drivers/md/raid1.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * raid1.c : Multiple Devices driver for Linux
   4  *
   5  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   6  *
   7  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   8  *
   9  * RAID-1 management functions.
  10  *
  11  * Better read-balancing code written by Mika Kuoppala <[email protected]>, 2000
  12  *
  13  * Fixes to reconstruction by Jakob Østergaard" <[email protected]>
  14  * Various fixes by Neil Brown <[email protected]>
  15  *
  16  * Changes by Peter T. Breuer <[email protected]> 31/1/2003 to support
  17  * bitmapped intelligence in resync:
  18  *
  19  *      - bitmap marked during normal i/o
  20  *      - bitmap used to skip nondirty blocks during sync
  21  *
  22  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  23  * - persistent bitmap code
  24  */
  25
  26 #include <linux/slab.h>
  27 #include <linux/delay.h>
  28 #include <linux/blkdev.h>
  29 #include <linux/module.h>
  30 #include <linux/seq_file.h>
  31 #include <linux/ratelimit.h>
  32 #include <linux/interval_tree_generic.h>
  33
  34 #include <trace/events/block.h>
  35
  36 #include "md.h"
  37 #include "raid1.h"
  38 #include "md-bitmap.h"
  39
  40 #define UNSUPPORTED_MDDEV_FLAGS         \
  41         ((1L << MD_HAS_JOURNAL) |       \
  42          (1L << MD_JOURNAL_CLEAN) |     \
  43          (1L << MD_HAS_PPL) |           \
  44          (1L << MD_HAS_MULTIPLE_PPLS))
  45
  46 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
  47 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
  48
  49 #define RAID_1_10_NAME "raid1"
  50 #include "raid1-10.c"
  51
  52 #define START(node) ((node)->start)
  53 #define LAST(node) ((node)->last)
  54 INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
  55                      START, LAST, static inline, raid1_rb);
  56
  57 static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
  58                                 struct serial_info *si, int idx)
  59 {
  60         unsigned long flags;
  61         int ret = 0;
  62         sector_t lo = r1_bio->sector;
  63         sector_t hi = lo + r1_bio->sectors;
  64         struct serial_in_rdev *serial = &rdev->serial[idx];
  65
  66         spin_lock_irqsave(&serial->serial_lock, flags);
  67         /* collision happened */
  68         if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
  69                 ret = -EBUSY;
  70         else {
  71                 si->start = lo;
  72                 si->last = hi;
  73                 raid1_rb_insert(si, &serial->serial_rb);
  74         }
  75         spin_unlock_irqrestore(&serial->serial_lock, flags);
  76
  77         return ret;
  78 }
  79
  80 static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
  81 {
  82         struct mddev *mddev = rdev->mddev;
  83         struct serial_info *si;
  84         int idx = sector_to_idx(r1_bio->sector);
  85         struct serial_in_rdev *serial = &rdev->serial[idx];
  86
  87         if (WARN_ON(!mddev->serial_info_pool))
  88                 return;
  89         si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
  90         wait_event(serial->serial_io_wait,
  91                    check_and_add_serial(rdev, r1_bio, si, idx) == 0);
  92 }
  93
  94 static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
  95 {
  96         struct serial_info *si;
  97         unsigned long flags;
  98         int found = 0;
  99         struct mddev *mddev = rdev->mddev;
 100         int idx = sector_to_idx(lo);
 101         struct serial_in_rdev *serial = &rdev->serial[idx];
 102
 103         spin_lock_irqsave(&serial->serial_lock, flags);
 104         for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
 105              si; si = raid1_rb_iter_next(si, lo, hi)) {
 106                 if (si->start == lo && si->last == hi) {
 107                         raid1_rb_remove(si, &serial->serial_rb);
 108                         mempool_free(si, mddev->serial_info_pool);
 109                         found = 1;
 110                         break;
 111                 }
 112         }
 113         if (!found)
 114                 WARN(1, "The write IO is not recorded for serialization\n");
 115         spin_unlock_irqrestore(&serial->serial_lock, flags);
 116         wake_up(&serial->serial_io_wait);
 117 }
 118
 119 /*
 120  * for resync bio, r1bio pointer can be retrieved from the per-bio
 121  * 'struct resync_pages'.
 122  */
 123 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
 124 {
 125         return get_resync_pages(bio)->raid_bio;
 126 }
 127
 128 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
 129 {
 130         struct pool_info *pi = data;
 131         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
 132
 133         /* allocate a r1bio with room for raid_disks entries in the bios array */
 134         return kzalloc(size, gfp_flags);
 135 }
 136
 137 #define RESYNC_DEPTH 32
 138 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 139 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 140 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 141 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 142 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 143
 144 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 145 {
 146         struct pool_info *pi = data;
 147         struct r1bio *r1_bio;
 148         struct bio *bio;
 149         int need_pages;
 150         int j;
 151         struct resync_pages *rps;
 152
 153         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 154         if (!r1_bio)
 155                 return NULL;
 156
 157         rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
 158                             gfp_flags);
 159         if (!rps)
 160                 goto out_free_r1bio;
 161
 162         /*
 163          * Allocate bios : 1 for reading, n-1 for writing
 164          */
 165         for (j = pi->raid_disks ; j-- ; ) {
 166                 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
 167                 if (!bio)
 168                         goto out_free_bio;
 169                 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
 170                 r1_bio->bios[j] = bio;
 171         }
 172         /*
 173          * Allocate RESYNC_PAGES data pages and attach them to
 174          * the first bio.
 175          * If this is a user-requested check/repair, allocate
 176          * RESYNC_PAGES for each bio.
 177          */
 178         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 179                 need_pages = pi->raid_disks;
 180         else
 181                 need_pages = 1;
 182         for (j = 0; j < pi->raid_disks; j++) {
 183                 struct resync_pages *rp = &rps[j];
 184
 185                 bio = r1_bio->bios[j];
 186
 187                 if (j < need_pages) {
 188                         if (resync_alloc_pages(rp, gfp_flags))
 189                                 goto out_free_pages;
 190                 } else {
 191                         memcpy(rp, &rps[0], sizeof(*rp));
 192                         resync_get_all_pages(rp);
 193                 }
 194
 195                 rp->raid_bio = r1_bio;
 196                 bio->bi_private = rp;
 197         }
 198
 199         r1_bio->master_bio = NULL;
 200
 201         return r1_bio;
 202
 203 out_free_pages:
 204         while (--j >= 0)
 205                 resync_free_pages(&rps[j]);
 206
 207 out_free_bio:
 208         while (++j < pi->raid_disks) {
 209                 bio_uninit(r1_bio->bios[j]);
 210                 kfree(r1_bio->bios[j]);
 211         }
 212         kfree(rps);
 213
 214 out_free_r1bio:
 215         rbio_pool_free(r1_bio, data);
 216         return NULL;
 217 }
 218
 219 static void r1buf_pool_free(void *__r1_bio, void *data)
 220 {
 221         struct pool_info *pi = data;
 222         int i;
 223         struct r1bio *r1bio = __r1_bio;
 224         struct resync_pages *rp = NULL;
 225
 226         for (i = pi->raid_disks; i--; ) {
 227                 rp = get_resync_pages(r1bio->bios[i]);
 228                 resync_free_pages(rp);
 229                 bio_uninit(r1bio->bios[i]);
 230                 kfree(r1bio->bios[i]);
 231         }
 232
 233         /* resync pages array stored in the 1st bio's .bi_private */
 234         kfree(rp);
 235
 236         rbio_pool_free(r1bio, data);
 237 }
 238
 239 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 240 {
 241         int i;
 242
 243         for (i = 0; i < conf->raid_disks * 2; i++) {
 244                 struct bio **bio = r1_bio->bios + i;
 245                 if (!BIO_SPECIAL(*bio))
 246                         bio_put(*bio);
 247                 *bio = NULL;
 248         }
 249 }
 250
 251 static void free_r1bio(struct r1bio *r1_bio)
 252 {
 253         struct r1conf *conf = r1_bio->mddev->private;
 254
 255         put_all_bios(conf, r1_bio);
 256         mempool_free(r1_bio, &conf->r1bio_pool);
 257 }
 258
 259 static void put_buf(struct r1bio *r1_bio)
 260 {
 261         struct r1conf *conf = r1_bio->mddev->private;
 262         sector_t sect = r1_bio->sector;
 263         int i;
 264
 265         for (i = 0; i < conf->raid_disks * 2; i++) {
 266                 struct bio *bio = r1_bio->bios[i];
 267                 if (bio->bi_end_io)
 268                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 269         }
 270
 271         mempool_free(r1_bio, &conf->r1buf_pool);
 272
 273         lower_barrier(conf, sect);
 274 }
 275
 276 static void reschedule_retry(struct r1bio *r1_bio)
 277 {
 278         unsigned long flags;
 279         struct mddev *mddev = r1_bio->mddev;
 280         struct r1conf *conf = mddev->private;
 281         int idx;
 282
 283         idx = sector_to_idx(r1_bio->sector);
 284         spin_lock_irqsave(&conf->device_lock, flags);
 285         list_add(&r1_bio->retry_list, &conf->retry_list);
 286         atomic_inc(&conf->nr_queued[idx]);
 287         spin_unlock_irqrestore(&conf->device_lock, flags);
 288
 289         wake_up(&conf->wait_barrier);
 290         md_wakeup_thread(mddev->thread);
 291 }
 292
 293 /*
 294  * raid_end_bio_io() is called when we have finished servicing a mirrored
 295  * operation and are ready to return a success/failure code to the buffer
 296  * cache layer.
