drivers/md/raid5.c

   1 /*
   2  * raid5.c : Multiple Devices driver for Linux
   3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   4  *         Copyright (C) 1999, 2000 Ingo Molnar
   5  *         Copyright (C) 2002, 2003 H. Peter Anvin
   6  *
   7  * RAID-4/5/6 management functions.
   8  * Thanks to Penguin Computing for making the RAID-6 development possible
   9  * by donating a test server!
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 /*
  22  * BITMAP UNPLUGGING:
  23  *
  24  * The sequencing for updating the bitmap reliably is a little
  25  * subtle (and I got it wrong the first time) so it deserves some
  26  * explanation.
  27  *
  28  * We group bitmap updates into batches.  Each batch has a number.
  29  * We may write out several batches at once, but that isn't very important.
  30  * conf->seq_write is the number of the last batch successfully written.
  31  * conf->seq_flush is the number of the last batch that was closed to
  32  *    new additions.
  33  * When we discover that we will need to write to any block in a stripe
  34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  35  * the number of the batch it will be in. This is seq_flush+1.
  36  * When we are ready to do a write, if that batch hasn't been written yet,
  37  *   we plug the array and queue the stripe for later.
  38  * When an unplug happens, we increment bm_flush, thus closing the current
  39  *   batch.
  40  * When we notice that bm_flush > bm_write, we write out all pending updates
  41  * to the bitmap, and advance bm_write to where bm_flush was.
  42  * This may occasionally write a bit out twice, but is sure never to
  43  * miss any bits.
  44  */
  45
  46 #include <linux/blkdev.h>
  47 #include <linux/kthread.h>
  48 #include <linux/raid/pq.h>
  49 #include <linux/async_tx.h>
  50 #include <linux/module.h>
  51 #include <linux/async.h>
  52 #include <linux/seq_file.h>
  53 #include <linux/cpu.h>
  54 #include <linux/slab.h>
  55 #include <linux/ratelimit.h>
  56 #include <linux/nodemask.h>
  57 #include <linux/flex_array.h>
  58 #include <linux/sched/signal.h>
  59
  60 #include <trace/events/block.h>
  61 #include <linux/list_sort.h>
  62
  63 #include "md.h"
  64 #include "raid5.h"
  65 #include "raid0.h"
  66 #include "bitmap.h"
  67 #include "raid5-log.h"
  68
  69 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
  70
  71 #define cpu_to_group(cpu) cpu_to_node(cpu)
  72 #define ANY_GROUP NUMA_NO_NODE
  73
  74 static bool devices_handle_discard_safely = false;
  75 module_param(devices_handle_discard_safely, bool, 0644);
  76 MODULE_PARM_DESC(devices_handle_discard_safely,
  77                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  78 static struct workqueue_struct *raid5_wq;
  79
  80 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  81 {
  82         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  83         return &conf->stripe_hashtbl[hash];
  84 }
  85
  86 static inline int stripe_hash_locks_hash(sector_t sect)
  87 {
  88         return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
  89 }
  90
  91 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  92 {
  93         spin_lock_irq(conf->hash_locks + hash);
  94         spin_lock(&conf->device_lock);
  95 }
  96
  97 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  98 {
  99         spin_unlock(&conf->device_lock);
 100         spin_unlock_irq(conf->hash_locks + hash);
 101 }
 102
 103 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 104 {
 105         int i;
 106         spin_lock_irq(conf->hash_locks);
 107         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 108                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 109         spin_lock(&conf->device_lock);
 110 }
 111
 112 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 113 {
 114         int i;
 115         spin_unlock(&conf->device_lock);
 116         for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 117                 spin_unlock(conf->hash_locks + i);
 118         spin_unlock_irq(conf->hash_locks);
 119 }
 120
 121 /* Find first data disk in a raid6 stripe */
 122 static inline int raid6_d0(struct stripe_head *sh)
 123 {
 124         if (sh->ddf_layout)
 125                 /* ddf always start from first device */
 126                 return 0;
 127         /* md starts just after Q block */
 128         if (sh->qd_idx == sh->disks - 1)
 129                 return 0;
 130         else
 131                 return sh->qd_idx + 1;
 132 }
 133 static inline int raid6_next_disk(int disk, int raid_disks)
 134 {
 135         disk++;
 136         return (disk < raid_disks) ? disk : 0;
 137 }
 138
 139 /* When walking through the disks in a raid5, starting at raid6_d0,
 140  * We need to map each disk to a 'slot', where the data disks are slot
 141  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 142  * is raid_disks-1.  This help does that mapping.
 143  */
 144 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 145                              int *count, int syndrome_disks)
 146 {
 147         int slot = *count;
 148
 149         if (sh->ddf_layout)
 150                 (*count)++;
 151         if (idx == sh->pd_idx)
 152                 return syndrome_disks;
 153         if (idx == sh->qd_idx)
 154                 return syndrome_disks + 1;
 155         if (!sh->ddf_layout)
 156                 (*count)++;
 157         return slot;
 158 }
 159
 160 static void print_raid5_conf (struct r5conf *conf);
 161
 162 static int stripe_operations_active(struct stripe_head *sh)
 163 {
 164         return sh->check_state || sh->reconstruct_state ||
 165                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 166                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 167 }
 168
 169 static bool stripe_is_lowprio(struct stripe_head *sh)
 170 {
 171         return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 172                 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 173                !test_bit(STRIPE_R5C_CACHING, &sh->state);
 174 }
 175
 176 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 177 {
 178         struct r5conf *conf = sh->raid_conf;
 179         struct r5worker_group *group;
 180         int thread_cnt;
 181         int i, cpu = sh->cpu;
 182
 183         if (!cpu_online(cpu)) {
 184                 cpu = cpumask_any(cpu_online_mask);
 185                 sh->cpu = cpu;
 186         }
 187
 188         if (list_empty(&sh->lru)) {
 189                 struct r5worker_group *group;
 190                 group = conf->worker_groups + cpu_to_group(cpu);
 191                 if (stripe_is_lowprio(sh))
 192                         list_add_tail(&sh->lru, &group->loprio_list);
 193                 else
 194                         list_add_tail(&sh->lru, &group->handle_list);
 195                 group->stripes_cnt++;
 196                 sh->group = group;
 197         }
 198
 199         if (conf->worker_cnt_per_group == 0) {
 200                 md_wakeup_thread(conf->mddev->thread);
 201                 return;
 202         }
 203
 204         group = conf->worker_groups + cpu_to_group(sh->cpu);
 205
 206         group->workers[0].working = true;
 207         /* at least one worker should run to avoid race */
 208         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 209
 210         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 211         /* wakeup more workers */
 212         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 213                 if (group->workers[i].working == false) {
 214                         group->workers[i].working = true;
 215                         queue_work_on(sh->cpu, raid5_wq,
 216                                       &group->workers[i].work);
 217                         thread_cnt--;
 218                 }
 219         }
 220 }
 221
 222 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 223                               struct list_head *temp_inactive_list)
 224 {
 225         int i;
 226         int injournal = 0;      /* number of date pages with R5_InJournal */
 227
 228         BUG_ON(!list_empty(&sh->lru));
 229         BUG_ON(atomic_read(&conf->active_stripes)==0);
 230
 231         if (r5c_is_writeback(conf->log))
 232                 for (i = sh->disks; i--; )
 233                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 234                                 injournal++;
 235         /*
 236          * In the following cases, the stripe cannot be released to cached
 237          * lists. Therefore, we make the stripe write out and set
 238          * STRIPE_HANDLE:
 239          *   1. when quiesce in r5c write back;
 240          *   2. when resync is requested fot the stripe.
 241          */
 242         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 243             (conf->quiesce && r5c_is_writeback(conf->log) &&
 244              !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 245                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 246                         r5c_make_stripe_write_out(sh);
 247                 set_bit(STRIPE_HANDLE, &sh->state);
 248         }
 249
 250         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 251                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 252                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 253                         list_add_tail(&sh->lru, &conf->delayed_list);
 254                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 255                            sh->bm_seq - conf->seq_write > 0)
 256                         list_add_tail(&sh->lru, &conf->bitmap_list);
 257                 else {
 258                         clear_bit(STRIPE_DELAYED, &sh->state);
 259                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 260                         if (conf->worker_cnt_per_group == 0) {
 261                                 if (stripe_is_lowprio(sh))
 262                                         list_add_tail(&sh->lru,
 263                                                         &conf->loprio_list);
 264                                 else
 265                                         list_add_tail(&sh->lru,
 266                                                         &conf->handle_list);
 267                         } else {
 268                                 raid5_wakeup_stripe_thread(sh);
 269                                 return;
 270                         }
 271                 }
 272                 md_wakeup_thread(conf->mddev->thread);
 273         } else {
 274                 BUG_ON(stripe_operations_active(sh));
 275                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 276                         if (atomic_dec_return(&conf->preread_active_stripes)
 277                             < IO_THRESHOLD)
 278                                 md_wakeup_thread(conf->mddev->thread);
 279                 atomic_dec(&conf->active_stripes);
 280                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 281                         if (!r5c_is_writeback(conf->log))
 282                                 list_add_tail(&sh->lru, temp_inactive_list);
 283                         else {
 284                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 285                                 if (injournal == 0)
 286                                         list_add_tail(&sh->lru, temp_inactive_list);
 287                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 288                                         /* full stripe */
 289                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 290                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 291                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 292                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 293                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 294                                         r5c_check_cached_full_stripe(conf);
 295                                 } else
 296                                         /*
 297                                          * STRIPE_R5C_PARTIAL_STRIPE is set in
 298                                          * r5c_try_caching_write(). No need to
 299                                          * set it again.
 300                                          */
 301                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 302                         }
 303                 }
 304         }
 305 }
 306
 307 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 308                              struct list_head *temp_inactive_list)
 309 {
 310         if (atomic_dec_and_test(&sh->count))
 311                 do_release_stripe(conf, sh, temp_inactive_list);
 312 }
 313
 314 /*
 315  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 316  *
 317  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 318  * given time. Adding stripes only takes device lock, while deleting stripes
 319  * only takes hash lock.
 320  */
 321 static void release_inactive_stripe_list(struct r5conf *conf,
 322                                          struct list_head *temp_inactive_list,
 323                                          int hash)
 324 {
 325         int size;
 326         bool do_wakeup = false;
 327         unsigned long flags;
 328
 329         if (hash == NR_STRIPE_HASH_LOCKS) {
 330                 size = NR_STRIPE_HASH_LOCKS;
 331                 hash = NR_STRIPE_HASH_LOCKS - 1;
 332         } else
 333                 size = 1;
 334         while (size) {
 335                 struct list_head *list = &temp_inactive_list[size - 1];
 336
 337                 /*
 338                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 339                  * remove stripes from the list
 340                  */
 341                 if (!list_empty_careful(list)) {
 342                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 343                         if (list_empty(conf->inactive_list + hash) &&
 344                             !list_empty(list))
 345                                 atomic_dec(&conf->empty_inactive_list_nr);
 346                         list_splice_tail_init(list, conf->inactive_list + hash);
 347                         do_wakeup = true;
 348                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 349                 }
 350                 size--;
 351                 hash--;
 352         }
 353
 354         if (do_wakeup) {
 355                 wake_up(&conf->wait_for_stripe);
 356                 if (atomic_read(&conf->active_stripes) == 0)
 357                         wake_up(&conf->wait_for_quiescent);
 358                 if (conf->retry_read_aligned)
 359                         md_wakeup_thread(conf->mddev->thread);
 360         }
 361 }
 362
 363 /* should hold conf->device_lock already */
 364 static int release_stripe_list(struct r5conf *conf,
 365                                struct list_head *temp_inactive_list)
 366 {
 367         struct stripe_head *sh, *t;
 368         int count = 0;
 369         struct llist_node *head;
 370
 371         head = llist_del_all(&conf->released_stripes);
 372         head = llist_reverse_order(head);
 373         llist_for_each_entry_safe(sh, t, head, release_list) {
 374                 int hash;
 375
 376                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 377                 smp_mb();
 378                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 379                 /*
 380                  * Don't worry the bit is set here, because if the bit is set
 381                  * again, the count is always > 1. This is true for
 382                  * STRIPE_ON_UNPLUG_LIST bit too.
 383                  */
 384                 hash = sh->hash_lock_index;
 385                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 386                 count++;
 387         }
 388
 389         return count;
 390 }
 391
 392 void raid5_release_stripe(struct stripe_head *sh)
 393 {
 394         struct r5conf *conf = sh->raid_conf;
 395         unsigned long flags;
 396         struct list_head list;
 397         int hash;
 398         bool wakeup;
 399
 400         /* Avoid release_list until the last reference.
 401          */
 402         if (atomic_add_unless(&sh->count, -1, 1))
 403                 return;
 404
 405         if (unlikely(!conf->mddev->thread) ||
 406                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 407                 goto slow_path;
 408         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 409         if (wakeup)
 410                 md_wakeup_thread(conf->mddev->thread);
 411         return;
 412 slow_path:
 413         local_irq_save(flags);
 414         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 415         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 416                 INIT_LIST_HEAD(&list);
 417                 hash = sh->hash_lock_index;
 418                 do_release_stripe(conf, sh, &list);
 419                 spin_unlock(&conf->device_lock);
 420                 release_inactive_stripe_list(conf, &list, hash);
 421         }
 422         local_irq_restore(flags);
 423 }
 424
 425 static inline void remove_hash(struct stripe_head *sh)
 426 {
 427         pr_debug("remove_hash(), stripe %llu\n",
 428                 (unsigned long long)sh->sector);
 429
 430         hlist_del_init(&sh->hash);
 431 }
 432
 433 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 434 {
 435         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 436
 437         pr_debug("insert_hash(), stripe %llu\n",
 438                 (unsigned long long)sh->sector);
 439
 440         hlist_add_head(&sh->hash, hp);
 441 }
 442
 443 /* find an idle stripe, make sure it is unhashed, and return it. */
 444 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 445 {
 446         struct stripe_head *sh = NULL;
 447         struct list_head *first;
 448
 449         if (list_empty(conf->inactive_list + hash))
 450                 goto out;
 451         first = (conf->inactive_list + hash)->next;
 452         sh = list_entry(first, struct stripe_head, lru);
 453         list_del_init(first);
 454         remove_hash(sh);
 455         atomic_inc(&conf->active_stripes);
 456         BUG_ON(hash != sh->hash_lock_index);
 457         if (list_empty(conf->inactive_list + hash))
 458                 atomic_inc(&conf->empty_inactive_list_nr);
 459 out:
 460         return sh;
 461 }
 462
 463 static void shrink_buffers(struct stripe_head *sh)
 464 {
 465         struct page *p;
 466         int i;
 467         int num = sh->raid_conf->pool_size;
 468
 469         for (i = 0; i < num ; i++) {
 470                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 471                 p = sh->dev[i].page;
 472                 if (!p)
 473                         continue;
 474                 sh->dev[i].page = NULL;
 475                 put_page(p);
 476         }
 477 }
 478
 479 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 480 {
 481         int i;
 482         int num = sh->raid_conf->pool_size;
 483
 484         for (i = 0; i < num; i++) {
 485                 struct page *page;
 486
 487                 if (!(page = alloc_page(gfp))) {
 488                         return 1;
 489                 }
 490                 sh->dev[i].page = page;
 491                 sh->dev[i].orig_page = page;
 492         }
 493
 494         return 0;
 495 }
 496
 497 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 498                             struct stripe_head *sh);
 499
 500 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 501 {
 502         struct r5conf *conf = sh->raid_conf;
 503         int i, seq;
 504
 505         BUG_ON(atomic_read(&sh->count) != 0);
 506         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 507         BUG_ON(stripe_operations_active(sh));
 508         BUG_ON(sh->batch_head);
 509
 510         pr_debug("init_stripe called, stripe %llu\n",
 511                 (unsigned long long)sector);
 512 retry:
 513         seq = read_seqcount_begin(&conf->gen_lock);
 514         sh->generation = conf->generation - previous;
 515         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 516         sh->sector = sector;
 517         stripe_set_idx(sector, conf, previous, sh);
 518         sh->state = 0;
 519
 520         for (i = sh->disks; i--; ) {
 521                 struct r5dev *dev = &sh->dev[i];
 522
 523                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 524                     test_bit(R5_LOCKED, &dev->flags)) {
 525                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 526                                (unsigned long long)sh->sector, i, dev->toread,
 527                                dev->read, dev->towrite, dev->written,
 528                                test_bit(R5_LOCKED, &dev->flags));
 529                         WARN_ON(1);
 530                 }
 531                 dev->flags = 0;
 532                 dev->sector = raid5_compute_blocknr(sh, i, previous);
 533         }
 534         if (read_seqcount_retry(&conf->gen_lock, seq))
 535                 goto retry;
 536         sh->overwrite_disks = 0;
 537         insert_hash(conf, sh);
 538         sh->cpu = smp_processor_id();
 539         set_bit(STRIPE_BATCH_READY, &sh->state);
 540 }
 541
 542 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 543                                          short generation)
 544 {
 545         struct stripe_head *sh;
 546
 547         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 548         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 549                 if (sh->sector == sector && sh->generation == generation)
 550                         return sh;
 551         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 552         return NULL;
 553 }
 554
 555 /*
 556  * Need to check if array has failed when deciding whether to:
 557  *  - start an array
 558  *  - remove non-faulty devices
 559  *  - add a spare
 560  *  - allow a reshape
 561  * This determination is simple when no reshape is happening.
 562  * However if there is a reshape, we need to carefully check
 563  * both the before and after sections.
 564  * This is because some failed devices may only affect one
 565  * of the two sections, and some non-in_sync devices may
 566  * be insync in the section most affected by failed devices.
 567  */
 568 int raid5_calc_degraded(struct r5conf *conf)
 569 {
 570         int degraded, degraded2;
 571         int i;
 572
 573         rcu_read_lock();
 574         degraded = 0;
 575         for (i = 0; i < conf->previous_raid_disks; i++) {
 576                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 577                 if (rdev && test_bit(Faulty, &rdev->flags))
 578                         rdev = rcu_dereference(conf->disks[i].replacement);
 579                 if (!rdev || test_bit(Faulty, &rdev->flags))
 580                         degraded++;
 581                 else if (test_bit(In_sync, &rdev->flags))
 582                         ;
 583                 else
 584                         /* not in-sync or faulty.
 585                          * If the reshape increases the number of devices,
 586                          * this is being recovered by the reshape, so
 587                          * this 'previous' section is not in_sync.
 588                          * If the number of devices is being reduced however,
 589                          * the device can only be part of the array if
 590                          * we are reverting a reshape, so this section will
 591                          * be in-sync.
 592                          */
 593                         if (conf->raid_disks >= conf->previous_raid_disks)
 594                                 degraded++;
 595         }
 596         rcu_read_unlock();
 597         if (conf->raid_disks == conf->previous_raid_disks)
 598                 return degraded;
 599         rcu_read_lock();
 600         degraded2 = 0;
 601         for (i = 0; i < conf->raid_disks; i++) {
 602                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 603                 if (rdev && test_bit(Faulty, &rdev->flags))
 604                         rdev = rcu_dereference(conf->disks[i].replacement);
 605                 if (!rdev || test_bit(Faulty, &rdev->flags))
 606                         degraded2++;
 607                 else if (test_bit(In_sync, &rdev->flags))
 608                         ;
 609                 else
 610                         /* not in-sync or faulty.
 611                          * If reshape increases the number of devices, this
 612                          * section has already been recovered, else it
 613                          * almost certainly hasn't.
 614                          */
 615                         if (conf->raid_disks <= conf->previous_raid_disks)
 616                                 degraded2++;
 617         }
 618         rcu_read_unlock();
 619         if (degraded2 > degraded)
 620                 return degraded2;
 621         return degraded;
 622 }
 623
 624 static int has_failed(struct r5conf *conf)
 625 {
 626         int degraded;
 627
 628         if (conf->mddev->reshape_position == MaxSector)
 629                 return conf->mddev->degraded > conf->max_degraded;
 630
 631         degraded = raid5_calc_degraded(conf);
 632         if (degraded > conf->max_degraded)
 633                 return 1;
 634         return 0;
 635 }
 636
 637 struct stripe_head *
 638 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 639                         int previous, int noblock, int noquiesce)
 640 {
 641         struct stripe_head *sh;
 642         int hash = stripe_hash_locks_hash(sector);
 643         int inc_empty_inactive_list_flag;
 644
 645         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 646
 647         spin_lock_irq(conf->hash_locks + hash);
 648
 649         do {
 650                 wait_event_lock_irq(conf->wait_for_quiescent,
 651                                     conf->quiesce == 0 || noquiesce,
 652                                     *(conf->hash_locks + hash));
 653                 sh = __find_stripe(conf, sector, conf->generation - previous);
 654                 if (!sh) {
 655                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 656                                 sh = get_free_stripe(conf, hash);
 657                                 if (!sh && !test_bit(R5_DID_ALLOC,
 658                                                      &conf->cache_state))
 659                                         set_bit(R5_ALLOC_MORE,
 660                                                 &conf->cache_state);
 661                         }
 662                         if (noblock && sh == NULL)
 663                                 break;
 664
 665                         r5c_check_stripe_cache_usage(conf);
 666                         if (!sh) {
 667                                 set_bit(R5_INACTIVE_BLOCKED,
 668                                         &conf->cache_state);
 669                                 r5l_wake_reclaim(conf->log, 0);
 670                                 wait_event_lock_irq(
 671                                         conf->wait_for_stripe,
 672                                         !list_empty(conf->inactive_list + hash) &&
 673                                         (atomic_read(&conf->active_stripes)
 674                                          < (conf->max_nr_stripes * 3 / 4)
 675                                          || !test_bit(R5_INACTIVE_BLOCKED,
 676                                                       &conf->cache_state)),
 677                                         *(conf->hash_locks + hash));
 678                                 clear_bit(R5_INACTIVE_BLOCKED,
 679                                           &conf->cache_state);
 680                         } else {
 681                                 init_stripe(sh, sector, previous);
 682                                 atomic_inc(&sh->count);
 683                         }
 684                 } else if (!atomic_inc_not_zero(&sh->count)) {
 685                         spin_lock(&conf->device_lock);
 686                         if (!atomic_read(&sh->count)) {
 687                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 688                                         atomic_inc(&conf->active_stripes);
 689                                 BUG_ON(list_empty(&sh->lru) &&
 690                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 691                                 inc_empty_inactive_list_flag = 0;
 692                                 if (!list_empty(conf->inactive_list + hash))
 693                                         inc_empty_inactive_list_flag = 1;
 694                                 list_del_init(&sh->lru);
 695                                 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 696                                         atomic_inc(&conf->empty_inactive_list_nr);
 697                                 if (sh->group) {
 698                                         sh->group->stripes_cnt--;
 699                                         sh->group = NULL;
 700                                 }
 701                         }
 702                         atomic_inc(&sh->count);
 703                         spin_unlock(&conf->device_lock);
 704                 }
 705         } while (sh == NULL);
 706
 707         spin_unlock_irq(conf->hash_locks + hash);
 708         return sh;
 709 }
 710
 711 static bool is_full_stripe_write(struct stripe_head *sh)
 712 {
 713         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 714         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 715 }
 716
 717 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 718 {
 719         if (sh1 > sh2) {
 720                 spin_lock_irq(&sh2->stripe_lock);
 721                 spin_lock_nested(&sh1->stripe_lock, 1);
 722         } else {
 723                 spin_lock_irq(&sh1->stripe_lock);
 724                 spin_lock_nested(&sh2->stripe_lock, 1);
 725         }
 726 }
 727
 728 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 729 {
 730         spin_unlock(&sh1->stripe_lock);
 731         spin_unlock_irq(&sh2->stripe_lock);
 732 }
 733
 734 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 735 static bool stripe_can_batch(struct stripe_head *sh)
 736 {
 737         struct r5conf *conf = sh->raid_conf;
 738
 739         if (conf->log || raid5_has_ppl(conf))
 740                 return false;
 741         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 742                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 743                 is_full_stripe_write(sh);
 744 }
 745
 746 /* we only do back search */
 747 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 748 {
 749         struct stripe_head *head;
 750         sector_t head_sector, tmp_sec;
 751         int hash;
 752         int dd_idx;
 753         int inc_empty_inactive_list_flag;
 754
 755         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 756         tmp_sec = sh->sector;
 757         if (!sector_div(tmp_sec, conf->chunk_sectors))
 758                 return;
 759         head_sector = sh->sector - STRIPE_SECTORS;
 760
 761         hash = stripe_hash_locks_hash(head_sector);
 762         spin_lock_irq(conf->hash_locks + hash);
 763         head = __find_stripe(conf, head_sector, conf->generation);
 764         if (head && !atomic_inc_not_zero(&head->count)) {
 765                 spin_lock(&conf->device_lock);
 766                 if (!atomic_read(&head->count)) {
 767                         if (!test_bit(STRIPE_HANDLE, &head->state))
 768                                 atomic_inc(&conf->active_stripes);
 769                         BUG_ON(list_empty(&head->lru) &&
 770                                !test_bit(STRIPE_EXPANDING, &head->state));
 771                         inc_empty_inactive_list_flag = 0;
 772                         if (!list_empty(conf->inactive_list + hash))
 773                                 inc_empty_inactive_list_flag = 1;
 774                         list_del_init(&head->lru);
 775                         if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 776                                 atomic_inc(&conf->empty_inactive_list_nr);
 777                         if (head->group) {
 778                                 head->group->stripes_cnt--;
 779                                 head->group = NULL;
 780                         }
 781                 }
 782                 atomic_inc(&head->count);
 783                 spin_unlock(&conf->device_lock);
 784         }
 785         spin_unlock_irq(conf->hash_locks + hash);
 786
 787         if (!head)
 788                 return;
 789         if (!stripe_can_batch(head))
 790                 goto out;
 791
 792         lock_two_stripes(head, sh);
 793         /* clear_batch_ready clear the flag */
 794         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 795                 goto unlock_out;
 796
 797         if (sh->batch_head)
 798                 goto unlock_out;
 799
 800         dd_idx = 0;
 801         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 802                 dd_idx++;
 803         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 804             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 805                 goto unlock_out;
 806
 807         if (head->batch_head) {
 808                 spin_lock(&head->batch_head->batch_lock);
 809                 /* This batch list is already running */
 810                 if (!stripe_can_batch(head)) {
 811                         spin_unlock(&head->batch_head->batch_lock);
 812                         goto unlock_out;
 813                 }
 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
 822                 sh->batch_head = head->batch_head;
 823         } else {
 824                 head->batch_head = head;
 825                 sh->batch_head = head->batch_head;
 826                 spin_lock(&head->batch_lock);
 827                 list_add_tail(&sh->batch_list, &head->batch_list);
 828                 spin_unlock(&head->batch_lock);
 829         }
 830
 831         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 832                 if (atomic_dec_return(&conf->preread_active_stripes)
 833                     < IO_THRESHOLD)
 834                         md_wakeup_thread(conf->mddev->thread);
 835
 836         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 837                 int seq = sh->bm_seq;
 838                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 839                     sh->batch_head->bm_seq > seq)
 840                         seq = sh->batch_head->bm_seq;
 841                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 842                 sh->batch_head->bm_seq = seq;
 843         }
 844
 845         atomic_inc(&sh->count);
 846 unlock_out:
 847         unlock_two_stripes(head, sh);
 848 out:
 849         raid5_release_stripe(head);
 850 }
 851
 852 /* Determine if 'data_offset' or 'new_data_offset' should be used
 853  * in this stripe_head.
 854  */
 855 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 856 {
 857         sector_t progress = conf->reshape_progress;
 858         /* Need a memory barrier to make sure we see the value
 859          * of conf->generation, or ->data_offset that was set before
 860          * reshape_progress was updated.
 861          */
 862         smp_rmb();
 863         if (progress == MaxSector)
 864                 return 0;
 865         if (sh->generation == conf->generation - 1)
 866                 return 0;
 867         /* We are in a reshape, and this is a new-generation stripe,
 868          * so use new_data_offset.
 869          */
 870         return 1;
 871 }
 872
 873 static void dispatch_bio_list(struct bio_list *tmp)
 874 {
 875         struct bio *bio;
 876
 877         while ((bio = bio_list_pop(tmp)))
 878                 generic_make_request(bio);
 879 }
 880
 881 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
 882 {
 883         const struct r5pending_data *da = list_entry(a,
 884                                 struct r5pending_data, sibling);
 885         const struct r5pending_data *db = list_entry(b,
 886                                 struct r5pending_data, sibling);
 887         if (da->sector > db->sector)
 888                 return 1;
 889         if (da->sector < db->sector)
 890                 return -1;
 891         return 0;
 892 }
 893
 894 static void dispatch_defer_bios(struct r5conf *conf, int target,
 895                                 struct bio_list *list)
 896 {
 897         struct r5pending_data *data;
 898         struct list_head *first, *next = NULL;
 899         int cnt = 0;
 900
 901         if (conf->pending_data_cnt == 0)
 902                 return;
 903
 904         list_sort(NULL, &conf->pending_list, cmp_stripe);
 905
 906         first = conf->pending_list.next;
 907
 908         /* temporarily move the head */
 909         if (conf->next_pending_data)
 910                 list_move_tail(&conf->pending_list,
 911                                 &conf->next_pending_data->sibling);
 912
 913         while (!list_empty(&conf->pending_list)) {
 914                 data = list_first_entry(&conf->pending_list,
 915                         struct r5pending_data, sibling);
 916                 if (&data->sibling == first)
 917                         first = data->sibling.next;
 918                 next = data->sibling.next;
 919
 920                 bio_list_merge(list, &data->bios);
 921                 list_move(&data->sibling, &conf->free_list);
 922                 cnt++;
 923                 if (cnt >= target)
 924                         break;
 925         }
 926         conf->pending_data_cnt -= cnt;
 927         BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
 928
 929         if (next != &conf->pending_list)
 930                 conf->next_pending_data = list_entry(next,
 931                                 struct r5pending_data, sibling);
 932         else
 933                 conf->next_pending_data = NULL;
 934         /* list isn't empty */
 935         if (first != &conf->pending_list)
 936                 list_move_tail(&conf->pending_list, first);
 937 }
 938
 939 static void flush_deferred_bios(struct r5conf *conf)
 940 {
 941         struct bio_list tmp = BIO_EMPTY_LIST;
 942
 943         if (conf->pending_data_cnt == 0)
 944                 return;
 945
 946         spin_lock(&conf->pending_bios_lock);
 947         dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
 948         BUG_ON(conf->pending_data_cnt != 0);
 949         spin_unlock(&conf->pending_bios_lock);
 950
 951         dispatch_bio_list(&tmp);
 952 }
 953
 954 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
 955                                 struct bio_list *bios)
 956 {
 957         struct bio_list tmp = BIO_EMPTY_LIST;
 958         struct r5pending_data *ent;
 959
 960         spin_lock(&conf->pending_bios_lock);
 961         ent = list_first_entry(&conf->free_list, struct r5pending_data,
 962                                                         sibling);
 963         list_move_tail(&ent->sibling, &conf->pending_list);
 964         ent->sector = sector;
 965         bio_list_init(&ent->bios);
 966         bio_list_merge(&ent->bios, bios);
 967         conf->pending_data_cnt++;
 968         if (conf->pending_data_cnt >= PENDING_IO_MAX)
 969                 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
 970
 971         spin_unlock(&conf->pending_bios_lock);
 972
 973         dispatch_bio_list(&tmp);
 974 }
 975
 976 static void
 977 raid5_end_read_request(struct bio *bi);
 978 static void
 979 raid5_end_write_request(struct bio *bi);
 980
 981 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 982 {
 983         struct r5conf *conf = sh->raid_conf;
 984         int i, disks = sh->disks;
 985         struct stripe_head *head_sh = sh;
 986         struct bio_list pending_bios = BIO_EMPTY_LIST;
 987         bool should_defer;
 988
 989         might_sleep();
 990
 991         if (log_stripe(sh, s) == 0)
 992                 return;
 993
 994         should_defer = conf->batch_bio_dispatch && conf->group_cnt;
 995
 996         for (i = disks; i--; ) {
 997                 int op, op_flags = 0;
 998                 int replace_only = 0;
 999                 struct bio *bi, *rbi;
1000                 struct md_rdev *rdev, *rrdev = NULL;
1001
1002                 sh = head_sh;
1003                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1004                         op = REQ_OP_WRITE;
1005                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1006                                 op_flags = REQ_FUA;
1007                         if (test_bit(R5_Discard, &sh->dev[i].flags))
1008                                 op = REQ_OP_DISCARD;
1009                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1010                         op = REQ_OP_READ;
1011                 else if (test_and_clear_bit(R5_WantReplace,
1012                                             &sh->dev[i].flags)) {
1013                         op = REQ_OP_WRITE;
1014                         replace_only = 1;
1015                 } else
1016                         continue;
1017                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1018                         op_flags |= REQ_SYNC;
1019
1020 again:
1021                 bi = &sh->dev[i].req;
1022                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1023
1024                 rcu_read_lock();
1025                 rrdev = rcu_dereference(conf->disks[i].replacement);
1026                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1027                 rdev = rcu_dereference(conf->disks[i].rdev);
1028                 if (!rdev) {
1029                         rdev = rrdev;
1030                         rrdev = NULL;
1031                 }
1032                 if (op_is_write(op)) {
1033                         if (replace_only)
1034                                 rdev = NULL;
1035                         if (rdev == rrdev)
1036                                 /* We raced and saw duplicates */
1037                                 rrdev = NULL;
1038                 } else {
1039                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1040                                 rdev = rrdev;
1041                         rrdev = NULL;
1042                 }
1043
1044                 if (rdev && test_bit(Faulty, &rdev->flags))
1045                         rdev = NULL;
1046                 if (rdev)
1047                         atomic_inc(&rdev->nr_pending);
1048                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1049                         rrdev = NULL;
1050                 if (rrdev)
1051                         atomic_inc(&rrdev->nr_pending);
1052                 rcu_read_unlock();
1053
1054                 /* We have already checked bad blocks for reads.  Now
1055                  * need to check for writes.  We never accept write errors
1056                  * on the replacement, so we don't to check rrdev.
