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 <trace/events/block.h>
  59
  60 #include "md.h"
  61 #include "raid5.h"
  62 #include "raid0.h"
  63 #include "bitmap.h"
  64
  65 #define cpu_to_group(cpu) cpu_to_node(cpu)
  66 #define ANY_GROUP NUMA_NO_NODE
  67
  68 static bool devices_handle_discard_safely = false;
  69 module_param(devices_handle_discard_safely, bool, 0644);
  70 MODULE_PARM_DESC(devices_handle_discard_safely,
  71                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  72 static struct workqueue_struct *raid5_wq;
  73 /*
  74  * Stripe cache
  75  */
  76
  77 #define NR_STRIPES              256
  78 #define STRIPE_SIZE             PAGE_SIZE
  79 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  80 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  81 #define IO_THRESHOLD            1
  82 #define BYPASS_THRESHOLD        1
  83 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
  84 #define HASH_MASK               (NR_HASH - 1)
  85 #define MAX_STRIPE_BATCH        8
  86
  87 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  88 {
  89         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  90         return &conf->stripe_hashtbl[hash];
  91 }
  92
  93 static inline int stripe_hash_locks_hash(sector_t sect)
  94 {
  95         return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
  96 }
  97
  98 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  99 {
 100         spin_lock_irq(conf->hash_locks + hash);
 101         spin_lock(&conf->device_lock);
 102 }
 103
 104 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
 105 {
 106         spin_unlock(&conf->device_lock);
 107         spin_unlock_irq(conf->hash_locks + hash);
 108 }
 109
 110 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
 111 {
 112         int i;
 113         spin_lock_irq(conf->hash_locks);
 114         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 115                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 116         spin_lock(&conf->device_lock);
 117 }
 118
 119 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 120 {
 121         int i;
 122         spin_unlock(&conf->device_lock);
 123         for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 124                 spin_unlock(conf->hash_locks + i);
 125         spin_unlock_irq(conf->hash_locks);
 126 }
 127
 128 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
 129  * order without overlap.  There may be several bio's per stripe+device, and
 130  * a bio could span several devices.
 131  * When walking this list for a particular stripe+device, we must never proceed
 132  * beyond a bio that extends past this device, as the next bio might no longer
 133  * be valid.
 134  * This function is used to determine the 'next' bio in the list, given the sector
 135  * of the current stripe+device
 136  */
 137 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
 138 {
 139         int sectors = bio_sectors(bio);
 140         if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
 141                 return bio->bi_next;
 142         else
 143                 return NULL;
 144 }
 145
 146 /*
 147  * We maintain a biased count of active stripes in the bottom 16 bits of
 148  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
 149  */
 150 static inline int raid5_bi_processed_stripes(struct bio *bio)
 151 {
 152         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 153         return (atomic_read(segments) >> 16) & 0xffff;
 154 }
 155
 156 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
 157 {
 158         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 159         return atomic_sub_return(1, segments) & 0xffff;
 160 }
 161
 162 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
 163 {
 164         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 165         atomic_inc(segments);
 166 }
 167
 168 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
 169         unsigned int cnt)
 170 {
 171         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 172         int old, new;
 173
 174         do {
 175                 old = atomic_read(segments);
 176                 new = (old & 0xffff) | (cnt << 16);
 177         } while (atomic_cmpxchg(segments, old, new) != old);
 178 }
 179
 180 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
 181 {
 182         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 183         atomic_set(segments, cnt);
 184 }
 185
 186 /* Find first data disk in a raid6 stripe */
 187 static inline int raid6_d0(struct stripe_head *sh)
 188 {
 189         if (sh->ddf_layout)
 190                 /* ddf always start from first device */
 191                 return 0;
 192         /* md starts just after Q block */
 193         if (sh->qd_idx == sh->disks - 1)
 194                 return 0;
 195         else
 196                 return sh->qd_idx + 1;
 197 }
 198 static inline int raid6_next_disk(int disk, int raid_disks)
 199 {
 200         disk++;
 201         return (disk < raid_disks) ? disk : 0;
 202 }
 203
 204 /* When walking through the disks in a raid5, starting at raid6_d0,
 205  * We need to map each disk to a 'slot', where the data disks are slot
 206  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 207  * is raid_disks-1.  This help does that mapping.
 208  */
 209 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 210                              int *count, int syndrome_disks)
 211 {
 212         int slot = *count;
 213
 214         if (sh->ddf_layout)
 215                 (*count)++;
 216         if (idx == sh->pd_idx)
 217                 return syndrome_disks;
 218         if (idx == sh->qd_idx)
 219                 return syndrome_disks + 1;
 220         if (!sh->ddf_layout)
 221                 (*count)++;
 222         return slot;
 223 }
 224
 225 static void return_io(struct bio_list *return_bi)
 226 {
 227         struct bio *bi;
 228         while ((bi = bio_list_pop(return_bi)) != NULL) {
 229                 bi->bi_iter.bi_size = 0;
 230                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
 231                                          bi, 0);
 232                 bio_endio(bi);
 233         }
 234 }
 235
 236 static void print_raid5_conf (struct r5conf *conf);
 237
 238 static int stripe_operations_active(struct stripe_head *sh)
 239 {
 240         return sh->check_state || sh->reconstruct_state ||
 241                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 242                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 243 }
 244
 245 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 246 {
 247         struct r5conf *conf = sh->raid_conf;
 248         struct r5worker_group *group;
 249         int thread_cnt;
 250         int i, cpu = sh->cpu;
 251
 252         if (!cpu_online(cpu)) {
 253                 cpu = cpumask_any(cpu_online_mask);
 254                 sh->cpu = cpu;
 255         }
 256
 257         if (list_empty(&sh->lru)) {
 258                 struct r5worker_group *group;
 259                 group = conf->worker_groups + cpu_to_group(cpu);
 260                 list_add_tail(&sh->lru, &group->handle_list);
 261                 group->stripes_cnt++;
 262                 sh->group = group;
 263         }
 264
 265         if (conf->worker_cnt_per_group == 0) {
 266                 md_wakeup_thread(conf->mddev->thread);
 267                 return;
 268         }
 269
 270         group = conf->worker_groups + cpu_to_group(sh->cpu);
 271
 272         group->workers[0].working = true;
 273         /* at least one worker should run to avoid race */
 274         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 275
 276         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 277         /* wakeup more workers */
 278         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 279                 if (group->workers[i].working == false) {
 280                         group->workers[i].working = true;
 281                         queue_work_on(sh->cpu, raid5_wq,
 282                                       &group->workers[i].work);
 283                         thread_cnt--;
 284                 }
 285         }
 286 }
 287
 288 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 289                               struct list_head *temp_inactive_list)
 290 {
 291         BUG_ON(!list_empty(&sh->lru));
 292         BUG_ON(atomic_read(&conf->active_stripes)==0);
 293         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 294                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 295                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 296                         list_add_tail(&sh->lru, &conf->delayed_list);
 297                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 298                            sh->bm_seq - conf->seq_write > 0)
 299                         list_add_tail(&sh->lru, &conf->bitmap_list);
 300                 else {
 301                         clear_bit(STRIPE_DELAYED, &sh->state);
 302                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 303                         if (conf->worker_cnt_per_group == 0) {
 304                                 list_add_tail(&sh->lru, &conf->handle_list);
 305                         } else {
 306                                 raid5_wakeup_stripe_thread(sh);
 307                                 return;
 308                         }
 309                 }
 310                 md_wakeup_thread(conf->mddev->thread);
 311         } else {
 312                 BUG_ON(stripe_operations_active(sh));
 313                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 314                         if (atomic_dec_return(&conf->preread_active_stripes)
 315                             < IO_THRESHOLD)
 316                                 md_wakeup_thread(conf->mddev->thread);
 317                 atomic_dec(&conf->active_stripes);
 318                 if (!test_bit(STRIPE_EXPANDING, &sh->state))
 319                         list_add_tail(&sh->lru, temp_inactive_list);
 320         }
 321 }
 322
 323 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 324                              struct list_head *temp_inactive_list)
 325 {
 326         if (atomic_dec_and_test(&sh->count))
 327                 do_release_stripe(conf, sh, temp_inactive_list);
 328 }
 329
 330 /*
 331  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 332  *
 333  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 334  * given time. Adding stripes only takes device lock, while deleting stripes
 335  * only takes hash lock.
 336  */
 337 static void release_inactive_stripe_list(struct r5conf *conf,
 338                                          struct list_head *temp_inactive_list,
 339                                          int hash)
 340 {
 341         int size;
 342         bool do_wakeup = false;
 343         unsigned long flags;
 344
 345         if (hash == NR_STRIPE_HASH_LOCKS) {
 346                 size = NR_STRIPE_HASH_LOCKS;
 347                 hash = NR_STRIPE_HASH_LOCKS - 1;
 348         } else
 349                 size = 1;
 350         while (size) {
 351                 struct list_head *list = &temp_inactive_list[size - 1];
 352
 353                 /*
 354                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 355                  * remove stripes from the list
 356                  */
 357                 if (!list_empty_careful(list)) {
 358                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 359                         if (list_empty(conf->inactive_list + hash) &&
 360                             !list_empty(list))
 361                                 atomic_dec(&conf->empty_inactive_list_nr);
 362                         list_splice_tail_init(list, conf->inactive_list + hash);
 363                         do_wakeup = true;
 364                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 365                 }
 366                 size--;
 367                 hash--;
 368         }
 369
 370         if (do_wakeup) {
 371                 wake_up(&conf->wait_for_stripe);
 372                 if (atomic_read(&conf->active_stripes) == 0)
 373                         wake_up(&conf->wait_for_quiescent);
 374                 if (conf->retry_read_aligned)
 375                         md_wakeup_thread(conf->mddev->thread);
 376         }
 377 }
 378
 379 /* should hold conf->device_lock already */
 380 static int release_stripe_list(struct r5conf *conf,
 381                                struct list_head *temp_inactive_list)
 382 {
 383         struct stripe_head *sh;
 384         int count = 0;
 385         struct llist_node *head;
 386
 387         head = llist_del_all(&conf->released_stripes);
 388         head = llist_reverse_order(head);
 389         while (head) {
 390                 int hash;
 391
 392                 sh = llist_entry(head, struct stripe_head, release_list);
 393                 head = llist_next(head);
 394                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 395                 smp_mb();
 396                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 397                 /*
 398                  * Don't worry the bit is set here, because if the bit is set
 399                  * again, the count is always > 1. This is true for
 400                  * STRIPE_ON_UNPLUG_LIST bit too.
 401                  */
 402                 hash = sh->hash_lock_index;
 403                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 404                 count++;
 405         }
 406
 407         return count;
 408 }
 409
 410 void raid5_release_stripe(struct stripe_head *sh)
 411 {
 412         struct r5conf *conf = sh->raid_conf;
 413         unsigned long flags;
 414         struct list_head list;
 415         int hash;
 416         bool wakeup;
 417
 418         /* Avoid release_list until the last reference.
 419          */
 420         if (atomic_add_unless(&sh->count, -1, 1))
 421                 return;
 422
 423         if (unlikely(!conf->mddev->thread) ||
 424                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 425                 goto slow_path;
 426         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 427         if (wakeup)
 428                 md_wakeup_thread(conf->mddev->thread);
 429         return;
 430 slow_path:
 431         local_irq_save(flags);
 432         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 433         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 434                 INIT_LIST_HEAD(&list);
 435                 hash = sh->hash_lock_index;
 436                 do_release_stripe(conf, sh, &list);
 437                 spin_unlock(&conf->device_lock);
 438                 release_inactive_stripe_list(conf, &list, hash);
 439         }
 440         local_irq_restore(flags);
 441 }
 442
 443 static inline void remove_hash(struct stripe_head *sh)
 444 {
 445         pr_debug("remove_hash(), stripe %llu\n",
 446                 (unsigned long long)sh->sector);
 447
 448         hlist_del_init(&sh->hash);
 449 }
 450
 451 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 452 {
 453         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 454
 455         pr_debug("insert_hash(), stripe %llu\n",
 456                 (unsigned long long)sh->sector);
 457
 458         hlist_add_head(&sh->hash, hp);
 459 }
 460
 461 /* find an idle stripe, make sure it is unhashed, and return it. */
 462 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 463 {
 464         struct stripe_head *sh = NULL;
 465         struct list_head *first;
 466
 467         if (list_empty(conf->inactive_list + hash))
 468                 goto out;
 469         first = (conf->inactive_list + hash)->next;
 470         sh = list_entry(first, struct stripe_head, lru);
 471         list_del_init(first);
 472         remove_hash(sh);
 473         atomic_inc(&conf->active_stripes);
 474         BUG_ON(hash != sh->hash_lock_index);
 475         if (list_empty(conf->inactive_list + hash))
 476                 atomic_inc(&conf->empty_inactive_list_nr);
 477 out:
 478         return sh;
 479 }
 480
 481 static void shrink_buffers(struct stripe_head *sh)
 482 {
 483         struct page *p;
 484         int i;
 485         int num = sh->raid_conf->pool_size;
 486
 487         for (i = 0; i < num ; i++) {
 488                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 489                 p = sh->dev[i].page;
 490                 if (!p)
 491                         continue;
 492                 sh->dev[i].page = NULL;
 493                 put_page(p);
 494         }
 495 }
 496
 497 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 498 {
 499         int i;
 500         int num = sh->raid_conf->pool_size;
 501
 502         for (i = 0; i < num; i++) {
 503                 struct page *page;
 504
 505                 if (!(page = alloc_page(gfp))) {
 506                         return 1;
 507                 }
 508                 sh->dev[i].page = page;
 509                 sh->dev[i].orig_page = page;
 510         }
 511         return 0;
 512 }
 513
 514 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 515 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 516                             struct stripe_head *sh);
 517
 518 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 519 {
 520         struct r5conf *conf = sh->raid_conf;
 521         int i, seq;
 522
 523         BUG_ON(atomic_read(&sh->count) != 0);
 524         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 525         BUG_ON(stripe_operations_active(sh));
 526         BUG_ON(sh->batch_head);
 527
 528         pr_debug("init_stripe called, stripe %llu\n",
 529                 (unsigned long long)sector);
 530 retry:
 531         seq = read_seqcount_begin(&conf->gen_lock);
 532         sh->generation = conf->generation - previous;
 533         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 534         sh->sector = sector;
 535         stripe_set_idx(sector, conf, previous, sh);
 536         sh->state = 0;
 537
 538         for (i = sh->disks; i--; ) {
 539                 struct r5dev *dev = &sh->dev[i];
 540
 541                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 542                     test_bit(R5_LOCKED, &dev->flags)) {
 543                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 544                                (unsigned long long)sh->sector, i, dev->toread,
 545                                dev->read, dev->towrite, dev->written,
 546                                test_bit(R5_LOCKED, &dev->flags));
 547                         WARN_ON(1);
 548                 }
 549                 dev->flags = 0;
 550                 raid5_build_block(sh, i, previous);
 551         }
 552         if (read_seqcount_retry(&conf->gen_lock, seq))
 553                 goto retry;
 554         sh->overwrite_disks = 0;
 555         insert_hash(conf, sh);
 556         sh->cpu = smp_processor_id();
 557         set_bit(STRIPE_BATCH_READY, &sh->state);
 558 }
 559
 560 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 561                                          short generation)
 562 {
 563         struct stripe_head *sh;
 564
 565         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 566         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 567                 if (sh->sector == sector && sh->generation == generation)
 568                         return sh;
 569         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 570         return NULL;
 571 }
 572
 573 /*
 574  * Need to check if array has failed when deciding whether to:
 575  *  - start an array
 576  *  - remove non-faulty devices
 577  *  - add a spare
 578  *  - allow a reshape
 579  * This determination is simple when no reshape is happening.
 580  * However if there is a reshape, we need to carefully check
 581  * both the before and after sections.
 582  * This is because some failed devices may only affect one
 583  * of the two sections, and some non-in_sync devices may
 584  * be insync in the section most affected by failed devices.
 585  */
 586 static int calc_degraded(struct r5conf *conf)
 587 {
 588         int degraded, degraded2;
 589         int i;
 590
 591         rcu_read_lock();
 592         degraded = 0;
 593         for (i = 0; i < conf->previous_raid_disks; i++) {
 594                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 595                 if (rdev && test_bit(Faulty, &rdev->flags))
 596                         rdev = rcu_dereference(conf->disks[i].replacement);
 597                 if (!rdev || test_bit(Faulty, &rdev->flags))
 598                         degraded++;
 599                 else if (test_bit(In_sync, &rdev->flags))
 600                         ;
 601                 else
 602                         /* not in-sync or faulty.
 603                          * If the reshape increases the number of devices,
 604                          * this is being recovered by the reshape, so
 605                          * this 'previous' section is not in_sync.
 606                          * If the number of devices is being reduced however,
 607                          * the device can only be part of the array if
 608                          * we are reverting a reshape, so this section will
 609                          * be in-sync.
 610                          */
 611                         if (conf->raid_disks >= conf->previous_raid_disks)
 612                                 degraded++;
 613         }
 614         rcu_read_unlock();
 615         if (conf->raid_disks == conf->previous_raid_disks)
 616                 return degraded;
 617         rcu_read_lock();
 618         degraded2 = 0;
 619         for (i = 0; i < conf->raid_disks; i++) {
 620                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 621                 if (rdev && test_bit(Faulty, &rdev->flags))
 622                         rdev = rcu_dereference(conf->disks[i].replacement);
 623                 if (!rdev || test_bit(Faulty, &rdev->flags))
 624                         degraded2++;
 625                 else if (test_bit(In_sync, &rdev->flags))
 626                         ;
 627                 else
 628                         /* not in-sync or faulty.
 629                          * If reshape increases the number of devices, this
 630                          * section has already been recovered, else it
 631                          * almost certainly hasn't.
 632                          */
 633                         if (conf->raid_disks <= conf->previous_raid_disks)
 634                                 degraded2++;
 635         }
 636         rcu_read_unlock();
 637         if (degraded2 > degraded)
 638                 return degraded2;
 639         return degraded;
 640 }
 641
 642 static int has_failed(struct r5conf *conf)
 643 {
 644         int degraded;
 645
 646         if (conf->mddev->reshape_position == MaxSector)
 647                 return conf->mddev->degraded > conf->max_degraded;
 648
 649         degraded = calc_degraded(conf);
 650         if (degraded > conf->max_degraded)
 651                 return 1;
 652         return 0;
 653 }
 654
 655 struct stripe_head *
 656 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 657                         int previous, int noblock, int noquiesce)
 658 {
 659         struct stripe_head *sh;
 660         int hash = stripe_hash_locks_hash(sector);
 661         int inc_empty_inactive_list_flag;
 662
 663         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 664
 665         spin_lock_irq(conf->hash_locks + hash);
 666
 667         do {
 668                 wait_event_lock_irq(conf->wait_for_quiescent,
 669                                     conf->quiesce == 0 || noquiesce,
 670                                     *(conf->hash_locks + hash));
 671                 sh = __find_stripe(conf, sector, conf->generation - previous);
 672                 if (!sh) {
 673                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 674                                 sh = get_free_stripe(conf, hash);
 675                                 if (!sh && !test_bit(R5_DID_ALLOC,
 676                                                      &conf->cache_state))
 677                                         set_bit(R5_ALLOC_MORE,
 678                                                 &conf->cache_state);
 679                         }
 680                         if (noblock && sh == NULL)
 681                                 break;
 682                         if (!sh) {
 683                                 set_bit(R5_INACTIVE_BLOCKED,
 684                                         &conf->cache_state);
 685                                 wait_event_lock_irq(
 686                                         conf->wait_for_stripe,
 687                                         !list_empty(conf->inactive_list + hash) &&
 688                                         (atomic_read(&conf->active_stripes)
 689                                          < (conf->max_nr_stripes * 3 / 4)
 690                                          || !test_bit(R5_INACTIVE_BLOCKED,
 691                                                       &conf->cache_state)),
 692                                         *(conf->hash_locks + hash));
 693                                 clear_bit(R5_INACTIVE_BLOCKED,
 694                                           &conf->cache_state);
 695                         } else {
 696                                 init_stripe(sh, sector, previous);
 697                                 atomic_inc(&sh->count);
 698                         }
 699                 } else if (!atomic_inc_not_zero(&sh->count)) {
 700                         spin_lock(&conf->device_lock);
 701                         if (!atomic_read(&sh->count)) {
 702                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 703                                         atomic_inc(&conf->active_stripes);
 704                                 BUG_ON(list_empty(&sh->lru) &&
 705                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 706                                 inc_empty_inactive_list_flag = 0;
 707                                 if (!list_empty(conf->inactive_list + hash))
 708                                         inc_empty_inactive_list_flag = 1;
 709                                 list_del_init(&sh->lru);
 710                                 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 711                                         atomic_inc(&conf->empty_inactive_list_nr);
 712                                 if (sh->group) {
 713                                         sh->group->stripes_cnt--;
 714                                         sh->group = NULL;
 715                                 }
 716                         }
 717                         atomic_inc(&sh->count);
 718                         spin_unlock(&conf->device_lock);
 719                 }
 720         } while (sh == NULL);
 721
 722         spin_unlock_irq(conf->hash_locks + hash);
 723         return sh;
 724 }
 725
 726 static bool is_full_stripe_write(struct stripe_head *sh)
 727 {
 728         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 729         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 730 }
 731
 732 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 733 {
 734         if (sh1 > sh2) {
 735                 spin_lock_irq(&sh2->stripe_lock);
 736                 spin_lock_nested(&sh1->stripe_lock, 1);
 737         } else {
 738                 spin_lock_irq(&sh1->stripe_lock);
 739                 spin_lock_nested(&sh2->stripe_lock, 1);
 740         }
 741 }
 742
 743 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 744 {
 745         spin_unlock(&sh1->stripe_lock);
 746         spin_unlock_irq(&sh2->stripe_lock);
 747 }
 748
 749 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 750 static bool stripe_can_batch(struct stripe_head *sh)
 751 {
 752         struct r5conf *conf = sh->raid_conf;
 753
 754         if (conf->log)
 755                 return false;
 756         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 757                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 758                 is_full_stripe_write(sh);
 759 }
 760
 761 /* we only do back search */
 762 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 763 {
 764         struct stripe_head *head;
 765         sector_t head_sector, tmp_sec;
 766         int hash;
 767         int dd_idx;
 768         int inc_empty_inactive_list_flag;
 769
 770         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 771         tmp_sec = sh->sector;
 772         if (!sector_div(tmp_sec, conf->chunk_sectors))
 773                 return;
 774         head_sector = sh->sector - STRIPE_SECTORS;
 775
 776         hash = stripe_hash_locks_hash(head_sector);
 777         spin_lock_irq(conf->hash_locks + hash);
 778         head = __find_stripe(conf, head_sector, conf->generation);
 779         if (head && !atomic_inc_not_zero(&head->count)) {
 780                 spin_lock(&conf->device_lock);
 781                 if (!atomic_read(&head->count)) {
 782                         if (!test_bit(STRIPE_HANDLE, &head->state))
 783                                 atomic_inc(&conf->active_stripes);
 784                         BUG_ON(list_empty(&head->lru) &&
 785                                !test_bit(STRIPE_EXPANDING, &head->state));
 786                         inc_empty_inactive_list_flag = 0;
 787                         if (!list_empty(conf->inactive_list + hash))
 788                                 inc_empty_inactive_list_flag = 1;
 789                         list_del_init(&head->lru);
 790                         if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 791                                 atomic_inc(&conf->empty_inactive_list_nr);
 792                         if (head->group) {
 793                                 head->group->stripes_cnt--;
 794                                 head->group = NULL;
 795                         }
 796                 }
 797                 atomic_inc(&head->count);
 798                 spin_unlock(&conf->device_lock);
 799         }
 800         spin_unlock_irq(conf->hash_locks + hash);
 801
 802         if (!head)
 803                 return;
 804         if (!stripe_can_batch(head))
 805                 goto out;
 806
 807         lock_two_stripes(head, sh);
 808         /* clear_batch_ready clear the flag */
 809         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 810                 goto unlock_out;
 811
 812         if (sh->batch_head)
 813                 goto unlock_out;
 814
 815         dd_idx = 0;
 816         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 817                 dd_idx++;
 818         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 819             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 820                 goto unlock_out;
 821
 822         if (head->batch_head) {
 823                 spin_lock(&head->batch_head->batch_lock);
 824                 /* This batch list is already running */
 825                 if (!stripe_can_batch(head)) {
 826                         spin_unlock(&head->batch_head->batch_lock);
 827                         goto unlock_out;
 828                 }
 829                 /*
 830                  * We must assign batch_head of this stripe within the
 831                  * batch_lock, otherwise clear_batch_ready of batch head
 832                  * stripe could clear BATCH_READY bit of this stripe and
 833                  * this stripe->batch_head doesn't get assigned, which
 834                  * could confuse clear_batch_ready for this stripe
 835                  */
 836                 sh->batch_head = head->batch_head;
 837
 838                 /*
 839                  * at this point, head's BATCH_READY could be cleared, but we
 840                  * can still add the stripe to batch list
 841                  */
 842                 list_add(&sh->batch_list, &head->batch_list);
 843                 spin_unlock(&head->batch_head->batch_lock);
 844         } else {
 845                 head->batch_head = head;
 846                 sh->batch_head = head->batch_head;
 847                 spin_lock(&head->batch_lock);
 848                 list_add_tail(&sh->batch_list, &head->batch_list);
 849                 spin_unlock(&head->batch_lock);
 850         }
 851
 852         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 853                 if (atomic_dec_return(&conf->preread_active_stripes)
 854                     < IO_THRESHOLD)
 855                         md_wakeup_thread(conf->mddev->thread);
 856
 857         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 858                 int seq = sh->bm_seq;
 859                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 860                     sh->batch_head->bm_seq > seq)
 861                         seq = sh->batch_head->bm_seq;
 862                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 863                 sh->batch_head->bm_seq = seq;
 864         }
 865
 866         atomic_inc(&sh->count);
 867 unlock_out:
 868         unlock_two_stripes(head, sh);
 869 out:
 870         raid5_release_stripe(head);
 871 }
 872
 873 /* Determine if 'data_offset' or 'new_data_offset' should be used
 874  * in this stripe_head.
 875  */
 876 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 877 {
 878         sector_t progress = conf->reshape_progress;
 879         /* Need a memory barrier to make sure we see the value
 880          * of conf->generation, or ->data_offset that was set before
 881          * reshape_progress was updated.
 882          */
 883         smp_rmb();
 884         if (progress == MaxSector)
 885                 return 0;
 886         if (sh->generation == conf->generation - 1)
 887                 return 0;
 888         /* We are in a reshape, and this is a new-generation stripe,
 889          * so use new_data_offset.
 890          */
 891         return 1;
 892 }
 893
 894 static void
 895 raid5_end_read_request(struct bio *bi);
 896 static void
 897 raid5_end_write_request(struct bio *bi);
 898
 899 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 900 {
 901         struct r5conf *conf = sh->raid_conf;
 902         int i, disks = sh->disks;
 903         struct stripe_head *head_sh = sh;
 904
 905         might_sleep();
 906
 907         if (r5l_write_stripe(conf->log, sh) == 0)
 908                 return;
 909         for (i = disks; i--; ) {
 910                 int op, op_flags = 0;
 911                 int replace_only = 0;
 912                 struct bio *bi, *rbi;
 913                 struct md_rdev *rdev, *rrdev = NULL;
 914
 915                 sh = head_sh;
 916                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 917                         op = REQ_OP_WRITE;
 918                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
 919                                 op_flags = WRITE_FUA;
 920                         if (test_bit(R5_Discard, &sh->dev[i].flags))
 921                                 op = REQ_OP_DISCARD;
 922                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 923                         op = REQ_OP_READ;
 924                 else if (test_and_clear_bit(R5_WantReplace,
 925                                             &sh->dev[i].flags)) {
 926                         op = REQ_OP_WRITE;
 927                         replace_only = 1;
 928                 } else
 929                         continue;
 930                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
 931                         op_flags |= REQ_SYNC;
 932
 933 again:
 934                 bi = &sh->dev[i].req;
 935                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
 936
 937                 rcu_read_lock();
 938                 rrdev = rcu_dereference(conf->disks[i].replacement);
 939                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
 940                 rdev = rcu_dereference(conf->disks[i].rdev);
 941                 if (!rdev) {
 942                         rdev = rrdev;
 943                         rrdev = NULL;
 944                 }
 945                 if (op_is_write(op)) {
 946                         if (replace_only)
 947                                 rdev = NULL;
 948                         if (rdev == rrdev)
 949                                 /* We raced and saw duplicates */
 950                                 rrdev = NULL;
 951                 } else {
 952                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
 953                                 rdev = rrdev;
 954                         rrdev = NULL;
 955                 }
 956
 957                 if (rdev && test_bit(Faulty, &rdev->flags))
 958                         rdev = NULL;
 959                 if (rdev)
 960                         atomic_inc(&rdev->nr_pending);
 961                 if (rrdev && test_bit(Faulty, &rrdev->flags))
 962                         rrdev = NULL;
 963                 if (rrdev)
 964                         atomic_inc(&rrdev->nr_pending);
 965                 rcu_read_unlock();
 966
 967                 /* We have already checked bad blocks for reads.  Now
 968                  * need to check for writes.  We never accept write errors
 969                  * on the replacement, so we don't to check rrdev.
 970                  */
 971                 while (op_is_write(op) && rdev &&
 972                        test_bit(WriteErrorSeen, &rdev->flags)) {
 973                         sector_t first_bad;
 974                         int bad_sectors;
 975                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
 976                                               &first_bad, &bad_sectors);
 977                         if (!bad)
 978                                 break;
 979
 980                         if (bad < 0) {
 981                                 set_bit(BlockedBadBlocks, &rdev->flags);
 982                                 if (!conf->mddev->external &&
 983                                     conf->mddev->flags) {
 984                                         /* It is very unlikely, but we might
 985                                          * still need to write out the
 986                                          * bad block log - better give it
 987                                          * a chance*/
 988                                         md_check_recovery(conf->mddev);
 989                                 }
 990                                 /*
 991                                  * Because md_wait_for_blocked_rdev
 992                                  * will dec nr_pending, we must
 993                                  * increment it first.