 297  */
 298 static void call_bio_endio(struct r1bio *r1_bio)
 299 {
 300         struct bio *bio = r1_bio->master_bio;
 301
 302         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 303                 bio->bi_status = BLK_STS_IOERR;
 304
 305         bio_endio(bio);
 306 }
 307
 308 static void raid_end_bio_io(struct r1bio *r1_bio)
 309 {
 310         struct bio *bio = r1_bio->master_bio;
 311         struct r1conf *conf = r1_bio->mddev->private;
 312         sector_t sector = r1_bio->sector;
 313
 314         /* if nobody has done the final endio yet, do it now */
 315         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 316                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 317                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 318                          (unsigned long long) bio->bi_iter.bi_sector,
 319                          (unsigned long long) bio_end_sector(bio) - 1);
 320
 321                 call_bio_endio(r1_bio);
 322         }
 323
 324         free_r1bio(r1_bio);
 325         /*
 326          * Wake up any possible resync thread that waits for the device
 327          * to go idle.  All I/Os, even write-behind writes, are done.
 328          */
 329         allow_barrier(conf, sector);
 330 }
 331
 332 /*
 333  * Update disk head position estimator based on IRQ completion info.
 334  */
 335 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 336 {
 337         struct r1conf *conf = r1_bio->mddev->private;
 338
 339         conf->mirrors[disk].head_position =
 340                 r1_bio->sector + (r1_bio->sectors);
 341 }
 342
 343 /*
 344  * Find the disk number which triggered given bio
 345  */
 346 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 347 {
 348         int mirror;
 349         struct r1conf *conf = r1_bio->mddev->private;
 350         int raid_disks = conf->raid_disks;
 351
 352         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 353                 if (r1_bio->bios[mirror] == bio)
 354                         break;
 355
 356         BUG_ON(mirror == raid_disks * 2);
 357         update_head_pos(mirror, r1_bio);
 358
 359         return mirror;
 360 }
 361
 362 static void raid1_end_read_request(struct bio *bio)
 363 {
 364         int uptodate = !bio->bi_status;
 365         struct r1bio *r1_bio = bio->bi_private;
 366         struct r1conf *conf = r1_bio->mddev->private;
 367         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 368
 369         /*
 370          * this branch is our 'one mirror IO has finished' event handler:
 371          */
 372         update_head_pos(r1_bio->read_disk, r1_bio);
 373
 374         if (uptodate)
 375                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 376         else if (test_bit(FailFast, &rdev->flags) &&
 377                  test_bit(R1BIO_FailFast, &r1_bio->state))
 378                 /* This was a fail-fast read so we definitely
 379                  * want to retry */
 380                 ;
 381         else {
 382                 /* If all other devices have failed, we want to return
 383                  * the error upwards rather than fail the last device.
 384                  * Here we redefine "uptodate" to mean "Don't want to retry"
 385                  */
 386                 unsigned long flags;
 387                 spin_lock_irqsave(&conf->device_lock, flags);
 388                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 389                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 390                      test_bit(In_sync, &rdev->flags)))
 391                         uptodate = 1;
 392                 spin_unlock_irqrestore(&conf->device_lock, flags);
 393         }
 394
 395         if (uptodate) {
 396                 raid_end_bio_io(r1_bio);
 397                 rdev_dec_pending(rdev, conf->mddev);
 398         } else {
 399                 /*
 400                  * oops, read error:
 401                  */
 402                 pr_err_ratelimited("md/raid1:%s: %pg: rescheduling sector %llu\n",
 403                                    mdname(conf->mddev),
 404                                    rdev->bdev,
 405                                    (unsigned long long)r1_bio->sector);
 406                 set_bit(R1BIO_ReadError, &r1_bio->state);
 407                 reschedule_retry(r1_bio);
 408                 /* don't drop the reference on read_disk yet */
 409         }
 410 }
 411
 412 static void close_write(struct r1bio *r1_bio)
 413 {
 414         struct mddev *mddev = r1_bio->mddev;
 415
 416         /* it really is the end of this request */
 417         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 418                 bio_free_pages(r1_bio->behind_master_bio);
 419                 bio_put(r1_bio->behind_master_bio);
 420                 r1_bio->behind_master_bio = NULL;
 421         }
 422
 423         /* clear the bitmap if all writes complete successfully */
 424         mddev->bitmap_ops->endwrite(mddev, r1_bio->sector, r1_bio->sectors,
 425                                     !test_bit(R1BIO_Degraded, &r1_bio->state),
 426                                     test_bit(R1BIO_BehindIO, &r1_bio->state));
 427         md_write_end(mddev);
 428 }
 429
 430 static void r1_bio_write_done(struct r1bio *r1_bio)
 431 {
 432         if (!atomic_dec_and_test(&r1_bio->remaining))
 433                 return;
 434
 435         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 436                 reschedule_retry(r1_bio);
 437         else {
 438                 close_write(r1_bio);
 439                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 440                         reschedule_retry(r1_bio);
 441                 else
 442                         raid_end_bio_io(r1_bio);
 443         }
 444 }
 445
 446 static void raid1_end_write_request(struct bio *bio)
 447 {
 448         struct r1bio *r1_bio = bio->bi_private;
 449         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 450         struct r1conf *conf = r1_bio->mddev->private;
 451         struct bio *to_put = NULL;
 452         int mirror = find_bio_disk(r1_bio, bio);
 453         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 454         bool discard_error;
 455         sector_t lo = r1_bio->sector;
 456         sector_t hi = r1_bio->sector + r1_bio->sectors;
 457
 458         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 459
 460         /*
 461          * 'one mirror IO has finished' event handler:
 462          */
 463         if (bio->bi_status && !discard_error) {
 464                 set_bit(WriteErrorSeen, &rdev->flags);
 465                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
 466                         set_bit(MD_RECOVERY_NEEDED, &
 467                                 conf->mddev->recovery);
 468
 469                 if (test_bit(FailFast, &rdev->flags) &&
 470                     (bio->bi_opf & MD_FAILFAST) &&
 471                     /* We never try FailFast to WriteMostly devices */
 472                     !test_bit(WriteMostly, &rdev->flags)) {
 473                         md_error(r1_bio->mddev, rdev);
 474                 }
 475
 476                 /*
 477                  * When the device is faulty, it is not necessary to
 478                  * handle write error.
 479                  */
 480                 if (!test_bit(Faulty, &rdev->flags))
 481                         set_bit(R1BIO_WriteError, &r1_bio->state);
 482                 else {
 483                         /* Fail the request */
 484                         set_bit(R1BIO_Degraded, &r1_bio->state);
 485                         /* Finished with this branch */
 486                         r1_bio->bios[mirror] = NULL;
 487                         to_put = bio;
 488                 }
 489         } else {
 490                 /*
 491                  * Set R1BIO_Uptodate in our master bio, so that we
 492                  * will return a good error code for to the higher
 493                  * levels even if IO on some other mirrored buffer
 494                  * fails.
 495                  *
 496                  * The 'master' represents the composite IO operation
 497                  * to user-side. So if something waits for IO, then it
 498                  * will wait for the 'master' bio.
