kernel/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 *return_bi)
 227 {
 228         struct bio *bi = return_bi;
 229         while (bi) {
 230
 231                 return_bi = bi->bi_next;
 232                 bi->bi_next = NULL;
 233                 bi->bi_iter.bi_size = 0;
 234                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
 235                                          bi, 0);
 236                 bio_endio(bi, 0);
 237                 bi = return_bi;
 238         }
 239 }
 240
 241 static void print_raid5_conf (struct r5conf *conf);
 242
 243 static int stripe_operations_active(struct stripe_head *sh)
 244 {
 245         return sh->check_state || sh->reconstruct_state ||
 246                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 247                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 248 }
 249
 250 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 251 {
 252         struct r5conf *conf = sh->raid_conf;
 253         struct r5worker_group *group;
 254         int thread_cnt;
 255         int i, cpu = sh->cpu;
 256
 257         if (!cpu_online(cpu)) {
 258                 cpu = cpumask_any(cpu_online_mask);
 259                 sh->cpu = cpu;
 260         }
 261
 262         if (list_empty(&sh->lru)) {
 263                 struct r5worker_group *group;
 264                 group = conf->worker_groups + cpu_to_group(cpu);
 265                 list_add_tail(&sh->lru, &group->handle_list);
 266                 group->stripes_cnt++;
 267                 sh->group = group;
 268         }
 269
 270         if (conf->worker_cnt_per_group == 0) {
 271                 md_wakeup_thread(conf->mddev->thread);
 272                 return;
 273         }
 274
 275         group = conf->worker_groups + cpu_to_group(sh->cpu);
 276
 277         group->workers[0].working = true;
 278         /* at least one worker should run to avoid race */
 279         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 280
 281         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 282         /* wakeup more workers */
 283         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 284                 if (group->workers[i].working == false) {
 285                         group->workers[i].working = true;
 286                         queue_work_on(sh->cpu, raid5_wq,
 287                                       &group->workers[i].work);
 288                         thread_cnt--;
 289                 }
 290         }
 291 }
 292
 293 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 294                               struct list_head *temp_inactive_list)
 295 {
 296         BUG_ON(!list_empty(&sh->lru));
 297         BUG_ON(atomic_read(&conf->active_stripes)==0);
 298         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 299                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 300                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 301                         list_add_tail(&sh->lru, &conf->delayed_list);
 302                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 303                            sh->bm_seq - conf->seq_write > 0)
 304                         list_add_tail(&sh->lru, &conf->bitmap_list);
 305                 else {
 306                         clear_bit(STRIPE_DELAYED, &sh->state);
 307                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 308                         if (conf->worker_cnt_per_group == 0) {
 309                                 list_add_tail(&sh->lru, &conf->handle_list);
 310                         } else {
 311                                 raid5_wakeup_stripe_thread(sh);
 312                                 return;
 313                         }
 314                 }
 315                 md_wakeup_thread(conf->mddev->thread);
 316         } else {
 317                 BUG_ON(stripe_operations_active(sh));
 318                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 319                         if (atomic_dec_return(&conf->preread_active_stripes)
 320                             < IO_THRESHOLD)
 321                                 md_wakeup_thread(conf->mddev->thread);
 322                 atomic_dec(&conf->active_stripes);
 323                 if (!test_bit(STRIPE_EXPANDING, &sh->state))
 324                         list_add_tail(&sh->lru, temp_inactive_list);
 325         }
 326 }
 327
 328 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 329                              struct list_head *temp_inactive_list)
 330 {
 331         if (atomic_dec_and_test(&sh->count))
 332                 do_release_stripe(conf, sh, temp_inactive_list);
 333 }
 334
 335 /*
 336  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 337  *
 338  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 339  * given time. Adding stripes only takes device lock, while deleting stripes
 340  * only takes hash lock.
 341  */
 342 static void release_inactive_stripe_list(struct r5conf *conf,
 343                                          struct list_head *temp_inactive_list,
 344                                          int hash)
 345 {
 346         int size;
 347         bool do_wakeup = false;
 348         unsigned long flags;
 349
 350         if (hash == NR_STRIPE_HASH_LOCKS) {
 351                 size = NR_STRIPE_HASH_LOCKS;
 352                 hash = NR_STRIPE_HASH_LOCKS - 1;
 353         } else
 354                 size = 1;
 355         while (size) {
 356                 struct list_head *list = &temp_inactive_list[size - 1];
 357
 358                 /*
 359                  * We don't hold any lock here yet, get_active_stripe() might
 360                  * remove stripes from the list
 361                  */
 362                 if (!list_empty_careful(list)) {
 363                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 364                         if (list_empty(conf->inactive_list + hash) &&
 365                             !list_empty(list))
 366                                 atomic_dec(&conf->empty_inactive_list_nr);
 367                         list_splice_tail_init(list, conf->inactive_list + hash);
 368                         do_wakeup = true;
 369                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 370                 }
 371                 size--;
 372                 hash--;
 373         }
 374
 375         if (do_wakeup) {
 376                 wake_up(&conf->wait_for_stripe);
 377                 if (conf->retry_read_aligned)
 378                         md_wakeup_thread(conf->mddev->thread);
 379         }
 380 }
 381
 382 /* should hold conf->device_lock already */
 383 static int release_stripe_list(struct r5conf *conf,
 384                                struct list_head *temp_inactive_list)
 385 {
 386         struct stripe_head *sh;
 387         int count = 0;
 388         struct llist_node *head;
 389
 390         head = llist_del_all(&conf->released_stripes);
 391         head = llist_reverse_order(head);
 392         while (head) {
 393                 int hash;
 394
 395                 sh = llist_entry(head, struct stripe_head, release_list);
 396                 head = llist_next(head);
 397                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 398                 smp_mb();
 399                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 400                 /*
 401                  * Don't worry the bit is set here, because if the bit is set
 402                  * again, the count is always > 1. This is true for
 403                  * STRIPE_ON_UNPLUG_LIST bit too.
 404                  */
 405                 hash = sh->hash_lock_index;
 406                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 407                 count++;
 408         }
 409
 410         return count;
 411 }
 412
 413 static void release_stripe(struct stripe_head *sh)
 414 {
 415         struct r5conf *conf = sh->raid_conf;
 416         unsigned long flags;
 417         struct list_head list;
 418         int hash;
 419         bool wakeup;
 420
 421         /* Avoid release_list until the last reference.
 422          */
 423         if (atomic_add_unless(&sh->count, -1, 1))
 424                 return;
 425
 426         if (unlikely(!conf->mddev->thread) ||
 427                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 428                 goto slow_path;
 429         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 430         if (wakeup)
 431                 md_wakeup_thread(conf->mddev->thread);
 432         return;
 433 slow_path:
 434         local_irq_save(flags);
 435         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 436         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 437                 INIT_LIST_HEAD(&list);
 438                 hash = sh->hash_lock_index;
 439                 do_release_stripe(conf, sh, &list);
 440                 spin_unlock(&conf->device_lock);
 441                 release_inactive_stripe_list(conf, &list, hash);
 442         }
 443         local_irq_restore(flags);
 444 }
 445
 446 static inline void remove_hash(struct stripe_head *sh)
 447 {
 448         pr_debug("remove_hash(), stripe %llu\n",
 449                 (unsigned long long)sh->sector);
 450
 451         hlist_del_init(&sh->hash);
 452 }
 453
 454 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 455 {
 456         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 457
 458         pr_debug("insert_hash(), stripe %llu\n",
 459                 (unsigned long long)sh->sector);
 460
 461         hlist_add_head(&sh->hash, hp);
 462 }
 463
 464 /* find an idle stripe, make sure it is unhashed, and return it. */
 465 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 466 {
 467         struct stripe_head *sh = NULL;
 468         struct list_head *first;
 469
 470         if (list_empty(conf->inactive_list + hash))
 471                 goto out;
 472         first = (conf->inactive_list + hash)->next;
 473         sh = list_entry(first, struct stripe_head, lru);
 474         list_del_init(first);
 475         remove_hash(sh);
 476         atomic_inc(&conf->active_stripes);
 477         BUG_ON(hash != sh->hash_lock_index);
 478         if (list_empty(conf->inactive_list + hash))
 479                 atomic_inc(&conf->empty_inactive_list_nr);
 480 out:
 481         return sh;
 482 }
 483
 484 static void shrink_buffers(struct stripe_head *sh)
 485 {
 486         struct page *p;
 487         int i;
 488         int num = sh->raid_conf->pool_size;
 489
 490         for (i = 0; i < num ; i++) {
 491                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 492                 p = sh->dev[i].page;
 493                 if (!p)
 494                         continue;
 495                 sh->dev[i].page = NULL;
 496                 put_page(p);
 497         }
 498 }
 499
 500 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 501 {
 502         int i;
 503         int num = sh->raid_conf->pool_size;
 504
 505         for (i = 0; i < num; i++) {
 506                 struct page *page;
 507
 508                 if (!(page = alloc_page(gfp))) {
 509                         return 1;
 510                 }
 511                 sh->dev[i].page = page;
 512                 sh->dev[i].orig_page = page;
 513         }
 514         return 0;
 515 }
 516
 517 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 518 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 519                             struct stripe_head *sh);
 520
 521 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 522 {
 523         struct r5conf *conf = sh->raid_conf;
 524         int i, seq;
 525
 526         BUG_ON(atomic_read(&sh->count) != 0);
 527         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 528         BUG_ON(stripe_operations_active(sh));
 529         BUG_ON(sh->batch_head);
 530
 531         pr_debug("init_stripe called, stripe %llu\n",
 532                 (unsigned long long)sector);
 533 retry:
 534         seq = read_seqcount_begin(&conf->gen_lock);
 535         sh->generation = conf->generation - previous;
 536         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 537         sh->sector = sector;
 538         stripe_set_idx(sector, conf, previous, sh);
 539         sh->state = 0;
 540
 541         for (i = sh->disks; i--; ) {
 542                 struct r5dev *dev = &sh->dev[i];
 543
 544                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 545                     test_bit(R5_LOCKED, &dev->flags)) {
 546                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 547                                (unsigned long long)sh->sector, i, dev->toread,
 548                                dev->read, dev->towrite, dev->written,
 549                                test_bit(R5_LOCKED, &dev->flags));
 550                         WARN_ON(1);
 551                 }
 552                 dev->flags = 0;
 553                 raid5_build_block(sh, i, previous);
 554         }
 555         if (read_seqcount_retry(&conf->gen_lock, seq))
 556                 goto retry;
 557         sh->overwrite_disks = 0;
 558         insert_hash(conf, sh);
 559         sh->cpu = smp_processor_id();
 560         set_bit(STRIPE_BATCH_READY, &sh->state);
 561 }
 562
 563 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 564                                          short generation)
 565 {
 566         struct stripe_head *sh;
 567
 568         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 569         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 570                 if (sh->sector == sector && sh->generation == generation)
 571                         return sh;
 572         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 573         return NULL;
 574 }
 575
 576 /*
 577  * Need to check if array has failed when deciding whether to:
 578  *  - start an array
 579  *  - remove non-faulty devices
 580  *  - add a spare
 581  *  - allow a reshape
 582  * This determination is simple when no reshape is happening.
 583  * However if there is a reshape, we need to carefully check
 584  * both the before and after sections.
 585  * This is because some failed devices may only affect one
 586  * of the two sections, and some non-in_sync devices may
 587  * be insync in the section most affected by failed devices.
 588  */
 589 static int calc_degraded(struct r5conf *conf)
 590 {
 591         int degraded, degraded2;
 592         int i;
 593
 594         rcu_read_lock();
 595         degraded = 0;
 596         for (i = 0; i < conf->previous_raid_disks; i++) {
 597                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 598                 if (rdev && test_bit(Faulty, &rdev->flags))
 599                         rdev = rcu_dereference(conf->disks[i].replacement);
 600                 if (!rdev || test_bit(Faulty, &rdev->flags))
 601                         degraded++;
 602                 else if (test_bit(In_sync, &rdev->flags))
 603                         ;
 604                 else
 605                         /* not in-sync or faulty.
 606                          * If the reshape increases the number of devices,
 607                          * this is being recovered by the reshape, so
 608                          * this 'previous' section is not in_sync.
 609                          * If the number of devices is being reduced however,
 610                          * the device can only be part of the array if
 611                          * we are reverting a reshape, so this section will
 612                          * be in-sync.
 613                          */
 614                         if (conf->raid_disks >= conf->previous_raid_disks)
 615                                 degraded++;
 616         }
 617         rcu_read_unlock();
 618         if (conf->raid_disks == conf->previous_raid_disks)
 619                 return degraded;
 620         rcu_read_lock();
 621         degraded2 = 0;
 622         for (i = 0; i < conf->raid_disks; i++) {
 623                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 624                 if (rdev && test_bit(Faulty, &rdev->flags))
 625                         rdev = rcu_dereference(conf->disks[i].replacement);
 626                 if (!rdev || test_bit(Faulty, &rdev->flags))
 627                         degraded2++;
 628                 else if (test_bit(In_sync, &rdev->flags))
 629                         ;
 630                 else
 631                         /* not in-sync or faulty.
 632                          * If reshape increases the number of devices, this
 633                          * section has already been recovered, else it
 634                          * almost certainly hasn't.
 635                          */
 636                         if (conf->raid_disks <= conf->previous_raid_disks)
 637                                 degraded2++;
 638         }
 639         rcu_read_unlock();
 640         if (degraded2 > degraded)
 641                 return degraded2;
 642         return degraded;
 643 }
 644
 645 static int has_failed(struct r5conf *conf)
 646 {
 647         int degraded;
 648
 649         if (conf->mddev->reshape_position == MaxSector)
 650                 return conf->mddev->degraded > conf->max_degraded;
 651
 652         degraded = calc_degraded(conf);
 653         if (degraded > conf->max_degraded)
 654                 return 1;
 655         return 0;
 656 }
 657
 658 static struct stripe_head *
 659 get_active_stripe(struct r5conf *conf, sector_t sector,
 660                   int previous, int noblock, int noquiesce)
 661 {
 662         struct stripe_head *sh;
 663         int hash = stripe_hash_locks_hash(sector);
 664
 665         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 666
 667         spin_lock_irq(conf->hash_locks + hash);
 668
 669         do {
 670                 wait_event_lock_irq(conf->wait_for_stripe,
 671                                     conf->quiesce == 0 || noquiesce,
 672                                     *(conf->hash_locks + hash));
 673                 sh = __find_stripe(conf, sector, conf->generation - previous);
 674                 if (!sh) {
 675                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 676                                 sh = get_free_stripe(conf, hash);
 677                                 if (!sh && llist_empty(&conf->released_stripes) &&
 678                                     !test_bit(R5_DID_ALLOC, &conf->cache_state))
 679                                         set_bit(R5_ALLOC_MORE,
 680                                                 &conf->cache_state);
 681                         }
 682                         if (noblock && sh == NULL)
 683                                 break;
 684                         if (!sh) {
 685                                 set_bit(R5_INACTIVE_BLOCKED,
 686                                         &conf->cache_state);
 687                                 wait_event_lock_irq(
 688                                         conf->wait_for_stripe,
 689                                         !list_empty(conf->inactive_list + hash) &&
 690                                         (atomic_read(&conf->active_stripes)
 691                                          < (conf->max_nr_stripes * 3 / 4)
 692                                          || !test_bit(R5_INACTIVE_BLOCKED,
 693                                                       &conf->cache_state)),
 694                                         *(conf->hash_locks + hash));
 695                                 clear_bit(R5_INACTIVE_BLOCKED,
 696                                           &conf->cache_state);
 697                         } else {
 698                                 init_stripe(sh, sector, previous);
 699                                 atomic_inc(&sh->count);
 700                         }
 701                 } else if (!atomic_inc_not_zero(&sh->count)) {
 702                         spin_lock(&conf->device_lock);
 703                         if (!atomic_read(&sh->count)) {
 704                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 705                                         atomic_inc(&conf->active_stripes);
 706                                 BUG_ON(list_empty(&sh->lru) &&
 707                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 708                                 list_del_init(&sh->lru);
 709                                 if (sh->group) {
 710                                         sh->group->stripes_cnt--;
 711                                         sh->group = NULL;
 712                                 }
 713                         }
 714                         atomic_inc(&sh->count);
 715                         spin_unlock(&conf->device_lock);
 716                 }
 717         } while (sh == NULL);
 718
 719         spin_unlock_irq(conf->hash_locks + hash);
 720         return sh;
 721 }
 722
 723 static bool is_full_stripe_write(struct stripe_head *sh)
 724 {
 725         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 726         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 727 }
 728
 729 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 730 {
 731         local_irq_disable();
 732         if (sh1 > sh2) {
 733                 spin_lock(&sh2->stripe_lock);
 734                 spin_lock_nested(&sh1->stripe_lock, 1);
 735         } else {
 736                 spin_lock(&sh1->stripe_lock);
 737                 spin_lock_nested(&sh2->stripe_lock, 1);
 738         }
 739 }
 740
 741 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 742 {
 743         spin_unlock(&sh1->stripe_lock);
 744         spin_unlock(&sh2->stripe_lock);
 745         local_irq_enable();
 746 }
 747
 748 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 749 static bool stripe_can_batch(struct stripe_head *sh)
 750 {
 751         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 752                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 753                 is_full_stripe_write(sh);
 754 }
 755
 756 /* we only do back search */
 757 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 758 {
 759         struct stripe_head *head;
 760         sector_t head_sector, tmp_sec;
 761         int hash;
 762         int dd_idx;
 763
 764         if (!stripe_can_batch(sh))
 765                 return;
 766         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 767         tmp_sec = sh->sector;
 768         if (!sector_div(tmp_sec, conf->chunk_sectors))
 769                 return;
 770         head_sector = sh->sector - STRIPE_SECTORS;
 771
 772         hash = stripe_hash_locks_hash(head_sector);
 773         spin_lock_irq(conf->hash_locks + hash);
 774         head = __find_stripe(conf, head_sector, conf->generation);
 775         if (head && !atomic_inc_not_zero(&head->count)) {
 776                 spin_lock(&conf->device_lock);
 777                 if (!atomic_read(&head->count)) {
 778                         if (!test_bit(STRIPE_HANDLE, &head->state))
 779                                 atomic_inc(&conf->active_stripes);
 780                         BUG_ON(list_empty(&head->lru) &&
 781                                !test_bit(STRIPE_EXPANDING, &head->state));
 782                         list_del_init(&head->lru);
 783                         if (head->group) {
 784                                 head->group->stripes_cnt--;
 785                                 head->group = NULL;
 786                         }
 787                 }
 788                 atomic_inc(&head->count);
 789                 spin_unlock(&conf->device_lock);
 790         }
 791         spin_unlock_irq(conf->hash_locks + hash);
 792
 793         if (!head)
 794                 return;
 795         if (!stripe_can_batch(head))
 796                 goto out;
 797
 798         lock_two_stripes(head, sh);
 799         /* clear_batch_ready clear the flag */
 800         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 801                 goto unlock_out;
 802
 803         if (sh->batch_head)
 804                 goto unlock_out;
 805
 806         dd_idx = 0;
 807         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 808                 dd_idx++;
 809         if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
 810                 goto unlock_out;
 811
 812         if (head->batch_head) {
 813                 spin_lock(&head->batch_head->batch_lock);
 814                 /* This batch list is already running */
 815                 if (!stripe_can_batch(head)) {
 816                         spin_unlock(&head->batch_head->batch_lock);
 817                         goto unlock_out;
 818                 }
 819
 820                 /*
 821                  * at this point, head's BATCH_READY could be cleared, but we
 822                  * can still add the stripe to batch list
 823                  */
 824                 list_add(&sh->batch_list, &head->batch_list);
 825                 spin_unlock(&head->batch_head->batch_lock);
 826
 827                 sh->batch_head = head->batch_head;
 828         } else {
 829                 head->batch_head = head;
 830                 sh->batch_head = head->batch_head;
 831                 spin_lock(&head->batch_lock);
 832                 list_add_tail(&sh->batch_list, &head->batch_list);
 833                 spin_unlock(&head->batch_lock);
 834         }
 835
 836         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 837                 if (atomic_dec_return(&conf->preread_active_stripes)
 838                     < IO_THRESHOLD)
 839                         md_wakeup_thread(conf->mddev->thread);
 840
 841         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 842                 int seq = sh->bm_seq;
 843                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 844                     sh->batch_head->bm_seq > seq)
 845                         seq = sh->batch_head->bm_seq;
 846                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 847                 sh->batch_head->bm_seq = seq;
 848         }
 849
 850         atomic_inc(&sh->count);
 851 unlock_out:
 852         unlock_two_stripes(head, sh);
 853 out:
 854         release_stripe(head);
 855 }
 856
 857 /* Determine if 'data_offset' or 'new_data_offset' should be used
 858  * in this stripe_head.
 859  */
 860 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 861 {
 862         sector_t progress = conf->reshape_progress;
 863         /* Need a memory barrier to make sure we see the value
 864          * of conf->generation, or ->data_offset that was set before
 865          * reshape_progress was updated.
 866          */
 867         smp_rmb();
 868         if (progress == MaxSector)
 869                 return 0;
 870         if (sh->generation == conf->generation - 1)
 871                 return 0;
 872         /* We are in a reshape, and this is a new-generation stripe,
 873          * so use new_data_offset.
 874          */
 875         return 1;
 876 }
 877
 878 static void
 879 raid5_end_read_request(struct bio *bi, int error);
 880 static void
 881 raid5_end_write_request(struct bio *bi, int error);
 882
 883 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 884 {
 885         struct r5conf *conf = sh->raid_conf;
 886         int i, disks = sh->disks;
 887         struct stripe_head *head_sh = sh;
 888
 889         might_sleep();
 890
 891         for (i = disks; i--; ) {
 892                 int rw;
 893                 int replace_only = 0;
 894                 struct bio *bi, *rbi;
 895                 struct md_rdev *rdev, *rrdev = NULL;
 896
 897                 sh = head_sh;
 898                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 899                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
 900                                 rw = WRITE_FUA;
 901                         else
 902                                 rw = WRITE;
 903                         if (test_bit(R5_Discard, &sh->dev[i].flags))
 904                                 rw |= REQ_DISCARD;
 905                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 906                         rw = READ;
 907                 else if (test_and_clear_bit(R5_WantReplace,
 908                                             &sh->dev[i].flags)) {
 909                         rw = WRITE;
 910                         replace_only = 1;
 911                 } else
 912                         continue;
 913                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
 914                         rw |= REQ_SYNC;
 915
 916 again:
 917                 bi = &sh->dev[i].req;
 918                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
 919
 920                 rcu_read_lock();
 921                 rrdev = rcu_dereference(conf->disks[i].replacement);
 922                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
 923                 rdev = rcu_dereference(conf->disks[i].rdev);
 924                 if (!rdev) {
 925                         rdev = rrdev;
 926                         rrdev = NULL;
 927                 }
 928                 if (rw & WRITE) {
 929                         if (replace_only)
 930                                 rdev = NULL;
 931                         if (rdev == rrdev)
 932                                 /* We raced and saw duplicates */
 933                                 rrdev = NULL;
 934                 } else {
 935                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
 936                                 rdev = rrdev;
 937                         rrdev = NULL;
 938                 }
 939
 940                 if (rdev && test_bit(Faulty, &rdev->flags))
 941                         rdev = NULL;
 942                 if (rdev)
 943                         atomic_inc(&rdev->nr_pending);
 944                 if (rrdev && test_bit(Faulty, &rrdev->flags))
 945                         rrdev = NULL;
 946                 if (rrdev)
 947                         atomic_inc(&rrdev->nr_pending);
 948                 rcu_read_unlock();
 949
 950                 /* We have already checked bad blocks for reads.  Now
 951                  * need to check for writes.  We never accept write errors
 952                  * on the replacement, so we don't to check rrdev.
 953                  */
 954                 while ((rw & WRITE) && rdev &&
 955                        test_bit(WriteErrorSeen, &rdev->flags)) {
 956                         sector_t first_bad;
 957                         int bad_sectors;
 958                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
 959                                               &first_bad, &bad_sectors);
 960                         if (!bad)
 961                                 break;
 962
 963                         if (bad < 0) {
 964                                 set_bit(BlockedBadBlocks, &rdev->flags);
 965                                 if (!conf->mddev->external &&
 966                                     conf->mddev->flags) {
 967                                         /* It is very unlikely, but we might
 968                                          * still need to write out the
 969                                          * bad block log - better give it
 970                                          * a chance*/
 971                                         md_check_recovery(conf->mddev);
 972                                 }
 973                                 /*
 974                                  * Because md_wait_for_blocked_rdev
 975                                  * will dec nr_pending, we must
 976                                  * increment it first.
 977                                  */
 978                                 atomic_inc(&rdev->nr_pending);
 979                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
 980                         } else {
 981                                 /* Acknowledged bad block - skip the write */
 982                                 rdev_dec_pending(rdev, conf->mddev);
 983                                 rdev = NULL;
 984                         }
 985                 }
 986
 987                 if (rdev) {
 988                         if (s->syncing || s->expanding || s->expanded
 989                             || s->replacing)
 990                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
 991
 992                         set_bit(STRIPE_IO_STARTED, &sh->state);
 993
 994                         bio_reset(bi);
 995                         bi->bi_bdev = rdev->bdev;
 996                         bi->bi_rw = rw;
 997                         bi->bi_end_io = (rw & WRITE)
 998                                 ? raid5_end_write_request
 999                                 : raid5_end_read_request;
1000                         bi->bi_private = sh;
1001
1002                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1003                                 __func__, (unsigned long long)sh->sector,
1004                                 bi->bi_rw, i);
1005                         atomic_inc(&sh->count);
1006                         if (sh != head_sh)
1007                                 atomic_inc(&head_sh->count);
1008                         if (use_new_offset(conf, sh))
1009                                 bi->bi_iter.bi_sector = (sh->sector
1010                                                  + rdev->new_data_offset);
1011                         else
1012                                 bi->bi_iter.bi_sector = (sh->sector
1013                                                  + rdev->data_offset);
1014                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1015                                 bi->bi_rw |= REQ_NOMERGE;
1016
1017                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1018                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1019                         sh->dev[i].vec.bv_page = sh->dev[i].page;
1020                         bi->bi_vcnt = 1;
1021                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1022                         bi->bi_io_vec[0].bv_offset = 0;
1023                         bi->bi_iter.bi_size = STRIPE_SIZE;
1024                         /*
1025                          * If this is discard request, set bi_vcnt 0. We don't
1026                          * want to confuse SCSI because SCSI will replace payload
1027                          */
1028                         if (rw & REQ_DISCARD)
1029                                 bi->bi_vcnt = 0;
1030                         if (rrdev)
1031                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1032
1033                         if (conf->mddev->gendisk)
1034                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1035                                                       bi, disk_devt(conf->mddev->gendisk),
1036                                                       sh->dev[i].sector);
1037                         generic_make_request(bi);
1038                 }
1039                 if (rrdev) {
1040                         if (s->syncing || s->expanding || s->expanded
1041                             || s->replacing)
1042                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1043
1044                         set_bit(STRIPE_IO_STARTED, &sh->state);
1045
1046                         bio_reset(rbi);
1047                         rbi->bi_bdev = rrdev->bdev;
1048                         rbi->bi_rw = rw;
1049                         BUG_ON(!(rw & WRITE));
1050                         rbi->bi_end_io = raid5_end_write_request;
1051                         rbi->bi_private = sh;
1052
1053                         pr_debug("%s: for %llu schedule op %ld on "
1054                                  "replacement disc %d\n",
1055                                 __func__, (unsigned long long)sh->sector,
1056                                 rbi->bi_rw, i);
1057                         atomic_inc(&sh->count);
1058                         if (sh != head_sh)
1059                                 atomic_inc(&head_sh->count);
1060                         if (use_new_offset(conf, sh))
1061                                 rbi->bi_iter.bi_sector = (sh->sector
1062                                                   + rrdev->new_data_offset);
1063                         else
1064                                 rbi->bi_iter.bi_sector = (sh->sector
1065                                                   + rrdev->data_offset);
1066                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1067                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1068                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1069                         rbi->bi_vcnt = 1;
1070                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1071                         rbi->bi_io_vec[0].bv_offset = 0;
1072                         rbi->bi_iter.bi_size = STRIPE_SIZE;
1073                         /*
1074                          * If this is discard request, set bi_vcnt 0. We don't
1075                          * want to confuse SCSI because SCSI will replace payload
1076                          */
1077                         if (rw & REQ_DISCARD)
1078                                 rbi->bi_vcnt = 0;
1079                         if (conf->mddev->gendisk)
1080                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1081                                                       rbi, disk_devt(conf->mddev->gendisk),
1082                                                       sh->dev[i].sector);
1083                         generic_make_request(rbi);
1084                 }
1085                 if (!rdev && !rrdev) {
1086                         if (rw & WRITE)
1087                                 set_bit(STRIPE_DEGRADED, &sh->state);
1088                         pr_debug("skip op %ld on disc %d for sector %llu\n",
1089                                 bi->bi_rw, i, (unsigned long long)sh->sector);
1090                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1091                         set_bit(STRIPE_HANDLE, &sh->state);
1092                 }
1093
1094                 if (!head_sh->batch_head)
1095                         continue;
1096                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1097                                       batch_list);
1098                 if (sh != head_sh)
1099                         goto again;
1100         }
1101 }
1102
1103 static struct dma_async_tx_descriptor *
1104 async_copy_data(int frombio, struct bio *bio, struct page **page,
1105         sector_t sector, struct dma_async_tx_descriptor *tx,
1106         struct stripe_head *sh)
1107 {
1108         struct bio_vec bvl;
1109         struct bvec_iter iter;
1110         struct page *bio_page;
1111         int page_offset;
1112         struct async_submit_ctl submit;
1113         enum async_tx_flags flags = 0;
1114
1115         if (bio->bi_iter.bi_sector >= sector)
1116                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1117         else
1118                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1119
1120         if (frombio)
1121                 flags |= ASYNC_TX_FENCE;
1122         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1123
1124         bio_for_each_segment(bvl, bio, iter) {
1125                 int len = bvl.bv_len;
1126                 int clen;
1127                 int b_offset = 0;
1128
1129                 if (page_offset < 0) {
1130                         b_offset = -page_offset;
1131                         page_offset += b_offset;
1132                         len -= b_offset;
1133                 }
1134
1135                 if (len > 0 && page_offset + len > STRIPE_SIZE)
1136                         clen = STRIPE_SIZE - page_offset;
1137                 else
1138                         clen = len;
1139
1140                 if (clen > 0) {
1141                         b_offset += bvl.bv_offset;
1142                         bio_page = bvl.bv_page;
1143                         if (frombio) {
1144                                 if (sh->raid_conf->skip_copy &&
1145                                     b_offset == 0 && page_offset == 0 &&
1146                                     clen == STRIPE_SIZE)
1147                                         *page = bio_page;
1148                                 else
1149                                         tx = async_memcpy(*page, bio_page, page_offset,
1150                                                   b_offset, clen, &submit);
1151                         } else
1152                                 tx = async_memcpy(bio_page, *page, b_offset,
1153                                                   page_offset, clen, &submit);
1154                 }
1155                 /* chain the operations */
1156                 submit.depend_tx = tx;
1157
1158                 if (clen < len) /* hit end of page */
1159                         break;
1160                 page_offset +=  len;
1161         }
1162
1163         return tx;
1164 }
1165
1166 static void ops_complete_biofill(void *stripe_head_ref)
1167 {
1168         struct stripe_head *sh = stripe_head_ref;
1169         struct bio *return_bi = NULL;
1170         int i;
1171
1172         pr_debug("%s: stripe %llu\n", __func__,
1173                 (unsigned long long)sh->sector);
1174
1175         /* clear completed biofills */
1176         for (i = sh->disks; i--; ) {
1177                 struct r5dev *dev = &sh->dev[i];
1178
1179                 /* acknowledge completion of a biofill operation */
1180                 /* and check if we need to reply to a read request,
1181                  * new R5_Wantfill requests are held off until
1182                  * !STRIPE_BIOFILL_RUN
1183                  */
1184                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1185                         struct bio *rbi, *rbi2;
1186
1187                         BUG_ON(!dev->read);
1188                         rbi = dev->read;
1189                         dev->read = NULL;
1190                         while (rbi && rbi->bi_iter.bi_sector <
1191                                 dev->sector + STRIPE_SECTORS) {
1192                                 rbi2 = r5_next_bio(rbi, dev->sector);
1193                                 if (!raid5_dec_bi_active_stripes(rbi)) {
1194                                         rbi->bi_next = return_bi;
1195                                         return_bi = rbi;
1196                                 }
1197                                 rbi = rbi2;
1198                         }
1199                 }
1200         }
1201         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1202
1203         return_io(return_bi);
1204
1205         set_bit(STRIPE_HANDLE, &sh->state);
1206         release_stripe(sh);
1207 }
1208
1209 static void ops_run_biofill(struct stripe_head *sh)
1210 {
1211         struct dma_async_tx_descriptor *tx = NULL;
1212         struct async_submit_ctl submit;
1213         int i;
1214
1215         BUG_ON(sh->batch_head);
1216         pr_debug("%s: stripe %llu\n", __func__,
1217                 (unsigned long long)sh->sector);
1218
1219         for (i = sh->disks; i--; ) {
1220                 struct r5dev *dev = &sh->dev[i];
1221                 if (test_bit(R5_Wantfill, &dev->flags)) {
1222                         struct bio *rbi;
1223                         spin_lock_irq(&sh->stripe_lock);
1224                         dev->read = rbi = dev->toread;
1225                         dev->toread = NULL;
1226                         spin_unlock_irq(&sh->stripe_lock);
1227                         while (rbi && rbi->bi_iter.bi_sector <
1228                                 dev->sector + STRIPE_SECTORS) {
1229                                 tx = async_copy_data(0, rbi, &dev->page,
1230                                         dev->sector, tx, sh);
1231                                 rbi = r5_next_bio(rbi, dev->sector);
1232                         }
1233                 }
1234         }
1235
1236         atomic_inc(&sh->count);
1237         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1238         async_trigger_callback(&submit);
1239 }
1240
1241 static void mark_target_uptodate(struct stripe_head *sh, int target)
1242 {
1243         struct r5dev *tgt;
1244
1245         if (target < 0)
1246                 return;
1247
1248         tgt = &sh->dev[target];
1249         set_bit(R5_UPTODATE, &tgt->flags);
1250         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1251         clear_bit(R5_Wantcompute, &tgt->flags);
1252 }
1253
1254 static void ops_complete_compute(void *stripe_head_ref)
1255 {
1256         struct stripe_head *sh = stripe_head_ref;
1257
1258         pr_debug("%s: stripe %llu\n", __func__,
1259                 (unsigned long long)sh->sector);
1260
1261         /* mark the computed target(s) as uptodate */
1262         mark_target_uptodate(sh, sh->ops.target);
1263         mark_target_uptodate(sh, sh->ops.target2);
1264
1265         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1266         if (sh->check_state == check_state_compute_run)
1267                 sh->check_state = check_state_compute_result;
1268         set_bit(STRIPE_HANDLE, &sh->state);
1269         release_stripe(sh);
1270 }
1271
1272 /* return a pointer to the address conversion region of the scribble buffer */
1273 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1274                                  struct raid5_percpu *percpu, int i)
1275 {
1276         void *addr;
1277
1278         addr = flex_array_get(percpu->scribble, i);
1279         return addr + sizeof(struct page *) * (sh->disks + 2);
1280 }
1281
1282 /* return a pointer to the address conversion region of the scribble buffer */
1283 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1284 {
1285         void *addr;
1286
1287         addr = flex_array_get(percpu->scribble, i);
1288         return addr;
1289 }
1290
1291 static struct dma_async_tx_descriptor *
1292 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1293 {
1294         int disks = sh->disks;
1295         struct page **xor_srcs = to_addr_page(percpu, 0);
1296         int target = sh->ops.target;
1297         struct r5dev *tgt = &sh->dev[target];
1298         struct page *xor_dest = tgt->page;
1299         int count = 0;
1300         struct dma_async_tx_descriptor *tx;
1301         struct async_submit_ctl submit;
1302         int i;
1303
1304         BUG_ON(sh->batch_head);
1305
1306         pr_debug("%s: stripe %llu block: %d\n",
1307                 __func__, (unsigned long long)sh->sector, target);
1308         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1309
1310         for (i = disks; i--; )
1311                 if (i != target)
1312                         xor_srcs[count++] = sh->dev[i].page;
1313
1314         atomic_inc(&sh->count);
1315
1316         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1317                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1318         if (unlikely(count == 1))
1319                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1320         else
1321                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1322
1323         return tx;
1324 }
1325
1326 /* set_syndrome_sources - populate source buffers for gen_syndrome
1327  * @srcs - (struct page *) array of size sh->disks
1328  * @sh - stripe_head to parse
1329  *
1330  * Populates srcs in proper layout order for the stripe and returns the
1331  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1332  * destination buffer is recorded in srcs[count] and the Q destination
1333  * is recorded in srcs[count+1]].