 994                                  */
 995                                 atomic_inc(&rdev->nr_pending);
 996                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
 997                         } else {
 998                                 /* Acknowledged bad block - skip the write */
 999                                 rdev_dec_pending(rdev, conf->mddev);
1000                                 rdev = NULL;
1001                         }
1002                 }
1003
1004                 if (rdev) {
1005                         if (s->syncing || s->expanding || s->expanded
1006                             || s->replacing)
1007                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1008
1009                         set_bit(STRIPE_IO_STARTED, &sh->state);
1010
1011                         bi->bi_bdev = rdev->bdev;
1012                         bio_set_op_attrs(bi, op, op_flags);
1013                         bi->bi_end_io = op_is_write(op)
1014                                 ? raid5_end_write_request
1015                                 : raid5_end_read_request;
1016                         bi->bi_private = sh;
1017
1018                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1019                                 __func__, (unsigned long long)sh->sector,
1020                                 bi->bi_opf, i);
1021                         atomic_inc(&sh->count);
1022                         if (sh != head_sh)
1023                                 atomic_inc(&head_sh->count);
1024                         if (use_new_offset(conf, sh))
1025                                 bi->bi_iter.bi_sector = (sh->sector
1026                                                  + rdev->new_data_offset);
1027                         else
1028                                 bi->bi_iter.bi_sector = (sh->sector
1029                                                  + rdev->data_offset);
1030                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1031                                 bi->bi_opf |= REQ_NOMERGE;
1032
1033                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1034                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1035                         sh->dev[i].vec.bv_page = sh->dev[i].page;
1036                         bi->bi_vcnt = 1;
1037                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1038                         bi->bi_io_vec[0].bv_offset = 0;
1039                         bi->bi_iter.bi_size = STRIPE_SIZE;
1040                         /*
1041                          * If this is discard request, set bi_vcnt 0. We don't
1042                          * want to confuse SCSI because SCSI will replace payload
1043                          */
1044                         if (op == REQ_OP_DISCARD)
1045                                 bi->bi_vcnt = 0;
1046                         if (rrdev)
1047                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1048
1049                         if (conf->mddev->gendisk)
1050                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1051                                                       bi, disk_devt(conf->mddev->gendisk),
1052                                                       sh->dev[i].sector);
1053                         generic_make_request(bi);
1054                 }
1055                 if (rrdev) {
1056                         if (s->syncing || s->expanding || s->expanded
1057                             || s->replacing)
1058                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1059
1060                         set_bit(STRIPE_IO_STARTED, &sh->state);
1061
1062                         rbi->bi_bdev = rrdev->bdev;
1063                         bio_set_op_attrs(rbi, op, op_flags);
1064                         BUG_ON(!op_is_write(op));
1065                         rbi->bi_end_io = raid5_end_write_request;
1066                         rbi->bi_private = sh;
1067
1068                         pr_debug("%s: for %llu schedule op %d on "
1069                                  "replacement disc %d\n",
1070                                 __func__, (unsigned long long)sh->sector,
1071                                 rbi->bi_opf, i);
1072                         atomic_inc(&sh->count);
1073                         if (sh != head_sh)
1074                                 atomic_inc(&head_sh->count);
1075                         if (use_new_offset(conf, sh))
1076                                 rbi->bi_iter.bi_sector = (sh->sector
1077                                                   + rrdev->new_data_offset);
1078                         else
1079                                 rbi->bi_iter.bi_sector = (sh->sector
1080                                                   + rrdev->data_offset);
1081                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1082                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1083                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1084                         rbi->bi_vcnt = 1;
1085                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1086                         rbi->bi_io_vec[0].bv_offset = 0;
1087                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1088                         /*
1089                          * If this is discard request, set bi_vcnt 0. We don't
1090                          * want to confuse SCSI because SCSI will replace payload
1091                          */
1092                         if (op == REQ_OP_DISCARD)
1093                                 rbi->bi_vcnt = 0;
1094                         if (conf->mddev->gendisk)
1095                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1096                                                       rbi, disk_devt(conf->mddev->gendisk),
1097                                                       sh->dev[i].sector);
1098                         generic_make_request(rbi);
1099                 }
1100                 if (!rdev && !rrdev) {
1101                         if (op_is_write(op))
1102                                 set_bit(STRIPE_DEGRADED, &sh->state);
1103                         pr_debug("skip op %d on disc %d for sector %llu\n",
1104                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1105                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1106                         set_bit(STRIPE_HANDLE, &sh->state);
1107                 }
1108
1109                 if (!head_sh->batch_head)
1110                         continue;
1111                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1112                                       batch_list);
1113                 if (sh != head_sh)
1114                         goto again;
1115         }
1116 }
1117
1118 static struct dma_async_tx_descriptor *
1119 async_copy_data(int frombio, struct bio *bio, struct page **page,
1120         sector_t sector, struct dma_async_tx_descriptor *tx,
1121         struct stripe_head *sh)
1122 {
1123         struct bio_vec bvl;
1124         struct bvec_iter iter;
1125         struct page *bio_page;
1126         int page_offset;
1127         struct async_submit_ctl submit;
1128         enum async_tx_flags flags = 0;
1129
1130         if (bio->bi_iter.bi_sector >= sector)
1131                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1132         else
1133                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1134
1135         if (frombio)
1136                 flags |= ASYNC_TX_FENCE;
1137         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1138
1139         bio_for_each_segment(bvl, bio, iter) {
1140                 int len = bvl.bv_len;
1141                 int clen;
1142                 int b_offset = 0;
1143
1144                 if (page_offset < 0) {
1145                         b_offset = -page_offset;
1146                         page_offset += b_offset;
1147                         len -= b_offset;
1148                 }
1149
1150                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1151                         clen = STRIPE_SIZE - page_offset;
1152                 else
1153                         clen = len;
1154
1155                 if (clen > 0) {
1156                         b_offset += bvl.bv_offset;
1157                         bio_page = bvl.bv_page;
1158                         if (frombio) {
1159                                 if (sh->raid_conf->skip_copy &&
1160                                     b_offset == 0 && page_offset == 0 &&
1161                                     clen == STRIPE_SIZE)
1162                                         *page = bio_page;
1163                                 else
1164                                         tx = async_memcpy(*page, bio_page, page_offset,
1165                                                   b_offset, clen, &submit);
1166                         } else
1167                                 tx = async_memcpy(bio_page, *page, b_offset,
1168                                                   page_offset, clen, &submit);
1169                 }
1170                 /* chain the operations */
1171                 submit.depend_tx = tx;
1172
1173                 if (clen < len) /* hit end of page */
1174                         break;
1175                 page_offset +=  len;
1176         }
1177
1178         return tx;
1179 }
1180
1181 static void ops_complete_biofill(void *stripe_head_ref)
1182 {
1183         struct stripe_head *sh = stripe_head_ref;
1184         struct bio_list return_bi = BIO_EMPTY_LIST;
1185         int i;
1186
1187         pr_debug("%s: stripe %llu\n", __func__,
1188                 (unsigned long long)sh->sector);
1189
1190         /* clear completed biofills */
1191         for (i = sh->disks; i--; ) {
1192                 struct r5dev *dev = &sh->dev[i];
1193
1194                 /* acknowledge completion of a biofill operation */
1195                 /* and check if we need to reply to a read request,
1196                  * new R5_Wantfill requests are held off until
1197                  * !STRIPE_BIOFILL_RUN
1198                  */
1199                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1200                         struct bio *rbi, *rbi2;
1201
1202                         BUG_ON(!dev->read);
1203                         rbi = dev->read;
1204                         dev->read = NULL;
1205                         while (rbi && rbi->bi_iter.bi_sector <
1206                                 dev->sector + STRIPE_SECTORS) {
1207                                 rbi2 = r5_next_bio(rbi, dev->sector);
1208                                 if (!raid5_dec_bi_active_stripes(rbi))
1209                                         bio_list_add(&return_bi, rbi);
1210                                 rbi = rbi2;
1211                         }
1212                 }
1213         }
1214         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1215
1216         return_io(&return_bi);
1217
1218         set_bit(STRIPE_HANDLE, &sh->state);
1219         raid5_release_stripe(sh);
1220 }
1221
1222 static void ops_run_biofill(struct stripe_head *sh)
1223 {
1224         struct dma_async_tx_descriptor *tx = NULL;
1225         struct async_submit_ctl submit;
1226         int i;
1227
1228         BUG_ON(sh->batch_head);
1229         pr_debug("%s: stripe %llu\n", __func__,
1230                 (unsigned long long)sh->sector);
1231
1232         for (i = sh->disks; i--; ) {
1233                 struct r5dev *dev = &sh->dev[i];
1234                 if (test_bit(R5_Wantfill, &dev->flags)) {
1235                         struct bio *rbi;
1236                         spin_lock_irq(&sh->stripe_lock);
1237                         dev->read = rbi = dev->toread;
1238                         dev->toread = NULL;
1239                         spin_unlock_irq(&sh->stripe_lock);
1240                         while (rbi && rbi->bi_iter.bi_sector <
1241                                 dev->sector + STRIPE_SECTORS) {
1242                                 tx = async_copy_data(0, rbi, &dev->page,
1243                                         dev->sector, tx, sh);
1244                                 rbi = r5_next_bio(rbi, dev->sector);
1245                         }
1246                 }
1247         }
1248
1249         atomic_inc(&sh->count);
1250         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1251         async_trigger_callback(&submit);
1252 }
1253
1254 static void mark_target_uptodate(struct stripe_head *sh, int target)
1255 {
1256         struct r5dev *tgt;
1257
1258         if (target < 0)
1259                 return;
1260
1261         tgt = &sh->dev[target];
1262         set_bit(R5_UPTODATE, &tgt->flags);
1263         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1264         clear_bit(R5_Wantcompute, &tgt->flags);
1265 }
1266
1267 static void ops_complete_compute(void *stripe_head_ref)
1268 {
1269         struct stripe_head *sh = stripe_head_ref;
1270
1271         pr_debug("%s: stripe %llu\n", __func__,
1272                 (unsigned long long)sh->sector);
1273
1274         /* mark the computed target(s) as uptodate */
1275         mark_target_uptodate(sh, sh->ops.target);
1276         mark_target_uptodate(sh, sh->ops.target2);
1277
1278         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1279         if (sh->check_state == check_state_compute_run)
1280                 sh->check_state = check_state_compute_result;
1281         set_bit(STRIPE_HANDLE, &sh->state);
1282         raid5_release_stripe(sh);
1283 }
1284
1285 /* return a pointer to the address conversion region of the scribble buffer */
1286 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1287                                  struct raid5_percpu *percpu, int i)
1288 {
1289         void *addr;
1290
1291         addr = flex_array_get(percpu->scribble, i);
1292         return addr + sizeof(struct page *) * (sh->disks + 2);
1293 }
1294
1295 /* return a pointer to the address conversion region of the scribble buffer */
1296 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1297 {
1298         void *addr;
1299
1300         addr = flex_array_get(percpu->scribble, i);
1301         return addr;
1302 }
1303
1304 static struct dma_async_tx_descriptor *
1305 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1306 {
1307         int disks = sh->disks;
1308         struct page **xor_srcs = to_addr_page(percpu, 0);
1309         int target = sh->ops.target;
1310         struct r5dev *tgt = &sh->dev[target];
1311         struct page *xor_dest = tgt->page;
1312         int count = 0;
1313         struct dma_async_tx_descriptor *tx;
1314         struct async_submit_ctl submit;
1315         int i;
1316
1317         BUG_ON(sh->batch_head);
1318
1319         pr_debug("%s: stripe %llu block: %d\n",
1320                 __func__, (unsigned long long)sh->sector, target);
1321         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1322
1323         for (i = disks; i--; )
1324                 if (i != target)
1325                         xor_srcs[count++] = sh->dev[i].page;
1326
1327         atomic_inc(&sh->count);
1328
1329         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1330                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1331         if (unlikely(count == 1))
1332                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1333         else
1334                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1335
1336         return tx;
1337 }
1338
1339 /* set_syndrome_sources - populate source buffers for gen_syndrome
1340  * @srcs - (struct page *) array of size sh->disks
1341  * @sh - stripe_head to parse
1342  *
1343  * Populates srcs in proper layout order for the stripe and returns the
1344  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1345  * destination buffer is recorded in srcs[count] and the Q destination
1346  * is recorded in srcs[count+1]].
1347  */
1348 static int set_syndrome_sources(struct page **srcs,
1349                                 struct stripe_head *sh,
1350                                 int srctype)
1351 {
1352         int disks = sh->disks;
1353         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1354         int d0_idx = raid6_d0(sh);
1355         int count;
1356         int i;
1357
1358         for (i = 0; i < disks; i++)
1359                 srcs[i] = NULL;
1360
1361         count = 0;
1362         i = d0_idx;
1363         do {
1364                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1365                 struct r5dev *dev = &sh->dev[i];
1366
1367                 if (i == sh->qd_idx || i == sh->pd_idx ||
1368                     (srctype == SYNDROME_SRC_ALL) ||
1369                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1370                      test_bit(R5_Wantdrain, &dev->flags)) ||
1371                     (srctype == SYNDROME_SRC_WRITTEN &&
1372                      dev->written))
1373                         srcs[slot] = sh->dev[i].page;
1374                 i = raid6_next_disk(i, disks);
1375         } while (i != d0_idx);
1376
1377         return syndrome_disks;
1378 }
1379
1380 static struct dma_async_tx_descriptor *
1381 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1382 {
1383         int disks = sh->disks;
1384         struct page **blocks = to_addr_page(percpu, 0);
1385         int target;
1386         int qd_idx = sh->qd_idx;
1387         struct dma_async_tx_descriptor *tx;
1388         struct async_submit_ctl submit;
1389         struct r5dev *tgt;
1390         struct page *dest;
1391         int i;
1392         int count;
1393
1394         BUG_ON(sh->batch_head);
1395         if (sh->ops.target < 0)
1396                 target = sh->ops.target2;
1397         else if (sh->ops.target2 < 0)
1398                 target = sh->ops.target;
1399         else
1400                 /* we should only have one valid target */
1401                 BUG();
1402         BUG_ON(target < 0);
1403         pr_debug("%s: stripe %llu block: %d\n",
1404                 __func__, (unsigned long long)sh->sector, target);
1405
1406         tgt = &sh->dev[target];
1407         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1408         dest = tgt->page;
1409
1410         atomic_inc(&sh->count);
1411
1412         if (target == qd_idx) {
1413                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1414                 blocks[count] = NULL; /* regenerating p is not necessary */
1415                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1416                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1417                                   ops_complete_compute, sh,
1418                                   to_addr_conv(sh, percpu, 0));
1419                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1420         } else {
1421                 /* Compute any data- or p-drive using XOR */
1422                 count = 0;
1423                 for (i = disks; i-- ; ) {
1424                         if (i == target || i == qd_idx)
1425                                 continue;
1426                         blocks[count++] = sh->dev[i].page;
1427                 }
1428
1429                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1430                                   NULL, ops_complete_compute, sh,
1431                                   to_addr_conv(sh, percpu, 0));
1432                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1433         }
1434
1435         return tx;
1436 }
1437
1438 static struct dma_async_tx_descriptor *
1439 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1440 {
1441         int i, count, disks = sh->disks;
1442         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1443         int d0_idx = raid6_d0(sh);
1444         int faila = -1, failb = -1;
1445         int target = sh->ops.target;
1446         int target2 = sh->ops.target2;
1447         struct r5dev *tgt = &sh->dev[target];
1448         struct r5dev *tgt2 = &sh->dev[target2];
1449         struct dma_async_tx_descriptor *tx;
1450         struct page **blocks = to_addr_page(percpu, 0);
1451         struct async_submit_ctl submit;
1452
1453         BUG_ON(sh->batch_head);
1454         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1455                  __func__, (unsigned long long)sh->sector, target, target2);
1456         BUG_ON(target < 0 || target2 < 0);
1457         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1458         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1459
1460         /* we need to open-code set_syndrome_sources to handle the
1461          * slot number conversion for 'faila' and 'failb'
1462          */
1463         for (i = 0; i < disks ; i++)
1464                 blocks[i] = NULL;
1465         count = 0;
1466         i = d0_idx;
1467         do {
1468                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1469
1470                 blocks[slot] = sh->dev[i].page;
1471
1472                 if (i == target)
1473                         faila = slot;
1474                 if (i == target2)
1475                         failb = slot;
1476                 i = raid6_next_disk(i, disks);
1477         } while (i != d0_idx);
1478
1479         BUG_ON(faila == failb);
1480         if (failb < faila)
1481                 swap(faila, failb);
1482         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1483                  __func__, (unsigned long long)sh->sector, faila, failb);
1484
1485         atomic_inc(&sh->count);
1486
1487         if (failb == syndrome_disks+1) {
1488                 /* Q disk is one of the missing disks */
1489                 if (faila == syndrome_disks) {
1490                         /* Missing P+Q, just recompute */
1491                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1492                                           ops_complete_compute, sh,
1493                                           to_addr_conv(sh, percpu, 0));
1494                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1495                                                   STRIPE_SIZE, &submit);
1496                 } else {
1497                         struct page *dest;
1498                         int data_target;
1499                         int qd_idx = sh->qd_idx;
1500
1501                         /* Missing D+Q: recompute D from P, then recompute Q */
1502                         if (target == qd_idx)
1503                                 data_target = target2;
1504                         else
1505                                 data_target = target;
1506
1507                         count = 0;
1508                         for (i = disks; i-- ; ) {
1509                                 if (i == data_target || i == qd_idx)
1510                                         continue;
1511                                 blocks[count++] = sh->dev[i].page;
1512                         }
1513                         dest = sh->dev[data_target].page;
1514                         init_async_submit(&submit,
1515                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1516                                           NULL, NULL, NULL,
1517                                           to_addr_conv(sh, percpu, 0));
1518                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1519                                        &submit);
1520
1521                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1522                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1523                                           ops_complete_compute, sh,
1524                                           to_addr_conv(sh, percpu, 0));
1525                         return async_gen_syndrome(blocks, 0, count+2,
1526                                                   STRIPE_SIZE, &submit);
1527                 }
1528         } else {
1529                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1530                                   ops_complete_compute, sh,
1531                                   to_addr_conv(sh, percpu, 0));
1532                 if (failb == syndrome_disks) {
1533                         /* We're missing D+P. */
1534                         return async_raid6_datap_recov(syndrome_disks+2,
1535                                                        STRIPE_SIZE, faila,
1536                                                        blocks, &submit);
1537                 } else {
1538                         /* We're missing D+D. */
1539                         return async_raid6_2data_recov(syndrome_disks+2,
1540                                                        STRIPE_SIZE, faila, failb,
1541                                                        blocks, &submit);
1542                 }
1543         }
1544 }
1545
1546 static void ops_complete_prexor(void *stripe_head_ref)
1547 {
1548         struct stripe_head *sh = stripe_head_ref;
1549
1550         pr_debug("%s: stripe %llu\n", __func__,
1551                 (unsigned long long)sh->sector);
1552 }
1553
1554 static struct dma_async_tx_descriptor *
1555 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1556                 struct dma_async_tx_descriptor *tx)
1557 {
1558         int disks = sh->disks;
1559         struct page **xor_srcs = to_addr_page(percpu, 0);
1560         int count = 0, pd_idx = sh->pd_idx, i;
1561         struct async_submit_ctl submit;
1562
1563         /* existing parity data subtracted */
1564         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1565
1566         BUG_ON(sh->batch_head);
1567         pr_debug("%s: stripe %llu\n", __func__,
1568                 (unsigned long long)sh->sector);
1569
1570         for (i = disks; i--; ) {
1571                 struct r5dev *dev = &sh->dev[i];
1572                 /* Only process blocks that are known to be uptodate */
1573                 if (test_bit(R5_Wantdrain, &dev->flags))
1574                         xor_srcs[count++] = dev->page;
1575         }
1576
1577         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1578                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1579         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1580
1581         return tx;
1582 }
1583
1584 static struct dma_async_tx_descriptor *
1585 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1586                 struct dma_async_tx_descriptor *tx)
1587 {
1588         struct page **blocks = to_addr_page(percpu, 0);
1589         int count;
1590         struct async_submit_ctl submit;
1591
1592         pr_debug("%s: stripe %llu\n", __func__,
1593                 (unsigned long long)sh->sector);
1594
1595         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1596
1597         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1598                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1599         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1600
1601         return tx;
1602 }
1603
1604 static struct dma_async_tx_descriptor *
1605 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1606 {
1607         int disks = sh->disks;
1608         int i;
1609         struct stripe_head *head_sh = sh;
1610
1611         pr_debug("%s: stripe %llu\n", __func__,
1612                 (unsigned long long)sh->sector);
1613
1614         for (i = disks; i--; ) {
1615                 struct r5dev *dev;
1616                 struct bio *chosen;
1617
1618                 sh = head_sh;
1619                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1620                         struct bio *wbi;
1621
1622 again:
1623                         dev = &sh->dev[i];
1624                         spin_lock_irq(&sh->stripe_lock);
1625                         chosen = dev->towrite;
1626                         dev->towrite = NULL;
1627                         sh->overwrite_disks = 0;
1628                         BUG_ON(dev->written);
1629                         wbi = dev->written = chosen;
1630                         spin_unlock_irq(&sh->stripe_lock);
1631                         WARN_ON(dev->page != dev->orig_page);
1632
1633                         while (wbi && wbi->bi_iter.bi_sector <
1634                                 dev->sector + STRIPE_SECTORS) {
1635                                 if (wbi->bi_opf & REQ_FUA)
1636                                         set_bit(R5_WantFUA, &dev->flags);
1637                                 if (wbi->bi_opf & REQ_SYNC)
1638                                         set_bit(R5_SyncIO, &dev->flags);
1639                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1640                                         set_bit(R5_Discard, &dev->flags);
1641                                 else {
1642                                         tx = async_copy_data(1, wbi, &dev->page,
1643                                                 dev->sector, tx, sh);
1644                                         if (dev->page != dev->orig_page) {
1645                                                 set_bit(R5_SkipCopy, &dev->flags);
1646                                                 clear_bit(R5_UPTODATE, &dev->flags);
1647                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1648                                         }
1649                                 }
1650                                 wbi = r5_next_bio(wbi, dev->sector);
1651                         }
1652
1653                         if (head_sh->batch_head) {
1654                                 sh = list_first_entry(&sh->batch_list,
1655                                                       struct stripe_head,
1656                                                       batch_list);
1657                                 if (sh == head_sh)
1658                                         continue;
1659                                 goto again;
1660                         }
1661                 }
1662         }
1663
1664         return tx;
1665 }
1666
1667 static void ops_complete_reconstruct(void *stripe_head_ref)
1668 {
1669         struct stripe_head *sh = stripe_head_ref;
1670         int disks = sh->disks;
1671         int pd_idx = sh->pd_idx;
1672         int qd_idx = sh->qd_idx;
1673         int i;
1674         bool fua = false, sync = false, discard = false;
1675
1676         pr_debug("%s: stripe %llu\n", __func__,
1677                 (unsigned long long)sh->sector);
1678
1679         for (i = disks; i--; ) {
1680                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1681                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1682                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1683         }
1684
1685         for (i = disks; i--; ) {
1686                 struct r5dev *dev = &sh->dev[i];
1687
1688                 if (dev->written || i == pd_idx || i == qd_idx) {
1689                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1690                                 set_bit(R5_UPTODATE, &dev->flags);
1691                                 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1692                                         set_bit(R5_Expanded, &dev->flags);
1693                         }
1694                         if (fua)
1695                                 set_bit(R5_WantFUA, &dev->flags);
1696                         if (sync)
1697                                 set_bit(R5_SyncIO, &dev->flags);
1698                 }
1699         }
1700
1701         if (sh->reconstruct_state == reconstruct_state_drain_run)
1702                 sh->reconstruct_state = reconstruct_state_drain_result;
1703         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1704                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1705         else {
1706                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1707                 sh->reconstruct_state = reconstruct_state_result;
1708         }
1709
1710         set_bit(STRIPE_HANDLE, &sh->state);
1711         raid5_release_stripe(sh);
1712 }
1713
1714 static void
1715 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1716                      struct dma_async_tx_descriptor *tx)
1717 {
1718         int disks = sh->disks;
1719         struct page **xor_srcs;
1720         struct async_submit_ctl submit;
1721         int count, pd_idx = sh->pd_idx, i;
1722         struct page *xor_dest;
1723         int prexor = 0;
1724         unsigned long flags;
1725         int j = 0;
1726         struct stripe_head *head_sh = sh;
1727         int last_stripe;
1728
1729         pr_debug("%s: stripe %llu\n", __func__,
1730                 (unsigned long long)sh->sector);
1731
1732         for (i = 0; i < sh->disks; i++) {
1733                 if (pd_idx == i)
1734                         continue;
1735                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1736                         break;
1737         }
1738         if (i >= sh->disks) {
1739                 atomic_inc(&sh->count);
1740                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1741                 ops_complete_reconstruct(sh);
1742                 return;
1743         }
1744 again:
1745         count = 0;
1746         xor_srcs = to_addr_page(percpu, j);
1747         /* check if prexor is active which means only process blocks
1748          * that are part of a read-modify-write (written)
1749          */
1750         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1751                 prexor = 1;
1752                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1753                 for (i = disks; i--; ) {
1754                         struct r5dev *dev = &sh->dev[i];
1755                         if (head_sh->dev[i].written)
1756                                 xor_srcs[count++] = dev->page;
1757                 }
1758         } else {
1759                 xor_dest = sh->dev[pd_idx].page;
1760                 for (i = disks; i--; ) {
1761                         struct r5dev *dev = &sh->dev[i];
1762                         if (i != pd_idx)
1763                                 xor_srcs[count++] = dev->page;
1764                 }
1765         }
1766
1767         /* 1/ if we prexor'd then the dest is reused as a source
1768          * 2/ if we did not prexor then we are redoing the parity
1769          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1770          * for the synchronous xor case
1771          */
1772         last_stripe = !head_sh->batch_head ||
1773                 list_first_entry(&sh->batch_list,
1774                                  struct stripe_head, batch_list) == head_sh;
1775         if (last_stripe) {
1776                 flags = ASYNC_TX_ACK |
1777                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1778
1779                 atomic_inc(&head_sh->count);
1780                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1781                                   to_addr_conv(sh, percpu, j));
1782         } else {
1783                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1784                 init_async_submit(&submit, flags, tx, NULL, NULL,
1785                                   to_addr_conv(sh, percpu, j));
1786         }
1787
1788         if (unlikely(count == 1))
1789                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1790         else
1791                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1792         if (!last_stripe) {
1793                 j++;
1794                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1795                                       batch_list);
1796                 goto again;
1797         }
1798 }
1799
1800 static void
1801 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1802                      struct dma_async_tx_descriptor *tx)
1803 {
1804         struct async_submit_ctl submit;
1805         struct page **blocks;
1806         int count, i, j = 0;
1807         struct stripe_head *head_sh = sh;
1808         int last_stripe;
1809         int synflags;
1810         unsigned long txflags;
1811
1812         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1813
1814         for (i = 0; i < sh->disks; i++) {
1815                 if (sh->pd_idx == i || sh->qd_idx == i)
1816                         continue;
1817                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1818                         break;
1819         }
1820         if (i >= sh->disks) {
1821                 atomic_inc(&sh->count);
1822                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1823                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1824                 ops_complete_reconstruct(sh);
1825                 return;
1826         }
1827
1828 again:
1829         blocks = to_addr_page(percpu, j);
1830
1831         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1832                 synflags = SYNDROME_SRC_WRITTEN;
1833                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1834         } else {
1835                 synflags = SYNDROME_SRC_ALL;
1836                 txflags = ASYNC_TX_ACK;
1837         }
1838
1839         count = set_syndrome_sources(blocks, sh, synflags);
1840         last_stripe = !head_sh->batch_head ||
1841                 list_first_entry(&sh->batch_list,
1842                                  struct stripe_head, batch_list) == head_sh;
1843
1844         if (last_stripe) {
1845                 atomic_inc(&head_sh->count);
1846                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1847                                   head_sh, to_addr_conv(sh, percpu, j));
1848         } else
1849                 init_async_submit(&submit, 0, tx, NULL, NULL,
1850                                   to_addr_conv(sh, percpu, j));
1851         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1852         if (!last_stripe) {
1853                 j++;
1854                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1855                                       batch_list);
1856                 goto again;
1857         }
1858 }
1859
1860 static void ops_complete_check(void *stripe_head_ref)
1861 {
1862         struct stripe_head *sh = stripe_head_ref;
1863
1864         pr_debug("%s: stripe %llu\n", __func__,
1865                 (unsigned long long)sh->sector);
1866
1867         sh->check_state = check_state_check_result;
1868         set_bit(STRIPE_HANDLE, &sh->state);
1869         raid5_release_stripe(sh);
1870 }
1871
1872 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1873 {
1874         int disks = sh->disks;
1875         int pd_idx = sh->pd_idx;
1876         int qd_idx = sh->qd_idx;
1877         struct page *xor_dest;
1878         struct page **xor_srcs = to_addr_page(percpu, 0);
1879         struct dma_async_tx_descriptor *tx;
1880         struct async_submit_ctl submit;
1881         int count;
1882         int i;
1883
1884         pr_debug("%s: stripe %llu\n", __func__,
1885                 (unsigned long long)sh->sector);
1886
1887         BUG_ON(sh->batch_head);
1888         count = 0;
1889         xor_dest = sh->dev[pd_idx].page;
1890         xor_srcs[count++] = xor_dest;
1891         for (i = disks; i--; ) {
1892                 if (i == pd_idx || i == qd_idx)
1893                         continue;
1894                 xor_srcs[count++] = sh->dev[i].page;
1895         }
1896
1897         init_async_submit(&submit, 0, NULL, NULL, NULL,
1898                           to_addr_conv(sh, percpu, 0));
1899         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1900                            &sh->ops.zero_sum_result, &submit);
1901
1902         atomic_inc(&sh->count);
1903         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1904         tx = async_trigger_callback(&submit);
1905 }
1906
1907 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1908 {
1909         struct page **srcs = to_addr_page(percpu, 0);
1910         struct async_submit_ctl submit;
1911         int count;
1912
1913         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1914                 (unsigned long long)sh->sector, checkp);
1915
1916         BUG_ON(sh->batch_head);
1917         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1918         if (!checkp)
1919                 srcs[count] = NULL;
1920
1921         atomic_inc(&sh->count);
1922         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1923                           sh, to_addr_conv(sh, percpu, 0));
1924         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1925                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1926 }
1927
1928 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1929 {
1930         int overlap_clear = 0, i, disks = sh->disks;
1931         struct dma_async_tx_descriptor *tx = NULL;
1932         struct r5conf *conf = sh->raid_conf;
1933         int level = conf->level;
1934         struct raid5_percpu *percpu;
1935         unsigned long cpu;
1936
1937         cpu = get_cpu();
1938         percpu = per_cpu_ptr(conf->percpu, cpu);
1939         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1940                 ops_run_biofill(sh);
1941                 overlap_clear++;
1942         }
1943
1944         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1945                 if (level < 6)
1946                         tx = ops_run_compute5(sh, percpu);
1947                 else {
1948                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1949                                 tx = ops_run_compute6_1(sh, percpu);
1950                         else
1951                                 tx = ops_run_compute6_2(sh, percpu);
1952                 }
1953                 /* terminate the chain if reconstruct is not set to be run */
1954                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1955                         async_tx_ack(tx);
1956         }
1957
1958         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1959                 if (level < 6)
1960                         tx = ops_run_prexor5(sh, percpu, tx);
1961                 else
1962                         tx = ops_run_prexor6(sh, percpu, tx);
1963         }
1964
1965         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1966                 tx = ops_run_biodrain(sh, tx);
1967                 overlap_clear++;
1968         }
1969
1970         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1971                 if (level < 6)
1972                         ops_run_reconstruct5(sh, percpu, tx);
1973                 else
1974                         ops_run_reconstruct6(sh, percpu, tx);
1975         }
1976
1977         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1978                 if (sh->check_state == check_state_run)
1979                         ops_run_check_p(sh, percpu);
1980                 else if (sh->check_state == check_state_run_q)
1981                         ops_run_check_pq(sh, percpu, 0);
1982                 else if (sh->check_state == check_state_run_pq)
1983                         ops_run_check_pq(sh, percpu, 1);
1984                 else
1985                         BUG();
1986         }
1987
1988         if (overlap_clear && !sh->batch_head)
1989                 for (i = disks; i--; ) {
1990                         struct r5dev *dev = &sh->dev[i];
1991                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1992                                 wake_up(&sh->raid_conf->wait_for_overlap);
1993                 }
1994         put_cpu();
1995 }
1996
1997 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
1998         int disks)
1999 {
2000         struct stripe_head *sh;
2001         int i;
2002
2003         sh = kmem_cache_zalloc(sc, gfp);
2004         if (sh) {
2005                 spin_lock_init(&sh->stripe_lock);
2006                 spin_lock_init(&sh->batch_lock);
2007                 INIT_LIST_HEAD(&sh->batch_list);
2008                 INIT_LIST_HEAD(&sh->lru);
2009                 atomic_set(&sh->count, 1);
2010                 for (i = 0; i < disks; i++) {
2011                         struct r5dev *dev = &sh->dev[i];
2012
2013                         bio_init(&dev->req);
2014                         dev->req.bi_io_vec = &dev->vec;
2015                         dev->req.bi_max_vecs = 1;
2016
2017                         bio_init(&dev->rreq);
2018                         dev->rreq.bi_io_vec = &dev->rvec;
2019                         dev->rreq.bi_max_vecs = 1;
2020                 }
2021         }
2022         return sh;
2023 }
2024 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2025 {
2026         struct stripe_head *sh;
2027
2028         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size);
2029         if (!sh)
2030                 return 0;
2031
2032         sh->raid_conf = conf;
2033
2034         if (grow_buffers(sh, gfp)) {
2035                 shrink_buffers(sh);
2036                 kmem_cache_free(conf->slab_cache, sh);
2037                 return 0;
2038         }
2039         sh->hash_lock_index =
2040                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2041         /* we just created an active stripe so... */
2042         atomic_inc(&conf->active_stripes);
2043
2044         raid5_release_stripe(sh);
2045         conf->max_nr_stripes++;
2046         return 1;
2047 }
2048
2049 static int grow_stripes(struct r5conf *conf, int num)
2050 {
2051         struct kmem_cache *sc;
2052         size_t namelen = sizeof(conf->cache_name[0]);
2053         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2054
2055         if (conf->mddev->gendisk)
2056                 snprintf(conf->cache_name[0], namelen,
2057                         "raid%d-%s", conf->level, mdname(conf->mddev));
2058         else
2059                 snprintf(conf->cache_name[0], namelen,
2060                         "raid%d-%p", conf->level, conf->mddev);
2061         snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2062
2063         conf->active_name = 0;
2064         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2065                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2066                                0, 0, NULL);
2067         if (!sc)
2068                 return 1;
2069         conf->slab_cache = sc;
2070         conf->pool_size = devs;
2071         while (num--)
2072                 if (!grow_one_stripe(conf, GFP_KERNEL))
2073                         return 1;
2074
2075         return 0;
2076 }
2077
2078 /**
2079  * scribble_len - return the required size of the scribble region
2080  * @num - total number of disks in the array
2081  *
2082  * The size must be enough to contain:
2083  * 1/ a struct page pointer for each device in the array +2
2084  * 2/ room to convert each entry in (1) to its corresponding dma
2085  *    (dma_map_page()) or page (page_address()) address.
2086  *
2087  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2088  * calculate over all devices (not just the data blocks), using zeros in place
2089  * of the P and Q blocks.
2090  */
2091 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2092 {
2093         struct flex_array *ret;
2094         size_t len;
2095
2096         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2097         ret = flex_array_alloc(len, cnt, flags);
2098         if (!ret)
2099                 return NULL;
2100         /* always prealloc all elements, so no locking is required */
2101         if (flex_array_prealloc(ret, 0, cnt, flags)) {
2102                 flex_array_free(ret);
2103                 return NULL;
2104         }
2105         return ret;
2106 }
2107
2108 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2109 {
2110         unsigned long cpu;
2111         int err = 0;
2112
2113         /*
2114          * Never shrink. And mddev_suspend() could deadlock if this is called
2115          * from raid5d. In that case, scribble_disks and scribble_sectors
2116          * should equal to new_disks and new_sectors
2117          */
2118         if (conf->scribble_disks >= new_disks &&
2119             conf->scribble_sectors >= new_sectors)
2120                 return 0;
2121         mddev_suspend(conf->mddev);
2122         get_online_cpus();
2123         for_each_present_cpu(cpu) {
2124                 struct raid5_percpu *percpu;
2125                 struct flex_array *scribble;
2126
2127                 percpu = per_cpu_ptr(conf->percpu, cpu);
2128                 scribble = scribble_alloc(new_disks,
2129                                           new_sectors / STRIPE_SECTORS,
2130                                           GFP_NOIO);
2131
2132                 if (scribble) {
2133                         flex_array_free(percpu->scribble);
2134                         percpu->scribble = scribble;
2135                 } else {
2136                         err = -ENOMEM;
2137                         break;
2138                 }
2139         }
2140         put_online_cpus();
2141         mddev_resume(conf->mddev);
2142         if (!err) {
2143                 conf->scribble_disks = new_disks;
2144                 conf->scribble_sectors = new_sectors;
2145         }
2146         return err;
2147 }
2148
2149 static int resize_stripes(struct r5conf *conf, int newsize)
2150 {
2151         /* Make all the stripes able to hold 'newsize' devices.