 499                  */
 500                 r1_bio->bios[mirror] = NULL;
 501                 to_put = bio;
 502                 /*
 503                  * Do not set R1BIO_Uptodate if the current device is
 504                  * rebuilding or Faulty. This is because we cannot use
 505                  * such device for properly reading the data back (we could
 506                  * potentially use it, if the current write would have felt
 507                  * before rdev->recovery_offset, but for simplicity we don't
 508                  * check this here.
 509                  */
 510                 if (test_bit(In_sync, &rdev->flags) &&
 511                     !test_bit(Faulty, &rdev->flags))
 512                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 513
 514                 /* Maybe we can clear some bad blocks. */
 515                 if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors) &&
 516                     !discard_error) {
 517                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 518                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 519                 }
 520         }
 521
 522         if (behind) {
 523                 if (test_bit(CollisionCheck, &rdev->flags))
 524                         remove_serial(rdev, lo, hi);
 525                 if (test_bit(WriteMostly, &rdev->flags))
 526                         atomic_dec(&r1_bio->behind_remaining);
 527
 528                 /*
 529                  * In behind mode, we ACK the master bio once the I/O
 530                  * has safely reached all non-writemostly
 531                  * disks. Setting the Returned bit ensures that this
 532                  * gets done only once -- we don't ever want to return
 533                  * -EIO here, instead we'll wait
 534                  */
 535                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 536                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 537                         /* Maybe we can return now */
 538                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 539                                 struct bio *mbio = r1_bio->master_bio;
 540                                 pr_debug("raid1: behind end write sectors"
 541                                          " %llu-%llu\n",
 542                                          (unsigned long long) mbio->bi_iter.bi_sector,
 543                                          (unsigned long long) bio_end_sector(mbio) - 1);
 544                                 call_bio_endio(r1_bio);
 545                         }
 546                 }
 547         } else if (rdev->mddev->serialize_policy)
 548                 remove_serial(rdev, lo, hi);
 549         if (r1_bio->bios[mirror] == NULL)
 550                 rdev_dec_pending(rdev, conf->mddev);
 551
 552         /*
 553          * Let's see if all mirrored write operations have finished
 554          * already.
 555          */
 556         r1_bio_write_done(r1_bio);
 557
 558         if (to_put)
 559                 bio_put(to_put);
 560 }
 561
 562 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 563                                           sector_t sectors)
 564 {
 565         sector_t len;
 566
 567         WARN_ON(sectors == 0);
 568         /*
 569          * len is the number of sectors from start_sector to end of the
 570          * barrier unit which start_sector belongs to.
 571          */
 572         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 573               start_sector;
 574
 575         if (len > sectors)
 576                 len = sectors;
 577
 578         return len;
 579 }
 580
 581 static void update_read_sectors(struct r1conf *conf, int disk,
 582                                 sector_t this_sector, int len)
 583 {
 584         struct raid1_info *info = &conf->mirrors[disk];
 585
 586         atomic_inc(&info->rdev->nr_pending);
 587         if (info->next_seq_sect != this_sector)
 588                 info->seq_start = this_sector;
 589         info->next_seq_sect = this_sector + len;
 590 }
 591
 592 static int choose_first_rdev(struct r1conf *conf, struct r1bio *r1_bio,
 593                              int *max_sectors)
 594 {
 595         sector_t this_sector = r1_bio->sector;
 596         int len = r1_bio->sectors;
 597         int disk;
 598
 599         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 600                 struct md_rdev *rdev;
 601                 int read_len;
 602
 603                 if (r1_bio->bios[disk] == IO_BLOCKED)
 604                         continue;
 605
 606                 rdev = conf->mirrors[disk].rdev;
 607                 if (!rdev || test_bit(Faulty, &rdev->flags))
 608                         continue;
 609
 610                 /* choose the first disk even if it has some bad blocks. */
 611                 read_len = raid1_check_read_range(rdev, this_sector, &len);
 612                 if (read_len > 0) {
 613                         update_read_sectors(conf, disk, this_sector, read_len);
 614                         *max_sectors = read_len;
 615                         return disk;
 616                 }
 617         }
 618
 619         return -1;
 620 }
 621
 622 static bool rdev_in_recovery(struct md_rdev *rdev, struct r1bio *r1_bio)
 623 {
 624         return !test_bit(In_sync, &rdev->flags) &&
 625                rdev->recovery_offset < r1_bio->sector + r1_bio->sectors;
 626 }
 627
 628 static int choose_bb_rdev(struct r1conf *conf, struct r1bio *r1_bio,
 629                           int *max_sectors)
 630 {
 631         sector_t this_sector = r1_bio->sector;
 632         int best_disk = -1;
 633         int best_len = 0;
 634         int disk;
 635
 636         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 637                 struct md_rdev *rdev;
 638                 int len;
 639                 int read_len;
 640
 641                 if (r1_bio->bios[disk] == IO_BLOCKED)
 642                         continue;
 643
 644                 rdev = conf->mirrors[disk].rdev;
 645                 if (!rdev || test_bit(Faulty, &rdev->flags) ||
 646                     rdev_in_recovery(rdev, r1_bio) ||
 647                     test_bit(WriteMostly, &rdev->flags))
 648                         continue;
 649
 650                 /* keep track of the disk with the most readable sectors. */
 651                 len = r1_bio->sectors;
 652                 read_len = raid1_check_read_range(rdev, this_sector, &len);
 653                 if (read_len > best_len) {
 654                         best_disk = disk;
 655                         best_len = read_len;
 656                 }
 657         }
 658
 659         if (best_disk != -1) {
 660                 *max_sectors = best_len;
 661                 update_read_sectors(conf, best_disk, this_sector, best_len);
 662         }
 663
 664         return best_disk;
 665 }
 666
 667 static int choose_slow_rdev(struct r1conf *conf, struct r1bio *r1_bio,
 668                             int *max_sectors)
 669 {
 670         sector_t this_sector = r1_bio->sector;
 671         int bb_disk = -1;
 672         int bb_read_len = 0;
 673         int disk;
 674
 675         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 676                 struct md_rdev *rdev;
 677                 int len;
 678                 int read_len;
 679
 680                 if (r1_bio->bios[disk] == IO_BLOCKED)
 681                         continue;
 682
 683                 rdev = conf->mirrors[disk].rdev;
 684                 if (!rdev || test_bit(Faulty, &rdev->flags) ||
 685                     !test_bit(WriteMostly, &rdev->flags) ||
 686                     rdev_in_recovery(rdev, r1_bio))
 687                         continue;
 688
 689                 /* there are no bad blocks, we can use this disk */
 690                 len = r1_bio->sectors;
 691                 read_len = raid1_check_read_range(rdev, this_sector, &len);
 692                 if (read_len == r1_bio->sectors) {
 693                         *max_sectors = read_len;
 694                         update_read_sectors(conf, disk, this_sector, read_len);
 695                         return disk;
 696                 }
 697
 698                 /*
 699                  * there are partial bad blocks, choose the rdev with largest
 700                  * read length.
 701                  */
 702                 if (read_len > bb_read_len) {
 703                         bb_disk = disk;
 704                         bb_read_len = read_len;
 705                 }
 706         }
 707
 708         if (bb_disk != -1) {
 709                 *max_sectors = bb_read_len;
 710                 update_read_sectors(conf, bb_disk, this_sector, bb_read_len);
 711         }
 712
 713         return bb_disk;
 714 }
 715
 716 static bool is_sequential(struct r1conf *conf, int disk, struct r1bio *r1_bio)
 717 {
 718         /* TODO: address issues with this check and concurrency. */
 719         return conf->mirrors[disk].next_seq_sect == r1_bio->sector ||
 720                conf->mirrors[disk].head_position == r1_bio->sector;
 721 }
 722
 723 /*
 724  * If buffered sequential IO size exceeds optimal iosize, check if there is idle
 725  * disk. If yes, choose the idle disk.