1057                  */
1058                 while (op_is_write(op) && rdev &&
1059                        test_bit(WriteErrorSeen, &rdev->flags)) {
1060                         sector_t first_bad;
1061                         int bad_sectors;
1062                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1063                                               &first_bad, &bad_sectors);
1064                         if (!bad)
1065                                 break;
1066
1067                         if (bad < 0) {
1068                                 set_bit(BlockedBadBlocks, &rdev->flags);
1069                                 if (!conf->mddev->external &&
1070                                     conf->mddev->sb_flags) {
1071                                         /* It is very unlikely, but we might
1072                                          * still need to write out the
1073                                          * bad block log - better give it
1074                                          * a chance*/
1075                                         md_check_recovery(conf->mddev);
1076                                 }
1077                                 /*
1078                                  * Because md_wait_for_blocked_rdev
1079                                  * will dec nr_pending, we must
1080                                  * increment it first.
1081                                  */
1082                                 atomic_inc(&rdev->nr_pending);
1083                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
1084                         } else {
1085                                 /* Acknowledged bad block - skip the write */
1086                                 rdev_dec_pending(rdev, conf->mddev);
1087                                 rdev = NULL;
1088                         }
1089                 }
1090
1091                 if (rdev) {
1092                         if (s->syncing || s->expanding || s->expanded
1093                             || s->replacing)
1094                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1095
1096                         set_bit(STRIPE_IO_STARTED, &sh->state);
1097
1098                         bi->bi_bdev = rdev->bdev;
1099                         bio_set_op_attrs(bi, op, op_flags);
1100                         bi->bi_end_io = op_is_write(op)
1101                                 ? raid5_end_write_request
1102                                 : raid5_end_read_request;
1103                         bi->bi_private = sh;
1104
1105                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1106                                 __func__, (unsigned long long)sh->sector,
1107                                 bi->bi_opf, i);
1108                         atomic_inc(&sh->count);
1109                         if (sh != head_sh)
1110                                 atomic_inc(&head_sh->count);
1111                         if (use_new_offset(conf, sh))
1112                                 bi->bi_iter.bi_sector = (sh->sector
1113                                                  + rdev->new_data_offset);
1114                         else
1115                                 bi->bi_iter.bi_sector = (sh->sector
1116                                                  + rdev->data_offset);
1117                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1118                                 bi->bi_opf |= REQ_NOMERGE;
1119
1120                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1121                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1122
1123                         if (!op_is_write(op) &&
1124                             test_bit(R5_InJournal, &sh->dev[i].flags))
1125                                 /*
1126                                  * issuing read for a page in journal, this
1127                                  * must be preparing for prexor in rmw; read
1128                                  * the data into orig_page
1129                                  */
1130                                 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1131                         else
1132                                 sh->dev[i].vec.bv_page = sh->dev[i].page;
1133                         bi->bi_vcnt = 1;
1134                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1135                         bi->bi_io_vec[0].bv_offset = 0;
1136                         bi->bi_iter.bi_size = STRIPE_SIZE;
1137                         /*
1138                          * If this is discard request, set bi_vcnt 0. We don't
1139                          * want to confuse SCSI because SCSI will replace payload
1140                          */
1141                         if (op == REQ_OP_DISCARD)
1142                                 bi->bi_vcnt = 0;
1143                         if (rrdev)
1144                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1145
1146                         if (conf->mddev->gendisk)
1147                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1148                                                       bi, disk_devt(conf->mddev->gendisk),
1149                                                       sh->dev[i].sector);
1150                         if (should_defer && op_is_write(op))
1151                                 bio_list_add(&pending_bios, bi);
1152                         else
1153                                 generic_make_request(bi);
1154                 }
1155                 if (rrdev) {
1156                         if (s->syncing || s->expanding || s->expanded
1157                             || s->replacing)
1158                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1159
1160                         set_bit(STRIPE_IO_STARTED, &sh->state);
1161
1162                         rbi->bi_bdev = rrdev->bdev;
1163                         bio_set_op_attrs(rbi, op, op_flags);
1164                         BUG_ON(!op_is_write(op));
1165                         rbi->bi_end_io = raid5_end_write_request;
1166                         rbi->bi_private = sh;
1167
1168                         pr_debug("%s: for %llu schedule op %d on "
1169                                  "replacement disc %d\n",
1170                                 __func__, (unsigned long long)sh->sector,
1171                                 rbi->bi_opf, i);
1172                         atomic_inc(&sh->count);
1173                         if (sh != head_sh)
1174                                 atomic_inc(&head_sh->count);
1175                         if (use_new_offset(conf, sh))
1176                                 rbi->bi_iter.bi_sector = (sh->sector
1177                                                   + rrdev->new_data_offset);
1178                         else
1179                                 rbi->bi_iter.bi_sector = (sh->sector
1180                                                   + rrdev->data_offset);
1181                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1182                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1183                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1184                         rbi->bi_vcnt = 1;
1185                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1186                         rbi->bi_io_vec[0].bv_offset = 0;
1187                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1188                         /*
1189                          * If this is discard request, set bi_vcnt 0. We don't
1190                          * want to confuse SCSI because SCSI will replace payload
1191                          */
1192                         if (op == REQ_OP_DISCARD)
1193                                 rbi->bi_vcnt = 0;
1194                         if (conf->mddev->gendisk)
1195                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1196                                                       rbi, disk_devt(conf->mddev->gendisk),
1197                                                       sh->dev[i].sector);
1198                         if (should_defer && op_is_write(op))
1199                                 bio_list_add(&pending_bios, rbi);
1200                         else
1201                                 generic_make_request(rbi);
1202                 }
1203                 if (!rdev && !rrdev) {
1204                         if (op_is_write(op))
1205                                 set_bit(STRIPE_DEGRADED, &sh->state);
1206                         pr_debug("skip op %d on disc %d for sector %llu\n",
1207                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1208                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1209                         set_bit(STRIPE_HANDLE, &sh->state);
1210                 }
1211
1212                 if (!head_sh->batch_head)
1213                         continue;
1214                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1215                                       batch_list);
1216                 if (sh != head_sh)
1217                         goto again;
1218         }
1219
1220         if (should_defer && !bio_list_empty(&pending_bios))
1221                 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1222 }
1223
1224 static struct dma_async_tx_descriptor *
1225 async_copy_data(int frombio, struct bio *bio, struct page **page,
1226         sector_t sector, struct dma_async_tx_descriptor *tx,
1227         struct stripe_head *sh, int no_skipcopy)
1228 {
1229         struct bio_vec bvl;
1230         struct bvec_iter iter;
1231         struct page *bio_page;
1232         int page_offset;
1233         struct async_submit_ctl submit;
1234         enum async_tx_flags flags = 0;
1235
1236         if (bio->bi_iter.bi_sector >= sector)
1237                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1238         else
1239                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1240
1241         if (frombio)
1242                 flags |= ASYNC_TX_FENCE;
1243         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1244
1245         bio_for_each_segment(bvl, bio, iter) {
1246                 int len = bvl.bv_len;
1247                 int clen;
1248                 int b_offset = 0;
1249
1250                 if (page_offset < 0) {
1251                         b_offset = -page_offset;
1252                         page_offset += b_offset;
1253                         len -= b_offset;
1254                 }
1255
1256                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1257                         clen = STRIPE_SIZE - page_offset;
1258                 else
1259                         clen = len;
1260
1261                 if (clen > 0) {
1262                         b_offset += bvl.bv_offset;
1263                         bio_page = bvl.bv_page;
1264                         if (frombio) {
1265                                 if (sh->raid_conf->skip_copy &&
1266                                     b_offset == 0 && page_offset == 0 &&
1267                                     clen == STRIPE_SIZE &&
1268                                     !no_skipcopy)
1269                                         *page = bio_page;
1270                                 else
1271                                         tx = async_memcpy(*page, bio_page, page_offset,
1272                                                   b_offset, clen, &submit);
1273                         } else
1274                                 tx = async_memcpy(bio_page, *page, b_offset,
1275                                                   page_offset, clen, &submit);
1276                 }
1277                 /* chain the operations */
1278                 submit.depend_tx = tx;
1279
1280                 if (clen < len) /* hit end of page */
1281                         break;
1282                 page_offset +=  len;
1283         }
1284
1285         return tx;
1286 }
1287
1288 static void ops_complete_biofill(void *stripe_head_ref)
1289 {
1290         struct stripe_head *sh = stripe_head_ref;
1291         int i;
1292
1293         pr_debug("%s: stripe %llu\n", __func__,
1294                 (unsigned long long)sh->sector);
1295
1296         /* clear completed biofills */
1297         for (i = sh->disks; i--; ) {
1298                 struct r5dev *dev = &sh->dev[i];
1299
1300                 /* acknowledge completion of a biofill operation */
1301                 /* and check if we need to reply to a read request,
1302                  * new R5_Wantfill requests are held off until
1303                  * !STRIPE_BIOFILL_RUN
1304                  */
1305                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1306                         struct bio *rbi, *rbi2;
1307
1308                         BUG_ON(!dev->read);
1309                         rbi = dev->read;
1310                         dev->read = NULL;
1311                         while (rbi && rbi->bi_iter.bi_sector <
1312                                 dev->sector + STRIPE_SECTORS) {
1313                                 rbi2 = r5_next_bio(rbi, dev->sector);
1314                                 bio_endio(rbi);
1315                                 rbi = rbi2;
1316                         }
1317                 }
1318         }
1319         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1320
1321         set_bit(STRIPE_HANDLE, &sh->state);
1322         raid5_release_stripe(sh);
1323 }
1324
1325 static void ops_run_biofill(struct stripe_head *sh)
1326 {
1327         struct dma_async_tx_descriptor *tx = NULL;
1328         struct async_submit_ctl submit;
1329         int i;
1330
1331         BUG_ON(sh->batch_head);
1332         pr_debug("%s: stripe %llu\n", __func__,
1333                 (unsigned long long)sh->sector);
1334
1335         for (i = sh->disks; i--; ) {
1336                 struct r5dev *dev = &sh->dev[i];
1337                 if (test_bit(R5_Wantfill, &dev->flags)) {
1338                         struct bio *rbi;
1339                         spin_lock_irq(&sh->stripe_lock);
1340                         dev->read = rbi = dev->toread;
1341                         dev->toread = NULL;
1342                         spin_unlock_irq(&sh->stripe_lock);
1343                         while (rbi && rbi->bi_iter.bi_sector <
1344                                 dev->sector + STRIPE_SECTORS) {
1345                                 tx = async_copy_data(0, rbi, &dev->page,
1346                                                      dev->sector, tx, sh, 0);
1347                                 rbi = r5_next_bio(rbi, dev->sector);
1348                         }
1349                 }
1350         }
1351
1352         atomic_inc(&sh->count);
1353         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1354         async_trigger_callback(&submit);
1355 }
1356
1357 static void mark_target_uptodate(struct stripe_head *sh, int target)
1358 {
1359         struct r5dev *tgt;
1360
1361         if (target < 0)
1362                 return;
1363
1364         tgt = &sh->dev[target];
1365         set_bit(R5_UPTODATE, &tgt->flags);
1366         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1367         clear_bit(R5_Wantcompute, &tgt->flags);
1368 }
1369
1370 static void ops_complete_compute(void *stripe_head_ref)
1371 {
1372         struct stripe_head *sh = stripe_head_ref;
1373
1374         pr_debug("%s: stripe %llu\n", __func__,
1375                 (unsigned long long)sh->sector);
1376
1377         /* mark the computed target(s) as uptodate */
1378         mark_target_uptodate(sh, sh->ops.target);
1379         mark_target_uptodate(sh, sh->ops.target2);
1380
1381         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1382         if (sh->check_state == check_state_compute_run)
1383                 sh->check_state = check_state_compute_result;
1384         set_bit(STRIPE_HANDLE, &sh->state);
1385         raid5_release_stripe(sh);
1386 }
1387
1388 /* return a pointer to the address conversion region of the scribble buffer */
1389 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1390                                  struct raid5_percpu *percpu, int i)
1391 {
1392         void *addr;
1393
1394         addr = flex_array_get(percpu->scribble, i);
1395         return addr + sizeof(struct page *) * (sh->disks + 2);
1396 }
1397
1398 /* return a pointer to the address conversion region of the scribble buffer */
1399 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1400 {
1401         void *addr;
1402
1403         addr = flex_array_get(percpu->scribble, i);
1404         return addr;
1405 }
1406
1407 static struct dma_async_tx_descriptor *
1408 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1409 {
1410         int disks = sh->disks;
1411         struct page **xor_srcs = to_addr_page(percpu, 0);
1412         int target = sh->ops.target;
1413         struct r5dev *tgt = &sh->dev[target];
1414         struct page *xor_dest = tgt->page;
1415         int count = 0;
1416         struct dma_async_tx_descriptor *tx;
1417         struct async_submit_ctl submit;
1418         int i;
1419
1420         BUG_ON(sh->batch_head);
1421
1422         pr_debug("%s: stripe %llu block: %d\n",
1423                 __func__, (unsigned long long)sh->sector, target);
1424         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1425
1426         for (i = disks; i--; )
1427                 if (i != target)
1428                         xor_srcs[count++] = sh->dev[i].page;
1429
1430         atomic_inc(&sh->count);
1431
1432         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1433                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1434         if (unlikely(count == 1))
1435                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1436         else
1437                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1438
1439         return tx;
1440 }
1441
1442 /* set_syndrome_sources - populate source buffers for gen_syndrome
1443  * @srcs - (struct page *) array of size sh->disks
1444  * @sh - stripe_head to parse
1445  *
1446  * Populates srcs in proper layout order for the stripe and returns the
1447  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1448  * destination buffer is recorded in srcs[count] and the Q destination
1449  * is recorded in srcs[count+1]].
1450  */
1451 static int set_syndrome_sources(struct page **srcs,
1452                                 struct stripe_head *sh,
1453                                 int srctype)
1454 {
1455         int disks = sh->disks;
1456         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1457         int d0_idx = raid6_d0(sh);
1458         int count;
1459         int i;
1460
1461         for (i = 0; i < disks; i++)
1462                 srcs[i] = NULL;
1463
1464         count = 0;
1465         i = d0_idx;
1466         do {
1467                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1468                 struct r5dev *dev = &sh->dev[i];
1469
1470                 if (i == sh->qd_idx || i == sh->pd_idx ||
1471                     (srctype == SYNDROME_SRC_ALL) ||
1472                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1473                      (test_bit(R5_Wantdrain, &dev->flags) ||
1474                       test_bit(R5_InJournal, &dev->flags))) ||
1475                     (srctype == SYNDROME_SRC_WRITTEN &&
1476                      (dev->written ||
1477                       test_bit(R5_InJournal, &dev->flags)))) {
1478                         if (test_bit(R5_InJournal, &dev->flags))
1479                                 srcs[slot] = sh->dev[i].orig_page;
1480                         else
1481                                 srcs[slot] = sh->dev[i].page;
1482                 }
1483                 i = raid6_next_disk(i, disks);
1484         } while (i != d0_idx);
1485
1486         return syndrome_disks;
1487 }
1488
1489 static struct dma_async_tx_descriptor *
1490 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1491 {
1492         int disks = sh->disks;
1493         struct page **blocks = to_addr_page(percpu, 0);
1494         int target;
1495         int qd_idx = sh->qd_idx;
1496         struct dma_async_tx_descriptor *tx;
1497         struct async_submit_ctl submit;
1498         struct r5dev *tgt;
1499         struct page *dest;
1500         int i;
1501         int count;
1502
1503         BUG_ON(sh->batch_head);
1504         if (sh->ops.target < 0)
1505                 target = sh->ops.target2;
1506         else if (sh->ops.target2 < 0)
1507                 target = sh->ops.target;
1508         else
1509                 /* we should only have one valid target */
1510                 BUG();
1511         BUG_ON(target < 0);
1512         pr_debug("%s: stripe %llu block: %d\n",
1513                 __func__, (unsigned long long)sh->sector, target);
1514
1515         tgt = &sh->dev[target];
1516         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1517         dest = tgt->page;
1518
1519         atomic_inc(&sh->count);
1520
1521         if (target == qd_idx) {
1522                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1523                 blocks[count] = NULL; /* regenerating p is not necessary */
1524                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1525                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1526                                   ops_complete_compute, sh,
1527                                   to_addr_conv(sh, percpu, 0));
1528                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1529         } else {
1530                 /* Compute any data- or p-drive using XOR */
1531                 count = 0;
1532                 for (i = disks; i-- ; ) {
1533                         if (i == target || i == qd_idx)
1534                                 continue;
1535                         blocks[count++] = sh->dev[i].page;
1536                 }
1537
1538                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1539                                   NULL, ops_complete_compute, sh,
1540                                   to_addr_conv(sh, percpu, 0));
1541                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1542         }
1543
1544         return tx;
1545 }
1546
1547 static struct dma_async_tx_descriptor *
1548 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1549 {
1550         int i, count, disks = sh->disks;
1551         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1552         int d0_idx = raid6_d0(sh);
1553         int faila = -1, failb = -1;
1554         int target = sh->ops.target;
1555         int target2 = sh->ops.target2;
1556         struct r5dev *tgt = &sh->dev[target];
1557         struct r5dev *tgt2 = &sh->dev[target2];
1558         struct dma_async_tx_descriptor *tx;
1559         struct page **blocks = to_addr_page(percpu, 0);
1560         struct async_submit_ctl submit;
1561
1562         BUG_ON(sh->batch_head);
1563         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1564                  __func__, (unsigned long long)sh->sector, target, target2);
1565         BUG_ON(target < 0 || target2 < 0);
1566         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1567         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1568
1569         /* we need to open-code set_syndrome_sources to handle the
1570          * slot number conversion for 'faila' and 'failb'
1571          */
1572         for (i = 0; i < disks ; i++)
1573                 blocks[i] = NULL;
1574         count = 0;
1575         i = d0_idx;
1576         do {
1577                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1578
1579                 blocks[slot] = sh->dev[i].page;
1580
1581                 if (i == target)
1582                         faila = slot;
1583                 if (i == target2)
1584                         failb = slot;
1585                 i = raid6_next_disk(i, disks);
1586         } while (i != d0_idx);
1587
1588         BUG_ON(faila == failb);
1589         if (failb < faila)
1590                 swap(faila, failb);
1591         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1592                  __func__, (unsigned long long)sh->sector, faila, failb);
1593
1594         atomic_inc(&sh->count);
1595
1596         if (failb == syndrome_disks+1) {
1597                 /* Q disk is one of the missing disks */
1598                 if (faila == syndrome_disks) {
1599                         /* Missing P+Q, just recompute */
1600                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1601                                           ops_complete_compute, sh,
1602                                           to_addr_conv(sh, percpu, 0));
1603                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1604                                                   STRIPE_SIZE, &submit);
1605                 } else {
1606                         struct page *dest;
1607                         int data_target;
1608                         int qd_idx = sh->qd_idx;
1609
1610                         /* Missing D+Q: recompute D from P, then recompute Q */
1611                         if (target == qd_idx)
1612                                 data_target = target2;
1613                         else
1614                                 data_target = target;
1615
1616                         count = 0;
1617                         for (i = disks; i-- ; ) {
1618                                 if (i == data_target || i == qd_idx)
1619                                         continue;
1620                                 blocks[count++] = sh->dev[i].page;
1621                         }
1622                         dest = sh->dev[data_target].page;
1623                         init_async_submit(&submit,
1624                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1625                                           NULL, NULL, NULL,
1626                                           to_addr_conv(sh, percpu, 0));
1627                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1628                                        &submit);
1629
1630                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1631                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1632                                           ops_complete_compute, sh,
1633                                           to_addr_conv(sh, percpu, 0));
1634                         return async_gen_syndrome(blocks, 0, count+2,
1635                                                   STRIPE_SIZE, &submit);
1636                 }
1637         } else {
1638                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1639                                   ops_complete_compute, sh,
1640                                   to_addr_conv(sh, percpu, 0));
1641                 if (failb == syndrome_disks) {
1642                         /* We're missing D+P. */
1643                         return async_raid6_datap_recov(syndrome_disks+2,
1644                                                        STRIPE_SIZE, faila,
1645                                                        blocks, &submit);
1646                 } else {
1647                         /* We're missing D+D. */
1648                         return async_raid6_2data_recov(syndrome_disks+2,
1649                                                        STRIPE_SIZE, faila, failb,
1650                                                        blocks, &submit);
1651                 }
1652         }
1653 }
1654
1655 static void ops_complete_prexor(void *stripe_head_ref)
1656 {
1657         struct stripe_head *sh = stripe_head_ref;
1658
1659         pr_debug("%s: stripe %llu\n", __func__,
1660                 (unsigned long long)sh->sector);
1661
1662         if (r5c_is_writeback(sh->raid_conf->log))
1663                 /*
1664                  * raid5-cache write back uses orig_page during prexor.
1665                  * After prexor, it is time to free orig_page
1666                  */
1667                 r5c_release_extra_page(sh);
1668 }
1669
1670 static struct dma_async_tx_descriptor *
1671 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1672                 struct dma_async_tx_descriptor *tx)
1673 {
1674         int disks = sh->disks;
1675         struct page **xor_srcs = to_addr_page(percpu, 0);
1676         int count = 0, pd_idx = sh->pd_idx, i;
1677         struct async_submit_ctl submit;
1678
1679         /* existing parity data subtracted */
1680         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1681
1682         BUG_ON(sh->batch_head);
1683         pr_debug("%s: stripe %llu\n", __func__,
1684                 (unsigned long long)sh->sector);
1685
1686         for (i = disks; i--; ) {
1687                 struct r5dev *dev = &sh->dev[i];
1688                 /* Only process blocks that are known to be uptodate */
1689                 if (test_bit(R5_InJournal, &dev->flags))
1690                         xor_srcs[count++] = dev->orig_page;
1691                 else if (test_bit(R5_Wantdrain, &dev->flags))
1692                         xor_srcs[count++] = dev->page;
1693         }
1694
1695         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1696                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1697         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1698
1699         return tx;
1700 }
1701
1702 static struct dma_async_tx_descriptor *
1703 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1704                 struct dma_async_tx_descriptor *tx)
1705 {
1706         struct page **blocks = to_addr_page(percpu, 0);
1707         int count;
1708         struct async_submit_ctl submit;
1709
1710         pr_debug("%s: stripe %llu\n", __func__,
1711                 (unsigned long long)sh->sector);
1712
1713         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1714
1715         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1716                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1717         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1718
1719         return tx;
1720 }
1721
1722 static struct dma_async_tx_descriptor *
1723 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1724 {
1725         struct r5conf *conf = sh->raid_conf;
1726         int disks = sh->disks;
1727         int i;
1728         struct stripe_head *head_sh = sh;
1729
1730         pr_debug("%s: stripe %llu\n", __func__,
1731                 (unsigned long long)sh->sector);
1732
1733         for (i = disks; i--; ) {
1734                 struct r5dev *dev;
1735                 struct bio *chosen;
1736
1737                 sh = head_sh;
1738                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1739                         struct bio *wbi;
1740
1741 again:
1742                         dev = &sh->dev[i];
1743                         /*
1744                          * clear R5_InJournal, so when rewriting a page in
1745                          * journal, it is not skipped by r5l_log_stripe()
1746                          */
1747                         clear_bit(R5_InJournal, &dev->flags);
1748                         spin_lock_irq(&sh->stripe_lock);
1749                         chosen = dev->towrite;
1750                         dev->towrite = NULL;
1751                         sh->overwrite_disks = 0;
1752                         BUG_ON(dev->written);
1753                         wbi = dev->written = chosen;
1754                         spin_unlock_irq(&sh->stripe_lock);
1755                         WARN_ON(dev->page != dev->orig_page);
1756
1757                         while (wbi && wbi->bi_iter.bi_sector <
1758                                 dev->sector + STRIPE_SECTORS) {
1759                                 if (wbi->bi_opf & REQ_FUA)
1760                                         set_bit(R5_WantFUA, &dev->flags);
1761                                 if (wbi->bi_opf & REQ_SYNC)
1762                                         set_bit(R5_SyncIO, &dev->flags);
1763                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1764                                         set_bit(R5_Discard, &dev->flags);
1765                                 else {
1766                                         tx = async_copy_data(1, wbi, &dev->page,
1767                                                              dev->sector, tx, sh,
1768                                                              r5c_is_writeback(conf->log));
1769                                         if (dev->page != dev->orig_page &&
1770                                             !r5c_is_writeback(conf->log)) {
1771                                                 set_bit(R5_SkipCopy, &dev->flags);
1772                                                 clear_bit(R5_UPTODATE, &dev->flags);
1773                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1774                                         }
1775                                 }
1776                                 wbi = r5_next_bio(wbi, dev->sector);
1777                         }
1778
1779                         if (head_sh->batch_head) {
1780                                 sh = list_first_entry(&sh->batch_list,
1781                                                       struct stripe_head,
1782                                                       batch_list);
1783                                 if (sh == head_sh)
1784                                         continue;
1785                                 goto again;
1786                         }
1787                 }
1788         }
1789
1790         return tx;
1791 }
1792
1793 static void ops_complete_reconstruct(void *stripe_head_ref)
1794 {
1795         struct stripe_head *sh = stripe_head_ref;
1796         int disks = sh->disks;
1797         int pd_idx = sh->pd_idx;
1798         int qd_idx = sh->qd_idx;
1799         int i;
1800         bool fua = false, sync = false, discard = false;
1801
1802         pr_debug("%s: stripe %llu\n", __func__,
1803                 (unsigned long long)sh->sector);
1804
1805         for (i = disks; i--; ) {
1806                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1807                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1808                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1809         }
1810
1811         for (i = disks; i--; ) {
1812                 struct r5dev *dev = &sh->dev[i];
1813
1814                 if (dev->written || i == pd_idx || i == qd_idx) {
1815                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1816                                 set_bit(R5_UPTODATE, &dev->flags);
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         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2192
2193         if (conf->mddev->gendisk)
2194                 sprintf(conf->cache_name[0],
2195                         "raid%d-%s", conf->level, mdname(conf->mddev));
2196         else
2197                 sprintf(conf->cache_name[0],
2198                         "raid%d-%p", conf->level, conf->mddev);
2199         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2200
2201         conf->active_name = 0;
2202         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2203                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2204                                0, 0, NULL);
2205         if (!sc)
2206                 return 1;
2207         conf->slab_cache = sc;
2208         conf->pool_size = devs;
2209         while (num--)
2210                 if (!grow_one_stripe(conf, GFP_KERNEL))
2211                         return 1;
2212
2213         return 0;
2214 }
2215
2216 /**
2217  * scribble_len - return the required size of the scribble region
2218  * @num - total number of disks in the array
2219  *
2220  * The size must be enough to contain:
2221  * 1/ a struct page pointer for each device in the array +2
2222  * 2/ room to convert each entry in (1) to its corresponding dma
2223  *    (dma_map_page()) or page (page_address()) address.
2224  *
2225  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2226  * calculate over all devices (not just the data blocks), using zeros in place
2227  * of the P and Q blocks.
2228  */
2229 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2230 {
2231         struct flex_array *ret;
2232         size_t len;
2233
2234         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2235         ret = flex_array_alloc(len, cnt, flags);
2236         if (!ret)
2237                 return NULL;
2238         /* always prealloc all elements, so no locking is required */
2239         if (flex_array_prealloc(ret, 0, cnt, flags)) {
2240                 flex_array_free(ret);
2241                 return NULL;
2242         }
2243         return ret;
2244 }
2245
2246 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2247 {
2248         unsigned long cpu;
2249         int err = 0;
2250
2251         /*
2252          * Never shrink. And mddev_suspend() could deadlock if this is called
2253          * from raid5d. In that case, scribble_disks and scribble_sectors
2254          * should equal to new_disks and new_sectors
2255          */
2256         if (conf->scribble_disks >= new_disks &&
2257             conf->scribble_sectors >= new_sectors)
2258                 return 0;
2259         mddev_suspend(conf->mddev);
2260         get_online_cpus();
2261         for_each_present_cpu(cpu) {
2262                 struct raid5_percpu *percpu;
2263                 struct flex_array *scribble;
2264
2265                 percpu = per_cpu_ptr(conf->percpu, cpu);
2266                 scribble = scribble_alloc(new_disks,
2267                                           new_sectors / STRIPE_SECTORS,
2268                                           GFP_NOIO);
2269
2270                 if (scribble) {
2271                         flex_array_free(percpu->scribble);
2272                         percpu->scribble = scribble;
2273                 } else {
2274                         err = -ENOMEM;
2275                         break;
2276                 }
2277         }
2278         put_online_cpus();
2279         mddev_resume(conf->mddev);
2280         if (!err) {
2281                 conf->scribble_disks = new_disks;
2282                 conf->scribble_sectors = new_sectors;
2283         }
2284         return err;
2285 }
2286
2287 static int resize_stripes(struct r5conf *conf, int newsize)
2288 {
2289         /* Make all the stripes able to hold 'newsize' devices.
2290          * New slots in each stripe get 'page' set to a new page.
2291          *
2292          * This happens in stages:
2293          * 1/ create a new kmem_cache and allocate the required number of
2294          *    stripe_heads.
2295          * 2/ gather all the old stripe_heads and transfer the pages across
2296          *    to the new stripe_heads.  This will have the side effect of
2297          *    freezing the array as once all stripe_heads have been collected,
2298          *    no IO will be possible.  Old stripe heads are freed once their
2299          *    pages have been transferred over, and the old kmem_cache is
2300          *    freed when all stripes are done.
2301          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2302          *    we simple return a failure status - no need to clean anything up.
2303          * 4/ allocate new pages for the new slots in the new stripe_heads.
2304          *    If this fails, we don't bother trying the shrink the
2305          *    stripe_heads down again, we just leave them as they are.
2306          *    As each stripe_head is processed the new one is released into
2307          *    active service.
2308          *
2309          * Once step2 is started, we cannot afford to wait for a write,
2310          * so we use GFP_NOIO allocations.
2311          */
2312         struct stripe_head *osh, *nsh;
2313         LIST_HEAD(newstripes);
2314         struct disk_info *ndisks;
2315         int err = 0;
2316         struct kmem_cache *sc;
2317         int i;
2318         int hash, cnt;
2319
2320         md_allow_write(conf->mddev);
2321
2322         /* Step 1 */
2323         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2324                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2325                                0, 0, NULL);
2326         if (!sc)
2327                 return -ENOMEM;
2328
2329         /* Need to ensure auto-resizing doesn't interfere */
2330         mutex_lock(&conf->cache_size_mutex);
2331
2332         for (i = conf->max_nr_stripes; i; i--) {
2333                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2334                 if (!nsh)
2335                         break;
2336
2337                 list_add(&nsh->lru, &newstripes);
2338         }
2339         if (i) {
2340                 /* didn't get enough, give up */
2341                 while (!list_empty(&newstripes)) {
2342                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2343                         list_del(&nsh->lru);
2344                         free_stripe(sc, nsh);
2345                 }
2346                 kmem_cache_destroy(sc);
2347                 mutex_unlock(&conf->cache_size_mutex);
2348                 return -ENOMEM;
2349         }
2350         /* Step 2 - Must use GFP_NOIO now.
2351          * OK, we have enough stripes, start collecting inactive
2352          * stripes and copying them over
2353          */
2354         hash = 0;
2355         cnt = 0;
2356         list_for_each_entry(nsh, &newstripes, lru) {
2357                 lock_device_hash_lock(conf, hash);
2358                 wait_event_cmd(conf->wait_for_stripe,
2359                                     !list_empty(conf->inactive_list + hash),
2360                                     unlock_device_hash_lock(conf, hash),
2361                                     lock_device_hash_lock(conf, hash));
2362                 osh = get_free_stripe(conf, hash);
2363                 unlock_device_hash_lock(conf, hash);
2364
2365                 for(i=0; i<conf->pool_size; i++) {
2366                         nsh->dev[i].page = osh->dev[i].page;
2367                         nsh->dev[i].orig_page = osh->dev[i].page;
2368                 }
2369                 nsh->hash_lock_index = hash;
2370                 free_stripe(conf->slab_cache, osh);
2371                 cnt++;
2372                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2373                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2374                         hash++;
2375                         cnt = 0;
2376                 }
2377         }
2378         kmem_cache_destroy(conf->slab_cache);
2379
2380         /* Step 3.
