147
5148         bi->bi_next = conf->retry_read_aligned_list;
5149         conf->retry_read_aligned_list = bi;
5150
5151         spin_unlock_irqrestore(&conf->device_lock, flags);
5152         md_wakeup_thread(conf->mddev->thread);
5153 }
5154
5155 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5156                                          unsigned int *offset)
5157 {
5158         struct bio *bi;
5159
5160         bi = conf->retry_read_aligned;
5161         if (bi) {
5162                 *offset = conf->retry_read_offset;
5163                 conf->retry_read_aligned = NULL;
5164                 return bi;
5165         }
5166         bi = conf->retry_read_aligned_list;
5167         if(bi) {
5168                 conf->retry_read_aligned_list = bi->bi_next;
5169                 bi->bi_next = NULL;
5170                 *offset = 0;
5171         }
5172
5173         return bi;
5174 }
5175
5176 /*
5177  *  The "raid5_align_endio" should check if the read succeeded and if it
5178  *  did, call bio_endio on the original bio (having bio_put the new bio
5179  *  first).
5180  *  If the read failed..
5181  */
5182 static void raid5_align_endio(struct bio *bi)
5183 {
5184         struct bio* raid_bi  = bi->bi_private;
5185         struct mddev *mddev;
5186         struct r5conf *conf;
5187         struct md_rdev *rdev;
5188         blk_status_t error = bi->bi_status;
5189
5190         bio_put(bi);
5191
5192         rdev = (void*)raid_bi->bi_next;
5193         raid_bi->bi_next = NULL;
5194         mddev = rdev->mddev;
5195         conf = mddev->private;
5196
5197         rdev_dec_pending(rdev, conf->mddev);
5198
5199         if (!error) {
5200                 bio_endio(raid_bi);
5201                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5202                         wake_up(&conf->wait_for_quiescent);
5203                 return;
5204         }
5205
5206         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5207
5208         add_bio_to_retry(raid_bi, conf);
5209 }
5210
5211 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5212 {
5213         struct r5conf *conf = mddev->private;
5214         int dd_idx;
5215         struct bio* align_bi;
5216         struct md_rdev *rdev;
5217         sector_t end_sector;
5218
5219         if (!in_chunk_boundary(mddev, raid_bio)) {
5220                 pr_debug("%s: non aligned\n", __func__);
5221                 return 0;
5222         }
5223         /*
5224          * use bio_clone_fast to make a copy of the bio
5225          */
5226         align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5227         if (!align_bi)
5228                 return 0;
5229         /*
5230          *   set bi_end_io to a new function, and set bi_private to the
5231          *     original bio.
5232          */
5233         align_bi->bi_end_io  = raid5_align_endio;
5234         align_bi->bi_private = raid_bio;
5235         /*
5236          *      compute position
5237          */
5238         align_bi->bi_iter.bi_sector =
5239                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5240                                      0, &dd_idx, NULL);
5241
5242         end_sector = bio_end_sector(align_bi);
5243         rcu_read_lock();
5244         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5245         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5246             rdev->recovery_offset < end_sector) {
5247                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5248                 if (rdev &&
5249                     (test_bit(Faulty, &rdev->flags) ||
5250                     !(test_bit(In_sync, &rdev->flags) ||
5251                       rdev->recovery_offset >= end_sector)))
5252                         rdev = NULL;
5253         }
5254
5255         if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5256                 rcu_read_unlock();
5257                 bio_put(align_bi);
5258                 return 0;
5259         }
5260
5261         if (rdev) {
5262                 sector_t first_bad;
5263                 int bad_sectors;
5264
5265                 atomic_inc(&rdev->nr_pending);
5266                 rcu_read_unlock();
5267                 raid_bio->bi_next = (void*)rdev;
5268                 bio_set_dev(align_bi, rdev->bdev);
5269
5270                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5271                                 bio_sectors(align_bi),
5272                                 &first_bad, &bad_sectors)) {
5273                         bio_put(align_bi);
5274                         rdev_dec_pending(rdev, mddev);
5275                         return 0;
5276                 }
5277
5278                 /* No reshape active, so we can trust rdev->data_offset */
5279                 align_bi->bi_iter.bi_sector += rdev->data_offset;
5280
5281                 spin_lock_irq(&conf->device_lock);
5282                 wait_event_lock_irq(conf->wait_for_quiescent,
5283                                     conf->quiesce == 0,
5284                                     conf->device_lock);
5285                 atomic_inc(&conf->active_aligned_reads);
5286                 spin_unlock_irq(&conf->device_lock);
5287
5288                 if (mddev->gendisk)
5289                         trace_block_bio_remap(align_bi->bi_disk->queue,
5290                                               align_bi, disk_devt(mddev->gendisk),
5291                                               raid_bio->bi_iter.bi_sector);
5292                 generic_make_request(align_bi);
5293                 return 1;
5294         } else {
5295                 rcu_read_unlock();
5296                 bio_put(align_bi);
5297                 return 0;
5298         }
5299 }
5300
5301 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5302 {
5303         struct bio *split;
5304         sector_t sector = raid_bio->bi_iter.bi_sector;
5305         unsigned chunk_sects = mddev->chunk_sectors;
5306         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5307
5308         if (sectors < bio_sectors(raid_bio)) {
5309                 struct r5conf *conf = mddev->private;
5310                 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5311                 bio_chain(split, raid_bio);
5312                 generic_make_request(raid_bio);
5313                 raid_bio = split;
5314         }
5315
5316         if (!raid5_read_one_chunk(mddev, raid_bio))
5317                 return raid_bio;
5318
5319         return NULL;
5320 }
5321
5322 /* __get_priority_stripe - get the next stripe to process
5323  *
5324  * Full stripe writes are allowed to pass preread active stripes up until
5325  * the bypass_threshold is exceeded.  In general the bypass_count
5326  * increments when the handle_list is handled before the hold_list; however, it
5327  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5328  * stripe with in flight i/o.  The bypass_count will be reset when the
5329  * head of the hold_list has changed, i.e. the head was promoted to the
5330  * handle_list.
5331  */
5332 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5333 {
5334         struct stripe_head *sh, *tmp;
5335         struct list_head *handle_list = NULL;
5336         struct r5worker_group *wg;
5337         bool second_try = !r5c_is_writeback(conf->log) &&
5338                 !r5l_log_disk_error(conf);
5339         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5340                 r5l_log_disk_error(conf);
5341
5342 again:
5343         wg = NULL;
5344         sh = NULL;
5345         if (conf->worker_cnt_per_group == 0) {
5346                 handle_list = try_loprio ? &conf->loprio_list :
5347                                         &conf->handle_list;
5348         } else if (group != ANY_GROUP) {
5349                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5350                                 &conf->worker_groups[group].handle_list;
5351                 wg = &conf->worker_groups[group];
5352         } else {
5353                 int i;
5354                 for (i = 0; i < conf->group_cnt; i++) {
5355                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5356                                 &conf->worker_groups[i].handle_list;
5357                         wg = &conf->worker_groups[i];
5358                         if (!list_empty(handle_list))
5359                                 break;
5360                 }
5361         }
5362
5363         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5364                   __func__,
5365                   list_empty(handle_list) ? "empty" : "busy",
5366                   list_empty(&conf->hold_list) ? "empty" : "busy",
5367                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5368
5369         if (!list_empty(handle_list)) {
5370                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5371
5372                 if (list_empty(&conf->hold_list))
5373                         conf->bypass_count = 0;
5374                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5375                         if (conf->hold_list.next == conf->last_hold)
5376                                 conf->bypass_count++;
5377                         else {
5378                                 conf->last_hold = conf->hold_list.next;
5379                                 conf->bypass_count -= conf->bypass_threshold;
5380                                 if (conf->bypass_count < 0)
5381                                         conf->bypass_count = 0;
5382                         }
5383                 }
5384         } else if (!list_empty(&conf->hold_list) &&
5385                    ((conf->bypass_threshold &&
5386                      conf->bypass_count > conf->bypass_threshold) ||
5387                     atomic_read(&conf->pending_full_writes) == 0)) {
5388
5389                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5390                         if (conf->worker_cnt_per_group == 0 ||
5391                             group == ANY_GROUP ||
5392                             !cpu_online(tmp->cpu) ||
5393                             cpu_to_group(tmp->cpu) == group) {
5394                                 sh = tmp;
5395                                 break;
5396                         }
5397                 }
5398
5399                 if (sh) {
5400                         conf->bypass_count -= conf->bypass_threshold;
5401                         if (conf->bypass_count < 0)
5402                                 conf->bypass_count = 0;
5403                 }
5404                 wg = NULL;
5405         }
5406
5407         if (!sh) {
5408                 if (second_try)
5409                         return NULL;
5410                 second_try = true;
5411                 try_loprio = !try_loprio;
5412                 goto again;
5413         }
5414
5415         if (wg) {
5416                 wg->stripes_cnt--;
5417                 sh->group = NULL;
5418         }
5419         list_del_init(&sh->lru);
5420         BUG_ON(atomic_inc_return(&sh->count) != 1);
5421         return sh;
5422 }
5423
5424 struct raid5_plug_cb {
5425         struct blk_plug_cb      cb;
5426         struct list_head        list;
5427         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5428 };
5429
5430 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5431 {
5432         struct raid5_plug_cb *cb = container_of(
5433                 blk_cb, struct raid5_plug_cb, cb);
5434         struct stripe_head *sh;
5435         struct mddev *mddev = cb->cb.data;
5436         struct r5conf *conf = mddev->private;
5437         int cnt = 0;
5438         int hash;
5439
5440         if (cb->list.next && !list_empty(&cb->list)) {
5441                 spin_lock_irq(&conf->device_lock);
5442                 while (!list_empty(&cb->list)) {
5443                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5444                         list_del_init(&sh->lru);
5445                         /*
5446                          * avoid race release_stripe_plug() sees
5447                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5448                          * is still in our list
5449                          */
5450                         smp_mb__before_atomic();
5451                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5452                         /*
5453                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5454                          * case, the count is always > 1 here
5455                          */
5456                         hash = sh->hash_lock_index;
5457                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5458                         cnt++;
5459                 }
5460                 spin_unlock_irq(&conf->device_lock);
5461         }
5462         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5463                                      NR_STRIPE_HASH_LOCKS);
5464         if (mddev->queue)
5465                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5466         kfree(cb);
5467 }
5468
5469 static void release_stripe_plug(struct mddev *mddev,
5470                                 struct stripe_head *sh)
5471 {
5472         struct blk_plug_cb *blk_cb = blk_check_plugged(
5473                 raid5_unplug, mddev,
5474                 sizeof(struct raid5_plug_cb));
5475         struct raid5_plug_cb *cb;
5476
5477         if (!blk_cb) {
5478                 raid5_release_stripe(sh);
5479                 return;
5480         }
5481
5482         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5483
5484         if (cb->list.next == NULL) {
5485                 int i;
5486                 INIT_LIST_HEAD(&cb->list);
5487                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5488                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5489         }
5490
5491         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5492                 list_add_tail(&sh->lru, &cb->list);
5493         else
5494                 raid5_release_stripe(sh);
5495 }
5496
5497 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5498 {
5499         struct r5conf *conf = mddev->private;
5500         sector_t logical_sector, last_sector;
5501         struct stripe_head *sh;
5502         int stripe_sectors;
5503
5504         if (mddev->reshape_position != MaxSector)
5505                 /* Skip discard while reshape is happening */
5506                 return;
5507
5508         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5509         last_sector = bio_end_sector(bi);
5510
5511         bi->bi_next = NULL;
5512
5513         stripe_sectors = conf->chunk_sectors *
5514                 (conf->raid_disks - conf->max_degraded);
5515         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5516                                                stripe_sectors);
5517         sector_div(last_sector, stripe_sectors);
5518
5519         logical_sector *= conf->chunk_sectors;
5520         last_sector *= conf->chunk_sectors;
5521
5522         for (; logical_sector < last_sector;
5523              logical_sector += STRIPE_SECTORS) {
5524                 DEFINE_WAIT(w);
5525                 int d;
5526         again:
5527                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5528                 prepare_to_wait(&conf->wait_for_overlap, &w,
5529                                 TASK_UNINTERRUPTIBLE);
5530                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5531                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5532                         raid5_release_stripe(sh);
5533                         schedule();
5534                         goto again;
5535                 }
5536                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5537                 spin_lock_irq(&sh->stripe_lock);
5538                 for (d = 0; d < conf->raid_disks; d++) {
5539                         if (d == sh->pd_idx || d == sh->qd_idx)
5540                                 continue;
5541                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5542                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5543                                 spin_unlock_irq(&sh->stripe_lock);
5544                                 raid5_release_stripe(sh);
5545                                 schedule();
5546                                 goto again;
5547                         }
5548                 }
5549                 set_bit(STRIPE_DISCARD, &sh->state);
5550                 finish_wait(&conf->wait_for_overlap, &w);
5551                 sh->overwrite_disks = 0;
5552                 for (d = 0; d < conf->raid_disks; d++) {
5553                         if (d == sh->pd_idx || d == sh->qd_idx)
5554                                 continue;
5555                         sh->dev[d].towrite = bi;
5556                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5557                         bio_inc_remaining(bi);
5558                         md_write_inc(mddev, bi);
5559                         sh->overwrite_disks++;
5560                 }
5561                 spin_unlock_irq(&sh->stripe_lock);
5562                 if (conf->mddev->bitmap) {
5563                         for (d = 0;
5564                              d < conf->raid_disks - conf->max_degraded;
5565                              d++)
5566                                 md_bitmap_startwrite(mddev->bitmap,
5567                                                      sh->sector,
5568                                                      STRIPE_SECTORS,
5569                                                      0);
5570                         sh->bm_seq = conf->seq_flush + 1;
5571                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5572                 }
5573
5574                 set_bit(STRIPE_HANDLE, &sh->state);
5575                 clear_bit(STRIPE_DELAYED, &sh->state);
5576                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5577                         atomic_inc(&conf->preread_active_stripes);
5578                 release_stripe_plug(mddev, sh);
5579         }
5580
5581         bio_endio(bi);
5582 }
5583
5584 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5585 {
5586         struct r5conf *conf = mddev->private;
5587         int dd_idx;
5588         sector_t new_sector;
5589         sector_t logical_sector, last_sector;
5590         struct stripe_head *sh;
5591         const int rw = bio_data_dir(bi);
5592         DEFINE_WAIT(w);
5593         bool do_prepare;
5594         bool do_flush = false;
5595
5596         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5597                 int ret = log_handle_flush_request(conf, bi);
5598
5599                 if (ret == 0)
5600                         return true;
5601                 if (ret == -ENODEV) {
5602                         if (md_flush_request(mddev, bi))
5603                                 return true;
5604                 }
5605                 /* ret == -EAGAIN, fallback */
5606                 /*
5607                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5608                  * we need to flush journal device
5609                  */
5610                 do_flush = bi->bi_opf & REQ_PREFLUSH;
5611         }
5612
5613         if (!md_write_start(mddev, bi))
5614                 return false;
5615         /*
5616          * If array is degraded, better not do chunk aligned read because
5617          * later we might have to read it again in order to reconstruct
5618          * data on failed drives.
