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         local_irq_disable();
 114         spin_lock(conf->hash_locks);
 115         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
 116                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
 117         spin_lock(&conf->device_lock);
 118 }
 119
 120 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 121 {
 122         int i;
 123         spin_unlock(&conf->device_lock);
 124         for (i = NR_STRIPE_HASH_LOCKS; i; i--)
 125                 spin_unlock(conf->hash_locks + i - 1);
 126         local_irq_enable();
 127 }
 128
 129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
 130  * order without overlap.  There may be several bio's per stripe+device, and
 131  * a bio could span several devices.
 132  * When walking this list for a particular stripe+device, we must never proceed
 133  * beyond a bio that extends past this device, as the next bio might no longer
 134  * be valid.
 135  * This function is used to determine the 'next' bio in the list, given the sector
 136  * of the current stripe+device
 137  */
 138 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
 139 {
 140         int sectors = bio_sectors(bio);
 141         if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
 142                 return bio->bi_next;
 143         else
 144                 return NULL;
 145 }
 146
 147 /*
 148  * We maintain a biased count of active stripes in the bottom 16 bits of
 149  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
 150  */
 151 static inline int raid5_bi_processed_stripes(struct bio *bio)
 152 {
 153         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 154         return (atomic_read(segments) >> 16) & 0xffff;
 155 }
 156
 157 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
 158 {
 159         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 160         return atomic_sub_return(1, segments) & 0xffff;
 161 }
 162
 163 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
 164 {
 165         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 166         atomic_inc(segments);
 167 }
 168
 169 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
 170         unsigned int cnt)
 171 {
 172         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 173         int old, new;
 174
 175         do {
 176                 old = atomic_read(segments);
 177                 new = (old & 0xffff) | (cnt << 16);
 178         } while (atomic_cmpxchg(segments, old, new) != old);
 179 }
 180
 181 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
 182 {
 183         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 184         atomic_set(segments, cnt);
 185 }
 186
 187 /* Find first data disk in a raid6 stripe */
 188 static inline int raid6_d0(struct stripe_head *sh)
 189 {
 190         if (sh->ddf_layout)
 191                 /* ddf always start from first device */
 192                 return 0;
 193         /* md starts just after Q block */
 194         if (sh->qd_idx == sh->disks - 1)
 195                 return 0;
 196         else
 197                 return sh->qd_idx + 1;
 198 }
 199 static inline int raid6_next_disk(int disk, int raid_disks)
 200 {
 201         disk++;
 202         return (disk < raid_disks) ? disk : 0;
 203 }
 204
 205 /* When walking through the disks in a raid5, starting at raid6_d0,
 206  * We need to map each disk to a 'slot', where the data disks are slot
 207  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 208  * is raid_disks-1.  This help does that mapping.
 209  */
 210 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 211                              int *count, int syndrome_disks)
 212 {
 213         int slot = *count;
 214
 215         if (sh->ddf_layout)
 216                 (*count)++;
 217         if (idx == sh->pd_idx)
 218                 return syndrome_disks;
 219         if (idx == sh->qd_idx)
 220                 return syndrome_disks + 1;
 221         if (!sh->ddf_layout)
 222                 (*count)++;
 223         return slot;
 224 }
 225
 226 static void return_io(struct bio_list *return_bi)
 227 {
 228         struct bio *bi;
 229         while ((bi = bio_list_pop(return_bi)) != NULL) {
 230                 bi->bi_iter.bi_size = 0;
 231                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
 232                                          bi, 0);
 233                 bio_endio(bi);
 234         }
 235 }
 236
 237 static void print_raid5_conf (struct r5conf *conf);
 238
 239 static int stripe_operations_active(struct stripe_head *sh)
 240 {
 241         return sh->check_state || sh->reconstruct_state ||
 242                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 243                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 244 }
 245
 246 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 247 {
 248         struct r5conf *conf = sh->raid_conf;
 249         struct r5worker_group *group;
 250         int thread_cnt;
 251         int i, cpu = sh->cpu;
 252
 253         if (!cpu_online(cpu)) {
 254                 cpu = cpumask_any(cpu_online_mask);
 255                 sh->cpu = cpu;
 256         }
 257
 258         if (list_empty(&sh->lru)) {
 259                 struct r5worker_group *group;
 260                 group = conf->worker_groups + cpu_to_group(cpu);
 261                 list_add_tail(&sh->lru, &group->handle_list);
 262                 group->stripes_cnt++;
 263                 sh->group = group;
 264         }
 265
 266         if (conf->worker_cnt_per_group == 0) {
 267                 md_wakeup_thread(conf->mddev->thread);
 268                 return;
 269         }
 270
 271         group = conf->worker_groups + cpu_to_group(sh->cpu);
 272
 273         group->workers[0].working = true;
 274         /* at least one worker should run to avoid race */
 275         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 276
 277         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 278         /* wakeup more workers */
 279         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 280                 if (group->workers[i].working == false) {
 281                         group->workers[i].working = true;
 282                         queue_work_on(sh->cpu, raid5_wq,
 283                                       &group->workers[i].work);
 284                         thread_cnt--;
 285                 }
 286         }
 287 }
 288
 289 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 290                               struct list_head *temp_inactive_list)
 291 {
 292         BUG_ON(!list_empty(&sh->lru));
 293         BUG_ON(atomic_read(&conf->active_stripes)==0);
 294         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 295                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 296                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 297                         list_add_tail(&sh->lru, &conf->delayed_list);
 298                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 299                            sh->bm_seq - conf->seq_write > 0)
 300                         list_add_tail(&sh->lru, &conf->bitmap_list);
 301                 else {
 302                         clear_bit(STRIPE_DELAYED, &sh->state);
 303                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 304                         if (conf->worker_cnt_per_group == 0) {
 305                                 list_add_tail(&sh->lru, &conf->handle_list);
 306                         } else {
 307                                 raid5_wakeup_stripe_thread(sh);
 308                                 return;
 309                         }
 310                 }
 311                 md_wakeup_thread(conf->mddev->thread);
 312         } else {
 313                 BUG_ON(stripe_operations_active(sh));
 314                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 315                         if (atomic_dec_return(&conf->preread_active_stripes)
 316                             < IO_THRESHOLD)
 317                                 md_wakeup_thread(conf->mddev->thread);
 318                 atomic_dec(&conf->active_stripes);
 319                 if (!test_bit(STRIPE_EXPANDING, &sh->state))
 320                         list_add_tail(&sh->lru, temp_inactive_list);
 321         }
 322 }
 323
 324 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 325                              struct list_head *temp_inactive_list)
 326 {
 327         if (atomic_dec_and_test(&sh->count))
 328                 do_release_stripe(conf, sh, temp_inactive_list);
 329 }
 330
 331 /*
 332  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 333  *
 334  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 335  * given time. Adding stripes only takes device lock, while deleting stripes
 336  * only takes hash lock.
 337  */
 338 static void release_inactive_stripe_list(struct r5conf *conf,
 339                                          struct list_head *temp_inactive_list,
 340                                          int hash)
 341 {
 342         int size;
 343         unsigned long do_wakeup = 0;
 344         int i = 0;
 345         unsigned long flags;
 346
 347         if (hash == NR_STRIPE_HASH_LOCKS) {
 348                 size = NR_STRIPE_HASH_LOCKS;
 349                 hash = NR_STRIPE_HASH_LOCKS - 1;
 350         } else
 351                 size = 1;
 352         while (size) {
 353                 struct list_head *list = &temp_inactive_list[size - 1];
 354
 355                 /*
 356                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 357                  * remove stripes from the list
 358                  */
 359                 if (!list_empty_careful(list)) {
 360                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 361                         if (list_empty(conf->inactive_list + hash) &&
 362                             !list_empty(list))
 363                                 atomic_dec(&conf->empty_inactive_list_nr);
 364                         list_splice_tail_init(list, conf->inactive_list + hash);
 365                         do_wakeup |= 1 << hash;
 366                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 367                 }
 368                 size--;
 369                 hash--;
 370         }
 371
 372         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
 373                 if (do_wakeup & (1 << i))
 374                         wake_up(&conf->wait_for_stripe[i]);
 375         }
 376
 377         if (do_wakeup) {
 378                 if (atomic_read(&conf->active_stripes) == 0)
 379                         wake_up(&conf->wait_for_quiescent);
 380                 if (conf->retry_read_aligned)
 381                         md_wakeup_thread(conf->mddev->thread);
 382         }
 383 }
 384
 385 /* should hold conf->device_lock already */
 386 static int release_stripe_list(struct r5conf *conf,
 387                                struct list_head *temp_inactive_list)
 388 {
 389         struct stripe_head *sh;
 390         int count = 0;
 391         struct llist_node *head;
 392
 393         head = llist_del_all(&conf->released_stripes);
 394         head = llist_reverse_order(head);
 395         while (head) {
 396                 int hash;
 397
 398                 sh = llist_entry(head, struct stripe_head, release_list);
 399                 head = llist_next(head);
 400                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 401                 smp_mb();
 402                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 403                 /*
 404                  * Don't worry the bit is set here, because if the bit is set
 405                  * again, the count is always > 1. This is true for
 406                  * STRIPE_ON_UNPLUG_LIST bit too.
 407                  */
 408                 hash = sh->hash_lock_index;
 409                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 410                 count++;
 411         }
 412
 413         return count;
 414 }
 415
 416 void raid5_release_stripe(struct stripe_head *sh)
 417 {
 418         struct r5conf *conf = sh->raid_conf;
 419         unsigned long flags;
 420         struct list_head list;
 421         int hash;
 422         bool wakeup;
 423
 424         /* Avoid release_list until the last reference.
 425          */
 426         if (atomic_add_unless(&sh->count, -1, 1))
 427                 return;
 428
 429         if (unlikely(!conf->mddev->thread) ||
 430                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 431                 goto slow_path;
 432         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 433         if (wakeup)
 434                 md_wakeup_thread(conf->mddev->thread);
 435         return;
 436 slow_path:
 437         local_irq_save(flags);
 438         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 439         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 440                 INIT_LIST_HEAD(&list);
 441                 hash = sh->hash_lock_index;
 442                 do_release_stripe(conf, sh, &list);
 443                 spin_unlock(&conf->device_lock);
 444                 release_inactive_stripe_list(conf, &list, hash);
 445         }
 446         local_irq_restore(flags);
 447 }
 448
 449 static inline void remove_hash(struct stripe_head *sh)
 450 {
 451         pr_debug("remove_hash(), stripe %llu\n",
 452                 (unsigned long long)sh->sector);
 453
 454         hlist_del_init(&sh->hash);
 455 }
 456
 457 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 458 {
 459         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 460
 461         pr_debug("insert_hash(), stripe %llu\n",
 462                 (unsigned long long)sh->sector);
 463
 464         hlist_add_head(&sh->hash, hp);
 465 }
 466
 467 /* find an idle stripe, make sure it is unhashed, and return it. */
 468 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 469 {
 470         struct stripe_head *sh = NULL;
 471         struct list_head *first;
 472
 473         if (list_empty(conf->inactive_list + hash))
 474                 goto out;
 475         first = (conf->inactive_list + hash)->next;
 476         sh = list_entry(first, struct stripe_head, lru);
 477         list_del_init(first);
 478         remove_hash(sh);
 479         atomic_inc(&conf->active_stripes);
 480         BUG_ON(hash != sh->hash_lock_index);
 481         if (list_empty(conf->inactive_list + hash))
 482                 atomic_inc(&conf->empty_inactive_list_nr);
 483 out:
 484         return sh;
 485 }
 486
 487 static void shrink_buffers(struct stripe_head *sh)
 488 {
 489         struct page *p;
 490         int i;
 491         int num = sh->raid_conf->pool_size;
 492
 493         for (i = 0; i < num ; i++) {
 494                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 495                 p = sh->dev[i].page;
 496                 if (!p)
 497                         continue;
 498                 sh->dev[i].page = NULL;
 499                 put_page(p);
 500         }
 501 }
 502
 503 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 504 {
 505         int i;
 506         int num = sh->raid_conf->pool_size;
 507
 508         for (i = 0; i < num; i++) {
 509                 struct page *page;
 510
 511                 if (!(page = alloc_page(gfp))) {
 512                         return 1;
 513                 }
 514                 sh->dev[i].page = page;
 515                 sh->dev[i].orig_page = page;
 516         }
 517         return 0;
 518 }
 519
 520 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 521 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 522                             struct stripe_head *sh);
 523
 524 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 525 {
 526         struct r5conf *conf = sh->raid_conf;
 527         int i, seq;
 528
 529         BUG_ON(atomic_read(&sh->count) != 0);
 530         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 531         BUG_ON(stripe_operations_active(sh));
 532         BUG_ON(sh->batch_head);
 533
 534         pr_debug("init_stripe called, stripe %llu\n",
 535                 (unsigned long long)sector);
 536 retry:
 537         seq = read_seqcount_begin(&conf->gen_lock);
 538         sh->generation = conf->generation - previous;
 539         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 540         sh->sector = sector;
 541         stripe_set_idx(sector, conf, previous, sh);
 542         sh->state = 0;
 543
 544         for (i = sh->disks; i--; ) {
 545                 struct r5dev *dev = &sh->dev[i];
 546
 547                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 548                     test_bit(R5_LOCKED, &dev->flags)) {
 549                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 550                                (unsigned long long)sh->sector, i, dev->toread,
 551                                dev->read, dev->towrite, dev->written,
 552                                test_bit(R5_LOCKED, &dev->flags));
 553                         WARN_ON(1);
 554                 }
 555                 dev->flags = 0;
 556                 raid5_build_block(sh, i, previous);
 557         }
 558         if (read_seqcount_retry(&conf->gen_lock, seq))
 559                 goto retry;
 560         sh->overwrite_disks = 0;
 561         insert_hash(conf, sh);
 562         sh->cpu = smp_processor_id();
 563         set_bit(STRIPE_BATCH_READY, &sh->state);
 564 }
 565
 566 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 567                                          short generation)
 568 {
 569         struct stripe_head *sh;
 570
 571         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 572         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 573                 if (sh->sector == sector && sh->generation == generation)
 574                         return sh;
 575         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 576         return NULL;
 577 }
 578
 579 /*
 580  * Need to check if array has failed when deciding whether to:
 581  *  - start an array
 582  *  - remove non-faulty devices
 583  *  - add a spare
 584  *  - allow a reshape
 585  * This determination is simple when no reshape is happening.
 586  * However if there is a reshape, we need to carefully check
 587  * both the before and after sections.
 588  * This is because some failed devices may only affect one
 589  * of the two sections, and some non-in_sync devices may
 590  * be insync in the section most affected by failed devices.
 591  */
 592 static int calc_degraded(struct r5conf *conf)
 593 {
 594         int degraded, degraded2;
 595         int i;
 596
 597         rcu_read_lock();
 598         degraded = 0;
 599         for (i = 0; i < conf->previous_raid_disks; i++) {
 600                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 601                 if (rdev && test_bit(Faulty, &rdev->flags))
 602                         rdev = rcu_dereference(conf->disks[i].replacement);
 603                 if (!rdev || test_bit(Faulty, &rdev->flags))
 604                         degraded++;
 605                 else if (test_bit(In_sync, &rdev->flags))
 606                         ;
 607                 else
 608                         /* not in-sync or faulty.
 609                          * If the reshape increases the number of devices,
 610                          * this is being recovered by the reshape, so
 611                          * this 'previous' section is not in_sync.
 612                          * If the number of devices is being reduced however,
 613                          * the device can only be part of the array if
 614                          * we are reverting a reshape, so this section will
 615                          * be in-sync.
 616                          */
 617                         if (conf->raid_disks >= conf->previous_raid_disks)
 618                                 degraded++;
 619         }
 620         rcu_read_unlock();
 621         if (conf->raid_disks == conf->previous_raid_disks)
 622                 return degraded;
 623         rcu_read_lock();
 624         degraded2 = 0;
 625         for (i = 0; i < conf->raid_disks; i++) {
 626                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 627                 if (rdev && test_bit(Faulty, &rdev->flags))
 628                         rdev = rcu_dereference(conf->disks[i].replacement);
 629                 if (!rdev || test_bit(Faulty, &rdev->flags))
 630                         degraded2++;
 631                 else if (test_bit(In_sync, &rdev->flags))
 632                         ;
 633                 else
 634                         /* not in-sync or faulty.
 635                          * If reshape increases the number of devices, this
 636                          * section has already been recovered, else it
 637                          * almost certainly hasn't.
 638                          */
 639                         if (conf->raid_disks <= conf->previous_raid_disks)
 640                                 degraded2++;
 641         }
 642         rcu_read_unlock();
 643         if (degraded2 > degraded)
 644                 return degraded2;
 645         return degraded;
 646 }
 647
 648 static int has_failed(struct r5conf *conf)
 649 {
 650         int degraded;
 651
 652         if (conf->mddev->reshape_position == MaxSector)
 653                 return conf->mddev->degraded > conf->max_degraded;
 654
 655         degraded = calc_degraded(conf);
 656         if (degraded > conf->max_degraded)
 657                 return 1;
 658         return 0;
 659 }
 660
 661 struct stripe_head *
 662 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 663                         int previous, int noblock, int noquiesce)
 664 {
 665         struct stripe_head *sh;
 666         int hash = stripe_hash_locks_hash(sector);
 667
 668         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 669
 670         spin_lock_irq(conf->hash_locks + hash);
 671
 672         do {
 673                 wait_event_lock_irq(conf->wait_for_quiescent,
 674                                     conf->quiesce == 0 || noquiesce,
 675                                     *(conf->hash_locks + hash));
 676                 sh = __find_stripe(conf, sector, conf->generation - previous);
 677                 if (!sh) {
 678                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 679                                 sh = get_free_stripe(conf, hash);
 680                                 if (!sh && !test_bit(R5_DID_ALLOC,
 681                                                      &conf->cache_state))
 682                                         set_bit(R5_ALLOC_MORE,
 683                                                 &conf->cache_state);
 684                         }
 685                         if (noblock && sh == NULL)
 686                                 break;
 687                         if (!sh) {
 688                                 set_bit(R5_INACTIVE_BLOCKED,
 689                                         &conf->cache_state);
 690                                 wait_event_exclusive_cmd(
 691                                         conf->wait_for_stripe[hash],
 692                                         !list_empty(conf->inactive_list + hash) &&
 693                                         (atomic_read(&conf->active_stripes)
 694                                          < (conf->max_nr_stripes * 3 / 4)
 695                                          || !test_bit(R5_INACTIVE_BLOCKED,
 696                                                       &conf->cache_state)),
 697                                         spin_unlock_irq(conf->hash_locks + hash),
 698                                         spin_lock_irq(conf->hash_locks + hash));
 699                                 clear_bit(R5_INACTIVE_BLOCKED,
 700                                           &conf->cache_state);
 701                         } else {
 702                                 init_stripe(sh, sector, previous);
 703                                 atomic_inc(&sh->count);
 704                         }
 705                 } else if (!atomic_inc_not_zero(&sh->count)) {
 706                         spin_lock(&conf->device_lock);
 707                         if (!atomic_read(&sh->count)) {
 708                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 709                                         atomic_inc(&conf->active_stripes);
 710                                 BUG_ON(list_empty(&sh->lru) &&
 711                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 712                                 list_del_init(&sh->lru);
 713                                 if (sh->group) {
 714                                         sh->group->stripes_cnt--;
 715                                         sh->group = NULL;
 716                                 }
 717                         }
 718                         atomic_inc(&sh->count);
 719                         spin_unlock(&conf->device_lock);
 720                 }
 721         } while (sh == NULL);
 722
 723         if (!list_empty(conf->inactive_list + hash))
 724                 wake_up(&conf->wait_for_stripe[hash]);
 725
 726         spin_unlock_irq(conf->hash_locks + hash);
 727         return sh;
 728 }
 729
 730 static bool is_full_stripe_write(struct stripe_head *sh)
 731 {
 732         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 733         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 734 }
 735
 736 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 737 {
 738         local_irq_disable();
 739         if (sh1 > sh2) {
 740                 spin_lock(&sh2->stripe_lock);
 741                 spin_lock_nested(&sh1->stripe_lock, 1);
 742         } else {
 743                 spin_lock(&sh1->stripe_lock);
 744                 spin_lock_nested(&sh2->stripe_lock, 1);
 745         }
 746 }
 747
 748 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 749 {
 750         spin_unlock(&sh1->stripe_lock);
 751         spin_unlock(&sh2->stripe_lock);
 752         local_irq_enable();
 753 }
 754
 755 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 756 static bool stripe_can_batch(struct stripe_head *sh)
 757 {
 758         struct r5conf *conf = sh->raid_conf;
 759
 760         if (conf->log)
 761                 return false;
 762         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 763                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 764                 is_full_stripe_write(sh);
 765 }
 766
 767 /* we only do back search */
 768 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 769 {
 770         struct stripe_head *head;
 771         sector_t head_sector, tmp_sec;
 772         int hash;
 773         int dd_idx;
 774
 775         if (!stripe_can_batch(sh))
 776                 return;
 777         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 778         tmp_sec = sh->sector;
 779         if (!sector_div(tmp_sec, conf->chunk_sectors))
 780                 return;
 781         head_sector = sh->sector - STRIPE_SECTORS;
 782
 783         hash = stripe_hash_locks_hash(head_sector);
 784         spin_lock_irq(conf->hash_locks + hash);
 785         head = __find_stripe(conf, head_sector, conf->generation);
 786         if (head && !atomic_inc_not_zero(&head->count)) {
 787                 spin_lock(&conf->device_lock);
 788                 if (!atomic_read(&head->count)) {
 789                         if (!test_bit(STRIPE_HANDLE, &head->state))
 790                                 atomic_inc(&conf->active_stripes);
 791                         BUG_ON(list_empty(&head->lru) &&
 792                                !test_bit(STRIPE_EXPANDING, &head->state));
 793                         list_del_init(&head->lru);
 794                         if (head->group) {
 795                                 head->group->stripes_cnt--;
 796                                 head->group = NULL;
 797                         }
 798                 }
 799                 atomic_inc(&head->count);
 800                 spin_unlock(&conf->device_lock);
 801         }
 802         spin_unlock_irq(conf->hash_locks + hash);
 803
 804         if (!head)
 805                 return;
 806         if (!stripe_can_batch(head))
 807                 goto out;
 808
 809         lock_two_stripes(head, sh);
 810         /* clear_batch_ready clear the flag */
 811         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 812                 goto unlock_out;
 813
 814         if (sh->batch_head)
 815                 goto unlock_out;
 816
 817         dd_idx = 0;
 818         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 819                 dd_idx++;
 820         if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
 821                 goto unlock_out;
 822
 823         if (head->batch_head) {
 824                 spin_lock(&head->batch_head->batch_lock);
 825                 /* This batch list is already running */
 826                 if (!stripe_can_batch(head)) {
 827                         spin_unlock(&head->batch_head->batch_lock);
 828                         goto unlock_out;
 829                 }
 830
 831                 /*
 832                  * at this point, head's BATCH_READY could be cleared, but we
 833                  * can still add the stripe to batch list
 834                  */
 835                 list_add(&sh->batch_list, &head->batch_list);
 836                 spin_unlock(&head->batch_head->batch_lock);
 837
 838                 sh->batch_head = head->batch_head;
 839         } else {
 840                 head->batch_head = head;
 841                 sh->batch_head = head->batch_head;
 842                 spin_lock(&head->batch_lock);
 843                 list_add_tail(&sh->batch_list, &head->batch_list);
 844                 spin_unlock(&head->batch_lock);
 845         }
 846
 847         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 848                 if (atomic_dec_return(&conf->preread_active_stripes)
 849                     < IO_THRESHOLD)
 850                         md_wakeup_thread(conf->mddev->thread);
 851
 852         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 853                 int seq = sh->bm_seq;
 854                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 855                     sh->batch_head->bm_seq > seq)
 856                         seq = sh->batch_head->bm_seq;
 857                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 858                 sh->batch_head->bm_seq = seq;
 859         }
 860
 861         atomic_inc(&sh->count);
 862 unlock_out:
 863         unlock_two_stripes(head, sh);
 864 out:
 865         raid5_release_stripe(head);
 866 }
 867
 868 /* Determine if 'data_offset' or 'new_data_offset' should be used
 869  * in this stripe_head.
 870  */
 871 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 872 {
 873         sector_t progress = conf->reshape_progress;
 874         /* Need a memory barrier to make sure we see the value
 875          * of conf->generation, or ->data_offset that was set before
 876          * reshape_progress was updated.
 877          */
 878         smp_rmb();
 879         if (progress == MaxSector)
 880                 return 0;
 881         if (sh->generation == conf->generation - 1)
 882                 return 0;
 883         /* We are in a reshape, and this is a new-generation stripe,
 884          * so use new_data_offset.
 885          */
 886         return 1;
 887 }
 888
 889 static void
 890 raid5_end_read_request(struct bio *bi);
 891 static void
 892 raid5_end_write_request(struct bio *bi);
 893
 894 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 895 {
 896         struct r5conf *conf = sh->raid_conf;
 897         int i, disks = sh->disks;
 898         struct stripe_head *head_sh = sh;
 899
 900         might_sleep();
 901
 902         if (r5l_write_stripe(conf->log, sh) == 0)
 903                 return;
 904         for (i = disks; i--; ) {
 905                 int rw;
 906                 int replace_only = 0;
 907                 struct bio *bi, *rbi;
 908                 struct md_rdev *rdev, *rrdev = NULL;
 909
 910                 sh = head_sh;
 911                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 912                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
 913                                 rw = WRITE_FUA;
 914                         else
 915                                 rw = WRITE;
 916                         if (test_bit(R5_Discard, &sh->dev[i].flags))
 917                                 rw |= REQ_DISCARD;
 918                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 919                         rw = READ;
 920                 else if (test_and_clear_bit(R5_WantReplace,
 921                                             &sh->dev[i].flags)) {
 922                         rw = WRITE;
 923                         replace_only = 1;
 924                 } else
 925                         continue;
 926                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
 927                         rw |= REQ_SYNC;
 928
 929 again:
 930                 bi = &sh->dev[i].req;
 931                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
 932
 933                 rcu_read_lock();
 934                 rrdev = rcu_dereference(conf->disks[i].replacement);
 935                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
 936                 rdev = rcu_dereference(conf->disks[i].rdev);
 937                 if (!rdev) {
 938                         rdev = rrdev;
 939                         rrdev = NULL;
 940                 }
 941                 if (rw & WRITE) {
 942                         if (replace_only)
 943                                 rdev = NULL;
 944                         if (rdev == rrdev)
 945                                 /* We raced and saw duplicates */
 946                                 rrdev = NULL;
 947                 } else {
 948                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
 949                                 rdev = rrdev;
 950                         rrdev = NULL;
 951                 }
 952
 953                 if (rdev && test_bit(Faulty, &rdev->flags))
 954                         rdev = NULL;
 955                 if (rdev)
 956                         atomic_inc(&rdev->nr_pending);
 957                 if (rrdev && test_bit(Faulty, &rrdev->flags))
 958                         rrdev = NULL;
 959                 if (rrdev)
 960                         atomic_inc(&rrdev->nr_pending);
 961                 rcu_read_unlock();
 962
 963                 /* We have already checked bad blocks for reads.  Now
 964                  * need to check for writes.  We never accept write errors
 965                  * on the replacement, so we don't to check rrdev.
 966                  */
 967                 while ((rw & WRITE) && rdev &&
 968                        test_bit(WriteErrorSeen, &rdev->flags)) {
 969                         sector_t first_bad;
 970                         int bad_sectors;
 971                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
 972                                               &first_bad, &bad_sectors);
 973                         if (!bad)
 974                                 break;
 975
 976                         if (bad < 0) {
 977                                 set_bit(BlockedBadBlocks, &rdev->flags);
 978                                 if (!conf->mddev->external &&
 979                                     conf->mddev->flags) {
 980                                         /* It is very unlikely, but we might
 981                                          * still need to write out the
 982                                          * bad block log - better give it
 983                                          * a chance*/
 984                                         md_check_recovery(conf->mddev);
 985                                 }
 986                                 /*
 987                                  * Because md_wait_for_blocked_rdev
 988                                  * will dec nr_pending, we must
 989                                  * increment it first.