2381          * At this point, we are holding all the stripes so the array
2382          * is completely stalled, so now is a good time to resize
2383          * conf->disks and the scribble region
2384          */
2385         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2386         if (ndisks) {
2387                 for (i = 0; i < conf->pool_size; i++)
2388                         ndisks[i] = conf->disks[i];
2389
2390                 for (i = conf->pool_size; i < newsize; i++) {
2391                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2392                         if (!ndisks[i].extra_page)
2393                                 err = -ENOMEM;
2394                 }
2395
2396                 if (err) {
2397                         for (i = conf->pool_size; i < newsize; i++)
2398                                 if (ndisks[i].extra_page)
2399                                         put_page(ndisks[i].extra_page);
2400                         kfree(ndisks);
2401                 } else {
2402                         kfree(conf->disks);
2403                         conf->disks = ndisks;
2404                 }
2405         } else
2406                 err = -ENOMEM;
2407
2408         mutex_unlock(&conf->cache_size_mutex);
2409
2410         conf->slab_cache = sc;
2411         conf->active_name = 1-conf->active_name;
2412
2413         /* Step 4, return new stripes to service */
2414         while(!list_empty(&newstripes)) {
2415                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2416                 list_del_init(&nsh->lru);
2417
2418                 for (i=conf->raid_disks; i < newsize; i++)
2419                         if (nsh->dev[i].page == NULL) {
2420                                 struct page *p = alloc_page(GFP_NOIO);
2421                                 nsh->dev[i].page = p;
2422                                 nsh->dev[i].orig_page = p;
2423                                 if (!p)
2424                                         err = -ENOMEM;
2425                         }
2426                 raid5_release_stripe(nsh);
2427         }
2428         /* critical section pass, GFP_NOIO no longer needed */
2429
2430         if (!err)
2431                 conf->pool_size = newsize;
2432         return err;
2433 }
2434
2435 static int drop_one_stripe(struct r5conf *conf)
2436 {
2437         struct stripe_head *sh;
2438         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2439
2440         spin_lock_irq(conf->hash_locks + hash);
2441         sh = get_free_stripe(conf, hash);
2442         spin_unlock_irq(conf->hash_locks + hash);
2443         if (!sh)
2444                 return 0;
2445         BUG_ON(atomic_read(&sh->count));
2446         shrink_buffers(sh);
2447         free_stripe(conf->slab_cache, sh);
2448         atomic_dec(&conf->active_stripes);
2449         conf->max_nr_stripes--;
2450         return 1;
2451 }
2452
2453 static void shrink_stripes(struct r5conf *conf)
2454 {
2455         while (conf->max_nr_stripes &&
2456                drop_one_stripe(conf))
2457                 ;
2458
2459         kmem_cache_destroy(conf->slab_cache);
2460         conf->slab_cache = NULL;
2461 }
2462
2463 static void raid5_end_read_request(struct bio * bi)
2464 {
2465         struct stripe_head *sh = bi->bi_private;
2466         struct r5conf *conf = sh->raid_conf;
2467         int disks = sh->disks, i;
2468         char b[BDEVNAME_SIZE];
2469         struct md_rdev *rdev = NULL;
2470         sector_t s;
2471
2472         for (i=0 ; i<disks; i++)
2473                 if (bi == &sh->dev[i].req)
2474                         break;
2475
2476         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2477                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2478                 bi->bi_status);
2479         if (i == disks) {
2480                 bio_reset(bi);
2481                 BUG();
2482                 return;
2483         }
2484         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2485                 /* If replacement finished while this request was outstanding,
2486                  * 'replacement' might be NULL already.
2487                  * In that case it moved down to 'rdev'.
2488                  * rdev is not removed until all requests are finished.
2489                  */
2490                 rdev = conf->disks[i].replacement;
2491         if (!rdev)
2492                 rdev = conf->disks[i].rdev;
2493
2494         if (use_new_offset(conf, sh))
2495                 s = sh->sector + rdev->new_data_offset;
2496         else
2497                 s = sh->sector + rdev->data_offset;
2498         if (!bi->bi_status) {
2499                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2500                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2501                         /* Note that this cannot happen on a
2502                          * replacement device.  We just fail those on
2503                          * any error
2504                          */
2505                         pr_info_ratelimited(
2506                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2507                                 mdname(conf->mddev), STRIPE_SECTORS,
2508                                 (unsigned long long)s,
2509                                 bdevname(rdev->bdev, b));
2510                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2511                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2512                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2513                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2514                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2515
2516                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2517                         /*
2518                          * end read for a page in journal, this
2519                          * must be preparing for prexor in rmw
2520                          */
2521                         set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2522
2523                 if (atomic_read(&rdev->read_errors))
2524                         atomic_set(&rdev->read_errors, 0);
2525         } else {
2526                 const char *bdn = bdevname(rdev->bdev, b);
2527                 int retry = 0;
2528                 int set_bad = 0;
2529
2530                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2531                 atomic_inc(&rdev->read_errors);
2532                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2533                         pr_warn_ratelimited(
2534                                 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2535                                 mdname(conf->mddev),
2536                                 (unsigned long long)s,
2537                                 bdn);
2538                 else if (conf->mddev->degraded >= conf->max_degraded) {
2539                         set_bad = 1;
2540                         pr_warn_ratelimited(
2541                                 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2542                                 mdname(conf->mddev),
2543                                 (unsigned long long)s,
2544                                 bdn);
2545                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2546                         /* Oh, no!!! */
2547                         set_bad = 1;
2548                         pr_warn_ratelimited(
2549                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2550                                 mdname(conf->mddev),
2551                                 (unsigned long long)s,
2552                                 bdn);
2553                 } else if (atomic_read(&rdev->read_errors)
2554                          > conf->max_nr_stripes)
2555                         pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2556                                mdname(conf->mddev), bdn);
2557                 else
2558                         retry = 1;
2559                 if (set_bad && test_bit(In_sync, &rdev->flags)
2560                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2561                         retry = 1;
2562                 if (retry)
2563                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2564                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2565                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2566                         } else
2567                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2568                 else {
2569                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2570                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2571                         if (!(set_bad
2572                               && test_bit(In_sync, &rdev->flags)
2573                               && rdev_set_badblocks(
2574                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2575                                 md_error(conf->mddev, rdev);
2576                 }
2577         }
2578         rdev_dec_pending(rdev, conf->mddev);
2579         bio_reset(bi);
2580         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2581         set_bit(STRIPE_HANDLE, &sh->state);
2582         raid5_release_stripe(sh);
2583 }
2584
2585 static void raid5_end_write_request(struct bio *bi)
2586 {
2587         struct stripe_head *sh = bi->bi_private;
2588         struct r5conf *conf = sh->raid_conf;
2589         int disks = sh->disks, i;
2590         struct md_rdev *uninitialized_var(rdev);
2591         sector_t first_bad;
2592         int bad_sectors;
2593         int replacement = 0;
2594
2595         for (i = 0 ; i < disks; i++) {
2596                 if (bi == &sh->dev[i].req) {
2597                         rdev = conf->disks[i].rdev;
2598                         break;
2599                 }
2600                 if (bi == &sh->dev[i].rreq) {
2601                         rdev = conf->disks[i].replacement;
2602                         if (rdev)
2603                                 replacement = 1;
2604                         else
2605                                 /* rdev was removed and 'replacement'
2606                                  * replaced it.  rdev is not removed
2607                                  * until all requests are finished.
2608                                  */
2609                                 rdev = conf->disks[i].rdev;
2610                         break;
2611                 }
2612         }
2613         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2614                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2615                 bi->bi_status);
2616         if (i == disks) {
2617                 bio_reset(bi);
2618                 BUG();
2619                 return;
2620         }
2621
2622         if (replacement) {
2623                 if (bi->bi_status)
2624                         md_error(conf->mddev, rdev);
2625                 else if (is_badblock(rdev, sh->sector,
2626                                      STRIPE_SECTORS,
2627                                      &first_bad, &bad_sectors))
2628                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2629         } else {
2630                 if (bi->bi_status) {
2631                         set_bit(STRIPE_DEGRADED, &sh->state);
2632                         set_bit(WriteErrorSeen, &rdev->flags);
2633                         set_bit(R5_WriteError, &sh->dev[i].flags);
2634                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2635                                 set_bit(MD_RECOVERY_NEEDED,
2636                                         &rdev->mddev->recovery);
2637                 } else if (is_badblock(rdev, sh->sector,
2638                                        STRIPE_SECTORS,
2639                                        &first_bad, &bad_sectors)) {
2640                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2641                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2642                                 /* That was a successful write so make
2643                                  * sure it looks like we already did
2644                                  * a re-write.
2645                                  */
2646                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2647                 }
2648         }
2649         rdev_dec_pending(rdev, conf->mddev);
2650
2651         if (sh->batch_head && bi->bi_status && !replacement)
2652                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2653
2654         bio_reset(bi);
2655         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2656                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2657         set_bit(STRIPE_HANDLE, &sh->state);
2658         raid5_release_stripe(sh);
2659
2660         if (sh->batch_head && sh != sh->batch_head)
2661                 raid5_release_stripe(sh->batch_head);
2662 }
2663
2664 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2665 {
2666         char b[BDEVNAME_SIZE];
2667         struct r5conf *conf = mddev->private;
2668         unsigned long flags;
2669         pr_debug("raid456: error called\n");
2670
2671         spin_lock_irqsave(&conf->device_lock, flags);
2672         clear_bit(In_sync, &rdev->flags);
2673         mddev->degraded = raid5_calc_degraded(conf);
2674         spin_unlock_irqrestore(&conf->device_lock, flags);
2675         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2676
2677         set_bit(Blocked, &rdev->flags);
2678         set_bit(Faulty, &rdev->flags);
2679         set_mask_bits(&mddev->sb_flags, 0,
2680                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2681         pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2682                 "md/raid:%s: Operation continuing on %d devices.\n",
2683                 mdname(mddev),
2684                 bdevname(rdev->bdev, b),
2685                 mdname(mddev),
2686                 conf->raid_disks - mddev->degraded);
2687         r5c_update_on_rdev_error(mddev, rdev);
2688 }
2689
2690 /*
2691  * Input: a 'big' sector number,
2692  * Output: index of the data and parity disk, and the sector # in them.
2693  */
2694 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2695                               int previous, int *dd_idx,
2696                               struct stripe_head *sh)
2697 {
2698         sector_t stripe, stripe2;
2699         sector_t chunk_number;
2700         unsigned int chunk_offset;
2701         int pd_idx, qd_idx;
2702         int ddf_layout = 0;
2703         sector_t new_sector;
2704         int algorithm = previous ? conf->prev_algo
2705                                  : conf->algorithm;
2706         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2707                                          : conf->chunk_sectors;
2708         int raid_disks = previous ? conf->previous_raid_disks
2709                                   : conf->raid_disks;
2710         int data_disks = raid_disks - conf->max_degraded;
2711
2712         /* First compute the information on this sector */
2713
2714         /*
2715          * Compute the chunk number and the sector offset inside the chunk
2716          */
2717         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2718         chunk_number = r_sector;
2719
2720         /*
2721          * Compute the stripe number
2722          */
2723         stripe = chunk_number;
2724         *dd_idx = sector_div(stripe, data_disks);
2725         stripe2 = stripe;
2726         /*
2727          * Select the parity disk based on the user selected algorithm.
2728          */
2729         pd_idx = qd_idx = -1;
2730         switch(conf->level) {
2731         case 4:
2732                 pd_idx = data_disks;
2733                 break;
2734         case 5:
2735                 switch (algorithm) {
2736                 case ALGORITHM_LEFT_ASYMMETRIC:
2737                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2738                         if (*dd_idx >= pd_idx)
2739                                 (*dd_idx)++;
2740                         break;
2741                 case ALGORITHM_RIGHT_ASYMMETRIC:
2742                         pd_idx = sector_div(stripe2, raid_disks);
2743                         if (*dd_idx >= pd_idx)
2744                                 (*dd_idx)++;
2745                         break;
2746                 case ALGORITHM_LEFT_SYMMETRIC:
2747                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2748                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2749                         break;
2750                 case ALGORITHM_RIGHT_SYMMETRIC:
2751                         pd_idx = sector_div(stripe2, raid_disks);
2752                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2753                         break;
2754                 case ALGORITHM_PARITY_0:
2755                         pd_idx = 0;
2756                         (*dd_idx)++;
2757                         break;
2758                 case ALGORITHM_PARITY_N:
2759                         pd_idx = data_disks;
2760                         break;
2761                 default:
2762                         BUG();
2763                 }
2764                 break;
2765         case 6:
2766
2767                 switch (algorithm) {
2768                 case ALGORITHM_LEFT_ASYMMETRIC:
2769                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2770                         qd_idx = pd_idx + 1;
2771                         if (pd_idx == raid_disks-1) {
2772                                 (*dd_idx)++;    /* Q D D D P */
2773                                 qd_idx = 0;
2774                         } else if (*dd_idx >= pd_idx)
2775                                 (*dd_idx) += 2; /* D D P Q D */
2776                         break;
2777                 case ALGORITHM_RIGHT_ASYMMETRIC:
2778                         pd_idx = sector_div(stripe2, raid_disks);
2779                         qd_idx = pd_idx + 1;
2780                         if (pd_idx == raid_disks-1) {
2781                                 (*dd_idx)++;    /* Q D D D P */
2782                                 qd_idx = 0;
2783                         } else if (*dd_idx >= pd_idx)
2784                                 (*dd_idx) += 2; /* D D P Q D */
2785                         break;
2786                 case ALGORITHM_LEFT_SYMMETRIC:
2787                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2788                         qd_idx = (pd_idx + 1) % raid_disks;
2789                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2790                         break;
2791                 case ALGORITHM_RIGHT_SYMMETRIC:
2792                         pd_idx = sector_div(stripe2, raid_disks);
2793                         qd_idx = (pd_idx + 1) % raid_disks;
2794                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2795                         break;
2796
2797                 case ALGORITHM_PARITY_0:
2798                         pd_idx = 0;
2799                         qd_idx = 1;
2800                         (*dd_idx) += 2;
2801                         break;
2802                 case ALGORITHM_PARITY_N:
2803                         pd_idx = data_disks;
2804                         qd_idx = data_disks + 1;
2805                         break;
2806
2807                 case ALGORITHM_ROTATING_ZERO_RESTART:
2808                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2809                          * of blocks for computing Q is different.
2810                          */
2811                         pd_idx = sector_div(stripe2, raid_disks);
2812                         qd_idx = pd_idx + 1;
2813                         if (pd_idx == raid_disks-1) {
2814                                 (*dd_idx)++;    /* Q D D D P */
2815                                 qd_idx = 0;
2816                         } else if (*dd_idx >= pd_idx)
2817                                 (*dd_idx) += 2; /* D D P Q D */
2818                         ddf_layout = 1;
2819                         break;
2820
2821                 case ALGORITHM_ROTATING_N_RESTART:
2822                         /* Same a left_asymmetric, by first stripe is
2823                          * D D D P Q  rather than
2824                          * Q D D D P
2825                          */
2826                         stripe2 += 1;
2827                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2828                         qd_idx = pd_idx + 1;
2829                         if (pd_idx == raid_disks-1) {
2830                                 (*dd_idx)++;    /* Q D D D P */
2831                                 qd_idx = 0;
2832                         } else if (*dd_idx >= pd_idx)
2833                                 (*dd_idx) += 2; /* D D P Q D */
2834                         ddf_layout = 1;
2835                         break;
2836
2837                 case ALGORITHM_ROTATING_N_CONTINUE:
2838                         /* Same as left_symmetric but Q is before P */
2839                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2840                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2841                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2842                         ddf_layout = 1;
2843                         break;
2844
2845                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2846                         /* RAID5 left_asymmetric, with Q on last device */
2847                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2848                         if (*dd_idx >= pd_idx)
2849                                 (*dd_idx)++;
2850                         qd_idx = raid_disks - 1;
2851                         break;
2852
2853                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2854                         pd_idx = sector_div(stripe2, raid_disks-1);
2855                         if (*dd_idx >= pd_idx)
2856                                 (*dd_idx)++;
2857                         qd_idx = raid_disks - 1;
2858                         break;
2859
2860                 case ALGORITHM_LEFT_SYMMETRIC_6:
2861                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2862                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2863                         qd_idx = raid_disks - 1;
2864                         break;
2865
2866                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2867                         pd_idx = sector_div(stripe2, raid_disks-1);
2868                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2869                         qd_idx = raid_disks - 1;
2870                         break;
2871
2872                 case ALGORITHM_PARITY_0_6:
2873                         pd_idx = 0;
2874                         (*dd_idx)++;
2875                         qd_idx = raid_disks - 1;
2876                         break;
2877
2878                 default:
2879                         BUG();
2880                 }
2881                 break;
2882         }
2883
2884         if (sh) {
2885                 sh->pd_idx = pd_idx;
2886                 sh->qd_idx = qd_idx;
2887                 sh->ddf_layout = ddf_layout;
2888         }
2889         /*
2890          * Finally, compute the new sector number
2891          */
2892         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2893         return new_sector;
2894 }
2895
2896 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2897 {
2898         struct r5conf *conf = sh->raid_conf;
2899         int raid_disks = sh->disks;
2900         int data_disks = raid_disks - conf->max_degraded;
2901         sector_t new_sector = sh->sector, check;
2902         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2903                                          : conf->chunk_sectors;
2904         int algorithm = previous ? conf->prev_algo
2905                                  : conf->algorithm;
2906         sector_t stripe;
2907         int chunk_offset;
2908         sector_t chunk_number;
2909         int dummy1, dd_idx = i;
2910         sector_t r_sector;
2911         struct stripe_head sh2;
2912
2913         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2914         stripe = new_sector;
2915
2916         if (i == sh->pd_idx)
2917                 return 0;
2918         switch(conf->level) {
2919         case 4: break;
2920         case 5:
2921                 switch (algorithm) {
2922                 case ALGORITHM_LEFT_ASYMMETRIC:
2923                 case ALGORITHM_RIGHT_ASYMMETRIC:
2924                         if (i > sh->pd_idx)
2925                                 i--;
2926                         break;
2927                 case ALGORITHM_LEFT_SYMMETRIC:
2928                 case ALGORITHM_RIGHT_SYMMETRIC:
2929                         if (i < sh->pd_idx)
2930                                 i += raid_disks;
2931                         i -= (sh->pd_idx + 1);
2932                         break;
2933                 case ALGORITHM_PARITY_0:
2934                         i -= 1;
2935                         break;
2936                 case ALGORITHM_PARITY_N:
2937                         break;
2938                 default:
2939                         BUG();
2940                 }
2941                 break;
2942         case 6:
2943                 if (i == sh->qd_idx)
2944                         return 0; /* It is the Q disk */
2945                 switch (algorithm) {
2946                 case ALGORITHM_LEFT_ASYMMETRIC:
2947                 case ALGORITHM_RIGHT_ASYMMETRIC:
2948                 case ALGORITHM_ROTATING_ZERO_RESTART:
2949                 case ALGORITHM_ROTATING_N_RESTART:
2950                         if (sh->pd_idx == raid_disks-1)
2951                                 i--;    /* Q D D D P */
2952                         else if (i > sh->pd_idx)
2953                                 i -= 2; /* D D P Q D */
2954                         break;
2955                 case ALGORITHM_LEFT_SYMMETRIC:
2956                 case ALGORITHM_RIGHT_SYMMETRIC:
2957                         if (sh->pd_idx == raid_disks-1)
2958                                 i--; /* Q D D D P */
2959                         else {
2960                                 /* D D P Q D */
2961                                 if (i < sh->pd_idx)
2962                                         i += raid_disks;
2963                                 i -= (sh->pd_idx + 2);
2964                         }
2965                         break;
2966                 case ALGORITHM_PARITY_0:
2967                         i -= 2;
2968                         break;
2969                 case ALGORITHM_PARITY_N:
2970                         break;
2971                 case ALGORITHM_ROTATING_N_CONTINUE:
2972                         /* Like left_symmetric, but P is before Q */
2973                         if (sh->pd_idx == 0)
2974                                 i--;    /* P D D D Q */
2975                         else {
2976                                 /* D D Q P D */
2977                                 if (i < sh->pd_idx)
2978                                         i += raid_disks;
2979                                 i -= (sh->pd_idx + 1);
2980                         }
2981                         break;
2982                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2983                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2984                         if (i > sh->pd_idx)
2985                                 i--;
2986                         break;
2987                 case ALGORITHM_LEFT_SYMMETRIC_6:
2988                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2989                         if (i < sh->pd_idx)
2990                                 i += data_disks + 1;
2991                         i -= (sh->pd_idx + 1);
2992                         break;
2993                 case ALGORITHM_PARITY_0_6:
2994                         i -= 1;
2995                         break;
2996                 default:
2997                         BUG();
2998                 }
2999                 break;
3000         }
3001
3002         chunk_number = stripe * data_disks + i;
3003         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3004
3005         check = raid5_compute_sector(conf, r_sector,
3006                                      previous, &dummy1, &sh2);
3007         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3008                 || sh2.qd_idx != sh->qd_idx) {
3009                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3010                         mdname(conf->mddev));
3011                 return 0;
3012         }
3013         return r_sector;
3014 }
3015
3016 /*
3017  * There are cases where we want handle_stripe_dirtying() and
3018  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3019  *
3020  * This function checks whether we want to delay the towrite. Specifically,
3021  * we delay the towrite when:
3022  *
3023  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3024  *      stripe has data in journal (for other devices).
3025  *
3026  *      In this case, when reading data for the non-overwrite dev, it is
3027  *      necessary to handle complex rmw of write back cache (prexor with
3028  *      orig_page, and xor with page). To keep read path simple, we would
3029  *      like to flush data in journal to RAID disks first, so complex rmw
3030  *      is handled in the write patch (handle_stripe_dirtying).
3031  *
3032  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3033  *
3034  *      It is important to be able to flush all stripes in raid5-cache.
3035  *      Therefore, we need reserve some space on the journal device for
3036  *      these flushes. If flush operation includes pending writes to the
3037  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3038  *      for the flush out. If we exclude these pending writes from flush
3039  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3040  *      Therefore, excluding pending writes in these cases enables more
3041  *      efficient use of the journal device.
3042  *
3043  *      Note: To make sure the stripe makes progress, we only delay
3044  *      towrite for stripes with data already in journal (injournal > 0).
3045  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3046  *      no_space_stripes list.
3047  *
3048  *   3. during journal failure
3049  *      In journal failure, we try to flush all cached data to raid disks
3050  *      based on data in stripe cache. The array is read-only to upper
3051  *      layers, so we would skip all pending writes.
3052  *
3053  */
3054 static inline bool delay_towrite(struct r5conf *conf,
3055                                  struct r5dev *dev,
3056                                  struct stripe_head_state *s)
3057 {
3058         /* case 1 above */
3059         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3060             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3061                 return true;
3062         /* case 2 above */
3063         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3064             s->injournal > 0)
3065                 return true;
3066         /* case 3 above */
3067         if (s->log_failed && s->injournal)
3068                 return true;
3069         return false;
3070 }
3071
3072 static void
3073 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3074                          int rcw, int expand)
3075 {
3076         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3077         struct r5conf *conf = sh->raid_conf;
3078         int level = conf->level;
3079
3080         if (rcw) {
3081                 /*
3082                  * In some cases, handle_stripe_dirtying initially decided to
3083                  * run rmw and allocates extra page for prexor. However, rcw is
3084                  * cheaper later on. We need to free the extra page now,
3085                  * because we won't be able to do that in ops_complete_prexor().
3086                  */
3087                 r5c_release_extra_page(sh);
3088
3089                 for (i = disks; i--; ) {
3090                         struct r5dev *dev = &sh->dev[i];
3091
3092                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3093                                 set_bit(R5_LOCKED, &dev->flags);
3094                                 set_bit(R5_Wantdrain, &dev->flags);
3095                                 if (!expand)
3096                                         clear_bit(R5_UPTODATE, &dev->flags);
3097                                 s->locked++;
3098                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3099                                 set_bit(R5_LOCKED, &dev->flags);
3100                                 s->locked++;
3101                         }
3102                 }
3103                 /* if we are not expanding this is a proper write request, and
3104                  * there will be bios with new data to be drained into the
3105                  * stripe cache
3106                  */
3107                 if (!expand) {
3108                         if (!s->locked)
3109                                 /* False alarm, nothing to do */
3110                                 return;
3111                         sh->reconstruct_state = reconstruct_state_drain_run;
3112                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3113                 } else
3114                         sh->reconstruct_state = reconstruct_state_run;
3115
3116                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3117
3118                 if (s->locked + conf->max_degraded == disks)
3119                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3120                                 atomic_inc(&conf->pending_full_writes);
3121         } else {
3122                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3123                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3124                 BUG_ON(level == 6 &&
3125                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3126                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3127
3128                 for (i = disks; i--; ) {
3129                         struct r5dev *dev = &sh->dev[i];
3130                         if (i == pd_idx || i == qd_idx)
3131                                 continue;
3132
3133                         if (dev->towrite &&
3134                             (test_bit(R5_UPTODATE, &dev->flags) ||
3135                              test_bit(R5_Wantcompute, &dev->flags))) {
3136                                 set_bit(R5_Wantdrain, &dev->flags);
3137                                 set_bit(R5_LOCKED, &dev->flags);
3138                                 clear_bit(R5_UPTODATE, &dev->flags);
3139                                 s->locked++;
3140                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3141                                 set_bit(R5_LOCKED, &dev->flags);
3142                                 s->locked++;
3143                         }
3144                 }
3145                 if (!s->locked)
3146                         /* False alarm - nothing to do */
3147                         return;
3148                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3149                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3150                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3151                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3152         }
3153
3154         /* keep the parity disk(s) locked while asynchronous operations
3155          * are in flight
3156          */
3157         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3158         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3159         s->locked++;
3160
3161         if (level == 6) {
3162                 int qd_idx = sh->qd_idx;
3163                 struct r5dev *dev = &sh->dev[qd_idx];
3164
3165                 set_bit(R5_LOCKED, &dev->flags);
3166                 clear_bit(R5_UPTODATE, &dev->flags);
3167                 s->locked++;
3168         }
3169
3170         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3171             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3172             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3173             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3174                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3175
3176         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3177                 __func__, (unsigned long long)sh->sector,
3178                 s->locked, s->ops_request);
3179 }
3180
3181 /*
3182  * Each stripe/dev can have one or more bion attached.
3183  * toread/towrite point to the first in a chain.
3184  * The bi_next chain must be in order.
3185  */
3186 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3187                           int forwrite, int previous)
3188 {
3189         struct bio **bip;
3190         struct r5conf *conf = sh->raid_conf;
3191         int firstwrite=0;
3192
3193         pr_debug("adding bi b#%llu to stripe s#%llu\n",
3194                 (unsigned long long)bi->bi_iter.bi_sector,
3195                 (unsigned long long)sh->sector);
3196
3197         spin_lock_irq(&sh->stripe_lock);
3198         /* Don't allow new IO added to stripes in batch list */
3199         if (sh->batch_head)
3200                 goto overlap;
3201         if (forwrite) {
3202                 bip = &sh->dev[dd_idx].towrite;
3203                 if (*bip == NULL)
3204                         firstwrite = 1;
3205         } else
3206                 bip = &sh->dev[dd_idx].toread;
3207         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3208                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3209                         goto overlap;
3210                 bip = & (*bip)->bi_next;
3211         }
3212         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3213                 goto overlap;
3214
3215         if (forwrite && raid5_has_ppl(conf)) {
3216                 /*
3217                  * With PPL only writes to consecutive data chunks within a
3218                  * stripe are allowed because for a single stripe_head we can
3219                  * only have one PPL entry at a time, which describes one data
3220                  * range. Not really an overlap, but wait_for_overlap can be
3221                  * used to handle this.
3222                  */
3223                 sector_t sector;
3224                 sector_t first = 0;
3225                 sector_t last = 0;
3226                 int count = 0;
3227                 int i;
3228
3229                 for (i = 0; i < sh->disks; i++) {
3230                         if (i != sh->pd_idx &&
3231                             (i == dd_idx || sh->dev[i].towrite)) {
3232                                 sector = sh->dev[i].sector;
3233                                 if (count == 0 || sector < first)
3234                                         first = sector;
3235                                 if (sector > last)
3236                                         last = sector;
3237                                 count++;
3238                         }
3239                 }
3240
3241                 if (first + conf->chunk_sectors * (count - 1) != last)
3242                         goto overlap;
3243         }
3244
3245         if (!forwrite || previous)
3246                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3247
3248         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3249         if (*bip)
3250                 bi->bi_next = *bip;
3251         *bip = bi;
3252         bio_inc_remaining(bi);
3253         md_write_inc(conf->mddev, bi);
3254
3255         if (forwrite) {
3256                 /* check if page is covered */
3257                 sector_t sector = sh->dev[dd_idx].sector;
3258                 for (bi=sh->dev[dd_idx].towrite;
3259                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3260                              bi && bi->bi_iter.bi_sector <= sector;
3261                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3262                         if (bio_end_sector(bi) >= sector)
3263                                 sector = bio_end_sector(bi);
3264                 }
3265                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3266                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3267                                 sh->overwrite_disks++;
3268         }
3269
3270         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3271                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3272                 (unsigned long long)sh->sector, dd_idx);
3273
3274         if (conf->mddev->bitmap && firstwrite) {
3275                 /* Cannot hold spinlock over bitmap_startwrite,
3276                  * but must ensure this isn't added to a batch until
3277                  * we have added to the bitmap and set bm_seq.
3278                  * So set STRIPE_BITMAP_PENDING to prevent
3279                  * batching.
3280                  * If multiple add_stripe_bio() calls race here they
3281                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3282                  * to complete "bitmap_startwrite" gets to set
3283                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3284                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3285                  * any more.
3286                  */
3287                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3288                 spin_unlock_irq(&sh->stripe_lock);
3289                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3290                                   STRIPE_SECTORS, 0);
3291                 spin_lock_irq(&sh->stripe_lock);
3292                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3293                 if (!sh->batch_head) {
3294                         sh->bm_seq = conf->seq_flush+1;
3295                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3296                 }
3297         }
3298         spin_unlock_irq(&sh->stripe_lock);
3299
3300         if (stripe_can_batch(sh))
3301                 stripe_add_to_batch_list(conf, sh);
3302         return 1;
3303
3304  overlap:
3305         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3306         spin_unlock_irq(&sh->stripe_lock);
3307         return 0;
3308 }
3309
3310 static void end_reshape(struct r5conf *conf);
3311
3312 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3313                             struct stripe_head *sh)
3314 {
3315         int sectors_per_chunk =
3316                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3317         int dd_idx;
3318         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3319         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3320
3321         raid5_compute_sector(conf,
3322                              stripe * (disks - conf->max_degraded)
3323                              *sectors_per_chunk + chunk_offset,
3324                              previous,
3325                              &dd_idx, sh);
3326 }
3327
3328 static void
3329 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3330                      struct stripe_head_state *s, int disks)
3331 {
3332         int i;
3333         BUG_ON(sh->batch_head);
3334         for (i = disks; i--; ) {
3335                 struct bio *bi;
3336                 int bitmap_end = 0;
3337
3338                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3339                         struct md_rdev *rdev;
3340                         rcu_read_lock();
3341                         rdev = rcu_dereference(conf->disks[i].rdev);
3342                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3343                             !test_bit(Faulty, &rdev->flags))
3344                                 atomic_inc(&rdev->nr_pending);
3345                         else
3346                                 rdev = NULL;
3347                         rcu_read_unlock();
3348                         if (rdev) {
3349                                 if (!rdev_set_badblocks(
3350                                             rdev,
3351                                             sh->sector,
3352                                             STRIPE_SECTORS, 0))
3353                                         md_error(conf->mddev, rdev);
3354                                 rdev_dec_pending(rdev, conf->mddev);
3355                         }
3356                 }
3357                 spin_lock_irq(&sh->stripe_lock);
3358                 /* fail all writes first */
3359                 bi = sh->dev[i].towrite;
3360                 sh->dev[i].towrite = NULL;
3361                 sh->overwrite_disks = 0;
3362                 spin_unlock_irq(&sh->stripe_lock);
3363                 if (bi)
3364                         bitmap_end = 1;
3365
3366                 log_stripe_write_finished(sh);
3367
3368                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3369                         wake_up(&conf->wait_for_overlap);
3370
3371                 while (bi && bi->bi_iter.bi_sector <
3372                         sh->dev[i].sector + STRIPE_SECTORS) {
3373                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3374
3375                         md_write_end(conf->mddev);
3376                         bio_io_error(bi);
3377                         bi = nextbi;
3378                 }
3379                 if (bitmap_end)
3380                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3381                                 STRIPE_SECTORS, 0, 0);
3382                 bitmap_end = 0;
3383                 /* and fail all 'written' */
3384                 bi = sh->dev[i].written;
3385                 sh->dev[i].written = NULL;
3386                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3387                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3388                         sh->dev[i].page = sh->dev[i].orig_page;
3389                 }
3390
3391                 if (bi) bitmap_end = 1;
3392                 while (bi && bi->bi_iter.bi_sector <
3393                        sh->dev[i].sector + STRIPE_SECTORS) {
3394                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3395
3396                         md_write_end(conf->mddev);
3397                         bio_io_error(bi);
3398                         bi = bi2;
3399                 }
3400
3401                 /* fail any reads if this device is non-operational and
3402                  * the data has not reached the cache yet.
3403                  */
3404                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3405                     s->failed > conf->max_degraded &&
3406                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3407                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3408                         spin_lock_irq(&sh->stripe_lock);
3409                         bi = sh->dev[i].toread;
3410                         sh->dev[i].toread = NULL;
3411                         spin_unlock_irq(&sh->stripe_lock);
3412                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3413                                 wake_up(&conf->wait_for_overlap);
3414                         if (bi)
3415                                 s->to_read--;
3416                         while (bi && bi->bi_iter.bi_sector <
3417                                sh->dev[i].sector + STRIPE_SECTORS) {
3418                                 struct bio *nextbi =
3419                                         r5_next_bio(bi, sh->dev[i].sector);
3420
3421                                 bio_io_error(bi);
3422                                 bi = nextbi;
3423                         }
3424                 }
3425                 if (bitmap_end)
3426                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3427                                         STRIPE_SECTORS, 0, 0);
3428                 /* If we were in the middle of a write the parity block might
3429                  * still be locked - so just clear all R5_LOCKED flags
3430                  */
3431                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3432         }
3433         s->to_write = 0;
3434         s->written = 0;
3435
3436         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3437                 if (atomic_dec_and_test(&conf->pending_full_writes))
3438                         md_wakeup_thread(conf->mddev->thread);
3439 }
3440
3441 static void
3442 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3443                    struct stripe_head_state *s)
3444 {
3445         int abort = 0;
3446         int i;
3447
3448         BUG_ON(sh->batch_head);
3449         clear_bit(STRIPE_SYNCING, &sh->state);
3450         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3451                 wake_up(&conf->wait_for_overlap);
3452         s->syncing = 0;
3453         s->replacing = 0;
3454         /* There is nothing more to do for sync/check/repair.
3455          * Don't even need to abort as that is handled elsewhere
3456          * if needed, and not always wanted e.g. if there is a known
3457          * bad block here.