5619          */
5620         if (rw == READ && mddev->degraded == 0 &&
5621             mddev->reshape_position == MaxSector) {
5622                 bi = chunk_aligned_read(mddev, bi);
5623                 if (!bi)
5624                         return true;
5625         }
5626
5627         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5628                 make_discard_request(mddev, bi);
5629                 md_write_end(mddev);
5630                 return true;
5631         }
5632
5633         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5634         last_sector = bio_end_sector(bi);
5635         bi->bi_next = NULL;
5636
5637         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5638         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5639                 int previous;
5640                 int seq;
5641
5642                 do_prepare = false;
5643         retry:
5644                 seq = read_seqcount_begin(&conf->gen_lock);
5645                 previous = 0;
5646                 if (do_prepare)
5647                         prepare_to_wait(&conf->wait_for_overlap, &w,
5648                                 TASK_UNINTERRUPTIBLE);
5649                 if (unlikely(conf->reshape_progress != MaxSector)) {
5650                         /* spinlock is needed as reshape_progress may be
5651                          * 64bit on a 32bit platform, and so it might be
5652                          * possible to see a half-updated value
5653                          * Of course reshape_progress could change after
5654                          * the lock is dropped, so once we get a reference
5655                          * to the stripe that we think it is, we will have
5656                          * to check again.
5657                          */
5658                         spin_lock_irq(&conf->device_lock);
5659                         if (mddev->reshape_backwards
5660                             ? logical_sector < conf->reshape_progress
5661                             : logical_sector >= conf->reshape_progress) {
5662                                 previous = 1;
5663                         } else {
5664                                 if (mddev->reshape_backwards
5665                                     ? logical_sector < conf->reshape_safe
5666                                     : logical_sector >= conf->reshape_safe) {
5667                                         spin_unlock_irq(&conf->device_lock);
5668                                         schedule();
5669                                         do_prepare = true;
5670                                         goto retry;
5671                                 }
5672                         }
5673                         spin_unlock_irq(&conf->device_lock);
5674                 }
5675
5676                 new_sector = raid5_compute_sector(conf, logical_sector,
5677                                                   previous,
5678                                                   &dd_idx, NULL);
5679                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5680                         (unsigned long long)new_sector,
5681                         (unsigned long long)logical_sector);
5682
5683                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5684                                        (bi->bi_opf & REQ_RAHEAD), 0);
5685                 if (sh) {
5686                         if (unlikely(previous)) {
5687                                 /* expansion might have moved on while waiting for a
5688                                  * stripe, so we must do the range check again.
5689                                  * Expansion could still move past after this
5690                                  * test, but as we are holding a reference to
5691                                  * 'sh', we know that if that happens,
5692                                  *  STRIPE_EXPANDING will get set and the expansion
5693                                  * won't proceed until we finish with the stripe.
5694                                  */
5695                                 int must_retry = 0;
5696                                 spin_lock_irq(&conf->device_lock);
5697                                 if (mddev->reshape_backwards
5698                                     ? logical_sector >= conf->reshape_progress
5699                                     : logical_sector < conf->reshape_progress)
5700                                         /* mismatch, need to try again */
5701                                         must_retry = 1;
5702                                 spin_unlock_irq(&conf->device_lock);
5703                                 if (must_retry) {
5704                                         raid5_release_stripe(sh);
5705                                         schedule();
5706                                         do_prepare = true;
5707                                         goto retry;
5708                                 }
5709                         }
5710                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5711                                 /* Might have got the wrong stripe_head
5712                                  * by accident
5713                                  */
5714                                 raid5_release_stripe(sh);
5715                                 goto retry;
5716                         }
5717
5718                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5719                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5720                                 /* Stripe is busy expanding or
5721                                  * add failed due to overlap.  Flush everything
5722                                  * and wait a while
5723                                  */
5724                                 md_wakeup_thread(mddev->thread);
5725                                 raid5_release_stripe(sh);
5726                                 schedule();
5727                                 do_prepare = true;
5728                                 goto retry;
5729                         }
5730                         if (do_flush) {
5731                                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5732                                 /* we only need flush for one stripe */
5733                                 do_flush = false;
5734                         }
5735
5736                         if (!sh->batch_head || sh == sh->batch_head)
5737                                 set_bit(STRIPE_HANDLE, &sh->state);
5738                         clear_bit(STRIPE_DELAYED, &sh->state);
5739                         if ((!sh->batch_head || sh == sh->batch_head) &&
5740                             (bi->bi_opf & REQ_SYNC) &&
5741                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5742                                 atomic_inc(&conf->preread_active_stripes);
5743                         release_stripe_plug(mddev, sh);
5744                 } else {
5745                         /* cannot get stripe for read-ahead, just give-up */
5746                         bi->bi_status = BLK_STS_IOERR;
5747                         break;
5748                 }
5749         }
5750         finish_wait(&conf->wait_for_overlap, &w);
5751
5752         if (rw == WRITE)
5753                 md_write_end(mddev);
5754         bio_endio(bi);
5755         return true;
5756 }
5757
5758 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5759
5760 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5761 {
5762         /* reshaping is quite different to recovery/resync so it is
5763          * handled quite separately ... here.
5764          *
5765          * On each call to sync_request, we gather one chunk worth of
5766          * destination stripes and flag them as expanding.
5767          * Then we find all the source stripes and request reads.
5768          * As the reads complete, handle_stripe will copy the data
5769          * into the destination stripe and release that stripe.
5770          */
5771         struct r5conf *conf = mddev->private;
5772         struct stripe_head *sh;
5773         struct md_rdev *rdev;
5774         sector_t first_sector, last_sector;
5775         int raid_disks = conf->previous_raid_disks;
5776         int data_disks = raid_disks - conf->max_degraded;
5777         int new_data_disks = conf->raid_disks - conf->max_degraded;
5778         int i;
5779         int dd_idx;
5780         sector_t writepos, readpos, safepos;
5781         sector_t stripe_addr;
5782         int reshape_sectors;
5783         struct list_head stripes;
5784         sector_t retn;
5785
5786         if (sector_nr == 0) {
5787                 /* If restarting in the middle, skip the initial sectors */
5788                 if (mddev->reshape_backwards &&
5789                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5790                         sector_nr = raid5_size(mddev, 0, 0)
5791                                 - conf->reshape_progress;
5792                 } else if (mddev->reshape_backwards &&
5793                            conf->reshape_progress == MaxSector) {
5794                         /* shouldn't happen, but just in case, finish up.*/
5795                         sector_nr = MaxSector;
5796                 } else if (!mddev->reshape_backwards &&
5797                            conf->reshape_progress > 0)
5798                         sector_nr = conf->reshape_progress;
5799                 sector_div(sector_nr, new_data_disks);
5800                 if (sector_nr) {
5801                         mddev->curr_resync_completed = sector_nr;
5802                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5803                         *skipped = 1;
5804                         retn = sector_nr;
5805                         goto finish;
5806                 }
5807         }
5808
5809         /* We need to process a full chunk at a time.
5810          * If old and new chunk sizes differ, we need to process the
5811          * largest of these
5812          */
5813
5814         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5815
5816         /* We update the metadata at least every 10 seconds, or when
5817          * the data about to be copied would over-write the source of
5818          * the data at the front of the range.  i.e. one new_stripe
5819          * along from reshape_progress new_maps to after where
5820          * reshape_safe old_maps to
5821          */
5822         writepos = conf->reshape_progress;
5823         sector_div(writepos, new_data_disks);
5824         readpos = conf->reshape_progress;
5825         sector_div(readpos, data_disks);
5826         safepos = conf->reshape_safe;
5827         sector_div(safepos, data_disks);
5828         if (mddev->reshape_backwards) {
5829                 BUG_ON(writepos < reshape_sectors);
5830                 writepos -= reshape_sectors;
5831                 readpos += reshape_sectors;
5832                 safepos += reshape_sectors;
5833         } else {
5834                 writepos += reshape_sectors;
5835                 /* readpos and safepos are worst-case calculations.
5836                  * A negative number is overly pessimistic, and causes
5837                  * obvious problems for unsigned storage.  So clip to 0.
5838                  */
5839                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5840                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5841         }
5842
5843         /* Having calculated the 'writepos' possibly use it
5844          * to set 'stripe_addr' which is where we will write to.
5845          */
5846         if (mddev->reshape_backwards) {
5847                 BUG_ON(conf->reshape_progress == 0);
5848                 stripe_addr = writepos;
5849                 BUG_ON((mddev->dev_sectors &
5850                         ~((sector_t)reshape_sectors - 1))
5851                        - reshape_sectors - stripe_addr
5852                        != sector_nr);
5853         } else {
5854                 BUG_ON(writepos != sector_nr + reshape_sectors);
5855                 stripe_addr = sector_nr;
5856         }
5857
5858         /* 'writepos' is the most advanced device address we might write.
5859          * 'readpos' is the least advanced device address we might read.
5860          * 'safepos' is the least address recorded in the metadata as having
5861          *     been reshaped.
5862          * If there is a min_offset_diff, these are adjusted either by
5863          * increasing the safepos/readpos if diff is negative, or
5864          * increasing writepos if diff is positive.
5865          * If 'readpos' is then behind 'writepos', there is no way that we can
5866          * ensure safety in the face of a crash - that must be done by userspace
5867          * making a backup of the data.  So in that case there is no particular
5868          * rush to update metadata.
5869          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5870          * update the metadata to advance 'safepos' to match 'readpos' so that
5871          * we can be safe in the event of a crash.
5872          * So we insist on updating metadata if safepos is behind writepos and
5873          * readpos is beyond writepos.
5874          * In any case, update the metadata every 10 seconds.
5875          * Maybe that number should be configurable, but I'm not sure it is
5876          * worth it.... maybe it could be a multiple of safemode_delay???
5877          */
5878         if (conf->min_offset_diff < 0) {
5879                 safepos += -conf->min_offset_diff;
5880                 readpos += -conf->min_offset_diff;
5881         } else
5882                 writepos += conf->min_offset_diff;
5883
5884         if ((mddev->reshape_backwards
5885              ? (safepos > writepos && readpos < writepos)
5886              : (safepos < writepos && readpos > writepos)) ||
5887             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5888                 /* Cannot proceed until we've updated the superblock... */
5889                 wait_event(conf->wait_for_overlap,
5890                            atomic_read(&conf->reshape_stripes)==0
5891                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5892                 if (atomic_read(&conf->reshape_stripes) != 0)
5893                         return 0;
5894                 mddev->reshape_position = conf->reshape_progress;
5895                 mddev->curr_resync_completed = sector_nr;
5896                 if (!mddev->reshape_backwards)
5897                         /* Can update recovery_offset */
5898                         rdev_for_each(rdev, mddev)
5899                                 if (rdev->raid_disk >= 0 &&
5900                                     !test_bit(Journal, &rdev->flags) &&
5901                                     !test_bit(In_sync, &rdev->flags) &&
5902                                     rdev->recovery_offset < sector_nr)
5903                                         rdev->recovery_offset = sector_nr;
5904
5905                 conf->reshape_checkpoint = jiffies;
5906                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5907                 md_wakeup_thread(mddev->thread);
5908                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5909                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5910                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5911                         return 0;
5912                 spin_lock_irq(&conf->device_lock);
5913                 conf->reshape_safe = mddev->reshape_position;
5914                 spin_unlock_irq(&conf->device_lock);
5915                 wake_up(&conf->wait_for_overlap);
5916                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5917         }
5918
5919         INIT_LIST_HEAD(&stripes);
5920         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5921                 int j;
5922                 int skipped_disk = 0;
5923                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5924                 set_bit(STRIPE_EXPANDING, &sh->state);
5925                 atomic_inc(&conf->reshape_stripes);
5926                 /* If any of this stripe is beyond the end of the old
5927                  * array, then we need to zero those blocks
5928                  */
5929                 for (j=sh->disks; j--;) {
5930                         sector_t s;
5931                         if (j == sh->pd_idx)
5932                                 continue;
5933                         if (conf->level == 6 &&
5934                             j == sh->qd_idx)
5935                                 continue;
5936                         s = raid5_compute_blocknr(sh, j, 0);
5937                         if (s < raid5_size(mddev, 0, 0)) {
5938                                 skipped_disk = 1;
5939                                 continue;
5940                         }
5941                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5942                         set_bit(R5_Expanded, &sh->dev[j].flags);
5943                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5944                 }
5945                 if (!skipped_disk) {
5946                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5947                         set_bit(STRIPE_HANDLE, &sh->state);
5948                 }
5949                 list_add(&sh->lru, &stripes);
5950         }
5951         spin_lock_irq(&conf->device_lock);
5952         if (mddev->reshape_backwards)
5953                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5954         else
5955                 conf->reshape_progress += reshape_sectors * new_data_disks;
5956         spin_unlock_irq(&conf->device_lock);
5957         /* Ok, those stripe are ready. We can start scheduling
5958          * reads on the source stripes.
5959          * The source stripes are determined by mapping the first and last
5960          * block on the destination stripes.