 990                                  */
 991                                 atomic_inc(&rdev->nr_pending);
 992                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
 993                         } else {
 994                                 /* Acknowledged bad block - skip the write */
 995                                 rdev_dec_pending(rdev, conf->mddev);
 996                                 rdev = NULL;
 997                         }
 998                 }
 999
1000                 if (rdev) {
1001                         if (s->syncing || s->expanding || s->expanded
1002                             || s->replacing)
1003                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1004
1005                         set_bit(STRIPE_IO_STARTED, &sh->state);
1006
1007                         bio_reset(bi);
1008                         bi->bi_bdev = rdev->bdev;
1009                         bi->bi_rw = rw;
1010                         bi->bi_end_io = (rw & WRITE)
1011                                 ? raid5_end_write_request
1012                                 : raid5_end_read_request;
1013                         bi->bi_private = sh;
1014
1015                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1016                                 __func__, (unsigned long long)sh->sector,
1017                                 bi->bi_rw, i);
1018                         atomic_inc(&sh->count);
1019                         if (sh != head_sh)
1020                                 atomic_inc(&head_sh->count);
1021                         if (use_new_offset(conf, sh))
1022                                 bi->bi_iter.bi_sector = (sh->sector
1023                                                  + rdev->new_data_offset);
1024                         else
1025                                 bi->bi_iter.bi_sector = (sh->sector
1026                                                  + rdev->data_offset);
1027                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1028                                 bi->bi_rw |= REQ_NOMERGE;
1029
1030                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1031                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1032                         sh->dev[i].vec.bv_page = sh->dev[i].page;
1033                         bi->bi_vcnt = 1;
1034                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1035                         bi->bi_io_vec[0].bv_offset = 0;
1036                         bi->bi_iter.bi_size = STRIPE_SIZE;
1037                         /*
1038                          * If this is discard request, set bi_vcnt 0. We don't
1039                          * want to confuse SCSI because SCSI will replace payload
1040                          */
1041                         if (rw & REQ_DISCARD)
1042                                 bi->bi_vcnt = 0;
1043                         if (rrdev)
1044                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1045
1046                         if (conf->mddev->gendisk)
1047                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1048                                                       bi, disk_devt(conf->mddev->gendisk),
1049                                                       sh->dev[i].sector);
1050                         generic_make_request(bi);
1051                 }
1052                 if (rrdev) {
1053                         if (s->syncing || s->expanding || s->expanded
1054                             || s->replacing)
1055                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1056
1057                         set_bit(STRIPE_IO_STARTED, &sh->state);
1058
1059                         bio_reset(rbi);
1060                         rbi->bi_bdev = rrdev->bdev;
1061                         rbi->bi_rw = rw;
1062                         BUG_ON(!(rw & WRITE));
1063                         rbi->bi_end_io = raid5_end_write_request;
1064                         rbi->bi_private = sh;
1065
1066                         pr_debug("%s: for %llu schedule op %ld on "
1067                                  "replacement disc %d\n",
1068                                 __func__, (unsigned long long)sh->sector,
1069                                 rbi->bi_rw, i);
1070                         atomic_inc(&sh->count);
1071                         if (sh != head_sh)
1072                                 atomic_inc(&head_sh->count);
1073                         if (use_new_offset(conf, sh))
1074                                 rbi->bi_iter.bi_sector = (sh->sector
1075                                                   + rrdev->new_data_offset);
1076                         else
1077                                 rbi->bi_iter.bi_sector = (sh->sector
1078                                                   + rrdev->data_offset);
1079                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1080                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1081                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1082                         rbi->bi_vcnt = 1;
1083                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1084                         rbi->bi_io_vec[0].bv_offset = 0;
1085                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1086                         /*
1087                          * If this is discard request, set bi_vcnt 0. We don't
1088                          * want to confuse SCSI because SCSI will replace payload
1089                          */
1090                         if (rw & REQ_DISCARD)
1091                                 rbi->bi_vcnt = 0;
1092                         if (conf->mddev->gendisk)
1093                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1094                                                       rbi, disk_devt(conf->mddev->gendisk),
1095                                                       sh->dev[i].sector);
1096                         generic_make_request(rbi);
1097                 }
1098                 if (!rdev && !rrdev) {
1099                         if (rw & WRITE)
1100                                 set_bit(STRIPE_DEGRADED, &sh->state);
1101                         pr_debug("skip op %ld on disc %d for sector %llu\n",
1102                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1103                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1104                         set_bit(STRIPE_HANDLE, &sh->state);
1105                 }
1106
1107                 if (!head_sh->batch_head)
1108                         continue;
1109                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1110                                       batch_list);
1111                 if (sh != head_sh)
1112                         goto again;
1113         }
1114 }
1115
1116 static struct dma_async_tx_descriptor *
1117 async_copy_data(int frombio, struct bio *bio, struct page **page,
1118         sector_t sector, struct dma_async_tx_descriptor *tx,
1119         struct stripe_head *sh)
1120 {
1121         struct bio_vec bvl;
1122         struct bvec_iter iter;
1123         struct page *bio_page;
1124         int page_offset;
1125         struct async_submit_ctl submit;
1126         enum async_tx_flags flags = 0;
1127
1128         if (bio->bi_iter.bi_sector >= sector)
1129                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1130         else
1131                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1132
1133         if (frombio)
1134                 flags |= ASYNC_TX_FENCE;
1135         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1136
1137         bio_for_each_segment(bvl, bio, iter) {
1138                 int len = bvl.bv_len;
1139                 int clen;
1140                 int b_offset = 0;
1141
1142                 if (page_offset < 0) {
1143                         b_offset = -page_offset;
1144                         page_offset += b_offset;
1145                         len -= b_offset;
1146                 }
1147
1148                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1149                         clen = STRIPE_SIZE - page_offset;
1150                 else
1151                         clen = len;
1152
1153                 if (clen > 0) {
1154                         b_offset += bvl.bv_offset;
1155                         bio_page = bvl.bv_page;
1156                         if (frombio) {
1157                                 if (sh->raid_conf->skip_copy &&
1158                                     b_offset == 0 && page_offset == 0 &&
1159                                     clen == STRIPE_SIZE)
1160                                         *page = bio_page;
1161                                 else
1162                                         tx = async_memcpy(*page, bio_page, page_offset,
1163                                                   b_offset, clen, &submit);
1164                         } else
1165                                 tx = async_memcpy(bio_page, *page, b_offset,
1166                                                   page_offset, clen, &submit);
1167                 }
1168                 /* chain the operations */
1169                 submit.depend_tx = tx;
1170
1171                 if (clen < len) /* hit end of page */
1172                         break;
1173                 page_offset +=  len;
1174         }
1175
1176         return tx;
1177 }
1178
1179 static void ops_complete_biofill(void *stripe_head_ref)
1180 {
1181         struct stripe_head *sh = stripe_head_ref;
1182         struct bio_list return_bi = BIO_EMPTY_LIST;
1183         int i;
1184
1185         pr_debug("%s: stripe %llu\n", __func__,
1186                 (unsigned long long)sh->sector);
1187
1188         /* clear completed biofills */
1189         for (i = sh->disks; i--; ) {
1190                 struct r5dev *dev = &sh->dev[i];
1191
1192                 /* acknowledge completion of a biofill operation */
1193                 /* and check if we need to reply to a read request,
1194                  * new R5_Wantfill requests are held off until
1195                  * !STRIPE_BIOFILL_RUN
1196                  */
1197                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1198                         struct bio *rbi, *rbi2;
1199
1200                         BUG_ON(!dev->read);
1201                         rbi = dev->read;
1202                         dev->read = NULL;
1203                         while (rbi && rbi->bi_iter.bi_sector <
1204                                 dev->sector + STRIPE_SECTORS) {
1205                                 rbi2 = r5_next_bio(rbi, dev->sector);
1206                                 if (!raid5_dec_bi_active_stripes(rbi))
1207                                         bio_list_add(&return_bi, rbi);
1208                                 rbi = rbi2;
1209                         }
1210                 }
1211         }
1212         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1213
1214         return_io(&return_bi);
1215
1216         set_bit(STRIPE_HANDLE, &sh->state);
1217         raid5_release_stripe(sh);
1218 }
1219
1220 static void ops_run_biofill(struct stripe_head *sh)
1221 {
1222         struct dma_async_tx_descriptor *tx = NULL;
1223         struct async_submit_ctl submit;
1224         int i;
1225
1226         BUG_ON(sh->batch_head);
1227         pr_debug("%s: stripe %llu\n", __func__,
1228                 (unsigned long long)sh->sector);
1229
1230         for (i = sh->disks; i--; ) {
1231                 struct r5dev *dev = &sh->dev[i];
1232                 if (test_bit(R5_Wantfill, &dev->flags)) {
1233                         struct bio *rbi;
1234                         spin_lock_irq(&sh->stripe_lock);
1235                         dev->read = rbi = dev->toread;
1236                         dev->toread = NULL;
1237                         spin_unlock_irq(&sh->stripe_lock);
1238                         while (rbi && rbi->bi_iter.bi_sector <
1239                                 dev->sector + STRIPE_SECTORS) {
1240                                 tx = async_copy_data(0, rbi, &dev->page,
1241                                         dev->sector, tx, sh);
1242                                 rbi = r5_next_bio(rbi, dev->sector);
1243                         }
1244                 }
1245         }
1246
1247         atomic_inc(&sh->count);
1248         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1249         async_trigger_callback(&submit);
1250 }
1251
1252 static void mark_target_uptodate(struct stripe_head *sh, int target)
1253 {
1254         struct r5dev *tgt;
1255
1256         if (target < 0)
1257                 return;
1258
1259         tgt = &sh->dev[target];
1260         set_bit(R5_UPTODATE, &tgt->flags);
1261         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1262         clear_bit(R5_Wantcompute, &tgt->flags);
1263 }
1264
1265 static void ops_complete_compute(void *stripe_head_ref)
1266 {
1267         struct stripe_head *sh = stripe_head_ref;
1268
1269         pr_debug("%s: stripe %llu\n", __func__,
1270                 (unsigned long long)sh->sector);
1271
1272         /* mark the computed target(s) as uptodate */
1273         mark_target_uptodate(sh, sh->ops.target);
1274         mark_target_uptodate(sh, sh->ops.target2);
1275
1276         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1277         if (sh->check_state == check_state_compute_run)
1278                 sh->check_state = check_state_compute_result;
1279         set_bit(STRIPE_HANDLE, &sh->state);
1280         raid5_release_stripe(sh);
1281 }
1282
1283 /* return a pointer to the address conversion region of the scribble buffer */
1284 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1285                                  struct raid5_percpu *percpu, int i)
1286 {
1287         void *addr;
1288
1289         addr = flex_array_get(percpu->scribble, i);
1290         return addr + sizeof(struct page *) * (sh->disks + 2);
1291 }
1292
1293 /* return a pointer to the address conversion region of the scribble buffer */
1294 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1295 {
1296         void *addr;
1297
1298         addr = flex_array_get(percpu->scribble, i);
1299         return addr;
1300 }
1301
1302 static struct dma_async_tx_descriptor *
1303 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1304 {
1305         int disks = sh->disks;
1306         struct page **xor_srcs = to_addr_page(percpu, 0);
1307         int target = sh->ops.target;
1308         struct r5dev *tgt = &sh->dev[target];
1309         struct page *xor_dest = tgt->page;
1310         int count = 0;
1311         struct dma_async_tx_descriptor *tx;
1312         struct async_submit_ctl submit;
1313         int i;
1314
1315         BUG_ON(sh->batch_head);
1316
1317         pr_debug("%s: stripe %llu block: %d\n",
1318                 __func__, (unsigned long long)sh->sector, target);
1319         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1320
1321         for (i = disks; i--; )
1322                 if (i != target)
1323                         xor_srcs[count++] = sh->dev[i].page;
1324
1325         atomic_inc(&sh->count);
1326
1327         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1328                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1329         if (unlikely(count == 1))
1330                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1331         else
1332                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1333
1334         return tx;
1335 }
1336
1337 /* set_syndrome_sources - populate source buffers for gen_syndrome
1338  * @srcs - (struct page *) array of size sh->disks
1339  * @sh - stripe_head to parse
1340  *
1341  * Populates srcs in proper layout order for the stripe and returns the
1342  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1343  * destination buffer is recorded in srcs[count] and the Q destination
1344  * is recorded in srcs[count+1]].
1345  */
1346 static int set_syndrome_sources(struct page **srcs,
1347                                 struct stripe_head *sh,
1348                                 int srctype)
1349 {
1350         int disks = sh->disks;
1351         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1352         int d0_idx = raid6_d0(sh);
1353         int count;
1354         int i;
1355
1356         for (i = 0; i < disks; i++)
1357                 srcs[i] = NULL;
1358
1359         count = 0;
1360         i = d0_idx;
1361         do {
1362                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1363                 struct r5dev *dev = &sh->dev[i];
1364
1365                 if (i == sh->qd_idx || i == sh->pd_idx ||
1366                     (srctype == SYNDROME_SRC_ALL) ||
1367                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1368                      test_bit(R5_Wantdrain, &dev->flags)) ||
1369                     (srctype == SYNDROME_SRC_WRITTEN &&
1370                      dev->written))
1371                         srcs[slot] = sh->dev[i].page;
1372                 i = raid6_next_disk(i, disks);
1373         } while (i != d0_idx);
1374
1375         return syndrome_disks;
1376 }
1377
1378 static struct dma_async_tx_descriptor *
1379 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1380 {
1381         int disks = sh->disks;
1382         struct page **blocks = to_addr_page(percpu, 0);
1383         int target;
1384         int qd_idx = sh->qd_idx;
1385         struct dma_async_tx_descriptor *tx;
1386         struct async_submit_ctl submit;
1387         struct r5dev *tgt;
1388         struct page *dest;
1389         int i;
1390         int count;
1391
1392         BUG_ON(sh->batch_head);
1393         if (sh->ops.target < 0)
1394                 target = sh->ops.target2;
1395         else if (sh->ops.target2 < 0)
1396                 target = sh->ops.target;
1397         else
1398                 /* we should only have one valid target */
1399                 BUG();
1400         BUG_ON(target < 0);
1401         pr_debug("%s: stripe %llu block: %d\n",
1402                 __func__, (unsigned long long)sh->sector, target);
1403
1404         tgt = &sh->dev[target];
1405         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1406         dest = tgt->page;
1407
1408         atomic_inc(&sh->count);
1409
1410         if (target == qd_idx) {
1411                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1412                 blocks[count] = NULL; /* regenerating p is not necessary */
1413                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1414                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1415                                   ops_complete_compute, sh,
1416                                   to_addr_conv(sh, percpu, 0));
1417                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1418         } else {
1419                 /* Compute any data- or p-drive using XOR */
1420                 count = 0;
1421                 for (i = disks; i-- ; ) {
1422                         if (i == target || i == qd_idx)
1423                                 continue;
1424                         blocks[count++] = sh->dev[i].page;
1425                 }
1426
1427                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1428                                   NULL, ops_complete_compute, sh,
1429                                   to_addr_conv(sh, percpu, 0));
1430                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1431         }
1432
1433         return tx;
1434 }
1435
1436 static struct dma_async_tx_descriptor *
1437 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1438 {
1439         int i, count, disks = sh->disks;
1440         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1441         int d0_idx = raid6_d0(sh);
1442         int faila = -1, failb = -1;
1443         int target = sh->ops.target;
1444         int target2 = sh->ops.target2;
1445         struct r5dev *tgt = &sh->dev[target];
1446         struct r5dev *tgt2 = &sh->dev[target2];
1447         struct dma_async_tx_descriptor *tx;
1448         struct page **blocks = to_addr_page(percpu, 0);
1449         struct async_submit_ctl submit;
1450
1451         BUG_ON(sh->batch_head);
1452         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1453                  __func__, (unsigned long long)sh->sector, target, target2);
1454         BUG_ON(target < 0 || target2 < 0);
1455         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1456         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1457
1458         /* we need to open-code set_syndrome_sources to handle the
1459          * slot number conversion for 'faila' and 'failb'
1460          */
1461         for (i = 0; i < disks ; i++)
1462                 blocks[i] = NULL;
1463         count = 0;
1464         i = d0_idx;
1465         do {
1466                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1467
1468                 blocks[slot] = sh->dev[i].page;
1469
1470                 if (i == target)
1471                         faila = slot;
1472                 if (i == target2)
1473                         failb = slot;
1474                 i = raid6_next_disk(i, disks);
1475         } while (i != d0_idx);
1476
1477         BUG_ON(faila == failb);
1478         if (failb < faila)
1479                 swap(faila, failb);
1480         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1481                  __func__, (unsigned long long)sh->sector, faila, failb);
1482
1483         atomic_inc(&sh->count);
1484
1485         if (failb == syndrome_disks+1) {
1486                 /* Q disk is one of the missing disks */
1487                 if (faila == syndrome_disks) {
1488                         /* Missing P+Q, just recompute */
1489                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1490                                           ops_complete_compute, sh,
1491                                           to_addr_conv(sh, percpu, 0));
1492                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1493                                                   STRIPE_SIZE, &submit);
1494                 } else {
1495                         struct page *dest;
1496                         int data_target;
1497                         int qd_idx = sh->qd_idx;
1498
1499                         /* Missing D+Q: recompute D from P, then recompute Q */
1500                         if (target == qd_idx)
1501                                 data_target = target2;
1502                         else
1503                                 data_target = target;
1504
1505                         count = 0;
1506                         for (i = disks; i-- ; ) {
1507                                 if (i == data_target || i == qd_idx)
1508                                         continue;
1509                                 blocks[count++] = sh->dev[i].page;
1510                         }
1511                         dest = sh->dev[data_target].page;
1512                         init_async_submit(&submit,
1513                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1514                                           NULL, NULL, NULL,
1515                                           to_addr_conv(sh, percpu, 0));
1516                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1517                                        &submit);
1518
1519                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1520                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1521                                           ops_complete_compute, sh,
1522                                           to_addr_conv(sh, percpu, 0));
1523                         return async_gen_syndrome(blocks, 0, count+2,
1524                                                   STRIPE_SIZE, &submit);
1525                 }
1526         } else {
1527                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1528                                   ops_complete_compute, sh,
1529                                   to_addr_conv(sh, percpu, 0));
1530                 if (failb == syndrome_disks) {
1531                         /* We're missing D+P. */
1532                         return async_raid6_datap_recov(syndrome_disks+2,
1533                                                        STRIPE_SIZE, faila,
1534                                                        blocks, &submit);
1535                 } else {
1536                         /* We're missing D+D. */
1537                         return async_raid6_2data_recov(syndrome_disks+2,
1538                                                        STRIPE_SIZE, faila, failb,
1539                                                        blocks, &submit);
1540                 }
1541         }
1542 }
1543
1544 static void ops_complete_prexor(void *stripe_head_ref)
1545 {
1546         struct stripe_head *sh = stripe_head_ref;
1547
1548         pr_debug("%s: stripe %llu\n", __func__,
1549                 (unsigned long long)sh->sector);
1550 }
1551
1552 static struct dma_async_tx_descriptor *
1553 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1554                 struct dma_async_tx_descriptor *tx)
1555 {
1556         int disks = sh->disks;
1557         struct page **xor_srcs = to_addr_page(percpu, 0);
1558         int count = 0, pd_idx = sh->pd_idx, i;
1559         struct async_submit_ctl submit;
1560
1561         /* existing parity data subtracted */
1562         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1563
1564         BUG_ON(sh->batch_head);
1565         pr_debug("%s: stripe %llu\n", __func__,
1566                 (unsigned long long)sh->sector);
1567
1568         for (i = disks; i--; ) {
1569                 struct r5dev *dev = &sh->dev[i];
1570                 /* Only process blocks that are known to be uptodate */
1571                 if (test_bit(R5_Wantdrain, &dev->flags))
1572                         xor_srcs[count++] = dev->page;
1573         }
1574
1575         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1576                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1577         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1578
1579         return tx;
1580 }
1581
1582 static struct dma_async_tx_descriptor *
1583 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1584                 struct dma_async_tx_descriptor *tx)
1585 {
1586         struct page **blocks = to_addr_page(percpu, 0);
1587         int count;
1588         struct async_submit_ctl submit;
1589
1590         pr_debug("%s: stripe %llu\n", __func__,
1591                 (unsigned long long)sh->sector);
1592
1593         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1594
1595         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1596                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1597         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1598
1599         return tx;
1600 }
1601
1602 static struct dma_async_tx_descriptor *
1603 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1604 {
1605         int disks = sh->disks;
1606         int i;
1607         struct stripe_head *head_sh = sh;
1608
1609         pr_debug("%s: stripe %llu\n", __func__,
1610                 (unsigned long long)sh->sector);
1611
1612         for (i = disks; i--; ) {
1613                 struct r5dev *dev;
1614                 struct bio *chosen;
1615
1616                 sh = head_sh;
1617                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1618                         struct bio *wbi;
1619
1620 again:
1621                         dev = &sh->dev[i];
1622                         spin_lock_irq(&sh->stripe_lock);
1623                         chosen = dev->towrite;
1624                         dev->towrite = NULL;
1625                         sh->overwrite_disks = 0;
1626                         BUG_ON(dev->written);
1627                         wbi = dev->written = chosen;
1628                         spin_unlock_irq(&sh->stripe_lock);
1629                         WARN_ON(dev->page != dev->orig_page);
1630
1631                         while (wbi && wbi->bi_iter.bi_sector <
1632                                 dev->sector + STRIPE_SECTORS) {
1633                                 if (wbi->bi_rw & REQ_FUA)
1634                                         set_bit(R5_WantFUA, &dev->flags);
1635                                 if (wbi->bi_rw & REQ_SYNC)
1636                                         set_bit(R5_SyncIO, &dev->flags);
1637                                 if (wbi->bi_rw & REQ_DISCARD)
1638                                         set_bit(R5_Discard, &dev->flags);
1639                                 else {
1640                                         tx = async_copy_data(1, wbi, &dev->page,
1641                                                 dev->sector, tx, sh);
1642                                         if (dev->page != dev->orig_page) {
1643                                                 set_bit(R5_SkipCopy, &dev->flags);
1644                                                 clear_bit(R5_UPTODATE, &dev->flags);
1645                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1646                                         }
1647                                 }
1648                                 wbi = r5_next_bio(wbi, dev->sector);
1649                         }
1650
1651                         if (head_sh->batch_head) {
1652                                 sh = list_first_entry(&sh->batch_list,
1653                                                       struct stripe_head,
1654                                                       batch_list);
1655                                 if (sh == head_sh)
1656                                         continue;
1657                                 goto again;
1658                         }
1659                 }
1660         }
1661
1662         return tx;
1663 }
1664
1665 static void ops_complete_reconstruct(void *stripe_head_ref)
1666 {
1667         struct stripe_head *sh = stripe_head_ref;
1668         int disks = sh->disks;
1669         int pd_idx = sh->pd_idx;
1670         int qd_idx = sh->qd_idx;
1671         int i;
1672         bool fua = false, sync = false, discard = false;
1673
1674         pr_debug("%s: stripe %llu\n", __func__,
1675                 (unsigned long long)sh->sector);
1676
1677         for (i = disks; i--; ) {
1678                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1679                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1680                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1681         }
1682
1683         for (i = disks; i--; ) {
1684                 struct r5dev *dev = &sh->dev[i];
1685
1686                 if (dev->written || i == pd_idx || i == qd_idx) {
1687                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1688                                 set_bit(R5_UPTODATE, &dev->flags);
1689                         if (fua)
1690                                 set_bit(R5_WantFUA, &dev->flags);
1691                         if (sync)
1692                                 set_bit(R5_SyncIO, &dev->flags);
1693                 }
1694         }
1695
1696         if (sh->reconstruct_state == reconstruct_state_drain_run)
1697                 sh->reconstruct_state = reconstruct_state_drain_result;
1698         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1699                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1700         else {
1701                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1702                 sh->reconstruct_state = reconstruct_state_result;
1703         }
1704
1705         set_bit(STRIPE_HANDLE, &sh->state);
1706         raid5_release_stripe(sh);
1707 }
1708
1709 static void
1710 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1711                      struct dma_async_tx_descriptor *tx)
1712 {
1713         int disks = sh->disks;
1714         struct page **xor_srcs;
1715         struct async_submit_ctl submit;
1716         int count, pd_idx = sh->pd_idx, i;
1717         struct page *xor_dest;
1718         int prexor = 0;
1719         unsigned long flags;
1720         int j = 0;
1721         struct stripe_head *head_sh = sh;
1722         int last_stripe;
1723
1724         pr_debug("%s: stripe %llu\n", __func__,
1725                 (unsigned long long)sh->sector);
1726
1727         for (i = 0; i < sh->disks; i++) {
1728                 if (pd_idx == i)
1729                         continue;
1730                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1731                         break;
1732         }
1733         if (i >= sh->disks) {
1734                 atomic_inc(&sh->count);
1735                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1736                 ops_complete_reconstruct(sh);
1737                 return;
1738         }
1739 again:
1740         count = 0;
1741         xor_srcs = to_addr_page(percpu, j);
1742         /* check if prexor is active which means only process blocks
1743          * that are part of a read-modify-write (written)
1744          */
1745         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1746                 prexor = 1;
1747                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1748                 for (i = disks; i--; ) {
1749                         struct r5dev *dev = &sh->dev[i];
1750                         if (head_sh->dev[i].written)
1751                                 xor_srcs[count++] = dev->page;
1752                 }
1753         } else {
1754                 xor_dest = sh->dev[pd_idx].page;
1755                 for (i = disks; i--; ) {
1756                         struct r5dev *dev = &sh->dev[i];
1757                         if (i != pd_idx)
1758                                 xor_srcs[count++] = dev->page;
1759                 }
1760         }
1761
1762         /* 1/ if we prexor'd then the dest is reused as a source
1763          * 2/ if we did not prexor then we are redoing the parity
1764          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1765          * for the synchronous xor case
1766          */
1767         last_stripe = !head_sh->batch_head ||
1768                 list_first_entry(&sh->batch_list,
1769                                  struct stripe_head, batch_list) == head_sh;
1770         if (last_stripe) {
1771                 flags = ASYNC_TX_ACK |
1772                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1773
1774                 atomic_inc(&head_sh->count);
1775                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1776                                   to_addr_conv(sh, percpu, j));
1777         } else {
1778                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1779                 init_async_submit(&submit, flags, tx, NULL, NULL,
1780                                   to_addr_conv(sh, percpu, j));
1781         }
1782
1783         if (unlikely(count == 1))
1784                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1785         else
1786                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1787         if (!last_stripe) {
1788                 j++;
1789                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1790                                       batch_list);
1791                 goto again;
1792         }
1793 }
1794
1795 static void
1796 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1797                      struct dma_async_tx_descriptor *tx)
1798 {
1799         struct async_submit_ctl submit;
1800         struct page **blocks;
1801         int count, i, j = 0;
1802         struct stripe_head *head_sh = sh;
1803         int last_stripe;
1804         int synflags;
1805         unsigned long txflags;
1806
1807         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1808
1809         for (i = 0; i < sh->disks; i++) {
1810                 if (sh->pd_idx == i || sh->qd_idx == i)
1811                         continue;
1812                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1813                         break;
1814         }
1815         if (i >= sh->disks) {
1816                 atomic_inc(&sh->count);
1817                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1818                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1819                 ops_complete_reconstruct(sh);
1820                 return;
1821         }
1822
1823 again:
1824         blocks = to_addr_page(percpu, j);
1825
1826         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1827                 synflags = SYNDROME_SRC_WRITTEN;
1828                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1829         } else {
1830                 synflags = SYNDROME_SRC_ALL;
1831                 txflags = ASYNC_TX_ACK;
1832         }
1833
1834         count = set_syndrome_sources(blocks, sh, synflags);
1835         last_stripe = !head_sh->batch_head ||
1836                 list_first_entry(&sh->batch_list,
1837                                  struct stripe_head, batch_list) == head_sh;
1838
1839         if (last_stripe) {
1840                 atomic_inc(&head_sh->count);
1841                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1842                                   head_sh, to_addr_conv(sh, percpu, j));
1843         } else
1844                 init_async_submit(&submit, 0, tx, NULL, NULL,
1845                                   to_addr_conv(sh, percpu, j));
1846         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1847         if (!last_stripe) {
1848                 j++;
1849                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1850                                       batch_list);
1851                 goto again;
1852         }
1853 }
1854
1855 static void ops_complete_check(void *stripe_head_ref)
1856 {
1857         struct stripe_head *sh = stripe_head_ref;
1858
1859         pr_debug("%s: stripe %llu\n", __func__,
1860                 (unsigned long long)sh->sector);
1861
1862         sh->check_state = check_state_check_result;
1863         set_bit(STRIPE_HANDLE, &sh->state);
1864         raid5_release_stripe(sh);
1865 }
1866
1867 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1868 {
1869         int disks = sh->disks;
1870         int pd_idx = sh->pd_idx;
1871         int qd_idx = sh->qd_idx;
1872         struct page *xor_dest;
1873         struct page **xor_srcs = to_addr_page(percpu, 0);
1874         struct dma_async_tx_descriptor *tx;
1875         struct async_submit_ctl submit;
1876         int count;
1877         int i;
1878
1879         pr_debug("%s: stripe %llu\n", __func__,
1880                 (unsigned long long)sh->sector);
1881
1882         BUG_ON(sh->batch_head);
1883         count = 0;
1884         xor_dest = sh->dev[pd_idx].page;
1885         xor_srcs[count++] = xor_dest;
1886         for (i = disks; i--; ) {
1887                 if (i == pd_idx || i == qd_idx)
1888                         continue;
1889                 xor_srcs[count++] = sh->dev[i].page;
1890         }
1891
1892         init_async_submit(&submit, 0, NULL, NULL, NULL,
1893                           to_addr_conv(sh, percpu, 0));
1894         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1895                            &sh->ops.zero_sum_result, &submit);
1896
1897         atomic_inc(&sh->count);
1898         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1899         tx = async_trigger_callback(&submit);
1900 }
1901
1902 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1903 {
1904         struct page **srcs = to_addr_page(percpu, 0);
1905         struct async_submit_ctl submit;
1906         int count;
1907
1908         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1909                 (unsigned long long)sh->sector, checkp);
1910
1911         BUG_ON(sh->batch_head);
1912         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1913         if (!checkp)
1914                 srcs[count] = NULL;
1915
1916         atomic_inc(&sh->count);
1917         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1918                           sh, to_addr_conv(sh, percpu, 0));
1919         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1920                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1921 }
1922
1923 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1924 {
1925         int overlap_clear = 0, i, disks = sh->disks;
1926         struct dma_async_tx_descriptor *tx = NULL;
1927         struct r5conf *conf = sh->raid_conf;
1928         int level = conf->level;
1929         struct raid5_percpu *percpu;
1930         unsigned long cpu;
1931
1932         cpu = get_cpu();
1933         percpu = per_cpu_ptr(conf->percpu, cpu);
1934         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1935                 ops_run_biofill(sh);
1936                 overlap_clear++;
1937         }
1938
1939         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1940                 if (level < 6)
1941                         tx = ops_run_compute5(sh, percpu);
1942                 else {
1943                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1944                                 tx = ops_run_compute6_1(sh, percpu);
1945                         else
1946                                 tx = ops_run_compute6_2(sh, percpu);
1947                 }
1948                 /* terminate the chain if reconstruct is not set to be run */
1949                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1950                         async_tx_ack(tx);
1951         }
1952
1953         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1954                 if (level < 6)
1955                         tx = ops_run_prexor5(sh, percpu, tx);
1956                 else
1957                         tx = ops_run_prexor6(sh, percpu, tx);
1958         }
1959
1960         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1961                 tx = ops_run_biodrain(sh, tx);
1962                 overlap_clear++;
1963         }
1964
1965         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1966                 if (level < 6)
1967                         ops_run_reconstruct5(sh, percpu, tx);
1968                 else
1969                         ops_run_reconstruct6(sh, percpu, tx);
1970         }
1971
1972         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1973                 if (sh->check_state == check_state_run)
1974                         ops_run_check_p(sh, percpu);
1975                 else if (sh->check_state == check_state_run_q)
1976                         ops_run_check_pq(sh, percpu, 0);
1977                 else if (sh->check_state == check_state_run_pq)
1978                         ops_run_check_pq(sh, percpu, 1);
1979                 else
1980                         BUG();
1981         }
1982
1983         if (overlap_clear && !sh->batch_head)
1984                 for (i = disks; i--; ) {
1985                         struct r5dev *dev = &sh->dev[i];
1986                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1987                                 wake_up(&sh->raid_conf->wait_for_overlap);
1988                 }
1989         put_cpu();
1990 }
1991
1992 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1993 {
1994         struct stripe_head *sh;
1995
1996         sh = kmem_cache_zalloc(sc, gfp);
1997         if (sh) {
1998                 spin_lock_init(&sh->stripe_lock);
1999                 spin_lock_init(&sh->batch_lock);
2000                 INIT_LIST_HEAD(&sh->batch_list);
2001                 INIT_LIST_HEAD(&sh->lru);
2002                 atomic_set(&sh->count, 1);
2003         }
2004         return sh;
2005 }
2006 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2007 {
2008         struct stripe_head *sh;
2009
2010         sh = alloc_stripe(conf->slab_cache, gfp);
2011         if (!sh)
2012                 return 0;
2013
2014         sh->raid_conf = conf;
2015
2016         if (grow_buffers(sh, gfp)) {
2017                 shrink_buffers(sh);
2018                 kmem_cache_free(conf->slab_cache, sh);
2019                 return 0;
2020         }
2021         sh->hash_lock_index =
2022                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2023         /* we just created an active stripe so... */
2024         atomic_inc(&conf->active_stripes);
2025
2026         raid5_release_stripe(sh);
2027         conf->max_nr_stripes++;
2028         return 1;
2029 }
2030
2031 static int grow_stripes(struct r5conf *conf, int num)
2032 {
2033         struct kmem_cache *sc;
2034         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2035
2036         if (conf->mddev->gendisk)
2037                 sprintf(conf->cache_name[0],
2038                         "raid%d-%s", conf->level, mdname(conf->mddev));
2039         else
2040                 sprintf(conf->cache_name[0],
2041                         "raid%d-%p", conf->level, conf->mddev);
2042         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2043
2044         conf->active_name = 0;
2045         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2046                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2047                                0, 0, NULL);
2048         if (!sc)
2049                 return 1;
2050         conf->slab_cache = sc;
2051         conf->pool_size = devs;
2052         while (num--)
2053                 if (!grow_one_stripe(conf, GFP_KERNEL))
2054                         return 1;
2055
2056         return 0;
2057 }
2058
2059 /**
2060  * scribble_len - return the required size of the scribble region
2061  * @num - total number of disks in the array
2062  *
2063  * The size must be enough to contain:
2064  * 1/ a struct page pointer for each device in the array +2
2065  * 2/ room to convert each entry in (1) to its corresponding dma
2066  *    (dma_map_page()) or page (page_address()) address.
2067  *
2068  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2069  * calculate over all devices (not just the data blocks), using zeros in place
2070  * of the P and Q blocks.
2071  */
2072 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2073 {
2074         struct flex_array *ret;
2075         size_t len;
2076
2077         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2078         ret = flex_array_alloc(len, cnt, flags);
2079         if (!ret)
2080                 return NULL;
2081         /* always prealloc all elements, so no locking is required */
2082         if (flex_array_prealloc(ret, 0, cnt, flags)) {
2083                 flex_array_free(ret);
2084                 return NULL;
2085         }
2086         return ret;
2087 }
2088
2089 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2090 {
2091         unsigned long cpu;
2092         int err = 0;
2093
2094         /*
2095          * Never shrink. And mddev_suspend() could deadlock if this is called
2096          * from raid5d. In that case, scribble_disks and scribble_sectors
2097          * should equal to new_disks and new_sectors
2098          */
2099         if (conf->scribble_disks >= new_disks &&
2100             conf->scribble_sectors >= new_sectors)
2101                 return 0;
2102         mddev_suspend(conf->mddev);
2103         get_online_cpus();
2104         for_each_present_cpu(cpu) {
2105                 struct raid5_percpu *percpu;
2106                 struct flex_array *scribble;
2107
2108                 percpu = per_cpu_ptr(conf->percpu, cpu);
2109                 scribble = scribble_alloc(new_disks,
2110                                           new_sectors / STRIPE_SECTORS,
2111                                           GFP_NOIO);
2112
2113                 if (scribble) {
2114                         flex_array_free(percpu->scribble);
2115                         percpu->scribble = scribble;
2116                 } else {
2117                         err = -ENOMEM;
2118                         break;
2119                 }
2120         }
2121         put_online_cpus();
2122         mddev_resume(conf->mddev);
2123         if (!err) {
2124                 conf->scribble_disks = new_disks;
2125                 conf->scribble_sectors = new_sectors;
2126         }
2127         return err;
2128 }
2129
2130 static int resize_stripes(struct r5conf *conf, int newsize)
2131 {
2132         /* Make all the stripes able to hold 'newsize' devices.
2133          * New slots in each stripe get 'page' set to a new page.
2134          *
2135          * This happens in stages:
2136          * 1/ create a new kmem_cache and allocate the required number of
2137          *    stripe_heads.
2138          * 2/ gather all the old stripe_heads and transfer the pages across
2139          *    to the new stripe_heads.  This will have the side effect of
2140          *    freezing the array as once all stripe_heads have been collected,
2141          *    no IO will be possible.  Old stripe heads are freed once their
2142          *    pages have been transferred over, and the old kmem_cache is
2143          *    freed when all stripes are done.
2144          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2145          *    we simple return a failre status - no need to clean anything up.
2146          * 4/ allocate new pages for the new slots in the new stripe_heads.
2147          *    If this fails, we don't bother trying the shrink the
2148          *    stripe_heads down again, we just leave them as they are.
2149          *    As each stripe_head is processed the new one is released into
2150          *    active service.
2151          *
2152          * Once step2 is started, we cannot afford to wait for a write,
2153          * so we use GFP_NOIO allocations.