1334  */
1335 static int set_syndrome_sources(struct page **srcs,
1336                                 struct stripe_head *sh,
1337                                 int srctype)
1338 {
1339         int disks = sh->disks;
1340         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1341         int d0_idx = raid6_d0(sh);
1342         int count;
1343         int i;
1344
1345         for (i = 0; i < disks; i++)
1346                 srcs[i] = NULL;
1347
1348         count = 0;
1349         i = d0_idx;
1350         do {
1351                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1352                 struct r5dev *dev = &sh->dev[i];
1353
1354                 if (i == sh->qd_idx || i == sh->pd_idx ||
1355                     (srctype == SYNDROME_SRC_ALL) ||
1356                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1357                      test_bit(R5_Wantdrain, &dev->flags)) ||
1358                     (srctype == SYNDROME_SRC_WRITTEN &&
1359                      dev->written))
1360                         srcs[slot] = sh->dev[i].page;
1361                 i = raid6_next_disk(i, disks);
1362         } while (i != d0_idx);
1363
1364         return syndrome_disks;
1365 }
1366
1367 static struct dma_async_tx_descriptor *
1368 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1369 {
1370         int disks = sh->disks;
1371         struct page **blocks = to_addr_page(percpu, 0);
1372         int target;
1373         int qd_idx = sh->qd_idx;
1374         struct dma_async_tx_descriptor *tx;
1375         struct async_submit_ctl submit;
1376         struct r5dev *tgt;
1377         struct page *dest;
1378         int i;
1379         int count;
1380
1381         BUG_ON(sh->batch_head);
1382         if (sh->ops.target < 0)
1383                 target = sh->ops.target2;
1384         else if (sh->ops.target2 < 0)
1385                 target = sh->ops.target;
1386         else
1387                 /* we should only have one valid target */
1388                 BUG();
1389         BUG_ON(target < 0);
1390         pr_debug("%s: stripe %llu block: %d\n",
1391                 __func__, (unsigned long long)sh->sector, target);
1392
1393         tgt = &sh->dev[target];
1394         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1395         dest = tgt->page;
1396
1397         atomic_inc(&sh->count);
1398
1399         if (target == qd_idx) {
1400                 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1401                 blocks[count] = NULL; /* regenerating p is not necessary */
1402                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1403                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1404                                   ops_complete_compute, sh,
1405                                   to_addr_conv(sh, percpu, 0));
1406                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1407         } else {
1408                 /* Compute any data- or p-drive using XOR */
1409                 count = 0;
1410                 for (i = disks; i-- ; ) {
1411                         if (i == target || i == qd_idx)
1412                                 continue;
1413                         blocks[count++] = sh->dev[i].page;
1414                 }
1415
1416                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1417                                   NULL, ops_complete_compute, sh,
1418                                   to_addr_conv(sh, percpu, 0));
1419                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1420         }
1421
1422         return tx;
1423 }
1424
1425 static struct dma_async_tx_descriptor *
1426 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1427 {
1428         int i, count, disks = sh->disks;
1429         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1430         int d0_idx = raid6_d0(sh);
1431         int faila = -1, failb = -1;
1432         int target = sh->ops.target;
1433         int target2 = sh->ops.target2;
1434         struct r5dev *tgt = &sh->dev[target];
1435         struct r5dev *tgt2 = &sh->dev[target2];
1436         struct dma_async_tx_descriptor *tx;
1437         struct page **blocks = to_addr_page(percpu, 0);
1438         struct async_submit_ctl submit;
1439
1440         BUG_ON(sh->batch_head);
1441         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1442                  __func__, (unsigned long long)sh->sector, target, target2);
1443         BUG_ON(target < 0 || target2 < 0);
1444         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1445         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1446
1447         /* we need to open-code set_syndrome_sources to handle the
1448          * slot number conversion for 'faila' and 'failb'
1449          */
1450         for (i = 0; i < disks ; i++)
1451                 blocks[i] = NULL;
1452         count = 0;
1453         i = d0_idx;
1454         do {
1455                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1456
1457                 blocks[slot] = sh->dev[i].page;
1458
1459                 if (i == target)
1460                         faila = slot;
1461                 if (i == target2)
1462                         failb = slot;
1463                 i = raid6_next_disk(i, disks);
1464         } while (i != d0_idx);
1465
1466         BUG_ON(faila == failb);
1467         if (failb < faila)
1468                 swap(faila, failb);
1469         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1470                  __func__, (unsigned long long)sh->sector, faila, failb);
1471
1472         atomic_inc(&sh->count);
1473
1474         if (failb == syndrome_disks+1) {
1475                 /* Q disk is one of the missing disks */
1476                 if (faila == syndrome_disks) {
1477                         /* Missing P+Q, just recompute */
1478                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1479                                           ops_complete_compute, sh,
1480                                           to_addr_conv(sh, percpu, 0));
1481                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1482                                                   STRIPE_SIZE, &submit);
1483                 } else {
1484                         struct page *dest;
1485                         int data_target;
1486                         int qd_idx = sh->qd_idx;
1487
1488                         /* Missing D+Q: recompute D from P, then recompute Q */
1489                         if (target == qd_idx)
1490                                 data_target = target2;
1491                         else
1492                                 data_target = target;
1493
1494                         count = 0;
1495                         for (i = disks; i-- ; ) {
1496                                 if (i == data_target || i == qd_idx)
1497                                         continue;
1498                                 blocks[count++] = sh->dev[i].page;
1499                         }
1500                         dest = sh->dev[data_target].page;
1501                         init_async_submit(&submit,
1502                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1503                                           NULL, NULL, NULL,
1504                                           to_addr_conv(sh, percpu, 0));
1505                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1506                                        &submit);
1507
1508                         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1509                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1510                                           ops_complete_compute, sh,
1511                                           to_addr_conv(sh, percpu, 0));
1512                         return async_gen_syndrome(blocks, 0, count+2,
1513                                                   STRIPE_SIZE, &submit);
1514                 }
1515         } else {
1516                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1517                                   ops_complete_compute, sh,
1518                                   to_addr_conv(sh, percpu, 0));
1519                 if (failb == syndrome_disks) {
1520                         /* We're missing D+P. */
1521                         return async_raid6_datap_recov(syndrome_disks+2,
1522                                                        STRIPE_SIZE, faila,
1523                                                        blocks, &submit);
1524                 } else {
1525                         /* We're missing D+D. */
1526                         return async_raid6_2data_recov(syndrome_disks+2,
1527                                                        STRIPE_SIZE, faila, failb,
1528                                                        blocks, &submit);
1529                 }
1530         }
1531 }
1532
1533 static void ops_complete_prexor(void *stripe_head_ref)
1534 {
1535         struct stripe_head *sh = stripe_head_ref;
1536
1537         pr_debug("%s: stripe %llu\n", __func__,
1538                 (unsigned long long)sh->sector);
1539 }
1540
1541 static struct dma_async_tx_descriptor *
1542 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1543                 struct dma_async_tx_descriptor *tx)
1544 {
1545         int disks = sh->disks;
1546         struct page **xor_srcs = to_addr_page(percpu, 0);
1547         int count = 0, pd_idx = sh->pd_idx, i;
1548         struct async_submit_ctl submit;
1549
1550         /* existing parity data subtracted */
1551         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1552
1553         BUG_ON(sh->batch_head);
1554         pr_debug("%s: stripe %llu\n", __func__,
1555                 (unsigned long long)sh->sector);
1556
1557         for (i = disks; i--; ) {
1558                 struct r5dev *dev = &sh->dev[i];
1559                 /* Only process blocks that are known to be uptodate */
1560                 if (test_bit(R5_Wantdrain, &dev->flags))
1561                         xor_srcs[count++] = dev->page;
1562         }
1563
1564         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1565                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1566         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1567
1568         return tx;
1569 }
1570
1571 static struct dma_async_tx_descriptor *
1572 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1573                 struct dma_async_tx_descriptor *tx)
1574 {
1575         struct page **blocks = to_addr_page(percpu, 0);
1576         int count;
1577         struct async_submit_ctl submit;
1578
1579         pr_debug("%s: stripe %llu\n", __func__,
1580                 (unsigned long long)sh->sector);
1581
1582         count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1583
1584         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1585                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1586         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1587
1588         return tx;
1589 }
1590
1591 static struct dma_async_tx_descriptor *
1592 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1593 {
1594         int disks = sh->disks;
1595         int i;
1596         struct stripe_head *head_sh = sh;
1597
1598         pr_debug("%s: stripe %llu\n", __func__,
1599                 (unsigned long long)sh->sector);
1600
1601         for (i = disks; i--; ) {
1602                 struct r5dev *dev;
1603                 struct bio *chosen;
1604
1605                 sh = head_sh;
1606                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1607                         struct bio *wbi;
1608
1609 again:
1610                         dev = &sh->dev[i];
1611                         spin_lock_irq(&sh->stripe_lock);
1612                         chosen = dev->towrite;
1613                         dev->towrite = NULL;
1614                         sh->overwrite_disks = 0;
1615                         BUG_ON(dev->written);
1616                         wbi = dev->written = chosen;
1617                         spin_unlock_irq(&sh->stripe_lock);
1618                         WARN_ON(dev->page != dev->orig_page);
1619
1620                         while (wbi && wbi->bi_iter.bi_sector <
1621                                 dev->sector + STRIPE_SECTORS) {
1622                                 if (wbi->bi_rw & REQ_FUA)
1623                                         set_bit(R5_WantFUA, &dev->flags);
1624                                 if (wbi->bi_rw & REQ_SYNC)
1625                                         set_bit(R5_SyncIO, &dev->flags);
1626                                 if (wbi->bi_rw & REQ_DISCARD)
1627                                         set_bit(R5_Discard, &dev->flags);
1628                                 else {
1629                                         tx = async_copy_data(1, wbi, &dev->page,
1630                                                 dev->sector, tx, sh);
1631                                         if (dev->page != dev->orig_page) {
1632                                                 set_bit(R5_SkipCopy, &dev->flags);
1633                                                 clear_bit(R5_UPTODATE, &dev->flags);
1634                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1635                                         }
1636                                 }
1637                                 wbi = r5_next_bio(wbi, dev->sector);
1638                         }
1639
1640                         if (head_sh->batch_head) {
1641                                 sh = list_first_entry(&sh->batch_list,
1642                                                       struct stripe_head,
1643                                                       batch_list);
1644                                 if (sh == head_sh)
1645                                         continue;
1646                                 goto again;
1647                         }
1648                 }
1649         }
1650
1651         return tx;
1652 }
1653
1654 static void ops_complete_reconstruct(void *stripe_head_ref)
1655 {
1656         struct stripe_head *sh = stripe_head_ref;
1657         int disks = sh->disks;
1658         int pd_idx = sh->pd_idx;
1659         int qd_idx = sh->qd_idx;
1660         int i;
1661         bool fua = false, sync = false, discard = false;
1662
1663         pr_debug("%s: stripe %llu\n", __func__,
1664                 (unsigned long long)sh->sector);
1665
1666         for (i = disks; i--; ) {
1667                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1668                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1669                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1670         }
1671
1672         for (i = disks; i--; ) {
1673                 struct r5dev *dev = &sh->dev[i];
1674
1675                 if (dev->written || i == pd_idx || i == qd_idx) {
1676                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1677                                 set_bit(R5_UPTODATE, &dev->flags);
1678                         if (fua)
1679                                 set_bit(R5_WantFUA, &dev->flags);
1680                         if (sync)
1681                                 set_bit(R5_SyncIO, &dev->flags);
1682                 }
1683         }
1684
1685         if (sh->reconstruct_state == reconstruct_state_drain_run)
1686                 sh->reconstruct_state = reconstruct_state_drain_result;
1687         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1688                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1689         else {
1690                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1691                 sh->reconstruct_state = reconstruct_state_result;
1692         }
1693
1694         set_bit(STRIPE_HANDLE, &sh->state);
1695         release_stripe(sh);
1696 }
1697
1698 static void
1699 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1700                      struct dma_async_tx_descriptor *tx)
1701 {
1702         int disks = sh->disks;
1703         struct page **xor_srcs;
1704         struct async_submit_ctl submit;
1705         int count, pd_idx = sh->pd_idx, i;
1706         struct page *xor_dest;
1707         int prexor = 0;
1708         unsigned long flags;
1709         int j = 0;
1710         struct stripe_head *head_sh = sh;
1711         int last_stripe;
1712
1713         pr_debug("%s: stripe %llu\n", __func__,
1714                 (unsigned long long)sh->sector);
1715
1716         for (i = 0; i < sh->disks; i++) {
1717                 if (pd_idx == i)
1718                         continue;
1719                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1720                         break;
1721         }
1722         if (i >= sh->disks) {
1723                 atomic_inc(&sh->count);
1724                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1725                 ops_complete_reconstruct(sh);
1726                 return;
1727         }
1728 again:
1729         count = 0;
1730         xor_srcs = to_addr_page(percpu, j);
1731         /* check if prexor is active which means only process blocks
1732          * that are part of a read-modify-write (written)
1733          */
1734         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1735                 prexor = 1;
1736                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1737                 for (i = disks; i--; ) {
1738                         struct r5dev *dev = &sh->dev[i];
1739                         if (head_sh->dev[i].written)
1740                                 xor_srcs[count++] = dev->page;
1741                 }
1742         } else {
1743                 xor_dest = sh->dev[pd_idx].page;
1744                 for (i = disks; i--; ) {
1745                         struct r5dev *dev = &sh->dev[i];
1746                         if (i != pd_idx)
1747                                 xor_srcs[count++] = dev->page;
1748                 }
1749         }
1750
1751         /* 1/ if we prexor'd then the dest is reused as a source
1752          * 2/ if we did not prexor then we are redoing the parity
1753          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1754          * for the synchronous xor case
1755          */
1756         last_stripe = !head_sh->batch_head ||
1757                 list_first_entry(&sh->batch_list,
1758                                  struct stripe_head, batch_list) == head_sh;
1759         if (last_stripe) {
1760                 flags = ASYNC_TX_ACK |
1761                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1762
1763                 atomic_inc(&head_sh->count);
1764                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1765                                   to_addr_conv(sh, percpu, j));
1766         } else {
1767                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1768                 init_async_submit(&submit, flags, tx, NULL, NULL,
1769                                   to_addr_conv(sh, percpu, j));
1770         }
1771
1772         if (unlikely(count == 1))
1773                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1774         else
1775                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1776         if (!last_stripe) {
1777                 j++;
1778                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1779                                       batch_list);
1780                 goto again;
1781         }
1782 }
1783
1784 static void
1785 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1786                      struct dma_async_tx_descriptor *tx)
1787 {
1788         struct async_submit_ctl submit;
1789         struct page **blocks;
1790         int count, i, j = 0;
1791         struct stripe_head *head_sh = sh;
1792         int last_stripe;
1793         int synflags;
1794         unsigned long txflags;
1795
1796         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1797
1798         for (i = 0; i < sh->disks; i++) {
1799                 if (sh->pd_idx == i || sh->qd_idx == i)
1800                         continue;
1801                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1802                         break;
1803         }
1804         if (i >= sh->disks) {
1805                 atomic_inc(&sh->count);
1806                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1807                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1808                 ops_complete_reconstruct(sh);
1809                 return;
1810         }
1811
1812 again:
1813         blocks = to_addr_page(percpu, j);
1814
1815         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1816                 synflags = SYNDROME_SRC_WRITTEN;
1817                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1818         } else {
1819                 synflags = SYNDROME_SRC_ALL;
1820                 txflags = ASYNC_TX_ACK;
1821         }
1822
1823         count = set_syndrome_sources(blocks, sh, synflags);
1824         last_stripe = !head_sh->batch_head ||
1825                 list_first_entry(&sh->batch_list,
1826                                  struct stripe_head, batch_list) == head_sh;
1827
1828         if (last_stripe) {
1829                 atomic_inc(&head_sh->count);
1830                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1831                                   head_sh, to_addr_conv(sh, percpu, j));
1832         } else
1833                 init_async_submit(&submit, 0, tx, NULL, NULL,
1834                                   to_addr_conv(sh, percpu, j));
1835         tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1836         if (!last_stripe) {
1837                 j++;
1838                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1839                                       batch_list);
1840                 goto again;
1841         }
1842 }
1843
1844 static void ops_complete_check(void *stripe_head_ref)
1845 {
1846         struct stripe_head *sh = stripe_head_ref;
1847
1848         pr_debug("%s: stripe %llu\n", __func__,
1849                 (unsigned long long)sh->sector);
1850
1851         sh->check_state = check_state_check_result;
1852         set_bit(STRIPE_HANDLE, &sh->state);
1853         release_stripe(sh);
1854 }
1855
1856 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1857 {
1858         int disks = sh->disks;
1859         int pd_idx = sh->pd_idx;
1860         int qd_idx = sh->qd_idx;
1861         struct page *xor_dest;
1862         struct page **xor_srcs = to_addr_page(percpu, 0);
1863         struct dma_async_tx_descriptor *tx;
1864         struct async_submit_ctl submit;
1865         int count;
1866         int i;
1867
1868         pr_debug("%s: stripe %llu\n", __func__,
1869                 (unsigned long long)sh->sector);
1870
1871         BUG_ON(sh->batch_head);
1872         count = 0;
1873         xor_dest = sh->dev[pd_idx].page;
1874         xor_srcs[count++] = xor_dest;
1875         for (i = disks; i--; ) {
1876                 if (i == pd_idx || i == qd_idx)
1877                         continue;
1878                 xor_srcs[count++] = sh->dev[i].page;
1879         }
1880
1881         init_async_submit(&submit, 0, NULL, NULL, NULL,
1882                           to_addr_conv(sh, percpu, 0));
1883         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1884                            &sh->ops.zero_sum_result, &submit);
1885
1886         atomic_inc(&sh->count);
1887         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1888         tx = async_trigger_callback(&submit);
1889 }
1890
1891 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1892 {
1893         struct page **srcs = to_addr_page(percpu, 0);
1894         struct async_submit_ctl submit;
1895         int count;
1896
1897         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1898                 (unsigned long long)sh->sector, checkp);
1899
1900         BUG_ON(sh->batch_head);
1901         count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1902         if (!checkp)
1903                 srcs[count] = NULL;
1904
1905         atomic_inc(&sh->count);
1906         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1907                           sh, to_addr_conv(sh, percpu, 0));
1908         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1909                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1910 }
1911
1912 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1913 {
1914         int overlap_clear = 0, i, disks = sh->disks;
1915         struct dma_async_tx_descriptor *tx = NULL;
1916         struct r5conf *conf = sh->raid_conf;
1917         int level = conf->level;
1918         struct raid5_percpu *percpu;
1919         unsigned long cpu;
1920
1921         cpu = get_cpu_light();
1922         percpu = per_cpu_ptr(conf->percpu, cpu);
1923         spin_lock(&percpu->lock);
1924         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1925                 ops_run_biofill(sh);
1926                 overlap_clear++;
1927         }
1928
1929         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1930                 if (level < 6)
1931                         tx = ops_run_compute5(sh, percpu);
1932                 else {
1933                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1934                                 tx = ops_run_compute6_1(sh, percpu);
1935                         else
1936                                 tx = ops_run_compute6_2(sh, percpu);
1937                 }
1938                 /* terminate the chain if reconstruct is not set to be run */
1939                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1940                         async_tx_ack(tx);
1941         }
1942
1943         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1944                 if (level < 6)
1945                         tx = ops_run_prexor5(sh, percpu, tx);
1946                 else
1947                         tx = ops_run_prexor6(sh, percpu, tx);
1948         }
1949
1950         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1951                 tx = ops_run_biodrain(sh, tx);
1952                 overlap_clear++;
1953         }
1954
1955         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1956                 if (level < 6)
1957                         ops_run_reconstruct5(sh, percpu, tx);
1958                 else
1959                         ops_run_reconstruct6(sh, percpu, tx);
1960         }
1961
1962         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1963                 if (sh->check_state == check_state_run)
1964                         ops_run_check_p(sh, percpu);
1965                 else if (sh->check_state == check_state_run_q)
1966                         ops_run_check_pq(sh, percpu, 0);
1967                 else if (sh->check_state == check_state_run_pq)
1968                         ops_run_check_pq(sh, percpu, 1);
1969                 else
1970                         BUG();
1971         }
1972
1973         if (overlap_clear && !sh->batch_head)
1974                 for (i = disks; i--; ) {
1975                         struct r5dev *dev = &sh->dev[i];
1976                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1977                                 wake_up(&sh->raid_conf->wait_for_overlap);
1978                 }
1979         spin_unlock(&percpu->lock);
1980         put_cpu_light();
1981 }
1982
1983 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1984 {
1985         struct stripe_head *sh;
1986
1987         sh = kmem_cache_zalloc(sc, gfp);
1988         if (sh) {
1989                 spin_lock_init(&sh->stripe_lock);
1990                 spin_lock_init(&sh->batch_lock);
1991                 INIT_LIST_HEAD(&sh->batch_list);
1992                 INIT_LIST_HEAD(&sh->lru);
1993                 atomic_set(&sh->count, 1);
1994         }
1995         return sh;
1996 }
1997 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
1998 {
1999         struct stripe_head *sh;
2000
2001         sh = alloc_stripe(conf->slab_cache, gfp);
2002         if (!sh)
2003                 return 0;
2004
2005         sh->raid_conf = conf;
2006
2007         if (grow_buffers(sh, gfp)) {
2008                 shrink_buffers(sh);
2009                 kmem_cache_free(conf->slab_cache, sh);
2010                 return 0;
2011         }
2012         sh->hash_lock_index =
2013                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2014         /* we just created an active stripe so... */
2015         atomic_inc(&conf->active_stripes);
2016
2017         release_stripe(sh);
2018         conf->max_nr_stripes++;
2019         return 1;
2020 }
2021
2022 static int grow_stripes(struct r5conf *conf, int num)
2023 {
2024         struct kmem_cache *sc;
2025         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2026
2027         if (conf->mddev->gendisk)
2028                 sprintf(conf->cache_name[0],
2029                         "raid%d-%s", conf->level, mdname(conf->mddev));
2030         else
2031                 sprintf(conf->cache_name[0],
2032                         "raid%d-%p", conf->level, conf->mddev);
2033         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2034
2035         conf->active_name = 0;
2036         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2037                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2038                                0, 0, NULL);
2039         if (!sc)
2040                 return 1;
2041         conf->slab_cache = sc;
2042         conf->pool_size = devs;
2043         while (num--)
2044                 if (!grow_one_stripe(conf, GFP_KERNEL))
2045                         return 1;
2046
2047         return 0;
2048 }
2049
2050 /**
2051  * scribble_len - return the required size of the scribble region
2052  * @num - total number of disks in the array
2053  *
2054  * The size must be enough to contain:
2055  * 1/ a struct page pointer for each device in the array +2
2056  * 2/ room to convert each entry in (1) to its corresponding dma
2057  *    (dma_map_page()) or page (page_address()) address.
2058  *
2059  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2060  * calculate over all devices (not just the data blocks), using zeros in place
2061  * of the P and Q blocks.
2062  */
2063 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2064 {
2065         struct flex_array *ret;
2066         size_t len;
2067
2068         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2069         ret = flex_array_alloc(len, cnt, flags);
2070         if (!ret)
2071                 return NULL;
2072         /* always prealloc all elements, so no locking is required */
2073         if (flex_array_prealloc(ret, 0, cnt, flags)) {
2074                 flex_array_free(ret);
2075                 return NULL;
2076         }
2077         return ret;
2078 }
2079
2080 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2081 {
2082         unsigned long cpu;
2083         int err = 0;
2084
2085         mddev_suspend(conf->mddev);
2086         get_online_cpus();
2087         for_each_present_cpu(cpu) {
2088                 struct raid5_percpu *percpu;
2089                 struct flex_array *scribble;
2090
2091                 percpu = per_cpu_ptr(conf->percpu, cpu);
2092                 scribble = scribble_alloc(new_disks,
2093                                           new_sectors / STRIPE_SECTORS,
2094                                           GFP_NOIO);
2095
2096                 if (scribble) {
2097                         flex_array_free(percpu->scribble);
2098                         percpu->scribble = scribble;
2099                 } else {
2100                         err = -ENOMEM;
2101                         break;
2102                 }
2103         }
2104         put_online_cpus();
2105         mddev_resume(conf->mddev);
2106         return err;
2107 }
2108
2109 static int resize_stripes(struct r5conf *conf, int newsize)
2110 {
2111         /* Make all the stripes able to hold 'newsize' devices.
2112          * New slots in each stripe get 'page' set to a new page.
2113          *
2114          * This happens in stages:
2115          * 1/ create a new kmem_cache and allocate the required number of
2116          *    stripe_heads.
2117          * 2/ gather all the old stripe_heads and transfer the pages across
2118          *    to the new stripe_heads.  This will have the side effect of
2119          *    freezing the array as once all stripe_heads have been collected,
2120          *    no IO will be possible.  Old stripe heads are freed once their
2121          *    pages have been transferred over, and the old kmem_cache is
2122          *    freed when all stripes are done.
2123          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2124          *    we simple return a failre status - no need to clean anything up.
2125          * 4/ allocate new pages for the new slots in the new stripe_heads.
2126          *    If this fails, we don't bother trying the shrink the
2127          *    stripe_heads down again, we just leave them as they are.
2128          *    As each stripe_head is processed the new one is released into
2129          *    active service.
2130          *
2131          * Once step2 is started, we cannot afford to wait for a write,
2132          * so we use GFP_NOIO allocations.
2133          */
2134         struct stripe_head *osh, *nsh;
2135         LIST_HEAD(newstripes);
2136         struct disk_info *ndisks;
2137         int err;
2138         struct kmem_cache *sc;
2139         int i;
2140         int hash, cnt;
2141
2142         if (newsize <= conf->pool_size)
2143                 return 0; /* never bother to shrink */
2144
2145         err = md_allow_write(conf->mddev);
2146         if (err)
2147                 return err;
2148
2149         /* Step 1 */
2150         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2151                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2152                                0, 0, NULL);
2153         if (!sc)
2154                 return -ENOMEM;
2155
2156         /* Need to ensure auto-resizing doesn't interfere */
2157         mutex_lock(&conf->cache_size_mutex);
2158
2159         for (i = conf->max_nr_stripes; i; i--) {
2160                 nsh = alloc_stripe(sc, GFP_KERNEL);
2161                 if (!nsh)
2162                         break;
2163
2164                 nsh->raid_conf = conf;
2165                 list_add(&nsh->lru, &newstripes);
2166         }
2167         if (i) {
2168                 /* didn't get enough, give up */
2169                 while (!list_empty(&newstripes)) {
2170                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2171                         list_del(&nsh->lru);
2172                         kmem_cache_free(sc, nsh);
2173                 }
2174                 kmem_cache_destroy(sc);
2175                 mutex_unlock(&conf->cache_size_mutex);
2176                 return -ENOMEM;
2177         }
2178         /* Step 2 - Must use GFP_NOIO now.
2179          * OK, we have enough stripes, start collecting inactive
2180          * stripes and copying them over
2181          */
2182         hash = 0;
2183         cnt = 0;
2184         list_for_each_entry(nsh, &newstripes, lru) {
2185                 lock_device_hash_lock(conf, hash);
2186                 wait_event_cmd(conf->wait_for_stripe,
2187                                     !list_empty(conf->inactive_list + hash),
2188                                     unlock_device_hash_lock(conf, hash),
2189                                     lock_device_hash_lock(conf, hash));
2190                 osh = get_free_stripe(conf, hash);
2191                 unlock_device_hash_lock(conf, hash);
2192
2193                 for(i=0; i<conf->pool_size; i++) {
2194                         nsh->dev[i].page = osh->dev[i].page;
2195                         nsh->dev[i].orig_page = osh->dev[i].page;
2196                 }
2197                 nsh->hash_lock_index = hash;
2198                 kmem_cache_free(conf->slab_cache, osh);
2199                 cnt++;
2200                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2201                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2202                         hash++;
2203                         cnt = 0;
2204                 }
2205         }
2206         kmem_cache_destroy(conf->slab_cache);
2207
2208         /* Step 3.