5961          */
5962         first_sector =
5963                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5964                                      1, &dd_idx, NULL);
5965         last_sector =
5966                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5967                                             * new_data_disks - 1),
5968                                      1, &dd_idx, NULL);
5969         if (last_sector >= mddev->dev_sectors)
5970                 last_sector = mddev->dev_sectors - 1;
5971         while (first_sector <= last_sector) {
5972                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5973                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5974                 set_bit(STRIPE_HANDLE, &sh->state);
5975                 raid5_release_stripe(sh);
5976                 first_sector += STRIPE_SECTORS;
5977         }
5978         /* Now that the sources are clearly marked, we can release
5979          * the destination stripes
5980          */
5981         while (!list_empty(&stripes)) {
5982                 sh = list_entry(stripes.next, struct stripe_head, lru);
5983                 list_del_init(&sh->lru);
5984                 raid5_release_stripe(sh);
5985         }
5986         /* If this takes us to the resync_max point where we have to pause,
5987          * then we need to write out the superblock.
5988          */
5989         sector_nr += reshape_sectors;
5990         retn = reshape_sectors;
5991 finish:
5992         if (mddev->curr_resync_completed > mddev->resync_max ||
5993             (sector_nr - mddev->curr_resync_completed) * 2
5994             >= mddev->resync_max - mddev->curr_resync_completed) {
5995                 /* Cannot proceed until we've updated the superblock... */
5996                 wait_event(conf->wait_for_overlap,
5997                            atomic_read(&conf->reshape_stripes) == 0
5998                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5999                 if (atomic_read(&conf->reshape_stripes) != 0)
6000                         goto ret;
6001                 mddev->reshape_position = conf->reshape_progress;
6002                 mddev->curr_resync_completed = sector_nr;
6003                 if (!mddev->reshape_backwards)
6004                         /* Can update recovery_offset */
6005                         rdev_for_each(rdev, mddev)
6006                                 if (rdev->raid_disk >= 0 &&
6007                                     !test_bit(Journal, &rdev->flags) &&
6008                                     !test_bit(In_sync, &rdev->flags) &&
6009                                     rdev->recovery_offset < sector_nr)
6010                                         rdev->recovery_offset = sector_nr;
6011                 conf->reshape_checkpoint = jiffies;
6012                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6013                 md_wakeup_thread(mddev->thread);
6014                 wait_event(mddev->sb_wait,
6015                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6016                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6017                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6018                         goto ret;
6019                 spin_lock_irq(&conf->device_lock);
6020                 conf->reshape_safe = mddev->reshape_position;
6021                 spin_unlock_irq(&conf->device_lock);
6022                 wake_up(&conf->wait_for_overlap);
6023                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6024         }
6025 ret:
6026         return retn;
6027 }
6028
6029 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6030                                           int *skipped)
6031 {
6032         struct r5conf *conf = mddev->private;
6033         struct stripe_head *sh;
6034         sector_t max_sector = mddev->dev_sectors;
6035         sector_t sync_blocks;
6036         int still_degraded = 0;
6037         int i;
6038
6039         if (sector_nr >= max_sector) {
6040                 /* just being told to finish up .. nothing much to do */
6041
6042                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6043                         end_reshape(conf);
6044                         return 0;
6045                 }
6046
6047                 if (mddev->curr_resync < max_sector) /* aborted */
6048                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6049                                            &sync_blocks, 1);
6050                 else /* completed sync */
6051                         conf->fullsync = 0;
6052                 md_bitmap_close_sync(mddev->bitmap);
6053
6054                 return 0;
6055         }
6056
6057         /* Allow raid5_quiesce to complete */
6058         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6059
6060         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6061                 return reshape_request(mddev, sector_nr, skipped);
6062
6063         /* No need to check resync_max as we never do more than one
6064          * stripe, and as resync_max will always be on a chunk boundary,
6065          * if the check in md_do_sync didn't fire, there is no chance
6066          * of overstepping resync_max here
6067          */
6068
6069         /* if there is too many failed drives and we are trying
6070          * to resync, then assert that we are finished, because there is
6071          * nothing we can do.
6072          */
6073         if (mddev->degraded >= conf->max_degraded &&
6074             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6075                 sector_t rv = mddev->dev_sectors - sector_nr;
6076                 *skipped = 1;
6077                 return rv;
6078         }
6079         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6080             !conf->fullsync &&
6081             !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6082             sync_blocks >= STRIPE_SECTORS) {
6083                 /* we can skip this block, and probably more */
6084                 sync_blocks /= STRIPE_SECTORS;
6085                 *skipped = 1;
6086                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6087         }
6088
6089         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6090
6091         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6092         if (sh == NULL) {
6093                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6094                 /* make sure we don't swamp the stripe cache if someone else
6095                  * is trying to get access
6096                  */
6097                 schedule_timeout_uninterruptible(1);
6098         }
6099         /* Need to check if array will still be degraded after recovery/resync
6100          * Note in case of > 1 drive failures it's possible we're rebuilding
6101          * one drive while leaving another faulty drive in array.
6102          */
6103         rcu_read_lock();
6104         for (i = 0; i < conf->raid_disks; i++) {
6105                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6106
6107                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6108                         still_degraded = 1;
6109         }
6110         rcu_read_unlock();
6111
6112         md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6113
6114         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6115         set_bit(STRIPE_HANDLE, &sh->state);
6116
6117         raid5_release_stripe(sh);
6118
6119         return STRIPE_SECTORS;
6120 }
6121
6122 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6123                                unsigned int offset)
6124 {
6125         /* We may not be able to submit a whole bio at once as there
6126          * may not be enough stripe_heads available.
6127          * We cannot pre-allocate enough stripe_heads as we may need
6128          * more than exist in the cache (if we allow ever large chunks).
6129          * So we do one stripe head at a time and record in
6130          * ->bi_hw_segments how many have been done.
6131          *
6132          * We *know* that this entire raid_bio is in one chunk, so
6133          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6134          */
6135         struct stripe_head *sh;
6136         int dd_idx;
6137         sector_t sector, logical_sector, last_sector;
6138         int scnt = 0;
6139         int handled = 0;
6140
6141         logical_sector = raid_bio->bi_iter.bi_sector &
6142                 ~((sector_t)STRIPE_SECTORS-1);
6143         sector = raid5_compute_sector(conf, logical_sector,
6144                                       0, &dd_idx, NULL);
6145         last_sector = bio_end_sector(raid_bio);
6146
6147         for (; logical_sector < last_sector;
6148              logical_sector += STRIPE_SECTORS,
6149                      sector += STRIPE_SECTORS,
6150                      scnt++) {
6151
6152                 if (scnt < offset)
6153                         /* already done this stripe */
6154                         continue;
6155
6156                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6157
6158                 if (!sh) {
6159                         /* failed to get a stripe - must wait */
6160                         conf->retry_read_aligned = raid_bio;
6161                         conf->retry_read_offset = scnt;
6162                         return handled;
6163                 }
6164
6165                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6166                         raid5_release_stripe(sh);
6167                         conf->retry_read_aligned = raid_bio;
6168                         conf->retry_read_offset = scnt;
6169                         return handled;
6170                 }
6171
6172                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6173                 handle_stripe(sh);
6174                 raid5_release_stripe(sh);
6175                 handled++;
6176         }
6177
6178         bio_endio(raid_bio);
6179
6180         if (atomic_dec_and_test(&conf->active_aligned_reads))
6181                 wake_up(&conf->wait_for_quiescent);
6182         return handled;
6183 }
6184
6185 static int handle_active_stripes(struct r5conf *conf, int group,
6186                                  struct r5worker *worker,
6187                                  struct list_head *temp_inactive_list)
6188                 __releases(&conf->device_lock)
6189                 __acquires(&conf->device_lock)
6190 {
6191         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6192         int i, batch_size = 0, hash;
6193         bool release_inactive = false;
6194
6195         while (batch_size < MAX_STRIPE_BATCH &&
6196                         (sh = __get_priority_stripe(conf, group)) != NULL)
6197                 batch[batch_size++] = sh;
6198
6199         if (batch_size == 0) {
6200                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6201                         if (!list_empty(temp_inactive_list + i))
6202                                 break;
6203                 if (i == NR_STRIPE_HASH_LOCKS) {
6204                         spin_unlock_irq(&conf->device_lock);
6205                         log_flush_stripe_to_raid(conf);
6206                         spin_lock_irq(&conf->device_lock);
6207                         return batch_size;
6208                 }
6209                 release_inactive = true;
6210         }
6211         spin_unlock_irq(&conf->device_lock);
6212
6213         release_inactive_stripe_list(conf, temp_inactive_list,
6214                                      NR_STRIPE_HASH_LOCKS);
6215
6216         r5l_flush_stripe_to_raid(conf->log);
6217         if (release_inactive) {
6218                 spin_lock_irq(&conf->device_lock);
6219                 return 0;
6220         }
6221
6222         for (i = 0; i < batch_size; i++)
6223                 handle_stripe(batch[i]);
6224         log_write_stripe_run(conf);
6225
6226         cond_resched();
6227
6228         spin_lock_irq(&conf->device_lock);
6229         for (i = 0; i < batch_size; i++) {
6230                 hash = batch[i]->hash_lock_index;
6231                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6232         }
6233         return batch_size;
6234 }
6235
6236 static void raid5_do_work(struct work_struct *work)
6237 {
6238         struct r5worker *worker = container_of(work, struct r5worker, work);
6239         struct r5worker_group *group = worker->group;
6240         struct r5conf *conf = group->conf;
6241         struct mddev *mddev = conf->mddev;
6242         int group_id = group - conf->worker_groups;
6243         int handled;
6244         struct blk_plug plug;
6245
6246         pr_debug("+++ raid5worker active\n");
6247
6248         blk_start_plug(&plug);
6249         handled = 0;
6250         spin_lock_irq(&conf->device_lock);
6251         while (1) {
6252                 int batch_size, released;
6253
6254                 released = release_stripe_list(conf, worker->temp_inactive_list);
6255
6256                 batch_size = handle_active_stripes(conf, group_id, worker,
6257                                                    worker->temp_inactive_list);
6258                 worker->working = false;
6259                 if (!batch_size && !released)
6260                         break;
6261                 handled += batch_size;
6262                 wait_event_lock_irq(mddev->sb_wait,
6263                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6264                         conf->device_lock);
6265         }
6266         pr_debug("%d stripes handled\n", handled);
6267
6268         spin_unlock_irq(&conf->device_lock);
6269
6270         flush_deferred_bios(conf);
6271
6272         r5l_flush_stripe_to_raid(conf->log);
6273
6274         async_tx_issue_pending_all();
6275         blk_finish_plug(&plug);
6276
6277         pr_debug("--- raid5worker inactive\n");
6278 }
6279
6280 /*
6281  * This is our raid5 kernel thread.
6282  *
6283  * We scan the hash table for stripes which can be handled now.
6284  * During the scan, completed stripes are saved for us by the interrupt
6285  * handler, so that they will not have to wait for our next wakeup.