2154          */
2155         struct stripe_head *osh, *nsh;
2156         LIST_HEAD(newstripes);
2157         struct disk_info *ndisks;
2158         int err;
2159         struct kmem_cache *sc;
2160         int i;
2161         int hash, cnt;
2162
2163         if (newsize <= conf->pool_size)
2164                 return 0; /* never bother to shrink */
2165
2166         err = md_allow_write(conf->mddev);
2167         if (err)
2168                 return err;
2169
2170         /* Step 1 */
2171         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2172                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2173                                0, 0, NULL);
2174         if (!sc)
2175                 return -ENOMEM;
2176
2177         /* Need to ensure auto-resizing doesn't interfere */
2178         mutex_lock(&conf->cache_size_mutex);
2179
2180         for (i = conf->max_nr_stripes; i; i--) {
2181                 nsh = alloc_stripe(sc, GFP_KERNEL);
2182                 if (!nsh)
2183                         break;
2184
2185                 nsh->raid_conf = conf;
2186                 list_add(&nsh->lru, &newstripes);
2187         }
2188         if (i) {
2189                 /* didn't get enough, give up */
2190                 while (!list_empty(&newstripes)) {
2191                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2192                         list_del(&nsh->lru);
2193                         kmem_cache_free(sc, nsh);
2194                 }
2195                 kmem_cache_destroy(sc);
2196                 mutex_unlock(&conf->cache_size_mutex);
2197                 return -ENOMEM;
2198         }
2199         /* Step 2 - Must use GFP_NOIO now.
2200          * OK, we have enough stripes, start collecting inactive
2201          * stripes and copying them over
2202          */
2203         hash = 0;
2204         cnt = 0;
2205         list_for_each_entry(nsh, &newstripes, lru) {
2206                 lock_device_hash_lock(conf, hash);
2207                 wait_event_exclusive_cmd(conf->wait_for_stripe[hash],
2208                                     !list_empty(conf->inactive_list + hash),
2209                                     unlock_device_hash_lock(conf, hash),
2210                                     lock_device_hash_lock(conf, hash));
2211                 osh = get_free_stripe(conf, hash);
2212                 unlock_device_hash_lock(conf, hash);
2213
2214                 for(i=0; i<conf->pool_size; i++) {
2215                         nsh->dev[i].page = osh->dev[i].page;
2216                         nsh->dev[i].orig_page = osh->dev[i].page;
2217                 }
2218                 nsh->hash_lock_index = hash;
2219                 kmem_cache_free(conf->slab_cache, osh);
2220                 cnt++;
2221                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2222                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2223                         hash++;
2224                         cnt = 0;
2225                 }
2226         }
2227         kmem_cache_destroy(conf->slab_cache);
2228
2229         /* Step 3.
2230          * At this point, we are holding all the stripes so the array
2231          * is completely stalled, so now is a good time to resize
2232          * conf->disks and the scribble region
2233          */
2234         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2235         if (ndisks) {
2236                 for (i=0; i<conf->raid_disks; i++)
2237                         ndisks[i] = conf->disks[i];
2238                 kfree(conf->disks);
2239                 conf->disks = ndisks;
2240         } else
2241                 err = -ENOMEM;
2242
2243         mutex_unlock(&conf->cache_size_mutex);
2244         /* Step 4, return new stripes to service */
2245         while(!list_empty(&newstripes)) {
2246                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2247                 list_del_init(&nsh->lru);
2248
2249                 for (i=conf->raid_disks; i < newsize; i++)
2250                         if (nsh->dev[i].page == NULL) {
2251                                 struct page *p = alloc_page(GFP_NOIO);
2252                                 nsh->dev[i].page = p;
2253                                 nsh->dev[i].orig_page = p;
2254                                 if (!p)
2255                                         err = -ENOMEM;
2256                         }
2257                 raid5_release_stripe(nsh);
2258         }
2259         /* critical section pass, GFP_NOIO no longer needed */
2260
2261         conf->slab_cache = sc;
2262         conf->active_name = 1-conf->active_name;
2263         if (!err)
2264                 conf->pool_size = newsize;
2265         return err;
2266 }
2267
2268 static int drop_one_stripe(struct r5conf *conf)
2269 {
2270         struct stripe_head *sh;
2271         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2272
2273         spin_lock_irq(conf->hash_locks + hash);
2274         sh = get_free_stripe(conf, hash);
2275         spin_unlock_irq(conf->hash_locks + hash);
2276         if (!sh)
2277                 return 0;
2278         BUG_ON(atomic_read(&sh->count));
2279         shrink_buffers(sh);
2280         kmem_cache_free(conf->slab_cache, sh);
2281         atomic_dec(&conf->active_stripes);
2282         conf->max_nr_stripes--;
2283         return 1;
2284 }
2285
2286 static void shrink_stripes(struct r5conf *conf)
2287 {
2288         while (conf->max_nr_stripes &&
2289                drop_one_stripe(conf))
2290                 ;
2291
2292         kmem_cache_destroy(conf->slab_cache);
2293         conf->slab_cache = NULL;
2294 }
2295
2296 static void raid5_end_read_request(struct bio * bi)
2297 {
2298         struct stripe_head *sh = bi->bi_private;
2299         struct r5conf *conf = sh->raid_conf;
2300         int disks = sh->disks, i;
2301         char b[BDEVNAME_SIZE];
2302         struct md_rdev *rdev = NULL;
2303         sector_t s;
2304
2305         for (i=0 ; i<disks; i++)
2306                 if (bi == &sh->dev[i].req)
2307                         break;
2308
2309         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2310                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2311                 bi->bi_error);
2312         if (i == disks) {
2313                 BUG();
2314                 return;
2315         }
2316         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2317                 /* If replacement finished while this request was outstanding,
2318                  * 'replacement' might be NULL already.
2319                  * In that case it moved down to 'rdev'.
2320                  * rdev is not removed until all requests are finished.
2321                  */
2322                 rdev = conf->disks[i].replacement;
2323         if (!rdev)
2324                 rdev = conf->disks[i].rdev;
2325
2326         if (use_new_offset(conf, sh))
2327                 s = sh->sector + rdev->new_data_offset;
2328         else
2329                 s = sh->sector + rdev->data_offset;
2330         if (!bi->bi_error) {
2331                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2332                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2333                         /* Note that this cannot happen on a
2334                          * replacement device.  We just fail those on
2335                          * any error
2336                          */
2337                         printk_ratelimited(
2338                                 KERN_INFO
2339                                 "md/raid:%s: read error corrected"
2340                                 " (%lu sectors at %llu on %s)\n",
2341                                 mdname(conf->mddev), STRIPE_SECTORS,
2342                                 (unsigned long long)s,
2343                                 bdevname(rdev->bdev, b));
2344                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2345                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2346                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2347                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2348                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2349
2350                 if (atomic_read(&rdev->read_errors))
2351                         atomic_set(&rdev->read_errors, 0);
2352         } else {
2353                 const char *bdn = bdevname(rdev->bdev, b);
2354                 int retry = 0;
2355                 int set_bad = 0;
2356
2357                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2358                 atomic_inc(&rdev->read_errors);
2359                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2360                         printk_ratelimited(
2361                                 KERN_WARNING
2362                                 "md/raid:%s: read error on replacement device "
2363                                 "(sector %llu on %s).\n",
2364                                 mdname(conf->mddev),
2365                                 (unsigned long long)s,
2366                                 bdn);
2367                 else if (conf->mddev->degraded >= conf->max_degraded) {
2368                         set_bad = 1;
2369                         printk_ratelimited(
2370                                 KERN_WARNING
2371                                 "md/raid:%s: read error not correctable "
2372                                 "(sector %llu on %s).\n",
2373                                 mdname(conf->mddev),
2374                                 (unsigned long long)s,
2375                                 bdn);
2376                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2377                         /* Oh, no!!! */
2378                         set_bad = 1;
2379                         printk_ratelimited(
2380                                 KERN_WARNING
2381                                 "md/raid:%s: read error NOT corrected!! "
2382                                 "(sector %llu on %s).\n",
2383                                 mdname(conf->mddev),
2384                                 (unsigned long long)s,
2385                                 bdn);
2386                 } else if (atomic_read(&rdev->read_errors)
2387                          > conf->max_nr_stripes)
2388                         printk(KERN_WARNING
2389                                "md/raid:%s: Too many read errors, failing device %s.\n",
2390                                mdname(conf->mddev), bdn);
2391                 else
2392                         retry = 1;
2393                 if (set_bad && test_bit(In_sync, &rdev->flags)
2394                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2395                         retry = 1;
2396                 if (retry)
2397                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2398                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2399                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2400                         } else
2401                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2402                 else {
2403                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2404                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2405                         if (!(set_bad
2406                               && test_bit(In_sync, &rdev->flags)
2407                               && rdev_set_badblocks(
2408                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2409                                 md_error(conf->mddev, rdev);
2410                 }
2411         }
2412         rdev_dec_pending(rdev, conf->mddev);
2413         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2414         set_bit(STRIPE_HANDLE, &sh->state);
2415         raid5_release_stripe(sh);
2416 }
2417
2418 static void raid5_end_write_request(struct bio *bi)
2419 {
2420         struct stripe_head *sh = bi->bi_private;
2421         struct r5conf *conf = sh->raid_conf;
2422         int disks = sh->disks, i;
2423         struct md_rdev *uninitialized_var(rdev);
2424         sector_t first_bad;
2425         int bad_sectors;
2426         int replacement = 0;
2427
2428         for (i = 0 ; i < disks; i++) {
2429                 if (bi == &sh->dev[i].req) {
2430                         rdev = conf->disks[i].rdev;
2431                         break;
2432                 }
2433                 if (bi == &sh->dev[i].rreq) {
2434                         rdev = conf->disks[i].replacement;
2435                         if (rdev)
2436                                 replacement = 1;
2437                         else
2438                                 /* rdev was removed and 'replacement'
2439                                  * replaced it.  rdev is not removed
2440                                  * until all requests are finished.
2441                                  */
2442                                 rdev = conf->disks[i].rdev;
2443                         break;
2444                 }
2445         }
2446         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2447                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2448                 bi->bi_error);
2449         if (i == disks) {
2450                 BUG();
2451                 return;
2452         }
2453
2454         if (replacement) {
2455                 if (bi->bi_error)
2456                         md_error(conf->mddev, rdev);
2457                 else if (is_badblock(rdev, sh->sector,
2458                                      STRIPE_SECTORS,
2459                                      &first_bad, &bad_sectors))
2460                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2461         } else {
2462                 if (bi->bi_error) {
2463                         set_bit(STRIPE_DEGRADED, &sh->state);
2464                         set_bit(WriteErrorSeen, &rdev->flags);
2465                         set_bit(R5_WriteError, &sh->dev[i].flags);
2466                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2467                                 set_bit(MD_RECOVERY_NEEDED,
2468                                         &rdev->mddev->recovery);
2469                 } else if (is_badblock(rdev, sh->sector,
2470                                        STRIPE_SECTORS,
2471                                        &first_bad, &bad_sectors)) {
2472                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2473                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2474                                 /* That was a successful write so make
2475                                  * sure it looks like we already did
2476                                  * a re-write.
2477                                  */
2478                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2479                 }
2480         }
2481         rdev_dec_pending(rdev, conf->mddev);
2482
2483         if (sh->batch_head && bi->bi_error && !replacement)
2484                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2485
2486         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2487                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2488         set_bit(STRIPE_HANDLE, &sh->state);
2489         raid5_release_stripe(sh);
2490
2491         if (sh->batch_head && sh != sh->batch_head)
2492                 raid5_release_stripe(sh->batch_head);
2493 }
2494
2495 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2496 {
2497         struct r5dev *dev = &sh->dev[i];
2498
2499         bio_init(&dev->req);
2500         dev->req.bi_io_vec = &dev->vec;
2501         dev->req.bi_max_vecs = 1;
2502         dev->req.bi_private = sh;
2503
2504         bio_init(&dev->rreq);
2505         dev->rreq.bi_io_vec = &dev->rvec;
2506         dev->rreq.bi_max_vecs = 1;
2507         dev->rreq.bi_private = sh;
2508
2509         dev->flags = 0;
2510         dev->sector = raid5_compute_blocknr(sh, i, previous);
2511 }
2512
2513 static void error(struct mddev *mddev, struct md_rdev *rdev)
2514 {
2515         char b[BDEVNAME_SIZE];
2516         struct r5conf *conf = mddev->private;
2517         unsigned long flags;
2518         pr_debug("raid456: error called\n");
2519
2520         spin_lock_irqsave(&conf->device_lock, flags);
2521         clear_bit(In_sync, &rdev->flags);
2522         mddev->degraded = calc_degraded(conf);
2523         spin_unlock_irqrestore(&conf->device_lock, flags);
2524         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2525
2526         set_bit(Blocked, &rdev->flags);
2527         set_bit(Faulty, &rdev->flags);
2528         set_bit(MD_CHANGE_DEVS, &mddev->flags);
2529         set_bit(MD_CHANGE_PENDING, &mddev->flags);
2530         printk(KERN_ALERT
2531                "md/raid:%s: Disk failure on %s, disabling device.\n"
2532                "md/raid:%s: Operation continuing on %d devices.\n",
2533                mdname(mddev),
2534                bdevname(rdev->bdev, b),
2535                mdname(mddev),
2536                conf->raid_disks - mddev->degraded);
2537 }
2538
2539 /*
2540  * Input: a 'big' sector number,
2541  * Output: index of the data and parity disk, and the sector # in them.
2542  */
2543 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2544                               int previous, int *dd_idx,
2545                               struct stripe_head *sh)
2546 {
2547         sector_t stripe, stripe2;
2548         sector_t chunk_number;
2549         unsigned int chunk_offset;
2550         int pd_idx, qd_idx;
2551         int ddf_layout = 0;
2552         sector_t new_sector;
2553         int algorithm = previous ? conf->prev_algo
2554                                  : conf->algorithm;
2555         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2556                                          : conf->chunk_sectors;
2557         int raid_disks = previous ? conf->previous_raid_disks
2558                                   : conf->raid_disks;
2559         int data_disks = raid_disks - conf->max_degraded;
2560
2561         /* First compute the information on this sector */
2562
2563         /*
2564          * Compute the chunk number and the sector offset inside the chunk
2565          */
2566         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2567         chunk_number = r_sector;
2568
2569         /*
2570          * Compute the stripe number
2571          */
2572         stripe = chunk_number;
2573         *dd_idx = sector_div(stripe, data_disks);
2574         stripe2 = stripe;
2575         /*
2576          * Select the parity disk based on the user selected algorithm.
2577          */
2578         pd_idx = qd_idx = -1;
2579         switch(conf->level) {
2580         case 4:
2581                 pd_idx = data_disks;
2582                 break;
2583         case 5:
2584                 switch (algorithm) {
2585                 case ALGORITHM_LEFT_ASYMMETRIC:
2586                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2587                         if (*dd_idx >= pd_idx)
2588                                 (*dd_idx)++;
2589                         break;
2590                 case ALGORITHM_RIGHT_ASYMMETRIC:
2591                         pd_idx = sector_div(stripe2, raid_disks);
2592                         if (*dd_idx >= pd_idx)
2593                                 (*dd_idx)++;
2594                         break;
2595                 case ALGORITHM_LEFT_SYMMETRIC:
2596                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2597                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2598                         break;
2599                 case ALGORITHM_RIGHT_SYMMETRIC:
2600                         pd_idx = sector_div(stripe2, raid_disks);
2601                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2602                         break;
2603                 case ALGORITHM_PARITY_0:
2604                         pd_idx = 0;
2605                         (*dd_idx)++;
2606                         break;
2607                 case ALGORITHM_PARITY_N:
2608                         pd_idx = data_disks;
2609                         break;
2610                 default:
2611                         BUG();
2612                 }
2613                 break;
2614         case 6:
2615
2616                 switch (algorithm) {
2617                 case ALGORITHM_LEFT_ASYMMETRIC:
2618                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2619                         qd_idx = pd_idx + 1;
2620                         if (pd_idx == raid_disks-1) {
2621                                 (*dd_idx)++;    /* Q D D D P */
2622                                 qd_idx = 0;
2623                         } else if (*dd_idx >= pd_idx)
2624                                 (*dd_idx) += 2; /* D D P Q D */
2625                         break;
2626                 case ALGORITHM_RIGHT_ASYMMETRIC:
2627                         pd_idx = sector_div(stripe2, raid_disks);
2628                         qd_idx = pd_idx + 1;
2629                         if (pd_idx == raid_disks-1) {
2630                                 (*dd_idx)++;    /* Q D D D P */
2631                                 qd_idx = 0;
2632                         } else if (*dd_idx >= pd_idx)
2633                                 (*dd_idx) += 2; /* D D P Q D */
2634                         break;
2635                 case ALGORITHM_LEFT_SYMMETRIC:
2636                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2637                         qd_idx = (pd_idx + 1) % raid_disks;
2638                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2639                         break;
2640                 case ALGORITHM_RIGHT_SYMMETRIC:
2641                         pd_idx = sector_div(stripe2, raid_disks);
2642                         qd_idx = (pd_idx + 1) % raid_disks;
2643                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2644                         break;
2645
2646                 case ALGORITHM_PARITY_0:
2647                         pd_idx = 0;
2648                         qd_idx = 1;
2649                         (*dd_idx) += 2;
2650                         break;
2651                 case ALGORITHM_PARITY_N:
2652                         pd_idx = data_disks;
2653                         qd_idx = data_disks + 1;
2654                         break;
2655
2656                 case ALGORITHM_ROTATING_ZERO_RESTART:
2657                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2658                          * of blocks for computing Q is different.
2659                          */
2660                         pd_idx = sector_div(stripe2, raid_disks);
2661                         qd_idx = pd_idx + 1;
2662                         if (pd_idx == raid_disks-1) {
2663                                 (*dd_idx)++;    /* Q D D D P */
2664                                 qd_idx = 0;
2665                         } else if (*dd_idx >= pd_idx)
2666                                 (*dd_idx) += 2; /* D D P Q D */
2667                         ddf_layout = 1;
2668                         break;
2669
2670                 case ALGORITHM_ROTATING_N_RESTART:
2671                         /* Same a left_asymmetric, by first stripe is
2672                          * D D D P Q  rather than
2673                          * Q D D D P
2674                          */
2675                         stripe2 += 1;
2676                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2677                         qd_idx = pd_idx + 1;
2678                         if (pd_idx == raid_disks-1) {
2679                                 (*dd_idx)++;    /* Q D D D P */
2680                                 qd_idx = 0;
2681                         } else if (*dd_idx >= pd_idx)
2682                                 (*dd_idx) += 2; /* D D P Q D */
2683                         ddf_layout = 1;
2684                         break;
2685
2686                 case ALGORITHM_ROTATING_N_CONTINUE:
2687                         /* Same as left_symmetric but Q is before P */
2688                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2689                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2690                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2691                         ddf_layout = 1;
2692                         break;
2693
2694                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2695                         /* RAID5 left_asymmetric, with Q on last device */
2696                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2697                         if (*dd_idx >= pd_idx)
2698                                 (*dd_idx)++;
2699                         qd_idx = raid_disks - 1;
2700                         break;
2701
2702                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2703                         pd_idx = sector_div(stripe2, raid_disks-1);
2704                         if (*dd_idx >= pd_idx)
2705                                 (*dd_idx)++;
2706                         qd_idx = raid_disks - 1;
2707                         break;
2708
2709                 case ALGORITHM_LEFT_SYMMETRIC_6:
2710                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2711                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2712                         qd_idx = raid_disks - 1;
2713                         break;
2714
2715                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2716                         pd_idx = sector_div(stripe2, raid_disks-1);
2717                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2718                         qd_idx = raid_disks - 1;
2719                         break;
2720
2721                 case ALGORITHM_PARITY_0_6:
2722                         pd_idx = 0;
2723                         (*dd_idx)++;
2724                         qd_idx = raid_disks - 1;
2725                         break;
2726
2727                 default:
2728                         BUG();
2729                 }
2730                 break;
2731         }
2732
2733         if (sh) {
2734                 sh->pd_idx = pd_idx;
2735                 sh->qd_idx = qd_idx;
2736                 sh->ddf_layout = ddf_layout;
2737         }
2738         /*
2739          * Finally, compute the new sector number
2740          */
2741         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2742         return new_sector;
2743 }
2744
2745 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2746 {
2747         struct r5conf *conf = sh->raid_conf;
2748         int raid_disks = sh->disks;
2749         int data_disks = raid_disks - conf->max_degraded;
2750         sector_t new_sector = sh->sector, check;
2751         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2752                                          : conf->chunk_sectors;
2753         int algorithm = previous ? conf->prev_algo
2754                                  : conf->algorithm;
2755         sector_t stripe;
2756         int chunk_offset;
2757         sector_t chunk_number;
2758         int dummy1, dd_idx = i;
2759         sector_t r_sector;
2760         struct stripe_head sh2;
2761
2762         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2763         stripe = new_sector;
2764
2765         if (i == sh->pd_idx)
2766                 return 0;
2767         switch(conf->level) {
2768         case 4: break;
2769         case 5:
2770                 switch (algorithm) {
2771                 case ALGORITHM_LEFT_ASYMMETRIC:
2772                 case ALGORITHM_RIGHT_ASYMMETRIC:
2773                         if (i > sh->pd_idx)
2774                                 i--;
2775                         break;
2776                 case ALGORITHM_LEFT_SYMMETRIC:
2777                 case ALGORITHM_RIGHT_SYMMETRIC:
2778                         if (i < sh->pd_idx)
2779                                 i += raid_disks;
2780                         i -= (sh->pd_idx + 1);
2781                         break;
2782                 case ALGORITHM_PARITY_0:
2783                         i -= 1;
2784                         break;
2785                 case ALGORITHM_PARITY_N:
2786                         break;
2787                 default:
2788                         BUG();
2789                 }
2790                 break;
2791         case 6:
2792                 if (i == sh->qd_idx)
2793                         return 0; /* It is the Q disk */
2794                 switch (algorithm) {
2795                 case ALGORITHM_LEFT_ASYMMETRIC:
2796                 case ALGORITHM_RIGHT_ASYMMETRIC:
2797                 case ALGORITHM_ROTATING_ZERO_RESTART:
2798                 case ALGORITHM_ROTATING_N_RESTART:
2799                         if (sh->pd_idx == raid_disks-1)
2800                                 i--;    /* Q D D D P */
2801                         else if (i > sh->pd_idx)
2802                                 i -= 2; /* D D P Q D */
2803                         break;
2804                 case ALGORITHM_LEFT_SYMMETRIC:
2805                 case ALGORITHM_RIGHT_SYMMETRIC:
2806                         if (sh->pd_idx == raid_disks-1)
2807                                 i--; /* Q D D D P */
2808                         else {
2809                                 /* D D P Q D */
2810                                 if (i < sh->pd_idx)
2811                                         i += raid_disks;
2812                                 i -= (sh->pd_idx + 2);
2813                         }
2814                         break;
2815                 case ALGORITHM_PARITY_0:
2816                         i -= 2;
2817                         break;
2818                 case ALGORITHM_PARITY_N:
2819                         break;
2820                 case ALGORITHM_ROTATING_N_CONTINUE:
2821                         /* Like left_symmetric, but P is before Q */
2822                         if (sh->pd_idx == 0)
2823                                 i--;    /* P D D D Q */
2824                         else {
2825                                 /* D D Q P D */
2826                                 if (i < sh->pd_idx)
2827                                         i += raid_disks;
2828                                 i -= (sh->pd_idx + 1);
2829                         }
2830                         break;
2831                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2832                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2833                         if (i > sh->pd_idx)
2834                                 i--;
2835                         break;
2836                 case ALGORITHM_LEFT_SYMMETRIC_6:
2837                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2838                         if (i < sh->pd_idx)
2839                                 i += data_disks + 1;
2840                         i -= (sh->pd_idx + 1);
2841                         break;
2842                 case ALGORITHM_PARITY_0_6:
2843                         i -= 1;
2844                         break;
2845                 default:
2846                         BUG();
2847                 }
2848                 break;
2849         }
2850
2851         chunk_number = stripe * data_disks + i;
2852         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2853
2854         check = raid5_compute_sector(conf, r_sector,
2855                                      previous, &dummy1, &sh2);
2856         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2857                 || sh2.qd_idx != sh->qd_idx) {
2858                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2859                        mdname(conf->mddev));
2860                 return 0;
2861         }
2862         return r_sector;
2863 }
2864
2865 static void
2866 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2867                          int rcw, int expand)
2868 {
2869         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2870         struct r5conf *conf = sh->raid_conf;
2871         int level = conf->level;
2872
2873         if (rcw) {
2874
2875                 for (i = disks; i--; ) {
2876                         struct r5dev *dev = &sh->dev[i];
2877
2878                         if (dev->towrite) {
2879                                 set_bit(R5_LOCKED, &dev->flags);
2880                                 set_bit(R5_Wantdrain, &dev->flags);
2881                                 if (!expand)
2882                                         clear_bit(R5_UPTODATE, &dev->flags);
2883                                 s->locked++;
2884                         }
2885                 }
2886                 /* if we are not expanding this is a proper write request, and
2887                  * there will be bios with new data to be drained into the
2888                  * stripe cache
2889                  */
2890                 if (!expand) {
2891                         if (!s->locked)
2892                                 /* False alarm, nothing to do */
2893                                 return;
2894                         sh->reconstruct_state = reconstruct_state_drain_run;
2895                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2896                 } else
2897                         sh->reconstruct_state = reconstruct_state_run;
2898
2899                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2900
2901                 if (s->locked + conf->max_degraded == disks)
2902                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2903                                 atomic_inc(&conf->pending_full_writes);
2904         } else {
2905                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2906                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2907                 BUG_ON(level == 6 &&
2908                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2909                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2910
2911                 for (i = disks; i--; ) {
2912                         struct r5dev *dev = &sh->dev[i];
2913                         if (i == pd_idx || i == qd_idx)
2914                                 continue;
2915
2916                         if (dev->towrite &&
2917                             (test_bit(R5_UPTODATE, &dev->flags) ||
2918                              test_bit(R5_Wantcompute, &dev->flags))) {
2919                                 set_bit(R5_Wantdrain, &dev->flags);
2920                                 set_bit(R5_LOCKED, &dev->flags);
2921                                 clear_bit(R5_UPTODATE, &dev->flags);
2922                                 s->locked++;
2923                         }
2924                 }
2925                 if (!s->locked)
2926                         /* False alarm - nothing to do */
2927                         return;
2928                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2929                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2930                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2931                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2932         }
2933
2934         /* keep the parity disk(s) locked while asynchronous operations
2935          * are in flight
2936          */
2937         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2938         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2939         s->locked++;
2940
2941         if (level == 6) {
2942                 int qd_idx = sh->qd_idx;
2943                 struct r5dev *dev = &sh->dev[qd_idx];
2944
2945                 set_bit(R5_LOCKED, &dev->flags);
2946                 clear_bit(R5_UPTODATE, &dev->flags);
2947                 s->locked++;
2948         }
2949
2950         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2951                 __func__, (unsigned long long)sh->sector,
2952                 s->locked, s->ops_request);
2953 }
2954
2955 /*
2956  * Each stripe/dev can have one or more bion attached.
2957  * toread/towrite point to the first in a chain.
2958  * The bi_next chain must be in order.
2959  */
2960 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2961                           int forwrite, int previous)
2962 {
2963         struct bio **bip;
2964         struct r5conf *conf = sh->raid_conf;
2965         int firstwrite=0;
2966
2967         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2968                 (unsigned long long)bi->bi_iter.bi_sector,
2969                 (unsigned long long)sh->sector);
2970
2971         /*
2972          * If several bio share a stripe. The bio bi_phys_segments acts as a
2973          * reference count to avoid race. The reference count should already be
2974          * increased before this function is called (for example, in
2975          * make_request()), so other bio sharing this stripe will not free the
2976          * stripe. If a stripe is owned by one stripe, the stripe lock will
2977          * protect it.
2978          */
2979         spin_lock_irq(&sh->stripe_lock);
2980         /* Don't allow new IO added to stripes in batch list */
2981         if (sh->batch_head)
2982                 goto overlap;
2983         if (forwrite) {
2984                 bip = &sh->dev[dd_idx].towrite;
2985                 if (*bip == NULL)
2986                         firstwrite = 1;
2987         } else
2988                 bip = &sh->dev[dd_idx].toread;
2989         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2990                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2991                         goto overlap;
2992                 bip = & (*bip)->bi_next;
2993         }
2994         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2995                 goto overlap;
2996
2997         if (!forwrite || previous)
2998                 clear_bit(STRIPE_BATCH_READY, &sh->state);
2999
3000         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3001         if (*bip)
3002                 bi->bi_next = *bip;
3003         *bip = bi;
3004         raid5_inc_bi_active_stripes(bi);
3005
3006         if (forwrite) {
3007                 /* check if page is covered */
3008                 sector_t sector = sh->dev[dd_idx].sector;
3009                 for (bi=sh->dev[dd_idx].towrite;
3010                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3011                              bi && bi->bi_iter.bi_sector <= sector;
3012                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3013                         if (bio_end_sector(bi) >= sector)
3014                                 sector = bio_end_sector(bi);
3015                 }
3016                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3017                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3018                                 sh->overwrite_disks++;
3019         }
3020
3021         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3022                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3023                 (unsigned long long)sh->sector, dd_idx);
3024
3025         if (conf->mddev->bitmap && firstwrite) {
3026                 /* Cannot hold spinlock over bitmap_startwrite,
3027                  * but must ensure this isn't added to a batch until
3028                  * we have added to the bitmap and set bm_seq.
3029                  * So set STRIPE_BITMAP_PENDING to prevent
3030                  * batching.
3031                  * If multiple add_stripe_bio() calls race here they
3032                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3033                  * to complete "bitmap_startwrite" gets to set
3034                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3035                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3036                  * any more.
3037                  */
3038                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3039                 spin_unlock_irq(&sh->stripe_lock);
3040                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3041                                   STRIPE_SECTORS, 0);
3042                 spin_lock_irq(&sh->stripe_lock);
3043                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3044                 if (!sh->batch_head) {
3045                         sh->bm_seq = conf->seq_flush+1;
3046                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3047                 }
3048         }
3049         spin_unlock_irq(&sh->stripe_lock);
3050
3051         if (stripe_can_batch(sh))
3052                 stripe_add_to_batch_list(conf, sh);
3053         return 1;
3054
3055  overlap:
3056         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3057         spin_unlock_irq(&sh->stripe_lock);
3058         return 0;
3059 }
3060
3061 static void end_reshape(struct r5conf *conf);
3062
3063 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3064                             struct stripe_head *sh)
3065 {
3066         int sectors_per_chunk =
3067                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3068         int dd_idx;
3069         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3070         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3071
3072         raid5_compute_sector(conf,
3073                              stripe * (disks - conf->max_degraded)
3074                              *sectors_per_chunk + chunk_offset,
3075                              previous,
3076                              &dd_idx, sh);
3077 }
3078
3079 static void
3080 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3081                                 struct stripe_head_state *s, int disks,
3082                                 struct bio_list *return_bi)
3083 {
3084         int i;
3085         BUG_ON(sh->batch_head);
3086         for (i = disks; i--; ) {
3087                 struct bio *bi;
3088                 int bitmap_end = 0;
3089
3090                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3091                         struct md_rdev *rdev;
3092                         rcu_read_lock();
3093                         rdev = rcu_dereference(conf->disks[i].rdev);
3094                         if (rdev && test_bit(In_sync, &rdev->flags))
3095                                 atomic_inc(&rdev->nr_pending);
3096                         else
3097                                 rdev = NULL;
3098                         rcu_read_unlock();
3099                         if (rdev) {
3100                                 if (!rdev_set_badblocks(
3101                                             rdev,
3102                                             sh->sector,
3103                                             STRIPE_SECTORS, 0))
3104                                         md_error(conf->mddev, rdev);
3105                                 rdev_dec_pending(rdev, conf->mddev);
3106                         }
3107                 }
3108                 spin_lock_irq(&sh->stripe_lock);
3109                 /* fail all writes first */
3110                 bi = sh->dev[i].towrite;
3111                 sh->dev[i].towrite = NULL;
3112                 sh->overwrite_disks = 0;
3113                 spin_unlock_irq(&sh->stripe_lock);
3114                 if (bi)
3115                         bitmap_end = 1;
3116
3117                 r5l_stripe_write_finished(sh);
3118
3119                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3120                         wake_up(&conf->wait_for_overlap);
3121
3122                 while (bi && bi->bi_iter.bi_sector <
3123                         sh->dev[i].sector + STRIPE_SECTORS) {
3124                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3125
3126                         bi->bi_error = -EIO;
3127                         if (!raid5_dec_bi_active_stripes(bi)) {
3128                                 md_write_end(conf->mddev);
3129                                 bio_list_add(return_bi, bi);
3130                         }
3131                         bi = nextbi;
3132                 }
3133                 if (bitmap_end)
3134                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3135                                 STRIPE_SECTORS, 0, 0);
3136                 bitmap_end = 0;
3137                 /* and fail all 'written' */
3138                 bi = sh->dev[i].written;
3139                 sh->dev[i].written = NULL;
3140                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3141                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3142                         sh->dev[i].page = sh->dev[i].orig_page;
3143                 }
3144
3145                 if (bi) bitmap_end = 1;
3146                 while (bi && bi->bi_iter.bi_sector <
3147                        sh->dev[i].sector + STRIPE_SECTORS) {
3148                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3149
3150                         bi->bi_error = -EIO;
3151                         if (!raid5_dec_bi_active_stripes(bi)) {
3152                                 md_write_end(conf->mddev);
3153                                 bio_list_add(return_bi, bi);
3154                         }
3155                         bi = bi2;
3156                 }
3157
3158                 /* fail any reads if this device is non-operational and
3159                  * the data has not reached the cache yet.
