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 raid5_build_block(struct stripe_head *sh, int i, int previous);
 498 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 499                             struct stripe_head *sh);
 500
 501 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 502 {
 503         struct r5conf *conf = sh->raid_conf;
 504         int i, seq;
 505
 506         BUG_ON(atomic_read(&sh->count) != 0);
 507         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 508         BUG_ON(stripe_operations_active(sh));
 509         BUG_ON(sh->batch_head);
 510
 511         pr_debug("init_stripe called, stripe %llu\n",
 512                 (unsigned long long)sector);
 513 retry:
 514         seq = read_seqcount_begin(&conf->gen_lock);
 515         sh->generation = conf->generation - previous;
 516         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 517         sh->sector = sector;
 518         stripe_set_idx(sector, conf, previous, sh);
 519         sh->state = 0;
 520
 521         for (i = sh->disks; i--; ) {
 522                 struct r5dev *dev = &sh->dev[i];
 523
 524                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 525                     test_bit(R5_LOCKED, &dev->flags)) {
 526                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 527                                (unsigned long long)sh->sector, i, dev->toread,
 528                                dev->read, dev->towrite, dev->written,
 529                                test_bit(R5_LOCKED, &dev->flags));
 530                         WARN_ON(1);
 531                 }
 532                 dev->flags = 0;
 533                 raid5_build_block(sh, i, previous);
 534         }
 535         if (read_seqcount_retry(&conf->gen_lock, seq))
 536                 goto retry;
 537         sh->overwrite_disks = 0;
 538         insert_hash(conf, sh);
 539         sh->cpu = smp_processor_id();
 540         set_bit(STRIPE_BATCH_READY, &sh->state);
 541 }
 542
 543 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 544                                          short generation)
 545 {
 546         struct stripe_head *sh;
 547
 548         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 549         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 550                 if (sh->sector == sector && sh->generation == generation)
 551                         return sh;
 552         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 553         return NULL;
 554 }
 555
 556 /*
 557  * Need to check if array has failed when deciding whether to:
 558  *  - start an array
 559  *  - remove non-faulty devices
 560  *  - add a spare
 561  *  - allow a reshape
 562  * This determination is simple when no reshape is happening.
 563  * However if there is a reshape, we need to carefully check
 564  * both the before and after sections.
 565  * This is because some failed devices may only affect one
 566  * of the two sections, and some non-in_sync devices may
 567  * be insync in the section most affected by failed devices.
 568  */
 569 int raid5_calc_degraded(struct r5conf *conf)
 570 {
 571         int degraded, degraded2;
 572         int i;
 573
 574         rcu_read_lock();
 575         degraded = 0;
 576         for (i = 0; i < conf->previous_raid_disks; i++) {
 577                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 578                 if (rdev && test_bit(Faulty, &rdev->flags))
 579                         rdev = rcu_dereference(conf->disks[i].replacement);
 580                 if (!rdev || test_bit(Faulty, &rdev->flags))
 581                         degraded++;
 582                 else if (test_bit(In_sync, &rdev->flags))
 583                         ;
 584                 else
 585                         /* not in-sync or faulty.
 586                          * If the reshape increases the number of devices,
 587                          * this is being recovered by the reshape, so
 588                          * this 'previous' section is not in_sync.
 589                          * If the number of devices is being reduced however,
 590                          * the device can only be part of the array if
 591                          * we are reverting a reshape, so this section will
 592                          * be in-sync.
 593                          */
 594                         if (conf->raid_disks >= conf->previous_raid_disks)
 595                                 degraded++;
 596         }
 597         rcu_read_unlock();
 598         if (conf->raid_disks == conf->previous_raid_disks)
 599                 return degraded;
 600         rcu_read_lock();
 601         degraded2 = 0;
 602         for (i = 0; i < conf->raid_disks; i++) {
 603                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 604                 if (rdev && test_bit(Faulty, &rdev->flags))
 605                         rdev = rcu_dereference(conf->disks[i].replacement);
 606                 if (!rdev || test_bit(Faulty, &rdev->flags))
 607                         degraded2++;
 608                 else if (test_bit(In_sync, &rdev->flags))
 609                         ;
 610                 else
 611                         /* not in-sync or faulty.
 612                          * If reshape increases the number of devices, this
 613                          * section has already been recovered, else it
 614                          * almost certainly hasn't.
 615                          */
 616                         if (conf->raid_disks <= conf->previous_raid_disks)
 617                                 degraded2++;
 618         }
 619         rcu_read_unlock();
 620         if (degraded2 > degraded)
 621                 return degraded2;
 622         return degraded;
 623 }
 624
 625 static int has_failed(struct r5conf *conf)
 626 {
 627         int degraded;
 628
 629         if (conf->mddev->reshape_position == MaxSector)
 630                 return conf->mddev->degraded > conf->max_degraded;
 631
 632         degraded = raid5_calc_degraded(conf);
 633         if (degraded > conf->max_degraded)
 634                 return 1;
 635         return 0;
 636 }
 637
 638 struct stripe_head *
 639 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 640                         int previous, int noblock, int noquiesce)
 641 {
 642         struct stripe_head *sh;
 643         int hash = stripe_hash_locks_hash(sector);
 644         int inc_empty_inactive_list_flag;
 645
 646         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 647
 648         spin_lock_irq(conf->hash_locks + hash);
 649
 650         do {
 651                 wait_event_lock_irq(conf->wait_for_quiescent,
 652                                     conf->quiesce == 0 || noquiesce,
 653                                     *(conf->hash_locks + hash));
 654                 sh = __find_stripe(conf, sector, conf->generation - previous);
 655                 if (!sh) {
 656                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 657                                 sh = get_free_stripe(conf, hash);
 658                                 if (!sh && !test_bit(R5_DID_ALLOC,
 659                                                      &conf->cache_state))
 660                                         set_bit(R5_ALLOC_MORE,
 661                                                 &conf->cache_state);
 662                         }
 663                         if (noblock && sh == NULL)
 664                                 break;
 665
 666                         r5c_check_stripe_cache_usage(conf);
 667                         if (!sh) {
 668                                 set_bit(R5_INACTIVE_BLOCKED,
 669                                         &conf->cache_state);
 670                                 r5l_wake_reclaim(conf->log, 0);
 671                                 wait_event_lock_irq(
 672                                         conf->wait_for_stripe,
 673                                         !list_empty(conf->inactive_list + hash) &&
 674                                         (atomic_read(&conf->active_stripes)
 675                                          < (conf->max_nr_stripes * 3 / 4)
 676                                          || !test_bit(R5_INACTIVE_BLOCKED,
 677                                                       &conf->cache_state)),
 678                                         *(conf->hash_locks + hash));
 679                                 clear_bit(R5_INACTIVE_BLOCKED,
 680                                           &conf->cache_state);
 681                         } else {
 682                                 init_stripe(sh, sector, previous);
 683                                 atomic_inc(&sh->count);
 684                         }
 685                 } else if (!atomic_inc_not_zero(&sh->count)) {
 686                         spin_lock(&conf->device_lock);
 687                         if (!atomic_read(&sh->count)) {
 688                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 689                                         atomic_inc(&conf->active_stripes);
 690                                 BUG_ON(list_empty(&sh->lru) &&
 691                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 692                                 inc_empty_inactive_list_flag = 0;
 693                                 if (!list_empty(conf->inactive_list + hash))
 694                                         inc_empty_inactive_list_flag = 1;
 695                                 list_del_init(&sh->lru);
 696                                 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 697                                         atomic_inc(&conf->empty_inactive_list_nr);
 698                                 if (sh->group) {
 699                                         sh->group->stripes_cnt--;
 700                                         sh->group = NULL;
 701                                 }
 702                         }
 703                         atomic_inc(&sh->count);
 704                         spin_unlock(&conf->device_lock);
 705                 }
 706         } while (sh == NULL);
 707
 708         spin_unlock_irq(conf->hash_locks + hash);
 709         return sh;
 710 }
 711
 712 static bool is_full_stripe_write(struct stripe_head *sh)
 713 {
 714         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 715         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 716 }
 717
 718 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 719 {
 720         if (sh1 > sh2) {
 721                 spin_lock_irq(&sh2->stripe_lock);
 722                 spin_lock_nested(&sh1->stripe_lock, 1);
 723         } else {
 724                 spin_lock_irq(&sh1->stripe_lock);
 725                 spin_lock_nested(&sh2->stripe_lock, 1);
 726         }
 727 }
 728
 729 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 730 {
 731         spin_unlock(&sh1->stripe_lock);
 732         spin_unlock_irq(&sh2->stripe_lock);
 733 }
 734
 735 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 736 static bool stripe_can_batch(struct stripe_head *sh)
 737 {
 738         struct r5conf *conf = sh->raid_conf;
 739
 740         if (conf->log || raid5_has_ppl(conf))
 741                 return false;
 742         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 743                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 744                 is_full_stripe_write(sh);
 745 }
 746
 747 /* we only do back search */
 748 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 749 {
 750         struct stripe_head *head;
 751         sector_t head_sector, tmp_sec;
 752         int hash;
 753         int dd_idx;
 754         int inc_empty_inactive_list_flag;
 755
 756         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 757         tmp_sec = sh->sector;
 758         if (!sector_div(tmp_sec, conf->chunk_sectors))
 759                 return;
 760         head_sector = sh->sector - STRIPE_SECTORS;
 761
 762         hash = stripe_hash_locks_hash(head_sector);
 763         spin_lock_irq(conf->hash_locks + hash);
 764         head = __find_stripe(conf, head_sector, conf->generation);
 765         if (head && !atomic_inc_not_zero(&head->count)) {
 766                 spin_lock(&conf->device_lock);
 767                 if (!atomic_read(&head->count)) {
 768                         if (!test_bit(STRIPE_HANDLE, &head->state))
 769                                 atomic_inc(&conf->active_stripes);
 770                         BUG_ON(list_empty(&head->lru) &&
 771                                !test_bit(STRIPE_EXPANDING, &head->state));
 772                         inc_empty_inactive_list_flag = 0;
 773                         if (!list_empty(conf->inactive_list + hash))
 774                                 inc_empty_inactive_list_flag = 1;
 775                         list_del_init(&head->lru);
 776                         if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 777                                 atomic_inc(&conf->empty_inactive_list_nr);
 778                         if (head->group) {
 779                                 head->group->stripes_cnt--;
 780                                 head->group = NULL;
 781                         }
 782                 }
 783                 atomic_inc(&head->count);
 784                 spin_unlock(&conf->device_lock);
 785         }
 786         spin_unlock_irq(conf->hash_locks + hash);
 787
 788         if (!head)
 789                 return;
 790         if (!stripe_can_batch(head))
 791                 goto out;
 792
 793         lock_two_stripes(head, sh);
 794         /* clear_batch_ready clear the flag */
 795         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 796                 goto unlock_out;
 797
 798         if (sh->batch_head)
 799                 goto unlock_out;
 800
 801         dd_idx = 0;
 802         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 803                 dd_idx++;
 804         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 805             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 806                 goto unlock_out;
 807
 808         if (head->batch_head) {
 809                 spin_lock(&head->batch_head->batch_lock);
 810                 /* This batch list is already running */
 811                 if (!stripe_can_batch(head)) {
 812                         spin_unlock(&head->batch_head->batch_lock);
 813                         goto unlock_out;
 814                 }
 815
 816                 /*
 817                  * at this point, head's BATCH_READY could be cleared, but we
 818                  * can still add the stripe to batch list
 819                  */
 820                 list_add(&sh->batch_list, &head->batch_list);
 821                 spin_unlock(&head->batch_head->batch_lock);
 822
 823                 sh->batch_head = head->batch_head;
 824         } else {
 825                 head->batch_head = head;
 826                 sh->batch_head = head->batch_head;
 827                 spin_lock(&head->batch_lock);
 828                 list_add_tail(&sh->batch_list, &head->batch_list);
 829                 spin_unlock(&head->batch_lock);
 830         }
 831
 832         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 833                 if (atomic_dec_return(&conf->preread_active_stripes)
 834                     < IO_THRESHOLD)
 835                         md_wakeup_thread(conf->mddev->thread);
 836
 837         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 838                 int seq = sh->bm_seq;
 839                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 840                     sh->batch_head->bm_seq > seq)
 841                         seq = sh->batch_head->bm_seq;
 842                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 843                 sh->batch_head->bm_seq = seq;
 844         }
 845
 846         atomic_inc(&sh->count);
 847 unlock_out:
 848         unlock_two_stripes(head, sh);
 849 out:
 850         raid5_release_stripe(head);
 851 }
 852
 853 /* Determine if 'data_offset' or 'new_data_offset' should be used
 854  * in this stripe_head.
 855  */
 856 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 857 {
 858         sector_t progress = conf->reshape_progress;
 859         /* Need a memory barrier to make sure we see the value
 860          * of conf->generation, or ->data_offset that was set before
 861          * reshape_progress was updated.
 862          */
 863         smp_rmb();
 864         if (progress == MaxSector)
 865                 return 0;
 866         if (sh->generation == conf->generation - 1)
 867                 return 0;
 868         /* We are in a reshape, and this is a new-generation stripe,
 869          * so use new_data_offset.
 870          */
 871         return 1;
 872 }
 873
 874 static void dispatch_bio_list(struct bio_list *tmp)
 875 {
 876         struct bio *bio;
 877
 878         while ((bio = bio_list_pop(tmp)))
 879                 generic_make_request(bio);
 880 }
 881
 882 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
 883 {
 884         const struct r5pending_data *da = list_entry(a,
 885                                 struct r5pending_data, sibling);
 886         const struct r5pending_data *db = list_entry(b,
 887                                 struct r5pending_data, sibling);
 888         if (da->sector > db->sector)
 889                 return 1;
 890         if (da->sector < db->sector)
 891                 return -1;
 892         return 0;
 893 }
 894
 895 static void dispatch_defer_bios(struct r5conf *conf, int target,
 896                                 struct bio_list *list)
 897 {
 898         struct r5pending_data *data;
 899         struct list_head *first, *next = NULL;
 900         int cnt = 0;
 901
 902         if (conf->pending_data_cnt == 0)
 903                 return;
 904
 905         list_sort(NULL, &conf->pending_list, cmp_stripe);
 906
 907         first = conf->pending_list.next;
 908
 909         /* temporarily move the head */
 910         if (conf->next_pending_data)
 911                 list_move_tail(&conf->pending_list,
 912                                 &conf->next_pending_data->sibling);
 913
 914         while (!list_empty(&conf->pending_list)) {
 915                 data = list_first_entry(&conf->pending_list,
 916                         struct r5pending_data, sibling);
 917                 if (&data->sibling == first)
 918                         first = data->sibling.next;
 919                 next = data->sibling.next;
 920
 921                 bio_list_merge(list, &data->bios);
 922                 list_move(&data->sibling, &conf->free_list);
 923                 cnt++;
 924                 if (cnt >= target)
 925                         break;
 926         }
 927         conf->pending_data_cnt -= cnt;
 928         BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
 929
 930         if (next != &conf->pending_list)
 931                 conf->next_pending_data = list_entry(next,
 932                                 struct r5pending_data, sibling);
 933         else
 934                 conf->next_pending_data = NULL;
 935         /* list isn't empty */
 936         if (first != &conf->pending_list)
 937                 list_move_tail(&conf->pending_list, first);
 938 }
 939
 940 static void flush_deferred_bios(struct r5conf *conf)
 941 {
 942         struct bio_list tmp = BIO_EMPTY_LIST;
 943
 944         if (conf->pending_data_cnt == 0)
 945                 return;
 946
 947         spin_lock(&conf->pending_bios_lock);
 948         dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
 949         BUG_ON(conf->pending_data_cnt != 0);
 950         spin_unlock(&conf->pending_bios_lock);
 951
 952         dispatch_bio_list(&tmp);
 953 }
 954
 955 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
 956                                 struct bio_list *bios)
 957 {
 958         struct bio_list tmp = BIO_EMPTY_LIST;
 959         struct r5pending_data *ent;
 960
 961         spin_lock(&conf->pending_bios_lock);
 962         ent = list_first_entry(&conf->free_list, struct r5pending_data,
 963                                                         sibling);
 964         list_move_tail(&ent->sibling, &conf->pending_list);
 965         ent->sector = sector;
 966         bio_list_init(&ent->bios);
 967         bio_list_merge(&ent->bios, bios);
 968         conf->pending_data_cnt++;
 969         if (conf->pending_data_cnt >= PENDING_IO_MAX)
 970                 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
 971
 972         spin_unlock(&conf->pending_bios_lock);
 973
 974         dispatch_bio_list(&tmp);
 975 }
 976
 977 static void
 978 raid5_end_read_request(struct bio *bi);
 979 static void
 980 raid5_end_write_request(struct bio *bi);
 981
 982 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 983 {
 984         struct r5conf *conf = sh->raid_conf;
 985         int i, disks = sh->disks;
 986         struct stripe_head *head_sh = sh;
 987         struct bio_list pending_bios = BIO_EMPTY_LIST;
 988         bool should_defer;
 989
 990         might_sleep();
 991
 992         if (log_stripe(sh, s) == 0)
 993                 return;
 994
 995         should_defer = conf->batch_bio_dispatch && conf->group_cnt;
 996
 997         for (i = disks; i--; ) {
 998                 int op, op_flags = 0;
 999                 int replace_only = 0;
1000                 struct bio *bi, *rbi;
1001                 struct md_rdev *rdev, *rrdev = NULL;
1002
1003                 sh = head_sh;
1004                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1005                         op = REQ_OP_WRITE;
1006                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1007                                 op_flags = REQ_FUA;
1008                         if (test_bit(R5_Discard, &sh->dev[i].flags))
1009                                 op = REQ_OP_DISCARD;
1010                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1011                         op = REQ_OP_READ;
1012                 else if (test_and_clear_bit(R5_WantReplace,
1013                                             &sh->dev[i].flags)) {
1014                         op = REQ_OP_WRITE;
1015                         replace_only = 1;
1016                 } else
1017                         continue;
1018                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1019                         op_flags |= REQ_SYNC;
1020
1021 again:
1022                 bi = &sh->dev[i].req;
1023                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1024
1025                 rcu_read_lock();
1026                 rrdev = rcu_dereference(conf->disks[i].replacement);
1027                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1028                 rdev = rcu_dereference(conf->disks[i].rdev);
1029                 if (!rdev) {
1030                         rdev = rrdev;
1031                         rrdev = NULL;
1032                 }
1033                 if (op_is_write(op)) {
1034                         if (replace_only)
1035                                 rdev = NULL;
1036                         if (rdev == rrdev)
1037                                 /* We raced and saw duplicates */
1038                                 rrdev = NULL;
1039                 } else {
1040                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1041                                 rdev = rrdev;
1042                         rrdev = NULL;
1043                 }
1044
1045                 if (rdev && test_bit(Faulty, &rdev->flags))
1046                         rdev = NULL;
1047                 if (rdev)
1048                         atomic_inc(&rdev->nr_pending);
1049                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1050                         rrdev = NULL;
1051                 if (rrdev)
1052                         atomic_inc(&rrdev->nr_pending);
1053                 rcu_read_unlock();
1054
1055                 /* We have already checked bad blocks for reads.  Now
1056                  * need to check for writes.  We never accept write errors
1057                  * on the replacement, so we don't to check rrdev.
1058                  */
1059                 while (op_is_write(op) && rdev &&
1060                        test_bit(WriteErrorSeen, &rdev->flags)) {
1061                         sector_t first_bad;
1062                         int bad_sectors;
1063                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1064                                               &first_bad, &bad_sectors);
1065                         if (!bad)
1066                                 break;
1067
1068                         if (bad < 0) {
1069                                 set_bit(BlockedBadBlocks, &rdev->flags);
1070                                 if (!conf->mddev->external &&
1071                                     conf->mddev->sb_flags) {
1072                                         /* It is very unlikely, but we might
1073                                          * still need to write out the
1074                                          * bad block log - better give it
1075                                          * a chance*/
1076                                         md_check_recovery(conf->mddev);
1077                                 }
1078                                 /*
1079                                  * Because md_wait_for_blocked_rdev
1080                                  * will dec nr_pending, we must
1081                                  * increment it first.
1082                                  */
1083                                 atomic_inc(&rdev->nr_pending);
1084                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
1085                         } else {
1086                                 /* Acknowledged bad block - skip the write */
1087                                 rdev_dec_pending(rdev, conf->mddev);
1088                                 rdev = NULL;
1089                         }
1090                 }
1091
1092                 if (rdev) {
1093                         if (s->syncing || s->expanding || s->expanded
1094                             || s->replacing)
1095                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1096
1097                         set_bit(STRIPE_IO_STARTED, &sh->state);
1098
1099                         bi->bi_bdev = rdev->bdev;
1100                         bio_set_op_attrs(bi, op, op_flags);
1101                         bi->bi_end_io = op_is_write(op)
1102                                 ? raid5_end_write_request
1103                                 : raid5_end_read_request;
1104                         bi->bi_private = sh;
1105
1106                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1107                                 __func__, (unsigned long long)sh->sector,
1108                                 bi->bi_opf, i);
1109                         atomic_inc(&sh->count);
1110                         if (sh != head_sh)
1111                                 atomic_inc(&head_sh->count);
1112                         if (use_new_offset(conf, sh))
1113                                 bi->bi_iter.bi_sector = (sh->sector
1114                                                  + rdev->new_data_offset);
1115                         else
1116                                 bi->bi_iter.bi_sector = (sh->sector
1117                                                  + rdev->data_offset);
1118                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1119                                 bi->bi_opf |= REQ_NOMERGE;
1120
1121                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1122                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1123
1124                         if (!op_is_write(op) &&
1125                             test_bit(R5_InJournal, &sh->dev[i].flags))
1126                                 /*
1127                                  * issuing read for a page in journal, this
1128                                  * must be preparing for prexor in rmw; read
1129                                  * the data into orig_page
1130                                  */
1131                                 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1132                         else
1133                                 sh->dev[i].vec.bv_page = sh->dev[i].page;
1134                         bi->bi_vcnt = 1;
1135                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1136                         bi->bi_io_vec[0].bv_offset = 0;
1137                         bi->bi_iter.bi_size = STRIPE_SIZE;
1138                         /*
1139                          * If this is discard request, set bi_vcnt 0. We don't
1140                          * want to confuse SCSI because SCSI will replace payload
1141                          */
1142                         if (op == REQ_OP_DISCARD)
1143                                 bi->bi_vcnt = 0;
1144                         if (rrdev)
1145                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1146
1147                         if (conf->mddev->gendisk)
1148                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1149                                                       bi, disk_devt(conf->mddev->gendisk),
1150                                                       sh->dev[i].sector);
1151                         if (should_defer && op_is_write(op))
1152                                 bio_list_add(&pending_bios, bi);
1153                         else
1154                                 generic_make_request(bi);
1155                 }
1156                 if (rrdev) {
1157                         if (s->syncing || s->expanding || s->expanded
1158                             || s->replacing)
1159                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1160
1161                         set_bit(STRIPE_IO_STARTED, &sh->state);
1162
1163                         rbi->bi_bdev = rrdev->bdev;
1164                         bio_set_op_attrs(rbi, op, op_flags);
1165                         BUG_ON(!op_is_write(op));
1166                         rbi->bi_end_io = raid5_end_write_request;
1167                         rbi->bi_private = sh;
1168
1169                         pr_debug("%s: for %llu schedule op %d on "
1170                                  "replacement disc %d\n",
1171                                 __func__, (unsigned long long)sh->sector,
1172                                 rbi->bi_opf, i);
1173                         atomic_inc(&sh->count);
1174                         if (sh != head_sh)
1175                                 atomic_inc(&head_sh->count);
1176                         if (use_new_offset(conf, sh))
1177                                 rbi->bi_iter.bi_sector = (sh->sector
1178                                                   + rrdev->new_data_offset);
1179                         else
1180                                 rbi->bi_iter.bi_sector = (sh->sector
1181                                                   + rrdev->data_offset);
1182                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1183                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1184                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1185                         rbi->bi_vcnt = 1;
1186                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1187                         rbi->bi_io_vec[0].bv_offset = 0;
1188                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1189                         /*
1190                          * If this is discard request, set bi_vcnt 0. We don't
1191                          * want to confuse SCSI because SCSI will replace payload
1192                          */
1193                         if (op == REQ_OP_DISCARD)
1194                                 rbi->bi_vcnt = 0;
1195                         if (conf->mddev->gendisk)
1196                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1197                                                       rbi, disk_devt(conf->mddev->gendisk),
1198                                                       sh->dev[i].sector);
1199                         if (should_defer && op_is_write(op))
1200                                 bio_list_add(&pending_bios, rbi);
1201                         else
1202                                 generic_make_request(rbi);
1203                 }
1204                 if (!rdev && !rrdev) {
1205                         if (op_is_write(op))
1206                                 set_bit(STRIPE_DEGRADED, &sh->state);
1207                         pr_debug("skip op %d on disc %d for sector %llu\n",
1208                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1209                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1210                         set_bit(STRIPE_HANDLE, &sh->state);
1211                 }
1212
1213                 if (!head_sh->batch_head)
1214                         continue;
1215                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1216                                       batch_list);
1217                 if (sh != head_sh)
1218                         goto again;
1219         }
1220
1221         if (should_defer && !bio_list_empty(&pending_bios))
1222                 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1223 }
1224
1225 static struct dma_async_tx_descriptor *
1226 async_copy_data(int frombio, struct bio *bio, struct page **page,
1227         sector_t sector, struct dma_async_tx_descriptor *tx,
1228         struct stripe_head *sh, int no_skipcopy)
1229 {
1230         struct bio_vec bvl;
1231         struct bvec_iter iter;
1232         struct page *bio_page;
1233         int page_offset;
1234         struct async_submit_ctl submit;
1235         enum async_tx_flags flags = 0;
1236
1237         if (bio->bi_iter.bi_sector >= sector)
1238                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1239         else
1240                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1241
1242         if (frombio)
1243                 flags |= ASYNC_TX_FENCE;
1244         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1245
1246         bio_for_each_segment(bvl, bio, iter) {
1247                 int len = bvl.bv_len;
1248                 int clen;
1249                 int b_offset = 0;
1250
1251                 if (page_offset < 0) {
1252                         b_offset = -page_offset;
1253                         page_offset += b_offset;
1254                         len -= b_offset;
1255                 }
1256
1257                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1258                         clen = STRIPE_SIZE - page_offset;
1259                 else
1260                         clen = len;
1261
1262                 if (clen > 0) {
1263                         b_offset += bvl.bv_offset;
1264                         bio_page = bvl.bv_page;
1265                         if (frombio) {
1266                                 if (sh->raid_conf->skip_copy &&
1267                                     b_offset == 0 && page_offset == 0 &&
1268                                     clen == STRIPE_SIZE &&
1269                                     !no_skipcopy)
1270                                         *page = bio_page;
1271                                 else
1272                                         tx = async_memcpy(*page, bio_page, page_offset,
1273                                                   b_offset, clen, &submit);
1274                         } else
1275                                 tx = async_memcpy(bio_page, *page, b_offset,
1276                                                   page_offset, clen, &submit);
1277                 }
1278                 /* chain the operations */
1279                 submit.depend_tx = tx;
1280
1281                 if (clen < len) /* hit end of page */
1282                         break;
1283                 page_offset +=  len;
1284         }
1285
1286         return tx;
1287 }
1288
1289 static void ops_complete_biofill(void *stripe_head_ref)
1290 {
1291         struct stripe_head *sh = stripe_head_ref;
1292         int i;
1293
1294         pr_debug("%s: stripe %llu\n", __func__,
1295                 (unsigned long long)sh->sector);
1296
1297         /* clear completed biofills */
1298         for (i = sh->disks; i--; ) {
1299                 struct r5dev *dev = &sh->dev[i];
1300
1301                 /* acknowledge completion of a biofill operation */
1302                 /* and check if we need to reply to a read request,
1303                  * new R5_Wantfill requests are held off until
1304                  * !STRIPE_BIOFILL_RUN
1305                  */
1306                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1307                         struct bio *rbi, *rbi2;
1308
1309                         BUG_ON(!dev->read);
1310                         rbi = dev->read;
1311                         dev->read = NULL;
1312                         while (rbi && rbi->bi_iter.bi_sector <
1313                                 dev->sector + STRIPE_SECTORS) {
1314                                 rbi2 = r5_next_bio(rbi, dev->sector);
1315                                 bio_endio(rbi);
1316                                 rbi = rbi2;
1317                         }
1318                 }
1319         }
1320         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1321
1322         set_bit(STRIPE_HANDLE, &sh->state);
1323         raid5_release_stripe(sh);
1324 }
1325
1326 static void ops_run_biofill(struct stripe_head *sh)
1327 {
1328         struct dma_async_tx_descriptor *tx = NULL;
1329         struct async_submit_ctl submit;
1330         int i;
1331
1332         BUG_ON(sh->batch_head);
1333         pr_debug("%s: stripe %llu\n", __func__,
1334                 (unsigned long long)sh->sector);
1335
1336         for (i = sh->disks; i--; ) {
1337                 struct r5dev *dev = &sh->dev[i];
1338                 if (test_bit(R5_Wantfill, &dev->flags)) {
1339                         struct bio *rbi;
1340                         spin_lock_irq(&sh->stripe_lock);
1341                         dev->read = rbi = dev->toread;
1342                         dev->toread = NULL;
1343                         spin_unlock_irq(&sh->stripe_lock);
1344                         while (rbi && rbi->bi_iter.bi_sector <
1345                                 dev->sector + STRIPE_SECTORS) {
1346                                 tx = async_copy_data(0, rbi, &dev->page,
1347                                                      dev->sector, tx, sh, 0);
1348                                 rbi = r5_next_bio(rbi, dev->sector);
1349                         }
1350                 }
1351         }
1352
1353         atomic_inc(&sh->count);
1354         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1355         async_trigger_callback(&submit);
1356 }
1357
1358 static void mark_target_uptodate(struct stripe_head *sh, int target)
1359 {
1360         struct r5dev *tgt;
1361
1362         if (target < 0)
1363                 return;
1364
1365         tgt = &sh->dev[target];
1366         set_bit(R5_UPTODATE, &tgt->flags);
1367         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1368         clear_bit(R5_Wantcompute, &tgt->flags);
1369 }
1370
1371 static void ops_complete_compute(void *stripe_head_ref)
1372 {
1373         struct stripe_head *sh = stripe_head_ref;
1374
1375         pr_debug("%s: stripe %llu\n", __func__,
1376                 (unsigned long long)sh->sector);
1377
1378         /* mark the computed target(s) as uptodate */
1379         mark_target_uptodate(sh, sh->ops.target);
1380         mark_target_uptodate(sh, sh->ops.target2);
1381
1382         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1383         if (sh->check_state == check_state_compute_run)
1384                 sh->check_state = check_state_compute_result;
1385         set_bit(STRIPE_HANDLE, &sh->state);
1386         raid5_release_stripe(sh);
1387 }
1388
1389 /* return a pointer to the address conversion region of the scribble buffer */
1390 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1391                                  struct raid5_percpu *percpu, int i)
1392 {
1393         void *addr;
1394
1395         addr = flex_array_get(percpu->scribble, i);
1396         return addr + sizeof(struct page *) * (sh->disks + 2);
1397 }
1398
1399 /* return a pointer to the address conversion region of the scribble buffer */
1400 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1401 {
1402         void *addr;
1403
1404         addr = flex_array_get(percpu->scribble, i);
1405         return addr;
1406 }
1407
1408 static struct dma_async_tx_descriptor *
1409 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1410 {
1411         int disks = sh->disks;
1412         struct page **xor_srcs = to_addr_page(percpu, 0);
1413         int target = sh->ops.target;
1414         struct r5dev *tgt = &sh->dev[target];
1415         struct page *xor_dest = tgt->page;
1416         int count = 0;
1417         struct dma_async_tx_descriptor *tx;
1418         struct async_submit_ctl submit;
1419         int i;
1420
1421         BUG_ON(sh->batch_head);
1422
1423         pr_debug("%s: stripe %llu block: %d\n",
1424                 __func__, (unsigned long long)sh->sector, target);
1425         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1426
1427         for (i = disks; i--; )
1428                 if (i != target)
1429                         xor_srcs[count++] = sh->dev[i].page;
1430
1431         atomic_inc(&sh->count);
1432
1433         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1434                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1435         if (unlikely(count == 1))
1436                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1437         else
1438                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1439
1440         return tx;
1441 }
1442
1443 /* set_syndrome_sources - populate source buffers for gen_syndrome
1444  * @srcs - (struct page *) array of size sh->disks
1445  * @sh - stripe_head to parse
1446  *
1447  * Populates srcs in proper layout order for the stripe and returns the
1448  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1449  * destination buffer is recorded in srcs[count] and the Q destination
1450  * is recorded in srcs[count+1]].