 726  */
 727 static bool should_choose_next(struct r1conf *conf, int disk)
 728 {
 729         struct raid1_info *mirror = &conf->mirrors[disk];
 730         int opt_iosize;
 731
 732         if (!test_bit(Nonrot, &mirror->rdev->flags))
 733                 return false;
 734
 735         opt_iosize = bdev_io_opt(mirror->rdev->bdev) >> 9;
 736         return opt_iosize > 0 && mirror->seq_start != MaxSector &&
 737                mirror->next_seq_sect > opt_iosize &&
 738                mirror->next_seq_sect - opt_iosize >= mirror->seq_start;
 739 }
 740
 741 static bool rdev_readable(struct md_rdev *rdev, struct r1bio *r1_bio)
 742 {
 743         if (!rdev || test_bit(Faulty, &rdev->flags))
 744                 return false;
 745
 746         if (rdev_in_recovery(rdev, r1_bio))
 747                 return false;
 748
 749         /* don't read from slow disk unless have to */
 750         if (test_bit(WriteMostly, &rdev->flags))
 751                 return false;
 752
 753         /* don't split IO for bad blocks unless have to */
 754         if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors))
 755                 return false;
 756
 757         return true;
 758 }
 759
 760 struct read_balance_ctl {
 761         sector_t closest_dist;
 762         int closest_dist_disk;
 763         int min_pending;
 764         int min_pending_disk;
 765         int sequential_disk;
 766         int readable_disks;
 767 };
 768
 769 static int choose_best_rdev(struct r1conf *conf, struct r1bio *r1_bio)
 770 {
 771         int disk;
 772         struct read_balance_ctl ctl = {
 773                 .closest_dist_disk      = -1,
 774                 .closest_dist           = MaxSector,
 775                 .min_pending_disk       = -1,
 776                 .min_pending            = UINT_MAX,
 777                 .sequential_disk        = -1,
 778         };
 779
 780         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 781                 struct md_rdev *rdev;
 782                 sector_t dist;
 783                 unsigned int pending;
 784
 785                 if (r1_bio->bios[disk] == IO_BLOCKED)
 786                         continue;
 787
 788                 rdev = conf->mirrors[disk].rdev;
 789                 if (!rdev_readable(rdev, r1_bio))
 790                         continue;
 791
 792                 /* At least two disks to choose from so failfast is OK */
 793                 if (ctl.readable_disks++ == 1)
 794                         set_bit(R1BIO_FailFast, &r1_bio->state);
 795
 796                 pending = atomic_read(&rdev->nr_pending);
 797                 dist = abs(r1_bio->sector - conf->mirrors[disk].head_position);
 798
 799                 /* Don't change to another disk for sequential reads */
 800                 if (is_sequential(conf, disk, r1_bio)) {
 801                         if (!should_choose_next(conf, disk))
 802                                 return disk;
 803
 804                         /*
 805                          * Add 'pending' to avoid choosing this disk if
 806                          * there is other idle disk.
 807                          */
 808                         pending++;
 809                         /*
 810                          * If there is no other idle disk, this disk
 811                          * will be chosen.
 812                          */
 813                         ctl.sequential_disk = disk;
 814                 }
 815
 816                 if (ctl.min_pending > pending) {
 817                         ctl.min_pending = pending;
 818                         ctl.min_pending_disk = disk;
 819                 }
 820
 821                 if (ctl.closest_dist > dist) {
 822                         ctl.closest_dist = dist;
 823                         ctl.closest_dist_disk = disk;
 824                 }
 825         }
 826
 827         /*
 828          * sequential IO size exceeds optimal iosize, however, there is no other
 829          * idle disk, so choose the sequential disk.
 830          */
 831         if (ctl.sequential_disk != -1 && ctl.min_pending != 0)
 832                 return ctl.sequential_disk;
 833
 834         /*
 835          * If all disks are rotational, choose the closest disk. If any disk is
 836          * non-rotational, choose the disk with less pending request even the
 837          * disk is rotational, which might/might not be optimal for raids with
 838          * mixed ratation/non-rotational disks depending on workload.
 839          */
 840         if (ctl.min_pending_disk != -1 &&
 841             (READ_ONCE(conf->nonrot_disks) || ctl.min_pending == 0))
 842                 return ctl.min_pending_disk;
 843         else
 844                 return ctl.closest_dist_disk;
 845 }
 846
 847 /*
 848  * This routine returns the disk from which the requested read should be done.
 849  *
 850  * 1) If resync is in progress, find the first usable disk and use it even if it
 851  * has some bad blocks.
 852  *
 853  * 2) Now that there is no resync, loop through all disks and skipping slow
 854  * disks and disks with bad blocks for now. Only pay attention to key disk
 855  * choice.
 856  *
 857  * 3) If we've made it this far, now look for disks with bad blocks and choose
 858  * the one with most number of sectors.
 859  *
 860  * 4) If we are all the way at the end, we have no choice but to use a disk even
 861  * if it is write mostly.
 862  *
 863  * The rdev for the device selected will have nr_pending incremented.
 864  */
 865 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio,
 866                         int *max_sectors)
 867 {
 868         int disk;
 869
 870         clear_bit(R1BIO_FailFast, &r1_bio->state);
 871
 872         if (raid1_should_read_first(conf->mddev, r1_bio->sector,
 873                                     r1_bio->sectors))
 874                 return choose_first_rdev(conf, r1_bio, max_sectors);
 875
 876         disk = choose_best_rdev(conf, r1_bio);
 877         if (disk >= 0) {
 878                 *max_sectors = r1_bio->sectors;
 879                 update_read_sectors(conf, disk, r1_bio->sector,
 880                                     r1_bio->sectors);
 881                 return disk;
 882         }
 883
 884         /*
 885          * If we are here it means we didn't find a perfectly good disk so
 886          * now spend a bit more time trying to find one with the most good
 887          * sectors.
 888          */
 889         disk = choose_bb_rdev(conf, r1_bio, max_sectors);
 890         if (disk >= 0)
 891                 return disk;
 892
 893         return choose_slow_rdev(conf, r1_bio, max_sectors);
 894 }
 895
 896 static void wake_up_barrier(struct r1conf *conf)
 897 {
 898         if (wq_has_sleeper(&conf->wait_barrier))
 899                 wake_up(&conf->wait_barrier);
 900 }
 901
 902 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 903 {
 904         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 905         raid1_prepare_flush_writes(conf->mddev);
 906         wake_up_barrier(conf);
 907
 908         while (bio) { /* submit pending writes */
 909                 struct bio *next = bio->bi_next;
 910
 911                 raid1_submit_write(bio);
 912                 bio = next;
 913                 cond_resched();
 914         }
 915 }
 916
 917 static void flush_pending_writes(struct r1conf *conf)
 918 {
 919         /* Any writes that have been queued but are awaiting
 920          * bitmap updates get flushed here.
 921          */
 922         spin_lock_irq(&conf->device_lock);
 923
 924         if (conf->pending_bio_list.head) {
 925                 struct blk_plug plug;
 926                 struct bio *bio;
 927
 928                 bio = bio_list_get(&conf->pending_bio_list);
 929                 spin_unlock_irq(&conf->device_lock);
 930
 931                 /*
 932                  * As this is called in a wait_event() loop (see freeze_array),
 933                  * current->state might be TASK_UNINTERRUPTIBLE which will
 934                  * cause a warning when we prepare to wait again.  As it is
 935                  * rare that this path is taken, it is perfectly safe to force
 936                  * us to go around the wait_event() loop again, so the warning
 937                  * is a false-positive.  Silence the warning by resetting
 938                  * thread state
 939                  */
 940                 __set_current_state(TASK_RUNNING);
 941                 blk_start_plug(&plug);
 942                 flush_bio_list(conf, bio);
 943                 blk_finish_plug(&plug);
 944         } else
 945                 spin_unlock_irq(&conf->device_lock);
 946 }
 947
 948 /* Barriers....
 949  * Sometimes we need to suspend IO while we do something else,
 950  * either some resync/recovery, or reconfigure the array.
 951  * To do this we raise a 'barrier'.
 952  * The 'barrier' is a counter that can be raised multiple times
 953  * to count how many activities are happening which preclude
 954  * normal IO.
 955  * We can only raise the barrier if there is no pending IO.
 956  * i.e. if nr_pending == 0.
 957  * We choose only to raise the barrier if no-one is waiting for the
 958  * barrier to go down.  This means that as soon as an IO request
 959  * is ready, no other operations which require a barrier will start
 960  * until the IO request has had a chance.
 961  *
 962  * So: regular IO calls 'wait_barrier'.  When that returns there
 963  *    is no backgroup IO happening,  It must arrange to call
 964  *    allow_barrier when it has finished its IO.
 965  * backgroup IO calls must call raise_barrier.  Once that returns
 966  *    there is no normal IO happeing.  It must arrange to call
 967  *    lower_barrier when the particular background IO completes.
 968  *
 969  * If resync/recovery is interrupted, returns -EINTR;
 970  * Otherwise, returns 0.