3458          * For recover/replace we need to record a bad block on all
3459          * non-sync devices, or abort the recovery
3460          */
3461         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3462                 /* During recovery devices cannot be removed, so
3463                  * locking and refcounting of rdevs is not needed
3464                  */
3465                 rcu_read_lock();
3466                 for (i = 0; i < conf->raid_disks; i++) {
3467                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3468                         if (rdev
3469                             && !test_bit(Faulty, &rdev->flags)
3470                             && !test_bit(In_sync, &rdev->flags)
3471                             && !rdev_set_badblocks(rdev, sh->sector,
3472                                                    STRIPE_SECTORS, 0))
3473                                 abort = 1;
3474                         rdev = rcu_dereference(conf->disks[i].replacement);
3475                         if (rdev
3476                             && !test_bit(Faulty, &rdev->flags)
3477                             && !test_bit(In_sync, &rdev->flags)
3478                             && !rdev_set_badblocks(rdev, sh->sector,
3479                                                    STRIPE_SECTORS, 0))
3480                                 abort = 1;
3481                 }
3482                 rcu_read_unlock();
3483                 if (abort)
3484                         conf->recovery_disabled =
3485                                 conf->mddev->recovery_disabled;
3486         }
3487         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3488 }
3489
3490 static int want_replace(struct stripe_head *sh, int disk_idx)
3491 {
3492         struct md_rdev *rdev;
3493         int rv = 0;
3494
3495         rcu_read_lock();
3496         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3497         if (rdev
3498             && !test_bit(Faulty, &rdev->flags)
3499             && !test_bit(In_sync, &rdev->flags)
3500             && (rdev->recovery_offset <= sh->sector
3501                 || rdev->mddev->recovery_cp <= sh->sector))
3502                 rv = 1;
3503         rcu_read_unlock();
3504         return rv;
3505 }
3506
3507 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3508                            int disk_idx, int disks)
3509 {
3510         struct r5dev *dev = &sh->dev[disk_idx];
3511         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3512                                   &sh->dev[s->failed_num[1]] };
3513         int i;
3514
3515
3516         if (test_bit(R5_LOCKED, &dev->flags) ||
3517             test_bit(R5_UPTODATE, &dev->flags))
3518                 /* No point reading this as we already have it or have
3519                  * decided to get it.
3520                  */
3521                 return 0;
3522
3523         if (dev->toread ||
3524             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3525                 /* We need this block to directly satisfy a request */
3526                 return 1;
3527
3528         if (s->syncing || s->expanding ||
3529             (s->replacing && want_replace(sh, disk_idx)))
3530                 /* When syncing, or expanding we read everything.
3531                  * When replacing, we need the replaced block.
3532                  */
3533                 return 1;
3534
3535         if ((s->failed >= 1 && fdev[0]->toread) ||
3536             (s->failed >= 2 && fdev[1]->toread))
3537                 /* If we want to read from a failed device, then
3538                  * we need to actually read every other device.
3539                  */
3540                 return 1;
3541
3542         /* Sometimes neither read-modify-write nor reconstruct-write
3543          * cycles can work.  In those cases we read every block we
3544          * can.  Then the parity-update is certain to have enough to
3545          * work with.
3546          * This can only be a problem when we need to write something,
3547          * and some device has failed.  If either of those tests
3548          * fail we need look no further.
3549          */
3550         if (!s->failed || !s->to_write)
3551                 return 0;
3552
3553         if (test_bit(R5_Insync, &dev->flags) &&
3554             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3555                 /* Pre-reads at not permitted until after short delay
3556                  * to gather multiple requests.  However if this
3557                  * device is no Insync, the block could only be computed
3558                  * and there is no need to delay that.
3559                  */
3560                 return 0;
3561
3562         for (i = 0; i < s->failed && i < 2; i++) {
3563                 if (fdev[i]->towrite &&
3564                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3565                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3566                         /* If we have a partial write to a failed
3567                          * device, then we will need to reconstruct
3568                          * the content of that device, so all other
3569                          * devices must be read.
3570                          */
3571                         return 1;
3572         }
3573
3574         /* If we are forced to do a reconstruct-write, either because
3575          * the current RAID6 implementation only supports that, or
3576          * because parity cannot be trusted and we are currently
3577          * recovering it, there is extra need to be careful.
3578          * If one of the devices that we would need to read, because
3579          * it is not being overwritten (and maybe not written at all)
3580          * is missing/faulty, then we need to read everything we can.
3581          */
3582         if (sh->raid_conf->level != 6 &&
3583             sh->sector < sh->raid_conf->mddev->recovery_cp)
3584                 /* reconstruct-write isn't being forced */
3585                 return 0;
3586         for (i = 0; i < s->failed && i < 2; i++) {
3587                 if (s->failed_num[i] != sh->pd_idx &&
3588                     s->failed_num[i] != sh->qd_idx &&
3589                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3590                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3591                         return 1;
3592         }
3593
3594         return 0;
3595 }
3596
3597 /* fetch_block - checks the given member device to see if its data needs
3598  * to be read or computed to satisfy a request.
3599  *
3600  * Returns 1 when no more member devices need to be checked, otherwise returns
3601  * 0 to tell the loop in handle_stripe_fill to continue
3602  */
3603 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3604                        int disk_idx, int disks)
3605 {
3606         struct r5dev *dev = &sh->dev[disk_idx];
3607
3608         /* is the data in this block needed, and can we get it? */
3609         if (need_this_block(sh, s, disk_idx, disks)) {
3610                 /* we would like to get this block, possibly by computing it,
3611                  * otherwise read it if the backing disk is insync
3612                  */
3613                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3614                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3615                 BUG_ON(sh->batch_head);
3616
3617                 /*
3618                  * In the raid6 case if the only non-uptodate disk is P
3619                  * then we already trusted P to compute the other failed
3620                  * drives. It is safe to compute rather than re-read P.
3621                  * In other cases we only compute blocks from failed
3622                  * devices, otherwise check/repair might fail to detect
3623                  * a real inconsistency.
3624                  */
3625
3626                 if ((s->uptodate == disks - 1) &&
3627                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3628                     (s->failed && (disk_idx == s->failed_num[0] ||
3629                                    disk_idx == s->failed_num[1])))) {
3630                         /* have disk failed, and we're requested to fetch it;
3631                          * do compute it
3632                          */
3633                         pr_debug("Computing stripe %llu block %d\n",
3634                                (unsigned long long)sh->sector, disk_idx);
3635                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3636                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3637                         set_bit(R5_Wantcompute, &dev->flags);
3638                         sh->ops.target = disk_idx;
3639                         sh->ops.target2 = -1; /* no 2nd target */
3640                         s->req_compute = 1;
3641                         /* Careful: from this point on 'uptodate' is in the eye
3642                          * of raid_run_ops which services 'compute' operations
3643                          * before writes. R5_Wantcompute flags a block that will
3644                          * be R5_UPTODATE by the time it is needed for a
3645                          * subsequent operation.
3646                          */
3647                         s->uptodate++;
3648                         return 1;
3649                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3650                         /* Computing 2-failure is *very* expensive; only
3651                          * do it if failed >= 2
3652                          */
3653                         int other;
3654                         for (other = disks; other--; ) {
3655                                 if (other == disk_idx)
3656                                         continue;
3657                                 if (!test_bit(R5_UPTODATE,
3658                                       &sh->dev[other].flags))
3659                                         break;
3660                         }
3661                         BUG_ON(other < 0);
3662                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3663                                (unsigned long long)sh->sector,
3664                                disk_idx, other);
3665                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3666                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3667                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3668                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3669                         sh->ops.target = disk_idx;
3670                         sh->ops.target2 = other;
3671                         s->uptodate += 2;
3672                         s->req_compute = 1;
3673                         return 1;
3674                 } else if (test_bit(R5_Insync, &dev->flags)) {
3675                         set_bit(R5_LOCKED, &dev->flags);
3676                         set_bit(R5_Wantread, &dev->flags);
3677                         s->locked++;
3678                         pr_debug("Reading block %d (sync=%d)\n",
3679                                 disk_idx, s->syncing);
3680                 }
3681         }
3682
3683         return 0;
3684 }
3685
3686 /**
3687  * handle_stripe_fill - read or compute data to satisfy pending requests.
3688  */
3689 static void handle_stripe_fill(struct stripe_head *sh,
3690                                struct stripe_head_state *s,
3691                                int disks)
3692 {
3693         int i;
3694
3695         /* look for blocks to read/compute, skip this if a compute
3696          * is already in flight, or if the stripe contents are in the
3697          * midst of changing due to a write
3698          */
3699         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3700             !sh->reconstruct_state) {
3701
3702                 /*
3703                  * For degraded stripe with data in journal, do not handle
3704                  * read requests yet, instead, flush the stripe to raid
3705                  * disks first, this avoids handling complex rmw of write
3706                  * back cache (prexor with orig_page, and then xor with
3707                  * page) in the read path
3708                  */
3709                 if (s->injournal && s->failed) {
3710                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3711                                 r5c_make_stripe_write_out(sh);
3712                         goto out;
3713                 }
3714
3715                 for (i = disks; i--; )
3716                         if (fetch_block(sh, s, i, disks))
3717                                 break;
3718         }
3719 out:
3720         set_bit(STRIPE_HANDLE, &sh->state);
3721 }
3722
3723 static void break_stripe_batch_list(struct stripe_head *head_sh,
3724                                     unsigned long handle_flags);
3725 /* handle_stripe_clean_event
3726  * any written block on an uptodate or failed drive can be returned.
3727  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3728  * never LOCKED, so we don't need to test 'failed' directly.
3729  */
3730 static void handle_stripe_clean_event(struct r5conf *conf,
3731         struct stripe_head *sh, int disks)
3732 {
3733         int i;
3734         struct r5dev *dev;
3735         int discard_pending = 0;
3736         struct stripe_head *head_sh = sh;
3737         bool do_endio = false;
3738
3739         for (i = disks; i--; )
3740                 if (sh->dev[i].written) {
3741                         dev = &sh->dev[i];
3742                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3743                             (test_bit(R5_UPTODATE, &dev->flags) ||
3744                              test_bit(R5_Discard, &dev->flags) ||
3745                              test_bit(R5_SkipCopy, &dev->flags))) {
3746                                 /* We can return any write requests */
3747                                 struct bio *wbi, *wbi2;
3748                                 pr_debug("Return write for disc %d\n", i);
3749                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3750                                         clear_bit(R5_UPTODATE, &dev->flags);
3751                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3752                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3753                                 }
3754                                 do_endio = true;
3755
3756 returnbi:
3757                                 dev->page = dev->orig_page;
3758                                 wbi = dev->written;
3759                                 dev->written = NULL;
3760                                 while (wbi && wbi->bi_iter.bi_sector <
3761                                         dev->sector + STRIPE_SECTORS) {
3762                                         wbi2 = r5_next_bio(wbi, dev->sector);
3763                                         md_write_end(conf->mddev);
3764                                         bio_endio(wbi);
3765                                         wbi = wbi2;
3766                                 }
3767                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3768                                                 STRIPE_SECTORS,
3769                                          !test_bit(STRIPE_DEGRADED, &sh->state),
3770                                                 0);
3771                                 if (head_sh->batch_head) {
3772                                         sh = list_first_entry(&sh->batch_list,
3773                                                               struct stripe_head,
3774                                                               batch_list);
3775                                         if (sh != head_sh) {
3776                                                 dev = &sh->dev[i];
3777                                                 goto returnbi;
3778                                         }
3779                                 }
3780                                 sh = head_sh;
3781                                 dev = &sh->dev[i];
3782                         } else if (test_bit(R5_Discard, &dev->flags))
3783                                 discard_pending = 1;
3784                 }
3785
3786         log_stripe_write_finished(sh);
3787
3788         if (!discard_pending &&
3789             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3790                 int hash;
3791                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3792                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3793                 if (sh->qd_idx >= 0) {
3794                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3795                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3796                 }
3797                 /* now that discard is done we can proceed with any sync */
3798                 clear_bit(STRIPE_DISCARD, &sh->state);
3799                 /*
3800                  * SCSI discard will change some bio fields and the stripe has
3801                  * no updated data, so remove it from hash list and the stripe
3802                  * will be reinitialized
3803                  */
3804 unhash:
3805                 hash = sh->hash_lock_index;
3806                 spin_lock_irq(conf->hash_locks + hash);
3807                 remove_hash(sh);
3808                 spin_unlock_irq(conf->hash_locks + hash);
3809                 if (head_sh->batch_head) {
3810                         sh = list_first_entry(&sh->batch_list,
3811                                               struct stripe_head, batch_list);
3812                         if (sh != head_sh)
3813                                         goto unhash;
3814                 }
3815                 sh = head_sh;
3816
3817                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3818                         set_bit(STRIPE_HANDLE, &sh->state);
3819
3820         }
3821
3822         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3823                 if (atomic_dec_and_test(&conf->pending_full_writes))
3824                         md_wakeup_thread(conf->mddev->thread);
3825
3826         if (head_sh->batch_head && do_endio)
3827                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3828 }
3829
3830 /*
3831  * For RMW in write back cache, we need extra page in prexor to store the
3832  * old data. This page is stored in dev->orig_page.
3833  *
3834  * This function checks whether we have data for prexor. The exact logic
3835  * is:
3836  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3837  */
3838 static inline bool uptodate_for_rmw(struct r5dev *dev)
3839 {
3840         return (test_bit(R5_UPTODATE, &dev->flags)) &&
3841                 (!test_bit(R5_InJournal, &dev->flags) ||
3842                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3843 }
3844
3845 static int handle_stripe_dirtying(struct r5conf *conf,
3846                                   struct stripe_head *sh,
3847                                   struct stripe_head_state *s,
3848                                   int disks)
3849 {
3850         int rmw = 0, rcw = 0, i;
3851         sector_t recovery_cp = conf->mddev->recovery_cp;
3852
3853         /* Check whether resync is now happening or should start.
3854          * If yes, then the array is dirty (after unclean shutdown or
3855          * initial creation), so parity in some stripes might be inconsistent.
3856          * In this case, we need to always do reconstruct-write, to ensure
3857          * that in case of drive failure or read-error correction, we
3858          * generate correct data from the parity.
3859          */
3860         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3861             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3862              s->failed == 0)) {
3863                 /* Calculate the real rcw later - for now make it
3864                  * look like rcw is cheaper
3865                  */
3866                 rcw = 1; rmw = 2;
3867                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3868                          conf->rmw_level, (unsigned long long)recovery_cp,
3869                          (unsigned long long)sh->sector);
3870         } else for (i = disks; i--; ) {
3871                 /* would I have to read this buffer for read_modify_write */
3872                 struct r5dev *dev = &sh->dev[i];
3873                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3874                      i == sh->pd_idx || i == sh->qd_idx ||
3875                      test_bit(R5_InJournal, &dev->flags)) &&
3876                     !test_bit(R5_LOCKED, &dev->flags) &&
3877                     !(uptodate_for_rmw(dev) ||
3878                       test_bit(R5_Wantcompute, &dev->flags))) {
3879                         if (test_bit(R5_Insync, &dev->flags))
3880                                 rmw++;
3881                         else
3882                                 rmw += 2*disks;  /* cannot read it */
3883                 }
3884                 /* Would I have to read this buffer for reconstruct_write */
3885                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3886                     i != sh->pd_idx && i != sh->qd_idx &&
3887                     !test_bit(R5_LOCKED, &dev->flags) &&
3888                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3889                       test_bit(R5_Wantcompute, &dev->flags))) {
3890                         if (test_bit(R5_Insync, &dev->flags))
3891                                 rcw++;
3892                         else
3893                                 rcw += 2*disks;
3894                 }
3895         }
3896
3897         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3898                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
3899         set_bit(STRIPE_HANDLE, &sh->state);
3900         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3901                 /* prefer read-modify-write, but need to get some data */
3902                 if (conf->mddev->queue)
3903                         blk_add_trace_msg(conf->mddev->queue,
3904                                           "raid5 rmw %llu %d",
3905                                           (unsigned long long)sh->sector, rmw);
3906                 for (i = disks; i--; ) {
3907                         struct r5dev *dev = &sh->dev[i];
3908                         if (test_bit(R5_InJournal, &dev->flags) &&
3909                             dev->page == dev->orig_page &&
3910                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3911                                 /* alloc page for prexor */
3912                                 struct page *p = alloc_page(GFP_NOIO);
3913
3914                                 if (p) {
3915                                         dev->orig_page = p;
3916                                         continue;
3917                                 }
3918
3919                                 /*
3920                                  * alloc_page() failed, try use
3921                                  * disk_info->extra_page
3922                                  */
3923                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3924                                                       &conf->cache_state)) {
3925                                         r5c_use_extra_page(sh);
3926                                         break;
3927                                 }
3928
3929                                 /* extra_page in use, add to delayed_list */
3930                                 set_bit(STRIPE_DELAYED, &sh->state);
3931                                 s->waiting_extra_page = 1;
3932                                 return -EAGAIN;
3933                         }
3934                 }
3935
3936                 for (i = disks; i--; ) {
3937                         struct r5dev *dev = &sh->dev[i];
3938                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3939                              i == sh->pd_idx || i == sh->qd_idx ||
3940                              test_bit(R5_InJournal, &dev->flags)) &&
3941                             !test_bit(R5_LOCKED, &dev->flags) &&
3942                             !(uptodate_for_rmw(dev) ||
3943                               test_bit(R5_Wantcompute, &dev->flags)) &&
3944                             test_bit(R5_Insync, &dev->flags)) {
3945                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3946                                              &sh->state)) {
3947                                         pr_debug("Read_old block %d for r-m-w\n",
3948                                                  i);
3949                                         set_bit(R5_LOCKED, &dev->flags);
3950                                         set_bit(R5_Wantread, &dev->flags);
3951                                         s->locked++;
3952                                 } else {
3953                                         set_bit(STRIPE_DELAYED, &sh->state);
3954                                         set_bit(STRIPE_HANDLE, &sh->state);
3955                                 }
3956                         }
3957                 }
3958         }
3959         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3960                 /* want reconstruct write, but need to get some data */
3961                 int qread =0;
3962                 rcw = 0;
3963                 for (i = disks; i--; ) {
3964                         struct r5dev *dev = &sh->dev[i];
3965                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3966                             i != sh->pd_idx && i != sh->qd_idx &&
3967                             !test_bit(R5_LOCKED, &dev->flags) &&
3968                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3969                               test_bit(R5_Wantcompute, &dev->flags))) {
3970                                 rcw++;
3971                                 if (test_bit(R5_Insync, &dev->flags) &&
3972                                     test_bit(STRIPE_PREREAD_ACTIVE,
3973                                              &sh->state)) {
3974                                         pr_debug("Read_old block "
3975                                                 "%d for Reconstruct\n", i);
3976                                         set_bit(R5_LOCKED, &dev->flags);
3977                                         set_bit(R5_Wantread, &dev->flags);
3978                                         s->locked++;
3979                                         qread++;
3980                                 } else {
3981                                         set_bit(STRIPE_DELAYED, &sh->state);
3982                                         set_bit(STRIPE_HANDLE, &sh->state);
3983                                 }
3984                         }
3985                 }
3986                 if (rcw && conf->mddev->queue)
3987                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3988                                           (unsigned long long)sh->sector,
3989                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3990         }
3991
3992         if (rcw > disks && rmw > disks &&
3993             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3994                 set_bit(STRIPE_DELAYED, &sh->state);
3995
3996         /* now if nothing is locked, and if we have enough data,
3997          * we can start a write request
3998          */
3999         /* since handle_stripe can be called at any time we need to handle the
4000          * case where a compute block operation has been submitted and then a
4001          * subsequent call wants to start a write request.  raid_run_ops only
4002          * handles the case where compute block and reconstruct are requested
4003          * simultaneously.  If this is not the case then new writes need to be
4004          * held off until the compute completes.
4005          */
4006         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4007             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4008              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4009                 schedule_reconstruction(sh, s, rcw == 0, 0);
4010         return 0;
4011 }
4012
4013 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4014                                 struct stripe_head_state *s, int disks)
4015 {
4016         struct r5dev *dev = NULL;
4017
4018         BUG_ON(sh->batch_head);
4019         set_bit(STRIPE_HANDLE, &sh->state);
4020
4021         switch (sh->check_state) {
4022         case check_state_idle:
4023                 /* start a new check operation if there are no failures */
4024                 if (s->failed == 0) {
4025                         BUG_ON(s->uptodate != disks);
4026                         sh->check_state = check_state_run;
4027                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4028                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4029                         s->uptodate--;
4030                         break;
4031                 }
4032                 dev = &sh->dev[s->failed_num[0]];
4033                 /* fall through */
4034         case check_state_compute_result:
4035                 sh->check_state = check_state_idle;
4036                 if (!dev)
4037                         dev = &sh->dev[sh->pd_idx];
4038
4039                 /* check that a write has not made the stripe insync */
4040                 if (test_bit(STRIPE_INSYNC, &sh->state))
4041                         break;
4042
4043                 /* either failed parity check, or recovery is happening */
4044                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4045                 BUG_ON(s->uptodate != disks);
4046
4047                 set_bit(R5_LOCKED, &dev->flags);
4048                 s->locked++;
4049                 set_bit(R5_Wantwrite, &dev->flags);
4050
4051                 clear_bit(STRIPE_DEGRADED, &sh->state);
4052                 set_bit(STRIPE_INSYNC, &sh->state);
4053                 break;
4054         case check_state_run:
4055                 break; /* we will be called again upon completion */
4056         case check_state_check_result:
4057                 sh->check_state = check_state_idle;
4058
4059                 /* if a failure occurred during the check operation, leave
4060                  * STRIPE_INSYNC not set and let the stripe be handled again
4061                  */
4062                 if (s->failed)
4063                         break;
4064
4065                 /* handle a successful check operation, if parity is correct
4066                  * we are done.  Otherwise update the mismatch count and repair
4067                  * parity if !MD_RECOVERY_CHECK
4068                  */
4069                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4070                         /* parity is correct (on disc,
4071                          * not in buffer any more)
4072                          */
4073                         set_bit(STRIPE_INSYNC, &sh->state);
4074                 else {
4075                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4076                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4077                                 /* don't try to repair!! */
4078                                 set_bit(STRIPE_INSYNC, &sh->state);
4079                                 pr_warn_ratelimited("%s: mismatch sector in range "
4080                                                     "%llu-%llu\n", mdname(conf->mddev),
4081                                                     (unsigned long long) sh->sector,
4082                                                     (unsigned long long) sh->sector +
4083                                                     STRIPE_SECTORS);
4084                         } else {
4085                                 sh->check_state = check_state_compute_run;
4086                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4087                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4088                                 set_bit(R5_Wantcompute,
4089                                         &sh->dev[sh->pd_idx].flags);
4090                                 sh->ops.target = sh->pd_idx;
4091                                 sh->ops.target2 = -1;
4092                                 s->uptodate++;
4093                         }
4094                 }
4095                 break;
4096         case check_state_compute_run:
4097                 break;
4098         default:
4099                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4100                        __func__, sh->check_state,
4101                        (unsigned long long) sh->sector);
4102                 BUG();
4103         }
4104 }
4105
4106 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4107                                   struct stripe_head_state *s,
4108                                   int disks)
4109 {
4110         int pd_idx = sh->pd_idx;
4111         int qd_idx = sh->qd_idx;
4112         struct r5dev *dev;
4113
4114         BUG_ON(sh->batch_head);
4115         set_bit(STRIPE_HANDLE, &sh->state);
4116
4117         BUG_ON(s->failed > 2);
4118
4119         /* Want to check and possibly repair P and Q.
4120          * However there could be one 'failed' device, in which
4121          * case we can only check one of them, possibly using the
4122          * other to generate missing data
4123          */
4124
4125         switch (sh->check_state) {
4126         case check_state_idle:
4127                 /* start a new check operation if there are < 2 failures */
4128                 if (s->failed == s->q_failed) {
4129                         /* The only possible failed device holds Q, so it
4130                          * makes sense to check P (If anything else were failed,
4131                          * we would have used P to recreate it).
4132                          */
4133                         sh->check_state = check_state_run;
4134                 }
4135                 if (!s->q_failed && s->failed < 2) {
4136                         /* Q is not failed, and we didn't use it to generate
4137                          * anything, so it makes sense to check it
4138                          */
4139                         if (sh->check_state == check_state_run)
4140                                 sh->check_state = check_state_run_pq;
4141                         else
4142                                 sh->check_state = check_state_run_q;
4143                 }
4144
4145                 /* discard potentially stale zero_sum_result */
4146                 sh->ops.zero_sum_result = 0;
4147
4148                 if (sh->check_state == check_state_run) {
4149                         /* async_xor_zero_sum destroys the contents of P */
4150                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4151                         s->uptodate--;
4152                 }
4153                 if (sh->check_state >= check_state_run &&
4154                     sh->check_state <= check_state_run_pq) {
4155                         /* async_syndrome_zero_sum preserves P and Q, so
4156                          * no need to mark them !uptodate here
4157                          */
4158                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4159                         break;
4160                 }
4161
4162                 /* we have 2-disk failure */
4163                 BUG_ON(s->failed != 2);
4164                 /* fall through */
4165         case check_state_compute_result:
4166                 sh->check_state = check_state_idle;
4167
4168                 /* check that a write has not made the stripe insync */
4169                 if (test_bit(STRIPE_INSYNC, &sh->state))
4170                         break;
4171
4172                 /* now write out any block on a failed drive,
4173                  * or P or Q if they were recomputed
4174                  */
4175                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
4176                 if (s->failed == 2) {
4177                         dev = &sh->dev[s->failed_num[1]];
4178                         s->locked++;
4179                         set_bit(R5_LOCKED, &dev->flags);
4180                         set_bit(R5_Wantwrite, &dev->flags);
4181                 }
4182                 if (s->failed >= 1) {
4183                         dev = &sh->dev[s->failed_num[0]];
4184                         s->locked++;
4185                         set_bit(R5_LOCKED, &dev->flags);
4186                         set_bit(R5_Wantwrite, &dev->flags);
4187                 }
4188                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4189                         dev = &sh->dev[pd_idx];
4190                         s->locked++;
4191                         set_bit(R5_LOCKED, &dev->flags);
4192                         set_bit(R5_Wantwrite, &dev->flags);
4193                 }
4194                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4195                         dev = &sh->dev[qd_idx];
4196                         s->locked++;
4197                         set_bit(R5_LOCKED, &dev->flags);
4198                         set_bit(R5_Wantwrite, &dev->flags);
4199                 }
4200                 clear_bit(STRIPE_DEGRADED, &sh->state);
4201
4202                 set_bit(STRIPE_INSYNC, &sh->state);
4203                 break;
4204         case check_state_run:
4205         case check_state_run_q:
4206         case check_state_run_pq:
4207                 break; /* we will be called again upon completion */
4208         case check_state_check_result:
4209                 sh->check_state = check_state_idle;
4210
4211                 /* handle a successful check operation, if parity is correct
4212                  * we are done.  Otherwise update the mismatch count and repair
4213                  * parity if !MD_RECOVERY_CHECK
4214                  */
4215                 if (sh->ops.zero_sum_result == 0) {
4216                         /* both parities are correct */
4217                         if (!s->failed)
4218                                 set_bit(STRIPE_INSYNC, &sh->state);
4219                         else {
4220                                 /* in contrast to the raid5 case we can validate
4221                                  * parity, but still have a failure to write
4222                                  * back
4223                                  */
4224                                 sh->check_state = check_state_compute_result;
4225                                 /* Returning at this point means that we may go
4226                                  * off and bring p and/or q uptodate again so
4227                                  * we make sure to check zero_sum_result again
4228                                  * to verify if p or q need writeback
4229                                  */
4230                         }
4231                 } else {
4232                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4233                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4234                                 /* don't try to repair!! */
4235                                 set_bit(STRIPE_INSYNC, &sh->state);
4236                                 pr_warn_ratelimited("%s: mismatch sector in range "
4237                                                     "%llu-%llu\n", mdname(conf->mddev),
4238                                                     (unsigned long long) sh->sector,
4239                                                     (unsigned long long) sh->sector +
4240                                                     STRIPE_SECTORS);
4241                         } else {
4242                                 int *target = &sh->ops.target;
4243
4244                                 sh->ops.target = -1;
4245                                 sh->ops.target2 = -1;
4246                                 sh->check_state = check_state_compute_run;
4247                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4248                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4249                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4250                                         set_bit(R5_Wantcompute,
4251                                                 &sh->dev[pd_idx].flags);
4252                                         *target = pd_idx;
4253                                         target = &sh->ops.target2;
4254                                         s->uptodate++;
4255                                 }
4256                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4257                                         set_bit(R5_Wantcompute,
4258                                                 &sh->dev[qd_idx].flags);
4259                                         *target = qd_idx;
4260                                         s->uptodate++;
4261                                 }
4262                         }
4263                 }
4264                 break;
4265         case check_state_compute_run:
4266                 break;
4267         default:
4268                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4269                         __func__, sh->check_state,
4270                         (unsigned long long) sh->sector);
4271                 BUG();
4272         }
4273 }
4274
4275 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4276 {
4277         int i;
4278
4279         /* We have read all the blocks in this stripe and now we need to
4280          * copy some of them into a target stripe for expand.
4281          */
4282         struct dma_async_tx_descriptor *tx = NULL;
4283         BUG_ON(sh->batch_head);
4284         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4285         for (i = 0; i < sh->disks; i++)
4286                 if (i != sh->pd_idx && i != sh->qd_idx) {
4287                         int dd_idx, j;
4288                         struct stripe_head *sh2;
4289                         struct async_submit_ctl submit;
4290
4291                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4292                         sector_t s = raid5_compute_sector(conf, bn, 0,
4293                                                           &dd_idx, NULL);
4294                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4295                         if (sh2 == NULL)
4296                                 /* so far only the early blocks of this stripe
4297                                  * have been requested.  When later blocks
4298                                  * get requested, we will try again
4299                                  */
4300                                 continue;
4301                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4302                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4303                                 /* must have already done this block */
4304                                 raid5_release_stripe(sh2);
4305                                 continue;
4306                         }
4307
4308                         /* place all the copies on one channel */
4309                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4310                         tx = async_memcpy(sh2->dev[dd_idx].page,
4311                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
4312                                           &submit);
4313
4314                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4315                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4316                         for (j = 0; j < conf->raid_disks; j++)
4317                                 if (j != sh2->pd_idx &&
4318                                     j != sh2->qd_idx &&
4319                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4320                                         break;
4321                         if (j == conf->raid_disks) {
4322                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4323                                 set_bit(STRIPE_HANDLE, &sh2->state);
4324                         }
4325                         raid5_release_stripe(sh2);
4326
4327                 }
4328         /* done submitting copies, wait for them to complete */
4329         async_tx_quiesce(&tx);
4330 }
4331
4332 /*
4333  * handle_stripe - do things to a stripe.
4334  *
4335  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4336  * state of various bits to see what needs to be done.
4337  * Possible results:
4338  *    return some read requests which now have data
4339  *    return some write requests which are safely on storage
4340  *    schedule a read on some buffers
4341  *    schedule a write of some buffers
4342  *    return confirmation of parity correctness
4343  *
4344  */
4345
4346 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4347 {
4348         struct r5conf *conf = sh->raid_conf;
4349         int disks = sh->disks;
4350         struct r5dev *dev;
4351         int i;
4352         int do_recovery = 0;
4353
4354         memset(s, 0, sizeof(*s));
4355
4356         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4357         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4358         s->failed_num[0] = -1;
4359         s->failed_num[1] = -1;
4360         s->log_failed = r5l_log_disk_error(conf);
4361
4362         /* Now to look around and see what can be done */
4363         rcu_read_lock();
4364         for (i=disks; i--; ) {
4365                 struct md_rdev *rdev;
4366                 sector_t first_bad;
4367                 int bad_sectors;
4368                 int is_bad = 0;
4369
4370                 dev = &sh->dev[i];
4371
4372                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4373                          i, dev->flags,
4374                          dev->toread, dev->towrite, dev->written);
4375                 /* maybe we can reply to a read
4376                  *
4377                  * new wantfill requests are only permitted while
4378                  * ops_complete_biofill is guaranteed to be inactive
4379                  */
4380                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4381                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4382                         set_bit(R5_Wantfill, &dev->flags);
4383
4384                 /* now count some things */
4385                 if (test_bit(R5_LOCKED, &dev->flags))
4386                         s->locked++;
4387                 if (test_bit(R5_UPTODATE, &dev->flags))
4388                         s->uptodate++;
4389                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4390                         s->compute++;
4391                         BUG_ON(s->compute > 2);
4392                 }
4393
4394                 if (test_bit(R5_Wantfill, &dev->flags))
4395                         s->to_fill++;
4396                 else if (dev->toread)
4397                         s->to_read++;
4398                 if (dev->towrite) {
4399                         s->to_write++;
4400                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4401                                 s->non_overwrite++;
4402                 }
4403                 if (dev->written)
4404                         s->written++;
4405                 /* Prefer to use the replacement for reads, but only
4406                  * if it is recovered enough and has no bad blocks.