2152          * New slots in each stripe get 'page' set to a new page.
2153          *
2154          * This happens in stages:
2155          * 1/ create a new kmem_cache and allocate the required number of
2156          *    stripe_heads.
2157          * 2/ gather all the old stripe_heads and transfer the pages across
2158          *    to the new stripe_heads.  This will have the side effect of
2159          *    freezing the array as once all stripe_heads have been collected,
2160          *    no IO will be possible.  Old stripe heads are freed once their
2161          *    pages have been transferred over, and the old kmem_cache is
2162          *    freed when all stripes are done.
2163          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2164          *    we simple return a failre status - no need to clean anything up.
2165          * 4/ allocate new pages for the new slots in the new stripe_heads.
2166          *    If this fails, we don't bother trying the shrink the
2167          *    stripe_heads down again, we just leave them as they are.
2168          *    As each stripe_head is processed the new one is released into
2169          *    active service.
2170          *
2171          * Once step2 is started, we cannot afford to wait for a write,
2172          * so we use GFP_NOIO allocations.
2173          */
2174         struct stripe_head *osh, *nsh;
2175         LIST_HEAD(newstripes);
2176         struct disk_info *ndisks;
2177         int err;
2178         struct kmem_cache *sc;
2179         int i;
2180         int hash, cnt;
2181
2182         if (newsize <= conf->pool_size)
2183                 return 0; /* never bother to shrink */
2184
2185         err = md_allow_write(conf->mddev);
2186         if (err)
2187                 return err;
2188
2189         /* Step 1 */
2190         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2191                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2192                                0, 0, NULL);
2193         if (!sc)
2194                 return -ENOMEM;
2195
2196         /* Need to ensure auto-resizing doesn't interfere */
2197         mutex_lock(&conf->cache_size_mutex);
2198
2199         for (i = conf->max_nr_stripes; i; i--) {
2200                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize);
2201                 if (!nsh)
2202                         break;
2203
2204                 nsh->raid_conf = conf;
2205                 list_add(&nsh->lru, &newstripes);
2206         }
2207         if (i) {
2208                 /* didn't get enough, give up */
2209                 while (!list_empty(&newstripes)) {
2210                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2211                         list_del(&nsh->lru);
2212                         kmem_cache_free(sc, nsh);
2213                 }
2214                 kmem_cache_destroy(sc);
2215                 mutex_unlock(&conf->cache_size_mutex);
2216                 return -ENOMEM;
2217         }
2218         /* Step 2 - Must use GFP_NOIO now.
2219          * OK, we have enough stripes, start collecting inactive
2220          * stripes and copying them over
2221          */
2222         hash = 0;
2223         cnt = 0;
2224         list_for_each_entry(nsh, &newstripes, lru) {
2225                 lock_device_hash_lock(conf, hash);
2226                 wait_event_cmd(conf->wait_for_stripe,
2227                                     !list_empty(conf->inactive_list + hash),
2228                                     unlock_device_hash_lock(conf, hash),
2229                                     lock_device_hash_lock(conf, hash));
2230                 osh = get_free_stripe(conf, hash);
2231                 unlock_device_hash_lock(conf, hash);
2232
2233                 for(i=0; i<conf->pool_size; i++) {
2234                         nsh->dev[i].page = osh->dev[i].page;
2235                         nsh->dev[i].orig_page = osh->dev[i].page;
2236                 }
2237                 nsh->hash_lock_index = hash;
2238                 kmem_cache_free(conf->slab_cache, osh);
2239                 cnt++;
2240                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2241                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2242                         hash++;
2243                         cnt = 0;
2244                 }
2245         }
2246         kmem_cache_destroy(conf->slab_cache);
2247
2248         /* Step 3.
2249          * At this point, we are holding all the stripes so the array
2250          * is completely stalled, so now is a good time to resize
2251          * conf->disks and the scribble region
2252          */
2253         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2254         if (ndisks) {
2255                 for (i=0; i<conf->raid_disks; i++)
2256                         ndisks[i] = conf->disks[i];
2257                 kfree(conf->disks);
2258                 conf->disks = ndisks;
2259         } else
2260                 err = -ENOMEM;
2261
2262         conf->slab_cache = sc;
2263         conf->active_name = 1-conf->active_name;
2264
2265         /* Step 4, return new stripes to service */
2266         while(!list_empty(&newstripes)) {
2267                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2268                 list_del_init(&nsh->lru);
2269
2270                 for (i=conf->raid_disks; i < newsize; i++)
2271                         if (nsh->dev[i].page == NULL) {
2272                                 struct page *p = alloc_page(GFP_NOIO);
2273                                 nsh->dev[i].page = p;
2274                                 nsh->dev[i].orig_page = p;
2275                                 if (!p)
2276                                         err = -ENOMEM;
2277                         }
2278                 raid5_release_stripe(nsh);
2279         }
2280         /* critical section pass, GFP_NOIO no longer needed */
2281
2282         if (!err)
2283                 conf->pool_size = newsize;
2284         mutex_unlock(&conf->cache_size_mutex);
2285
2286         return err;
2287 }
2288
2289 static int drop_one_stripe(struct r5conf *conf)
2290 {
2291         struct stripe_head *sh;
2292         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2293
2294         spin_lock_irq(conf->hash_locks + hash);
2295         sh = get_free_stripe(conf, hash);
2296         spin_unlock_irq(conf->hash_locks + hash);
2297         if (!sh)
2298                 return 0;
2299         BUG_ON(atomic_read(&sh->count));
2300         shrink_buffers(sh);
2301         kmem_cache_free(conf->slab_cache, sh);
2302         atomic_dec(&conf->active_stripes);
2303         conf->max_nr_stripes--;
2304         return 1;
2305 }
2306
2307 static void shrink_stripes(struct r5conf *conf)
2308 {
2309         while (conf->max_nr_stripes &&
2310                drop_one_stripe(conf))
2311                 ;
2312
2313         kmem_cache_destroy(conf->slab_cache);
2314         conf->slab_cache = NULL;
2315 }
2316
2317 static void raid5_end_read_request(struct bio * bi)
2318 {
2319         struct stripe_head *sh = bi->bi_private;
2320         struct r5conf *conf = sh->raid_conf;
2321         int disks = sh->disks, i;
2322         char b[BDEVNAME_SIZE];
2323         struct md_rdev *rdev = NULL;
2324         sector_t s;
2325
2326         for (i=0 ; i<disks; i++)
2327                 if (bi == &sh->dev[i].req)
2328                         break;
2329
2330         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2331                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2332                 bi->bi_error);
2333         if (i == disks) {
2334                 bio_reset(bi);
2335                 BUG();
2336                 return;
2337         }
2338         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2339                 /* If replacement finished while this request was outstanding,
2340                  * 'replacement' might be NULL already.
2341                  * In that case it moved down to 'rdev'.
2342                  * rdev is not removed until all requests are finished.
2343                  */
2344                 rdev = conf->disks[i].replacement;
2345         if (!rdev)
2346                 rdev = conf->disks[i].rdev;
2347
2348         if (use_new_offset(conf, sh))
2349                 s = sh->sector + rdev->new_data_offset;
2350         else
2351                 s = sh->sector + rdev->data_offset;
2352         if (!bi->bi_error) {
2353                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2354                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2355                         /* Note that this cannot happen on a
2356                          * replacement device.  We just fail those on
2357                          * any error
2358                          */
2359                         printk_ratelimited(
2360                                 KERN_INFO
2361                                 "md/raid:%s: read error corrected"
2362                                 " (%lu sectors at %llu on %s)\n",
2363                                 mdname(conf->mddev), STRIPE_SECTORS,
2364                                 (unsigned long long)s,
2365                                 bdevname(rdev->bdev, b));
2366                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2367                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2368                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2369                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2370                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2371
2372                 if (atomic_read(&rdev->read_errors))
2373                         atomic_set(&rdev->read_errors, 0);
2374         } else {
2375                 const char *bdn = bdevname(rdev->bdev, b);
2376                 int retry = 0;
2377                 int set_bad = 0;
2378
2379                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2380                 atomic_inc(&rdev->read_errors);
2381                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2382                         printk_ratelimited(
2383                                 KERN_WARNING
2384                                 "md/raid:%s: read error on replacement device "
2385                                 "(sector %llu on %s).\n",
2386                                 mdname(conf->mddev),
2387                                 (unsigned long long)s,
2388                                 bdn);
2389                 else if (conf->mddev->degraded >= conf->max_degraded) {
2390                         set_bad = 1;
2391                         printk_ratelimited(
2392                                 KERN_WARNING
2393                                 "md/raid:%s: read error not correctable "
2394                                 "(sector %llu on %s).\n",
2395                                 mdname(conf->mddev),
2396                                 (unsigned long long)s,
2397                                 bdn);
2398                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2399                         /* Oh, no!!! */
2400                         set_bad = 1;
2401                         printk_ratelimited(
2402                                 KERN_WARNING
2403                                 "md/raid:%s: read error NOT corrected!! "
2404                                 "(sector %llu on %s).\n",
2405                                 mdname(conf->mddev),
2406                                 (unsigned long long)s,
2407                                 bdn);
2408                 } else if (atomic_read(&rdev->read_errors)
2409                          > conf->max_nr_stripes)
2410                         printk(KERN_WARNING
2411                                "md/raid:%s: Too many read errors, failing device %s.\n",
2412                                mdname(conf->mddev), bdn);
2413                 else
2414                         retry = 1;
2415                 if (set_bad && test_bit(In_sync, &rdev->flags)
2416                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2417                         retry = 1;
2418                 if (retry)
2419                         if (sh->qd_idx >= 0 && sh->pd_idx == i)
2420                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2421                         else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2422                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2423                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2424                         } else
2425                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2426                 else {
2427                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2428                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2429                         if (!(set_bad
2430                               && test_bit(In_sync, &rdev->flags)
2431                               && rdev_set_badblocks(
2432                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2433                                 md_error(conf->mddev, rdev);
2434                 }
2435         }
2436         rdev_dec_pending(rdev, conf->mddev);
2437         bio_reset(bi);
2438         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2439         set_bit(STRIPE_HANDLE, &sh->state);
2440         raid5_release_stripe(sh);
2441 }
2442
2443 static void raid5_end_write_request(struct bio *bi)
2444 {
2445         struct stripe_head *sh = bi->bi_private;
2446         struct r5conf *conf = sh->raid_conf;
2447         int disks = sh->disks, i;
2448         struct md_rdev *uninitialized_var(rdev);
2449         sector_t first_bad;
2450         int bad_sectors;
2451         int replacement = 0;
2452
2453         for (i = 0 ; i < disks; i++) {
2454                 if (bi == &sh->dev[i].req) {
2455                         rdev = conf->disks[i].rdev;
2456                         break;
2457                 }
2458                 if (bi == &sh->dev[i].rreq) {
2459                         rdev = conf->disks[i].replacement;
2460                         if (rdev)
2461                                 replacement = 1;
2462                         else
2463                                 /* rdev was removed and 'replacement'
2464                                  * replaced it.  rdev is not removed
2465                                  * until all requests are finished.
2466                                  */
2467                                 rdev = conf->disks[i].rdev;
2468                         break;
2469                 }
2470         }
2471         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2472                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2473                 bi->bi_error);
2474         if (i == disks) {
2475                 bio_reset(bi);
2476                 BUG();
2477                 return;
2478         }
2479
2480         if (replacement) {
2481                 if (bi->bi_error)
2482                         md_error(conf->mddev, rdev);
2483                 else if (is_badblock(rdev, sh->sector,
2484                                      STRIPE_SECTORS,
2485                                      &first_bad, &bad_sectors))
2486                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2487         } else {
2488                 if (bi->bi_error) {
2489                         set_bit(STRIPE_DEGRADED, &sh->state);
2490                         set_bit(WriteErrorSeen, &rdev->flags);
2491                         set_bit(R5_WriteError, &sh->dev[i].flags);
2492                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2493                                 set_bit(MD_RECOVERY_NEEDED,
2494                                         &rdev->mddev->recovery);
2495                 } else if (is_badblock(rdev, sh->sector,
2496                                        STRIPE_SECTORS,
2497                                        &first_bad, &bad_sectors)) {
2498                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2499                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2500                                 /* That was a successful write so make
2501                                  * sure it looks like we already did
2502                                  * a re-write.
2503                                  */
2504                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2505                 }
2506         }
2507         rdev_dec_pending(rdev, conf->mddev);
2508
2509         if (sh->batch_head && bi->bi_error && !replacement)
2510                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2511
2512         bio_reset(bi);
2513         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2514                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2515         set_bit(STRIPE_HANDLE, &sh->state);
2516         raid5_release_stripe(sh);
2517
2518         if (sh->batch_head && sh != sh->batch_head)
2519                 raid5_release_stripe(sh->batch_head);
2520 }
2521
2522 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2523 {
2524         struct r5dev *dev = &sh->dev[i];
2525
2526         dev->flags = 0;
2527         dev->sector = raid5_compute_blocknr(sh, i, previous);
2528 }
2529
2530 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2531 {
2532         char b[BDEVNAME_SIZE];
2533         struct r5conf *conf = mddev->private;
2534         unsigned long flags;
2535         pr_debug("raid456: error called\n");
2536
2537         spin_lock_irqsave(&conf->device_lock, flags);
2538         clear_bit(In_sync, &rdev->flags);
2539         mddev->degraded = calc_degraded(conf);
2540         spin_unlock_irqrestore(&conf->device_lock, flags);
2541         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2542
2543         set_bit(Blocked, &rdev->flags);
2544         set_bit(Faulty, &rdev->flags);
2545         set_mask_bits(&mddev->flags, 0,
2546                       BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
2547         printk(KERN_ALERT
2548                "md/raid:%s: Disk failure on %s, disabling device.\n"
2549                "md/raid:%s: Operation continuing on %d devices.\n",
2550                mdname(mddev),
2551                bdevname(rdev->bdev, b),
2552                mdname(mddev),
2553                conf->raid_disks - mddev->degraded);
2554 }
2555
2556 /*
2557  * Input: a 'big' sector number,
2558  * Output: index of the data and parity disk, and the sector # in them.
2559  */
2560 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2561                               int previous, int *dd_idx,
2562                               struct stripe_head *sh)
2563 {
2564         sector_t stripe, stripe2;
2565         sector_t chunk_number;
2566         unsigned int chunk_offset;
2567         int pd_idx, qd_idx;
2568         int ddf_layout = 0;
2569         sector_t new_sector;
2570         int algorithm = previous ? conf->prev_algo
2571                                  : conf->algorithm;
2572         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2573                                          : conf->chunk_sectors;
2574         int raid_disks = previous ? conf->previous_raid_disks
2575                                   : conf->raid_disks;
2576         int data_disks = raid_disks - conf->max_degraded;
2577
2578         /* First compute the information on this sector */
2579
2580         /*
2581          * Compute the chunk number and the sector offset inside the chunk
2582          */
2583         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2584         chunk_number = r_sector;
2585
2586         /*
2587          * Compute the stripe number
2588          */
2589         stripe = chunk_number;
2590         *dd_idx = sector_div(stripe, data_disks);
2591         stripe2 = stripe;
2592         /*
2593          * Select the parity disk based on the user selected algorithm.
2594          */
2595         pd_idx = qd_idx = -1;
2596         switch(conf->level) {
2597         case 4:
2598                 pd_idx = data_disks;
2599                 break;
2600         case 5:
2601                 switch (algorithm) {
2602                 case ALGORITHM_LEFT_ASYMMETRIC:
2603                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2604                         if (*dd_idx >= pd_idx)
2605                                 (*dd_idx)++;
2606                         break;
2607                 case ALGORITHM_RIGHT_ASYMMETRIC:
2608                         pd_idx = sector_div(stripe2, raid_disks);
2609                         if (*dd_idx >= pd_idx)
2610                                 (*dd_idx)++;
2611                         break;
2612                 case ALGORITHM_LEFT_SYMMETRIC:
2613                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2614                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2615                         break;
2616                 case ALGORITHM_RIGHT_SYMMETRIC:
2617                         pd_idx = sector_div(stripe2, raid_disks);
2618                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2619                         break;
2620                 case ALGORITHM_PARITY_0:
2621                         pd_idx = 0;
2622                         (*dd_idx)++;
2623                         break;
2624                 case ALGORITHM_PARITY_N:
2625                         pd_idx = data_disks;
2626                         break;
2627                 default:
2628                         BUG();
2629                 }
2630                 break;
2631         case 6:
2632
2633                 switch (algorithm) {
2634                 case ALGORITHM_LEFT_ASYMMETRIC:
2635                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2636                         qd_idx = pd_idx + 1;
2637                         if (pd_idx == raid_disks-1) {
2638                                 (*dd_idx)++;    /* Q D D D P */
2639                                 qd_idx = 0;
2640                         } else if (*dd_idx >= pd_idx)
2641                                 (*dd_idx) += 2; /* D D P Q D */
2642                         break;
2643                 case ALGORITHM_RIGHT_ASYMMETRIC:
2644                         pd_idx = sector_div(stripe2, raid_disks);
2645                         qd_idx = pd_idx + 1;
2646                         if (pd_idx == raid_disks-1) {
2647                                 (*dd_idx)++;    /* Q D D D P */
2648                                 qd_idx = 0;
2649                         } else if (*dd_idx >= pd_idx)
2650                                 (*dd_idx) += 2; /* D D P Q D */
2651                         break;
2652                 case ALGORITHM_LEFT_SYMMETRIC:
2653                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2654                         qd_idx = (pd_idx + 1) % raid_disks;
2655                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2656                         break;
2657                 case ALGORITHM_RIGHT_SYMMETRIC:
2658                         pd_idx = sector_div(stripe2, raid_disks);
2659                         qd_idx = (pd_idx + 1) % raid_disks;
2660                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2661                         break;
2662
2663                 case ALGORITHM_PARITY_0:
2664                         pd_idx = 0;
2665                         qd_idx = 1;
2666                         (*dd_idx) += 2;
2667                         break;
2668                 case ALGORITHM_PARITY_N:
2669                         pd_idx = data_disks;
2670                         qd_idx = data_disks + 1;
2671                         break;
2672
2673                 case ALGORITHM_ROTATING_ZERO_RESTART:
2674                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2675                          * of blocks for computing Q is different.
2676                          */
2677                         pd_idx = sector_div(stripe2, raid_disks);
2678                         qd_idx = pd_idx + 1;
2679                         if (pd_idx == raid_disks-1) {
2680                                 (*dd_idx)++;    /* Q D D D P */
2681                                 qd_idx = 0;
2682                         } else if (*dd_idx >= pd_idx)
2683                                 (*dd_idx) += 2; /* D D P Q D */
2684                         ddf_layout = 1;
2685                         break;
2686
2687                 case ALGORITHM_ROTATING_N_RESTART:
2688                         /* Same a left_asymmetric, by first stripe is
2689                          * D D D P Q  rather than
2690                          * Q D D D P
2691                          */
2692                         stripe2 += 1;
2693                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2694                         qd_idx = pd_idx + 1;
2695                         if (pd_idx == raid_disks-1) {
2696                                 (*dd_idx)++;    /* Q D D D P */
2697                                 qd_idx = 0;
2698                         } else if (*dd_idx >= pd_idx)
2699                                 (*dd_idx) += 2; /* D D P Q D */
2700                         ddf_layout = 1;
2701                         break;
2702
2703                 case ALGORITHM_ROTATING_N_CONTINUE:
2704                         /* Same as left_symmetric but Q is before P */
2705                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2706                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2707                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2708                         ddf_layout = 1;
2709                         break;
2710
2711                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2712                         /* RAID5 left_asymmetric, with Q on last device */
2713                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2714                         if (*dd_idx >= pd_idx)
2715                                 (*dd_idx)++;
2716                         qd_idx = raid_disks - 1;
2717                         break;
2718
2719                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2720                         pd_idx = sector_div(stripe2, raid_disks-1);
2721                         if (*dd_idx >= pd_idx)
2722                                 (*dd_idx)++;
2723                         qd_idx = raid_disks - 1;
2724                         break;
2725
2726                 case ALGORITHM_LEFT_SYMMETRIC_6:
2727                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2728                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2729                         qd_idx = raid_disks - 1;
2730                         break;
2731
2732                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2733                         pd_idx = sector_div(stripe2, raid_disks-1);
2734                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2735                         qd_idx = raid_disks - 1;
2736                         break;
2737
2738                 case ALGORITHM_PARITY_0_6:
2739                         pd_idx = 0;
2740                         (*dd_idx)++;
2741                         qd_idx = raid_disks - 1;
2742                         break;
2743
2744                 default:
2745                         BUG();
2746                 }
2747                 break;
2748         }
2749
2750         if (sh) {
2751                 sh->pd_idx = pd_idx;
2752                 sh->qd_idx = qd_idx;
2753                 sh->ddf_layout = ddf_layout;
2754         }
2755         /*
2756          * Finally, compute the new sector number
2757          */
2758         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2759         return new_sector;
2760 }
2761
2762 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2763 {
2764         struct r5conf *conf = sh->raid_conf;
2765         int raid_disks = sh->disks;
2766         int data_disks = raid_disks - conf->max_degraded;
2767         sector_t new_sector = sh->sector, check;
2768         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2769                                          : conf->chunk_sectors;
2770         int algorithm = previous ? conf->prev_algo
2771                                  : conf->algorithm;
2772         sector_t stripe;
2773         int chunk_offset;
2774         sector_t chunk_number;
2775         int dummy1, dd_idx = i;
2776         sector_t r_sector;
2777         struct stripe_head sh2;
2778
2779         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2780         stripe = new_sector;
2781
2782         if (i == sh->pd_idx)
2783                 return 0;
2784         switch(conf->level) {
2785         case 4: break;
2786         case 5:
2787                 switch (algorithm) {
2788                 case ALGORITHM_LEFT_ASYMMETRIC:
2789                 case ALGORITHM_RIGHT_ASYMMETRIC:
2790                         if (i > sh->pd_idx)
2791                                 i--;
2792                         break;
2793                 case ALGORITHM_LEFT_SYMMETRIC:
2794                 case ALGORITHM_RIGHT_SYMMETRIC:
2795                         if (i < sh->pd_idx)
2796                                 i += raid_disks;
2797                         i -= (sh->pd_idx + 1);
2798                         break;
2799                 case ALGORITHM_PARITY_0:
2800                         i -= 1;
2801                         break;
2802                 case ALGORITHM_PARITY_N:
2803                         break;
2804                 default:
2805                         BUG();
2806                 }
2807                 break;
2808         case 6:
2809                 if (i == sh->qd_idx)
2810                         return 0; /* It is the Q disk */
2811                 switch (algorithm) {
2812                 case ALGORITHM_LEFT_ASYMMETRIC:
2813                 case ALGORITHM_RIGHT_ASYMMETRIC:
2814                 case ALGORITHM_ROTATING_ZERO_RESTART:
2815                 case ALGORITHM_ROTATING_N_RESTART:
2816                         if (sh->pd_idx == raid_disks-1)
2817                                 i--;    /* Q D D D P */
2818                         else if (i > sh->pd_idx)
2819                                 i -= 2; /* D D P Q D */
2820                         break;
2821                 case ALGORITHM_LEFT_SYMMETRIC:
2822                 case ALGORITHM_RIGHT_SYMMETRIC:
2823                         if (sh->pd_idx == raid_disks-1)
2824                                 i--; /* Q D D D P */
2825                         else {
2826                                 /* D D P Q D */
2827                                 if (i < sh->pd_idx)
2828                                         i += raid_disks;
2829                                 i -= (sh->pd_idx + 2);
2830                         }
2831                         break;
2832                 case ALGORITHM_PARITY_0:
2833                         i -= 2;
2834                         break;
2835                 case ALGORITHM_PARITY_N:
2836                         break;
2837                 case ALGORITHM_ROTATING_N_CONTINUE:
2838                         /* Like left_symmetric, but P is before Q */
2839                         if (sh->pd_idx == 0)
2840                                 i--;    /* P D D D Q */
2841                         else {
2842                                 /* D D Q P D */
2843                                 if (i < sh->pd_idx)
2844                                         i += raid_disks;
2845                                 i -= (sh->pd_idx + 1);
2846                         }
2847                         break;
2848                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2849                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2850                         if (i > sh->pd_idx)
2851                                 i--;
2852                         break;
2853                 case ALGORITHM_LEFT_SYMMETRIC_6:
2854                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2855                         if (i < sh->pd_idx)
2856                                 i += data_disks + 1;
2857                         i -= (sh->pd_idx + 1);
2858                         break;
2859                 case ALGORITHM_PARITY_0_6:
2860                         i -= 1;
2861                         break;
2862                 default:
2863                         BUG();
2864                 }
2865                 break;
2866         }
2867
2868         chunk_number = stripe * data_disks + i;
2869         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2870
2871         check = raid5_compute_sector(conf, r_sector,
2872                                      previous, &dummy1, &sh2);
2873         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2874                 || sh2.qd_idx != sh->qd_idx) {
2875                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2876                        mdname(conf->mddev));
2877                 return 0;
2878         }
2879         return r_sector;
2880 }
2881
2882 static void
2883 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2884                          int rcw, int expand)
2885 {
2886         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2887         struct r5conf *conf = sh->raid_conf;
2888         int level = conf->level;
2889
2890         if (rcw) {
2891
2892                 for (i = disks; i--; ) {
2893                         struct r5dev *dev = &sh->dev[i];
2894
2895                         if (dev->towrite) {
2896                                 set_bit(R5_LOCKED, &dev->flags);
2897                                 set_bit(R5_Wantdrain, &dev->flags);
2898                                 if (!expand)
2899                                         clear_bit(R5_UPTODATE, &dev->flags);
2900                                 s->locked++;
2901                         }
2902                 }
2903                 /* if we are not expanding this is a proper write request, and
2904                  * there will be bios with new data to be drained into the
2905                  * stripe cache
2906                  */
2907                 if (!expand) {
2908                         if (!s->locked)
2909                                 /* False alarm, nothing to do */
2910                                 return;
2911                         sh->reconstruct_state = reconstruct_state_drain_run;
2912                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2913                 } else
2914                         sh->reconstruct_state = reconstruct_state_run;
2915
2916                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2917
2918                 if (s->locked + conf->max_degraded == disks)
2919                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2920                                 atomic_inc(&conf->pending_full_writes);
2921         } else {
2922                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2923                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2924                 BUG_ON(level == 6 &&
2925                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2926                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2927
2928                 for (i = disks; i--; ) {
2929                         struct r5dev *dev = &sh->dev[i];
2930                         if (i == pd_idx || i == qd_idx)
2931                                 continue;
2932
2933                         if (dev->towrite &&
2934                             (test_bit(R5_UPTODATE, &dev->flags) ||
2935                              test_bit(R5_Wantcompute, &dev->flags))) {
2936                                 set_bit(R5_Wantdrain, &dev->flags);
2937                                 set_bit(R5_LOCKED, &dev->flags);
2938                                 clear_bit(R5_UPTODATE, &dev->flags);
2939                                 s->locked++;
2940                         }
2941                 }
2942                 if (!s->locked)
2943                         /* False alarm - nothing to do */
2944                         return;
2945                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2946                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2947                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2948                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2949         }
2950
2951         /* keep the parity disk(s) locked while asynchronous operations
2952          * are in flight
2953          */
2954         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2955         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2956         s->locked++;
2957
2958         if (level == 6) {
2959                 int qd_idx = sh->qd_idx;
2960                 struct r5dev *dev = &sh->dev[qd_idx];
2961
2962                 set_bit(R5_LOCKED, &dev->flags);
2963                 clear_bit(R5_UPTODATE, &dev->flags);
2964                 s->locked++;
2965         }
2966
2967         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2968                 __func__, (unsigned long long)sh->sector,
2969                 s->locked, s->ops_request);
2970 }
2971
2972 /*
2973  * Each stripe/dev can have one or more bion attached.
2974  * toread/towrite point to the first in a chain.
2975  * The bi_next chain must be in order.
2976  */
2977 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2978                           int forwrite, int previous)
2979 {
2980         struct bio **bip;
2981         struct r5conf *conf = sh->raid_conf;
2982         int firstwrite=0;
2983
2984         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2985                 (unsigned long long)bi->bi_iter.bi_sector,
2986                 (unsigned long long)sh->sector);
2987
2988         /*
2989          * If several bio share a stripe. The bio bi_phys_segments acts as a
2990          * reference count to avoid race. The reference count should already be
2991          * increased before this function is called (for example, in
2992          * raid5_make_request()), so other bio sharing this stripe will not free the
2993          * stripe. If a stripe is owned by one stripe, the stripe lock will
2994          * protect it.
2995          */
2996         spin_lock_irq(&sh->stripe_lock);
2997         /* Don't allow new IO added to stripes in batch list */
2998         if (sh->batch_head)
2999                 goto overlap;
3000         if (forwrite) {
3001                 bip = &sh->dev[dd_idx].towrite;
3002                 if (*bip == NULL)
3003                         firstwrite = 1;
3004         } else
3005                 bip = &sh->dev[dd_idx].toread;
3006         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3007                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3008                         goto overlap;
3009                 bip = & (*bip)->bi_next;
3010         }
3011         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3012                 goto overlap;
3013
3014         if (!forwrite || previous)
3015                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3016
3017         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3018         if (*bip)
3019                 bi->bi_next = *bip;
3020         *bip = bi;
3021         raid5_inc_bi_active_stripes(bi);
3022
3023         if (forwrite) {
3024                 /* check if page is covered */
3025                 sector_t sector = sh->dev[dd_idx].sector;
3026                 for (bi=sh->dev[dd_idx].towrite;
3027                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3028                              bi && bi->bi_iter.bi_sector <= sector;
3029                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3030                         if (bio_end_sector(bi) >= sector)
3031                                 sector = bio_end_sector(bi);
3032                 }
3033                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3034                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3035                                 sh->overwrite_disks++;
3036         }
3037
3038         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3039                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3040                 (unsigned long long)sh->sector, dd_idx);
3041
3042         if (conf->mddev->bitmap && firstwrite) {
3043                 /* Cannot hold spinlock over bitmap_startwrite,
3044                  * but must ensure this isn't added to a batch until
3045                  * we have added to the bitmap and set bm_seq.
3046                  * So set STRIPE_BITMAP_PENDING to prevent
3047                  * batching.
3048                  * If multiple add_stripe_bio() calls race here they
3049                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3050                  * to complete "bitmap_startwrite" gets to set
3051                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3052                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3053                  * any more.
3054                  */
3055                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3056                 spin_unlock_irq(&sh->stripe_lock);
3057                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3058                                   STRIPE_SECTORS, 0);
3059                 spin_lock_irq(&sh->stripe_lock);
3060                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3061                 if (!sh->batch_head) {
3062                         sh->bm_seq = conf->seq_flush+1;
3063                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3064                 }
3065         }
3066         spin_unlock_irq(&sh->stripe_lock);
3067
3068         if (stripe_can_batch(sh))
3069                 stripe_add_to_batch_list(conf, sh);
3070         return 1;
3071
3072  overlap:
3073         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3074         spin_unlock_irq(&sh->stripe_lock);
3075         return 0;
3076 }
3077
3078 static void end_reshape(struct r5conf *conf);
3079
3080 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3081                             struct stripe_head *sh)
3082 {
3083         int sectors_per_chunk =
3084                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3085         int dd_idx;
3086         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3087         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3088
3089         raid5_compute_sector(conf,
3090                              stripe * (disks - conf->max_degraded)
3091                              *sectors_per_chunk + chunk_offset,
3092                              previous,
3093                              &dd_idx, sh);
3094 }
3095
3096 static void
3097 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3098                                 struct stripe_head_state *s, int disks,
3099                                 struct bio_list *return_bi)
3100 {
3101         int i;
3102         BUG_ON(sh->batch_head);
3103         for (i = disks; i--; ) {
3104                 struct bio *bi;
3105                 int bitmap_end = 0;
3106
3107                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3108                         struct md_rdev *rdev;
3109                         rcu_read_lock();
3110                         rdev = rcu_dereference(conf->disks[i].rdev);
3111                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3112                             !test_bit(Faulty, &rdev->flags))
3113                                 atomic_inc(&rdev->nr_pending);
3114                         else
3115                                 rdev = NULL;
3116                         rcu_read_unlock();
3117                         if (rdev) {
3118                                 if (!rdev_set_badblocks(
3119                                             rdev,
3120                                             sh->sector,
3121                                             STRIPE_SECTORS, 0))
3122                                         md_error(conf->mddev, rdev);
3123                                 rdev_dec_pending(rdev, conf->mddev);
3124                         }
3125                 }
3126                 spin_lock_irq(&sh->stripe_lock);
3127                 /* fail all writes first */
3128                 bi = sh->dev[i].towrite;
3129                 sh->dev[i].towrite = NULL;
3130                 sh->overwrite_disks = 0;
3131                 spin_unlock_irq(&sh->stripe_lock);
3132                 if (bi)
3133                         bitmap_end = 1;
3134
3135                 r5l_stripe_write_finished(sh);
3136
3137                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3138                         wake_up(&conf->wait_for_overlap);
3139
3140                 while (bi && bi->bi_iter.bi_sector <
3141                         sh->dev[i].sector + STRIPE_SECTORS) {
3142                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3143
3144                         bi->bi_error = -EIO;
3145                         if (!raid5_dec_bi_active_stripes(bi)) {
3146                                 md_write_end(conf->mddev);
3147                                 bio_list_add(return_bi, bi);
3148                         }
3149                         bi = nextbi;
3150                 }
3151                 if (bitmap_end)
3152                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3153                                 STRIPE_SECTORS, 0, 0);
3154                 bitmap_end = 0;
3155                 /* and fail all 'written' */
3156                 bi = sh->dev[i].written;
3157                 sh->dev[i].written = NULL;
3158                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3159                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3160                         sh->dev[i].page = sh->dev[i].orig_page;
3161                 }
3162
3163                 if (bi) bitmap_end = 1;
3164                 while (bi && bi->bi_iter.bi_sector <
3165                        sh->dev[i].sector + STRIPE_SECTORS) {
3166                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3167
3168                         bi->bi_error = -EIO;
3169                         if (!raid5_dec_bi_active_stripes(bi)) {
3170                                 md_write_end(conf->mddev);
3171                                 bio_list_add(return_bi, bi);
3172                         }
3173                         bi = bi2;
3174                 }
3175
3176                 /* fail any reads if this device is non-operational and
3177                  * the data has not reached the cache yet.