3160                  */
3161                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3162                     s->failed > conf->max_degraded &&
3163                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3164                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3165                         spin_lock_irq(&sh->stripe_lock);
3166                         bi = sh->dev[i].toread;
3167                         sh->dev[i].toread = NULL;
3168                         spin_unlock_irq(&sh->stripe_lock);
3169                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3170                                 wake_up(&conf->wait_for_overlap);
3171                         if (bi)
3172                                 s->to_read--;
3173                         while (bi && bi->bi_iter.bi_sector <
3174                                sh->dev[i].sector + STRIPE_SECTORS) {
3175                                 struct bio *nextbi =
3176                                         r5_next_bio(bi, sh->dev[i].sector);
3177
3178                                 bi->bi_error = -EIO;
3179                                 if (!raid5_dec_bi_active_stripes(bi))
3180                                         bio_list_add(return_bi, bi);
3181                                 bi = nextbi;
3182                         }
3183                 }
3184                 if (bitmap_end)
3185                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3186                                         STRIPE_SECTORS, 0, 0);
3187                 /* If we were in the middle of a write the parity block might
3188                  * still be locked - so just clear all R5_LOCKED flags
3189                  */
3190                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3191         }
3192         s->to_write = 0;
3193         s->written = 0;
3194
3195         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3196                 if (atomic_dec_and_test(&conf->pending_full_writes))
3197                         md_wakeup_thread(conf->mddev->thread);
3198 }
3199
3200 static void
3201 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3202                    struct stripe_head_state *s)
3203 {
3204         int abort = 0;
3205         int i;
3206
3207         BUG_ON(sh->batch_head);
3208         clear_bit(STRIPE_SYNCING, &sh->state);
3209         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3210                 wake_up(&conf->wait_for_overlap);
3211         s->syncing = 0;
3212         s->replacing = 0;
3213         /* There is nothing more to do for sync/check/repair.
3214          * Don't even need to abort as that is handled elsewhere
3215          * if needed, and not always wanted e.g. if there is a known
3216          * bad block here.
3217          * For recover/replace we need to record a bad block on all
3218          * non-sync devices, or abort the recovery
3219          */
3220         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3221                 /* During recovery devices cannot be removed, so
3222                  * locking and refcounting of rdevs is not needed
3223                  */
3224                 for (i = 0; i < conf->raid_disks; i++) {
3225                         struct md_rdev *rdev = conf->disks[i].rdev;
3226                         if (rdev
3227                             && !test_bit(Faulty, &rdev->flags)
3228                             && !test_bit(In_sync, &rdev->flags)
3229                             && !rdev_set_badblocks(rdev, sh->sector,
3230                                                    STRIPE_SECTORS, 0))
3231                                 abort = 1;
3232                         rdev = conf->disks[i].replacement;
3233                         if (rdev
3234                             && !test_bit(Faulty, &rdev->flags)
3235                             && !test_bit(In_sync, &rdev->flags)
3236                             && !rdev_set_badblocks(rdev, sh->sector,
3237                                                    STRIPE_SECTORS, 0))
3238                                 abort = 1;
3239                 }
3240                 if (abort)
3241                         conf->recovery_disabled =
3242                                 conf->mddev->recovery_disabled;
3243         }
3244         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3245 }
3246
3247 static int want_replace(struct stripe_head *sh, int disk_idx)
3248 {
3249         struct md_rdev *rdev;
3250         int rv = 0;
3251         /* Doing recovery so rcu locking not required */
3252         rdev = sh->raid_conf->disks[disk_idx].replacement;
3253         if (rdev
3254             && !test_bit(Faulty, &rdev->flags)
3255             && !test_bit(In_sync, &rdev->flags)
3256             && (rdev->recovery_offset <= sh->sector
3257                 || rdev->mddev->recovery_cp <= sh->sector))
3258                 rv = 1;
3259
3260         return rv;
3261 }
3262
3263 /* fetch_block - checks the given member device to see if its data needs
3264  * to be read or computed to satisfy a request.
3265  *
3266  * Returns 1 when no more member devices need to be checked, otherwise returns
3267  * 0 to tell the loop in handle_stripe_fill to continue
3268  */
3269
3270 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3271                            int disk_idx, int disks)
3272 {
3273         struct r5dev *dev = &sh->dev[disk_idx];
3274         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3275                                   &sh->dev[s->failed_num[1]] };
3276         int i;
3277
3278
3279         if (test_bit(R5_LOCKED, &dev->flags) ||
3280             test_bit(R5_UPTODATE, &dev->flags))
3281                 /* No point reading this as we already have it or have
3282                  * decided to get it.
3283                  */
3284                 return 0;
3285
3286         if (dev->toread ||
3287             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3288                 /* We need this block to directly satisfy a request */
3289                 return 1;
3290
3291         if (s->syncing || s->expanding ||
3292             (s->replacing && want_replace(sh, disk_idx)))
3293                 /* When syncing, or expanding we read everything.
3294                  * When replacing, we need the replaced block.
3295                  */
3296                 return 1;
3297
3298         if ((s->failed >= 1 && fdev[0]->toread) ||
3299             (s->failed >= 2 && fdev[1]->toread))
3300                 /* If we want to read from a failed device, then
3301                  * we need to actually read every other device.
3302                  */
3303                 return 1;
3304
3305         /* Sometimes neither read-modify-write nor reconstruct-write
3306          * cycles can work.  In those cases we read every block we
3307          * can.  Then the parity-update is certain to have enough to
3308          * work with.
3309          * This can only be a problem when we need to write something,
3310          * and some device has failed.  If either of those tests
3311          * fail we need look no further.
3312          */
3313         if (!s->failed || !s->to_write)
3314                 return 0;
3315
3316         if (test_bit(R5_Insync, &dev->flags) &&
3317             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3318                 /* Pre-reads at not permitted until after short delay
3319                  * to gather multiple requests.  However if this
3320                  * device is no Insync, the block could only be be computed
3321                  * and there is no need to delay that.
3322                  */
3323                 return 0;
3324
3325         for (i = 0; i < s->failed && i < 2; i++) {
3326                 if (fdev[i]->towrite &&
3327                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3328                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3329                         /* If we have a partial write to a failed
3330                          * device, then we will need to reconstruct
3331                          * the content of that device, so all other
3332                          * devices must be read.
3333                          */
3334                         return 1;
3335         }
3336
3337         /* If we are forced to do a reconstruct-write, either because
3338          * the current RAID6 implementation only supports that, or
3339          * or because parity cannot be trusted and we are currently
3340          * recovering it, there is extra need to be careful.
3341          * If one of the devices that we would need to read, because
3342          * it is not being overwritten (and maybe not written at all)
3343          * is missing/faulty, then we need to read everything we can.
3344          */
3345         if (sh->raid_conf->level != 6 &&
3346             sh->sector < sh->raid_conf->mddev->recovery_cp)
3347                 /* reconstruct-write isn't being forced */
3348                 return 0;
3349         for (i = 0; i < s->failed && i < 2; i++) {
3350                 if (s->failed_num[i] != sh->pd_idx &&
3351                     s->failed_num[i] != sh->qd_idx &&
3352                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3353                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3354                         return 1;
3355         }
3356
3357         return 0;
3358 }
3359
3360 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3361                        int disk_idx, int disks)
3362 {
3363         struct r5dev *dev = &sh->dev[disk_idx];
3364
3365         /* is the data in this block needed, and can we get it? */
3366         if (need_this_block(sh, s, disk_idx, disks)) {
3367                 /* we would like to get this block, possibly by computing it,
3368                  * otherwise read it if the backing disk is insync
3369                  */
3370                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3371                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3372                 BUG_ON(sh->batch_head);
3373                 if ((s->uptodate == disks - 1) &&
3374                     (s->failed && (disk_idx == s->failed_num[0] ||
3375                                    disk_idx == s->failed_num[1]))) {
3376                         /* have disk failed, and we're requested to fetch it;
3377                          * do compute it
3378                          */
3379                         pr_debug("Computing stripe %llu block %d\n",
3380                                (unsigned long long)sh->sector, disk_idx);
3381                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3382                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3383                         set_bit(R5_Wantcompute, &dev->flags);
3384                         sh->ops.target = disk_idx;
3385                         sh->ops.target2 = -1; /* no 2nd target */
3386                         s->req_compute = 1;
3387                         /* Careful: from this point on 'uptodate' is in the eye
3388                          * of raid_run_ops which services 'compute' operations
3389                          * before writes. R5_Wantcompute flags a block that will
3390                          * be R5_UPTODATE by the time it is needed for a
3391                          * subsequent operation.
3392                          */
3393                         s->uptodate++;
3394                         return 1;
3395                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3396                         /* Computing 2-failure is *very* expensive; only
3397                          * do it if failed >= 2
3398                          */
3399                         int other;
3400                         for (other = disks; other--; ) {
3401                                 if (other == disk_idx)
3402                                         continue;
3403                                 if (!test_bit(R5_UPTODATE,
3404                                       &sh->dev[other].flags))
3405                                         break;
3406                         }
3407                         BUG_ON(other < 0);
3408                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3409                                (unsigned long long)sh->sector,
3410                                disk_idx, other);
3411                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3412                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3413                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3414                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3415                         sh->ops.target = disk_idx;
3416                         sh->ops.target2 = other;
3417                         s->uptodate += 2;
3418                         s->req_compute = 1;
3419                         return 1;
3420                 } else if (test_bit(R5_Insync, &dev->flags)) {
3421                         set_bit(R5_LOCKED, &dev->flags);
3422                         set_bit(R5_Wantread, &dev->flags);
3423                         s->locked++;
3424                         pr_debug("Reading block %d (sync=%d)\n",
3425                                 disk_idx, s->syncing);
3426                 }
3427         }
3428
3429         return 0;
3430 }
3431
3432 /**
3433  * handle_stripe_fill - read or compute data to satisfy pending requests.
3434  */
3435 static void handle_stripe_fill(struct stripe_head *sh,
3436                                struct stripe_head_state *s,
3437                                int disks)
3438 {
3439         int i;
3440
3441         /* look for blocks to read/compute, skip this if a compute
3442          * is already in flight, or if the stripe contents are in the
3443          * midst of changing due to a write
3444          */
3445         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3446             !sh->reconstruct_state)
3447                 for (i = disks; i--; )
3448                         if (fetch_block(sh, s, i, disks))
3449                                 break;
3450         set_bit(STRIPE_HANDLE, &sh->state);
3451 }
3452
3453 static void break_stripe_batch_list(struct stripe_head *head_sh,
3454                                     unsigned long handle_flags);
3455 /* handle_stripe_clean_event
3456  * any written block on an uptodate or failed drive can be returned.
3457  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3458  * never LOCKED, so we don't need to test 'failed' directly.
3459  */
3460 static void handle_stripe_clean_event(struct r5conf *conf,
3461         struct stripe_head *sh, int disks, struct bio_list *return_bi)
3462 {
3463         int i;
3464         struct r5dev *dev;
3465         int discard_pending = 0;
3466         struct stripe_head *head_sh = sh;
3467         bool do_endio = false;
3468
3469         for (i = disks; i--; )
3470                 if (sh->dev[i].written) {
3471                         dev = &sh->dev[i];
3472                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3473                             (test_bit(R5_UPTODATE, &dev->flags) ||
3474                              test_bit(R5_Discard, &dev->flags) ||
3475                              test_bit(R5_SkipCopy, &dev->flags))) {
3476                                 /* We can return any write requests */
3477                                 struct bio *wbi, *wbi2;
3478                                 pr_debug("Return write for disc %d\n", i);
3479                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3480                                         clear_bit(R5_UPTODATE, &dev->flags);
3481                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3482                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3483                                 }
3484                                 do_endio = true;
3485
3486 returnbi:
3487                                 dev->page = dev->orig_page;
3488                                 wbi = dev->written;
3489                                 dev->written = NULL;
3490                                 while (wbi && wbi->bi_iter.bi_sector <
3491                                         dev->sector + STRIPE_SECTORS) {
3492                                         wbi2 = r5_next_bio(wbi, dev->sector);
3493                                         if (!raid5_dec_bi_active_stripes(wbi)) {
3494                                                 md_write_end(conf->mddev);
3495                                                 bio_list_add(return_bi, wbi);
3496                                         }
3497                                         wbi = wbi2;
3498                                 }
3499                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3500                                                 STRIPE_SECTORS,
3501                                          !test_bit(STRIPE_DEGRADED, &sh->state),
3502                                                 0);
3503                                 if (head_sh->batch_head) {
3504                                         sh = list_first_entry(&sh->batch_list,
3505                                                               struct stripe_head,
3506                                                               batch_list);
3507                                         if (sh != head_sh) {
3508                                                 dev = &sh->dev[i];
3509                                                 goto returnbi;
3510                                         }
3511                                 }
3512                                 sh = head_sh;
3513                                 dev = &sh->dev[i];
3514                         } else if (test_bit(R5_Discard, &dev->flags))
3515                                 discard_pending = 1;
3516                         WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3517                         WARN_ON(dev->page != dev->orig_page);
3518                 }
3519
3520         r5l_stripe_write_finished(sh);
3521
3522         if (!discard_pending &&
3523             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3524                 int hash;
3525                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3526                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3527                 if (sh->qd_idx >= 0) {
3528                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3529                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3530                 }
3531                 /* now that discard is done we can proceed with any sync */
3532                 clear_bit(STRIPE_DISCARD, &sh->state);
3533                 /*
3534                  * SCSI discard will change some bio fields and the stripe has
3535                  * no updated data, so remove it from hash list and the stripe
3536                  * will be reinitialized
3537                  */
3538 unhash:
3539                 hash = sh->hash_lock_index;
3540                 spin_lock_irq(conf->hash_locks + hash);
3541                 remove_hash(sh);
3542                 spin_unlock_irq(conf->hash_locks + hash);
3543                 if (head_sh->batch_head) {
3544                         sh = list_first_entry(&sh->batch_list,
3545                                               struct stripe_head, batch_list);
3546                         if (sh != head_sh)
3547                                         goto unhash;
3548                 }
3549                 sh = head_sh;
3550
3551                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3552                         set_bit(STRIPE_HANDLE, &sh->state);
3553
3554         }
3555
3556         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3557                 if (atomic_dec_and_test(&conf->pending_full_writes))
3558                         md_wakeup_thread(conf->mddev->thread);
3559
3560         if (head_sh->batch_head && do_endio)
3561                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3562 }
3563
3564 static void handle_stripe_dirtying(struct r5conf *conf,
3565                                    struct stripe_head *sh,
3566                                    struct stripe_head_state *s,
3567                                    int disks)
3568 {
3569         int rmw = 0, rcw = 0, i;
3570         sector_t recovery_cp = conf->mddev->recovery_cp;
3571
3572         /* Check whether resync is now happening or should start.
3573          * If yes, then the array is dirty (after unclean shutdown or
3574          * initial creation), so parity in some stripes might be inconsistent.
3575          * In this case, we need to always do reconstruct-write, to ensure
3576          * that in case of drive failure or read-error correction, we
3577          * generate correct data from the parity.
3578          */
3579         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3580             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3581              s->failed == 0)) {
3582                 /* Calculate the real rcw later - for now make it
3583                  * look like rcw is cheaper
3584                  */
3585                 rcw = 1; rmw = 2;
3586                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3587                          conf->rmw_level, (unsigned long long)recovery_cp,
3588                          (unsigned long long)sh->sector);
3589         } else for (i = disks; i--; ) {
3590                 /* would I have to read this buffer for read_modify_write */
3591                 struct r5dev *dev = &sh->dev[i];
3592                 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3593                     !test_bit(R5_LOCKED, &dev->flags) &&
3594                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3595                       test_bit(R5_Wantcompute, &dev->flags))) {
3596                         if (test_bit(R5_Insync, &dev->flags))
3597                                 rmw++;
3598                         else
3599                                 rmw += 2*disks;  /* cannot read it */
3600                 }
3601                 /* Would I have to read this buffer for reconstruct_write */
3602                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3603                     i != sh->pd_idx && i != sh->qd_idx &&
3604                     !test_bit(R5_LOCKED, &dev->flags) &&
3605                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3606                     test_bit(R5_Wantcompute, &dev->flags))) {
3607                         if (test_bit(R5_Insync, &dev->flags))
3608                                 rcw++;
3609                         else
3610                                 rcw += 2*disks;
3611                 }
3612         }
3613         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3614                 (unsigned long long)sh->sector, rmw, rcw);
3615         set_bit(STRIPE_HANDLE, &sh->state);
3616         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3617                 /* prefer read-modify-write, but need to get some data */
3618                 if (conf->mddev->queue)
3619                         blk_add_trace_msg(conf->mddev->queue,
3620                                           "raid5 rmw %llu %d",
3621                                           (unsigned long long)sh->sector, rmw);
3622                 for (i = disks; i--; ) {
3623                         struct r5dev *dev = &sh->dev[i];
3624                         if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3625                             !test_bit(R5_LOCKED, &dev->flags) &&
3626                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3627                             test_bit(R5_Wantcompute, &dev->flags)) &&
3628                             test_bit(R5_Insync, &dev->flags)) {
3629                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3630                                              &sh->state)) {
3631                                         pr_debug("Read_old block %d for r-m-w\n",
3632                                                  i);
3633                                         set_bit(R5_LOCKED, &dev->flags);
3634                                         set_bit(R5_Wantread, &dev->flags);
3635                                         s->locked++;
3636                                 } else {
3637                                         set_bit(STRIPE_DELAYED, &sh->state);
3638                                         set_bit(STRIPE_HANDLE, &sh->state);
3639                                 }
3640                         }
3641                 }
3642         }
3643         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3644                 /* want reconstruct write, but need to get some data */
3645                 int qread =0;
3646                 rcw = 0;
3647                 for (i = disks; i--; ) {
3648                         struct r5dev *dev = &sh->dev[i];
3649                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3650                             i != sh->pd_idx && i != sh->qd_idx &&
3651                             !test_bit(R5_LOCKED, &dev->flags) &&
3652                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3653                               test_bit(R5_Wantcompute, &dev->flags))) {
3654                                 rcw++;
3655                                 if (test_bit(R5_Insync, &dev->flags) &&
3656                                     test_bit(STRIPE_PREREAD_ACTIVE,
3657                                              &sh->state)) {
3658                                         pr_debug("Read_old block "
3659                                                 "%d for Reconstruct\n", i);
3660                                         set_bit(R5_LOCKED, &dev->flags);
3661                                         set_bit(R5_Wantread, &dev->flags);
3662                                         s->locked++;
3663                                         qread++;
3664                                 } else {
3665                                         set_bit(STRIPE_DELAYED, &sh->state);
3666                                         set_bit(STRIPE_HANDLE, &sh->state);
3667                                 }
3668                         }
3669                 }
3670                 if (rcw && conf->mddev->queue)
3671                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3672                                           (unsigned long long)sh->sector,
3673                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3674         }
3675
3676         if (rcw > disks && rmw > disks &&
3677             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3678                 set_bit(STRIPE_DELAYED, &sh->state);
3679
3680         /* now if nothing is locked, and if we have enough data,
3681          * we can start a write request
3682          */
3683         /* since handle_stripe can be called at any time we need to handle the
3684          * case where a compute block operation has been submitted and then a
3685          * subsequent call wants to start a write request.  raid_run_ops only
3686          * handles the case where compute block and reconstruct are requested
3687          * simultaneously.  If this is not the case then new writes need to be
3688          * held off until the compute completes.
3689          */
3690         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3691             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3692             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3693                 schedule_reconstruction(sh, s, rcw == 0, 0);
3694 }
3695
3696 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3697                                 struct stripe_head_state *s, int disks)
3698 {
3699         struct r5dev *dev = NULL;
3700
3701         BUG_ON(sh->batch_head);
3702         set_bit(STRIPE_HANDLE, &sh->state);
3703
3704         switch (sh->check_state) {
3705         case check_state_idle:
3706                 /* start a new check operation if there are no failures */
3707                 if (s->failed == 0) {
3708                         BUG_ON(s->uptodate != disks);
3709                         sh->check_state = check_state_run;
3710                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3711                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3712                         s->uptodate--;
3713                         break;
3714                 }
3715                 dev = &sh->dev[s->failed_num[0]];
3716                 /* fall through */
3717         case check_state_compute_result:
3718                 sh->check_state = check_state_idle;
3719                 if (!dev)
3720                         dev = &sh->dev[sh->pd_idx];
3721
3722                 /* check that a write has not made the stripe insync */
3723                 if (test_bit(STRIPE_INSYNC, &sh->state))
3724                         break;
3725
3726                 /* either failed parity check, or recovery is happening */
3727                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3728                 BUG_ON(s->uptodate != disks);
3729
3730                 set_bit(R5_LOCKED, &dev->flags);
3731                 s->locked++;
3732                 set_bit(R5_Wantwrite, &dev->flags);
3733
3734                 clear_bit(STRIPE_DEGRADED, &sh->state);
3735                 set_bit(STRIPE_INSYNC, &sh->state);
3736                 break;
3737         case check_state_run:
3738                 break; /* we will be called again upon completion */
3739         case check_state_check_result:
3740                 sh->check_state = check_state_idle;
3741
3742                 /* if a failure occurred during the check operation, leave
3743                  * STRIPE_INSYNC not set and let the stripe be handled again
3744                  */
3745                 if (s->failed)
3746                         break;
3747
3748                 /* handle a successful check operation, if parity is correct
3749                  * we are done.  Otherwise update the mismatch count and repair
3750                  * parity if !MD_RECOVERY_CHECK
3751                  */
3752                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3753                         /* parity is correct (on disc,
3754                          * not in buffer any more)
3755                          */
3756                         set_bit(STRIPE_INSYNC, &sh->state);
3757                 else {
3758                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3759                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3760                                 /* don't try to repair!! */
3761                                 set_bit(STRIPE_INSYNC, &sh->state);
3762                         else {
3763                                 sh->check_state = check_state_compute_run;
3764                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3765                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3766                                 set_bit(R5_Wantcompute,
3767                                         &sh->dev[sh->pd_idx].flags);
3768                                 sh->ops.target = sh->pd_idx;
3769                                 sh->ops.target2 = -1;
3770                                 s->uptodate++;
3771                         }
3772                 }
3773                 break;
3774         case check_state_compute_run:
3775                 break;
3776         default:
3777                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3778                        __func__, sh->check_state,
3779                        (unsigned long long) sh->sector);
3780                 BUG();
3781         }
3782 }
3783
3784 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3785                                   struct stripe_head_state *s,
3786                                   int disks)
3787 {
3788         int pd_idx = sh->pd_idx;
3789         int qd_idx = sh->qd_idx;
3790         struct r5dev *dev;
3791
3792         BUG_ON(sh->batch_head);
3793         set_bit(STRIPE_HANDLE, &sh->state);
3794
3795         BUG_ON(s->failed > 2);
3796
3797         /* Want to check and possibly repair P and Q.
3798          * However there could be one 'failed' device, in which
3799          * case we can only check one of them, possibly using the
3800          * other to generate missing data
3801          */
3802
3803         switch (sh->check_state) {
3804         case check_state_idle:
3805                 /* start a new check operation if there are < 2 failures */
3806                 if (s->failed == s->q_failed) {
3807                         /* The only possible failed device holds Q, so it
3808                          * makes sense to check P (If anything else were failed,
3809                          * we would have used P to recreate it).
3810                          */
3811                         sh->check_state = check_state_run;
3812                 }
3813                 if (!s->q_failed && s->failed < 2) {
3814                         /* Q is not failed, and we didn't use it to generate
3815                          * anything, so it makes sense to check it
3816                          */
3817                         if (sh->check_state == check_state_run)
3818                                 sh->check_state = check_state_run_pq;
3819                         else
3820                                 sh->check_state = check_state_run_q;
3821                 }
3822
3823                 /* discard potentially stale zero_sum_result */
3824                 sh->ops.zero_sum_result = 0;
3825
3826                 if (sh->check_state == check_state_run) {
3827                         /* async_xor_zero_sum destroys the contents of P */
3828                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3829                         s->uptodate--;
3830                 }
3831                 if (sh->check_state >= check_state_run &&
3832                     sh->check_state <= check_state_run_pq) {
3833                         /* async_syndrome_zero_sum preserves P and Q, so
3834                          * no need to mark them !uptodate here
3835                          */
3836                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3837                         break;
3838                 }
3839
3840                 /* we have 2-disk failure */
3841                 BUG_ON(s->failed != 2);
3842                 /* fall through */
3843         case check_state_compute_result:
3844                 sh->check_state = check_state_idle;
3845
3846                 /* check that a write has not made the stripe insync */
3847                 if (test_bit(STRIPE_INSYNC, &sh->state))
3848                         break;
3849
3850                 /* now write out any block on a failed drive,
3851                  * or P or Q if they were recomputed
3852                  */
3853                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3854                 if (s->failed == 2) {
3855                         dev = &sh->dev[s->failed_num[1]];
3856                         s->locked++;
3857                         set_bit(R5_LOCKED, &dev->flags);
3858                         set_bit(R5_Wantwrite, &dev->flags);
3859                 }
3860                 if (s->failed >= 1) {
3861                         dev = &sh->dev[s->failed_num[0]];
3862                         s->locked++;
3863                         set_bit(R5_LOCKED, &dev->flags);
3864                         set_bit(R5_Wantwrite, &dev->flags);
3865                 }
3866                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3867                         dev = &sh->dev[pd_idx];
3868                         s->locked++;
3869                         set_bit(R5_LOCKED, &dev->flags);
3870                         set_bit(R5_Wantwrite, &dev->flags);
3871                 }
3872                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3873                         dev = &sh->dev[qd_idx];
3874                         s->locked++;
3875                         set_bit(R5_LOCKED, &dev->flags);
3876                         set_bit(R5_Wantwrite, &dev->flags);
3877                 }
3878                 clear_bit(STRIPE_DEGRADED, &sh->state);
3879
3880                 set_bit(STRIPE_INSYNC, &sh->state);
3881                 break;
3882         case check_state_run:
3883         case check_state_run_q:
3884         case check_state_run_pq:
3885                 break; /* we will be called again upon completion */
3886         case check_state_check_result:
3887                 sh->check_state = check_state_idle;
3888
3889                 /* handle a successful check operation, if parity is correct
3890                  * we are done.  Otherwise update the mismatch count and repair
3891                  * parity if !MD_RECOVERY_CHECK
3892                  */
3893                 if (sh->ops.zero_sum_result == 0) {
3894                         /* both parities are correct */
3895                         if (!s->failed)
3896                                 set_bit(STRIPE_INSYNC, &sh->state);
3897                         else {
3898                                 /* in contrast to the raid5 case we can validate
3899                                  * parity, but still have a failure to write
3900                                  * back
3901                                  */
3902                                 sh->check_state = check_state_compute_result;
3903                                 /* Returning at this point means that we may go
3904                                  * off and bring p and/or q uptodate again so
3905                                  * we make sure to check zero_sum_result again
3906                                  * to verify if p or q need writeback
3907                                  */
3908                         }
3909                 } else {
3910                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3911                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3912                                 /* don't try to repair!! */
3913                                 set_bit(STRIPE_INSYNC, &sh->state);
3914                         else {
3915                                 int *target = &sh->ops.target;
3916
3917                                 sh->ops.target = -1;
3918                                 sh->ops.target2 = -1;
3919                                 sh->check_state = check_state_compute_run;
3920                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3921                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3922                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3923                                         set_bit(R5_Wantcompute,
3924                                                 &sh->dev[pd_idx].flags);
3925                                         *target = pd_idx;
3926                                         target = &sh->ops.target2;
3927                                         s->uptodate++;
3928                                 }
3929                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3930                                         set_bit(R5_Wantcompute,
3931                                                 &sh->dev[qd_idx].flags);
3932                                         *target = qd_idx;
3933                                         s->uptodate++;
3934                                 }
3935                         }
3936                 }
3937                 break;
3938         case check_state_compute_run:
3939                 break;
3940         default:
3941                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3942                        __func__, sh->check_state,
3943                        (unsigned long long) sh->sector);
3944                 BUG();
3945         }
3946 }
3947
3948 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3949 {
3950         int i;
3951
3952         /* We have read all the blocks in this stripe and now we need to
3953          * copy some of them into a target stripe for expand.
3954          */
3955         struct dma_async_tx_descriptor *tx = NULL;
3956         BUG_ON(sh->batch_head);
3957         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3958         for (i = 0; i < sh->disks; i++)
3959                 if (i != sh->pd_idx && i != sh->qd_idx) {
3960                         int dd_idx, j;
3961                         struct stripe_head *sh2;
3962                         struct async_submit_ctl submit;
3963
3964                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
3965                         sector_t s = raid5_compute_sector(conf, bn, 0,
3966                                                           &dd_idx, NULL);
3967                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
3968                         if (sh2 == NULL)
3969                                 /* so far only the early blocks of this stripe
3970                                  * have been requested.  When later blocks
3971                                  * get requested, we will try again
3972                                  */
3973                                 continue;
3974                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3975                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3976                                 /* must have already done this block */
3977                                 raid5_release_stripe(sh2);
3978                                 continue;
3979                         }
3980
3981                         /* place all the copies on one channel */
3982                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3983                         tx = async_memcpy(sh2->dev[dd_idx].page,
3984                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
3985                                           &submit);
3986
3987                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3988                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3989                         for (j = 0; j < conf->raid_disks; j++)
3990                                 if (j != sh2->pd_idx &&
3991                                     j != sh2->qd_idx &&
3992                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
3993                                         break;
3994                         if (j == conf->raid_disks) {
3995                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3996                                 set_bit(STRIPE_HANDLE, &sh2->state);
3997                         }
3998                         raid5_release_stripe(sh2);
3999
4000                 }
4001         /* done submitting copies, wait for them to complete */
4002         async_tx_quiesce(&tx);
4003 }
4004
4005 /*
4006  * handle_stripe - do things to a stripe.
4007  *
4008  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4009  * state of various bits to see what needs to be done.
4010  * Possible results:
4011  *    return some read requests which now have data
4012  *    return some write requests which are safely on storage
4013  *    schedule a read on some buffers
4014  *    schedule a write of some buffers
4015  *    return confirmation of parity correctness
4016  *
4017  */
4018
4019 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4020 {
4021         struct r5conf *conf = sh->raid_conf;
4022         int disks = sh->disks;
4023         struct r5dev *dev;
4024         int i;
4025         int do_recovery = 0;
4026
4027         memset(s, 0, sizeof(*s));
4028
4029         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4030         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4031         s->failed_num[0] = -1;
4032         s->failed_num[1] = -1;
4033         s->log_failed = r5l_log_disk_error(conf);
4034
4035         /* Now to look around and see what can be done */
4036         rcu_read_lock();
4037         for (i=disks; i--; ) {
4038                 struct md_rdev *rdev;
4039                 sector_t first_bad;
4040                 int bad_sectors;
4041                 int is_bad = 0;
4042
4043                 dev = &sh->dev[i];
4044
4045                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4046                          i, dev->flags,
4047                          dev->toread, dev->towrite, dev->written);
4048                 /* maybe we can reply to a read
4049                  *
4050                  * new wantfill requests are only permitted while
4051                  * ops_complete_biofill is guaranteed to be inactive
4052                  */
4053                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4054                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4055                         set_bit(R5_Wantfill, &dev->flags);
4056
4057                 /* now count some things */
4058                 if (test_bit(R5_LOCKED, &dev->flags))
4059                         s->locked++;
4060                 if (test_bit(R5_UPTODATE, &dev->flags))
4061                         s->uptodate++;
4062                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4063                         s->compute++;
4064                         BUG_ON(s->compute > 2);
4065                 }
4066
4067                 if (test_bit(R5_Wantfill, &dev->flags))
4068                         s->to_fill++;
4069                 else if (dev->toread)
4070                         s->to_read++;
4071                 if (dev->towrite) {
4072                         s->to_write++;
4073                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4074                                 s->non_overwrite++;
4075                 }
4076                 if (dev->written)
4077                         s->written++;
4078                 /* Prefer to use the replacement for reads, but only
4079                  * if it is recovered enough and has no bad blocks.