1451  */
1452 static int set_syndrome_sources(struct page **srcs,
1453                                 struct stripe_head *sh,
1454                                 int srctype)
1455 {
1456         int disks = sh->disks;
1457         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1458         int d0_idx = raid6_d0(sh);
1459         int count;
1460         int i;
1461
1462         for (i = 0; i < disks; i++)
1463                 srcs[i] = NULL;
1464
1465         count = 0;
1466         i = d0_idx;
1467         do {
1468                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1469                 struct r5dev *dev = &sh->dev[i];
1470
1471                 if (i == sh->qd_idx || i == sh->pd_idx ||
1472                     (srctype == SYNDROME_SRC_ALL) ||
1473                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1474                      (test_bit(R5_Wantdrain, &dev->flags) ||
1475                       test_bit(R5_InJournal, &dev->flags))) ||
1476                     (srctype == SYNDROME_SRC_WRITTEN &&
1477                      (dev->written ||
1478                       test_bit(R5_InJournal, &dev->flags)))) {
1479                         if (test_bit(R5_InJournal, &dev->flags))
1480                                 srcs[slot] = sh->dev[i].orig_page;
1481                         else
1482                                 srcs[slot] = sh->dev[i].page;
1483                 }
1484                 i = raid6_next_disk(i, disks);
1485         } while (i != d0_idx);
1486
1487         return syndrome_disks;
1488 }
1489
1490 static struct dma_async_tx_descriptor *
1491 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1492 {
1493         int disks = sh->disks;
1494         struct page **blocks = to_addr_page(percpu, 0);
1495         int target;
1496         int qd_idx = sh->qd_idx;
1497         struct dma_async_tx_descriptor *tx;
1498         struct async_submit_ctl submit;
1499         struct r5dev *tgt;
1500         struct page *dest;
1501         int i;
1502         int count;
1503
1504         BUG_ON(sh->batch_head);
1505         if (sh->ops.target < 0)
1506                 target = sh->ops.target2;
1507         else if (sh->ops.target2 < 0)
1508                 target = sh->ops.target;
1509         else
1510                 /* we should only have one valid target */
1511                 BUG();
1512         BUG_ON(target < 0);
1513         pr_debug("%s: stripe %llu block: %d\n",
1514                 __func__, (unsigned long long)sh->sector, target);
1515
1516         tgt = &sh->dev[target];
1517         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1518         dest = tgt->page;
1519
1520         atomic_inc(&sh->count);
1521
1522         if (target == qd_idx) {
1523                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1524                 blocks[count] = NULL; /* regenerating p is not necessary */
1525                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1526                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1527                                   ops_complete_compute, sh,
1528                                   to_addr_conv(sh, percpu, 0));
1529                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1530         } else {
1531                 /* Compute any data- or p-drive using XOR */
1532                 count = 0;
1533                 for (i = disks; i-- ; ) {
1534                         if (i == target || i == qd_idx)
1535                                 continue;
1536                         blocks[count++] = sh->dev[i].page;
1537                 }
1538
1539                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1540                                   NULL, ops_complete_compute, sh,
1541                                   to_addr_conv(sh, percpu, 0));
1542                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1543         }
1544
1545         return tx;
1546 }
1547
1548 static struct dma_async_tx_descriptor *
1549 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1550 {
1551         int i, count, disks = sh->disks;
1552         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1553         int d0_idx = raid6_d0(sh);
1554         int faila = -1, failb = -1;
1555         int target = sh->ops.target;
1556         int target2 = sh->ops.target2;
1557         struct r5dev *tgt = &sh->dev[target];
1558         struct r5dev *tgt2 = &sh->dev[target2];
1559         struct dma_async_tx_descriptor *tx;
1560         struct page **blocks = to_addr_page(percpu, 0);
1561         struct async_submit_ctl submit;
1562
1563         BUG_ON(sh->batch_head);
1564         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1565                  __func__, (unsigned long long)sh->sector, target, target2);
1566         BUG_ON(target < 0 || target2 < 0);
1567         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1568         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1569
1570         /* we need to open-code set_syndrome_sources to handle the
1571          * slot number conversion for 'faila' and 'failb'
1572          */
1573         for (i = 0; i < disks ; i++)
1574                 blocks[i] = NULL;
1575         count = 0;
1576         i = d0_idx;
1577         do {
1578                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1579
1580                 blocks[slot] = sh->dev[i].page;
1581
1582                 if (i == target)
1583                         faila = slot;
1584                 if (i == target2)
1585                         failb = slot;
1586                 i = raid6_next_disk(i, disks);
1587         } while (i != d0_idx);
1588
1589         BUG_ON(faila == failb);
1590         if (failb < faila)
1591                 swap(faila, failb);
1592         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1593                  __func__, (unsigned long long)sh->sector, faila, failb);
1594
1595         atomic_inc(&sh->count);
1596
1597         if (failb == syndrome_disks+1) {
1598                 /* Q disk is one of the missing disks */
1599                 if (faila == syndrome_disks) {
1600                         /* Missing P+Q, just recompute */
1601                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1602                                           ops_complete_compute, sh,
1603                                           to_addr_conv(sh, percpu, 0));
1604                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1605                                                   STRIPE_SIZE, &submit);
1606                 } else {
1607                         struct page *dest;
1608                         int data_target;
1609                         int qd_idx = sh->qd_idx;
1610
1611                         /* Missing D+Q: recompute D from P, then recompute Q */
1612                         if (target == qd_idx)
1613                                 data_target = target2;
1614                         else
1615                                 data_target = target;
1616
1617                         count = 0;
1618                         for (i = disks; i-- ; ) {
1619                                 if (i == data_target || i == qd_idx)
1620                                         continue;
1621                                 blocks[count++] = sh->dev[i].page;
1622                         }
1623                         dest = sh->dev[data_target].page;
1624                         init_async_submit(&submit,
1625                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1626                                           NULL, NULL, NULL,
1627                                           to_addr_conv(sh, percpu, 0));
1628                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1629                                        &submit);
1630
1631                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1632                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1633                                           ops_complete_compute, sh,
1634                                           to_addr_conv(sh, percpu, 0));
1635                         return async_gen_syndrome(blocks, 0, count+2,
1636                                                   STRIPE_SIZE, &submit);
1637                 }
1638         } else {
1639                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1640                                   ops_complete_compute, sh,
1641                                   to_addr_conv(sh, percpu, 0));
1642                 if (failb == syndrome_disks) {
1643                         /* We're missing D+P. */
1644                         return async_raid6_datap_recov(syndrome_disks+2,
1645                                                        STRIPE_SIZE, faila,
1646                                                        blocks, &submit);
1647                 } else {
1648                         /* We're missing D+D. */
1649                         return async_raid6_2data_recov(syndrome_disks+2,
1650                                                        STRIPE_SIZE, faila, failb,
1651                                                        blocks, &submit);
1652                 }
1653         }
1654 }
1655
1656 static void ops_complete_prexor(void *stripe_head_ref)
1657 {
1658         struct stripe_head *sh = stripe_head_ref;
1659
1660         pr_debug("%s: stripe %llu\n", __func__,
1661                 (unsigned long long)sh->sector);
1662
1663         if (r5c_is_writeback(sh->raid_conf->log))
1664                 /*
1665                  * raid5-cache write back uses orig_page during prexor.
1666                  * After prexor, it is time to free orig_page
1667                  */
1668                 r5c_release_extra_page(sh);
1669 }
1670
1671 static struct dma_async_tx_descriptor *
1672 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1673                 struct dma_async_tx_descriptor *tx)
1674 {
1675         int disks = sh->disks;
1676         struct page **xor_srcs = to_addr_page(percpu, 0);
1677         int count = 0, pd_idx = sh->pd_idx, i;
1678         struct async_submit_ctl submit;
1679
1680         /* existing parity data subtracted */
1681         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1682
1683         BUG_ON(sh->batch_head);
1684         pr_debug("%s: stripe %llu\n", __func__,
1685                 (unsigned long long)sh->sector);
1686
1687         for (i = disks; i--; ) {
1688                 struct r5dev *dev = &sh->dev[i];
1689                 /* Only process blocks that are known to be uptodate */
1690                 if (test_bit(R5_InJournal, &dev->flags))
1691                         xor_srcs[count++] = dev->orig_page;
1692                 else if (test_bit(R5_Wantdrain, &dev->flags))
1693                         xor_srcs[count++] = dev->page;
1694         }
1695
1696         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1697                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1698         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1699
1700         return tx;
1701 }
1702
1703 static struct dma_async_tx_descriptor *
1704 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1705                 struct dma_async_tx_descriptor *tx)
1706 {
1707         struct page **blocks = to_addr_page(percpu, 0);
1708         int count;
1709         struct async_submit_ctl submit;
1710
1711         pr_debug("%s: stripe %llu\n", __func__,
1712                 (unsigned long long)sh->sector);
1713
1714         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1715
1716         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1717                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1718         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1719
1720         return tx;
1721 }
1722
1723 static struct dma_async_tx_descriptor *
1724 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1725 {
1726         struct r5conf *conf = sh->raid_conf;
1727         int disks = sh->disks;
1728         int i;
1729         struct stripe_head *head_sh = sh;
1730
1731         pr_debug("%s: stripe %llu\n", __func__,
1732                 (unsigned long long)sh->sector);
1733
1734         for (i = disks; i--; ) {
1735                 struct r5dev *dev;
1736                 struct bio *chosen;
1737
1738                 sh = head_sh;
1739                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1740                         struct bio *wbi;
1741
1742 again:
1743                         dev = &sh->dev[i];
1744                         /*
1745                          * clear R5_InJournal, so when rewriting a page in
1746                          * journal, it is not skipped by r5l_log_stripe()
1747                          */
1748                         clear_bit(R5_InJournal, &dev->flags);
1749                         spin_lock_irq(&sh->stripe_lock);
1750                         chosen = dev->towrite;
1751                         dev->towrite = NULL;
1752                         sh->overwrite_disks = 0;
1753                         BUG_ON(dev->written);
1754                         wbi = dev->written = chosen;
1755                         spin_unlock_irq(&sh->stripe_lock);
1756                         WARN_ON(dev->page != dev->orig_page);
1757
1758                         while (wbi && wbi->bi_iter.bi_sector <
1759                                 dev->sector + STRIPE_SECTORS) {
1760                                 if (wbi->bi_opf & REQ_FUA)
1761                                         set_bit(R5_WantFUA, &dev->flags);
1762                                 if (wbi->bi_opf & REQ_SYNC)
1763                                         set_bit(R5_SyncIO, &dev->flags);
1764                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1765                                         set_bit(R5_Discard, &dev->flags);
1766                                 else {
1767                                         tx = async_copy_data(1, wbi, &dev->page,
1768                                                              dev->sector, tx, sh,
1769                                                              r5c_is_writeback(conf->log));
1770                                         if (dev->page != dev->orig_page &&
1771                                             !r5c_is_writeback(conf->log)) {
1772                                                 set_bit(R5_SkipCopy, &dev->flags);
1773                                                 clear_bit(R5_UPTODATE, &dev->flags);
1774                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1775                                         }
1776                                 }
1777                                 wbi = r5_next_bio(wbi, dev->sector);
1778                         }
1779
1780                         if (head_sh->batch_head) {
1781                                 sh = list_first_entry(&sh->batch_list,
1782                                                       struct stripe_head,
1783                                                       batch_list);
1784                                 if (sh == head_sh)
1785                                         continue;
1786                                 goto again;
1787                         }
1788                 }
1789         }
1790
1791         return tx;
1792 }
1793
1794 static void ops_complete_reconstruct(void *stripe_head_ref)
1795 {
1796         struct stripe_head *sh = stripe_head_ref;
1797         int disks = sh->disks;
1798         int pd_idx = sh->pd_idx;
1799         int qd_idx = sh->qd_idx;
1800         int i;
1801         bool fua = false, sync = false, discard = false;
1802
1803         pr_debug("%s: stripe %llu\n", __func__,
1804                 (unsigned long long)sh->sector);
1805
1806         for (i = disks; i--; ) {
1807                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1808                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1809                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1810         }
1811
1812         for (i = disks; i--; ) {
1813                 struct r5dev *dev = &sh->dev[i];
1814
1815                 if (dev->written || i == pd_idx || i == qd_idx) {
1816                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1817                                 set_bit(R5_UPTODATE, &dev->flags);
1818                         if (fua)
1819                                 set_bit(R5_WantFUA, &dev->flags);
1820                         if (sync)
1821                                 set_bit(R5_SyncIO, &dev->flags);
1822                 }
1823         }
1824
1825         if (sh->reconstruct_state == reconstruct_state_drain_run)
1826                 sh->reconstruct_state = reconstruct_state_drain_result;
1827         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1828                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1829         else {
1830                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1831                 sh->reconstruct_state = reconstruct_state_result;
1832         }
1833
1834         set_bit(STRIPE_HANDLE, &sh->state);
1835         raid5_release_stripe(sh);
1836 }
1837
1838 static void
1839 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1840                      struct dma_async_tx_descriptor *tx)
1841 {
1842         int disks = sh->disks;
1843         struct page **xor_srcs;
1844         struct async_submit_ctl submit;
1845         int count, pd_idx = sh->pd_idx, i;
1846         struct page *xor_dest;
1847         int prexor = 0;
1848         unsigned long flags;
1849         int j = 0;
1850         struct stripe_head *head_sh = sh;
1851         int last_stripe;
1852
1853         pr_debug("%s: stripe %llu\n", __func__,
1854                 (unsigned long long)sh->sector);
1855
1856         for (i = 0; i < sh->disks; i++) {
1857                 if (pd_idx == i)
1858                         continue;
1859                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1860                         break;
1861         }
1862         if (i >= sh->disks) {
1863                 atomic_inc(&sh->count);
1864                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1865                 ops_complete_reconstruct(sh);
1866                 return;
1867         }
1868 again:
1869         count = 0;
1870         xor_srcs = to_addr_page(percpu, j);
1871         /* check if prexor is active which means only process blocks
1872          * that are part of a read-modify-write (written)
1873          */
1874         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1875                 prexor = 1;
1876                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1877                 for (i = disks; i--; ) {
1878                         struct r5dev *dev = &sh->dev[i];
1879                         if (head_sh->dev[i].written ||
1880                             test_bit(R5_InJournal, &head_sh->dev[i].flags))
1881                                 xor_srcs[count++] = dev->page;
1882                 }
1883         } else {
1884                 xor_dest = sh->dev[pd_idx].page;
1885                 for (i = disks; i--; ) {
1886                         struct r5dev *dev = &sh->dev[i];
1887                         if (i != pd_idx)
1888                                 xor_srcs[count++] = dev->page;
1889                 }
1890         }
1891
1892         /* 1/ if we prexor'd then the dest is reused as a source
1893          * 2/ if we did not prexor then we are redoing the parity
1894          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1895          * for the synchronous xor case
1896          */
1897         last_stripe = !head_sh->batch_head ||
1898                 list_first_entry(&sh->batch_list,
1899                                  struct stripe_head, batch_list) == head_sh;
1900         if (last_stripe) {
1901                 flags = ASYNC_TX_ACK |
1902                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1903
1904                 atomic_inc(&head_sh->count);
1905                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1906                                   to_addr_conv(sh, percpu, j));
1907         } else {
1908                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1909                 init_async_submit(&submit, flags, tx, NULL, NULL,
1910                                   to_addr_conv(sh, percpu, j));
1911         }
1912
1913         if (unlikely(count == 1))
1914                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1915         else
1916                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1917         if (!last_stripe) {
1918                 j++;
1919                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1920                                       batch_list);
1921                 goto again;
1922         }
1923 }
1924
1925 static void
1926 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1927                      struct dma_async_tx_descriptor *tx)
1928 {
1929         struct async_submit_ctl submit;
1930         struct page **blocks;
1931         int count, i, j = 0;
1932         struct stripe_head *head_sh = sh;
1933         int last_stripe;
1934         int synflags;
1935         unsigned long txflags;
1936
1937         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1938
1939         for (i = 0; i < sh->disks; i++) {
1940                 if (sh->pd_idx == i || sh->qd_idx == i)
1941                         continue;
1942                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1943                         break;
1944         }
1945         if (i >= sh->disks) {
1946                 atomic_inc(&sh->count);
1947                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1948                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1949                 ops_complete_reconstruct(sh);
1950                 return;
1951         }
1952
1953 again:
1954         blocks = to_addr_page(percpu, j);
1955
1956         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1957                 synflags = SYNDROME_SRC_WRITTEN;
1958                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1959         } else {
1960                 synflags = SYNDROME_SRC_ALL;
1961                 txflags = ASYNC_TX_ACK;
1962         }
1963
1964         count = set_syndrome_sources(blocks, sh, synflags);
1965         last_stripe = !head_sh->batch_head ||
1966                 list_first_entry(&sh->batch_list,
1967                                  struct stripe_head, batch_list) == head_sh;
1968
1969         if (last_stripe) {
1970                 atomic_inc(&head_sh->count);
1971                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1972                                   head_sh, to_addr_conv(sh, percpu, j));
1973         } else
1974                 init_async_submit(&submit, 0, tx, NULL, NULL,
1975                                   to_addr_conv(sh, percpu, j));
1976         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1977         if (!last_stripe) {
1978                 j++;
1979                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1980                                       batch_list);
1981                 goto again;
1982         }
1983 }
1984
1985 static void ops_complete_check(void *stripe_head_ref)
1986 {
1987         struct stripe_head *sh = stripe_head_ref;
1988
1989         pr_debug("%s: stripe %llu\n", __func__,
1990                 (unsigned long long)sh->sector);
1991
1992         sh->check_state = check_state_check_result;
1993         set_bit(STRIPE_HANDLE, &sh->state);
1994         raid5_release_stripe(sh);
1995 }
1996
1997 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1998 {
1999         int disks = sh->disks;
2000         int pd_idx = sh->pd_idx;
2001         int qd_idx = sh->qd_idx;
2002         struct page *xor_dest;
2003         struct page **xor_srcs = to_addr_page(percpu, 0);
2004         struct dma_async_tx_descriptor *tx;
2005         struct async_submit_ctl submit;
2006         int count;
2007         int i;
2008
2009         pr_debug("%s: stripe %llu\n", __func__,
2010                 (unsigned long long)sh->sector);
2011
2012         BUG_ON(sh->batch_head);
2013         count = 0;
2014         xor_dest = sh->dev[pd_idx].page;
2015         xor_srcs[count++] = xor_dest;
2016         for (i = disks; i--; ) {
2017                 if (i == pd_idx || i == qd_idx)
2018                         continue;
2019                 xor_srcs[count++] = sh->dev[i].page;
2020         }
2021
2022         init_async_submit(&submit, 0, NULL, NULL, NULL,
2023                           to_addr_conv(sh, percpu, 0));
2024         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2025                            &sh->ops.zero_sum_result, &submit);
2026
2027         atomic_inc(&sh->count);
2028         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2029         tx = async_trigger_callback(&submit);
2030 }
2031
2032 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2033 {
2034         struct page **srcs = to_addr_page(percpu, 0);
2035         struct async_submit_ctl submit;
2036         int count;
2037
2038         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2039                 (unsigned long long)sh->sector, checkp);
2040
2041         BUG_ON(sh->batch_head);
2042         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2043         if (!checkp)
2044                 srcs[count] = NULL;
2045
2046         atomic_inc(&sh->count);
2047         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2048                           sh, to_addr_conv(sh, percpu, 0));
2049         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2050                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2051 }
2052
2053 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2054 {
2055         int overlap_clear = 0, i, disks = sh->disks;
2056         struct dma_async_tx_descriptor *tx = NULL;
2057         struct r5conf *conf = sh->raid_conf;
2058         int level = conf->level;
2059         struct raid5_percpu *percpu;
2060         unsigned long cpu;
2061
2062         cpu = get_cpu();
2063         percpu = per_cpu_ptr(conf->percpu, cpu);
2064         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2065                 ops_run_biofill(sh);
2066                 overlap_clear++;
2067         }
2068
2069         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2070                 if (level < 6)
2071                         tx = ops_run_compute5(sh, percpu);
2072                 else {
2073                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
2074                                 tx = ops_run_compute6_1(sh, percpu);
2075                         else
2076                                 tx = ops_run_compute6_2(sh, percpu);
2077                 }
2078                 /* terminate the chain if reconstruct is not set to be run */
2079                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2080                         async_tx_ack(tx);
2081         }
2082
2083         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2084                 if (level < 6)
2085                         tx = ops_run_prexor5(sh, percpu, tx);
2086                 else
2087                         tx = ops_run_prexor6(sh, percpu, tx);
2088         }
2089
2090         if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2091                 tx = ops_run_partial_parity(sh, percpu, tx);
2092
2093         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2094                 tx = ops_run_biodrain(sh, tx);
2095                 overlap_clear++;
2096         }
2097
2098         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2099                 if (level < 6)
2100                         ops_run_reconstruct5(sh, percpu, tx);
2101                 else
2102                         ops_run_reconstruct6(sh, percpu, tx);
2103         }
2104
2105         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2106                 if (sh->check_state == check_state_run)
2107                         ops_run_check_p(sh, percpu);
2108                 else if (sh->check_state == check_state_run_q)
2109                         ops_run_check_pq(sh, percpu, 0);
2110                 else if (sh->check_state == check_state_run_pq)
2111                         ops_run_check_pq(sh, percpu, 1);
2112                 else
2113                         BUG();
2114         }
2115
2116         if (overlap_clear && !sh->batch_head)
2117                 for (i = disks; i--; ) {
2118                         struct r5dev *dev = &sh->dev[i];
2119                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2120                                 wake_up(&sh->raid_conf->wait_for_overlap);
2121                 }
2122         put_cpu();
2123 }
2124
2125 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2126 {
2127         if (sh->ppl_page)
2128                 __free_page(sh->ppl_page);
2129         kmem_cache_free(sc, sh);
2130 }
2131
2132 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2133         int disks, struct r5conf *conf)
2134 {
2135         struct stripe_head *sh;
2136         int i;
2137
2138         sh = kmem_cache_zalloc(sc, gfp);
2139         if (sh) {
2140                 spin_lock_init(&sh->stripe_lock);
2141                 spin_lock_init(&sh->batch_lock);
2142                 INIT_LIST_HEAD(&sh->batch_list);
2143                 INIT_LIST_HEAD(&sh->lru);
2144                 INIT_LIST_HEAD(&sh->r5c);
2145                 INIT_LIST_HEAD(&sh->log_list);
2146                 atomic_set(&sh->count, 1);
2147                 sh->raid_conf = conf;
2148                 sh->log_start = MaxSector;
2149                 for (i = 0; i < disks; i++) {
2150                         struct r5dev *dev = &sh->dev[i];
2151
2152                         bio_init(&dev->req, &dev->vec, 1);
2153                         bio_init(&dev->rreq, &dev->rvec, 1);
2154                 }
2155
2156                 if (raid5_has_ppl(conf)) {
2157                         sh->ppl_page = alloc_page(gfp);
2158                         if (!sh->ppl_page) {
2159                                 free_stripe(sc, sh);
2160                                 sh = NULL;
2161                         }
2162                 }
2163         }
2164         return sh;
2165 }
2166 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2167 {
2168         struct stripe_head *sh;
2169
2170         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2171         if (!sh)
2172                 return 0;
2173
2174         if (grow_buffers(sh, gfp)) {
2175                 shrink_buffers(sh);
2176                 free_stripe(conf->slab_cache, sh);
2177                 return 0;
2178         }
2179         sh->hash_lock_index =
2180                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2181         /* we just created an active stripe so... */
2182         atomic_inc(&conf->active_stripes);
2183
2184         raid5_release_stripe(sh);
2185         conf->max_nr_stripes++;
2186         return 1;
2187 }
2188
2189 static int grow_stripes(struct r5conf *conf, int num)
2190 {
2191         struct kmem_cache *sc;
2192         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2193
2194         if (conf->mddev->gendisk)
2195                 sprintf(conf->cache_name[0],
2196                         "raid%d-%s", conf->level, mdname(conf->mddev));
2197         else
2198                 sprintf(conf->cache_name[0],
2199                         "raid%d-%p", conf->level, conf->mddev);
2200         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2201
2202         conf->active_name = 0;
2203         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2204                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2205                                0, 0, NULL);
2206         if (!sc)
2207                 return 1;
2208         conf->slab_cache = sc;
2209         conf->pool_size = devs;
2210         while (num--)
2211                 if (!grow_one_stripe(conf, GFP_KERNEL))
2212                         return 1;
2213
2214         return 0;
2215 }
2216
2217 /**
2218  * scribble_len - return the required size of the scribble region
2219  * @num - total number of disks in the array
2220  *
2221  * The size must be enough to contain:
2222  * 1/ a struct page pointer for each device in the array +2
2223  * 2/ room to convert each entry in (1) to its corresponding dma
2224  *    (dma_map_page()) or page (page_address()) address.
2225  *
2226  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2227  * calculate over all devices (not just the data blocks), using zeros in place
2228  * of the P and Q blocks.
2229  */
2230 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2231 {
2232         struct flex_array *ret;
2233         size_t len;
2234
2235         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2236         ret = flex_array_alloc(len, cnt, flags);
2237         if (!ret)
2238                 return NULL;
2239         /* always prealloc all elements, so no locking is required */
2240         if (flex_array_prealloc(ret, 0, cnt, flags)) {
2241                 flex_array_free(ret);
2242                 return NULL;
2243         }
2244         return ret;
2245 }
2246
2247 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2248 {
2249         unsigned long cpu;
2250         int err = 0;
2251
2252         /*
2253          * Never shrink. And mddev_suspend() could deadlock if this is called
2254          * from raid5d. In that case, scribble_disks and scribble_sectors
2255          * should equal to new_disks and new_sectors
2256          */
2257         if (conf->scribble_disks >= new_disks &&
2258             conf->scribble_sectors >= new_sectors)
2259                 return 0;
2260         mddev_suspend(conf->mddev);
2261         get_online_cpus();
2262         for_each_present_cpu(cpu) {
2263                 struct raid5_percpu *percpu;
2264                 struct flex_array *scribble;
2265
2266                 percpu = per_cpu_ptr(conf->percpu, cpu);
2267                 scribble = scribble_alloc(new_disks,
2268                                           new_sectors / STRIPE_SECTORS,
2269                                           GFP_NOIO);
2270
2271                 if (scribble) {
2272                         flex_array_free(percpu->scribble);
2273                         percpu->scribble = scribble;
2274                 } else {
2275                         err = -ENOMEM;
2276                         break;
2277                 }
2278         }
2279         put_online_cpus();
2280         mddev_resume(conf->mddev);
2281         if (!err) {
2282                 conf->scribble_disks = new_disks;
2283                 conf->scribble_sectors = new_sectors;
2284         }
2285         return err;
2286 }
2287
2288 static int resize_stripes(struct r5conf *conf, int newsize)
2289 {
2290         /* Make all the stripes able to hold 'newsize' devices.
2291          * New slots in each stripe get 'page' set to a new page.
2292          *
2293          * This happens in stages:
2294          * 1/ create a new kmem_cache and allocate the required number of
2295          *    stripe_heads.
2296          * 2/ gather all the old stripe_heads and transfer the pages across
2297          *    to the new stripe_heads.  This will have the side effect of
2298          *    freezing the array as once all stripe_heads have been collected,
2299          *    no IO will be possible.  Old stripe heads are freed once their
2300          *    pages have been transferred over, and the old kmem_cache is
2301          *    freed when all stripes are done.
2302          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2303          *    we simple return a failure status - no need to clean anything up.
2304          * 4/ allocate new pages for the new slots in the new stripe_heads.
2305          *    If this fails, we don't bother trying the shrink the
2306          *    stripe_heads down again, we just leave them as they are.
2307          *    As each stripe_head is processed the new one is released into
2308          *    active service.
2309          *
2310          * Once step2 is started, we cannot afford to wait for a write,
2311          * so we use GFP_NOIO allocations.
2312          */
2313         struct stripe_head *osh, *nsh;
2314         LIST_HEAD(newstripes);
2315         struct disk_info *ndisks;
2316         int err = 0;
2317         struct kmem_cache *sc;
2318         int i;
2319         int hash, cnt;
2320
2321         md_allow_write(conf->mddev);
2322
2323         /* Step 1 */
2324         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2325                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2326                                0, 0, NULL);
2327         if (!sc)
2328                 return -ENOMEM;
2329
2330         /* Need to ensure auto-resizing doesn't interfere */
2331         mutex_lock(&conf->cache_size_mutex);
2332
2333         for (i = conf->max_nr_stripes; i; i--) {
2334                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2335                 if (!nsh)
2336                         break;
2337
2338                 list_add(&nsh->lru, &newstripes);
2339         }
2340         if (i) {
2341                 /* didn't get enough, give up */
2342                 while (!list_empty(&newstripes)) {
2343                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2344                         list_del(&nsh->lru);
2345                         free_stripe(sc, nsh);
2346                 }
2347                 kmem_cache_destroy(sc);
2348                 mutex_unlock(&conf->cache_size_mutex);
2349                 return -ENOMEM;
2350         }
2351         /* Step 2 - Must use GFP_NOIO now.
2352          * OK, we have enough stripes, start collecting inactive
2353          * stripes and copying them over
2354          */
2355         hash = 0;
2356         cnt = 0;
2357         list_for_each_entry(nsh, &newstripes, lru) {
2358                 lock_device_hash_lock(conf, hash);
2359                 wait_event_cmd(conf->wait_for_stripe,
2360                                     !list_empty(conf->inactive_list + hash),
2361                                     unlock_device_hash_lock(conf, hash),
2362                                     lock_device_hash_lock(conf, hash));
2363                 osh = get_free_stripe(conf, hash);
2364                 unlock_device_hash_lock(conf, hash);
2365
2366                 for(i=0; i<conf->pool_size; i++) {
2367                         nsh->dev[i].page = osh->dev[i].page;
2368                         nsh->dev[i].orig_page = osh->dev[i].page;
2369                 }
2370                 nsh->hash_lock_index = hash;
2371                 free_stripe(conf->slab_cache, osh);
2372                 cnt++;
2373                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2374                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2375                         hash++;
2376                         cnt = 0;
2377                 }
2378         }
2379         kmem_cache_destroy(conf->slab_cache);
2380
2381         /* Step 3.
2382          * At this point, we are holding all the stripes so the array
2383          * is completely stalled, so now is a good time to resize
2384          * conf->disks and the scribble region
2385          */
2386         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2387         if (ndisks) {
2388                 for (i = 0; i < conf->pool_size; i++)
2389                         ndisks[i] = conf->disks[i];
2390
2391                 for (i = conf->pool_size; i < newsize; i++) {
2392                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2393                         if (!ndisks[i].extra_page)
2394                                 err = -ENOMEM;
2395                 }
2396
2397                 if (err) {
2398                         for (i = conf->pool_size; i < newsize; i++)
2399                                 if (ndisks[i].extra_page)
2400                                         put_page(ndisks[i].extra_page);
2401                         kfree(ndisks);
2402                 } else {
2403                         kfree(conf->disks);
2404                         conf->disks = ndisks;
2405                 }
2406         } else
2407                 err = -ENOMEM;
2408
2409         mutex_unlock(&conf->cache_size_mutex);
2410
2411         conf->slab_cache = sc;
2412         conf->active_name = 1-conf->active_name;
2413
2414         /* Step 4, return new stripes to service */
2415         while(!list_empty(&newstripes)) {
2416                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2417                 list_del_init(&nsh->lru);
2418
2419                 for (i=conf->raid_disks; i < newsize; i++)
2420                         if (nsh->dev[i].page == NULL) {
2421                                 struct page *p = alloc_page(GFP_NOIO);
2422                                 nsh->dev[i].page = p;
2423                                 nsh->dev[i].orig_page = p;
2424                                 if (!p)
2425                                         err = -ENOMEM;
2426                         }
2427                 raid5_release_stripe(nsh);
2428         }
2429         /* critical section pass, GFP_NOIO no longer needed */
2430
2431         if (!err)
2432                 conf->pool_size = newsize;
2433         return err;
2434 }
2435
2436 static int drop_one_stripe(struct r5conf *conf)
2437 {
2438         struct stripe_head *sh;
2439         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2440
2441         spin_lock_irq(conf->hash_locks + hash);
2442         sh = get_free_stripe(conf, hash);
2443         spin_unlock_irq(conf->hash_locks + hash);
2444         if (!sh)
2445                 return 0;
2446         BUG_ON(atomic_read(&sh->count));
2447         shrink_buffers(sh);
2448         free_stripe(conf->slab_cache, sh);
2449         atomic_dec(&conf->active_stripes);
2450         conf->max_nr_stripes--;
2451         return 1;
2452 }
2453
2454 static void shrink_stripes(struct r5conf *conf)
2455 {
2456         while (conf->max_nr_stripes &&
2457                drop_one_stripe(conf))
2458                 ;
2459
2460         kmem_cache_destroy(conf->slab_cache);
2461         conf->slab_cache = NULL;
2462 }
2463
2464 static void raid5_end_read_request(struct bio * bi)
2465 {
2466         struct stripe_head *sh = bi->bi_private;
2467         struct r5conf *conf = sh->raid_conf;
2468         int disks = sh->disks, i;
2469         char b[BDEVNAME_SIZE];
2470         struct md_rdev *rdev = NULL;
2471         sector_t s;
2472
2473         for (i=0 ; i<disks; i++)
2474                 if (bi == &sh->dev[i].req)
2475                         break;
2476
2477         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2478                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2479                 bi->bi_error);
2480         if (i == disks) {
2481                 bio_reset(bi);
2482                 BUG();
2483                 return;
2484         }
2485         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2486                 /* If replacement finished while this request was outstanding,
2487                  * 'replacement' might be NULL already.
2488                  * In that case it moved down to 'rdev'.
2489                  * rdev is not removed until all requests are finished.