3178                  */
3179                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3180                     s->failed > conf->max_degraded &&
3181                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3182                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3183                         spin_lock_irq(&sh->stripe_lock);
3184                         bi = sh->dev[i].toread;
3185                         sh->dev[i].toread = NULL;
3186                         spin_unlock_irq(&sh->stripe_lock);
3187                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3188                                 wake_up(&conf->wait_for_overlap);
3189                         if (bi)
3190                                 s->to_read--;
3191                         while (bi && bi->bi_iter.bi_sector <
3192                                sh->dev[i].sector + STRIPE_SECTORS) {
3193                                 struct bio *nextbi =
3194                                         r5_next_bio(bi, sh->dev[i].sector);
3195
3196                                 bi->bi_error = -EIO;
3197                                 if (!raid5_dec_bi_active_stripes(bi))
3198                                         bio_list_add(return_bi, bi);
3199                                 bi = nextbi;
3200                         }
3201                 }
3202                 if (bitmap_end)
3203                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3204                                         STRIPE_SECTORS, 0, 0);
3205                 /* If we were in the middle of a write the parity block might
3206                  * still be locked - so just clear all R5_LOCKED flags
3207                  */
3208                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3209         }
3210         s->to_write = 0;
3211         s->written = 0;
3212
3213         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3214                 if (atomic_dec_and_test(&conf->pending_full_writes))
3215                         md_wakeup_thread(conf->mddev->thread);
3216 }
3217
3218 static void
3219 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3220                    struct stripe_head_state *s)
3221 {
3222         int abort = 0;
3223         int i;
3224
3225         BUG_ON(sh->batch_head);
3226         clear_bit(STRIPE_SYNCING, &sh->state);
3227         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3228                 wake_up(&conf->wait_for_overlap);
3229         s->syncing = 0;
3230         s->replacing = 0;
3231         /* There is nothing more to do for sync/check/repair.
3232          * Don't even need to abort as that is handled elsewhere
3233          * if needed, and not always wanted e.g. if there is a known
3234          * bad block here.
3235          * For recover/replace we need to record a bad block on all
3236          * non-sync devices, or abort the recovery
3237          */
3238         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3239                 /* During recovery devices cannot be removed, so
3240                  * locking and refcounting of rdevs is not needed
3241                  */
3242                 rcu_read_lock();
3243                 for (i = 0; i < conf->raid_disks; i++) {
3244                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3245                         if (rdev
3246                             && !test_bit(Faulty, &rdev->flags)
3247                             && !test_bit(In_sync, &rdev->flags)
3248                             && !rdev_set_badblocks(rdev, sh->sector,
3249                                                    STRIPE_SECTORS, 0))
3250                                 abort = 1;
3251                         rdev = rcu_dereference(conf->disks[i].replacement);
3252                         if (rdev
3253                             && !test_bit(Faulty, &rdev->flags)
3254                             && !test_bit(In_sync, &rdev->flags)
3255                             && !rdev_set_badblocks(rdev, sh->sector,
3256                                                    STRIPE_SECTORS, 0))
3257                                 abort = 1;
3258                 }
3259                 rcu_read_unlock();
3260                 if (abort)
3261                         conf->recovery_disabled =
3262                                 conf->mddev->recovery_disabled;
3263         }
3264         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3265 }
3266
3267 static int want_replace(struct stripe_head *sh, int disk_idx)
3268 {
3269         struct md_rdev *rdev;
3270         int rv = 0;
3271
3272         rcu_read_lock();
3273         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3274         if (rdev
3275             && !test_bit(Faulty, &rdev->flags)
3276             && !test_bit(In_sync, &rdev->flags)
3277             && (rdev->recovery_offset <= sh->sector
3278                 || rdev->mddev->recovery_cp <= sh->sector))
3279                 rv = 1;
3280         rcu_read_unlock();
3281         return rv;
3282 }
3283
3284 /* fetch_block - checks the given member device to see if its data needs
3285  * to be read or computed to satisfy a request.
3286  *
3287  * Returns 1 when no more member devices need to be checked, otherwise returns
3288  * 0 to tell the loop in handle_stripe_fill to continue
3289  */
3290
3291 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3292                            int disk_idx, int disks)
3293 {
3294         struct r5dev *dev = &sh->dev[disk_idx];
3295         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3296                                   &sh->dev[s->failed_num[1]] };
3297         int i;
3298
3299
3300         if (test_bit(R5_LOCKED, &dev->flags) ||
3301             test_bit(R5_UPTODATE, &dev->flags))
3302                 /* No point reading this as we already have it or have
3303                  * decided to get it.
3304                  */
3305                 return 0;
3306
3307         if (dev->toread ||
3308             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3309                 /* We need this block to directly satisfy a request */
3310                 return 1;
3311
3312         if (s->syncing || s->expanding ||
3313             (s->replacing && want_replace(sh, disk_idx)))
3314                 /* When syncing, or expanding we read everything.
3315                  * When replacing, we need the replaced block.
3316                  */
3317                 return 1;
3318
3319         if ((s->failed >= 1 && fdev[0]->toread) ||
3320             (s->failed >= 2 && fdev[1]->toread))
3321                 /* If we want to read from a failed device, then
3322                  * we need to actually read every other device.
3323                  */
3324                 return 1;
3325
3326         /* Sometimes neither read-modify-write nor reconstruct-write
3327          * cycles can work.  In those cases we read every block we
3328          * can.  Then the parity-update is certain to have enough to
3329          * work with.
3330          * This can only be a problem when we need to write something,
3331          * and some device has failed.  If either of those tests
3332          * fail we need look no further.
3333          */
3334         if (!s->failed || !s->to_write)
3335                 return 0;
3336
3337         if (test_bit(R5_Insync, &dev->flags) &&
3338             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3339                 /* Pre-reads at not permitted until after short delay
3340                  * to gather multiple requests.  However if this
3341                  * device is no Insync, the block could only be be computed
3342                  * and there is no need to delay that.
3343                  */
3344                 return 0;
3345
3346         for (i = 0; i < s->failed && i < 2; i++) {
3347                 if (fdev[i]->towrite &&
3348                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3349                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3350                         /* If we have a partial write to a failed
3351                          * device, then we will need to reconstruct
3352                          * the content of that device, so all other
3353                          * devices must be read.
3354                          */
3355                         return 1;
3356         }
3357
3358         /* If we are forced to do a reconstruct-write, either because
3359          * the current RAID6 implementation only supports that, or
3360          * or because parity cannot be trusted and we are currently
3361          * recovering it, there is extra need to be careful.
3362          * If one of the devices that we would need to read, because
3363          * it is not being overwritten (and maybe not written at all)
3364          * is missing/faulty, then we need to read everything we can.
3365          */
3366         if (sh->raid_conf->level != 6 &&
3367             sh->raid_conf->rmw_level != PARITY_DISABLE_RMW &&
3368             sh->sector < sh->raid_conf->mddev->recovery_cp)
3369                 /* reconstruct-write isn't being forced */
3370                 return 0;
3371         for (i = 0; i < s->failed && i < 2; i++) {
3372                 if (s->failed_num[i] != sh->pd_idx &&
3373                     s->failed_num[i] != sh->qd_idx &&
3374                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3375                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3376                         return 1;
3377         }
3378
3379         return 0;
3380 }
3381
3382 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3383                        int disk_idx, int disks)
3384 {
3385         struct r5dev *dev = &sh->dev[disk_idx];
3386
3387         /* is the data in this block needed, and can we get it? */
3388         if (need_this_block(sh, s, disk_idx, disks)) {
3389                 /* we would like to get this block, possibly by computing it,
3390                  * otherwise read it if the backing disk is insync
3391                  */
3392                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3393                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3394                 BUG_ON(sh->batch_head);
3395
3396                 /*
3397                  * In the raid6 case if the only non-uptodate disk is P
3398                  * then we already trusted P to compute the other failed
3399                  * drives. It is safe to compute rather than re-read P.
3400                  * In other cases we only compute blocks from failed
3401                  * devices, otherwise check/repair might fail to detect
3402                  * a real inconsistency.
3403                  */
3404
3405                 if ((s->uptodate == disks - 1) &&
3406                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3407                     (s->failed && (disk_idx == s->failed_num[0] ||
3408                                    disk_idx == s->failed_num[1])))) {
3409                         /* have disk failed, and we're requested to fetch it;
3410                          * do compute it
3411                          */
3412                         pr_debug("Computing stripe %llu block %d\n",
3413                                (unsigned long long)sh->sector, disk_idx);
3414                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3415                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3416                         set_bit(R5_Wantcompute, &dev->flags);
3417                         sh->ops.target = disk_idx;
3418                         sh->ops.target2 = -1; /* no 2nd target */
3419                         s->req_compute = 1;
3420                         /* Careful: from this point on 'uptodate' is in the eye
3421                          * of raid_run_ops which services 'compute' operations
3422                          * before writes. R5_Wantcompute flags a block that will
3423                          * be R5_UPTODATE by the time it is needed for a
3424                          * subsequent operation.
3425                          */
3426                         s->uptodate++;
3427                         return 1;
3428                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3429                         /* Computing 2-failure is *very* expensive; only
3430                          * do it if failed >= 2
3431                          */
3432                         int other;
3433                         for (other = disks; other--; ) {
3434                                 if (other == disk_idx)
3435                                         continue;
3436                                 if (!test_bit(R5_UPTODATE,
3437                                       &sh->dev[other].flags))
3438                                         break;
3439                         }
3440                         BUG_ON(other < 0);
3441                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3442                                (unsigned long long)sh->sector,
3443                                disk_idx, other);
3444                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3445                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3446                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3447                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3448                         sh->ops.target = disk_idx;
3449                         sh->ops.target2 = other;
3450                         s->uptodate += 2;
3451                         s->req_compute = 1;
3452                         return 1;
3453                 } else if (test_bit(R5_Insync, &dev->flags)) {
3454                         set_bit(R5_LOCKED, &dev->flags);
3455                         set_bit(R5_Wantread, &dev->flags);
3456                         s->locked++;
3457                         pr_debug("Reading block %d (sync=%d)\n",
3458                                 disk_idx, s->syncing);
3459                 }
3460         }
3461
3462         return 0;
3463 }
3464
3465 /**
3466  * handle_stripe_fill - read or compute data to satisfy pending requests.
3467  */
3468 static void handle_stripe_fill(struct stripe_head *sh,
3469                                struct stripe_head_state *s,
3470                                int disks)
3471 {
3472         int i;
3473
3474         /* look for blocks to read/compute, skip this if a compute
3475          * is already in flight, or if the stripe contents are in the
3476          * midst of changing due to a write
3477          */
3478         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3479             !sh->reconstruct_state)
3480                 for (i = disks; i--; )
3481                         if (fetch_block(sh, s, i, disks))
3482                                 break;
3483         set_bit(STRIPE_HANDLE, &sh->state);
3484 }
3485
3486 static void break_stripe_batch_list(struct stripe_head *head_sh,
3487                                     unsigned long handle_flags);
3488 /* handle_stripe_clean_event
3489  * any written block on an uptodate or failed drive can be returned.
3490  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3491  * never LOCKED, so we don't need to test 'failed' directly.
3492  */
3493 static void handle_stripe_clean_event(struct r5conf *conf,
3494         struct stripe_head *sh, int disks, struct bio_list *return_bi)
3495 {
3496         int i;
3497         struct r5dev *dev;
3498         int discard_pending = 0;
3499         struct stripe_head *head_sh = sh;
3500         bool do_endio = false;
3501
3502         for (i = disks; i--; )
3503                 if (sh->dev[i].written) {
3504                         dev = &sh->dev[i];
3505                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3506                             (test_bit(R5_UPTODATE, &dev->flags) ||
3507                              test_bit(R5_Discard, &dev->flags) ||
3508                              test_bit(R5_SkipCopy, &dev->flags))) {
3509                                 /* We can return any write requests */
3510                                 struct bio *wbi, *wbi2;
3511                                 pr_debug("Return write for disc %d\n", i);
3512                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3513                                         clear_bit(R5_UPTODATE, &dev->flags);
3514                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3515                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3516                                 }
3517                                 do_endio = true;
3518
3519 returnbi:
3520                                 dev->page = dev->orig_page;
3521                                 wbi = dev->written;
3522                                 dev->written = NULL;
3523                                 while (wbi && wbi->bi_iter.bi_sector <
3524                                         dev->sector + STRIPE_SECTORS) {
3525                                         wbi2 = r5_next_bio(wbi, dev->sector);
3526                                         if (!raid5_dec_bi_active_stripes(wbi)) {
3527                                                 md_write_end(conf->mddev);
3528                                                 bio_list_add(return_bi, wbi);
3529                                         }
3530                                         wbi = wbi2;
3531                                 }
3532                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3533                                                 STRIPE_SECTORS,
3534                                          !test_bit(STRIPE_DEGRADED, &sh->state),
3535                                                 0);
3536                                 if (head_sh->batch_head) {
3537                                         sh = list_first_entry(&sh->batch_list,
3538                                                               struct stripe_head,
3539                                                               batch_list);
3540                                         if (sh != head_sh) {
3541                                                 dev = &sh->dev[i];
3542                                                 goto returnbi;
3543                                         }
3544                                 }
3545                                 sh = head_sh;
3546                                 dev = &sh->dev[i];
3547                         } else if (test_bit(R5_Discard, &dev->flags))
3548                                 discard_pending = 1;
3549                 }
3550
3551         r5l_stripe_write_finished(sh);
3552
3553         if (!discard_pending &&
3554             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3555                 int hash;
3556                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3557                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3558                 if (sh->qd_idx >= 0) {
3559                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3560                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3561                 }
3562                 /* now that discard is done we can proceed with any sync */
3563                 clear_bit(STRIPE_DISCARD, &sh->state);
3564                 /*
3565                  * SCSI discard will change some bio fields and the stripe has
3566                  * no updated data, so remove it from hash list and the stripe
3567                  * will be reinitialized
3568                  */
3569 unhash:
3570                 hash = sh->hash_lock_index;
3571                 spin_lock_irq(conf->hash_locks + hash);
3572                 remove_hash(sh);
3573                 spin_unlock_irq(conf->hash_locks + hash);
3574                 if (head_sh->batch_head) {
3575                         sh = list_first_entry(&sh->batch_list,
3576                                               struct stripe_head, batch_list);
3577                         if (sh != head_sh)
3578                                         goto unhash;
3579                 }
3580                 sh = head_sh;
3581
3582                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3583                         set_bit(STRIPE_HANDLE, &sh->state);
3584
3585         }
3586
3587         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3588                 if (atomic_dec_and_test(&conf->pending_full_writes))
3589                         md_wakeup_thread(conf->mddev->thread);
3590
3591         if (head_sh->batch_head && do_endio)
3592                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3593 }
3594
3595 static void handle_stripe_dirtying(struct r5conf *conf,
3596                                    struct stripe_head *sh,
3597                                    struct stripe_head_state *s,
3598                                    int disks)
3599 {
3600         int rmw = 0, rcw = 0, i;
3601         sector_t recovery_cp = conf->mddev->recovery_cp;
3602
3603         /* Check whether resync is now happening or should start.
3604          * If yes, then the array is dirty (after unclean shutdown or
3605          * initial creation), so parity in some stripes might be inconsistent.
3606          * In this case, we need to always do reconstruct-write, to ensure
3607          * that in case of drive failure or read-error correction, we
3608          * generate correct data from the parity.
3609          */
3610         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3611             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3612              s->failed == 0)) {
3613                 /* Calculate the real rcw later - for now make it
3614                  * look like rcw is cheaper
3615                  */
3616                 rcw = 1; rmw = 2;
3617                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3618                          conf->rmw_level, (unsigned long long)recovery_cp,
3619                          (unsigned long long)sh->sector);
3620         } else for (i = disks; i--; ) {
3621                 /* would I have to read this buffer for read_modify_write */
3622                 struct r5dev *dev = &sh->dev[i];
3623                 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3624                     !test_bit(R5_LOCKED, &dev->flags) &&
3625                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3626                       test_bit(R5_Wantcompute, &dev->flags))) {
3627                         if (test_bit(R5_Insync, &dev->flags))
3628                                 rmw++;
3629                         else
3630                                 rmw += 2*disks;  /* cannot read it */
3631                 }
3632                 /* Would I have to read this buffer for reconstruct_write */
3633                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3634                     i != sh->pd_idx && i != sh->qd_idx &&
3635                     !test_bit(R5_LOCKED, &dev->flags) &&
3636                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3637                     test_bit(R5_Wantcompute, &dev->flags))) {
3638                         if (test_bit(R5_Insync, &dev->flags))
3639                                 rcw++;
3640                         else
3641                                 rcw += 2*disks;
3642                 }
3643         }
3644         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3645                 (unsigned long long)sh->sector, rmw, rcw);
3646         set_bit(STRIPE_HANDLE, &sh->state);
3647         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3648                 /* prefer read-modify-write, but need to get some data */
3649                 if (conf->mddev->queue)
3650                         blk_add_trace_msg(conf->mddev->queue,
3651                                           "raid5 rmw %llu %d",
3652                                           (unsigned long long)sh->sector, rmw);
3653                 for (i = disks; i--; ) {
3654                         struct r5dev *dev = &sh->dev[i];
3655                         if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3656                             !test_bit(R5_LOCKED, &dev->flags) &&
3657                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3658                             test_bit(R5_Wantcompute, &dev->flags)) &&
3659                             test_bit(R5_Insync, &dev->flags)) {
3660                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3661                                              &sh->state)) {
3662                                         pr_debug("Read_old block %d for r-m-w\n",
3663                                                  i);
3664                                         set_bit(R5_LOCKED, &dev->flags);
3665                                         set_bit(R5_Wantread, &dev->flags);
3666                                         s->locked++;
3667                                 } else {
3668                                         set_bit(STRIPE_DELAYED, &sh->state);
3669                                         set_bit(STRIPE_HANDLE, &sh->state);
3670                                 }
3671                         }
3672                 }
3673         }
3674         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3675                 /* want reconstruct write, but need to get some data */
3676                 int qread =0;
3677                 rcw = 0;
3678                 for (i = disks; i--; ) {
3679                         struct r5dev *dev = &sh->dev[i];
3680                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3681                             i != sh->pd_idx && i != sh->qd_idx &&
3682                             !test_bit(R5_LOCKED, &dev->flags) &&
3683                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3684                               test_bit(R5_Wantcompute, &dev->flags))) {
3685                                 rcw++;
3686                                 if (test_bit(R5_Insync, &dev->flags) &&
3687                                     test_bit(STRIPE_PREREAD_ACTIVE,
3688                                              &sh->state)) {
3689                                         pr_debug("Read_old block "
3690                                                 "%d for Reconstruct\n", i);
3691                                         set_bit(R5_LOCKED, &dev->flags);
3692                                         set_bit(R5_Wantread, &dev->flags);
3693                                         s->locked++;
3694                                         qread++;
3695                                 } else {
3696                                         set_bit(STRIPE_DELAYED, &sh->state);
3697                                         set_bit(STRIPE_HANDLE, &sh->state);
3698                                 }
3699                         }
3700                 }
3701                 if (rcw && conf->mddev->queue)
3702                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3703                                           (unsigned long long)sh->sector,
3704                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3705         }
3706
3707         if (rcw > disks && rmw > disks &&
3708             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3709                 set_bit(STRIPE_DELAYED, &sh->state);
3710
3711         /* now if nothing is locked, and if we have enough data,
3712          * we can start a write request
3713          */
3714         /* since handle_stripe can be called at any time we need to handle the
3715          * case where a compute block operation has been submitted and then a
3716          * subsequent call wants to start a write request.  raid_run_ops only
3717          * handles the case where compute block and reconstruct are requested
3718          * simultaneously.  If this is not the case then new writes need to be
3719          * held off until the compute completes.
3720          */
3721         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3722             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3723             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3724                 schedule_reconstruction(sh, s, rcw == 0, 0);
3725 }
3726
3727 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3728                                 struct stripe_head_state *s, int disks)
3729 {
3730         struct r5dev *dev = NULL;
3731
3732         BUG_ON(sh->batch_head);
3733         set_bit(STRIPE_HANDLE, &sh->state);
3734
3735         switch (sh->check_state) {
3736         case check_state_idle:
3737                 /* start a new check operation if there are no failures */
3738                 if (s->failed == 0) {
3739                         BUG_ON(s->uptodate != disks);
3740                         sh->check_state = check_state_run;
3741                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3742                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3743                         s->uptodate--;
3744                         break;
3745                 }
3746                 dev = &sh->dev[s->failed_num[0]];
3747                 /* fall through */
3748         case check_state_compute_result:
3749                 sh->check_state = check_state_idle;
3750                 if (!dev)
3751                         dev = &sh->dev[sh->pd_idx];
3752
3753                 /* check that a write has not made the stripe insync */
3754                 if (test_bit(STRIPE_INSYNC, &sh->state))
3755                         break;
3756
3757                 /* either failed parity check, or recovery is happening */
3758                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3759                 BUG_ON(s->uptodate != disks);
3760
3761                 set_bit(R5_LOCKED, &dev->flags);
3762                 s->locked++;
3763                 set_bit(R5_Wantwrite, &dev->flags);
3764
3765                 clear_bit(STRIPE_DEGRADED, &sh->state);
3766                 set_bit(STRIPE_INSYNC, &sh->state);
3767                 break;
3768         case check_state_run:
3769                 break; /* we will be called again upon completion */
3770         case check_state_check_result:
3771                 sh->check_state = check_state_idle;
3772
3773                 /* if a failure occurred during the check operation, leave
3774                  * STRIPE_INSYNC not set and let the stripe be handled again
3775                  */
3776                 if (s->failed)
3777                         break;
3778
3779                 /* handle a successful check operation, if parity is correct
3780                  * we are done.  Otherwise update the mismatch count and repair
3781                  * parity if !MD_RECOVERY_CHECK
3782                  */
3783                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3784                         /* parity is correct (on disc,
3785                          * not in buffer any more)
3786                          */
3787                         set_bit(STRIPE_INSYNC, &sh->state);
3788                 else {
3789                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3790                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3791                                 /* don't try to repair!! */
3792                                 set_bit(STRIPE_INSYNC, &sh->state);
3793                         else {
3794                                 sh->check_state = check_state_compute_run;
3795                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3796                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3797                                 set_bit(R5_Wantcompute,
3798                                         &sh->dev[sh->pd_idx].flags);
3799                                 sh->ops.target = sh->pd_idx;
3800                                 sh->ops.target2 = -1;
3801                                 s->uptodate++;
3802                         }
3803                 }
3804                 break;
3805         case check_state_compute_run:
3806                 break;
3807         default:
3808                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3809                        __func__, sh->check_state,
3810                        (unsigned long long) sh->sector);
3811                 BUG();
3812         }
3813 }
3814
3815 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3816                                   struct stripe_head_state *s,
3817                                   int disks)
3818 {
3819         int pd_idx = sh->pd_idx;
3820         int qd_idx = sh->qd_idx;
3821         struct r5dev *dev;
3822
3823         BUG_ON(sh->batch_head);
3824         set_bit(STRIPE_HANDLE, &sh->state);
3825
3826         BUG_ON(s->failed > 2);
3827
3828         /* Want to check and possibly repair P and Q.
3829          * However there could be one 'failed' device, in which
3830          * case we can only check one of them, possibly using the
3831          * other to generate missing data
3832          */
3833
3834         switch (sh->check_state) {
3835         case check_state_idle:
3836                 /* start a new check operation if there are < 2 failures */
3837                 if (s->failed == s->q_failed) {
3838                         /* The only possible failed device holds Q, so it
3839                          * makes sense to check P (If anything else were failed,
3840                          * we would have used P to recreate it).
3841                          */
3842                         sh->check_state = check_state_run;
3843                 }
3844                 if (!s->q_failed && s->failed < 2) {
3845                         /* Q is not failed, and we didn't use it to generate
3846                          * anything, so it makes sense to check it
3847                          */
3848                         if (sh->check_state == check_state_run)
3849                                 sh->check_state = check_state_run_pq;
3850                         else
3851                                 sh->check_state = check_state_run_q;
3852                 }
3853
3854                 /* discard potentially stale zero_sum_result */
3855                 sh->ops.zero_sum_result = 0;
3856
3857                 if (sh->check_state == check_state_run) {
3858                         /* async_xor_zero_sum destroys the contents of P */
3859                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3860                         s->uptodate--;
3861                 }
3862                 if (sh->check_state >= check_state_run &&
3863                     sh->check_state <= check_state_run_pq) {
3864                         /* async_syndrome_zero_sum preserves P and Q, so
3865                          * no need to mark them !uptodate here
3866                          */
3867                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3868                         break;
3869                 }
3870
3871                 /* we have 2-disk failure */
3872                 BUG_ON(s->failed != 2);
3873                 /* fall through */
3874         case check_state_compute_result:
3875                 sh->check_state = check_state_idle;
3876
3877                 /* check that a write has not made the stripe insync */
3878                 if (test_bit(STRIPE_INSYNC, &sh->state))
3879                         break;
3880
3881                 /* now write out any block on a failed drive,
3882                  * or P or Q if they were recomputed
3883                  */
3884                 dev = NULL;
3885                 if (s->failed == 2) {
3886                         dev = &sh->dev[s->failed_num[1]];
3887                         s->locked++;
3888                         set_bit(R5_LOCKED, &dev->flags);
3889                         set_bit(R5_Wantwrite, &dev->flags);
3890                 }
3891                 if (s->failed >= 1) {
3892                         dev = &sh->dev[s->failed_num[0]];
3893                         s->locked++;
3894                         set_bit(R5_LOCKED, &dev->flags);
3895                         set_bit(R5_Wantwrite, &dev->flags);
3896                 }
3897                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3898                         dev = &sh->dev[pd_idx];
3899                         s->locked++;
3900                         set_bit(R5_LOCKED, &dev->flags);
3901                         set_bit(R5_Wantwrite, &dev->flags);
3902                 }
3903                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3904                         dev = &sh->dev[qd_idx];
3905                         s->locked++;
3906                         set_bit(R5_LOCKED, &dev->flags);
3907                         set_bit(R5_Wantwrite, &dev->flags);
3908                 }
3909                 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
3910                               "%s: disk%td not up to date\n",
3911                               mdname(conf->mddev),
3912                               dev - (struct r5dev *) &sh->dev)) {
3913                         clear_bit(R5_LOCKED, &dev->flags);
3914                         clear_bit(R5_Wantwrite, &dev->flags);
3915                         s->locked--;
3916                 }
3917                 clear_bit(STRIPE_DEGRADED, &sh->state);
3918
3919                 set_bit(STRIPE_INSYNC, &sh->state);
3920                 break;
3921         case check_state_run:
3922         case check_state_run_q:
3923         case check_state_run_pq:
3924                 break; /* we will be called again upon completion */
3925         case check_state_check_result:
3926                 sh->check_state = check_state_idle;
3927
3928                 /* handle a successful check operation, if parity is correct
3929                  * we are done.  Otherwise update the mismatch count and repair
3930                  * parity if !MD_RECOVERY_CHECK
3931                  */
3932                 if (sh->ops.zero_sum_result == 0) {
3933                         /* both parities are correct */
3934                         if (!s->failed)
3935                                 set_bit(STRIPE_INSYNC, &sh->state);
3936                         else {
3937                                 /* in contrast to the raid5 case we can validate
3938                                  * parity, but still have a failure to write
3939                                  * back
3940                                  */
3941                                 sh->check_state = check_state_compute_result;
3942                                 /* Returning at this point means that we may go
3943                                  * off and bring p and/or q uptodate again so
3944                                  * we make sure to check zero_sum_result again
3945                                  * to verify if p or q need writeback
3946                                  */
3947                         }
3948                 } else {
3949                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3950                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3951                                 /* don't try to repair!! */
3952                                 set_bit(STRIPE_INSYNC, &sh->state);
3953                         else {
3954                                 int *target = &sh->ops.target;
3955
3956                                 sh->ops.target = -1;
3957                                 sh->ops.target2 = -1;
3958                                 sh->check_state = check_state_compute_run;
3959                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3960                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3961                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3962                                         set_bit(R5_Wantcompute,
3963                                                 &sh->dev[pd_idx].flags);
3964                                         *target = pd_idx;
3965                                         target = &sh->ops.target2;
3966                                         s->uptodate++;
3967                                 }
3968                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3969                                         set_bit(R5_Wantcompute,
3970                                                 &sh->dev[qd_idx].flags);
3971                                         *target = qd_idx;
3972                                         s->uptodate++;
3973                                 }
3974                         }
3975                 }
3976                 break;
3977         case check_state_compute_run:
3978                 break;
3979         default:
3980                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3981                        __func__, sh->check_state,
3982                        (unsigned long long) sh->sector);
3983                 BUG();
3984         }
3985 }
3986
3987 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3988 {
3989         int i;
3990
3991         /* We have read all the blocks in this stripe and now we need to
3992          * copy some of them into a target stripe for expand.
3993          */
3994         struct dma_async_tx_descriptor *tx = NULL;
3995         BUG_ON(sh->batch_head);
3996         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3997         for (i = 0; i < sh->disks; i++)
3998                 if (i != sh->pd_idx && i != sh->qd_idx) {
3999                         int dd_idx, j;
4000                         struct stripe_head *sh2;
4001                         struct async_submit_ctl submit;
4002
4003                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4004                         sector_t s = raid5_compute_sector(conf, bn, 0,
4005                                                           &dd_idx, NULL);
4006                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4007                         if (sh2 == NULL)
4008                                 /* so far only the early blocks of this stripe
4009                                  * have been requested.  When later blocks
4010                                  * get requested, we will try again
4011                                  */
4012                                 continue;
4013                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4014                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4015                                 /* must have already done this block */
4016                                 raid5_release_stripe(sh2);
4017                                 continue;
4018                         }
4019
4020                         /* place all the copies on one channel */
4021                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4022                         tx = async_memcpy(sh2->dev[dd_idx].page,
4023                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
4024                                           &submit);
4025
4026                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4027                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4028                         for (j = 0; j < conf->raid_disks; j++)
4029                                 if (j != sh2->pd_idx &&
4030                                     j != sh2->qd_idx &&
4031                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4032                                         break;
4033                         if (j == conf->raid_disks) {
4034                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4035                                 set_bit(STRIPE_HANDLE, &sh2->state);
4036                         }
4037                         raid5_release_stripe(sh2);
4038
4039                 }
4040         /* done submitting copies, wait for them to complete */
4041         async_tx_quiesce(&tx);
4042 }
4043
4044 /*
4045  * handle_stripe - do things to a stripe.
4046  *
4047  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4048  * state of various bits to see what needs to be done.
4049  * Possible results:
4050  *    return some read requests which now have data
4051  *    return some write requests which are safely on storage
4052  *    schedule a read on some buffers
4053  *    schedule a write of some buffers
4054  *    return confirmation of parity correctness
4055  *
4056  */
4057
4058 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4059 {
4060         struct r5conf *conf = sh->raid_conf;
4061         int disks = sh->disks;
4062         struct r5dev *dev;
4063         int i;
4064         int do_recovery = 0;
4065
4066         memset(s, 0, sizeof(*s));
4067
4068         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4069         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4070         s->failed_num[0] = -1;
4071         s->failed_num[1] = -1;
4072         s->log_failed = r5l_log_disk_error(conf);
4073
4074         /* Now to look around and see what can be done */
4075         rcu_read_lock();
4076         for (i=disks; i--; ) {
4077                 struct md_rdev *rdev;
4078                 sector_t first_bad;
4079                 int bad_sectors;
4080                 int is_bad = 0;
4081
4082                 dev = &sh->dev[i];
4083
4084                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4085                          i, dev->flags,
4086                          dev->toread, dev->towrite, dev->written);
4087                 /* maybe we can reply to a read
4088                  *
4089                  * new wantfill requests are only permitted while
4090                  * ops_complete_biofill is guaranteed to be inactive
4091                  */
4092                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4093                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4094                         set_bit(R5_Wantfill, &dev->flags);
4095
4096                 /* now count some things */
4097                 if (test_bit(R5_LOCKED, &dev->flags))
4098                         s->locked++;
4099                 if (test_bit(R5_UPTODATE, &dev->flags))
4100                         s->uptodate++;
4101                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4102                         s->compute++;
4103                         BUG_ON(s->compute > 2);
4104                 }
4105
4106                 if (test_bit(R5_Wantfill, &dev->flags))
4107                         s->to_fill++;
4108                 else if (dev->toread)
4109                         s->to_read++;
4110                 if (dev->towrite) {
4111                         s->to_write++;
4112                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4113                                 s->non_overwrite++;
4114                 }
4115                 if (dev->written)
4116                         s->written++;
4117                 /* Prefer to use the replacement for reads, but only
4118                  * if it is recovered enough and has no bad blocks.