6286  */
6287 static void raid5d(struct md_thread *thread)
6288 {
6289         struct mddev *mddev = thread->mddev;
6290         struct r5conf *conf = mddev->private;
6291         int handled;
6292         struct blk_plug plug;
6293
6294         pr_debug("+++ raid5d active\n");
6295
6296         md_check_recovery(mddev);
6297
6298         blk_start_plug(&plug);
6299         handled = 0;
6300         spin_lock_irq(&conf->device_lock);
6301         while (1) {
6302                 struct bio *bio;
6303                 int batch_size, released;
6304                 unsigned int offset;
6305
6306                 released = release_stripe_list(conf, conf->temp_inactive_list);
6307                 if (released)
6308                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6309
6310                 if (
6311                     !list_empty(&conf->bitmap_list)) {
6312                         /* Now is a good time to flush some bitmap updates */
6313                         conf->seq_flush++;
6314                         spin_unlock_irq(&conf->device_lock);
6315                         md_bitmap_unplug(mddev->bitmap);
6316                         spin_lock_irq(&conf->device_lock);
6317                         conf->seq_write = conf->seq_flush;
6318                         activate_bit_delay(conf, conf->temp_inactive_list);
6319                 }
6320                 raid5_activate_delayed(conf);
6321
6322                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6323                         int ok;
6324                         spin_unlock_irq(&conf->device_lock);
6325                         ok = retry_aligned_read(conf, bio, offset);
6326                         spin_lock_irq(&conf->device_lock);
6327                         if (!ok)
6328                                 break;
6329                         handled++;
6330                 }
6331
6332                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6333                                                    conf->temp_inactive_list);
6334                 if (!batch_size && !released)
6335                         break;
6336                 handled += batch_size;
6337
6338                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6339                         spin_unlock_irq(&conf->device_lock);
6340                         md_check_recovery(mddev);
6341                         spin_lock_irq(&conf->device_lock);
6342                 }
6343         }
6344         pr_debug("%d stripes handled\n", handled);
6345
6346         spin_unlock_irq(&conf->device_lock);
6347         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6348             mutex_trylock(&conf->cache_size_mutex)) {
6349                 grow_one_stripe(conf, __GFP_NOWARN);
6350                 /* Set flag even if allocation failed.  This helps
6351                  * slow down allocation requests when mem is short
6352                  */
6353                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6354                 mutex_unlock(&conf->cache_size_mutex);
6355         }
6356
6357         flush_deferred_bios(conf);
6358
6359         r5l_flush_stripe_to_raid(conf->log);
6360
6361         async_tx_issue_pending_all();
6362         blk_finish_plug(&plug);
6363
6364         pr_debug("--- raid5d inactive\n");
6365 }
6366
6367 static ssize_t
6368 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6369 {
6370         struct r5conf *conf;
6371         int ret = 0;
6372         spin_lock(&mddev->lock);
6373         conf = mddev->private;
6374         if (conf)
6375                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6376         spin_unlock(&mddev->lock);
6377         return ret;
6378 }
6379
6380 int
6381 raid5_set_cache_size(struct mddev *mddev, int size)
6382 {
6383         int result = 0;
6384         struct r5conf *conf = mddev->private;
6385
6386         if (size <= 16 || size > 32768)
6387                 return -EINVAL;
6388
6389         conf->min_nr_stripes = size;
6390         mutex_lock(&conf->cache_size_mutex);
6391         while (size < conf->max_nr_stripes &&
6392                drop_one_stripe(conf))
6393                 ;
6394         mutex_unlock(&conf->cache_size_mutex);
6395
6396         md_allow_write(mddev);
6397
6398         mutex_lock(&conf->cache_size_mutex);
6399         while (size > conf->max_nr_stripes)
6400                 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6401                         conf->min_nr_stripes = conf->max_nr_stripes;
6402                         result = -ENOMEM;
6403                         break;
6404                 }
6405         mutex_unlock(&conf->cache_size_mutex);
6406
6407         return result;
6408 }
6409 EXPORT_SYMBOL(raid5_set_cache_size);
6410
6411 static ssize_t
6412 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6413 {
6414         struct r5conf *conf;
6415         unsigned long new;
6416         int err;
6417
6418         if (len >= PAGE_SIZE)
6419                 return -EINVAL;
6420         if (kstrtoul(page, 10, &new))
6421                 return -EINVAL;
6422         err = mddev_lock(mddev);
6423         if (err)
6424                 return err;
6425         conf = mddev->private;
6426         if (!conf)
6427                 err = -ENODEV;
6428         else
6429                 err = raid5_set_cache_size(mddev, new);
6430         mddev_unlock(mddev);
6431
6432         return err ?: len;
6433 }
6434
6435 static struct md_sysfs_entry
6436 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6437                                 raid5_show_stripe_cache_size,
6438                                 raid5_store_stripe_cache_size);
6439
6440 static ssize_t
6441 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6442 {
6443         struct r5conf *conf = mddev->private;
6444         if (conf)
6445                 return sprintf(page, "%d\n", conf->rmw_level);
6446         else
6447                 return 0;
6448 }
6449
6450 static ssize_t
6451 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6452 {
6453         struct r5conf *conf = mddev->private;
6454         unsigned long new;
6455
6456         if (!conf)
6457                 return -ENODEV;
6458
6459         if (len >= PAGE_SIZE)
6460                 return -EINVAL;
6461
6462         if (kstrtoul(page, 10, &new))
6463                 return -EINVAL;
6464
6465         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6466                 return -EINVAL;
6467
6468         if (new != PARITY_DISABLE_RMW &&
6469             new != PARITY_ENABLE_RMW &&
6470             new != PARITY_PREFER_RMW)
6471                 return -EINVAL;
6472
6473         conf->rmw_level = new;
6474         return len;
6475 }
6476
6477 static struct md_sysfs_entry
6478 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6479                          raid5_show_rmw_level,
6480                          raid5_store_rmw_level);
6481
6482
6483 static ssize_t
6484 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6485 {
6486         struct r5conf *conf;
6487         int ret = 0;
6488         spin_lock(&mddev->lock);
6489         conf = mddev->private;
6490         if (conf)
6491                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6492         spin_unlock(&mddev->lock);
6493         return ret;
6494 }
6495
6496 static ssize_t
6497 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6498 {
6499         struct r5conf *conf;
6500         unsigned long new;
6501         int err;
6502
6503         if (len >= PAGE_SIZE)
6504                 return -EINVAL;
6505         if (kstrtoul(page, 10, &new))
6506                 return -EINVAL;
6507
6508         err = mddev_lock(mddev);
6509         if (err)
6510                 return err;
6511         conf = mddev->private;
6512         if (!conf)
6513                 err = -ENODEV;
6514         else if (new > conf->min_nr_stripes)
6515                 err = -EINVAL;
6516         else
6517                 conf->bypass_threshold = new;
6518         mddev_unlock(mddev);
6519         return err ?: len;
6520 }
6521
6522 static struct md_sysfs_entry
6523 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6524                                         S_IRUGO | S_IWUSR,
6525                                         raid5_show_preread_threshold,
6526                                         raid5_store_preread_threshold);
6527
6528 static ssize_t
6529 raid5_show_skip_copy(struct mddev *mddev, char *page)
6530 {
6531         struct r5conf *conf;
6532         int ret = 0;
6533         spin_lock(&mddev->lock);
6534         conf = mddev->private;
6535         if (conf)
6536                 ret = sprintf(page, "%d\n", conf->skip_copy);
6537         spin_unlock(&mddev->lock);
6538         return ret;
6539 }
6540
6541 static ssize_t
6542 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6543 {
6544         struct r5conf *conf;
6545         unsigned long new;
6546         int err;
6547
6548         if (len >= PAGE_SIZE)
6549                 return -EINVAL;
6550         if (kstrtoul(page, 10, &new))
6551                 return -EINVAL;
6552         new = !!new;
6553
6554         err = mddev_lock(mddev);
6555         if (err)
6556                 return err;
6557         conf = mddev->private;
6558         if (!conf)
6559                 err = -ENODEV;
6560         else if (new != conf->skip_copy) {
6561                 mddev_suspend(mddev);
6562                 conf->skip_copy = new;
6563                 if (new)
6564                         mddev->queue->backing_dev_info->capabilities |=
6565                                 BDI_CAP_STABLE_WRITES;
6566                 else
6567                         mddev->queue->backing_dev_info->capabilities &=
6568                                 ~BDI_CAP_STABLE_WRITES;
6569                 mddev_resume(mddev);
6570         }
6571         mddev_unlock(mddev);
6572         return err ?: len;
6573 }
6574
6575 static struct md_sysfs_entry
6576 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6577                                         raid5_show_skip_copy,
6578                                         raid5_store_skip_copy);
6579
6580 static ssize_t
6581 stripe_cache_active_show(struct mddev *mddev, char *page)
6582 {
6583         struct r5conf *conf = mddev->private;
6584         if (conf)
6585                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6586         else
6587                 return 0;
6588 }
6589
6590 static struct md_sysfs_entry
6591 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6592
6593 static ssize_t
6594 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6595 {
6596         struct r5conf *conf;
6597         int ret = 0;
6598         spin_lock(&mddev->lock);
6599         conf = mddev->private;
6600         if (conf)
6601                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6602         spin_unlock(&mddev->lock);
6603         return ret;
6604 }
6605
6606 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6607                                int *group_cnt,
6608                                struct r5worker_group **worker_groups);
6609 static ssize_t
6610 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6611 {
6612         struct r5conf *conf;
6613         unsigned int new;
6614         int err;
6615         struct r5worker_group *new_groups, *old_groups;
6616         int group_cnt;
6617
6618         if (len >= PAGE_SIZE)
6619                 return -EINVAL;
6620         if (kstrtouint(page, 10, &new))
6621                 return -EINVAL;
6622         /* 8192 should be big enough */
6623         if (new > 8192)
6624                 return -EINVAL;
6625
6626         err = mddev_lock(mddev);
6627         if (err)
6628                 return err;
6629         conf = mddev->private;
6630         if (!conf)
6631                 err = -ENODEV;
6632         else if (new != conf->worker_cnt_per_group) {
6633                 mddev_suspend(mddev);
6634
6635                 old_groups = conf->worker_groups;
6636                 if (old_groups)
6637                         flush_workqueue(raid5_wq);
6638
6639                 err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
6640                 if (!err) {
6641                         spin_lock_irq(&conf->device_lock);
6642                         conf->group_cnt = group_cnt;
6643                         conf->worker_cnt_per_group = new;
6644                         conf->worker_groups = new_groups;
6645                         spin_unlock_irq(&conf->device_lock);
6646
6647                         if (old_groups)
6648                                 kfree(old_groups[0].workers);
6649                         kfree(old_groups);
6650                 }
6651                 mddev_resume(mddev);
6652         }
6653         mddev_unlock(mddev);
6654
6655         return err ?: len;
6656 }
6657
6658 static struct md_sysfs_entry
6659 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6660                                 raid5_show_group_thread_cnt,
6661                                 raid5_store_group_thread_cnt);
6662
6663 static struct attribute *raid5_attrs[] =  {
6664         &raid5_stripecache_size.attr,
6665         &raid5_stripecache_active.attr,
6666         &raid5_preread_bypass_threshold.attr,
6667         &raid5_group_thread_cnt.attr,
6668         &raid5_skip_copy.attr,
6669         &raid5_rmw_level.attr,
6670         &r5c_journal_mode.attr,
6671         &ppl_write_hint.attr,
6672         NULL,
6673 };
6674 static struct attribute_group raid5_attrs_group = {
6675         .name = NULL,
6676         .attrs = raid5_attrs,
6677 };
6678
6679 static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
6680                                struct r5worker_group **worker_groups)
6681 {
6682         int i, j, k;
6683         ssize_t size;
6684         struct r5worker *workers;
6685
6686         if (cnt == 0) {
6687                 *group_cnt = 0;
6688                 *worker_groups = NULL;
6689                 return 0;
6690         }
6691         *group_cnt = num_possible_nodes();
6692         size = sizeof(struct r5worker) * cnt;
6693         workers = kcalloc(size, *group_cnt, GFP_NOIO);
6694         *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6695                                  GFP_NOIO);
6696         if (!*worker_groups || !workers) {
6697                 kfree(workers);
6698                 kfree(*worker_groups);
6699                 return -ENOMEM;
6700         }
6701
6702         for (i = 0; i < *group_cnt; i++) {
6703                 struct r5worker_group *group;
6704
6705                 group = &(*worker_groups)[i];
6706                 INIT_LIST_HEAD(&group->handle_list);
6707                 INIT_LIST_HEAD(&group->loprio_list);
6708                 group->conf = conf;
6709                 group->workers = workers + i * cnt;
6710
6711                 for (j = 0; j < cnt; j++) {
6712                         struct r5worker *worker = group->workers + j;
6713                         worker->group = group;
6714                         INIT_WORK(&worker->work, raid5_do_work);
6715
6716                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6717                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6718                 }
6719         }
6720
6721         return 0;
6722 }
6723
6724 static void free_thread_groups(struct r5conf *conf)
6725 {
6726         if (conf->worker_groups)
6727                 kfree(conf->worker_groups[0].workers);
6728         kfree(conf->worker_groups);
6729         conf->worker_groups = NULL;
6730 }
6731
6732 static sector_t
6733 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6734 {
6735         struct r5conf *conf = mddev->private;
6736
6737         if (!sectors)
6738                 sectors = mddev->dev_sectors;
6739         if (!raid_disks)
6740                 /* size is defined by the smallest of previous and new size */
6741                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6742
6743         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6744         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6745         return sectors * (raid_disks - conf->max_degraded);
6746 }
6747
6748 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6749 {
6750         safe_put_page(percpu->spare_page);
6751         percpu->spare_page = NULL;
6752         kvfree(percpu->scribble);
6753         percpu->scribble = NULL;
6754 }
6755
6756 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6757 {
6758         if (conf->level == 6 && !percpu->spare_page) {
6759                 percpu->spare_page = alloc_page(GFP_KERNEL);
6760                 if (!percpu->spare_page)
6761                         return -ENOMEM;
6762         }
6763
6764         if (scribble_alloc(percpu,
6765                            max(conf->raid_disks,
6766                                conf->previous_raid_disks),
6767                            max(conf->chunk_sectors,
6768                                conf->prev_chunk_sectors)
6769                            / STRIPE_SECTORS)) {
6770                 free_scratch_buffer(conf, percpu);
6771                 return -ENOMEM;
6772         }
6773
6774         return 0;
6775 }
6776
6777 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6778 {
6779         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6780
6781         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6782         return 0;
6783 }
6784
6785 static void raid5_free_percpu(struct r5conf *conf)
6786 {
6787         if (!conf->percpu)
6788                 return;
6789
6790         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6791         free_percpu(conf->percpu);
6792 }
6793
6794 static void free_conf(struct r5conf *conf)
6795 {
6796         int i;
6797
6798         log_exit(conf);
6799
6800         unregister_shrinker(&conf->shrinker);
6801         free_thread_groups(conf);
6802         shrink_stripes(conf);
6803         raid5_free_percpu(conf);
6804         for (i = 0; i < conf->pool_size; i++)
6805                 if (conf->disks[i].extra_page)
6806                         put_page(conf->disks[i].extra_page);
6807         kfree(conf->disks);
6808         bioset_exit(&conf->bio_split);
6809         kfree(conf->stripe_hashtbl);
6810         kfree(conf->pending_data);
6811         kfree(conf);
6812 }
6813
6814 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6815 {
6816         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6817         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6818
6819         if (alloc_scratch_buffer(conf, percpu)) {
6820                 pr_warn("%s: failed memory allocation for cpu%u\n",
6821                         __func__, cpu);
6822                 return -ENOMEM;
6823         }
6824         return 0;
6825 }
6826
6827 static int raid5_alloc_percpu(struct r5conf *conf)
6828 {
6829         int err = 0;
6830
6831         conf->percpu = alloc_percpu(struct raid5_percpu);
6832         if (!conf->percpu)
6833                 return -ENOMEM;
6834
6835         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6836         if (!err) {
6837                 conf->scribble_disks = max(conf->raid_disks,
6838                         conf->previous_raid_disks);
6839                 conf->scribble_sectors = max(conf->chunk_sectors,
6840                         conf->prev_chunk_sectors);
6841         }
6842         return err;
6843 }
6844
6845 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6846                                       struct shrink_control *sc)
6847 {
6848         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6849         unsigned long ret = SHRINK_STOP;
6850
6851         if (mutex_trylock(&conf->cache_size_mutex)) {
6852                 ret= 0;
6853                 while (ret < sc->nr_to_scan &&
6854                        conf->max_nr_stripes > conf->min_nr_stripes) {
6855                         if (drop_one_stripe(conf) == 0) {
6856                                 ret = SHRINK_STOP;
6857                                 break;
6858                         }
6859                         ret++;
6860                 }
6861                 mutex_unlock(&conf->cache_size_mutex);
6862         }
6863         return ret;
6864 }
6865
6866 static unsigned long raid5_cache_count(struct shrinker *shrink,
6867                                        struct shrink_control *sc)
6868 {
6869         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6870
6871         if (conf->max_nr_stripes < conf->min_nr_stripes)
6872                 /* unlikely, but not impossible */
6873                 return 0;
6874         return conf->max_nr_stripes - conf->min_nr_stripes;
6875 }
6876
6877 static struct r5conf *setup_conf(struct mddev *mddev)
6878 {
6879         struct r5conf *conf;
6880         int raid_disk, memory, max_disks;
6881         struct md_rdev *rdev;
6882         struct disk_info *disk;
6883         char pers_name[6];
6884         int i;
6885         int group_cnt;
6886         struct r5worker_group *new_group;
6887         int ret;
6888
6889         if (mddev->new_level != 5
6890             && mddev->new_level != 4
6891             && mddev->new_level != 6) {
6892                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6893                         mdname(mddev), mddev->new_level);
6894                 return ERR_PTR(-EIO);
6895         }
6896         if ((mddev->new_level == 5
6897              && !algorithm_valid_raid5(mddev->new_layout)) ||
6898             (mddev->new_level == 6
6899              && !algorithm_valid_raid6(mddev->new_layout))) {
6900                 pr_warn("md/raid:%s: layout %d not supported\n",
6901                         mdname(mddev), mddev->new_layout);
6902                 return ERR_PTR(-EIO);
6903         }
6904         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6905                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6906                         mdname(mddev), mddev->raid_disks);
6907                 return ERR_PTR(-EINVAL);
6908         }
6909
6910         if (!mddev->new_chunk_sectors ||
6911             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6912             !is_power_of_2(mddev->new_chunk_sectors)) {
6913                 pr_warn("md/raid:%s: invalid chunk size %d\n",
6914                         mdname(mddev), mddev->new_chunk_sectors << 9);
6915                 return ERR_PTR(-EINVAL);
6916         }
6917
6918         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6919         if (conf == NULL)
6920                 goto abort;
6921         INIT_LIST_HEAD(&conf->free_list);
6922         INIT_LIST_HEAD(&conf->pending_list);
6923         conf->pending_data = kcalloc(PENDING_IO_MAX,
6924                                      sizeof(struct r5pending_data),
6925                                      GFP_KERNEL);
6926         if (!conf->pending_data)
6927                 goto abort;
6928         for (i = 0; i < PENDING_IO_MAX; i++)
6929                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6930         /* Don't enable multi-threading by default*/
6931         if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
6932                 conf->group_cnt = group_cnt;
6933                 conf->worker_cnt_per_group = 0;
6934                 conf->worker_groups = new_group;
6935         } else
6936                 goto abort;
6937         spin_lock_init(&conf->device_lock);
6938         seqcount_init(&conf->gen_lock);
6939         mutex_init(&conf->cache_size_mutex);
6940         init_waitqueue_head(&conf->wait_for_quiescent);
6941         init_waitqueue_head(&conf->wait_for_stripe);
6942         init_waitqueue_head(&conf->wait_for_overlap);
6943         INIT_LIST_HEAD(&conf->handle_list);
6944         INIT_LIST_HEAD(&conf->loprio_list);
6945         INIT_LIST_HEAD(&conf->hold_list);
6946         INIT_LIST_HEAD(&conf->delayed_list);
6947         INIT_LIST_HEAD(&conf->bitmap_list);
6948         init_llist_head(&conf->released_stripes);
6949         atomic_set(&conf->active_stripes, 0);
6950         atomic_set(&conf->preread_active_stripes, 0);
6951         atomic_set(&conf->active_aligned_reads, 0);
6952         spin_lock_init(&conf->pending_bios_lock);
6953         conf->batch_bio_dispatch = true;
6954         rdev_for_each(rdev, mddev) {
6955                 if (test_bit(Journal, &rdev->flags))
6956                         continue;
6957                 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6958                         conf->batch_bio_dispatch = false;
6959                         break;
6960                 }
6961         }
6962
6963         conf->bypass_threshold = BYPASS_THRESHOLD;
6964         conf->recovery_disabled = mddev->recovery_disabled - 1;
6965
6966         conf->raid_disks = mddev->raid_disks;
6967         if (mddev->reshape_position == MaxSector)
6968                 conf->previous_raid_disks = mddev->raid_disks;
6969         else
6970                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6971         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6972
6973         conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
6974                               GFP_KERNEL);
6975
6976         if (!conf->disks)
6977                 goto abort;
6978
6979         for (i = 0; i < max_disks; i++) {
6980                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6981                 if (!conf->disks[i].extra_page)
6982                         goto abort;
6983         }
6984
6985         ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
6986         if (ret)
6987                 goto abort;
6988         conf->mddev = mddev;
6989
6990         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6991                 goto abort;
6992
6993         /* We init hash_locks[0] separately to that it can be used
6994          * as the reference lock in the spin_lock_nest_lock() call
6995          * in lock_all_device_hash_locks_irq in order to convince
6996          * lockdep that we know what we are doing.