2490                  */
2491                 rdev = conf->disks[i].replacement;
2492         if (!rdev)
2493                 rdev = conf->disks[i].rdev;
2494
2495         if (use_new_offset(conf, sh))
2496                 s = sh->sector + rdev->new_data_offset;
2497         else
2498                 s = sh->sector + rdev->data_offset;
2499         if (!bi->bi_error) {
2500                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2501                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2502                         /* Note that this cannot happen on a
2503                          * replacement device.  We just fail those on
2504                          * any error
2505                          */
2506                         pr_info_ratelimited(
2507                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2508                                 mdname(conf->mddev), STRIPE_SECTORS,
2509                                 (unsigned long long)s,
2510                                 bdevname(rdev->bdev, b));
2511                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2512                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2513                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2514                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2515                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2516
2517                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2518                         /*
2519                          * end read for a page in journal, this
2520                          * must be preparing for prexor in rmw
2521                          */
2522                         set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2523
2524                 if (atomic_read(&rdev->read_errors))
2525                         atomic_set(&rdev->read_errors, 0);
2526         } else {
2527                 const char *bdn = bdevname(rdev->bdev, b);
2528                 int retry = 0;
2529                 int set_bad = 0;
2530
2531                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2532                 atomic_inc(&rdev->read_errors);
2533                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2534                         pr_warn_ratelimited(
2535                                 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2536                                 mdname(conf->mddev),
2537                                 (unsigned long long)s,
2538                                 bdn);
2539                 else if (conf->mddev->degraded >= conf->max_degraded) {
2540                         set_bad = 1;
2541                         pr_warn_ratelimited(
2542                                 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2543                                 mdname(conf->mddev),
2544                                 (unsigned long long)s,
2545                                 bdn);
2546                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2547                         /* Oh, no!!! */
2548                         set_bad = 1;
2549                         pr_warn_ratelimited(
2550                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2551                                 mdname(conf->mddev),
2552                                 (unsigned long long)s,
2553                                 bdn);
2554                 } else if (atomic_read(&rdev->read_errors)
2555                          > conf->max_nr_stripes)
2556                         pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2557                                mdname(conf->mddev), bdn);
2558                 else
2559                         retry = 1;
2560                 if (set_bad && test_bit(In_sync, &rdev->flags)
2561                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2562                         retry = 1;
2563                 if (retry)
2564                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2565                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2566                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2567                         } else
2568                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2569                 else {
2570                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2571                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2572                         if (!(set_bad
2573                               && test_bit(In_sync, &rdev->flags)
2574                               && rdev_set_badblocks(
2575                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2576                                 md_error(conf->mddev, rdev);
2577                 }
2578         }
2579         rdev_dec_pending(rdev, conf->mddev);
2580         bio_reset(bi);
2581         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2582         set_bit(STRIPE_HANDLE, &sh->state);
2583         raid5_release_stripe(sh);
2584 }
2585
2586 static void raid5_end_write_request(struct bio *bi)
2587 {
2588         struct stripe_head *sh = bi->bi_private;
2589         struct r5conf *conf = sh->raid_conf;
2590         int disks = sh->disks, i;
2591         struct md_rdev *uninitialized_var(rdev);
2592         sector_t first_bad;
2593         int bad_sectors;
2594         int replacement = 0;
2595
2596         for (i = 0 ; i < disks; i++) {
2597                 if (bi == &sh->dev[i].req) {
2598                         rdev = conf->disks[i].rdev;
2599                         break;
2600                 }
2601                 if (bi == &sh->dev[i].rreq) {
2602                         rdev = conf->disks[i].replacement;
2603                         if (rdev)
2604                                 replacement = 1;
2605                         else
2606                                 /* rdev was removed and 'replacement'
2607                                  * replaced it.  rdev is not removed
2608                                  * until all requests are finished.
2609                                  */
2610                                 rdev = conf->disks[i].rdev;
2611                         break;
2612                 }
2613         }
2614         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2615                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2616                 bi->bi_error);
2617         if (i == disks) {
2618                 bio_reset(bi);
2619                 BUG();
2620                 return;
2621         }
2622
2623         if (replacement) {
2624                 if (bi->bi_error)
2625                         md_error(conf->mddev, rdev);
2626                 else if (is_badblock(rdev, sh->sector,
2627                                      STRIPE_SECTORS,
2628                                      &first_bad, &bad_sectors))
2629                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2630         } else {
2631                 if (bi->bi_error) {
2632                         set_bit(STRIPE_DEGRADED, &sh->state);
2633                         set_bit(WriteErrorSeen, &rdev->flags);
2634                         set_bit(R5_WriteError, &sh->dev[i].flags);
2635                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2636                                 set_bit(MD_RECOVERY_NEEDED,
2637                                         &rdev->mddev->recovery);
2638                 } else if (is_badblock(rdev, sh->sector,
2639                                        STRIPE_SECTORS,
2640                                        &first_bad, &bad_sectors)) {
2641                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2642                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2643                                 /* That was a successful write so make
2644                                  * sure it looks like we already did
2645                                  * a re-write.
2646                                  */
2647                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2648                 }
2649         }
2650         rdev_dec_pending(rdev, conf->mddev);
2651
2652         if (sh->batch_head && bi->bi_error && !replacement)
2653                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2654
2655         bio_reset(bi);
2656         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2657                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2658         set_bit(STRIPE_HANDLE, &sh->state);
2659         raid5_release_stripe(sh);
2660
2661         if (sh->batch_head && sh != sh->batch_head)
2662                 raid5_release_stripe(sh->batch_head);
2663 }
2664
2665 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2666 {
2667         struct r5dev *dev = &sh->dev[i];
2668
2669         dev->flags = 0;
2670         dev->sector = raid5_compute_blocknr(sh, i, previous);
2671 }
2672
2673 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2674 {
2675         char b[BDEVNAME_SIZE];
2676         struct r5conf *conf = mddev->private;
2677         unsigned long flags;
2678         pr_debug("raid456: error called\n");
2679
2680         spin_lock_irqsave(&conf->device_lock, flags);
2681         clear_bit(In_sync, &rdev->flags);
2682         mddev->degraded = raid5_calc_degraded(conf);
2683         spin_unlock_irqrestore(&conf->device_lock, flags);
2684         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2685
2686         set_bit(Blocked, &rdev->flags);
2687         set_bit(Faulty, &rdev->flags);
2688         set_mask_bits(&mddev->sb_flags, 0,
2689                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2690         pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2691                 "md/raid:%s: Operation continuing on %d devices.\n",
2692                 mdname(mddev),
2693                 bdevname(rdev->bdev, b),
2694                 mdname(mddev),
2695                 conf->raid_disks - mddev->degraded);
2696         r5c_update_on_rdev_error(mddev, rdev);
2697 }
2698
2699 /*
2700  * Input: a 'big' sector number,
2701  * Output: index of the data and parity disk, and the sector # in them.
2702  */
2703 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2704                               int previous, int *dd_idx,
2705                               struct stripe_head *sh)
2706 {
2707         sector_t stripe, stripe2;
2708         sector_t chunk_number;
2709         unsigned int chunk_offset;
2710         int pd_idx, qd_idx;
2711         int ddf_layout = 0;
2712         sector_t new_sector;
2713         int algorithm = previous ? conf->prev_algo
2714                                  : conf->algorithm;
2715         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2716                                          : conf->chunk_sectors;
2717         int raid_disks = previous ? conf->previous_raid_disks
2718                                   : conf->raid_disks;
2719         int data_disks = raid_disks - conf->max_degraded;
2720
2721         /* First compute the information on this sector */
2722
2723         /*
2724          * Compute the chunk number and the sector offset inside the chunk
2725          */
2726         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2727         chunk_number = r_sector;
2728
2729         /*
2730          * Compute the stripe number
2731          */
2732         stripe = chunk_number;
2733         *dd_idx = sector_div(stripe, data_disks);
2734         stripe2 = stripe;
2735         /*
2736          * Select the parity disk based on the user selected algorithm.
2737          */
2738         pd_idx = qd_idx = -1;
2739         switch(conf->level) {
2740         case 4:
2741                 pd_idx = data_disks;
2742                 break;
2743         case 5:
2744                 switch (algorithm) {
2745                 case ALGORITHM_LEFT_ASYMMETRIC:
2746                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2747                         if (*dd_idx >= pd_idx)
2748                                 (*dd_idx)++;
2749                         break;
2750                 case ALGORITHM_RIGHT_ASYMMETRIC:
2751                         pd_idx = sector_div(stripe2, raid_disks);
2752                         if (*dd_idx >= pd_idx)
2753                                 (*dd_idx)++;
2754                         break;
2755                 case ALGORITHM_LEFT_SYMMETRIC:
2756                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2757                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2758                         break;
2759                 case ALGORITHM_RIGHT_SYMMETRIC:
2760                         pd_idx = sector_div(stripe2, raid_disks);
2761                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2762                         break;
2763                 case ALGORITHM_PARITY_0:
2764                         pd_idx = 0;
2765                         (*dd_idx)++;
2766                         break;
2767                 case ALGORITHM_PARITY_N:
2768                         pd_idx = data_disks;
2769                         break;
2770                 default:
2771                         BUG();
2772                 }
2773                 break;
2774         case 6:
2775
2776                 switch (algorithm) {
2777                 case ALGORITHM_LEFT_ASYMMETRIC:
2778                         pd_idx = raid_disks - 1 - 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_RIGHT_ASYMMETRIC:
2787                         pd_idx = sector_div(stripe2, raid_disks);
2788                         qd_idx = pd_idx + 1;
2789                         if (pd_idx == raid_disks-1) {
2790                                 (*dd_idx)++;    /* Q D D D P */
2791                                 qd_idx = 0;
2792                         } else if (*dd_idx >= pd_idx)
2793                                 (*dd_idx) += 2; /* D D P Q D */
2794                         break;
2795                 case ALGORITHM_LEFT_SYMMETRIC:
2796                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2797                         qd_idx = (pd_idx + 1) % raid_disks;
2798                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2799                         break;
2800                 case ALGORITHM_RIGHT_SYMMETRIC:
2801                         pd_idx = sector_div(stripe2, raid_disks);
2802                         qd_idx = (pd_idx + 1) % raid_disks;
2803                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2804                         break;
2805
2806                 case ALGORITHM_PARITY_0:
2807                         pd_idx = 0;
2808                         qd_idx = 1;
2809                         (*dd_idx) += 2;
2810                         break;
2811                 case ALGORITHM_PARITY_N:
2812                         pd_idx = data_disks;
2813                         qd_idx = data_disks + 1;
2814                         break;
2815
2816                 case ALGORITHM_ROTATING_ZERO_RESTART:
2817                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2818                          * of blocks for computing Q is different.
2819                          */
2820                         pd_idx = sector_div(stripe2, raid_disks);
2821                         qd_idx = pd_idx + 1;
2822                         if (pd_idx == raid_disks-1) {
2823                                 (*dd_idx)++;    /* Q D D D P */
2824                                 qd_idx = 0;
2825                         } else if (*dd_idx >= pd_idx)
2826                                 (*dd_idx) += 2; /* D D P Q D */
2827                         ddf_layout = 1;
2828                         break;
2829
2830                 case ALGORITHM_ROTATING_N_RESTART:
2831                         /* Same a left_asymmetric, by first stripe is
2832                          * D D D P Q  rather than
2833                          * Q D D D P
2834                          */
2835                         stripe2 += 1;
2836                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2837                         qd_idx = pd_idx + 1;
2838                         if (pd_idx == raid_disks-1) {
2839                                 (*dd_idx)++;    /* Q D D D P */
2840                                 qd_idx = 0;
2841                         } else if (*dd_idx >= pd_idx)
2842                                 (*dd_idx) += 2; /* D D P Q D */
2843                         ddf_layout = 1;
2844                         break;
2845
2846                 case ALGORITHM_ROTATING_N_CONTINUE:
2847                         /* Same as left_symmetric but Q is before P */
2848                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2849                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2850                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2851                         ddf_layout = 1;
2852                         break;
2853
2854                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2855                         /* RAID5 left_asymmetric, with Q on last device */
2856                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2857                         if (*dd_idx >= pd_idx)
2858                                 (*dd_idx)++;
2859                         qd_idx = raid_disks - 1;
2860                         break;
2861
2862                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2863                         pd_idx = sector_div(stripe2, raid_disks-1);
2864                         if (*dd_idx >= pd_idx)
2865                                 (*dd_idx)++;
2866                         qd_idx = raid_disks - 1;
2867                         break;
2868
2869                 case ALGORITHM_LEFT_SYMMETRIC_6:
2870                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2871                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2872                         qd_idx = raid_disks - 1;
2873                         break;
2874
2875                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2876                         pd_idx = sector_div(stripe2, raid_disks-1);
2877                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2878                         qd_idx = raid_disks - 1;
2879                         break;
2880
2881                 case ALGORITHM_PARITY_0_6:
2882                         pd_idx = 0;
2883                         (*dd_idx)++;
2884                         qd_idx = raid_disks - 1;
2885                         break;
2886
2887                 default:
2888                         BUG();
2889                 }
2890                 break;
2891         }
2892
2893         if (sh) {
2894                 sh->pd_idx = pd_idx;
2895                 sh->qd_idx = qd_idx;
2896                 sh->ddf_layout = ddf_layout;
2897         }
2898         /*
2899          * Finally, compute the new sector number
2900          */
2901         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2902         return new_sector;
2903 }
2904
2905 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2906 {
2907         struct r5conf *conf = sh->raid_conf;
2908         int raid_disks = sh->disks;
2909         int data_disks = raid_disks - conf->max_degraded;
2910         sector_t new_sector = sh->sector, check;
2911         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2912                                          : conf->chunk_sectors;
2913         int algorithm = previous ? conf->prev_algo
2914                                  : conf->algorithm;
2915         sector_t stripe;
2916         int chunk_offset;
2917         sector_t chunk_number;
2918         int dummy1, dd_idx = i;
2919         sector_t r_sector;
2920         struct stripe_head sh2;
2921
2922         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2923         stripe = new_sector;
2924
2925         if (i == sh->pd_idx)
2926                 return 0;
2927         switch(conf->level) {
2928         case 4: break;
2929         case 5:
2930                 switch (algorithm) {
2931                 case ALGORITHM_LEFT_ASYMMETRIC:
2932                 case ALGORITHM_RIGHT_ASYMMETRIC:
2933                         if (i > sh->pd_idx)
2934                                 i--;
2935                         break;
2936                 case ALGORITHM_LEFT_SYMMETRIC:
2937                 case ALGORITHM_RIGHT_SYMMETRIC:
2938                         if (i < sh->pd_idx)
2939                                 i += raid_disks;
2940                         i -= (sh->pd_idx + 1);
2941                         break;
2942                 case ALGORITHM_PARITY_0:
2943                         i -= 1;
2944                         break;
2945                 case ALGORITHM_PARITY_N:
2946                         break;
2947                 default:
2948                         BUG();
2949                 }
2950                 break;
2951         case 6:
2952                 if (i == sh->qd_idx)
2953                         return 0; /* It is the Q disk */
2954                 switch (algorithm) {
2955                 case ALGORITHM_LEFT_ASYMMETRIC:
2956                 case ALGORITHM_RIGHT_ASYMMETRIC:
2957                 case ALGORITHM_ROTATING_ZERO_RESTART:
2958                 case ALGORITHM_ROTATING_N_RESTART:
2959                         if (sh->pd_idx == raid_disks-1)
2960                                 i--;    /* Q D D D P */
2961                         else if (i > sh->pd_idx)
2962                                 i -= 2; /* D D P Q D */
2963                         break;
2964                 case ALGORITHM_LEFT_SYMMETRIC:
2965                 case ALGORITHM_RIGHT_SYMMETRIC:
2966                         if (sh->pd_idx == raid_disks-1)
2967                                 i--; /* Q D D D P */
2968                         else {
2969                                 /* D D P Q D */
2970                                 if (i < sh->pd_idx)
2971                                         i += raid_disks;
2972                                 i -= (sh->pd_idx + 2);
2973                         }
2974                         break;
2975                 case ALGORITHM_PARITY_0:
2976                         i -= 2;
2977                         break;
2978                 case ALGORITHM_PARITY_N:
2979                         break;
2980                 case ALGORITHM_ROTATING_N_CONTINUE:
2981                         /* Like left_symmetric, but P is before Q */
2982                         if (sh->pd_idx == 0)
2983                                 i--;    /* P D D D Q */
2984                         else {
2985                                 /* D D Q P D */
2986                                 if (i < sh->pd_idx)
2987                                         i += raid_disks;
2988                                 i -= (sh->pd_idx + 1);
2989                         }
2990                         break;
2991                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2992                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2993                         if (i > sh->pd_idx)
2994                                 i--;
2995                         break;
2996                 case ALGORITHM_LEFT_SYMMETRIC_6:
2997                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2998                         if (i < sh->pd_idx)
2999                                 i += data_disks + 1;
3000                         i -= (sh->pd_idx + 1);
3001                         break;
3002                 case ALGORITHM_PARITY_0_6:
3003                         i -= 1;
3004                         break;
3005                 default:
3006                         BUG();
3007                 }
3008                 break;
3009         }
3010
3011         chunk_number = stripe * data_disks + i;
3012         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3013
3014         check = raid5_compute_sector(conf, r_sector,
3015                                      previous, &dummy1, &sh2);
3016         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3017                 || sh2.qd_idx != sh->qd_idx) {
3018                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3019                         mdname(conf->mddev));
3020                 return 0;
3021         }
3022         return r_sector;
3023 }
3024
3025 /*
3026  * There are cases where we want handle_stripe_dirtying() and
3027  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3028  *
3029  * This function checks whether we want to delay the towrite. Specifically,
3030  * we delay the towrite when:
3031  *
3032  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3033  *      stripe has data in journal (for other devices).
3034  *
3035  *      In this case, when reading data for the non-overwrite dev, it is
3036  *      necessary to handle complex rmw of write back cache (prexor with
3037  *      orig_page, and xor with page). To keep read path simple, we would
3038  *      like to flush data in journal to RAID disks first, so complex rmw
3039  *      is handled in the write patch (handle_stripe_dirtying).
3040  *
3041  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3042  *
3043  *      It is important to be able to flush all stripes in raid5-cache.
3044  *      Therefore, we need reserve some space on the journal device for
3045  *      these flushes. If flush operation includes pending writes to the
3046  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3047  *      for the flush out. If we exclude these pending writes from flush
3048  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3049  *      Therefore, excluding pending writes in these cases enables more
3050  *      efficient use of the journal device.
3051  *
3052  *      Note: To make sure the stripe makes progress, we only delay
3053  *      towrite for stripes with data already in journal (injournal > 0).
3054  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3055  *      no_space_stripes list.
3056  *
3057  *   3. during journal failure
3058  *      In journal failure, we try to flush all cached data to raid disks
3059  *      based on data in stripe cache. The array is read-only to upper
3060  *      layers, so we would skip all pending writes.
3061  *
3062  */
3063 static inline bool delay_towrite(struct r5conf *conf,
3064                                  struct r5dev *dev,
3065                                  struct stripe_head_state *s)
3066 {
3067         /* case 1 above */
3068         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3069             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3070                 return true;
3071         /* case 2 above */
3072         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3073             s->injournal > 0)
3074                 return true;
3075         /* case 3 above */
3076         if (s->log_failed && s->injournal)
3077                 return true;
3078         return false;
3079 }
3080
3081 static void
3082 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3083                          int rcw, int expand)
3084 {
3085         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3086         struct r5conf *conf = sh->raid_conf;
3087         int level = conf->level;
3088
3089         if (rcw) {
3090                 /*
3091                  * In some cases, handle_stripe_dirtying initially decided to
3092                  * run rmw and allocates extra page for prexor. However, rcw is
3093                  * cheaper later on. We need to free the extra page now,
3094                  * because we won't be able to do that in ops_complete_prexor().
3095                  */
3096                 r5c_release_extra_page(sh);
3097
3098                 for (i = disks; i--; ) {
3099                         struct r5dev *dev = &sh->dev[i];
3100
3101                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3102                                 set_bit(R5_LOCKED, &dev->flags);
3103                                 set_bit(R5_Wantdrain, &dev->flags);
3104                                 if (!expand)
3105                                         clear_bit(R5_UPTODATE, &dev->flags);
3106                                 s->locked++;
3107                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3108                                 set_bit(R5_LOCKED, &dev->flags);
3109                                 s->locked++;
3110                         }
3111                 }
3112                 /* if we are not expanding this is a proper write request, and
3113                  * there will be bios with new data to be drained into the
3114                  * stripe cache
3115                  */
3116                 if (!expand) {
3117                         if (!s->locked)
3118                                 /* False alarm, nothing to do */
3119                                 return;
3120                         sh->reconstruct_state = reconstruct_state_drain_run;
3121                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3122                 } else
3123                         sh->reconstruct_state = reconstruct_state_run;
3124
3125                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3126
3127                 if (s->locked + conf->max_degraded == disks)
3128                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3129                                 atomic_inc(&conf->pending_full_writes);
3130         } else {
3131                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3132                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3133                 BUG_ON(level == 6 &&
3134                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3135                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3136
3137                 for (i = disks; i--; ) {
3138                         struct r5dev *dev = &sh->dev[i];
3139                         if (i == pd_idx || i == qd_idx)
3140                                 continue;
3141
3142                         if (dev->towrite &&
3143                             (test_bit(R5_UPTODATE, &dev->flags) ||
3144                              test_bit(R5_Wantcompute, &dev->flags))) {
3145                                 set_bit(R5_Wantdrain, &dev->flags);
3146                                 set_bit(R5_LOCKED, &dev->flags);
3147                                 clear_bit(R5_UPTODATE, &dev->flags);
3148                                 s->locked++;
3149                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3150                                 set_bit(R5_LOCKED, &dev->flags);
3151                                 s->locked++;
3152                         }
3153                 }
3154                 if (!s->locked)
3155                         /* False alarm - nothing to do */
3156                         return;
3157                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3158                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3159                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3160                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3161         }
3162
3163         /* keep the parity disk(s) locked while asynchronous operations
3164          * are in flight
3165          */
3166         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3167         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3168         s->locked++;
3169
3170         if (level == 6) {
3171                 int qd_idx = sh->qd_idx;
3172                 struct r5dev *dev = &sh->dev[qd_idx];
3173
3174                 set_bit(R5_LOCKED, &dev->flags);
3175                 clear_bit(R5_UPTODATE, &dev->flags);
3176                 s->locked++;
3177         }
3178
3179         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3180             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3181             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3182             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3183                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3184
3185         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3186                 __func__, (unsigned long long)sh->sector,
3187                 s->locked, s->ops_request);
3188 }
3189
3190 /*
3191  * Each stripe/dev can have one or more bion attached.
3192  * toread/towrite point to the first in a chain.
3193  * The bi_next chain must be in order.
3194  */
3195 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3196                           int forwrite, int previous)
3197 {
3198         struct bio **bip;
3199         struct r5conf *conf = sh->raid_conf;
3200         int firstwrite=0;
3201
3202         pr_debug("adding bi b#%llu to stripe s#%llu\n",
3203                 (unsigned long long)bi->bi_iter.bi_sector,
3204                 (unsigned long long)sh->sector);
3205
3206         spin_lock_irq(&sh->stripe_lock);
3207         /* Don't allow new IO added to stripes in batch list */
3208         if (sh->batch_head)
3209                 goto overlap;
3210         if (forwrite) {
3211                 bip = &sh->dev[dd_idx].towrite;
3212                 if (*bip == NULL)
3213                         firstwrite = 1;
3214         } else
3215                 bip = &sh->dev[dd_idx].toread;
3216         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3217                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3218                         goto overlap;
3219                 bip = & (*bip)->bi_next;
3220         }
3221         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3222                 goto overlap;
3223
3224         if (forwrite && raid5_has_ppl(conf)) {
3225                 /*
3226                  * With PPL only writes to consecutive data chunks within a
3227                  * stripe are allowed because for a single stripe_head we can
3228                  * only have one PPL entry at a time, which describes one data
3229                  * range. Not really an overlap, but wait_for_overlap can be
3230                  * used to handle this.
3231                  */
3232                 sector_t sector;
3233                 sector_t first = 0;
3234                 sector_t last = 0;
3235                 int count = 0;
3236                 int i;
3237
3238                 for (i = 0; i < sh->disks; i++) {
3239                         if (i != sh->pd_idx &&
3240                             (i == dd_idx || sh->dev[i].towrite)) {
3241                                 sector = sh->dev[i].sector;
3242                                 if (count == 0 || sector < first)
3243                                         first = sector;
3244                                 if (sector > last)
3245                                         last = sector;
3246                                 count++;
3247                         }
3248                 }
3249
3250                 if (first + conf->chunk_sectors * (count - 1) != last)
3251                         goto overlap;
3252         }
3253
3254         if (!forwrite || previous)
3255                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3256
3257         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3258         if (*bip)
3259                 bi->bi_next = *bip;
3260         *bip = bi;
3261         bio_inc_remaining(bi);
3262         md_write_inc(conf->mddev, bi);
3263
3264         if (forwrite) {
3265                 /* check if page is covered */
3266                 sector_t sector = sh->dev[dd_idx].sector;
3267                 for (bi=sh->dev[dd_idx].towrite;
3268                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3269                              bi && bi->bi_iter.bi_sector <= sector;
3270                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3271                         if (bio_end_sector(bi) >= sector)
3272                                 sector = bio_end_sector(bi);
3273                 }
3274                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3275                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3276                                 sh->overwrite_disks++;
3277         }
3278
3279         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3280                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3281                 (unsigned long long)sh->sector, dd_idx);
3282
3283         if (conf->mddev->bitmap && firstwrite) {
3284                 /* Cannot hold spinlock over bitmap_startwrite,
3285                  * but must ensure this isn't added to a batch until
3286                  * we have added to the bitmap and set bm_seq.
3287                  * So set STRIPE_BITMAP_PENDING to prevent
3288                  * batching.
3289                  * If multiple add_stripe_bio() calls race here they
3290                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3291                  * to complete "bitmap_startwrite" gets to set
3292                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3293                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3294                  * any more.
3295                  */
3296                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3297                 spin_unlock_irq(&sh->stripe_lock);
3298                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3299                                   STRIPE_SECTORS, 0);
3300                 spin_lock_irq(&sh->stripe_lock);
3301                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3302                 if (!sh->batch_head) {
3303                         sh->bm_seq = conf->seq_flush+1;
3304                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3305                 }
3306         }
3307         spin_unlock_irq(&sh->stripe_lock);
3308
3309         if (stripe_can_batch(sh))
3310                 stripe_add_to_batch_list(conf, sh);
3311         return 1;
3312
3313  overlap:
3314         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3315         spin_unlock_irq(&sh->stripe_lock);
3316         return 0;
3317 }
3318
3319 static void end_reshape(struct r5conf *conf);
3320
3321 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3322                             struct stripe_head *sh)
3323 {
3324         int sectors_per_chunk =
3325                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3326         int dd_idx;
3327         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3328         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3329
3330         raid5_compute_sector(conf,
3331                              stripe * (disks - conf->max_degraded)
3332                              *sectors_per_chunk + chunk_offset,
3333                              previous,
3334                              &dd_idx, sh);
3335 }
3336
3337 static void
3338 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3339                      struct stripe_head_state *s, int disks)
3340 {
3341         int i;
3342         BUG_ON(sh->batch_head);
3343         for (i = disks; i--; ) {
3344                 struct bio *bi;
3345                 int bitmap_end = 0;
3346
3347                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3348                         struct md_rdev *rdev;
3349                         rcu_read_lock();
3350                         rdev = rcu_dereference(conf->disks[i].rdev);
3351                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3352                             !test_bit(Faulty, &rdev->flags))
3353                                 atomic_inc(&rdev->nr_pending);
3354                         else
3355                                 rdev = NULL;
3356                         rcu_read_unlock();
3357                         if (rdev) {
3358                                 if (!rdev_set_badblocks(
3359                                             rdev,
3360                                             sh->sector,
3361                                             STRIPE_SECTORS, 0))
3362                                         md_error(conf->mddev, rdev);
3363                                 rdev_dec_pending(rdev, conf->mddev);
3364                         }
3365                 }
3366                 spin_lock_irq(&sh->stripe_lock);
3367                 /* fail all writes first */
3368                 bi = sh->dev[i].towrite;
3369                 sh->dev[i].towrite = NULL;
3370                 sh->overwrite_disks = 0;
3371                 spin_unlock_irq(&sh->stripe_lock);
3372                 if (bi)
3373                         bitmap_end = 1;
3374
3375                 log_stripe_write_finished(sh);
3376
3377                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3378                         wake_up(&conf->wait_for_overlap);
3379
3380                 while (bi && bi->bi_iter.bi_sector <
3381                         sh->dev[i].sector + STRIPE_SECTORS) {
3382                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3383
3384                         bi->bi_error = -EIO;
3385                         md_write_end(conf->mddev);
3386                         bio_endio(bi);
3387                         bi = nextbi;
3388                 }
3389                 if (bitmap_end)
3390                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3391                                 STRIPE_SECTORS, 0, 0);
3392                 bitmap_end = 0;
3393                 /* and fail all 'written' */
3394                 bi = sh->dev[i].written;
3395                 sh->dev[i].written = NULL;
3396                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3397                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3398                         sh->dev[i].page = sh->dev[i].orig_page;
3399                 }
3400
3401                 if (bi) bitmap_end = 1;
3402                 while (bi && bi->bi_iter.bi_sector <
3403                        sh->dev[i].sector + STRIPE_SECTORS) {
3404                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3405
3406                         bi->bi_error = -EIO;
3407                         md_write_end(conf->mddev);
3408                         bio_endio(bi);
3409                         bi = bi2;
3410                 }
3411
3412                 /* fail any reads if this device is non-operational and
3413                  * the data has not reached the cache yet.
3414                  */
3415                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3416                     s->failed > conf->max_degraded &&
3417                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3418                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3419                         spin_lock_irq(&sh->stripe_lock);
3420                         bi = sh->dev[i].toread;
3421                         sh->dev[i].toread = NULL;
3422                         spin_unlock_irq(&sh->stripe_lock);
3423                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3424                                 wake_up(&conf->wait_for_overlap);
3425                         if (bi)
3426                                 s->to_read--;
3427                         while (bi && bi->bi_iter.bi_sector <
3428                                sh->dev[i].sector + STRIPE_SECTORS) {
3429                                 struct bio *nextbi =
3430                                         r5_next_bio(bi, sh->dev[i].sector);
3431
3432                                 bi->bi_error = -EIO;
3433                                 bio_endio(bi);
3434                                 bi = nextbi;
3435                         }
3436                 }
3437                 if (bitmap_end)
3438                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3439                                         STRIPE_SECTORS, 0, 0);
3440                 /* If we were in the middle of a write the parity block might
3441                  * still be locked - so just clear all R5_LOCKED flags
3442                  */
3443                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3444         }
3445         s->to_write = 0;
3446         s->written = 0;
3447
3448         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3449                 if (atomic_dec_and_test(&conf->pending_full_writes))
3450                         md_wakeup_thread(conf->mddev->thread);
3451 }
3452
3453 static void
3454 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3455                    struct stripe_head_state *s)
3456 {
3457         int abort = 0;
3458         int i;
3459
3460         BUG_ON(sh->batch_head);
3461         clear_bit(STRIPE_SYNCING, &sh->state);
3462         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3463                 wake_up(&conf->wait_for_overlap);
3464         s->syncing = 0;
3465         s->replacing = 0;
3466         /* There is nothing more to do for sync/check/repair.
3467          * Don't even need to abort as that is handled elsewhere
3468          * if needed, and not always wanted e.g. if there is a known
3469          * bad block here.
3470          * For recover/replace we need to record a bad block on all
3471          * non-sync devices, or abort the recovery
3472          */
3473         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3474                 /* During recovery devices cannot be removed, so
3475                  * locking and refcounting of rdevs is not needed
3476                  */
3477                 rcu_read_lock();
3478                 for (i = 0; i < conf->raid_disks; i++) {
3479                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3480                         if (rdev
3481                             && !test_bit(Faulty, &rdev->flags)
3482                             && !test_bit(In_sync, &rdev->flags)
3483                             && !rdev_set_badblocks(rdev, sh->sector,
3484                                                    STRIPE_SECTORS, 0))
3485                                 abort = 1;
3486                         rdev = rcu_dereference(conf->disks[i].replacement);
3487                         if (rdev
3488                             && !test_bit(Faulty, &rdev->flags)
3489                             && !test_bit(In_sync, &rdev->flags)
3490                             && !rdev_set_badblocks(rdev, sh->sector,
3491                                                    STRIPE_SECTORS, 0))
3492                                 abort = 1;
3493                 }
3494                 rcu_read_unlock();
3495                 if (abort)
3496                         conf->recovery_disabled =
3497                                 conf->mddev->recovery_disabled;
3498         }
3499         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3500 }
3501
3502 static int want_replace(struct stripe_head *sh, int disk_idx)
3503 {
3504         struct md_rdev *rdev;
3505         int rv = 0;
3506
3507         rcu_read_lock();
3508         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3509         if (rdev
3510             && !test_bit(Faulty, &rdev->flags)
3511             && !test_bit(In_sync, &rdev->flags)
3512             && (rdev->recovery_offset <= sh->sector
3513                 || rdev->mddev->recovery_cp <= sh->sector))
3514                 rv = 1;
3515         rcu_read_unlock();
3516         return rv;
3517 }
3518
3519 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3520                            int disk_idx, int disks)
3521 {
3522         struct r5dev *dev = &sh->dev[disk_idx];
3523         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3524                                   &sh->dev[s->failed_num[1]] };
3525         int i;
3526
3527
3528         if (test_bit(R5_LOCKED, &dev->flags) ||
3529             test_bit(R5_UPTODATE, &dev->flags))
3530                 /* No point reading this as we already have it or have
3531                  * decided to get it.