4119                  */
4120                 rdev = rcu_dereference(conf->disks[i].replacement);
4121                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4122                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4123                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4124                                  &first_bad, &bad_sectors))
4125                         set_bit(R5_ReadRepl, &dev->flags);
4126                 else {
4127                         if (rdev && !test_bit(Faulty, &rdev->flags))
4128                                 set_bit(R5_NeedReplace, &dev->flags);
4129                         else
4130                                 clear_bit(R5_NeedReplace, &dev->flags);
4131                         rdev = rcu_dereference(conf->disks[i].rdev);
4132                         clear_bit(R5_ReadRepl, &dev->flags);
4133                 }
4134                 if (rdev && test_bit(Faulty, &rdev->flags))
4135                         rdev = NULL;
4136                 if (rdev) {
4137                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4138                                              &first_bad, &bad_sectors);
4139                         if (s->blocked_rdev == NULL
4140                             && (test_bit(Blocked, &rdev->flags)
4141                                 || is_bad < 0)) {
4142                                 if (is_bad < 0)
4143                                         set_bit(BlockedBadBlocks,
4144                                                 &rdev->flags);
4145                                 s->blocked_rdev = rdev;
4146                                 atomic_inc(&rdev->nr_pending);
4147                         }
4148                 }
4149                 clear_bit(R5_Insync, &dev->flags);
4150                 if (!rdev)
4151                         /* Not in-sync */;
4152                 else if (is_bad) {
4153                         /* also not in-sync */
4154                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4155                             test_bit(R5_UPTODATE, &dev->flags)) {
4156                                 /* treat as in-sync, but with a read error
4157                                  * which we can now try to correct
4158                                  */
4159                                 set_bit(R5_Insync, &dev->flags);
4160                                 set_bit(R5_ReadError, &dev->flags);
4161                         }
4162                 } else if (test_bit(In_sync, &rdev->flags))
4163                         set_bit(R5_Insync, &dev->flags);
4164                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4165                         /* in sync if before recovery_offset */
4166                         set_bit(R5_Insync, &dev->flags);
4167                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4168                          test_bit(R5_Expanded, &dev->flags))
4169                         /* If we've reshaped into here, we assume it is Insync.
4170                          * We will shortly update recovery_offset to make
4171                          * it official.
4172                          */
4173                         set_bit(R5_Insync, &dev->flags);
4174
4175                 if (test_bit(R5_WriteError, &dev->flags)) {
4176                         /* This flag does not apply to '.replacement'
4177                          * only to .rdev, so make sure to check that*/
4178                         struct md_rdev *rdev2 = rcu_dereference(
4179                                 conf->disks[i].rdev);
4180                         if (rdev2 == rdev)
4181                                 clear_bit(R5_Insync, &dev->flags);
4182                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4183                                 s->handle_bad_blocks = 1;
4184                                 atomic_inc(&rdev2->nr_pending);
4185                         } else
4186                                 clear_bit(R5_WriteError, &dev->flags);
4187                 }
4188                 if (test_bit(R5_MadeGood, &dev->flags)) {
4189                         /* This flag does not apply to '.replacement'
4190                          * only to .rdev, so make sure to check that*/
4191                         struct md_rdev *rdev2 = rcu_dereference(
4192                                 conf->disks[i].rdev);
4193                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4194                                 s->handle_bad_blocks = 1;
4195                                 atomic_inc(&rdev2->nr_pending);
4196                         } else
4197                                 clear_bit(R5_MadeGood, &dev->flags);
4198                 }
4199                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4200                         struct md_rdev *rdev2 = rcu_dereference(
4201                                 conf->disks[i].replacement);
4202                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4203                                 s->handle_bad_blocks = 1;
4204                                 atomic_inc(&rdev2->nr_pending);
4205                         } else
4206                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4207                 }
4208                 if (!test_bit(R5_Insync, &dev->flags)) {
4209                         /* The ReadError flag will just be confusing now */
4210                         clear_bit(R5_ReadError, &dev->flags);
4211                         clear_bit(R5_ReWrite, &dev->flags);
4212                 }
4213                 if (test_bit(R5_ReadError, &dev->flags))
4214                         clear_bit(R5_Insync, &dev->flags);
4215                 if (!test_bit(R5_Insync, &dev->flags)) {
4216                         if (s->failed < 2)
4217                                 s->failed_num[s->failed] = i;
4218                         s->failed++;
4219                         if (rdev && !test_bit(Faulty, &rdev->flags))
4220                                 do_recovery = 1;
4221                         else if (!rdev) {
4222                                 rdev = rcu_dereference(
4223                                     conf->disks[i].replacement);
4224                                 if (rdev && !test_bit(Faulty, &rdev->flags))
4225                                         do_recovery = 1;
4226                         }
4227                 }
4228         }
4229         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4230                 /* If there is a failed device being replaced,
4231                  *     we must be recovering.
4232                  * else if we are after recovery_cp, we must be syncing
4233                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4234                  * else we can only be replacing
4235                  * sync and recovery both need to read all devices, and so
4236                  * use the same flag.
4237                  */
4238                 if (do_recovery ||
4239                     sh->sector >= conf->mddev->recovery_cp ||
4240                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4241                         s->syncing = 1;
4242                 else
4243                         s->replacing = 1;
4244         }
4245         rcu_read_unlock();
4246 }
4247
4248 static int clear_batch_ready(struct stripe_head *sh)
4249 {
4250         /* Return '1' if this is a member of batch, or
4251          * '0' if it is a lone stripe or a head which can now be
4252          * handled.
4253          */
4254         struct stripe_head *tmp;
4255         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4256                 return (sh->batch_head && sh->batch_head != sh);
4257         spin_lock(&sh->stripe_lock);
4258         if (!sh->batch_head) {
4259                 spin_unlock(&sh->stripe_lock);
4260                 return 0;
4261         }
4262
4263         /*
4264          * this stripe could be added to a batch list before we check
4265          * BATCH_READY, skips it
4266          */
4267         if (sh->batch_head != sh) {
4268                 spin_unlock(&sh->stripe_lock);
4269                 return 1;
4270         }
4271         spin_lock(&sh->batch_lock);
4272         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4273                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4274         spin_unlock(&sh->batch_lock);
4275         spin_unlock(&sh->stripe_lock);
4276
4277         /*
4278          * BATCH_READY is cleared, no new stripes can be added.
4279          * batch_list can be accessed without lock
4280          */
4281         return 0;
4282 }
4283
4284 static void break_stripe_batch_list(struct stripe_head *head_sh,
4285                                     unsigned long handle_flags)
4286 {
4287         struct stripe_head *sh, *next;
4288         int i;
4289         int do_wakeup = 0;
4290
4291         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4292
4293                 list_del_init(&sh->batch_list);
4294
4295                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4296                                           (1 << STRIPE_SYNCING) |
4297                                           (1 << STRIPE_REPLACED) |
4298                                           (1 << STRIPE_DELAYED) |
4299                                           (1 << STRIPE_BIT_DELAY) |
4300                                           (1 << STRIPE_FULL_WRITE) |
4301                                           (1 << STRIPE_BIOFILL_RUN) |
4302                                           (1 << STRIPE_COMPUTE_RUN)  |
4303                                           (1 << STRIPE_OPS_REQ_PENDING) |
4304                                           (1 << STRIPE_DISCARD) |
4305                                           (1 << STRIPE_BATCH_READY) |
4306                                           (1 << STRIPE_BATCH_ERR) |
4307                                           (1 << STRIPE_BITMAP_PENDING)),
4308                         "stripe state: %lx\n", sh->state);
4309                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4310                                               (1 << STRIPE_REPLACED)),
4311                         "head stripe state: %lx\n", head_sh->state);
4312
4313                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4314                                             (1 << STRIPE_PREREAD_ACTIVE) |
4315                                             (1 << STRIPE_DEGRADED) |
4316                                             (1 << STRIPE_ON_UNPLUG_LIST)),
4317                               head_sh->state & (1 << STRIPE_INSYNC));
4318
4319                 sh->check_state = head_sh->check_state;
4320                 sh->reconstruct_state = head_sh->reconstruct_state;
4321                 for (i = 0; i < sh->disks; i++) {
4322                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4323                                 do_wakeup = 1;
4324                         sh->dev[i].flags = head_sh->dev[i].flags &
4325                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4326                 }
4327                 spin_lock_irq(&sh->stripe_lock);
4328                 sh->batch_head = NULL;
4329                 spin_unlock_irq(&sh->stripe_lock);
4330                 if (handle_flags == 0 ||
4331                     sh->state & handle_flags)
4332                         set_bit(STRIPE_HANDLE, &sh->state);
4333                 raid5_release_stripe(sh);
4334         }
4335         spin_lock_irq(&head_sh->stripe_lock);
4336         head_sh->batch_head = NULL;
4337         spin_unlock_irq(&head_sh->stripe_lock);
4338         for (i = 0; i < head_sh->disks; i++)
4339                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4340                         do_wakeup = 1;
4341         if (head_sh->state & handle_flags)
4342                 set_bit(STRIPE_HANDLE, &head_sh->state);
4343
4344         if (do_wakeup)
4345                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4346 }
4347
4348 static void handle_stripe(struct stripe_head *sh)
4349 {
4350         struct stripe_head_state s;
4351         struct r5conf *conf = sh->raid_conf;
4352         int i;
4353         int prexor;
4354         int disks = sh->disks;
4355         struct r5dev *pdev, *qdev;
4356
4357         clear_bit(STRIPE_HANDLE, &sh->state);
4358         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4359                 /* already being handled, ensure it gets handled
4360                  * again when current action finishes */
4361                 set_bit(STRIPE_HANDLE, &sh->state);
4362                 return;
4363         }
4364
4365         if (clear_batch_ready(sh) ) {
4366                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4367                 return;
4368         }
4369
4370         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4371                 break_stripe_batch_list(sh, 0);
4372
4373         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4374                 spin_lock(&sh->stripe_lock);
4375                 /* Cannot process 'sync' concurrently with 'discard' */
4376                 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4377                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4378                         set_bit(STRIPE_SYNCING, &sh->state);
4379                         clear_bit(STRIPE_INSYNC, &sh->state);
4380                         clear_bit(STRIPE_REPLACED, &sh->state);
4381                 }
4382                 spin_unlock(&sh->stripe_lock);
4383         }
4384         clear_bit(STRIPE_DELAYED, &sh->state);
4385
4386         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4387                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4388                (unsigned long long)sh->sector, sh->state,
4389                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4390                sh->check_state, sh->reconstruct_state);
4391
4392         analyse_stripe(sh, &s);
4393
4394         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4395                 goto finish;
4396
4397         if (s.handle_bad_blocks) {
4398                 set_bit(STRIPE_HANDLE, &sh->state);
4399                 goto finish;
4400         }
4401
4402         if (unlikely(s.blocked_rdev)) {
4403                 if (s.syncing || s.expanding || s.expanded ||
4404                     s.replacing || s.to_write || s.written) {
4405                         set_bit(STRIPE_HANDLE, &sh->state);
4406                         goto finish;
4407                 }
4408                 /* There is nothing for the blocked_rdev to block */
4409                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4410                 s.blocked_rdev = NULL;
4411         }
4412
4413         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4414                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4415                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4416         }
4417
4418         pr_debug("locked=%d uptodate=%d to_read=%d"
4419                " to_write=%d failed=%d failed_num=%d,%d\n",
4420                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4421                s.failed_num[0], s.failed_num[1]);
4422         /* check if the array has lost more than max_degraded devices and,
4423          * if so, some requests might need to be failed.
4424          */
4425         if (s.failed > conf->max_degraded || s.log_failed) {
4426                 sh->check_state = 0;
4427                 sh->reconstruct_state = 0;
4428                 break_stripe_batch_list(sh, 0);
4429                 if (s.to_read+s.to_write+s.written)
4430                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4431                 if (s.syncing + s.replacing)
4432                         handle_failed_sync(conf, sh, &s);
4433         }
4434
4435         /* Now we check to see if any write operations have recently
4436          * completed
4437          */
4438         prexor = 0;
4439         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4440                 prexor = 1;
4441         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4442             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4443                 sh->reconstruct_state = reconstruct_state_idle;
4444
4445                 /* All the 'written' buffers and the parity block are ready to
4446                  * be written back to disk
4447                  */
4448                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4449                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4450                 BUG_ON(sh->qd_idx >= 0 &&
4451                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4452                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4453                 for (i = disks; i--; ) {
4454                         struct r5dev *dev = &sh->dev[i];
4455                         if (test_bit(R5_LOCKED, &dev->flags) &&
4456                                 (i == sh->pd_idx || i == sh->qd_idx ||
4457                                  dev->written)) {
4458                                 pr_debug("Writing block %d\n", i);
4459                                 set_bit(R5_Wantwrite, &dev->flags);
4460                                 if (prexor)
4461                                         continue;
4462                                 if (s.failed > 1)
4463                                         continue;
4464                                 if (!test_bit(R5_Insync, &dev->flags) ||
4465                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4466                                      s.failed == 0))
4467                                         set_bit(STRIPE_INSYNC, &sh->state);
4468                         }
4469                 }
4470                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4471                         s.dec_preread_active = 1;
4472         }
4473
4474         /*
4475          * might be able to return some write requests if the parity blocks
4476          * are safe, or on a failed drive
4477          */
4478         pdev = &sh->dev[sh->pd_idx];
4479         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4480                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4481         qdev = &sh->dev[sh->qd_idx];
4482         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4483                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4484                 || conf->level < 6;
4485
4486         if (s.written &&
4487             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4488                              && !test_bit(R5_LOCKED, &pdev->flags)
4489                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4490                                  test_bit(R5_Discard, &pdev->flags))))) &&
4491             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4492                              && !test_bit(R5_LOCKED, &qdev->flags)
4493                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4494                                  test_bit(R5_Discard, &qdev->flags))))))
4495                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4496
4497         /* Now we might consider reading some blocks, either to check/generate
4498          * parity, or to satisfy requests
4499          * or to load a block that is being partially written.
4500          */
4501         if (s.to_read || s.non_overwrite
4502             || (s.to_write && s.failed)
4503             || (s.syncing && (s.uptodate + s.compute < disks))
4504             || s.replacing
4505             || s.expanding)
4506                 handle_stripe_fill(sh, &s, disks);
4507
4508         /* Now to consider new write requests and what else, if anything
4509          * should be read.  We do not handle new writes when:
4510          * 1/ A 'write' operation (copy+xor) is already in flight.
4511          * 2/ A 'check' operation is in flight, as it may clobber the parity
4512          *    block.
4513          */
4514         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4515                 handle_stripe_dirtying(conf, sh, &s, disks);
4516
4517         /* maybe we need to check and possibly fix the parity for this stripe
4518          * Any reads will already have been scheduled, so we just see if enough
4519          * data is available.  The parity check is held off while parity
4520          * dependent operations are in flight.
4521          */
4522         if (sh->check_state ||
4523             (s.syncing && s.locked == 0 &&
4524              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4525              !test_bit(STRIPE_INSYNC, &sh->state))) {
4526                 if (conf->level == 6)
4527                         handle_parity_checks6(conf, sh, &s, disks);
4528                 else
4529                         handle_parity_checks5(conf, sh, &s, disks);
4530         }
4531
4532         if ((s.replacing || s.syncing) && s.locked == 0
4533             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4534             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4535                 /* Write out to replacement devices where possible */
4536                 for (i = 0; i < conf->raid_disks; i++)
4537                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4538                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4539                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4540                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4541                                 s.locked++;
4542                         }
4543                 if (s.replacing)
4544                         set_bit(STRIPE_INSYNC, &sh->state);
4545                 set_bit(STRIPE_REPLACED, &sh->state);
4546         }
4547         if ((s.syncing || s.replacing) && s.locked == 0 &&
4548             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4549             test_bit(STRIPE_INSYNC, &sh->state)) {
4550                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4551                 clear_bit(STRIPE_SYNCING, &sh->state);
4552                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4553                         wake_up(&conf->wait_for_overlap);
4554         }
4555
4556         /* If the failed drives are just a ReadError, then we might need
4557          * to progress the repair/check process
4558          */
4559         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4560                 for (i = 0; i < s.failed; i++) {
4561                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4562                         if (test_bit(R5_ReadError, &dev->flags)
4563                             && !test_bit(R5_LOCKED, &dev->flags)
4564                             && test_bit(R5_UPTODATE, &dev->flags)
4565                                 ) {
4566                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4567                                         set_bit(R5_Wantwrite, &dev->flags);
4568                                         set_bit(R5_ReWrite, &dev->flags);
4569                                         set_bit(R5_LOCKED, &dev->flags);
4570                                         s.locked++;
4571                                 } else {
4572                                         /* let's read it back */
4573                                         set_bit(R5_Wantread, &dev->flags);
4574                                         set_bit(R5_LOCKED, &dev->flags);
4575                                         s.locked++;
4576                                 }
4577                         }
4578                 }
4579
4580         /* Finish reconstruct operations initiated by the expansion process */
4581         if (sh->reconstruct_state == reconstruct_state_result) {
4582                 struct stripe_head *sh_src
4583                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4584                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4585                         /* sh cannot be written until sh_src has been read.
4586                          * so arrange for sh to be delayed a little
4587                          */
4588                         set_bit(STRIPE_DELAYED, &sh->state);
4589                         set_bit(STRIPE_HANDLE, &sh->state);
4590                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4591                                               &sh_src->state))
4592                                 atomic_inc(&conf->preread_active_stripes);
4593                         raid5_release_stripe(sh_src);
4594                         goto finish;
4595                 }
4596                 if (sh_src)
4597                         raid5_release_stripe(sh_src);
4598
4599                 sh->reconstruct_state = reconstruct_state_idle;
4600                 clear_bit(STRIPE_EXPANDING, &sh->state);
4601                 for (i = conf->raid_disks; i--; ) {
4602                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4603                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4604                         s.locked++;
4605                 }
4606         }
4607
4608         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4609             !sh->reconstruct_state) {
4610                 /* Need to write out all blocks after computing parity */
4611                 sh->disks = conf->raid_disks;
4612                 stripe_set_idx(sh->sector, conf, 0, sh);
4613                 schedule_reconstruction(sh, &s, 1, 1);
4614         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4615                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4616                 atomic_dec(&conf->reshape_stripes);
4617                 wake_up(&conf->wait_for_overlap);
4618                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4619         }
4620
4621         if (s.expanding && s.locked == 0 &&
4622             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4623                 handle_stripe_expansion(conf, sh);
4624
4625 finish:
4626         /* wait for this device to become unblocked */
4627         if (unlikely(s.blocked_rdev)) {
4628                 if (conf->mddev->external)
4629                         md_wait_for_blocked_rdev(s.blocked_rdev,
4630                                                  conf->mddev);
4631                 else
4632                         /* Internal metadata will immediately
4633                          * be written by raid5d, so we don't
4634                          * need to wait here.
4635                          */
4636                         rdev_dec_pending(s.blocked_rdev,
4637                                          conf->mddev);
4638         }
4639
4640         if (s.handle_bad_blocks)
4641                 for (i = disks; i--; ) {
4642                         struct md_rdev *rdev;
4643                         struct r5dev *dev = &sh->dev[i];
4644                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4645                                 /* We own a safe reference to the rdev */
4646                                 rdev = conf->disks[i].rdev;
4647                                 if (!rdev_set_badblocks(rdev, sh->sector,
4648                                                         STRIPE_SECTORS, 0))
4649                                         md_error(conf->mddev, rdev);
4650                                 rdev_dec_pending(rdev, conf->mddev);
4651                         }
4652                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4653                                 rdev = conf->disks[i].rdev;
4654                                 rdev_clear_badblocks(rdev, sh->sector,
4655                                                      STRIPE_SECTORS, 0);
4656                                 rdev_dec_pending(rdev, conf->mddev);
4657                         }
4658                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4659                                 rdev = conf->disks[i].replacement;
4660                                 if (!rdev)
4661                                         /* rdev have been moved down */
4662                                         rdev = conf->disks[i].rdev;
4663                                 rdev_clear_badblocks(rdev, sh->sector,
4664                                                      STRIPE_SECTORS, 0);
4665                                 rdev_dec_pending(rdev, conf->mddev);
4666                         }
4667                 }
4668
4669         if (s.ops_request)
4670                 raid_run_ops(sh, s.ops_request);
4671
4672         ops_run_io(sh, &s);
4673
4674         if (s.dec_preread_active) {
4675                 /* We delay this until after ops_run_io so that if make_request
4676                  * is waiting on a flush, it won't continue until the writes
4677                  * have actually been submitted.
4678                  */
4679                 atomic_dec(&conf->preread_active_stripes);
4680                 if (atomic_read(&conf->preread_active_stripes) <
4681                     IO_THRESHOLD)
4682                         md_wakeup_thread(conf->mddev->thread);
4683         }
4684
4685         if (!bio_list_empty(&s.return_bi)) {
4686                 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags) &&
4687                                 (s.failed <= conf->max_degraded ||
4688                                         conf->mddev->external == 0)) {
4689                         spin_lock_irq(&conf->device_lock);
4690                         bio_list_merge(&conf->return_bi, &s.return_bi);
4691                         spin_unlock_irq(&conf->device_lock);
4692                         md_wakeup_thread(conf->mddev->thread);
4693                 } else
4694                         return_io(&s.return_bi);
4695         }
4696
4697         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4698 }
4699
4700 static void raid5_activate_delayed(struct r5conf *conf)
4701 {
4702         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4703                 while (!list_empty(&conf->delayed_list)) {
4704                         struct list_head *l = conf->delayed_list.next;
4705                         struct stripe_head *sh;
4706                         sh = list_entry(l, struct stripe_head, lru);
4707                         list_del_init(l);
4708                         clear_bit(STRIPE_DELAYED, &sh->state);
4709                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4710                                 atomic_inc(&conf->preread_active_stripes);
4711                         list_add_tail(&sh->lru, &conf->hold_list);
4712                         raid5_wakeup_stripe_thread(sh);
4713                 }
4714         }
4715 }
4716
4717 static void activate_bit_delay(struct r5conf *conf,
4718         struct list_head *temp_inactive_list)
4719 {
4720         /* device_lock is held */
4721         struct list_head head;
4722         list_add(&head, &conf->bitmap_list);
4723         list_del_init(&conf->bitmap_list);
4724         while (!list_empty(&head)) {
4725                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4726                 int hash;
4727                 list_del_init(&sh->lru);
4728                 atomic_inc(&sh->count);
4729                 hash = sh->hash_lock_index;
4730                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4731         }
4732 }
4733
4734 static int raid5_congested(struct mddev *mddev, int bits)
4735 {
4736         struct r5conf *conf = mddev->private;
4737
4738         /* No difference between reads and writes.  Just check
4739          * how busy the stripe_cache is
4740          */
4741
4742         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4743                 return 1;
4744         if (conf->quiesce)
4745                 return 1;
4746         if (atomic_read(&conf->empty_inactive_list_nr))
4747                 return 1;
4748
4749         return 0;
4750 }
4751
4752 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4753 {
4754         struct r5conf *conf = mddev->private;
4755         sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4756         unsigned int chunk_sectors;
4757         unsigned int bio_sectors = bio_sectors(bio);
4758
4759         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4760         return  chunk_sectors >=
4761                 ((sector & (chunk_sectors - 1)) + bio_sectors);
4762 }
4763
4764 /*
4765  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
4766  *  later sampled by raid5d.
4767  */
4768 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4769 {
4770         unsigned long flags;
4771
4772         spin_lock_irqsave(&conf->device_lock, flags);
4773
4774         bi->bi_next = conf->retry_read_aligned_list;
4775         conf->retry_read_aligned_list = bi;
4776
4777         spin_unlock_irqrestore(&conf->device_lock, flags);
4778         md_wakeup_thread(conf->mddev->thread);
4779 }
4780
4781 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4782 {
4783         struct bio *bi;
4784
4785         bi = conf->retry_read_aligned;
4786         if (bi) {
4787                 conf->retry_read_aligned = NULL;
4788                 return bi;
4789         }
4790         bi = conf->retry_read_aligned_list;
4791         if(bi) {
4792                 conf->retry_read_aligned_list = bi->bi_next;
4793                 bi->bi_next = NULL;
4794                 /*
4795                  * this sets the active strip count to 1 and the processed
4796                  * strip count to zero (upper 8 bits)
4797                  */
4798                 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4799         }
4800
4801         return bi;
4802 }
4803
4804 /*
4805  *  The "raid5_align_endio" should check if the read succeeded and if it
4806  *  did, call bio_endio on the original bio (having bio_put the new bio
4807  *  first).
4808  *  If the read failed..
4809  */
4810 static void raid5_align_endio(struct bio *bi)
4811 {
4812         struct bio* raid_bi  = bi->bi_private;
4813         struct mddev *mddev;
4814         struct r5conf *conf;
4815         struct md_rdev *rdev;
4816         int error = bi->bi_error;
4817
4818         bio_put(bi);
4819
4820         rdev = (void*)raid_bi->bi_next;
4821         raid_bi->bi_next = NULL;
4822         mddev = rdev->mddev;
4823         conf = mddev->private;
4824
4825         rdev_dec_pending(rdev, conf->mddev);
4826
4827         if (!error) {
4828                 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4829                                          raid_bi, 0);
4830                 bio_endio(raid_bi);
4831                 if (atomic_dec_and_test(&conf->active_aligned_reads))
4832                         wake_up(&conf->wait_for_quiescent);
4833                 return;
4834         }
4835
4836         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4837
4838         add_bio_to_retry(raid_bi, conf);
4839 }
4840
4841 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4842 {
4843         struct r5conf *conf = mddev->private;
4844         int dd_idx;
4845         struct bio* align_bi;
4846         struct md_rdev *rdev;
4847         sector_t end_sector;
4848
4849         if (!in_chunk_boundary(mddev, raid_bio)) {
4850                 pr_debug("%s: non aligned\n", __func__);
4851                 return 0;
4852         }
4853         /*
4854          * use bio_clone_mddev to make a copy of the bio
4855          */
4856         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4857         if (!align_bi)
4858                 return 0;
4859         /*
4860          *   set bi_end_io to a new function, and set bi_private to the
4861          *     original bio.
4862          */
4863         align_bi->bi_end_io  = raid5_align_endio;
4864         align_bi->bi_private = raid_bio;
4865         /*
4866          *      compute position
4867          */
4868         align_bi->bi_iter.bi_sector =
4869                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4870                                      0, &dd_idx, NULL);
4871
4872         end_sector = bio_end_sector(align_bi);
4873         rcu_read_lock();
4874         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4875         if (!rdev || test_bit(Faulty, &rdev->flags) ||
4876             rdev->recovery_offset < end_sector) {
4877                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4878                 if (rdev &&
4879                     (test_bit(Faulty, &rdev->flags) ||
4880                     !(test_bit(In_sync, &rdev->flags) ||
4881                       rdev->recovery_offset >= end_sector)))
4882                         rdev = NULL;
4883         }
4884         if (rdev) {
4885                 sector_t first_bad;
4886                 int bad_sectors;
4887
4888                 atomic_inc(&rdev->nr_pending);
4889                 rcu_read_unlock();
4890                 raid_bio->bi_next = (void*)rdev;
4891                 align_bi->bi_bdev =  rdev->bdev;
4892                 bio_clear_flag(align_bi, BIO_SEG_VALID);
4893
4894                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4895                                 bio_sectors(align_bi),
4896                                 &first_bad, &bad_sectors)) {
4897                         bio_put(align_bi);
4898                         rdev_dec_pending(rdev, mddev);
4899                         return 0;
4900                 }
4901
4902                 /* No reshape active, so we can trust rdev->data_offset */
4903                 align_bi->bi_iter.bi_sector += rdev->data_offset;
4904
4905                 spin_lock_irq(&conf->device_lock);
4906                 wait_event_lock_irq(conf->wait_for_quiescent,
4907                                     conf->quiesce == 0,
4908                                     conf->device_lock);
4909                 atomic_inc(&conf->active_aligned_reads);
4910                 spin_unlock_irq(&conf->device_lock);
4911
4912                 if (mddev->gendisk)
4913                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4914                                               align_bi, disk_devt(mddev->gendisk),
4915                                               raid_bio->bi_iter.bi_sector);
4916                 generic_make_request(align_bi);
4917                 return 1;
4918         } else {
4919                 rcu_read_unlock();
4920                 bio_put(align_bi);
4921                 return 0;
4922         }
4923 }
4924
4925 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4926 {
4927         struct bio *split;
4928
4929         do {
4930                 sector_t sector = raid_bio->bi_iter.bi_sector;
4931                 unsigned chunk_sects = mddev->chunk_sectors;
4932                 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4933
4934                 if (sectors < bio_sectors(raid_bio)) {
4935                         split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4936                         bio_chain(split, raid_bio);
4937                 } else
4938                         split = raid_bio;
4939
4940                 if (!raid5_read_one_chunk(mddev, split)) {
4941                         if (split != raid_bio)
4942                                 generic_make_request(raid_bio);
4943                         return split;
4944                 }
4945         } while (split != raid_bio);
4946
4947         return NULL;
4948 }
4949
4950 /* __get_priority_stripe - get the next stripe to process
4951  *
4952  * Full stripe writes are allowed to pass preread active stripes up until
4953  * the bypass_threshold is exceeded.  In general the bypass_count
4954  * increments when the handle_list is handled before the hold_list; however, it
4955  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4956  * stripe with in flight i/o.  The bypass_count will be reset when the
4957  * head of the hold_list has changed, i.e. the head was promoted to the
4958  * handle_list.