2209          * At this point, we are holding all the stripes so the array
2210          * is completely stalled, so now is a good time to resize
2211          * conf->disks and the scribble region
2212          */
2213         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2214         if (ndisks) {
2215                 for (i=0; i<conf->raid_disks; i++)
2216                         ndisks[i] = conf->disks[i];
2217                 kfree(conf->disks);
2218                 conf->disks = ndisks;
2219         } else
2220                 err = -ENOMEM;
2221
2222         mutex_unlock(&conf->cache_size_mutex);
2223         /* Step 4, return new stripes to service */
2224         while(!list_empty(&newstripes)) {
2225                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2226                 list_del_init(&nsh->lru);
2227
2228                 for (i=conf->raid_disks; i < newsize; i++)
2229                         if (nsh->dev[i].page == NULL) {
2230                                 struct page *p = alloc_page(GFP_NOIO);
2231                                 nsh->dev[i].page = p;
2232                                 nsh->dev[i].orig_page = p;
2233                                 if (!p)
2234                                         err = -ENOMEM;
2235                         }
2236                 release_stripe(nsh);
2237         }
2238         /* critical section pass, GFP_NOIO no longer needed */
2239
2240         conf->slab_cache = sc;
2241         conf->active_name = 1-conf->active_name;
2242         if (!err)
2243                 conf->pool_size = newsize;
2244         return err;
2245 }
2246
2247 static int drop_one_stripe(struct r5conf *conf)
2248 {
2249         struct stripe_head *sh;
2250         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2251
2252         spin_lock_irq(conf->hash_locks + hash);
2253         sh = get_free_stripe(conf, hash);
2254         spin_unlock_irq(conf->hash_locks + hash);
2255         if (!sh)
2256                 return 0;
2257         BUG_ON(atomic_read(&sh->count));
2258         shrink_buffers(sh);
2259         kmem_cache_free(conf->slab_cache, sh);
2260         atomic_dec(&conf->active_stripes);
2261         conf->max_nr_stripes--;
2262         return 1;
2263 }
2264
2265 static void shrink_stripes(struct r5conf *conf)
2266 {
2267         while (conf->max_nr_stripes &&
2268                drop_one_stripe(conf))
2269                 ;
2270
2271         if (conf->slab_cache)
2272                 kmem_cache_destroy(conf->slab_cache);
2273         conf->slab_cache = NULL;
2274 }
2275
2276 static void raid5_end_read_request(struct bio * bi, int error)
2277 {
2278         struct stripe_head *sh = bi->bi_private;
2279         struct r5conf *conf = sh->raid_conf;
2280         int disks = sh->disks, i;
2281         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2282         char b[BDEVNAME_SIZE];
2283         struct md_rdev *rdev = NULL;
2284         sector_t s;
2285
2286         for (i=0 ; i<disks; i++)
2287                 if (bi == &sh->dev[i].req)
2288                         break;
2289
2290         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2291                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2292                 uptodate);
2293         if (i == disks) {
2294                 BUG();
2295                 return;
2296         }
2297         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2298                 /* If replacement finished while this request was outstanding,
2299                  * 'replacement' might be NULL already.
2300                  * In that case it moved down to 'rdev'.
2301                  * rdev is not removed until all requests are finished.
2302                  */
2303                 rdev = conf->disks[i].replacement;
2304         if (!rdev)
2305                 rdev = conf->disks[i].rdev;
2306
2307         if (use_new_offset(conf, sh))
2308                 s = sh->sector + rdev->new_data_offset;
2309         else
2310                 s = sh->sector + rdev->data_offset;
2311         if (uptodate) {
2312                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2313                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2314                         /* Note that this cannot happen on a
2315                          * replacement device.  We just fail those on
2316                          * any error
2317                          */
2318                         printk_ratelimited(
2319                                 KERN_INFO
2320                                 "md/raid:%s: read error corrected"
2321                                 " (%lu sectors at %llu on %s)\n",
2322                                 mdname(conf->mddev), STRIPE_SECTORS,
2323                                 (unsigned long long)s,
2324                                 bdevname(rdev->bdev, b));
2325                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2326                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2327                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2328                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2329                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2330
2331                 if (atomic_read(&rdev->read_errors))
2332                         atomic_set(&rdev->read_errors, 0);
2333         } else {
2334                 const char *bdn = bdevname(rdev->bdev, b);
2335                 int retry = 0;
2336                 int set_bad = 0;
2337
2338                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2339                 atomic_inc(&rdev->read_errors);
2340                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2341                         printk_ratelimited(
2342                                 KERN_WARNING
2343                                 "md/raid:%s: read error on replacement device "
2344                                 "(sector %llu on %s).\n",
2345                                 mdname(conf->mddev),
2346                                 (unsigned long long)s,
2347                                 bdn);
2348                 else if (conf->mddev->degraded >= conf->max_degraded) {
2349                         set_bad = 1;
2350                         printk_ratelimited(
2351                                 KERN_WARNING
2352                                 "md/raid:%s: read error not correctable "
2353                                 "(sector %llu on %s).\n",
2354                                 mdname(conf->mddev),
2355                                 (unsigned long long)s,
2356                                 bdn);
2357                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2358                         /* Oh, no!!! */
2359                         set_bad = 1;
2360                         printk_ratelimited(
2361                                 KERN_WARNING
2362                                 "md/raid:%s: read error NOT corrected!! "
2363                                 "(sector %llu on %s).\n",
2364                                 mdname(conf->mddev),
2365                                 (unsigned long long)s,
2366                                 bdn);
2367                 } else if (atomic_read(&rdev->read_errors)
2368                          > conf->max_nr_stripes)
2369                         printk(KERN_WARNING
2370                                "md/raid:%s: Too many read errors, failing device %s.\n",
2371                                mdname(conf->mddev), bdn);
2372                 else
2373                         retry = 1;
2374                 if (set_bad && test_bit(In_sync, &rdev->flags)
2375                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2376                         retry = 1;
2377                 if (retry)
2378                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2379                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2380                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2381                         } else
2382                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2383                 else {
2384                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2385                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2386                         if (!(set_bad
2387                               && test_bit(In_sync, &rdev->flags)
2388                               && rdev_set_badblocks(
2389                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
2390                                 md_error(conf->mddev, rdev);
2391                 }
2392         }
2393         rdev_dec_pending(rdev, conf->mddev);
2394         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2395         set_bit(STRIPE_HANDLE, &sh->state);
2396         release_stripe(sh);
2397 }
2398
2399 static void raid5_end_write_request(struct bio *bi, int error)
2400 {
2401         struct stripe_head *sh = bi->bi_private;
2402         struct r5conf *conf = sh->raid_conf;
2403         int disks = sh->disks, i;
2404         struct md_rdev *uninitialized_var(rdev);
2405         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2406         sector_t first_bad;
2407         int bad_sectors;
2408         int replacement = 0;
2409
2410         for (i = 0 ; i < disks; i++) {
2411                 if (bi == &sh->dev[i].req) {
2412                         rdev = conf->disks[i].rdev;
2413                         break;
2414                 }
2415                 if (bi == &sh->dev[i].rreq) {
2416                         rdev = conf->disks[i].replacement;
2417                         if (rdev)
2418                                 replacement = 1;
2419                         else
2420                                 /* rdev was removed and 'replacement'
2421                                  * replaced it.  rdev is not removed
2422                                  * until all requests are finished.
2423                                  */
2424                                 rdev = conf->disks[i].rdev;
2425                         break;
2426                 }
2427         }
2428         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2429                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2430                 uptodate);
2431         if (i == disks) {
2432                 BUG();
2433                 return;
2434         }
2435
2436         if (replacement) {
2437                 if (!uptodate)
2438                         md_error(conf->mddev, rdev);
2439                 else if (is_badblock(rdev, sh->sector,
2440                                      STRIPE_SECTORS,
2441                                      &first_bad, &bad_sectors))
2442                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2443         } else {
2444                 if (!uptodate) {
2445                         set_bit(STRIPE_DEGRADED, &sh->state);
2446                         set_bit(WriteErrorSeen, &rdev->flags);
2447                         set_bit(R5_WriteError, &sh->dev[i].flags);
2448                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2449                                 set_bit(MD_RECOVERY_NEEDED,
2450                                         &rdev->mddev->recovery);
2451                 } else if (is_badblock(rdev, sh->sector,
2452                                        STRIPE_SECTORS,
2453                                        &first_bad, &bad_sectors)) {
2454                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2455                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2456                                 /* That was a successful write so make
2457                                  * sure it looks like we already did
2458                                  * a re-write.
2459                                  */
2460                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2461                 }
2462         }
2463         rdev_dec_pending(rdev, conf->mddev);
2464
2465         if (sh->batch_head && !uptodate && !replacement)
2466                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2467
2468         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2469                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2470         set_bit(STRIPE_HANDLE, &sh->state);
2471         release_stripe(sh);
2472
2473         if (sh->batch_head && sh != sh->batch_head)
2474                 release_stripe(sh->batch_head);
2475 }
2476
2477 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2478
2479 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2480 {
2481         struct r5dev *dev = &sh->dev[i];
2482
2483         bio_init(&dev->req);
2484         dev->req.bi_io_vec = &dev->vec;
2485         dev->req.bi_max_vecs = 1;
2486         dev->req.bi_private = sh;
2487
2488         bio_init(&dev->rreq);
2489         dev->rreq.bi_io_vec = &dev->rvec;
2490         dev->rreq.bi_max_vecs = 1;
2491         dev->rreq.bi_private = sh;
2492
2493         dev->flags = 0;
2494         dev->sector = compute_blocknr(sh, i, previous);
2495 }
2496
2497 static void error(struct mddev *mddev, struct md_rdev *rdev)
2498 {
2499         char b[BDEVNAME_SIZE];
2500         struct r5conf *conf = mddev->private;
2501         unsigned long flags;
2502         pr_debug("raid456: error called\n");
2503
2504         spin_lock_irqsave(&conf->device_lock, flags);
2505         clear_bit(In_sync, &rdev->flags);
2506         mddev->degraded = calc_degraded(conf);
2507         spin_unlock_irqrestore(&conf->device_lock, flags);
2508         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2509
2510         set_bit(Blocked, &rdev->flags);
2511         set_bit(Faulty, &rdev->flags);
2512         set_bit(MD_CHANGE_DEVS, &mddev->flags);
2513         printk(KERN_ALERT
2514                "md/raid:%s: Disk failure on %s, disabling device.\n"
2515                "md/raid:%s: Operation continuing on %d devices.\n",
2516                mdname(mddev),
2517                bdevname(rdev->bdev, b),
2518                mdname(mddev),
2519                conf->raid_disks - mddev->degraded);
2520 }
2521
2522 /*
2523  * Input: a 'big' sector number,
2524  * Output: index of the data and parity disk, and the sector # in them.
2525  */
2526 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2527                                      int previous, int *dd_idx,
2528                                      struct stripe_head *sh)
2529 {
2530         sector_t stripe, stripe2;
2531         sector_t chunk_number;
2532         unsigned int chunk_offset;
2533         int pd_idx, qd_idx;
2534         int ddf_layout = 0;
2535         sector_t new_sector;
2536         int algorithm = previous ? conf->prev_algo
2537                                  : conf->algorithm;
2538         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2539                                          : conf->chunk_sectors;
2540         int raid_disks = previous ? conf->previous_raid_disks
2541                                   : conf->raid_disks;
2542         int data_disks = raid_disks - conf->max_degraded;
2543
2544         /* First compute the information on this sector */
2545
2546         /*
2547          * Compute the chunk number and the sector offset inside the chunk
2548          */
2549         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2550         chunk_number = r_sector;
2551
2552         /*
2553          * Compute the stripe number
2554          */
2555         stripe = chunk_number;
2556         *dd_idx = sector_div(stripe, data_disks);
2557         stripe2 = stripe;
2558         /*
2559          * Select the parity disk based on the user selected algorithm.
2560          */
2561         pd_idx = qd_idx = -1;
2562         switch(conf->level) {
2563         case 4:
2564                 pd_idx = data_disks;
2565                 break;
2566         case 5:
2567                 switch (algorithm) {
2568                 case ALGORITHM_LEFT_ASYMMETRIC:
2569                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2570                         if (*dd_idx >= pd_idx)
2571                                 (*dd_idx)++;
2572                         break;
2573                 case ALGORITHM_RIGHT_ASYMMETRIC:
2574                         pd_idx = sector_div(stripe2, raid_disks);
2575                         if (*dd_idx >= pd_idx)
2576                                 (*dd_idx)++;
2577                         break;
2578                 case ALGORITHM_LEFT_SYMMETRIC:
2579                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2580                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2581                         break;
2582                 case ALGORITHM_RIGHT_SYMMETRIC:
2583                         pd_idx = sector_div(stripe2, raid_disks);
2584                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2585                         break;
2586                 case ALGORITHM_PARITY_0:
2587                         pd_idx = 0;
2588                         (*dd_idx)++;
2589                         break;
2590                 case ALGORITHM_PARITY_N:
2591                         pd_idx = data_disks;
2592                         break;
2593                 default:
2594                         BUG();
2595                 }
2596                 break;
2597         case 6:
2598
2599                 switch (algorithm) {
2600                 case ALGORITHM_LEFT_ASYMMETRIC:
2601                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2602                         qd_idx = pd_idx + 1;
2603                         if (pd_idx == raid_disks-1) {
2604                                 (*dd_idx)++;    /* Q D D D P */
2605                                 qd_idx = 0;
2606                         } else if (*dd_idx >= pd_idx)
2607                                 (*dd_idx) += 2; /* D D P Q D */
2608                         break;
2609                 case ALGORITHM_RIGHT_ASYMMETRIC:
2610                         pd_idx = sector_div(stripe2, raid_disks);
2611                         qd_idx = pd_idx + 1;
2612                         if (pd_idx == raid_disks-1) {
2613                                 (*dd_idx)++;    /* Q D D D P */
2614                                 qd_idx = 0;
2615                         } else if (*dd_idx >= pd_idx)
2616                                 (*dd_idx) += 2; /* D D P Q D */
2617                         break;
2618                 case ALGORITHM_LEFT_SYMMETRIC:
2619                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2620                         qd_idx = (pd_idx + 1) % raid_disks;
2621                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2622                         break;
2623                 case ALGORITHM_RIGHT_SYMMETRIC:
2624                         pd_idx = sector_div(stripe2, raid_disks);
2625                         qd_idx = (pd_idx + 1) % raid_disks;
2626                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2627                         break;
2628
2629                 case ALGORITHM_PARITY_0:
2630                         pd_idx = 0;
2631                         qd_idx = 1;
2632                         (*dd_idx) += 2;
2633                         break;
2634                 case ALGORITHM_PARITY_N:
2635                         pd_idx = data_disks;
2636                         qd_idx = data_disks + 1;
2637                         break;
2638
2639                 case ALGORITHM_ROTATING_ZERO_RESTART:
2640                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2641                          * of blocks for computing Q is different.
2642                          */
2643                         pd_idx = sector_div(stripe2, raid_disks);
2644                         qd_idx = pd_idx + 1;
2645                         if (pd_idx == raid_disks-1) {
2646                                 (*dd_idx)++;    /* Q D D D P */
2647                                 qd_idx = 0;
2648                         } else if (*dd_idx >= pd_idx)
2649                                 (*dd_idx) += 2; /* D D P Q D */
2650                         ddf_layout = 1;
2651                         break;
2652
2653                 case ALGORITHM_ROTATING_N_RESTART:
2654                         /* Same a left_asymmetric, by first stripe is
2655                          * D D D P Q  rather than
2656                          * Q D D D P
2657                          */
2658                         stripe2 += 1;
2659                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2660                         qd_idx = pd_idx + 1;
2661                         if (pd_idx == raid_disks-1) {
2662                                 (*dd_idx)++;    /* Q D D D P */
2663                                 qd_idx = 0;
2664                         } else if (*dd_idx >= pd_idx)
2665                                 (*dd_idx) += 2; /* D D P Q D */
2666                         ddf_layout = 1;
2667                         break;
2668
2669                 case ALGORITHM_ROTATING_N_CONTINUE:
2670                         /* Same as left_symmetric but Q is before P */
2671                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2672                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2673                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2674                         ddf_layout = 1;
2675                         break;
2676
2677                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2678                         /* RAID5 left_asymmetric, with Q on last device */
2679                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2680                         if (*dd_idx >= pd_idx)
2681                                 (*dd_idx)++;
2682                         qd_idx = raid_disks - 1;
2683                         break;
2684
2685                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2686                         pd_idx = sector_div(stripe2, raid_disks-1);
2687                         if (*dd_idx >= pd_idx)
2688                                 (*dd_idx)++;
2689                         qd_idx = raid_disks - 1;
2690                         break;
2691
2692                 case ALGORITHM_LEFT_SYMMETRIC_6:
2693                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2694                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2695                         qd_idx = raid_disks - 1;
2696                         break;
2697
2698                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2699                         pd_idx = sector_div(stripe2, raid_disks-1);
2700                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2701                         qd_idx = raid_disks - 1;
2702                         break;
2703
2704                 case ALGORITHM_PARITY_0_6:
2705                         pd_idx = 0;
2706                         (*dd_idx)++;
2707                         qd_idx = raid_disks - 1;
2708                         break;
2709
2710                 default:
2711                         BUG();
2712                 }
2713                 break;
2714         }
2715
2716         if (sh) {
2717                 sh->pd_idx = pd_idx;
2718                 sh->qd_idx = qd_idx;
2719                 sh->ddf_layout = ddf_layout;
2720         }
2721         /*
2722          * Finally, compute the new sector number
2723          */
2724         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2725         return new_sector;
2726 }
2727
2728 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2729 {
2730         struct r5conf *conf = sh->raid_conf;
2731         int raid_disks = sh->disks;
2732         int data_disks = raid_disks - conf->max_degraded;
2733         sector_t new_sector = sh->sector, check;
2734         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2735                                          : conf->chunk_sectors;
2736         int algorithm = previous ? conf->prev_algo
2737                                  : conf->algorithm;
2738         sector_t stripe;
2739         int chunk_offset;
2740         sector_t chunk_number;
2741         int dummy1, dd_idx = i;
2742         sector_t r_sector;
2743         struct stripe_head sh2;
2744
2745         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2746         stripe = new_sector;
2747
2748         if (i == sh->pd_idx)
2749                 return 0;
2750         switch(conf->level) {
2751         case 4: break;
2752         case 5:
2753                 switch (algorithm) {
2754                 case ALGORITHM_LEFT_ASYMMETRIC:
2755                 case ALGORITHM_RIGHT_ASYMMETRIC:
2756                         if (i > sh->pd_idx)
2757                                 i--;
2758                         break;
2759                 case ALGORITHM_LEFT_SYMMETRIC:
2760                 case ALGORITHM_RIGHT_SYMMETRIC:
2761                         if (i < sh->pd_idx)
2762                                 i += raid_disks;
2763                         i -= (sh->pd_idx + 1);
2764                         break;
2765                 case ALGORITHM_PARITY_0:
2766                         i -= 1;
2767                         break;
2768                 case ALGORITHM_PARITY_N:
2769                         break;
2770                 default:
2771                         BUG();
2772                 }
2773                 break;
2774         case 6:
2775                 if (i == sh->qd_idx)
2776                         return 0; /* It is the Q disk */
2777                 switch (algorithm) {
2778                 case ALGORITHM_LEFT_ASYMMETRIC:
2779                 case ALGORITHM_RIGHT_ASYMMETRIC:
2780                 case ALGORITHM_ROTATING_ZERO_RESTART:
2781                 case ALGORITHM_ROTATING_N_RESTART:
2782                         if (sh->pd_idx == raid_disks-1)
2783                                 i--;    /* Q D D D P */
2784                         else if (i > sh->pd_idx)
2785                                 i -= 2; /* D D P Q D */
2786                         break;
2787                 case ALGORITHM_LEFT_SYMMETRIC:
2788                 case ALGORITHM_RIGHT_SYMMETRIC:
2789                         if (sh->pd_idx == raid_disks-1)
2790                                 i--; /* Q D D D P */
2791                         else {
2792                                 /* D D P Q D */
2793                                 if (i < sh->pd_idx)
2794                                         i += raid_disks;
2795                                 i -= (sh->pd_idx + 2);
2796                         }
2797                         break;
2798                 case ALGORITHM_PARITY_0:
2799                         i -= 2;
2800                         break;
2801                 case ALGORITHM_PARITY_N:
2802                         break;
2803                 case ALGORITHM_ROTATING_N_CONTINUE:
2804                         /* Like left_symmetric, but P is before Q */
2805                         if (sh->pd_idx == 0)
2806                                 i--;    /* P D D D Q */
2807                         else {
2808                                 /* D D Q P D */
2809                                 if (i < sh->pd_idx)
2810                                         i += raid_disks;
2811                                 i -= (sh->pd_idx + 1);
2812                         }
2813                         break;
2814                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2815                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2816                         if (i > sh->pd_idx)
2817                                 i--;
2818                         break;
2819                 case ALGORITHM_LEFT_SYMMETRIC_6:
2820                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2821                         if (i < sh->pd_idx)
2822                                 i += data_disks + 1;
2823                         i -= (sh->pd_idx + 1);
2824                         break;
2825                 case ALGORITHM_PARITY_0_6:
2826                         i -= 1;
2827                         break;
2828                 default:
2829                         BUG();
2830                 }
2831                 break;
2832         }
2833
2834         chunk_number = stripe * data_disks + i;
2835         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2836
2837         check = raid5_compute_sector(conf, r_sector,
2838                                      previous, &dummy1, &sh2);
2839         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2840                 || sh2.qd_idx != sh->qd_idx) {
2841                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2842                        mdname(conf->mddev));
2843                 return 0;
2844         }
2845         return r_sector;
2846 }
2847
2848 static void
2849 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2850                          int rcw, int expand)
2851 {
2852         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2853         struct r5conf *conf = sh->raid_conf;
2854         int level = conf->level;
2855
2856         if (rcw) {
2857
2858                 for (i = disks; i--; ) {
2859                         struct r5dev *dev = &sh->dev[i];
2860
2861                         if (dev->towrite) {
2862                                 set_bit(R5_LOCKED, &dev->flags);
2863                                 set_bit(R5_Wantdrain, &dev->flags);
2864                                 if (!expand)
2865                                         clear_bit(R5_UPTODATE, &dev->flags);
2866                                 s->locked++;
2867                         }
2868                 }
2869                 /* if we are not expanding this is a proper write request, and
2870                  * there will be bios with new data to be drained into the
2871                  * stripe cache
2872                  */
2873                 if (!expand) {
2874                         if (!s->locked)
2875                                 /* False alarm, nothing to do */
2876                                 return;
2877                         sh->reconstruct_state = reconstruct_state_drain_run;
2878                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2879                 } else
2880                         sh->reconstruct_state = reconstruct_state_run;
2881
2882                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2883
2884                 if (s->locked + conf->max_degraded == disks)
2885                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2886                                 atomic_inc(&conf->pending_full_writes);
2887         } else {
2888                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2889                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2890                 BUG_ON(level == 6 &&
2891                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2892                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2893
2894                 for (i = disks; i--; ) {
2895                         struct r5dev *dev = &sh->dev[i];
2896                         if (i == pd_idx || i == qd_idx)
2897                                 continue;
2898
2899                         if (dev->towrite &&
2900                             (test_bit(R5_UPTODATE, &dev->flags) ||
2901                              test_bit(R5_Wantcompute, &dev->flags))) {
2902                                 set_bit(R5_Wantdrain, &dev->flags);
2903                                 set_bit(R5_LOCKED, &dev->flags);
2904                                 clear_bit(R5_UPTODATE, &dev->flags);
2905                                 s->locked++;
2906                         }
2907                 }
2908                 if (!s->locked)
2909                         /* False alarm - nothing to do */
2910                         return;
2911                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2912                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2913                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2914                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2915         }
2916
2917         /* keep the parity disk(s) locked while asynchronous operations
2918          * are in flight
2919          */
2920         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2921         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2922         s->locked++;
2923
2924         if (level == 6) {
2925                 int qd_idx = sh->qd_idx;
2926                 struct r5dev *dev = &sh->dev[qd_idx];
2927
2928                 set_bit(R5_LOCKED, &dev->flags);
2929                 clear_bit(R5_UPTODATE, &dev->flags);
2930                 s->locked++;
2931         }
2932
2933         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2934                 __func__, (unsigned long long)sh->sector,
2935                 s->locked, s->ops_request);
2936 }
2937
2938 /*
2939  * Each stripe/dev can have one or more bion attached.
2940  * toread/towrite point to the first in a chain.
2941  * The bi_next chain must be in order.
2942  */
2943 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2944                           int forwrite, int previous)
2945 {
2946         struct bio **bip;
2947         struct r5conf *conf = sh->raid_conf;
2948         int firstwrite=0;
2949
2950         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2951                 (unsigned long long)bi->bi_iter.bi_sector,
2952                 (unsigned long long)sh->sector);
2953
2954         /*
2955          * If several bio share a stripe. The bio bi_phys_segments acts as a
2956          * reference count to avoid race. The reference count should already be
2957          * increased before this function is called (for example, in
2958          * make_request()), so other bio sharing this stripe will not free the
2959          * stripe. If a stripe is owned by one stripe, the stripe lock will
2960          * protect it.
2961          */
2962         spin_lock_irq(&sh->stripe_lock);
2963         /* Don't allow new IO added to stripes in batch list */
2964         if (sh->batch_head)
2965                 goto overlap;
2966         if (forwrite) {
2967                 bip = &sh->dev[dd_idx].towrite;
2968                 if (*bip == NULL)
2969                         firstwrite = 1;
2970         } else
2971                 bip = &sh->dev[dd_idx].toread;
2972         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2973                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2974                         goto overlap;
2975                 bip = & (*bip)->bi_next;
2976         }
2977         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2978                 goto overlap;
2979
2980         if (!forwrite || previous)
2981                 clear_bit(STRIPE_BATCH_READY, &sh->state);
2982
2983         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2984         if (*bip)
2985                 bi->bi_next = *bip;
2986         *bip = bi;
2987         raid5_inc_bi_active_stripes(bi);
2988
2989         if (forwrite) {
2990                 /* check if page is covered */
2991                 sector_t sector = sh->dev[dd_idx].sector;
2992                 for (bi=sh->dev[dd_idx].towrite;
2993                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2994                              bi && bi->bi_iter.bi_sector <= sector;
2995                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2996                         if (bio_end_sector(bi) >= sector)
2997                                 sector = bio_end_sector(bi);
2998                 }
2999                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3000                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3001                                 sh->overwrite_disks++;
3002         }
3003
3004         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3005                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3006                 (unsigned long long)sh->sector, dd_idx);
3007
3008         if (conf->mddev->bitmap && firstwrite) {
3009                 /* Cannot hold spinlock over bitmap_startwrite,
3010                  * but must ensure this isn't added to a batch until
3011                  * we have added to the bitmap and set bm_seq.
3012                  * So set STRIPE_BITMAP_PENDING to prevent
3013                  * batching.
3014                  * If multiple add_stripe_bio() calls race here they
3015                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3016                  * to complete "bitmap_startwrite" gets to set
3017                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3018                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3019                  * any more.
3020                  */
3021                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3022                 spin_unlock_irq(&sh->stripe_lock);
3023                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3024                                   STRIPE_SECTORS, 0);
3025                 spin_lock_irq(&sh->stripe_lock);
3026                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3027                 if (!sh->batch_head) {
3028                         sh->bm_seq = conf->seq_flush+1;
3029                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3030                 }
3031         }
3032         spin_unlock_irq(&sh->stripe_lock);
3033
3034         if (stripe_can_batch(sh))
3035                 stripe_add_to_batch_list(conf, sh);
3036         return 1;
3037
3038  overlap:
3039         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3040         spin_unlock_irq(&sh->stripe_lock);
3041         return 0;
3042 }
3043
3044 static void end_reshape(struct r5conf *conf);
3045
3046 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3047                             struct stripe_head *sh)
3048 {
3049         int sectors_per_chunk =
3050                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3051         int dd_idx;
3052         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3053         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3054
3055         raid5_compute_sector(conf,
3056                              stripe * (disks - conf->max_degraded)
3057                              *sectors_per_chunk + chunk_offset,
3058                              previous,
3059                              &dd_idx, sh);
3060 }
3061
3062 static void
3063 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3064                                 struct stripe_head_state *s, int disks,
3065                                 struct bio **return_bi)
3066 {
3067         int i;
3068         BUG_ON(sh->batch_head);
3069         for (i = disks; i--; ) {
3070                 struct bio *bi;
3071                 int bitmap_end = 0;
3072
3073                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3074                         struct md_rdev *rdev;
3075                         rcu_read_lock();
3076                         rdev = rcu_dereference(conf->disks[i].rdev);
3077                         if (rdev && test_bit(In_sync, &rdev->flags))
3078                                 atomic_inc(&rdev->nr_pending);
3079                         else
3080                                 rdev = NULL;
3081                         rcu_read_unlock();
3082                         if (rdev) {
3083                                 if (!rdev_set_badblocks(
3084                                             rdev,
3085                                             sh->sector,
3086                                             STRIPE_SECTORS, 0))
3087                                         md_error(conf->mddev, rdev);
3088                                 rdev_dec_pending(rdev, conf->mddev);
3089                         }
3090                 }
3091                 spin_lock_irq(&sh->stripe_lock);
3092                 /* fail all writes first */
3093                 bi = sh->dev[i].towrite;
3094                 sh->dev[i].towrite = NULL;
3095                 sh->overwrite_disks = 0;
3096                 spin_unlock_irq(&sh->stripe_lock);
3097                 if (bi)
3098                         bitmap_end = 1;
3099
3100                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3101                         wake_up(&conf->wait_for_overlap);
3102
3103                 while (bi && bi->bi_iter.bi_sector <
3104                         sh->dev[i].sector + STRIPE_SECTORS) {
3105                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3106                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3107                         if (!raid5_dec_bi_active_stripes(bi)) {
3108                                 md_write_end(conf->mddev);
3109                                 bi->bi_next = *return_bi;
3110                                 *return_bi = bi;
3111                         }
3112                         bi = nextbi;
3113                 }
3114                 if (bitmap_end)
3115                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3116                                 STRIPE_SECTORS, 0, 0);
3117                 bitmap_end = 0;
3118                 /* and fail all 'written' */
3119                 bi = sh->dev[i].written;
3120                 sh->dev[i].written = NULL;
3121                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3122                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3123                         sh->dev[i].page = sh->dev[i].orig_page;
3124                 }
3125
3126                 if (bi) bitmap_end = 1;
3127                 while (bi && bi->bi_iter.bi_sector <
3128                        sh->dev[i].sector + STRIPE_SECTORS) {
3129                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3130                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
3131                         if (!raid5_dec_bi_active_stripes(bi)) {
3132                                 md_write_end(conf->mddev);
3133                                 bi->bi_next = *return_bi;
3134                                 *return_bi = bi;
3135                         }
3136                         bi = bi2;
3137                 }
3138
3139                 /* fail any reads if this device is non-operational and
3140                  * the data has not reached the cache yet.
3141                  */
3142                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3143                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3144                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3145                         spin_lock_irq(&sh->stripe_lock);
3146                         bi = sh->dev[i].toread;
3147                         sh->dev[i].toread = NULL;
3148                         spin_unlock_irq(&sh->stripe_lock);
3149                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3150                                 wake_up(&conf->wait_for_overlap);
3151                         while (bi && bi->bi_iter.bi_sector <
3152                                sh->dev[i].sector + STRIPE_SECTORS) {
3153                                 struct bio *nextbi =
3154                                         r5_next_bio(bi, sh->dev[i].sector);
3155                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3156                                 if (!raid5_dec_bi_active_stripes(bi)) {
3157                                         bi->bi_next = *return_bi;
3158                                         *return_bi = bi;
3159                                 }
3160                                 bi = nextbi;
3161                         }
3162                 }
3163                 if (bitmap_end)
3164                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3165                                         STRIPE_SECTORS, 0, 0);
3166                 /* If we were in the middle of a write the parity block might
3167                  * still be locked - so just clear all R5_LOCKED flags
3168                  */
3169                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3170         }
3171
3172         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3173                 if (atomic_dec_and_test(&conf->pending_full_writes))
3174                         md_wakeup_thread(conf->mddev->thread);
3175 }
3176
3177 static void
3178 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3179                    struct stripe_head_state *s)
3180 {
3181         int abort = 0;
3182         int i;
3183
3184         BUG_ON(sh->batch_head);
3185         clear_bit(STRIPE_SYNCING, &sh->state);
3186         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3187                 wake_up(&conf->wait_for_overlap);
3188         s->syncing = 0;
3189         s->replacing = 0;
3190         /* There is nothing more to do for sync/check/repair.