4407                  */
4408                 rdev = rcu_dereference(conf->disks[i].replacement);
4409                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4410                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4411                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4412                                  &first_bad, &bad_sectors))
4413                         set_bit(R5_ReadRepl, &dev->flags);
4414                 else {
4415                         if (rdev && !test_bit(Faulty, &rdev->flags))
4416                                 set_bit(R5_NeedReplace, &dev->flags);
4417                         else
4418                                 clear_bit(R5_NeedReplace, &dev->flags);
4419                         rdev = rcu_dereference(conf->disks[i].rdev);
4420                         clear_bit(R5_ReadRepl, &dev->flags);
4421                 }
4422                 if (rdev && test_bit(Faulty, &rdev->flags))
4423                         rdev = NULL;
4424                 if (rdev) {
4425                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4426                                              &first_bad, &bad_sectors);
4427                         if (s->blocked_rdev == NULL
4428                             && (test_bit(Blocked, &rdev->flags)
4429                                 || is_bad < 0)) {
4430                                 if (is_bad < 0)
4431                                         set_bit(BlockedBadBlocks,
4432                                                 &rdev->flags);
4433                                 s->blocked_rdev = rdev;
4434                                 atomic_inc(&rdev->nr_pending);
4435                         }
4436                 }
4437                 clear_bit(R5_Insync, &dev->flags);
4438                 if (!rdev)
4439                         /* Not in-sync */;
4440                 else if (is_bad) {
4441                         /* also not in-sync */
4442                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4443                             test_bit(R5_UPTODATE, &dev->flags)) {
4444                                 /* treat as in-sync, but with a read error
4445                                  * which we can now try to correct
4446                                  */
4447                                 set_bit(R5_Insync, &dev->flags);
4448                                 set_bit(R5_ReadError, &dev->flags);
4449                         }
4450                 } else if (test_bit(In_sync, &rdev->flags))
4451                         set_bit(R5_Insync, &dev->flags);
4452                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4453                         /* in sync if before recovery_offset */
4454                         set_bit(R5_Insync, &dev->flags);
4455                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4456                          test_bit(R5_Expanded, &dev->flags))
4457                         /* If we've reshaped into here, we assume it is Insync.
4458                          * We will shortly update recovery_offset to make
4459                          * it official.
4460                          */
4461                         set_bit(R5_Insync, &dev->flags);
4462
4463                 if (test_bit(R5_WriteError, &dev->flags)) {
4464                         /* This flag does not apply to '.replacement'
4465                          * only to .rdev, so make sure to check that*/
4466                         struct md_rdev *rdev2 = rcu_dereference(
4467                                 conf->disks[i].rdev);
4468                         if (rdev2 == rdev)
4469                                 clear_bit(R5_Insync, &dev->flags);
4470                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4471                                 s->handle_bad_blocks = 1;
4472                                 atomic_inc(&rdev2->nr_pending);
4473                         } else
4474                                 clear_bit(R5_WriteError, &dev->flags);
4475                 }
4476                 if (test_bit(R5_MadeGood, &dev->flags)) {
4477                         /* This flag does not apply to '.replacement'
4478                          * only to .rdev, so make sure to check that*/
4479                         struct md_rdev *rdev2 = rcu_dereference(
4480                                 conf->disks[i].rdev);
4481                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4482                                 s->handle_bad_blocks = 1;
4483                                 atomic_inc(&rdev2->nr_pending);
4484                         } else
4485                                 clear_bit(R5_MadeGood, &dev->flags);
4486                 }
4487                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4488                         struct md_rdev *rdev2 = rcu_dereference(
4489                                 conf->disks[i].replacement);
4490                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4491                                 s->handle_bad_blocks = 1;
4492                                 atomic_inc(&rdev2->nr_pending);
4493                         } else
4494                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4495                 }
4496                 if (!test_bit(R5_Insync, &dev->flags)) {
4497                         /* The ReadError flag will just be confusing now */
4498                         clear_bit(R5_ReadError, &dev->flags);
4499                         clear_bit(R5_ReWrite, &dev->flags);
4500                 }
4501                 if (test_bit(R5_ReadError, &dev->flags))
4502                         clear_bit(R5_Insync, &dev->flags);
4503                 if (!test_bit(R5_Insync, &dev->flags)) {
4504                         if (s->failed < 2)
4505                                 s->failed_num[s->failed] = i;
4506                         s->failed++;
4507                         if (rdev && !test_bit(Faulty, &rdev->flags))
4508                                 do_recovery = 1;
4509                 }
4510
4511                 if (test_bit(R5_InJournal, &dev->flags))
4512                         s->injournal++;
4513                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4514                         s->just_cached++;
4515         }
4516         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4517                 /* If there is a failed device being replaced,
4518                  *     we must be recovering.
4519                  * else if we are after recovery_cp, we must be syncing
4520                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4521                  * else we can only be replacing
4522                  * sync and recovery both need to read all devices, and so
4523                  * use the same flag.
4524                  */
4525                 if (do_recovery ||
4526                     sh->sector >= conf->mddev->recovery_cp ||
4527                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4528                         s->syncing = 1;
4529                 else
4530                         s->replacing = 1;
4531         }
4532         rcu_read_unlock();
4533 }
4534
4535 static int clear_batch_ready(struct stripe_head *sh)
4536 {
4537         /* Return '1' if this is a member of batch, or
4538          * '0' if it is a lone stripe or a head which can now be
4539          * handled.
4540          */
4541         struct stripe_head *tmp;
4542         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4543                 return (sh->batch_head && sh->batch_head != sh);
4544         spin_lock(&sh->stripe_lock);
4545         if (!sh->batch_head) {
4546                 spin_unlock(&sh->stripe_lock);
4547                 return 0;
4548         }
4549
4550         /*
4551          * this stripe could be added to a batch list before we check
4552          * BATCH_READY, skips it
4553          */
4554         if (sh->batch_head != sh) {
4555                 spin_unlock(&sh->stripe_lock);
4556                 return 1;
4557         }
4558         spin_lock(&sh->batch_lock);
4559         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4560                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4561         spin_unlock(&sh->batch_lock);
4562         spin_unlock(&sh->stripe_lock);
4563
4564         /*
4565          * BATCH_READY is cleared, no new stripes can be added.
4566          * batch_list can be accessed without lock
4567          */
4568         return 0;
4569 }
4570
4571 static void break_stripe_batch_list(struct stripe_head *head_sh,
4572                                     unsigned long handle_flags)
4573 {
4574         struct stripe_head *sh, *next;
4575         int i;
4576         int do_wakeup = 0;
4577
4578         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4579
4580                 list_del_init(&sh->batch_list);
4581
4582                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4583                                           (1 << STRIPE_SYNCING) |
4584                                           (1 << STRIPE_REPLACED) |
4585                                           (1 << STRIPE_DELAYED) |
4586                                           (1 << STRIPE_BIT_DELAY) |
4587                                           (1 << STRIPE_FULL_WRITE) |
4588                                           (1 << STRIPE_BIOFILL_RUN) |
4589                                           (1 << STRIPE_COMPUTE_RUN)  |
4590                                           (1 << STRIPE_OPS_REQ_PENDING) |
4591                                           (1 << STRIPE_DISCARD) |
4592                                           (1 << STRIPE_BATCH_READY) |
4593                                           (1 << STRIPE_BATCH_ERR) |
4594                                           (1 << STRIPE_BITMAP_PENDING)),
4595                         "stripe state: %lx\n", sh->state);
4596                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4597                                               (1 << STRIPE_REPLACED)),
4598                         "head stripe state: %lx\n", head_sh->state);
4599
4600                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4601                                             (1 << STRIPE_PREREAD_ACTIVE) |
4602                                             (1 << STRIPE_DEGRADED)),
4603                               head_sh->state & (1 << STRIPE_INSYNC));
4604
4605                 sh->check_state = head_sh->check_state;
4606                 sh->reconstruct_state = head_sh->reconstruct_state;
4607                 for (i = 0; i < sh->disks; i++) {
4608                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4609                                 do_wakeup = 1;
4610                         sh->dev[i].flags = head_sh->dev[i].flags &
4611                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4612                 }
4613                 spin_lock_irq(&sh->stripe_lock);
4614                 sh->batch_head = NULL;
4615                 spin_unlock_irq(&sh->stripe_lock);
4616                 if (handle_flags == 0 ||
4617                     sh->state & handle_flags)
4618                         set_bit(STRIPE_HANDLE, &sh->state);
4619                 raid5_release_stripe(sh);
4620         }
4621         spin_lock_irq(&head_sh->stripe_lock);
4622         head_sh->batch_head = NULL;
4623         spin_unlock_irq(&head_sh->stripe_lock);
4624         for (i = 0; i < head_sh->disks; i++)
4625                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4626                         do_wakeup = 1;
4627         if (head_sh->state & handle_flags)
4628                 set_bit(STRIPE_HANDLE, &head_sh->state);
4629
4630         if (do_wakeup)
4631                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4632 }
4633
4634 static void handle_stripe(struct stripe_head *sh)
4635 {
4636         struct stripe_head_state s;
4637         struct r5conf *conf = sh->raid_conf;
4638         int i;
4639         int prexor;
4640         int disks = sh->disks;
4641         struct r5dev *pdev, *qdev;
4642
4643         clear_bit(STRIPE_HANDLE, &sh->state);
4644         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4645                 /* already being handled, ensure it gets handled
4646                  * again when current action finishes */
4647                 set_bit(STRIPE_HANDLE, &sh->state);
4648                 return;
4649         }
4650
4651         if (clear_batch_ready(sh) ) {
4652                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4653                 return;
4654         }
4655
4656         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4657                 break_stripe_batch_list(sh, 0);
4658
4659         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4660                 spin_lock(&sh->stripe_lock);
4661                 /*
4662                  * Cannot process 'sync' concurrently with 'discard'.
4663                  * Flush data in r5cache before 'sync'.
4664                  */
4665                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4666                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4667                     !test_bit(STRIPE_DISCARD, &sh->state) &&
4668                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4669                         set_bit(STRIPE_SYNCING, &sh->state);
4670                         clear_bit(STRIPE_INSYNC, &sh->state);
4671                         clear_bit(STRIPE_REPLACED, &sh->state);
4672                 }
4673                 spin_unlock(&sh->stripe_lock);
4674         }
4675         clear_bit(STRIPE_DELAYED, &sh->state);
4676
4677         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4678                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4679                (unsigned long long)sh->sector, sh->state,
4680                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4681                sh->check_state, sh->reconstruct_state);
4682
4683         analyse_stripe(sh, &s);
4684
4685         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4686                 goto finish;
4687
4688         if (s.handle_bad_blocks ||
4689             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4690                 set_bit(STRIPE_HANDLE, &sh->state);
4691                 goto finish;
4692         }
4693
4694         if (unlikely(s.blocked_rdev)) {
4695                 if (s.syncing || s.expanding || s.expanded ||
4696                     s.replacing || s.to_write || s.written) {
4697                         set_bit(STRIPE_HANDLE, &sh->state);
4698                         goto finish;
4699                 }
4700                 /* There is nothing for the blocked_rdev to block */
4701                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4702                 s.blocked_rdev = NULL;
4703         }
4704
4705         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4706                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4707                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4708         }
4709
4710         pr_debug("locked=%d uptodate=%d to_read=%d"
4711                " to_write=%d failed=%d failed_num=%d,%d\n",
4712                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4713                s.failed_num[0], s.failed_num[1]);
4714         /*
4715          * check if the array has lost more than max_degraded devices and,
4716          * if so, some requests might need to be failed.
4717          *
4718          * When journal device failed (log_failed), we will only process
4719          * the stripe if there is data need write to raid disks
4720          */
4721         if (s.failed > conf->max_degraded ||
4722             (s.log_failed && s.injournal == 0)) {
4723                 sh->check_state = 0;
4724                 sh->reconstruct_state = 0;
4725                 break_stripe_batch_list(sh, 0);
4726                 if (s.to_read+s.to_write+s.written)
4727                         handle_failed_stripe(conf, sh, &s, disks);
4728                 if (s.syncing + s.replacing)
4729                         handle_failed_sync(conf, sh, &s);
4730         }
4731
4732         /* Now we check to see if any write operations have recently
4733          * completed
4734          */
4735         prexor = 0;
4736         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4737                 prexor = 1;
4738         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4739             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4740                 sh->reconstruct_state = reconstruct_state_idle;
4741
4742                 /* All the 'written' buffers and the parity block are ready to
4743                  * be written back to disk
4744                  */
4745                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4746                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4747                 BUG_ON(sh->qd_idx >= 0 &&
4748                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4749                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4750                 for (i = disks; i--; ) {
4751                         struct r5dev *dev = &sh->dev[i];
4752                         if (test_bit(R5_LOCKED, &dev->flags) &&
4753                                 (i == sh->pd_idx || i == sh->qd_idx ||
4754                                  dev->written || test_bit(R5_InJournal,
4755                                                           &dev->flags))) {
4756                                 pr_debug("Writing block %d\n", i);
4757                                 set_bit(R5_Wantwrite, &dev->flags);
4758                                 if (prexor)
4759                                         continue;
4760                                 if (s.failed > 1)
4761                                         continue;
4762                                 if (!test_bit(R5_Insync, &dev->flags) ||
4763                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4764                                      s.failed == 0))
4765                                         set_bit(STRIPE_INSYNC, &sh->state);
4766                         }
4767                 }
4768                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4769                         s.dec_preread_active = 1;
4770         }
4771
4772         /*
4773          * might be able to return some write requests if the parity blocks
4774          * are safe, or on a failed drive
4775          */
4776         pdev = &sh->dev[sh->pd_idx];
4777         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4778                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4779         qdev = &sh->dev[sh->qd_idx];
4780         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4781                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4782                 || conf->level < 6;
4783
4784         if (s.written &&
4785             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4786                              && !test_bit(R5_LOCKED, &pdev->flags)
4787                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4788                                  test_bit(R5_Discard, &pdev->flags))))) &&
4789             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4790                              && !test_bit(R5_LOCKED, &qdev->flags)
4791                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4792                                  test_bit(R5_Discard, &qdev->flags))))))
4793                 handle_stripe_clean_event(conf, sh, disks);
4794
4795         if (s.just_cached)
4796                 r5c_handle_cached_data_endio(conf, sh, disks);
4797         log_stripe_write_finished(sh);
4798
4799         /* Now we might consider reading some blocks, either to check/generate
4800          * parity, or to satisfy requests
4801          * or to load a block that is being partially written.
4802          */
4803         if (s.to_read || s.non_overwrite
4804             || (conf->level == 6 && s.to_write && s.failed)
4805             || (s.syncing && (s.uptodate + s.compute < disks))
4806             || s.replacing
4807             || s.expanding)
4808                 handle_stripe_fill(sh, &s, disks);
4809
4810         /*
4811          * When the stripe finishes full journal write cycle (write to journal
4812          * and raid disk), this is the clean up procedure so it is ready for
4813          * next operation.
4814          */
4815         r5c_finish_stripe_write_out(conf, sh, &s);
4816
4817         /*
4818          * Now to consider new write requests, cache write back and what else,
4819          * if anything should be read.  We do not handle new writes when:
4820          * 1/ A 'write' operation (copy+xor) is already in flight.
4821          * 2/ A 'check' operation is in flight, as it may clobber the parity
4822          *    block.
4823          * 3/ A r5c cache log write is in flight.
4824          */
4825
4826         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4827                 if (!r5c_is_writeback(conf->log)) {
4828                         if (s.to_write)
4829                                 handle_stripe_dirtying(conf, sh, &s, disks);
4830                 } else { /* write back cache */
4831                         int ret = 0;
4832
4833                         /* First, try handle writes in caching phase */
4834                         if (s.to_write)
4835                                 ret = r5c_try_caching_write(conf, sh, &s,
4836                                                             disks);
4837                         /*
4838                          * If caching phase failed: ret == -EAGAIN
4839                          *    OR
4840                          * stripe under reclaim: !caching && injournal
4841                          *
4842                          * fall back to handle_stripe_dirtying()
4843                          */
4844                         if (ret == -EAGAIN ||
4845                             /* stripe under reclaim: !caching && injournal */
4846                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4847                              s.injournal > 0)) {
4848                                 ret = handle_stripe_dirtying(conf, sh, &s,
4849                                                              disks);
4850                                 if (ret == -EAGAIN)
4851                                         goto finish;
4852                         }
4853                 }
4854         }
4855
4856         /* maybe we need to check and possibly fix the parity for this stripe
4857          * Any reads will already have been scheduled, so we just see if enough
4858          * data is available.  The parity check is held off while parity
4859          * dependent operations are in flight.
4860          */
4861         if (sh->check_state ||
4862             (s.syncing && s.locked == 0 &&
4863              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4864              !test_bit(STRIPE_INSYNC, &sh->state))) {
4865                 if (conf->level == 6)
4866                         handle_parity_checks6(conf, sh, &s, disks);
4867                 else
4868                         handle_parity_checks5(conf, sh, &s, disks);
4869         }
4870
4871         if ((s.replacing || s.syncing) && s.locked == 0
4872             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4873             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4874                 /* Write out to replacement devices where possible */
4875                 for (i = 0; i < conf->raid_disks; i++)
4876                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4877                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4878                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4879                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4880                                 s.locked++;
4881                         }
4882                 if (s.replacing)
4883                         set_bit(STRIPE_INSYNC, &sh->state);
4884                 set_bit(STRIPE_REPLACED, &sh->state);
4885         }
4886         if ((s.syncing || s.replacing) && s.locked == 0 &&
4887             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4888             test_bit(STRIPE_INSYNC, &sh->state)) {
4889                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4890                 clear_bit(STRIPE_SYNCING, &sh->state);
4891                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4892                         wake_up(&conf->wait_for_overlap);
4893         }
4894
4895         /* If the failed drives are just a ReadError, then we might need
4896          * to progress the repair/check process
4897          */
4898         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4899                 for (i = 0; i < s.failed; i++) {
4900                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4901                         if (test_bit(R5_ReadError, &dev->flags)
4902                             && !test_bit(R5_LOCKED, &dev->flags)
4903                             && test_bit(R5_UPTODATE, &dev->flags)
4904                                 ) {
4905                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4906                                         set_bit(R5_Wantwrite, &dev->flags);
4907                                         set_bit(R5_ReWrite, &dev->flags);
4908                                         set_bit(R5_LOCKED, &dev->flags);
4909                                         s.locked++;
4910                                 } else {
4911                                         /* let's read it back */
4912                                         set_bit(R5_Wantread, &dev->flags);
4913                                         set_bit(R5_LOCKED, &dev->flags);
4914                                         s.locked++;
4915                                 }
4916                         }
4917                 }
4918
4919         /* Finish reconstruct operations initiated by the expansion process */
4920         if (sh->reconstruct_state == reconstruct_state_result) {
4921                 struct stripe_head *sh_src
4922                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4923                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4924                         /* sh cannot be written until sh_src has been read.
4925                          * so arrange for sh to be delayed a little
4926                          */
4927                         set_bit(STRIPE_DELAYED, &sh->state);
4928                         set_bit(STRIPE_HANDLE, &sh->state);
4929                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4930                                               &sh_src->state))
4931                                 atomic_inc(&conf->preread_active_stripes);
4932                         raid5_release_stripe(sh_src);
4933                         goto finish;
4934                 }
4935                 if (sh_src)
4936                         raid5_release_stripe(sh_src);
4937
4938                 sh->reconstruct_state = reconstruct_state_idle;
4939                 clear_bit(STRIPE_EXPANDING, &sh->state);
4940                 for (i = conf->raid_disks; i--; ) {
4941                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4942                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4943                         s.locked++;
4944                 }
4945         }
4946
4947         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4948             !sh->reconstruct_state) {
4949                 /* Need to write out all blocks after computing parity */
4950                 sh->disks = conf->raid_disks;
4951                 stripe_set_idx(sh->sector, conf, 0, sh);
4952                 schedule_reconstruction(sh, &s, 1, 1);
4953         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4954                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4955                 atomic_dec(&conf->reshape_stripes);
4956                 wake_up(&conf->wait_for_overlap);
4957                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4958         }
4959
4960         if (s.expanding && s.locked == 0 &&
4961             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4962                 handle_stripe_expansion(conf, sh);
4963
4964 finish:
4965         /* wait for this device to become unblocked */
4966         if (unlikely(s.blocked_rdev)) {
4967                 if (conf->mddev->external)
4968                         md_wait_for_blocked_rdev(s.blocked_rdev,
4969                                                  conf->mddev);
4970                 else
4971                         /* Internal metadata will immediately
4972                          * be written by raid5d, so we don't
4973                          * need to wait here.
4974                          */
4975                         rdev_dec_pending(s.blocked_rdev,
4976                                          conf->mddev);
4977         }
4978
4979         if (s.handle_bad_blocks)
4980                 for (i = disks; i--; ) {
4981                         struct md_rdev *rdev;
4982                         struct r5dev *dev = &sh->dev[i];
4983                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4984                                 /* We own a safe reference to the rdev */
4985                                 rdev = conf->disks[i].rdev;
4986                                 if (!rdev_set_badblocks(rdev, sh->sector,
4987                                                         STRIPE_SECTORS, 0))
4988                                         md_error(conf->mddev, rdev);
4989                                 rdev_dec_pending(rdev, conf->mddev);
4990                         }
4991                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4992                                 rdev = conf->disks[i].rdev;
4993                                 rdev_clear_badblocks(rdev, sh->sector,
4994                                                      STRIPE_SECTORS, 0);
4995                                 rdev_dec_pending(rdev, conf->mddev);
4996                         }
4997                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4998                                 rdev = conf->disks[i].replacement;
4999                                 if (!rdev)
5000                                         /* rdev have been moved down */
5001                                         rdev = conf->disks[i].rdev;
5002                                 rdev_clear_badblocks(rdev, sh->sector,
5003                                                      STRIPE_SECTORS, 0);
5004                                 rdev_dec_pending(rdev, conf->mddev);
5005                         }
5006                 }
5007
5008         if (s.ops_request)
5009                 raid_run_ops(sh, s.ops_request);
5010
5011         ops_run_io(sh, &s);
5012
5013         if (s.dec_preread_active) {
5014                 /* We delay this until after ops_run_io so that if make_request
5015                  * is waiting on a flush, it won't continue until the writes
5016                  * have actually been submitted.
5017                  */
5018                 atomic_dec(&conf->preread_active_stripes);
5019                 if (atomic_read(&conf->preread_active_stripes) <
5020                     IO_THRESHOLD)
5021                         md_wakeup_thread(conf->mddev->thread);
5022         }
5023
5024         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5025 }
5026
5027 static void raid5_activate_delayed(struct r5conf *conf)
5028 {
5029         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5030                 while (!list_empty(&conf->delayed_list)) {
5031                         struct list_head *l = conf->delayed_list.next;
5032                         struct stripe_head *sh;
5033                         sh = list_entry(l, struct stripe_head, lru);
5034                         list_del_init(l);
5035                         clear_bit(STRIPE_DELAYED, &sh->state);
5036                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5037                                 atomic_inc(&conf->preread_active_stripes);
5038                         list_add_tail(&sh->lru, &conf->hold_list);
5039                         raid5_wakeup_stripe_thread(sh);
5040                 }
5041         }
5042 }
5043
5044 static void activate_bit_delay(struct r5conf *conf,
5045         struct list_head *temp_inactive_list)
5046 {
5047         /* device_lock is held */
5048         struct list_head head;
5049         list_add(&head, &conf->bitmap_list);
5050         list_del_init(&conf->bitmap_list);
5051         while (!list_empty(&head)) {
5052                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5053                 int hash;
5054                 list_del_init(&sh->lru);
5055                 atomic_inc(&sh->count);
5056                 hash = sh->hash_lock_index;
5057                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5058         }
5059 }
5060
5061 static int raid5_congested(struct mddev *mddev, int bits)
5062 {
5063         struct r5conf *conf = mddev->private;
5064
5065         /* No difference between reads and writes.  Just check
5066          * how busy the stripe_cache is
5067          */
5068
5069         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5070                 return 1;
5071
5072         /* Also checks whether there is pressure on r5cache log space */
5073         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5074                 return 1;
5075         if (conf->quiesce)
5076                 return 1;
5077         if (atomic_read(&conf->empty_inactive_list_nr))
5078                 return 1;
5079
5080         return 0;
5081 }
5082
5083 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5084 {
5085         struct r5conf *conf = mddev->private;
5086         sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
5087         unsigned int chunk_sectors;
5088         unsigned int bio_sectors = bio_sectors(bio);
5089
5090         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5091         return  chunk_sectors >=
5092                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5093 }
5094
5095 /*
5096  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5097  *  later sampled by raid5d.
5098  */
5099 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5100 {
5101         unsigned long flags;
5102
5103         spin_lock_irqsave(&conf->device_lock, flags);
5104
5105         bi->bi_next = conf->retry_read_aligned_list;
5106         conf->retry_read_aligned_list = bi;
5107
5108         spin_unlock_irqrestore(&conf->device_lock, flags);
5109         md_wakeup_thread(conf->mddev->thread);
5110 }
5111
5112 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5113                                          unsigned int *offset)
5114 {
5115         struct bio *bi;
5116
5117         bi = conf->retry_read_aligned;
5118         if (bi) {
5119                 *offset = conf->retry_read_offset;
5120                 conf->retry_read_aligned = NULL;
5121                 return bi;
5122         }
5123         bi = conf->retry_read_aligned_list;
5124         if(bi) {
5125                 conf->retry_read_aligned_list = bi->bi_next;
5126                 bi->bi_next = NULL;
5127                 *offset = 0;
5128         }
5129
5130         return bi;
5131 }
5132
5133 /*
5134  *  The "raid5_align_endio" should check if the read succeeded and if it
5135  *  did, call bio_endio on the original bio (having bio_put the new bio
5136  *  first).
5137  *  If the read failed..
5138  */
5139 static void raid5_align_endio(struct bio *bi)
5140 {
5141         struct bio* raid_bi  = bi->bi_private;
5142         struct mddev *mddev;
5143         struct r5conf *conf;
5144         struct md_rdev *rdev;
5145         blk_status_t error = bi->bi_status;
5146
5147         bio_put(bi);
5148
5149         rdev = (void*)raid_bi->bi_next;
5150         raid_bi->bi_next = NULL;
5151         mddev = rdev->mddev;
5152         conf = mddev->private;
5153
5154         rdev_dec_pending(rdev, conf->mddev);
5155
5156         if (!error) {
5157                 bio_endio(raid_bi);
5158                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5159                         wake_up(&conf->wait_for_quiescent);
5160                 return;
5161         }
5162
5163         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5164
5165         add_bio_to_retry(raid_bi, conf);
5166 }
5167
5168 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5169 {
5170         struct r5conf *conf = mddev->private;
5171         int dd_idx;
5172         struct bio* align_bi;
5173         struct md_rdev *rdev;
5174         sector_t end_sector;
5175
5176         if (!in_chunk_boundary(mddev, raid_bio)) {
5177                 pr_debug("%s: non aligned\n", __func__);
5178                 return 0;
5179         }
5180         /*
5181          * use bio_clone_fast to make a copy of the bio
5182          */
5183         align_bi = bio_clone_fast(raid_bio, GFP_NOIO, mddev->bio_set);
5184         if (!align_bi)
5185                 return 0;
5186         /*
5187          *   set bi_end_io to a new function, and set bi_private to the
5188          *     original bio.
5189          */
5190         align_bi->bi_end_io  = raid5_align_endio;
5191         align_bi->bi_private = raid_bio;
5192         /*
5193          *      compute position
5194          */
5195         align_bi->bi_iter.bi_sector =
5196                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5197                                      0, &dd_idx, NULL);
5198
5199         end_sector = bio_end_sector(align_bi);
5200         rcu_read_lock();
5201         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5202         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5203             rdev->recovery_offset < end_sector) {
5204                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5205                 if (rdev &&
5206                     (test_bit(Faulty, &rdev->flags) ||
5207                     !(test_bit(In_sync, &rdev->flags) ||
5208                       rdev->recovery_offset >= end_sector)))
5209                         rdev = NULL;
5210         }
5211
5212         if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5213                 rcu_read_unlock();
5214                 bio_put(align_bi);
5215                 return 0;
5216         }
5217
5218         if (rdev) {
5219                 sector_t first_bad;
5220                 int bad_sectors;
5221
5222                 atomic_inc(&rdev->nr_pending);
5223                 rcu_read_unlock();
5224                 raid_bio->bi_next = (void*)rdev;
5225                 align_bi->bi_bdev =  rdev->bdev;
5226                 bio_clear_flag(align_bi, BIO_SEG_VALID);
5227
5228                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5229                                 bio_sectors(align_bi),
5230                                 &first_bad, &bad_sectors)) {
5231                         bio_put(align_bi);
5232                         rdev_dec_pending(rdev, mddev);
5233                         return 0;
5234                 }
5235
5236                 /* No reshape active, so we can trust rdev->data_offset */
5237                 align_bi->bi_iter.bi_sector += rdev->data_offset;
5238
5239                 spin_lock_irq(&conf->device_lock);
5240                 wait_event_lock_irq(conf->wait_for_quiescent,
5241                                     conf->quiesce == 0,
5242                                     conf->device_lock);
5243                 atomic_inc(&conf->active_aligned_reads);
5244                 spin_unlock_irq(&conf->device_lock);
5245
5246                 if (mddev->gendisk)
5247                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
5248                                               align_bi, disk_devt(mddev->gendisk),
5249                                               raid_bio->bi_iter.bi_sector);
5250                 generic_make_request(align_bi);
5251                 return 1;
5252         } else {
5253                 rcu_read_unlock();
5254                 bio_put(align_bi);
5255                 return 0;
5256         }
5257 }
5258
5259 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5260 {
5261         struct bio *split;
5262         sector_t sector = raid_bio->bi_iter.bi_sector;
5263         unsigned chunk_sects = mddev->chunk_sectors;
5264         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5265
5266         if (sectors < bio_sectors(raid_bio)) {
5267                 struct r5conf *conf = mddev->private;
5268                 split = bio_split(raid_bio, sectors, GFP_NOIO, conf->bio_split);
5269                 bio_chain(split, raid_bio);
5270                 generic_make_request(raid_bio);
5271                 raid_bio = split;
5272         }
5273
5274         if (!raid5_read_one_chunk(mddev, raid_bio))
5275                 return raid_bio;
5276
5277         return NULL;
5278 }
5279
5280 /* __get_priority_stripe - get the next stripe to process
5281  *
5282  * Full stripe writes are allowed to pass preread active stripes up until
5283  * the bypass_threshold is exceeded.  In general the bypass_count
5284  * increments when the handle_list is handled before the hold_list; however, it
5285  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5286  * stripe with in flight i/o.  The bypass_count will be reset when the
5287  * head of the hold_list has changed, i.e. the head was promoted to the
5288  * handle_list.