4959  */
4960 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4961 {
4962         struct stripe_head *sh = NULL, *tmp;
4963         struct list_head *handle_list = NULL;
4964         struct r5worker_group *wg = NULL;
4965
4966         if (conf->worker_cnt_per_group == 0) {
4967                 handle_list = &conf->handle_list;
4968         } else if (group != ANY_GROUP) {
4969                 handle_list = &conf->worker_groups[group].handle_list;
4970                 wg = &conf->worker_groups[group];
4971         } else {
4972                 int i;
4973                 for (i = 0; i < conf->group_cnt; i++) {
4974                         handle_list = &conf->worker_groups[i].handle_list;
4975                         wg = &conf->worker_groups[i];
4976                         if (!list_empty(handle_list))
4977                                 break;
4978                 }
4979         }
4980
4981         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4982                   __func__,
4983                   list_empty(handle_list) ? "empty" : "busy",
4984                   list_empty(&conf->hold_list) ? "empty" : "busy",
4985                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
4986
4987         if (!list_empty(handle_list)) {
4988                 sh = list_entry(handle_list->next, typeof(*sh), lru);
4989
4990                 if (list_empty(&conf->hold_list))
4991                         conf->bypass_count = 0;
4992                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4993                         if (conf->hold_list.next == conf->last_hold)
4994                                 conf->bypass_count++;
4995                         else {
4996                                 conf->last_hold = conf->hold_list.next;
4997                                 conf->bypass_count -= conf->bypass_threshold;
4998                                 if (conf->bypass_count < 0)
4999                                         conf->bypass_count = 0;
5000                         }
5001                 }
5002         } else if (!list_empty(&conf->hold_list) &&
5003                    ((conf->bypass_threshold &&
5004                      conf->bypass_count > conf->bypass_threshold) ||
5005                     atomic_read(&conf->pending_full_writes) == 0)) {
5006
5007                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5008                         if (conf->worker_cnt_per_group == 0 ||
5009                             group == ANY_GROUP ||
5010                             !cpu_online(tmp->cpu) ||
5011                             cpu_to_group(tmp->cpu) == group) {
5012                                 sh = tmp;
5013                                 break;
5014                         }
5015                 }
5016
5017                 if (sh) {
5018                         conf->bypass_count -= conf->bypass_threshold;
5019                         if (conf->bypass_count < 0)
5020                                 conf->bypass_count = 0;
5021                 }
5022                 wg = NULL;
5023         }
5024
5025         if (!sh)
5026                 return NULL;
5027
5028         if (wg) {
5029                 wg->stripes_cnt--;
5030                 sh->group = NULL;
5031         }
5032         list_del_init(&sh->lru);
5033         BUG_ON(atomic_inc_return(&sh->count) != 1);
5034         return sh;
5035 }
5036
5037 struct raid5_plug_cb {
5038         struct blk_plug_cb      cb;
5039         struct list_head        list;
5040         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5041 };
5042
5043 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5044 {
5045         struct raid5_plug_cb *cb = container_of(
5046                 blk_cb, struct raid5_plug_cb, cb);
5047         struct stripe_head *sh;
5048         struct mddev *mddev = cb->cb.data;
5049         struct r5conf *conf = mddev->private;
5050         int cnt = 0;
5051         int hash;
5052
5053         if (cb->list.next && !list_empty(&cb->list)) {
5054                 spin_lock_irq(&conf->device_lock);
5055                 while (!list_empty(&cb->list)) {
5056                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5057                         list_del_init(&sh->lru);
5058                         /*
5059                          * avoid race release_stripe_plug() sees
5060                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5061                          * is still in our list
5062                          */
5063                         smp_mb__before_atomic();
5064                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5065                         /*
5066                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5067                          * case, the count is always > 1 here
5068                          */
5069                         hash = sh->hash_lock_index;
5070                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5071                         cnt++;
5072                 }
5073                 spin_unlock_irq(&conf->device_lock);
5074         }
5075         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5076                                      NR_STRIPE_HASH_LOCKS);
5077         if (mddev->queue)
5078                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5079         kfree(cb);
5080 }
5081
5082 static void release_stripe_plug(struct mddev *mddev,
5083                                 struct stripe_head *sh)
5084 {
5085         struct blk_plug_cb *blk_cb = blk_check_plugged(
5086                 raid5_unplug, mddev,
5087                 sizeof(struct raid5_plug_cb));
5088         struct raid5_plug_cb *cb;
5089
5090         if (!blk_cb) {
5091                 raid5_release_stripe(sh);
5092                 return;
5093         }
5094
5095         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5096
5097         if (cb->list.next == NULL) {
5098                 int i;
5099                 INIT_LIST_HEAD(&cb->list);
5100                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5101                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5102         }
5103
5104         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5105                 list_add_tail(&sh->lru, &cb->list);
5106         else
5107                 raid5_release_stripe(sh);
5108 }
5109
5110 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5111 {
5112         struct r5conf *conf = mddev->private;
5113         sector_t logical_sector, last_sector;
5114         struct stripe_head *sh;
5115         int remaining;
5116         int stripe_sectors;
5117
5118         if (mddev->reshape_position != MaxSector)
5119                 /* Skip discard while reshape is happening */
5120                 return;
5121
5122         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5123         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5124
5125         bi->bi_next = NULL;
5126         bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5127
5128         stripe_sectors = conf->chunk_sectors *
5129                 (conf->raid_disks - conf->max_degraded);
5130         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5131                                                stripe_sectors);
5132         sector_div(last_sector, stripe_sectors);
5133
5134         logical_sector *= conf->chunk_sectors;
5135         last_sector *= conf->chunk_sectors;
5136
5137         for (; logical_sector < last_sector;
5138              logical_sector += STRIPE_SECTORS) {
5139                 DEFINE_WAIT(w);
5140                 int d;
5141         again:
5142                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5143                 prepare_to_wait(&conf->wait_for_overlap, &w,
5144                                 TASK_UNINTERRUPTIBLE);
5145                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5146                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5147                         raid5_release_stripe(sh);
5148                         schedule();
5149                         goto again;
5150                 }
5151                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5152                 spin_lock_irq(&sh->stripe_lock);
5153                 for (d = 0; d < conf->raid_disks; d++) {
5154                         if (d == sh->pd_idx || d == sh->qd_idx)
5155                                 continue;
5156                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5157                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5158                                 spin_unlock_irq(&sh->stripe_lock);
5159                                 raid5_release_stripe(sh);
5160                                 schedule();
5161                                 goto again;
5162                         }
5163                 }
5164                 set_bit(STRIPE_DISCARD, &sh->state);
5165                 finish_wait(&conf->wait_for_overlap, &w);
5166                 sh->overwrite_disks = 0;
5167                 for (d = 0; d < conf->raid_disks; d++) {
5168                         if (d == sh->pd_idx || d == sh->qd_idx)
5169                                 continue;
5170                         sh->dev[d].towrite = bi;
5171                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5172                         raid5_inc_bi_active_stripes(bi);
5173                         sh->overwrite_disks++;
5174                 }
5175                 spin_unlock_irq(&sh->stripe_lock);
5176                 if (conf->mddev->bitmap) {
5177                         for (d = 0;
5178                              d < conf->raid_disks - conf->max_degraded;
5179                              d++)
5180                                 bitmap_startwrite(mddev->bitmap,
5181                                                   sh->sector,
5182                                                   STRIPE_SECTORS,
5183                                                   0);
5184                         sh->bm_seq = conf->seq_flush + 1;
5185                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5186                 }
5187
5188                 set_bit(STRIPE_HANDLE, &sh->state);
5189                 clear_bit(STRIPE_DELAYED, &sh->state);
5190                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5191                         atomic_inc(&conf->preread_active_stripes);
5192                 release_stripe_plug(mddev, sh);
5193         }
5194
5195         remaining = raid5_dec_bi_active_stripes(bi);
5196         if (remaining == 0) {
5197                 md_write_end(mddev);
5198                 bio_endio(bi);
5199         }
5200 }
5201
5202 static void raid5_make_request(struct mddev *mddev, struct bio * bi)
5203 {
5204         struct r5conf *conf = mddev->private;
5205         int dd_idx;
5206         sector_t new_sector;
5207         sector_t logical_sector, last_sector;
5208         struct stripe_head *sh;
5209         const int rw = bio_data_dir(bi);
5210         int remaining;
5211         DEFINE_WAIT(w);
5212         bool do_prepare;
5213
5214         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5215                 int ret = r5l_handle_flush_request(conf->log, bi);
5216
5217                 if (ret == 0)
5218                         return;
5219                 if (ret == -ENODEV) {
5220                         md_flush_request(mddev, bi);
5221                         return;
5222                 }
5223                 /* ret == -EAGAIN, fallback */
5224         }
5225
5226         md_write_start(mddev, bi);
5227
5228         /*
5229          * If array is degraded, better not do chunk aligned read because
5230          * later we might have to read it again in order to reconstruct
5231          * data on failed drives.
5232          */
5233         if (rw == READ && mddev->degraded == 0 &&
5234             mddev->reshape_position == MaxSector) {
5235                 bi = chunk_aligned_read(mddev, bi);
5236                 if (!bi)
5237                         return;
5238         }
5239
5240         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5241                 make_discard_request(mddev, bi);
5242                 return;
5243         }
5244
5245         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5246         last_sector = bio_end_sector(bi);
5247         bi->bi_next = NULL;
5248         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
5249
5250         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5251         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5252                 int previous;
5253                 int seq;
5254
5255                 do_prepare = false;
5256         retry:
5257                 seq = read_seqcount_begin(&conf->gen_lock);
5258                 previous = 0;
5259                 if (do_prepare)
5260                         prepare_to_wait(&conf->wait_for_overlap, &w,
5261                                 TASK_UNINTERRUPTIBLE);
5262                 if (unlikely(conf->reshape_progress != MaxSector)) {
5263                         /* spinlock is needed as reshape_progress may be
5264                          * 64bit on a 32bit platform, and so it might be
5265                          * possible to see a half-updated value
5266                          * Of course reshape_progress could change after
5267                          * the lock is dropped, so once we get a reference
5268                          * to the stripe that we think it is, we will have
5269                          * to check again.
5270                          */
5271                         spin_lock_irq(&conf->device_lock);
5272                         if (mddev->reshape_backwards
5273                             ? logical_sector < conf->reshape_progress
5274                             : logical_sector >= conf->reshape_progress) {
5275                                 previous = 1;
5276                         } else {
5277                                 if (mddev->reshape_backwards
5278                                     ? logical_sector < conf->reshape_safe
5279                                     : logical_sector >= conf->reshape_safe) {
5280                                         spin_unlock_irq(&conf->device_lock);
5281                                         schedule();
5282                                         do_prepare = true;
5283                                         goto retry;
5284                                 }
5285                         }
5286                         spin_unlock_irq(&conf->device_lock);
5287                 }
5288
5289                 new_sector = raid5_compute_sector(conf, logical_sector,
5290                                                   previous,
5291                                                   &dd_idx, NULL);
5292                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5293                         (unsigned long long)new_sector,
5294                         (unsigned long long)logical_sector);
5295
5296                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5297                                        (bi->bi_opf & REQ_RAHEAD), 0);
5298                 if (sh) {
5299                         if (unlikely(previous)) {
5300                                 /* expansion might have moved on while waiting for a
5301                                  * stripe, so we must do the range check again.
5302                                  * Expansion could still move past after this
5303                                  * test, but as we are holding a reference to
5304                                  * 'sh', we know that if that happens,
5305                                  *  STRIPE_EXPANDING will get set and the expansion
5306                                  * won't proceed until we finish with the stripe.
5307                                  */
5308                                 int must_retry = 0;
5309                                 spin_lock_irq(&conf->device_lock);
5310                                 if (mddev->reshape_backwards
5311                                     ? logical_sector >= conf->reshape_progress
5312                                     : logical_sector < conf->reshape_progress)
5313                                         /* mismatch, need to try again */
5314                                         must_retry = 1;
5315                                 spin_unlock_irq(&conf->device_lock);
5316                                 if (must_retry) {
5317                                         raid5_release_stripe(sh);
5318                                         schedule();
5319                                         do_prepare = true;
5320                                         goto retry;
5321                                 }
5322                         }
5323                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5324                                 /* Might have got the wrong stripe_head
5325                                  * by accident
5326                                  */
5327                                 raid5_release_stripe(sh);
5328                                 goto retry;
5329                         }
5330
5331                         if (rw == WRITE &&
5332                             logical_sector >= mddev->suspend_lo &&
5333                             logical_sector < mddev->suspend_hi) {
5334                                 raid5_release_stripe(sh);
5335                                 /* As the suspend_* range is controlled by
5336                                  * userspace, we want an interruptible
5337                                  * wait.
5338                                  */
5339                                 prepare_to_wait(&conf->wait_for_overlap,
5340                                                 &w, TASK_INTERRUPTIBLE);
5341                                 if (logical_sector >= mddev->suspend_lo &&
5342                                     logical_sector < mddev->suspend_hi) {
5343                                         sigset_t full, old;
5344                                         sigfillset(&full);
5345                                         sigprocmask(SIG_BLOCK, &full, &old);
5346                                         schedule();
5347                                         sigprocmask(SIG_SETMASK, &old, NULL);
5348                                         do_prepare = true;
5349                                 }
5350                                 goto retry;
5351                         }
5352
5353                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5354                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5355                                 /* Stripe is busy expanding or
5356                                  * add failed due to overlap.  Flush everything
5357                                  * and wait a while
5358                                  */
5359                                 md_wakeup_thread(mddev->thread);
5360                                 raid5_release_stripe(sh);
5361                                 schedule();
5362                                 do_prepare = true;
5363                                 goto retry;
5364                         }
5365                         set_bit(STRIPE_HANDLE, &sh->state);
5366                         clear_bit(STRIPE_DELAYED, &sh->state);
5367                         if ((!sh->batch_head || sh == sh->batch_head) &&
5368                             (bi->bi_opf & REQ_SYNC) &&
5369                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5370                                 atomic_inc(&conf->preread_active_stripes);
5371                         release_stripe_plug(mddev, sh);
5372                 } else {
5373                         /* cannot get stripe for read-ahead, just give-up */
5374                         bi->bi_error = -EIO;
5375                         break;
5376                 }
5377         }
5378         finish_wait(&conf->wait_for_overlap, &w);
5379
5380         remaining = raid5_dec_bi_active_stripes(bi);
5381         if (remaining == 0) {
5382
5383                 if ( rw == WRITE )
5384                         md_write_end(mddev);
5385
5386                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5387                                          bi, 0);
5388                 bio_endio(bi);
5389         }
5390 }
5391
5392 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5393
5394 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5395 {
5396         /* reshaping is quite different to recovery/resync so it is
5397          * handled quite separately ... here.
5398          *
5399          * On each call to sync_request, we gather one chunk worth of
5400          * destination stripes and flag them as expanding.
5401          * Then we find all the source stripes and request reads.
5402          * As the reads complete, handle_stripe will copy the data
5403          * into the destination stripe and release that stripe.
5404          */
5405         struct r5conf *conf = mddev->private;
5406         struct stripe_head *sh;
5407         sector_t first_sector, last_sector;
5408         int raid_disks = conf->previous_raid_disks;
5409         int data_disks = raid_disks - conf->max_degraded;
5410         int new_data_disks = conf->raid_disks - conf->max_degraded;
5411         int i;
5412         int dd_idx;
5413         sector_t writepos, readpos, safepos;
5414         sector_t stripe_addr;
5415         int reshape_sectors;
5416         struct list_head stripes;
5417         sector_t retn;
5418
5419         if (sector_nr == 0) {
5420                 /* If restarting in the middle, skip the initial sectors */
5421                 if (mddev->reshape_backwards &&
5422                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5423                         sector_nr = raid5_size(mddev, 0, 0)
5424                                 - conf->reshape_progress;
5425                 } else if (mddev->reshape_backwards &&
5426                            conf->reshape_progress == MaxSector) {
5427                         /* shouldn't happen, but just in case, finish up.*/
5428                         sector_nr = MaxSector;
5429                 } else if (!mddev->reshape_backwards &&
5430                            conf->reshape_progress > 0)
5431                         sector_nr = conf->reshape_progress;
5432                 sector_div(sector_nr, new_data_disks);
5433                 if (sector_nr) {
5434                         mddev->curr_resync_completed = sector_nr;
5435                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5436                         *skipped = 1;
5437                         retn = sector_nr;
5438                         goto finish;
5439                 }
5440         }
5441
5442         /* We need to process a full chunk at a time.
5443          * If old and new chunk sizes differ, we need to process the
5444          * largest of these
5445          */
5446
5447         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5448
5449         /* We update the metadata at least every 10 seconds, or when
5450          * the data about to be copied would over-write the source of
5451          * the data at the front of the range.  i.e. one new_stripe
5452          * along from reshape_progress new_maps to after where
5453          * reshape_safe old_maps to
5454          */
5455         writepos = conf->reshape_progress;
5456         sector_div(writepos, new_data_disks);
5457         readpos = conf->reshape_progress;
5458         sector_div(readpos, data_disks);
5459         safepos = conf->reshape_safe;
5460         sector_div(safepos, data_disks);
5461         if (mddev->reshape_backwards) {
5462                 BUG_ON(writepos < reshape_sectors);
5463                 writepos -= reshape_sectors;
5464                 readpos += reshape_sectors;
5465                 safepos += reshape_sectors;
5466         } else {
5467                 writepos += reshape_sectors;
5468                 /* readpos and safepos are worst-case calculations.
5469                  * A negative number is overly pessimistic, and causes
5470                  * obvious problems for unsigned storage.  So clip to 0.
5471                  */
5472                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5473                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5474         }
5475
5476         /* Having calculated the 'writepos' possibly use it
5477          * to set 'stripe_addr' which is where we will write to.
5478          */
5479         if (mddev->reshape_backwards) {
5480                 BUG_ON(conf->reshape_progress == 0);
5481                 stripe_addr = writepos;
5482                 BUG_ON((mddev->dev_sectors &
5483                         ~((sector_t)reshape_sectors - 1))
5484                        - reshape_sectors - stripe_addr
5485                        != sector_nr);
5486         } else {
5487                 BUG_ON(writepos != sector_nr + reshape_sectors);
5488                 stripe_addr = sector_nr;
5489         }
5490
5491         /* 'writepos' is the most advanced device address we might write.
5492          * 'readpos' is the least advanced device address we might read.
5493          * 'safepos' is the least address recorded in the metadata as having
5494          *     been reshaped.
5495          * If there is a min_offset_diff, these are adjusted either by
5496          * increasing the safepos/readpos if diff is negative, or
5497          * increasing writepos if diff is positive.
5498          * If 'readpos' is then behind 'writepos', there is no way that we can
5499          * ensure safety in the face of a crash - that must be done by userspace
5500          * making a backup of the data.  So in that case there is no particular
5501          * rush to update metadata.
5502          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5503          * update the metadata to advance 'safepos' to match 'readpos' so that
5504          * we can be safe in the event of a crash.
5505          * So we insist on updating metadata if safepos is behind writepos and
5506          * readpos is beyond writepos.
5507          * In any case, update the metadata every 10 seconds.
5508          * Maybe that number should be configurable, but I'm not sure it is
5509          * worth it.... maybe it could be a multiple of safemode_delay???
5510          */
5511         if (conf->min_offset_diff < 0) {
5512                 safepos += -conf->min_offset_diff;
5513                 readpos += -conf->min_offset_diff;
5514         } else
5515                 writepos += conf->min_offset_diff;
5516
5517         if ((mddev->reshape_backwards
5518              ? (safepos > writepos && readpos < writepos)
5519              : (safepos < writepos && readpos > writepos)) ||
5520             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5521                 /* Cannot proceed until we've updated the superblock... */
5522                 wait_event(conf->wait_for_overlap,
5523                            atomic_read(&conf->reshape_stripes)==0
5524                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5525                 if (atomic_read(&conf->reshape_stripes) != 0)
5526                         return 0;
5527                 mddev->reshape_position = conf->reshape_progress;
5528                 mddev->curr_resync_completed = sector_nr;
5529                 conf->reshape_checkpoint = jiffies;
5530                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5531                 md_wakeup_thread(mddev->thread);
5532                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5533                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5534                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5535                         return 0;
5536                 spin_lock_irq(&conf->device_lock);
5537                 conf->reshape_safe = mddev->reshape_position;
5538                 spin_unlock_irq(&conf->device_lock);
5539                 wake_up(&conf->wait_for_overlap);
5540                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5541         }
5542
5543         INIT_LIST_HEAD(&stripes);
5544         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5545                 int j;
5546                 int skipped_disk = 0;
5547                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5548                 set_bit(STRIPE_EXPANDING, &sh->state);
5549                 atomic_inc(&conf->reshape_stripes);
5550                 /* If any of this stripe is beyond the end of the old
5551                  * array, then we need to zero those blocks
5552                  */
5553                 for (j=sh->disks; j--;) {
5554                         sector_t s;
5555                         if (j == sh->pd_idx)
5556                                 continue;
5557                         if (conf->level == 6 &&
5558                             j == sh->qd_idx)
5559                                 continue;
5560                         s = raid5_compute_blocknr(sh, j, 0);
5561                         if (s < raid5_size(mddev, 0, 0)) {
5562                                 skipped_disk = 1;
5563                                 continue;
5564                         }
5565                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5566                         set_bit(R5_Expanded, &sh->dev[j].flags);
5567                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5568                 }
5569                 if (!skipped_disk) {
5570                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5571                         set_bit(STRIPE_HANDLE, &sh->state);
5572                 }
5573                 list_add(&sh->lru, &stripes);
5574         }
5575         spin_lock_irq(&conf->device_lock);
5576         if (mddev->reshape_backwards)
5577                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5578         else
5579                 conf->reshape_progress += reshape_sectors * new_data_disks;
5580         spin_unlock_irq(&conf->device_lock);
5581         /* Ok, those stripe are ready. We can start scheduling
5582          * reads on the source stripes.
5583          * The source stripes are determined by mapping the first and last
5584          * block on the destination stripes.
5585          */
5586         first_sector =
5587                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5588                                      1, &dd_idx, NULL);
5589         last_sector =
5590                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5591                                             * new_data_disks - 1),
5592                                      1, &dd_idx, NULL);
5593         if (last_sector >= mddev->dev_sectors)
5594                 last_sector = mddev->dev_sectors - 1;
5595         while (first_sector <= last_sector) {
5596                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5597                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5598                 set_bit(STRIPE_HANDLE, &sh->state);
5599                 raid5_release_stripe(sh);
5600                 first_sector += STRIPE_SECTORS;
5601         }
5602         /* Now that the sources are clearly marked, we can release
5603          * the destination stripes
5604          */
5605         while (!list_empty(&stripes)) {
5606                 sh = list_entry(stripes.next, struct stripe_head, lru);
5607                 list_del_init(&sh->lru);
5608                 raid5_release_stripe(sh);
5609         }
5610         /* If this takes us to the resync_max point where we have to pause,
5611          * then we need to write out the superblock.
5612          */
5613         sector_nr += reshape_sectors;
5614         retn = reshape_sectors;
5615 finish:
5616         if (mddev->curr_resync_completed > mddev->resync_max ||
5617             (sector_nr - mddev->curr_resync_completed) * 2
5618             >= mddev->resync_max - mddev->curr_resync_completed) {
5619                 /* Cannot proceed until we've updated the superblock... */
5620                 wait_event(conf->wait_for_overlap,
5621                            atomic_read(&conf->reshape_stripes) == 0
5622                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5623                 if (atomic_read(&conf->reshape_stripes) != 0)
5624                         goto ret;
5625                 mddev->reshape_position = conf->reshape_progress;
5626                 mddev->curr_resync_completed = sector_nr;
5627                 conf->reshape_checkpoint = jiffies;
5628                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5629                 md_wakeup_thread(mddev->thread);
5630                 wait_event(mddev->sb_wait,
5631                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5632                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5633                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5634                         goto ret;
5635                 spin_lock_irq(&conf->device_lock);
5636                 conf->reshape_safe = mddev->reshape_position;
5637                 spin_unlock_irq(&conf->device_lock);
5638                 wake_up(&conf->wait_for_overlap);
5639                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5640         }
5641 ret:
5642         return retn;
5643 }
5644
5645 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
5646                                           int *skipped)
5647 {
5648         struct r5conf *conf = mddev->private;
5649         struct stripe_head *sh;
5650         sector_t max_sector = mddev->dev_sectors;
5651         sector_t sync_blocks;
5652         int still_degraded = 0;
5653         int i;
5654
5655         if (sector_nr >= max_sector) {
5656                 /* just being told to finish up .. nothing much to do */
5657
5658                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5659                         end_reshape(conf);
5660                         return 0;
5661                 }
5662
5663                 if (mddev->curr_resync < max_sector) /* aborted */
5664                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5665                                         &sync_blocks, 1);
5666                 else /* completed sync */
5667                         conf->fullsync = 0;
5668                 bitmap_close_sync(mddev->bitmap);
5669
5670                 return 0;
5671         }
5672
5673         /* Allow raid5_quiesce to complete */
5674         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5675
5676         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5677                 return reshape_request(mddev, sector_nr, skipped);
5678
5679         /* No need to check resync_max as we never do more than one
5680          * stripe, and as resync_max will always be on a chunk boundary,
5681          * if the check in md_do_sync didn't fire, there is no chance
5682          * of overstepping resync_max here
5683          */
5684
5685         /* if there is too many failed drives and we are trying
5686          * to resync, then assert that we are finished, because there is
5687          * nothing we can do.
5688          */
5689         if (mddev->degraded >= conf->max_degraded &&
5690             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5691                 sector_t rv = mddev->dev_sectors - sector_nr;
5692                 *skipped = 1;
5693                 return rv;
5694         }
5695         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5696             !conf->fullsync &&
5697             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5698             sync_blocks >= STRIPE_SECTORS) {
5699                 /* we can skip this block, and probably more */
5700                 sync_blocks /= STRIPE_SECTORS;
5701                 *skipped = 1;
5702                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5703         }
5704
5705         bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5706
5707         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5708         if (sh == NULL) {
5709                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5710                 /* make sure we don't swamp the stripe cache if someone else
5711                  * is trying to get access
5712                  */
5713                 schedule_timeout_uninterruptible(1);
5714         }
5715         /* Need to check if array will still be degraded after recovery/resync
5716          * Note in case of > 1 drive failures it's possible we're rebuilding
5717          * one drive while leaving another faulty drive in array.
5718          */
5719         rcu_read_lock();
5720         for (i = 0; i < conf->raid_disks; i++) {
5721                 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5722
5723                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5724                         still_degraded = 1;
5725         }
5726         rcu_read_unlock();
5727
5728         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5729
5730         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5731         set_bit(STRIPE_HANDLE, &sh->state);
5732
5733         raid5_release_stripe(sh);
5734
5735         return STRIPE_SECTORS;
5736 }
5737
5738 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5739 {
5740         /* We may not be able to submit a whole bio at once as there
5741          * may not be enough stripe_heads available.
5742          * We cannot pre-allocate enough stripe_heads as we may need
5743          * more than exist in the cache (if we allow ever large chunks).
5744          * So we do one stripe head at a time and record in
5745          * ->bi_hw_segments how many have been done.
5746          *
5747          * We *know* that this entire raid_bio is in one chunk, so
5748          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5749          */
5750         struct stripe_head *sh;
5751         int dd_idx;
5752         sector_t sector, logical_sector, last_sector;
5753         int scnt = 0;
5754         int remaining;
5755         int handled = 0;
5756
5757         logical_sector = raid_bio->bi_iter.bi_sector &
5758                 ~((sector_t)STRIPE_SECTORS-1);
5759         sector = raid5_compute_sector(conf, logical_sector,
5760                                       0, &dd_idx, NULL);
5761         last_sector = bio_end_sector(raid_bio);
5762
5763         for (; logical_sector < last_sector;
5764              logical_sector += STRIPE_SECTORS,
5765                      sector += STRIPE_SECTORS,
5766                      scnt++) {
5767
5768                 if (scnt < raid5_bi_processed_stripes(raid_bio))
5769                         /* already done this stripe */
5770                         continue;
5771
5772                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5773
5774                 if (!sh) {
5775                         /* failed to get a stripe - must wait */
5776                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5777                         conf->retry_read_aligned = raid_bio;
5778                         return handled;
5779                 }
5780
5781                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5782                         raid5_release_stripe(sh);
5783                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5784                         conf->retry_read_aligned = raid_bio;
5785                         return handled;
5786                 }
5787
5788                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5789                 handle_stripe(sh);
5790                 raid5_release_stripe(sh);
5791                 handled++;
5792         }
5793         remaining = raid5_dec_bi_active_stripes(raid_bio);
5794         if (remaining == 0) {
5795                 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5796                                          raid_bio, 0);
5797                 bio_endio(raid_bio);
5798         }
5799         if (atomic_dec_and_test(&conf->active_aligned_reads))
5800                 wake_up(&conf->wait_for_quiescent);
5801         return handled;
5802 }
5803
5804 static int handle_active_stripes(struct r5conf *conf, int group,
5805                                  struct r5worker *worker,
5806                                  struct list_head *temp_inactive_list)
5807 {
5808         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5809         int i, batch_size = 0, hash;
5810         bool release_inactive = false;
5811
5812         while (batch_size < MAX_STRIPE_BATCH &&
5813                         (sh = __get_priority_stripe(conf, group)) != NULL)
5814                 batch[batch_size++] = sh;
5815
5816         if (batch_size == 0) {
5817                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5818                         if (!list_empty(temp_inactive_list + i))
5819                                 break;
5820                 if (i == NR_STRIPE_HASH_LOCKS) {
5821                         spin_unlock_irq(&conf->device_lock);
5822                         r5l_flush_stripe_to_raid(conf->log);
5823                         spin_lock_irq(&conf->device_lock);
5824                         return batch_size;
5825                 }
5826                 release_inactive = true;
5827         }
5828         spin_unlock_irq(&conf->device_lock);
5829
5830         release_inactive_stripe_list(conf, temp_inactive_list,
5831                                      NR_STRIPE_HASH_LOCKS);
5832
5833         r5l_flush_stripe_to_raid(conf->log);
5834         if (release_inactive) {
5835                 spin_lock_irq(&conf->device_lock);
5836                 return 0;
5837         }
5838
5839         for (i = 0; i < batch_size; i++)
5840                 handle_stripe(batch[i]);
5841         r5l_write_stripe_run(conf->log);
5842
5843         cond_resched();
5844
5845         spin_lock_irq(&conf->device_lock);
5846         for (i = 0; i < batch_size; i++) {
5847                 hash = batch[i]->hash_lock_index;
5848                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5849         }
5850         return batch_size;
5851 }
5852
5853 static void raid5_do_work(struct work_struct *work)
5854 {
5855         struct r5worker *worker = container_of(work, struct r5worker, work);
5856         struct r5worker_group *group = worker->group;
5857         struct r5conf *conf = group->conf;
5858         int group_id = group - conf->worker_groups;
5859         int handled;
5860         struct blk_plug plug;
5861
5862         pr_debug("+++ raid5worker active\n");
5863
5864         blk_start_plug(&plug);
5865         handled = 0;
5866         spin_lock_irq(&conf->device_lock);
5867         while (1) {
5868                 int batch_size, released;
5869
5870                 released = release_stripe_list(conf, worker->temp_inactive_list);
5871
5872                 batch_size = handle_active_stripes(conf, group_id, worker,
5873                                                    worker->temp_inactive_list);
5874                 worker->working = false;
5875                 if (!batch_size && !released)
5876                         break;
5877                 handled += batch_size;
5878         }
5879         pr_debug("%d stripes handled\n", handled);
5880
5881         spin_unlock_irq(&conf->device_lock);
5882
5883         r5l_flush_stripe_to_raid(conf->log);
5884
5885         async_tx_issue_pending_all();
5886         blk_finish_plug(&plug);
5887
5888         pr_debug("--- raid5worker inactive\n");
5889 }
5890
5891 /*
5892  * This is our raid5 kernel thread.
5893  *
5894  * We scan the hash table for stripes which can be handled now.
5895  * During the scan, completed stripes are saved for us by the interrupt
5896  * handler, so that they will not have to wait for our next wakeup.