3532                  */
3533                 return 0;
3534
3535         if (dev->toread ||
3536             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3537                 /* We need this block to directly satisfy a request */
3538                 return 1;
3539
3540         if (s->syncing || s->expanding ||
3541             (s->replacing && want_replace(sh, disk_idx)))
3542                 /* When syncing, or expanding we read everything.
3543                  * When replacing, we need the replaced block.
3544                  */
3545                 return 1;
3546
3547         if ((s->failed >= 1 && fdev[0]->toread) ||
3548             (s->failed >= 2 && fdev[1]->toread))
3549                 /* If we want to read from a failed device, then
3550                  * we need to actually read every other device.
3551                  */
3552                 return 1;
3553
3554         /* Sometimes neither read-modify-write nor reconstruct-write
3555          * cycles can work.  In those cases we read every block we
3556          * can.  Then the parity-update is certain to have enough to
3557          * work with.
3558          * This can only be a problem when we need to write something,
3559          * and some device has failed.  If either of those tests
3560          * fail we need look no further.
3561          */
3562         if (!s->failed || !s->to_write)
3563                 return 0;
3564
3565         if (test_bit(R5_Insync, &dev->flags) &&
3566             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3567                 /* Pre-reads at not permitted until after short delay
3568                  * to gather multiple requests.  However if this
3569                  * device is no Insync, the block could only be computed
3570                  * and there is no need to delay that.
3571                  */
3572                 return 0;
3573
3574         for (i = 0; i < s->failed && i < 2; i++) {
3575                 if (fdev[i]->towrite &&
3576                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3577                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3578                         /* If we have a partial write to a failed
3579                          * device, then we will need to reconstruct
3580                          * the content of that device, so all other
3581                          * devices must be read.
3582                          */
3583                         return 1;
3584         }
3585
3586         /* If we are forced to do a reconstruct-write, either because
3587          * the current RAID6 implementation only supports that, or
3588          * because parity cannot be trusted and we are currently
3589          * recovering it, there is extra need to be careful.
3590          * If one of the devices that we would need to read, because
3591          * it is not being overwritten (and maybe not written at all)
3592          * is missing/faulty, then we need to read everything we can.
3593          */
3594         if (sh->raid_conf->level != 6 &&
3595             sh->sector < sh->raid_conf->mddev->recovery_cp)
3596                 /* reconstruct-write isn't being forced */
3597                 return 0;
3598         for (i = 0; i < s->failed && i < 2; i++) {
3599                 if (s->failed_num[i] != sh->pd_idx &&
3600                     s->failed_num[i] != sh->qd_idx &&
3601                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3602                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3603                         return 1;
3604         }
3605
3606         return 0;
3607 }
3608
3609 /* fetch_block - checks the given member device to see if its data needs
3610  * to be read or computed to satisfy a request.
3611  *
3612  * Returns 1 when no more member devices need to be checked, otherwise returns
3613  * 0 to tell the loop in handle_stripe_fill to continue
3614  */
3615 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3616                        int disk_idx, int disks)
3617 {
3618         struct r5dev *dev = &sh->dev[disk_idx];
3619
3620         /* is the data in this block needed, and can we get it? */
3621         if (need_this_block(sh, s, disk_idx, disks)) {
3622                 /* we would like to get this block, possibly by computing it,
3623                  * otherwise read it if the backing disk is insync
3624                  */
3625                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3626                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3627                 BUG_ON(sh->batch_head);
3628
3629                 /*
3630                  * In the raid6 case if the only non-uptodate disk is P
3631                  * then we already trusted P to compute the other failed
3632                  * drives. It is safe to compute rather than re-read P.
3633                  * In other cases we only compute blocks from failed
3634                  * devices, otherwise check/repair might fail to detect
3635                  * a real inconsistency.
3636                  */
3637
3638                 if ((s->uptodate == disks - 1) &&
3639                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3640                     (s->failed && (disk_idx == s->failed_num[0] ||
3641                                    disk_idx == s->failed_num[1])))) {
3642                         /* have disk failed, and we're requested to fetch it;
3643                          * do compute it
3644                          */
3645                         pr_debug("Computing stripe %llu block %d\n",
3646                                (unsigned long long)sh->sector, disk_idx);
3647                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3648                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3649                         set_bit(R5_Wantcompute, &dev->flags);
3650                         sh->ops.target = disk_idx;
3651                         sh->ops.target2 = -1; /* no 2nd target */
3652                         s->req_compute = 1;
3653                         /* Careful: from this point on 'uptodate' is in the eye
3654                          * of raid_run_ops which services 'compute' operations
3655                          * before writes. R5_Wantcompute flags a block that will
3656                          * be R5_UPTODATE by the time it is needed for a
3657                          * subsequent operation.
3658                          */
3659                         s->uptodate++;
3660                         return 1;
3661                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3662                         /* Computing 2-failure is *very* expensive; only
3663                          * do it if failed >= 2
3664                          */
3665                         int other;
3666                         for (other = disks; other--; ) {
3667                                 if (other == disk_idx)
3668                                         continue;
3669                                 if (!test_bit(R5_UPTODATE,
3670                                       &sh->dev[other].flags))
3671                                         break;
3672                         }
3673                         BUG_ON(other < 0);
3674                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3675                                (unsigned long long)sh->sector,
3676                                disk_idx, other);
3677                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3678                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3679                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3680                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3681                         sh->ops.target = disk_idx;
3682                         sh->ops.target2 = other;
3683                         s->uptodate += 2;
3684                         s->req_compute = 1;
3685                         return 1;
3686                 } else if (test_bit(R5_Insync, &dev->flags)) {
3687                         set_bit(R5_LOCKED, &dev->flags);
3688                         set_bit(R5_Wantread, &dev->flags);
3689                         s->locked++;
3690                         pr_debug("Reading block %d (sync=%d)\n",
3691                                 disk_idx, s->syncing);
3692                 }
3693         }
3694
3695         return 0;
3696 }
3697
3698 /**
3699  * handle_stripe_fill - read or compute data to satisfy pending requests.
3700  */
3701 static void handle_stripe_fill(struct stripe_head *sh,
3702                                struct stripe_head_state *s,
3703                                int disks)
3704 {
3705         int i;
3706
3707         /* look for blocks to read/compute, skip this if a compute
3708          * is already in flight, or if the stripe contents are in the
3709          * midst of changing due to a write
3710          */
3711         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3712             !sh->reconstruct_state) {
3713
3714                 /*
3715                  * For degraded stripe with data in journal, do not handle
3716                  * read requests yet, instead, flush the stripe to raid
3717                  * disks first, this avoids handling complex rmw of write
3718                  * back cache (prexor with orig_page, and then xor with
3719                  * page) in the read path
3720                  */
3721                 if (s->injournal && s->failed) {
3722                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3723                                 r5c_make_stripe_write_out(sh);
3724                         goto out;
3725                 }
3726
3727                 for (i = disks; i--; )
3728                         if (fetch_block(sh, s, i, disks))
3729                                 break;
3730         }
3731 out:
3732         set_bit(STRIPE_HANDLE, &sh->state);
3733 }
3734
3735 static void break_stripe_batch_list(struct stripe_head *head_sh,
3736                                     unsigned long handle_flags);
3737 /* handle_stripe_clean_event
3738  * any written block on an uptodate or failed drive can be returned.
3739  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3740  * never LOCKED, so we don't need to test 'failed' directly.
3741  */
3742 static void handle_stripe_clean_event(struct r5conf *conf,
3743         struct stripe_head *sh, int disks)
3744 {
3745         int i;
3746         struct r5dev *dev;
3747         int discard_pending = 0;
3748         struct stripe_head *head_sh = sh;
3749         bool do_endio = false;
3750
3751         for (i = disks; i--; )
3752                 if (sh->dev[i].written) {
3753                         dev = &sh->dev[i];
3754                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3755                             (test_bit(R5_UPTODATE, &dev->flags) ||
3756                              test_bit(R5_Discard, &dev->flags) ||
3757                              test_bit(R5_SkipCopy, &dev->flags))) {
3758                                 /* We can return any write requests */
3759                                 struct bio *wbi, *wbi2;
3760                                 pr_debug("Return write for disc %d\n", i);
3761                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3762                                         clear_bit(R5_UPTODATE, &dev->flags);
3763                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3764                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3765                                 }
3766                                 do_endio = true;
3767
3768 returnbi:
3769                                 dev->page = dev->orig_page;
3770                                 wbi = dev->written;
3771                                 dev->written = NULL;
3772                                 while (wbi && wbi->bi_iter.bi_sector <
3773                                         dev->sector + STRIPE_SECTORS) {
3774                                         wbi2 = r5_next_bio(wbi, dev->sector);
3775                                         md_write_end(conf->mddev);
3776                                         bio_endio(wbi);
3777                                         wbi = wbi2;
3778                                 }
3779                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3780                                                 STRIPE_SECTORS,
3781                                          !test_bit(STRIPE_DEGRADED, &sh->state),
3782                                                 0);
3783                                 if (head_sh->batch_head) {
3784                                         sh = list_first_entry(&sh->batch_list,
3785                                                               struct stripe_head,
3786                                                               batch_list);
3787                                         if (sh != head_sh) {
3788                                                 dev = &sh->dev[i];
3789                                                 goto returnbi;
3790                                         }
3791                                 }
3792                                 sh = head_sh;
3793                                 dev = &sh->dev[i];
3794                         } else if (test_bit(R5_Discard, &dev->flags))
3795                                 discard_pending = 1;
3796                 }
3797
3798         log_stripe_write_finished(sh);
3799
3800         if (!discard_pending &&
3801             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3802                 int hash;
3803                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3804                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3805                 if (sh->qd_idx >= 0) {
3806                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3807                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3808                 }
3809                 /* now that discard is done we can proceed with any sync */
3810                 clear_bit(STRIPE_DISCARD, &sh->state);
3811                 /*
3812                  * SCSI discard will change some bio fields and the stripe has
3813                  * no updated data, so remove it from hash list and the stripe
3814                  * will be reinitialized
3815                  */
3816 unhash:
3817                 hash = sh->hash_lock_index;
3818                 spin_lock_irq(conf->hash_locks + hash);
3819                 remove_hash(sh);
3820                 spin_unlock_irq(conf->hash_locks + hash);
3821                 if (head_sh->batch_head) {
3822                         sh = list_first_entry(&sh->batch_list,
3823                                               struct stripe_head, batch_list);
3824                         if (sh != head_sh)
3825                                         goto unhash;
3826                 }
3827                 sh = head_sh;
3828
3829                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3830                         set_bit(STRIPE_HANDLE, &sh->state);
3831
3832         }
3833
3834         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3835                 if (atomic_dec_and_test(&conf->pending_full_writes))
3836                         md_wakeup_thread(conf->mddev->thread);
3837
3838         if (head_sh->batch_head && do_endio)
3839                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3840 }
3841
3842 /*
3843  * For RMW in write back cache, we need extra page in prexor to store the
3844  * old data. This page is stored in dev->orig_page.
3845  *
3846  * This function checks whether we have data for prexor. The exact logic
3847  * is:
3848  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3849  */
3850 static inline bool uptodate_for_rmw(struct r5dev *dev)
3851 {
3852         return (test_bit(R5_UPTODATE, &dev->flags)) &&
3853                 (!test_bit(R5_InJournal, &dev->flags) ||
3854                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3855 }
3856
3857 static int handle_stripe_dirtying(struct r5conf *conf,
3858                                   struct stripe_head *sh,
3859                                   struct stripe_head_state *s,
3860                                   int disks)
3861 {
3862         int rmw = 0, rcw = 0, i;
3863         sector_t recovery_cp = conf->mddev->recovery_cp;
3864
3865         /* Check whether resync is now happening or should start.
3866          * If yes, then the array is dirty (after unclean shutdown or
3867          * initial creation), so parity in some stripes might be inconsistent.
3868          * In this case, we need to always do reconstruct-write, to ensure
3869          * that in case of drive failure or read-error correction, we
3870          * generate correct data from the parity.
3871          */
3872         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3873             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3874              s->failed == 0)) {
3875                 /* Calculate the real rcw later - for now make it
3876                  * look like rcw is cheaper
3877                  */
3878                 rcw = 1; rmw = 2;
3879                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3880                          conf->rmw_level, (unsigned long long)recovery_cp,
3881                          (unsigned long long)sh->sector);
3882         } else for (i = disks; i--; ) {
3883                 /* would I have to read this buffer for read_modify_write */
3884                 struct r5dev *dev = &sh->dev[i];
3885                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3886                      i == sh->pd_idx || i == sh->qd_idx ||
3887                      test_bit(R5_InJournal, &dev->flags)) &&
3888                     !test_bit(R5_LOCKED, &dev->flags) &&
3889                     !(uptodate_for_rmw(dev) ||
3890                       test_bit(R5_Wantcompute, &dev->flags))) {
3891                         if (test_bit(R5_Insync, &dev->flags))
3892                                 rmw++;
3893                         else
3894                                 rmw += 2*disks;  /* cannot read it */
3895                 }
3896                 /* Would I have to read this buffer for reconstruct_write */
3897                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3898                     i != sh->pd_idx && i != sh->qd_idx &&
3899                     !test_bit(R5_LOCKED, &dev->flags) &&
3900                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3901                       test_bit(R5_Wantcompute, &dev->flags))) {
3902                         if (test_bit(R5_Insync, &dev->flags))
3903                                 rcw++;
3904                         else
3905                                 rcw += 2*disks;
3906                 }
3907         }
3908
3909         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3910                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
3911         set_bit(STRIPE_HANDLE, &sh->state);
3912         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3913                 /* prefer read-modify-write, but need to get some data */
3914                 if (conf->mddev->queue)
3915                         blk_add_trace_msg(conf->mddev->queue,
3916                                           "raid5 rmw %llu %d",
3917                                           (unsigned long long)sh->sector, rmw);
3918                 for (i = disks; i--; ) {
3919                         struct r5dev *dev = &sh->dev[i];
3920                         if (test_bit(R5_InJournal, &dev->flags) &&
3921                             dev->page == dev->orig_page &&
3922                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3923                                 /* alloc page for prexor */
3924                                 struct page *p = alloc_page(GFP_NOIO);
3925
3926                                 if (p) {
3927                                         dev->orig_page = p;
3928                                         continue;
3929                                 }
3930
3931                                 /*
3932                                  * alloc_page() failed, try use
3933                                  * disk_info->extra_page
3934                                  */
3935                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3936                                                       &conf->cache_state)) {
3937                                         r5c_use_extra_page(sh);
3938                                         break;
3939                                 }
3940
3941                                 /* extra_page in use, add to delayed_list */
3942                                 set_bit(STRIPE_DELAYED, &sh->state);
3943                                 s->waiting_extra_page = 1;
3944                                 return -EAGAIN;
3945                         }
3946                 }
3947
3948                 for (i = disks; i--; ) {
3949                         struct r5dev *dev = &sh->dev[i];
3950                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3951                              i == sh->pd_idx || i == sh->qd_idx ||
3952                              test_bit(R5_InJournal, &dev->flags)) &&
3953                             !test_bit(R5_LOCKED, &dev->flags) &&
3954                             !(uptodate_for_rmw(dev) ||
3955                               test_bit(R5_Wantcompute, &dev->flags)) &&
3956                             test_bit(R5_Insync, &dev->flags)) {
3957                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3958                                              &sh->state)) {
3959                                         pr_debug("Read_old block %d for r-m-w\n",
3960                                                  i);
3961                                         set_bit(R5_LOCKED, &dev->flags);
3962                                         set_bit(R5_Wantread, &dev->flags);
3963                                         s->locked++;
3964                                 } else {
3965                                         set_bit(STRIPE_DELAYED, &sh->state);
3966                                         set_bit(STRIPE_HANDLE, &sh->state);
3967                                 }
3968                         }
3969                 }
3970         }
3971         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3972                 /* want reconstruct write, but need to get some data */
3973                 int qread =0;
3974                 rcw = 0;
3975                 for (i = disks; i--; ) {
3976                         struct r5dev *dev = &sh->dev[i];
3977                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3978                             i != sh->pd_idx && i != sh->qd_idx &&
3979                             !test_bit(R5_LOCKED, &dev->flags) &&
3980                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3981                               test_bit(R5_Wantcompute, &dev->flags))) {
3982                                 rcw++;
3983                                 if (test_bit(R5_Insync, &dev->flags) &&
3984                                     test_bit(STRIPE_PREREAD_ACTIVE,
3985                                              &sh->state)) {
3986                                         pr_debug("Read_old block "
3987                                                 "%d for Reconstruct\n", i);
3988                                         set_bit(R5_LOCKED, &dev->flags);
3989                                         set_bit(R5_Wantread, &dev->flags);
3990                                         s->locked++;
3991                                         qread++;
3992                                 } else {
3993                                         set_bit(STRIPE_DELAYED, &sh->state);
3994                                         set_bit(STRIPE_HANDLE, &sh->state);
3995                                 }
3996                         }
3997                 }
3998                 if (rcw && conf->mddev->queue)
3999                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4000                                           (unsigned long long)sh->sector,
4001                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4002         }
4003
4004         if (rcw > disks && rmw > disks &&
4005             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4006                 set_bit(STRIPE_DELAYED, &sh->state);
4007
4008         /* now if nothing is locked, and if we have enough data,
4009          * we can start a write request
4010          */
4011         /* since handle_stripe can be called at any time we need to handle the
4012          * case where a compute block operation has been submitted and then a
4013          * subsequent call wants to start a write request.  raid_run_ops only
4014          * handles the case where compute block and reconstruct are requested
4015          * simultaneously.  If this is not the case then new writes need to be
4016          * held off until the compute completes.
4017          */
4018         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4019             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4020              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4021                 schedule_reconstruction(sh, s, rcw == 0, 0);
4022         return 0;
4023 }
4024
4025 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4026                                 struct stripe_head_state *s, int disks)
4027 {
4028         struct r5dev *dev = NULL;
4029
4030         BUG_ON(sh->batch_head);
4031         set_bit(STRIPE_HANDLE, &sh->state);
4032
4033         switch (sh->check_state) {
4034         case check_state_idle:
4035                 /* start a new check operation if there are no failures */
4036                 if (s->failed == 0) {
4037                         BUG_ON(s->uptodate != disks);
4038                         sh->check_state = check_state_run;
4039                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4040                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4041                         s->uptodate--;
4042                         break;
4043                 }
4044                 dev = &sh->dev[s->failed_num[0]];
4045                 /* fall through */
4046         case check_state_compute_result:
4047                 sh->check_state = check_state_idle;
4048                 if (!dev)
4049                         dev = &sh->dev[sh->pd_idx];
4050
4051                 /* check that a write has not made the stripe insync */
4052                 if (test_bit(STRIPE_INSYNC, &sh->state))
4053                         break;
4054
4055                 /* either failed parity check, or recovery is happening */
4056                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4057                 BUG_ON(s->uptodate != disks);
4058
4059                 set_bit(R5_LOCKED, &dev->flags);
4060                 s->locked++;
4061                 set_bit(R5_Wantwrite, &dev->flags);
4062
4063                 clear_bit(STRIPE_DEGRADED, &sh->state);
4064                 set_bit(STRIPE_INSYNC, &sh->state);
4065                 break;
4066         case check_state_run:
4067                 break; /* we will be called again upon completion */
4068         case check_state_check_result:
4069                 sh->check_state = check_state_idle;
4070
4071                 /* if a failure occurred during the check operation, leave
4072                  * STRIPE_INSYNC not set and let the stripe be handled again
4073                  */
4074                 if (s->failed)
4075                         break;
4076
4077                 /* handle a successful check operation, if parity is correct
4078                  * we are done.  Otherwise update the mismatch count and repair
4079                  * parity if !MD_RECOVERY_CHECK
4080                  */
4081                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4082                         /* parity is correct (on disc,
4083                          * not in buffer any more)
4084                          */
4085                         set_bit(STRIPE_INSYNC, &sh->state);
4086                 else {
4087                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4088                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
4089                                 /* don't try to repair!! */
4090                                 set_bit(STRIPE_INSYNC, &sh->state);
4091                         else {
4092                                 sh->check_state = check_state_compute_run;
4093                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4094                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4095                                 set_bit(R5_Wantcompute,
4096                                         &sh->dev[sh->pd_idx].flags);
4097                                 sh->ops.target = sh->pd_idx;
4098                                 sh->ops.target2 = -1;
4099                                 s->uptodate++;
4100                         }
4101                 }
4102                 break;
4103         case check_state_compute_run:
4104                 break;
4105         default:
4106                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4107                        __func__, sh->check_state,
4108                        (unsigned long long) sh->sector);
4109                 BUG();
4110         }
4111 }
4112
4113 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4114                                   struct stripe_head_state *s,
4115                                   int disks)
4116 {
4117         int pd_idx = sh->pd_idx;
4118         int qd_idx = sh->qd_idx;
4119         struct r5dev *dev;
4120
4121         BUG_ON(sh->batch_head);
4122         set_bit(STRIPE_HANDLE, &sh->state);
4123
4124         BUG_ON(s->failed > 2);
4125
4126         /* Want to check and possibly repair P and Q.
4127          * However there could be one 'failed' device, in which
4128          * case we can only check one of them, possibly using the
4129          * other to generate missing data
4130          */
4131
4132         switch (sh->check_state) {
4133         case check_state_idle:
4134                 /* start a new check operation if there are < 2 failures */
4135                 if (s->failed == s->q_failed) {
4136                         /* The only possible failed device holds Q, so it
4137                          * makes sense to check P (If anything else were failed,
4138                          * we would have used P to recreate it).
4139                          */
4140                         sh->check_state = check_state_run;
4141                 }
4142                 if (!s->q_failed && s->failed < 2) {
4143                         /* Q is not failed, and we didn't use it to generate
4144                          * anything, so it makes sense to check it
4145                          */
4146                         if (sh->check_state == check_state_run)
4147                                 sh->check_state = check_state_run_pq;
4148                         else
4149                                 sh->check_state = check_state_run_q;
4150                 }
4151
4152                 /* discard potentially stale zero_sum_result */
4153                 sh->ops.zero_sum_result = 0;
4154
4155                 if (sh->check_state == check_state_run) {
4156                         /* async_xor_zero_sum destroys the contents of P */
4157                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4158                         s->uptodate--;
4159                 }
4160                 if (sh->check_state >= check_state_run &&
4161                     sh->check_state <= check_state_run_pq) {
4162                         /* async_syndrome_zero_sum preserves P and Q, so
4163                          * no need to mark them !uptodate here
4164                          */
4165                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4166                         break;
4167                 }
4168
4169                 /* we have 2-disk failure */
4170                 BUG_ON(s->failed != 2);
4171                 /* fall through */
4172         case check_state_compute_result:
4173                 sh->check_state = check_state_idle;
4174
4175                 /* check that a write has not made the stripe insync */
4176                 if (test_bit(STRIPE_INSYNC, &sh->state))
4177                         break;
4178
4179                 /* now write out any block on a failed drive,
4180                  * or P or Q if they were recomputed
4181                  */
4182                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
4183                 if (s->failed == 2) {
4184                         dev = &sh->dev[s->failed_num[1]];
4185                         s->locked++;
4186                         set_bit(R5_LOCKED, &dev->flags);
4187                         set_bit(R5_Wantwrite, &dev->flags);
4188                 }
4189                 if (s->failed >= 1) {
4190                         dev = &sh->dev[s->failed_num[0]];
4191                         s->locked++;
4192                         set_bit(R5_LOCKED, &dev->flags);
4193                         set_bit(R5_Wantwrite, &dev->flags);
4194                 }
4195                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4196                         dev = &sh->dev[pd_idx];
4197                         s->locked++;
4198                         set_bit(R5_LOCKED, &dev->flags);
4199                         set_bit(R5_Wantwrite, &dev->flags);
4200                 }
4201                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4202                         dev = &sh->dev[qd_idx];
4203                         s->locked++;
4204                         set_bit(R5_LOCKED, &dev->flags);
4205                         set_bit(R5_Wantwrite, &dev->flags);
4206                 }
4207                 clear_bit(STRIPE_DEGRADED, &sh->state);
4208
4209                 set_bit(STRIPE_INSYNC, &sh->state);
4210                 break;
4211         case check_state_run:
4212         case check_state_run_q:
4213         case check_state_run_pq:
4214                 break; /* we will be called again upon completion */
4215         case check_state_check_result:
4216                 sh->check_state = check_state_idle;
4217
4218                 /* handle a successful check operation, if parity is correct
4219                  * we are done.  Otherwise update the mismatch count and repair
4220                  * parity if !MD_RECOVERY_CHECK
4221                  */
4222                 if (sh->ops.zero_sum_result == 0) {
4223                         /* both parities are correct */
4224                         if (!s->failed)
4225                                 set_bit(STRIPE_INSYNC, &sh->state);
4226                         else {
4227                                 /* in contrast to the raid5 case we can validate
4228                                  * parity, but still have a failure to write
4229                                  * back
4230                                  */
4231                                 sh->check_state = check_state_compute_result;
4232                                 /* Returning at this point means that we may go
4233                                  * off and bring p and/or q uptodate again so
4234                                  * we make sure to check zero_sum_result again
4235                                  * to verify if p or q need writeback
4236                                  */
4237                         }
4238                 } else {
4239                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4240                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
4241                                 /* don't try to repair!! */
4242                                 set_bit(STRIPE_INSYNC, &sh->state);
4243                         else {
4244                                 int *target = &sh->ops.target;
4245
4246                                 sh->ops.target = -1;
4247                                 sh->ops.target2 = -1;
4248                                 sh->check_state = check_state_compute_run;
4249                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4250                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4251                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4252                                         set_bit(R5_Wantcompute,
4253                                                 &sh->dev[pd_idx].flags);
4254                                         *target = pd_idx;
4255                                         target = &sh->ops.target2;
4256                                         s->uptodate++;
4257                                 }
4258                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4259                                         set_bit(R5_Wantcompute,
4260                                                 &sh->dev[qd_idx].flags);
4261                                         *target = qd_idx;
4262                                         s->uptodate++;
4263                                 }
4264                         }
4265                 }
4266                 break;
4267         case check_state_compute_run:
4268                 break;
4269         default:
4270                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4271                         __func__, sh->check_state,
4272                         (unsigned long long) sh->sector);
4273                 BUG();
4274         }
4275 }
4276
4277 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4278 {
4279         int i;
4280
4281         /* We have read all the blocks in this stripe and now we need to
4282          * copy some of them into a target stripe for expand.
4283          */
4284         struct dma_async_tx_descriptor *tx = NULL;
4285         BUG_ON(sh->batch_head);
4286         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4287         for (i = 0; i < sh->disks; i++)
4288                 if (i != sh->pd_idx && i != sh->qd_idx) {
4289                         int dd_idx, j;
4290                         struct stripe_head *sh2;
4291                         struct async_submit_ctl submit;
4292
4293                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4294                         sector_t s = raid5_compute_sector(conf, bn, 0,
4295                                                           &dd_idx, NULL);
4296                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4297                         if (sh2 == NULL)
4298                                 /* so far only the early blocks of this stripe
4299                                  * have been requested.  When later blocks
4300                                  * get requested, we will try again
4301                                  */
4302                                 continue;
4303                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4304                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4305                                 /* must have already done this block */
4306                                 raid5_release_stripe(sh2);
4307                                 continue;
4308                         }
4309
4310                         /* place all the copies on one channel */
4311                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4312                         tx = async_memcpy(sh2->dev[dd_idx].page,
4313                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
4314                                           &submit);
4315
4316                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4317                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4318                         for (j = 0; j < conf->raid_disks; j++)
4319                                 if (j != sh2->pd_idx &&
4320                                     j != sh2->qd_idx &&
4321                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4322                                         break;
4323                         if (j == conf->raid_disks) {
4324                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4325                                 set_bit(STRIPE_HANDLE, &sh2->state);
4326                         }
4327                         raid5_release_stripe(sh2);
4328
4329                 }
4330         /* done submitting copies, wait for them to complete */
4331         async_tx_quiesce(&tx);
4332 }
4333
4334 /*
4335  * handle_stripe - do things to a stripe.
4336  *
4337  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4338  * state of various bits to see what needs to be done.
4339  * Possible results:
4340  *    return some read requests which now have data
4341  *    return some write requests which are safely on storage
4342  *    schedule a read on some buffers
4343  *    schedule a write of some buffers
4344  *    return confirmation of parity correctness
4345  *
4346  */
4347
4348 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4349 {
4350         struct r5conf *conf = sh->raid_conf;
4351         int disks = sh->disks;
4352         struct r5dev *dev;
4353         int i;
4354         int do_recovery = 0;
4355
4356         memset(s, 0, sizeof(*s));
4357
4358         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4359         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4360         s->failed_num[0] = -1;
4361         s->failed_num[1] = -1;
4362         s->log_failed = r5l_log_disk_error(conf);
4363
4364         /* Now to look around and see what can be done */
4365         rcu_read_lock();
4366         for (i=disks; i--; ) {
4367                 struct md_rdev *rdev;
4368                 sector_t first_bad;
4369                 int bad_sectors;
4370                 int is_bad = 0;
4371
4372                 dev = &sh->dev[i];
4373
4374                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4375                          i, dev->flags,
4376                          dev->toread, dev->towrite, dev->written);
4377                 /* maybe we can reply to a read
4378                  *
4379                  * new wantfill requests are only permitted while
4380                  * ops_complete_biofill is guaranteed to be inactive
4381                  */
4382                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4383                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4384                         set_bit(R5_Wantfill, &dev->flags);
4385
4386                 /* now count some things */
4387                 if (test_bit(R5_LOCKED, &dev->flags))
4388                         s->locked++;
4389                 if (test_bit(R5_UPTODATE, &dev->flags))
4390                         s->uptodate++;
4391                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4392                         s->compute++;
4393                         BUG_ON(s->compute > 2);
4394                 }
4395
4396                 if (test_bit(R5_Wantfill, &dev->flags))
4397                         s->to_fill++;
4398                 else if (dev->toread)
4399                         s->to_read++;
4400                 if (dev->towrite) {
4401                         s->to_write++;
4402                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4403                                 s->non_overwrite++;
4404                 }
4405                 if (dev->written)
4406                         s->written++;
4407                 /* Prefer to use the replacement for reads, but only
4408                  * if it is recovered enough and has no bad blocks.
4409                  */
4410                 rdev = rcu_dereference(conf->disks[i].replacement);
4411                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4412                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4413                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4414                                  &first_bad, &bad_sectors))
4415                         set_bit(R5_ReadRepl, &dev->flags);
4416                 else {
4417                         if (rdev && !test_bit(Faulty, &rdev->flags))
4418                                 set_bit(R5_NeedReplace, &dev->flags);
4419                         else
4420                                 clear_bit(R5_NeedReplace, &dev->flags);
4421                         rdev = rcu_dereference(conf->disks[i].rdev);
4422                         clear_bit(R5_ReadRepl, &dev->flags);
4423                 }
4424                 if (rdev && test_bit(Faulty, &rdev->flags))
4425                         rdev = NULL;
4426                 if (rdev) {
4427                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4428                                              &first_bad, &bad_sectors);
4429                         if (s->blocked_rdev == NULL
4430                             && (test_bit(Blocked, &rdev->flags)
4431                                 || is_bad < 0)) {
4432                                 if (is_bad < 0)
4433                                         set_bit(BlockedBadBlocks,
4434                                                 &rdev->flags);
4435                                 s->blocked_rdev = rdev;
4436                                 atomic_inc(&rdev->nr_pending);
4437                         }
4438                 }
4439                 clear_bit(R5_Insync, &dev->flags);
4440                 if (!rdev)
4441                         /* Not in-sync */;
4442                 else if (is_bad) {
4443                         /* also not in-sync */
4444                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4445                             test_bit(R5_UPTODATE, &dev->flags)) {
4446                                 /* treat as in-sync, but with a read error
4447                                  * which we can now try to correct
4448                                  */
4449                                 set_bit(R5_Insync, &dev->flags);
4450                                 set_bit(R5_ReadError, &dev->flags);
4451                         }
4452                 } else if (test_bit(In_sync, &rdev->flags))
4453                         set_bit(R5_Insync, &dev->flags);
4454                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4455                         /* in sync if before recovery_offset */
4456                         set_bit(R5_Insync, &dev->flags);
4457                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4458                          test_bit(R5_Expanded, &dev->flags))
4459                         /* If we've reshaped into here, we assume it is Insync.
4460                          * We will shortly update recovery_offset to make
4461                          * it official.