3191          * Don't even need to abort as that is handled elsewhere
3192          * if needed, and not always wanted e.g. if there is a known
3193          * bad block here.
3194          * For recover/replace we need to record a bad block on all
3195          * non-sync devices, or abort the recovery
3196          */
3197         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3198                 /* During recovery devices cannot be removed, so
3199                  * locking and refcounting of rdevs is not needed
3200                  */
3201                 for (i = 0; i < conf->raid_disks; i++) {
3202                         struct md_rdev *rdev = conf->disks[i].rdev;
3203                         if (rdev
3204                             && !test_bit(Faulty, &rdev->flags)
3205                             && !test_bit(In_sync, &rdev->flags)
3206                             && !rdev_set_badblocks(rdev, sh->sector,
3207                                                    STRIPE_SECTORS, 0))
3208                                 abort = 1;
3209                         rdev = conf->disks[i].replacement;
3210                         if (rdev
3211                             && !test_bit(Faulty, &rdev->flags)
3212                             && !test_bit(In_sync, &rdev->flags)
3213                             && !rdev_set_badblocks(rdev, sh->sector,
3214                                                    STRIPE_SECTORS, 0))
3215                                 abort = 1;
3216                 }
3217                 if (abort)
3218                         conf->recovery_disabled =
3219                                 conf->mddev->recovery_disabled;
3220         }
3221         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3222 }
3223
3224 static int want_replace(struct stripe_head *sh, int disk_idx)
3225 {
3226         struct md_rdev *rdev;
3227         int rv = 0;
3228         /* Doing recovery so rcu locking not required */
3229         rdev = sh->raid_conf->disks[disk_idx].replacement;
3230         if (rdev
3231             && !test_bit(Faulty, &rdev->flags)
3232             && !test_bit(In_sync, &rdev->flags)
3233             && (rdev->recovery_offset <= sh->sector
3234                 || rdev->mddev->recovery_cp <= sh->sector))
3235                 rv = 1;
3236
3237         return rv;
3238 }
3239
3240 /* fetch_block - checks the given member device to see if its data needs
3241  * to be read or computed to satisfy a request.
3242  *
3243  * Returns 1 when no more member devices need to be checked, otherwise returns
3244  * 0 to tell the loop in handle_stripe_fill to continue
3245  */
3246
3247 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3248                            int disk_idx, int disks)
3249 {
3250         struct r5dev *dev = &sh->dev[disk_idx];
3251         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3252                                   &sh->dev[s->failed_num[1]] };
3253         int i;
3254
3255
3256         if (test_bit(R5_LOCKED, &dev->flags) ||
3257             test_bit(R5_UPTODATE, &dev->flags))
3258                 /* No point reading this as we already have it or have
3259                  * decided to get it.
3260                  */
3261                 return 0;
3262
3263         if (dev->toread ||
3264             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3265                 /* We need this block to directly satisfy a request */
3266                 return 1;
3267
3268         if (s->syncing || s->expanding ||
3269             (s->replacing && want_replace(sh, disk_idx)))
3270                 /* When syncing, or expanding we read everything.
3271                  * When replacing, we need the replaced block.
3272                  */
3273                 return 1;
3274
3275         if ((s->failed >= 1 && fdev[0]->toread) ||
3276             (s->failed >= 2 && fdev[1]->toread))
3277                 /* If we want to read from a failed device, then
3278                  * we need to actually read every other device.
3279                  */
3280                 return 1;
3281
3282         /* Sometimes neither read-modify-write nor reconstruct-write
3283          * cycles can work.  In those cases we read every block we
3284          * can.  Then the parity-update is certain to have enough to
3285          * work with.
3286          * This can only be a problem when we need to write something,
3287          * and some device has failed.  If either of those tests
3288          * fail we need look no further.
3289          */
3290         if (!s->failed || !s->to_write)
3291                 return 0;
3292
3293         if (test_bit(R5_Insync, &dev->flags) &&
3294             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3295                 /* Pre-reads at not permitted until after short delay
3296                  * to gather multiple requests.  However if this
3297                  * device is no Insync, the block could only be be computed
3298                  * and there is no need to delay that.
3299                  */
3300                 return 0;
3301
3302         for (i = 0; i < s->failed; i++) {
3303                 if (fdev[i]->towrite &&
3304                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3305                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3306                         /* If we have a partial write to a failed
3307                          * device, then we will need to reconstruct
3308                          * the content of that device, so all other
3309                          * devices must be read.
3310                          */
3311                         return 1;
3312         }
3313
3314         /* If we are forced to do a reconstruct-write, either because
3315          * the current RAID6 implementation only supports that, or
3316          * or because parity cannot be trusted and we are currently
3317          * recovering it, there is extra need to be careful.
3318          * If one of the devices that we would need to read, because
3319          * it is not being overwritten (and maybe not written at all)
3320          * is missing/faulty, then we need to read everything we can.
3321          */
3322         if (sh->raid_conf->level != 6 &&
3323             sh->sector < sh->raid_conf->mddev->recovery_cp)
3324                 /* reconstruct-write isn't being forced */
3325                 return 0;
3326         for (i = 0; i < s->failed; i++) {
3327                 if (s->failed_num[i] != sh->pd_idx &&
3328                     s->failed_num[i] != sh->qd_idx &&
3329                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3330                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3331                         return 1;
3332         }
3333
3334         return 0;
3335 }
3336
3337 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3338                        int disk_idx, int disks)
3339 {
3340         struct r5dev *dev = &sh->dev[disk_idx];
3341
3342         /* is the data in this block needed, and can we get it? */
3343         if (need_this_block(sh, s, disk_idx, disks)) {
3344                 /* we would like to get this block, possibly by computing it,
3345                  * otherwise read it if the backing disk is insync
3346                  */
3347                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3348                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3349                 BUG_ON(sh->batch_head);
3350                 if ((s->uptodate == disks - 1) &&
3351                     (s->failed && (disk_idx == s->failed_num[0] ||
3352                                    disk_idx == s->failed_num[1]))) {
3353                         /* have disk failed, and we're requested to fetch it;
3354                          * do compute it
3355                          */
3356                         pr_debug("Computing stripe %llu block %d\n",
3357                                (unsigned long long)sh->sector, disk_idx);
3358                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3359                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3360                         set_bit(R5_Wantcompute, &dev->flags);
3361                         sh->ops.target = disk_idx;
3362                         sh->ops.target2 = -1; /* no 2nd target */
3363                         s->req_compute = 1;
3364                         /* Careful: from this point on 'uptodate' is in the eye
3365                          * of raid_run_ops which services 'compute' operations
3366                          * before writes. R5_Wantcompute flags a block that will
3367                          * be R5_UPTODATE by the time it is needed for a
3368                          * subsequent operation.
3369                          */
3370                         s->uptodate++;
3371                         return 1;
3372                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3373                         /* Computing 2-failure is *very* expensive; only
3374                          * do it if failed >= 2
3375                          */
3376                         int other;
3377                         for (other = disks; other--; ) {
3378                                 if (other == disk_idx)
3379                                         continue;
3380                                 if (!test_bit(R5_UPTODATE,
3381                                       &sh->dev[other].flags))
3382                                         break;
3383                         }
3384                         BUG_ON(other < 0);
3385                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3386                                (unsigned long long)sh->sector,
3387                                disk_idx, other);
3388                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3389                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3390                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3391                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3392                         sh->ops.target = disk_idx;
3393                         sh->ops.target2 = other;
3394                         s->uptodate += 2;
3395                         s->req_compute = 1;
3396                         return 1;
3397                 } else if (test_bit(R5_Insync, &dev->flags)) {
3398                         set_bit(R5_LOCKED, &dev->flags);
3399                         set_bit(R5_Wantread, &dev->flags);
3400                         s->locked++;
3401                         pr_debug("Reading block %d (sync=%d)\n",
3402                                 disk_idx, s->syncing);
3403                 }
3404         }
3405
3406         return 0;
3407 }
3408
3409 /**
3410  * handle_stripe_fill - read or compute data to satisfy pending requests.
3411  */
3412 static void handle_stripe_fill(struct stripe_head *sh,
3413                                struct stripe_head_state *s,
3414                                int disks)
3415 {
3416         int i;
3417
3418         /* look for blocks to read/compute, skip this if a compute
3419          * is already in flight, or if the stripe contents are in the
3420          * midst of changing due to a write
3421          */
3422         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3423             !sh->reconstruct_state)
3424                 for (i = disks; i--; )
3425                         if (fetch_block(sh, s, i, disks))
3426                                 break;
3427         set_bit(STRIPE_HANDLE, &sh->state);
3428 }
3429
3430 static void break_stripe_batch_list(struct stripe_head *head_sh,
3431                                     unsigned long handle_flags);
3432 /* handle_stripe_clean_event
3433  * any written block on an uptodate or failed drive can be returned.
3434  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3435  * never LOCKED, so we don't need to test 'failed' directly.
3436  */
3437 static void handle_stripe_clean_event(struct r5conf *conf,
3438         struct stripe_head *sh, int disks, struct bio **return_bi)
3439 {
3440         int i;
3441         struct r5dev *dev;
3442         int discard_pending = 0;
3443         struct stripe_head *head_sh = sh;
3444         bool do_endio = false;
3445
3446         for (i = disks; i--; )
3447                 if (sh->dev[i].written) {
3448                         dev = &sh->dev[i];
3449                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3450                             (test_bit(R5_UPTODATE, &dev->flags) ||
3451                              test_bit(R5_Discard, &dev->flags) ||
3452                              test_bit(R5_SkipCopy, &dev->flags))) {
3453                                 /* We can return any write requests */
3454                                 struct bio *wbi, *wbi2;
3455                                 pr_debug("Return write for disc %d\n", i);
3456                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3457                                         clear_bit(R5_UPTODATE, &dev->flags);
3458                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3459                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3460                                 }
3461                                 do_endio = true;
3462
3463 returnbi:
3464                                 dev->page = dev->orig_page;
3465                                 wbi = dev->written;
3466                                 dev->written = NULL;
3467                                 while (wbi && wbi->bi_iter.bi_sector <
3468                                         dev->sector + STRIPE_SECTORS) {
3469                                         wbi2 = r5_next_bio(wbi, dev->sector);
3470                                         if (!raid5_dec_bi_active_stripes(wbi)) {
3471                                                 md_write_end(conf->mddev);
3472                                                 wbi->bi_next = *return_bi;
3473                                                 *return_bi = wbi;
3474                                         }
3475                                         wbi = wbi2;
3476                                 }
3477                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3478                                                 STRIPE_SECTORS,
3479                                          !test_bit(STRIPE_DEGRADED, &sh->state),
3480                                                 0);
3481                                 if (head_sh->batch_head) {
3482                                         sh = list_first_entry(&sh->batch_list,
3483                                                               struct stripe_head,
3484                                                               batch_list);
3485                                         if (sh != head_sh) {
3486                                                 dev = &sh->dev[i];
3487                                                 goto returnbi;
3488                                         }
3489                                 }
3490                                 sh = head_sh;
3491                                 dev = &sh->dev[i];
3492                         } else if (test_bit(R5_Discard, &dev->flags))
3493                                 discard_pending = 1;
3494                         WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3495                         WARN_ON(dev->page != dev->orig_page);
3496                 }
3497         if (!discard_pending &&
3498             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3499                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3500                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3501                 if (sh->qd_idx >= 0) {
3502                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3503                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3504                 }
3505                 /* now that discard is done we can proceed with any sync */
3506                 clear_bit(STRIPE_DISCARD, &sh->state);
3507                 /*
3508                  * SCSI discard will change some bio fields and the stripe has
3509                  * no updated data, so remove it from hash list and the stripe
3510                  * will be reinitialized
3511                  */
3512                 spin_lock_irq(&conf->device_lock);
3513 unhash:
3514                 remove_hash(sh);
3515                 if (head_sh->batch_head) {
3516                         sh = list_first_entry(&sh->batch_list,
3517                                               struct stripe_head, batch_list);
3518                         if (sh != head_sh)
3519                                         goto unhash;
3520                 }
3521                 spin_unlock_irq(&conf->device_lock);
3522                 sh = head_sh;
3523
3524                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3525                         set_bit(STRIPE_HANDLE, &sh->state);
3526
3527         }
3528
3529         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3530                 if (atomic_dec_and_test(&conf->pending_full_writes))
3531                         md_wakeup_thread(conf->mddev->thread);
3532
3533         if (head_sh->batch_head && do_endio)
3534                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3535 }
3536
3537 static void handle_stripe_dirtying(struct r5conf *conf,
3538                                    struct stripe_head *sh,
3539                                    struct stripe_head_state *s,
3540                                    int disks)
3541 {
3542         int rmw = 0, rcw = 0, i;
3543         sector_t recovery_cp = conf->mddev->recovery_cp;
3544
3545         /* Check whether resync is now happening or should start.
3546          * If yes, then the array is dirty (after unclean shutdown or
3547          * initial creation), so parity in some stripes might be inconsistent.
3548          * In this case, we need to always do reconstruct-write, to ensure
3549          * that in case of drive failure or read-error correction, we
3550          * generate correct data from the parity.
3551          */
3552         if (conf->rmw_level == PARITY_DISABLE_RMW ||
3553             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3554              s->failed == 0)) {
3555                 /* Calculate the real rcw later - for now make it
3556                  * look like rcw is cheaper
3557                  */
3558                 rcw = 1; rmw = 2;
3559                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3560                          conf->rmw_level, (unsigned long long)recovery_cp,
3561                          (unsigned long long)sh->sector);
3562         } else for (i = disks; i--; ) {
3563                 /* would I have to read this buffer for read_modify_write */
3564                 struct r5dev *dev = &sh->dev[i];
3565                 if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3566                     !test_bit(R5_LOCKED, &dev->flags) &&
3567                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3568                       test_bit(R5_Wantcompute, &dev->flags))) {
3569                         if (test_bit(R5_Insync, &dev->flags))
3570                                 rmw++;
3571                         else
3572                                 rmw += 2*disks;  /* cannot read it */
3573                 }
3574                 /* Would I have to read this buffer for reconstruct_write */
3575                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3576                     i != sh->pd_idx && i != sh->qd_idx &&
3577                     !test_bit(R5_LOCKED, &dev->flags) &&
3578                     !(test_bit(R5_UPTODATE, &dev->flags) ||
3579                     test_bit(R5_Wantcompute, &dev->flags))) {
3580                         if (test_bit(R5_Insync, &dev->flags))
3581                                 rcw++;
3582                         else
3583                                 rcw += 2*disks;
3584                 }
3585         }
3586         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3587                 (unsigned long long)sh->sector, rmw, rcw);
3588         set_bit(STRIPE_HANDLE, &sh->state);
3589         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3590                 /* prefer read-modify-write, but need to get some data */
3591                 if (conf->mddev->queue)
3592                         blk_add_trace_msg(conf->mddev->queue,
3593                                           "raid5 rmw %llu %d",
3594                                           (unsigned long long)sh->sector, rmw);
3595                 for (i = disks; i--; ) {
3596                         struct r5dev *dev = &sh->dev[i];
3597                         if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3598                             !test_bit(R5_LOCKED, &dev->flags) &&
3599                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3600                             test_bit(R5_Wantcompute, &dev->flags)) &&
3601                             test_bit(R5_Insync, &dev->flags)) {
3602                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
3603                                              &sh->state)) {
3604                                         pr_debug("Read_old block %d for r-m-w\n",
3605                                                  i);
3606                                         set_bit(R5_LOCKED, &dev->flags);
3607                                         set_bit(R5_Wantread, &dev->flags);
3608                                         s->locked++;
3609                                 } else {
3610                                         set_bit(STRIPE_DELAYED, &sh->state);
3611                                         set_bit(STRIPE_HANDLE, &sh->state);
3612                                 }
3613                         }
3614                 }
3615         }
3616         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3617                 /* want reconstruct write, but need to get some data */
3618                 int qread =0;
3619                 rcw = 0;
3620                 for (i = disks; i--; ) {
3621                         struct r5dev *dev = &sh->dev[i];
3622                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3623                             i != sh->pd_idx && i != sh->qd_idx &&
3624                             !test_bit(R5_LOCKED, &dev->flags) &&
3625                             !(test_bit(R5_UPTODATE, &dev->flags) ||
3626                               test_bit(R5_Wantcompute, &dev->flags))) {
3627                                 rcw++;
3628                                 if (test_bit(R5_Insync, &dev->flags) &&
3629                                     test_bit(STRIPE_PREREAD_ACTIVE,
3630                                              &sh->state)) {
3631                                         pr_debug("Read_old block "
3632                                                 "%d for Reconstruct\n", i);
3633                                         set_bit(R5_LOCKED, &dev->flags);
3634                                         set_bit(R5_Wantread, &dev->flags);
3635                                         s->locked++;
3636                                         qread++;
3637                                 } else {
3638                                         set_bit(STRIPE_DELAYED, &sh->state);
3639                                         set_bit(STRIPE_HANDLE, &sh->state);
3640                                 }
3641                         }
3642                 }
3643                 if (rcw && conf->mddev->queue)
3644                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3645                                           (unsigned long long)sh->sector,
3646                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3647         }
3648
3649         if (rcw > disks && rmw > disks &&
3650             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3651                 set_bit(STRIPE_DELAYED, &sh->state);
3652
3653         /* now if nothing is locked, and if we have enough data,
3654          * we can start a write request
3655          */
3656         /* since handle_stripe can be called at any time we need to handle the
3657          * case where a compute block operation has been submitted and then a
3658          * subsequent call wants to start a write request.  raid_run_ops only
3659          * handles the case where compute block and reconstruct are requested
3660          * simultaneously.  If this is not the case then new writes need to be
3661          * held off until the compute completes.
3662          */
3663         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3664             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3665             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3666                 schedule_reconstruction(sh, s, rcw == 0, 0);
3667 }
3668
3669 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3670                                 struct stripe_head_state *s, int disks)
3671 {
3672         struct r5dev *dev = NULL;
3673
3674         BUG_ON(sh->batch_head);
3675         set_bit(STRIPE_HANDLE, &sh->state);
3676
3677         switch (sh->check_state) {
3678         case check_state_idle:
3679                 /* start a new check operation if there are no failures */
3680                 if (s->failed == 0) {
3681                         BUG_ON(s->uptodate != disks);
3682                         sh->check_state = check_state_run;
3683                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3684                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3685                         s->uptodate--;
3686                         break;
3687                 }
3688                 dev = &sh->dev[s->failed_num[0]];
3689                 /* fall through */
3690         case check_state_compute_result:
3691                 sh->check_state = check_state_idle;
3692                 if (!dev)
3693                         dev = &sh->dev[sh->pd_idx];
3694
3695                 /* check that a write has not made the stripe insync */
3696                 if (test_bit(STRIPE_INSYNC, &sh->state))
3697                         break;
3698
3699                 /* either failed parity check, or recovery is happening */
3700                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3701                 BUG_ON(s->uptodate != disks);
3702
3703                 set_bit(R5_LOCKED, &dev->flags);
3704                 s->locked++;
3705                 set_bit(R5_Wantwrite, &dev->flags);
3706
3707                 clear_bit(STRIPE_DEGRADED, &sh->state);
3708                 set_bit(STRIPE_INSYNC, &sh->state);
3709                 break;
3710         case check_state_run:
3711                 break; /* we will be called again upon completion */
3712         case check_state_check_result:
3713                 sh->check_state = check_state_idle;
3714
3715                 /* if a failure occurred during the check operation, leave
3716                  * STRIPE_INSYNC not set and let the stripe be handled again
3717                  */
3718                 if (s->failed)
3719                         break;
3720
3721                 /* handle a successful check operation, if parity is correct
3722                  * we are done.  Otherwise update the mismatch count and repair
3723                  * parity if !MD_RECOVERY_CHECK
3724                  */
3725                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3726                         /* parity is correct (on disc,
3727                          * not in buffer any more)
3728                          */
3729                         set_bit(STRIPE_INSYNC, &sh->state);
3730                 else {
3731                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3732                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3733                                 /* don't try to repair!! */
3734                                 set_bit(STRIPE_INSYNC, &sh->state);
3735                         else {
3736                                 sh->check_state = check_state_compute_run;
3737                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3738                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3739                                 set_bit(R5_Wantcompute,
3740                                         &sh->dev[sh->pd_idx].flags);
3741                                 sh->ops.target = sh->pd_idx;
3742                                 sh->ops.target2 = -1;
3743                                 s->uptodate++;
3744                         }
3745                 }
3746                 break;
3747         case check_state_compute_run:
3748                 break;
3749         default:
3750                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3751                        __func__, sh->check_state,
3752                        (unsigned long long) sh->sector);
3753                 BUG();
3754         }
3755 }
3756
3757 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3758                                   struct stripe_head_state *s,
3759                                   int disks)
3760 {
3761         int pd_idx = sh->pd_idx;
3762         int qd_idx = sh->qd_idx;
3763         struct r5dev *dev;
3764
3765         BUG_ON(sh->batch_head);
3766         set_bit(STRIPE_HANDLE, &sh->state);
3767
3768         BUG_ON(s->failed > 2);
3769
3770         /* Want to check and possibly repair P and Q.
3771          * However there could be one 'failed' device, in which
3772          * case we can only check one of them, possibly using the
3773          * other to generate missing data
3774          */
3775
3776         switch (sh->check_state) {
3777         case check_state_idle:
3778                 /* start a new check operation if there are < 2 failures */
3779                 if (s->failed == s->q_failed) {
3780                         /* The only possible failed device holds Q, so it
3781                          * makes sense to check P (If anything else were failed,
3782                          * we would have used P to recreate it).
3783                          */
3784                         sh->check_state = check_state_run;
3785                 }
3786                 if (!s->q_failed && s->failed < 2) {
3787                         /* Q is not failed, and we didn't use it to generate
3788                          * anything, so it makes sense to check it
3789                          */
3790                         if (sh->check_state == check_state_run)
3791                                 sh->check_state = check_state_run_pq;
3792                         else
3793                                 sh->check_state = check_state_run_q;
3794                 }
3795
3796                 /* discard potentially stale zero_sum_result */
3797                 sh->ops.zero_sum_result = 0;
3798
3799                 if (sh->check_state == check_state_run) {
3800                         /* async_xor_zero_sum destroys the contents of P */
3801                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3802                         s->uptodate--;
3803                 }
3804                 if (sh->check_state >= check_state_run &&
3805                     sh->check_state <= check_state_run_pq) {
3806                         /* async_syndrome_zero_sum preserves P and Q, so
3807                          * no need to mark them !uptodate here
3808                          */
3809                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3810                         break;
3811                 }
3812
3813                 /* we have 2-disk failure */
3814                 BUG_ON(s->failed != 2);
3815                 /* fall through */
3816         case check_state_compute_result:
3817                 sh->check_state = check_state_idle;
3818
3819                 /* check that a write has not made the stripe insync */
3820                 if (test_bit(STRIPE_INSYNC, &sh->state))
3821                         break;
3822
3823                 /* now write out any block on a failed drive,
3824                  * or P or Q if they were recomputed
3825                  */
3826                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3827                 if (s->failed == 2) {
3828                         dev = &sh->dev[s->failed_num[1]];
3829                         s->locked++;
3830                         set_bit(R5_LOCKED, &dev->flags);
3831                         set_bit(R5_Wantwrite, &dev->flags);
3832                 }
3833                 if (s->failed >= 1) {
3834                         dev = &sh->dev[s->failed_num[0]];
3835                         s->locked++;
3836                         set_bit(R5_LOCKED, &dev->flags);
3837                         set_bit(R5_Wantwrite, &dev->flags);
3838                 }
3839                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3840                         dev = &sh->dev[pd_idx];
3841                         s->locked++;
3842                         set_bit(R5_LOCKED, &dev->flags);
3843                         set_bit(R5_Wantwrite, &dev->flags);
3844                 }
3845                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3846                         dev = &sh->dev[qd_idx];
3847                         s->locked++;
3848                         set_bit(R5_LOCKED, &dev->flags);
3849                         set_bit(R5_Wantwrite, &dev->flags);
3850                 }
3851                 clear_bit(STRIPE_DEGRADED, &sh->state);
3852
3853                 set_bit(STRIPE_INSYNC, &sh->state);
3854                 break;
3855         case check_state_run:
3856         case check_state_run_q:
3857         case check_state_run_pq:
3858                 break; /* we will be called again upon completion */
3859         case check_state_check_result:
3860                 sh->check_state = check_state_idle;
3861
3862                 /* handle a successful check operation, if parity is correct
3863                  * we are done.  Otherwise update the mismatch count and repair
3864                  * parity if !MD_RECOVERY_CHECK
3865                  */
3866                 if (sh->ops.zero_sum_result == 0) {
3867                         /* both parities are correct */
3868                         if (!s->failed)
3869                                 set_bit(STRIPE_INSYNC, &sh->state);
3870                         else {
3871                                 /* in contrast to the raid5 case we can validate
3872                                  * parity, but still have a failure to write
3873                                  * back
3874                                  */
3875                                 sh->check_state = check_state_compute_result;
3876                                 /* Returning at this point means that we may go
3877                                  * off and bring p and/or q uptodate again so
3878                                  * we make sure to check zero_sum_result again
3879                                  * to verify if p or q need writeback
3880                                  */
3881                         }
3882                 } else {
3883                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3884                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3885                                 /* don't try to repair!! */
3886                                 set_bit(STRIPE_INSYNC, &sh->state);
3887                         else {
3888                                 int *target = &sh->ops.target;
3889
3890                                 sh->ops.target = -1;
3891                                 sh->ops.target2 = -1;
3892                                 sh->check_state = check_state_compute_run;
3893                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3894                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3895                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3896                                         set_bit(R5_Wantcompute,
3897                                                 &sh->dev[pd_idx].flags);
3898                                         *target = pd_idx;
3899                                         target = &sh->ops.target2;
3900                                         s->uptodate++;
3901                                 }
3902                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3903                                         set_bit(R5_Wantcompute,
3904                                                 &sh->dev[qd_idx].flags);
3905                                         *target = qd_idx;
3906                                         s->uptodate++;
3907                                 }
3908                         }
3909                 }
3910                 break;
3911         case check_state_compute_run:
3912                 break;
3913         default:
3914                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3915                        __func__, sh->check_state,
3916                        (unsigned long long) sh->sector);
3917                 BUG();
3918         }
3919 }
3920
3921 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3922 {
3923         int i;
3924
3925         /* We have read all the blocks in this stripe and now we need to
3926          * copy some of them into a target stripe for expand.
3927          */
3928         struct dma_async_tx_descriptor *tx = NULL;
3929         BUG_ON(sh->batch_head);
3930         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3931         for (i = 0; i < sh->disks; i++)
3932                 if (i != sh->pd_idx && i != sh->qd_idx) {
3933                         int dd_idx, j;
3934                         struct stripe_head *sh2;
3935                         struct async_submit_ctl submit;
3936
3937                         sector_t bn = compute_blocknr(sh, i, 1);
3938                         sector_t s = raid5_compute_sector(conf, bn, 0,
3939                                                           &dd_idx, NULL);
3940                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
3941                         if (sh2 == NULL)
3942                                 /* so far only the early blocks of this stripe
3943                                  * have been requested.  When later blocks
3944                                  * get requested, we will try again
3945                                  */
3946                                 continue;
3947                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3948                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3949                                 /* must have already done this block */
3950                                 release_stripe(sh2);
3951                                 continue;
3952                         }
3953
3954                         /* place all the copies on one channel */
3955                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3956                         tx = async_memcpy(sh2->dev[dd_idx].page,
3957                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
3958                                           &submit);
3959
3960                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3961                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3962                         for (j = 0; j < conf->raid_disks; j++)
3963                                 if (j != sh2->pd_idx &&
3964                                     j != sh2->qd_idx &&
3965                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
3966                                         break;
3967                         if (j == conf->raid_disks) {
3968                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3969                                 set_bit(STRIPE_HANDLE, &sh2->state);
3970                         }
3971                         release_stripe(sh2);
3972
3973                 }
3974         /* done submitting copies, wait for them to complete */
3975         async_tx_quiesce(&tx);
3976 }
3977
3978 /*
3979  * handle_stripe - do things to a stripe.
3980  *
3981  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3982  * state of various bits to see what needs to be done.
3983  * Possible results:
3984  *    return some read requests which now have data
3985  *    return some write requests which are safely on storage
3986  *    schedule a read on some buffers
3987  *    schedule a write of some buffers
3988  *    return confirmation of parity correctness
3989  *
3990  */
3991
3992 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3993 {
3994         struct r5conf *conf = sh->raid_conf;
3995         int disks = sh->disks;
3996         struct r5dev *dev;
3997         int i;
3998         int do_recovery = 0;
3999
4000         memset(s, 0, sizeof(*s));
4001
4002         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4003         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4004         s->failed_num[0] = -1;
4005         s->failed_num[1] = -1;
4006
4007         /* Now to look around and see what can be done */
4008         rcu_read_lock();
4009         for (i=disks; i--; ) {
4010                 struct md_rdev *rdev;
4011                 sector_t first_bad;
4012                 int bad_sectors;
4013                 int is_bad = 0;
4014
4015                 dev = &sh->dev[i];
4016
4017                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4018                          i, dev->flags,
4019                          dev->toread, dev->towrite, dev->written);
4020                 /* maybe we can reply to a read
4021                  *
4022                  * new wantfill requests are only permitted while
4023                  * ops_complete_biofill is guaranteed to be inactive
4024                  */
4025                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4026                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4027                         set_bit(R5_Wantfill, &dev->flags);
4028
4029                 /* now count some things */
4030                 if (test_bit(R5_LOCKED, &dev->flags))
4031                         s->locked++;
4032                 if (test_bit(R5_UPTODATE, &dev->flags))
4033                         s->uptodate++;
4034                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4035                         s->compute++;
4036                         BUG_ON(s->compute > 2);
4037                 }
4038
4039                 if (test_bit(R5_Wantfill, &dev->flags))
4040                         s->to_fill++;
4041                 else if (dev->toread)
4042                         s->to_read++;
4043                 if (dev->towrite) {
4044                         s->to_write++;
4045                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4046                                 s->non_overwrite++;
4047                 }
4048                 if (dev->written)
4049                         s->written++;
4050                 /* Prefer to use the replacement for reads, but only
4051                  * if it is recovered enough and has no bad blocks.
4052                  */
4053                 rdev = rcu_dereference(conf->disks[i].replacement);
4054                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4055                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4056                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4057                                  &first_bad, &bad_sectors))
4058                         set_bit(R5_ReadRepl, &dev->flags);
4059                 else {
4060                         if (rdev)
4061                                 set_bit(R5_NeedReplace, &dev->flags);
4062                         rdev = rcu_dereference(conf->disks[i].rdev);
4063                         clear_bit(R5_ReadRepl, &dev->flags);
4064                 }
4065                 if (rdev && test_bit(Faulty, &rdev->flags))
4066                         rdev = NULL;
4067                 if (rdev) {
4068                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4069                                              &first_bad, &bad_sectors);
4070                         if (s->blocked_rdev == NULL
4071                             && (test_bit(Blocked, &rdev->flags)
4072                                 || is_bad < 0)) {
4073                                 if (is_bad < 0)
4074                                         set_bit(BlockedBadBlocks,
4075                                                 &rdev->flags);
4076                                 s->blocked_rdev = rdev;
4077                                 atomic_inc(&rdev->nr_pending);
4078                         }
4079                 }
4080                 clear_bit(R5_Insync, &dev->flags);
4081                 if (!rdev)
4082                         /* Not in-sync */;
4083                 else if (is_bad) {
4084                         /* also not in-sync */
4085                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4086                             test_bit(R5_UPTODATE, &dev->flags)) {
4087                                 /* treat as in-sync, but with a read error
4088                                  * which we can now try to correct
4089                                  */
4090                                 set_bit(R5_Insync, &dev->flags);
4091                                 set_bit(R5_ReadError, &dev->flags);
4092                         }
4093                 } else if (test_bit(In_sync, &rdev->flags))
4094                         set_bit(R5_Insync, &dev->flags);
4095                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4096                         /* in sync if before recovery_offset */
4097                         set_bit(R5_Insync, &dev->flags);
4098                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4099                          test_bit(R5_Expanded, &dev->flags))
4100                         /* If we've reshaped into here, we assume it is Insync.