5897  */
5898 static void raid5d(struct md_thread *thread)
5899 {
5900         struct mddev *mddev = thread->mddev;
5901         struct r5conf *conf = mddev->private;
5902         int handled;
5903         struct blk_plug plug;
5904
5905         pr_debug("+++ raid5d active\n");
5906
5907         md_check_recovery(mddev);
5908
5909         if (!bio_list_empty(&conf->return_bi) &&
5910             !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5911                 struct bio_list tmp = BIO_EMPTY_LIST;
5912                 spin_lock_irq(&conf->device_lock);
5913                 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5914                         bio_list_merge(&tmp, &conf->return_bi);
5915                         bio_list_init(&conf->return_bi);
5916                 }
5917                 spin_unlock_irq(&conf->device_lock);
5918                 return_io(&tmp);
5919         }
5920
5921         blk_start_plug(&plug);
5922         handled = 0;
5923         spin_lock_irq(&conf->device_lock);
5924         while (1) {
5925                 struct bio *bio;
5926                 int batch_size, released;
5927
5928                 released = release_stripe_list(conf, conf->temp_inactive_list);
5929                 if (released)
5930                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
5931
5932                 if (
5933                     !list_empty(&conf->bitmap_list)) {
5934                         /* Now is a good time to flush some bitmap updates */
5935                         conf->seq_flush++;
5936                         spin_unlock_irq(&conf->device_lock);
5937                         bitmap_unplug(mddev->bitmap);
5938                         spin_lock_irq(&conf->device_lock);
5939                         conf->seq_write = conf->seq_flush;
5940                         activate_bit_delay(conf, conf->temp_inactive_list);
5941                 }
5942                 raid5_activate_delayed(conf);
5943
5944                 while ((bio = remove_bio_from_retry(conf))) {
5945                         int ok;
5946                         spin_unlock_irq(&conf->device_lock);
5947                         ok = retry_aligned_read(conf, bio);
5948                         spin_lock_irq(&conf->device_lock);
5949                         if (!ok)
5950                                 break;
5951                         handled++;
5952                 }
5953
5954                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5955                                                    conf->temp_inactive_list);
5956                 if (!batch_size && !released)
5957                         break;
5958                 handled += batch_size;
5959
5960                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5961                         spin_unlock_irq(&conf->device_lock);
5962                         md_check_recovery(mddev);
5963                         spin_lock_irq(&conf->device_lock);
5964                 }
5965         }
5966         pr_debug("%d stripes handled\n", handled);
5967
5968         spin_unlock_irq(&conf->device_lock);
5969         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5970             mutex_trylock(&conf->cache_size_mutex)) {
5971                 grow_one_stripe(conf, __GFP_NOWARN);
5972                 /* Set flag even if allocation failed.  This helps
5973                  * slow down allocation requests when mem is short
5974                  */
5975                 set_bit(R5_DID_ALLOC, &conf->cache_state);
5976                 mutex_unlock(&conf->cache_size_mutex);
5977         }
5978
5979         r5l_flush_stripe_to_raid(conf->log);
5980
5981         async_tx_issue_pending_all();
5982         blk_finish_plug(&plug);
5983
5984         pr_debug("--- raid5d inactive\n");
5985 }
5986
5987 static ssize_t
5988 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5989 {
5990         struct r5conf *conf;
5991         int ret = 0;
5992         spin_lock(&mddev->lock);
5993         conf = mddev->private;
5994         if (conf)
5995                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5996         spin_unlock(&mddev->lock);
5997         return ret;
5998 }
5999
6000 int
6001 raid5_set_cache_size(struct mddev *mddev, int size)
6002 {
6003         struct r5conf *conf = mddev->private;
6004         int err;
6005
6006         if (size <= 16 || size > 32768)
6007                 return -EINVAL;
6008
6009         conf->min_nr_stripes = size;
6010         mutex_lock(&conf->cache_size_mutex);
6011         while (size < conf->max_nr_stripes &&
6012                drop_one_stripe(conf))
6013                 ;
6014         mutex_unlock(&conf->cache_size_mutex);
6015
6016
6017         err = md_allow_write(mddev);
6018         if (err)
6019                 return err;
6020
6021         mutex_lock(&conf->cache_size_mutex);
6022         while (size > conf->max_nr_stripes)
6023                 if (!grow_one_stripe(conf, GFP_KERNEL))
6024                         break;
6025         mutex_unlock(&conf->cache_size_mutex);
6026
6027         return 0;
6028 }
6029 EXPORT_SYMBOL(raid5_set_cache_size);
6030
6031 static ssize_t
6032 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6033 {
6034         struct r5conf *conf;
6035         unsigned long new;
6036         int err;
6037
6038         if (len >= PAGE_SIZE)
6039                 return -EINVAL;
6040         if (kstrtoul(page, 10, &new))
6041                 return -EINVAL;
6042         err = mddev_lock(mddev);
6043         if (err)
6044                 return err;
6045         conf = mddev->private;
6046         if (!conf)
6047                 err = -ENODEV;
6048         else
6049                 err = raid5_set_cache_size(mddev, new);
6050         mddev_unlock(mddev);
6051
6052         return err ?: len;
6053 }
6054
6055 static struct md_sysfs_entry
6056 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6057                                 raid5_show_stripe_cache_size,
6058                                 raid5_store_stripe_cache_size);
6059
6060 static ssize_t
6061 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6062 {
6063         struct r5conf *conf = mddev->private;
6064         if (conf)
6065                 return sprintf(page, "%d\n", conf->rmw_level);
6066         else
6067                 return 0;
6068 }
6069
6070 static ssize_t
6071 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6072 {
6073         struct r5conf *conf = mddev->private;
6074         unsigned long new;
6075
6076         if (!conf)
6077                 return -ENODEV;
6078
6079         if (len >= PAGE_SIZE)
6080                 return -EINVAL;
6081
6082         if (kstrtoul(page, 10, &new))
6083                 return -EINVAL;
6084
6085         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6086                 return -EINVAL;
6087
6088         if (new != PARITY_DISABLE_RMW &&
6089             new != PARITY_ENABLE_RMW &&
6090             new != PARITY_PREFER_RMW)
6091                 return -EINVAL;
6092
6093         conf->rmw_level = new;
6094         return len;
6095 }
6096
6097 static struct md_sysfs_entry
6098 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6099                          raid5_show_rmw_level,
6100                          raid5_store_rmw_level);
6101
6102
6103 static ssize_t
6104 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6105 {
6106         struct r5conf *conf;
6107         int ret = 0;
6108         spin_lock(&mddev->lock);
6109         conf = mddev->private;
6110         if (conf)
6111                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6112         spin_unlock(&mddev->lock);
6113         return ret;
6114 }
6115
6116 static ssize_t
6117 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6118 {
6119         struct r5conf *conf;
6120         unsigned long new;
6121         int err;
6122
6123         if (len >= PAGE_SIZE)
6124                 return -EINVAL;
6125         if (kstrtoul(page, 10, &new))
6126                 return -EINVAL;
6127
6128         err = mddev_lock(mddev);
6129         if (err)
6130                 return err;
6131         conf = mddev->private;
6132         if (!conf)
6133                 err = -ENODEV;
6134         else if (new > conf->min_nr_stripes)
6135                 err = -EINVAL;
6136         else
6137                 conf->bypass_threshold = new;
6138         mddev_unlock(mddev);
6139         return err ?: len;
6140 }
6141
6142 static struct md_sysfs_entry
6143 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6144                                         S_IRUGO | S_IWUSR,
6145                                         raid5_show_preread_threshold,
6146                                         raid5_store_preread_threshold);
6147
6148 static ssize_t
6149 raid5_show_skip_copy(struct mddev *mddev, char *page)
6150 {
6151         struct r5conf *conf;
6152         int ret = 0;
6153         spin_lock(&mddev->lock);
6154         conf = mddev->private;
6155         if (conf)
6156                 ret = sprintf(page, "%d\n", conf->skip_copy);
6157         spin_unlock(&mddev->lock);
6158         return ret;
6159 }
6160
6161 static ssize_t
6162 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6163 {
6164         struct r5conf *conf;
6165         unsigned long new;
6166         int err;
6167
6168         if (len >= PAGE_SIZE)
6169                 return -EINVAL;
6170         if (kstrtoul(page, 10, &new))
6171                 return -EINVAL;
6172         new = !!new;
6173
6174         err = mddev_lock(mddev);
6175         if (err)
6176                 return err;
6177         conf = mddev->private;
6178         if (!conf)
6179                 err = -ENODEV;
6180         else if (new != conf->skip_copy) {
6181                 mddev_suspend(mddev);
6182                 conf->skip_copy = new;
6183                 if (new)
6184                         mddev->queue->backing_dev_info.capabilities |=
6185                                 BDI_CAP_STABLE_WRITES;
6186                 else
6187                         mddev->queue->backing_dev_info.capabilities &=
6188                                 ~BDI_CAP_STABLE_WRITES;
6189                 mddev_resume(mddev);
6190         }
6191         mddev_unlock(mddev);
6192         return err ?: len;
6193 }
6194
6195 static struct md_sysfs_entry
6196 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6197                                         raid5_show_skip_copy,
6198                                         raid5_store_skip_copy);
6199
6200 static ssize_t
6201 stripe_cache_active_show(struct mddev *mddev, char *page)
6202 {
6203         struct r5conf *conf = mddev->private;
6204         if (conf)
6205                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6206         else
6207                 return 0;
6208 }
6209
6210 static struct md_sysfs_entry
6211 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6212
6213 static ssize_t
6214 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6215 {
6216         struct r5conf *conf;
6217         int ret = 0;
6218         spin_lock(&mddev->lock);
6219         conf = mddev->private;
6220         if (conf)
6221                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6222         spin_unlock(&mddev->lock);
6223         return ret;
6224 }
6225
6226 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6227                                int *group_cnt,
6228                                int *worker_cnt_per_group,
6229                                struct r5worker_group **worker_groups);
6230 static ssize_t
6231 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6232 {
6233         struct r5conf *conf;
6234         unsigned long new;
6235         int err;
6236         struct r5worker_group *new_groups, *old_groups;
6237         int group_cnt, worker_cnt_per_group;
6238
6239         if (len >= PAGE_SIZE)
6240                 return -EINVAL;
6241         if (kstrtoul(page, 10, &new))
6242                 return -EINVAL;
6243
6244         err = mddev_lock(mddev);
6245         if (err)
6246                 return err;
6247         conf = mddev->private;
6248         if (!conf)
6249                 err = -ENODEV;
6250         else if (new != conf->worker_cnt_per_group) {
6251                 mddev_suspend(mddev);
6252
6253                 old_groups = conf->worker_groups;
6254                 if (old_groups)
6255                         flush_workqueue(raid5_wq);
6256
6257                 err = alloc_thread_groups(conf, new,
6258                                           &group_cnt, &worker_cnt_per_group,
6259                                           &new_groups);
6260                 if (!err) {
6261                         spin_lock_irq(&conf->device_lock);
6262                         conf->group_cnt = group_cnt;
6263                         conf->worker_cnt_per_group = worker_cnt_per_group;
6264                         conf->worker_groups = new_groups;
6265                         spin_unlock_irq(&conf->device_lock);
6266
6267                         if (old_groups)
6268                                 kfree(old_groups[0].workers);
6269                         kfree(old_groups);
6270                 }
6271                 mddev_resume(mddev);
6272         }
6273         mddev_unlock(mddev);
6274
6275         return err ?: len;
6276 }
6277
6278 static struct md_sysfs_entry
6279 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6280                                 raid5_show_group_thread_cnt,
6281                                 raid5_store_group_thread_cnt);
6282
6283 static struct attribute *raid5_attrs[] =  {
6284         &raid5_stripecache_size.attr,
6285         &raid5_stripecache_active.attr,
6286         &raid5_preread_bypass_threshold.attr,
6287         &raid5_group_thread_cnt.attr,
6288         &raid5_skip_copy.attr,
6289         &raid5_rmw_level.attr,
6290         NULL,
6291 };
6292 static struct attribute_group raid5_attrs_group = {
6293         .name = NULL,
6294         .attrs = raid5_attrs,
6295 };
6296
6297 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6298                                int *group_cnt,
6299                                int *worker_cnt_per_group,
6300                                struct r5worker_group **worker_groups)
6301 {
6302         int i, j, k;
6303         ssize_t size;
6304         struct r5worker *workers;
6305
6306         *worker_cnt_per_group = cnt;
6307         if (cnt == 0) {
6308                 *group_cnt = 0;
6309                 *worker_groups = NULL;
6310                 return 0;
6311         }
6312         *group_cnt = num_possible_nodes();
6313         size = sizeof(struct r5worker) * cnt;
6314         workers = kzalloc(size * *group_cnt, GFP_NOIO);
6315         *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6316                                 *group_cnt, GFP_NOIO);
6317         if (!*worker_groups || !workers) {
6318                 kfree(workers);
6319                 kfree(*worker_groups);
6320                 return -ENOMEM;
6321         }
6322
6323         for (i = 0; i < *group_cnt; i++) {
6324                 struct r5worker_group *group;
6325
6326                 group = &(*worker_groups)[i];
6327                 INIT_LIST_HEAD(&group->handle_list);
6328                 group->conf = conf;
6329                 group->workers = workers + i * cnt;
6330
6331                 for (j = 0; j < cnt; j++) {
6332                         struct r5worker *worker = group->workers + j;
6333                         worker->group = group;
6334                         INIT_WORK(&worker->work, raid5_do_work);
6335
6336                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6337                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6338                 }
6339         }
6340
6341         return 0;
6342 }
6343
6344 static void free_thread_groups(struct r5conf *conf)
6345 {
6346         if (conf->worker_groups)
6347                 kfree(conf->worker_groups[0].workers);
6348         kfree(conf->worker_groups);
6349         conf->worker_groups = NULL;
6350 }
6351
6352 static sector_t
6353 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6354 {
6355         struct r5conf *conf = mddev->private;
6356
6357         if (!sectors)
6358                 sectors = mddev->dev_sectors;
6359         if (!raid_disks)
6360                 /* size is defined by the smallest of previous and new size */
6361                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6362
6363         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6364         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6365         return sectors * (raid_disks - conf->max_degraded);
6366 }
6367
6368 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6369 {
6370         safe_put_page(percpu->spare_page);
6371         if (percpu->scribble)
6372                 flex_array_free(percpu->scribble);
6373         percpu->spare_page = NULL;
6374         percpu->scribble = NULL;
6375 }
6376
6377 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6378 {
6379         if (conf->level == 6 && !percpu->spare_page)
6380                 percpu->spare_page = alloc_page(GFP_KERNEL);
6381         if (!percpu->scribble)
6382                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6383                                                       conf->previous_raid_disks),
6384                                                   max(conf->chunk_sectors,
6385                                                       conf->prev_chunk_sectors)
6386                                                    / STRIPE_SECTORS,
6387                                                   GFP_KERNEL);
6388
6389         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6390                 free_scratch_buffer(conf, percpu);
6391                 return -ENOMEM;
6392         }
6393
6394         return 0;
6395 }
6396
6397 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6398 {
6399         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6400
6401         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6402         return 0;
6403 }
6404
6405 static void raid5_free_percpu(struct r5conf *conf)
6406 {
6407         if (!conf->percpu)
6408                 return;
6409
6410         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6411         free_percpu(conf->percpu);
6412 }
6413
6414 static void free_conf(struct r5conf *conf)
6415 {
6416         if (conf->log)
6417                 r5l_exit_log(conf->log);
6418         if (conf->shrinker.nr_deferred)
6419                 unregister_shrinker(&conf->shrinker);
6420
6421         free_thread_groups(conf);
6422         shrink_stripes(conf);
6423         raid5_free_percpu(conf);
6424         kfree(conf->disks);
6425         kfree(conf->stripe_hashtbl);
6426         kfree(conf);
6427 }
6428
6429 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6430 {
6431         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6432         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6433
6434         if (alloc_scratch_buffer(conf, percpu)) {
6435                 pr_err("%s: failed memory allocation for cpu%u\n",
6436                        __func__, cpu);
6437                 return -ENOMEM;
6438         }
6439         return 0;
6440 }
6441
6442 static int raid5_alloc_percpu(struct r5conf *conf)
6443 {
6444         int err = 0;
6445
6446         conf->percpu = alloc_percpu(struct raid5_percpu);
6447         if (!conf->percpu)
6448                 return -ENOMEM;
6449
6450         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6451         if (!err) {
6452                 conf->scribble_disks = max(conf->raid_disks,
6453                         conf->previous_raid_disks);
6454                 conf->scribble_sectors = max(conf->chunk_sectors,
6455                         conf->prev_chunk_sectors);
6456         }
6457         return err;
6458 }
6459
6460 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6461                                       struct shrink_control *sc)
6462 {
6463         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6464         unsigned long ret = SHRINK_STOP;
6465
6466         if (mutex_trylock(&conf->cache_size_mutex)) {
6467                 ret= 0;
6468                 while (ret < sc->nr_to_scan &&
6469                        conf->max_nr_stripes > conf->min_nr_stripes) {
6470                         if (drop_one_stripe(conf) == 0) {
6471                                 ret = SHRINK_STOP;
6472                                 break;
6473                         }
6474                         ret++;
6475                 }
6476                 mutex_unlock(&conf->cache_size_mutex);
6477         }
6478         return ret;
6479 }
6480
6481 static unsigned long raid5_cache_count(struct shrinker *shrink,
6482                                        struct shrink_control *sc)
6483 {
6484         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6485
6486         if (conf->max_nr_stripes < conf->min_nr_stripes)
6487                 /* unlikely, but not impossible */
6488                 return 0;
6489         return conf->max_nr_stripes - conf->min_nr_stripes;
6490 }
6491
6492 static struct r5conf *setup_conf(struct mddev *mddev)
6493 {
6494         struct r5conf *conf;
6495         int raid_disk, memory, max_disks;
6496         struct md_rdev *rdev;
6497         struct disk_info *disk;
6498         char pers_name[6];
6499         int i;
6500         int group_cnt, worker_cnt_per_group;
6501         struct r5worker_group *new_group;
6502
6503         if (mddev->new_level != 5
6504             && mddev->new_level != 4
6505             && mddev->new_level != 6) {
6506                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6507                        mdname(mddev), mddev->new_level);
6508                 return ERR_PTR(-EIO);
6509         }
6510         if ((mddev->new_level == 5
6511              && !algorithm_valid_raid5(mddev->new_layout)) ||
6512             (mddev->new_level == 6
6513              && !algorithm_valid_raid6(mddev->new_layout))) {
6514                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6515                        mdname(mddev), mddev->new_layout);
6516                 return ERR_PTR(-EIO);
6517         }
6518         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6519                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6520                        mdname(mddev), mddev->raid_disks);
6521                 return ERR_PTR(-EINVAL);
6522         }
6523
6524         if (!mddev->new_chunk_sectors ||
6525             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6526             !is_power_of_2(mddev->new_chunk_sectors)) {
6527                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6528                        mdname(mddev), mddev->new_chunk_sectors << 9);
6529                 return ERR_PTR(-EINVAL);
6530         }
6531
6532         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6533         if (conf == NULL)
6534                 goto abort;
6535         /* Don't enable multi-threading by default*/
6536         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6537                                  &new_group)) {
6538                 conf->group_cnt = group_cnt;
6539                 conf->worker_cnt_per_group = worker_cnt_per_group;
6540                 conf->worker_groups = new_group;
6541         } else
6542                 goto abort;
6543         spin_lock_init(&conf->device_lock);
6544         seqcount_init(&conf->gen_lock);
6545         mutex_init(&conf->cache_size_mutex);
6546         init_waitqueue_head(&conf->wait_for_quiescent);
6547         init_waitqueue_head(&conf->wait_for_stripe);
6548         init_waitqueue_head(&conf->wait_for_overlap);
6549         INIT_LIST_HEAD(&conf->handle_list);
6550         INIT_LIST_HEAD(&conf->hold_list);
6551         INIT_LIST_HEAD(&conf->delayed_list);
6552         INIT_LIST_HEAD(&conf->bitmap_list);
6553         bio_list_init(&conf->return_bi);
6554         init_llist_head(&conf->released_stripes);
6555         atomic_set(&conf->active_stripes, 0);
6556         atomic_set(&conf->preread_active_stripes, 0);
6557         atomic_set(&conf->active_aligned_reads, 0);
6558         conf->bypass_threshold = BYPASS_THRESHOLD;
6559         conf->recovery_disabled = mddev->recovery_disabled - 1;
6560
6561         conf->raid_disks = mddev->raid_disks;
6562         if (mddev->reshape_position == MaxSector)
6563                 conf->previous_raid_disks = mddev->raid_disks;
6564         else
6565                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6566         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6567
6568         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6569                               GFP_KERNEL);
6570         if (!conf->disks)
6571                 goto abort;
6572
6573         conf->mddev = mddev;
6574
6575         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6576                 goto abort;
6577
6578         /* We init hash_locks[0] separately to that it can be used
6579          * as the reference lock in the spin_lock_nest_lock() call
6580          * in lock_all_device_hash_locks_irq in order to convince
6581          * lockdep that we know what we are doing.
6582          */
6583         spin_lock_init(conf->hash_locks);
6584         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6585                 spin_lock_init(conf->hash_locks + i);
6586
6587         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6588                 INIT_LIST_HEAD(conf->inactive_list + i);
6589
6590         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6591                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6592
6593         conf->level = mddev->new_level;
6594         conf->chunk_sectors = mddev->new_chunk_sectors;
6595         if (raid5_alloc_percpu(conf) != 0)
6596                 goto abort;
6597
6598         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6599
6600         rdev_for_each(rdev, mddev) {
6601                 raid_disk = rdev->raid_disk;
6602                 if (raid_disk >= max_disks
6603                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6604                         continue;
6605                 disk = conf->disks + raid_disk;
6606
6607                 if (test_bit(Replacement, &rdev->flags)) {
6608                         if (disk->replacement)
6609                                 goto abort;
6610                         disk->replacement = rdev;
6611                 } else {
6612                         if (disk->rdev)
6613                                 goto abort;
6614                         disk->rdev = rdev;
6615                 }
6616
6617                 if (test_bit(In_sync, &rdev->flags)) {
6618                         char b[BDEVNAME_SIZE];
6619                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6620                                " disk %d\n",
6621                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6622                 } else if (rdev->saved_raid_disk != raid_disk)
6623                         /* Cannot rely on bitmap to complete recovery */
6624                         conf->fullsync = 1;
6625         }
6626
6627         conf->level = mddev->new_level;
6628         if (conf->level == 6) {
6629                 conf->max_degraded = 2;
6630                 if (raid6_call.xor_syndrome)
6631                         conf->rmw_level = PARITY_ENABLE_RMW;
6632                 else
6633                         conf->rmw_level = PARITY_DISABLE_RMW;
6634         } else {
6635                 conf->max_degraded = 1;
6636                 conf->rmw_level = PARITY_ENABLE_RMW;
6637         }
6638         conf->algorithm = mddev->new_layout;
6639         conf->reshape_progress = mddev->reshape_position;
6640         if (conf->reshape_progress != MaxSector) {
6641                 conf->prev_chunk_sectors = mddev->chunk_sectors;
6642                 conf->prev_algo = mddev->layout;
6643         } else {
6644                 conf->prev_chunk_sectors = conf->chunk_sectors;
6645                 conf->prev_algo = conf->algorithm;
6646         }
6647
6648         conf->min_nr_stripes = NR_STRIPES;
6649         if (mddev->reshape_position != MaxSector) {
6650                 int stripes = max_t(int,
6651                         ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
6652                         ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
6653                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
6654                 if (conf->min_nr_stripes != NR_STRIPES)
6655                         printk(KERN_INFO
6656                                 "md/raid:%s: force stripe size %d for reshape\n",
6657                                 mdname(mddev), conf->min_nr_stripes);
6658         }
6659         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6660                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6661         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6662         if (grow_stripes(conf, conf->min_nr_stripes)) {
6663                 printk(KERN_ERR
6664                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
6665                        mdname(mddev), memory);
6666                 goto abort;
6667         } else
6668                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6669                        mdname(mddev), memory);
6670         /*
6671          * Losing a stripe head costs more than the time to refill it,
6672          * it reduces the queue depth and so can hurt throughput.
6673          * So set it rather large, scaled by number of devices.
6674          */
6675         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6676         conf->shrinker.scan_objects = raid5_cache_scan;
6677         conf->shrinker.count_objects = raid5_cache_count;
6678         conf->shrinker.batch = 128;
6679         conf->shrinker.flags = 0;
6680         if (register_shrinker(&conf->shrinker)) {
6681                 printk(KERN_ERR
6682                        "md/raid:%s: couldn't register shrinker.\n",
6683                        mdname(mddev));
6684                 goto abort;
6685         }
6686
6687         sprintf(pers_name, "raid%d", mddev->new_level);
6688         conf->thread = md_register_thread(raid5d, mddev, pers_name);
6689         if (!conf->thread) {
6690                 printk(KERN_ERR
6691                        "md/raid:%s: couldn't allocate thread.\n",
6692                        mdname(mddev));
6693                 goto abort;
6694         }
6695
6696         return conf;
6697
6698  abort:
6699         if (conf) {
6700                 free_conf(conf);
6701                 return ERR_PTR(-EIO);
6702         } else
6703                 return ERR_PTR(-ENOMEM);
6704 }
6705
6706 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6707 {
6708         switch (algo) {
6709         case ALGORITHM_PARITY_0:
6710                 if (raid_disk < max_degraded)
6711                         return 1;
6712                 break;
6713         case ALGORITHM_PARITY_N:
6714                 if (raid_disk >= raid_disks - max_degraded)
6715                         return 1;
6716                 break;
6717         case ALGORITHM_PARITY_0_6:
6718                 if (raid_disk == 0 ||
6719                     raid_disk == raid_disks - 1)
6720                         return 1;
6721                 break;
6722         case ALGORITHM_LEFT_ASYMMETRIC_6:
6723         case ALGORITHM_RIGHT_ASYMMETRIC_6:
6724         case ALGORITHM_LEFT_SYMMETRIC_6:
6725         case ALGORITHM_RIGHT_SYMMETRIC_6:
6726                 if (raid_disk == raid_disks - 1)
6727                         return 1;
6728         }
6729         return 0;
6730 }
6731
6732 static int raid5_run(struct mddev *mddev)
6733 {
6734         struct r5conf *conf;
6735         int working_disks = 0;
6736         int dirty_parity_disks = 0;
6737         struct md_rdev *rdev;
6738         struct md_rdev *journal_dev = NULL;
6739         sector_t reshape_offset = 0;
6740         int i;
6741         long long min_offset_diff = 0;
6742         int first = 1;
6743
6744         if (mddev->recovery_cp != MaxSector)
6745                 printk(KERN_NOTICE "md/raid:%s: not clean"
6746                        " -- starting background reconstruction\n",
6747                        mdname(mddev));
6748
6749         rdev_for_each(rdev, mddev) {
6750                 long long diff;
6751
6752                 if (test_bit(Journal, &rdev->flags)) {
6753                         journal_dev = rdev;
6754                         continue;
6755                 }
6756                 if (rdev->raid_disk < 0)
6757                         continue;
6758                 diff = (rdev->new_data_offset - rdev->data_offset);
6759                 if (first) {
6760                         min_offset_diff = diff;
6761                         first = 0;
6762                 } else if (mddev->reshape_backwards &&
6763                          diff < min_offset_diff)
6764                         min_offset_diff = diff;
6765                 else if (!mddev->reshape_backwards &&
6766                          diff > min_offset_diff)
6767                         min_offset_diff = diff;
6768         }
6769
6770         if (mddev->reshape_position != MaxSector) {
6771                 /* Check that we can continue the reshape.
6772                  * Difficulties arise if the stripe we would write to
6773                  * next is at or after the stripe we would read from next.
6774                  * For a reshape that changes the number of devices, this
6775                  * is only possible for a very short time, and mdadm makes
6776                  * sure that time appears to have past before assembling
6777                  * the array.  So we fail if that time hasn't passed.
6778                  * For a reshape that keeps the number of devices the same
6779                  * mdadm must be monitoring the reshape can keeping the
6780                  * critical areas read-only and backed up.  It will start
6781                  * the array in read-only mode, so we check for that.
6782                  */
6783                 sector_t here_new, here_old;
6784                 int old_disks;
6785                 int max_degraded = (mddev->level == 6 ? 2 : 1);
6786                 int chunk_sectors;
6787                 int new_data_disks;
6788
6789                 if (journal_dev) {
6790                         printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6791                                mdname(mddev));
6792                         return -EINVAL;
6793                 }
6794
6795                 if (mddev->new_level != mddev->level) {
6796                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
6797                                "required - aborting.\n",
6798                                mdname(mddev));
6799                         return -EINVAL;
6800                 }
6801                 old_disks = mddev->raid_disks - mddev->delta_disks;
6802                 /* reshape_position must be on a new-stripe boundary, and one
6803                  * further up in new geometry must map after here in old
6804                  * geometry.
6805                  * If the chunk sizes are different, then as we perform reshape
6806                  * in units of the largest of the two, reshape_position needs
6807                  * be a multiple of the largest chunk size times new data disks.
6808                  */
6809                 here_new = mddev->reshape_position;
6810                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6811                 new_data_disks = mddev->raid_disks - max_degraded;
6812                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6813                         printk(KERN_ERR "md/raid:%s: reshape_position not "
6814                                "on a stripe boundary\n", mdname(mddev));
6815                         return -EINVAL;
6816                 }
6817                 reshape_offset = here_new * chunk_sectors;
6818                 /* here_new is the stripe we will write to */
6819                 here_old = mddev->reshape_position;
6820                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6821                 /* here_old is the first stripe that we might need to read
6822                  * from */
6823                 if (mddev->delta_disks == 0) {
6824                         /* We cannot be sure it is safe to start an in-place
6825                          * reshape.  It is only safe if user-space is monitoring
6826                          * and taking constant backups.
6827                          * mdadm always starts a situation like this in
6828                          * readonly mode so it can take control before
6829                          * allowing any writes.  So just check for that.
6830                          */
6831                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6832                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
6833                                 /* not really in-place - so OK */;
6834                         else if (mddev->ro == 0) {
6835                                 printk(KERN_ERR "md/raid:%s: in-place reshape "
6836                                        "must be started in read-only mode "
6837                                        "- aborting\n",
6838                                        mdname(mddev));
6839                                 return -EINVAL;
6840                         }
6841                 } else if (mddev->reshape_backwards
6842                     ? (here_new * chunk_sectors + min_offset_diff <=
6843                        here_old * chunk_sectors)
6844                     : (here_new * chunk_sectors >=
6845                        here_old * chunk_sectors + (-min_offset_diff))) {
6846                         /* Reading from the same stripe as writing to - bad */
6847                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6848                                "auto-recovery - aborting.\n",
6849                                mdname(mddev));
6850                         return -EINVAL;
6851                 }
6852                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6853                        mdname(mddev));
6854                 /* OK, we should be able to continue; */
6855         } else {
6856                 BUG_ON(mddev->level != mddev->new_level);
6857                 BUG_ON(mddev->layout != mddev->new_layout);
6858                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6859                 BUG_ON(mddev->delta_disks != 0);
6860         }
6861
6862         if (mddev->private == NULL)
6863                 conf = setup_conf(mddev);
6864         else
6865                 conf = mddev->private;
6866
6867         if (IS_ERR(conf))
6868                 return PTR_ERR(conf);
6869
6870         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
6871                 if (!journal_dev) {
6872                         pr_err("md/raid:%s: journal disk is missing, force array readonly\n",
6873                                mdname(mddev));
6874                         mddev->ro = 1;
6875                         set_disk_ro(mddev->gendisk, 1);
6876                 } else if (mddev->recovery_cp == MaxSector)
6877                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
6878         }
6879
6880         conf->min_offset_diff = min_offset_diff;
6881         mddev->thread = conf->thread;
6882         conf->thread = NULL;
6883         mddev->private = conf;
6884
6885         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6886              i++) {
6887                 rdev = conf->disks[i].rdev;
6888                 if (!rdev && conf->disks[i].replacement) {
6889                         /* The replacement is all we have yet */
6890                         rdev = conf->disks[i].replacement;
6891                         conf->disks[i].replacement = NULL;
6892                         clear_bit(Replacement, &rdev->flags);
6893                         conf->disks[i].rdev = rdev;
6894                 }
6895                 if (!rdev)
6896                         continue;
6897                 if (conf->disks[i].replacement &&
6898                     conf->reshape_progress != MaxSector) {
6899                         /* replacements and reshape simply do not mix. */
6900                         printk(KERN_ERR "md: cannot handle concurrent "
6901                                "replacement and reshape.\n");
6902                         goto abort;
6903                 }
6904                 if (test_bit(In_sync, &rdev->flags)) {
6905                         working_disks++;
6906                         continue;
6907                 }
6908                 /* This disc is not fully in-sync.  However if it
6909                  * just stored parity (beyond the recovery_offset),
6910                  * when we don't need to be concerned about the
6911                  * array being dirty.
6912                  * When reshape goes 'backwards', we never have
6913                  * partially completed devices, so we only need
6914                  * to worry about reshape going forwards.
6915                  */
6916                 /* Hack because v0.91 doesn't store recovery_offset properly. */
6917                 if (mddev->major_version == 0 &&
6918                     mddev->minor_version > 90)
6919                         rdev->recovery_offset = reshape_offset;
6920
6921                 if (rdev->recovery_offset < reshape_offset) {
6922                         /* We need to check old and new layout */
6923                         if (!only_parity(rdev->raid_disk,
6924                                          conf->algorithm,
6925                                          conf->raid_disks,
6926                                          conf->max_degraded))
6927                                 continue;
6928                 }
6929                 if (!only_parity(rdev->raid_disk,
6930                                  conf->prev_algo,
6931                                  conf->previous_raid_disks,
6932                                  conf->max_degraded))
6933                         continue;
6934                 dirty_parity_disks++;
6935         }
6936
6937         /*
6938          * 0 for a fully functional array, 1 or 2 for a degraded array.
6939          */
6940         mddev->degraded = calc_degraded(conf);
6941
6942         if (has_failed(conf)) {
6943                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6944                         " (%d/%d failed)\n",
6945                         mdname(mddev), mddev->degraded, conf->raid_disks);
6946                 goto abort;
6947         }
6948
6949         /* device size must be a multiple of chunk size */
6950         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6951         mddev->resync_max_sectors = mddev->dev_sectors;
6952
6953         if (mddev->degraded > dirty_parity_disks &&
6954             mddev->recovery_cp != MaxSector) {
6955                 if (mddev->ok_start_degraded)
6956                         printk(KERN_WARNING
6957                                "md/raid:%s: starting dirty degraded array"
6958                                " - data corruption possible.\n",
6959                                mdname(mddev));
6960                 else {
6961                         printk(KERN_ERR
6962                                "md/raid:%s: cannot start dirty degraded array.\n",
6963                                mdname(mddev));
6964                         goto abort;
6965                 }
6966         }
6967
6968         if (mddev->degraded == 0)
6969                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6970                        " devices, algorithm %d\n", mdname(mddev), conf->level,
6971                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6972                        mddev->new_layout);
6973         else
6974                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6975                        " out of %d devices, algorithm %d\n",
6976                        mdname(mddev), conf->level,
6977                        mddev->raid_disks - mddev->degraded,
6978                        mddev->raid_disks, mddev->new_layout);
6979
6980         print_raid5_conf(conf);
6981
6982         if (conf->reshape_progress != MaxSector) {
6983                 conf->reshape_safe = conf->reshape_progress;
6984                 atomic_set(&conf->reshape_stripes, 0);
6985                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6986                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6987                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6988                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6989                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6990                                                         "reshape");
6991                 if (!mddev->sync_thread)
6992                         goto abort;
6993         }
6994
6995         /* Ok, everything is just fine now */
6996         if (mddev->to_remove == &raid5_attrs_group)
6997                 mddev->to_remove = NULL;
6998         else if (mddev->kobj.sd &&
6999             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7000                 printk(KERN_WARNING
7001                        "raid5: failed to create sysfs attributes for %s\n",
7002                        mdname(mddev));
7003         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7004
7005         if (mddev->queue) {
7006                 int chunk_size;
7007                 bool discard_supported = true;
7008                 /* read-ahead size must cover two whole stripes, which
7009                  * is 2 * (datadisks) * chunksize where 'n' is the
7010                  * number of raid devices
7011                  */
7012                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7013                 int stripe = data_disks *
7014                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7015                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7016                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7017
7018                 chunk_size = mddev->chunk_sectors << 9;
7019                 blk_queue_io_min(mddev->queue, chunk_size);
7020                 blk_queue_io_opt(mddev->queue, chunk_size *
7021                                  (conf->raid_disks - conf->max_degraded));
7022                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7023                 /*
7024                  * We can only discard a whole stripe. It doesn't make sense to
7025                  * discard data disk but write parity disk
7026                  */
7027                 stripe = stripe * PAGE_SIZE;
7028                 /* Round up to power of 2, as discard handling
7029                  * currently assumes that */
7030                 while ((stripe-1) & stripe)
7031                         stripe = (stripe | (stripe-1)) + 1;
7032                 mddev->queue->limits.discard_alignment = stripe;
7033                 mddev->queue->limits.discard_granularity = stripe;
7034
7035                 /*
7036                  * We use 16-bit counter of active stripes in bi_phys_segments
7037                  * (minus one for over-loaded initialization)
7038                  */
7039                 blk_queue_max_hw_sectors(mddev->queue, 0xfffe * STRIPE_SECTORS);
7040                 blk_queue_max_discard_sectors(mddev->queue,
7041                                               0xfffe * STRIPE_SECTORS);
7042
7043                 /*
7044                  * unaligned part of discard request will be ignored, so can't
7045                  * guarantee discard_zeroes_data
7046                  */
7047                 mddev->queue->limits.discard_zeroes_data = 0;
7048
7049                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7050
7051                 rdev_for_each(rdev, mddev) {
7052                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7053                                           rdev->data_offset << 9);
7054                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7055                                           rdev->new_data_offset << 9);
7056                         /*
7057                          * discard_zeroes_data is required, otherwise data
7058                          * could be lost. Consider a scenario: discard a stripe
7059                          * (the stripe could be inconsistent if
7060                          * discard_zeroes_data is 0); write one disk of the
7061                          * stripe (the stripe could be inconsistent again
7062                          * depending on which disks are used to calculate
7063                          * parity); the disk is broken; The stripe data of this
7064                          * disk is lost.
7065                          */
7066                         if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7067                             !bdev_get_queue(rdev->bdev)->
7068                                                 limits.discard_zeroes_data)
7069                                 discard_supported = false;
7070                         /* Unfortunately, discard_zeroes_data is not currently
7071                          * a guarantee - just a hint.  So we only allow DISCARD
7072                          * if the sysadmin has confirmed that only safe devices
7073                          * are in use by setting a module parameter.