 971  */
 972 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 973 {
 974         int idx = sector_to_idx(sector_nr);
 975
 976         spin_lock_irq(&conf->resync_lock);
 977
 978         /* Wait until no block IO is waiting */
 979         wait_event_lock_irq(conf->wait_barrier,
 980                             !atomic_read(&conf->nr_waiting[idx]),
 981                             conf->resync_lock);
 982
 983         /* block any new IO from starting */
 984         atomic_inc(&conf->barrier[idx]);
 985         /*
 986          * In raise_barrier() we firstly increase conf->barrier[idx] then
 987          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 988          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 989          * A memory barrier here to make sure conf->nr_pending[idx] won't
 990          * be fetched before conf->barrier[idx] is increased. Otherwise
 991          * there will be a race between raise_barrier() and _wait_barrier().
 992          */
 993         smp_mb__after_atomic();
 994
 995         /* For these conditions we must wait:
 996          * A: while the array is in frozen state
 997          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 998          *    existing in corresponding I/O barrier bucket.
 999          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
1000          *    max resync count which allowed on current I/O barrier bucket.
1001          */
1002         wait_event_lock_irq(conf->wait_barrier,
1003                             (!conf->array_frozen &&
1004                              !atomic_read(&conf->nr_pending[idx]) &&
1005                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
1006                                 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
1007                             conf->resync_lock);
1008
1009         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
1010                 atomic_dec(&conf->barrier[idx]);
1011                 spin_unlock_irq(&conf->resync_lock);
1012                 wake_up(&conf->wait_barrier);
1013                 return -EINTR;
1014         }
1015
1016         atomic_inc(&conf->nr_sync_pending);
1017         spin_unlock_irq(&conf->resync_lock);
1018
1019         return 0;
1020 }
1021
1022 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
1023 {
1024         int idx = sector_to_idx(sector_nr);
1025
1026         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
1027
1028         atomic_dec(&conf->barrier[idx]);
1029         atomic_dec(&conf->nr_sync_pending);
1030         wake_up(&conf->wait_barrier);
1031 }
1032
1033 static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
1034 {
1035         bool ret = true;
1036
1037         /*
1038          * We need to increase conf->nr_pending[idx] very early here,
1039          * then raise_barrier() can be blocked when it waits for
1040          * conf->nr_pending[idx] to be 0. Then we can avoid holding
1041          * conf->resync_lock when there is no barrier raised in same
1042          * barrier unit bucket. Also if the array is frozen, I/O
1043          * should be blocked until array is unfrozen.
1044          */
1045         atomic_inc(&conf->nr_pending[idx]);
1046         /*
1047          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
1048          * check conf->barrier[idx]. In raise_barrier() we firstly increase
1049          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
1050          * barrier is necessary here to make sure conf->barrier[idx] won't be
1051          * fetched before conf->nr_pending[idx] is increased. Otherwise there
1052          * will be a race between _wait_barrier() and raise_barrier().
1053          */
1054         smp_mb__after_atomic();
1055
1056         /*
1057          * Don't worry about checking two atomic_t variables at same time
1058          * here. If during we check conf->barrier[idx], the array is
1059          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
1060          * 0, it is safe to return and make the I/O continue. Because the
1061          * array is frozen, all I/O returned here will eventually complete
1062          * or be queued, no race will happen. See code comment in
1063          * frozen_array().
1064          */
1065         if (!READ_ONCE(conf->array_frozen) &&
1066             !atomic_read(&conf->barrier[idx]))
1067                 return ret;
1068
1069         /*
1070          * After holding conf->resync_lock, conf->nr_pending[idx]
1071          * should be decreased before waiting for barrier to drop.
1072          * Otherwise, we may encounter a race condition because
1073          * raise_barrer() might be waiting for conf->nr_pending[idx]
1074          * to be 0 at same time.
1075          */
1076         spin_lock_irq(&conf->resync_lock);
1077         atomic_inc(&conf->nr_waiting[idx]);
1078         atomic_dec(&conf->nr_pending[idx]);
1079         /*
1080          * In case freeze_array() is waiting for
1081          * get_unqueued_pending() == extra
1082          */
1083         wake_up_barrier(conf);
1084         /* Wait for the barrier in same barrier unit bucket to drop. */
1085
1086         /* Return false when nowait flag is set */
1087         if (nowait) {
1088                 ret = false;
1089         } else {
1090                 wait_event_lock_irq(conf->wait_barrier,
1091                                 !conf->array_frozen &&
1092                                 !atomic_read(&conf->barrier[idx]),
1093                                 conf->resync_lock);
1094                 atomic_inc(&conf->nr_pending[idx]);
1095         }
1096
1097         atomic_dec(&conf->nr_waiting[idx]);
1098         spin_unlock_irq(&conf->resync_lock);
1099         return ret;
1100 }
1101
1102 static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1103 {
1104         int idx = sector_to_idx(sector_nr);
1105         bool ret = true;
1106
1107         /*
1108          * Very similar to _wait_barrier(). The difference is, for read
1109          * I/O we don't need wait for sync I/O, but if the whole array
1110          * is frozen, the read I/O still has to wait until the array is
1111          * unfrozen. Since there is no ordering requirement with
1112          * conf->barrier[idx] here, memory barrier is unnecessary as well.
1113          */
1114         atomic_inc(&conf->nr_pending[idx]);
1115
1116         if (!READ_ONCE(conf->array_frozen))
1117                 return ret;
1118
1119         spin_lock_irq(&conf->resync_lock);
1120         atomic_inc(&conf->nr_waiting[idx]);
1121         atomic_dec(&conf->nr_pending[idx]);
1122         /*
1123          * In case freeze_array() is waiting for
1124          * get_unqueued_pending() == extra
1125          */
1126         wake_up_barrier(conf);
1127         /* Wait for array to be unfrozen */
1128
1129         /* Return false when nowait flag is set */
1130         if (nowait) {
1131                 /* Return false when nowait flag is set */
1132                 ret = false;
1133         } else {
1134                 wait_event_lock_irq(conf->wait_barrier,
1135                                 !conf->array_frozen,
1136                                 conf->resync_lock);
1137                 atomic_inc(&conf->nr_pending[idx]);
1138         }
1139
1140         atomic_dec(&conf->nr_waiting[idx]);
1141         spin_unlock_irq(&conf->resync_lock);
1142         return ret;
1143 }
1144
1145 static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1146 {
1147         int idx = sector_to_idx(sector_nr);
1148
1149         return _wait_barrier(conf, idx, nowait);
1150 }
1151
1152 static void _allow_barrier(struct r1conf *conf, int idx)
1153 {
1154         atomic_dec(&conf->nr_pending[idx]);
1155         wake_up_barrier(conf);
1156 }
1157
1158 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1159 {
1160         int idx = sector_to_idx(sector_nr);
1161
1162         _allow_barrier(conf, idx);
1163 }
1164
1165 /* conf->resync_lock should be held */
1166 static int get_unqueued_pending(struct r1conf *conf)
1167 {
1168         int idx, ret;
1169
1170         ret = atomic_read(&conf->nr_sync_pending);
1171         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1172                 ret += atomic_read(&conf->nr_pending[idx]) -
1173                         atomic_read(&conf->nr_queued[idx]);
1174
1175         return ret;
1176 }
1177
1178 static void freeze_array(struct r1conf *conf, int extra)
1179 {
1180         /* Stop sync I/O and normal I/O and wait for everything to
1181          * go quiet.
1182          * This is called in two situations:
1183          * 1) management command handlers (reshape, remove disk, quiesce).
1184          * 2) one normal I/O request failed.
1185
1186          * After array_frozen is set to 1, new sync IO will be blocked at
1187          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1188          * or wait_read_barrier(). The flying I/Os will either complete or be
1189          * queued. When everything goes quite, there are only queued I/Os left.
1190
1191          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1192          * barrier bucket index which this I/O request hits. When all sync and
1193          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1194          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1195          * in handle_read_error(), we may call freeze_array() before trying to
1196          * fix the read error. In this case, the error read I/O is not queued,
1197          * so get_unqueued_pending() == 1.
1198          *
1199          * Therefore before this function returns, we need to wait until
1200          * get_unqueued_pendings(conf) gets equal to extra. For
1201          * normal I/O context, extra is 1, in rested situations extra is 0.