5289  */
5290 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5291 {
5292         struct stripe_head *sh, *tmp;
5293         struct list_head *handle_list = NULL;
5294         struct r5worker_group *wg;
5295         bool second_try = !r5c_is_writeback(conf->log) &&
5296                 !r5l_log_disk_error(conf);
5297         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5298                 r5l_log_disk_error(conf);
5299
5300 again:
5301         wg = NULL;
5302         sh = NULL;
5303         if (conf->worker_cnt_per_group == 0) {
5304                 handle_list = try_loprio ? &conf->loprio_list :
5305                                         &conf->handle_list;
5306         } else if (group != ANY_GROUP) {
5307                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5308                                 &conf->worker_groups[group].handle_list;
5309                 wg = &conf->worker_groups[group];
5310         } else {
5311                 int i;
5312                 for (i = 0; i < conf->group_cnt; i++) {
5313                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5314                                 &conf->worker_groups[i].handle_list;
5315                         wg = &conf->worker_groups[i];
5316                         if (!list_empty(handle_list))
5317                                 break;
5318                 }
5319         }
5320
5321         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5322                   __func__,
5323                   list_empty(handle_list) ? "empty" : "busy",
5324                   list_empty(&conf->hold_list) ? "empty" : "busy",
5325                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5326
5327         if (!list_empty(handle_list)) {
5328                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5329
5330                 if (list_empty(&conf->hold_list))
5331                         conf->bypass_count = 0;
5332                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5333                         if (conf->hold_list.next == conf->last_hold)
5334                                 conf->bypass_count++;
5335                         else {
5336                                 conf->last_hold = conf->hold_list.next;
5337                                 conf->bypass_count -= conf->bypass_threshold;
5338                                 if (conf->bypass_count < 0)
5339                                         conf->bypass_count = 0;
5340                         }
5341                 }
5342         } else if (!list_empty(&conf->hold_list) &&
5343                    ((conf->bypass_threshold &&
5344                      conf->bypass_count > conf->bypass_threshold) ||
5345                     atomic_read(&conf->pending_full_writes) == 0)) {
5346
5347                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5348                         if (conf->worker_cnt_per_group == 0 ||
5349                             group == ANY_GROUP ||
5350                             !cpu_online(tmp->cpu) ||
5351                             cpu_to_group(tmp->cpu) == group) {
5352                                 sh = tmp;
5353                                 break;
5354                         }
5355                 }
5356
5357                 if (sh) {
5358                         conf->bypass_count -= conf->bypass_threshold;
5359                         if (conf->bypass_count < 0)
5360                                 conf->bypass_count = 0;
5361                 }
5362                 wg = NULL;
5363         }
5364
5365         if (!sh) {
5366                 if (second_try)
5367                         return NULL;
5368                 second_try = true;
5369                 try_loprio = !try_loprio;
5370                 goto again;
5371         }
5372
5373         if (wg) {
5374                 wg->stripes_cnt--;
5375                 sh->group = NULL;
5376         }
5377         list_del_init(&sh->lru);
5378         BUG_ON(atomic_inc_return(&sh->count) != 1);
5379         return sh;
5380 }
5381
5382 struct raid5_plug_cb {
5383         struct blk_plug_cb      cb;
5384         struct list_head        list;
5385         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5386 };
5387
5388 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5389 {
5390         struct raid5_plug_cb *cb = container_of(
5391                 blk_cb, struct raid5_plug_cb, cb);
5392         struct stripe_head *sh;
5393         struct mddev *mddev = cb->cb.data;
5394         struct r5conf *conf = mddev->private;
5395         int cnt = 0;
5396         int hash;
5397
5398         if (cb->list.next && !list_empty(&cb->list)) {
5399                 spin_lock_irq(&conf->device_lock);
5400                 while (!list_empty(&cb->list)) {
5401                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5402                         list_del_init(&sh->lru);
5403                         /*
5404                          * avoid race release_stripe_plug() sees
5405                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5406                          * is still in our list
5407                          */
5408                         smp_mb__before_atomic();
5409                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5410                         /*
5411                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5412                          * case, the count is always > 1 here
5413                          */
5414                         hash = sh->hash_lock_index;
5415                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5416                         cnt++;
5417                 }
5418                 spin_unlock_irq(&conf->device_lock);
5419         }
5420         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5421                                      NR_STRIPE_HASH_LOCKS);
5422         if (mddev->queue)
5423                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5424         kfree(cb);
5425 }
5426
5427 static void release_stripe_plug(struct mddev *mddev,
5428                                 struct stripe_head *sh)
5429 {
5430         struct blk_plug_cb *blk_cb = blk_check_plugged(
5431                 raid5_unplug, mddev,
5432                 sizeof(struct raid5_plug_cb));
5433         struct raid5_plug_cb *cb;
5434
5435         if (!blk_cb) {
5436                 raid5_release_stripe(sh);
5437                 return;
5438         }
5439
5440         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5441
5442         if (cb->list.next == NULL) {
5443                 int i;
5444                 INIT_LIST_HEAD(&cb->list);
5445                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5446                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5447         }
5448
5449         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5450                 list_add_tail(&sh->lru, &cb->list);
5451         else
5452                 raid5_release_stripe(sh);
5453 }
5454
5455 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5456 {
5457         struct r5conf *conf = mddev->private;
5458         sector_t logical_sector, last_sector;
5459         struct stripe_head *sh;
5460         int stripe_sectors;
5461
5462         if (mddev->reshape_position != MaxSector)
5463                 /* Skip discard while reshape is happening */
5464                 return;
5465
5466         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5467         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5468
5469         bi->bi_next = NULL;
5470
5471         stripe_sectors = conf->chunk_sectors *
5472                 (conf->raid_disks - conf->max_degraded);
5473         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5474                                                stripe_sectors);
5475         sector_div(last_sector, stripe_sectors);
5476
5477         logical_sector *= conf->chunk_sectors;
5478         last_sector *= conf->chunk_sectors;
5479
5480         for (; logical_sector < last_sector;
5481              logical_sector += STRIPE_SECTORS) {
5482                 DEFINE_WAIT(w);
5483                 int d;
5484         again:
5485                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5486                 prepare_to_wait(&conf->wait_for_overlap, &w,
5487                                 TASK_UNINTERRUPTIBLE);
5488                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5489                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5490                         raid5_release_stripe(sh);
5491                         schedule();
5492                         goto again;
5493                 }
5494                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5495                 spin_lock_irq(&sh->stripe_lock);
5496                 for (d = 0; d < conf->raid_disks; d++) {
5497                         if (d == sh->pd_idx || d == sh->qd_idx)
5498                                 continue;
5499                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5500                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5501                                 spin_unlock_irq(&sh->stripe_lock);
5502                                 raid5_release_stripe(sh);
5503                                 schedule();
5504                                 goto again;
5505                         }
5506                 }
5507                 set_bit(STRIPE_DISCARD, &sh->state);
5508                 finish_wait(&conf->wait_for_overlap, &w);
5509                 sh->overwrite_disks = 0;
5510                 for (d = 0; d < conf->raid_disks; d++) {
5511                         if (d == sh->pd_idx || d == sh->qd_idx)
5512                                 continue;
5513                         sh->dev[d].towrite = bi;
5514                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5515                         bio_inc_remaining(bi);
5516                         md_write_inc(mddev, bi);
5517                         sh->overwrite_disks++;
5518                 }
5519                 spin_unlock_irq(&sh->stripe_lock);
5520                 if (conf->mddev->bitmap) {
5521                         for (d = 0;
5522                              d < conf->raid_disks - conf->max_degraded;
5523                              d++)
5524                                 bitmap_startwrite(mddev->bitmap,
5525                                                   sh->sector,
5526                                                   STRIPE_SECTORS,
5527                                                   0);
5528                         sh->bm_seq = conf->seq_flush + 1;
5529                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5530                 }
5531
5532                 set_bit(STRIPE_HANDLE, &sh->state);
5533                 clear_bit(STRIPE_DELAYED, &sh->state);
5534                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5535                         atomic_inc(&conf->preread_active_stripes);
5536                 release_stripe_plug(mddev, sh);
5537         }
5538
5539         bio_endio(bi);
5540 }
5541
5542 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5543 {
5544         struct r5conf *conf = mddev->private;
5545         int dd_idx;
5546         sector_t new_sector;
5547         sector_t logical_sector, last_sector;
5548         struct stripe_head *sh;
5549         const int rw = bio_data_dir(bi);
5550         DEFINE_WAIT(w);
5551         bool do_prepare;
5552         bool do_flush = false;
5553
5554         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5555                 int ret = r5l_handle_flush_request(conf->log, bi);
5556
5557                 if (ret == 0)
5558                         return true;
5559                 if (ret == -ENODEV) {
5560                         md_flush_request(mddev, bi);
5561                         return true;
5562                 }
5563                 /* ret == -EAGAIN, fallback */
5564                 /*
5565                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5566                  * we need to flush journal device
5567                  */
5568                 do_flush = bi->bi_opf & REQ_PREFLUSH;
5569         }
5570
5571         if (!md_write_start(mddev, bi))
5572                 return false;
5573         /*
5574          * If array is degraded, better not do chunk aligned read because
5575          * later we might have to read it again in order to reconstruct
5576          * data on failed drives.
5577          */
5578         if (rw == READ && mddev->degraded == 0 &&
5579             mddev->reshape_position == MaxSector) {
5580                 bi = chunk_aligned_read(mddev, bi);
5581                 if (!bi)
5582                         return true;
5583         }
5584
5585         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5586                 make_discard_request(mddev, bi);
5587                 md_write_end(mddev);
5588                 return true;
5589         }
5590
5591         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5592         last_sector = bio_end_sector(bi);
5593         bi->bi_next = NULL;
5594
5595         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5596         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5597                 int previous;
5598                 int seq;
5599
5600                 do_prepare = false;
5601         retry:
5602                 seq = read_seqcount_begin(&conf->gen_lock);
5603                 previous = 0;
5604                 if (do_prepare)
5605                         prepare_to_wait(&conf->wait_for_overlap, &w,
5606                                 TASK_UNINTERRUPTIBLE);
5607                 if (unlikely(conf->reshape_progress != MaxSector)) {
5608                         /* spinlock is needed as reshape_progress may be
5609                          * 64bit on a 32bit platform, and so it might be
5610                          * possible to see a half-updated value
5611                          * Of course reshape_progress could change after
5612                          * the lock is dropped, so once we get a reference
5613                          * to the stripe that we think it is, we will have
5614                          * to check again.
5615                          */
5616                         spin_lock_irq(&conf->device_lock);
5617                         if (mddev->reshape_backwards
5618                             ? logical_sector < conf->reshape_progress
5619                             : logical_sector >= conf->reshape_progress) {
5620                                 previous = 1;
5621                         } else {
5622                                 if (mddev->reshape_backwards
5623                                     ? logical_sector < conf->reshape_safe
5624                                     : logical_sector >= conf->reshape_safe) {
5625                                         spin_unlock_irq(&conf->device_lock);
5626                                         schedule();
5627                                         do_prepare = true;
5628                                         goto retry;
5629                                 }
5630                         }
5631                         spin_unlock_irq(&conf->device_lock);
5632                 }
5633
5634                 new_sector = raid5_compute_sector(conf, logical_sector,
5635                                                   previous,
5636                                                   &dd_idx, NULL);
5637                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5638                         (unsigned long long)new_sector,
5639                         (unsigned long long)logical_sector);
5640
5641                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5642                                        (bi->bi_opf & REQ_RAHEAD), 0);
5643                 if (sh) {
5644                         if (unlikely(previous)) {
5645                                 /* expansion might have moved on while waiting for a
5646                                  * stripe, so we must do the range check again.
5647                                  * Expansion could still move past after this
5648                                  * test, but as we are holding a reference to
5649                                  * 'sh', we know that if that happens,
5650                                  *  STRIPE_EXPANDING will get set and the expansion
5651                                  * won't proceed until we finish with the stripe.
5652                                  */
5653                                 int must_retry = 0;
5654                                 spin_lock_irq(&conf->device_lock);
5655                                 if (mddev->reshape_backwards
5656                                     ? logical_sector >= conf->reshape_progress
5657                                     : logical_sector < conf->reshape_progress)
5658                                         /* mismatch, need to try again */
5659                                         must_retry = 1;
5660                                 spin_unlock_irq(&conf->device_lock);
5661                                 if (must_retry) {
5662                                         raid5_release_stripe(sh);
5663                                         schedule();
5664                                         do_prepare = true;
5665                                         goto retry;
5666                                 }
5667                         }
5668                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5669                                 /* Might have got the wrong stripe_head
5670                                  * by accident
5671                                  */
5672                                 raid5_release_stripe(sh);
5673                                 goto retry;
5674                         }
5675
5676                         if (rw == WRITE &&
5677                             logical_sector >= mddev->suspend_lo &&
5678                             logical_sector < mddev->suspend_hi) {
5679                                 raid5_release_stripe(sh);
5680                                 /* As the suspend_* range is controlled by
5681                                  * userspace, we want an interruptible
5682                                  * wait.
5683                                  */
5684                                 prepare_to_wait(&conf->wait_for_overlap,
5685                                                 &w, TASK_INTERRUPTIBLE);
5686                                 if (logical_sector >= mddev->suspend_lo &&
5687                                     logical_sector < mddev->suspend_hi) {
5688                                         sigset_t full, old;
5689                                         sigfillset(&full);
5690                                         sigprocmask(SIG_BLOCK, &full, &old);
5691                                         schedule();
5692                                         sigprocmask(SIG_SETMASK, &old, NULL);
5693                                         do_prepare = true;
5694                                 }
5695                                 goto retry;
5696                         }
5697
5698                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5699                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5700                                 /* Stripe is busy expanding or
5701                                  * add failed due to overlap.  Flush everything
5702                                  * and wait a while
5703                                  */
5704                                 md_wakeup_thread(mddev->thread);
5705                                 raid5_release_stripe(sh);
5706                                 schedule();
5707                                 do_prepare = true;
5708                                 goto retry;
5709                         }
5710                         if (do_flush) {
5711                                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5712                                 /* we only need flush for one stripe */
5713                                 do_flush = false;
5714                         }
5715
5716                         set_bit(STRIPE_HANDLE, &sh->state);
5717                         clear_bit(STRIPE_DELAYED, &sh->state);
5718                         if ((!sh->batch_head || sh == sh->batch_head) &&
5719                             (bi->bi_opf & REQ_SYNC) &&
5720                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5721                                 atomic_inc(&conf->preread_active_stripes);
5722                         release_stripe_plug(mddev, sh);
5723                 } else {
5724                         /* cannot get stripe for read-ahead, just give-up */
5725                         bi->bi_status = BLK_STS_IOERR;
5726                         break;
5727                 }
5728         }
5729         finish_wait(&conf->wait_for_overlap, &w);
5730
5731         if (rw == WRITE)
5732                 md_write_end(mddev);
5733         bio_endio(bi);
5734         return true;
5735 }
5736
5737 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5738
5739 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5740 {
5741         /* reshaping is quite different to recovery/resync so it is
5742          * handled quite separately ... here.
5743          *
5744          * On each call to sync_request, we gather one chunk worth of
5745          * destination stripes and flag them as expanding.
5746          * Then we find all the source stripes and request reads.
5747          * As the reads complete, handle_stripe will copy the data
5748          * into the destination stripe and release that stripe.
5749          */
5750         struct r5conf *conf = mddev->private;
5751         struct stripe_head *sh;
5752         sector_t first_sector, last_sector;
5753         int raid_disks = conf->previous_raid_disks;
5754         int data_disks = raid_disks - conf->max_degraded;
5755         int new_data_disks = conf->raid_disks - conf->max_degraded;
5756         int i;
5757         int dd_idx;
5758         sector_t writepos, readpos, safepos;
5759         sector_t stripe_addr;
5760         int reshape_sectors;
5761         struct list_head stripes;
5762         sector_t retn;
5763
5764         if (sector_nr == 0) {
5765                 /* If restarting in the middle, skip the initial sectors */
5766                 if (mddev->reshape_backwards &&
5767                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5768                         sector_nr = raid5_size(mddev, 0, 0)
5769                                 - conf->reshape_progress;
5770                 } else if (mddev->reshape_backwards &&
5771                            conf->reshape_progress == MaxSector) {
5772                         /* shouldn't happen, but just in case, finish up.*/
5773                         sector_nr = MaxSector;
5774                 } else if (!mddev->reshape_backwards &&
5775                            conf->reshape_progress > 0)
5776                         sector_nr = conf->reshape_progress;
5777                 sector_div(sector_nr, new_data_disks);
5778                 if (sector_nr) {
5779                         mddev->curr_resync_completed = sector_nr;
5780                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5781                         *skipped = 1;
5782                         retn = sector_nr;
5783                         goto finish;
5784                 }
5785         }
5786
5787         /* We need to process a full chunk at a time.
5788          * If old and new chunk sizes differ, we need to process the
5789          * largest of these
5790          */
5791
5792         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5793
5794         /* We update the metadata at least every 10 seconds, or when
5795          * the data about to be copied would over-write the source of
5796          * the data at the front of the range.  i.e. one new_stripe
5797          * along from reshape_progress new_maps to after where
5798          * reshape_safe old_maps to
5799          */
5800         writepos = conf->reshape_progress;
5801         sector_div(writepos, new_data_disks);
5802         readpos = conf->reshape_progress;
5803         sector_div(readpos, data_disks);
5804         safepos = conf->reshape_safe;
5805         sector_div(safepos, data_disks);
5806         if (mddev->reshape_backwards) {
5807                 BUG_ON(writepos < reshape_sectors);
5808                 writepos -= reshape_sectors;
5809                 readpos += reshape_sectors;
5810                 safepos += reshape_sectors;
5811         } else {
5812                 writepos += reshape_sectors;
5813                 /* readpos and safepos are worst-case calculations.
5814                  * A negative number is overly pessimistic, and causes
5815                  * obvious problems for unsigned storage.  So clip to 0.
5816                  */
5817                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5818                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5819         }
5820
5821         /* Having calculated the 'writepos' possibly use it
5822          * to set 'stripe_addr' which is where we will write to.
5823          */
5824         if (mddev->reshape_backwards) {
5825                 BUG_ON(conf->reshape_progress == 0);
5826                 stripe_addr = writepos;
5827                 BUG_ON((mddev->dev_sectors &
5828                         ~((sector_t)reshape_sectors - 1))
5829                        - reshape_sectors - stripe_addr
5830                        != sector_nr);
5831         } else {
5832                 BUG_ON(writepos != sector_nr + reshape_sectors);
5833                 stripe_addr = sector_nr;
5834         }
5835
5836         /* 'writepos' is the most advanced device address we might write.
5837          * 'readpos' is the least advanced device address we might read.
5838          * 'safepos' is the least address recorded in the metadata as having
5839          *     been reshaped.
5840          * If there is a min_offset_diff, these are adjusted either by
5841          * increasing the safepos/readpos if diff is negative, or
5842          * increasing writepos if diff is positive.
5843          * If 'readpos' is then behind 'writepos', there is no way that we can
5844          * ensure safety in the face of a crash - that must be done by userspace
5845          * making a backup of the data.  So in that case there is no particular
5846          * rush to update metadata.
5847          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5848          * update the metadata to advance 'safepos' to match 'readpos' so that
5849          * we can be safe in the event of a crash.
5850          * So we insist on updating metadata if safepos is behind writepos and
5851          * readpos is beyond writepos.
5852          * In any case, update the metadata every 10 seconds.
5853          * Maybe that number should be configurable, but I'm not sure it is
5854          * worth it.... maybe it could be a multiple of safemode_delay???
5855          */
5856         if (conf->min_offset_diff < 0) {
5857                 safepos += -conf->min_offset_diff;
5858                 readpos += -conf->min_offset_diff;
5859         } else
5860                 writepos += conf->min_offset_diff;
5861
5862         if ((mddev->reshape_backwards
5863              ? (safepos > writepos && readpos < writepos)
5864              : (safepos < writepos && readpos > writepos)) ||
5865             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5866                 /* Cannot proceed until we've updated the superblock... */
5867                 wait_event(conf->wait_for_overlap,
5868                            atomic_read(&conf->reshape_stripes)==0
5869                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5870                 if (atomic_read(&conf->reshape_stripes) != 0)
5871                         return 0;
5872                 mddev->reshape_position = conf->reshape_progress;
5873                 mddev->curr_resync_completed = sector_nr;
5874                 conf->reshape_checkpoint = jiffies;
5875                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5876                 md_wakeup_thread(mddev->thread);
5877                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5878                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5879                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5880                         return 0;
5881                 spin_lock_irq(&conf->device_lock);
5882                 conf->reshape_safe = mddev->reshape_position;
5883                 spin_unlock_irq(&conf->device_lock);
5884                 wake_up(&conf->wait_for_overlap);
5885                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5886         }
5887
5888         INIT_LIST_HEAD(&stripes);
5889         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5890                 int j;
5891                 int skipped_disk = 0;
5892                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5893                 set_bit(STRIPE_EXPANDING, &sh->state);
5894                 atomic_inc(&conf->reshape_stripes);
5895                 /* If any of this stripe is beyond the end of the old
5896                  * array, then we need to zero those blocks
5897                  */
5898                 for (j=sh->disks; j--;) {
5899                         sector_t s;
5900                         if (j == sh->pd_idx)
5901                                 continue;
5902                         if (conf->level == 6 &&
5903                             j == sh->qd_idx)
5904                                 continue;
5905                         s = raid5_compute_blocknr(sh, j, 0);
5906                         if (s < raid5_size(mddev, 0, 0)) {
5907                                 skipped_disk = 1;
5908                                 continue;
5909                         }
5910                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5911                         set_bit(R5_Expanded, &sh->dev[j].flags);
5912                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5913                 }
5914                 if (!skipped_disk) {
5915                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5916                         set_bit(STRIPE_HANDLE, &sh->state);
5917                 }
5918                 list_add(&sh->lru, &stripes);
5919         }
5920         spin_lock_irq(&conf->device_lock);
5921         if (mddev->reshape_backwards)
5922                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5923         else
5924                 conf->reshape_progress += reshape_sectors * new_data_disks;
5925         spin_unlock_irq(&conf->device_lock);
5926         /* Ok, those stripe are ready. We can start scheduling
5927          * reads on the source stripes.
5928          * The source stripes are determined by mapping the first and last
5929          * block on the destination stripes.
5930          */
5931         first_sector =
5932                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5933                                      1, &dd_idx, NULL);
5934         last_sector =
5935                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5936                                             * new_data_disks - 1),
5937                                      1, &dd_idx, NULL);
5938         if (last_sector >= mddev->dev_sectors)
5939                 last_sector = mddev->dev_sectors - 1;
5940         while (first_sector <= last_sector) {
5941                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5942                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5943                 set_bit(STRIPE_HANDLE, &sh->state);
5944                 raid5_release_stripe(sh);
5945                 first_sector += STRIPE_SECTORS;
5946         }
5947         /* Now that the sources are clearly marked, we can release
5948          * the destination stripes
5949          */
5950         while (!list_empty(&stripes)) {
5951                 sh = list_entry(stripes.next, struct stripe_head, lru);
5952                 list_del_init(&sh->lru);
5953                 raid5_release_stripe(sh);
5954         }
5955         /* If this takes us to the resync_max point where we have to pause,
5956          * then we need to write out the superblock.
5957          */
5958         sector_nr += reshape_sectors;
5959         retn = reshape_sectors;
5960 finish:
5961         if (mddev->curr_resync_completed > mddev->resync_max ||
5962             (sector_nr - mddev->curr_resync_completed) * 2
5963             >= mddev->resync_max - mddev->curr_resync_completed) {
5964                 /* Cannot proceed until we've updated the superblock... */
5965                 wait_event(conf->wait_for_overlap,
5966                            atomic_read(&conf->reshape_stripes) == 0
5967                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5968                 if (atomic_read(&conf->reshape_stripes) != 0)
5969                         goto ret;
5970                 mddev->reshape_position = conf->reshape_progress;
5971                 mddev->curr_resync_completed = sector_nr;
5972                 conf->reshape_checkpoint = jiffies;
5973                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5974                 md_wakeup_thread(mddev->thread);
5975                 wait_event(mddev->sb_wait,
5976                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
5977                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5978                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5979                         goto ret;
5980                 spin_lock_irq(&conf->device_lock);
5981                 conf->reshape_safe = mddev->reshape_position;
5982                 spin_unlock_irq(&conf->device_lock);
5983                 wake_up(&conf->wait_for_overlap);
5984                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5985         }
5986 ret:
5987         return retn;
5988 }
5989
5990 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
5991                                           int *skipped)
5992 {
5993         struct r5conf *conf = mddev->private;
5994         struct stripe_head *sh;
5995         sector_t max_sector = mddev->dev_sectors;
5996         sector_t sync_blocks;
5997         int still_degraded = 0;
5998         int i;
5999
6000         if (sector_nr >= max_sector) {
6001                 /* just being told to finish up .. nothing much to do */
6002
6003                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6004                         end_reshape(conf);
6005                         return 0;
6006                 }
6007
6008                 if (mddev->curr_resync < max_sector) /* aborted */
6009                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6010                                         &sync_blocks, 1);
6011                 else /* completed sync */
6012                         conf->fullsync = 0;
6013                 bitmap_close_sync(mddev->bitmap);
6014
6015                 return 0;
6016         }
6017
6018         /* Allow raid5_quiesce to complete */
6019         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6020
6021         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6022                 return reshape_request(mddev, sector_nr, skipped);
6023
6024         /* No need to check resync_max as we never do more than one
6025          * stripe, and as resync_max will always be on a chunk boundary,
6026          * if the check in md_do_sync didn't fire, there is no chance
6027          * of overstepping resync_max here
6028          */
6029
6030         /* if there is too many failed drives and we are trying
6031          * to resync, then assert that we are finished, because there is
6032          * nothing we can do.
6033          */
6034         if (mddev->degraded >= conf->max_degraded &&
6035             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6036                 sector_t rv = mddev->dev_sectors - sector_nr;
6037                 *skipped = 1;
6038                 return rv;
6039         }
6040         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6041             !conf->fullsync &&
6042             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6043             sync_blocks >= STRIPE_SECTORS) {
6044                 /* we can skip this block, and probably more */
6045                 sync_blocks /= STRIPE_SECTORS;
6046                 *skipped = 1;
6047                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6048         }
6049
6050         bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6051
6052         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6053         if (sh == NULL) {
6054                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6055                 /* make sure we don't swamp the stripe cache if someone else
6056                  * is trying to get access
6057                  */
6058                 schedule_timeout_uninterruptible(1);
6059         }
6060         /* Need to check if array will still be degraded after recovery/resync
6061          * Note in case of > 1 drive failures it's possible we're rebuilding
6062          * one drive while leaving another faulty drive in array.
6063          */
6064         rcu_read_lock();
6065         for (i = 0; i < conf->raid_disks; i++) {
6066                 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
6067
6068                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6069                         still_degraded = 1;
6070         }
6071         rcu_read_unlock();
6072
6073         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6074
6075         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6076         set_bit(STRIPE_HANDLE, &sh->state);
6077
6078         raid5_release_stripe(sh);
6079
6080         return STRIPE_SECTORS;
6081 }
6082
6083 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6084                                unsigned int offset)
6085 {
6086         /* We may not be able to submit a whole bio at once as there
6087          * may not be enough stripe_heads available.
6088          * We cannot pre-allocate enough stripe_heads as we may need
6089          * more than exist in the cache (if we allow ever large chunks).
6090          * So we do one stripe head at a time and record in
6091          * ->bi_hw_segments how many have been done.
6092          *
6093          * We *know* that this entire raid_bio is in one chunk, so
6094          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6095          */
6096         struct stripe_head *sh;
6097         int dd_idx;
6098         sector_t sector, logical_sector, last_sector;
6099         int scnt = 0;
6100         int handled = 0;
6101
6102         logical_sector = raid_bio->bi_iter.bi_sector &
6103                 ~((sector_t)STRIPE_SECTORS-1);
6104         sector = raid5_compute_sector(conf, logical_sector,
6105                                       0, &dd_idx, NULL);
6106         last_sector = bio_end_sector(raid_bio);
6107
6108         for (; logical_sector < last_sector;
6109              logical_sector += STRIPE_SECTORS,
6110                      sector += STRIPE_SECTORS,
6111                      scnt++) {
6112
6113                 if (scnt < offset)
6114                         /* already done this stripe */
6115                         continue;
6116
6117                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6118
6119                 if (!sh) {
6120                         /* failed to get a stripe - must wait */
6121                         conf->retry_read_aligned = raid_bio;
6122                         conf->retry_read_offset = scnt;
6123                         return handled;
6124                 }
6125
6126                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6127                         raid5_release_stripe(sh);
6128                         conf->retry_read_aligned = raid_bio;
6129                         conf->retry_read_offset = scnt;
6130                         return handled;
6131                 }
6132
6133                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6134                 handle_stripe(sh);
6135                 raid5_release_stripe(sh);
6136                 handled++;
6137         }
6138
6139         bio_endio(raid_bio);
6140
6141         if (atomic_dec_and_test(&conf->active_aligned_reads))
6142                 wake_up(&conf->wait_for_quiescent);
6143         return handled;
6144 }
6145
6146 static int handle_active_stripes(struct r5conf *conf, int group,
6147                                  struct r5worker *worker,
6148                                  struct list_head *temp_inactive_list)
6149 {
6150         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6151         int i, batch_size = 0, hash;
6152         bool release_inactive = false;
6153
6154         while (batch_size < MAX_STRIPE_BATCH &&
6155                         (sh = __get_priority_stripe(conf, group)) != NULL)
6156                 batch[batch_size++] = sh;
6157
6158         if (batch_size == 0) {
6159                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6160                         if (!list_empty(temp_inactive_list + i))
6161                                 break;
6162                 if (i == NR_STRIPE_HASH_LOCKS) {
6163                         spin_unlock_irq(&conf->device_lock);
6164                         r5l_flush_stripe_to_raid(conf->log);
6165                         spin_lock_irq(&conf->device_lock);
6166                         return batch_size;
6167                 }
6168                 release_inactive = true;
6169         }
6170         spin_unlock_irq(&conf->device_lock);
6171
6172         release_inactive_stripe_list(conf, temp_inactive_list,
6173                                      NR_STRIPE_HASH_LOCKS);
6174
6175         r5l_flush_stripe_to_raid(conf->log);
6176         if (release_inactive) {
6177                 spin_lock_irq(&conf->device_lock);
6178                 return 0;
6179         }
6180
6181         for (i = 0; i < batch_size; i++)
6182                 handle_stripe(batch[i]);
6183         log_write_stripe_run(conf);
6184
6185         cond_resched();
6186
6187         spin_lock_irq(&conf->device_lock);
6188         for (i = 0; i < batch_size; i++) {
6189                 hash = batch[i]->hash_lock_index;
6190                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6191         }
6192         return batch_size;
6193 }
6194
6195 static void raid5_do_work(struct work_struct *work)
6196 {
6197         struct r5worker *worker = container_of(work, struct r5worker, work);
6198         struct r5worker_group *group = worker->group;
6199         struct r5conf *conf = group->conf;
6200         struct mddev *mddev = conf->mddev;
6201         int group_id = group - conf->worker_groups;
6202         int handled;
6203         struct blk_plug plug;
6204
6205         pr_debug("+++ raid5worker active\n");
6206
6207         blk_start_plug(&plug);
6208         handled = 0;
6209         spin_lock_irq(&conf->device_lock);
6210         while (1) {
6211                 int batch_size, released;
6212
6213                 released = release_stripe_list(conf, worker->temp_inactive_list);
6214
6215                 batch_size = handle_active_stripes(conf, group_id, worker,
6216                                                    worker->temp_inactive_list);
6217                 worker->working = false;
6218                 if (!batch_size && !released)
6219                         break;
6220                 handled += batch_size;
6221                 wait_event_lock_irq(mddev->sb_wait,
6222                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6223                         conf->device_lock);
6224         }
6225         pr_debug("%d stripes handled\n", handled);
6226
6227         spin_unlock_irq(&conf->device_lock);
6228
6229         flush_deferred_bios(conf);
6230
6231         r5l_flush_stripe_to_raid(conf->log);
6232
6233         async_tx_issue_pending_all();
6234         blk_finish_plug(&plug);
6235
6236         pr_debug("--- raid5worker inactive\n");
6237 }
6238
6239 /*
6240  * This is our raid5 kernel thread.
6241  *
6242  * We scan the hash table for stripes which can be handled now.
6243  * During the scan, completed stripes are saved for us by the interrupt
6244  * handler, so that they will not have to wait for our next wakeup.