6997          */
6998         spin_lock_init(conf->hash_locks);
6999         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7000                 spin_lock_init(conf->hash_locks + i);
7001
7002         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7003                 INIT_LIST_HEAD(conf->inactive_list + i);
7004
7005         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7006                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7007
7008         atomic_set(&conf->r5c_cached_full_stripes, 0);
7009         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7010         atomic_set(&conf->r5c_cached_partial_stripes, 0);
7011         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7012         atomic_set(&conf->r5c_flushing_full_stripes, 0);
7013         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7014
7015         conf->level = mddev->new_level;
7016         conf->chunk_sectors = mddev->new_chunk_sectors;
7017         if (raid5_alloc_percpu(conf) != 0)
7018                 goto abort;
7019
7020         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7021
7022         rdev_for_each(rdev, mddev) {
7023                 raid_disk = rdev->raid_disk;
7024                 if (raid_disk >= max_disks
7025                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7026                         continue;
7027                 disk = conf->disks + raid_disk;
7028
7029                 if (test_bit(Replacement, &rdev->flags)) {
7030                         if (disk->replacement)
7031                                 goto abort;
7032                         disk->replacement = rdev;
7033                 } else {
7034                         if (disk->rdev)
7035                                 goto abort;
7036                         disk->rdev = rdev;
7037                 }
7038
7039                 if (test_bit(In_sync, &rdev->flags)) {
7040                         char b[BDEVNAME_SIZE];
7041                         pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7042                                 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7043                 } else if (rdev->saved_raid_disk != raid_disk)
7044                         /* Cannot rely on bitmap to complete recovery */
7045                         conf->fullsync = 1;
7046         }
7047
7048         conf->level = mddev->new_level;
7049         if (conf->level == 6) {
7050                 conf->max_degraded = 2;
7051                 if (raid6_call.xor_syndrome)
7052                         conf->rmw_level = PARITY_ENABLE_RMW;
7053                 else
7054                         conf->rmw_level = PARITY_DISABLE_RMW;
7055         } else {
7056                 conf->max_degraded = 1;
7057                 conf->rmw_level = PARITY_ENABLE_RMW;
7058         }
7059         conf->algorithm = mddev->new_layout;
7060         conf->reshape_progress = mddev->reshape_position;
7061         if (conf->reshape_progress != MaxSector) {
7062                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7063                 conf->prev_algo = mddev->layout;
7064         } else {
7065                 conf->prev_chunk_sectors = conf->chunk_sectors;
7066                 conf->prev_algo = conf->algorithm;
7067         }
7068
7069         conf->min_nr_stripes = NR_STRIPES;
7070         if (mddev->reshape_position != MaxSector) {
7071                 int stripes = max_t(int,
7072                         ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7073                         ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7074                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7075                 if (conf->min_nr_stripes != NR_STRIPES)
7076                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7077                                 mdname(mddev), conf->min_nr_stripes);
7078         }
7079         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7080                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7081         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7082         if (grow_stripes(conf, conf->min_nr_stripes)) {
7083                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7084                         mdname(mddev), memory);
7085                 goto abort;
7086         } else
7087                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7088         /*
7089          * Losing a stripe head costs more than the time to refill it,
7090          * it reduces the queue depth and so can hurt throughput.
7091          * So set it rather large, scaled by number of devices.
7092          */
7093         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7094         conf->shrinker.scan_objects = raid5_cache_scan;
7095         conf->shrinker.count_objects = raid5_cache_count;
7096         conf->shrinker.batch = 128;
7097         conf->shrinker.flags = 0;
7098         if (register_shrinker(&conf->shrinker)) {
7099                 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7100                         mdname(mddev));
7101                 goto abort;
7102         }
7103
7104         sprintf(pers_name, "raid%d", mddev->new_level);
7105         conf->thread = md_register_thread(raid5d, mddev, pers_name);
7106         if (!conf->thread) {
7107                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7108                         mdname(mddev));
7109                 goto abort;
7110         }
7111
7112         return conf;
7113
7114  abort:
7115         if (conf) {
7116                 free_conf(conf);
7117                 return ERR_PTR(-EIO);
7118         } else
7119                 return ERR_PTR(-ENOMEM);
7120 }
7121
7122 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7123 {
7124         switch (algo) {
7125         case ALGORITHM_PARITY_0:
7126                 if (raid_disk < max_degraded)
7127                         return 1;
7128                 break;
7129         case ALGORITHM_PARITY_N:
7130                 if (raid_disk >= raid_disks - max_degraded)
7131                         return 1;
7132                 break;
7133         case ALGORITHM_PARITY_0_6:
7134                 if (raid_disk == 0 ||
7135                     raid_disk == raid_disks - 1)
7136                         return 1;
7137                 break;
7138         case ALGORITHM_LEFT_ASYMMETRIC_6:
7139         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7140         case ALGORITHM_LEFT_SYMMETRIC_6:
7141         case ALGORITHM_RIGHT_SYMMETRIC_6:
7142                 if (raid_disk == raid_disks - 1)
7143                         return 1;
7144         }
7145         return 0;
7146 }
7147
7148 static int raid5_run(struct mddev *mddev)
7149 {
7150         struct r5conf *conf;
7151         int working_disks = 0;
7152         int dirty_parity_disks = 0;
7153         struct md_rdev *rdev;
7154         struct md_rdev *journal_dev = NULL;
7155         sector_t reshape_offset = 0;
7156         int i;
7157         long long min_offset_diff = 0;
7158         int first = 1;
7159
7160         if (mddev_init_writes_pending(mddev) < 0)
7161                 return -ENOMEM;
7162
7163         if (mddev->recovery_cp != MaxSector)
7164                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7165                           mdname(mddev));
7166
7167         rdev_for_each(rdev, mddev) {
7168                 long long diff;
7169
7170                 if (test_bit(Journal, &rdev->flags)) {
7171                         journal_dev = rdev;
7172                         continue;
7173                 }
7174                 if (rdev->raid_disk < 0)
7175                         continue;
7176                 diff = (rdev->new_data_offset - rdev->data_offset);
7177                 if (first) {
7178                         min_offset_diff = diff;
7179                         first = 0;
7180                 } else if (mddev->reshape_backwards &&
7181                          diff < min_offset_diff)
7182                         min_offset_diff = diff;
7183                 else if (!mddev->reshape_backwards &&
7184                          diff > min_offset_diff)
7185                         min_offset_diff = diff;
7186         }
7187
7188         if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7189             (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7190                 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7191                           mdname(mddev));
7192                 return -EINVAL;
7193         }
7194
7195         if (mddev->reshape_position != MaxSector) {
7196                 /* Check that we can continue the reshape.
7197                  * Difficulties arise if the stripe we would write to
7198                  * next is at or after the stripe we would read from next.
7199                  * For a reshape that changes the number of devices, this
7200                  * is only possible for a very short time, and mdadm makes
7201                  * sure that time appears to have past before assembling
7202                  * the array.  So we fail if that time hasn't passed.
7203                  * For a reshape that keeps the number of devices the same
7204                  * mdadm must be monitoring the reshape can keeping the
7205                  * critical areas read-only and backed up.  It will start
7206                  * the array in read-only mode, so we check for that.
7207                  */
7208                 sector_t here_new, here_old;
7209                 int old_disks;
7210                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7211                 int chunk_sectors;
7212                 int new_data_disks;
7213
7214                 if (journal_dev) {
7215                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7216                                 mdname(mddev));
7217                         return -EINVAL;
7218                 }
7219
7220                 if (mddev->new_level != mddev->level) {
7221                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7222                                 mdname(mddev));
7223                         return -EINVAL;
7224                 }
7225                 old_disks = mddev->raid_disks - mddev->delta_disks;
7226                 /* reshape_position must be on a new-stripe boundary, and one
7227                  * further up in new geometry must map after here in old
7228                  * geometry.
7229                  * If the chunk sizes are different, then as we perform reshape
7230                  * in units of the largest of the two, reshape_position needs
7231                  * be a multiple of the largest chunk size times new data disks.
7232                  */
7233                 here_new = mddev->reshape_position;
7234                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7235                 new_data_disks = mddev->raid_disks - max_degraded;
7236                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7237                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7238                                 mdname(mddev));
7239                         return -EINVAL;
7240                 }
7241                 reshape_offset = here_new * chunk_sectors;
7242                 /* here_new is the stripe we will write to */
7243                 here_old = mddev->reshape_position;
7244                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7245                 /* here_old is the first stripe that we might need to read
7246                  * from */
7247                 if (mddev->delta_disks == 0) {
7248                         /* We cannot be sure it is safe to start an in-place
7249                          * reshape.  It is only safe if user-space is monitoring
7250                          * and taking constant backups.
7251                          * mdadm always starts a situation like this in
7252                          * readonly mode so it can take control before
7253                          * allowing any writes.  So just check for that.