4080                  */
4081                 rdev = rcu_dereference(conf->disks[i].replacement);
4082                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4083                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4084                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4085                                  &first_bad, &bad_sectors))
4086                         set_bit(R5_ReadRepl, &dev->flags);
4087                 else {
4088                         if (rdev && !test_bit(Faulty, &rdev->flags))
4089                                 set_bit(R5_NeedReplace, &dev->flags);
4090                         else
4091                                 clear_bit(R5_NeedReplace, &dev->flags);
4092                         rdev = rcu_dereference(conf->disks[i].rdev);
4093                         clear_bit(R5_ReadRepl, &dev->flags);
4094                 }
4095                 if (rdev && test_bit(Faulty, &rdev->flags))
4096                         rdev = NULL;
4097                 if (rdev) {
4098                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4099                                              &first_bad, &bad_sectors);
4100                         if (s->blocked_rdev == NULL
4101                             && (test_bit(Blocked, &rdev->flags)
4102                                 || is_bad < 0)) {
4103                                 if (is_bad < 0)
4104                                         set_bit(BlockedBadBlocks,
4105                                                 &rdev->flags);
4106                                 s->blocked_rdev = rdev;
4107                                 atomic_inc(&rdev->nr_pending);
4108                         }
4109                 }
4110                 clear_bit(R5_Insync, &dev->flags);
4111                 if (!rdev)
4112                         /* Not in-sync */;
4113                 else if (is_bad) {
4114                         /* also not in-sync */
4115                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4116                             test_bit(R5_UPTODATE, &dev->flags)) {
4117                                 /* treat as in-sync, but with a read error
4118                                  * which we can now try to correct
4119                                  */
4120                                 set_bit(R5_Insync, &dev->flags);
4121                                 set_bit(R5_ReadError, &dev->flags);
4122                         }
4123                 } else if (test_bit(In_sync, &rdev->flags))
4124                         set_bit(R5_Insync, &dev->flags);
4125                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4126                         /* in sync if before recovery_offset */
4127                         set_bit(R5_Insync, &dev->flags);
4128                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4129                          test_bit(R5_Expanded, &dev->flags))
4130                         /* If we've reshaped into here, we assume it is Insync.
4131                          * We will shortly update recovery_offset to make
4132                          * it official.
4133                          */
4134                         set_bit(R5_Insync, &dev->flags);
4135
4136                 if (test_bit(R5_WriteError, &dev->flags)) {
4137                         /* This flag does not apply to '.replacement'
4138                          * only to .rdev, so make sure to check that*/
4139                         struct md_rdev *rdev2 = rcu_dereference(
4140                                 conf->disks[i].rdev);
4141                         if (rdev2 == rdev)
4142                                 clear_bit(R5_Insync, &dev->flags);
4143                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4144                                 s->handle_bad_blocks = 1;
4145                                 atomic_inc(&rdev2->nr_pending);
4146                         } else
4147                                 clear_bit(R5_WriteError, &dev->flags);
4148                 }
4149                 if (test_bit(R5_MadeGood, &dev->flags)) {
4150                         /* This flag does not apply to '.replacement'
4151                          * only to .rdev, so make sure to check that*/
4152                         struct md_rdev *rdev2 = rcu_dereference(
4153                                 conf->disks[i].rdev);
4154                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4155                                 s->handle_bad_blocks = 1;
4156                                 atomic_inc(&rdev2->nr_pending);
4157                         } else
4158                                 clear_bit(R5_MadeGood, &dev->flags);
4159                 }
4160                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4161                         struct md_rdev *rdev2 = rcu_dereference(
4162                                 conf->disks[i].replacement);
4163                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4164                                 s->handle_bad_blocks = 1;
4165                                 atomic_inc(&rdev2->nr_pending);
4166                         } else
4167                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4168                 }
4169                 if (!test_bit(R5_Insync, &dev->flags)) {
4170                         /* The ReadError flag will just be confusing now */
4171                         clear_bit(R5_ReadError, &dev->flags);
4172                         clear_bit(R5_ReWrite, &dev->flags);
4173                 }
4174                 if (test_bit(R5_ReadError, &dev->flags))
4175                         clear_bit(R5_Insync, &dev->flags);
4176                 if (!test_bit(R5_Insync, &dev->flags)) {
4177                         if (s->failed < 2)
4178                                 s->failed_num[s->failed] = i;
4179                         s->failed++;
4180                         if (rdev && !test_bit(Faulty, &rdev->flags))
4181                                 do_recovery = 1;
4182                 }
4183         }
4184         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4185                 /* If there is a failed device being replaced,
4186                  *     we must be recovering.
4187                  * else if we are after recovery_cp, we must be syncing
4188                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4189                  * else we can only be replacing
4190                  * sync and recovery both need to read all devices, and so
4191                  * use the same flag.
4192                  */
4193                 if (do_recovery ||
4194                     sh->sector >= conf->mddev->recovery_cp ||
4195                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4196                         s->syncing = 1;
4197                 else
4198                         s->replacing = 1;
4199         }
4200         rcu_read_unlock();
4201 }
4202
4203 static int clear_batch_ready(struct stripe_head *sh)
4204 {
4205         /* Return '1' if this is a member of batch, or
4206          * '0' if it is a lone stripe or a head which can now be
4207          * handled.
4208          */
4209         struct stripe_head *tmp;
4210         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4211                 return (sh->batch_head && sh->batch_head != sh);
4212         spin_lock(&sh->stripe_lock);
4213         if (!sh->batch_head) {
4214                 spin_unlock(&sh->stripe_lock);
4215                 return 0;
4216         }
4217
4218         /*
4219          * this stripe could be added to a batch list before we check
4220          * BATCH_READY, skips it
4221          */
4222         if (sh->batch_head != sh) {
4223                 spin_unlock(&sh->stripe_lock);
4224                 return 1;
4225         }
4226         spin_lock(&sh->batch_lock);
4227         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4228                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4229         spin_unlock(&sh->batch_lock);
4230         spin_unlock(&sh->stripe_lock);
4231
4232         /*
4233          * BATCH_READY is cleared, no new stripes can be added.
4234          * batch_list can be accessed without lock
4235          */
4236         return 0;
4237 }
4238
4239 static void break_stripe_batch_list(struct stripe_head *head_sh,
4240                                     unsigned long handle_flags)
4241 {
4242         struct stripe_head *sh, *next;
4243         int i;
4244         int do_wakeup = 0;
4245
4246         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4247
4248                 list_del_init(&sh->batch_list);
4249
4250                 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4251                                           (1 << STRIPE_SYNCING) |
4252                                           (1 << STRIPE_REPLACED) |
4253                                           (1 << STRIPE_PREREAD_ACTIVE) |
4254                                           (1 << STRIPE_DELAYED) |
4255                                           (1 << STRIPE_BIT_DELAY) |
4256                                           (1 << STRIPE_FULL_WRITE) |
4257                                           (1 << STRIPE_BIOFILL_RUN) |
4258                                           (1 << STRIPE_COMPUTE_RUN)  |
4259                                           (1 << STRIPE_OPS_REQ_PENDING) |
4260                                           (1 << STRIPE_DISCARD) |
4261                                           (1 << STRIPE_BATCH_READY) |
4262                                           (1 << STRIPE_BATCH_ERR) |
4263                                           (1 << STRIPE_BITMAP_PENDING)));
4264                 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4265                                               (1 << STRIPE_REPLACED)));
4266
4267                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4268                                             (1 << STRIPE_DEGRADED)),
4269                               head_sh->state & (1 << STRIPE_INSYNC));
4270
4271                 sh->check_state = head_sh->check_state;
4272                 sh->reconstruct_state = head_sh->reconstruct_state;
4273                 for (i = 0; i < sh->disks; i++) {
4274                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4275                                 do_wakeup = 1;
4276                         sh->dev[i].flags = head_sh->dev[i].flags &
4277                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4278                 }
4279                 spin_lock_irq(&sh->stripe_lock);
4280                 sh->batch_head = NULL;
4281                 spin_unlock_irq(&sh->stripe_lock);
4282                 if (handle_flags == 0 ||
4283                     sh->state & handle_flags)
4284                         set_bit(STRIPE_HANDLE, &sh->state);
4285                 raid5_release_stripe(sh);
4286         }
4287         spin_lock_irq(&head_sh->stripe_lock);
4288         head_sh->batch_head = NULL;
4289         spin_unlock_irq(&head_sh->stripe_lock);
4290         for (i = 0; i < head_sh->disks; i++)
4291                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4292                         do_wakeup = 1;
4293         if (head_sh->state & handle_flags)
4294                 set_bit(STRIPE_HANDLE, &head_sh->state);
4295
4296         if (do_wakeup)
4297                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4298 }
4299
4300 static void handle_stripe(struct stripe_head *sh)
4301 {
4302         struct stripe_head_state s;
4303         struct r5conf *conf = sh->raid_conf;
4304         int i;
4305         int prexor;
4306         int disks = sh->disks;
4307         struct r5dev *pdev, *qdev;
4308
4309         clear_bit(STRIPE_HANDLE, &sh->state);
4310         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4311                 /* already being handled, ensure it gets handled
4312                  * again when current action finishes */
4313                 set_bit(STRIPE_HANDLE, &sh->state);
4314                 return;
4315         }
4316
4317         if (clear_batch_ready(sh) ) {
4318                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4319                 return;
4320         }
4321
4322         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4323                 break_stripe_batch_list(sh, 0);
4324
4325         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4326                 spin_lock(&sh->stripe_lock);
4327                 /* Cannot process 'sync' concurrently with 'discard' */
4328                 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4329                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4330                         set_bit(STRIPE_SYNCING, &sh->state);
4331                         clear_bit(STRIPE_INSYNC, &sh->state);
4332                         clear_bit(STRIPE_REPLACED, &sh->state);
4333                 }
4334                 spin_unlock(&sh->stripe_lock);
4335         }
4336         clear_bit(STRIPE_DELAYED, &sh->state);
4337
4338         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4339                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4340                (unsigned long long)sh->sector, sh->state,
4341                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4342                sh->check_state, sh->reconstruct_state);
4343
4344         analyse_stripe(sh, &s);
4345
4346         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4347                 goto finish;
4348
4349         if (s.handle_bad_blocks) {
4350                 set_bit(STRIPE_HANDLE, &sh->state);
4351                 goto finish;
4352         }
4353
4354         if (unlikely(s.blocked_rdev)) {
4355                 if (s.syncing || s.expanding || s.expanded ||
4356                     s.replacing || s.to_write || s.written) {
4357                         set_bit(STRIPE_HANDLE, &sh->state);
4358                         goto finish;
4359                 }
4360                 /* There is nothing for the blocked_rdev to block */
4361                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4362                 s.blocked_rdev = NULL;
4363         }
4364
4365         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4366                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4367                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4368         }
4369
4370         pr_debug("locked=%d uptodate=%d to_read=%d"
4371                " to_write=%d failed=%d failed_num=%d,%d\n",
4372                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4373                s.failed_num[0], s.failed_num[1]);
4374         /* check if the array has lost more than max_degraded devices and,
4375          * if so, some requests might need to be failed.
4376          */
4377         if (s.failed > conf->max_degraded || s.log_failed) {
4378                 sh->check_state = 0;
4379                 sh->reconstruct_state = 0;
4380                 break_stripe_batch_list(sh, 0);
4381                 if (s.to_read+s.to_write+s.written)
4382                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4383                 if (s.syncing + s.replacing)
4384                         handle_failed_sync(conf, sh, &s);
4385         }
4386
4387         /* Now we check to see if any write operations have recently
4388          * completed
4389          */
4390         prexor = 0;
4391         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4392                 prexor = 1;
4393         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4394             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4395                 sh->reconstruct_state = reconstruct_state_idle;
4396
4397                 /* All the 'written' buffers and the parity block are ready to
4398                  * be written back to disk
4399                  */
4400                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4401                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4402                 BUG_ON(sh->qd_idx >= 0 &&
4403                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4404                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4405                 for (i = disks; i--; ) {
4406                         struct r5dev *dev = &sh->dev[i];
4407                         if (test_bit(R5_LOCKED, &dev->flags) &&
4408                                 (i == sh->pd_idx || i == sh->qd_idx ||
4409                                  dev->written)) {
4410                                 pr_debug("Writing block %d\n", i);
4411                                 set_bit(R5_Wantwrite, &dev->flags);
4412                                 if (prexor)
4413                                         continue;
4414                                 if (s.failed > 1)
4415                                         continue;
4416                                 if (!test_bit(R5_Insync, &dev->flags) ||
4417                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4418                                      s.failed == 0))
4419                                         set_bit(STRIPE_INSYNC, &sh->state);
4420                         }
4421                 }
4422                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4423                         s.dec_preread_active = 1;
4424         }
4425
4426         /*
4427          * might be able to return some write requests if the parity blocks
4428          * are safe, or on a failed drive
4429          */
4430         pdev = &sh->dev[sh->pd_idx];
4431         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4432                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4433         qdev = &sh->dev[sh->qd_idx];
4434         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4435                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4436                 || conf->level < 6;
4437
4438         if (s.written &&
4439             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4440                              && !test_bit(R5_LOCKED, &pdev->flags)
4441                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4442                                  test_bit(R5_Discard, &pdev->flags))))) &&
4443             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4444                              && !test_bit(R5_LOCKED, &qdev->flags)
4445                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4446                                  test_bit(R5_Discard, &qdev->flags))))))
4447                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4448
4449         /* Now we might consider reading some blocks, either to check/generate
4450          * parity, or to satisfy requests
4451          * or to load a block that is being partially written.
4452          */
4453         if (s.to_read || s.non_overwrite
4454             || (conf->level == 6 && s.to_write && s.failed)
4455             || (s.syncing && (s.uptodate + s.compute < disks))
4456             || s.replacing
4457             || s.expanding)
4458                 handle_stripe_fill(sh, &s, disks);
4459
4460         /* Now to consider new write requests and what else, if anything
4461          * should be read.  We do not handle new writes when:
4462          * 1/ A 'write' operation (copy+xor) is already in flight.
4463          * 2/ A 'check' operation is in flight, as it may clobber the parity
4464          *    block.
4465          */
4466         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4467                 handle_stripe_dirtying(conf, sh, &s, disks);
4468
4469         /* maybe we need to check and possibly fix the parity for this stripe
4470          * Any reads will already have been scheduled, so we just see if enough
4471          * data is available.  The parity check is held off while parity
4472          * dependent operations are in flight.
4473          */
4474         if (sh->check_state ||
4475             (s.syncing && s.locked == 0 &&
4476              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4477              !test_bit(STRIPE_INSYNC, &sh->state))) {
4478                 if (conf->level == 6)
4479                         handle_parity_checks6(conf, sh, &s, disks);
4480                 else
4481                         handle_parity_checks5(conf, sh, &s, disks);
4482         }
4483
4484         if ((s.replacing || s.syncing) && s.locked == 0
4485             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4486             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4487                 /* Write out to replacement devices where possible */
4488                 for (i = 0; i < conf->raid_disks; i++)
4489                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4490                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4491                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4492                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4493                                 s.locked++;
4494                         }
4495                 if (s.replacing)
4496                         set_bit(STRIPE_INSYNC, &sh->state);
4497                 set_bit(STRIPE_REPLACED, &sh->state);
4498         }
4499         if ((s.syncing || s.replacing) && s.locked == 0 &&
4500             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4501             test_bit(STRIPE_INSYNC, &sh->state)) {
4502                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4503                 clear_bit(STRIPE_SYNCING, &sh->state);
4504                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4505                         wake_up(&conf->wait_for_overlap);
4506         }
4507
4508         /* If the failed drives are just a ReadError, then we might need
4509          * to progress the repair/check process
4510          */
4511         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4512                 for (i = 0; i < s.failed; i++) {
4513                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4514                         if (test_bit(R5_ReadError, &dev->flags)
4515                             && !test_bit(R5_LOCKED, &dev->flags)
4516                             && test_bit(R5_UPTODATE, &dev->flags)
4517                                 ) {
4518                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4519                                         set_bit(R5_Wantwrite, &dev->flags);
4520                                         set_bit(R5_ReWrite, &dev->flags);
4521                                         set_bit(R5_LOCKED, &dev->flags);
4522                                         s.locked++;
4523                                 } else {
4524                                         /* let's read it back */
4525                                         set_bit(R5_Wantread, &dev->flags);
4526                                         set_bit(R5_LOCKED, &dev->flags);
4527                                         s.locked++;
4528                                 }
4529                         }
4530                 }
4531
4532         /* Finish reconstruct operations initiated by the expansion process */
4533         if (sh->reconstruct_state == reconstruct_state_result) {
4534                 struct stripe_head *sh_src
4535                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4536                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4537                         /* sh cannot be written until sh_src has been read.
4538                          * so arrange for sh to be delayed a little
4539                          */
4540                         set_bit(STRIPE_DELAYED, &sh->state);
4541                         set_bit(STRIPE_HANDLE, &sh->state);
4542                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4543                                               &sh_src->state))
4544                                 atomic_inc(&conf->preread_active_stripes);
4545                         raid5_release_stripe(sh_src);
4546                         goto finish;
4547                 }
4548                 if (sh_src)
4549                         raid5_release_stripe(sh_src);
4550
4551                 sh->reconstruct_state = reconstruct_state_idle;
4552                 clear_bit(STRIPE_EXPANDING, &sh->state);
4553                 for (i = conf->raid_disks; i--; ) {
4554                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4555                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4556                         s.locked++;
4557                 }
4558         }
4559
4560         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4561             !sh->reconstruct_state) {
4562                 /* Need to write out all blocks after computing parity */
4563                 sh->disks = conf->raid_disks;
4564                 stripe_set_idx(sh->sector, conf, 0, sh);
4565                 schedule_reconstruction(sh, &s, 1, 1);
4566         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4567                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4568                 atomic_dec(&conf->reshape_stripes);
4569                 wake_up(&conf->wait_for_overlap);
4570                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4571         }
4572
4573         if (s.expanding && s.locked == 0 &&
4574             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4575                 handle_stripe_expansion(conf, sh);
4576
4577 finish:
4578         /* wait for this device to become unblocked */
4579         if (unlikely(s.blocked_rdev)) {
4580                 if (conf->mddev->external)
4581                         md_wait_for_blocked_rdev(s.blocked_rdev,
4582                                                  conf->mddev);
4583                 else
4584                         /* Internal metadata will immediately
4585                          * be written by raid5d, so we don't
4586                          * need to wait here.
4587                          */
4588                         rdev_dec_pending(s.blocked_rdev,
4589                                          conf->mddev);
4590         }
4591
4592         if (s.handle_bad_blocks)
4593                 for (i = disks; i--; ) {
4594                         struct md_rdev *rdev;
4595                         struct r5dev *dev = &sh->dev[i];
4596                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4597                                 /* We own a safe reference to the rdev */
4598                                 rdev = conf->disks[i].rdev;
4599                                 if (!rdev_set_badblocks(rdev, sh->sector,
4600                                                         STRIPE_SECTORS, 0))
4601                                         md_error(conf->mddev, rdev);
4602                                 rdev_dec_pending(rdev, conf->mddev);
4603                         }
4604                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4605                                 rdev = conf->disks[i].rdev;
4606                                 rdev_clear_badblocks(rdev, sh->sector,
4607                                                      STRIPE_SECTORS, 0);
4608                                 rdev_dec_pending(rdev, conf->mddev);
4609                         }
4610                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4611                                 rdev = conf->disks[i].replacement;
4612                                 if (!rdev)
4613                                         /* rdev have been moved down */
4614                                         rdev = conf->disks[i].rdev;
4615                                 rdev_clear_badblocks(rdev, sh->sector,
4616                                                      STRIPE_SECTORS, 0);
4617                                 rdev_dec_pending(rdev, conf->mddev);
4618                         }
4619                 }
4620
4621         if (s.ops_request)
4622                 raid_run_ops(sh, s.ops_request);
4623
4624         ops_run_io(sh, &s);
4625
4626         if (s.dec_preread_active) {
4627                 /* We delay this until after ops_run_io so that if make_request
4628                  * is waiting on a flush, it won't continue until the writes
4629                  * have actually been submitted.
4630                  */
4631                 atomic_dec(&conf->preread_active_stripes);
4632                 if (atomic_read(&conf->preread_active_stripes) <
4633                     IO_THRESHOLD)
4634                         md_wakeup_thread(conf->mddev->thread);
4635         }
4636
4637         if (!bio_list_empty(&s.return_bi)) {
4638                 if (test_bit(MD_CHANGE_PENDING, &conf->mddev->flags)) {
4639                         spin_lock_irq(&conf->device_lock);
4640                         bio_list_merge(&conf->return_bi, &s.return_bi);
4641                         spin_unlock_irq(&conf->device_lock);
4642                         md_wakeup_thread(conf->mddev->thread);
4643                 } else
4644                         return_io(&s.return_bi);
4645         }
4646
4647         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4648 }
4649
4650 static void raid5_activate_delayed(struct r5conf *conf)
4651 {
4652         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4653                 while (!list_empty(&conf->delayed_list)) {
4654                         struct list_head *l = conf->delayed_list.next;
4655                         struct stripe_head *sh;
4656                         sh = list_entry(l, struct stripe_head, lru);
4657                         list_del_init(l);
4658                         clear_bit(STRIPE_DELAYED, &sh->state);
4659                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4660                                 atomic_inc(&conf->preread_active_stripes);
4661                         list_add_tail(&sh->lru, &conf->hold_list);
4662                         raid5_wakeup_stripe_thread(sh);
4663                 }
4664         }
4665 }
4666
4667 static void activate_bit_delay(struct r5conf *conf,
4668         struct list_head *temp_inactive_list)
4669 {
4670         /* device_lock is held */
4671         struct list_head head;
4672         list_add(&head, &conf->bitmap_list);
4673         list_del_init(&conf->bitmap_list);
4674         while (!list_empty(&head)) {
4675                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4676                 int hash;
4677                 list_del_init(&sh->lru);
4678                 atomic_inc(&sh->count);
4679                 hash = sh->hash_lock_index;
4680                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4681         }
4682 }
4683
4684 static int raid5_congested(struct mddev *mddev, int bits)
4685 {
4686         struct r5conf *conf = mddev->private;
4687
4688         /* No difference between reads and writes.  Just check
4689          * how busy the stripe_cache is
4690          */
4691
4692         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4693                 return 1;
4694         if (conf->quiesce)
4695                 return 1;
4696         if (atomic_read(&conf->empty_inactive_list_nr))
4697                 return 1;
4698
4699         return 0;
4700 }
4701
4702 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4703 {
4704         struct r5conf *conf = mddev->private;
4705         sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4706         unsigned int chunk_sectors;
4707         unsigned int bio_sectors = bio_sectors(bio);
4708
4709         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
4710         return  chunk_sectors >=
4711                 ((sector & (chunk_sectors - 1)) + bio_sectors);
4712 }
4713
4714 /*
4715  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
4716  *  later sampled by raid5d.
4717  */
4718 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4719 {
4720         unsigned long flags;
4721
4722         spin_lock_irqsave(&conf->device_lock, flags);
4723
4724         bi->bi_next = conf->retry_read_aligned_list;
4725         conf->retry_read_aligned_list = bi;
4726
4727         spin_unlock_irqrestore(&conf->device_lock, flags);
4728         md_wakeup_thread(conf->mddev->thread);
4729 }
4730
4731 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4732 {
4733         struct bio *bi;
4734
4735         bi = conf->retry_read_aligned;
4736         if (bi) {
4737                 conf->retry_read_aligned = NULL;
4738                 return bi;
4739         }
4740         bi = conf->retry_read_aligned_list;
4741         if(bi) {
4742                 conf->retry_read_aligned_list = bi->bi_next;
4743                 bi->bi_next = NULL;
4744                 /*
4745                  * this sets the active strip count to 1 and the processed
4746                  * strip count to zero (upper 8 bits)
4747                  */
4748                 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4749         }
4750
4751         return bi;
4752 }
4753
4754 /*
4755  *  The "raid5_align_endio" should check if the read succeeded and if it
4756  *  did, call bio_endio on the original bio (having bio_put the new bio
4757  *  first).
4758  *  If the read failed..
4759  */
4760 static void raid5_align_endio(struct bio *bi)
4761 {
4762         struct bio* raid_bi  = bi->bi_private;
4763         struct mddev *mddev;
4764         struct r5conf *conf;
4765         struct md_rdev *rdev;
4766         int error = bi->bi_error;
4767
4768         bio_put(bi);
4769
4770         rdev = (void*)raid_bi->bi_next;
4771         raid_bi->bi_next = NULL;
4772         mddev = rdev->mddev;
4773         conf = mddev->private;
4774
4775         rdev_dec_pending(rdev, conf->mddev);
4776
4777         if (!error) {
4778                 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4779                                          raid_bi, 0);
4780                 bio_endio(raid_bi);
4781                 if (atomic_dec_and_test(&conf->active_aligned_reads))
4782                         wake_up(&conf->wait_for_quiescent);
4783                 return;
4784         }
4785
4786         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4787
4788         add_bio_to_retry(raid_bi, conf);
4789 }
4790
4791 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
4792 {
4793         struct r5conf *conf = mddev->private;
4794         int dd_idx;
4795         struct bio* align_bi;
4796         struct md_rdev *rdev;
4797         sector_t end_sector;
4798
4799         if (!in_chunk_boundary(mddev, raid_bio)) {
4800                 pr_debug("%s: non aligned\n", __func__);
4801                 return 0;
4802         }
4803         /*
4804          * use bio_clone_mddev to make a copy of the bio
4805          */
4806         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4807         if (!align_bi)
4808                 return 0;
4809         /*
4810          *   set bi_end_io to a new function, and set bi_private to the
4811          *     original bio.
4812          */
4813         align_bi->bi_end_io  = raid5_align_endio;
4814         align_bi->bi_private = raid_bio;
4815         /*
4816          *      compute position
4817          */
4818         align_bi->bi_iter.bi_sector =
4819                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4820                                      0, &dd_idx, NULL);
4821
4822         end_sector = bio_end_sector(align_bi);
4823         rcu_read_lock();
4824         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4825         if (!rdev || test_bit(Faulty, &rdev->flags) ||
4826             rdev->recovery_offset < end_sector) {
4827                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4828                 if (rdev &&
4829                     (test_bit(Faulty, &rdev->flags) ||
4830                     !(test_bit(In_sync, &rdev->flags) ||
4831                       rdev->recovery_offset >= end_sector)))
4832                         rdev = NULL;
4833         }
4834         if (rdev) {
4835                 sector_t first_bad;
4836                 int bad_sectors;
4837
4838                 atomic_inc(&rdev->nr_pending);
4839                 rcu_read_unlock();
4840                 raid_bio->bi_next = (void*)rdev;
4841                 align_bi->bi_bdev =  rdev->bdev;
4842                 bio_clear_flag(align_bi, BIO_SEG_VALID);
4843
4844                 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
4845                                 bio_sectors(align_bi),
4846                                 &first_bad, &bad_sectors)) {
4847                         bio_put(align_bi);
4848                         rdev_dec_pending(rdev, mddev);
4849                         return 0;
4850                 }
4851
4852                 /* No reshape active, so we can trust rdev->data_offset */
4853                 align_bi->bi_iter.bi_sector += rdev->data_offset;
4854
4855                 spin_lock_irq(&conf->device_lock);
4856                 wait_event_lock_irq(conf->wait_for_quiescent,
4857                                     conf->quiesce == 0,
4858                                     conf->device_lock);
4859                 atomic_inc(&conf->active_aligned_reads);
4860                 spin_unlock_irq(&conf->device_lock);
4861
4862                 if (mddev->gendisk)
4863                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4864                                               align_bi, disk_devt(mddev->gendisk),
4865                                               raid_bio->bi_iter.bi_sector);
4866                 generic_make_request(align_bi);
4867                 return 1;
4868         } else {
4869                 rcu_read_unlock();
4870                 bio_put(align_bi);
4871                 return 0;
4872         }
4873 }
4874
4875 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
4876 {
4877         struct bio *split;
4878
4879         do {
4880                 sector_t sector = raid_bio->bi_iter.bi_sector;
4881                 unsigned chunk_sects = mddev->chunk_sectors;
4882                 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
4883
4884                 if (sectors < bio_sectors(raid_bio)) {
4885                         split = bio_split(raid_bio, sectors, GFP_NOIO, fs_bio_set);
4886                         bio_chain(split, raid_bio);
4887                 } else
4888                         split = raid_bio;
4889
4890                 if (!raid5_read_one_chunk(mddev, split)) {
4891                         if (split != raid_bio)
4892                                 generic_make_request(raid_bio);
4893                         return split;
4894                 }
4895         } while (split != raid_bio);
4896
4897         return NULL;
4898 }
4899
4900 /* __get_priority_stripe - get the next stripe to process
4901  *
4902  * Full stripe writes are allowed to pass preread active stripes up until
4903  * the bypass_threshold is exceeded.  In general the bypass_count
4904  * increments when the handle_list is handled before the hold_list; however, it
4905  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4906  * stripe with in flight i/o.  The bypass_count will be reset when the
4907  * head of the hold_list has changed, i.e. the head was promoted to the
4908  * handle_list.
4909  */
4910 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4911 {
4912         struct stripe_head *sh = NULL, *tmp;
4913         struct list_head *handle_list = NULL;
4914         struct r5worker_group *wg = NULL;
4915
4916         if (conf->worker_cnt_per_group == 0) {
4917                 handle_list = &conf->handle_list;
4918         } else if (group != ANY_GROUP) {
4919                 handle_list = &conf->worker_groups[group].handle_list;
4920                 wg = &conf->worker_groups[group];
4921         } else {
4922                 int i;
4923                 for (i = 0; i < conf->group_cnt; i++) {
4924                         handle_list = &conf->worker_groups[i].handle_list;
4925                         wg = &conf->worker_groups[i];
4926                         if (!list_empty(handle_list))
4927                                 break;
4928                 }
4929         }
4930
4931         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4932                   __func__,
4933                   list_empty(handle_list) ? "empty" : "busy",
4934                   list_empty(&conf->hold_list) ? "empty" : "busy",
4935                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
4936
4937         if (!list_empty(handle_list)) {
4938                 sh = list_entry(handle_list->next, typeof(*sh), lru);
4939
4940                 if (list_empty(&conf->hold_list))
4941                         conf->bypass_count = 0;
4942                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4943                         if (conf->hold_list.next == conf->last_hold)
4944                                 conf->bypass_count++;
4945                         else {
4946                                 conf->last_hold = conf->hold_list.next;
4947                                 conf->bypass_count -= conf->bypass_threshold;
4948                                 if (conf->bypass_count < 0)
4949                                         conf->bypass_count = 0;
4950                         }
4951                 }
4952         } else if (!list_empty(&conf->hold_list) &&
4953                    ((conf->bypass_threshold &&
4954                      conf->bypass_count > conf->bypass_threshold) ||
4955                     atomic_read(&conf->pending_full_writes) == 0)) {
4956
4957                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
4958                         if (conf->worker_cnt_per_group == 0 ||
4959                             group == ANY_GROUP ||
4960                             !cpu_online(tmp->cpu) ||
4961                             cpu_to_group(tmp->cpu) == group) {
4962                                 sh = tmp;
4963                                 break;
4964                         }
4965                 }
4966
4967                 if (sh) {
4968                         conf->bypass_count -= conf->bypass_threshold;
4969                         if (conf->bypass_count < 0)
4970                                 conf->bypass_count = 0;
4971                 }
4972                 wg = NULL;
4973         }
4974
4975         if (!sh)
4976                 return NULL;
4977
4978         if (wg) {
4979                 wg->stripes_cnt--;
4980                 sh->group = NULL;
4981         }
4982         list_del_init(&sh->lru);
4983         BUG_ON(atomic_inc_return(&sh->count) != 1);
4984         return sh;
4985 }
4986
4987 struct raid5_plug_cb {
4988         struct blk_plug_cb      cb;
4989         struct list_head        list;
4990         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4991 };
4992
4993 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4994 {
4995         struct raid5_plug_cb *cb = container_of(
4996                 blk_cb, struct raid5_plug_cb, cb);
4997         struct stripe_head *sh;
4998         struct mddev *mddev = cb->cb.data;
4999         struct r5conf *conf = mddev->private;
5000         int cnt = 0;
5001         int hash;
5002
5003         if (cb->list.next && !list_empty(&cb->list)) {
5004                 spin_lock_irq(&conf->device_lock);
5005                 while (!list_empty(&cb->list)) {
5006                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5007                         list_del_init(&sh->lru);
5008                         /*
5009                          * avoid race release_stripe_plug() sees
5010                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5011                          * is still in our list
5012                          */
5013                         smp_mb__before_atomic();
5014                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5015                         /*
5016                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5017                          * case, the count is always > 1 here
5018                          */
5019                         hash = sh->hash_lock_index;
5020                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5021                         cnt++;
5022                 }
5023                 spin_unlock_irq(&conf->device_lock);
5024         }
5025         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5026                                      NR_STRIPE_HASH_LOCKS);
5027         if (mddev->queue)
5028                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5029         kfree(cb);
5030 }
5031
5032 static void release_stripe_plug(struct mddev *mddev,
5033                                 struct stripe_head *sh)
5034 {
5035         struct blk_plug_cb *blk_cb = blk_check_plugged(
5036                 raid5_unplug, mddev,
5037                 sizeof(struct raid5_plug_cb));
5038         struct raid5_plug_cb *cb;
5039
5040         if (!blk_cb) {
5041                 raid5_release_stripe(sh);
5042                 return;
5043         }
5044
5045         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5046
5047         if (cb->list.next == NULL) {
5048                 int i;
5049                 INIT_LIST_HEAD(&cb->list);
5050                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5051                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5052         }
5053
5054         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5055                 list_add_tail(&sh->lru, &cb->list);
5056         else
5057                 raid5_release_stripe(sh);
5058 }
5059
5060 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5061 {
5062         struct r5conf *conf = mddev->private;
5063         sector_t logical_sector, last_sector;
5064         struct stripe_head *sh;
5065         int remaining;
5066         int stripe_sectors;
5067
5068         if (mddev->reshape_position != MaxSector)
5069                 /* Skip discard while reshape is happening */
5070                 return;
5071
5072         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5073         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5074
5075         bi->bi_next = NULL;
5076         bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5077
5078         stripe_sectors = conf->chunk_sectors *
5079                 (conf->raid_disks - conf->max_degraded);
5080         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5081                                                stripe_sectors);
5082         sector_div(last_sector, stripe_sectors);
5083
5084         logical_sector *= conf->chunk_sectors;
5085         last_sector *= conf->chunk_sectors;
5086
5087         for (; logical_sector < last_sector;
5088              logical_sector += STRIPE_SECTORS) {
5089                 DEFINE_WAIT(w);
5090                 int d;
5091         again:
5092                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5093                 prepare_to_wait(&conf->wait_for_overlap, &w,
5094                                 TASK_UNINTERRUPTIBLE);
5095                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5096                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5097                         raid5_release_stripe(sh);
5098                         schedule();
5099                         goto again;
5100                 }
5101                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5102                 spin_lock_irq(&sh->stripe_lock);
5103                 for (d = 0; d < conf->raid_disks; d++) {
5104                         if (d == sh->pd_idx || d == sh->qd_idx)
5105                                 continue;
5106                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5107                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5108                                 spin_unlock_irq(&sh->stripe_lock);
5109                                 raid5_release_stripe(sh);
5110                                 schedule();
5111                                 goto again;
5112                         }
5113                 }
5114                 set_bit(STRIPE_DISCARD, &sh->state);
5115                 finish_wait(&conf->wait_for_overlap, &w);
5116                 sh->overwrite_disks = 0;
5117                 for (d = 0; d < conf->raid_disks; d++) {
5118                         if (d == sh->pd_idx || d == sh->qd_idx)
5119                                 continue;
5120                         sh->dev[d].towrite = bi;
5121                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5122                         raid5_inc_bi_active_stripes(bi);
5123                         sh->overwrite_disks++;
5124                 }
5125                 spin_unlock_irq(&sh->stripe_lock);
5126                 if (conf->mddev->bitmap) {
5127                         for (d = 0;
5128                              d < conf->raid_disks - conf->max_degraded;
5129                              d++)
5130                                 bitmap_startwrite(mddev->bitmap,
5131                                                   sh->sector,
5132                                                   STRIPE_SECTORS,
5133                                                   0);
5134                         sh->bm_seq = conf->seq_flush + 1;
5135                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5136                 }
5137
5138                 set_bit(STRIPE_HANDLE, &sh->state);
5139                 clear_bit(STRIPE_DELAYED, &sh->state);
5140                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5141                         atomic_inc(&conf->preread_active_stripes);
5142                 release_stripe_plug(mddev, sh);
5143         }
5144
5145         remaining = raid5_dec_bi_active_stripes(bi);
5146         if (remaining == 0) {
5147                 md_write_end(mddev);
5148                 bio_endio(bi);
5149         }
5150 }
5151
5152 static void make_request(struct mddev *mddev, struct bio * bi)
5153 {
5154         struct r5conf *conf = mddev->private;
5155         int dd_idx;
5156         sector_t new_sector;
5157         sector_t logical_sector, last_sector;
5158         struct stripe_head *sh;
5159         const int rw = bio_data_dir(bi);
5160         int remaining;
5161         DEFINE_WAIT(w);
5162         bool do_prepare;
5163
5164         if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5165                 int ret = r5l_handle_flush_request(conf->log, bi);
5166
5167                 if (ret == 0)
5168                         return;
5169                 if (ret == -ENODEV) {
5170                         md_flush_request(mddev, bi);
5171                         return;
5172                 }
5173                 /* ret == -EAGAIN, fallback */
5174         }
5175
5176         md_write_start(mddev, bi);
5177
5178         /*
5179          * If array is degraded, better not do chunk aligned read because
5180          * later we might have to read it again in order to reconstruct
5181          * data on failed drives.