4101                          * We will shortly update recovery_offset to make
4102                          * it official.
4103                          */
4104                         set_bit(R5_Insync, &dev->flags);
4105
4106                 if (test_bit(R5_WriteError, &dev->flags)) {
4107                         /* This flag does not apply to '.replacement'
4108                          * only to .rdev, so make sure to check that*/
4109                         struct md_rdev *rdev2 = rcu_dereference(
4110                                 conf->disks[i].rdev);
4111                         if (rdev2 == rdev)
4112                                 clear_bit(R5_Insync, &dev->flags);
4113                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4114                                 s->handle_bad_blocks = 1;
4115                                 atomic_inc(&rdev2->nr_pending);
4116                         } else
4117                                 clear_bit(R5_WriteError, &dev->flags);
4118                 }
4119                 if (test_bit(R5_MadeGood, &dev->flags)) {
4120                         /* This flag does not apply to '.replacement'
4121                          * only to .rdev, so make sure to check that*/
4122                         struct md_rdev *rdev2 = rcu_dereference(
4123                                 conf->disks[i].rdev);
4124                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4125                                 s->handle_bad_blocks = 1;
4126                                 atomic_inc(&rdev2->nr_pending);
4127                         } else
4128                                 clear_bit(R5_MadeGood, &dev->flags);
4129                 }
4130                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4131                         struct md_rdev *rdev2 = rcu_dereference(
4132                                 conf->disks[i].replacement);
4133                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4134                                 s->handle_bad_blocks = 1;
4135                                 atomic_inc(&rdev2->nr_pending);
4136                         } else
4137                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4138                 }
4139                 if (!test_bit(R5_Insync, &dev->flags)) {
4140                         /* The ReadError flag will just be confusing now */
4141                         clear_bit(R5_ReadError, &dev->flags);
4142                         clear_bit(R5_ReWrite, &dev->flags);
4143                 }
4144                 if (test_bit(R5_ReadError, &dev->flags))
4145                         clear_bit(R5_Insync, &dev->flags);
4146                 if (!test_bit(R5_Insync, &dev->flags)) {
4147                         if (s->failed < 2)
4148                                 s->failed_num[s->failed] = i;
4149                         s->failed++;
4150                         if (rdev && !test_bit(Faulty, &rdev->flags))
4151                                 do_recovery = 1;
4152                 }
4153         }
4154         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4155                 /* If there is a failed device being replaced,
4156                  *     we must be recovering.
4157                  * else if we are after recovery_cp, we must be syncing
4158                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4159                  * else we can only be replacing
4160                  * sync and recovery both need to read all devices, and so
4161                  * use the same flag.
4162                  */
4163                 if (do_recovery ||
4164                     sh->sector >= conf->mddev->recovery_cp ||
4165                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4166                         s->syncing = 1;
4167                 else
4168                         s->replacing = 1;
4169         }
4170         rcu_read_unlock();
4171 }
4172
4173 static int clear_batch_ready(struct stripe_head *sh)
4174 {
4175         /* Return '1' if this is a member of batch, or
4176          * '0' if it is a lone stripe or a head which can now be
4177          * handled.
4178          */
4179         struct stripe_head *tmp;
4180         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4181                 return (sh->batch_head && sh->batch_head != sh);
4182         spin_lock(&sh->stripe_lock);
4183         if (!sh->batch_head) {
4184                 spin_unlock(&sh->stripe_lock);
4185                 return 0;
4186         }
4187
4188         /*
4189          * this stripe could be added to a batch list before we check
4190          * BATCH_READY, skips it
4191          */
4192         if (sh->batch_head != sh) {
4193                 spin_unlock(&sh->stripe_lock);
4194                 return 1;
4195         }
4196         spin_lock(&sh->batch_lock);
4197         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4198                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4199         spin_unlock(&sh->batch_lock);
4200         spin_unlock(&sh->stripe_lock);
4201
4202         /*
4203          * BATCH_READY is cleared, no new stripes can be added.
4204          * batch_list can be accessed without lock
4205          */
4206         return 0;
4207 }
4208
4209 static void break_stripe_batch_list(struct stripe_head *head_sh,
4210                                     unsigned long handle_flags)
4211 {
4212         struct stripe_head *sh, *next;
4213         int i;
4214         int do_wakeup = 0;
4215
4216         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4217
4218                 list_del_init(&sh->batch_list);
4219
4220                 WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4221                                           (1 << STRIPE_SYNCING) |
4222                                           (1 << STRIPE_REPLACED) |
4223                                           (1 << STRIPE_PREREAD_ACTIVE) |
4224                                           (1 << STRIPE_DELAYED) |
4225                                           (1 << STRIPE_BIT_DELAY) |
4226                                           (1 << STRIPE_FULL_WRITE) |
4227                                           (1 << STRIPE_BIOFILL_RUN) |
4228                                           (1 << STRIPE_COMPUTE_RUN)  |
4229                                           (1 << STRIPE_OPS_REQ_PENDING) |
4230                                           (1 << STRIPE_DISCARD) |
4231                                           (1 << STRIPE_BATCH_READY) |
4232                                           (1 << STRIPE_BATCH_ERR) |
4233                                           (1 << STRIPE_BITMAP_PENDING)));
4234                 WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4235                                               (1 << STRIPE_REPLACED)));
4236
4237                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4238                                             (1 << STRIPE_DEGRADED)),
4239                               head_sh->state & (1 << STRIPE_INSYNC));
4240
4241                 sh->check_state = head_sh->check_state;
4242                 sh->reconstruct_state = head_sh->reconstruct_state;
4243                 for (i = 0; i < sh->disks; i++) {
4244                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4245                                 do_wakeup = 1;
4246                         sh->dev[i].flags = head_sh->dev[i].flags &
4247                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4248                 }
4249                 spin_lock_irq(&sh->stripe_lock);
4250                 sh->batch_head = NULL;
4251                 spin_unlock_irq(&sh->stripe_lock);
4252                 if (handle_flags == 0 ||
4253                     sh->state & handle_flags)
4254                         set_bit(STRIPE_HANDLE, &sh->state);
4255                 release_stripe(sh);
4256         }
4257         spin_lock_irq(&head_sh->stripe_lock);
4258         head_sh->batch_head = NULL;
4259         spin_unlock_irq(&head_sh->stripe_lock);
4260         for (i = 0; i < head_sh->disks; i++)
4261                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4262                         do_wakeup = 1;
4263         if (head_sh->state & handle_flags)
4264                 set_bit(STRIPE_HANDLE, &head_sh->state);
4265
4266         if (do_wakeup)
4267                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4268 }
4269
4270 static void handle_stripe(struct stripe_head *sh)
4271 {
4272         struct stripe_head_state s;
4273         struct r5conf *conf = sh->raid_conf;
4274         int i;
4275         int prexor;
4276         int disks = sh->disks;
4277         struct r5dev *pdev, *qdev;
4278
4279         clear_bit(STRIPE_HANDLE, &sh->state);
4280         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4281                 /* already being handled, ensure it gets handled
4282                  * again when current action finishes */
4283                 set_bit(STRIPE_HANDLE, &sh->state);
4284                 return;
4285         }
4286
4287         if (clear_batch_ready(sh) ) {
4288                 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4289                 return;
4290         }
4291
4292         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4293                 break_stripe_batch_list(sh, 0);
4294
4295         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4296                 spin_lock(&sh->stripe_lock);
4297                 /* Cannot process 'sync' concurrently with 'discard' */
4298                 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4299                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4300                         set_bit(STRIPE_SYNCING, &sh->state);
4301                         clear_bit(STRIPE_INSYNC, &sh->state);
4302                         clear_bit(STRIPE_REPLACED, &sh->state);
4303                 }
4304                 spin_unlock(&sh->stripe_lock);
4305         }
4306         clear_bit(STRIPE_DELAYED, &sh->state);
4307
4308         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4309                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4310                (unsigned long long)sh->sector, sh->state,
4311                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4312                sh->check_state, sh->reconstruct_state);
4313
4314         analyse_stripe(sh, &s);
4315
4316         if (s.handle_bad_blocks) {
4317                 set_bit(STRIPE_HANDLE, &sh->state);
4318                 goto finish;
4319         }
4320
4321         if (unlikely(s.blocked_rdev)) {
4322                 if (s.syncing || s.expanding || s.expanded ||
4323                     s.replacing || s.to_write || s.written) {
4324                         set_bit(STRIPE_HANDLE, &sh->state);
4325                         goto finish;
4326                 }
4327                 /* There is nothing for the blocked_rdev to block */
4328                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4329                 s.blocked_rdev = NULL;
4330         }
4331
4332         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4333                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4334                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4335         }
4336
4337         pr_debug("locked=%d uptodate=%d to_read=%d"
4338                " to_write=%d failed=%d failed_num=%d,%d\n",
4339                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4340                s.failed_num[0], s.failed_num[1]);
4341         /* check if the array has lost more than max_degraded devices and,
4342          * if so, some requests might need to be failed.
4343          */
4344         if (s.failed > conf->max_degraded) {
4345                 sh->check_state = 0;
4346                 sh->reconstruct_state = 0;
4347                 break_stripe_batch_list(sh, 0);
4348                 if (s.to_read+s.to_write+s.written)
4349                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4350                 if (s.syncing + s.replacing)
4351                         handle_failed_sync(conf, sh, &s);
4352         }
4353
4354         /* Now we check to see if any write operations have recently
4355          * completed
4356          */
4357         prexor = 0;
4358         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4359                 prexor = 1;
4360         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4361             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4362                 sh->reconstruct_state = reconstruct_state_idle;
4363
4364                 /* All the 'written' buffers and the parity block are ready to
4365                  * be written back to disk
4366                  */
4367                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4368                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4369                 BUG_ON(sh->qd_idx >= 0 &&
4370                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4371                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4372                 for (i = disks; i--; ) {
4373                         struct r5dev *dev = &sh->dev[i];
4374                         if (test_bit(R5_LOCKED, &dev->flags) &&
4375                                 (i == sh->pd_idx || i == sh->qd_idx ||
4376                                  dev->written)) {
4377                                 pr_debug("Writing block %d\n", i);
4378                                 set_bit(R5_Wantwrite, &dev->flags);
4379                                 if (prexor)
4380                                         continue;
4381                                 if (s.failed > 1)
4382                                         continue;
4383                                 if (!test_bit(R5_Insync, &dev->flags) ||
4384                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
4385                                      s.failed == 0))
4386                                         set_bit(STRIPE_INSYNC, &sh->state);
4387                         }
4388                 }
4389                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4390                         s.dec_preread_active = 1;
4391         }
4392
4393         /*
4394          * might be able to return some write requests if the parity blocks
4395          * are safe, or on a failed drive
4396          */
4397         pdev = &sh->dev[sh->pd_idx];
4398         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4399                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4400         qdev = &sh->dev[sh->qd_idx];
4401         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4402                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4403                 || conf->level < 6;
4404
4405         if (s.written &&
4406             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4407                              && !test_bit(R5_LOCKED, &pdev->flags)
4408                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
4409                                  test_bit(R5_Discard, &pdev->flags))))) &&
4410             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4411                              && !test_bit(R5_LOCKED, &qdev->flags)
4412                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
4413                                  test_bit(R5_Discard, &qdev->flags))))))
4414                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4415
4416         /* Now we might consider reading some blocks, either to check/generate
4417          * parity, or to satisfy requests
4418          * or to load a block that is being partially written.
4419          */
4420         if (s.to_read || s.non_overwrite
4421             || (conf->level == 6 && s.to_write && s.failed)
4422             || (s.syncing && (s.uptodate + s.compute < disks))
4423             || s.replacing
4424             || s.expanding)
4425                 handle_stripe_fill(sh, &s, disks);
4426
4427         /* Now to consider new write requests and what else, if anything
4428          * should be read.  We do not handle new writes when:
4429          * 1/ A 'write' operation (copy+xor) is already in flight.
4430          * 2/ A 'check' operation is in flight, as it may clobber the parity
4431          *    block.
4432          */
4433         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4434                 handle_stripe_dirtying(conf, sh, &s, disks);
4435
4436         /* maybe we need to check and possibly fix the parity for this stripe
4437          * Any reads will already have been scheduled, so we just see if enough
4438          * data is available.  The parity check is held off while parity
4439          * dependent operations are in flight.
4440          */
4441         if (sh->check_state ||
4442             (s.syncing && s.locked == 0 &&
4443              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4444              !test_bit(STRIPE_INSYNC, &sh->state))) {
4445                 if (conf->level == 6)
4446                         handle_parity_checks6(conf, sh, &s, disks);
4447                 else
4448                         handle_parity_checks5(conf, sh, &s, disks);
4449         }
4450
4451         if ((s.replacing || s.syncing) && s.locked == 0
4452             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4453             && !test_bit(STRIPE_REPLACED, &sh->state)) {
4454                 /* Write out to replacement devices where possible */
4455                 for (i = 0; i < conf->raid_disks; i++)
4456                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4457                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4458                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
4459                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
4460                                 s.locked++;
4461                         }
4462                 if (s.replacing)
4463                         set_bit(STRIPE_INSYNC, &sh->state);
4464                 set_bit(STRIPE_REPLACED, &sh->state);
4465         }
4466         if ((s.syncing || s.replacing) && s.locked == 0 &&
4467             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4468             test_bit(STRIPE_INSYNC, &sh->state)) {
4469                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4470                 clear_bit(STRIPE_SYNCING, &sh->state);
4471                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4472                         wake_up(&conf->wait_for_overlap);
4473         }
4474
4475         /* If the failed drives are just a ReadError, then we might need
4476          * to progress the repair/check process
4477          */
4478         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4479                 for (i = 0; i < s.failed; i++) {
4480                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
4481                         if (test_bit(R5_ReadError, &dev->flags)
4482                             && !test_bit(R5_LOCKED, &dev->flags)
4483                             && test_bit(R5_UPTODATE, &dev->flags)
4484                                 ) {
4485                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
4486                                         set_bit(R5_Wantwrite, &dev->flags);
4487                                         set_bit(R5_ReWrite, &dev->flags);
4488                                         set_bit(R5_LOCKED, &dev->flags);
4489                                         s.locked++;
4490                                 } else {
4491                                         /* let's read it back */
4492                                         set_bit(R5_Wantread, &dev->flags);
4493                                         set_bit(R5_LOCKED, &dev->flags);
4494                                         s.locked++;
4495                                 }
4496                         }
4497                 }
4498
4499         /* Finish reconstruct operations initiated by the expansion process */
4500         if (sh->reconstruct_state == reconstruct_state_result) {
4501                 struct stripe_head *sh_src
4502                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
4503                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4504                         /* sh cannot be written until sh_src has been read.
4505                          * so arrange for sh to be delayed a little
4506                          */
4507                         set_bit(STRIPE_DELAYED, &sh->state);
4508                         set_bit(STRIPE_HANDLE, &sh->state);
4509                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4510                                               &sh_src->state))
4511                                 atomic_inc(&conf->preread_active_stripes);
4512                         release_stripe(sh_src);
4513                         goto finish;
4514                 }
4515                 if (sh_src)
4516                         release_stripe(sh_src);
4517
4518                 sh->reconstruct_state = reconstruct_state_idle;
4519                 clear_bit(STRIPE_EXPANDING, &sh->state);
4520                 for (i = conf->raid_disks; i--; ) {
4521                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
4522                         set_bit(R5_LOCKED, &sh->dev[i].flags);
4523                         s.locked++;
4524                 }
4525         }
4526
4527         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4528             !sh->reconstruct_state) {
4529                 /* Need to write out all blocks after computing parity */
4530                 sh->disks = conf->raid_disks;
4531                 stripe_set_idx(sh->sector, conf, 0, sh);
4532                 schedule_reconstruction(sh, &s, 1, 1);
4533         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4534                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4535                 atomic_dec(&conf->reshape_stripes);
4536                 wake_up(&conf->wait_for_overlap);
4537                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4538         }
4539
4540         if (s.expanding && s.locked == 0 &&
4541             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4542                 handle_stripe_expansion(conf, sh);
4543
4544 finish:
4545         /* wait for this device to become unblocked */
4546         if (unlikely(s.blocked_rdev)) {
4547                 if (conf->mddev->external)
4548                         md_wait_for_blocked_rdev(s.blocked_rdev,
4549                                                  conf->mddev);
4550                 else
4551                         /* Internal metadata will immediately
4552                          * be written by raid5d, so we don't
4553                          * need to wait here.
4554                          */
4555                         rdev_dec_pending(s.blocked_rdev,
4556                                          conf->mddev);
4557         }
4558
4559         if (s.handle_bad_blocks)
4560                 for (i = disks; i--; ) {
4561                         struct md_rdev *rdev;
4562                         struct r5dev *dev = &sh->dev[i];
4563                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4564                                 /* We own a safe reference to the rdev */
4565                                 rdev = conf->disks[i].rdev;
4566                                 if (!rdev_set_badblocks(rdev, sh->sector,
4567                                                         STRIPE_SECTORS, 0))
4568                                         md_error(conf->mddev, rdev);
4569                                 rdev_dec_pending(rdev, conf->mddev);
4570                         }
4571                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4572                                 rdev = conf->disks[i].rdev;
4573                                 rdev_clear_badblocks(rdev, sh->sector,
4574                                                      STRIPE_SECTORS, 0);
4575                                 rdev_dec_pending(rdev, conf->mddev);
4576                         }
4577                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4578                                 rdev = conf->disks[i].replacement;
4579                                 if (!rdev)
4580                                         /* rdev have been moved down */
4581                                         rdev = conf->disks[i].rdev;
4582                                 rdev_clear_badblocks(rdev, sh->sector,
4583                                                      STRIPE_SECTORS, 0);
4584                                 rdev_dec_pending(rdev, conf->mddev);
4585                         }
4586                 }
4587
4588         if (s.ops_request)
4589                 raid_run_ops(sh, s.ops_request);
4590
4591         ops_run_io(sh, &s);
4592
4593         if (s.dec_preread_active) {
4594                 /* We delay this until after ops_run_io so that if make_request
4595                  * is waiting on a flush, it won't continue until the writes
4596                  * have actually been submitted.
4597                  */
4598                 atomic_dec(&conf->preread_active_stripes);
4599                 if (atomic_read(&conf->preread_active_stripes) <
4600                     IO_THRESHOLD)
4601                         md_wakeup_thread(conf->mddev->thread);
4602         }
4603
4604         return_io(s.return_bi);
4605
4606         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4607 }
4608
4609 static void raid5_activate_delayed(struct r5conf *conf)
4610 {
4611         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4612                 while (!list_empty(&conf->delayed_list)) {
4613                         struct list_head *l = conf->delayed_list.next;
4614                         struct stripe_head *sh;
4615                         sh = list_entry(l, struct stripe_head, lru);
4616                         list_del_init(l);
4617                         clear_bit(STRIPE_DELAYED, &sh->state);
4618                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4619                                 atomic_inc(&conf->preread_active_stripes);
4620                         list_add_tail(&sh->lru, &conf->hold_list);
4621                         raid5_wakeup_stripe_thread(sh);
4622                 }
4623         }
4624 }
4625
4626 static void activate_bit_delay(struct r5conf *conf,
4627         struct list_head *temp_inactive_list)
4628 {
4629         /* device_lock is held */
4630         struct list_head head;
4631         list_add(&head, &conf->bitmap_list);
4632         list_del_init(&conf->bitmap_list);
4633         while (!list_empty(&head)) {
4634                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4635                 int hash;
4636                 list_del_init(&sh->lru);
4637                 atomic_inc(&sh->count);
4638                 hash = sh->hash_lock_index;
4639                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
4640         }
4641 }
4642
4643 static int raid5_congested(struct mddev *mddev, int bits)
4644 {
4645         struct r5conf *conf = mddev->private;
4646
4647         /* No difference between reads and writes.  Just check
4648          * how busy the stripe_cache is
4649          */
4650
4651         if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4652                 return 1;
4653         if (conf->quiesce)
4654                 return 1;
4655         if (atomic_read(&conf->empty_inactive_list_nr))
4656                 return 1;
4657
4658         return 0;
4659 }
4660
4661 /* We want read requests to align with chunks where possible,
4662  * but write requests don't need to.
4663  */
4664 static int raid5_mergeable_bvec(struct mddev *mddev,
4665                                 struct bvec_merge_data *bvm,
4666                                 struct bio_vec *biovec)
4667 {
4668         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4669         int max;
4670         unsigned int chunk_sectors = mddev->chunk_sectors;
4671         unsigned int bio_sectors = bvm->bi_size >> 9;
4672
4673         /*
4674          * always allow writes to be mergeable, read as well if array
4675          * is degraded as we'll go through stripe cache anyway.
4676          */
4677         if ((bvm->bi_rw & 1) == WRITE || mddev->degraded)
4678                 return biovec->bv_len;
4679
4680         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4681                 chunk_sectors = mddev->new_chunk_sectors;
4682         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4683         if (max < 0) max = 0;
4684         if (max <= biovec->bv_len && bio_sectors == 0)
4685                 return biovec->bv_len;
4686         else
4687                 return max;
4688 }
4689
4690 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4691 {
4692         sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4693         unsigned int chunk_sectors = mddev->chunk_sectors;
4694         unsigned int bio_sectors = bio_sectors(bio);
4695
4696         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4697                 chunk_sectors = mddev->new_chunk_sectors;
4698         return  chunk_sectors >=
4699                 ((sector & (chunk_sectors - 1)) + bio_sectors);
4700 }
4701
4702 /*
4703  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
4704  *  later sampled by raid5d.
4705  */
4706 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4707 {
4708         unsigned long flags;
4709
4710         spin_lock_irqsave(&conf->device_lock, flags);
4711
4712         bi->bi_next = conf->retry_read_aligned_list;
4713         conf->retry_read_aligned_list = bi;
4714
4715         spin_unlock_irqrestore(&conf->device_lock, flags);
4716         md_wakeup_thread(conf->mddev->thread);
4717 }
4718
4719 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4720 {
4721         struct bio *bi;
4722
4723         bi = conf->retry_read_aligned;
4724         if (bi) {
4725                 conf->retry_read_aligned = NULL;
4726                 return bi;
4727         }
4728         bi = conf->retry_read_aligned_list;
4729         if(bi) {
4730                 conf->retry_read_aligned_list = bi->bi_next;
4731                 bi->bi_next = NULL;
4732                 /*
4733                  * this sets the active strip count to 1 and the processed
4734                  * strip count to zero (upper 8 bits)
4735                  */
4736                 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4737         }
4738
4739         return bi;
4740 }
4741
4742 /*
4743  *  The "raid5_align_endio" should check if the read succeeded and if it
4744  *  did, call bio_endio on the original bio (having bio_put the new bio
4745  *  first).
4746  *  If the read failed..
4747  */
4748 static void raid5_align_endio(struct bio *bi, int error)
4749 {
4750         struct bio* raid_bi  = bi->bi_private;
4751         struct mddev *mddev;
4752         struct r5conf *conf;
4753         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4754         struct md_rdev *rdev;
4755
4756         bio_put(bi);
4757
4758         rdev = (void*)raid_bi->bi_next;
4759         raid_bi->bi_next = NULL;
4760         mddev = rdev->mddev;
4761         conf = mddev->private;
4762
4763         rdev_dec_pending(rdev, conf->mddev);
4764
4765         if (!error && uptodate) {
4766                 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4767                                          raid_bi, 0);
4768                 bio_endio(raid_bi, 0);
4769                 if (atomic_dec_and_test(&conf->active_aligned_reads))
4770                         wake_up(&conf->wait_for_stripe);
4771                 return;
4772         }
4773
4774         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4775
4776         add_bio_to_retry(raid_bi, conf);
4777 }
4778
4779 static int bio_fits_rdev(struct bio *bi)
4780 {
4781         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4782
4783         if (bio_sectors(bi) > queue_max_sectors(q))
4784                 return 0;
4785         blk_recount_segments(q, bi);
4786         if (bi->bi_phys_segments > queue_max_segments(q))
4787                 return 0;
4788
4789         if (q->merge_bvec_fn)
4790                 /* it's too hard to apply the merge_bvec_fn at this stage,
4791                  * just just give up
4792                  */
4793                 return 0;
4794
4795         return 1;
4796 }
4797
4798 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4799 {
4800         struct r5conf *conf = mddev->private;
4801         int dd_idx;
4802         struct bio* align_bi;
4803         struct md_rdev *rdev;
4804         sector_t end_sector;
4805
4806         if (!in_chunk_boundary(mddev, raid_bio)) {
4807                 pr_debug("chunk_aligned_read : non aligned\n");
4808                 return 0;
4809         }
4810         /*
4811          * use bio_clone_mddev to make a copy of the bio
4812          */
4813         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4814         if (!align_bi)
4815                 return 0;
4816         /*
4817          *   set bi_end_io to a new function, and set bi_private to the
4818          *     original bio.
4819          */
4820         align_bi->bi_end_io  = raid5_align_endio;
4821         align_bi->bi_private = raid_bio;
4822         /*
4823          *      compute position
4824          */
4825         align_bi->bi_iter.bi_sector =
4826                 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4827                                      0, &dd_idx, NULL);
4828
4829         end_sector = bio_end_sector(align_bi);
4830         rcu_read_lock();
4831         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4832         if (!rdev || test_bit(Faulty, &rdev->flags) ||
4833             rdev->recovery_offset < end_sector) {
4834                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4835                 if (rdev &&
4836                     (test_bit(Faulty, &rdev->flags) ||
4837                     !(test_bit(In_sync, &rdev->flags) ||
4838                       rdev->recovery_offset >= end_sector)))
4839                         rdev = NULL;
4840         }
4841         if (rdev) {
4842                 sector_t first_bad;
4843                 int bad_sectors;
4844
4845                 atomic_inc(&rdev->nr_pending);
4846                 rcu_read_unlock();
4847                 raid_bio->bi_next = (void*)rdev;
4848                 align_bi->bi_bdev =  rdev->bdev;
4849                 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
4850
4851                 if (!bio_fits_rdev(align_bi) ||
4852                     is_badblock(rdev, align_bi->bi_iter.bi_sector,
4853                                 bio_sectors(align_bi),
4854                                 &first_bad, &bad_sectors)) {
4855                         /* too big in some way, or has a known bad block */
4856                         bio_put(align_bi);
4857                         rdev_dec_pending(rdev, mddev);
4858                         return 0;
4859                 }
4860
4861                 /* No reshape active, so we can trust rdev->data_offset */
4862                 align_bi->bi_iter.bi_sector += rdev->data_offset;
4863
4864                 spin_lock_irq(&conf->device_lock);
4865                 wait_event_lock_irq(conf->wait_for_stripe,
4866                                     conf->quiesce == 0,
4867                                     conf->device_lock);
4868                 atomic_inc(&conf->active_aligned_reads);
4869                 spin_unlock_irq(&conf->device_lock);
4870
4871                 if (mddev->gendisk)
4872                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4873                                               align_bi, disk_devt(mddev->gendisk),
4874                                               raid_bio->bi_iter.bi_sector);
4875                 generic_make_request(align_bi);
4876                 return 1;
4877         } else {
4878                 rcu_read_unlock();
4879                 bio_put(align_bi);
4880                 return 0;
4881         }
4882 }
4883
4884 /* __get_priority_stripe - get the next stripe to process
4885  *
4886  * Full stripe writes are allowed to pass preread active stripes up until
4887  * the bypass_threshold is exceeded.  In general the bypass_count
4888  * increments when the handle_list is handled before the hold_list; however, it
4889  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4890  * stripe with in flight i/o.  The bypass_count will be reset when the
4891  * head of the hold_list has changed, i.e. the head was promoted to the
4892  * handle_list.
4893  */
4894 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4895 {
4896         struct stripe_head *sh = NULL, *tmp;
4897         struct list_head *handle_list = NULL;
4898         struct r5worker_group *wg = NULL;
4899
4900         if (conf->worker_cnt_per_group == 0) {
4901                 handle_list = &conf->handle_list;
4902         } else if (group != ANY_GROUP) {
4903                 handle_list = &conf->worker_groups[group].handle_list;
4904                 wg = &conf->worker_groups[group];
4905         } else {
4906                 int i;
4907                 for (i = 0; i < conf->group_cnt; i++) {
4908                         handle_list = &conf->worker_groups[i].handle_list;
4909                         wg = &conf->worker_groups[i];
4910                         if (!list_empty(handle_list))
4911                                 break;
4912                 }
4913         }
4914
4915         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4916                   __func__,
4917                   list_empty(handle_list) ? "empty" : "busy",
4918                   list_empty(&conf->hold_list) ? "empty" : "busy",
4919                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
4920
4921         if (!list_empty(handle_list)) {
4922                 sh = list_entry(handle_list->next, typeof(*sh), lru);
4923
4924                 if (list_empty(&conf->hold_list))
4925                         conf->bypass_count = 0;
4926                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4927                         if (conf->hold_list.next == conf->last_hold)
4928                                 conf->bypass_count++;
4929                         else {
4930                                 conf->last_hold = conf->hold_list.next;
4931                                 conf->bypass_count -= conf->bypass_threshold;
4932                                 if (conf->bypass_count < 0)
4933                                         conf->bypass_count = 0;
4934                         }
4935                 }
4936         } else if (!list_empty(&conf->hold_list) &&
4937                    ((conf->bypass_threshold &&
4938                      conf->bypass_count > conf->bypass_threshold) ||
4939                     atomic_read(&conf->pending_full_writes) == 0)) {
4940
4941                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
4942                         if (conf->worker_cnt_per_group == 0 ||
4943                             group == ANY_GROUP ||
4944                             !cpu_online(tmp->cpu) ||
4945                             cpu_to_group(tmp->cpu) == group) {
4946                                 sh = tmp;
4947                                 break;
4948                         }
4949                 }
4950
4951                 if (sh) {
4952                         conf->bypass_count -= conf->bypass_threshold;
4953                         if (conf->bypass_count < 0)
4954                                 conf->bypass_count = 0;
4955                 }
4956                 wg = NULL;
4957         }
4958
4959         if (!sh)
4960                 return NULL;
4961
4962         if (wg) {
4963                 wg->stripes_cnt--;
4964                 sh->group = NULL;
4965         }
4966         list_del_init(&sh->lru);
4967         BUG_ON(atomic_inc_return(&sh->count) != 1);
4968         return sh;
4969 }
4970
4971 struct raid5_plug_cb {
4972         struct blk_plug_cb      cb;
4973         struct list_head        list;
4974         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4975 };
4976
4977 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4978 {
4979         struct raid5_plug_cb *cb = container_of(
4980                 blk_cb, struct raid5_plug_cb, cb);
4981         struct stripe_head *sh;
4982         struct mddev *mddev = cb->cb.data;
4983         struct r5conf *conf = mddev->private;
4984         int cnt = 0;
4985         int hash;
4986
4987         if (cb->list.next && !list_empty(&cb->list)) {
4988                 spin_lock_irq(&conf->device_lock);
4989                 while (!list_empty(&cb->list)) {
4990                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
4991                         list_del_init(&sh->lru);
4992                         /*
4993                          * avoid race release_stripe_plug() sees
4994                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
4995                          * is still in our list
4996                          */
4997                         smp_mb__before_atomic();
4998                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4999                         /*
5000                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5001                          * case, the count is always > 1 here
5002                          */
5003                         hash = sh->hash_lock_index;
5004                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5005                         cnt++;
5006                 }
5007                 spin_unlock_irq(&conf->device_lock);
5008         }
5009         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5010                                      NR_STRIPE_HASH_LOCKS);
5011         if (mddev->queue)
5012                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5013         kfree(cb);
5014 }
5015
5016 static void release_stripe_plug(struct mddev *mddev,
5017                                 struct stripe_head *sh)
5018 {
5019         struct blk_plug_cb *blk_cb = blk_check_plugged(
5020                 raid5_unplug, mddev,
5021                 sizeof(struct raid5_plug_cb));
5022         struct raid5_plug_cb *cb;
5023
5024         if (!blk_cb) {
5025                 release_stripe(sh);
5026                 return;
5027         }
5028
5029         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5030
5031         if (cb->list.next == NULL) {
5032                 int i;
5033                 INIT_LIST_HEAD(&cb->list);
5034                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5035                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5036         }
5037
5038         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5039                 list_add_tail(&sh->lru, &cb->list);
5040         else
5041                 release_stripe(sh);
5042 }
5043
5044 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5045 {
5046         struct r5conf *conf = mddev->private;
5047         sector_t logical_sector, last_sector;
5048         struct stripe_head *sh;
5049         int remaining;
5050         int stripe_sectors;
5051
5052         if (mddev->reshape_position != MaxSector)
5053                 /* Skip discard while reshape is happening */
5054                 return;
5055
5056         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5057         last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5058
5059         bi->bi_next = NULL;
5060         bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5061
5062         stripe_sectors = conf->chunk_sectors *
5063                 (conf->raid_disks - conf->max_degraded);
5064         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5065                                                stripe_sectors);
5066         sector_div(last_sector, stripe_sectors);
5067
5068         logical_sector *= conf->chunk_sectors;
5069         last_sector *= conf->chunk_sectors;
5070
5071         for (; logical_sector < last_sector;
5072              logical_sector += STRIPE_SECTORS) {
5073                 DEFINE_WAIT(w);
5074                 int d;
5075         again:
5076                 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
5077                 prepare_to_wait(&conf->wait_for_overlap, &w,
5078                                 TASK_UNINTERRUPTIBLE);
5079                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5080                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5081                         release_stripe(sh);
5082                         schedule();
5083                         goto again;
5084                 }
5085                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5086                 spin_lock_irq(&sh->stripe_lock);
5087                 for (d = 0; d < conf->raid_disks; d++) {
5088                         if (d == sh->pd_idx || d == sh->qd_idx)
5089                                 continue;
5090                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5091                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5092                                 spin_unlock_irq(&sh->stripe_lock);
5093                                 release_stripe(sh);
5094                                 schedule();
5095                                 goto again;
5096                         }
5097                 }
5098                 set_bit(STRIPE_DISCARD, &sh->state);
5099                 finish_wait(&conf->wait_for_overlap, &w);
5100                 sh->overwrite_disks = 0;
5101                 for (d = 0; d < conf->raid_disks; d++) {
5102                         if (d == sh->pd_idx || d == sh->qd_idx)
5103                                 continue;
5104                         sh->dev[d].towrite = bi;
5105                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5106                         raid5_inc_bi_active_stripes(bi);
5107                         sh->overwrite_disks++;
5108                 }
5109                 spin_unlock_irq(&sh->stripe_lock);
5110                 if (conf->mddev->bitmap) {
5111                         for (d = 0;
5112                              d < conf->raid_disks - conf->max_degraded;
5113                              d++)
5114                                 bitmap_startwrite(mddev->bitmap,
5115                                                   sh->sector,
5116                                                   STRIPE_SECTORS,
5117                                                   0);
5118                         sh->bm_seq = conf->seq_flush + 1;
5119                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5120                 }
5121
5122                 set_bit(STRIPE_HANDLE, &sh->state);
5123                 clear_bit(STRIPE_DELAYED, &sh->state);
5124                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5125                         atomic_inc(&conf->preread_active_stripes);
5126                 release_stripe_plug(mddev, sh);
5127         }
5128
5129         remaining = raid5_dec_bi_active_stripes(bi);
5130         if (remaining == 0) {
5131                 md_write_end(mddev);
5132                 bio_endio(bi, 0);
5133         }
5134 }
5135
5136 static void make_request(struct mddev *mddev, struct bio * bi)
5137 {
5138         struct r5conf *conf = mddev->private;
5139         int dd_idx;
5140         sector_t new_sector;
5141         sector_t logical_sector, last_sector;
5142         struct stripe_head *sh;
5143         const int rw = bio_data_dir(bi);
5144         int remaining;
5145         DEFINE_WAIT(w);
5146         bool do_prepare;
5147
5148         if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5149                 md_flush_request(mddev, bi);
5150                 return;
5151         }
5152
5153         md_write_start(mddev, bi);
5154
5155         /*
5156          * If array is degraded, better not do chunk aligned read because
5157          * later we might have to read it again in order to reconstruct
5158          * data on failed drives.