7254                          */
7255                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7256                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7257                                 /* not really in-place - so OK */;
7258                         else if (mddev->ro == 0) {
7259                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7260                                         mdname(mddev));
7261                                 return -EINVAL;
7262                         }
7263                 } else if (mddev->reshape_backwards
7264                     ? (here_new * chunk_sectors + min_offset_diff <=
7265                        here_old * chunk_sectors)
7266                     : (here_new * chunk_sectors >=
7267                        here_old * chunk_sectors + (-min_offset_diff))) {
7268                         /* Reading from the same stripe as writing to - bad */
7269                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7270                                 mdname(mddev));
7271                         return -EINVAL;
7272                 }
7273                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7274                 /* OK, we should be able to continue; */
7275         } else {
7276                 BUG_ON(mddev->level != mddev->new_level);
7277                 BUG_ON(mddev->layout != mddev->new_layout);
7278                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7279                 BUG_ON(mddev->delta_disks != 0);
7280         }
7281
7282         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7283             test_bit(MD_HAS_PPL, &mddev->flags)) {
7284                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7285                         mdname(mddev));
7286                 clear_bit(MD_HAS_PPL, &mddev->flags);
7287                 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7288         }
7289
7290         if (mddev->private == NULL)
7291                 conf = setup_conf(mddev);
7292         else
7293                 conf = mddev->private;
7294
7295         if (IS_ERR(conf))
7296                 return PTR_ERR(conf);
7297
7298         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7299                 if (!journal_dev) {
7300                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7301                                 mdname(mddev));
7302                         mddev->ro = 1;
7303                         set_disk_ro(mddev->gendisk, 1);
7304                 } else if (mddev->recovery_cp == MaxSector)
7305                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7306         }
7307
7308         conf->min_offset_diff = min_offset_diff;
7309         mddev->thread = conf->thread;
7310         conf->thread = NULL;
7311         mddev->private = conf;
7312
7313         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7314              i++) {
7315                 rdev = conf->disks[i].rdev;
7316                 if (!rdev && conf->disks[i].replacement) {
7317                         /* The replacement is all we have yet */
7318                         rdev = conf->disks[i].replacement;
7319                         conf->disks[i].replacement = NULL;
7320                         clear_bit(Replacement, &rdev->flags);
7321                         conf->disks[i].rdev = rdev;
7322                 }
7323                 if (!rdev)
7324                         continue;
7325                 if (conf->disks[i].replacement &&
7326                     conf->reshape_progress != MaxSector) {
7327                         /* replacements and reshape simply do not mix. */
7328                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7329                         goto abort;
7330                 }
7331                 if (test_bit(In_sync, &rdev->flags)) {
7332                         working_disks++;
7333                         continue;
7334                 }
7335                 /* This disc is not fully in-sync.  However if it
7336                  * just stored parity (beyond the recovery_offset),
7337                  * when we don't need to be concerned about the
7338                  * array being dirty.
7339                  * When reshape goes 'backwards', we never have
7340                  * partially completed devices, so we only need
7341                  * to worry about reshape going forwards.
7342                  */
7343                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7344                 if (mddev->major_version == 0 &&
7345                     mddev->minor_version > 90)
7346                         rdev->recovery_offset = reshape_offset;
7347
7348                 if (rdev->recovery_offset < reshape_offset) {
7349                         /* We need to check old and new layout */
7350                         if (!only_parity(rdev->raid_disk,
7351                                          conf->algorithm,
7352                                          conf->raid_disks,
7353                                          conf->max_degraded))
7354                                 continue;
7355                 }
7356                 if (!only_parity(rdev->raid_disk,
7357                                  conf->prev_algo,
7358                                  conf->previous_raid_disks,
7359                                  conf->max_degraded))
7360                         continue;
7361                 dirty_parity_disks++;
7362         }
7363
7364         /*
7365          * 0 for a fully functional array, 1 or 2 for a degraded array.
7366          */
7367         mddev->degraded = raid5_calc_degraded(conf);
7368
7369         if (has_failed(conf)) {
7370                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7371                         mdname(mddev), mddev->degraded, conf->raid_disks);
7372                 goto abort;
7373         }
7374
7375         /* device size must be a multiple of chunk size */
7376         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7377         mddev->resync_max_sectors = mddev->dev_sectors;
7378
7379         if (mddev->degraded > dirty_parity_disks &&
7380             mddev->recovery_cp != MaxSector) {
7381                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7382                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7383                                 mdname(mddev));
7384                 else if (mddev->ok_start_degraded)
7385                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7386                                 mdname(mddev));
7387                 else {
7388                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7389                                 mdname(mddev));
7390                         goto abort;
7391                 }
7392         }
7393
7394         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7395                 mdname(mddev), conf->level,
7396                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7397                 mddev->new_layout);
7398
7399         print_raid5_conf(conf);
7400
7401         if (conf->reshape_progress != MaxSector) {
7402                 conf->reshape_safe = conf->reshape_progress;
7403                 atomic_set(&conf->reshape_stripes, 0);
7404                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7405                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7406                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7407                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7408                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7409                                                         "reshape");
7410                 if (!mddev->sync_thread)
7411                         goto abort;
7412         }
7413
7414         /* Ok, everything is just fine now */
7415         if (mddev->to_remove == &raid5_attrs_group)
7416                 mddev->to_remove = NULL;
7417         else if (mddev->kobj.sd &&
7418             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7419                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7420                         mdname(mddev));
7421         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7422
7423         if (mddev->queue) {
7424                 int chunk_size;
7425                 /* read-ahead size must cover two whole stripes, which
7426                  * is 2 * (datadisks) * chunksize where 'n' is the
7427                  * number of raid devices
7428                  */
7429                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7430                 int stripe = data_disks *
7431                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7432                 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7433                         mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7434
7435                 chunk_size = mddev->chunk_sectors << 9;
7436                 blk_queue_io_min(mddev->queue, chunk_size);
7437                 blk_queue_io_opt(mddev->queue, chunk_size *
7438                                  (conf->raid_disks - conf->max_degraded));
7439                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7440                 /*
7441                  * We can only discard a whole stripe. It doesn't make sense to
7442                  * discard data disk but write parity disk
7443                  */
7444                 stripe = stripe * PAGE_SIZE;
7445                 /* Round up to power of 2, as discard handling
7446                  * currently assumes that */
7447                 while ((stripe-1) & stripe)
7448                         stripe = (stripe | (stripe-1)) + 1;
7449                 mddev->queue->limits.discard_alignment = stripe;
7450                 mddev->queue->limits.discard_granularity = stripe;
7451
7452                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7453                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7454
7455                 rdev_for_each(rdev, mddev) {
7456                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7457                                           rdev->data_offset << 9);
7458                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7459                                           rdev->new_data_offset << 9);
7460                 }
7461
7462                 /*
7463                  * zeroing is required, otherwise data
7464                  * could be lost. Consider a scenario: discard a stripe
7465                  * (the stripe could be inconsistent if
7466                  * discard_zeroes_data is 0); write one disk of the
7467                  * stripe (the stripe could be inconsistent again
7468                  * depending on which disks are used to calculate
7469                  * parity); the disk is broken; The stripe data of this
7470                  * disk is lost.
7471                  *
7472                  * We only allow DISCARD if the sysadmin has confirmed that
7473                  * only safe devices are in use by setting a module parameter.
7474                  * A better idea might be to turn DISCARD into WRITE_ZEROES
7475                  * requests, as that is required to be safe.
7476                  */
7477                 if (devices_handle_discard_safely &&
7478                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7479                     mddev->queue->limits.discard_granularity >= stripe)
7480                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7481                                                 mddev->queue);
7482                 else
7483                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7484                                                 mddev->queue);
7485
7486                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7487         }
7488
7489         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7490                 goto abort;
7491
7492         return 0;
7493 abort:
7494         md_unregister_thread(&mddev->thread);
7495         print_raid5_conf(conf);
7496         free_conf(conf);
7497         mddev->private = NULL;
7498         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7499         return -EIO;
7500 }
7501
7502 static void raid5_free(struct mddev *mddev, void *priv)
7503 {
7504         struct r5conf *conf = priv;
7505
7506         free_conf(conf);
7507         mddev->to_remove = &raid5_attrs_group;
7508 }
7509
7510 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7511 {
7512         struct r5conf *conf = mddev->private;
7513         int i;
7514
7515         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7516                 conf->chunk_sectors / 2, mddev->layout);
7517         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7518         rcu_read_lock();
7519         for (i = 0; i < conf->raid_disks; i++) {
7520                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7521                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7522         }
7523         rcu_read_unlock();
7524         seq_printf (seq, "]");
7525 }
7526
7527 static void print_raid5_conf (struct r5conf *conf)
7528 {
7529         int i;
7530         struct disk_info *tmp;
7531
7532         pr_debug("RAID conf printout:\n");
7533         if (!conf) {
7534                 pr_debug("(conf==NULL)\n");
7535                 return;
7536         }
7537         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7538                conf->raid_disks,
7539                conf->raid_disks - conf->mddev->degraded);
7540
7541         for (i = 0; i < conf->raid_disks; i++) {
7542                 char b[BDEVNAME_SIZE];
7543                 tmp = conf->disks + i;
7544                 if (tmp->rdev)
7545                         pr_debug(" disk %d, o:%d, dev:%s\n",
7546                                i, !test_bit(Faulty, &tmp->rdev->flags),
7547                                bdevname(tmp->rdev->bdev, b));
7548         }
7549 }
7550
7551 static int raid5_spare_active(struct mddev *mddev)
7552 {
7553         int i;
7554         struct r5conf *conf = mddev->private;
7555         struct disk_info *tmp;
7556         int count = 0;
7557         unsigned long flags;
7558
7559         for (i = 0; i < conf->raid_disks; i++) {
7560                 tmp = conf->disks + i;
7561                 if (tmp->replacement
7562                     && tmp->replacement->recovery_offset == MaxSector
7563                     && !test_bit(Faulty, &tmp->replacement->flags)
7564                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7565                         /* Replacement has just become active. */
7566                         if (!tmp->rdev
7567                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7568                                 count++;
7569                         if (tmp->rdev) {
7570                                 /* Replaced device not technically faulty,
7571                                  * but we need to be sure it gets removed
7572                                  * and never re-added.
7573                                  */
7574                                 set_bit(Faulty, &tmp->rdev->flags);
7575                                 sysfs_notify_dirent_safe(
7576                                         tmp->rdev->sysfs_state);
7577                         }
7578                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7579                 } else if (tmp->rdev
7580                     && tmp->rdev->recovery_offset == MaxSector
7581                     && !test_bit(Faulty, &tmp->rdev->flags)
7582                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7583                         count++;
7584                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7585                 }
7586         }
7587         spin_lock_irqsave(&conf->device_lock, flags);
7588         mddev->degraded = raid5_calc_degraded(conf);
7589         spin_unlock_irqrestore(&conf->device_lock, flags);
7590         print_raid5_conf(conf);
7591         return count;
7592 }
7593
7594 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7595 {
7596         struct r5conf *conf = mddev->private;
7597         int err = 0;
7598         int number = rdev->raid_disk;
7599         struct md_rdev **rdevp;
7600         struct disk_info *p = conf->disks + number;
7601
7602         print_raid5_conf(conf);
7603         if (test_bit(Journal, &rdev->flags) && conf->log) {
7604                 /*
7605                  * we can't wait pending write here, as this is called in
7606                  * raid5d, wait will deadlock.
7607                  * neilb: there is no locking about new writes here,
7608                  * so this cannot be safe.
7609                  */
7610                 if (atomic_read(&conf->active_stripes) ||
7611                     atomic_read(&conf->r5c_cached_full_stripes) ||
7612                     atomic_read(&conf->r5c_cached_partial_stripes)) {
7613                         return -EBUSY;
7614                 }
7615                 log_exit(conf);
7616                 return 0;
7617         }
7618         if (rdev == p->rdev)
7619                 rdevp = &p->rdev;
7620         else if (rdev == p->replacement)
7621                 rdevp = &p->replacement;
7622         else
7623                 return 0;
7624
7625         if (number >= conf->raid_disks &&
7626             conf->reshape_progress == MaxSector)
7627                 clear_bit(In_sync, &rdev->flags);
7628
7629         if (test_bit(In_sync, &rdev->flags) ||
7630             atomic_read(&rdev->nr_pending)) {
7631                 err = -EBUSY;
7632                 goto abort;
7633         }
7634         /* Only remove non-faulty devices if recovery
7635          * isn't possible.
7636          */
7637         if (!test_bit(Faulty, &rdev->flags) &&
7638             mddev->recovery_disabled != conf->recovery_disabled &&
7639             !has_failed(conf) &&
7640             (!p->replacement || p->replacement == rdev) &&
7641             number < conf->raid_disks) {
7642                 err = -EBUSY;
7643                 goto abort;
7644         }
7645         *rdevp = NULL;
7646         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7647                 synchronize_rcu();
7648                 if (atomic_read(&rdev->nr_pending)) {
7649                         /* lost the race, try later */
7650                         err = -EBUSY;
7651                         *rdevp = rdev;
7652                 }
7653         }
7654         if (!err) {
7655                 err = log_modify(conf, rdev, false);
7656                 if (err)
7657                         goto abort;
7658         }
7659         if (p->replacement) {
7660                 /* We must have just cleared 'rdev' */
7661                 p->rdev = p->replacement;
7662                 clear_bit(Replacement, &p->replacement->flags);
7663                 smp_mb(); /* Make sure other CPUs may see both as identical
7664                            * but will never see neither - if they are careful
7665                            */
7666                 p->replacement = NULL;
7667
7668                 if (!err)
7669                         err = log_modify(conf, p->rdev, true);
7670         }
7671
7672         clear_bit(WantReplacement, &rdev->flags);
7673 abort:
7674
7675         print_raid5_conf(conf);
7676         return err;
7677 }
7678
7679 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7680 {
7681         struct r5conf *conf = mddev->private;
7682         int ret, err = -EEXIST;
7683         int disk;
7684         struct disk_info *p;
7685         int first = 0;
7686         int last = conf->raid_disks - 1;
7687
7688         if (test_bit(Journal, &rdev->flags)) {
7689                 if (conf->log)
7690                         return -EBUSY;
7691
7692                 rdev->raid_disk = 0;
7693                 /*
7694                  * The array is in readonly mode if journal is missing, so no
7695                  * write requests running. We should be safe
7696                  */
7697                 ret = log_init(conf, rdev, false);
7698                 if (ret)
7699                         return ret;
7700
7701                 ret = r5l_start(conf->log);
7702                 if (ret)
7703                         return ret;
7704
7705                 return 0;
7706         }
7707         if (mddev->recovery_disabled == conf->recovery_disabled)
7708                 return -EBUSY;
7709
7710         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7711                 /* no point adding a device */
7712                 return -EINVAL;
7713
7714         if (rdev->raid_disk >= 0)
7715                 first = last = rdev->raid_disk;
7716
7717         /*
7718          * find the disk ... but prefer rdev->saved_raid_disk
7719          * if possible.