5182          */
5183         if (rw == READ && mddev->degraded == 0 &&
5184             mddev->reshape_position == MaxSector) {
5185                 bi = chunk_aligned_read(mddev, bi);
5186                 if (!bi)
5187                         return;
5188         }
5189
5190         if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5191                 make_discard_request(mddev, bi);
5192                 return;
5193         }
5194
5195         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5196         last_sector = bio_end_sector(bi);
5197         bi->bi_next = NULL;
5198         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
5199
5200         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5201         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5202                 int previous;
5203                 int seq;
5204
5205                 do_prepare = false;
5206         retry:
5207                 seq = read_seqcount_begin(&conf->gen_lock);
5208                 previous = 0;
5209                 if (do_prepare)
5210                         prepare_to_wait(&conf->wait_for_overlap, &w,
5211                                 TASK_UNINTERRUPTIBLE);
5212                 if (unlikely(conf->reshape_progress != MaxSector)) {
5213                         /* spinlock is needed as reshape_progress may be
5214                          * 64bit on a 32bit platform, and so it might be
5215                          * possible to see a half-updated value
5216                          * Of course reshape_progress could change after
5217                          * the lock is dropped, so once we get a reference
5218                          * to the stripe that we think it is, we will have
5219                          * to check again.
5220                          */
5221                         spin_lock_irq(&conf->device_lock);
5222                         if (mddev->reshape_backwards
5223                             ? logical_sector < conf->reshape_progress
5224                             : logical_sector >= conf->reshape_progress) {
5225                                 previous = 1;
5226                         } else {
5227                                 if (mddev->reshape_backwards
5228                                     ? logical_sector < conf->reshape_safe
5229                                     : logical_sector >= conf->reshape_safe) {
5230                                         spin_unlock_irq(&conf->device_lock);
5231                                         schedule();
5232                                         do_prepare = true;
5233                                         goto retry;
5234                                 }
5235                         }
5236                         spin_unlock_irq(&conf->device_lock);
5237                 }
5238
5239                 new_sector = raid5_compute_sector(conf, logical_sector,
5240                                                   previous,
5241                                                   &dd_idx, NULL);
5242                 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5243                         (unsigned long long)new_sector,
5244                         (unsigned long long)logical_sector);
5245
5246                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5247                                        (bi->bi_rw&RWA_MASK), 0);
5248                 if (sh) {
5249                         if (unlikely(previous)) {
5250                                 /* expansion might have moved on while waiting for a
5251                                  * stripe, so we must do the range check again.
5252                                  * Expansion could still move past after this
5253                                  * test, but as we are holding a reference to
5254                                  * 'sh', we know that if that happens,
5255                                  *  STRIPE_EXPANDING will get set and the expansion
5256                                  * won't proceed until we finish with the stripe.
5257                                  */
5258                                 int must_retry = 0;
5259                                 spin_lock_irq(&conf->device_lock);
5260                                 if (mddev->reshape_backwards
5261                                     ? logical_sector >= conf->reshape_progress
5262                                     : logical_sector < conf->reshape_progress)
5263                                         /* mismatch, need to try again */
5264                                         must_retry = 1;
5265                                 spin_unlock_irq(&conf->device_lock);
5266                                 if (must_retry) {
5267                                         raid5_release_stripe(sh);
5268                                         schedule();
5269                                         do_prepare = true;
5270                                         goto retry;
5271                                 }
5272                         }
5273                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5274                                 /* Might have got the wrong stripe_head
5275                                  * by accident
5276                                  */
5277                                 raid5_release_stripe(sh);
5278                                 goto retry;
5279                         }
5280
5281                         if (rw == WRITE &&
5282                             logical_sector >= mddev->suspend_lo &&
5283                             logical_sector < mddev->suspend_hi) {
5284                                 raid5_release_stripe(sh);
5285                                 /* As the suspend_* range is controlled by
5286                                  * userspace, we want an interruptible
5287                                  * wait.
5288                                  */
5289                                 flush_signals(current);
5290                                 prepare_to_wait(&conf->wait_for_overlap,
5291                                                 &w, TASK_INTERRUPTIBLE);
5292                                 if (logical_sector >= mddev->suspend_lo &&
5293                                     logical_sector < mddev->suspend_hi) {
5294                                         schedule();
5295                                         do_prepare = true;
5296                                 }
5297                                 goto retry;
5298                         }
5299
5300                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5301                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5302                                 /* Stripe is busy expanding or
5303                                  * add failed due to overlap.  Flush everything
5304                                  * and wait a while
5305                                  */
5306                                 md_wakeup_thread(mddev->thread);
5307                                 raid5_release_stripe(sh);
5308                                 schedule();
5309                                 do_prepare = true;
5310                                 goto retry;
5311                         }
5312                         set_bit(STRIPE_HANDLE, &sh->state);
5313                         clear_bit(STRIPE_DELAYED, &sh->state);
5314                         if ((!sh->batch_head || sh == sh->batch_head) &&
5315                             (bi->bi_rw & REQ_SYNC) &&
5316                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5317                                 atomic_inc(&conf->preread_active_stripes);
5318                         release_stripe_plug(mddev, sh);
5319                 } else {
5320                         /* cannot get stripe for read-ahead, just give-up */
5321                         bi->bi_error = -EIO;
5322                         break;
5323                 }
5324         }
5325         finish_wait(&conf->wait_for_overlap, &w);
5326
5327         remaining = raid5_dec_bi_active_stripes(bi);
5328         if (remaining == 0) {
5329
5330                 if ( rw == WRITE )
5331                         md_write_end(mddev);
5332
5333                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5334                                          bi, 0);
5335                 bio_endio(bi);
5336         }
5337 }
5338
5339 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5340
5341 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5342 {
5343         /* reshaping is quite different to recovery/resync so it is
5344          * handled quite separately ... here.
5345          *
5346          * On each call to sync_request, we gather one chunk worth of
5347          * destination stripes and flag them as expanding.
5348          * Then we find all the source stripes and request reads.
5349          * As the reads complete, handle_stripe will copy the data
5350          * into the destination stripe and release that stripe.
5351          */
5352         struct r5conf *conf = mddev->private;
5353         struct stripe_head *sh;
5354         sector_t first_sector, last_sector;
5355         int raid_disks = conf->previous_raid_disks;
5356         int data_disks = raid_disks - conf->max_degraded;
5357         int new_data_disks = conf->raid_disks - conf->max_degraded;
5358         int i;
5359         int dd_idx;
5360         sector_t writepos, readpos, safepos;
5361         sector_t stripe_addr;
5362         int reshape_sectors;
5363         struct list_head stripes;
5364         sector_t retn;
5365
5366         if (sector_nr == 0) {
5367                 /* If restarting in the middle, skip the initial sectors */
5368                 if (mddev->reshape_backwards &&
5369                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5370                         sector_nr = raid5_size(mddev, 0, 0)
5371                                 - conf->reshape_progress;
5372                 } else if (mddev->reshape_backwards &&
5373                            conf->reshape_progress == MaxSector) {
5374                         /* shouldn't happen, but just in case, finish up.*/
5375                         sector_nr = MaxSector;
5376                 } else if (!mddev->reshape_backwards &&
5377                            conf->reshape_progress > 0)
5378                         sector_nr = conf->reshape_progress;
5379                 sector_div(sector_nr, new_data_disks);
5380                 if (sector_nr) {
5381                         mddev->curr_resync_completed = sector_nr;
5382                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5383                         *skipped = 1;
5384                         retn = sector_nr;
5385                         goto finish;
5386                 }
5387         }
5388
5389         /* We need to process a full chunk at a time.
5390          * If old and new chunk sizes differ, we need to process the
5391          * largest of these
5392          */
5393
5394         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5395
5396         /* We update the metadata at least every 10 seconds, or when
5397          * the data about to be copied would over-write the source of
5398          * the data at the front of the range.  i.e. one new_stripe
5399          * along from reshape_progress new_maps to after where
5400          * reshape_safe old_maps to
5401          */
5402         writepos = conf->reshape_progress;
5403         sector_div(writepos, new_data_disks);
5404         readpos = conf->reshape_progress;
5405         sector_div(readpos, data_disks);
5406         safepos = conf->reshape_safe;
5407         sector_div(safepos, data_disks);
5408         if (mddev->reshape_backwards) {
5409                 BUG_ON(writepos < reshape_sectors);
5410                 writepos -= reshape_sectors;
5411                 readpos += reshape_sectors;
5412                 safepos += reshape_sectors;
5413         } else {
5414                 writepos += reshape_sectors;
5415                 /* readpos and safepos are worst-case calculations.
5416                  * A negative number is overly pessimistic, and causes
5417                  * obvious problems for unsigned storage.  So clip to 0.
5418                  */
5419                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5420                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5421         }
5422
5423         /* Having calculated the 'writepos' possibly use it
5424          * to set 'stripe_addr' which is where we will write to.
5425          */
5426         if (mddev->reshape_backwards) {
5427                 BUG_ON(conf->reshape_progress == 0);
5428                 stripe_addr = writepos;
5429                 BUG_ON((mddev->dev_sectors &
5430                         ~((sector_t)reshape_sectors - 1))
5431                        - reshape_sectors - stripe_addr
5432                        != sector_nr);
5433         } else {
5434                 BUG_ON(writepos != sector_nr + reshape_sectors);
5435                 stripe_addr = sector_nr;
5436         }
5437
5438         /* 'writepos' is the most advanced device address we might write.
5439          * 'readpos' is the least advanced device address we might read.
5440          * 'safepos' is the least address recorded in the metadata as having
5441          *     been reshaped.
5442          * If there is a min_offset_diff, these are adjusted either by
5443          * increasing the safepos/readpos if diff is negative, or
5444          * increasing writepos if diff is positive.
5445          * If 'readpos' is then behind 'writepos', there is no way that we can
5446          * ensure safety in the face of a crash - that must be done by userspace
5447          * making a backup of the data.  So in that case there is no particular
5448          * rush to update metadata.
5449          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5450          * update the metadata to advance 'safepos' to match 'readpos' so that
5451          * we can be safe in the event of a crash.
5452          * So we insist on updating metadata if safepos is behind writepos and
5453          * readpos is beyond writepos.
5454          * In any case, update the metadata every 10 seconds.
5455          * Maybe that number should be configurable, but I'm not sure it is
5456          * worth it.... maybe it could be a multiple of safemode_delay???
5457          */
5458         if (conf->min_offset_diff < 0) {
5459                 safepos += -conf->min_offset_diff;
5460                 readpos += -conf->min_offset_diff;
5461         } else
5462                 writepos += conf->min_offset_diff;
5463
5464         if ((mddev->reshape_backwards
5465              ? (safepos > writepos && readpos < writepos)
5466              : (safepos < writepos && readpos > writepos)) ||
5467             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5468                 /* Cannot proceed until we've updated the superblock... */
5469                 wait_event(conf->wait_for_overlap,
5470                            atomic_read(&conf->reshape_stripes)==0
5471                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5472                 if (atomic_read(&conf->reshape_stripes) != 0)
5473                         return 0;
5474                 mddev->reshape_position = conf->reshape_progress;
5475                 mddev->curr_resync_completed = sector_nr;
5476                 conf->reshape_checkpoint = jiffies;
5477                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5478                 md_wakeup_thread(mddev->thread);
5479                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5480                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5481                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5482                         return 0;
5483                 spin_lock_irq(&conf->device_lock);
5484                 conf->reshape_safe = mddev->reshape_position;
5485                 spin_unlock_irq(&conf->device_lock);
5486                 wake_up(&conf->wait_for_overlap);
5487                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5488         }
5489
5490         INIT_LIST_HEAD(&stripes);
5491         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5492                 int j;
5493                 int skipped_disk = 0;
5494                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5495                 set_bit(STRIPE_EXPANDING, &sh->state);
5496                 atomic_inc(&conf->reshape_stripes);
5497                 /* If any of this stripe is beyond the end of the old
5498                  * array, then we need to zero those blocks
5499                  */
5500                 for (j=sh->disks; j--;) {
5501                         sector_t s;
5502                         if (j == sh->pd_idx)
5503                                 continue;
5504                         if (conf->level == 6 &&
5505                             j == sh->qd_idx)
5506                                 continue;
5507                         s = raid5_compute_blocknr(sh, j, 0);
5508                         if (s < raid5_size(mddev, 0, 0)) {
5509                                 skipped_disk = 1;
5510                                 continue;
5511                         }
5512                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5513                         set_bit(R5_Expanded, &sh->dev[j].flags);
5514                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5515                 }
5516                 if (!skipped_disk) {
5517                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5518                         set_bit(STRIPE_HANDLE, &sh->state);
5519                 }
5520                 list_add(&sh->lru, &stripes);
5521         }
5522         spin_lock_irq(&conf->device_lock);
5523         if (mddev->reshape_backwards)
5524                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5525         else
5526                 conf->reshape_progress += reshape_sectors * new_data_disks;
5527         spin_unlock_irq(&conf->device_lock);
5528         /* Ok, those stripe are ready. We can start scheduling
5529          * reads on the source stripes.
5530          * The source stripes are determined by mapping the first and last
5531          * block on the destination stripes.
5532          */
5533         first_sector =
5534                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5535                                      1, &dd_idx, NULL);
5536         last_sector =
5537                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5538                                             * new_data_disks - 1),
5539                                      1, &dd_idx, NULL);
5540         if (last_sector >= mddev->dev_sectors)
5541                 last_sector = mddev->dev_sectors - 1;
5542         while (first_sector <= last_sector) {
5543                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5544                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5545                 set_bit(STRIPE_HANDLE, &sh->state);
5546                 raid5_release_stripe(sh);
5547                 first_sector += STRIPE_SECTORS;
5548         }
5549         /* Now that the sources are clearly marked, we can release
5550          * the destination stripes
5551          */
5552         while (!list_empty(&stripes)) {
5553                 sh = list_entry(stripes.next, struct stripe_head, lru);
5554                 list_del_init(&sh->lru);
5555                 raid5_release_stripe(sh);
5556         }
5557         /* If this takes us to the resync_max point where we have to pause,
5558          * then we need to write out the superblock.
5559          */
5560         sector_nr += reshape_sectors;
5561         retn = reshape_sectors;
5562 finish:
5563         if (mddev->curr_resync_completed > mddev->resync_max ||
5564             (sector_nr - mddev->curr_resync_completed) * 2
5565             >= mddev->resync_max - mddev->curr_resync_completed) {
5566                 /* Cannot proceed until we've updated the superblock... */
5567                 wait_event(conf->wait_for_overlap,
5568                            atomic_read(&conf->reshape_stripes) == 0
5569                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5570                 if (atomic_read(&conf->reshape_stripes) != 0)
5571                         goto ret;
5572                 mddev->reshape_position = conf->reshape_progress;
5573                 mddev->curr_resync_completed = sector_nr;
5574                 conf->reshape_checkpoint = jiffies;
5575                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5576                 md_wakeup_thread(mddev->thread);
5577                 wait_event(mddev->sb_wait,
5578                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5579                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5580                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5581                         goto ret;
5582                 spin_lock_irq(&conf->device_lock);
5583                 conf->reshape_safe = mddev->reshape_position;
5584                 spin_unlock_irq(&conf->device_lock);
5585                 wake_up(&conf->wait_for_overlap);
5586                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5587         }
5588 ret:
5589         return retn;
5590 }
5591
5592 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5593 {
5594         struct r5conf *conf = mddev->private;
5595         struct stripe_head *sh;
5596         sector_t max_sector = mddev->dev_sectors;
5597         sector_t sync_blocks;
5598         int still_degraded = 0;
5599         int i;
5600
5601         if (sector_nr >= max_sector) {
5602                 /* just being told to finish up .. nothing much to do */
5603
5604                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5605                         end_reshape(conf);
5606                         return 0;
5607                 }
5608
5609                 if (mddev->curr_resync < max_sector) /* aborted */
5610                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5611                                         &sync_blocks, 1);
5612                 else /* completed sync */
5613                         conf->fullsync = 0;
5614                 bitmap_close_sync(mddev->bitmap);
5615
5616                 return 0;
5617         }
5618
5619         /* Allow raid5_quiesce to complete */
5620         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5621
5622         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5623                 return reshape_request(mddev, sector_nr, skipped);
5624
5625         /* No need to check resync_max as we never do more than one
5626          * stripe, and as resync_max will always be on a chunk boundary,
5627          * if the check in md_do_sync didn't fire, there is no chance
5628          * of overstepping resync_max here
5629          */
5630
5631         /* if there is too many failed drives and we are trying
5632          * to resync, then assert that we are finished, because there is
5633          * nothing we can do.
5634          */
5635         if (mddev->degraded >= conf->max_degraded &&
5636             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5637                 sector_t rv = mddev->dev_sectors - sector_nr;
5638                 *skipped = 1;
5639                 return rv;
5640         }
5641         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5642             !conf->fullsync &&
5643             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5644             sync_blocks >= STRIPE_SECTORS) {
5645                 /* we can skip this block, and probably more */
5646                 sync_blocks /= STRIPE_SECTORS;
5647                 *skipped = 1;
5648                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5649         }
5650
5651         bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
5652
5653         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
5654         if (sh == NULL) {
5655                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
5656                 /* make sure we don't swamp the stripe cache if someone else
5657                  * is trying to get access
5658                  */
5659                 schedule_timeout_uninterruptible(1);
5660         }
5661         /* Need to check if array will still be degraded after recovery/resync
5662          * Note in case of > 1 drive failures it's possible we're rebuilding
5663          * one drive while leaving another faulty drive in array.
5664          */
5665         rcu_read_lock();
5666         for (i = 0; i < conf->raid_disks; i++) {
5667                 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5668
5669                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5670                         still_degraded = 1;
5671         }
5672         rcu_read_unlock();
5673
5674         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5675
5676         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5677         set_bit(STRIPE_HANDLE, &sh->state);
5678
5679         raid5_release_stripe(sh);
5680
5681         return STRIPE_SECTORS;
5682 }
5683
5684 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5685 {
5686         /* We may not be able to submit a whole bio at once as there
5687          * may not be enough stripe_heads available.
5688          * We cannot pre-allocate enough stripe_heads as we may need
5689          * more than exist in the cache (if we allow ever large chunks).
5690          * So we do one stripe head at a time and record in
5691          * ->bi_hw_segments how many have been done.
5692          *
5693          * We *know* that this entire raid_bio is in one chunk, so
5694          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5695          */
5696         struct stripe_head *sh;
5697         int dd_idx;
5698         sector_t sector, logical_sector, last_sector;
5699         int scnt = 0;
5700         int remaining;
5701         int handled = 0;
5702
5703         logical_sector = raid_bio->bi_iter.bi_sector &
5704                 ~((sector_t)STRIPE_SECTORS-1);
5705         sector = raid5_compute_sector(conf, logical_sector,
5706                                       0, &dd_idx, NULL);
5707         last_sector = bio_end_sector(raid_bio);
5708
5709         for (; logical_sector < last_sector;
5710              logical_sector += STRIPE_SECTORS,
5711                      sector += STRIPE_SECTORS,
5712                      scnt++) {
5713
5714                 if (scnt < raid5_bi_processed_stripes(raid_bio))
5715                         /* already done this stripe */
5716                         continue;
5717
5718                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
5719
5720                 if (!sh) {
5721                         /* failed to get a stripe - must wait */
5722                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5723                         conf->retry_read_aligned = raid_bio;
5724                         return handled;
5725                 }
5726
5727                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5728                         raid5_release_stripe(sh);
5729                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5730                         conf->retry_read_aligned = raid_bio;
5731                         return handled;
5732                 }
5733
5734                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5735                 handle_stripe(sh);
5736                 raid5_release_stripe(sh);
5737                 handled++;
5738         }
5739         remaining = raid5_dec_bi_active_stripes(raid_bio);
5740         if (remaining == 0) {
5741                 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5742                                          raid_bio, 0);
5743                 bio_endio(raid_bio);
5744         }
5745         if (atomic_dec_and_test(&conf->active_aligned_reads))
5746                 wake_up(&conf->wait_for_quiescent);
5747         return handled;
5748 }
5749
5750 static int handle_active_stripes(struct r5conf *conf, int group,
5751                                  struct r5worker *worker,
5752                                  struct list_head *temp_inactive_list)
5753 {
5754         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5755         int i, batch_size = 0, hash;
5756         bool release_inactive = false;
5757
5758         while (batch_size < MAX_STRIPE_BATCH &&
5759                         (sh = __get_priority_stripe(conf, group)) != NULL)
5760                 batch[batch_size++] = sh;
5761
5762         if (batch_size == 0) {
5763                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5764                         if (!list_empty(temp_inactive_list + i))
5765                                 break;
5766                 if (i == NR_STRIPE_HASH_LOCKS) {
5767                         spin_unlock_irq(&conf->device_lock);
5768                         r5l_flush_stripe_to_raid(conf->log);
5769                         spin_lock_irq(&conf->device_lock);
5770                         return batch_size;
5771                 }
5772                 release_inactive = true;
5773         }
5774         spin_unlock_irq(&conf->device_lock);
5775
5776         release_inactive_stripe_list(conf, temp_inactive_list,
5777                                      NR_STRIPE_HASH_LOCKS);
5778
5779         r5l_flush_stripe_to_raid(conf->log);
5780         if (release_inactive) {
5781                 spin_lock_irq(&conf->device_lock);
5782                 return 0;
5783         }
5784
5785         for (i = 0; i < batch_size; i++)
5786                 handle_stripe(batch[i]);
5787         r5l_write_stripe_run(conf->log);
5788
5789         cond_resched();
5790
5791         spin_lock_irq(&conf->device_lock);
5792         for (i = 0; i < batch_size; i++) {
5793                 hash = batch[i]->hash_lock_index;
5794                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5795         }
5796         return batch_size;
5797 }
5798
5799 static void raid5_do_work(struct work_struct *work)
5800 {
5801         struct r5worker *worker = container_of(work, struct r5worker, work);
5802         struct r5worker_group *group = worker->group;
5803         struct r5conf *conf = group->conf;
5804         int group_id = group - conf->worker_groups;
5805         int handled;
5806         struct blk_plug plug;
5807
5808         pr_debug("+++ raid5worker active\n");
5809
5810         blk_start_plug(&plug);
5811         handled = 0;
5812         spin_lock_irq(&conf->device_lock);
5813         while (1) {
5814                 int batch_size, released;
5815
5816                 released = release_stripe_list(conf, worker->temp_inactive_list);
5817
5818                 batch_size = handle_active_stripes(conf, group_id, worker,
5819                                                    worker->temp_inactive_list);
5820                 worker->working = false;
5821                 if (!batch_size && !released)
5822                         break;
5823                 handled += batch_size;
5824         }
5825         pr_debug("%d stripes handled\n", handled);
5826
5827         spin_unlock_irq(&conf->device_lock);
5828         blk_finish_plug(&plug);
5829
5830         pr_debug("--- raid5worker inactive\n");
5831 }
5832
5833 /*
5834  * This is our raid5 kernel thread.
5835  *
5836  * We scan the hash table for stripes which can be handled now.
5837  * During the scan, completed stripes are saved for us by the interrupt
5838  * handler, so that they will not have to wait for our next wakeup.