1202          */
1203         spin_lock_irq(&conf->resync_lock);
1204         conf->array_frozen = 1;
1205         mddev_add_trace_msg(conf->mddev, "raid1 wait freeze");
1206         wait_event_lock_irq_cmd(
1207                 conf->wait_barrier,
1208                 get_unqueued_pending(conf) == extra,
1209                 conf->resync_lock,
1210                 flush_pending_writes(conf));
1211         spin_unlock_irq(&conf->resync_lock);
1212 }
1213 static void unfreeze_array(struct r1conf *conf)
1214 {
1215         /* reverse the effect of the freeze */
1216         spin_lock_irq(&conf->resync_lock);
1217         conf->array_frozen = 0;
1218         spin_unlock_irq(&conf->resync_lock);
1219         wake_up(&conf->wait_barrier);
1220 }
1221
1222 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1223                                            struct bio *bio)
1224 {
1225         int size = bio->bi_iter.bi_size;
1226         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1227         int i = 0;
1228         struct bio *behind_bio = NULL;
1229
1230         behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1231                                       &r1_bio->mddev->bio_set);
1232
1233         /* discard op, we don't support writezero/writesame yet */
1234         if (!bio_has_data(bio)) {
1235                 behind_bio->bi_iter.bi_size = size;
1236                 goto skip_copy;
1237         }
1238
1239         while (i < vcnt && size) {
1240                 struct page *page;
1241                 int len = min_t(int, PAGE_SIZE, size);
1242
1243                 page = alloc_page(GFP_NOIO);
1244                 if (unlikely(!page))
1245                         goto free_pages;
1246
1247                 if (!bio_add_page(behind_bio, page, len, 0)) {
1248                         put_page(page);
1249                         goto free_pages;
1250                 }
1251
1252                 size -= len;
1253                 i++;
1254         }
1255
1256         bio_copy_data(behind_bio, bio);
1257 skip_copy:
1258         r1_bio->behind_master_bio = behind_bio;
1259         set_bit(R1BIO_BehindIO, &r1_bio->state);
1260
1261         return;
1262
1263 free_pages:
1264         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1265                  bio->bi_iter.bi_size);
1266         bio_free_pages(behind_bio);
1267         bio_put(behind_bio);
1268 }
1269
1270 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1271 {
1272         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1273                                                   cb);
1274         struct mddev *mddev = plug->cb.data;
1275         struct r1conf *conf = mddev->private;
1276         struct bio *bio;
1277
1278         if (from_schedule) {
1279                 spin_lock_irq(&conf->device_lock);
1280                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1281                 spin_unlock_irq(&conf->device_lock);
1282                 wake_up_barrier(conf);
1283                 md_wakeup_thread(mddev->thread);
1284                 kfree(plug);
1285                 return;
1286         }
1287
1288         /* we aren't scheduling, so we can do the write-out directly. */
1289         bio = bio_list_get(&plug->pending);
1290         flush_bio_list(conf, bio);
1291         kfree(plug);
1292 }
1293
1294 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1295 {
1296         r1_bio->master_bio = bio;
1297         r1_bio->sectors = bio_sectors(bio);
1298         r1_bio->state = 0;
1299         r1_bio->mddev = mddev;
1300         r1_bio->sector = bio->bi_iter.bi_sector;
1301 }
1302
1303 static inline struct r1bio *
1304 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1305 {
1306         struct r1conf *conf = mddev->private;
1307         struct r1bio *r1_bio;
1308
1309         r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1310         /* Ensure no bio records IO_BLOCKED */
1311         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1312         init_r1bio(r1_bio, mddev, bio);
1313         return r1_bio;
1314 }
1315
1316 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1317                                int max_read_sectors, struct r1bio *r1_bio)
1318 {
1319         struct r1conf *conf = mddev->private;
1320         struct raid1_info *mirror;
1321         struct bio *read_bio;
1322         const enum req_op op = bio_op(bio);
1323         const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1324         int max_sectors;
1325         int rdisk, error;
1326         bool r1bio_existed = !!r1_bio;
1327
1328         /*
1329          * If r1_bio is set, we are blocking the raid1d thread
1330          * so there is a tiny risk of deadlock.  So ask for
1331          * emergency memory if needed.
1332          */
1333         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1334
1335         /*
1336          * Still need barrier for READ in case that whole
1337          * array is frozen.
1338          */
1339         if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1340                                 bio->bi_opf & REQ_NOWAIT)) {
1341                 bio_wouldblock_error(bio);
1342                 return;
1343         }
1344
1345         if (!r1_bio)
1346                 r1_bio = alloc_r1bio(mddev, bio);
1347         else
1348                 init_r1bio(r1_bio, mddev, bio);
1349         r1_bio->sectors = max_read_sectors;
1350
1351         /*
1352          * make_request() can abort the operation when read-ahead is being
1353          * used and no empty request is available.
1354          */
1355         rdisk = read_balance(conf, r1_bio, &max_sectors);
1356         if (rdisk < 0) {
1357                 /* couldn't find anywhere to read from */
1358                 if (r1bio_existed)
1359                         pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
1360                                             mdname(mddev),
1361                                             conf->mirrors[r1_bio->read_disk].rdev->bdev,
1362                                             r1_bio->sector);
1363                 raid_end_bio_io(r1_bio);
1364                 return;
1365         }
1366         mirror = conf->mirrors + rdisk;
1367
1368         if (r1bio_existed)
1369                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %pg\n",
1370                                     mdname(mddev),
1371                                     (unsigned long long)r1_bio->sector,
1372                                     mirror->rdev->bdev);
1373
1374         if (test_bit(WriteMostly, &mirror->rdev->flags)) {
1375                 /*
1376                  * Reading from a write-mostly device must take care not to
1377                  * over-take any writes that are 'behind'
1378                  */
1379                 mddev_add_trace_msg(mddev, "raid1 wait behind writes");
1380                 mddev->bitmap_ops->wait_behind_writes(mddev);
1381         }
1382
1383         if (max_sectors < bio_sectors(bio)) {
1384                 struct bio *split = bio_split(bio, max_sectors,
1385                                               gfp, &conf->bio_split);
1386
1387                 if (IS_ERR(split)) {
1388                         error = PTR_ERR(split);
1389                         goto err_handle;
1390                 }
1391                 bio_chain(split, bio);
1392                 submit_bio_noacct(bio);
1393                 bio = split;
1394                 r1_bio->master_bio = bio;
1395                 r1_bio->sectors = max_sectors;
1396         }
1397
1398         r1_bio->read_disk = rdisk;
1399         if (!r1bio_existed) {
1400                 md_account_bio(mddev, &bio);
1401                 r1_bio->master_bio = bio;
1402         }
1403         read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1404                                    &mddev->bio_set);
1405
1406         r1_bio->bios[rdisk] = read_bio;
1407
1408         read_bio->bi_iter.bi_sector = r1_bio->sector +
1409                 mirror->rdev->data_offset;
1410         read_bio->bi_end_io = raid1_end_read_request;
1411         read_bio->bi_opf = op | do_sync;
1412         if (test_bit(FailFast, &mirror->rdev->flags) &&
1413             test_bit(R1BIO_FailFast, &r1_bio->state))
1414                 read_bio->bi_opf |= MD_FAILFAST;
1415         read_bio->bi_private = r1_bio;
1416         mddev_trace_remap(mddev, read_bio, r1_bio->sector);
1417         submit_bio_noacct(read_bio);
1418         return;
1419
1420 err_handle:
1421         atomic_dec(&mirror->rdev->nr_pending);
1422         bio->bi_status = errno_to_blk_status(error);
1423         set_bit(R1BIO_Uptodate, &r1_bio->state);
1424         raid_end_bio_io(r1_bio);
1425 }
1426
1427 static bool wait_blocked_rdev(struct mddev *mddev, struct bio *bio)
1428 {
1429         struct r1conf *conf = mddev->private;
1430         int disks = conf->raid_disks * 2;
1431         int i;
1432
1433 retry:
1434         for (i = 0; i < disks; i++) {
1435                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1436
1437                 if (!rdev)
1438                         continue;
1439
1440                 /* don't write here until the bad block is acknowledged */
1441                 if (test_bit(WriteErrorSeen, &rdev->flags) &&
1442                     rdev_has_badblock(rdev, bio->bi_iter.bi_sector,
1443                                       bio_sectors(bio)) < 0)
1444                         set_bit(BlockedBadBlocks, &rdev->flags);
1445
1446                 if (rdev_blocked(rdev)) {
1447                         if (bio->bi_opf & REQ_NOWAIT)
1448                                 return false;
1449
1450                         mddev_add_trace_msg(rdev->mddev, "raid1 wait rdev %d blocked",
1451                                             rdev->raid_disk);
1452                         atomic_inc(&rdev->nr_pending);
1453                         md_wait_for_blocked_rdev(rdev, rdev->mddev);
1454                         goto retry;
1455                 }
1456         }
1457
1458         return true;
1459 }
1460
1461 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1462                                 int max_write_sectors)
1463 {
1464         struct r1conf *conf = mddev->private;
1465         struct r1bio *r1_bio;
1466         int i, disks, k, error;
1467         unsigned long flags;
1468         int first_clone;
1469         int max_sectors;
1470         bool write_behind = false;
1471         bool is_discard = (bio_op(bio) == REQ_OP_DISCARD);
1472
1473         if (mddev_is_clustered(mddev) &&
1474              md_cluster_ops->area_resyncing(mddev, WRITE,
1475                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1476
1477                 DEFINE_WAIT(w);
1478                 if (bio->bi_opf & REQ_NOWAIT) {
1479                         bio_wouldblock_error(bio);
1480                         return;
1481                 }
1482                 for (;;) {
1483                         prepare_to_wait(&conf->wait_barrier,
1484                                         &w, TASK_IDLE);
1485                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1486                                                         bio->bi_iter.bi_sector,
1487                                                         bio_end_sector(bio)))
1488                                 break;
1489                         schedule();
1490                 }
1491                 finish_wait(&conf->wait_barrier, &w);
1492         }
1493
1494         /*
1495          * Register the new request and wait if the reconstruction
1496          * thread has put up a bar for new requests.