5159          */
5160         if (rw == READ && mddev->degraded == 0 &&
5161              mddev->reshape_position == MaxSector &&
5162              chunk_aligned_read(mddev,bi))
5163                 return;
5164
5165         if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5166                 make_discard_request(mddev, bi);
5167                 return;
5168         }
5169
5170         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5171         last_sector = bio_end_sector(bi);
5172         bi->bi_next = NULL;
5173         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
5174
5175         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5176         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5177                 int previous;
5178                 int seq;
5179
5180                 do_prepare = false;
5181         retry:
5182                 seq = read_seqcount_begin(&conf->gen_lock);
5183                 previous = 0;
5184                 if (do_prepare)
5185                         prepare_to_wait(&conf->wait_for_overlap, &w,
5186                                 TASK_UNINTERRUPTIBLE);
5187                 if (unlikely(conf->reshape_progress != MaxSector)) {
5188                         /* spinlock is needed as reshape_progress may be
5189                          * 64bit on a 32bit platform, and so it might be
5190                          * possible to see a half-updated value
5191                          * Of course reshape_progress could change after
5192                          * the lock is dropped, so once we get a reference
5193                          * to the stripe that we think it is, we will have
5194                          * to check again.
5195                          */
5196                         spin_lock_irq(&conf->device_lock);
5197                         if (mddev->reshape_backwards
5198                             ? logical_sector < conf->reshape_progress
5199                             : logical_sector >= conf->reshape_progress) {
5200                                 previous = 1;
5201                         } else {
5202                                 if (mddev->reshape_backwards
5203                                     ? logical_sector < conf->reshape_safe
5204                                     : logical_sector >= conf->reshape_safe) {
5205                                         spin_unlock_irq(&conf->device_lock);
5206                                         schedule();
5207                                         do_prepare = true;
5208                                         goto retry;
5209                                 }
5210                         }
5211                         spin_unlock_irq(&conf->device_lock);
5212                 }
5213
5214                 new_sector = raid5_compute_sector(conf, logical_sector,
5215                                                   previous,
5216                                                   &dd_idx, NULL);
5217                 pr_debug("raid456: make_request, sector %llu logical %llu\n",
5218                         (unsigned long long)new_sector,
5219                         (unsigned long long)logical_sector);
5220
5221                 sh = get_active_stripe(conf, new_sector, previous,
5222                                        (bi->bi_rw&RWA_MASK), 0);
5223                 if (sh) {
5224                         if (unlikely(previous)) {
5225                                 /* expansion might have moved on while waiting for a
5226                                  * stripe, so we must do the range check again.
5227                                  * Expansion could still move past after this
5228                                  * test, but as we are holding a reference to
5229                                  * 'sh', we know that if that happens,
5230                                  *  STRIPE_EXPANDING will get set and the expansion
5231                                  * won't proceed until we finish with the stripe.
5232                                  */
5233                                 int must_retry = 0;
5234                                 spin_lock_irq(&conf->device_lock);
5235                                 if (mddev->reshape_backwards
5236                                     ? logical_sector >= conf->reshape_progress
5237                                     : logical_sector < conf->reshape_progress)
5238                                         /* mismatch, need to try again */
5239                                         must_retry = 1;
5240                                 spin_unlock_irq(&conf->device_lock);
5241                                 if (must_retry) {
5242                                         release_stripe(sh);
5243                                         schedule();
5244                                         do_prepare = true;
5245                                         goto retry;
5246                                 }
5247                         }
5248                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5249                                 /* Might have got the wrong stripe_head
5250                                  * by accident
5251                                  */
5252                                 release_stripe(sh);
5253                                 goto retry;
5254                         }
5255
5256                         if (rw == WRITE &&
5257                             logical_sector >= mddev->suspend_lo &&
5258                             logical_sector < mddev->suspend_hi) {
5259                                 release_stripe(sh);
5260                                 /* As the suspend_* range is controlled by
5261                                  * userspace, we want an interruptible
5262                                  * wait.
5263                                  */
5264                                 flush_signals(current);
5265                                 prepare_to_wait(&conf->wait_for_overlap,
5266                                                 &w, TASK_INTERRUPTIBLE);
5267                                 if (logical_sector >= mddev->suspend_lo &&
5268                                     logical_sector < mddev->suspend_hi) {
5269                                         schedule();
5270                                         do_prepare = true;
5271                                 }
5272                                 goto retry;
5273                         }
5274
5275                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5276                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5277                                 /* Stripe is busy expanding or
5278                                  * add failed due to overlap.  Flush everything
5279                                  * and wait a while
5280                                  */
5281                                 md_wakeup_thread(mddev->thread);
5282                                 release_stripe(sh);
5283                                 schedule();
5284                                 do_prepare = true;
5285                                 goto retry;
5286                         }
5287                         set_bit(STRIPE_HANDLE, &sh->state);
5288                         clear_bit(STRIPE_DELAYED, &sh->state);
5289                         if ((!sh->batch_head || sh == sh->batch_head) &&
5290                             (bi->bi_rw & REQ_SYNC) &&
5291                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5292                                 atomic_inc(&conf->preread_active_stripes);
5293                         release_stripe_plug(mddev, sh);
5294                 } else {
5295                         /* cannot get stripe for read-ahead, just give-up */
5296                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
5297                         break;
5298                 }
5299         }
5300         finish_wait(&conf->wait_for_overlap, &w);
5301
5302         remaining = raid5_dec_bi_active_stripes(bi);
5303         if (remaining == 0) {
5304
5305                 if ( rw == WRITE )
5306                         md_write_end(mddev);
5307
5308                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5309                                          bi, 0);
5310                 bio_endio(bi, 0);
5311         }
5312 }
5313
5314 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5315
5316 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5317 {
5318         /* reshaping is quite different to recovery/resync so it is
5319          * handled quite separately ... here.
5320          *
5321          * On each call to sync_request, we gather one chunk worth of
5322          * destination stripes and flag them as expanding.
5323          * Then we find all the source stripes and request reads.
5324          * As the reads complete, handle_stripe will copy the data
5325          * into the destination stripe and release that stripe.
5326          */
5327         struct r5conf *conf = mddev->private;
5328         struct stripe_head *sh;
5329         sector_t first_sector, last_sector;
5330         int raid_disks = conf->previous_raid_disks;
5331         int data_disks = raid_disks - conf->max_degraded;
5332         int new_data_disks = conf->raid_disks - conf->max_degraded;
5333         int i;
5334         int dd_idx;
5335         sector_t writepos, readpos, safepos;
5336         sector_t stripe_addr;
5337         int reshape_sectors;
5338         struct list_head stripes;
5339
5340         if (sector_nr == 0) {
5341                 /* If restarting in the middle, skip the initial sectors */
5342                 if (mddev->reshape_backwards &&
5343                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5344                         sector_nr = raid5_size(mddev, 0, 0)
5345                                 - conf->reshape_progress;
5346                 } else if (!mddev->reshape_backwards &&
5347                            conf->reshape_progress > 0)
5348                         sector_nr = conf->reshape_progress;
5349                 sector_div(sector_nr, new_data_disks);
5350                 if (sector_nr) {
5351                         mddev->curr_resync_completed = sector_nr;
5352                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5353                         *skipped = 1;
5354                         return sector_nr;
5355                 }
5356         }
5357
5358         /* We need to process a full chunk at a time.
5359          * If old and new chunk sizes differ, we need to process the
5360          * largest of these
5361          */
5362         if (mddev->new_chunk_sectors > mddev->chunk_sectors)
5363                 reshape_sectors = mddev->new_chunk_sectors;
5364         else
5365                 reshape_sectors = mddev->chunk_sectors;
5366
5367         /* We update the metadata at least every 10 seconds, or when
5368          * the data about to be copied would over-write the source of
5369          * the data at the front of the range.  i.e. one new_stripe
5370          * along from reshape_progress new_maps to after where
5371          * reshape_safe old_maps to
5372          */
5373         writepos = conf->reshape_progress;
5374         sector_div(writepos, new_data_disks);
5375         readpos = conf->reshape_progress;
5376         sector_div(readpos, data_disks);
5377         safepos = conf->reshape_safe;
5378         sector_div(safepos, data_disks);
5379         if (mddev->reshape_backwards) {
5380                 writepos -= min_t(sector_t, reshape_sectors, writepos);
5381                 readpos += reshape_sectors;
5382                 safepos += reshape_sectors;
5383         } else {
5384                 writepos += reshape_sectors;
5385                 readpos -= min_t(sector_t, reshape_sectors, readpos);
5386                 safepos -= min_t(sector_t, reshape_sectors, safepos);
5387         }
5388
5389         /* Having calculated the 'writepos' possibly use it
5390          * to set 'stripe_addr' which is where we will write to.
5391          */
5392         if (mddev->reshape_backwards) {
5393                 BUG_ON(conf->reshape_progress == 0);
5394                 stripe_addr = writepos;
5395                 BUG_ON((mddev->dev_sectors &
5396                         ~((sector_t)reshape_sectors - 1))
5397                        - reshape_sectors - stripe_addr
5398                        != sector_nr);
5399         } else {
5400                 BUG_ON(writepos != sector_nr + reshape_sectors);
5401                 stripe_addr = sector_nr;
5402         }
5403
5404         /* 'writepos' is the most advanced device address we might write.
5405          * 'readpos' is the least advanced device address we might read.
5406          * 'safepos' is the least address recorded in the metadata as having
5407          *     been reshaped.
5408          * If there is a min_offset_diff, these are adjusted either by
5409          * increasing the safepos/readpos if diff is negative, or
5410          * increasing writepos if diff is positive.
5411          * If 'readpos' is then behind 'writepos', there is no way that we can
5412          * ensure safety in the face of a crash - that must be done by userspace
5413          * making a backup of the data.  So in that case there is no particular
5414          * rush to update metadata.
5415          * Otherwise if 'safepos' is behind 'writepos', then we really need to
5416          * update the metadata to advance 'safepos' to match 'readpos' so that
5417          * we can be safe in the event of a crash.
5418          * So we insist on updating metadata if safepos is behind writepos and
5419          * readpos is beyond writepos.
5420          * In any case, update the metadata every 10 seconds.
5421          * Maybe that number should be configurable, but I'm not sure it is
5422          * worth it.... maybe it could be a multiple of safemode_delay???
5423          */
5424         if (conf->min_offset_diff < 0) {
5425                 safepos += -conf->min_offset_diff;
5426                 readpos += -conf->min_offset_diff;
5427         } else
5428                 writepos += conf->min_offset_diff;
5429
5430         if ((mddev->reshape_backwards
5431              ? (safepos > writepos && readpos < writepos)
5432              : (safepos < writepos && readpos > writepos)) ||
5433             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5434                 /* Cannot proceed until we've updated the superblock... */
5435                 wait_event(conf->wait_for_overlap,
5436                            atomic_read(&conf->reshape_stripes)==0
5437                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5438                 if (atomic_read(&conf->reshape_stripes) != 0)
5439                         return 0;
5440                 mddev->reshape_position = conf->reshape_progress;
5441                 mddev->curr_resync_completed = sector_nr;
5442                 conf->reshape_checkpoint = jiffies;
5443                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5444                 md_wakeup_thread(mddev->thread);
5445                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
5446                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5447                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5448                         return 0;
5449                 spin_lock_irq(&conf->device_lock);
5450                 conf->reshape_safe = mddev->reshape_position;
5451                 spin_unlock_irq(&conf->device_lock);
5452                 wake_up(&conf->wait_for_overlap);
5453                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5454         }
5455
5456         INIT_LIST_HEAD(&stripes);
5457         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5458                 int j;
5459                 int skipped_disk = 0;
5460                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5461                 set_bit(STRIPE_EXPANDING, &sh->state);
5462                 atomic_inc(&conf->reshape_stripes);
5463                 /* If any of this stripe is beyond the end of the old
5464                  * array, then we need to zero those blocks
5465                  */
5466                 for (j=sh->disks; j--;) {
5467                         sector_t s;
5468                         if (j == sh->pd_idx)
5469                                 continue;
5470                         if (conf->level == 6 &&
5471                             j == sh->qd_idx)
5472                                 continue;
5473                         s = compute_blocknr(sh, j, 0);
5474                         if (s < raid5_size(mddev, 0, 0)) {
5475                                 skipped_disk = 1;
5476                                 continue;
5477                         }
5478                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5479                         set_bit(R5_Expanded, &sh->dev[j].flags);
5480                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
5481                 }
5482                 if (!skipped_disk) {
5483                         set_bit(STRIPE_EXPAND_READY, &sh->state);
5484                         set_bit(STRIPE_HANDLE, &sh->state);
5485                 }
5486                 list_add(&sh->lru, &stripes);
5487         }
5488         spin_lock_irq(&conf->device_lock);
5489         if (mddev->reshape_backwards)
5490                 conf->reshape_progress -= reshape_sectors * new_data_disks;
5491         else
5492                 conf->reshape_progress += reshape_sectors * new_data_disks;
5493         spin_unlock_irq(&conf->device_lock);
5494         /* Ok, those stripe are ready. We can start scheduling
5495          * reads on the source stripes.
5496          * The source stripes are determined by mapping the first and last
5497          * block on the destination stripes.
5498          */
5499         first_sector =
5500                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5501                                      1, &dd_idx, NULL);
5502         last_sector =
5503                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5504                                             * new_data_disks - 1),
5505                                      1, &dd_idx, NULL);
5506         if (last_sector >= mddev->dev_sectors)
5507                 last_sector = mddev->dev_sectors - 1;
5508         while (first_sector <= last_sector) {
5509                 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
5510                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5511                 set_bit(STRIPE_HANDLE, &sh->state);
5512                 release_stripe(sh);
5513                 first_sector += STRIPE_SECTORS;
5514         }
5515         /* Now that the sources are clearly marked, we can release
5516          * the destination stripes
5517          */
5518         while (!list_empty(&stripes)) {
5519                 sh = list_entry(stripes.next, struct stripe_head, lru);
5520                 list_del_init(&sh->lru);
5521                 release_stripe(sh);
5522         }
5523         /* If this takes us to the resync_max point where we have to pause,
5524          * then we need to write out the superblock.
5525          */
5526         sector_nr += reshape_sectors;
5527         if ((sector_nr - mddev->curr_resync_completed) * 2
5528             >= mddev->resync_max - mddev->curr_resync_completed) {
5529                 /* Cannot proceed until we've updated the superblock... */
5530                 wait_event(conf->wait_for_overlap,
5531                            atomic_read(&conf->reshape_stripes) == 0
5532                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5533                 if (atomic_read(&conf->reshape_stripes) != 0)
5534                         goto ret;
5535                 mddev->reshape_position = conf->reshape_progress;
5536                 mddev->curr_resync_completed = sector_nr;
5537                 conf->reshape_checkpoint = jiffies;
5538                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5539                 md_wakeup_thread(mddev->thread);
5540                 wait_event(mddev->sb_wait,
5541                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5542                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5543                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5544                         goto ret;
5545                 spin_lock_irq(&conf->device_lock);
5546                 conf->reshape_safe = mddev->reshape_position;
5547                 spin_unlock_irq(&conf->device_lock);
5548                 wake_up(&conf->wait_for_overlap);
5549                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5550         }
5551 ret:
5552         return reshape_sectors;
5553 }
5554
5555 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5556 {
5557         struct r5conf *conf = mddev->private;
5558         struct stripe_head *sh;
5559         sector_t max_sector = mddev->dev_sectors;
5560         sector_t sync_blocks;
5561         int still_degraded = 0;
5562         int i;
5563
5564         if (sector_nr >= max_sector) {
5565                 /* just being told to finish up .. nothing much to do */
5566
5567                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5568                         end_reshape(conf);
5569                         return 0;
5570                 }
5571
5572                 if (mddev->curr_resync < max_sector) /* aborted */
5573                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5574                                         &sync_blocks, 1);
5575                 else /* completed sync */
5576                         conf->fullsync = 0;
5577                 bitmap_close_sync(mddev->bitmap);
5578
5579                 return 0;
5580         }
5581
5582         /* Allow raid5_quiesce to complete */
5583         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5584
5585         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5586                 return reshape_request(mddev, sector_nr, skipped);
5587
5588         /* No need to check resync_max as we never do more than one
5589          * stripe, and as resync_max will always be on a chunk boundary,
5590          * if the check in md_do_sync didn't fire, there is no chance
5591          * of overstepping resync_max here
5592          */
5593
5594         /* if there is too many failed drives and we are trying
5595          * to resync, then assert that we are finished, because there is
5596          * nothing we can do.
5597          */
5598         if (mddev->degraded >= conf->max_degraded &&
5599             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5600                 sector_t rv = mddev->dev_sectors - sector_nr;
5601                 *skipped = 1;
5602                 return rv;
5603         }
5604         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5605             !conf->fullsync &&
5606             !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5607             sync_blocks >= STRIPE_SECTORS) {
5608                 /* we can skip this block, and probably more */
5609                 sync_blocks /= STRIPE_SECTORS;
5610                 *skipped = 1;
5611                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5612         }
5613
5614         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5615
5616         sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5617         if (sh == NULL) {
5618                 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5619                 /* make sure we don't swamp the stripe cache if someone else
5620                  * is trying to get access
5621                  */
5622                 schedule_timeout_uninterruptible(1);
5623         }
5624         /* Need to check if array will still be degraded after recovery/resync
5625          * Note in case of > 1 drive failures it's possible we're rebuilding
5626          * one drive while leaving another faulty drive in array.
5627          */
5628         rcu_read_lock();
5629         for (i = 0; i < conf->raid_disks; i++) {
5630                 struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5631
5632                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5633                         still_degraded = 1;
5634         }
5635         rcu_read_unlock();
5636
5637         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5638
5639         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5640         set_bit(STRIPE_HANDLE, &sh->state);
5641
5642         release_stripe(sh);
5643
5644         return STRIPE_SECTORS;
5645 }
5646
5647 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5648 {
5649         /* We may not be able to submit a whole bio at once as there
5650          * may not be enough stripe_heads available.
5651          * We cannot pre-allocate enough stripe_heads as we may need
5652          * more than exist in the cache (if we allow ever large chunks).
5653          * So we do one stripe head at a time and record in
5654          * ->bi_hw_segments how many have been done.
5655          *
5656          * We *know* that this entire raid_bio is in one chunk, so
5657          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5658          */
5659         struct stripe_head *sh;
5660         int dd_idx;
5661         sector_t sector, logical_sector, last_sector;
5662         int scnt = 0;
5663         int remaining;
5664         int handled = 0;
5665
5666         logical_sector = raid_bio->bi_iter.bi_sector &
5667                 ~((sector_t)STRIPE_SECTORS-1);
5668         sector = raid5_compute_sector(conf, logical_sector,
5669                                       0, &dd_idx, NULL);
5670         last_sector = bio_end_sector(raid_bio);
5671
5672         for (; logical_sector < last_sector;
5673              logical_sector += STRIPE_SECTORS,
5674                      sector += STRIPE_SECTORS,
5675                      scnt++) {
5676
5677                 if (scnt < raid5_bi_processed_stripes(raid_bio))
5678                         /* already done this stripe */
5679                         continue;
5680
5681                 sh = get_active_stripe(conf, sector, 0, 1, 1);
5682
5683                 if (!sh) {
5684                         /* failed to get a stripe - must wait */
5685                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5686                         conf->retry_read_aligned = raid_bio;
5687                         return handled;
5688                 }
5689
5690                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5691                         release_stripe(sh);
5692                         raid5_set_bi_processed_stripes(raid_bio, scnt);
5693                         conf->retry_read_aligned = raid_bio;
5694                         return handled;
5695                 }
5696
5697                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5698                 handle_stripe(sh);
5699                 release_stripe(sh);
5700                 handled++;
5701         }
5702         remaining = raid5_dec_bi_active_stripes(raid_bio);
5703         if (remaining == 0) {
5704                 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5705                                          raid_bio, 0);
5706                 bio_endio(raid_bio, 0);
5707         }
5708         if (atomic_dec_and_test(&conf->active_aligned_reads))
5709                 wake_up(&conf->wait_for_stripe);
5710         return handled;
5711 }
5712
5713 static int handle_active_stripes(struct r5conf *conf, int group,
5714                                  struct r5worker *worker,
5715                                  struct list_head *temp_inactive_list)
5716 {
5717         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5718         int i, batch_size = 0, hash;
5719         bool release_inactive = false;
5720
5721         while (batch_size < MAX_STRIPE_BATCH &&
5722                         (sh = __get_priority_stripe(conf, group)) != NULL)
5723                 batch[batch_size++] = sh;
5724
5725         if (batch_size == 0) {
5726                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5727                         if (!list_empty(temp_inactive_list + i))
5728                                 break;
5729                 if (i == NR_STRIPE_HASH_LOCKS)
5730                         return batch_size;
5731                 release_inactive = true;
5732         }
5733         spin_unlock_irq(&conf->device_lock);
5734
5735         release_inactive_stripe_list(conf, temp_inactive_list,
5736                                      NR_STRIPE_HASH_LOCKS);
5737
5738         if (release_inactive) {
5739                 spin_lock_irq(&conf->device_lock);
5740                 return 0;
5741         }
5742
5743         for (i = 0; i < batch_size; i++)
5744                 handle_stripe(batch[i]);
5745
5746         cond_resched();
5747
5748         spin_lock_irq(&conf->device_lock);
5749         for (i = 0; i < batch_size; i++) {
5750                 hash = batch[i]->hash_lock_index;
5751                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5752         }
5753         return batch_size;
5754 }
5755
5756 static void raid5_do_work(struct work_struct *work)
5757 {
5758         struct r5worker *worker = container_of(work, struct r5worker, work);
5759         struct r5worker_group *group = worker->group;
5760         struct r5conf *conf = group->conf;
5761         int group_id = group - conf->worker_groups;
5762         int handled;
5763         struct blk_plug plug;
5764
5765         pr_debug("+++ raid5worker active\n");
5766
5767         blk_start_plug(&plug);
5768         handled = 0;
5769         spin_lock_irq(&conf->device_lock);
5770         while (1) {
5771                 int batch_size, released;
5772
5773                 released = release_stripe_list(conf, worker->temp_inactive_list);
5774
5775                 batch_size = handle_active_stripes(conf, group_id, worker,
5776                                                    worker->temp_inactive_list);
5777                 worker->working = false;
5778                 if (!batch_size && !released)
5779                         break;
5780                 handled += batch_size;
5781         }
5782         pr_debug("%d stripes handled\n", handled);
5783
5784         spin_unlock_irq(&conf->device_lock);
5785         blk_finish_plug(&plug);
5786
5787         pr_debug("--- raid5worker inactive\n");
5788 }
5789
5790 /*
5791  * This is our raid5 kernel thread.
5792  *
5793  * We scan the hash table for stripes which can be handled now.
5794  * During the scan, completed stripes are saved for us by the interrupt
5795  * handler, so that they will not have to wait for our next wakeup.