5839  */
5840 static void raid5d(struct md_thread *thread)
5841 {
5842         struct mddev *mddev = thread->mddev;
5843         struct r5conf *conf = mddev->private;
5844         int handled;
5845         struct blk_plug plug;
5846
5847         pr_debug("+++ raid5d active\n");
5848
5849         md_check_recovery(mddev);
5850
5851         if (!bio_list_empty(&conf->return_bi) &&
5852             !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5853                 struct bio_list tmp = BIO_EMPTY_LIST;
5854                 spin_lock_irq(&conf->device_lock);
5855                 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
5856                         bio_list_merge(&tmp, &conf->return_bi);
5857                         bio_list_init(&conf->return_bi);
5858                 }
5859                 spin_unlock_irq(&conf->device_lock);
5860                 return_io(&tmp);
5861         }
5862
5863         blk_start_plug(&plug);
5864         handled = 0;
5865         spin_lock_irq(&conf->device_lock);
5866         while (1) {
5867                 struct bio *bio;
5868                 int batch_size, released;
5869
5870                 released = release_stripe_list(conf, conf->temp_inactive_list);
5871                 if (released)
5872                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
5873
5874                 if (
5875                     !list_empty(&conf->bitmap_list)) {
5876                         /* Now is a good time to flush some bitmap updates */
5877                         conf->seq_flush++;
5878                         spin_unlock_irq(&conf->device_lock);
5879                         bitmap_unplug(mddev->bitmap);
5880                         spin_lock_irq(&conf->device_lock);
5881                         conf->seq_write = conf->seq_flush;
5882                         activate_bit_delay(conf, conf->temp_inactive_list);
5883                 }
5884                 raid5_activate_delayed(conf);
5885
5886                 while ((bio = remove_bio_from_retry(conf))) {
5887                         int ok;
5888                         spin_unlock_irq(&conf->device_lock);
5889                         ok = retry_aligned_read(conf, bio);
5890                         spin_lock_irq(&conf->device_lock);
5891                         if (!ok)
5892                                 break;
5893                         handled++;
5894                 }
5895
5896                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5897                                                    conf->temp_inactive_list);
5898                 if (!batch_size && !released)
5899                         break;
5900                 handled += batch_size;
5901
5902                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5903                         spin_unlock_irq(&conf->device_lock);
5904                         md_check_recovery(mddev);
5905                         spin_lock_irq(&conf->device_lock);
5906                 }
5907         }
5908         pr_debug("%d stripes handled\n", handled);
5909
5910         spin_unlock_irq(&conf->device_lock);
5911         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5912             mutex_trylock(&conf->cache_size_mutex)) {
5913                 grow_one_stripe(conf, __GFP_NOWARN);
5914                 /* Set flag even if allocation failed.  This helps
5915                  * slow down allocation requests when mem is short
5916                  */
5917                 set_bit(R5_DID_ALLOC, &conf->cache_state);
5918                 mutex_unlock(&conf->cache_size_mutex);
5919         }
5920
5921         r5l_flush_stripe_to_raid(conf->log);
5922
5923         async_tx_issue_pending_all();
5924         blk_finish_plug(&plug);
5925
5926         pr_debug("--- raid5d inactive\n");
5927 }
5928
5929 static ssize_t
5930 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5931 {
5932         struct r5conf *conf;
5933         int ret = 0;
5934         spin_lock(&mddev->lock);
5935         conf = mddev->private;
5936         if (conf)
5937                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5938         spin_unlock(&mddev->lock);
5939         return ret;
5940 }
5941
5942 int
5943 raid5_set_cache_size(struct mddev *mddev, int size)
5944 {
5945         struct r5conf *conf = mddev->private;
5946         int err;
5947
5948         if (size <= 16 || size > 32768)
5949                 return -EINVAL;
5950
5951         conf->min_nr_stripes = size;
5952         mutex_lock(&conf->cache_size_mutex);
5953         while (size < conf->max_nr_stripes &&
5954                drop_one_stripe(conf))
5955                 ;
5956         mutex_unlock(&conf->cache_size_mutex);
5957
5958
5959         err = md_allow_write(mddev);
5960         if (err)
5961                 return err;
5962
5963         mutex_lock(&conf->cache_size_mutex);
5964         while (size > conf->max_nr_stripes)
5965                 if (!grow_one_stripe(conf, GFP_KERNEL))
5966                         break;
5967         mutex_unlock(&conf->cache_size_mutex);
5968
5969         return 0;
5970 }
5971 EXPORT_SYMBOL(raid5_set_cache_size);
5972
5973 static ssize_t
5974 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5975 {
5976         struct r5conf *conf;
5977         unsigned long new;
5978         int err;
5979
5980         if (len >= PAGE_SIZE)
5981                 return -EINVAL;
5982         if (kstrtoul(page, 10, &new))
5983                 return -EINVAL;
5984         err = mddev_lock(mddev);
5985         if (err)
5986                 return err;
5987         conf = mddev->private;
5988         if (!conf)
5989                 err = -ENODEV;
5990         else
5991                 err = raid5_set_cache_size(mddev, new);
5992         mddev_unlock(mddev);
5993
5994         return err ?: len;
5995 }
5996
5997 static struct md_sysfs_entry
5998 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5999                                 raid5_show_stripe_cache_size,
6000                                 raid5_store_stripe_cache_size);
6001
6002 static ssize_t
6003 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6004 {
6005         struct r5conf *conf = mddev->private;
6006         if (conf)
6007                 return sprintf(page, "%d\n", conf->rmw_level);
6008         else
6009                 return 0;
6010 }
6011
6012 static ssize_t
6013 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6014 {
6015         struct r5conf *conf = mddev->private;
6016         unsigned long new;
6017
6018         if (!conf)
6019                 return -ENODEV;
6020
6021         if (len >= PAGE_SIZE)
6022                 return -EINVAL;
6023
6024         if (kstrtoul(page, 10, &new))
6025                 return -EINVAL;
6026
6027         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6028                 return -EINVAL;
6029
6030         if (new != PARITY_DISABLE_RMW &&
6031             new != PARITY_ENABLE_RMW &&
6032             new != PARITY_PREFER_RMW)
6033                 return -EINVAL;
6034
6035         conf->rmw_level = new;
6036         return len;
6037 }
6038
6039 static struct md_sysfs_entry
6040 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6041                          raid5_show_rmw_level,
6042                          raid5_store_rmw_level);
6043
6044
6045 static ssize_t
6046 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6047 {
6048         struct r5conf *conf;
6049         int ret = 0;
6050         spin_lock(&mddev->lock);
6051         conf = mddev->private;
6052         if (conf)
6053                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6054         spin_unlock(&mddev->lock);
6055         return ret;
6056 }
6057
6058 static ssize_t
6059 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6060 {
6061         struct r5conf *conf;
6062         unsigned long new;
6063         int err;
6064
6065         if (len >= PAGE_SIZE)
6066                 return -EINVAL;
6067         if (kstrtoul(page, 10, &new))
6068                 return -EINVAL;
6069
6070         err = mddev_lock(mddev);
6071         if (err)
6072                 return err;
6073         conf = mddev->private;
6074         if (!conf)
6075                 err = -ENODEV;
6076         else if (new > conf->min_nr_stripes)
6077                 err = -EINVAL;
6078         else
6079                 conf->bypass_threshold = new;
6080         mddev_unlock(mddev);
6081         return err ?: len;
6082 }
6083
6084 static struct md_sysfs_entry
6085 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6086                                         S_IRUGO | S_IWUSR,
6087                                         raid5_show_preread_threshold,
6088                                         raid5_store_preread_threshold);
6089
6090 static ssize_t
6091 raid5_show_skip_copy(struct mddev *mddev, char *page)
6092 {
6093         struct r5conf *conf;
6094         int ret = 0;
6095         spin_lock(&mddev->lock);
6096         conf = mddev->private;
6097         if (conf)
6098                 ret = sprintf(page, "%d\n", conf->skip_copy);
6099         spin_unlock(&mddev->lock);
6100         return ret;
6101 }
6102
6103 static ssize_t
6104 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6105 {
6106         struct r5conf *conf;
6107         unsigned long new;
6108         int err;
6109
6110         if (len >= PAGE_SIZE)
6111                 return -EINVAL;
6112         if (kstrtoul(page, 10, &new))
6113                 return -EINVAL;
6114         new = !!new;
6115
6116         err = mddev_lock(mddev);
6117         if (err)
6118                 return err;
6119         conf = mddev->private;
6120         if (!conf)
6121                 err = -ENODEV;
6122         else if (new != conf->skip_copy) {
6123                 mddev_suspend(mddev);
6124                 conf->skip_copy = new;
6125                 if (new)
6126                         mddev->queue->backing_dev_info.capabilities |=
6127                                 BDI_CAP_STABLE_WRITES;
6128                 else
6129                         mddev->queue->backing_dev_info.capabilities &=
6130                                 ~BDI_CAP_STABLE_WRITES;
6131                 mddev_resume(mddev);
6132         }
6133         mddev_unlock(mddev);
6134         return err ?: len;
6135 }
6136
6137 static struct md_sysfs_entry
6138 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6139                                         raid5_show_skip_copy,
6140                                         raid5_store_skip_copy);
6141
6142 static ssize_t
6143 stripe_cache_active_show(struct mddev *mddev, char *page)
6144 {
6145         struct r5conf *conf = mddev->private;
6146         if (conf)
6147                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6148         else
6149                 return 0;
6150 }
6151
6152 static struct md_sysfs_entry
6153 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6154
6155 static ssize_t
6156 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6157 {
6158         struct r5conf *conf;
6159         int ret = 0;
6160         spin_lock(&mddev->lock);
6161         conf = mddev->private;
6162         if (conf)
6163                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6164         spin_unlock(&mddev->lock);
6165         return ret;
6166 }
6167
6168 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6169                                int *group_cnt,
6170                                int *worker_cnt_per_group,
6171                                struct r5worker_group **worker_groups);
6172 static ssize_t
6173 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6174 {
6175         struct r5conf *conf;
6176         unsigned long new;
6177         int err;
6178         struct r5worker_group *new_groups, *old_groups;
6179         int group_cnt, worker_cnt_per_group;
6180
6181         if (len >= PAGE_SIZE)
6182                 return -EINVAL;
6183         if (kstrtoul(page, 10, &new))
6184                 return -EINVAL;
6185
6186         err = mddev_lock(mddev);
6187         if (err)
6188                 return err;
6189         conf = mddev->private;
6190         if (!conf)
6191                 err = -ENODEV;
6192         else if (new != conf->worker_cnt_per_group) {
6193                 mddev_suspend(mddev);
6194
6195                 old_groups = conf->worker_groups;
6196                 if (old_groups)
6197                         flush_workqueue(raid5_wq);
6198
6199                 err = alloc_thread_groups(conf, new,
6200                                           &group_cnt, &worker_cnt_per_group,
6201                                           &new_groups);
6202                 if (!err) {
6203                         spin_lock_irq(&conf->device_lock);
6204                         conf->group_cnt = group_cnt;
6205                         conf->worker_cnt_per_group = worker_cnt_per_group;
6206                         conf->worker_groups = new_groups;
6207                         spin_unlock_irq(&conf->device_lock);
6208
6209                         if (old_groups)
6210                                 kfree(old_groups[0].workers);
6211                         kfree(old_groups);
6212                 }
6213                 mddev_resume(mddev);
6214         }
6215         mddev_unlock(mddev);
6216
6217         return err ?: len;
6218 }
6219
6220 static struct md_sysfs_entry
6221 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6222                                 raid5_show_group_thread_cnt,
6223                                 raid5_store_group_thread_cnt);
6224
6225 static struct attribute *raid5_attrs[] =  {
6226         &raid5_stripecache_size.attr,
6227         &raid5_stripecache_active.attr,
6228         &raid5_preread_bypass_threshold.attr,
6229         &raid5_group_thread_cnt.attr,
6230         &raid5_skip_copy.attr,
6231         &raid5_rmw_level.attr,
6232         NULL,
6233 };
6234 static struct attribute_group raid5_attrs_group = {
6235         .name = NULL,
6236         .attrs = raid5_attrs,
6237 };
6238
6239 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6240                                int *group_cnt,
6241                                int *worker_cnt_per_group,
6242                                struct r5worker_group **worker_groups)
6243 {
6244         int i, j, k;
6245         ssize_t size;
6246         struct r5worker *workers;
6247
6248         *worker_cnt_per_group = cnt;
6249         if (cnt == 0) {
6250                 *group_cnt = 0;
6251                 *worker_groups = NULL;
6252                 return 0;
6253         }
6254         *group_cnt = num_possible_nodes();
6255         size = sizeof(struct r5worker) * cnt;
6256         workers = kzalloc(size * *group_cnt, GFP_NOIO);
6257         *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6258                                 *group_cnt, GFP_NOIO);
6259         if (!*worker_groups || !workers) {
6260                 kfree(workers);
6261                 kfree(*worker_groups);
6262                 return -ENOMEM;
6263         }
6264
6265         for (i = 0; i < *group_cnt; i++) {
6266                 struct r5worker_group *group;
6267
6268                 group = &(*worker_groups)[i];
6269                 INIT_LIST_HEAD(&group->handle_list);
6270                 group->conf = conf;
6271                 group->workers = workers + i * cnt;
6272
6273                 for (j = 0; j < cnt; j++) {
6274                         struct r5worker *worker = group->workers + j;
6275                         worker->group = group;
6276                         INIT_WORK(&worker->work, raid5_do_work);
6277
6278                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6279                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6280                 }
6281         }
6282
6283         return 0;
6284 }
6285
6286 static void free_thread_groups(struct r5conf *conf)
6287 {
6288         if (conf->worker_groups)
6289                 kfree(conf->worker_groups[0].workers);
6290         kfree(conf->worker_groups);
6291         conf->worker_groups = NULL;
6292 }
6293
6294 static sector_t
6295 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6296 {
6297         struct r5conf *conf = mddev->private;
6298
6299         if (!sectors)
6300                 sectors = mddev->dev_sectors;
6301         if (!raid_disks)
6302                 /* size is defined by the smallest of previous and new size */
6303                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6304
6305         sectors &= ~((sector_t)conf->chunk_sectors - 1);
6306         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6307         return sectors * (raid_disks - conf->max_degraded);
6308 }
6309
6310 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6311 {
6312         safe_put_page(percpu->spare_page);
6313         if (percpu->scribble)
6314                 flex_array_free(percpu->scribble);
6315         percpu->spare_page = NULL;
6316         percpu->scribble = NULL;
6317 }
6318
6319 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6320 {
6321         if (conf->level == 6 && !percpu->spare_page)
6322                 percpu->spare_page = alloc_page(GFP_KERNEL);
6323         if (!percpu->scribble)
6324                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6325                                                       conf->previous_raid_disks),
6326                                                   max(conf->chunk_sectors,
6327                                                       conf->prev_chunk_sectors)
6328                                                    / STRIPE_SECTORS,
6329                                                   GFP_KERNEL);
6330
6331         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6332                 free_scratch_buffer(conf, percpu);
6333                 return -ENOMEM;
6334         }
6335
6336         return 0;
6337 }
6338
6339 static void raid5_free_percpu(struct r5conf *conf)
6340 {
6341         unsigned long cpu;
6342
6343         if (!conf->percpu)
6344                 return;
6345
6346 #ifdef CONFIG_HOTPLUG_CPU
6347         unregister_cpu_notifier(&conf->cpu_notify);
6348 #endif
6349
6350         get_online_cpus();
6351         for_each_possible_cpu(cpu)
6352                 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6353         put_online_cpus();
6354
6355         free_percpu(conf->percpu);
6356 }
6357
6358 static void free_conf(struct r5conf *conf)
6359 {
6360         if (conf->log)
6361                 r5l_exit_log(conf->log);
6362         if (conf->shrinker.seeks)
6363                 unregister_shrinker(&conf->shrinker);
6364
6365         free_thread_groups(conf);
6366         shrink_stripes(conf);
6367         raid5_free_percpu(conf);
6368         kfree(conf->disks);
6369         kfree(conf->stripe_hashtbl);
6370         kfree(conf);
6371 }
6372
6373 #ifdef CONFIG_HOTPLUG_CPU
6374 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6375                               void *hcpu)
6376 {
6377         struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6378         long cpu = (long)hcpu;
6379         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6380
6381         switch (action) {
6382         case CPU_UP_PREPARE:
6383         case CPU_UP_PREPARE_FROZEN:
6384                 if (alloc_scratch_buffer(conf, percpu)) {
6385                         pr_err("%s: failed memory allocation for cpu%ld\n",
6386                                __func__, cpu);
6387                         return notifier_from_errno(-ENOMEM);
6388                 }
6389                 break;
6390         case CPU_DEAD:
6391         case CPU_DEAD_FROZEN:
6392                 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6393                 break;
6394         default:
6395                 break;
6396         }
6397         return NOTIFY_OK;
6398 }
6399 #endif
6400
6401 static int raid5_alloc_percpu(struct r5conf *conf)
6402 {
6403         unsigned long cpu;
6404         int err = 0;
6405
6406         conf->percpu = alloc_percpu(struct raid5_percpu);
6407         if (!conf->percpu)
6408                 return -ENOMEM;
6409
6410 #ifdef CONFIG_HOTPLUG_CPU
6411         conf->cpu_notify.notifier_call = raid456_cpu_notify;
6412         conf->cpu_notify.priority = 0;
6413         err = register_cpu_notifier(&conf->cpu_notify);
6414         if (err)
6415                 return err;
6416 #endif
6417
6418         get_online_cpus();
6419         for_each_present_cpu(cpu) {
6420                 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6421                 if (err) {
6422                         pr_err("%s: failed memory allocation for cpu%ld\n",
6423                                __func__, cpu);
6424                         break;
6425                 }
6426         }
6427         put_online_cpus();
6428
6429         if (!err) {
6430                 conf->scribble_disks = max(conf->raid_disks,
6431                         conf->previous_raid_disks);
6432                 conf->scribble_sectors = max(conf->chunk_sectors,
6433                         conf->prev_chunk_sectors);
6434         }
6435         return err;
6436 }
6437
6438 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6439                                       struct shrink_control *sc)
6440 {
6441         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6442         unsigned long ret = SHRINK_STOP;
6443
6444         if (mutex_trylock(&conf->cache_size_mutex)) {
6445                 ret= 0;
6446                 while (ret < sc->nr_to_scan &&
6447                        conf->max_nr_stripes > conf->min_nr_stripes) {
6448                         if (drop_one_stripe(conf) == 0) {
6449                                 ret = SHRINK_STOP;
6450                                 break;
6451                         }
6452                         ret++;
6453                 }
6454                 mutex_unlock(&conf->cache_size_mutex);
6455         }
6456         return ret;
6457 }
6458
6459 static unsigned long raid5_cache_count(struct shrinker *shrink,
6460                                        struct shrink_control *sc)
6461 {
6462         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6463
6464         if (conf->max_nr_stripes < conf->min_nr_stripes)
6465                 /* unlikely, but not impossible */
6466                 return 0;
6467         return conf->max_nr_stripes - conf->min_nr_stripes;
6468 }
6469
6470 static struct r5conf *setup_conf(struct mddev *mddev)
6471 {
6472         struct r5conf *conf;
6473         int raid_disk, memory, max_disks;
6474         struct md_rdev *rdev;
6475         struct disk_info *disk;
6476         char pers_name[6];
6477         int i;
6478         int group_cnt, worker_cnt_per_group;
6479         struct r5worker_group *new_group;
6480
6481         if (mddev->new_level != 5
6482             && mddev->new_level != 4
6483             && mddev->new_level != 6) {
6484                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6485                        mdname(mddev), mddev->new_level);
6486                 return ERR_PTR(-EIO);
6487         }
6488         if ((mddev->new_level == 5
6489              && !algorithm_valid_raid5(mddev->new_layout)) ||
6490             (mddev->new_level == 6
6491              && !algorithm_valid_raid6(mddev->new_layout))) {
6492                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6493                        mdname(mddev), mddev->new_layout);
6494                 return ERR_PTR(-EIO);
6495         }
6496         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6497                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6498                        mdname(mddev), mddev->raid_disks);
6499                 return ERR_PTR(-EINVAL);
6500         }
6501
6502         if (!mddev->new_chunk_sectors ||
6503             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6504             !is_power_of_2(mddev->new_chunk_sectors)) {
6505                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6506                        mdname(mddev), mddev->new_chunk_sectors << 9);
6507                 return ERR_PTR(-EINVAL);
6508         }
6509
6510         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6511         if (conf == NULL)
6512                 goto abort;
6513         /* Don't enable multi-threading by default*/
6514         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6515                                  &new_group)) {
6516                 conf->group_cnt = group_cnt;
6517                 conf->worker_cnt_per_group = worker_cnt_per_group;
6518                 conf->worker_groups = new_group;
6519         } else
6520                 goto abort;
6521         spin_lock_init(&conf->device_lock);
6522         seqcount_init(&conf->gen_lock);
6523         mutex_init(&conf->cache_size_mutex);
6524         init_waitqueue_head(&conf->wait_for_quiescent);
6525         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++) {
6526                 init_waitqueue_head(&conf->wait_for_stripe[i]);
6527         }
6528         init_waitqueue_head(&conf->wait_for_overlap);
6529         INIT_LIST_HEAD(&conf->handle_list);
6530         INIT_LIST_HEAD(&conf->hold_list);
6531         INIT_LIST_HEAD(&conf->delayed_list);
6532         INIT_LIST_HEAD(&conf->bitmap_list);
6533         bio_list_init(&conf->return_bi);
6534         init_llist_head(&conf->released_stripes);
6535         atomic_set(&conf->active_stripes, 0);
6536         atomic_set(&conf->preread_active_stripes, 0);
6537         atomic_set(&conf->active_aligned_reads, 0);
6538         conf->bypass_threshold = BYPASS_THRESHOLD;
6539         conf->recovery_disabled = mddev->recovery_disabled - 1;
6540
6541         conf->raid_disks = mddev->raid_disks;
6542         if (mddev->reshape_position == MaxSector)
6543                 conf->previous_raid_disks = mddev->raid_disks;
6544         else
6545                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6546         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6547
6548         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6549                               GFP_KERNEL);
6550         if (!conf->disks)
6551                 goto abort;
6552
6553         conf->mddev = mddev;
6554
6555         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6556                 goto abort;
6557
6558         /* We init hash_locks[0] separately to that it can be used
6559          * as the reference lock in the spin_lock_nest_lock() call
6560          * in lock_all_device_hash_locks_irq in order to convince
6561          * lockdep that we know what we are doing.
6562          */
6563         spin_lock_init(conf->hash_locks);
6564         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6565                 spin_lock_init(conf->hash_locks + i);
6566
6567         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6568                 INIT_LIST_HEAD(conf->inactive_list + i);
6569
6570         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6571                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6572
6573         conf->level = mddev->new_level;
6574         conf->chunk_sectors = mddev->new_chunk_sectors;
6575         if (raid5_alloc_percpu(conf) != 0)
6576                 goto abort;
6577
6578         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6579
6580         rdev_for_each(rdev, mddev) {
6581                 raid_disk = rdev->raid_disk;
6582                 if (raid_disk >= max_disks
6583                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
6584                         continue;
6585                 disk = conf->disks + raid_disk;
6586
6587                 if (test_bit(Replacement, &rdev->flags)) {
6588                         if (disk->replacement)
6589                                 goto abort;
6590                         disk->replacement = rdev;
6591                 } else {
6592                         if (disk->rdev)
6593                                 goto abort;
6594                         disk->rdev = rdev;
6595                 }
6596
6597                 if (test_bit(In_sync, &rdev->flags)) {
6598                         char b[BDEVNAME_SIZE];
6599                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6600                                " disk %d\n",
6601                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6602                 } else if (rdev->saved_raid_disk != raid_disk)
6603                         /* Cannot rely on bitmap to complete recovery */
6604                         conf->fullsync = 1;
6605         }
6606
6607         conf->level = mddev->new_level;
6608         if (conf->level == 6) {
6609                 conf->max_degraded = 2;
6610                 if (raid6_call.xor_syndrome)
6611                         conf->rmw_level = PARITY_ENABLE_RMW;
6612                 else
6613                         conf->rmw_level = PARITY_DISABLE_RMW;
6614         } else {
6615                 conf->max_degraded = 1;
6616                 conf->rmw_level = PARITY_ENABLE_RMW;
6617         }
6618         conf->algorithm = mddev->new_layout;
6619         conf->reshape_progress = mddev->reshape_position;
6620         if (conf->reshape_progress != MaxSector) {
6621                 conf->prev_chunk_sectors = mddev->chunk_sectors;
6622                 conf->prev_algo = mddev->layout;
6623         } else {
6624                 conf->prev_chunk_sectors = conf->chunk_sectors;
6625                 conf->prev_algo = conf->algorithm;
6626         }
6627
6628         conf->min_nr_stripes = NR_STRIPES;
6629         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6630                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6631         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6632         if (grow_stripes(conf, conf->min_nr_stripes)) {
6633                 printk(KERN_ERR
6634                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
6635                        mdname(mddev), memory);
6636                 goto abort;
6637         } else
6638                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6639                        mdname(mddev), memory);
6640         /*
6641          * Losing a stripe head costs more than the time to refill it,
6642          * it reduces the queue depth and so can hurt throughput.
6643          * So set it rather large, scaled by number of devices.
6644          */
6645         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6646         conf->shrinker.scan_objects = raid5_cache_scan;
6647         conf->shrinker.count_objects = raid5_cache_count;
6648         conf->shrinker.batch = 128;
6649         conf->shrinker.flags = 0;
6650         register_shrinker(&conf->shrinker);
6651
6652         sprintf(pers_name, "raid%d", mddev->new_level);
6653         conf->thread = md_register_thread(raid5d, mddev, pers_name);
6654         if (!conf->thread) {
6655                 printk(KERN_ERR
6656                        "md/raid:%s: couldn't allocate thread.\n",
6657                        mdname(mddev));
6658                 goto abort;
6659         }
6660
6661         return conf;
6662
6663  abort:
6664         if (conf) {
6665                 free_conf(conf);
6666                 return ERR_PTR(-EIO);
6667         } else
6668                 return ERR_PTR(-ENOMEM);
6669 }
6670
6671 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6672 {
6673         switch (algo) {
6674         case ALGORITHM_PARITY_0:
6675                 if (raid_disk < max_degraded)
6676                         return 1;
6677                 break;
6678         case ALGORITHM_PARITY_N:
6679                 if (raid_disk >= raid_disks - max_degraded)
6680                         return 1;
6681                 break;
6682         case ALGORITHM_PARITY_0_6:
6683                 if (raid_disk == 0 ||
6684                     raid_disk == raid_disks - 1)
6685                         return 1;
6686                 break;
6687         case ALGORITHM_LEFT_ASYMMETRIC_6:
6688         case ALGORITHM_RIGHT_ASYMMETRIC_6:
6689         case ALGORITHM_LEFT_SYMMETRIC_6:
6690         case ALGORITHM_RIGHT_SYMMETRIC_6:
6691                 if (raid_disk == raid_disks - 1)
6692                         return 1;
6693         }
6694         return 0;
6695 }
6696
6697 static int run(struct mddev *mddev)
6698 {
6699         struct r5conf *conf;
6700         int working_disks = 0;
6701         int dirty_parity_disks = 0;
6702         struct md_rdev *rdev;
6703         struct md_rdev *journal_dev = NULL;
6704         sector_t reshape_offset = 0;
6705         int i;
6706         long long min_offset_diff = 0;
6707         int first = 1;
6708
6709         if (mddev->recovery_cp != MaxSector)
6710                 printk(KERN_NOTICE "md/raid:%s: not clean"
6711                        " -- starting background reconstruction\n",
6712                        mdname(mddev));
6713
6714         rdev_for_each(rdev, mddev) {
6715                 long long diff;
6716
6717                 if (test_bit(Journal, &rdev->flags)) {
6718                         journal_dev = rdev;
6719                         continue;
6720                 }
6721                 if (rdev->raid_disk < 0)
6722                         continue;
6723                 diff = (rdev->new_data_offset - rdev->data_offset);
6724                 if (first) {
6725                         min_offset_diff = diff;
6726                         first = 0;
6727                 } else if (mddev->reshape_backwards &&
6728                          diff < min_offset_diff)
6729                         min_offset_diff = diff;
6730                 else if (!mddev->reshape_backwards &&
6731                          diff > min_offset_diff)
6732                         min_offset_diff = diff;
6733         }
6734
6735         if (mddev->reshape_position != MaxSector) {
6736                 /* Check that we can continue the reshape.
6737                  * Difficulties arise if the stripe we would write to
6738                  * next is at or after the stripe we would read from next.
6739                  * For a reshape that changes the number of devices, this
6740                  * is only possible for a very short time, and mdadm makes
6741                  * sure that time appears to have past before assembling
6742                  * the array.  So we fail if that time hasn't passed.
6743                  * For a reshape that keeps the number of devices the same
6744                  * mdadm must be monitoring the reshape can keeping the
6745                  * critical areas read-only and backed up.  It will start
6746                  * the array in read-only mode, so we check for that.
6747                  */
6748                 sector_t here_new, here_old;
6749                 int old_disks;
6750                 int max_degraded = (mddev->level == 6 ? 2 : 1);
6751                 int chunk_sectors;
6752                 int new_data_disks;
6753
6754                 if (journal_dev) {
6755                         printk(KERN_ERR "md/raid:%s: don't support reshape with journal - aborting.\n",
6756                                mdname(mddev));
6757                         return -EINVAL;
6758                 }
6759
6760                 if (mddev->new_level != mddev->level) {
6761                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
6762                                "required - aborting.\n",
6763                                mdname(mddev));
6764                         return -EINVAL;
6765                 }
6766                 old_disks = mddev->raid_disks - mddev->delta_disks;
6767                 /* reshape_position must be on a new-stripe boundary, and one
6768                  * further up in new geometry must map after here in old
6769                  * geometry.
6770                  * If the chunk sizes are different, then as we perform reshape
6771                  * in units of the largest of the two, reshape_position needs
6772                  * be a multiple of the largest chunk size times new data disks.
6773                  */
6774                 here_new = mddev->reshape_position;
6775                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
6776                 new_data_disks = mddev->raid_disks - max_degraded;
6777                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
6778                         printk(KERN_ERR "md/raid:%s: reshape_position not "
6779                                "on a stripe boundary\n", mdname(mddev));
6780                         return -EINVAL;
6781                 }
6782                 reshape_offset = here_new * chunk_sectors;
6783                 /* here_new is the stripe we will write to */
6784                 here_old = mddev->reshape_position;
6785                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
6786                 /* here_old is the first stripe that we might need to read
6787                  * from */
6788                 if (mddev->delta_disks == 0) {
6789                         /* We cannot be sure it is safe to start an in-place
6790                          * reshape.  It is only safe if user-space is monitoring
6791                          * and taking constant backups.
6792                          * mdadm always starts a situation like this in
6793                          * readonly mode so it can take control before
6794                          * allowing any writes.  So just check for that.
6795                          */
6796                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6797                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
6798                                 /* not really in-place - so OK */;
6799                         else if (mddev->ro == 0) {
6800                                 printk(KERN_ERR "md/raid:%s: in-place reshape "
6801                                        "must be started in read-only mode "
6802                                        "- aborting\n",
6803                                        mdname(mddev));
6804                                 return -EINVAL;
6805                         }
6806                 } else if (mddev->reshape_backwards
6807                     ? (here_new * chunk_sectors + min_offset_diff <=
6808                        here_old * chunk_sectors)
6809                     : (here_new * chunk_sectors >=
6810                        here_old * chunk_sectors + (-min_offset_diff))) {
6811                         /* Reading from the same stripe as writing to - bad */
6812                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6813                                "auto-recovery - aborting.\n",
6814                                mdname(mddev));
6815                         return -EINVAL;
6816                 }
6817                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6818                        mdname(mddev));
6819                 /* OK, we should be able to continue; */
6820         } else {
6821                 BUG_ON(mddev->level != mddev->new_level);
6822                 BUG_ON(mddev->layout != mddev->new_layout);
6823                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6824                 BUG_ON(mddev->delta_disks != 0);
6825         }
6826
6827         if (mddev->private == NULL)
6828                 conf = setup_conf(mddev);
6829         else
6830                 conf = mddev->private;
6831
6832         if (IS_ERR(conf))
6833                 return PTR_ERR(conf);
6834
6835         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) && !journal_dev) {
6836                 printk(KERN_ERR "md/raid:%s: journal disk is missing, force array readonly\n",
6837                        mdname(mddev));
6838                 mddev->ro = 1;
6839                 set_disk_ro(mddev->gendisk, 1);
6840         }
6841
6842         conf->min_offset_diff = min_offset_diff;
6843         mddev->thread = conf->thread;
6844         conf->thread = NULL;
6845         mddev->private = conf;
6846
6847         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6848              i++) {
6849                 rdev = conf->disks[i].rdev;
6850                 if (!rdev && conf->disks[i].replacement) {
6851                         /* The replacement is all we have yet */
6852                         rdev = conf->disks[i].replacement;
6853                         conf->disks[i].replacement = NULL;
6854                         clear_bit(Replacement, &rdev->flags);
6855                         conf->disks[i].rdev = rdev;
6856                 }
6857                 if (!rdev)
6858                         continue;
6859                 if (conf->disks[i].replacement &&
6860                     conf->reshape_progress != MaxSector) {
6861                         /* replacements and reshape simply do not mix. */
6862                         printk(KERN_ERR "md: cannot handle concurrent "
6863                                "replacement and reshape.\n");
6864                         goto abort;
6865                 }
6866                 if (test_bit(In_sync, &rdev->flags)) {
6867                         working_disks++;
6868                         continue;
6869                 }
6870                 /* This disc is not fully in-sync.  However if it
6871                  * just stored parity (beyond the recovery_offset),
6872                  * when we don't need to be concerned about the
6873                  * array being dirty.
6874                  * When reshape goes 'backwards', we never have
6875                  * partially completed devices, so we only need
6876                  * to worry about reshape going forwards.
6877                  */
6878                 /* Hack because v0.91 doesn't store recovery_offset properly. */
6879                 if (mddev->major_version == 0 &&
6880                     mddev->minor_version > 90)
6881                         rdev->recovery_offset = reshape_offset;
6882
6883                 if (rdev->recovery_offset < reshape_offset) {
6884                         /* We need to check old and new layout */
6885                         if (!only_parity(rdev->raid_disk,
6886                                          conf->algorithm,
6887                                          conf->raid_disks,
6888                                          conf->max_degraded))
6889                                 continue;
6890                 }
6891                 if (!only_parity(rdev->raid_disk,
6892                                  conf->prev_algo,
6893                                  conf->previous_raid_disks,
6894                                  conf->max_degraded))
6895                         continue;
6896                 dirty_parity_disks++;
6897         }
6898
6899         /*
6900          * 0 for a fully functional array, 1 or 2 for a degraded array.
6901          */
6902         mddev->degraded = calc_degraded(conf);
6903
6904         if (has_failed(conf)) {
6905                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6906                         " (%d/%d failed)\n",
6907                         mdname(mddev), mddev->degraded, conf->raid_disks);
6908                 goto abort;
6909         }
6910
6911         /* device size must be a multiple of chunk size */
6912         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6913         mddev->resync_max_sectors = mddev->dev_sectors;
6914
6915         if (mddev->degraded > dirty_parity_disks &&
6916             mddev->recovery_cp != MaxSector) {
6917                 if (mddev->ok_start_degraded)
6918                         printk(KERN_WARNING
6919                                "md/raid:%s: starting dirty degraded array"
6920                                " - data corruption possible.\n",
6921                                mdname(mddev));
6922                 else {
6923                         printk(KERN_ERR
6924                                "md/raid:%s: cannot start dirty degraded array.\n",
6925                                mdname(mddev));
6926                         goto abort;
6927                 }
6928         }
6929
6930         if (mddev->degraded == 0)
6931                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6932                        " devices, algorithm %d\n", mdname(mddev), conf->level,
6933                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6934                        mddev->new_layout);
6935         else
6936                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6937                        " out of %d devices, algorithm %d\n",
6938                        mdname(mddev), conf->level,
6939                        mddev->raid_disks - mddev->degraded,
6940                        mddev->raid_disks, mddev->new_layout);
6941
6942         print_raid5_conf(conf);
6943
6944         if (conf->reshape_progress != MaxSector) {
6945                 conf->reshape_safe = conf->reshape_progress;
6946                 atomic_set(&conf->reshape_stripes, 0);
6947                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6948                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6949                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6950                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6951                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6952                                                         "reshape");
6953         }
6954
6955         /* Ok, everything is just fine now */
6956         if (mddev->to_remove == &raid5_attrs_group)
6957                 mddev->to_remove = NULL;
6958         else if (mddev->kobj.sd &&
6959             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6960                 printk(KERN_WARNING
6961                        "raid5: failed to create sysfs attributes for %s\n",
6962                        mdname(mddev));
6963         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6964
6965         if (mddev->queue) {
6966                 int chunk_size;
6967                 bool discard_supported = true;
6968                 /* read-ahead size must cover two whole stripes, which
6969                  * is 2 * (datadisks) * chunksize where 'n' is the
6970                  * number of raid devices
6971                  */
6972                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6973                 int stripe = data_disks *
6974                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6975                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6976                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6977
6978                 chunk_size = mddev->chunk_sectors << 9;
6979                 blk_queue_io_min(mddev->queue, chunk_size);
6980                 blk_queue_io_opt(mddev->queue, chunk_size *
6981                                  (conf->raid_disks - conf->max_degraded));
6982                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6983                 /*
6984                  * We can only discard a whole stripe. It doesn't make sense to
6985                  * discard data disk but write parity disk
6986                  */
6987                 stripe = stripe * PAGE_SIZE;
6988                 /* Round up to power of 2, as discard handling
6989                  * currently assumes that */
6990                 while ((stripe-1) & stripe)
6991                         stripe = (stripe | (stripe-1)) + 1;
6992                 mddev->queue->limits.discard_alignment = stripe;
6993                 mddev->queue->limits.discard_granularity = stripe;
6994                 /*
6995                  * unaligned part of discard request will be ignored, so can't
6996                  * guarantee discard_zeroes_data
6997                  */
6998                 mddev->queue->limits.discard_zeroes_data = 0;
6999
7000                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7001
7002                 rdev_for_each(rdev, mddev) {
7003                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7004                                           rdev->data_offset << 9);
7005                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7006                                           rdev->new_data_offset << 9);
7007                         /*
7008                          * discard_zeroes_data is required, otherwise data
7009                          * could be lost. Consider a scenario: discard a stripe
7010                          * (the stripe could be inconsistent if
7011                          * discard_zeroes_data is 0); write one disk of the
7012                          * stripe (the stripe could be inconsistent again
7013                          * depending on which disks are used to calculate
7014                          * parity); the disk is broken; The stripe data of this
7015                          * disk is lost.