4462                          */
4463                         set_bit(R5_Insync, &dev->flags);
4464
4465                 if (test_bit(R5_WriteError, &dev->flags)) {
4466                         /* This flag does not apply to '.replacement'
4467                          * only to .rdev, so make sure to check that*/
4468                         struct md_rdev *rdev2 = rcu_dereference(
4469                                 conf->disks[i].rdev);
4470                         if (rdev2 == rdev)
4471                                 clear_bit(R5_Insync, &dev->flags);
4472                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4473                                 s->handle_bad_blocks = 1;
4474                                 atomic_inc(&rdev2->nr_pending);
4475                         } else
4476                                 clear_bit(R5_WriteError, &dev->flags);
4477                 }
4478                 if (test_bit(R5_MadeGood, &dev->flags)) {
4479                         /* This flag does not apply to '.replacement'
4480                          * only to .rdev, so make sure to check that*/
4481                         struct md_rdev *rdev2 = rcu_dereference(
4482                                 conf->disks[i].rdev);
4483                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4484                                 s->handle_bad_blocks = 1;
4485                                 atomic_inc(&rdev2->nr_pending);
4486                         } else
4487                                 clear_bit(R5_MadeGood, &dev->flags);
4488                 }
4489                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4490                         struct md_rdev *rdev2 = rcu_dereference(
4491                                 conf->disks[i].replacement);
4492                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4493                                 s->handle_bad_blocks = 1;
4494                                 atomic_inc(&rdev2->nr_pending);
4495                         } else
4496                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4497                 }
4498                 if (!test_bit(R5_Insync, &dev->flags)) {
4499                         /* The ReadError flag will just be confusing now */
4500                         clear_bit(R5_ReadError, &dev->flags);
4501                         clear_bit(R5_ReWrite, &dev->flags);
4502                 }
4503                 if (test_bit(R5_ReadError, &dev->flags))
4504                         clear_bit(R5_Insync, &dev->flags);
4505                 if (!test_bit(R5_Insync, &dev->flags)) {
4506                         if (s->failed < 2)
4507                                 s->failed_num[s->failed] = i;
4508                         s->failed++;
4509                         if (rdev && !test_bit(Faulty, &rdev->flags))
4510                                 do_recovery = 1;
4511                 }
4512
4513                 if (test_bit(R5_InJournal, &dev->flags))
4514                         s->injournal++;
4515                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4516                         s->just_cached++;
4517         }
4518         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4519                 /* If there is a failed device being replaced,
4520                  *     we must be recovering.
4521                  * else if we are after recovery_cp, we must be syncing
4522                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4523                  * else we can only be replacing
4524                  * sync and recovery both need to read all devices, and so
4525                  * use the same flag.
4526                  */
4527                 if (do_recovery ||
4528                     sh->sector >= conf->mddev->recovery_cp ||
4529                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4530                         s->syncing = 1;
4531                 else
4532                         s->replacing = 1;
4533         }
4534         rcu_read_unlock();
4535 }
4536
4537 static int clear_batch_ready(struct stripe_head *sh)
4538 {
4539         /* Return '1' if this is a member of batch, or
4540          * '0' if it is a lone stripe or a head which can now be
4541          * handled.
4542          */
4543         struct stripe_head *tmp;
4544         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4545                 return (sh->batch_head && sh->batch_head != sh);
4546         spin_lock(&sh->stripe_lock);
4547         if (!sh->batch_head) {
4548                 spin_unlock(&sh->stripe_lock);
4549                 return 0;
4550         }
4551
4552         /*
4553          * this stripe could be added to a batch list before we check
4554          * BATCH_READY, skips it
4555          */
4556         if (sh->batch_head != sh) {
4557                 spin_unlock(&sh->stripe_lock);
4558                 return 1;
4559         }
4560         spin_lock(&sh->batch_lock);
4561         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4562                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4563         spin_unlock(&sh->batch_lock);
4564         spin_unlock(&sh->stripe_lock);
4565
4566         /*
4567          * BATCH_READY is cleared, no new stripes can be added.
4568          * batch_list can be accessed without lock
4569          */
4570         return 0;
4571 }
4572
4573 static void break_stripe_batch_list(struct stripe_head *head_sh,
4574                                     unsigned long handle_flags)
4575 {
4576         struct stripe_head *sh, *next;
4577         int i;
4578         int do_wakeup = 0;
4579
4580         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4581
4582                 list_del_init(&sh->batch_list);
4583
4584                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4585                                           (1 << STRIPE_SYNCING) |
4586                                           (1 << STRIPE_REPLACED) |
4587                                           (1 << STRIPE_DELAYED) |
4588                                           (1 << STRIPE_BIT_DELAY) |
4589                                           (1 << STRIPE_FULL_WRITE) |
4590                                           (1 << STRIPE_BIOFILL_RUN) |
4591                                           (1 << STRIPE_COMPUTE_RUN)  |
4592                                           (1 << STRIPE_OPS_REQ_PENDING) |
4593                                           (1 << STRIPE_DISCARD) |
4594                                           (1 << STRIPE_BATCH_READY) |
4595                                           (1 << STRIPE_BATCH_ERR) |
4596                                           (1 << STRIPE_BITMAP_PENDING)),
4597                         "stripe state: %lx\n", sh->state);
4598                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4599                                               (1 << STRIPE_REPLACED)),
4600                         "head stripe state: %lx\n", head_sh->state);
4601
4602                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4603                                             (1 << STRIPE_PREREAD_ACTIVE) |
4604                                             (1 << STRIPE_DEGRADED)),
4605                               head_sh->state & (1 << STRIPE_INSYNC));
4606
4607                 sh->check_state = head_sh->check_state;
4608                 sh->reconstruct_state = head_sh->reconstruct_state;
4609                 for (i = 0; i < sh->disks; i++) {
4610                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4611                                 do_wakeup = 1;
4612                         sh->dev[i].flags = head_sh->dev[i].flags &
4613                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4614                 }
4615                 spin_lock_irq(&sh->stripe_lock);
4616                 sh->batch_head = NULL;
4617                 spin_unlock_irq(&sh->stripe_lock);
4618                 if (handle_flags == 0 ||
4619                     sh->state & handle_flags)
4620                         set_bit(STRIPE_HANDLE, &sh->state);
4621                 raid5_release_stripe(sh);
4622         }
4623         spin_lock_irq(&head_sh->stripe_lock);
4624         head_sh->batch_head = NULL;
4625         spin_unlock_irq(&head_sh->stripe_lock);
4626         for (i = 0; i < head_sh->disks; i++)
4627                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4628                         do_wakeup = 1;
4629         if (head_sh->state & handle_flags)
4630                 set_bit(STRIPE_HANDLE, &head_sh->state);
4631
4632         if (do_wakeup)
4633                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4634 }
4635
4636 static void handle_stripe(struct stripe_head *sh)
4637 {
4638         struct stripe_head_state s;
4639         struct r5conf *conf = sh->raid_conf;
4640         int i;
4641         int prexor;
4642         int disks = sh->disks;
4643         struct r5dev *pdev, *qdev;
4644
4645         clear_bit(STRIPE_HANDLE, &sh->state);
4646         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4647                 /* already being handled, ensure it gets handled
4648                  * again when current action finishes */
4649                 set_bit(STRIPE_HANDLE, &sh->state);
4650                 return;
4651         }
4652
4653         if (clear_batch_ready(sh) ) {
4654                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4655                 return;
4656         }
4657
4658         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4659                 break_stripe_batch_list(sh, 0);
4660
4661         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4662                 spin_lock(&sh->stripe_lock);
4663                 /*
4664                  * Cannot process 'sync' concurrently with 'discard'.
4665                  * Flush data in r5cache before 'sync'.
4666                  */
4667                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4668                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4669                     !test_bit(STRIPE_DISCARD, &sh->state) &&
4670                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4671                         set_bit(STRIPE_SYNCING, &sh->state);
4672                         clear_bit(STRIPE_INSYNC, &sh->state);
4673                         clear_bit(STRIPE_REPLACED, &sh->state);
4674                 }
4675                 spin_unlock(&sh->stripe_lock);
4676         }
4677         clear_bit(STRIPE_DELAYED, &sh->state);
4678
4679         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4680                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4681                (unsigned long long)sh->sector, sh->state,
4682                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4683                sh->check_state, sh->reconstruct_state);
4684
4685         analyse_stripe(sh, &s);
4686
4687         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4688                 goto finish;
4689
4690         if (s.handle_bad_blocks ||
4691             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4692                 set_bit(STRIPE_HANDLE, &sh->state);
4693                 goto finish;
4694         }
4695
4696         if (unlikely(s.blocked_rdev)) {
4697                 if (s.syncing || s.expanding || s.expanded ||
4698                     s.replacing || s.to_write || s.written) {
4699                         set_bit(STRIPE_HANDLE, &sh->state);
4700                         goto finish;
4701                 }
4702                 /* There is nothing for the blocked_rdev to block */
4703                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4704                 s.blocked_rdev = NULL;
4705         }
4706
4707         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4708                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4709                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4710         }
4711
4712         pr_debug("locked=%d uptodate=%d to_read=%d"
4713                " to_write=%d failed=%d failed_num=%d,%d\n",
4714                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4715                s.failed_num[0], s.failed_num[1]);
4716         /*
4717          * check if the array has lost more than max_degraded devices and,
4718          * if so, some requests might need to be failed.
4719          *
4720          * When journal device failed (log_failed), we will only process
4721          * the stripe if there is data need write to raid disks
4722          */
4723         if (s.failed > conf->max_degraded ||
4724             (s.log_failed && s.injournal == 0)) {
4725                 sh->check_state = 0;
4726                 sh->reconstruct_state = 0;
4727                 break_stripe_batch_list(sh, 0);
4728                 if (s.to_read+s.to_write+s.written)
4729                         handle_failed_stripe(conf, sh, &s, disks);
4730                 if (s.syncing + s.replacing)
4731                         handle_failed_sync(conf, sh, &s);
4732         }
4733
4734         /* Now we check to see if any write operations have recently
4735          * completed
4736          */
4737         prexor = 0;
4738         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4739                 prexor = 1;
4740         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4741             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4742                 sh->reconstruct_state = reconstruct_state_idle;
4743
4744                 /* All the 'written' buffers and the parity block are ready to
4745                  * be written back to disk
4746                  */
4747                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4748                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4749                 BUG_ON(sh->qd_idx >= 0 &&
4750                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4751                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4752                 for (i = disks; i--; ) {
4753                         struct r5dev *dev = &sh->dev[i];
4754                         if (test_bit(R5_LOCKED, &dev->flags) &&
4755                                 (i == sh->pd_idx || i == sh->qd_idx ||
4756                                  dev->written || test_bit(R5_InJournal,
4757                                                           &dev->flags))) {
4758                                 pr_debug("Writing block %d\n", i);
4759                                 set_bit(R5_Wantwrite, &dev->flags);
4760                                 if (prexor)
4761                                         continue;
4762                                 if (s.failed > 1)
4763                                         continue;
4764                                 if (!test_bit(R5_Insync, &dev->flags) ||
4765                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4766                                      s.failed == 0))
4767                                         set_bit(STRIPE_INSYNC, &sh->state);
4768                         }
4769                 }
4770                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4771                         s.dec_preread_active = 1;
4772         }
4773
4774         /*
4775          * might be able to return some write requests if the parity blocks
4776          * are safe, or on a failed drive
4777          */
4778         pdev = &sh->dev[sh->pd_idx];
4779         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4780                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4781         qdev = &sh->dev[sh->qd_idx];
4782         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4783                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4784                 || conf->level < 6;
4785
4786         if (s.written &&
4787             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4788                              && !test_bit(R5_LOCKED, &pdev->flags)
4789                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4790                                  test_bit(R5_Discard, &pdev->flags))))) &&
4791             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4792                              && !test_bit(R5_LOCKED, &qdev->flags)
4793                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4794                                  test_bit(R5_Discard, &qdev->flags))))))
4795                 handle_stripe_clean_event(conf, sh, disks);
4796
4797         if (s.just_cached)
4798                 r5c_handle_cached_data_endio(conf, sh, disks);
4799         log_stripe_write_finished(sh);
4800
4801         /* Now we might consider reading some blocks, either to check/generate
4802          * parity, or to satisfy requests
4803          * or to load a block that is being partially written.
4804          */
4805         if (s.to_read || s.non_overwrite
4806             || (conf->level == 6 && s.to_write && s.failed)
4807             || (s.syncing && (s.uptodate + s.compute < disks))
4808             || s.replacing
4809             || s.expanding)
4810                 handle_stripe_fill(sh, &s, disks);
4811
4812         /*
4813          * When the stripe finishes full journal write cycle (write to journal
4814          * and raid disk), this is the clean up procedure so it is ready for
4815          * next operation.
4816          */
4817         r5c_finish_stripe_write_out(conf, sh, &s);
4818
4819         /*
4820          * Now to consider new write requests, cache write back and what else,
4821          * if anything should be read.  We do not handle new writes when:
4822          * 1/ A 'write' operation (copy+xor) is already in flight.
4823          * 2/ A 'check' operation is in flight, as it may clobber the parity
4824          *    block.
4825          * 3/ A r5c cache log write is in flight.
4826          */
4827
4828         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4829                 if (!r5c_is_writeback(conf->log)) {
4830                         if (s.to_write)
4831                                 handle_stripe_dirtying(conf, sh, &s, disks);
4832                 } else { /* write back cache */
4833                         int ret = 0;
4834
4835                         /* First, try handle writes in caching phase */
4836                         if (s.to_write)
4837                                 ret = r5c_try_caching_write(conf, sh, &s,
4838                                                             disks);
4839                         /*
4840                          * If caching phase failed: ret == -EAGAIN
4841                          *    OR
4842                          * stripe under reclaim: !caching && injournal
4843                          *
4844                          * fall back to handle_stripe_dirtying()
4845                          */
4846                         if (ret == -EAGAIN ||
4847                             /* stripe under reclaim: !caching && injournal */
4848                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4849                              s.injournal > 0)) {
4850                                 ret = handle_stripe_dirtying(conf, sh, &s,
4851                                                              disks);
4852                                 if (ret == -EAGAIN)
4853                                         goto finish;
4854                         }
4855                 }
4856         }
4857
4858         /* maybe we need to check and possibly fix the parity for this stripe
4859          * Any reads will already have been scheduled, so we just see if enough
4860          * data is available.  The parity check is held off while parity
4861          * dependent operations are in flight.
4862          */
4863         if (sh->check_state ||
4864             (s.syncing && s.locked == 0 &&
4865              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4866              !test_bit(STRIPE_INSYNC, &sh->state))) {
4867                 if (conf->level == 6)
4868                         handle_parity_checks6(conf, sh, &s, disks);
4869                 else
4870                         handle_parity_checks5(conf, sh, &s, disks);
4871         }
4872
4873         if ((s.replacing || s.syncing) && s.locked == 0
4874             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4875             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4876                 /* Write out to replacement devices where possible */
4877                 for (i = 0; i < conf->raid_disks; i++)
4878                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4879                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4880                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4881                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4882                                 s.locked++;
4883                         }
4884                 if (s.replacing)
4885                         set_bit(STRIPE_INSYNC, &sh->state);
4886                 set_bit(STRIPE_REPLACED, &sh->state);
4887         }
4888         if ((s.syncing || s.replacing) && s.locked == 0 &&
4889             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4890             test_bit(STRIPE_INSYNC, &sh->state)) {
4891                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4892                 clear_bit(STRIPE_SYNCING, &sh->state);
4893                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4894                         wake_up(&conf->wait_for_overlap);
4895         }
4896
4897         /* If the failed drives are just a ReadError, then we might need
4898          * to progress the repair/check process
4899          */
4900         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4901                 for (i = 0; i < s.failed; i++) {
4902                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4903                         if (test_bit(R5_ReadError, &dev->flags)
4904                             && !test_bit(R5_LOCKED, &dev->flags)
4905                             && test_bit(R5_UPTODATE, &dev->flags)
4906                                 ) {
4907                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4908                                         set_bit(R5_Wantwrite, &dev->flags);
4909                                         set_bit(R5_ReWrite, &dev->flags);
4910                                         set_bit(R5_LOCKED, &dev->flags);
4911                                         s.locked++;
4912                                 } else {
4913                                         /* let's read it back */
4914                                         set_bit(R5_Wantread, &dev->flags);
4915                                         set_bit(R5_LOCKED, &dev->flags);
4916                                         s.locked++;
4917                                 }
4918                         }
4919                 }
4920
4921         /* Finish reconstruct operations initiated by the expansion process */
4922         if (sh->reconstruct_state == reconstruct_state_result) {
4923                 struct stripe_head *sh_src
4924                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4925                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4926                         /* sh cannot be written until sh_src has been read.
4927                          * so arrange for sh to be delayed a little
4928                          */
4929                         set_bit(STRIPE_DELAYED, &sh->state);
4930                         set_bit(STRIPE_HANDLE, &sh->state);
4931                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4932                                               &sh_src->state))
4933                                 atomic_inc(&conf->preread_active_stripes);
4934                         raid5_release_stripe(sh_src);
4935                         goto finish;
4936                 }
4937                 if (sh_src)
4938                         raid5_release_stripe(sh_src);
4939
4940                 sh->reconstruct_state = reconstruct_state_idle;
4941                 clear_bit(STRIPE_EXPANDING, &sh->state);
4942                 for (i = conf->raid_disks; i--; ) {
4943                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4944                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4945                         s.locked++;
4946                 }
4947         }
4948
4949         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4950             !sh->reconstruct_state) {
4951                 /* Need to write out all blocks after computing parity */
4952                 sh->disks = conf->raid_disks;
4953                 stripe_set_idx(sh->sector, conf, 0, sh);
4954                 schedule_reconstruction(sh, &s, 1, 1);
4955         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4956                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4957                 atomic_dec(&conf->reshape_stripes);
4958                 wake_up(&conf->wait_for_overlap);
4959                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4960         }
4961
4962         if (s.expanding && s.locked == 0 &&
4963             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4964                 handle_stripe_expansion(conf, sh);
4965
4966 finish:
4967         /* wait for this device to become unblocked */
4968         if (unlikely(s.blocked_rdev)) {
4969                 if (conf->mddev->external)
4970                         md_wait_for_blocked_rdev(s.blocked_rdev,
4971                                                  conf->mddev);
4972                 else
4973                         /* Internal metadata will immediately
4974                          * be written by raid5d, so we don't
4975                          * need to wait here.
4976                          */
4977                         rdev_dec_pending(s.blocked_rdev,
4978                                          conf->mddev);
4979         }
4980
4981         if (s.handle_bad_blocks)
4982                 for (i = disks; i--; ) {
4983                         struct md_rdev *rdev;
4984                         struct r5dev *dev = &sh->dev[i];
4985                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4986                                 /* We own a safe reference to the rdev */
4987                                 rdev = conf->disks[i].rdev;
4988                                 if (!rdev_set_badblocks(rdev, sh->sector,
4989                                                         STRIPE_SECTORS, 0))
4990                                         md_error(conf->mddev, rdev);
4991                                 rdev_dec_pending(rdev, conf->mddev);
4992                         }
4993                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4994                                 rdev = conf->disks[i].rdev;
4995                                 rdev_clear_badblocks(rdev, sh->sector,
4996                                                      STRIPE_SECTORS, 0);
4997                                 rdev_dec_pending(rdev, conf->mddev);
4998                         }
4999                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5000                                 rdev = conf->disks[i].replacement;
5001                                 if (!rdev)
5002                                         /* rdev have been moved down */
5003                                         rdev = conf->disks[i].rdev;
5004                                 rdev_clear_badblocks(rdev, sh->sector,
5005                                                      STRIPE_SECTORS, 0);
5006                                 rdev_dec_pending(rdev, conf->mddev);
5007                         }
5008                 }
5009
5010         if (s.ops_request)
5011                 raid_run_ops(sh, s.ops_request);
5012
5013         ops_run_io(sh, &s);
5014
5015         if (s.dec_preread_active) {
5016                 /* We delay this until after ops_run_io so that if make_request
5017                  * is waiting on a flush, it won't continue until the writes
5018                  * have actually been submitted.
5019                  */
5020                 atomic_dec(&conf->preread_active_stripes);
5021                 if (atomic_read(&conf->preread_active_stripes) <
5022                     IO_THRESHOLD)
5023                         md_wakeup_thread(conf->mddev->thread);
5024         }
5025
5026         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5027 }
5028
5029 static void raid5_activate_delayed(struct r5conf *conf)
5030 {
5031         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5032                 while (!list_empty(&conf->delayed_list)) {
5033                         struct list_head *l = conf->delayed_list.next;
5034                         struct stripe_head *sh;
5035                         sh = list_entry(l, struct stripe_head, lru);
5036                         list_del_init(l);
5037                         clear_bit(STRIPE_DELAYED, &sh->state);
5038                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5039                                 atomic_inc(&conf->preread_active_stripes);
5040                         list_add_tail(&sh->lru, &conf->hold_list);
5041                         raid5_wakeup_stripe_thread(sh);
5042                 }
5043         }
5044 }
5045
5046 static void activate_bit_delay(struct r5conf *conf,
5047         struct list_head *temp_inactive_list)
5048 {
5049         /* device_lock is held */
5050         struct list_head head;
5051         list_add(&head, &conf->bitmap_list);
5052         list_del_init(&conf->bitmap_list);
5053         while (!list_empty(&head)) {
5054                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5055                 int hash;
5056                 list_del_init(&sh->lru);
5057                 atomic_inc(&sh->count);
5058                 hash = sh->hash_lock_index;
5059                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5060         }
5061 }
5062
5063 static int raid5_congested(struct mddev *mddev, int bits)
5064 {
5065         struct r5conf *conf = mddev->private;
5066
5067         /* No difference between reads and writes.  Just check
5068          * how busy the stripe_cache is
5069          */
5070
5071         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5072                 return 1;
5073
5074         /* Also checks whether there is pressure on r5cache log space */
5075         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5076                 return 1;
5077         if (conf->quiesce)
5078                 return 1;
5079         if (atomic_read(&conf->empty_inactive_list_nr))
5080                 return 1;
5081
5082         return 0;
5083 }
5084
5085 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5086 {
5087         struct r5conf *conf = mddev->private;
5088         sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
5089         unsigned int chunk_sectors;
5090         unsigned int bio_sectors = bio_sectors(bio);
5091
5092         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5093         return  chunk_sectors >=
5094                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5095 }
5096
5097 /*
5098  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5099  *  later sampled by raid5d.
5100  */
5101 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5102 {
5103         unsigned long flags;
5104
5105         spin_lock_irqsave(&conf->device_lock, flags);
5106
5107         bi->bi_next = conf->retry_read_aligned_list;
5108         conf->retry_read_aligned_list = bi;
5109
5110         spin_unlock_irqrestore(&conf->device_lock, flags);
5111         md_wakeup_thread(conf->mddev->thread);
5112 }
5113
5114 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5115                                          unsigned int *offset)
5116 {
5117         struct bio *bi;
5118
5119         bi = conf->retry_read_aligned;
5120         if (bi) {
5121                 *offset = conf->retry_read_offset;
5122                 conf->retry_read_aligned = NULL;
5123                 return bi;
5124         }
5125         bi = conf->retry_read_aligned_list;
5126         if(bi) {
5127                 conf->retry_read_aligned_list = bi->bi_next;
5128                 bi->bi_next = NULL;
5129                 *offset = 0;
5130         }
5131
5132         return bi;
5133 }
5134
5135 /*
5136  *  The "raid5_align_endio" should check if the read succeeded and if it
5137  *  did, call bio_endio on the original bio (having bio_put the new bio
5138  *  first).
5139  *  If the read failed..
5140  */
5141 static void raid5_align_endio(struct bio *bi)
5142 {
5143         struct bio* raid_bi  = bi->bi_private;
5144         struct mddev *mddev;
5145         struct r5conf *conf;
5146         struct md_rdev *rdev;
5147         int error = bi->bi_error;
5148
5149         bio_put(bi);
5150
5151         rdev = (void*)raid_bi->bi_next;
5152         raid_bi->bi_next = NULL;
5153         mddev = rdev->mddev;
5154         conf = mddev->private;
5155
5156         rdev_dec_pending(rdev, conf->mddev);
5157
5158         if (!error) {
5159                 bio_endio(raid_bi);
5160                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5161                         wake_up(&conf->wait_for_quiescent);
5162                 return;
5163         }
5164
5165         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5166
5167         add_bio_to_retry(raid_bi, conf);
5168 }
5169
5170 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5171 {
5172         struct r5conf *conf = mddev->private;
5173         int dd_idx;
5174         struct bio* align_bi;
5175         struct md_rdev *rdev;
5176         sector_t end_sector;
5177
5178         if (!in_chunk_boundary(mddev, raid_bio)) {
5179                 pr_debug("%s: non aligned\n", __func__);
5180                 return 0;
5181         }
5182         /*
5183          * use bio_clone_fast to make a copy of the bio
5184          */
5185         align_bi = bio_clone_fast(raid_bio, GFP_NOIO, mddev->bio_set);
5186         if (!align_bi)
5187                 return 0;
5188         /*
5189          *   set bi_end_io to a new function, and set bi_private to the
5190          *     original bio.
5191          */
5192         align_bi->bi_end_io  = raid5_align_endio;
5193         align_bi->bi_private = raid_bio;
5194         /*
5195          *      compute position
5196          */
5197         align_bi->bi_iter.bi_sector =
5198                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5199                                      0, &dd_idx, NULL);
5200
5201         end_sector = bio_end_sector(align_bi);
5202         rcu_read_lock();
5203         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5204         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5205             rdev->recovery_offset < end_sector) {
5206                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5207                 if (rdev &&
5208                     (test_bit(Faulty, &rdev->flags) ||
5209                     !(test_bit(In_sync, &rdev->flags) ||
5210                       rdev->recovery_offset >= end_sector)))
5211                         rdev = NULL;
5212         }
5213
5214         if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5215                 rcu_read_unlock();
5216                 bio_put(align_bi);
5217                 return 0;
5218         }
5219
5220         if (rdev) {
5221                 sector_t first_bad;
5222                 int bad_sectors;
5223
5224                 atomic_inc(&rdev->nr_pending);
5225                 rcu_read_unlock();
5226                 raid_bio->bi_next = (void*)rdev;
5227                 align_bi->bi_bdev =  rdev->bdev;
5228                 bio_clear_flag(align_bi, BIO_SEG_VALID);
5229
5230                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5231                                 bio_sectors(align_bi),
5232                                 &first_bad, &bad_sectors)) {
5233                         bio_put(align_bi);
5234                         rdev_dec_pending(rdev, mddev);
5235                         return 0;
5236                 }
5237
5238                 /* No reshape active, so we can trust rdev->data_offset */
5239                 align_bi->bi_iter.bi_sector += rdev->data_offset;
5240
5241                 spin_lock_irq(&conf->device_lock);
5242                 wait_event_lock_irq(conf->wait_for_quiescent,
5243                                     conf->quiesce == 0,
5244                                     conf->device_lock);
5245                 atomic_inc(&conf->active_aligned_reads);
5246                 spin_unlock_irq(&conf->device_lock);
5247
5248                 if (mddev->gendisk)
5249                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
5250                                               align_bi, disk_devt(mddev->gendisk),
5251                                               raid_bio->bi_iter.bi_sector);
5252                 generic_make_request(align_bi);
5253                 return 1;
5254         } else {
5255                 rcu_read_unlock();
5256                 bio_put(align_bi);
5257                 return 0;
5258         }
5259 }
5260
5261 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5262 {
5263         struct bio *split;
5264         sector_t sector = raid_bio->bi_iter.bi_sector;
5265         unsigned chunk_sects = mddev->chunk_sectors;
5266         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5267
5268         if (sectors < bio_sectors(raid_bio)) {
5269                 struct r5conf *conf = mddev->private;
5270                 split = bio_split(raid_bio, sectors, GFP_NOIO, conf->bio_split);
5271                 bio_chain(split, raid_bio);
5272                 generic_make_request(raid_bio);
5273                 raid_bio = split;
5274         }
5275
5276         if (!raid5_read_one_chunk(mddev, raid_bio))
5277                 return raid_bio;
5278
5279         return NULL;
5280 }
5281
5282 /* __get_priority_stripe - get the next stripe to process
5283  *
5284  * Full stripe writes are allowed to pass preread active stripes up until
5285  * the bypass_threshold is exceeded.  In general the bypass_count
5286  * increments when the handle_list is handled before the hold_list; however, it
5287  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5288  * stripe with in flight i/o.  The bypass_count will be reset when the
5289  * head of the hold_list has changed, i.e. the head was promoted to the
5290  * handle_list.
5291  */
5292 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5293 {
5294         struct stripe_head *sh, *tmp;
5295         struct list_head *handle_list = NULL;
5296         struct r5worker_group *wg;
5297         bool second_try = !r5c_is_writeback(conf->log) &&
5298                 !r5l_log_disk_error(conf);
5299         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5300                 r5l_log_disk_error(conf);
5301
5302 again:
5303         wg = NULL;
5304         sh = NULL;
5305         if (conf->worker_cnt_per_group == 0) {
5306                 handle_list = try_loprio ? &conf->loprio_list :
5307                                         &conf->handle_list;
5308         } else if (group != ANY_GROUP) {
5309                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5310                                 &conf->worker_groups[group].handle_list;
5311                 wg = &conf->worker_groups[group];
5312         } else {
5313                 int i;
5314                 for (i = 0; i < conf->group_cnt; i++) {
5315                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5316                                 &conf->worker_groups[i].handle_list;
5317                         wg = &conf->worker_groups[i];
5318                         if (!list_empty(handle_list))
5319                                 break;
5320                 }
5321         }
5322
5323         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5324                   __func__,
5325                   list_empty(handle_list) ? "empty" : "busy",
5326                   list_empty(&conf->hold_list) ? "empty" : "busy",
5327                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5328
5329         if (!list_empty(handle_list)) {
5330                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5331
5332                 if (list_empty(&conf->hold_list))
5333                         conf->bypass_count = 0;
5334                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5335                         if (conf->hold_list.next == conf->last_hold)
5336                                 conf->bypass_count++;
5337                         else {
5338                                 conf->last_hold = conf->hold_list.next;
5339                                 conf->bypass_count -= conf->bypass_threshold;
5340                                 if (conf->bypass_count < 0)
5341                                         conf->bypass_count = 0;
5342                         }
5343                 }
5344         } else if (!list_empty(&conf->hold_list) &&
5345                    ((conf->bypass_threshold &&
5346                      conf->bypass_count > conf->bypass_threshold) ||
5347                     atomic_read(&conf->pending_full_writes) == 0)) {
5348
5349                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5350                         if (conf->worker_cnt_per_group == 0 ||
5351                             group == ANY_GROUP ||
5352                             !cpu_online(tmp->cpu) ||
5353                             cpu_to_group(tmp->cpu) == group) {
5354                                 sh = tmp;
5355                                 break;
5356                         }
5357                 }
5358
5359                 if (sh) {
5360                         conf->bypass_count -= conf->bypass_threshold;
5361                         if (conf->bypass_count < 0)
5362                                 conf->bypass_count = 0;
5363                 }
5364                 wg = NULL;
5365         }
5366
5367         if (!sh) {
5368                 if (second_try)
5369                         return NULL;
5370                 second_try = true;
5371                 try_loprio = !try_loprio;
5372                 goto again;
5373         }
5374
5375         if (wg) {
5376                 wg->stripes_cnt--;
5377                 sh->group = NULL;
5378         }
5379         list_del_init(&sh->lru);
5380         BUG_ON(atomic_inc_return(&sh->count) != 1);
5381         return sh;
5382 }
5383
5384 struct raid5_plug_cb {
5385         struct blk_plug_cb      cb;
5386         struct list_head        list;
5387         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5388 };
5389
5390 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5391 {
5392         struct raid5_plug_cb *cb = container_of(
5393                 blk_cb, struct raid5_plug_cb, cb);
5394         struct stripe_head *sh;
5395         struct mddev *mddev = cb->cb.data;
5396         struct r5conf *conf = mddev->private;
5397         int cnt = 0;
5398         int hash;
5399
5400         if (cb->list.next && !list_empty(&cb->list)) {
5401                 spin_lock_irq(&conf->device_lock);
5402                 while (!list_empty(&cb->list)) {
5403                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5404                         list_del_init(&sh->lru);
5405                         /*
5406                          * avoid race release_stripe_plug() sees
5407                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5408                          * is still in our list
5409                          */
5410                         smp_mb__before_atomic();
5411                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5412                         /*
5413                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5414                          * case, the count is always > 1 here
5415                          */
5416                         hash = sh->hash_lock_index;
5417                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5418                         cnt++;
5419                 }
5420                 spin_unlock_irq(&conf->device_lock);
5421         }
5422         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5423                                      NR_STRIPE_HASH_LOCKS);
5424         if (mddev->queue)
5425                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5426         kfree(cb);
5427 }
5428
5429 static void release_stripe_plug(struct mddev *mddev,
5430                                 struct stripe_head *sh)
5431 {
5432         struct blk_plug_cb *blk_cb = blk_check_plugged(
5433                 raid5_unplug, mddev,
5434                 sizeof(struct raid5_plug_cb));
5435         struct raid5_plug_cb *cb;
5436
5437         if (!blk_cb) {
5438                 raid5_release_stripe(sh);
5439                 return;
5440         }
5441
5442         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5443
5444         if (cb->list.next == NULL) {
5445                 int i;
5446                 INIT_LIST_HEAD(&cb->list);
5447                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5448                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5449         }
5450
5451         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5452                 list_add_tail(&sh->lru, &cb->list);
5453         else
5454                 raid5_release_stripe(sh);
5455 }
5456
5457 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5458 {
5459         struct r5conf *conf = mddev->private;
5460         sector_t logical_sector, last_sector;
5461         struct stripe_head *sh;
5462         int stripe_sectors;
5463
5464         if (mddev->reshape_position != MaxSector)
5465                 /* Skip discard while reshape is happening */
5466                 return;
5467
5468         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5469         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5470
5471         bi->bi_next = NULL;
5472         md_write_start(mddev, bi);
5473
5474         stripe_sectors = conf->chunk_sectors *
5475                 (conf->raid_disks - conf->max_degraded);
5476         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5477                                                stripe_sectors);
5478         sector_div(last_sector, stripe_sectors);
5479
5480         logical_sector *= conf->chunk_sectors;
5481         last_sector *= conf->chunk_sectors;
5482
5483         for (; logical_sector < last_sector;
5484              logical_sector += STRIPE_SECTORS) {
5485                 DEFINE_WAIT(w);
5486                 int d;
5487         again:
5488                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5489                 prepare_to_wait(&conf->wait_for_overlap, &w,
5490                                 TASK_UNINTERRUPTIBLE);
5491                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5492                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5493                         raid5_release_stripe(sh);
5494                         schedule();
5495                         goto again;
5496                 }
5497                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5498                 spin_lock_irq(&sh->stripe_lock);
5499                 for (d = 0; d < conf->raid_disks; d++) {
5500                         if (d == sh->pd_idx || d == sh->qd_idx)
5501                                 continue;
5502                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5503                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5504                                 spin_unlock_irq(&sh->stripe_lock);
5505                                 raid5_release_stripe(sh);
5506                                 schedule();
5507                                 goto again;
5508                         }
5509                 }
5510                 set_bit(STRIPE_DISCARD, &sh->state);
5511                 finish_wait(&conf->wait_for_overlap, &w);
5512                 sh->overwrite_disks = 0;
5513                 for (d = 0; d < conf->raid_disks; d++) {
5514                         if (d == sh->pd_idx || d == sh->qd_idx)
5515                                 continue;
5516                         sh->dev[d].towrite = bi;
5517                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5518                         bio_inc_remaining(bi);
5519                         md_write_inc(mddev, bi);
5520                         sh->overwrite_disks++;
5521                 }
5522                 spin_unlock_irq(&sh->stripe_lock);
5523                 if (conf->mddev->bitmap) {
5524                         for (d = 0;
5525                              d < conf->raid_disks - conf->max_degraded;
5526                              d++)
5527                                 bitmap_startwrite(mddev->bitmap,
5528                                                   sh->sector,
5529                                                   STRIPE_SECTORS,
5530                                                   0);
5531                         sh->bm_seq = conf->seq_flush + 1;
5532                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5533                 }
5534
5535                 set_bit(STRIPE_HANDLE, &sh->state);
5536                 clear_bit(STRIPE_DELAYED, &sh->state);
5537                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5538                         atomic_inc(&conf->preread_active_stripes);
5539                 release_stripe_plug(mddev, sh);
5540         }
5541
5542         md_write_end(mddev);
5543         bio_endio(bi);
5544 }
5545
5546 static void raid5_make_request(struct mddev *mddev, struct bio * bi)
5547 {
5548         struct r5conf *conf = mddev->private;
5549         int dd_idx;
5550         sector_t new_sector;
5551         sector_t logical_sector, last_sector;
5552         struct stripe_head *sh;
5553         const int rw = bio_data_dir(bi);
5554         DEFINE_WAIT(w);
5555         bool do_prepare;
5556         bool do_flush = false;
5557
5558         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5559                 int ret = r5l_handle_flush_request(conf->log, bi);
5560
5561                 if (ret == 0)
5562                         return;
5563                 if (ret == -ENODEV) {
5564                         md_flush_request(mddev, bi);
5565                         return;
5566                 }
5567                 /* ret == -EAGAIN, fallback */
5568                 /*
5569                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5570                  * we need to flush journal device
5571                  */
5572                 do_flush = bi->bi_opf & REQ_PREFLUSH;
5573         }
5574
5575         /*
5576          * If array is degraded, better not do chunk aligned read because
5577          * later we might have to read it again in order to reconstruct
5578          * data on failed drives.