6245  */
6246 static void raid5d(struct md_thread *thread)
6247 {
6248         struct mddev *mddev = thread->mddev;
6249         struct r5conf *conf = mddev->private;
6250         int handled;
6251         struct blk_plug plug;
6252
6253         pr_debug("+++ raid5d active\n");
6254
6255         md_check_recovery(mddev);
6256
6257         blk_start_plug(&plug);
6258         handled = 0;
6259         spin_lock_irq(&conf->device_lock);
6260         while (1) {
6261                 struct bio *bio;
6262                 int batch_size, released;
6263                 unsigned int offset;
6264
6265                 released = release_stripe_list(conf, conf->temp_inactive_list);
6266                 if (released)
6267                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6268
6269                 if (
6270                     !list_empty(&conf->bitmap_list)) {
6271                         /* Now is a good time to flush some bitmap updates */
6272                         conf->seq_flush++;
6273                         spin_unlock_irq(&conf->device_lock);
6274                         bitmap_unplug(mddev->bitmap);
6275                         spin_lock_irq(&conf->device_lock);
6276                         conf->seq_write = conf->seq_flush;
6277                         activate_bit_delay(conf, conf->temp_inactive_list);
6278                 }
6279                 raid5_activate_delayed(conf);
6280
6281                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6282                         int ok;
6283                         spin_unlock_irq(&conf->device_lock);
6284                         ok = retry_aligned_read(conf, bio, offset);
6285                         spin_lock_irq(&conf->device_lock);
6286                         if (!ok)
6287                                 break;
6288                         handled++;
6289                 }
6290
6291                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6292                                                    conf->temp_inactive_list);
6293                 if (!batch_size && !released)
6294                         break;
6295                 handled += batch_size;
6296
6297                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6298                         spin_unlock_irq(&conf->device_lock);
6299                         md_check_recovery(mddev);
6300                         spin_lock_irq(&conf->device_lock);
6301                 }
6302         }
6303         pr_debug("%d stripes handled\n", handled);
6304
6305         spin_unlock_irq(&conf->device_lock);
6306         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6307             mutex_trylock(&conf->cache_size_mutex)) {
6308                 grow_one_stripe(conf, __GFP_NOWARN);
6309                 /* Set flag even if allocation failed.  This helps
6310                  * slow down allocation requests when mem is short
6311                  */
6312                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6313                 mutex_unlock(&conf->cache_size_mutex);
6314         }
6315
6316         flush_deferred_bios(conf);
6317
6318         r5l_flush_stripe_to_raid(conf->log);
6319
6320         async_tx_issue_pending_all();
6321         blk_finish_plug(&plug);
6322
6323         pr_debug("--- raid5d inactive\n");
6324 }
6325
6326 static ssize_t
6327 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6328 {
6329         struct r5conf *conf;
6330         int ret = 0;
6331         spin_lock(&mddev->lock);
6332         conf = mddev->private;
6333         if (conf)
6334                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6335         spin_unlock(&mddev->lock);
6336         return ret;
6337 }
6338
6339 int
6340 raid5_set_cache_size(struct mddev *mddev, int size)
6341 {
6342         struct r5conf *conf = mddev->private;
6343
6344         if (size <= 16 || size > 32768)
6345                 return -EINVAL;
6346
6347         conf->min_nr_stripes = size;
6348         mutex_lock(&conf->cache_size_mutex);
6349         while (size < conf->max_nr_stripes &&
6350                drop_one_stripe(conf))
6351                 ;
6352         mutex_unlock(&conf->cache_size_mutex);
6353
6354         md_allow_write(mddev);
6355
6356         mutex_lock(&conf->cache_size_mutex);
6357         while (size > conf->max_nr_stripes)
6358                 if (!grow_one_stripe(conf, GFP_KERNEL))
6359                         break;
6360         mutex_unlock(&conf->cache_size_mutex);
6361
6362         return 0;
6363 }
6364 EXPORT_SYMBOL(raid5_set_cache_size);
6365
6366 static ssize_t
6367 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6368 {
6369         struct r5conf *conf;
6370         unsigned long new;
6371         int err;
6372
6373         if (len >= PAGE_SIZE)
6374                 return -EINVAL;
6375         if (kstrtoul(page, 10, &new))
6376                 return -EINVAL;
6377         err = mddev_lock(mddev);
6378         if (err)
6379                 return err;
6380         conf = mddev->private;
6381         if (!conf)
6382                 err = -ENODEV;
6383         else
6384                 err = raid5_set_cache_size(mddev, new);
6385         mddev_unlock(mddev);
6386
6387         return err ?: len;
6388 }
6389
6390 static struct md_sysfs_entry
6391 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6392                                 raid5_show_stripe_cache_size,
6393                                 raid5_store_stripe_cache_size);
6394
6395 static ssize_t
6396 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6397 {
6398         struct r5conf *conf = mddev->private;
6399         if (conf)
6400                 return sprintf(page, "%d\n", conf->rmw_level);
6401         else
6402                 return 0;
6403 }
6404
6405 static ssize_t
6406 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6407 {
6408         struct r5conf *conf = mddev->private;
6409         unsigned long new;
6410
6411         if (!conf)
6412                 return -ENODEV;
6413
6414         if (len >= PAGE_SIZE)
6415                 return -EINVAL;
6416
6417         if (kstrtoul(page, 10, &new))
6418                 return -EINVAL;
6419
6420         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6421                 return -EINVAL;
6422
6423         if (new != PARITY_DISABLE_RMW &&
6424             new != PARITY_ENABLE_RMW &&
6425             new != PARITY_PREFER_RMW)
6426                 return -EINVAL;
6427
6428         conf->rmw_level = new;
6429         return len;
6430 }
6431
6432 static struct md_sysfs_entry
6433 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6434                          raid5_show_rmw_level,
6435                          raid5_store_rmw_level);
6436
6437
6438 static ssize_t
6439 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6440 {
6441         struct r5conf *conf;
6442         int ret = 0;
6443         spin_lock(&mddev->lock);
6444         conf = mddev->private;
6445         if (conf)
6446                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6447         spin_unlock(&mddev->lock);
6448         return ret;
6449 }
6450
6451 static ssize_t
6452 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6453 {
6454         struct r5conf *conf;
6455         unsigned long new;
6456         int err;
6457
6458         if (len >= PAGE_SIZE)
6459                 return -EINVAL;
6460         if (kstrtoul(page, 10, &new))
6461                 return -EINVAL;
6462
6463         err = mddev_lock(mddev);
6464         if (err)
6465                 return err;
6466         conf = mddev->private;
6467         if (!conf)
6468                 err = -ENODEV;
6469         else if (new > conf->min_nr_stripes)
6470                 err = -EINVAL;
6471         else
6472                 conf->bypass_threshold = new;
6473         mddev_unlock(mddev);
6474         return err ?: len;
6475 }
6476
6477 static struct md_sysfs_entry
6478 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6479                                         S_IRUGO | S_IWUSR,
6480                                         raid5_show_preread_threshold,
6481                                         raid5_store_preread_threshold);
6482
6483 static ssize_t
6484 raid5_show_skip_copy(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->skip_copy);
6492         spin_unlock(&mddev->lock);
6493         return ret;
6494 }
6495
6496 static ssize_t
6497 raid5_store_skip_copy(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         new = !!new;
6508
6509         err = mddev_lock(mddev);
6510         if (err)
6511                 return err;
6512         conf = mddev->private;
6513         if (!conf)
6514                 err = -ENODEV;
6515         else if (new != conf->skip_copy) {
6516                 mddev_suspend(mddev);
6517                 conf->skip_copy = new;
6518                 if (new)
6519                         mddev->queue->backing_dev_info->capabilities |=
6520                                 BDI_CAP_STABLE_WRITES;
6521                 else
6522                         mddev->queue->backing_dev_info->capabilities &=
6523                                 ~BDI_CAP_STABLE_WRITES;
6524                 mddev_resume(mddev);
6525         }
6526         mddev_unlock(mddev);
6527         return err ?: len;
6528 }
6529
6530 static struct md_sysfs_entry
6531 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6532                                         raid5_show_skip_copy,
6533                                         raid5_store_skip_copy);
6534
6535 static ssize_t
6536 stripe_cache_active_show(struct mddev *mddev, char *page)
6537 {
6538         struct r5conf *conf = mddev->private;
6539         if (conf)
6540                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6541         else
6542                 return 0;
6543 }
6544
6545 static struct md_sysfs_entry
6546 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6547
6548 static ssize_t
6549 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6550 {
6551         struct r5conf *conf;
6552         int ret = 0;
6553         spin_lock(&mddev->lock);
6554         conf = mddev->private;
6555         if (conf)
6556                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6557         spin_unlock(&mddev->lock);
6558         return ret;
6559 }
6560
6561 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6562                                int *group_cnt,
6563                                int *worker_cnt_per_group,
6564                                struct r5worker_group **worker_groups);
6565 static ssize_t
6566 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6567 {
6568         struct r5conf *conf;
6569         unsigned long new;
6570         int err;
6571         struct r5worker_group *new_groups, *old_groups;
6572         int group_cnt, worker_cnt_per_group;
6573
6574         if (len >= PAGE_SIZE)
6575                 return -EINVAL;
6576         if (kstrtoul(page, 10, &new))
6577                 return -EINVAL;
6578
6579         err = mddev_lock(mddev);
6580         if (err)
6581                 return err;
6582         conf = mddev->private;
6583         if (!conf)
6584                 err = -ENODEV;
6585         else if (new != conf->worker_cnt_per_group) {
6586                 mddev_suspend(mddev);
6587
6588                 old_groups = conf->worker_groups;
6589                 if (old_groups)
6590                         flush_workqueue(raid5_wq);
6591
6592                 err = alloc_thread_groups(conf, new,
6593                                           &group_cnt, &worker_cnt_per_group,
6594                                           &new_groups);
6595                 if (!err) {
6596                         spin_lock_irq(&conf->device_lock);
6597                         conf->group_cnt = group_cnt;
6598                         conf->worker_cnt_per_group = worker_cnt_per_group;
6599                         conf->worker_groups = new_groups;
6600                         spin_unlock_irq(&conf->device_lock);
6601
6602                         if (old_groups)
6603                                 kfree(old_groups[0].workers);
6604                         kfree(old_groups);
6605                 }
6606                 mddev_resume(mddev);
6607         }
6608         mddev_unlock(mddev);
6609
6610         return err ?: len;
6611 }
6612
6613 static struct md_sysfs_entry
6614 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6615                                 raid5_show_group_thread_cnt,
6616                                 raid5_store_group_thread_cnt);
6617
6618 static struct attribute *raid5_attrs[] =  {
6619         &raid5_stripecache_size.attr,
6620         &raid5_stripecache_active.attr,
6621         &raid5_preread_bypass_threshold.attr,
6622         &raid5_group_thread_cnt.attr,
6623         &raid5_skip_copy.attr,
6624         &raid5_rmw_level.attr,
6625         &r5c_journal_mode.attr,
6626         NULL,
6627 };
6628 static struct attribute_group raid5_attrs_group = {
6629         .name = NULL,
6630         .attrs = raid5_attrs,
6631 };
6632
6633 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6634                                int *group_cnt,
6635                                int *worker_cnt_per_group,
6636                                struct r5worker_group **worker_groups)
6637 {
6638         int i, j, k;
6639         ssize_t size;
6640         struct r5worker *workers;
6641
6642         *worker_cnt_per_group = cnt;
6643         if (cnt == 0) {
6644                 *group_cnt = 0;
6645                 *worker_groups = NULL;
6646                 return 0;
6647         }
6648         *group_cnt = num_possible_nodes();
6649         size = sizeof(struct r5worker) * cnt;
6650         workers = kzalloc(size * *group_cnt, GFP_NOIO);
6651         *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6652                                 *group_cnt, GFP_NOIO);
6653         if (!*worker_groups || !workers) {
6654                 kfree(workers);
6655                 kfree(*worker_groups);
6656                 return -ENOMEM;
6657         }
6658
6659         for (i = 0; i < *group_cnt; i++) {
6660                 struct r5worker_group *group;
6661
6662                 group = &(*worker_groups)[i];
6663                 INIT_LIST_HEAD(&group->handle_list);
6664                 INIT_LIST_HEAD(&group->loprio_list);
6665                 group->conf = conf;
6666                 group->workers = workers + i * cnt;
6667
6668                 for (j = 0; j < cnt; j++) {
6669                         struct r5worker *worker = group->workers + j;
6670                         worker->group = group;
6671                         INIT_WORK(&worker->work, raid5_do_work);
6672
6673                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6674                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6675                 }
6676         }
6677
6678         return 0;
6679 }
6680
6681 static void free_thread_groups(struct r5conf *conf)
6682 {
6683         if (conf->worker_groups)
6684                 kfree(conf->worker_groups[0].workers);
6685         kfree(conf->worker_groups);
6686         conf->worker_groups = NULL;
6687 }
6688
6689 static sector_t
6690 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6691 {
6692         struct r5conf *conf = mddev->private;
6693
6694         if (!sectors)
6695                 sectors = mddev->dev_sectors;
6696         if (!raid_disks)
6697                 /* size is defined by the smallest of previous and new size */
6698                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6699
6700         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6701         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6702         return sectors * (raid_disks - conf->max_degraded);
6703 }
6704
6705 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6706 {
6707         safe_put_page(percpu->spare_page);
6708         if (percpu->scribble)
6709                 flex_array_free(percpu->scribble);
6710         percpu->spare_page = NULL;
6711         percpu->scribble = NULL;
6712 }
6713
6714 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6715 {
6716         if (conf->level == 6 && !percpu->spare_page)
6717                 percpu->spare_page = alloc_page(GFP_KERNEL);
6718         if (!percpu->scribble)
6719                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6720                                                       conf->previous_raid_disks),
6721                                                   max(conf->chunk_sectors,
6722                                                       conf->prev_chunk_sectors)
6723                                                    / STRIPE_SECTORS,
6724                                                   GFP_KERNEL);
6725
6726         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6727                 free_scratch_buffer(conf, percpu);
6728                 return -ENOMEM;
6729         }
6730
6731         return 0;
6732 }
6733
6734 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6735 {
6736         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6737
6738         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6739         return 0;
6740 }
6741
6742 static void raid5_free_percpu(struct r5conf *conf)
6743 {
6744         if (!conf->percpu)
6745                 return;
6746
6747         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6748         free_percpu(conf->percpu);
6749 }
6750
6751 static void free_conf(struct r5conf *conf)
6752 {
6753         int i;
6754
6755         log_exit(conf);
6756
6757         if (conf->shrinker.nr_deferred)
6758                 unregister_shrinker(&conf->shrinker);
6759
6760         free_thread_groups(conf);
6761         shrink_stripes(conf);
6762         raid5_free_percpu(conf);
6763         for (i = 0; i < conf->pool_size; i++)
6764                 if (conf->disks[i].extra_page)
6765                         put_page(conf->disks[i].extra_page);
6766         kfree(conf->disks);
6767         if (conf->bio_split)
6768                 bioset_free(conf->bio_split);
6769         kfree(conf->stripe_hashtbl);
6770         kfree(conf->pending_data);
6771         kfree(conf);
6772 }
6773
6774 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6775 {
6776         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6777         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6778
6779         if (alloc_scratch_buffer(conf, percpu)) {
6780                 pr_warn("%s: failed memory allocation for cpu%u\n",
6781                         __func__, cpu);
6782                 return -ENOMEM;
6783         }
6784         return 0;
6785 }
6786
6787 static int raid5_alloc_percpu(struct r5conf *conf)
6788 {
6789         int err = 0;
6790
6791         conf->percpu = alloc_percpu(struct raid5_percpu);
6792         if (!conf->percpu)
6793                 return -ENOMEM;
6794
6795         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6796         if (!err) {
6797                 conf->scribble_disks = max(conf->raid_disks,
6798                         conf->previous_raid_disks);
6799                 conf->scribble_sectors = max(conf->chunk_sectors,
6800                         conf->prev_chunk_sectors);
6801         }
6802         return err;
6803 }
6804
6805 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6806                                       struct shrink_control *sc)
6807 {
6808         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6809         unsigned long ret = SHRINK_STOP;
6810
6811         if (mutex_trylock(&conf->cache_size_mutex)) {
6812                 ret= 0;
6813                 while (ret < sc->nr_to_scan &&
6814                        conf->max_nr_stripes > conf->min_nr_stripes) {
6815                         if (drop_one_stripe(conf) == 0) {
6816                                 ret = SHRINK_STOP;
6817                                 break;
6818                         }
6819                         ret++;
6820                 }
6821                 mutex_unlock(&conf->cache_size_mutex);
6822         }
6823         return ret;
6824 }
6825
6826 static unsigned long raid5_cache_count(struct shrinker *shrink,
6827                                        struct shrink_control *sc)
6828 {
6829         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6830
6831         if (conf->max_nr_stripes < conf->min_nr_stripes)
6832                 /* unlikely, but not impossible */
6833                 return 0;
6834         return conf->max_nr_stripes - conf->min_nr_stripes;
6835 }
6836
6837 static struct r5conf *setup_conf(struct mddev *mddev)
6838 {
6839         struct r5conf *conf;
6840         int raid_disk, memory, max_disks;
6841         struct md_rdev *rdev;
6842         struct disk_info *disk;
6843         char pers_name[6];
6844         int i;
6845         int group_cnt, worker_cnt_per_group;
6846         struct r5worker_group *new_group;
6847
6848         if (mddev->new_level != 5
6849             && mddev->new_level != 4
6850             && mddev->new_level != 6) {
6851                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6852                         mdname(mddev), mddev->new_level);
6853                 return ERR_PTR(-EIO);
6854         }
6855         if ((mddev->new_level == 5
6856              && !algorithm_valid_raid5(mddev->new_layout)) ||
6857             (mddev->new_level == 6
6858              && !algorithm_valid_raid6(mddev->new_layout))) {
6859                 pr_warn("md/raid:%s: layout %d not supported\n",
6860                         mdname(mddev), mddev->new_layout);
6861                 return ERR_PTR(-EIO);
6862         }
6863         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6864                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6865                         mdname(mddev), mddev->raid_disks);
6866                 return ERR_PTR(-EINVAL);
6867         }
6868
6869         if (!mddev->new_chunk_sectors ||
6870             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6871             !is_power_of_2(mddev->new_chunk_sectors)) {
6872                 pr_warn("md/raid:%s: invalid chunk size %d\n",
6873                         mdname(mddev), mddev->new_chunk_sectors << 9);
6874                 return ERR_PTR(-EINVAL);
6875         }
6876
6877         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6878         if (conf == NULL)
6879                 goto abort;
6880         INIT_LIST_HEAD(&conf->free_list);
6881         INIT_LIST_HEAD(&conf->pending_list);
6882         conf->pending_data = kzalloc(sizeof(struct r5pending_data) *
6883                 PENDING_IO_MAX, GFP_KERNEL);
6884         if (!conf->pending_data)
6885                 goto abort;
6886         for (i = 0; i < PENDING_IO_MAX; i++)
6887                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6888         /* Don't enable multi-threading by default*/
6889         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6890                                  &new_group)) {
6891                 conf->group_cnt = group_cnt;
6892                 conf->worker_cnt_per_group = worker_cnt_per_group;
6893                 conf->worker_groups = new_group;
6894         } else
6895                 goto abort;
6896         spin_lock_init(&conf->device_lock);
6897         seqcount_init(&conf->gen_lock);
6898         mutex_init(&conf->cache_size_mutex);
6899         init_waitqueue_head(&conf->wait_for_quiescent);
6900         init_waitqueue_head(&conf->wait_for_stripe);
6901         init_waitqueue_head(&conf->wait_for_overlap);
6902         INIT_LIST_HEAD(&conf->handle_list);
6903         INIT_LIST_HEAD(&conf->loprio_list);
6904         INIT_LIST_HEAD(&conf->hold_list);
6905         INIT_LIST_HEAD(&conf->delayed_list);
6906         INIT_LIST_HEAD(&conf->bitmap_list);
6907         init_llist_head(&conf->released_stripes);
6908         atomic_set(&conf->active_stripes, 0);
6909         atomic_set(&conf->preread_active_stripes, 0);
6910         atomic_set(&conf->active_aligned_reads, 0);
6911         spin_lock_init(&conf->pending_bios_lock);
6912         conf->batch_bio_dispatch = true;
6913         rdev_for_each(rdev, mddev) {
6914                 if (test_bit(Journal, &rdev->flags))
6915                         continue;
6916                 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6917                         conf->batch_bio_dispatch = false;
6918                         break;
6919                 }
6920         }
6921
6922         conf->bypass_threshold = BYPASS_THRESHOLD;
6923         conf->recovery_disabled = mddev->recovery_disabled - 1;
6924
6925         conf->raid_disks = mddev->raid_disks;
6926         if (mddev->reshape_position == MaxSector)
6927                 conf->previous_raid_disks = mddev->raid_disks;
6928         else
6929                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6930         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6931
6932         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6933                               GFP_KERNEL);
6934
6935         if (!conf->disks)
6936                 goto abort;
6937
6938         for (i = 0; i < max_disks; i++) {
6939                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6940                 if (!conf->disks[i].extra_page)
6941                         goto abort;
6942         }
6943
6944         conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
6945         if (!conf->bio_split)
6946                 goto abort;
6947         conf->mddev = mddev;
6948
6949         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6950                 goto abort;
6951
6952         /* We init hash_locks[0] separately to that it can be used
6953          * as the reference lock in the spin_lock_nest_lock() call
6954          * in lock_all_device_hash_locks_irq in order to convince
6955          * lockdep that we know what we are doing.
6956          */
6957         spin_lock_init(conf->hash_locks);
6958         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6959                 spin_lock_init(conf->hash_locks + i);
6960
6961         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6962                 INIT_LIST_HEAD(conf->inactive_list + i);
6963
6964         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6965                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6966
6967         atomic_set(&conf->r5c_cached_full_stripes, 0);
6968         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
6969         atomic_set(&conf->r5c_cached_partial_stripes, 0);
6970         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
6971         atomic_set(&conf->r5c_flushing_full_stripes, 0);
6972         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
6973
6974         conf->level = mddev->new_level;
6975         conf->chunk_sectors = mddev->new_chunk_sectors;
6976         if (raid5_alloc_percpu(conf) != 0)
6977                 goto abort;
6978
6979         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6980
6981         rdev_for_each(rdev, mddev) {
6982                 raid_disk = rdev->raid_disk;
6983                 if (raid_disk >= max_disks
6984                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6985                         continue;
6986                 disk = conf->disks + raid_disk;
6987
6988                 if (test_bit(Replacement, &rdev->flags)) {
6989                         if (disk->replacement)
6990                                 goto abort;
6991                         disk->replacement = rdev;
6992                 } else {
6993                         if (disk->rdev)
6994                                 goto abort;
6995                         disk->rdev = rdev;
6996                 }
6997
6998                 if (test_bit(In_sync, &rdev->flags)) {
6999                         char b[BDEVNAME_SIZE];
7000                         pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7001                                 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7002                 } else if (rdev->saved_raid_disk != raid_disk)
7003                         /* Cannot rely on bitmap to complete recovery */
7004                         conf->fullsync = 1;
7005         }
7006
7007         conf->level = mddev->new_level;
7008         if (conf->level == 6) {
7009                 conf->max_degraded = 2;
7010                 if (raid6_call.xor_syndrome)
7011                         conf->rmw_level = PARITY_ENABLE_RMW;
7012                 else
7013                         conf->rmw_level = PARITY_DISABLE_RMW;
7014         } else {
7015                 conf->max_degraded = 1;
7016                 conf->rmw_level = PARITY_ENABLE_RMW;
7017         }
7018         conf->algorithm = mddev->new_layout;
7019         conf->reshape_progress = mddev->reshape_position;
7020         if (conf->reshape_progress != MaxSector) {
7021                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7022                 conf->prev_algo = mddev->layout;
7023         } else {
7024                 conf->prev_chunk_sectors = conf->chunk_sectors;
7025                 conf->prev_algo = conf->algorithm;
7026         }
7027
7028         conf->min_nr_stripes = NR_STRIPES;
7029         if (mddev->reshape_position != MaxSector) {
7030                 int stripes = max_t(int,
7031                         ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7032                         ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7033                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7034                 if (conf->min_nr_stripes != NR_STRIPES)
7035                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7036                                 mdname(mddev), conf->min_nr_stripes);
7037         }
7038         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7039                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7040         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7041         if (grow_stripes(conf, conf->min_nr_stripes)) {
7042                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7043                         mdname(mddev), memory);
7044                 goto abort;
7045         } else
7046                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7047         /*
7048          * Losing a stripe head costs more than the time to refill it,
7049          * it reduces the queue depth and so can hurt throughput.
7050          * So set it rather large, scaled by number of devices.
7051          */
7052         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7053         conf->shrinker.scan_objects = raid5_cache_scan;
7054         conf->shrinker.count_objects = raid5_cache_count;
7055         conf->shrinker.batch = 128;
7056         conf->shrinker.flags = 0;
7057         if (register_shrinker(&conf->shrinker)) {
7058                 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7059                         mdname(mddev));
7060                 goto abort;
7061         }
7062
7063         sprintf(pers_name, "raid%d", mddev->new_level);
7064         conf->thread = md_register_thread(raid5d, mddev, pers_name);
7065         if (!conf->thread) {
7066                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7067                         mdname(mddev));
7068                 goto abort;
7069         }
7070
7071         return conf;
7072
7073  abort:
7074         if (conf) {
7075                 free_conf(conf);
7076                 return ERR_PTR(-EIO);
7077         } else
7078                 return ERR_PTR(-ENOMEM);
7079 }
7080
7081 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7082 {
7083         switch (algo) {
7084         case ALGORITHM_PARITY_0:
7085                 if (raid_disk < max_degraded)
7086                         return 1;
7087                 break;
7088         case ALGORITHM_PARITY_N:
7089                 if (raid_disk >= raid_disks - max_degraded)
7090                         return 1;
7091                 break;
7092         case ALGORITHM_PARITY_0_6:
7093                 if (raid_disk == 0 ||
7094                     raid_disk == raid_disks - 1)
7095                         return 1;
7096                 break;
7097         case ALGORITHM_LEFT_ASYMMETRIC_6:
7098         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7099         case ALGORITHM_LEFT_SYMMETRIC_6:
7100         case ALGORITHM_RIGHT_SYMMETRIC_6:
7101                 if (raid_disk == raid_disks - 1)
7102                         return 1;
7103         }
7104         return 0;
7105 }
7106
7107 static int raid5_run(struct mddev *mddev)
7108 {
7109         struct r5conf *conf;
7110         int working_disks = 0;
7111         int dirty_parity_disks = 0;
7112         struct md_rdev *rdev;
7113         struct md_rdev *journal_dev = NULL;
7114         sector_t reshape_offset = 0;
7115         int i;
7116         long long min_offset_diff = 0;
7117         int first = 1;
7118
7119         if (mddev_init_writes_pending(mddev) < 0)
7120                 return -ENOMEM;
7121
7122         if (mddev->recovery_cp != MaxSector)
7123                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7124                           mdname(mddev));
7125
7126         rdev_for_each(rdev, mddev) {
7127                 long long diff;
7128
7129                 if (test_bit(Journal, &rdev->flags)) {
7130                         journal_dev = rdev;
7131                         continue;
7132                 }
7133                 if (rdev->raid_disk < 0)
7134                         continue;
7135                 diff = (rdev->new_data_offset - rdev->data_offset);
7136                 if (first) {
7137                         min_offset_diff = diff;
7138                         first = 0;
7139                 } else if (mddev->reshape_backwards &&
7140                          diff < min_offset_diff)
7141                         min_offset_diff = diff;
7142                 else if (!mddev->reshape_backwards &&
7143                          diff > min_offset_diff)
7144                         min_offset_diff = diff;
7145         }
7146
7147         if (mddev->reshape_position != MaxSector) {
7148                 /* Check that we can continue the reshape.
7149                  * Difficulties arise if the stripe we would write to
7150                  * next is at or after the stripe we would read from next.
7151                  * For a reshape that changes the number of devices, this
7152                  * is only possible for a very short time, and mdadm makes
7153                  * sure that time appears to have past before assembling
7154                  * the array.  So we fail if that time hasn't passed.
7155                  * For a reshape that keeps the number of devices the same
7156                  * mdadm must be monitoring the reshape can keeping the
7157                  * critical areas read-only and backed up.  It will start
7158                  * the array in read-only mode, so we check for that.
7159                  */
7160                 sector_t here_new, here_old;
7161                 int old_disks;
7162                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7163                 int chunk_sectors;
7164                 int new_data_disks;
7165
7166                 if (journal_dev) {
7167                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7168                                 mdname(mddev));
7169                         return -EINVAL;
7170                 }
7171
7172                 if (mddev->new_level != mddev->level) {
7173                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7174                                 mdname(mddev));
7175                         return -EINVAL;
7176                 }
7177                 old_disks = mddev->raid_disks - mddev->delta_disks;
7178                 /* reshape_position must be on a new-stripe boundary, and one
7179                  * further up in new geometry must map after here in old
7180                  * geometry.
7181                  * If the chunk sizes are different, then as we perform reshape
7182                  * in units of the largest of the two, reshape_position needs
7183                  * be a multiple of the largest chunk size times new data disks.
7184                  */
7185                 here_new = mddev->reshape_position;
7186                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7187                 new_data_disks = mddev->raid_disks - max_degraded;
7188                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7189                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7190                                 mdname(mddev));
7191                         return -EINVAL;
7192                 }
7193                 reshape_offset = here_new * chunk_sectors;
7194                 /* here_new is the stripe we will write to */
7195                 here_old = mddev->reshape_position;
7196                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7197                 /* here_old is the first stripe that we might need to read
7198                  * from */
7199                 if (mddev->delta_disks == 0) {
7200                         /* We cannot be sure it is safe to start an in-place
7201                          * reshape.  It is only safe if user-space is monitoring
7202                          * and taking constant backups.
7203                          * mdadm always starts a situation like this in
7204                          * readonly mode so it can take control before
7205                          * allowing any writes.  So just check for that.
7206                          */
7207                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7208                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7209                                 /* not really in-place - so OK */;
7210                         else if (mddev->ro == 0) {
7211                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7212                                         mdname(mddev));
7213                                 return -EINVAL;
7214                         }
7215                 } else if (mddev->reshape_backwards
7216                     ? (here_new * chunk_sectors + min_offset_diff <=
7217                        here_old * chunk_sectors)
7218                     : (here_new * chunk_sectors >=
7219                        here_old * chunk_sectors + (-min_offset_diff))) {
7220                         /* Reading from the same stripe as writing to - bad */
7221                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7222                                 mdname(mddev));
7223                         return -EINVAL;
7224                 }
7225                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7226                 /* OK, we should be able to continue; */
7227         } else {
7228                 BUG_ON(mddev->level != mddev->new_level);
7229                 BUG_ON(mddev->layout != mddev->new_layout);
7230                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7231                 BUG_ON(mddev->delta_disks != 0);
7232         }
7233
7234         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7235             test_bit(MD_HAS_PPL, &mddev->flags)) {
7236                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7237                         mdname(mddev));
7238                 clear_bit(MD_HAS_PPL, &mddev->flags);
7239                 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7240         }
7241
7242         if (mddev->private == NULL)
7243                 conf = setup_conf(mddev);
7244         else
7245                 conf = mddev->private;
7246
7247         if (IS_ERR(conf))
7248                 return PTR_ERR(conf);
7249
7250         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7251                 if (!journal_dev) {
7252                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7253                                 mdname(mddev));
7254                         mddev->ro = 1;
7255                         set_disk_ro(mddev->gendisk, 1);
7256                 } else if (mddev->recovery_cp == MaxSector)
7257                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7258         }
7259
7260         conf->min_offset_diff = min_offset_diff;
7261         mddev->thread = conf->thread;
7262         conf->thread = NULL;
7263         mddev->private = conf;
7264
7265         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7266              i++) {
7267                 rdev = conf->disks[i].rdev;
7268                 if (!rdev && conf->disks[i].replacement) {
7269                         /* The replacement is all we have yet */
7270                         rdev = conf->disks[i].replacement;
7271                         conf->disks[i].replacement = NULL;
7272                         clear_bit(Replacement, &rdev->flags);
7273                         conf->disks[i].rdev = rdev;
7274                 }
7275                 if (!rdev)
7276                         continue;
7277                 if (conf->disks[i].replacement &&
7278                     conf->reshape_progress != MaxSector) {
7279                         /* replacements and reshape simply do not mix. */
7280                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7281                         goto abort;
7282                 }
7283                 if (test_bit(In_sync, &rdev->flags)) {
7284                         working_disks++;
7285                         continue;
7286                 }
7287                 /* This disc is not fully in-sync.  However if it
7288                  * just stored parity (beyond the recovery_offset),
7289                  * when we don't need to be concerned about the
7290                  * array being dirty.
7291                  * When reshape goes 'backwards', we never have
7292                  * partially completed devices, so we only need
7293                  * to worry about reshape going forwards.
7294                  */
7295                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7296                 if (mddev->major_version == 0 &&
7297                     mddev->minor_version > 90)
7298                         rdev->recovery_offset = reshape_offset;
7299
7300                 if (rdev->recovery_offset < reshape_offset) {
7301                         /* We need to check old and new layout */
7302                         if (!only_parity(rdev->raid_disk,
7303                                          conf->algorithm,
7304                                          conf->raid_disks,
7305                                          conf->max_degraded))
7306                                 continue;
7307                 }
7308                 if (!only_parity(rdev->raid_disk,
7309                                  conf->prev_algo,
7310                                  conf->previous_raid_disks,
7311                                  conf->max_degraded))
7312                         continue;
7313                 dirty_parity_disks++;
7314         }
7315
7316         /*
7317          * 0 for a fully functional array, 1 or 2 for a degraded array.
7318          */
7319         mddev->degraded = raid5_calc_degraded(conf);
7320
7321         if (has_failed(conf)) {
7322                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7323                         mdname(mddev), mddev->degraded, conf->raid_disks);
7324                 goto abort;
7325         }
7326
7327         /* device size must be a multiple of chunk size */
7328         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7329         mddev->resync_max_sectors = mddev->dev_sectors;
7330
7331         if (mddev->degraded > dirty_parity_disks &&
7332             mddev->recovery_cp != MaxSector) {
7333                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7334                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7335                                 mdname(mddev));
7336                 else if (mddev->ok_start_degraded)
7337                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7338                                 mdname(mddev));
7339                 else {
7340                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7341                                 mdname(mddev));
7342                         goto abort;
7343                 }
7344         }
7345
7346         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7347                 mdname(mddev), conf->level,
7348                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7349                 mddev->new_layout);
7350
7351         print_raid5_conf(conf);
7352
7353         if (conf->reshape_progress != MaxSector) {
7354                 conf->reshape_safe = conf->reshape_progress;
7355                 atomic_set(&conf->reshape_stripes, 0);
7356                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7357                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7358                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7359                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7360                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7361                                                         "reshape");
7362         }
7363
7364         /* Ok, everything is just fine now */
7365         if (mddev->to_remove == &raid5_attrs_group)
7366                 mddev->to_remove = NULL;
7367         else if (mddev->kobj.sd &&
7368             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7369                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7370                         mdname(mddev));
7371         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7372
7373         if (mddev->queue) {
7374                 int chunk_size;
7375                 /* read-ahead size must cover two whole stripes, which
7376                  * is 2 * (datadisks) * chunksize where 'n' is the
7377                  * number of raid devices
7378                  */
7379                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7380                 int stripe = data_disks *
7381                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7382                 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7383                         mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7384
7385                 chunk_size = mddev->chunk_sectors << 9;
7386                 blk_queue_io_min(mddev->queue, chunk_size);
7387                 blk_queue_io_opt(mddev->queue, chunk_size *
7388                                  (conf->raid_disks - conf->max_degraded));
7389                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7390                 /*
7391                  * We can only discard a whole stripe. It doesn't make sense to
7392                  * discard data disk but write parity disk
7393                  */
7394                 stripe = stripe * PAGE_SIZE;
7395                 /* Round up to power of 2, as discard handling
7396                  * currently assumes that */
7397                 while ((stripe-1) & stripe)
7398                         stripe = (stripe | (stripe-1)) + 1;
7399                 mddev->queue->limits.discard_alignment = stripe;
7400                 mddev->queue->limits.discard_granularity = stripe;
7401
7402                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7403                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7404
7405                 rdev_for_each(rdev, mddev) {
7406                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7407                                           rdev->data_offset << 9);
7408                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7409                                           rdev->new_data_offset << 9);
7410                 }
7411
7412                 /*
7413                  * zeroing is required, otherwise data
7414                  * could be lost. Consider a scenario: discard a stripe
7415                  * (the stripe could be inconsistent if
7416                  * discard_zeroes_data is 0); write one disk of the
7417                  * stripe (the stripe could be inconsistent again
7418                  * depending on which disks are used to calculate
7419                  * parity); the disk is broken; The stripe data of this
7420                  * disk is lost.
7421                  *
7422                  * We only allow DISCARD if the sysadmin has confirmed that
7423                  * only safe devices are in use by setting a module parameter.
7424                  * A better idea might be to turn DISCARD into WRITE_ZEROES
7425                  * requests, as that is required to be safe.
7426                  */
7427                 if (devices_handle_discard_safely &&
7428                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7429                     mddev->queue->limits.discard_granularity >= stripe)
7430                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7431                                                 mddev->queue);
7432                 else
7433                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7434                                                 mddev->queue);
7435
7436                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7437         }
7438
7439         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7440                 goto abort;
7441
7442         return 0;
7443 abort:
7444         md_unregister_thread(&mddev->thread);
7445         print_raid5_conf(conf);
7446         free_conf(conf);
7447         mddev->private = NULL;
7448         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7449         return -EIO;
7450 }
7451
7452 static void raid5_free(struct mddev *mddev, void *priv)
7453 {
7454         struct r5conf *conf = priv;
7455
7456         free_conf(conf);
7457         mddev->to_remove = &raid5_attrs_group;
7458 }
7459
7460 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7461 {
7462         struct r5conf *conf = mddev->private;
7463         int i;
7464
7465         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7466                 conf->chunk_sectors / 2, mddev->layout);
7467         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7468         rcu_read_lock();
7469         for (i = 0; i < conf->raid_disks; i++) {
7470                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7471                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7472         }
7473         rcu_read_unlock();
7474         seq_printf (seq, "]");
7475 }
7476
7477 static void print_raid5_conf (struct r5conf *conf)
7478 {
7479         int i;
7480         struct disk_info *tmp;
7481
7482         pr_debug("RAID conf printout:\n");
7483         if (!conf) {
7484                 pr_debug("(conf==NULL)\n");
7485                 return;
7486         }
7487         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7488                conf->raid_disks,
7489                conf->raid_disks - conf->mddev->degraded);
7490
7491         for (i = 0; i < conf->raid_disks; i++) {
7492                 char b[BDEVNAME_SIZE];
7493                 tmp = conf->disks + i;
7494                 if (tmp->rdev)
7495                         pr_debug(" disk %d, o:%d, dev:%s\n",
7496                                i, !test_bit(Faulty, &tmp->rdev->flags),
7497                                bdevname(tmp->rdev->bdev, b));
7498         }
7499 }
7500
7501 static int raid5_spare_active(struct mddev *mddev)
7502 {
7503         int i;
7504         struct r5conf *conf = mddev->private;
7505         struct disk_info *tmp;
7506         int count = 0;
7507         unsigned long flags;
7508
7509         for (i = 0; i < conf->raid_disks; i++) {
7510                 tmp = conf->disks + i;
7511                 if (tmp->replacement
7512                     && tmp->replacement->recovery_offset == MaxSector
7513                     && !test_bit(Faulty, &tmp->replacement->flags)
7514                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7515                         /* Replacement has just become active. */
7516                         if (!tmp->rdev
7517                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7518                                 count++;
7519                         if (tmp->rdev) {
7520                                 /* Replaced device not technically faulty,
7521                                  * but we need to be sure it gets removed
7522                                  * and never re-added.