7720          */
7721         if (rdev->saved_raid_disk >= 0 &&
7722             rdev->saved_raid_disk >= first &&
7723             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7724                 first = rdev->saved_raid_disk;
7725
7726         for (disk = first; disk <= last; disk++) {
7727                 p = conf->disks + disk;
7728                 if (p->rdev == NULL) {
7729                         clear_bit(In_sync, &rdev->flags);
7730                         rdev->raid_disk = disk;
7731                         if (rdev->saved_raid_disk != disk)
7732                                 conf->fullsync = 1;
7733                         rcu_assign_pointer(p->rdev, rdev);
7734
7735                         err = log_modify(conf, rdev, true);
7736
7737                         goto out;
7738                 }
7739         }
7740         for (disk = first; disk <= last; disk++) {
7741                 p = conf->disks + disk;
7742                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7743                     p->replacement == NULL) {
7744                         clear_bit(In_sync, &rdev->flags);
7745                         set_bit(Replacement, &rdev->flags);
7746                         rdev->raid_disk = disk;
7747                         err = 0;
7748                         conf->fullsync = 1;
7749                         rcu_assign_pointer(p->replacement, rdev);
7750                         break;
7751                 }
7752         }
7753 out:
7754         print_raid5_conf(conf);
7755         return err;
7756 }
7757
7758 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7759 {
7760         /* no resync is happening, and there is enough space
7761          * on all devices, so we can resize.
7762          * We need to make sure resync covers any new space.
7763          * If the array is shrinking we should possibly wait until
7764          * any io in the removed space completes, but it hardly seems
7765          * worth it.
7766          */
7767         sector_t newsize;
7768         struct r5conf *conf = mddev->private;
7769
7770         if (raid5_has_log(conf) || raid5_has_ppl(conf))
7771                 return -EINVAL;
7772         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7773         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7774         if (mddev->external_size &&
7775             mddev->array_sectors > newsize)
7776                 return -EINVAL;
7777         if (mddev->bitmap) {
7778                 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
7779                 if (ret)
7780                         return ret;
7781         }
7782         md_set_array_sectors(mddev, newsize);
7783         if (sectors > mddev->dev_sectors &&
7784             mddev->recovery_cp > mddev->dev_sectors) {
7785                 mddev->recovery_cp = mddev->dev_sectors;
7786                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7787         }
7788         mddev->dev_sectors = sectors;
7789         mddev->resync_max_sectors = sectors;
7790         return 0;
7791 }
7792
7793 static int check_stripe_cache(struct mddev *mddev)
7794 {
7795         /* Can only proceed if there are plenty of stripe_heads.
7796          * We need a minimum of one full stripe,, and for sensible progress
7797          * it is best to have about 4 times that.
7798          * If we require 4 times, then the default 256 4K stripe_heads will
7799          * allow for chunk sizes up to 256K, which is probably OK.
7800          * If the chunk size is greater, user-space should request more
7801          * stripe_heads first.
7802          */
7803         struct r5conf *conf = mddev->private;
7804         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7805             > conf->min_nr_stripes ||
7806             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7807             > conf->min_nr_stripes) {
7808                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7809                         mdname(mddev),
7810                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7811                          / STRIPE_SIZE)*4);
7812                 return 0;
7813         }
7814         return 1;
7815 }
7816
7817 static int check_reshape(struct mddev *mddev)
7818 {
7819         struct r5conf *conf = mddev->private;
7820
7821         if (raid5_has_log(conf) || raid5_has_ppl(conf))
7822                 return -EINVAL;
7823         if (mddev->delta_disks == 0 &&
7824             mddev->new_layout == mddev->layout &&
7825             mddev->new_chunk_sectors == mddev->chunk_sectors)
7826                 return 0; /* nothing to do */
7827         if (has_failed(conf))
7828                 return -EINVAL;
7829         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7830                 /* We might be able to shrink, but the devices must
7831                  * be made bigger first.
7832                  * For raid6, 4 is the minimum size.
7833                  * Otherwise 2 is the minimum
7834                  */
7835                 int min = 2;
7836                 if (mddev->level == 6)
7837                         min = 4;
7838                 if (mddev->raid_disks + mddev->delta_disks < min)
7839                         return -EINVAL;
7840         }
7841
7842         if (!check_stripe_cache(mddev))
7843                 return -ENOSPC;
7844
7845         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7846             mddev->delta_disks > 0)
7847                 if (resize_chunks(conf,
7848                                   conf->previous_raid_disks
7849                                   + max(0, mddev->delta_disks),
7850                                   max(mddev->new_chunk_sectors,
7851                                       mddev->chunk_sectors)
7852                             ) < 0)
7853                         return -ENOMEM;
7854
7855         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7856                 return 0; /* never bother to shrink */
7857         return resize_stripes(conf, (conf->previous_raid_disks
7858                                      + mddev->delta_disks));
7859 }
7860
7861 static int raid5_start_reshape(struct mddev *mddev)
7862 {
7863         struct r5conf *conf = mddev->private;
7864         struct md_rdev *rdev;
7865         int spares = 0;
7866         unsigned long flags;
7867
7868         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7869                 return -EBUSY;
7870
7871         if (!check_stripe_cache(mddev))
7872                 return -ENOSPC;
7873
7874         if (has_failed(conf))
7875                 return -EINVAL;
7876
7877         rdev_for_each(rdev, mddev) {
7878                 if (!test_bit(In_sync, &rdev->flags)
7879                     && !test_bit(Faulty, &rdev->flags))
7880                         spares++;
7881         }
7882
7883         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7884                 /* Not enough devices even to make a degraded array
7885                  * of that size
7886                  */
7887                 return -EINVAL;
7888
7889         /* Refuse to reduce size of the array.  Any reductions in
7890          * array size must be through explicit setting of array_size
7891          * attribute.
7892          */
7893         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7894             < mddev->array_sectors) {
7895                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7896                         mdname(mddev));
7897                 return -EINVAL;
7898         }
7899
7900         atomic_set(&conf->reshape_stripes, 0);
7901         spin_lock_irq(&conf->device_lock);
7902         write_seqcount_begin(&conf->gen_lock);
7903         conf->previous_raid_disks = conf->raid_disks;
7904         conf->raid_disks += mddev->delta_disks;
7905         conf->prev_chunk_sectors = conf->chunk_sectors;
7906         conf->chunk_sectors = mddev->new_chunk_sectors;
7907         conf->prev_algo = conf->algorithm;
7908         conf->algorithm = mddev->new_layout;
7909         conf->generation++;
7910         /* Code that selects data_offset needs to see the generation update
7911          * if reshape_progress has been set - so a memory barrier needed.
7912          */
7913         smp_mb();
7914         if (mddev->reshape_backwards)
7915                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7916         else
7917                 conf->reshape_progress = 0;
7918         conf->reshape_safe = conf->reshape_progress;
7919         write_seqcount_end(&conf->gen_lock);
7920         spin_unlock_irq(&conf->device_lock);
7921
7922         /* Now make sure any requests that proceeded on the assumption
7923          * the reshape wasn't running - like Discard or Read - have
7924          * completed.
7925          */
7926         mddev_suspend(mddev);
7927         mddev_resume(mddev);
7928
7929         /* Add some new drives, as many as will fit.
7930          * We know there are enough to make the newly sized array work.
7931          * Don't add devices if we are reducing the number of
7932          * devices in the array.  This is because it is not possible
7933          * to correctly record the "partially reconstructed" state of
7934          * such devices during the reshape and confusion could result.
7935          */
7936         if (mddev->delta_disks >= 0) {
7937                 rdev_for_each(rdev, mddev)
7938                         if (rdev->raid_disk < 0 &&
7939                             !test_bit(Faulty, &rdev->flags)) {
7940                                 if (raid5_add_disk(mddev, rdev) == 0) {
7941                                         if (rdev->raid_disk
7942                                             >= conf->previous_raid_disks)
7943                                                 set_bit(In_sync, &rdev->flags);
7944                                         else
7945                                                 rdev->recovery_offset = 0;
7946
7947                                         if (sysfs_link_rdev(mddev, rdev))
7948                                                 /* Failure here is OK */;
7949                                 }
7950                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7951                                    && !test_bit(Faulty, &rdev->flags)) {
7952                                 /* This is a spare that was manually added */
7953                                 set_bit(In_sync, &rdev->flags);
7954                         }
7955
7956                 /* When a reshape changes the number of devices,
7957                  * ->degraded is measured against the larger of the
7958                  * pre and post number of devices.
7959                  */
7960                 spin_lock_irqsave(&conf->device_lock, flags);
7961                 mddev->degraded = raid5_calc_degraded(conf);
7962                 spin_unlock_irqrestore(&conf->device_lock, flags);
7963         }
7964         mddev->raid_disks = conf->raid_disks;
7965         mddev->reshape_position = conf->reshape_progress;
7966         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7967
7968         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7969         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7970         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7971         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7972         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7973         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7974                                                 "reshape");
7975         if (!mddev->sync_thread) {
7976                 mddev->recovery = 0;
7977                 spin_lock_irq(&conf->device_lock);
7978                 write_seqcount_begin(&conf->gen_lock);
7979                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7980                 mddev->new_chunk_sectors =
7981                         conf->chunk_sectors = conf->prev_chunk_sectors;
7982                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7983                 rdev_for_each(rdev, mddev)
7984                         rdev->new_data_offset = rdev->data_offset;
7985                 smp_wmb();
7986                 conf->generation --;
7987                 conf->reshape_progress = MaxSector;
7988                 mddev->reshape_position = MaxSector;
7989                 write_seqcount_end(&conf->gen_lock);
7990                 spin_unlock_irq(&conf->device_lock);
7991                 return -EAGAIN;
7992         }
7993         conf->reshape_checkpoint = jiffies;
7994         md_wakeup_thread(mddev->sync_thread);
7995         md_new_event(mddev);
7996         return 0;
7997 }
7998
7999 /* This is called from the reshape thread and should make any
8000  * changes needed in 'conf'
8001  */
8002 static void end_reshape(struct r5conf *conf)
8003 {
8004
8005         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8006                 struct md_rdev *rdev;
8007
8008                 spin_lock_irq(&conf->device_lock);
8009                 conf->previous_raid_disks = conf->raid_disks;
8010                 md_finish_reshape(conf->mddev);
8011                 smp_wmb();
8012                 conf->reshape_progress = MaxSector;
8013                 conf->mddev->reshape_position = MaxSector;
8014                 rdev_for_each(rdev, conf->mddev)
8015                         if (rdev->raid_disk >= 0 &&
8016                             !test_bit(Journal, &rdev->flags) &&
8017                             !test_bit(In_sync, &rdev->flags))
8018                                 rdev->recovery_offset = MaxSector;
8019                 spin_unlock_irq(&conf->device_lock);
8020                 wake_up(&conf->wait_for_overlap);
8021
8022                 /* read-ahead size must cover two whole stripes, which is
8023                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
8024                  */
8025                 if (conf->mddev->queue) {
8026                         int data_disks = conf->raid_disks - conf->max_degraded;
8027                         int stripe = data_disks * ((conf->chunk_sectors << 9)
8028                                                    / PAGE_SIZE);
8029                         if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
8030                                 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
8031                 }
8032         }
8033 }
8034
8035 /* This is called from the raid5d thread with mddev_lock held.
8036  * It makes config changes to the device.
8037  */
8038 static void raid5_finish_reshape(struct mddev *mddev)
8039 {
8040         struct r5conf *conf = mddev->private;
8041
8042         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8043
8044                 if (mddev->delta_disks <= 0) {
8045                         int d;
8046                         spin_lock_irq(&conf->device_lock);
8047                         mddev->degraded = raid5_calc_degraded(conf);
8048                         spin_unlock_irq(&conf->device_lock);
8049                         for (d = conf->raid_disks ;
8050                              d < conf->raid_disks - mddev->delta_disks;
8051                              d++) {
8052                                 struct md_rdev *rdev = conf->disks[d].rdev;
8053                                 if (rdev)
8054                                         clear_bit(In_sync, &rdev->flags);
8055                                 rdev = conf->disks[d].replacement;
8056                                 if (rdev)
8057                                         clear_bit(In_sync, &rdev->flags);
8058                         }
8059                 }
8060                 mddev->layout = conf->algorithm;
8061                 mddev->chunk_sectors = conf->chunk_sectors;
8062                 mddev->reshape_position = MaxSector;
8063                 mddev->delta_disks = 0;
8064                 mddev->reshape_backwards = 0;
8065         }
8066 }
8067
8068 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8069 {
8070         struct r5conf *conf = mddev->private;
8071
8072         if (quiesce) {
8073                 /* stop all writes */
8074                 lock_all_device_hash_locks_irq(conf);
8075                 /* '2' tells resync/reshape to pause so that all
8076                  * active stripes can drain
8077                  */
8078                 r5c_flush_cache(conf, INT_MAX);
8079                 conf->quiesce = 2;
8080                 wait_event_cmd(conf->wait_for_quiescent,
8081                                     atomic_read(&conf->active_stripes) == 0 &&
8082                                     atomic_read(&conf->active_aligned_reads) == 0,
8083                                     unlock_all_device_hash_locks_irq(conf),
8084                                     lock_all_device_hash_locks_irq(conf));
8085                 conf->quiesce = 1;
8086                 unlock_all_device_hash_locks_irq(conf);
8087                 /* allow reshape to continue */
8088                 wake_up(&conf->wait_for_overlap);
8089         } else {
8090                 /* re-enable writes */
8091                 lock_all_device_hash_locks_irq(conf);
8092                 conf->quiesce = 0;
8093                 wake_up(&conf->wait_for_quiescent);
8094                 wake_up(&conf->wait_for_overlap);
8095                 unlock_all_device_hash_locks_irq(conf);
8096         }
8097         log_quiesce(conf, quiesce);
8098 }
8099
8100 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8101 {
8102         struct r0conf *raid0_conf = mddev->private;
8103         sector_t sectors;
8104
8105         /* for raid0 takeover only one zone is supported */
8106         if (raid0_conf->nr_strip_zones > 1) {
8107                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8108                         mdname(mddev));
8109                 return ERR_PTR(-EINVAL);
8110         }
8111
8112         sectors = raid0_conf->strip_zone[0].zone_end;
8113         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8114         mddev->dev_sectors = sectors;
8115         mddev->new_level = level;
8116         mddev->new_layout = ALGORITHM_PARITY_N;
8117         mddev->new_chunk_sectors = mddev->chunk_sectors;
8118         mddev->raid_disks += 1;
8119         mddev->delta_disks = 1;
8120         /* make sure it will be not marked as dirty */
8121         mddev->recovery_cp = MaxSector;
8122
8123         return setup_conf(mddev);
8124 }
8125
8126 static void *raid5_takeover_raid1(struct mddev *mddev)
8127 {
8128         int chunksect;
8129         void *ret;
8130
8131         if (mddev->raid_disks != 2 ||
8132             mddev->degraded > 1)
8133                 return ERR_PTR(-EINVAL);
8134
8135         /* Should check if there are write-behind devices? */
8136
8137         chunksect = 64*2; /* 64K by default */
8138
8139         /* The array must be an exact multiple of chunksize */
8140         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8141                 chunksect >>= 1;
8142
8143         if ((chunksect<<9) < STRIPE_SIZE)
8144                 /* array size does not allow a suitable chunk size */
8145                 return ERR_PTR(-EINVAL);
8146
8147         mddev->new_level = 5;
8148         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8149         mddev->new_chunk_sectors = chunksect;
8150
8151         ret = setup_conf(mddev);
8152         if (!IS_ERR(ret))
8153                 mddev_clear_unsupported_flags(mddev,
8154                         UNSUPPORTED_MDDEV_FLAGS);
8155         return ret;
8156 }
8157
8158 static void *raid5_takeover_raid6(struct mddev *mddev)
8159 {
8160         int new_layout;
8161
8162         switch (mddev->layout) {
8163         case ALGORITHM_LEFT_ASYMMETRIC_6:
8164                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8165                 break;
8166         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8167                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8168                 break;
8169         case ALGORITHM_LEFT_SYMMETRIC_6:
8170                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8171                 break;
8172         case ALGORITHM_RIGHT_SYMMETRIC_6:
8173                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8174                 break;
8175         case ALGORITHM_PARITY_0_6:
8176                 new_layout = ALGORITHM_PARITY_0;
8177                 break;
8178         case ALGORITHM_PARITY_N:
8179                 new_layout = ALGORITHM_PARITY_N;
8180                 break;
8181         default:
8182                 return ERR_PTR(-EINVAL);
8183         }
8184         mddev->new_level = 5;
8185         mddev->new_layout = new_layout;
8186         mddev->delta_disks = -1;
8187         mddev->raid_disks -= 1;
8188         return setup_conf(mddev);
8189 }
8190
8191 static int raid5_check_reshape(struct mddev *mddev)
8192 {
8193         /* For a 2-drive array, the layout and chunk size can be changed
8194          * immediately as not restriping is needed.