5579          */
5580         if (rw == READ && mddev->degraded == 0 &&
5581             mddev->reshape_position == MaxSector) {
5582                 bi = chunk_aligned_read(mddev, bi);
5583                 if (!bi)
5584                         return;
5585         }
5586
5587         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5588                 make_discard_request(mddev, bi);
5589                 return;
5590         }
5591
5592         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5593         last_sector = bio_end_sector(bi);
5594         bi->bi_next = NULL;
5595         md_write_start(mddev, bi);
5596
5597         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5598         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5599                 int previous;
5600                 int seq;
5601
5602                 do_prepare = false;
5603         retry:
5604                 seq = read_seqcount_begin(&conf->gen_lock);
5605                 previous = 0;
5606                 if (do_prepare)
5607                         prepare_to_wait(&conf->wait_for_overlap, &w,
5608                                 TASK_UNINTERRUPTIBLE);
5609                 if (unlikely(conf->reshape_progress != MaxSector)) {
5610                         /* spinlock is needed as reshape_progress may be
5611                          * 64bit on a 32bit platform, and so it might be
5612                          * possible to see a half-updated value
5613                          * Of course reshape_progress could change after
5614                          * the lock is dropped, so once we get a reference
5615                          * to the stripe that we think it is, we will have
5616                          * to check again.
5617                          */
5618                         spin_lock_irq(&conf->device_lock);
5619                         if (mddev->reshape_backwards
5620                             ? logical_sector < conf->reshape_progress
5621                             : logical_sector >= conf->reshape_progress) {
5622                                 previous = 1;
5623                         } else {
5624                                 if (mddev->reshape_backwards
5625                                     ? logical_sector < conf->reshape_safe
5626                                     : logical_sector >= conf->reshape_safe) {
5627                                         spin_unlock_irq(&conf->device_lock);
5628                                         schedule();
5629                                         do_prepare = true;
5630                                         goto retry;
5631                                 }
5632                         }
5633                         spin_unlock_irq(&conf->device_lock);
5634                 }
5635
5636                 new_sector = raid5_compute_sector(conf, logical_sector,
5637                                                   previous,
5638                                                   &dd_idx, NULL);
5639                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5640                         (unsigned long long)new_sector,
5641                         (unsigned long long)logical_sector);
5642
5643                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5644                                        (bi->bi_opf & REQ_RAHEAD), 0);
5645                 if (sh) {
5646                         if (unlikely(previous)) {
5647                                 /* expansion might have moved on while waiting for a
5648                                  * stripe, so we must do the range check again.
5649                                  * Expansion could still move past after this
5650                                  * test, but as we are holding a reference to
5651                                  * 'sh', we know that if that happens,
5652                                  *  STRIPE_EXPANDING will get set and the expansion
5653                                  * won't proceed until we finish with the stripe.
5654                                  */
5655                                 int must_retry = 0;
5656                                 spin_lock_irq(&conf->device_lock);
5657                                 if (mddev->reshape_backwards
5658                                     ? logical_sector >= conf->reshape_progress
5659                                     : logical_sector < conf->reshape_progress)
5660                                         /* mismatch, need to try again */
5661                                         must_retry = 1;
5662                                 spin_unlock_irq(&conf->device_lock);
5663                                 if (must_retry) {
5664                                         raid5_release_stripe(sh);
5665                                         schedule();
5666                                         do_prepare = true;
5667                                         goto retry;
5668                                 }
5669                         }
5670                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5671                                 /* Might have got the wrong stripe_head
5672                                  * by accident
5673                                  */
5674                                 raid5_release_stripe(sh);
5675                                 goto retry;
5676                         }
5677
5678                         if (rw == WRITE &&
5679                             logical_sector >= mddev->suspend_lo &&
5680                             logical_sector < mddev->suspend_hi) {
5681                                 raid5_release_stripe(sh);
5682                                 /* As the suspend_* range is controlled by
5683                                  * userspace, we want an interruptible
5684                                  * wait.
5685                                  */
5686                                 flush_signals(current);
5687                                 prepare_to_wait(&conf->wait_for_overlap,
5688                                                 &w, TASK_INTERRUPTIBLE);
5689                                 if (logical_sector >= mddev->suspend_lo &&
5690                                     logical_sector < mddev->suspend_hi) {
5691                                         schedule();
5692                                         do_prepare = true;
5693                                 }
5694                                 goto retry;
5695                         }
5696
5697                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5698                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5699                                 /* Stripe is busy expanding or
5700                                  * add failed due to overlap.  Flush everything
5701                                  * and wait a while
5702                                  */
5703                                 md_wakeup_thread(mddev->thread);
5704                                 raid5_release_stripe(sh);
5705                                 schedule();
5706                                 do_prepare = true;
5707                                 goto retry;
5708                         }
5709                         if (do_flush) {
5710                                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5711                                 /* we only need flush for one stripe */
5712                                 do_flush = false;
5713                         }
5714
5715                         set_bit(STRIPE_HANDLE, &sh->state);
5716                         clear_bit(STRIPE_DELAYED, &sh->state);
5717                         if ((!sh->batch_head || sh == sh->batch_head) &&
5718                             (bi->bi_opf & REQ_SYNC) &&
5719                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5720                                 atomic_inc(&conf->preread_active_stripes);
5721                         release_stripe_plug(mddev, sh);
5722                 } else {
5723                         /* cannot get stripe for read-ahead, just give-up */
5724                         bi->bi_error = -EIO;
5725                         break;
5726                 }
5727         }
5728         finish_wait(&conf->wait_for_overlap, &w);
5729
5730         if (rw == WRITE)
5731                 md_write_end(mddev);
5732         bio_endio(bi);
5733 }
5734
5735 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5736
5737 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5738 {
5739         /* reshaping is quite different to recovery/resync so it is
5740          * handled quite separately ... here.
5741          *
5742          * On each call to sync_request, we gather one chunk worth of
5743          * destination stripes and flag them as expanding.
5744          * Then we find all the source stripes and request reads.
5745          * As the reads complete, handle_stripe will copy the data
5746          * into the destination stripe and release that stripe.
5747          */
5748         struct r5conf *conf = mddev->private;
5749         struct stripe_head *sh;
5750         sector_t first_sector, last_sector;
5751         int raid_disks = conf->previous_raid_disks;
5752         int data_disks = raid_disks - conf->max_degraded;
5753         int new_data_disks = conf->raid_disks - conf->max_degraded;
5754         int i;
5755         int dd_idx;
5756         sector_t writepos, readpos, safepos;
5757         sector_t stripe_addr;
5758         int reshape_sectors;
5759         struct list_head stripes;
5760         sector_t retn;
5761
5762         if (sector_nr == 0) {
5763                 /* If restarting in the middle, skip the initial sectors */
5764                 if (mddev->reshape_backwards &&
5765                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5766                         sector_nr = raid5_size(mddev, 0, 0)
5767                                 - conf->reshape_progress;
5768                 } else if (mddev->reshape_backwards &&
5769                            conf->reshape_progress == MaxSector) {
5770                         /* shouldn't happen, but just in case, finish up.*/
5771                         sector_nr = MaxSector;
5772                 } else if (!mddev->reshape_backwards &&
5773                            conf->reshape_progress > 0)
5774                         sector_nr = conf->reshape_progress;
5775                 sector_div(sector_nr, new_data_disks);
5776                 if (sector_nr) {
5777                         mddev->curr_resync_completed = sector_nr;
5778                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5779                         *skipped = 1;
5780                         retn = sector_nr;
5781                         goto finish;
5782                 }
5783         }
5784
5785         /* We need to process a full chunk at a time.
5786          * If old and new chunk sizes differ, we need to process the
5787          * largest of these
5788          */
5789
5790         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5791
5792         /* We update the metadata at least every 10 seconds, or when
5793          * the data about to be copied would over-write the source of
5794          * the data at the front of the range.  i.e. one new_stripe
5795          * along from reshape_progress new_maps to after where
5796          * reshape_safe old_maps to
5797          */
5798         writepos = conf->reshape_progress;
5799         sector_div(writepos, new_data_disks);
5800         readpos = conf->reshape_progress;
5801         sector_div(readpos, data_disks);
5802         safepos = conf->reshape_safe;
5803         sector_div(safepos, data_disks);
5804         if (mddev->reshape_backwards) {
5805                 BUG_ON(writepos < reshape_sectors);
5806                 writepos -= reshape_sectors;
5807                 readpos += reshape_sectors;
5808                 safepos += reshape_sectors;
5809         } else {
5810                 writepos += reshape_sectors;
5811                 /* readpos and safepos are worst-case calculations.
5812                  * A negative number is overly pessimistic, and causes
5813                  * obvious problems for unsigned storage.  So clip to 0.
5814                  */
5815                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5816                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5817         }
5818
5819         /* Having calculated the 'writepos' possibly use it
5820          * to set 'stripe_addr' which is where we will write to.
5821          */
5822         if (mddev->reshape_backwards) {
5823                 BUG_ON(conf->reshape_progress == 0);
5824                 stripe_addr = writepos;
5825                 BUG_ON((mddev->dev_sectors &
5826                         ~((sector_t)reshape_sectors - 1))
5827                        - reshape_sectors - stripe_addr
5828                        != sector_nr);
5829         } else {
5830                 BUG_ON(writepos != sector_nr + reshape_sectors);
5831                 stripe_addr = sector_nr;
5832         }
5833
5834         /* 'writepos' is the most advanced device address we might write.
5835          * 'readpos' is the least advanced device address we might read.
5836          * 'safepos' is the least address recorded in the metadata as having
5837          *     been reshaped.
5838          * If there is a min_offset_diff, these are adjusted either by
5839          * increasing the safepos/readpos if diff is negative, or
5840          * increasing writepos if diff is positive.
5841          * If 'readpos' is then behind 'writepos', there is no way that we can
5842          * ensure safety in the face of a crash - that must be done by userspace
5843          * making a backup of the data.  So in that case there is no particular
5844          * rush to update metadata.
5845          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5846          * update the metadata to advance 'safepos' to match 'readpos' so that
5847          * we can be safe in the event of a crash.
5848          * So we insist on updating metadata if safepos is behind writepos and
5849          * readpos is beyond writepos.
5850          * In any case, update the metadata every 10 seconds.
5851          * Maybe that number should be configurable, but I'm not sure it is
5852          * worth it.... maybe it could be a multiple of safemode_delay???
5853          */
5854         if (conf->min_offset_diff < 0) {
5855                 safepos += -conf->min_offset_diff;
5856                 readpos += -conf->min_offset_diff;
5857         } else
5858                 writepos += conf->min_offset_diff;
5859
5860         if ((mddev->reshape_backwards
5861              ? (safepos > writepos && readpos < writepos)
5862              : (safepos < writepos && readpos > writepos)) ||
5863             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5864                 /* Cannot proceed until we've updated the superblock... */
5865                 wait_event(conf->wait_for_overlap,
5866                            atomic_read(&conf->reshape_stripes)==0
5867                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5868                 if (atomic_read(&conf->reshape_stripes) != 0)
5869                         return 0;
5870                 mddev->reshape_position = conf->reshape_progress;
5871                 mddev->curr_resync_completed = sector_nr;
5872                 conf->reshape_checkpoint = jiffies;
5873                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5874                 md_wakeup_thread(mddev->thread);
5875                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5876                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5877                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5878                         return 0;
5879                 spin_lock_irq(&conf->device_lock);
5880                 conf->reshape_safe = mddev->reshape_position;
5881                 spin_unlock_irq(&conf->device_lock);
5882                 wake_up(&conf->wait_for_overlap);
5883                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5884         }
5885
5886         INIT_LIST_HEAD(&stripes);
5887         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5888                 int j;
5889                 int skipped_disk = 0;
5890                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5891                 set_bit(STRIPE_EXPANDING, &sh->state);
5892                 atomic_inc(&conf->reshape_stripes);
5893                 /* If any of this stripe is beyond the end of the old
5894                  * array, then we need to zero those blocks
5895                  */
5896                 for (j=sh->disks; j--;) {
5897                         sector_t s;
5898                         if (j == sh->pd_idx)
5899                                 continue;
5900                         if (conf->level == 6 &&
5901                             j == sh->qd_idx)
5902                                 continue;
5903                         s = raid5_compute_blocknr(sh, j, 0);
5904                         if (s < raid5_size(mddev, 0, 0)) {
5905                                 skipped_disk = 1;
5906                                 continue;
5907                         }
5908                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5909                         set_bit(R5_Expanded, &sh->dev[j].flags);
5910                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5911                 }
5912                 if (!skipped_disk) {
5913                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5914                         set_bit(STRIPE_HANDLE, &sh->state);
5915                 }
5916                 list_add(&sh->lru, &stripes);
5917         }
5918         spin_lock_irq(&conf->device_lock);
5919         if (mddev->reshape_backwards)
5920                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5921         else
5922                 conf->reshape_progress += reshape_sectors * new_data_disks;
5923         spin_unlock_irq(&conf->device_lock);
5924         /* Ok, those stripe are ready. We can start scheduling
5925          * reads on the source stripes.
5926          * The source stripes are determined by mapping the first and last
5927          * block on the destination stripes.
5928          */
5929         first_sector =
5930                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5931                                      1, &dd_idx, NULL);
5932         last_sector =
5933                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5934                                             * new_data_disks - 1),
5935                                      1, &dd_idx, NULL);
5936         if (last_sector >= mddev->dev_sectors)
5937                 last_sector = mddev->dev_sectors - 1;
5938         while (first_sector <= last_sector) {
5939                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5940                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5941                 set_bit(STRIPE_HANDLE, &sh->state);
5942                 raid5_release_stripe(sh);
5943                 first_sector += STRIPE_SECTORS;
5944         }
5945         /* Now that the sources are clearly marked, we can release
5946          * the destination stripes
5947          */
5948         while (!list_empty(&stripes)) {
5949                 sh = list_entry(stripes.next, struct stripe_head, lru);
5950                 list_del_init(&sh->lru);
5951                 raid5_release_stripe(sh);
5952         }
5953         /* If this takes us to the resync_max point where we have to pause,
5954          * then we need to write out the superblock.
5955          */
5956         sector_nr += reshape_sectors;
5957         retn = reshape_sectors;
5958 finish:
5959         if (mddev->curr_resync_completed > mddev->resync_max ||
5960             (sector_nr - mddev->curr_resync_completed) * 2
5961             >= mddev->resync_max - mddev->curr_resync_completed) {
5962                 /* Cannot proceed until we've updated the superblock... */
5963                 wait_event(conf->wait_for_overlap,
5964                            atomic_read(&conf->reshape_stripes) == 0
5965                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5966                 if (atomic_read(&conf->reshape_stripes) != 0)
5967                         goto ret;
5968                 mddev->reshape_position = conf->reshape_progress;
5969                 mddev->curr_resync_completed = sector_nr;
5970                 conf->reshape_checkpoint = jiffies;
5971                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5972                 md_wakeup_thread(mddev->thread);
5973                 wait_event(mddev->sb_wait,
5974                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
5975                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5976                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5977                         goto ret;
5978                 spin_lock_irq(&conf->device_lock);
5979                 conf->reshape_safe = mddev->reshape_position;
5980                 spin_unlock_irq(&conf->device_lock);
5981                 wake_up(&conf->wait_for_overlap);
5982                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5983         }
5984 ret:
5985         return retn;
5986 }
5987
5988 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
5989                                           int *skipped)
5990 {
5991         struct r5conf *conf = mddev->private;
5992         struct stripe_head *sh;
5993         sector_t max_sector = mddev->dev_sectors;
5994         sector_t sync_blocks;
5995         int still_degraded = 0;
5996         int i;
5997
5998         if (sector_nr >= max_sector) {
5999                 /* just being told to finish up .. nothing much to do */
6000
6001                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6002                         end_reshape(conf);
6003                         return 0;
6004                 }
6005
6006                 if (mddev->curr_resync < max_sector) /* aborted */
6007                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6008                                         &sync_blocks, 1);
6009                 else /* completed sync */
6010                         conf->fullsync = 0;
6011                 bitmap_close_sync(mddev->bitmap);
6012
6013                 return 0;
6014         }
6015
6016         /* Allow raid5_quiesce to complete */
6017         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6018
6019         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6020                 return reshape_request(mddev, sector_nr, skipped);
6021
6022         /* No need to check resync_max as we never do more than one
6023          * stripe, and as resync_max will always be on a chunk boundary,
6024          * if the check in md_do_sync didn't fire, there is no chance
6025          * of overstepping resync_max here
6026          */
6027
6028         /* if there is too many failed drives and we are trying
6029          * to resync, then assert that we are finished, because there is
6030          * nothing we can do.
6031          */
6032         if (mddev->degraded >= conf->max_degraded &&
6033             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6034                 sector_t rv = mddev->dev_sectors - sector_nr;
6035                 *skipped = 1;
6036                 return rv;
6037         }
6038         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6039             !conf->fullsync &&
6040             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6041             sync_blocks >= STRIPE_SECTORS) {
6042                 /* we can skip this block, and probably more */
6043                 sync_blocks /= STRIPE_SECTORS;
6044                 *skipped = 1;
6045                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6046         }
6047
6048         bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6049
6050         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6051         if (sh == NULL) {
6052                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6053                 /* make sure we don't swamp the stripe cache if someone else
6054                  * is trying to get access
6055                  */
6056                 schedule_timeout_uninterruptible(1);
6057         }
6058         /* Need to check if array will still be degraded after recovery/resync
6059          * Note in case of > 1 drive failures it's possible we're rebuilding
6060          * one drive while leaving another faulty drive in array.
6061          */
6062         rcu_read_lock();
6063         for (i = 0; i < conf->raid_disks; i++) {
6064                 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
6065
6066                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6067                         still_degraded = 1;
6068         }
6069         rcu_read_unlock();
6070
6071         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6072
6073         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6074         set_bit(STRIPE_HANDLE, &sh->state);
6075
6076         raid5_release_stripe(sh);
6077
6078         return STRIPE_SECTORS;
6079 }
6080
6081 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6082                                unsigned int offset)
6083 {
6084         /* We may not be able to submit a whole bio at once as there
6085          * may not be enough stripe_heads available.
6086          * We cannot pre-allocate enough stripe_heads as we may need
6087          * more than exist in the cache (if we allow ever large chunks).
6088          * So we do one stripe head at a time and record in
6089          * ->bi_hw_segments how many have been done.
6090          *
6091          * We *know* that this entire raid_bio is in one chunk, so
6092          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6093          */
6094         struct stripe_head *sh;
6095         int dd_idx;
6096         sector_t sector, logical_sector, last_sector;
6097         int scnt = 0;
6098         int handled = 0;
6099
6100         logical_sector = raid_bio->bi_iter.bi_sector &
6101                 ~((sector_t)STRIPE_SECTORS-1);
6102         sector = raid5_compute_sector(conf, logical_sector,
6103                                       0, &dd_idx, NULL);
6104         last_sector = bio_end_sector(raid_bio);
6105
6106         for (; logical_sector < last_sector;
6107              logical_sector += STRIPE_SECTORS,
6108                      sector += STRIPE_SECTORS,
6109                      scnt++) {
6110
6111                 if (scnt < offset)
6112                         /* already done this stripe */
6113                         continue;
6114
6115                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6116
6117                 if (!sh) {
6118                         /* failed to get a stripe - must wait */
6119                         conf->retry_read_aligned = raid_bio;
6120                         conf->retry_read_offset = scnt;
6121                         return handled;
6122                 }
6123
6124                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6125                         raid5_release_stripe(sh);
6126                         conf->retry_read_aligned = raid_bio;
6127                         conf->retry_read_offset = scnt;
6128                         return handled;
6129                 }
6130
6131                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6132                 handle_stripe(sh);
6133                 raid5_release_stripe(sh);
6134                 handled++;
6135         }
6136
6137         bio_endio(raid_bio);
6138
6139         if (atomic_dec_and_test(&conf->active_aligned_reads))
6140                 wake_up(&conf->wait_for_quiescent);
6141         return handled;
6142 }
6143
6144 static int handle_active_stripes(struct r5conf *conf, int group,
6145                                  struct r5worker *worker,
6146                                  struct list_head *temp_inactive_list)
6147 {
6148         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6149         int i, batch_size = 0, hash;
6150         bool release_inactive = false;
6151
6152         while (batch_size < MAX_STRIPE_BATCH &&
6153                         (sh = __get_priority_stripe(conf, group)) != NULL)
6154                 batch[batch_size++] = sh;
6155
6156         if (batch_size == 0) {
6157                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6158                         if (!list_empty(temp_inactive_list + i))
6159                                 break;
6160                 if (i == NR_STRIPE_HASH_LOCKS) {
6161                         spin_unlock_irq(&conf->device_lock);
6162                         r5l_flush_stripe_to_raid(conf->log);
6163                         spin_lock_irq(&conf->device_lock);
6164                         return batch_size;
6165                 }
6166                 release_inactive = true;
6167         }
6168         spin_unlock_irq(&conf->device_lock);
6169
6170         release_inactive_stripe_list(conf, temp_inactive_list,
6171                                      NR_STRIPE_HASH_LOCKS);
6172
6173         r5l_flush_stripe_to_raid(conf->log);
6174         if (release_inactive) {
6175                 spin_lock_irq(&conf->device_lock);
6176                 return 0;
6177         }
6178
6179         for (i = 0; i < batch_size; i++)
6180                 handle_stripe(batch[i]);
6181         log_write_stripe_run(conf);
6182
6183         cond_resched();
6184
6185         spin_lock_irq(&conf->device_lock);
6186         for (i = 0; i < batch_size; i++) {
6187                 hash = batch[i]->hash_lock_index;
6188                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6189         }
6190         return batch_size;
6191 }
6192
6193 static void raid5_do_work(struct work_struct *work)
6194 {
6195         struct r5worker *worker = container_of(work, struct r5worker, work);
6196         struct r5worker_group *group = worker->group;
6197         struct r5conf *conf = group->conf;
6198         struct mddev *mddev = conf->mddev;
6199         int group_id = group - conf->worker_groups;
6200         int handled;
6201         struct blk_plug plug;
6202
6203         pr_debug("+++ raid5worker active\n");
6204
6205         blk_start_plug(&plug);
6206         handled = 0;
6207         spin_lock_irq(&conf->device_lock);
6208         while (1) {
6209                 int batch_size, released;
6210
6211                 released = release_stripe_list(conf, worker->temp_inactive_list);
6212
6213                 batch_size = handle_active_stripes(conf, group_id, worker,
6214                                                    worker->temp_inactive_list);
6215                 worker->working = false;
6216                 if (!batch_size && !released)
6217                         break;
6218                 handled += batch_size;
6219                 wait_event_lock_irq(mddev->sb_wait,
6220                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6221                         conf->device_lock);
6222         }
6223         pr_debug("%d stripes handled\n", handled);
6224
6225         spin_unlock_irq(&conf->device_lock);
6226         blk_finish_plug(&plug);
6227
6228         pr_debug("--- raid5worker inactive\n");
6229 }
6230
6231 /*
6232  * This is our raid5 kernel thread.
6233  *
6234  * We scan the hash table for stripes which can be handled now.
6235  * During the scan, completed stripes are saved for us by the interrupt
6236  * handler, so that they will not have to wait for our next wakeup.