5796  */
5797 static void raid5d(struct md_thread *thread)
5798 {
5799         struct mddev *mddev = thread->mddev;
5800         struct r5conf *conf = mddev->private;
5801         int handled;
5802         struct blk_plug plug;
5803
5804         pr_debug("+++ raid5d active\n");
5805
5806         md_check_recovery(mddev);
5807
5808         blk_start_plug(&plug);
5809         handled = 0;
5810         spin_lock_irq(&conf->device_lock);
5811         while (1) {
5812                 struct bio *bio;
5813                 int batch_size, released;
5814
5815                 released = release_stripe_list(conf, conf->temp_inactive_list);
5816                 if (released)
5817                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
5818
5819                 if (
5820                     !list_empty(&conf->bitmap_list)) {
5821                         /* Now is a good time to flush some bitmap updates */
5822                         conf->seq_flush++;
5823                         spin_unlock_irq(&conf->device_lock);
5824                         bitmap_unplug(mddev->bitmap);
5825                         spin_lock_irq(&conf->device_lock);
5826                         conf->seq_write = conf->seq_flush;
5827                         activate_bit_delay(conf, conf->temp_inactive_list);
5828                 }
5829                 raid5_activate_delayed(conf);
5830
5831                 while ((bio = remove_bio_from_retry(conf))) {
5832                         int ok;
5833                         spin_unlock_irq(&conf->device_lock);
5834                         ok = retry_aligned_read(conf, bio);
5835                         spin_lock_irq(&conf->device_lock);
5836                         if (!ok)
5837                                 break;
5838                         handled++;
5839                 }
5840
5841                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5842                                                    conf->temp_inactive_list);
5843                 if (!batch_size && !released)
5844                         break;
5845                 handled += batch_size;
5846
5847                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5848                         spin_unlock_irq(&conf->device_lock);
5849                         md_check_recovery(mddev);
5850                         spin_lock_irq(&conf->device_lock);
5851                 }
5852         }
5853         pr_debug("%d stripes handled\n", handled);
5854
5855         spin_unlock_irq(&conf->device_lock);
5856         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
5857             mutex_trylock(&conf->cache_size_mutex)) {
5858                 grow_one_stripe(conf, __GFP_NOWARN);
5859                 /* Set flag even if allocation failed.  This helps
5860                  * slow down allocation requests when mem is short
5861                  */
5862                 set_bit(R5_DID_ALLOC, &conf->cache_state);
5863                 mutex_unlock(&conf->cache_size_mutex);
5864         }
5865
5866         async_tx_issue_pending_all();
5867         blk_finish_plug(&plug);
5868
5869         pr_debug("--- raid5d inactive\n");
5870 }
5871
5872 static ssize_t
5873 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5874 {
5875         struct r5conf *conf;
5876         int ret = 0;
5877         spin_lock(&mddev->lock);
5878         conf = mddev->private;
5879         if (conf)
5880                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5881         spin_unlock(&mddev->lock);
5882         return ret;
5883 }
5884
5885 int
5886 raid5_set_cache_size(struct mddev *mddev, int size)
5887 {
5888         struct r5conf *conf = mddev->private;
5889         int err;
5890
5891         if (size <= 16 || size > 32768)
5892                 return -EINVAL;
5893
5894         conf->min_nr_stripes = size;
5895         mutex_lock(&conf->cache_size_mutex);
5896         while (size < conf->max_nr_stripes &&
5897                drop_one_stripe(conf))
5898                 ;
5899         mutex_unlock(&conf->cache_size_mutex);
5900
5901
5902         err = md_allow_write(mddev);
5903         if (err)
5904                 return err;
5905
5906         mutex_lock(&conf->cache_size_mutex);
5907         while (size > conf->max_nr_stripes)
5908                 if (!grow_one_stripe(conf, GFP_KERNEL))
5909                         break;
5910         mutex_unlock(&conf->cache_size_mutex);
5911
5912         return 0;
5913 }
5914 EXPORT_SYMBOL(raid5_set_cache_size);
5915
5916 static ssize_t
5917 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5918 {
5919         struct r5conf *conf;
5920         unsigned long new;
5921         int err;
5922
5923         if (len >= PAGE_SIZE)
5924                 return -EINVAL;
5925         if (kstrtoul(page, 10, &new))
5926                 return -EINVAL;
5927         err = mddev_lock(mddev);
5928         if (err)
5929                 return err;
5930         conf = mddev->private;
5931         if (!conf)
5932                 err = -ENODEV;
5933         else
5934                 err = raid5_set_cache_size(mddev, new);
5935         mddev_unlock(mddev);
5936
5937         return err ?: len;
5938 }
5939
5940 static struct md_sysfs_entry
5941 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5942                                 raid5_show_stripe_cache_size,
5943                                 raid5_store_stripe_cache_size);
5944
5945 static ssize_t
5946 raid5_show_rmw_level(struct mddev  *mddev, char *page)
5947 {
5948         struct r5conf *conf = mddev->private;
5949         if (conf)
5950                 return sprintf(page, "%d\n", conf->rmw_level);
5951         else
5952                 return 0;
5953 }
5954
5955 static ssize_t
5956 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
5957 {
5958         struct r5conf *conf = mddev->private;
5959         unsigned long new;
5960
5961         if (!conf)
5962                 return -ENODEV;
5963
5964         if (len >= PAGE_SIZE)
5965                 return -EINVAL;
5966
5967         if (kstrtoul(page, 10, &new))
5968                 return -EINVAL;
5969
5970         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
5971                 return -EINVAL;
5972
5973         if (new != PARITY_DISABLE_RMW &&
5974             new != PARITY_ENABLE_RMW &&
5975             new != PARITY_PREFER_RMW)
5976                 return -EINVAL;
5977
5978         conf->rmw_level = new;
5979         return len;
5980 }
5981
5982 static struct md_sysfs_entry
5983 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
5984                          raid5_show_rmw_level,
5985                          raid5_store_rmw_level);
5986
5987
5988 static ssize_t
5989 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5990 {
5991         struct r5conf *conf;
5992         int ret = 0;
5993         spin_lock(&mddev->lock);
5994         conf = mddev->private;
5995         if (conf)
5996                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
5997         spin_unlock(&mddev->lock);
5998         return ret;
5999 }
6000
6001 static ssize_t
6002 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6003 {
6004         struct r5conf *conf;
6005         unsigned long new;
6006         int err;
6007
6008         if (len >= PAGE_SIZE)
6009                 return -EINVAL;
6010         if (kstrtoul(page, 10, &new))
6011                 return -EINVAL;
6012
6013         err = mddev_lock(mddev);
6014         if (err)
6015                 return err;
6016         conf = mddev->private;
6017         if (!conf)
6018                 err = -ENODEV;
6019         else if (new > conf->min_nr_stripes)
6020                 err = -EINVAL;
6021         else
6022                 conf->bypass_threshold = new;
6023         mddev_unlock(mddev);
6024         return err ?: len;
6025 }
6026
6027 static struct md_sysfs_entry
6028 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6029                                         S_IRUGO | S_IWUSR,
6030                                         raid5_show_preread_threshold,
6031                                         raid5_store_preread_threshold);
6032
6033 static ssize_t
6034 raid5_show_skip_copy(struct mddev *mddev, char *page)
6035 {
6036         struct r5conf *conf;
6037         int ret = 0;
6038         spin_lock(&mddev->lock);
6039         conf = mddev->private;
6040         if (conf)
6041                 ret = sprintf(page, "%d\n", conf->skip_copy);
6042         spin_unlock(&mddev->lock);
6043         return ret;
6044 }
6045
6046 static ssize_t
6047 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6048 {
6049         struct r5conf *conf;
6050         unsigned long new;
6051         int err;
6052
6053         if (len >= PAGE_SIZE)
6054                 return -EINVAL;
6055         if (kstrtoul(page, 10, &new))
6056                 return -EINVAL;
6057         new = !!new;
6058
6059         err = mddev_lock(mddev);
6060         if (err)
6061                 return err;
6062         conf = mddev->private;
6063         if (!conf)
6064                 err = -ENODEV;
6065         else if (new != conf->skip_copy) {
6066                 mddev_suspend(mddev);
6067                 conf->skip_copy = new;
6068                 if (new)
6069                         mddev->queue->backing_dev_info.capabilities |=
6070                                 BDI_CAP_STABLE_WRITES;
6071                 else
6072                         mddev->queue->backing_dev_info.capabilities &=
6073                                 ~BDI_CAP_STABLE_WRITES;
6074                 mddev_resume(mddev);
6075         }
6076         mddev_unlock(mddev);
6077         return err ?: len;
6078 }
6079
6080 static struct md_sysfs_entry
6081 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6082                                         raid5_show_skip_copy,
6083                                         raid5_store_skip_copy);
6084
6085 static ssize_t
6086 stripe_cache_active_show(struct mddev *mddev, char *page)
6087 {
6088         struct r5conf *conf = mddev->private;
6089         if (conf)
6090                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6091         else
6092                 return 0;
6093 }
6094
6095 static struct md_sysfs_entry
6096 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6097
6098 static ssize_t
6099 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6100 {
6101         struct r5conf *conf;
6102         int ret = 0;
6103         spin_lock(&mddev->lock);
6104         conf = mddev->private;
6105         if (conf)
6106                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6107         spin_unlock(&mddev->lock);
6108         return ret;
6109 }
6110
6111 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6112                                int *group_cnt,
6113                                int *worker_cnt_per_group,
6114                                struct r5worker_group **worker_groups);
6115 static ssize_t
6116 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6117 {
6118         struct r5conf *conf;
6119         unsigned long new;
6120         int err;
6121         struct r5worker_group *new_groups, *old_groups;
6122         int group_cnt, worker_cnt_per_group;
6123
6124         if (len >= PAGE_SIZE)
6125                 return -EINVAL;
6126         if (kstrtoul(page, 10, &new))
6127                 return -EINVAL;
6128
6129         err = mddev_lock(mddev);
6130         if (err)
6131                 return err;
6132         conf = mddev->private;
6133         if (!conf)
6134                 err = -ENODEV;
6135         else if (new != conf->worker_cnt_per_group) {
6136                 mddev_suspend(mddev);
6137
6138                 old_groups = conf->worker_groups;
6139                 if (old_groups)
6140                         flush_workqueue(raid5_wq);
6141
6142                 err = alloc_thread_groups(conf, new,
6143                                           &group_cnt, &worker_cnt_per_group,
6144                                           &new_groups);
6145                 if (!err) {
6146                         spin_lock_irq(&conf->device_lock);
6147                         conf->group_cnt = group_cnt;
6148                         conf->worker_cnt_per_group = worker_cnt_per_group;
6149                         conf->worker_groups = new_groups;
6150                         spin_unlock_irq(&conf->device_lock);
6151
6152                         if (old_groups)
6153                                 kfree(old_groups[0].workers);
6154                         kfree(old_groups);
6155                 }
6156                 mddev_resume(mddev);
6157         }
6158         mddev_unlock(mddev);
6159
6160         return err ?: len;
6161 }
6162
6163 static struct md_sysfs_entry
6164 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6165                                 raid5_show_group_thread_cnt,
6166                                 raid5_store_group_thread_cnt);
6167
6168 static struct attribute *raid5_attrs[] =  {
6169         &raid5_stripecache_size.attr,
6170         &raid5_stripecache_active.attr,
6171         &raid5_preread_bypass_threshold.attr,
6172         &raid5_group_thread_cnt.attr,
6173         &raid5_skip_copy.attr,
6174         &raid5_rmw_level.attr,
6175         NULL,
6176 };
6177 static struct attribute_group raid5_attrs_group = {
6178         .name = NULL,
6179         .attrs = raid5_attrs,
6180 };
6181
6182 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6183                                int *group_cnt,
6184                                int *worker_cnt_per_group,
6185                                struct r5worker_group **worker_groups)
6186 {
6187         int i, j, k;
6188         ssize_t size;
6189         struct r5worker *workers;
6190
6191         *worker_cnt_per_group = cnt;
6192         if (cnt == 0) {
6193                 *group_cnt = 0;
6194                 *worker_groups = NULL;
6195                 return 0;
6196         }
6197         *group_cnt = num_possible_nodes();
6198         size = sizeof(struct r5worker) * cnt;
6199         workers = kzalloc(size * *group_cnt, GFP_NOIO);
6200         *worker_groups = kzalloc(sizeof(struct r5worker_group) *
6201                                 *group_cnt, GFP_NOIO);
6202         if (!*worker_groups || !workers) {
6203                 kfree(workers);
6204                 kfree(*worker_groups);
6205                 return -ENOMEM;
6206         }
6207
6208         for (i = 0; i < *group_cnt; i++) {
6209                 struct r5worker_group *group;
6210
6211                 group = &(*worker_groups)[i];
6212                 INIT_LIST_HEAD(&group->handle_list);
6213                 group->conf = conf;
6214                 group->workers = workers + i * cnt;
6215
6216                 for (j = 0; j < cnt; j++) {
6217                         struct r5worker *worker = group->workers + j;
6218                         worker->group = group;
6219                         INIT_WORK(&worker->work, raid5_do_work);
6220
6221                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6222                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6223                 }
6224         }
6225
6226         return 0;
6227 }
6228
6229 static void free_thread_groups(struct r5conf *conf)
6230 {
6231         if (conf->worker_groups)
6232                 kfree(conf->worker_groups[0].workers);
6233         kfree(conf->worker_groups);
6234         conf->worker_groups = NULL;
6235 }
6236
6237 static sector_t
6238 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6239 {
6240         struct r5conf *conf = mddev->private;
6241
6242         if (!sectors)
6243                 sectors = mddev->dev_sectors;
6244         if (!raid_disks)
6245                 /* size is defined by the smallest of previous and new size */
6246                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6247
6248         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6249         sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
6250         return sectors * (raid_disks - conf->max_degraded);
6251 }
6252
6253 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6254 {
6255         safe_put_page(percpu->spare_page);
6256         if (percpu->scribble)
6257                 flex_array_free(percpu->scribble);
6258         percpu->spare_page = NULL;
6259         percpu->scribble = NULL;
6260 }
6261
6262 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6263 {
6264         if (conf->level == 6 && !percpu->spare_page)
6265                 percpu->spare_page = alloc_page(GFP_KERNEL);
6266         if (!percpu->scribble)
6267                 percpu->scribble = scribble_alloc(max(conf->raid_disks,
6268                                                       conf->previous_raid_disks),
6269                                                   max(conf->chunk_sectors,
6270                                                       conf->prev_chunk_sectors)
6271                                                    / STRIPE_SECTORS,
6272                                                   GFP_KERNEL);
6273
6274         if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6275                 free_scratch_buffer(conf, percpu);
6276                 return -ENOMEM;
6277         }
6278
6279         return 0;
6280 }
6281
6282 static void raid5_free_percpu(struct r5conf *conf)
6283 {
6284         unsigned long cpu;
6285
6286         if (!conf->percpu)
6287                 return;
6288
6289 #ifdef CONFIG_HOTPLUG_CPU
6290         unregister_cpu_notifier(&conf->cpu_notify);
6291 #endif
6292
6293         get_online_cpus();
6294         for_each_possible_cpu(cpu)
6295                 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6296         put_online_cpus();
6297
6298         free_percpu(conf->percpu);
6299 }
6300
6301 static void free_conf(struct r5conf *conf)
6302 {
6303         if (conf->shrinker.seeks)
6304                 unregister_shrinker(&conf->shrinker);
6305         free_thread_groups(conf);
6306         shrink_stripes(conf);
6307         raid5_free_percpu(conf);
6308         kfree(conf->disks);
6309         kfree(conf->stripe_hashtbl);
6310         kfree(conf);
6311 }
6312
6313 #ifdef CONFIG_HOTPLUG_CPU
6314 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6315                               void *hcpu)
6316 {
6317         struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6318         long cpu = (long)hcpu;
6319         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6320
6321         switch (action) {
6322         case CPU_UP_PREPARE:
6323         case CPU_UP_PREPARE_FROZEN:
6324                 if (alloc_scratch_buffer(conf, percpu)) {
6325                         pr_err("%s: failed memory allocation for cpu%ld\n",
6326                                __func__, cpu);
6327                         return notifier_from_errno(-ENOMEM);
6328                 }
6329                 break;
6330         case CPU_DEAD:
6331         case CPU_DEAD_FROZEN:
6332                 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6333                 break;
6334         default:
6335                 break;
6336         }
6337         return NOTIFY_OK;
6338 }
6339 #endif
6340
6341 static int raid5_alloc_percpu(struct r5conf *conf)
6342 {
6343         unsigned long cpu;
6344         int err = 0;
6345
6346         conf->percpu = alloc_percpu(struct raid5_percpu);
6347         if (!conf->percpu)
6348                 return -ENOMEM;
6349
6350 #ifdef CONFIG_HOTPLUG_CPU
6351         conf->cpu_notify.notifier_call = raid456_cpu_notify;
6352         conf->cpu_notify.priority = 0;
6353         err = register_cpu_notifier(&conf->cpu_notify);
6354         if (err)
6355                 return err;
6356 #endif
6357
6358         get_online_cpus();
6359         for_each_present_cpu(cpu) {
6360                 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6361                 if (err) {
6362                         pr_err("%s: failed memory allocation for cpu%ld\n",
6363                                __func__, cpu);
6364                         break;
6365                 }
6366                 spin_lock_init(&per_cpu_ptr(conf->percpu, cpu)->lock);
6367         }
6368         put_online_cpus();
6369
6370         return err;
6371 }
6372
6373 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6374                                       struct shrink_control *sc)
6375 {
6376         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6377         unsigned long ret = SHRINK_STOP;
6378
6379         if (mutex_trylock(&conf->cache_size_mutex)) {
6380                 ret= 0;
6381                 while (ret < sc->nr_to_scan &&
6382                        conf->max_nr_stripes > conf->min_nr_stripes) {
6383                         if (drop_one_stripe(conf) == 0) {
6384                                 ret = SHRINK_STOP;
6385                                 break;
6386                         }
6387                         ret++;
6388                 }
6389                 mutex_unlock(&conf->cache_size_mutex);
6390         }
6391         return ret;
6392 }
6393
6394 static unsigned long raid5_cache_count(struct shrinker *shrink,
6395                                        struct shrink_control *sc)
6396 {
6397         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6398
6399         if (conf->max_nr_stripes < conf->min_nr_stripes)
6400                 /* unlikely, but not impossible */
6401                 return 0;
6402         return conf->max_nr_stripes - conf->min_nr_stripes;
6403 }
6404
6405 static struct r5conf *setup_conf(struct mddev *mddev)
6406 {
6407         struct r5conf *conf;
6408         int raid_disk, memory, max_disks;
6409         struct md_rdev *rdev;
6410         struct disk_info *disk;
6411         char pers_name[6];
6412         int i;
6413         int group_cnt, worker_cnt_per_group;
6414         struct r5worker_group *new_group;
6415
6416         if (mddev->new_level != 5
6417             && mddev->new_level != 4
6418             && mddev->new_level != 6) {
6419                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6420                        mdname(mddev), mddev->new_level);
6421                 return ERR_PTR(-EIO);
6422         }
6423         if ((mddev->new_level == 5
6424              && !algorithm_valid_raid5(mddev->new_layout)) ||
6425             (mddev->new_level == 6
6426              && !algorithm_valid_raid6(mddev->new_layout))) {
6427                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6428                        mdname(mddev), mddev->new_layout);
6429                 return ERR_PTR(-EIO);
6430         }
6431         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6432                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6433                        mdname(mddev), mddev->raid_disks);
6434                 return ERR_PTR(-EINVAL);
6435         }
6436
6437         if (!mddev->new_chunk_sectors ||
6438             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6439             !is_power_of_2(mddev->new_chunk_sectors)) {
6440                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6441                        mdname(mddev), mddev->new_chunk_sectors << 9);
6442                 return ERR_PTR(-EINVAL);
6443         }
6444
6445         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6446         if (conf == NULL)
6447                 goto abort;
6448         /* Don't enable multi-threading by default*/
6449         if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6450                                  &new_group)) {
6451                 conf->group_cnt = group_cnt;
6452                 conf->worker_cnt_per_group = worker_cnt_per_group;
6453                 conf->worker_groups = new_group;
6454         } else
6455                 goto abort;
6456         spin_lock_init(&conf->device_lock);
6457         seqcount_init(&conf->gen_lock);
6458         mutex_init(&conf->cache_size_mutex);
6459         init_waitqueue_head(&conf->wait_for_stripe);
6460         init_waitqueue_head(&conf->wait_for_overlap);
6461         INIT_LIST_HEAD(&conf->handle_list);
6462         INIT_LIST_HEAD(&conf->hold_list);
6463         INIT_LIST_HEAD(&conf->delayed_list);
6464         INIT_LIST_HEAD(&conf->bitmap_list);
6465         init_llist_head(&conf->released_stripes);
6466         atomic_set(&conf->active_stripes, 0);
6467         atomic_set(&conf->preread_active_stripes, 0);
6468         atomic_set(&conf->active_aligned_reads, 0);
6469         conf->bypass_threshold = BYPASS_THRESHOLD;
6470         conf->recovery_disabled = mddev->recovery_disabled - 1;
6471
6472         conf->raid_disks = mddev->raid_disks;
6473         if (mddev->reshape_position == MaxSector)
6474                 conf->previous_raid_disks = mddev->raid_disks;
6475         else
6476                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6477         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6478
6479         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6480                               GFP_KERNEL);
6481         if (!conf->disks)
6482                 goto abort;
6483
6484         conf->mddev = mddev;
6485
6486         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6487                 goto abort;
6488
6489         /* We init hash_locks[0] separately to that it can be used
6490          * as the reference lock in the spin_lock_nest_lock() call
6491          * in lock_all_device_hash_locks_irq in order to convince
6492          * lockdep that we know what we are doing.
6493          */
6494         spin_lock_init(conf->hash_locks);
6495         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6496                 spin_lock_init(conf->hash_locks + i);
6497
6498         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6499                 INIT_LIST_HEAD(conf->inactive_list + i);
6500
6501         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6502                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
6503
6504         conf->level = mddev->new_level;
6505         conf->chunk_sectors = mddev->new_chunk_sectors;
6506         if (raid5_alloc_percpu(conf) != 0)
6507                 goto abort;
6508
6509         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6510
6511         rdev_for_each(rdev, mddev) {
6512                 raid_disk = rdev->raid_disk;
6513                 if (raid_disk >= max_disks
6514                     || raid_disk < 0)
6515                         continue;
6516                 disk = conf->disks + raid_disk;
6517
6518                 if (test_bit(Replacement, &rdev->flags)) {
6519                         if (disk->replacement)
6520                                 goto abort;
6521                         disk->replacement = rdev;
6522                 } else {
6523                         if (disk->rdev)
6524                                 goto abort;
6525                         disk->rdev = rdev;
6526                 }
6527
6528                 if (test_bit(In_sync, &rdev->flags)) {
6529                         char b[BDEVNAME_SIZE];
6530                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6531                                " disk %d\n",
6532                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6533                 } else if (rdev->saved_raid_disk != raid_disk)
6534                         /* Cannot rely on bitmap to complete recovery */
6535                         conf->fullsync = 1;
6536         }
6537
6538         conf->level = mddev->new_level;
6539         if (conf->level == 6) {
6540                 conf->max_degraded = 2;
6541                 if (raid6_call.xor_syndrome)
6542                         conf->rmw_level = PARITY_ENABLE_RMW;
6543                 else
6544                         conf->rmw_level = PARITY_DISABLE_RMW;
6545         } else {
6546                 conf->max_degraded = 1;
6547                 conf->rmw_level = PARITY_ENABLE_RMW;
6548         }
6549         conf->algorithm = mddev->new_layout;
6550         conf->reshape_progress = mddev->reshape_position;
6551         if (conf->reshape_progress != MaxSector) {
6552                 conf->prev_chunk_sectors = mddev->chunk_sectors;
6553                 conf->prev_algo = mddev->layout;
6554         }
6555
6556         conf->min_nr_stripes = NR_STRIPES;
6557         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6558                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6559         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6560         if (grow_stripes(conf, conf->min_nr_stripes)) {
6561                 printk(KERN_ERR
6562                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
6563                        mdname(mddev), memory);
6564                 goto abort;
6565         } else
6566                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6567                        mdname(mddev), memory);
6568         /*
6569          * Losing a stripe head costs more than the time to refill it,
6570          * it reduces the queue depth and so can hurt throughput.
6571          * So set it rather large, scaled by number of devices.
6572          */
6573         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6574         conf->shrinker.scan_objects = raid5_cache_scan;
6575         conf->shrinker.count_objects = raid5_cache_count;
6576         conf->shrinker.batch = 128;
6577         conf->shrinker.flags = 0;
6578         register_shrinker(&conf->shrinker);
6579
6580         sprintf(pers_name, "raid%d", mddev->new_level);
6581         conf->thread = md_register_thread(raid5d, mddev, pers_name);
6582         if (!conf->thread) {
6583                 printk(KERN_ERR
6584                        "md/raid:%s: couldn't allocate thread.\n",
6585                        mdname(mddev));
6586                 goto abort;
6587         }
6588
6589         return conf;
6590
6591  abort:
6592         if (conf) {
6593                 free_conf(conf);
6594                 return ERR_PTR(-EIO);
6595         } else
6596                 return ERR_PTR(-ENOMEM);
6597 }
6598
6599 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6600 {
6601         switch (algo) {
6602         case ALGORITHM_PARITY_0:
6603                 if (raid_disk < max_degraded)
6604                         return 1;
6605                 break;
6606         case ALGORITHM_PARITY_N:
6607                 if (raid_disk >= raid_disks - max_degraded)
6608                         return 1;
6609                 break;
6610         case ALGORITHM_PARITY_0_6:
6611                 if (raid_disk == 0 ||
6612                     raid_disk == raid_disks - 1)
6613                         return 1;
6614                 break;
6615         case ALGORITHM_LEFT_ASYMMETRIC_6:
6616         case ALGORITHM_RIGHT_ASYMMETRIC_6:
6617         case ALGORITHM_LEFT_SYMMETRIC_6:
6618         case ALGORITHM_RIGHT_SYMMETRIC_6:
6619                 if (raid_disk == raid_disks - 1)
6620                         return 1;
6621         }
6622         return 0;
6623 }
6624
6625 static int run(struct mddev *mddev)
6626 {
6627         struct r5conf *conf;
6628         int working_disks = 0;
6629         int dirty_parity_disks = 0;
6630         struct md_rdev *rdev;
6631         sector_t reshape_offset = 0;
6632         int i;
6633         long long min_offset_diff = 0;
6634         int first = 1;
6635
6636         if (mddev->recovery_cp != MaxSector)
6637                 printk(KERN_NOTICE "md/raid:%s: not clean"
6638                        " -- starting background reconstruction\n",
6639                        mdname(mddev));
6640
6641         rdev_for_each(rdev, mddev) {
6642                 long long diff;
6643                 if (rdev->raid_disk < 0)
6644                         continue;
6645                 diff = (rdev->new_data_offset - rdev->data_offset);
6646                 if (first) {
6647                         min_offset_diff = diff;
6648                         first = 0;
6649                 } else if (mddev->reshape_backwards &&
6650                          diff < min_offset_diff)
6651                         min_offset_diff = diff;
6652                 else if (!mddev->reshape_backwards &&
6653                          diff > min_offset_diff)
6654                         min_offset_diff = diff;
6655         }
6656
6657         if (mddev->reshape_position != MaxSector) {
6658                 /* Check that we can continue the reshape.
6659                  * Difficulties arise if the stripe we would write to
6660                  * next is at or after the stripe we would read from next.
6661                  * For a reshape that changes the number of devices, this
6662                  * is only possible for a very short time, and mdadm makes
6663                  * sure that time appears to have past before assembling
6664                  * the array.  So we fail if that time hasn't passed.
6665                  * For a reshape that keeps the number of devices the same
6666                  * mdadm must be monitoring the reshape can keeping the
6667                  * critical areas read-only and backed up.  It will start
6668                  * the array in read-only mode, so we check for that.
6669                  */
6670                 sector_t here_new, here_old;
6671                 int old_disks;
6672                 int max_degraded = (mddev->level == 6 ? 2 : 1);
6673
6674                 if (mddev->new_level != mddev->level) {
6675                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
6676                                "required - aborting.\n",
6677                                mdname(mddev));
6678                         return -EINVAL;
6679                 }
6680                 old_disks = mddev->raid_disks - mddev->delta_disks;
6681                 /* reshape_position must be on a new-stripe boundary, and one
6682                  * further up in new geometry must map after here in old
6683                  * geometry.
6684                  */
6685                 here_new = mddev->reshape_position;
6686                 if (sector_div(here_new, mddev->new_chunk_sectors *
6687                                (mddev->raid_disks - max_degraded))) {
6688                         printk(KERN_ERR "md/raid:%s: reshape_position not "
6689                                "on a stripe boundary\n", mdname(mddev));
6690                         return -EINVAL;
6691                 }
6692                 reshape_offset = here_new * mddev->new_chunk_sectors;
6693                 /* here_new is the stripe we will write to */
6694                 here_old = mddev->reshape_position;
6695                 sector_div(here_old, mddev->chunk_sectors *
6696                            (old_disks-max_degraded));
6697                 /* here_old is the first stripe that we might need to read
6698                  * from */
6699                 if (mddev->delta_disks == 0) {
6700                         if ((here_new * mddev->new_chunk_sectors !=
6701                              here_old * mddev->chunk_sectors)) {
6702                                 printk(KERN_ERR "md/raid:%s: reshape position is"
6703                                        " confused - aborting\n", mdname(mddev));
6704                                 return -EINVAL;
6705                         }
6706                         /* We cannot be sure it is safe to start an in-place
6707                          * reshape.  It is only safe if user-space is monitoring
6708                          * and taking constant backups.
6709                          * mdadm always starts a situation like this in
6710                          * readonly mode so it can take control before
6711                          * allowing any writes.  So just check for that.
6712                          */
6713                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6714                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
6715                                 /* not really in-place - so OK */;
6716                         else if (mddev->ro == 0) {
6717                                 printk(KERN_ERR "md/raid:%s: in-place reshape "
6718                                        "must be started in read-only mode "
6719                                        "- aborting\n",
6720                                        mdname(mddev));
6721                                 return -EINVAL;
6722                         }
6723                 } else if (mddev->reshape_backwards
6724                     ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6725                        here_old * mddev->chunk_sectors)
6726                     : (here_new * mddev->new_chunk_sectors >=
6727                        here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6728                         /* Reading from the same stripe as writing to - bad */
6729                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6730                                "auto-recovery - aborting.\n",
6731                                mdname(mddev));
6732                         return -EINVAL;
6733                 }
6734                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6735                        mdname(mddev));
6736                 /* OK, we should be able to continue; */
6737         } else {
6738                 BUG_ON(mddev->level != mddev->new_level);
6739                 BUG_ON(mddev->layout != mddev->new_layout);
6740                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6741                 BUG_ON(mddev->delta_disks != 0);
6742         }
6743
6744         if (mddev->private == NULL)
6745                 conf = setup_conf(mddev);
6746         else
6747                 conf = mddev->private;
6748
6749         if (IS_ERR(conf))
6750                 return PTR_ERR(conf);
6751
6752         conf->min_offset_diff = min_offset_diff;
6753         mddev->thread = conf->thread;
6754         conf->thread = NULL;
6755         mddev->private = conf;
6756
6757         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6758              i++) {
6759                 rdev = conf->disks[i].rdev;
6760                 if (!rdev && conf->disks[i].replacement) {
6761                         /* The replacement is all we have yet */
6762                         rdev = conf->disks[i].replacement;
6763                         conf->disks[i].replacement = NULL;
6764                         clear_bit(Replacement, &rdev->flags);
6765                         conf->disks[i].rdev = rdev;
6766                 }
6767                 if (!rdev)
6768                         continue;
6769                 if (conf->disks[i].replacement &&
6770                     conf->reshape_progress != MaxSector) {
6771                         /* replacements and reshape simply do not mix. */
6772                         printk(KERN_ERR "md: cannot handle concurrent "
6773                                "replacement and reshape.\n");
6774                         goto abort;
6775                 }
6776                 if (test_bit(In_sync, &rdev->flags)) {
6777                         working_disks++;
6778                         continue;
6779                 }
6780                 /* This disc is not fully in-sync.  However if it
6781                  * just stored parity (beyond the recovery_offset),
6782                  * when we don't need to be concerned about the
6783                  * array being dirty.
6784                  * When reshape goes 'backwards', we never have
6785                  * partially completed devices, so we only need
6786                  * to worry about reshape going forwards.
6787                  */
6788                 /* Hack because v0.91 doesn't store recovery_offset properly. */
6789                 if (mddev->major_version == 0 &&
6790                     mddev->minor_version > 90)
6791                         rdev->recovery_offset = reshape_offset;
6792
6793                 if (rdev->recovery_offset < reshape_offset) {
6794                         /* We need to check old and new layout */
6795                         if (!only_parity(rdev->raid_disk,
6796                                          conf->algorithm,
6797                                          conf->raid_disks,
6798                                          conf->max_degraded))
6799                                 continue;
6800                 }
6801                 if (!only_parity(rdev->raid_disk,
6802                                  conf->prev_algo,
6803                                  conf->previous_raid_disks,
6804                                  conf->max_degraded))
6805                         continue;
6806                 dirty_parity_disks++;
6807         }
6808
6809         /*
6810          * 0 for a fully functional array, 1 or 2 for a degraded array.
6811          */
6812         mddev->degraded = calc_degraded(conf);
6813
6814         if (has_failed(conf)) {
6815                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
6816                         " (%d/%d failed)\n",
6817                         mdname(mddev), mddev->degraded, conf->raid_disks);
6818                 goto abort;
6819         }
6820
6821         /* device size must be a multiple of chunk size */
6822         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6823         mddev->resync_max_sectors = mddev->dev_sectors;
6824
6825         if (mddev->degraded > dirty_parity_disks &&
6826             mddev->recovery_cp != MaxSector) {
6827                 if (mddev->ok_start_degraded)
6828                         printk(KERN_WARNING
6829                                "md/raid:%s: starting dirty degraded array"
6830                                " - data corruption possible.\n",
6831                                mdname(mddev));
6832                 else {
6833                         printk(KERN_ERR
6834                                "md/raid:%s: cannot start dirty degraded array.\n",
6835                                mdname(mddev));
6836                         goto abort;
6837                 }
6838         }
6839
6840         if (mddev->degraded == 0)
6841                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6842                        " devices, algorithm %d\n", mdname(mddev), conf->level,
6843                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6844                        mddev->new_layout);
6845         else
6846                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6847                        " out of %d devices, algorithm %d\n",
6848                        mdname(mddev), conf->level,
6849                        mddev->raid_disks - mddev->degraded,
6850                        mddev->raid_disks, mddev->new_layout);
6851
6852         print_raid5_conf(conf);
6853
6854         if (conf->reshape_progress != MaxSector) {
6855                 conf->reshape_safe = conf->reshape_progress;
6856                 atomic_set(&conf->reshape_stripes, 0);
6857                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6858                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6859                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6860                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6861                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6862                                                         "reshape");
6863         }
6864
6865         /* Ok, everything is just fine now */
6866         if (mddev->to_remove == &raid5_attrs_group)
6867                 mddev->to_remove = NULL;
6868         else if (mddev->kobj.sd &&
6869             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6870                 printk(KERN_WARNING
6871                        "raid5: failed to create sysfs attributes for %s\n",
6872                        mdname(mddev));
6873         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6874
6875         if (mddev->queue) {
6876                 int chunk_size;
6877                 bool discard_supported = true;
6878                 /* read-ahead size must cover two whole stripes, which
6879                  * is 2 * (datadisks) * chunksize where 'n' is the
6880                  * number of raid devices
6881                  */
6882                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6883                 int stripe = data_disks *
6884                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6885                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6886                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6887
6888                 chunk_size = mddev->chunk_sectors << 9;
6889                 blk_queue_io_min(mddev->queue, chunk_size);
6890                 blk_queue_io_opt(mddev->queue, chunk_size *
6891                                  (conf->raid_disks - conf->max_degraded));
6892                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
6893                 /*
6894                  * We can only discard a whole stripe. It doesn't make sense to
6895                  * discard data disk but write parity disk
6896                  */
6897                 stripe = stripe * PAGE_SIZE;
6898                 /* Round up to power of 2, as discard handling
6899                  * currently assumes that */
6900                 while ((stripe-1) & stripe)
6901                         stripe = (stripe | (stripe-1)) + 1;
6902                 mddev->queue->limits.discard_alignment = stripe;
6903                 mddev->queue->limits.discard_granularity = stripe;
6904                 /*
6905                  * unaligned part of discard request will be ignored, so can't
6906                  * guarantee discard_zeroes_data
6907                  */
6908                 mddev->queue->limits.discard_zeroes_data = 0;
6909
6910                 blk_queue_max_write_same_sectors(mddev->queue, 0);
6911
6912                 rdev_for_each(rdev, mddev) {
6913                         disk_stack_limits(mddev->gendisk, rdev->bdev,
6914                                           rdev->data_offset << 9);
6915                         disk_stack_limits(mddev->gendisk, rdev->bdev,
6916                                           rdev->new_data_offset << 9);
6917                         /*
6918                          * discard_zeroes_data is required, otherwise data
6919                          * could be lost. Consider a scenario: discard a stripe
6920                          * (the stripe could be inconsistent if
6921                          * discard_zeroes_data is 0); write one disk of the
6922                          * stripe (the stripe could be inconsistent again
6923                          * depending on which disks are used to calculate
6924                          * parity); the disk is broken; The stripe data of this
6925                          * disk is lost.
6926                          */
6927                         if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6928                             !bdev_get_queue(rdev->bdev)->
6929                                                 limits.discard_zeroes_data)
6930                                 discard_supported = false;
6931                         /* Unfortunately, discard_zeroes_data is not currently
6932                          * a guarantee - just a hint.  So we only allow DISCARD
6933                          * if the sysadmin has confirmed that only safe devices
6934                          * are in use by setting a module parameter.