1497          * Continue immediately if no resync is active currently.
1498          */
1499         if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1500                                 bio->bi_opf & REQ_NOWAIT)) {
1501                 bio_wouldblock_error(bio);
1502                 return;
1503         }
1504
1505         if (!wait_blocked_rdev(mddev, bio)) {
1506                 bio_wouldblock_error(bio);
1507                 return;
1508         }
1509
1510         r1_bio = alloc_r1bio(mddev, bio);
1511         r1_bio->sectors = max_write_sectors;
1512
1513         /* first select target devices under rcu_lock and
1514          * inc refcount on their rdev.  Record them by setting
1515          * bios[x] to bio
1516          * If there are known/acknowledged bad blocks on any device on
1517          * which we have seen a write error, we want to avoid writing those
1518          * blocks.
1519          * This potentially requires several writes to write around
1520          * the bad blocks.  Each set of writes gets it's own r1bio
1521          * with a set of bios attached.
1522          */
1523
1524         disks = conf->raid_disks * 2;
1525         max_sectors = r1_bio->sectors;
1526         for (i = 0;  i < disks; i++) {
1527                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1528
1529                 /*
1530                  * The write-behind io is only attempted on drives marked as
1531                  * write-mostly, which means we could allocate write behind
1532                  * bio later.
1533                  */
1534                 if (!is_discard && rdev && test_bit(WriteMostly, &rdev->flags))
1535                         write_behind = true;
1536
1537                 r1_bio->bios[i] = NULL;
1538                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1539                         if (i < conf->raid_disks)
1540                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1541                         continue;
1542                 }
1543
1544                 atomic_inc(&rdev->nr_pending);
1545                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1546                         sector_t first_bad;
1547                         int bad_sectors;
1548                         int is_bad;
1549
1550                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1551                                              &first_bad, &bad_sectors);
1552                         if (is_bad && first_bad <= r1_bio->sector) {
1553                                 /* Cannot write here at all */
1554                                 bad_sectors -= (r1_bio->sector - first_bad);
1555                                 if (bad_sectors < max_sectors)
1556                                         /* mustn't write more than bad_sectors
1557                                          * to other devices yet
1558                                          */
1559                                         max_sectors = bad_sectors;
1560                                 rdev_dec_pending(rdev, mddev);
1561                                 /* We don't set R1BIO_Degraded as that
1562                                  * only applies if the disk is
1563                                  * missing, so it might be re-added,
1564                                  * and we want to know to recover this
1565                                  * chunk.
1566                                  * In this case the device is here,
1567                                  * and the fact that this chunk is not
1568                                  * in-sync is recorded in the bad
1569                                  * block log
1570                                  */
1571                                 continue;
1572                         }
1573                         if (is_bad) {
1574                                 int good_sectors;
1575
1576                                 /*
1577                                  * We cannot atomically write this, so just
1578                                  * error in that case. It could be possible to
1579                                  * atomically write other mirrors, but the
1580                                  * complexity of supporting that is not worth
1581                                  * the benefit.
1582                                  */
1583                                 if (bio->bi_opf & REQ_ATOMIC) {
1584                                         error = -EIO;
1585                                         goto err_handle;
1586                                 }
1587
1588                                 good_sectors = first_bad - r1_bio->sector;
1589                                 if (good_sectors < max_sectors)
1590                                         max_sectors = good_sectors;
1591                         }
1592                 }
1593                 r1_bio->bios[i] = bio;
1594         }
1595
1596         /*
1597          * When using a bitmap, we may call alloc_behind_master_bio below.
1598          * alloc_behind_master_bio allocates a copy of the data payload a page
1599          * at a time and thus needs a new bio that can fit the whole payload
1600          * this bio in page sized chunks.
1601          */
1602         if (write_behind && mddev->bitmap)
1603                 max_sectors = min_t(int, max_sectors,
1604                                     BIO_MAX_VECS * (PAGE_SIZE >> 9));
1605         if (max_sectors < bio_sectors(bio)) {
1606                 struct bio *split = bio_split(bio, max_sectors,
1607                                               GFP_NOIO, &conf->bio_split);
1608
1609                 if (IS_ERR(split)) {
1610                         error = PTR_ERR(split);
1611                         goto err_handle;
1612                 }
1613                 bio_chain(split, bio);
1614                 submit_bio_noacct(bio);
1615                 bio = split;
1616                 r1_bio->master_bio = bio;
1617                 r1_bio->sectors = max_sectors;
1618         }
1619
1620         md_account_bio(mddev, &bio);
1621         r1_bio->master_bio = bio;
1622         atomic_set(&r1_bio->remaining, 1);
1623         atomic_set(&r1_bio->behind_remaining, 0);
1624
1625         first_clone = 1;
1626
1627         for (i = 0; i < disks; i++) {
1628                 struct bio *mbio = NULL;
1629                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1630                 if (!r1_bio->bios[i])
1631                         continue;
1632
1633                 if (first_clone) {
1634                         unsigned long max_write_behind =
1635                                 mddev->bitmap_info.max_write_behind;
1636                         struct md_bitmap_stats stats;
1637                         int err;
1638
1639                         /* do behind I/O ?