6237  */
6238 static void raid5d(struct md_thread *thread)
6239 {
6240         struct mddev *mddev = thread->mddev;
6241         struct r5conf *conf = mddev->private;
6242         int handled;
6243         struct blk_plug plug;
6244
6245         pr_debug("+++ raid5d active\n");
6246
6247         md_check_recovery(mddev);
6248
6249         blk_start_plug(&plug);
6250         handled = 0;
6251         spin_lock_irq(&conf->device_lock);
6252         while (1) {
6253                 struct bio *bio;
6254                 int batch_size, released;
6255                 unsigned int offset;
6256
6257                 released = release_stripe_list(conf, conf->temp_inactive_list);
6258                 if (released)
6259                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6260
6261                 if (
6262                     !list_empty(&conf->bitmap_list)) {
6263                         /* Now is a good time to flush some bitmap updates */
6264                         conf->seq_flush++;
6265                         spin_unlock_irq(&conf->device_lock);
6266                         bitmap_unplug(mddev->bitmap);
6267                         spin_lock_irq(&conf->device_lock);
6268                         conf->seq_write = conf->seq_flush;
6269                         activate_bit_delay(conf, conf->temp_inactive_list);
6270                 }
6271                 raid5_activate_delayed(conf);
6272
6273                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6274                         int ok;
6275                         spin_unlock_irq(&conf->device_lock);
6276                         ok = retry_aligned_read(conf, bio, offset);
6277                         spin_lock_irq(&conf->device_lock);
6278                         if (!ok)
6279                                 break;
6280                         handled++;
6281                 }
6282
6283                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6284                                                    conf->temp_inactive_list);
6285                 if (!batch_size && !released)
6286                         break;
6287                 handled += batch_size;
6288
6289                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6290                         spin_unlock_irq(&conf->device_lock);
6291                         md_check_recovery(mddev);
6292                         spin_lock_irq(&conf->device_lock);
6293                 }
6294         }
6295         pr_debug("%d stripes handled\n", handled);
6296
6297         spin_unlock_irq(&conf->device_lock);
6298         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6299             mutex_trylock(&conf->cache_size_mutex)) {
6300                 grow_one_stripe(conf, __GFP_NOWARN);
6301                 /* Set flag even if allocation failed.  This helps
6302                  * slow down allocation requests when mem is short
6303                  */
6304                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6305                 mutex_unlock(&conf->cache_size_mutex);
6306         }
6307
6308         flush_deferred_bios(conf);
6309
6310         r5l_flush_stripe_to_raid(conf->log);
6311
6312         async_tx_issue_pending_all();
6313         blk_finish_plug(&plug);
6314
6315         pr_debug("--- raid5d inactive\n");
6316 }
6317
6318 static ssize_t
6319 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6320 {
6321         struct r5conf *conf;
6322         int ret = 0;
6323         spin_lock(&mddev->lock);
6324         conf = mddev->private;
6325         if (conf)
6326                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6327         spin_unlock(&mddev->lock);
6328         return ret;
6329 }
6330
6331 int
6332 raid5_set_cache_size(struct mddev *mddev, int size)
6333 {
6334         struct r5conf *conf = mddev->private;
6335
6336         if (size <= 16 || size > 32768)
6337                 return -EINVAL;
6338
6339         conf->min_nr_stripes = size;
6340         mutex_lock(&conf->cache_size_mutex);
6341         while (size < conf->max_nr_stripes &&
6342                drop_one_stripe(conf))
6343                 ;
6344         mutex_unlock(&conf->cache_size_mutex);
6345
6346         md_allow_write(mddev);
6347
6348         mutex_lock(&conf->cache_size_mutex);
6349         while (size > conf->max_nr_stripes)
6350                 if (!grow_one_stripe(conf, GFP_KERNEL))
6351                         break;
6352         mutex_unlock(&conf->cache_size_mutex);
6353
6354         return 0;
6355 }
6356 EXPORT_SYMBOL(raid5_set_cache_size);
6357
6358 static ssize_t
6359 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6360 {
6361         struct r5conf *conf;
6362         unsigned long new;
6363         int err;
6364
6365         if (len >= PAGE_SIZE)
6366                 return -EINVAL;
6367         if (kstrtoul(page, 10, &new))
6368                 return -EINVAL;
6369         err = mddev_lock(mddev);
6370         if (err)
6371                 return err;
6372         conf = mddev->private;
6373         if (!conf)
6374                 err = -ENODEV;
6375         else
6376                 err = raid5_set_cache_size(mddev, new);
6377         mddev_unlock(mddev);
6378
6379         return err ?: len;
6380 }
6381
6382 static struct md_sysfs_entry
6383 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6384                                 raid5_show_stripe_cache_size,
6385                                 raid5_store_stripe_cache_size);
6386
6387 static ssize_t
6388 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6389 {
6390         struct r5conf *conf = mddev->private;
6391         if (conf)
6392                 return sprintf(page, "%d\n", conf->rmw_level);
6393         else
6394                 return 0;
6395 }
6396
6397 static ssize_t
6398 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6399 {
6400         struct r5conf *conf = mddev->private;
6401         unsigned long new;
6402
6403         if (!conf)
6404                 return -ENODEV;
6405
6406         if (len >= PAGE_SIZE)
6407                 return -EINVAL;
6408
6409         if (kstrtoul(page, 10, &new))
6410                 return -EINVAL;
6411
6412         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6413                 return -EINVAL;
6414
6415         if (new != PARITY_DISABLE_RMW &&
6416             new != PARITY_ENABLE_RMW &&
6417             new != PARITY_PREFER_RMW)
6418                 return -EINVAL;
6419
6420         conf->rmw_level = new;
6421         return len;
6422 }
6423
6424 static struct md_sysfs_entry
6425 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6426                          raid5_show_rmw_level,
6427                          raid5_store_rmw_level);
6428
6429
6430 static ssize_t
6431 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6432 {
6433         struct r5conf *conf;
6434         int ret = 0;
6435         spin_lock(&mddev->lock);
6436         conf = mddev->private;
6437         if (conf)
6438                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6439         spin_unlock(&mddev->lock);
6440         return ret;
6441 }
6442
6443 static ssize_t
6444 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6445 {
6446         struct r5conf *conf;
6447         unsigned long new;
6448         int err;
6449
6450         if (len >= PAGE_SIZE)
6451                 return -EINVAL;
6452         if (kstrtoul(page, 10, &new))
6453                 return -EINVAL;
6454
6455         err = mddev_lock(mddev);
6456         if (err)
6457                 return err;
6458         conf = mddev->private;
6459         if (!conf)
6460                 err = -ENODEV;
6461         else if (new > conf->min_nr_stripes)
6462                 err = -EINVAL;
6463         else
6464                 conf->bypass_threshold = new;
6465         mddev_unlock(mddev);
6466         return err ?: len;
6467 }
6468
6469 static struct md_sysfs_entry
6470 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6471                                         S_IRUGO | S_IWUSR,
6472                                         raid5_show_preread_threshold,
6473                                         raid5_store_preread_threshold);
6474
6475 static ssize_t
6476 raid5_show_skip_copy(struct mddev *mddev, char *page)
6477 {
6478         struct r5conf *conf;
6479         int ret = 0;
6480         spin_lock(&mddev->lock);
6481         conf = mddev->private;
6482         if (conf)
6483                 ret = sprintf(page, "%d\n", conf->skip_copy);
6484         spin_unlock(&mddev->lock);
6485         return ret;
6486 }
6487
6488 static ssize_t
6489 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6490 {
6491         struct r5conf *conf;
6492         unsigned long new;
6493         int err;
6494
6495         if (len >= PAGE_SIZE)
6496                 return -EINVAL;
6497         if (kstrtoul(page, 10, &new))
6498                 return -EINVAL;
6499         new = !!new;
6500
6501         err = mddev_lock(mddev);
6502         if (err)
6503                 return err;
6504         conf = mddev->private;
6505         if (!conf)
6506                 err = -ENODEV;
6507         else if (new != conf->skip_copy) {
6508                 mddev_suspend(mddev);
6509                 conf->skip_copy = new;
6510                 if (new)
6511                         mddev->queue->backing_dev_info->capabilities |=
6512                                 BDI_CAP_STABLE_WRITES;
6513                 else
6514                         mddev->queue->backing_dev_info->capabilities &=
6515                                 ~BDI_CAP_STABLE_WRITES;
6516                 mddev_resume(mddev);
6517         }
6518         mddev_unlock(mddev);
6519         return err ?: len;
6520 }
6521
6522 static struct md_sysfs_entry
6523 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6524                                         raid5_show_skip_copy,
6525                                         raid5_store_skip_copy);
6526
6527 static ssize_t
6528 stripe_cache_active_show(struct mddev *mddev, char *page)
6529 {
6530         struct r5conf *conf = mddev->private;
6531         if (conf)
6532                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6533         else
6534                 return 0;
6535 }
6536
6537 static struct md_sysfs_entry
6538 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6539
6540 static ssize_t
6541 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6542 {
6543         struct r5conf *conf;
6544         int ret = 0;
6545         spin_lock(&mddev->lock);
6546         conf = mddev->private;
6547         if (conf)
6548                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6549         spin_unlock(&mddev->lock);
6550         return ret;
6551 }
6552
6553 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6554                                int *group_cnt,
6555                                int *worker_cnt_per_group,
6556                                struct r5worker_group **worker_groups);
6557 static ssize_t
6558 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6559 {
6560         struct r5conf *conf;
6561         unsigned long new;
6562         int err;
6563         struct r5worker_group *new_groups, *old_groups;
6564         int group_cnt, worker_cnt_per_group;
6565
6566         if (len >= PAGE_SIZE)
6567                 return -EINVAL;
6568         if (kstrtoul(page, 10, &new))
6569                 return -EINVAL;
6570
6571         err = mddev_lock(mddev);
6572         if (err)
6573                 return err;
6574         conf = mddev->private;
6575         if (!conf)
6576                 err = -ENODEV;
6577         else if (new != conf->worker_cnt_per_group) {
6578                 mddev_suspend(mddev);
6579
6580                 old_groups = conf->worker_groups;
6581                 if (old_groups)
6582                         flush_workqueue(raid5_wq);
6583
6584                 err = alloc_thread_groups(conf, new,
6585                                           &group_cnt, &worker_cnt_per_group,
6586                                           &new_groups);
6587                 if (!err) {
6588                         spin_lock_irq(&conf->device_lock);
6589                         conf->group_cnt = group_cnt;
6590                         conf->worker_cnt_per_group = worker_cnt_per_group;
6591                         conf->worker_groups = new_groups;
6592                         spin_unlock_irq(&conf->device_lock);
6593
6594                         if (old_groups)
6595                                 kfree(old_groups[0].workers);
6596                         kfree(old_groups);
6597                 }
6598                 mddev_resume(mddev);
6599         }
6600         mddev_unlock(mddev);
6601
6602         return err ?: len;
6603 }
6604
6605 static struct md_sysfs_entry
6606 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6607                                 raid5_show_group_thread_cnt,
6608                                 raid5_store_group_thread_cnt);
6609
6610 static struct attribute *raid5_attrs[] =  {
6611         &raid5_stripecache_size.attr,
6612         &raid5_stripecache_active.attr,
6613         &raid5_preread_bypass_threshold.attr,
6614         &raid5_group_thread_cnt.attr,
6615         &raid5_skip_copy.attr,
6616         &raid5_rmw_level.attr,
6617         &r5c_journal_mode.attr,
6618         NULL,
6619 };
6620 static struct attribute_group raid5_attrs_group = {
6621         .name = NULL,
6622         .attrs = raid5_attrs,
6623 };
6624
6625 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6626                                int *group_cnt,
6627                                int *worker_cnt_per_group,
6628                                struct r5worker_group **worker_groups)
6629 {
6630         int i, j, k;
6631         ssize_t size;
6632         struct r5worker *workers;
6633
6634         *worker_cnt_per_group = cnt;
6635         if (cnt == 0) {
6636                 *group_cnt = 0;
6637                 *worker_groups = NULL;
6638                 return 0;
6639         }
6640         *group_cnt = num_possible_nodes();
6641         size = sizeof(struct r5worker) * cnt;
6642         workers = kzalloc(size * *group_cnt, GFP_NOIO);
6643         *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6644                                 *group_cnt, GFP_NOIO);
6645         if (!*worker_groups || !workers) {
6646                 kfree(workers);
6647                 kfree(*worker_groups);
6648                 return -ENOMEM;
6649         }
6650
6651         for (i = 0; i < *group_cnt; i++) {
6652                 struct r5worker_group *group;
6653
6654                 group = &(*worker_groups)[i];
6655                 INIT_LIST_HEAD(&group->handle_list);
6656                 INIT_LIST_HEAD(&group->loprio_list);
6657                 group->conf = conf;
6658                 group->workers = workers + i * cnt;
6659
6660                 for (j = 0; j < cnt; j++) {
6661                         struct r5worker *worker = group->workers + j;
6662                         worker->group = group;
6663                         INIT_WORK(&worker->work, raid5_do_work);
6664
6665                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6666                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6667                 }
6668         }
6669
6670         return 0;
6671 }
6672
6673 static void free_thread_groups(struct r5conf *conf)
6674 {
6675         if (conf->worker_groups)
6676                 kfree(conf->worker_groups[0].workers);
6677         kfree(conf->worker_groups);
6678         conf->worker_groups = NULL;
6679 }
6680
6681 static sector_t
6682 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6683 {
6684         struct r5conf *conf = mddev->private;
6685
6686         if (!sectors)
6687                 sectors = mddev->dev_sectors;
6688         if (!raid_disks)
6689                 /* size is defined by the smallest of previous and new size */
6690                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6691
6692         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6693         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6694         return sectors * (raid_disks - conf->max_degraded);
6695 }
6696
6697 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6698 {
6699         safe_put_page(percpu->spare_page);
6700         if (percpu->scribble)
6701                 flex_array_free(percpu->scribble);
6702         percpu->spare_page = NULL;
6703         percpu->scribble = NULL;
6704 }
6705
6706 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6707 {
6708         if (conf->level == 6 && !percpu->spare_page)
6709                 percpu->spare_page = alloc_page(GFP_KERNEL);
6710         if (!percpu->scribble)
6711                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6712                                                       conf->previous_raid_disks),
6713                                                   max(conf->chunk_sectors,
6714                                                       conf->prev_chunk_sectors)
6715                                                    / STRIPE_SECTORS,
6716                                                   GFP_KERNEL);
6717
6718         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6719                 free_scratch_buffer(conf, percpu);
6720                 return -ENOMEM;
6721         }
6722
6723         return 0;
6724 }
6725
6726 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6727 {
6728         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6729
6730         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6731         return 0;
6732 }
6733
6734 static void raid5_free_percpu(struct r5conf *conf)
6735 {
6736         if (!conf->percpu)
6737                 return;
6738
6739         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6740         free_percpu(conf->percpu);
6741 }
6742
6743 static void free_conf(struct r5conf *conf)
6744 {
6745         int i;
6746
6747         log_exit(conf);
6748
6749         if (conf->shrinker.nr_deferred)
6750                 unregister_shrinker(&conf->shrinker);
6751
6752         free_thread_groups(conf);
6753         shrink_stripes(conf);
6754         raid5_free_percpu(conf);
6755         for (i = 0; i < conf->pool_size; i++)
6756                 if (conf->disks[i].extra_page)
6757                         put_page(conf->disks[i].extra_page);
6758         kfree(conf->disks);
6759         if (conf->bio_split)
6760                 bioset_free(conf->bio_split);
6761         kfree(conf->stripe_hashtbl);
6762         kfree(conf->pending_data);
6763         kfree(conf);
6764 }
6765
6766 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6767 {
6768         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6769         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6770
6771         if (alloc_scratch_buffer(conf, percpu)) {
6772                 pr_warn("%s: failed memory allocation for cpu%u\n",
6773                         __func__, cpu);
6774                 return -ENOMEM;
6775         }
6776         return 0;
6777 }
6778
6779 static int raid5_alloc_percpu(struct r5conf *conf)
6780 {
6781         int err = 0;
6782
6783         conf->percpu = alloc_percpu(struct raid5_percpu);
6784         if (!conf->percpu)
6785                 return -ENOMEM;
6786
6787         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6788         if (!err) {
6789                 conf->scribble_disks = max(conf->raid_disks,
6790                         conf->previous_raid_disks);
6791                 conf->scribble_sectors = max(conf->chunk_sectors,
6792                         conf->prev_chunk_sectors);
6793         }
6794         return err;
6795 }
6796
6797 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6798                                       struct shrink_control *sc)
6799 {
6800         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6801         unsigned long ret = SHRINK_STOP;
6802
6803         if (mutex_trylock(&conf->cache_size_mutex)) {
6804                 ret= 0;
6805                 while (ret < sc->nr_to_scan &&
6806                        conf->max_nr_stripes > conf->min_nr_stripes) {
6807                         if (drop_one_stripe(conf) == 0) {
6808                                 ret = SHRINK_STOP;
6809                                 break;
6810                         }
6811                         ret++;
6812                 }
6813                 mutex_unlock(&conf->cache_size_mutex);
6814         }
6815         return ret;
6816 }
6817
6818 static unsigned long raid5_cache_count(struct shrinker *shrink,
6819                                        struct shrink_control *sc)
6820 {
6821         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6822
6823         if (conf->max_nr_stripes < conf->min_nr_stripes)
6824                 /* unlikely, but not impossible */
6825                 return 0;
6826         return conf->max_nr_stripes - conf->min_nr_stripes;
6827 }
6828
6829 static struct r5conf *setup_conf(struct mddev *mddev)
6830 {
6831         struct r5conf *conf;
6832         int raid_disk, memory, max_disks;
6833         struct md_rdev *rdev;
6834         struct disk_info *disk;
6835         char pers_name[6];
6836         int i;
6837         int group_cnt, worker_cnt_per_group;
6838         struct r5worker_group *new_group;
6839
6840         if (mddev->new_level != 5
6841             && mddev->new_level != 4
6842             && mddev->new_level != 6) {
6843                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6844                         mdname(mddev), mddev->new_level);
6845                 return ERR_PTR(-EIO);
6846         }
6847         if ((mddev->new_level == 5
6848              && !algorithm_valid_raid5(mddev->new_layout)) ||
6849             (mddev->new_level == 6
6850              && !algorithm_valid_raid6(mddev->new_layout))) {
6851                 pr_warn("md/raid:%s: layout %d not supported\n",
6852                         mdname(mddev), mddev->new_layout);
6853                 return ERR_PTR(-EIO);
6854         }
6855         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6856                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6857                         mdname(mddev), mddev->raid_disks);
6858                 return ERR_PTR(-EINVAL);
6859         }
6860
6861         if (!mddev->new_chunk_sectors ||
6862             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6863             !is_power_of_2(mddev->new_chunk_sectors)) {
6864                 pr_warn("md/raid:%s: invalid chunk size %d\n",
6865                         mdname(mddev), mddev->new_chunk_sectors << 9);
6866                 return ERR_PTR(-EINVAL);
6867         }
6868
6869         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6870         if (conf == NULL)
6871                 goto abort;
6872         INIT_LIST_HEAD(&conf->free_list);
6873         INIT_LIST_HEAD(&conf->pending_list);
6874         conf->pending_data = kzalloc(sizeof(struct r5pending_data) *
6875                 PENDING_IO_MAX, GFP_KERNEL);
6876         if (!conf->pending_data)
6877                 goto abort;
6878         for (i = 0; i < PENDING_IO_MAX; i++)
6879                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6880         /* Don't enable multi-threading by default*/
6881         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6882                                  &new_group)) {
6883                 conf->group_cnt = group_cnt;
6884                 conf->worker_cnt_per_group = worker_cnt_per_group;
6885                 conf->worker_groups = new_group;
6886         } else
6887                 goto abort;
6888         spin_lock_init(&conf->device_lock);
6889         seqcount_init(&conf->gen_lock);
6890         mutex_init(&conf->cache_size_mutex);
6891         init_waitqueue_head(&conf->wait_for_quiescent);
6892         init_waitqueue_head(&conf->wait_for_stripe);
6893         init_waitqueue_head(&conf->wait_for_overlap);
6894         INIT_LIST_HEAD(&conf->handle_list);
6895         INIT_LIST_HEAD(&conf->loprio_list);
6896         INIT_LIST_HEAD(&conf->hold_list);
6897         INIT_LIST_HEAD(&conf->delayed_list);
6898         INIT_LIST_HEAD(&conf->bitmap_list);
6899         init_llist_head(&conf->released_stripes);
6900         atomic_set(&conf->active_stripes, 0);
6901         atomic_set(&conf->preread_active_stripes, 0);
6902         atomic_set(&conf->active_aligned_reads, 0);
6903         spin_lock_init(&conf->pending_bios_lock);
6904         conf->batch_bio_dispatch = true;
6905         rdev_for_each(rdev, mddev) {
6906                 if (test_bit(Journal, &rdev->flags))
6907                         continue;
6908                 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6909                         conf->batch_bio_dispatch = false;
6910                         break;
6911                 }
6912         }
6913
6914         conf->bypass_threshold = BYPASS_THRESHOLD;
6915         conf->recovery_disabled = mddev->recovery_disabled - 1;
6916
6917         conf->raid_disks = mddev->raid_disks;
6918         if (mddev->reshape_position == MaxSector)
6919                 conf->previous_raid_disks = mddev->raid_disks;
6920         else
6921                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6922         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6923
6924         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6925                               GFP_KERNEL);
6926
6927         if (!conf->disks)
6928                 goto abort;
6929
6930         for (i = 0; i < max_disks; i++) {
6931                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6932                 if (!conf->disks[i].extra_page)
6933                         goto abort;
6934         }
6935
6936         conf->bio_split = bioset_create(BIO_POOL_SIZE, 0);
6937         if (!conf->bio_split)
6938                 goto abort;
6939         conf->mddev = mddev;
6940
6941         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6942                 goto abort;
6943
6944         /* We init hash_locks[0] separately to that it can be used
6945          * as the reference lock in the spin_lock_nest_lock() call
6946          * in lock_all_device_hash_locks_irq in order to convince
6947          * lockdep that we know what we are doing.
6948          */
6949         spin_lock_init(conf->hash_locks);
6950         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6951                 spin_lock_init(conf->hash_locks + i);
6952
6953         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6954                 INIT_LIST_HEAD(conf->inactive_list + i);
6955
6956         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6957                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6958
6959         atomic_set(&conf->r5c_cached_full_stripes, 0);
6960         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
6961         atomic_set(&conf->r5c_cached_partial_stripes, 0);
6962         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
6963         atomic_set(&conf->r5c_flushing_full_stripes, 0);
6964         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
6965
6966         conf->level = mddev->new_level;
6967         conf->chunk_sectors = mddev->new_chunk_sectors;
6968         if (raid5_alloc_percpu(conf) != 0)
6969                 goto abort;
6970
6971         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6972
6973         rdev_for_each(rdev, mddev) {
6974                 raid_disk = rdev->raid_disk;
6975                 if (raid_disk >= max_disks
6976                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6977                         continue;
6978                 disk = conf->disks + raid_disk;
6979
6980                 if (test_bit(Replacement, &rdev->flags)) {
6981                         if (disk->replacement)
6982                                 goto abort;
6983                         disk->replacement = rdev;
6984                 } else {
6985                         if (disk->rdev)
6986                                 goto abort;
6987                         disk->rdev = rdev;
6988                 }
6989
6990                 if (test_bit(In_sync, &rdev->flags)) {
6991                         char b[BDEVNAME_SIZE];
6992                         pr_info("md/raid:%s: device %s operational as raid disk %d\n",
6993                                 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6994                 } else if (rdev->saved_raid_disk != raid_disk)
6995                         /* Cannot rely on bitmap to complete recovery */
6996                         conf->fullsync = 1;
6997         }
6998
6999         conf->level = mddev->new_level;
7000         if (conf->level == 6) {
7001                 conf->max_degraded = 2;
7002                 if (raid6_call.xor_syndrome)
7003                         conf->rmw_level = PARITY_ENABLE_RMW;
7004                 else
7005                         conf->rmw_level = PARITY_DISABLE_RMW;
7006         } else {
7007                 conf->max_degraded = 1;
7008                 conf->rmw_level = PARITY_ENABLE_RMW;
7009         }
7010         conf->algorithm = mddev->new_layout;
7011         conf->reshape_progress = mddev->reshape_position;
7012         if (conf->reshape_progress != MaxSector) {
7013                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7014                 conf->prev_algo = mddev->layout;
7015         } else {
7016                 conf->prev_chunk_sectors = conf->chunk_sectors;
7017                 conf->prev_algo = conf->algorithm;
7018         }
7019
7020         conf->min_nr_stripes = NR_STRIPES;
7021         if (mddev->reshape_position != MaxSector) {
7022                 int stripes = max_t(int,
7023                         ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7024                         ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7025                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7026                 if (conf->min_nr_stripes != NR_STRIPES)
7027                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7028                                 mdname(mddev), conf->min_nr_stripes);
7029         }
7030         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7031                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7032         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7033         if (grow_stripes(conf, conf->min_nr_stripes)) {
7034                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7035                         mdname(mddev), memory);
7036                 goto abort;
7037         } else
7038                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7039         /*
7040          * Losing a stripe head costs more than the time to refill it,
7041          * it reduces the queue depth and so can hurt throughput.
7042          * So set it rather large, scaled by number of devices.
7043          */
7044         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7045         conf->shrinker.scan_objects = raid5_cache_scan;
7046         conf->shrinker.count_objects = raid5_cache_count;
7047         conf->shrinker.batch = 128;
7048         conf->shrinker.flags = 0;
7049         if (register_shrinker(&conf->shrinker)) {
7050                 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7051                         mdname(mddev));
7052                 goto abort;
7053         }
7054
7055         sprintf(pers_name, "raid%d", mddev->new_level);
7056         conf->thread = md_register_thread(raid5d, mddev, pers_name);
7057         if (!conf->thread) {
7058                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7059                         mdname(mddev));
7060                 goto abort;
7061         }
7062
7063         return conf;
7064
7065  abort:
7066         if (conf) {
7067                 free_conf(conf);
7068                 return ERR_PTR(-EIO);
7069         } else
7070                 return ERR_PTR(-ENOMEM);
7071 }
7072
7073 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7074 {
7075         switch (algo) {
7076         case ALGORITHM_PARITY_0:
7077                 if (raid_disk < max_degraded)
7078                         return 1;
7079                 break;
7080         case ALGORITHM_PARITY_N:
7081                 if (raid_disk >= raid_disks - max_degraded)
7082                         return 1;
7083                 break;
7084         case ALGORITHM_PARITY_0_6:
7085                 if (raid_disk == 0 ||
7086                     raid_disk == raid_disks - 1)
7087                         return 1;
7088                 break;
7089         case ALGORITHM_LEFT_ASYMMETRIC_6:
7090         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7091         case ALGORITHM_LEFT_SYMMETRIC_6:
7092         case ALGORITHM_RIGHT_SYMMETRIC_6:
7093                 if (raid_disk == raid_disks - 1)
7094                         return 1;
7095         }
7096         return 0;
7097 }
7098
7099 static int raid5_run(struct mddev *mddev)
7100 {
7101         struct r5conf *conf;
7102         int working_disks = 0;
7103         int dirty_parity_disks = 0;
7104         struct md_rdev *rdev;
7105         struct md_rdev *journal_dev = NULL;
7106         sector_t reshape_offset = 0;
7107         int i;
7108         long long min_offset_diff = 0;
7109         int first = 1;
7110
7111         if (mddev->recovery_cp != MaxSector)
7112                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7113                           mdname(mddev));
7114
7115         rdev_for_each(rdev, mddev) {
7116                 long long diff;
7117
7118                 if (test_bit(Journal, &rdev->flags)) {
7119                         journal_dev = rdev;
7120                         continue;
7121                 }
7122                 if (rdev->raid_disk < 0)
7123                         continue;
7124                 diff = (rdev->new_data_offset - rdev->data_offset);
7125                 if (first) {
7126                         min_offset_diff = diff;
7127                         first = 0;
7128                 } else if (mddev->reshape_backwards &&
7129                          diff < min_offset_diff)
7130                         min_offset_diff = diff;
7131                 else if (!mddev->reshape_backwards &&
7132                          diff > min_offset_diff)
7133                         min_offset_diff = diff;
7134         }
7135
7136         if (mddev->reshape_position != MaxSector) {
7137                 /* Check that we can continue the reshape.
7138                  * Difficulties arise if the stripe we would write to
7139                  * next is at or after the stripe we would read from next.
7140                  * For a reshape that changes the number of devices, this
7141                  * is only possible for a very short time, and mdadm makes
7142                  * sure that time appears to have past before assembling
7143                  * the array.  So we fail if that time hasn't passed.
7144                  * For a reshape that keeps the number of devices the same
7145                  * mdadm must be monitoring the reshape can keeping the
7146                  * critical areas read-only and backed up.  It will start
7147                  * the array in read-only mode, so we check for that.
7148                  */
7149                 sector_t here_new, here_old;
7150                 int old_disks;
7151                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7152                 int chunk_sectors;
7153                 int new_data_disks;
7154
7155                 if (journal_dev) {
7156                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7157                                 mdname(mddev));
7158                         return -EINVAL;
7159                 }
7160
7161                 if (mddev->new_level != mddev->level) {
7162                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7163                                 mdname(mddev));
7164                         return -EINVAL;
7165                 }
7166                 old_disks = mddev->raid_disks - mddev->delta_disks;
7167                 /* reshape_position must be on a new-stripe boundary, and one
7168                  * further up in new geometry must map after here in old
7169                  * geometry.
7170                  * If the chunk sizes are different, then as we perform reshape
7171                  * in units of the largest of the two, reshape_position needs
7172                  * be a multiple of the largest chunk size times new data disks.
7173                  */
7174                 here_new = mddev->reshape_position;
7175                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7176                 new_data_disks = mddev->raid_disks - max_degraded;
7177                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7178                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7179                                 mdname(mddev));
7180                         return -EINVAL;
7181                 }
7182                 reshape_offset = here_new * chunk_sectors;
7183                 /* here_new is the stripe we will write to */
7184                 here_old = mddev->reshape_position;
7185                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7186                 /* here_old is the first stripe that we might need to read
7187                  * from */
7188                 if (mddev->delta_disks == 0) {
7189                         /* We cannot be sure it is safe to start an in-place
7190                          * reshape.  It is only safe if user-space is monitoring
7191                          * and taking constant backups.
7192                          * mdadm always starts a situation like this in
7193                          * readonly mode so it can take control before
7194                          * allowing any writes.  So just check for that.
7195                          */
7196                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7197                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7198                                 /* not really in-place - so OK */;
7199                         else if (mddev->ro == 0) {
7200                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7201                                         mdname(mddev));
7202                                 return -EINVAL;
7203                         }
7204                 } else if (mddev->reshape_backwards
7205                     ? (here_new * chunk_sectors + min_offset_diff <=
7206                        here_old * chunk_sectors)
7207                     : (here_new * chunk_sectors >=
7208                        here_old * chunk_sectors + (-min_offset_diff))) {
7209                         /* Reading from the same stripe as writing to - bad */
7210                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7211                                 mdname(mddev));
7212                         return -EINVAL;
7213                 }
7214                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7215                 /* OK, we should be able to continue; */
7216         } else {
7217                 BUG_ON(mddev->level != mddev->new_level);
7218                 BUG_ON(mddev->layout != mddev->new_layout);
7219                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7220                 BUG_ON(mddev->delta_disks != 0);
7221         }
7222
7223         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7224             test_bit(MD_HAS_PPL, &mddev->flags)) {
7225                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7226                         mdname(mddev));
7227                 clear_bit(MD_HAS_PPL, &mddev->flags);
7228         }
7229
7230         if (mddev->private == NULL)
7231                 conf = setup_conf(mddev);
7232         else
7233                 conf = mddev->private;
7234
7235         if (IS_ERR(conf))
7236                 return PTR_ERR(conf);
7237
7238         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7239                 if (!journal_dev) {
7240                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7241                                 mdname(mddev));
7242                         mddev->ro = 1;
7243                         set_disk_ro(mddev->gendisk, 1);
7244                 } else if (mddev->recovery_cp == MaxSector)
7245                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7246         }
7247
7248         conf->min_offset_diff = min_offset_diff;
7249         mddev->thread = conf->thread;
7250         conf->thread = NULL;
7251         mddev->private = conf;
7252
7253         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7254              i++) {
7255                 rdev = conf->disks[i].rdev;
7256                 if (!rdev && conf->disks[i].replacement) {
7257                         /* The replacement is all we have yet */
7258                         rdev = conf->disks[i].replacement;
7259                         conf->disks[i].replacement = NULL;
7260                         clear_bit(Replacement, &rdev->flags);
7261                         conf->disks[i].rdev = rdev;
7262                 }
7263                 if (!rdev)
7264                         continue;
7265                 if (conf->disks[i].replacement &&
7266                     conf->reshape_progress != MaxSector) {
7267                         /* replacements and reshape simply do not mix. */
7268                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7269                         goto abort;
7270                 }
7271                 if (test_bit(In_sync, &rdev->flags)) {
7272                         working_disks++;
7273                         continue;
7274                 }
7275                 /* This disc is not fully in-sync.  However if it
7276                  * just stored parity (beyond the recovery_offset),
7277                  * when we don't need to be concerned about the
7278                  * array being dirty.
7279                  * When reshape goes 'backwards', we never have
7280                  * partially completed devices, so we only need
7281                  * to worry about reshape going forwards.
7282                  */
7283                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7284                 if (mddev->major_version == 0 &&
7285                     mddev->minor_version > 90)
7286                         rdev->recovery_offset = reshape_offset;
7287
7288                 if (rdev->recovery_offset < reshape_offset) {
7289                         /* We need to check old and new layout */
7290                         if (!only_parity(rdev->raid_disk,
7291                                          conf->algorithm,
7292                                          conf->raid_disks,
7293                                          conf->max_degraded))
7294                                 continue;
7295                 }
7296                 if (!only_parity(rdev->raid_disk,
7297                                  conf->prev_algo,
7298                                  conf->previous_raid_disks,
7299                                  conf->max_degraded))
7300                         continue;
7301                 dirty_parity_disks++;
7302         }
7303
7304         /*
7305          * 0 for a fully functional array, 1 or 2 for a degraded array.
7306          */
7307         mddev->degraded = raid5_calc_degraded(conf);
7308
7309         if (has_failed(conf)) {
7310                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7311                         mdname(mddev), mddev->degraded, conf->raid_disks);
7312                 goto abort;
7313         }
7314
7315         /* device size must be a multiple of chunk size */
7316         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7317         mddev->resync_max_sectors = mddev->dev_sectors;
7318
7319         if (mddev->degraded > dirty_parity_disks &&
7320             mddev->recovery_cp != MaxSector) {
7321                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7322                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7323                                 mdname(mddev));
7324                 else if (mddev->ok_start_degraded)
7325                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7326                                 mdname(mddev));
7327                 else {
7328                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7329                                 mdname(mddev));
7330                         goto abort;
7331                 }
7332         }
7333
7334         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7335                 mdname(mddev), conf->level,
7336                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7337                 mddev->new_layout);
7338
7339         print_raid5_conf(conf);
7340
7341         if (conf->reshape_progress != MaxSector) {
7342                 conf->reshape_safe = conf->reshape_progress;
7343                 atomic_set(&conf->reshape_stripes, 0);
7344                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7345                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7346                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7347                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7348                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7349                                                         "reshape");
7350         }
7351
7352         /* Ok, everything is just fine now */
7353         if (mddev->to_remove == &raid5_attrs_group)
7354                 mddev->to_remove = NULL;
7355         else if (mddev->kobj.sd &&
7356             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7357                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7358                         mdname(mddev));
7359         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7360
7361         if (mddev->queue) {
7362                 int chunk_size;
7363                 /* read-ahead size must cover two whole stripes, which
7364                  * is 2 * (datadisks) * chunksize where 'n' is the
7365                  * number of raid devices
7366                  */
7367                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7368                 int stripe = data_disks *
7369                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7370                 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7371                         mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7372
7373                 chunk_size = mddev->chunk_sectors << 9;
7374                 blk_queue_io_min(mddev->queue, chunk_size);
7375                 blk_queue_io_opt(mddev->queue, chunk_size *
7376                                  (conf->raid_disks - conf->max_degraded));
7377                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7378                 /*
7379                  * We can only discard a whole stripe. It doesn't make sense to
7380                  * discard data disk but write parity disk
7381                  */
7382                 stripe = stripe * PAGE_SIZE;
7383                 /* Round up to power of 2, as discard handling
7384                  * currently assumes that */
7385                 while ((stripe-1) & stripe)
7386                         stripe = (stripe | (stripe-1)) + 1;
7387                 mddev->queue->limits.discard_alignment = stripe;
7388                 mddev->queue->limits.discard_granularity = stripe;
7389
7390                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7391                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7392
7393                 rdev_for_each(rdev, mddev) {
7394                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7395                                           rdev->data_offset << 9);
7396                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7397                                           rdev->new_data_offset << 9);
7398                 }
7399
7400                 /*
7401                  * zeroing is required, otherwise data
7402                  * could be lost. Consider a scenario: discard a stripe
7403                  * (the stripe could be inconsistent if
7404                  * discard_zeroes_data is 0); write one disk of the
7405                  * stripe (the stripe could be inconsistent again
7406                  * depending on which disks are used to calculate
7407                  * parity); the disk is broken; The stripe data of this
7408                  * disk is lost.
7409                  *
7410                  * We only allow DISCARD if the sysadmin has confirmed that
7411                  * only safe devices are in use by setting a module parameter.
7412                  * A better idea might be to turn DISCARD into WRITE_ZEROES
7413                  * requests, as that is required to be safe.
7414                  */
7415                 if (devices_handle_discard_safely &&
7416                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7417                     mddev->queue->limits.discard_granularity >= stripe)
7418                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7419                                                 mddev->queue);
7420                 else
7421                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7422                                                 mddev->queue);
7423
7424                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7425         }
7426
7427         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7428                 goto abort;
7429
7430         return 0;
7431 abort:
7432         md_unregister_thread(&mddev->thread);
7433         print_raid5_conf(conf);
7434         free_conf(conf);
7435         mddev->private = NULL;
7436         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7437         return -EIO;
7438 }
7439
7440 static void raid5_free(struct mddev *mddev, void *priv)
7441 {
7442         struct r5conf *conf = priv;
7443
7444         free_conf(conf);
7445         mddev->to_remove = &raid5_attrs_group;
7446 }
7447
7448 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7449 {
7450         struct r5conf *conf = mddev->private;
7451         int i;
7452
7453         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7454                 conf->chunk_sectors / 2, mddev->layout);
7455         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7456         rcu_read_lock();
7457         for (i = 0; i < conf->raid_disks; i++) {
7458                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7459                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7460         }
7461         rcu_read_unlock();
7462         seq_printf (seq, "]");
7463 }
7464
7465 static void print_raid5_conf (struct r5conf *conf)
7466 {
7467         int i;
7468         struct disk_info *tmp;
7469
7470         pr_debug("RAID conf printout:\n");
7471         if (!conf) {
7472                 pr_debug("(conf==NULL)\n");
7473                 return;
7474         }
7475         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7476                conf->raid_disks,
7477                conf->raid_disks - conf->mddev->degraded);
7478
7479         for (i = 0; i < conf->raid_disks; i++) {
7480                 char b[BDEVNAME_SIZE];
7481                 tmp = conf->disks + i;
7482                 if (tmp->rdev)
7483                         pr_debug(" disk %d, o:%d, dev:%s\n",
7484                                i, !test_bit(Faulty, &tmp->rdev->flags),
7485                                bdevname(tmp->rdev->bdev, b));
7486         }
7487 }
7488
7489 static int raid5_spare_active(struct mddev *mddev)
7490 {
7491         int i;
7492         struct r5conf *conf = mddev->private;
7493         struct disk_info *tmp;
7494         int count = 0;
7495         unsigned long flags;
7496
7497         for (i = 0; i < conf->raid_disks; i++) {
7498                 tmp = conf->disks + i;
7499                 if (tmp->replacement
7500                     && tmp->replacement->recovery_offset == MaxSector
7501                     && !test_bit(Faulty, &tmp->replacement->flags)
7502                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7503                         /* Replacement has just become active. */
7504                         if (!tmp->rdev
7505                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7506                                 count++;
7507                         if (tmp->rdev) {
7508                                 /* Replaced device not technically faulty,
7509                                  * but we need to be sure it gets removed
7510                                  * and never re-added.