7016                          */
7017                         if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
7018                             !bdev_get_queue(rdev->bdev)->
7019                                                 limits.discard_zeroes_data)
7020                                 discard_supported = false;
7021                         /* Unfortunately, discard_zeroes_data is not currently
7022                          * a guarantee - just a hint.  So we only allow DISCARD
7023                          * if the sysadmin has confirmed that only safe devices
7024                          * are in use by setting a module parameter.
7025                          */
7026                         if (!devices_handle_discard_safely) {
7027                                 if (discard_supported) {
7028                                         pr_info("md/raid456: discard support disabled due to uncertainty.\n");
7029                                         pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
7030                                 }
7031                                 discard_supported = false;
7032                         }
7033                 }
7034
7035                 if (discard_supported &&
7036                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7037                     mddev->queue->limits.discard_granularity >= stripe)
7038                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
7039                                                 mddev->queue);
7040                 else
7041                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
7042                                                 mddev->queue);
7043         }
7044
7045         if (journal_dev) {
7046                 char b[BDEVNAME_SIZE];
7047
7048                 printk(KERN_INFO"md/raid:%s: using device %s as journal\n",
7049                        mdname(mddev), bdevname(journal_dev->bdev, b));
7050                 r5l_init_log(conf, journal_dev);
7051         }
7052
7053         return 0;
7054 abort:
7055         md_unregister_thread(&mddev->thread);
7056         print_raid5_conf(conf);
7057         free_conf(conf);
7058         mddev->private = NULL;
7059         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
7060         return -EIO;
7061 }
7062
7063 static void raid5_free(struct mddev *mddev, void *priv)
7064 {
7065         struct r5conf *conf = priv;
7066
7067         free_conf(conf);
7068         mddev->to_remove = &raid5_attrs_group;
7069 }
7070
7071 static void status(struct seq_file *seq, struct mddev *mddev)
7072 {
7073         struct r5conf *conf = mddev->private;
7074         int i;
7075
7076         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7077                 conf->chunk_sectors / 2, mddev->layout);
7078         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7079         for (i = 0; i < conf->raid_disks; i++)
7080                 seq_printf (seq, "%s",
7081                                conf->disks[i].rdev &&
7082                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
7083         seq_printf (seq, "]");
7084 }
7085
7086 static void print_raid5_conf (struct r5conf *conf)
7087 {
7088         int i;
7089         struct disk_info *tmp;
7090
7091         printk(KERN_DEBUG "RAID conf printout:\n");
7092         if (!conf) {
7093                 printk("(conf==NULL)\n");
7094                 return;
7095         }
7096         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
7097                conf->raid_disks,
7098                conf->raid_disks - conf->mddev->degraded);
7099
7100         for (i = 0; i < conf->raid_disks; i++) {
7101                 char b[BDEVNAME_SIZE];
7102                 tmp = conf->disks + i;
7103                 if (tmp->rdev)
7104                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7105                                i, !test_bit(Faulty, &tmp->rdev->flags),
7106                                bdevname(tmp->rdev->bdev, b));
7107         }
7108 }
7109
7110 static int raid5_spare_active(struct mddev *mddev)
7111 {
7112         int i;
7113         struct r5conf *conf = mddev->private;
7114         struct disk_info *tmp;
7115         int count = 0;
7116         unsigned long flags;
7117
7118         for (i = 0; i < conf->raid_disks; i++) {
7119                 tmp = conf->disks + i;
7120                 if (tmp->replacement
7121                     && tmp->replacement->recovery_offset == MaxSector
7122                     && !test_bit(Faulty, &tmp->replacement->flags)
7123                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7124                         /* Replacement has just become active. */
7125                         if (!tmp->rdev
7126                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7127                                 count++;
7128                         if (tmp->rdev) {
7129                                 /* Replaced device not technically faulty,
7130                                  * but we need to be sure it gets removed
7131                                  * and never re-added.
7132                                  */
7133                                 set_bit(Faulty, &tmp->rdev->flags);
7134                                 sysfs_notify_dirent_safe(
7135                                         tmp->rdev->sysfs_state);
7136                         }
7137                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7138                 } else if (tmp->rdev
7139                     && tmp->rdev->recovery_offset == MaxSector
7140                     && !test_bit(Faulty, &tmp->rdev->flags)
7141                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7142                         count++;
7143                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7144                 }
7145         }
7146         spin_lock_irqsave(&conf->device_lock, flags);
7147         mddev->degraded = calc_degraded(conf);
7148         spin_unlock_irqrestore(&conf->device_lock, flags);
7149         print_raid5_conf(conf);
7150         return count;
7151 }
7152
7153 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7154 {
7155         struct r5conf *conf = mddev->private;
7156         int err = 0;
7157         int number = rdev->raid_disk;
7158         struct md_rdev **rdevp;
7159         struct disk_info *p = conf->disks + number;
7160
7161         print_raid5_conf(conf);
7162         if (test_bit(Journal, &rdev->flags)) {
7163                 /*
7164                  * journal disk is not removable, but we need give a chance to
7165                  * update superblock of other disks. Otherwise journal disk
7166                  * will be considered as 'fresh'
7167                  */
7168                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7169                 return -EINVAL;
7170         }
7171         if (rdev == p->rdev)
7172                 rdevp = &p->rdev;
7173         else if (rdev == p->replacement)
7174                 rdevp = &p->replacement;
7175         else
7176                 return 0;
7177
7178         if (number >= conf->raid_disks &&
7179             conf->reshape_progress == MaxSector)
7180                 clear_bit(In_sync, &rdev->flags);
7181
7182         if (test_bit(In_sync, &rdev->flags) ||
7183             atomic_read(&rdev->nr_pending)) {
7184                 err = -EBUSY;
7185                 goto abort;
7186         }
7187         /* Only remove non-faulty devices if recovery
7188          * isn't possible.
7189          */
7190         if (!test_bit(Faulty, &rdev->flags) &&
7191             mddev->recovery_disabled != conf->recovery_disabled &&
7192             !has_failed(conf) &&
7193             (!p->replacement || p->replacement == rdev) &&
7194             number < conf->raid_disks) {
7195                 err = -EBUSY;
7196                 goto abort;
7197         }
7198         *rdevp = NULL;
7199         synchronize_rcu();
7200         if (atomic_read(&rdev->nr_pending)) {
7201                 /* lost the race, try later */
7202                 err = -EBUSY;
7203                 *rdevp = rdev;
7204         } else if (p->replacement) {
7205                 /* We must have just cleared 'rdev' */
7206                 p->rdev = p->replacement;
7207                 clear_bit(Replacement, &p->replacement->flags);
7208                 smp_mb(); /* Make sure other CPUs may see both as identical
7209                            * but will never see neither - if they are careful
7210                            */
7211                 p->replacement = NULL;
7212                 clear_bit(WantReplacement, &rdev->flags);
7213         } else
7214                 /* We might have just removed the Replacement as faulty-
7215                  * clear the bit just in case
7216                  */
7217                 clear_bit(WantReplacement, &rdev->flags);
7218 abort:
7219
7220         print_raid5_conf(conf);
7221         return err;
7222 }
7223
7224 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7225 {
7226         struct r5conf *conf = mddev->private;
7227         int err = -EEXIST;
7228         int disk;
7229         struct disk_info *p;
7230         int first = 0;
7231         int last = conf->raid_disks - 1;
7232
7233         if (test_bit(Journal, &rdev->flags))
7234                 return -EINVAL;
7235         if (mddev->recovery_disabled == conf->recovery_disabled)
7236                 return -EBUSY;
7237
7238         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7239                 /* no point adding a device */
7240                 return -EINVAL;
7241
7242         if (rdev->raid_disk >= 0)
7243                 first = last = rdev->raid_disk;
7244
7245         /*
7246          * find the disk ... but prefer rdev->saved_raid_disk
7247          * if possible.
7248          */
7249         if (rdev->saved_raid_disk >= 0 &&
7250             rdev->saved_raid_disk >= first &&
7251             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7252                 first = rdev->saved_raid_disk;
7253
7254         for (disk = first; disk <= last; disk++) {
7255                 p = conf->disks + disk;
7256                 if (p->rdev == NULL) {
7257                         clear_bit(In_sync, &rdev->flags);
7258                         rdev->raid_disk = disk;
7259                         err = 0;
7260                         if (rdev->saved_raid_disk != disk)
7261                                 conf->fullsync = 1;
7262                         rcu_assign_pointer(p->rdev, rdev);
7263                         goto out;
7264                 }
7265         }
7266         for (disk = first; disk <= last; disk++) {
7267                 p = conf->disks + disk;
7268                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7269                     p->replacement == NULL) {
7270                         clear_bit(In_sync, &rdev->flags);
7271                         set_bit(Replacement, &rdev->flags);
7272                         rdev->raid_disk = disk;
7273                         err = 0;
7274                         conf->fullsync = 1;
7275                         rcu_assign_pointer(p->replacement, rdev);
7276                         break;
7277                 }
7278         }
7279 out:
7280         print_raid5_conf(conf);
7281         return err;
7282 }
7283
7284 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7285 {
7286         /* no resync is happening, and there is enough space
7287          * on all devices, so we can resize.
7288          * We need to make sure resync covers any new space.
7289          * If the array is shrinking we should possibly wait until
7290          * any io in the removed space completes, but it hardly seems
7291          * worth it.
7292          */
7293         sector_t newsize;
7294         struct r5conf *conf = mddev->private;
7295
7296         if (conf->log)
7297                 return -EINVAL;
7298         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7299         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7300         if (mddev->external_size &&
7301             mddev->array_sectors > newsize)
7302                 return -EINVAL;
7303         if (mddev->bitmap) {
7304                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7305                 if (ret)
7306                         return ret;
7307         }
7308         md_set_array_sectors(mddev, newsize);
7309         set_capacity(mddev->gendisk, mddev->array_sectors);
7310         revalidate_disk(mddev->gendisk);
7311         if (sectors > mddev->dev_sectors &&
7312             mddev->recovery_cp > mddev->dev_sectors) {
7313                 mddev->recovery_cp = mddev->dev_sectors;
7314                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7315         }
7316         mddev->dev_sectors = sectors;
7317         mddev->resync_max_sectors = sectors;
7318         return 0;
7319 }
7320
7321 static int check_stripe_cache(struct mddev *mddev)
7322 {
7323         /* Can only proceed if there are plenty of stripe_heads.
7324          * We need a minimum of one full stripe,, and for sensible progress
7325          * it is best to have about 4 times that.
7326          * If we require 4 times, then the default 256 4K stripe_heads will
7327          * allow for chunk sizes up to 256K, which is probably OK.
7328          * If the chunk size is greater, user-space should request more
7329          * stripe_heads first.
7330          */
7331         struct r5conf *conf = mddev->private;
7332         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7333             > conf->min_nr_stripes ||
7334             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7335             > conf->min_nr_stripes) {
7336                 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7337                        mdname(mddev),
7338                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7339                         / STRIPE_SIZE)*4);
7340                 return 0;
7341         }
7342         return 1;
7343 }
7344
7345 static int check_reshape(struct mddev *mddev)
7346 {
7347         struct r5conf *conf = mddev->private;
7348
7349         if (conf->log)
7350                 return -EINVAL;
7351         if (mddev->delta_disks == 0 &&
7352             mddev->new_layout == mddev->layout &&
7353             mddev->new_chunk_sectors == mddev->chunk_sectors)
7354                 return 0; /* nothing to do */
7355         if (has_failed(conf))
7356                 return -EINVAL;
7357         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7358                 /* We might be able to shrink, but the devices must
7359                  * be made bigger first.
7360                  * For raid6, 4 is the minimum size.
7361                  * Otherwise 2 is the minimum
7362                  */
7363                 int min = 2;
7364                 if (mddev->level == 6)
7365                         min = 4;
7366                 if (mddev->raid_disks + mddev->delta_disks < min)
7367                         return -EINVAL;
7368         }
7369
7370         if (!check_stripe_cache(mddev))
7371                 return -ENOSPC;
7372
7373         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7374             mddev->delta_disks > 0)
7375                 if (resize_chunks(conf,
7376                                   conf->previous_raid_disks
7377                                   + max(0, mddev->delta_disks),
7378                                   max(mddev->new_chunk_sectors,
7379                                       mddev->chunk_sectors)
7380                             ) < 0)
7381                         return -ENOMEM;
7382         return resize_stripes(conf, (conf->previous_raid_disks
7383                                      + mddev->delta_disks));
7384 }
7385
7386 static int raid5_start_reshape(struct mddev *mddev)
7387 {
7388         struct r5conf *conf = mddev->private;
7389         struct md_rdev *rdev;
7390         int spares = 0;
7391         unsigned long flags;
7392
7393         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7394                 return -EBUSY;
7395
7396         if (!check_stripe_cache(mddev))
7397                 return -ENOSPC;
7398
7399         if (has_failed(conf))
7400                 return -EINVAL;
7401
7402         rdev_for_each(rdev, mddev) {
7403                 if (!test_bit(In_sync, &rdev->flags)
7404                     && !test_bit(Faulty, &rdev->flags))
7405                         spares++;
7406         }
7407
7408         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7409                 /* Not enough devices even to make a degraded array
7410                  * of that size
7411                  */
7412                 return -EINVAL;
7413
7414         /* Refuse to reduce size of the array.  Any reductions in
7415          * array size must be through explicit setting of array_size
7416          * attribute.
7417          */
7418         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7419             < mddev->array_sectors) {
7420                 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7421                        "before number of disks\n", mdname(mddev));
7422                 return -EINVAL;
7423         }
7424
7425         atomic_set(&conf->reshape_stripes, 0);
7426         spin_lock_irq(&conf->device_lock);
7427         write_seqcount_begin(&conf->gen_lock);
7428         conf->previous_raid_disks = conf->raid_disks;
7429         conf->raid_disks += mddev->delta_disks;
7430         conf->prev_chunk_sectors = conf->chunk_sectors;
7431         conf->chunk_sectors = mddev->new_chunk_sectors;
7432         conf->prev_algo = conf->algorithm;
7433         conf->algorithm = mddev->new_layout;
7434         conf->generation++;
7435         /* Code that selects data_offset needs to see the generation update
7436          * if reshape_progress has been set - so a memory barrier needed.
7437          */
7438         smp_mb();
7439         if (mddev->reshape_backwards)
7440                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7441         else
7442                 conf->reshape_progress = 0;
7443         conf->reshape_safe = conf->reshape_progress;
7444         write_seqcount_end(&conf->gen_lock);
7445         spin_unlock_irq(&conf->device_lock);
7446
7447         /* Now make sure any requests that proceeded on the assumption
7448          * the reshape wasn't running - like Discard or Read - have
7449          * completed.
7450          */
7451         mddev_suspend(mddev);
7452         mddev_resume(mddev);
7453
7454         /* Add some new drives, as many as will fit.
7455          * We know there are enough to make the newly sized array work.
7456          * Don't add devices if we are reducing the number of
7457          * devices in the array.  This is because it is not possible
7458          * to correctly record the "partially reconstructed" state of
7459          * such devices during the reshape and confusion could result.
7460          */
7461         if (mddev->delta_disks >= 0) {
7462                 rdev_for_each(rdev, mddev)
7463                         if (rdev->raid_disk < 0 &&
7464                             !test_bit(Faulty, &rdev->flags)) {
7465                                 if (raid5_add_disk(mddev, rdev) == 0) {
7466                                         if (rdev->raid_disk
7467                                             >= conf->previous_raid_disks)
7468                                                 set_bit(In_sync, &rdev->flags);
7469                                         else
7470                                                 rdev->recovery_offset = 0;
7471
7472                                         if (sysfs_link_rdev(mddev, rdev))
7473                                                 /* Failure here is OK */;
7474                                 }
7475                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7476                                    && !test_bit(Faulty, &rdev->flags)) {
7477                                 /* This is a spare that was manually added */
7478                                 set_bit(In_sync, &rdev->flags);
7479                         }
7480
7481                 /* When a reshape changes the number of devices,
7482                  * ->degraded is measured against the larger of the
7483                  * pre and post number of devices.
7484                  */
7485                 spin_lock_irqsave(&conf->device_lock, flags);
7486                 mddev->degraded = calc_degraded(conf);
7487                 spin_unlock_irqrestore(&conf->device_lock, flags);
7488         }
7489         mddev->raid_disks = conf->raid_disks;
7490         mddev->reshape_position = conf->reshape_progress;
7491         set_bit(MD_CHANGE_DEVS, &mddev->flags);
7492
7493         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7494         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7495         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7496         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7497         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7498         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7499                                                 "reshape");
7500         if (!mddev->sync_thread) {
7501                 mddev->recovery = 0;
7502                 spin_lock_irq(&conf->device_lock);
7503                 write_seqcount_begin(&conf->gen_lock);
7504                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7505                 mddev->new_chunk_sectors =
7506                         conf->chunk_sectors = conf->prev_chunk_sectors;
7507                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7508                 rdev_for_each(rdev, mddev)
7509                         rdev->new_data_offset = rdev->data_offset;
7510                 smp_wmb();
7511                 conf->generation --;
7512                 conf->reshape_progress = MaxSector;
7513                 mddev->reshape_position = MaxSector;
7514                 write_seqcount_end(&conf->gen_lock);
7515                 spin_unlock_irq(&conf->device_lock);
7516                 return -EAGAIN;
7517         }
7518         conf->reshape_checkpoint = jiffies;
7519         md_wakeup_thread(mddev->sync_thread);
7520         md_new_event(mddev);
7521         return 0;
7522 }
7523
7524 /* This is called from the reshape thread and should make any
7525  * changes needed in 'conf'
7526  */
7527 static void end_reshape(struct r5conf *conf)
7528 {
7529
7530         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7531                 struct md_rdev *rdev;
7532
7533                 spin_lock_irq(&conf->device_lock);
7534                 conf->previous_raid_disks = conf->raid_disks;
7535                 rdev_for_each(rdev, conf->mddev)
7536                         rdev->data_offset = rdev->new_data_offset;
7537                 smp_wmb();
7538                 conf->reshape_progress = MaxSector;
7539                 conf->mddev->reshape_position = MaxSector;
7540                 spin_unlock_irq(&conf->device_lock);
7541                 wake_up(&conf->wait_for_overlap);
7542
7543                 /* read-ahead size must cover two whole stripes, which is
7544                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7545                  */
7546                 if (conf->mddev->queue) {
7547                         int data_disks = conf->raid_disks - conf->max_degraded;
7548                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7549                                                    / PAGE_SIZE);
7550                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7551                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7552                 }
7553         }
7554 }
7555
7556 /* This is called from the raid5d thread with mddev_lock held.
7557  * It makes config changes to the device.
7558  */
7559 static void raid5_finish_reshape(struct mddev *mddev)
7560 {
7561         struct r5conf *conf = mddev->private;
7562
7563         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7564
7565                 if (mddev->delta_disks > 0) {
7566                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7567                         set_capacity(mddev->gendisk, mddev->array_sectors);
7568                         revalidate_disk(mddev->gendisk);
7569                 } else {
7570                         int d;
7571                         spin_lock_irq(&conf->device_lock);
7572                         mddev->degraded = calc_degraded(conf);
7573                         spin_unlock_irq(&conf->device_lock);
7574                         for (d = conf->raid_disks ;
7575                              d < conf->raid_disks - mddev->delta_disks;
7576                              d++) {
7577                                 struct md_rdev *rdev = conf->disks[d].rdev;
7578                                 if (rdev)
7579                                         clear_bit(In_sync, &rdev->flags);
7580                                 rdev = conf->disks[d].replacement;
7581                                 if (rdev)
7582                                         clear_bit(In_sync, &rdev->flags);
7583                         }
7584                 }
7585                 mddev->layout = conf->algorithm;
7586                 mddev->chunk_sectors = conf->chunk_sectors;
7587                 mddev->reshape_position = MaxSector;
7588                 mddev->delta_disks = 0;
7589                 mddev->reshape_backwards = 0;
7590         }
7591 }
7592
7593 static void raid5_quiesce(struct mddev *mddev, int state)
7594 {
7595         struct r5conf *conf = mddev->private;
7596
7597         switch(state) {
7598         case 2: /* resume for a suspend */
7599                 wake_up(&conf->wait_for_overlap);
7600                 break;
7601
7602         case 1: /* stop all writes */
7603                 lock_all_device_hash_locks_irq(conf);
7604                 /* '2' tells resync/reshape to pause so that all
7605                  * active stripes can drain
7606                  */
7607                 conf->quiesce = 2;
7608                 wait_event_cmd(conf->wait_for_quiescent,
7609                                     atomic_read(&conf->active_stripes) == 0 &&
7610                                     atomic_read(&conf->active_aligned_reads) == 0,
7611                                     unlock_all_device_hash_locks_irq(conf),
7612                                     lock_all_device_hash_locks_irq(conf));
7613                 conf->quiesce = 1;
7614                 unlock_all_device_hash_locks_irq(conf);
7615                 /* allow reshape to continue */
7616                 wake_up(&conf->wait_for_overlap);
7617                 break;
7618
7619         case 0: /* re-enable writes */
7620                 lock_all_device_hash_locks_irq(conf);
7621                 conf->quiesce = 0;
7622                 wake_up(&conf->wait_for_quiescent);
7623                 wake_up(&conf->wait_for_overlap);
7624                 unlock_all_device_hash_locks_irq(conf);
7625                 break;
7626         }
7627         r5l_quiesce(conf->log, state);
7628 }
7629
7630 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7631 {
7632         struct r0conf *raid0_conf = mddev->private;
7633         sector_t sectors;
7634
7635         /* for raid0 takeover only one zone is supported */
7636         if (raid0_conf->nr_strip_zones > 1) {
7637                 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7638                        mdname(mddev));
7639                 return ERR_PTR(-EINVAL);
7640         }
7641
7642         sectors = raid0_conf->strip_zone[0].zone_end;
7643         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7644         mddev->dev_sectors = sectors;
7645         mddev->new_level = level;
7646         mddev->new_layout = ALGORITHM_PARITY_N;
7647         mddev->new_chunk_sectors = mddev->chunk_sectors;
7648         mddev->raid_disks += 1;
7649         mddev->delta_disks = 1;
7650         /* make sure it will be not marked as dirty */
7651         mddev->recovery_cp = MaxSector;
7652
7653         return setup_conf(mddev);
7654 }
7655
7656 static void *raid5_takeover_raid1(struct mddev *mddev)
7657 {
7658         int chunksect;
7659
7660         if (mddev->raid_disks != 2 ||
7661             mddev->degraded > 1)
7662                 return ERR_PTR(-EINVAL);
7663
7664         /* Should check if there are write-behind devices? */
7665
7666         chunksect = 64*2; /* 64K by default */
7667
7668         /* The array must be an exact multiple of chunksize */
7669         while (chunksect && (mddev->array_sectors & (chunksect-1)))
7670                 chunksect >>= 1;
7671
7672         if ((chunksect<<9) < STRIPE_SIZE)
7673                 /* array size does not allow a suitable chunk size */
7674                 return ERR_PTR(-EINVAL);
7675
7676         mddev->new_level = 5;
7677         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7678         mddev->new_chunk_sectors = chunksect;
7679
7680         return setup_conf(mddev);
7681 }
7682
7683 static void *raid5_takeover_raid6(struct mddev *mddev)
7684 {
7685         int new_layout;
7686
7687         switch (mddev->layout) {
7688         case ALGORITHM_LEFT_ASYMMETRIC_6:
7689                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7690                 break;
7691         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7692                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7693                 break;
7694         case ALGORITHM_LEFT_SYMMETRIC_6:
7695                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7696                 break;
7697         case ALGORITHM_RIGHT_SYMMETRIC_6:
7698                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7699                 break;
7700         case ALGORITHM_PARITY_0_6:
7701                 new_layout = ALGORITHM_PARITY_0;
7702                 break;
7703         case ALGORITHM_PARITY_N:
7704                 new_layout = ALGORITHM_PARITY_N;
7705                 break;
7706         default:
7707                 return ERR_PTR(-EINVAL);
7708         }
7709         mddev->new_level = 5;
7710         mddev->new_layout = new_layout;
7711         mddev->delta_disks = -1;
7712         mddev->raid_disks -= 1;
7713         return setup_conf(mddev);
7714 }
7715
7716 static int raid5_check_reshape(struct mddev *mddev)
7717 {
7718         /* For a 2-drive array, the layout and chunk size can be changed
7719          * immediately as not restriping is needed.
7720          * For larger arrays we record the new value - after validation
7721          * to be used by a reshape pass.
7722          */
7723         struct r5conf *conf = mddev->private;
7724         int new_chunk = mddev->new_chunk_sectors;
7725
7726         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7727                 return -EINVAL;
7728         if (new_chunk > 0) {
7729                 if (!is_power_of_2(new_chunk))
7730                         return -EINVAL;
7731                 if (new_chunk < (PAGE_SIZE>>9))
7732                         return -EINVAL;
7733                 if (mddev->array_sectors & (new_chunk-1))
7734                         /* not factor of array size */
7735                         return -EINVAL;
7736         }
7737
7738         /* They look valid */
7739
7740         if (mddev->raid_disks == 2) {
7741                 /* can make the change immediately */
7742                 if (mddev->new_layout >= 0) {
7743                         conf->algorithm = mddev->new_layout;
7744                         mddev->layout = mddev->new_layout;
7745                 }
7746                 if (new_chunk > 0) {
7747                         conf->chunk_sectors = new_chunk ;
7748                         mddev->chunk_sectors = new_chunk;
7749                 }
7750                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7751                 md_wakeup_thread(mddev->thread);
7752         }
7753         return check_reshape(mddev);
7754 }
7755
7756 static int raid6_check_reshape(struct mddev *mddev)
7757 {
7758         int new_chunk = mddev->new_chunk_sectors;
7759
7760         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7761                 return -EINVAL;
7762         if (new_chunk > 0) {
7763                 if (!is_power_of_2(new_chunk))
7764                         return -EINVAL;
7765                 if (new_chunk < (PAGE_SIZE >> 9))
7766                         return -EINVAL;
7767                 if (mddev->array_sectors & (new_chunk-1))
7768                         /* not factor of array size */
7769                         return -EINVAL;
7770         }
7771
7772         /* They look valid */
7773         return check_reshape(mddev);
7774 }
7775
7776 static void *raid5_takeover(struct mddev *mddev)
7777 {
7778         /* raid5 can take over:
7779          *  raid0 - if there is only one strip zone - make it a raid4 layout
7780          *  raid1 - if there are two drives.  We need to know the chunk size
7781          *  raid4 - trivial - just use a raid4 layout.
7782          *  raid6 - Providing it is a *_6 layout
7783          */
7784         if (mddev->level == 0)
7785                 return raid45_takeover_raid0(mddev, 5);
7786         if (mddev->level == 1)
7787                 return raid5_takeover_raid1(mddev);
7788         if (mddev->level == 4) {
7789                 mddev->new_layout = ALGORITHM_PARITY_N;
7790                 mddev->new_level = 5;
7791                 return setup_conf(mddev);
7792         }
7793         if (mddev->level == 6)
7794                 return raid5_takeover_raid6(mddev);
7795
7796         return ERR_PTR(-EINVAL);
7797 }
7798
7799 static void *raid4_takeover(struct mddev *mddev)
7800 {
7801         /* raid4 can take over:
7802          *  raid0 - if there is only one strip zone
7803          *  raid5 - if layout is right
7804          */
7805         if (mddev->level == 0)
7806                 return raid45_takeover_raid0(mddev, 4);
7807         if (mddev->level == 5 &&
7808             mddev->layout == ALGORITHM_PARITY_N) {
7809                 mddev->new_layout = 0;
7810                 mddev->new_level = 4;
7811                 return setup_conf(mddev);
7812         }
7813         return ERR_PTR(-EINVAL);
7814 }
7815
7816 static struct md_personality raid5_personality;
7817
7818 static void *raid6_takeover(struct mddev *mddev)
7819 {
7820         /* Currently can only take over a raid5.  We map the
7821          * personality to an equivalent raid6 personality
7822          * with the Q block at the end.
7823          */
7824         int new_layout;
7825
7826         if (mddev->pers != &raid5_personality)
7827                 return ERR_PTR(-EINVAL);
7828         if (mddev->degraded > 1)
7829                 return ERR_PTR(-EINVAL);
7830         if (mddev->raid_disks > 253)
7831                 return ERR_PTR(-EINVAL);
7832         if (mddev->raid_disks < 3)
7833                 return ERR_PTR(-EINVAL);
7834
7835         switch (mddev->layout) {
7836         case ALGORITHM_LEFT_ASYMMETRIC:
7837                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7838                 break;
7839         case ALGORITHM_RIGHT_ASYMMETRIC:
7840                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7841                 break;
7842         case ALGORITHM_LEFT_SYMMETRIC:
7843                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7844                 break;
7845         case ALGORITHM_RIGHT_SYMMETRIC:
7846                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7847                 break;
7848         case ALGORITHM_PARITY_0:
7849                 new_layout = ALGORITHM_PARITY_0_6;
7850                 break;
7851         case ALGORITHM_PARITY_N:
7852                 new_layout = ALGORITHM_PARITY_N;
7853                 break;
7854         default:
7855                 return ERR_PTR(-EINVAL);
7856         }
7857         mddev->new_level = 6;
7858         mddev->new_layout = new_layout;
7859         mddev->delta_disks = 1;
7860         mddev->raid_disks += 1;
7861         return setup_conf(mddev);
7862 }
7863
7864 static struct md_personality raid6_personality =
7865 {
7866         .name           = "raid6",
7867         .level          = 6,
7868         .owner          = THIS_MODULE,
7869         .make_request   = make_request,
7870         .run            = run,
7871         .free           = raid5_free,
7872         .status         = status,
7873         .error_handler  = error,
7874         .hot_add_disk   = raid5_add_disk,
7875         .hot_remove_disk= raid5_remove_disk,
7876         .spare_active   = raid5_spare_active,
7877         .sync_request   = sync_request,
7878         .resize         = raid5_resize,
7879         .size           = raid5_size,
7880         .check_reshape  = raid6_check_reshape,
7881         .start_reshape  = raid5_start_reshape,
7882         .finish_reshape = raid5_finish_reshape,
7883         .quiesce        = raid5_quiesce,
7884         .takeover       = raid6_takeover,
7885         .congested      = raid5_congested,
7886 };
7887 static struct md_personality raid5_personality =
7888 {
7889         .name           = "raid5",
7890         .level          = 5,
7891         .owner          = THIS_MODULE,
7892         .make_request   = make_request,
7893         .run            = run,
7894         .free           = raid5_free,
7895         .status         = status,
7896         .error_handler  = error,
7897         .hot_add_disk   = raid5_add_disk,
7898         .hot_remove_disk= raid5_remove_disk,
7899         .spare_active   = raid5_spare_active,
7900         .sync_request   = sync_request,
7901         .resize         = raid5_resize,
7902         .size           = raid5_size,
7903         .check_reshape  = raid5_check_reshape,
7904         .start_reshape  = raid5_start_reshape,
7905         .finish_reshape = raid5_finish_reshape,
7906         .quiesce        = raid5_quiesce,
7907         .takeover       = raid5_takeover,
7908         .congested      = raid5_congested,
7909 };
7910
7911 static struct md_personality raid4_personality =
7912 {
7913         .name           = "raid4",
7914         .level          = 4,
7915         .owner          = THIS_MODULE,
7916         .make_request   = make_request,
7917         .run            = run,
7918         .free           = raid5_free,
7919         .status         = status,
7920         .error_handler  = error,
7921         .hot_add_disk   = raid5_add_disk,
7922         .hot_remove_disk= raid5_remove_disk,
7923         .spare_active   = raid5_spare_active,
7924         .sync_request   = sync_request,
7925         .resize         = raid5_resize,
7926         .size           = raid5_size,
7927         .check_reshape  = raid5_check_reshape,
7928         .start_reshape  = raid5_start_reshape,
7929         .finish_reshape = raid5_finish_reshape,
7930         .quiesce        = raid5_quiesce,
7931         .takeover       = raid4_takeover,
7932         .congested      = raid5_congested,
7933 };
7934
7935 static int __init raid5_init(void)
7936 {
7937         raid5_wq = alloc_workqueue("raid5wq",
7938                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7939         if (!raid5_wq)
7940                 return -ENOMEM;
7941         register_md_personality(&raid6_personality);
7942         register_md_personality(&raid5_personality);
7943         register_md_personality(&raid4_personality);
7944         return 0;
7945 }
7946
7947 static void raid5_exit(void)
7948 {
7949         unregister_md_personality(&raid6_personality);
7950         unregister_md_personality(&raid5_personality);
7951         unregister_md_personality(&raid4_personality);
7952         destroy_workqueue(raid5_wq);
7953 }
7954
7955 module_init(raid5_init);
7956 module_exit(raid5_exit);
7957 MODULE_LICENSE("GPL");
7958 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7959 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7960 MODULE_ALIAS("md-raid5");
7961 MODULE_ALIAS("md-raid4");
7962 MODULE_ALIAS("md-level-5");
7963 MODULE_ALIAS("md-level-4");
7964 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7965 MODULE_ALIAS("md-raid6");
7966 MODULE_ALIAS("md-level-6");
7967
7968 /* This used to be two separate modules, they were: */
7969 MODULE_ALIAS("raid5");
7970 MODULE_ALIAS("raid6");