8195          * For larger arrays we record the new value - after validation
8196          * to be used by a reshape pass.
8197          */
8198         struct r5conf *conf = mddev->private;
8199         int new_chunk = mddev->new_chunk_sectors;
8200
8201         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8202                 return -EINVAL;
8203         if (new_chunk > 0) {
8204                 if (!is_power_of_2(new_chunk))
8205                         return -EINVAL;
8206                 if (new_chunk < (PAGE_SIZE>>9))
8207                         return -EINVAL;
8208                 if (mddev->array_sectors & (new_chunk-1))
8209                         /* not factor of array size */
8210                         return -EINVAL;
8211         }
8212
8213         /* They look valid */
8214
8215         if (mddev->raid_disks == 2) {
8216                 /* can make the change immediately */
8217                 if (mddev->new_layout >= 0) {
8218                         conf->algorithm = mddev->new_layout;
8219                         mddev->layout = mddev->new_layout;
8220                 }
8221                 if (new_chunk > 0) {
8222                         conf->chunk_sectors = new_chunk ;
8223                         mddev->chunk_sectors = new_chunk;
8224                 }
8225                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8226                 md_wakeup_thread(mddev->thread);
8227         }
8228         return check_reshape(mddev);
8229 }
8230
8231 static int raid6_check_reshape(struct mddev *mddev)
8232 {
8233         int new_chunk = mddev->new_chunk_sectors;
8234
8235         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8236                 return -EINVAL;
8237         if (new_chunk > 0) {
8238                 if (!is_power_of_2(new_chunk))
8239                         return -EINVAL;
8240                 if (new_chunk < (PAGE_SIZE >> 9))
8241                         return -EINVAL;
8242                 if (mddev->array_sectors & (new_chunk-1))
8243                         /* not factor of array size */
8244                         return -EINVAL;
8245         }
8246
8247         /* They look valid */
8248         return check_reshape(mddev);
8249 }
8250
8251 static void *raid5_takeover(struct mddev *mddev)
8252 {
8253         /* raid5 can take over:
8254          *  raid0 - if there is only one strip zone - make it a raid4 layout
8255          *  raid1 - if there are two drives.  We need to know the chunk size
8256          *  raid4 - trivial - just use a raid4 layout.
8257          *  raid6 - Providing it is a *_6 layout
8258          */
8259         if (mddev->level == 0)
8260                 return raid45_takeover_raid0(mddev, 5);
8261         if (mddev->level == 1)
8262                 return raid5_takeover_raid1(mddev);
8263         if (mddev->level == 4) {
8264                 mddev->new_layout = ALGORITHM_PARITY_N;
8265                 mddev->new_level = 5;
8266                 return setup_conf(mddev);
8267         }
8268         if (mddev->level == 6)
8269                 return raid5_takeover_raid6(mddev);
8270
8271         return ERR_PTR(-EINVAL);
8272 }
8273
8274 static void *raid4_takeover(struct mddev *mddev)
8275 {
8276         /* raid4 can take over:
8277          *  raid0 - if there is only one strip zone
8278          *  raid5 - if layout is right
8279          */
8280         if (mddev->level == 0)
8281                 return raid45_takeover_raid0(mddev, 4);
8282         if (mddev->level == 5 &&
8283             mddev->layout == ALGORITHM_PARITY_N) {
8284                 mddev->new_layout = 0;
8285                 mddev->new_level = 4;
8286                 return setup_conf(mddev);
8287         }
8288         return ERR_PTR(-EINVAL);
8289 }
8290
8291 static struct md_personality raid5_personality;
8292
8293 static void *raid6_takeover(struct mddev *mddev)
8294 {
8295         /* Currently can only take over a raid5.  We map the
8296          * personality to an equivalent raid6 personality
8297          * with the Q block at the end.
8298          */
8299         int new_layout;
8300
8301         if (mddev->pers != &raid5_personality)
8302                 return ERR_PTR(-EINVAL);
8303         if (mddev->degraded > 1)
8304                 return ERR_PTR(-EINVAL);
8305         if (mddev->raid_disks > 253)
8306                 return ERR_PTR(-EINVAL);
8307         if (mddev->raid_disks < 3)
8308                 return ERR_PTR(-EINVAL);
8309
8310         switch (mddev->layout) {
8311         case ALGORITHM_LEFT_ASYMMETRIC:
8312                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8313                 break;
8314         case ALGORITHM_RIGHT_ASYMMETRIC:
8315                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8316                 break;
8317         case ALGORITHM_LEFT_SYMMETRIC:
8318                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8319                 break;
8320         case ALGORITHM_RIGHT_SYMMETRIC:
8321                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8322                 break;
8323         case ALGORITHM_PARITY_0:
8324                 new_layout = ALGORITHM_PARITY_0_6;
8325                 break;
8326         case ALGORITHM_PARITY_N:
8327                 new_layout = ALGORITHM_PARITY_N;
8328                 break;
8329         default:
8330                 return ERR_PTR(-EINVAL);
8331         }
8332         mddev->new_level = 6;
8333         mddev->new_layout = new_layout;
8334         mddev->delta_disks = 1;
8335         mddev->raid_disks += 1;
8336         return setup_conf(mddev);
8337 }
8338
8339 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8340 {
8341         struct r5conf *conf;
8342         int err;
8343
8344         err = mddev_lock(mddev);
8345         if (err)
8346                 return err;
8347         conf = mddev->private;
8348         if (!conf) {
8349                 mddev_unlock(mddev);
8350                 return -ENODEV;
8351         }
8352
8353         if (strncmp(buf, "ppl", 3) == 0) {
8354                 /* ppl only works with RAID 5 */
8355                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8356                         err = log_init(conf, NULL, true);
8357                         if (!err) {
8358                                 err = resize_stripes(conf, conf->pool_size);
8359                                 if (err)
8360                                         log_exit(conf);
8361                         }
8362                 } else
8363                         err = -EINVAL;
8364         } else if (strncmp(buf, "resync", 6) == 0) {
8365                 if (raid5_has_ppl(conf)) {
8366                         mddev_suspend(mddev);
8367                         log_exit(conf);
8368                         mddev_resume(mddev);
8369                         err = resize_stripes(conf, conf->pool_size);
8370                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8371                            r5l_log_disk_error(conf)) {
8372                         bool journal_dev_exists = false;
8373                         struct md_rdev *rdev;
8374
8375                         rdev_for_each(rdev, mddev)
8376                                 if (test_bit(Journal, &rdev->flags)) {
8377                                         journal_dev_exists = true;
8378                                         break;
8379                                 }
8380
8381                         if (!journal_dev_exists) {
8382                                 mddev_suspend(mddev);
8383                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8384                                 mddev_resume(mddev);
8385                         } else  /* need remove journal device first */
8386                                 err = -EBUSY;
8387                 } else
8388                         err = -EINVAL;
8389         } else {
8390                 err = -EINVAL;
8391         }
8392
8393         if (!err)
8394                 md_update_sb(mddev, 1);
8395
8396         mddev_unlock(mddev);
8397
8398         return err;
8399 }
8400
8401 static int raid5_start(struct mddev *mddev)
8402 {
8403         struct r5conf *conf = mddev->private;
8404
8405         return r5l_start(conf->log);
8406 }
8407
8408 static struct md_personality raid6_personality =
8409 {
8410         .name           = "raid6",
8411         .level          = 6,
8412         .owner          = THIS_MODULE,
8413         .make_request   = raid5_make_request,
8414         .run            = raid5_run,
8415         .start          = raid5_start,
8416         .free           = raid5_free,
8417         .status         = raid5_status,
8418         .error_handler  = raid5_error,
8419         .hot_add_disk   = raid5_add_disk,
8420         .hot_remove_disk= raid5_remove_disk,
8421         .spare_active   = raid5_spare_active,
8422         .sync_request   = raid5_sync_request,
8423         .resize         = raid5_resize,
8424         .size           = raid5_size,
8425         .check_reshape  = raid6_check_reshape,
8426         .start_reshape  = raid5_start_reshape,
8427         .finish_reshape = raid5_finish_reshape,
8428         .quiesce        = raid5_quiesce,
8429         .takeover       = raid6_takeover,
8430         .congested      = raid5_congested,
8431         .change_consistency_policy = raid5_change_consistency_policy,
8432 };
8433 static struct md_personality raid5_personality =
8434 {
8435         .name           = "raid5",
8436         .level          = 5,
8437         .owner          = THIS_MODULE,
8438         .make_request   = raid5_make_request,
8439         .run            = raid5_run,
8440         .start          = raid5_start,
8441         .free           = raid5_free,
8442         .status         = raid5_status,
8443         .error_handler  = raid5_error,
8444         .hot_add_disk   = raid5_add_disk,
8445         .hot_remove_disk= raid5_remove_disk,
8446         .spare_active   = raid5_spare_active,
8447         .sync_request   = raid5_sync_request,
8448         .resize         = raid5_resize,
8449         .size           = raid5_size,
8450         .check_reshape  = raid5_check_reshape,
8451         .start_reshape  = raid5_start_reshape,
8452         .finish_reshape = raid5_finish_reshape,
8453         .quiesce        = raid5_quiesce,
8454         .takeover       = raid5_takeover,
8455         .congested      = raid5_congested,
8456         .change_consistency_policy = raid5_change_consistency_policy,
8457 };
8458
8459 static struct md_personality raid4_personality =
8460 {
8461         .name           = "raid4",
8462         .level          = 4,
8463         .owner          = THIS_MODULE,
8464         .make_request   = raid5_make_request,
8465         .run            = raid5_run,
8466         .start          = raid5_start,
8467         .free           = raid5_free,
8468         .status         = raid5_status,
8469         .error_handler  = raid5_error,
8470         .hot_add_disk   = raid5_add_disk,
8471         .hot_remove_disk= raid5_remove_disk,
8472         .spare_active   = raid5_spare_active,
8473         .sync_request   = raid5_sync_request,
8474         .resize         = raid5_resize,
8475         .size           = raid5_size,
8476         .check_reshape  = raid5_check_reshape,
8477         .start_reshape  = raid5_start_reshape,
8478         .finish_reshape = raid5_finish_reshape,
8479         .quiesce        = raid5_quiesce,
8480         .takeover       = raid4_takeover,
8481         .congested      = raid5_congested,
8482         .change_consistency_policy = raid5_change_consistency_policy,
8483 };
8484
8485 static int __init raid5_init(void)
8486 {
8487         int ret;
8488
8489         raid5_wq = alloc_workqueue("raid5wq",
8490                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8491         if (!raid5_wq)
8492                 return -ENOMEM;
8493
8494         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8495                                       "md/raid5:prepare",
8496                                       raid456_cpu_up_prepare,
8497                                       raid456_cpu_dead);
8498         if (ret) {
8499                 destroy_workqueue(raid5_wq);
8500                 return ret;
8501         }
8502         register_md_personality(&raid6_personality);
8503         register_md_personality(&raid5_personality);
8504         register_md_personality(&raid4_personality);
8505         return 0;
8506 }
8507
8508 static void raid5_exit(void)
8509 {
8510         unregister_md_personality(&raid6_personality);
8511         unregister_md_personality(&raid5_personality);
8512         unregister_md_personality(&raid4_personality);
8513         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8514         destroy_workqueue(raid5_wq);
8515 }
8516
8517 module_init(raid5_init);
8518 module_exit(raid5_exit);
8519 MODULE_LICENSE("GPL");
8520 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8521 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8522 MODULE_ALIAS("md-raid5");
8523 MODULE_ALIAS("md-raid4");
8524 MODULE_ALIAS("md-level-5");
8525 MODULE_ALIAS("md-level-4");
8526 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8527 MODULE_ALIAS("md-raid6");
8528 MODULE_ALIAS("md-level-6");
8529
8530 /* This used to be two separate modules, they were: */
8531 MODULE_ALIAS("raid5");
8532 MODULE_ALIAS("raid6");