7511                                  */
7512                                 set_bit(Faulty, &tmp->rdev->flags);
7513                                 sysfs_notify_dirent_safe(
7514                                         tmp->rdev->sysfs_state);
7515                         }
7516                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7517                 } else if (tmp->rdev
7518                     && tmp->rdev->recovery_offset == MaxSector
7519                     && !test_bit(Faulty, &tmp->rdev->flags)
7520                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7521                         count++;
7522                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7523                 }
7524         }
7525         spin_lock_irqsave(&conf->device_lock, flags);
7526         mddev->degraded = raid5_calc_degraded(conf);
7527         spin_unlock_irqrestore(&conf->device_lock, flags);
7528         print_raid5_conf(conf);
7529         return count;
7530 }
7531
7532 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7533 {
7534         struct r5conf *conf = mddev->private;
7535         int err = 0;
7536         int number = rdev->raid_disk;
7537         struct md_rdev **rdevp;
7538         struct disk_info *p = conf->disks + number;
7539
7540         print_raid5_conf(conf);
7541         if (test_bit(Journal, &rdev->flags) && conf->log) {
7542                 /*
7543                  * we can't wait pending write here, as this is called in
7544                  * raid5d, wait will deadlock.
7545                  * neilb: there is no locking about new writes here,
7546                  * so this cannot be safe.
7547                  */
7548                 if (atomic_read(&conf->active_stripes) ||
7549                     atomic_read(&conf->r5c_cached_full_stripes) ||
7550                     atomic_read(&conf->r5c_cached_partial_stripes)) {
7551                         return -EBUSY;
7552                 }
7553                 log_exit(conf);
7554                 return 0;
7555         }
7556         if (rdev == p->rdev)
7557                 rdevp = &p->rdev;
7558         else if (rdev == p->replacement)
7559                 rdevp = &p->replacement;
7560         else
7561                 return 0;
7562
7563         if (number >= conf->raid_disks &&
7564             conf->reshape_progress == MaxSector)
7565                 clear_bit(In_sync, &rdev->flags);
7566
7567         if (test_bit(In_sync, &rdev->flags) ||
7568             atomic_read(&rdev->nr_pending)) {
7569                 err = -EBUSY;
7570                 goto abort;
7571         }
7572         /* Only remove non-faulty devices if recovery
7573          * isn't possible.
7574          */
7575         if (!test_bit(Faulty, &rdev->flags) &&
7576             mddev->recovery_disabled != conf->recovery_disabled &&
7577             !has_failed(conf) &&
7578             (!p->replacement || p->replacement == rdev) &&
7579             number < conf->raid_disks) {
7580                 err = -EBUSY;
7581                 goto abort;
7582         }
7583         *rdevp = NULL;
7584         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7585                 synchronize_rcu();
7586                 if (atomic_read(&rdev->nr_pending)) {
7587                         /* lost the race, try later */
7588                         err = -EBUSY;
7589                         *rdevp = rdev;
7590                 }
7591         }
7592         if (!err) {
7593                 err = log_modify(conf, rdev, false);
7594                 if (err)
7595                         goto abort;
7596         }
7597         if (p->replacement) {
7598                 /* We must have just cleared 'rdev' */
7599                 p->rdev = p->replacement;
7600                 clear_bit(Replacement, &p->replacement->flags);
7601                 smp_mb(); /* Make sure other CPUs may see both as identical
7602                            * but will never see neither - if they are careful
7603                            */
7604                 p->replacement = NULL;
7605
7606                 if (!err)
7607                         err = log_modify(conf, p->rdev, true);
7608         }
7609
7610         clear_bit(WantReplacement, &rdev->flags);
7611 abort:
7612
7613         print_raid5_conf(conf);
7614         return err;
7615 }
7616
7617 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7618 {
7619         struct r5conf *conf = mddev->private;
7620         int err = -EEXIST;
7621         int disk;
7622         struct disk_info *p;
7623         int first = 0;
7624         int last = conf->raid_disks - 1;
7625
7626         if (test_bit(Journal, &rdev->flags)) {
7627                 if (conf->log)
7628                         return -EBUSY;
7629
7630                 rdev->raid_disk = 0;
7631                 /*
7632                  * The array is in readonly mode if journal is missing, so no
7633                  * write requests running. We should be safe
7634                  */
7635                 log_init(conf, rdev, false);
7636                 return 0;
7637         }
7638         if (mddev->recovery_disabled == conf->recovery_disabled)
7639                 return -EBUSY;
7640
7641         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7642                 /* no point adding a device */
7643                 return -EINVAL;
7644
7645         if (rdev->raid_disk >= 0)
7646                 first = last = rdev->raid_disk;
7647
7648         /*
7649          * find the disk ... but prefer rdev->saved_raid_disk
7650          * if possible.
7651          */
7652         if (rdev->saved_raid_disk >= 0 &&
7653             rdev->saved_raid_disk >= first &&
7654             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7655                 first = rdev->saved_raid_disk;
7656
7657         for (disk = first; disk <= last; disk++) {
7658                 p = conf->disks + disk;
7659                 if (p->rdev == NULL) {
7660                         clear_bit(In_sync, &rdev->flags);
7661                         rdev->raid_disk = disk;
7662                         if (rdev->saved_raid_disk != disk)
7663                                 conf->fullsync = 1;
7664                         rcu_assign_pointer(p->rdev, rdev);
7665
7666                         err = log_modify(conf, rdev, true);
7667
7668                         goto out;
7669                 }
7670         }
7671         for (disk = first; disk <= last; disk++) {
7672                 p = conf->disks + disk;
7673                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7674                     p->replacement == NULL) {
7675                         clear_bit(In_sync, &rdev->flags);
7676                         set_bit(Replacement, &rdev->flags);
7677                         rdev->raid_disk = disk;
7678                         err = 0;
7679                         conf->fullsync = 1;
7680                         rcu_assign_pointer(p->replacement, rdev);
7681                         break;
7682                 }
7683         }
7684 out:
7685         print_raid5_conf(conf);
7686         return err;
7687 }
7688
7689 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7690 {
7691         /* no resync is happening, and there is enough space
7692          * on all devices, so we can resize.
7693          * We need to make sure resync covers any new space.
7694          * If the array is shrinking we should possibly wait until
7695          * any io in the removed space completes, but it hardly seems
7696          * worth it.
7697          */
7698         sector_t newsize;
7699         struct r5conf *conf = mddev->private;
7700
7701         if (conf->log || raid5_has_ppl(conf))
7702                 return -EINVAL;
7703         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7704         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7705         if (mddev->external_size &&
7706             mddev->array_sectors > newsize)
7707                 return -EINVAL;
7708         if (mddev->bitmap) {
7709                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7710                 if (ret)
7711                         return ret;
7712         }
7713         md_set_array_sectors(mddev, newsize);
7714         if (sectors > mddev->dev_sectors &&
7715             mddev->recovery_cp > mddev->dev_sectors) {
7716                 mddev->recovery_cp = mddev->dev_sectors;
7717                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7718         }
7719         mddev->dev_sectors = sectors;
7720         mddev->resync_max_sectors = sectors;
7721         return 0;
7722 }
7723
7724 static int check_stripe_cache(struct mddev *mddev)
7725 {
7726         /* Can only proceed if there are plenty of stripe_heads.
7727          * We need a minimum of one full stripe,, and for sensible progress
7728          * it is best to have about 4 times that.
7729          * If we require 4 times, then the default 256 4K stripe_heads will
7730          * allow for chunk sizes up to 256K, which is probably OK.
7731          * If the chunk size is greater, user-space should request more
7732          * stripe_heads first.
7733          */
7734         struct r5conf *conf = mddev->private;
7735         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7736             > conf->min_nr_stripes ||
7737             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7738             > conf->min_nr_stripes) {
7739                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7740                         mdname(mddev),
7741                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7742                          / STRIPE_SIZE)*4);
7743                 return 0;
7744         }
7745         return 1;
7746 }
7747
7748 static int check_reshape(struct mddev *mddev)
7749 {
7750         struct r5conf *conf = mddev->private;
7751
7752         if (conf->log || raid5_has_ppl(conf))
7753                 return -EINVAL;
7754         if (mddev->delta_disks == 0 &&
7755             mddev->new_layout == mddev->layout &&
7756             mddev->new_chunk_sectors == mddev->chunk_sectors)
7757                 return 0; /* nothing to do */
7758         if (has_failed(conf))
7759                 return -EINVAL;
7760         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7761                 /* We might be able to shrink, but the devices must
7762                  * be made bigger first.
7763                  * For raid6, 4 is the minimum size.
7764                  * Otherwise 2 is the minimum
7765                  */
7766                 int min = 2;
7767                 if (mddev->level == 6)
7768                         min = 4;
7769                 if (mddev->raid_disks + mddev->delta_disks < min)
7770                         return -EINVAL;
7771         }
7772
7773         if (!check_stripe_cache(mddev))
7774                 return -ENOSPC;
7775
7776         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7777             mddev->delta_disks > 0)
7778                 if (resize_chunks(conf,
7779                                   conf->previous_raid_disks
7780                                   + max(0, mddev->delta_disks),
7781                                   max(mddev->new_chunk_sectors,
7782                                       mddev->chunk_sectors)
7783                             ) < 0)
7784                         return -ENOMEM;
7785
7786         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7787                 return 0; /* never bother to shrink */
7788         return resize_stripes(conf, (conf->previous_raid_disks
7789                                      + mddev->delta_disks));
7790 }
7791
7792 static int raid5_start_reshape(struct mddev *mddev)
7793 {
7794         struct r5conf *conf = mddev->private;
7795         struct md_rdev *rdev;
7796         int spares = 0;
7797         unsigned long flags;
7798
7799         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7800                 return -EBUSY;
7801
7802         if (!check_stripe_cache(mddev))
7803                 return -ENOSPC;
7804
7805         if (has_failed(conf))
7806                 return -EINVAL;
7807
7808         rdev_for_each(rdev, mddev) {
7809                 if (!test_bit(In_sync, &rdev->flags)
7810                     && !test_bit(Faulty, &rdev->flags))
7811                         spares++;
7812         }
7813
7814         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7815                 /* Not enough devices even to make a degraded array
7816                  * of that size
7817                  */
7818                 return -EINVAL;
7819
7820         /* Refuse to reduce size of the array.  Any reductions in
7821          * array size must be through explicit setting of array_size
7822          * attribute.
7823          */
7824         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7825             < mddev->array_sectors) {
7826                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7827                         mdname(mddev));
7828                 return -EINVAL;
7829         }
7830
7831         atomic_set(&conf->reshape_stripes, 0);
7832         spin_lock_irq(&conf->device_lock);
7833         write_seqcount_begin(&conf->gen_lock);
7834         conf->previous_raid_disks = conf->raid_disks;
7835         conf->raid_disks += mddev->delta_disks;
7836         conf->prev_chunk_sectors = conf->chunk_sectors;
7837         conf->chunk_sectors = mddev->new_chunk_sectors;
7838         conf->prev_algo = conf->algorithm;
7839         conf->algorithm = mddev->new_layout;
7840         conf->generation++;
7841         /* Code that selects data_offset needs to see the generation update
7842          * if reshape_progress has been set - so a memory barrier needed.
7843          */
7844         smp_mb();
7845         if (mddev->reshape_backwards)
7846                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7847         else
7848                 conf->reshape_progress = 0;
7849         conf->reshape_safe = conf->reshape_progress;
7850         write_seqcount_end(&conf->gen_lock);
7851         spin_unlock_irq(&conf->device_lock);
7852
7853         /* Now make sure any requests that proceeded on the assumption
7854          * the reshape wasn't running - like Discard or Read - have
7855          * completed.
7856          */
7857         mddev_suspend(mddev);
7858         mddev_resume(mddev);
7859
7860         /* Add some new drives, as many as will fit.
7861          * We know there are enough to make the newly sized array work.
7862          * Don't add devices if we are reducing the number of
7863          * devices in the array.  This is because it is not possible
7864          * to correctly record the "partially reconstructed" state of
7865          * such devices during the reshape and confusion could result.
7866          */
7867         if (mddev->delta_disks >= 0) {
7868                 rdev_for_each(rdev, mddev)
7869                         if (rdev->raid_disk < 0 &&
7870                             !test_bit(Faulty, &rdev->flags)) {
7871                                 if (raid5_add_disk(mddev, rdev) == 0) {
7872                                         if (rdev->raid_disk
7873                                             >= conf->previous_raid_disks)
7874                                                 set_bit(In_sync, &rdev->flags);
7875                                         else
7876                                                 rdev->recovery_offset = 0;
7877
7878                                         if (sysfs_link_rdev(mddev, rdev))
7879                                                 /* Failure here is OK */;
7880                                 }
7881                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7882                                    && !test_bit(Faulty, &rdev->flags)) {
7883                                 /* This is a spare that was manually added */
7884                                 set_bit(In_sync, &rdev->flags);
7885                         }
7886
7887                 /* When a reshape changes the number of devices,
7888                  * ->degraded is measured against the larger of the
7889                  * pre and post number of devices.
7890                  */
7891                 spin_lock_irqsave(&conf->device_lock, flags);
7892                 mddev->degraded = raid5_calc_degraded(conf);
7893                 spin_unlock_irqrestore(&conf->device_lock, flags);
7894         }
7895         mddev->raid_disks = conf->raid_disks;
7896         mddev->reshape_position = conf->reshape_progress;
7897         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7898
7899         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7900         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7901         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7902         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7903         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7904         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7905                                                 "reshape");
7906         if (!mddev->sync_thread) {
7907                 mddev->recovery = 0;
7908                 spin_lock_irq(&conf->device_lock);
7909                 write_seqcount_begin(&conf->gen_lock);
7910                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7911                 mddev->new_chunk_sectors =
7912                         conf->chunk_sectors = conf->prev_chunk_sectors;
7913                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7914                 rdev_for_each(rdev, mddev)
7915                         rdev->new_data_offset = rdev->data_offset;
7916                 smp_wmb();
7917                 conf->generation --;
7918                 conf->reshape_progress = MaxSector;
7919                 mddev->reshape_position = MaxSector;
7920                 write_seqcount_end(&conf->gen_lock);
7921                 spin_unlock_irq(&conf->device_lock);
7922                 return -EAGAIN;
7923         }
7924         conf->reshape_checkpoint = jiffies;
7925         md_wakeup_thread(mddev->sync_thread);
7926         md_new_event(mddev);
7927         return 0;
7928 }
7929
7930 /* This is called from the reshape thread and should make any
7931  * changes needed in 'conf'
7932  */
7933 static void end_reshape(struct r5conf *conf)
7934 {
7935
7936         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7937                 struct md_rdev *rdev;
7938
7939                 spin_lock_irq(&conf->device_lock);
7940                 conf->previous_raid_disks = conf->raid_disks;
7941                 rdev_for_each(rdev, conf->mddev)
7942                         rdev->data_offset = rdev->new_data_offset;
7943                 smp_wmb();
7944                 conf->reshape_progress = MaxSector;
7945                 conf->mddev->reshape_position = MaxSector;
7946                 spin_unlock_irq(&conf->device_lock);
7947                 wake_up(&conf->wait_for_overlap);
7948
7949                 /* read-ahead size must cover two whole stripes, which is
7950                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7951                  */
7952                 if (conf->mddev->queue) {
7953                         int data_disks = conf->raid_disks - conf->max_degraded;
7954                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7955                                                    / PAGE_SIZE);
7956                         if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7957                                 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7958                 }
7959         }
7960 }
7961
7962 /* This is called from the raid5d thread with mddev_lock held.
7963  * It makes config changes to the device.
7964  */
7965 static void raid5_finish_reshape(struct mddev *mddev)
7966 {
7967         struct r5conf *conf = mddev->private;
7968
7969         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7970
7971                 if (mddev->delta_disks > 0) {
7972                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7973                         if (mddev->queue) {
7974                                 set_capacity(mddev->gendisk, mddev->array_sectors);
7975                                 revalidate_disk(mddev->gendisk);
7976                         }
7977                 } else {
7978                         int d;
7979                         spin_lock_irq(&conf->device_lock);
7980                         mddev->degraded = raid5_calc_degraded(conf);
7981                         spin_unlock_irq(&conf->device_lock);
7982                         for (d = conf->raid_disks ;
7983                              d < conf->raid_disks - mddev->delta_disks;
7984                              d++) {
7985                                 struct md_rdev *rdev = conf->disks[d].rdev;
7986                                 if (rdev)
7987                                         clear_bit(In_sync, &rdev->flags);
7988                                 rdev = conf->disks[d].replacement;
7989                                 if (rdev)
7990                                         clear_bit(In_sync, &rdev->flags);
7991                         }
7992                 }
7993                 mddev->layout = conf->algorithm;
7994                 mddev->chunk_sectors = conf->chunk_sectors;
7995                 mddev->reshape_position = MaxSector;
7996                 mddev->delta_disks = 0;
7997                 mddev->reshape_backwards = 0;
7998         }
7999 }
8000
8001 static void raid5_quiesce(struct mddev *mddev, int state)
8002 {
8003         struct r5conf *conf = mddev->private;
8004
8005         switch(state) {
8006         case 2: /* resume for a suspend */
8007                 wake_up(&conf->wait_for_overlap);
8008                 break;
8009
8010         case 1: /* stop all writes */
8011                 lock_all_device_hash_locks_irq(conf);
8012                 /* '2' tells resync/reshape to pause so that all
8013                  * active stripes can drain
8014                  */
8015                 r5c_flush_cache(conf, INT_MAX);
8016                 conf->quiesce = 2;
8017                 wait_event_cmd(conf->wait_for_quiescent,
8018                                     atomic_read(&conf->active_stripes) == 0 &&
8019                                     atomic_read(&conf->active_aligned_reads) == 0,
8020                                     unlock_all_device_hash_locks_irq(conf),
8021                                     lock_all_device_hash_locks_irq(conf));
8022                 conf->quiesce = 1;
8023                 unlock_all_device_hash_locks_irq(conf);
8024                 /* allow reshape to continue */
8025                 wake_up(&conf->wait_for_overlap);
8026                 break;
8027
8028         case 0: /* re-enable writes */
8029                 lock_all_device_hash_locks_irq(conf);
8030                 conf->quiesce = 0;
8031                 wake_up(&conf->wait_for_quiescent);
8032                 wake_up(&conf->wait_for_overlap);
8033                 unlock_all_device_hash_locks_irq(conf);
8034                 break;
8035         }
8036         r5l_quiesce(conf->log, state);
8037 }
8038
8039 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8040 {
8041         struct r0conf *raid0_conf = mddev->private;
8042         sector_t sectors;
8043
8044         /* for raid0 takeover only one zone is supported */
8045         if (raid0_conf->nr_strip_zones > 1) {
8046                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8047                         mdname(mddev));
8048                 return ERR_PTR(-EINVAL);
8049         }
8050
8051         sectors = raid0_conf->strip_zone[0].zone_end;
8052         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8053         mddev->dev_sectors = sectors;
8054         mddev->new_level = level;
8055         mddev->new_layout = ALGORITHM_PARITY_N;
8056         mddev->new_chunk_sectors = mddev->chunk_sectors;
8057         mddev->raid_disks += 1;
8058         mddev->delta_disks = 1;
8059         /* make sure it will be not marked as dirty */
8060         mddev->recovery_cp = MaxSector;
8061
8062         return setup_conf(mddev);
8063 }
8064
8065 static void *raid5_takeover_raid1(struct mddev *mddev)
8066 {
8067         int chunksect;
8068         void *ret;
8069
8070         if (mddev->raid_disks != 2 ||
8071             mddev->degraded > 1)
8072                 return ERR_PTR(-EINVAL);
8073
8074         /* Should check if there are write-behind devices? */
8075
8076         chunksect = 64*2; /* 64K by default */
8077
8078         /* The array must be an exact multiple of chunksize */
8079         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8080                 chunksect >>= 1;
8081
8082         if ((chunksect<<9) < STRIPE_SIZE)
8083                 /* array size does not allow a suitable chunk size */
8084                 return ERR_PTR(-EINVAL);
8085
8086         mddev->new_level = 5;
8087         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8088         mddev->new_chunk_sectors = chunksect;
8089
8090         ret = setup_conf(mddev);
8091         if (!IS_ERR(ret))
8092                 mddev_clear_unsupported_flags(mddev,
8093                         UNSUPPORTED_MDDEV_FLAGS);
8094         return ret;
8095 }
8096
8097 static void *raid5_takeover_raid6(struct mddev *mddev)
8098 {
8099         int new_layout;
8100
8101         switch (mddev->layout) {
8102         case ALGORITHM_LEFT_ASYMMETRIC_6:
8103                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8104                 break;
8105         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8106                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8107                 break;
8108         case ALGORITHM_LEFT_SYMMETRIC_6:
8109                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8110                 break;
8111         case ALGORITHM_RIGHT_SYMMETRIC_6:
8112                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8113                 break;
8114         case ALGORITHM_PARITY_0_6:
8115                 new_layout = ALGORITHM_PARITY_0;
8116                 break;
8117         case ALGORITHM_PARITY_N:
8118                 new_layout = ALGORITHM_PARITY_N;
8119                 break;
8120         default:
8121                 return ERR_PTR(-EINVAL);
8122         }
8123         mddev->new_level = 5;
8124         mddev->new_layout = new_layout;
8125         mddev->delta_disks = -1;
8126         mddev->raid_disks -= 1;
8127         return setup_conf(mddev);
8128 }
8129
8130 static int raid5_check_reshape(struct mddev *mddev)
8131 {
8132         /* For a 2-drive array, the layout and chunk size can be changed
8133          * immediately as not restriping is needed.
8134          * For larger arrays we record the new value - after validation
8135          * to be used by a reshape pass.
8136          */
8137         struct r5conf *conf = mddev->private;
8138         int new_chunk = mddev->new_chunk_sectors;
8139
8140         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8141                 return -EINVAL;
8142         if (new_chunk > 0) {
8143                 if (!is_power_of_2(new_chunk))
8144                         return -EINVAL;
8145                 if (new_chunk < (PAGE_SIZE>>9))
8146                         return -EINVAL;
8147                 if (mddev->array_sectors & (new_chunk-1))
8148                         /* not factor of array size */
8149                         return -EINVAL;
8150         }
8151
8152         /* They look valid */
8153
8154         if (mddev->raid_disks == 2) {
8155                 /* can make the change immediately */
8156                 if (mddev->new_layout >= 0) {
8157                         conf->algorithm = mddev->new_layout;
8158                         mddev->layout = mddev->new_layout;
8159                 }
8160                 if (new_chunk > 0) {
8161                         conf->chunk_sectors = new_chunk ;
8162                         mddev->chunk_sectors = new_chunk;
8163                 }
8164                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8165                 md_wakeup_thread(mddev->thread);
8166         }
8167         return check_reshape(mddev);
8168 }
8169
8170 static int raid6_check_reshape(struct mddev *mddev)
8171 {
8172         int new_chunk = mddev->new_chunk_sectors;
8173
8174         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8175                 return -EINVAL;
8176         if (new_chunk > 0) {
8177                 if (!is_power_of_2(new_chunk))
8178                         return -EINVAL;
8179                 if (new_chunk < (PAGE_SIZE >> 9))
8180                         return -EINVAL;
8181                 if (mddev->array_sectors & (new_chunk-1))
8182                         /* not factor of array size */
8183                         return -EINVAL;
8184         }
8185
8186         /* They look valid */
8187         return check_reshape(mddev);
8188 }
8189
8190 static void *raid5_takeover(struct mddev *mddev)
8191 {
8192         /* raid5 can take over:
8193          *  raid0 - if there is only one strip zone - make it a raid4 layout
8194          *  raid1 - if there are two drives.  We need to know the chunk size
8195          *  raid4 - trivial - just use a raid4 layout.
8196          *  raid6 - Providing it is a *_6 layout
8197          */
8198         if (mddev->level == 0)
8199                 return raid45_takeover_raid0(mddev, 5);
8200         if (mddev->level == 1)
8201                 return raid5_takeover_raid1(mddev);
8202         if (mddev->level == 4) {
8203                 mddev->new_layout = ALGORITHM_PARITY_N;
8204                 mddev->new_level = 5;
8205                 return setup_conf(mddev);
8206         }
8207         if (mddev->level == 6)
8208                 return raid5_takeover_raid6(mddev);
8209
8210         return ERR_PTR(-EINVAL);
8211 }
8212
8213 static void *raid4_takeover(struct mddev *mddev)
8214 {
8215         /* raid4 can take over:
8216          *  raid0 - if there is only one strip zone
8217          *  raid5 - if layout is right
8218          */
8219         if (mddev->level == 0)
8220                 return raid45_takeover_raid0(mddev, 4);
8221         if (mddev->level == 5 &&
8222             mddev->layout == ALGORITHM_PARITY_N) {
8223                 mddev->new_layout = 0;
8224                 mddev->new_level = 4;
8225                 return setup_conf(mddev);
8226         }
8227         return ERR_PTR(-EINVAL);
8228 }
8229
8230 static struct md_personality raid5_personality;
8231
8232 static void *raid6_takeover(struct mddev *mddev)
8233 {
8234         /* Currently can only take over a raid5.  We map the
8235          * personality to an equivalent raid6 personality
8236          * with the Q block at the end.
8237          */
8238         int new_layout;
8239
8240         if (mddev->pers != &raid5_personality)
8241                 return ERR_PTR(-EINVAL);
8242         if (mddev->degraded > 1)
8243                 return ERR_PTR(-EINVAL);
8244         if (mddev->raid_disks > 253)
8245                 return ERR_PTR(-EINVAL);
8246         if (mddev->raid_disks < 3)
8247                 return ERR_PTR(-EINVAL);
8248
8249         switch (mddev->layout) {
8250         case ALGORITHM_LEFT_ASYMMETRIC:
8251                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8252                 break;
8253         case ALGORITHM_RIGHT_ASYMMETRIC:
8254                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8255                 break;
8256         case ALGORITHM_LEFT_SYMMETRIC:
8257                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8258                 break;
8259         case ALGORITHM_RIGHT_SYMMETRIC:
8260                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8261                 break;
8262         case ALGORITHM_PARITY_0:
8263                 new_layout = ALGORITHM_PARITY_0_6;
8264                 break;
8265         case ALGORITHM_PARITY_N:
8266                 new_layout = ALGORITHM_PARITY_N;
8267                 break;
8268         default:
8269                 return ERR_PTR(-EINVAL);
8270         }
8271         mddev->new_level = 6;
8272         mddev->new_layout = new_layout;
8273         mddev->delta_disks = 1;
8274         mddev->raid_disks += 1;
8275         return setup_conf(mddev);
8276 }
8277
8278 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8279 {
8280         struct r5conf *conf;
8281         int err;
8282
8283         err = mddev_lock(mddev);
8284         if (err)
8285                 return err;
8286         conf = mddev->private;
8287         if (!conf) {
8288                 mddev_unlock(mddev);
8289                 return -ENODEV;
8290         }
8291
8292         if (strncmp(buf, "ppl", 3) == 0) {
8293                 /* ppl only works with RAID 5 */
8294                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8295                         err = log_init(conf, NULL, true);
8296                         if (!err) {
8297                                 err = resize_stripes(conf, conf->pool_size);
8298                                 if (err)
8299                                         log_exit(conf);
8300                         }
8301                 } else
8302                         err = -EINVAL;
8303         } else if (strncmp(buf, "resync", 6) == 0) {
8304                 if (raid5_has_ppl(conf)) {
8305                         mddev_suspend(mddev);
8306                         log_exit(conf);
8307                         mddev_resume(mddev);
8308                         err = resize_stripes(conf, conf->pool_size);
8309                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8310                            r5l_log_disk_error(conf)) {
8311                         bool journal_dev_exists = false;
8312                         struct md_rdev *rdev;
8313
8314                         rdev_for_each(rdev, mddev)
8315                                 if (test_bit(Journal, &rdev->flags)) {
8316                                         journal_dev_exists = true;
8317                                         break;
8318                                 }
8319
8320                         if (!journal_dev_exists) {
8321                                 mddev_suspend(mddev);
8322                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8323                                 mddev_resume(mddev);
8324                         } else  /* need remove journal device first */
8325                                 err = -EBUSY;
8326                 } else
8327                         err = -EINVAL;
8328         } else {
8329                 err = -EINVAL;
8330         }
8331
8332         if (!err)
8333                 md_update_sb(mddev, 1);
8334
8335         mddev_unlock(mddev);
8336
8337         return err;
8338 }
8339
8340 static struct md_personality raid6_personality =
8341 {
8342         .name           = "raid6",
8343         .level          = 6,
8344         .owner          = THIS_MODULE,
8345         .make_request   = raid5_make_request,
8346         .run            = raid5_run,
8347         .free           = raid5_free,
8348         .status         = raid5_status,
8349         .error_handler  = raid5_error,
8350         .hot_add_disk   = raid5_add_disk,
8351         .hot_remove_disk= raid5_remove_disk,
8352         .spare_active   = raid5_spare_active,
8353         .sync_request   = raid5_sync_request,
8354         .resize         = raid5_resize,
8355         .size           = raid5_size,
8356         .check_reshape  = raid6_check_reshape,
8357         .start_reshape  = raid5_start_reshape,
8358         .finish_reshape = raid5_finish_reshape,
8359         .quiesce        = raid5_quiesce,
8360         .takeover       = raid6_takeover,
8361         .congested      = raid5_congested,
8362         .change_consistency_policy = raid5_change_consistency_policy,
8363 };
8364 static struct md_personality raid5_personality =
8365 {
8366         .name           = "raid5",
8367         .level          = 5,
8368         .owner          = THIS_MODULE,
8369         .make_request   = raid5_make_request,
8370         .run            = raid5_run,
8371         .free           = raid5_free,
8372         .status         = raid5_status,
8373         .error_handler  = raid5_error,
8374         .hot_add_disk   = raid5_add_disk,
8375         .hot_remove_disk= raid5_remove_disk,
8376         .spare_active   = raid5_spare_active,
8377         .sync_request   = raid5_sync_request,
8378         .resize         = raid5_resize,
8379         .size           = raid5_size,
8380         .check_reshape  = raid5_check_reshape,
8381         .start_reshape  = raid5_start_reshape,
8382         .finish_reshape = raid5_finish_reshape,
8383         .quiesce        = raid5_quiesce,
8384         .takeover       = raid5_takeover,
8385         .congested      = raid5_congested,
8386         .change_consistency_policy = raid5_change_consistency_policy,
8387 };
8388
8389 static struct md_personality raid4_personality =
8390 {
8391         .name           = "raid4",
8392         .level          = 4,
8393         .owner          = THIS_MODULE,
8394         .make_request   = raid5_make_request,
8395         .run            = raid5_run,
8396         .free           = raid5_free,
8397         .status         = raid5_status,
8398         .error_handler  = raid5_error,
8399         .hot_add_disk   = raid5_add_disk,
8400         .hot_remove_disk= raid5_remove_disk,
8401         .spare_active   = raid5_spare_active,
8402         .sync_request   = raid5_sync_request,
8403         .resize         = raid5_resize,
8404         .size           = raid5_size,
8405         .check_reshape  = raid5_check_reshape,
8406         .start_reshape  = raid5_start_reshape,
8407         .finish_reshape = raid5_finish_reshape,
8408         .quiesce        = raid5_quiesce,
8409         .takeover       = raid4_takeover,
8410         .congested      = raid5_congested,
8411         .change_consistency_policy = raid5_change_consistency_policy,
8412 };
8413
8414 static int __init raid5_init(void)
8415 {
8416         int ret;
8417
8418         raid5_wq = alloc_workqueue("raid5wq",
8419                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8420         if (!raid5_wq)
8421                 return -ENOMEM;
8422
8423         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8424                                       "md/raid5:prepare",
8425                                       raid456_cpu_up_prepare,
8426                                       raid456_cpu_dead);
8427         if (ret) {
8428                 destroy_workqueue(raid5_wq);
8429                 return ret;
8430         }
8431         register_md_personality(&raid6_personality);
8432         register_md_personality(&raid5_personality);
8433         register_md_personality(&raid4_personality);
8434         return 0;
8435 }
8436
8437 static void raid5_exit(void)
8438 {
8439         unregister_md_personality(&raid6_personality);
8440         unregister_md_personality(&raid5_personality);
8441         unregister_md_personality(&raid4_personality);
8442         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8443         destroy_workqueue(raid5_wq);
8444 }
8445
8446 module_init(raid5_init);
8447 module_exit(raid5_exit);
8448 MODULE_LICENSE("GPL");
8449 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8450 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8451 MODULE_ALIAS("md-raid5");
8452 MODULE_ALIAS("md-raid4");
8453 MODULE_ALIAS("md-level-5");
8454 MODULE_ALIAS("md-level-4");
8455 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8456 MODULE_ALIAS("md-raid6");
8457 MODULE_ALIAS("md-level-6");
8458
8459 /* This used to be two separate modules, they were: */
8460 MODULE_ALIAS("raid5");
8461 MODULE_ALIAS("raid6");