7074                          */
7075                         if (!devices_handle_discard_safely) {
7076                                 if (discard_supported) {
7077                                         pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7078                                         pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7079                                 }
7080                                 discard_supported = false;
7081                         }
7082                 }
7083
7084                 if (discard_supported &&
7085                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7086                     mddev->queue->limits.discard_granularity >= stripe)
7087                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7088                                                 mddev->queue);
7089                 else
7090                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7091                                                 mddev->queue);
7092
7093                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7094         }
7095
7096         if (journal_dev) {
7097                 char b[BDEVNAME_SIZE];
7098
7099                 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7100                        mdname(mddev), bdevname(journal_dev->bdev, b));
7101                 r5l_init_log(conf, journal_dev);
7102         }
7103
7104         return 0;
7105 abort:
7106         md_unregister_thread(&mddev->thread);
7107         print_raid5_conf(conf);
7108         free_conf(conf);
7109         mddev->private = NULL;
7110         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7111         return -EIO;
7112 }
7113
7114 static void raid5_free(struct mddev *mddev, void *priv)
7115 {
7116         struct r5conf *conf = priv;
7117
7118         free_conf(conf);
7119         mddev->to_remove = &raid5_attrs_group;
7120 }
7121
7122 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7123 {
7124         struct r5conf *conf = mddev->private;
7125         int i;
7126
7127         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7128                 conf->chunk_sectors / 2, mddev->layout);
7129         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7130         rcu_read_lock();
7131         for (i = 0; i < conf->raid_disks; i++) {
7132                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7133                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7134         }
7135         rcu_read_unlock();
7136         seq_printf (seq, "]");
7137 }
7138
7139 static void print_raid5_conf (struct r5conf *conf)
7140 {
7141         int i;
7142         struct disk_info *tmp;
7143
7144         printk(KERN_DEBUG "RAID conf printout:\n");
7145         if (!conf) {
7146                 printk("(conf==NULL)\n");
7147                 return;
7148         }
7149         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7150                conf->raid_disks,
7151                conf->raid_disks - conf->mddev->degraded);
7152
7153         for (i = 0; i < conf->raid_disks; i++) {
7154                 char b[BDEVNAME_SIZE];
7155                 tmp = conf->disks + i;
7156                 if (tmp->rdev)
7157                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7158                                i, !test_bit(Faulty, &tmp->rdev->flags),
7159                                bdevname(tmp->rdev->bdev, b));
7160         }
7161 }
7162
7163 static int raid5_spare_active(struct mddev *mddev)
7164 {
7165         int i;
7166         struct r5conf *conf = mddev->private;
7167         struct disk_info *tmp;
7168         int count = 0;
7169         unsigned long flags;
7170
7171         for (i = 0; i < conf->raid_disks; i++) {
7172                 tmp = conf->disks + i;
7173                 if (tmp->replacement
7174                     && tmp->replacement->recovery_offset == MaxSector
7175                     && !test_bit(Faulty, &tmp->replacement->flags)
7176                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7177                         /* Replacement has just become active. */
7178                         if (!tmp->rdev
7179                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7180                                 count++;
7181                         if (tmp->rdev) {
7182                                 /* Replaced device not technically faulty,
7183                                  * but we need to be sure it gets removed
7184                                  * and never re-added.
7185                                  */
7186                                 set_bit(Faulty, &tmp->rdev->flags);
7187                                 sysfs_notify_dirent_safe(
7188                                         tmp->rdev->sysfs_state);
7189                         }
7190                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7191                 } else if (tmp->rdev
7192                     && tmp->rdev->recovery_offset == MaxSector
7193                     && !test_bit(Faulty, &tmp->rdev->flags)
7194                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7195                         count++;
7196                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7197                 }
7198         }
7199         spin_lock_irqsave(&conf->device_lock, flags);
7200         mddev->degraded = calc_degraded(conf);
7201         spin_unlock_irqrestore(&conf->device_lock, flags);
7202         print_raid5_conf(conf);
7203         return count;
7204 }
7205
7206 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7207 {
7208         struct r5conf *conf = mddev->private;
7209         int err = 0;
7210         int number = rdev->raid_disk;
7211         struct md_rdev **rdevp;
7212         struct disk_info *p = conf->disks + number;
7213
7214         print_raid5_conf(conf);
7215         if (test_bit(Journal, &rdev->flags) && conf->log) {
7216                 struct r5l_log *log;
7217                 /*
7218                  * we can't wait pending write here, as this is called in
7219                  * raid5d, wait will deadlock.
7220                  */
7221                 if (atomic_read(&mddev->writes_pending))
7222                         return -EBUSY;
7223                 log = conf->log;
7224                 conf->log = NULL;
7225                 synchronize_rcu();
7226                 r5l_exit_log(log);
7227                 return 0;
7228         }
7229         if (rdev == p->rdev)
7230                 rdevp = &p->rdev;
7231         else if (rdev == p->replacement)
7232                 rdevp = &p->replacement;
7233         else
7234                 return 0;
7235
7236         if (number >= conf->raid_disks &&
7237             conf->reshape_progress == MaxSector)
7238                 clear_bit(In_sync, &rdev->flags);
7239
7240         if (test_bit(In_sync, &rdev->flags) ||
7241             atomic_read(&rdev->nr_pending)) {
7242                 err = -EBUSY;
7243                 goto abort;
7244         }
7245         /* Only remove non-faulty devices if recovery
7246          * isn't possible.
7247          */
7248         if (!test_bit(Faulty, &rdev->flags) &&
7249             mddev->recovery_disabled != conf->recovery_disabled &&
7250             !has_failed(conf) &&
7251             (!p->replacement || p->replacement == rdev) &&
7252             number < conf->raid_disks) {
7253                 err = -EBUSY;
7254                 goto abort;
7255         }
7256         *rdevp = NULL;
7257         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7258                 synchronize_rcu();
7259                 if (atomic_read(&rdev->nr_pending)) {
7260                         /* lost the race, try later */
7261                         err = -EBUSY;
7262                         *rdevp = rdev;
7263                 }
7264         }
7265         if (p->replacement) {
7266                 /* We must have just cleared 'rdev' */
7267                 p->rdev = p->replacement;
7268                 clear_bit(Replacement, &p->replacement->flags);
7269                 smp_mb(); /* Make sure other CPUs may see both as identical
7270                            * but will never see neither - if they are careful
7271                            */
7272                 p->replacement = NULL;
7273                 clear_bit(WantReplacement, &rdev->flags);
7274         } else
7275                 /* We might have just removed the Replacement as faulty-
7276                  * clear the bit just in case
7277                  */
7278                 clear_bit(WantReplacement, &rdev->flags);
7279 abort:
7280
7281         print_raid5_conf(conf);
7282         return err;
7283 }
7284
7285 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7286 {
7287         struct r5conf *conf = mddev->private;
7288         int err = -EEXIST;
7289         int disk;
7290         struct disk_info *p;
7291         int first = 0;
7292         int last = conf->raid_disks - 1;
7293
7294         if (test_bit(Journal, &rdev->flags)) {
7295                 char b[BDEVNAME_SIZE];
7296                 if (conf->log)
7297                         return -EBUSY;
7298
7299                 rdev->raid_disk = 0;
7300                 /*
7301                  * The array is in readonly mode if journal is missing, so no
7302                  * write requests running. We should be safe
7303                  */
7304                 r5l_init_log(conf, rdev);
7305                 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7306                        mdname(mddev), bdevname(rdev->bdev, b));
7307                 return 0;
7308         }
7309         if (mddev->recovery_disabled == conf->recovery_disabled)
7310                 return -EBUSY;
7311
7312         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7313                 /* no point adding a device */
7314                 return -EINVAL;
7315
7316         if (rdev->raid_disk >= 0)
7317                 first = last = rdev->raid_disk;
7318
7319         /*
7320          * find the disk ... but prefer rdev->saved_raid_disk
7321          * if possible.
7322          */
7323         if (rdev->saved_raid_disk >= 0 &&
7324             rdev->saved_raid_disk >= first &&
7325             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7326                 first = rdev->saved_raid_disk;
7327
7328         for (disk = first; disk <= last; disk++) {
7329                 p = conf->disks + disk;
7330                 if (p->rdev == NULL) {
7331                         clear_bit(In_sync, &rdev->flags);
7332                         rdev->raid_disk = disk;
7333                         err = 0;
7334                         if (rdev->saved_raid_disk != disk)
7335                                 conf->fullsync = 1;
7336                         rcu_assign_pointer(p->rdev, rdev);
7337                         goto out;
7338                 }
7339         }
7340         for (disk = first; disk <= last; disk++) {
7341                 p = conf->disks + disk;
7342                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7343                     p->replacement == NULL) {
7344                         clear_bit(In_sync, &rdev->flags);
7345                         set_bit(Replacement, &rdev->flags);
7346                         rdev->raid_disk = disk;
7347                         err = 0;
7348                         conf->fullsync = 1;
7349                         rcu_assign_pointer(p->replacement, rdev);
7350                         break;
7351                 }
7352         }
7353 out:
7354         print_raid5_conf(conf);
7355         return err;
7356 }
7357
7358 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7359 {
7360         /* no resync is happening, and there is enough space
7361          * on all devices, so we can resize.
7362          * We need to make sure resync covers any new space.
7363          * If the array is shrinking we should possibly wait until
7364          * any io in the removed space completes, but it hardly seems
7365          * worth it.
7366          */
7367         sector_t newsize;
7368         struct r5conf *conf = mddev->private;
7369
7370         if (conf->log)
7371                 return -EINVAL;
7372         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7373         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7374         if (mddev->external_size &&
7375             mddev->array_sectors > newsize)
7376                 return -EINVAL;
7377         if (mddev->bitmap) {
7378                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7379                 if (ret)
7380                         return ret;
7381         }
7382         md_set_array_sectors(mddev, newsize);
7383         set_capacity(mddev->gendisk, mddev->array_sectors);
7384         revalidate_disk(mddev->gendisk);
7385         if (sectors > mddev->dev_sectors &&
7386             mddev->recovery_cp > mddev->dev_sectors) {
7387                 mddev->recovery_cp = mddev->dev_sectors;
7388                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7389         }
7390         mddev->dev_sectors = sectors;
7391         mddev->resync_max_sectors = sectors;
7392         return 0;
7393 }
7394
7395 static int check_stripe_cache(struct mddev *mddev)
7396 {
7397         /* Can only proceed if there are plenty of stripe_heads.
7398          * We need a minimum of one full stripe,, and for sensible progress
7399          * it is best to have about 4 times that.
7400          * If we require 4 times, then the default 256 4K stripe_heads will
7401          * allow for chunk sizes up to 256K, which is probably OK.
7402          * If the chunk size is greater, user-space should request more
7403          * stripe_heads first.
7404          */
7405         struct r5conf *conf = mddev->private;
7406         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7407             > conf->min_nr_stripes ||
7408             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7409             > conf->min_nr_stripes) {
7410                 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7411                        mdname(mddev),
7412                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7413                         / STRIPE_SIZE)*4);
7414                 return 0;
7415         }
7416         return 1;
7417 }
7418
7419 static int check_reshape(struct mddev *mddev)
7420 {
7421         struct r5conf *conf = mddev->private;
7422
7423         if (conf->log)
7424                 return -EINVAL;
7425         if (mddev->delta_disks == 0 &&
7426             mddev->new_layout == mddev->layout &&
7427             mddev->new_chunk_sectors == mddev->chunk_sectors)
7428                 return 0; /* nothing to do */
7429         if (has_failed(conf))
7430                 return -EINVAL;
7431         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7432                 /* We might be able to shrink, but the devices must
7433                  * be made bigger first.
7434                  * For raid6, 4 is the minimum size.
7435                  * Otherwise 2 is the minimum
7436                  */
7437                 int min = 2;
7438                 if (mddev->level == 6)
7439                         min = 4;
7440                 if (mddev->raid_disks + mddev->delta_disks < min)
7441                         return -EINVAL;
7442         }
7443
7444         if (!check_stripe_cache(mddev))
7445                 return -ENOSPC;
7446
7447         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7448             mddev->delta_disks > 0)
7449                 if (resize_chunks(conf,
7450                                   conf->previous_raid_disks
7451                                   + max(0, mddev->delta_disks),
7452                                   max(mddev->new_chunk_sectors,
7453                                       mddev->chunk_sectors)
7454                             ) < 0)
7455                         return -ENOMEM;
7456         return resize_stripes(conf, (conf->previous_raid_disks
7457                                      + mddev->delta_disks));
7458 }
7459
7460 static int raid5_start_reshape(struct mddev *mddev)
7461 {
7462         struct r5conf *conf = mddev->private;
7463         struct md_rdev *rdev;
7464         int spares = 0;
7465         unsigned long flags;
7466
7467         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7468                 return -EBUSY;
7469
7470         if (!check_stripe_cache(mddev))
7471                 return -ENOSPC;
7472
7473         if (has_failed(conf))
7474                 return -EINVAL;
7475
7476         rdev_for_each(rdev, mddev) {
7477                 if (!test_bit(In_sync, &rdev->flags)
7478                     && !test_bit(Faulty, &rdev->flags))
7479                         spares++;
7480         }
7481
7482         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7483                 /* Not enough devices even to make a degraded array
7484                  * of that size
7485                  */
7486                 return -EINVAL;
7487
7488         /* Refuse to reduce size of the array.  Any reductions in
7489          * array size must be through explicit setting of array_size
7490          * attribute.
7491          */
7492         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7493             < mddev->array_sectors) {
7494                 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7495                        "before number of disks\n", mdname(mddev));
7496                 return -EINVAL;
7497         }
7498
7499         atomic_set(&conf->reshape_stripes, 0);
7500         spin_lock_irq(&conf->device_lock);
7501         write_seqcount_begin(&conf->gen_lock);
7502         conf->previous_raid_disks = conf->raid_disks;
7503         conf->raid_disks += mddev->delta_disks;
7504         conf->prev_chunk_sectors = conf->chunk_sectors;
7505         conf->chunk_sectors = mddev->new_chunk_sectors;
7506         conf->prev_algo = conf->algorithm;
7507         conf->algorithm = mddev->new_layout;
7508         conf->generation++;
7509         /* Code that selects data_offset needs to see the generation update
7510          * if reshape_progress has been set - so a memory barrier needed.
7511          */
7512         smp_mb();
7513         if (mddev->reshape_backwards)
7514                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7515         else
7516                 conf->reshape_progress = 0;
7517         conf->reshape_safe = conf->reshape_progress;
7518         write_seqcount_end(&conf->gen_lock);
7519         spin_unlock_irq(&conf->device_lock);
7520
7521         /* Now make sure any requests that proceeded on the assumption
7522          * the reshape wasn't running - like Discard or Read - have
7523          * completed.
7524          */
7525         mddev_suspend(mddev);
7526         mddev_resume(mddev);
7527
7528         /* Add some new drives, as many as will fit.
7529          * We know there are enough to make the newly sized array work.
7530          * Don't add devices if we are reducing the number of
7531          * devices in the array.  This is because it is not possible
7532          * to correctly record the "partially reconstructed" state of
7533          * such devices during the reshape and confusion could result.
7534          */
7535         if (mddev->delta_disks >= 0) {
7536                 rdev_for_each(rdev, mddev)
7537                         if (rdev->raid_disk < 0 &&
7538                             !test_bit(Faulty, &rdev->flags)) {
7539                                 if (raid5_add_disk(mddev, rdev) == 0) {
7540                                         if (rdev->raid_disk
7541                                             >= conf->previous_raid_disks)
7542                                                 set_bit(In_sync, &rdev->flags);
7543                                         else
7544                                                 rdev->recovery_offset = 0;
7545
7546                                         if (sysfs_link_rdev(mddev, rdev))
7547                                                 /* Failure here is OK */;
7548                                 }
7549                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7550                                    && !test_bit(Faulty, &rdev->flags)) {
7551                                 /* This is a spare that was manually added */
7552                                 set_bit(In_sync, &rdev->flags);
7553                         }
7554
7555                 /* When a reshape changes the number of devices,
7556                  * ->degraded is measured against the larger of the
7557                  * pre and post number of devices.
7558                  */
7559                 spin_lock_irqsave(&conf->device_lock, flags);
7560                 mddev->degraded = calc_degraded(conf);
7561                 spin_unlock_irqrestore(&conf->device_lock, flags);
7562         }
7563         mddev->raid_disks = conf->raid_disks;
7564         mddev->reshape_position = conf->reshape_progress;
7565         set_bit(MD_CHANGE_DEVS, &mddev->flags);
7566
7567         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7568         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7569         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7570         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7571         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7572         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7573                                                 "reshape");
7574         if (!mddev->sync_thread) {
7575                 mddev->recovery = 0;
7576                 spin_lock_irq(&conf->device_lock);
7577                 write_seqcount_begin(&conf->gen_lock);
7578                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7579                 mddev->new_chunk_sectors =
7580                         conf->chunk_sectors = conf->prev_chunk_sectors;
7581                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7582                 rdev_for_each(rdev, mddev)
7583                         rdev->new_data_offset = rdev->data_offset;
7584                 smp_wmb();
7585                 conf->generation --;
7586                 conf->reshape_progress = MaxSector;
7587                 mddev->reshape_position = MaxSector;
7588                 write_seqcount_end(&conf->gen_lock);
7589                 spin_unlock_irq(&conf->device_lock);
7590                 return -EAGAIN;
7591         }
7592         conf->reshape_checkpoint = jiffies;
7593         md_wakeup_thread(mddev->sync_thread);
7594         md_new_event(mddev);
7595         return 0;
7596 }
7597
7598 /* This is called from the reshape thread and should make any
7599  * changes needed in 'conf'
7600  */
7601 static void end_reshape(struct r5conf *conf)
7602 {
7603
7604         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7605
7606                 spin_lock_irq(&conf->device_lock);
7607                 conf->previous_raid_disks = conf->raid_disks;
7608                 md_finish_reshape(conf->mddev);
7609                 smp_wmb();
7610                 conf->reshape_progress = MaxSector;
7611                 conf->mddev->reshape_position = MaxSector;
7612                 spin_unlock_irq(&conf->device_lock);
7613                 wake_up(&conf->wait_for_overlap);
7614
7615                 /* read-ahead size must cover two whole stripes, which is
7616                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7617                  */
7618                 if (conf->mddev->queue) {
7619                         int data_disks = conf->raid_disks - conf->max_degraded;
7620                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7621                                                    / PAGE_SIZE);
7622                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7623                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7624                 }
7625         }
7626 }
7627
7628 /* This is called from the raid5d thread with mddev_lock held.
7629  * It makes config changes to the device.
7630  */
7631 static void raid5_finish_reshape(struct mddev *mddev)
7632 {
7633         struct r5conf *conf = mddev->private;
7634
7635         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7636
7637                 if (mddev->delta_disks <= 0) {
7638                         int d;
7639                         spin_lock_irq(&conf->device_lock);
7640                         mddev->degraded = calc_degraded(conf);
7641                         spin_unlock_irq(&conf->device_lock);
7642                         for (d = conf->raid_disks ;
7643                              d < conf->raid_disks - mddev->delta_disks;
7644                              d++) {
7645                                 struct md_rdev *rdev = conf->disks[d].rdev;
7646                                 if (rdev)
7647                                         clear_bit(In_sync, &rdev->flags);
7648                                 rdev = conf->disks[d].replacement;
7649                                 if (rdev)
7650                                         clear_bit(In_sync, &rdev->flags);
7651                         }
7652                 }
7653                 mddev->layout = conf->algorithm;
7654                 mddev->chunk_sectors = conf->chunk_sectors;
7655                 mddev->reshape_position = MaxSector;
7656                 mddev->delta_disks = 0;
7657                 mddev->reshape_backwards = 0;
7658         }
7659 }
7660
7661 static void raid5_quiesce(struct mddev *mddev, int state)
7662 {
7663         struct r5conf *conf = mddev->private;
7664
7665         switch(state) {
7666         case 2: /* resume for a suspend */
7667                 wake_up(&conf->wait_for_overlap);
7668                 break;
7669
7670         case 1: /* stop all writes */
7671                 lock_all_device_hash_locks_irq(conf);
7672                 /* '2' tells resync/reshape to pause so that all
7673                  * active stripes can drain
7674                  */
7675                 conf->quiesce = 2;
7676                 wait_event_cmd(conf->wait_for_quiescent,
7677                                     atomic_read(&conf->active_stripes) == 0 &&
7678                                     atomic_read(&conf->active_aligned_reads) == 0,
7679                                     unlock_all_device_hash_locks_irq(conf),
7680                                     lock_all_device_hash_locks_irq(conf));
7681                 conf->quiesce = 1;
7682                 unlock_all_device_hash_locks_irq(conf);
7683                 /* allow reshape to continue */
7684                 wake_up(&conf->wait_for_overlap);
7685                 break;
7686
7687         case 0: /* re-enable writes */
7688                 lock_all_device_hash_locks_irq(conf);
7689                 conf->quiesce = 0;
7690                 wake_up(&conf->wait_for_quiescent);
7691                 wake_up(&conf->wait_for_overlap);
7692                 unlock_all_device_hash_locks_irq(conf);
7693                 break;
7694         }
7695         r5l_quiesce(conf->log, state);
7696 }
7697
7698 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7699 {
7700         struct r0conf *raid0_conf = mddev->private;
7701         sector_t sectors;
7702
7703         /* for raid0 takeover only one zone is supported */
7704         if (raid0_conf->nr_strip_zones > 1) {
7705                 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7706                        mdname(mddev));
7707                 return ERR_PTR(-EINVAL);
7708         }
7709
7710         sectors = raid0_conf->strip_zone[0].zone_end;
7711         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7712         mddev->dev_sectors = sectors;
7713         mddev->new_level = level;
7714         mddev->new_layout = ALGORITHM_PARITY_N;
7715         mddev->new_chunk_sectors = mddev->chunk_sectors;
7716         mddev->raid_disks += 1;
7717         mddev->delta_disks = 1;
7718         /* make sure it will be not marked as dirty */
7719         mddev->recovery_cp = MaxSector;
7720
7721         return setup_conf(mddev);
7722 }
7723
7724 static void *raid5_takeover_raid1(struct mddev *mddev)
7725 {
7726         int chunksect;
7727
7728         if (mddev->raid_disks != 2 ||
7729             mddev->degraded > 1)
7730                 return ERR_PTR(-EINVAL);
7731
7732         /* Should check if there are write-behind devices? */
7733
7734         chunksect = 64*2; /* 64K by default */
7735
7736         /* The array must be an exact multiple of chunksize */
7737         while (chunksect && (mddev->array_sectors & (chunksect-1)))
7738                 chunksect >>= 1;
7739
7740         if ((chunksect<<9) < STRIPE_SIZE)
7741                 /* array size does not allow a suitable chunk size */
7742                 return ERR_PTR(-EINVAL);
7743
7744         mddev->new_level = 5;
7745         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7746         mddev->new_chunk_sectors = chunksect;
7747
7748         return setup_conf(mddev);
7749 }
7750
7751 static void *raid5_takeover_raid6(struct mddev *mddev)
7752 {
7753         int new_layout;
7754
7755         switch (mddev->layout) {
7756         case ALGORITHM_LEFT_ASYMMETRIC_6:
7757                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7758                 break;
7759         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7760                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7761                 break;
7762         case ALGORITHM_LEFT_SYMMETRIC_6:
7763                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7764                 break;
7765         case ALGORITHM_RIGHT_SYMMETRIC_6:
7766                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7767                 break;
7768         case ALGORITHM_PARITY_0_6:
7769                 new_layout = ALGORITHM_PARITY_0;
7770                 break;
7771         case ALGORITHM_PARITY_N:
7772                 new_layout = ALGORITHM_PARITY_N;
7773                 break;
7774         default:
7775                 return ERR_PTR(-EINVAL);
7776         }
7777         mddev->new_level = 5;
7778         mddev->new_layout = new_layout;
7779         mddev->delta_disks = -1;
7780         mddev->raid_disks -= 1;
7781         return setup_conf(mddev);
7782 }
7783
7784 static int raid5_check_reshape(struct mddev *mddev)
7785 {
7786         /* For a 2-drive array, the layout and chunk size can be changed
7787          * immediately as not restriping is needed.
7788          * For larger arrays we record the new value - after validation
7789          * to be used by a reshape pass.
7790          */
7791         struct r5conf *conf = mddev->private;
7792         int new_chunk = mddev->new_chunk_sectors;
7793
7794         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7795                 return -EINVAL;
7796         if (new_chunk > 0) {
7797                 if (!is_power_of_2(new_chunk))
7798                         return -EINVAL;
7799                 if (new_chunk < (PAGE_SIZE>>9))
7800                         return -EINVAL;
7801                 if (mddev->array_sectors & (new_chunk-1))
7802                         /* not factor of array size */
7803                         return -EINVAL;
7804         }
7805
7806         /* They look valid */
7807
7808         if (mddev->raid_disks == 2) {
7809                 /* can make the change immediately */
7810                 if (mddev->new_layout >= 0) {
7811                         conf->algorithm = mddev->new_layout;
7812                         mddev->layout = mddev->new_layout;
7813                 }
7814                 if (new_chunk > 0) {
7815                         conf->chunk_sectors = new_chunk ;
7816                         mddev->chunk_sectors = new_chunk;
7817                 }
7818                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7819                 md_wakeup_thread(mddev->thread);
7820         }
7821         return check_reshape(mddev);
7822 }
7823
7824 static int raid6_check_reshape(struct mddev *mddev)
7825 {
7826         int new_chunk = mddev->new_chunk_sectors;
7827
7828         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7829                 return -EINVAL;
7830         if (new_chunk > 0) {
7831                 if (!is_power_of_2(new_chunk))
7832                         return -EINVAL;
7833                 if (new_chunk < (PAGE_SIZE >> 9))
7834                         return -EINVAL;
7835                 if (mddev->array_sectors & (new_chunk-1))
7836                         /* not factor of array size */
7837                         return -EINVAL;
7838         }
7839
7840         /* They look valid */
7841         return check_reshape(mddev);
7842 }
7843
7844 static void *raid5_takeover(struct mddev *mddev)
7845 {
7846         /* raid5 can take over:
7847          *  raid0 - if there is only one strip zone - make it a raid4 layout
7848          *  raid1 - if there are two drives.  We need to know the chunk size
7849          *  raid4 - trivial - just use a raid4 layout.
7850          *  raid6 - Providing it is a *_6 layout
7851          */
7852         if (mddev->level == 0)
7853                 return raid45_takeover_raid0(mddev, 5);
7854         if (mddev->level == 1)
7855                 return raid5_takeover_raid1(mddev);
7856         if (mddev->level == 4) {
7857                 mddev->new_layout = ALGORITHM_PARITY_N;
7858                 mddev->new_level = 5;
7859                 return setup_conf(mddev);
7860         }
7861         if (mddev->level == 6)
7862                 return raid5_takeover_raid6(mddev);
7863
7864         return ERR_PTR(-EINVAL);
7865 }
7866
7867 static void *raid4_takeover(struct mddev *mddev)
7868 {
7869         /* raid4 can take over:
7870          *  raid0 - if there is only one strip zone
7871          *  raid5 - if layout is right
7872          */
7873         if (mddev->level == 0)
7874                 return raid45_takeover_raid0(mddev, 4);
7875         if (mddev->level == 5 &&
7876             mddev->layout == ALGORITHM_PARITY_N) {
7877                 mddev->new_layout = 0;
7878                 mddev->new_level = 4;
7879                 return setup_conf(mddev);
7880         }
7881         return ERR_PTR(-EINVAL);
7882 }
7883
7884 static struct md_personality raid5_personality;
7885
7886 static void *raid6_takeover(struct mddev *mddev)
7887 {
7888         /* Currently can only take over a raid5.  We map the
7889          * personality to an equivalent raid6 personality
7890          * with the Q block at the end.
7891          */
7892         int new_layout;
7893
7894         if (mddev->pers != &raid5_personality)
7895                 return ERR_PTR(-EINVAL);
7896         if (mddev->degraded > 1)
7897                 return ERR_PTR(-EINVAL);
7898         if (mddev->raid_disks > 253)
7899                 return ERR_PTR(-EINVAL);
7900         if (mddev->raid_disks < 3)
7901                 return ERR_PTR(-EINVAL);
7902
7903         switch (mddev->layout) {
7904         case ALGORITHM_LEFT_ASYMMETRIC:
7905                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7906                 break;
7907         case ALGORITHM_RIGHT_ASYMMETRIC:
7908                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7909                 break;
7910         case ALGORITHM_LEFT_SYMMETRIC:
7911                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7912                 break;
7913         case ALGORITHM_RIGHT_SYMMETRIC:
7914                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7915                 break;
7916         case ALGORITHM_PARITY_0:
7917                 new_layout = ALGORITHM_PARITY_0_6;
7918                 break;
7919         case ALGORITHM_PARITY_N:
7920                 new_layout = ALGORITHM_PARITY_N;
7921                 break;
7922         default:
7923                 return ERR_PTR(-EINVAL);
7924         }
7925         mddev->new_level = 6;
7926         mddev->new_layout = new_layout;
7927         mddev->delta_disks = 1;
7928         mddev->raid_disks += 1;
7929         return setup_conf(mddev);
7930 }
7931
7932 static struct md_personality raid6_personality =
7933 {
7934         .name           = "raid6",
7935         .level          = 6,
7936         .owner          = THIS_MODULE,
7937         .make_request   = raid5_make_request,
7938         .run            = raid5_run,
7939         .free           = raid5_free,
7940         .status         = raid5_status,
7941         .error_handler  = raid5_error,
7942         .hot_add_disk   = raid5_add_disk,
7943         .hot_remove_disk= raid5_remove_disk,
7944         .spare_active   = raid5_spare_active,
7945         .sync_request   = raid5_sync_request,
7946         .resize         = raid5_resize,
7947         .size           = raid5_size,
7948         .check_reshape  = raid6_check_reshape,
7949         .start_reshape  = raid5_start_reshape,
7950         .finish_reshape = raid5_finish_reshape,
7951         .quiesce        = raid5_quiesce,
7952         .takeover       = raid6_takeover,
7953         .congested      = raid5_congested,
7954 };
7955 static struct md_personality raid5_personality =
7956 {
7957         .name           = "raid5",
7958         .level          = 5,
7959         .owner          = THIS_MODULE,
7960         .make_request   = raid5_make_request,
7961         .run            = raid5_run,
7962         .free           = raid5_free,
7963         .status         = raid5_status,
7964         .error_handler  = raid5_error,
7965         .hot_add_disk   = raid5_add_disk,
7966         .hot_remove_disk= raid5_remove_disk,
7967         .spare_active   = raid5_spare_active,
7968         .sync_request   = raid5_sync_request,
7969         .resize         = raid5_resize,
7970         .size           = raid5_size,
7971         .check_reshape  = raid5_check_reshape,
7972         .start_reshape  = raid5_start_reshape,
7973         .finish_reshape = raid5_finish_reshape,
7974         .quiesce        = raid5_quiesce,
7975         .takeover       = raid5_takeover,
7976         .congested      = raid5_congested,
7977 };
7978
7979 static struct md_personality raid4_personality =
7980 {
7981         .name           = "raid4",
7982         .level          = 4,
7983         .owner          = THIS_MODULE,
7984         .make_request   = raid5_make_request,
7985         .run            = raid5_run,
7986         .free           = raid5_free,
7987         .status         = raid5_status,
7988         .error_handler  = raid5_error,
7989         .hot_add_disk   = raid5_add_disk,
7990         .hot_remove_disk= raid5_remove_disk,
7991         .spare_active   = raid5_spare_active,
7992         .sync_request   = raid5_sync_request,
7993         .resize         = raid5_resize,
7994         .size           = raid5_size,
7995         .check_reshape  = raid5_check_reshape,
7996         .start_reshape  = raid5_start_reshape,
7997         .finish_reshape = raid5_finish_reshape,
7998         .quiesce        = raid5_quiesce,
7999         .takeover       = raid4_takeover,
8000         .congested      = raid5_congested,
8001 };
8002
8003 static int __init raid5_init(void)
8004 {
8005         int ret;
8006
8007         raid5_wq = alloc_workqueue("raid5wq",
8008                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8009         if (!raid5_wq)
8010                 return -ENOMEM;
8011
8012         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8013                                       "md/raid5:prepare",
8014                                       raid456_cpu_up_prepare,
8015                                       raid456_cpu_dead);
8016         if (ret) {
8017                 destroy_workqueue(raid5_wq);
8018                 return ret;
8019         }
8020         register_md_personality(&raid6_personality);
8021         register_md_personality(&raid5_personality);
8022         register_md_personality(&raid4_personality);
8023         return 0;
8024 }
8025
8026 static void raid5_exit(void)
8027 {
8028         unregister_md_personality(&raid6_personality);
8029         unregister_md_personality(&raid5_personality);
8030         unregister_md_personality(&raid4_personality);
8031         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8032         destroy_workqueue(raid5_wq);
8033 }
8034
8035 module_init(raid5_init);
8036 module_exit(raid5_exit);
8037 MODULE_LICENSE("GPL");
8038 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8039 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8040 MODULE_ALIAS("md-raid5");
8041 MODULE_ALIAS("md-raid4");
8042 MODULE_ALIAS("md-level-5");
8043 MODULE_ALIAS("md-level-4");
8044 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8045 MODULE_ALIAS("md-raid6");
8046 MODULE_ALIAS("md-level-6");
8047
8048 /* This used to be two separate modules, they were: */
8049 MODULE_ALIAS("raid5");
8050 MODULE_ALIAS("raid6");