6935                          */
6936                         if (!devices_handle_discard_safely) {
6937                                 if (discard_supported) {
6938                                         pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6939                                         pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6940                                 }
6941                                 discard_supported = false;
6942                         }
6943                 }
6944
6945                 if (discard_supported &&
6946                    mddev->queue->limits.max_discard_sectors >= stripe &&
6947                    mddev->queue->limits.discard_granularity >= stripe)
6948                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6949                                                 mddev->queue);
6950                 else
6951                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6952                                                 mddev->queue);
6953         }
6954
6955         return 0;
6956 abort:
6957         md_unregister_thread(&mddev->thread);
6958         print_raid5_conf(conf);
6959         free_conf(conf);
6960         mddev->private = NULL;
6961         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6962         return -EIO;
6963 }
6964
6965 static void raid5_free(struct mddev *mddev, void *priv)
6966 {
6967         struct r5conf *conf = priv;
6968
6969         free_conf(conf);
6970         mddev->to_remove = &raid5_attrs_group;
6971 }
6972
6973 static void status(struct seq_file *seq, struct mddev *mddev)
6974 {
6975         struct r5conf *conf = mddev->private;
6976         int i;
6977
6978         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6979                 mddev->chunk_sectors / 2, mddev->layout);
6980         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6981         for (i = 0; i < conf->raid_disks; i++)
6982                 seq_printf (seq, "%s",
6983                                conf->disks[i].rdev &&
6984                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6985         seq_printf (seq, "]");
6986 }
6987
6988 static void print_raid5_conf (struct r5conf *conf)
6989 {
6990         int i;
6991         struct disk_info *tmp;
6992
6993         printk(KERN_DEBUG "RAID conf printout:\n");
6994         if (!conf) {
6995                 printk("(conf==NULL)\n");
6996                 return;
6997         }
6998         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6999                conf->raid_disks,
7000                conf->raid_disks - conf->mddev->degraded);
7001
7002         for (i = 0; i < conf->raid_disks; i++) {
7003                 char b[BDEVNAME_SIZE];
7004                 tmp = conf->disks + i;
7005                 if (tmp->rdev)
7006                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
7007                                i, !test_bit(Faulty, &tmp->rdev->flags),
7008                                bdevname(tmp->rdev->bdev, b));
7009         }
7010 }
7011
7012 static int raid5_spare_active(struct mddev *mddev)
7013 {
7014         int i;
7015         struct r5conf *conf = mddev->private;
7016         struct disk_info *tmp;
7017         int count = 0;
7018         unsigned long flags;
7019
7020         for (i = 0; i < conf->raid_disks; i++) {
7021                 tmp = conf->disks + i;
7022                 if (tmp->replacement
7023                     && tmp->replacement->recovery_offset == MaxSector
7024                     && !test_bit(Faulty, &tmp->replacement->flags)
7025                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7026                         /* Replacement has just become active. */
7027                         if (!tmp->rdev
7028                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7029                                 count++;
7030                         if (tmp->rdev) {
7031                                 /* Replaced device not technically faulty,
7032                                  * but we need to be sure it gets removed
7033                                  * and never re-added.
7034                                  */
7035                                 set_bit(Faulty, &tmp->rdev->flags);
7036                                 sysfs_notify_dirent_safe(
7037                                         tmp->rdev->sysfs_state);
7038                         }
7039                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7040                 } else if (tmp->rdev
7041                     && tmp->rdev->recovery_offset == MaxSector
7042                     && !test_bit(Faulty, &tmp->rdev->flags)
7043                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7044                         count++;
7045                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7046                 }
7047         }
7048         spin_lock_irqsave(&conf->device_lock, flags);
7049         mddev->degraded = calc_degraded(conf);
7050         spin_unlock_irqrestore(&conf->device_lock, flags);
7051         print_raid5_conf(conf);
7052         return count;
7053 }
7054
7055 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7056 {
7057         struct r5conf *conf = mddev->private;
7058         int err = 0;
7059         int number = rdev->raid_disk;
7060         struct md_rdev **rdevp;
7061         struct disk_info *p = conf->disks + number;
7062
7063         print_raid5_conf(conf);
7064         if (rdev == p->rdev)
7065                 rdevp = &p->rdev;
7066         else if (rdev == p->replacement)
7067                 rdevp = &p->replacement;
7068         else
7069                 return 0;
7070
7071         if (number >= conf->raid_disks &&
7072             conf->reshape_progress == MaxSector)
7073                 clear_bit(In_sync, &rdev->flags);
7074
7075         if (test_bit(In_sync, &rdev->flags) ||
7076             atomic_read(&rdev->nr_pending)) {
7077                 err = -EBUSY;
7078                 goto abort;
7079         }
7080         /* Only remove non-faulty devices if recovery
7081          * isn't possible.
7082          */
7083         if (!test_bit(Faulty, &rdev->flags) &&
7084             mddev->recovery_disabled != conf->recovery_disabled &&
7085             !has_failed(conf) &&
7086             (!p->replacement || p->replacement == rdev) &&
7087             number < conf->raid_disks) {
7088                 err = -EBUSY;
7089                 goto abort;
7090         }
7091         *rdevp = NULL;
7092         synchronize_rcu();
7093         if (atomic_read(&rdev->nr_pending)) {
7094                 /* lost the race, try later */
7095                 err = -EBUSY;
7096                 *rdevp = rdev;
7097         } else if (p->replacement) {
7098                 /* We must have just cleared 'rdev' */
7099                 p->rdev = p->replacement;
7100                 clear_bit(Replacement, &p->replacement->flags);
7101                 smp_mb(); /* Make sure other CPUs may see both as identical
7102                            * but will never see neither - if they are careful
7103                            */
7104                 p->replacement = NULL;
7105                 clear_bit(WantReplacement, &rdev->flags);
7106         } else
7107                 /* We might have just removed the Replacement as faulty-
7108                  * clear the bit just in case
7109                  */
7110                 clear_bit(WantReplacement, &rdev->flags);
7111 abort:
7112
7113         print_raid5_conf(conf);
7114         return err;
7115 }
7116
7117 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7118 {
7119         struct r5conf *conf = mddev->private;
7120         int err = -EEXIST;
7121         int disk;
7122         struct disk_info *p;
7123         int first = 0;
7124         int last = conf->raid_disks - 1;
7125
7126         if (mddev->recovery_disabled == conf->recovery_disabled)
7127                 return -EBUSY;
7128
7129         if (rdev->saved_raid_disk < 0 && has_failed(conf))
7130                 /* no point adding a device */
7131                 return -EINVAL;
7132
7133         if (rdev->raid_disk >= 0)
7134                 first = last = rdev->raid_disk;
7135
7136         /*
7137          * find the disk ... but prefer rdev->saved_raid_disk
7138          * if possible.
7139          */
7140         if (rdev->saved_raid_disk >= 0 &&
7141             rdev->saved_raid_disk >= first &&
7142             conf->disks[rdev->saved_raid_disk].rdev == NULL)
7143                 first = rdev->saved_raid_disk;
7144
7145         for (disk = first; disk <= last; disk++) {
7146                 p = conf->disks + disk;
7147                 if (p->rdev == NULL) {
7148                         clear_bit(In_sync, &rdev->flags);
7149                         rdev->raid_disk = disk;
7150                         err = 0;
7151                         if (rdev->saved_raid_disk != disk)
7152                                 conf->fullsync = 1;
7153                         rcu_assign_pointer(p->rdev, rdev);
7154                         goto out;
7155                 }
7156         }
7157         for (disk = first; disk <= last; disk++) {
7158                 p = conf->disks + disk;
7159                 if (test_bit(WantReplacement, &p->rdev->flags) &&
7160                     p->replacement == NULL) {
7161                         clear_bit(In_sync, &rdev->flags);
7162                         set_bit(Replacement, &rdev->flags);
7163                         rdev->raid_disk = disk;
7164                         err = 0;
7165                         conf->fullsync = 1;
7166                         rcu_assign_pointer(p->replacement, rdev);
7167                         break;
7168                 }
7169         }
7170 out:
7171         print_raid5_conf(conf);
7172         return err;
7173 }
7174
7175 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7176 {
7177         /* no resync is happening, and there is enough space
7178          * on all devices, so we can resize.
7179          * We need to make sure resync covers any new space.
7180          * If the array is shrinking we should possibly wait until
7181          * any io in the removed space completes, but it hardly seems
7182          * worth it.
7183          */
7184         sector_t newsize;
7185         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7186         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7187         if (mddev->external_size &&
7188             mddev->array_sectors > newsize)
7189                 return -EINVAL;
7190         if (mddev->bitmap) {
7191                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7192                 if (ret)
7193                         return ret;
7194         }
7195         md_set_array_sectors(mddev, newsize);
7196         set_capacity(mddev->gendisk, mddev->array_sectors);
7197         revalidate_disk(mddev->gendisk);
7198         if (sectors > mddev->dev_sectors &&
7199             mddev->recovery_cp > mddev->dev_sectors) {
7200                 mddev->recovery_cp = mddev->dev_sectors;
7201                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7202         }
7203         mddev->dev_sectors = sectors;
7204         mddev->resync_max_sectors = sectors;
7205         return 0;
7206 }
7207
7208 static int check_stripe_cache(struct mddev *mddev)
7209 {
7210         /* Can only proceed if there are plenty of stripe_heads.
7211          * We need a minimum of one full stripe,, and for sensible progress
7212          * it is best to have about 4 times that.
7213          * If we require 4 times, then the default 256 4K stripe_heads will
7214          * allow for chunk sizes up to 256K, which is probably OK.
7215          * If the chunk size is greater, user-space should request more
7216          * stripe_heads first.
7217          */
7218         struct r5conf *conf = mddev->private;
7219         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7220             > conf->min_nr_stripes ||
7221             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7222             > conf->min_nr_stripes) {
7223                 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7224                        mdname(mddev),
7225                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7226                         / STRIPE_SIZE)*4);
7227                 return 0;
7228         }
7229         return 1;
7230 }
7231
7232 static int check_reshape(struct mddev *mddev)
7233 {
7234         struct r5conf *conf = mddev->private;
7235
7236         if (mddev->delta_disks == 0 &&
7237             mddev->new_layout == mddev->layout &&
7238             mddev->new_chunk_sectors == mddev->chunk_sectors)
7239                 return 0; /* nothing to do */
7240         if (has_failed(conf))
7241                 return -EINVAL;
7242         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7243                 /* We might be able to shrink, but the devices must
7244                  * be made bigger first.
7245                  * For raid6, 4 is the minimum size.
7246                  * Otherwise 2 is the minimum
7247                  */
7248                 int min = 2;
7249                 if (mddev->level == 6)
7250                         min = 4;
7251                 if (mddev->raid_disks + mddev->delta_disks < min)
7252                         return -EINVAL;
7253         }
7254
7255         if (!check_stripe_cache(mddev))
7256                 return -ENOSPC;
7257
7258         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7259             mddev->delta_disks > 0)
7260                 if (resize_chunks(conf,
7261                                   conf->previous_raid_disks
7262                                   + max(0, mddev->delta_disks),
7263                                   max(mddev->new_chunk_sectors,
7264                                       mddev->chunk_sectors)
7265                             ) < 0)
7266                         return -ENOMEM;
7267         return resize_stripes(conf, (conf->previous_raid_disks
7268                                      + mddev->delta_disks));
7269 }
7270
7271 static int raid5_start_reshape(struct mddev *mddev)
7272 {
7273         struct r5conf *conf = mddev->private;
7274         struct md_rdev *rdev;
7275         int spares = 0;
7276         unsigned long flags;
7277
7278         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7279                 return -EBUSY;
7280
7281         if (!check_stripe_cache(mddev))
7282                 return -ENOSPC;
7283
7284         if (has_failed(conf))
7285                 return -EINVAL;
7286
7287         rdev_for_each(rdev, mddev) {
7288                 if (!test_bit(In_sync, &rdev->flags)
7289                     && !test_bit(Faulty, &rdev->flags))
7290                         spares++;
7291         }
7292
7293         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7294                 /* Not enough devices even to make a degraded array
7295                  * of that size
7296                  */
7297                 return -EINVAL;
7298
7299         /* Refuse to reduce size of the array.  Any reductions in
7300          * array size must be through explicit setting of array_size
7301          * attribute.
7302          */
7303         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7304             < mddev->array_sectors) {
7305                 printk(KERN_ERR "md/raid:%s: array size must be reduced "
7306                        "before number of disks\n", mdname(mddev));
7307                 return -EINVAL;
7308         }
7309
7310         atomic_set(&conf->reshape_stripes, 0);
7311         spin_lock_irq(&conf->device_lock);
7312         write_seqcount_begin(&conf->gen_lock);
7313         conf->previous_raid_disks = conf->raid_disks;
7314         conf->raid_disks += mddev->delta_disks;
7315         conf->prev_chunk_sectors = conf->chunk_sectors;
7316         conf->chunk_sectors = mddev->new_chunk_sectors;
7317         conf->prev_algo = conf->algorithm;
7318         conf->algorithm = mddev->new_layout;
7319         conf->generation++;
7320         /* Code that selects data_offset needs to see the generation update
7321          * if reshape_progress has been set - so a memory barrier needed.
7322          */
7323         smp_mb();
7324         if (mddev->reshape_backwards)
7325                 conf->reshape_progress = raid5_size(mddev, 0, 0);
7326         else
7327                 conf->reshape_progress = 0;
7328         conf->reshape_safe = conf->reshape_progress;
7329         write_seqcount_end(&conf->gen_lock);
7330         spin_unlock_irq(&conf->device_lock);
7331
7332         /* Now make sure any requests that proceeded on the assumption
7333          * the reshape wasn't running - like Discard or Read - have
7334          * completed.
7335          */
7336         mddev_suspend(mddev);
7337         mddev_resume(mddev);
7338
7339         /* Add some new drives, as many as will fit.
7340          * We know there are enough to make the newly sized array work.
7341          * Don't add devices if we are reducing the number of
7342          * devices in the array.  This is because it is not possible
7343          * to correctly record the "partially reconstructed" state of
7344          * such devices during the reshape and confusion could result.
7345          */
7346         if (mddev->delta_disks >= 0) {
7347                 rdev_for_each(rdev, mddev)
7348                         if (rdev->raid_disk < 0 &&
7349                             !test_bit(Faulty, &rdev->flags)) {
7350                                 if (raid5_add_disk(mddev, rdev) == 0) {
7351                                         if (rdev->raid_disk
7352                                             >= conf->previous_raid_disks)
7353                                                 set_bit(In_sync, &rdev->flags);
7354                                         else
7355                                                 rdev->recovery_offset = 0;
7356
7357                                         if (sysfs_link_rdev(mddev, rdev))
7358                                                 /* Failure here is OK */;
7359                                 }
7360                         } else if (rdev->raid_disk >= conf->previous_raid_disks
7361                                    && !test_bit(Faulty, &rdev->flags)) {
7362                                 /* This is a spare that was manually added */
7363                                 set_bit(In_sync, &rdev->flags);
7364                         }
7365
7366                 /* When a reshape changes the number of devices,
7367                  * ->degraded is measured against the larger of the
7368                  * pre and post number of devices.
7369                  */
7370                 spin_lock_irqsave(&conf->device_lock, flags);
7371                 mddev->degraded = calc_degraded(conf);
7372                 spin_unlock_irqrestore(&conf->device_lock, flags);
7373         }
7374         mddev->raid_disks = conf->raid_disks;
7375         mddev->reshape_position = conf->reshape_progress;
7376         set_bit(MD_CHANGE_DEVS, &mddev->flags);
7377
7378         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7379         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7380         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7381         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7382         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7383         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7384                                                 "reshape");
7385         if (!mddev->sync_thread) {
7386                 mddev->recovery = 0;
7387                 spin_lock_irq(&conf->device_lock);
7388                 write_seqcount_begin(&conf->gen_lock);
7389                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7390                 mddev->new_chunk_sectors =
7391                         conf->chunk_sectors = conf->prev_chunk_sectors;
7392                 mddev->new_layout = conf->algorithm = conf->prev_algo;
7393                 rdev_for_each(rdev, mddev)
7394                         rdev->new_data_offset = rdev->data_offset;
7395                 smp_wmb();
7396                 conf->generation --;
7397                 conf->reshape_progress = MaxSector;
7398                 mddev->reshape_position = MaxSector;
7399                 write_seqcount_end(&conf->gen_lock);
7400                 spin_unlock_irq(&conf->device_lock);
7401                 return -EAGAIN;
7402         }
7403         conf->reshape_checkpoint = jiffies;
7404         md_wakeup_thread(mddev->sync_thread);
7405         md_new_event(mddev);
7406         return 0;
7407 }
7408
7409 /* This is called from the reshape thread and should make any
7410  * changes needed in 'conf'
7411  */
7412 static void end_reshape(struct r5conf *conf)
7413 {
7414
7415         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7416                 struct md_rdev *rdev;
7417
7418                 spin_lock_irq(&conf->device_lock);
7419                 conf->previous_raid_disks = conf->raid_disks;
7420                 rdev_for_each(rdev, conf->mddev)
7421                         rdev->data_offset = rdev->new_data_offset;
7422                 smp_wmb();
7423                 conf->reshape_progress = MaxSector;
7424                 spin_unlock_irq(&conf->device_lock);
7425                 wake_up(&conf->wait_for_overlap);
7426
7427                 /* read-ahead size must cover two whole stripes, which is
7428                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7429                  */
7430                 if (conf->mddev->queue) {
7431                         int data_disks = conf->raid_disks - conf->max_degraded;
7432                         int stripe = data_disks * ((conf->chunk_sectors << 9)
7433                                                    / PAGE_SIZE);
7434                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7435                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7436                 }
7437         }
7438 }
7439
7440 /* This is called from the raid5d thread with mddev_lock held.
7441  * It makes config changes to the device.
7442  */
7443 static void raid5_finish_reshape(struct mddev *mddev)
7444 {
7445         struct r5conf *conf = mddev->private;
7446
7447         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7448
7449                 if (mddev->delta_disks > 0) {
7450                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7451                         set_capacity(mddev->gendisk, mddev->array_sectors);
7452                         revalidate_disk(mddev->gendisk);
7453                 } else {
7454                         int d;
7455                         spin_lock_irq(&conf->device_lock);
7456                         mddev->degraded = calc_degraded(conf);
7457                         spin_unlock_irq(&conf->device_lock);
7458                         for (d = conf->raid_disks ;
7459                              d < conf->raid_disks - mddev->delta_disks;
7460                              d++) {
7461                                 struct md_rdev *rdev = conf->disks[d].rdev;
7462                                 if (rdev)
7463                                         clear_bit(In_sync, &rdev->flags);
7464                                 rdev = conf->disks[d].replacement;
7465                                 if (rdev)
7466                                         clear_bit(In_sync, &rdev->flags);
7467                         }
7468                 }
7469                 mddev->layout = conf->algorithm;
7470                 mddev->chunk_sectors = conf->chunk_sectors;
7471                 mddev->reshape_position = MaxSector;
7472                 mddev->delta_disks = 0;
7473                 mddev->reshape_backwards = 0;
7474         }
7475 }
7476
7477 static void raid5_quiesce(struct mddev *mddev, int state)
7478 {
7479         struct r5conf *conf = mddev->private;
7480
7481         switch(state) {
7482         case 2: /* resume for a suspend */
7483                 wake_up(&conf->wait_for_overlap);
7484                 break;
7485
7486         case 1: /* stop all writes */
7487                 lock_all_device_hash_locks_irq(conf);
7488                 /* '2' tells resync/reshape to pause so that all
7489                  * active stripes can drain
7490                  */
7491                 conf->quiesce = 2;
7492                 wait_event_cmd(conf->wait_for_stripe,
7493                                     atomic_read(&conf->active_stripes) == 0 &&
7494                                     atomic_read(&conf->active_aligned_reads) == 0,
7495                                     unlock_all_device_hash_locks_irq(conf),
7496                                     lock_all_device_hash_locks_irq(conf));
7497                 conf->quiesce = 1;
7498                 unlock_all_device_hash_locks_irq(conf);
7499                 /* allow reshape to continue */
7500                 wake_up(&conf->wait_for_overlap);
7501                 break;
7502
7503         case 0: /* re-enable writes */
7504                 lock_all_device_hash_locks_irq(conf);
7505                 conf->quiesce = 0;
7506                 wake_up(&conf->wait_for_stripe);
7507                 wake_up(&conf->wait_for_overlap);
7508                 unlock_all_device_hash_locks_irq(conf);
7509                 break;
7510         }
7511 }
7512
7513 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7514 {
7515         struct r0conf *raid0_conf = mddev->private;
7516         sector_t sectors;
7517
7518         /* for raid0 takeover only one zone is supported */
7519         if (raid0_conf->nr_strip_zones > 1) {
7520                 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7521                        mdname(mddev));
7522                 return ERR_PTR(-EINVAL);
7523         }
7524
7525         sectors = raid0_conf->strip_zone[0].zone_end;
7526         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7527         mddev->dev_sectors = sectors;
7528         mddev->new_level = level;
7529         mddev->new_layout = ALGORITHM_PARITY_N;
7530         mddev->new_chunk_sectors = mddev->chunk_sectors;
7531         mddev->raid_disks += 1;
7532         mddev->delta_disks = 1;
7533         /* make sure it will be not marked as dirty */
7534         mddev->recovery_cp = MaxSector;
7535
7536         return setup_conf(mddev);
7537 }
7538
7539 static void *raid5_takeover_raid1(struct mddev *mddev)
7540 {
7541         int chunksect;
7542
7543         if (mddev->raid_disks != 2 ||
7544             mddev->degraded > 1)
7545                 return ERR_PTR(-EINVAL);
7546
7547         /* Should check if there are write-behind devices? */
7548
7549         chunksect = 64*2; /* 64K by default */
7550
7551         /* The array must be an exact multiple of chunksize */
7552         while (chunksect && (mddev->array_sectors & (chunksect-1)))
7553                 chunksect >>= 1;
7554
7555         if ((chunksect<<9) < STRIPE_SIZE)
7556                 /* array size does not allow a suitable chunk size */
7557                 return ERR_PTR(-EINVAL);
7558
7559         mddev->new_level = 5;
7560         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7561         mddev->new_chunk_sectors = chunksect;
7562
7563         return setup_conf(mddev);
7564 }
7565
7566 static void *raid5_takeover_raid6(struct mddev *mddev)
7567 {
7568         int new_layout;
7569
7570         switch (mddev->layout) {
7571         case ALGORITHM_LEFT_ASYMMETRIC_6:
7572                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7573                 break;
7574         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7575                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7576                 break;
7577         case ALGORITHM_LEFT_SYMMETRIC_6:
7578                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
7579                 break;
7580         case ALGORITHM_RIGHT_SYMMETRIC_6:
7581                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7582                 break;
7583         case ALGORITHM_PARITY_0_6:
7584                 new_layout = ALGORITHM_PARITY_0;
7585                 break;
7586         case ALGORITHM_PARITY_N:
7587                 new_layout = ALGORITHM_PARITY_N;
7588                 break;
7589         default:
7590                 return ERR_PTR(-EINVAL);
7591         }
7592         mddev->new_level = 5;
7593         mddev->new_layout = new_layout;
7594         mddev->delta_disks = -1;
7595         mddev->raid_disks -= 1;
7596         return setup_conf(mddev);
7597 }
7598
7599 static int raid5_check_reshape(struct mddev *mddev)
7600 {
7601         /* For a 2-drive array, the layout and chunk size can be changed
7602          * immediately as not restriping is needed.
7603          * For larger arrays we record the new value - after validation
7604          * to be used by a reshape pass.
7605          */
7606         struct r5conf *conf = mddev->private;
7607         int new_chunk = mddev->new_chunk_sectors;
7608
7609         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7610                 return -EINVAL;
7611         if (new_chunk > 0) {
7612                 if (!is_power_of_2(new_chunk))
7613                         return -EINVAL;
7614                 if (new_chunk < (PAGE_SIZE>>9))
7615                         return -EINVAL;
7616                 if (mddev->array_sectors & (new_chunk-1))
7617                         /* not factor of array size */
7618                         return -EINVAL;
7619         }
7620
7621         /* They look valid */
7622
7623         if (mddev->raid_disks == 2) {
7624                 /* can make the change immediately */
7625                 if (mddev->new_layout >= 0) {
7626                         conf->algorithm = mddev->new_layout;
7627                         mddev->layout = mddev->new_layout;
7628                 }
7629                 if (new_chunk > 0) {
7630                         conf->chunk_sectors = new_chunk ;
7631                         mddev->chunk_sectors = new_chunk;
7632                 }
7633                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
7634                 md_wakeup_thread(mddev->thread);
7635         }
7636         return check_reshape(mddev);
7637 }
7638
7639 static int raid6_check_reshape(struct mddev *mddev)
7640 {
7641         int new_chunk = mddev->new_chunk_sectors;
7642
7643         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7644                 return -EINVAL;
7645         if (new_chunk > 0) {
7646                 if (!is_power_of_2(new_chunk))
7647                         return -EINVAL;
7648                 if (new_chunk < (PAGE_SIZE >> 9))
7649                         return -EINVAL;
7650                 if (mddev->array_sectors & (new_chunk-1))
7651                         /* not factor of array size */
7652                         return -EINVAL;
7653         }
7654
7655         /* They look valid */
7656         return check_reshape(mddev);
7657 }
7658
7659 static void *raid5_takeover(struct mddev *mddev)
7660 {
7661         /* raid5 can take over:
7662          *  raid0 - if there is only one strip zone - make it a raid4 layout
7663          *  raid1 - if there are two drives.  We need to know the chunk size
7664          *  raid4 - trivial - just use a raid4 layout.
7665          *  raid6 - Providing it is a *_6 layout
7666          */
7667         if (mddev->level == 0)
7668                 return raid45_takeover_raid0(mddev, 5);
7669         if (mddev->level == 1)
7670                 return raid5_takeover_raid1(mddev);
7671         if (mddev->level == 4) {
7672                 mddev->new_layout = ALGORITHM_PARITY_N;
7673                 mddev->new_level = 5;
7674                 return setup_conf(mddev);
7675         }
7676         if (mddev->level == 6)
7677                 return raid5_takeover_raid6(mddev);
7678
7679         return ERR_PTR(-EINVAL);
7680 }
7681
7682 static void *raid4_takeover(struct mddev *mddev)
7683 {
7684         /* raid4 can take over:
7685          *  raid0 - if there is only one strip zone
7686          *  raid5 - if layout is right
7687          */
7688         if (mddev->level == 0)
7689                 return raid45_takeover_raid0(mddev, 4);
7690         if (mddev->level == 5 &&
7691             mddev->layout == ALGORITHM_PARITY_N) {
7692                 mddev->new_layout = 0;
7693                 mddev->new_level = 4;
7694                 return setup_conf(mddev);
7695         }
7696         return ERR_PTR(-EINVAL);
7697 }
7698
7699 static struct md_personality raid5_personality;
7700
7701 static void *raid6_takeover(struct mddev *mddev)
7702 {
7703         /* Currently can only take over a raid5.  We map the
7704          * personality to an equivalent raid6 personality
7705          * with the Q block at the end.
7706          */
7707         int new_layout;
7708
7709         if (mddev->pers != &raid5_personality)
7710                 return ERR_PTR(-EINVAL);
7711         if (mddev->degraded > 1)
7712                 return ERR_PTR(-EINVAL);
7713         if (mddev->raid_disks > 253)
7714                 return ERR_PTR(-EINVAL);
7715         if (mddev->raid_disks < 3)
7716                 return ERR_PTR(-EINVAL);
7717
7718         switch (mddev->layout) {
7719         case ALGORITHM_LEFT_ASYMMETRIC:
7720                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7721                 break;
7722         case ALGORITHM_RIGHT_ASYMMETRIC:
7723                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7724                 break;
7725         case ALGORITHM_LEFT_SYMMETRIC:
7726                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7727                 break;
7728         case ALGORITHM_RIGHT_SYMMETRIC:
7729                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7730                 break;
7731         case ALGORITHM_PARITY_0:
7732                 new_layout = ALGORITHM_PARITY_0_6;
7733                 break;
7734         case ALGORITHM_PARITY_N:
7735                 new_layout = ALGORITHM_PARITY_N;
7736                 break;
7737         default:
7738                 return ERR_PTR(-EINVAL);
7739         }
7740         mddev->new_level = 6;
7741         mddev->new_layout = new_layout;
7742         mddev->delta_disks = 1;
7743         mddev->raid_disks += 1;
7744         return setup_conf(mddev);
7745 }
7746
7747 static struct md_personality raid6_personality =
7748 {
7749         .name           = "raid6",
7750         .level          = 6,
7751         .owner          = THIS_MODULE,
7752         .make_request   = make_request,
7753         .run            = run,
7754         .free           = raid5_free,
7755         .status         = status,
7756         .error_handler  = error,
7757         .hot_add_disk   = raid5_add_disk,
7758         .hot_remove_disk= raid5_remove_disk,
7759         .spare_active   = raid5_spare_active,
7760         .sync_request   = sync_request,
7761         .resize         = raid5_resize,
7762         .size           = raid5_size,
7763         .check_reshape  = raid6_check_reshape,
7764         .start_reshape  = raid5_start_reshape,
7765         .finish_reshape = raid5_finish_reshape,
7766         .quiesce        = raid5_quiesce,
7767         .takeover       = raid6_takeover,
7768         .congested      = raid5_congested,
7769         .mergeable_bvec = raid5_mergeable_bvec,
7770 };
7771 static struct md_personality raid5_personality =
7772 {
7773         .name           = "raid5",
7774         .level          = 5,
7775         .owner          = THIS_MODULE,
7776         .make_request   = make_request,
7777         .run            = run,
7778         .free           = raid5_free,
7779         .status         = status,
7780         .error_handler  = error,
7781         .hot_add_disk   = raid5_add_disk,
7782         .hot_remove_disk= raid5_remove_disk,
7783         .spare_active   = raid5_spare_active,
7784         .sync_request   = sync_request,
7785         .resize         = raid5_resize,
7786         .size           = raid5_size,
7787         .check_reshape  = raid5_check_reshape,
7788         .start_reshape  = raid5_start_reshape,
7789         .finish_reshape = raid5_finish_reshape,
7790         .quiesce        = raid5_quiesce,
7791         .takeover       = raid5_takeover,
7792         .congested      = raid5_congested,
7793         .mergeable_bvec = raid5_mergeable_bvec,
7794 };
7795
7796 static struct md_personality raid4_personality =
7797 {
7798         .name           = "raid4",
7799         .level          = 4,
7800         .owner          = THIS_MODULE,
7801         .make_request   = make_request,
7802         .run            = run,
7803         .free           = raid5_free,
7804         .status         = status,
7805         .error_handler  = error,
7806         .hot_add_disk   = raid5_add_disk,
7807         .hot_remove_disk= raid5_remove_disk,
7808         .spare_active   = raid5_spare_active,
7809         .sync_request   = sync_request,
7810         .resize         = raid5_resize,
7811         .size           = raid5_size,
7812         .check_reshape  = raid5_check_reshape,
7813         .start_reshape  = raid5_start_reshape,
7814         .finish_reshape = raid5_finish_reshape,
7815         .quiesce        = raid5_quiesce,
7816         .takeover       = raid4_takeover,
7817         .congested      = raid5_congested,
7818         .mergeable_bvec = raid5_mergeable_bvec,
7819 };
7820
7821 static int __init raid5_init(void)
7822 {
7823         raid5_wq = alloc_workqueue("raid5wq",
7824                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7825         if (!raid5_wq)
7826                 return -ENOMEM;
7827         register_md_personality(&raid6_personality);
7828         register_md_personality(&raid5_personality);
7829         register_md_personality(&raid4_personality);
7830         return 0;
7831 }
7832
7833 static void raid5_exit(void)
7834 {
7835         unregister_md_personality(&raid6_personality);
7836         unregister_md_personality(&raid5_personality);
7837         unregister_md_personality(&raid4_personality);
7838         destroy_workqueue(raid5_wq);
7839 }
7840
7841 module_init(raid5_init);
7842 module_exit(raid5_exit);
7843 MODULE_LICENSE("GPL");
7844 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7845 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7846 MODULE_ALIAS("md-raid5");
7847 MODULE_ALIAS("md-raid4");
7848 MODULE_ALIAS("md-level-5");
7849 MODULE_ALIAS("md-level-4");
7850 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7851 MODULE_ALIAS("md-raid6");
7852 MODULE_ALIAS("md-level-6");
7853
7854 /* This used to be two separate modules, they were: */
7855 MODULE_ALIAS("raid5");
7856 MODULE_ALIAS("raid6");