7523                                  */
7524                                 set_bit(Faulty, &tmp->rdev->flags);
7525                                 sysfs_notify_dirent_safe(
7526                                         tmp->rdev->sysfs_state);
7527                         }
7528                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7529                 } else if (tmp->rdev
7530                     && tmp->rdev->recovery_offset == MaxSector
7531                     && !test_bit(Faulty, &tmp->rdev->flags)
7532                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7533                         count++;
7534                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7535                 }
7536         }
7537         spin_lock_irqsave(&conf->device_lock, flags);
7538         mddev->degraded = raid5_calc_degraded(conf);
7539         spin_unlock_irqrestore(&conf->device_lock, flags);
7540         print_raid5_conf(conf);
7541         return count;
7542 }
7543
7544 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7545 {
7546         struct r5conf *conf = mddev->private;
7547         int err = 0;
7548         int number = rdev->raid_disk;
7549         struct md_rdev **rdevp;
7550         struct disk_info *p = conf->disks + number;
7551
7552         print_raid5_conf(conf);
7553         if (test_bit(Journal, &rdev->flags) && conf->log) {
7554                 /*
7555                  * we can't wait pending write here, as this is called in
7556                  * raid5d, wait will deadlock.
7557                  * neilb: there is no locking about new writes here,
7558                  * so this cannot be safe.
7559                  */
7560                 if (atomic_read(&conf->active_stripes) ||
7561                     atomic_read(&conf->r5c_cached_full_stripes) ||
7562                     atomic_read(&conf->r5c_cached_partial_stripes)) {
7563                         return -EBUSY;
7564                 }
7565                 log_exit(conf);
7566                 return 0;
7567         }
7568         if (rdev == p->rdev)
7569                 rdevp = &p->rdev;
7570         else if (rdev == p->replacement)
7571                 rdevp = &p->replacement;
7572         else
7573                 return 0;
7574
7575         if (number >= conf->raid_disks &&
7576             conf->reshape_progress == MaxSector)
7577                 clear_bit(In_sync, &rdev->flags);
7578
7579         if (test_bit(In_sync, &rdev->flags) ||
7580             atomic_read(&rdev->nr_pending)) {
7581                 err = -EBUSY;
7582                 goto abort;
7583         }
7584         /* Only remove non-faulty devices if recovery
7585          * isn't possible.
7586          */
7587         if (!test_bit(Faulty, &rdev->flags) &&
7588             mddev->recovery_disabled != conf->recovery_disabled &&
7589             !has_failed(conf) &&
7590             (!p->replacement || p->replacement == rdev) &&
7591             number < conf->raid_disks) {
7592                 err = -EBUSY;
7593                 goto abort;
7594         }
7595         *rdevp = NULL;
7596         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7597                 synchronize_rcu();
7598                 if (atomic_read(&rdev->nr_pending)) {
7599                         /* lost the race, try later */
7600                         err = -EBUSY;
7601                         *rdevp = rdev;
7602                 }
7603         }
7604         if (!err) {
7605                 err = log_modify(conf, rdev, false);
7606                 if (err)
7607                         goto abort;
7608         }
7609         if (p->replacement) {
7610                 /* We must have just cleared 'rdev' */
7611                 p->rdev = p->replacement;
7612                 clear_bit(Replacement, &p->replacement->flags);
7613                 smp_mb(); /* Make sure other CPUs may see both as identical
7614                            * but will never see neither - if they are careful
7615                            */
7616                 p->replacement = NULL;
7617
7618                 if (!err)
7619                         err = log_modify(conf, p->rdev, true);
7620         }
7621
7622         clear_bit(WantReplacement, &rdev->flags);
7623 abort:
7624
7625         print_raid5_conf(conf);
7626         return err;
7627 }
7628
7629 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7630 {
7631         struct r5conf *conf = mddev->private;
7632         int err = -EEXIST;
7633         int disk;
7634         struct disk_info *p;
7635         int first = 0;
7636         int last = conf->raid_disks - 1;
7637
7638         if (test_bit(Journal, &rdev->flags)) {
7639                 if (conf->log)
7640                         return -EBUSY;
7641
7642                 rdev->raid_disk = 0;
7643                 /*
7644                  * The array is in readonly mode if journal is missing, so no
7645                  * write requests running. We should be safe
7646                  */
7647                 log_init(conf, rdev, false);
7648                 return 0;
7649         }
7650         if (mddev->recovery_disabled == conf->recovery_disabled)
7651                 return -EBUSY;
7652
7653         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7654                 /* no point adding a device */
7655                 return -EINVAL;
7656
7657         if (rdev->raid_disk >= 0)
7658                 first = last = rdev->raid_disk;
7659
7660         /*
7661          * find the disk ... but prefer rdev->saved_raid_disk
7662          * if possible.
7663          */
7664         if (rdev->saved_raid_disk >= 0 &&
7665             rdev->saved_raid_disk >= first &&
7666             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7667                 first = rdev->saved_raid_disk;
7668
7669         for (disk = first; disk <= last; disk++) {
7670                 p = conf->disks + disk;
7671                 if (p->rdev == NULL) {
7672                         clear_bit(In_sync, &rdev->flags);
7673                         rdev->raid_disk = disk;
7674                         if (rdev->saved_raid_disk != disk)
7675                                 conf->fullsync = 1;
7676                         rcu_assign_pointer(p->rdev, rdev);
7677
7678                         err = log_modify(conf, rdev, true);
7679
7680                         goto out;
7681                 }
7682         }
7683         for (disk = first; disk <= last; disk++) {
7684                 p = conf->disks + disk;
7685                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7686                     p->replacement == NULL) {
7687                         clear_bit(In_sync, &rdev->flags);
7688                         set_bit(Replacement, &rdev->flags);
7689                         rdev->raid_disk = disk;
7690                         err = 0;
7691                         conf->fullsync = 1;
7692                         rcu_assign_pointer(p->replacement, rdev);
7693                         break;
7694                 }
7695         }
7696 out:
7697         print_raid5_conf(conf);
7698         return err;
7699 }
7700
7701 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7702 {
7703         /* no resync is happening, and there is enough space
7704          * on all devices, so we can resize.
7705          * We need to make sure resync covers any new space.
7706          * If the array is shrinking we should possibly wait until
7707          * any io in the removed space completes, but it hardly seems
7708          * worth it.
7709          */
7710         sector_t newsize;
7711         struct r5conf *conf = mddev->private;
7712
7713         if (conf->log || raid5_has_ppl(conf))
7714                 return -EINVAL;
7715         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7716         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7717         if (mddev->external_size &&
7718             mddev->array_sectors > newsize)
7719                 return -EINVAL;
7720         if (mddev->bitmap) {
7721                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7722                 if (ret)
7723                         return ret;
7724         }
7725         md_set_array_sectors(mddev, newsize);
7726         if (sectors > mddev->dev_sectors &&
7727             mddev->recovery_cp > mddev->dev_sectors) {
7728                 mddev->recovery_cp = mddev->dev_sectors;
7729                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7730         }
7731         mddev->dev_sectors = sectors;
7732         mddev->resync_max_sectors = sectors;
7733         return 0;
7734 }
7735
7736 static int check_stripe_cache(struct mddev *mddev)
7737 {
7738         /* Can only proceed if there are plenty of stripe_heads.
7739          * We need a minimum of one full stripe,, and for sensible progress
7740          * it is best to have about 4 times that.
7741          * If we require 4 times, then the default 256 4K stripe_heads will
7742          * allow for chunk sizes up to 256K, which is probably OK.
7743          * If the chunk size is greater, user-space should request more
7744          * stripe_heads first.
7745          */
7746         struct r5conf *conf = mddev->private;
7747         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7748             > conf->min_nr_stripes ||
7749             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7750             > conf->min_nr_stripes) {
7751                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7752                         mdname(mddev),
7753                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7754                          / STRIPE_SIZE)*4);
7755                 return 0;
7756         }
7757         return 1;
7758 }
7759
7760 static int check_reshape(struct mddev *mddev)
7761 {
7762         struct r5conf *conf = mddev->private;
7763
7764         if (conf->log || raid5_has_ppl(conf))
7765                 return -EINVAL;
7766         if (mddev->delta_disks == 0 &&
7767             mddev->new_layout == mddev->layout &&
7768             mddev->new_chunk_sectors == mddev->chunk_sectors)
7769                 return 0; /* nothing to do */
7770         if (has_failed(conf))
7771                 return -EINVAL;
7772         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7773                 /* We might be able to shrink, but the devices must
7774                  * be made bigger first.
7775                  * For raid6, 4 is the minimum size.
7776                  * Otherwise 2 is the minimum
7777                  */
7778                 int min = 2;
7779                 if (mddev->level == 6)
7780                         min = 4;
7781                 if (mddev->raid_disks + mddev->delta_disks < min)
7782                         return -EINVAL;
7783         }
7784
7785         if (!check_stripe_cache(mddev))
7786                 return -ENOSPC;
7787
7788         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7789             mddev->delta_disks > 0)
7790                 if (resize_chunks(conf,
7791                                   conf->previous_raid_disks
7792                                   + max(0, mddev->delta_disks),
7793                                   max(mddev->new_chunk_sectors,
7794                                       mddev->chunk_sectors)
7795                             ) < 0)
7796                         return -ENOMEM;
7797
7798         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7799                 return 0; /* never bother to shrink */
7800         return resize_stripes(conf, (conf->previous_raid_disks
7801                                      + mddev->delta_disks));
7802 }
7803
7804 static int raid5_start_reshape(struct mddev *mddev)
7805 {
7806         struct r5conf *conf = mddev->private;
7807         struct md_rdev *rdev;
7808         int spares = 0;
7809         unsigned long flags;
7810
7811         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7812                 return -EBUSY;
7813
7814         if (!check_stripe_cache(mddev))
7815                 return -ENOSPC;
7816
7817         if (has_failed(conf))
7818                 return -EINVAL;
7819
7820         rdev_for_each(rdev, mddev) {
7821                 if (!test_bit(In_sync, &rdev->flags)
7822                     && !test_bit(Faulty, &rdev->flags))
7823                         spares++;
7824         }
7825
7826         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7827                 /* Not enough devices even to make a degraded array
7828                  * of that size
7829                  */
7830                 return -EINVAL;
7831
7832         /* Refuse to reduce size of the array.  Any reductions in
7833          * array size must be through explicit setting of array_size
7834          * attribute.
7835          */
7836         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7837             < mddev->array_sectors) {
7838                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7839                         mdname(mddev));
7840                 return -EINVAL;
7841         }
7842
7843         atomic_set(&conf->reshape_stripes, 0);
7844         spin_lock_irq(&conf->device_lock);
7845         write_seqcount_begin(&conf->gen_lock);
7846         conf->previous_raid_disks = conf->raid_disks;
7847         conf->raid_disks += mddev->delta_disks;
7848         conf->prev_chunk_sectors = conf->chunk_sectors;
7849         conf->chunk_sectors = mddev->new_chunk_sectors;
7850         conf->prev_algo = conf->algorithm;
7851         conf->algorithm = mddev->new_layout;
7852         conf->generation++;
7853         /* Code that selects data_offset needs to see the generation update
7854          * if reshape_progress has been set - so a memory barrier needed.
7855          */
7856         smp_mb();
7857         if (mddev->reshape_backwards)
7858                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7859         else
7860                 conf->reshape_progress = 0;
7861         conf->reshape_safe = conf->reshape_progress;
7862         write_seqcount_end(&conf->gen_lock);
7863         spin_unlock_irq(&conf->device_lock);
7864
7865         /* Now make sure any requests that proceeded on the assumption
7866          * the reshape wasn't running - like Discard or Read - have
7867          * completed.
7868          */
7869         mddev_suspend(mddev);
7870         mddev_resume(mddev);
7871
7872         /* Add some new drives, as many as will fit.
7873          * We know there are enough to make the newly sized array work.
7874          * Don't add devices if we are reducing the number of
7875          * devices in the array.  This is because it is not possible
7876          * to correctly record the "partially reconstructed" state of
7877          * such devices during the reshape and confusion could result.
7878          */
7879         if (mddev->delta_disks >= 0) {
7880                 rdev_for_each(rdev, mddev)
7881                         if (rdev->raid_disk < 0 &&
7882                             !test_bit(Faulty, &rdev->flags)) {
7883                                 if (raid5_add_disk(mddev, rdev) == 0) {
7884                                         if (rdev->raid_disk
7885                                             >= conf->previous_raid_disks)
7886                                                 set_bit(In_sync, &rdev->flags);
7887                                         else
7888                                                 rdev->recovery_offset = 0;
7889
7890                                         if (sysfs_link_rdev(mddev, rdev))
7891                                                 /* Failure here is OK */;
7892                                 }
7893                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7894                                    && !test_bit(Faulty, &rdev->flags)) {
7895                                 /* This is a spare that was manually added */
7896                                 set_bit(In_sync, &rdev->flags);
7897                         }
7898
7899                 /* When a reshape changes the number of devices,
7900                  * ->degraded is measured against the larger of the
7901                  * pre and post number of devices.
7902                  */
7903                 spin_lock_irqsave(&conf->device_lock, flags);
7904                 mddev->degraded = raid5_calc_degraded(conf);
7905                 spin_unlock_irqrestore(&conf->device_lock, flags);
7906         }
7907         mddev->raid_disks = conf->raid_disks;
7908         mddev->reshape_position = conf->reshape_progress;
7909         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7910
7911         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7912         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7913         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7914         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7915         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7916         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7917                                                 "reshape");
7918         if (!mddev->sync_thread) {
7919                 mddev->recovery = 0;
7920                 spin_lock_irq(&conf->device_lock);
7921                 write_seqcount_begin(&conf->gen_lock);
7922                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7923                 mddev->new_chunk_sectors =
7924                         conf->chunk_sectors = conf->prev_chunk_sectors;
7925                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7926                 rdev_for_each(rdev, mddev)
7927                         rdev->new_data_offset = rdev->data_offset;
7928                 smp_wmb();
7929                 conf->generation --;
7930                 conf->reshape_progress = MaxSector;
7931                 mddev->reshape_position = MaxSector;
7932                 write_seqcount_end(&conf->gen_lock);
7933                 spin_unlock_irq(&conf->device_lock);
7934                 return -EAGAIN;
7935         }
7936         conf->reshape_checkpoint = jiffies;
7937         md_wakeup_thread(mddev->sync_thread);
7938         md_new_event(mddev);
7939         return 0;
7940 }
7941
7942 /* This is called from the reshape thread and should make any
7943  * changes needed in 'conf'
7944  */
7945 static void end_reshape(struct r5conf *conf)
7946 {
7947
7948         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7949
7950                 spin_lock_irq(&conf->device_lock);
7951                 conf->previous_raid_disks = conf->raid_disks;
7952                 md_finish_reshape(conf->mddev);
7953                 smp_wmb();
7954                 conf->reshape_progress = MaxSector;
7955                 conf->mddev->reshape_position = MaxSector;
7956                 spin_unlock_irq(&conf->device_lock);
7957                 wake_up(&conf->wait_for_overlap);
7958
7959                 /* read-ahead size must cover two whole stripes, which is
7960                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7961                  */
7962                 if (conf->mddev->queue) {
7963                         int data_disks = conf->raid_disks - conf->max_degraded;
7964                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7965                                                    / PAGE_SIZE);
7966                         if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7967                                 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7968                 }
7969         }
7970 }
7971
7972 /* This is called from the raid5d thread with mddev_lock held.
7973  * It makes config changes to the device.
7974  */
7975 static void raid5_finish_reshape(struct mddev *mddev)
7976 {
7977         struct r5conf *conf = mddev->private;
7978
7979         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7980
7981                 if (mddev->delta_disks > 0) {
7982                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7983                         if (mddev->queue) {
7984                                 set_capacity(mddev->gendisk, mddev->array_sectors);
7985                                 revalidate_disk(mddev->gendisk);
7986                         }
7987                 } else {
7988                         int d;
7989                         spin_lock_irq(&conf->device_lock);
7990                         mddev->degraded = raid5_calc_degraded(conf);
7991                         spin_unlock_irq(&conf->device_lock);
7992                         for (d = conf->raid_disks ;
7993                              d < conf->raid_disks - mddev->delta_disks;
7994                              d++) {
7995                                 struct md_rdev *rdev = conf->disks[d].rdev;
7996                                 if (rdev)
7997                                         clear_bit(In_sync, &rdev->flags);
7998                                 rdev = conf->disks[d].replacement;
7999                                 if (rdev)
8000                                         clear_bit(In_sync, &rdev->flags);
8001                         }
8002                 }
8003                 mddev->layout = conf->algorithm;
8004                 mddev->chunk_sectors = conf->chunk_sectors;
8005                 mddev->reshape_position = MaxSector;
8006                 mddev->delta_disks = 0;
8007                 mddev->reshape_backwards = 0;
8008         }
8009 }
8010
8011 static void raid5_quiesce(struct mddev *mddev, int state)
8012 {
8013         struct r5conf *conf = mddev->private;
8014
8015         switch(state) {
8016         case 2: /* resume for a suspend */
8017                 wake_up(&conf->wait_for_overlap);
8018                 break;
8019
8020         case 1: /* stop all writes */
8021                 lock_all_device_hash_locks_irq(conf);
8022                 /* '2' tells resync/reshape to pause so that all
8023                  * active stripes can drain
8024                  */
8025                 r5c_flush_cache(conf, INT_MAX);
8026                 conf->quiesce = 2;
8027                 wait_event_cmd(conf->wait_for_quiescent,
8028                                     atomic_read(&conf->active_stripes) == 0 &&
8029                                     atomic_read(&conf->active_aligned_reads) == 0,
8030                                     unlock_all_device_hash_locks_irq(conf),
8031                                     lock_all_device_hash_locks_irq(conf));
8032                 conf->quiesce = 1;
8033                 unlock_all_device_hash_locks_irq(conf);
8034                 /* allow reshape to continue */
8035                 wake_up(&conf->wait_for_overlap);
8036                 break;
8037
8038         case 0: /* re-enable writes */
8039                 lock_all_device_hash_locks_irq(conf);
8040                 conf->quiesce = 0;
8041                 wake_up(&conf->wait_for_quiescent);
8042                 wake_up(&conf->wait_for_overlap);
8043                 unlock_all_device_hash_locks_irq(conf);
8044                 break;
8045         }
8046         r5l_quiesce(conf->log, state);
8047 }
8048
8049 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8050 {
8051         struct r0conf *raid0_conf = mddev->private;
8052         sector_t sectors;
8053
8054         /* for raid0 takeover only one zone is supported */
8055         if (raid0_conf->nr_strip_zones > 1) {
8056                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8057                         mdname(mddev));
8058                 return ERR_PTR(-EINVAL);
8059         }
8060
8061         sectors = raid0_conf->strip_zone[0].zone_end;
8062         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8063         mddev->dev_sectors = sectors;
8064         mddev->new_level = level;
8065         mddev->new_layout = ALGORITHM_PARITY_N;
8066         mddev->new_chunk_sectors = mddev->chunk_sectors;
8067         mddev->raid_disks += 1;
8068         mddev->delta_disks = 1;
8069         /* make sure it will be not marked as dirty */
8070         mddev->recovery_cp = MaxSector;
8071
8072         return setup_conf(mddev);
8073 }
8074
8075 static void *raid5_takeover_raid1(struct mddev *mddev)
8076 {
8077         int chunksect;
8078         void *ret;
8079
8080         if (mddev->raid_disks != 2 ||
8081             mddev->degraded > 1)
8082                 return ERR_PTR(-EINVAL);
8083
8084         /* Should check if there are write-behind devices? */
8085
8086         chunksect = 64*2; /* 64K by default */
8087
8088         /* The array must be an exact multiple of chunksize */
8089         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8090                 chunksect >>= 1;
8091
8092         if ((chunksect<<9) < STRIPE_SIZE)
8093                 /* array size does not allow a suitable chunk size */
8094                 return ERR_PTR(-EINVAL);
8095
8096         mddev->new_level = 5;
8097         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8098         mddev->new_chunk_sectors = chunksect;
8099
8100         ret = setup_conf(mddev);
8101         if (!IS_ERR(ret))
8102                 mddev_clear_unsupported_flags(mddev,
8103                         UNSUPPORTED_MDDEV_FLAGS);
8104         return ret;
8105 }
8106
8107 static void *raid5_takeover_raid6(struct mddev *mddev)
8108 {
8109         int new_layout;
8110
8111         switch (mddev->layout) {
8112         case ALGORITHM_LEFT_ASYMMETRIC_6:
8113                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8114                 break;
8115         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8116                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8117                 break;
8118         case ALGORITHM_LEFT_SYMMETRIC_6:
8119                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8120                 break;
8121         case ALGORITHM_RIGHT_SYMMETRIC_6:
8122                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8123                 break;
8124         case ALGORITHM_PARITY_0_6:
8125                 new_layout = ALGORITHM_PARITY_0;
8126                 break;
8127         case ALGORITHM_PARITY_N:
8128                 new_layout = ALGORITHM_PARITY_N;
8129                 break;
8130         default:
8131                 return ERR_PTR(-EINVAL);
8132         }
8133         mddev->new_level = 5;
8134         mddev->new_layout = new_layout;
8135         mddev->delta_disks = -1;
8136         mddev->raid_disks -= 1;
8137         return setup_conf(mddev);
8138 }
8139
8140 static int raid5_check_reshape(struct mddev *mddev)
8141 {
8142         /* For a 2-drive array, the layout and chunk size can be changed
8143          * immediately as not restriping is needed.
8144          * For larger arrays we record the new value - after validation
8145          * to be used by a reshape pass.
8146          */
8147         struct r5conf *conf = mddev->private;
8148         int new_chunk = mddev->new_chunk_sectors;
8149
8150         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8151                 return -EINVAL;
8152         if (new_chunk > 0) {
8153                 if (!is_power_of_2(new_chunk))
8154                         return -EINVAL;
8155                 if (new_chunk < (PAGE_SIZE>>9))
8156                         return -EINVAL;
8157                 if (mddev->array_sectors & (new_chunk-1))
8158                         /* not factor of array size */
8159                         return -EINVAL;
8160         }
8161
8162         /* They look valid */
8163
8164         if (mddev->raid_disks == 2) {
8165                 /* can make the change immediately */
8166                 if (mddev->new_layout >= 0) {
8167                         conf->algorithm = mddev->new_layout;
8168                         mddev->layout = mddev->new_layout;
8169                 }
8170                 if (new_chunk > 0) {
8171                         conf->chunk_sectors = new_chunk ;
8172                         mddev->chunk_sectors = new_chunk;
8173                 }
8174                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8175                 md_wakeup_thread(mddev->thread);
8176         }
8177         return check_reshape(mddev);
8178 }
8179
8180 static int raid6_check_reshape(struct mddev *mddev)
8181 {
8182         int new_chunk = mddev->new_chunk_sectors;
8183
8184         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8185                 return -EINVAL;
8186         if (new_chunk > 0) {
8187                 if (!is_power_of_2(new_chunk))
8188                         return -EINVAL;
8189                 if (new_chunk < (PAGE_SIZE >> 9))
8190                         return -EINVAL;
8191                 if (mddev->array_sectors & (new_chunk-1))
8192                         /* not factor of array size */
8193                         return -EINVAL;
8194         }
8195
8196         /* They look valid */
8197         return check_reshape(mddev);
8198 }
8199
8200 static void *raid5_takeover(struct mddev *mddev)
8201 {
8202         /* raid5 can take over:
8203          *  raid0 - if there is only one strip zone - make it a raid4 layout
8204          *  raid1 - if there are two drives.  We need to know the chunk size
8205          *  raid4 - trivial - just use a raid4 layout.
8206          *  raid6 - Providing it is a *_6 layout
8207          */
8208         if (mddev->level == 0)
8209                 return raid45_takeover_raid0(mddev, 5);
8210         if (mddev->level == 1)
8211                 return raid5_takeover_raid1(mddev);
8212         if (mddev->level == 4) {
8213                 mddev->new_layout = ALGORITHM_PARITY_N;
8214                 mddev->new_level = 5;
8215                 return setup_conf(mddev);
8216         }
8217         if (mddev->level == 6)
8218                 return raid5_takeover_raid6(mddev);
8219
8220         return ERR_PTR(-EINVAL);
8221 }
8222
8223 static void *raid4_takeover(struct mddev *mddev)
8224 {
8225         /* raid4 can take over:
8226          *  raid0 - if there is only one strip zone
8227          *  raid5 - if layout is right
8228          */
8229         if (mddev->level == 0)
8230                 return raid45_takeover_raid0(mddev, 4);
8231         if (mddev->level == 5 &&
8232             mddev->layout == ALGORITHM_PARITY_N) {
8233                 mddev->new_layout = 0;
8234                 mddev->new_level = 4;
8235                 return setup_conf(mddev);
8236         }
8237         return ERR_PTR(-EINVAL);
8238 }
8239
8240 static struct md_personality raid5_personality;
8241
8242 static void *raid6_takeover(struct mddev *mddev)
8243 {
8244         /* Currently can only take over a raid5.  We map the
8245          * personality to an equivalent raid6 personality
8246          * with the Q block at the end.
8247          */
8248         int new_layout;
8249
8250         if (mddev->pers != &raid5_personality)
8251                 return ERR_PTR(-EINVAL);
8252         if (mddev->degraded > 1)
8253                 return ERR_PTR(-EINVAL);
8254         if (mddev->raid_disks > 253)
8255                 return ERR_PTR(-EINVAL);
8256         if (mddev->raid_disks < 3)
8257                 return ERR_PTR(-EINVAL);
8258
8259         switch (mddev->layout) {
8260         case ALGORITHM_LEFT_ASYMMETRIC:
8261                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8262                 break;
8263         case ALGORITHM_RIGHT_ASYMMETRIC:
8264                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8265                 break;
8266         case ALGORITHM_LEFT_SYMMETRIC:
8267                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8268                 break;
8269         case ALGORITHM_RIGHT_SYMMETRIC:
8270                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8271                 break;
8272         case ALGORITHM_PARITY_0:
8273                 new_layout = ALGORITHM_PARITY_0_6;
8274                 break;
8275         case ALGORITHM_PARITY_N:
8276                 new_layout = ALGORITHM_PARITY_N;
8277                 break;
8278         default:
8279                 return ERR_PTR(-EINVAL);
8280         }
8281         mddev->new_level = 6;
8282         mddev->new_layout = new_layout;
8283         mddev->delta_disks = 1;
8284         mddev->raid_disks += 1;
8285         return setup_conf(mddev);
8286 }
8287
8288 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8289 {
8290         struct r5conf *conf;
8291         int err;
8292
8293         err = mddev_lock(mddev);
8294         if (err)
8295                 return err;
8296         conf = mddev->private;
8297         if (!conf) {
8298                 mddev_unlock(mddev);
8299                 return -ENODEV;
8300         }
8301
8302         if (strncmp(buf, "ppl", 3) == 0) {
8303                 /* ppl only works with RAID 5 */
8304                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8305                         err = log_init(conf, NULL, true);
8306                         if (!err) {
8307                                 err = resize_stripes(conf, conf->pool_size);
8308                                 if (err)
8309                                         log_exit(conf);
8310                         }
8311                 } else
8312                         err = -EINVAL;
8313         } else if (strncmp(buf, "resync", 6) == 0) {
8314                 if (raid5_has_ppl(conf)) {
8315                         mddev_suspend(mddev);
8316                         log_exit(conf);
8317                         mddev_resume(mddev);
8318                         err = resize_stripes(conf, conf->pool_size);
8319                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8320                            r5l_log_disk_error(conf)) {
8321                         bool journal_dev_exists = false;
8322                         struct md_rdev *rdev;
8323
8324                         rdev_for_each(rdev, mddev)
8325                                 if (test_bit(Journal, &rdev->flags)) {
8326                                         journal_dev_exists = true;
8327                                         break;
8328                                 }
8329
8330                         if (!journal_dev_exists) {
8331                                 mddev_suspend(mddev);
8332                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8333                                 mddev_resume(mddev);
8334                         } else  /* need remove journal device first */
8335                                 err = -EBUSY;
8336                 } else
8337                         err = -EINVAL;
8338         } else {
8339                 err = -EINVAL;
8340         }
8341
8342         if (!err)
8343                 md_update_sb(mddev, 1);
8344
8345         mddev_unlock(mddev);
8346
8347         return err;
8348 }
8349
8350 static struct md_personality raid6_personality =
8351 {
8352         .name           = "raid6",
8353         .level          = 6,
8354         .owner          = THIS_MODULE,
8355         .make_request   = raid5_make_request,
8356         .run            = raid5_run,
8357         .free           = raid5_free,
8358         .status         = raid5_status,
8359         .error_handler  = raid5_error,
8360         .hot_add_disk   = raid5_add_disk,
8361         .hot_remove_disk= raid5_remove_disk,
8362         .spare_active   = raid5_spare_active,
8363         .sync_request   = raid5_sync_request,
8364         .resize         = raid5_resize,
8365         .size           = raid5_size,
8366         .check_reshape  = raid6_check_reshape,
8367         .start_reshape  = raid5_start_reshape,
8368         .finish_reshape = raid5_finish_reshape,
8369         .quiesce        = raid5_quiesce,
8370         .takeover       = raid6_takeover,
8371         .congested      = raid5_congested,
8372         .change_consistency_policy = raid5_change_consistency_policy,
8373 };
8374 static struct md_personality raid5_personality =
8375 {
8376         .name           = "raid5",
8377         .level          = 5,
8378         .owner          = THIS_MODULE,
8379         .make_request   = raid5_make_request,
8380         .run            = raid5_run,
8381         .free           = raid5_free,
8382         .status         = raid5_status,
8383         .error_handler  = raid5_error,
8384         .hot_add_disk   = raid5_add_disk,
8385         .hot_remove_disk= raid5_remove_disk,
8386         .spare_active   = raid5_spare_active,
8387         .sync_request   = raid5_sync_request,
8388         .resize         = raid5_resize,
8389         .size           = raid5_size,
8390         .check_reshape  = raid5_check_reshape,
8391         .start_reshape  = raid5_start_reshape,
8392         .finish_reshape = raid5_finish_reshape,
8393         .quiesce        = raid5_quiesce,
8394         .takeover       = raid5_takeover,
8395         .congested      = raid5_congested,
8396         .change_consistency_policy = raid5_change_consistency_policy,
8397 };
8398
8399 static struct md_personality raid4_personality =
8400 {
8401         .name           = "raid4",
8402         .level          = 4,
8403         .owner          = THIS_MODULE,
8404         .make_request   = raid5_make_request,
8405         .run            = raid5_run,
8406         .free           = raid5_free,
8407         .status         = raid5_status,
8408         .error_handler  = raid5_error,
8409         .hot_add_disk   = raid5_add_disk,
8410         .hot_remove_disk= raid5_remove_disk,
8411         .spare_active   = raid5_spare_active,
8412         .sync_request   = raid5_sync_request,
8413         .resize         = raid5_resize,
8414         .size           = raid5_size,
8415         .check_reshape  = raid5_check_reshape,
8416         .start_reshape  = raid5_start_reshape,
8417         .finish_reshape = raid5_finish_reshape,
8418         .quiesce        = raid5_quiesce,
8419         .takeover       = raid4_takeover,
8420         .congested      = raid5_congested,
8421         .change_consistency_policy = raid5_change_consistency_policy,
8422 };
8423
8424 static int __init raid5_init(void)
8425 {
8426         int ret;
8427
8428         raid5_wq = alloc_workqueue("raid5wq",
8429                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8430         if (!raid5_wq)
8431                 return -ENOMEM;
8432
8433         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8434                                       "md/raid5:prepare",
8435                                       raid456_cpu_up_prepare,
8436                                       raid456_cpu_dead);
8437         if (ret) {
8438                 destroy_workqueue(raid5_wq);
8439                 return ret;
8440         }
8441         register_md_personality(&raid6_personality);
8442         register_md_personality(&raid5_personality);
8443         register_md_personality(&raid4_personality);
8444         return 0;
8445 }
8446
8447 static void raid5_exit(void)
8448 {
8449         unregister_md_personality(&raid6_personality);
8450         unregister_md_personality(&raid5_personality);
8451         unregister_md_personality(&raid4_personality);
8452         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8453         destroy_workqueue(raid5_wq);
8454 }
8455
8456 module_init(raid5_init);
8457 module_exit(raid5_exit);
8458 MODULE_LICENSE("GPL");
8459 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8460 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8461 MODULE_ALIAS("md-raid5");
8462 MODULE_ALIAS("md-raid4");
8463 MODULE_ALIAS("md-level-5");
8464 MODULE_ALIAS("md-level-4");
8465 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8466 MODULE_ALIAS("md-raid6");
8467 MODULE_ALIAS("md-level-6");
8468
8469 /* This used to be two separate modules, they were: */
8470 MODULE_ALIAS("raid5");
8471 MODULE_ALIAS("raid6");