1640                          * Not if there are too many, or cannot
1641                          * allocate memory, or a reader on WriteMostly
1642                          * is waiting for behind writes to flush */
1643                         err = mddev->bitmap_ops->get_stats(mddev->bitmap, &stats);
1644                         if (!err && write_behind && !stats.behind_wait &&
1645                             stats.behind_writes < max_write_behind)
1646                                 alloc_behind_master_bio(r1_bio, bio);
1647
1648                         mddev->bitmap_ops->startwrite(
1649                                 mddev, r1_bio->sector, r1_bio->sectors,
1650                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
1651                         first_clone = 0;
1652                 }
1653
1654                 if (r1_bio->behind_master_bio) {
1655                         mbio = bio_alloc_clone(rdev->bdev,
1656                                                r1_bio->behind_master_bio,
1657                                                GFP_NOIO, &mddev->bio_set);
1658                         if (test_bit(CollisionCheck, &rdev->flags))
1659                                 wait_for_serialization(rdev, r1_bio);
1660                         if (test_bit(WriteMostly, &rdev->flags))
1661                                 atomic_inc(&r1_bio->behind_remaining);
1662                 } else {
1663                         mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1664                                                &mddev->bio_set);
1665
1666                         if (mddev->serialize_policy)
1667                                 wait_for_serialization(rdev, r1_bio);
1668                 }
1669
1670                 r1_bio->bios[i] = mbio;
1671
1672                 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1673                 mbio->bi_end_io = raid1_end_write_request;
1674                 mbio->bi_opf = bio_op(bio) |
1675                         (bio->bi_opf & (REQ_SYNC | REQ_FUA | REQ_ATOMIC));
1676                 if (test_bit(FailFast, &rdev->flags) &&
1677                     !test_bit(WriteMostly, &rdev->flags) &&
1678                     conf->raid_disks - mddev->degraded > 1)
1679                         mbio->bi_opf |= MD_FAILFAST;
1680                 mbio->bi_private = r1_bio;
1681
1682                 atomic_inc(&r1_bio->remaining);
1683                 mddev_trace_remap(mddev, mbio, r1_bio->sector);
1684                 /* flush_pending_writes() needs access to the rdev so...*/
1685                 mbio->bi_bdev = (void *)rdev;
1686                 if (!raid1_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) {
1687                         spin_lock_irqsave(&conf->device_lock, flags);
1688                         bio_list_add(&conf->pending_bio_list, mbio);
1689                         spin_unlock_irqrestore(&conf->device_lock, flags);
1690                         md_wakeup_thread(mddev->thread);
1691                 }
1692         }
1693
1694         r1_bio_write_done(r1_bio);
1695
1696         /* In case raid1d snuck in to freeze_array */
1697         wake_up_barrier(conf);
1698         return;
1699 err_handle:
1700         for (k = 0; k < i; k++) {
1701                 if (r1_bio->bios[k]) {
1702                         rdev_dec_pending(conf->mirrors[k].rdev, mddev);
1703                         r1_bio->bios[k] = NULL;
1704                 }
1705         }
1706
1707         bio->bi_status = errno_to_blk_status(error);
1708         set_bit(R1BIO_Uptodate, &r1_bio->state);
1709         raid_end_bio_io(r1_bio);
1710 }
1711
1712 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1713 {
1714         sector_t sectors;
1715
1716         if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1717             && md_flush_request(mddev, bio))
1718                 return true;
1719
1720         /*
1721          * There is a limit to the maximum size, but
1722          * the read/write handler might find a lower limit
1723          * due to bad blocks.  To avoid multiple splits,
1724          * we pass the maximum number of sectors down
1725          * and let the lower level perform the split.
1726          */
1727         sectors = align_to_barrier_unit_end(
1728                 bio->bi_iter.bi_sector, bio_sectors(bio));
1729
1730         if (bio_data_dir(bio) == READ)
1731                 raid1_read_request(mddev, bio, sectors, NULL);
1732         else {
1733                 md_write_start(mddev,bio);
1734                 raid1_write_request(mddev, bio, sectors);
1735         }
1736         return true;
1737 }
1738
1739 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1740 {
1741         struct r1conf *conf = mddev->private;
1742         int i;
1743
1744         lockdep_assert_held(&mddev->lock);
1745
1746         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1747                    conf->raid_disks - mddev->degraded);
1748         for (i = 0; i < conf->raid_disks; i++) {
1749                 struct md_rdev *rdev = READ_ONCE(conf->mirrors[i].rdev);
1750
1751                 seq_printf(seq, "%s",
1752                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1753         }
1754         seq_printf(seq, "]");
1755 }
1756
1757 /**
1758  * raid1_error() - RAID1 error handler.
1759  * @mddev: affected md device.
1760  * @rdev: member device to fail.
1761  *
1762  * The routine acknowledges &rdev failure and determines new @mddev state.
1763  * If it failed, then:
1764  *      - &MD_BROKEN flag is set in &mddev->flags.
1765  *      - recovery is disabled.
1766  * Otherwise, it must be degraded:
1767  *      - recovery is interrupted.
1768  *      - &mddev->degraded is bumped.
1769  *
1770  * @rdev is marked as &Faulty excluding case when array is failed and
1771  * &mddev->fail_last_dev is off.
1772  */
1773 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1774 {
1775         struct r1conf *conf = mddev->private;
1776         unsigned long flags;
1777
1778         spin_lock_irqsave(&conf->device_lock, flags);
1779
1780         if (test_bit(In_sync, &rdev->flags) &&
1781             (conf->raid_disks - mddev->degraded) == 1) {
1782                 set_bit(MD_BROKEN, &mddev->flags);
1783
1784                 if (!mddev->fail_last_dev) {
1785                         conf->recovery_disabled = mddev->recovery_disabled;
1786                         spin_unlock_irqrestore(&conf->device_lock, flags);
1787                         return;
1788                 }
1789         }
1790         set_bit(Blocked, &rdev->flags);
1791         if (test_and_clear_bit(In_sync, &rdev->flags))
1792                 mddev->degraded++;
1793         set_bit(Faulty, &rdev->flags);
1794         spin_unlock_irqrestore(&conf->device_lock, flags);
1795         /*
1796          * if recovery is running, make sure it aborts.
1797          */
1798         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1799         set_mask_bits(&mddev->sb_flags, 0,
1800                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1801         pr_crit("md/raid1:%s: Disk failure on %pg, disabling device.\n"
1802                 "md/raid1:%s: Operation continuing on %d devices.\n",
1803                 mdname(mddev), rdev->bdev,
1804                 mdname(mddev), conf->raid_disks - mddev->degraded);
1805 }
1806
1807 static void print_conf(struct r1conf *conf)
1808 {
1809         int i;
1810
1811         pr_debug("RAID1 conf printout:\n");
1812         if (!conf) {
1813                 pr_debug("(!conf)\n");
1814                 return;
1815         }
1816         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1817                  conf->raid_disks);
1818
1819         lockdep_assert_held(&conf->mddev->reconfig_mutex);
1820         for (i = 0; i < conf->raid_disks; i++) {
1821                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1822                 if (rdev)
1823                         pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
1824                                  i, !test_bit(In_sync, &rdev->flags),
1825                                  !test_bit(Faulty, &rdev->flags),
1826                                  rdev->bdev);
1827         }
1828 }
1829
1830 static void close_sync(struct r1conf *conf)
1831 {
1832         int idx;
1833
1834         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1835                 _wait_barrier(conf, idx, false);
1836                 _allow_barrier(conf, idx);
1837         }
1838
1839         mempool_exit(&conf->r1buf_pool);
1840 }
1841
1842 static int raid1_spare_active(struct mddev *mddev)
1843 {
1844         int i;
1845         struct r1conf *conf = mddev->private;
1846         int count = 0;
1847         unsigned long flags;
1848
1849         /*
1850          * Find all failed disks within the RAID1 configuration
1851          * and mark them readable.
1852          * Called under mddev lock, so rcu protection not needed.
1853          * device_lock used to avoid races with raid1_end_read_request
1854          * which expects 'In_sync' flags and ->degraded to be consistent.
1855          */
1856         spin_lock_irqsave(&conf->device_lock, flags);
1857         for (i = 0; i < conf->raid_disks; i++) {
1858                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1859                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1860                 if (repl
1861                     && !test_bit(Candidate, &repl->flags)
1862                     && repl->recovery_offset == MaxSector
1863                     && !test_bit(Faulty, &repl->flags)
1864                     && !test_and_set_bit(In_sync, &repl->flags)) {
1865                         /* replacement has just become active */
1866                         if (!rdev ||
1867                             !test_and_clear_bit(In_sync, &rdev->flags))
1868                                 count++;
1869                         if (rdev) {
1870                                 /* Replaced device not technically
1871                                  * faulty, but we need to be sure
1872                                  * it gets removed and never re-added
1873                                  */
1874                                 set_bit(Faulty, &rdev->flags);
1875                                 sysfs_notify_dirent_safe(
1876                                         rdev->sysfs_state);
1877                         }
1878                 }
1879                 if (rdev
1880                     && rdev->recovery_offset == MaxSector
1881                     && !test_bit(Faulty, &rdev->flags)
1882                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1883                         count++;
1884                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1885                 }
1886         }
1887         mddev->degraded -= count;
1888         spin_unlock_irqrestore(&conf->device_lock, flags);
1889
1890         print_conf(conf);
1891         return count;
1892 }
1893
1894 static bool raid1_add_conf(struct r1conf *conf, struct md_rdev *rdev, int disk,
1895                            bool replacement)
1896 {
1897         struct raid1_info *info = conf->mirrors + disk;
1898
1899         if (replacement)
1900                 info += conf->raid_disks;
1901
1902         if (info->rdev)
1903                 return false;
1904
1905         if (bdev_nonrot(rdev->bdev)) {
1906                 set_bit(Nonrot, &rdev->flags);
1907                 WRITE_ONCE(conf->nonrot_disks, conf->nonrot_disks + 1);
1908         }
1909
1910         rdev->raid_disk = disk;
1911         info->head_position = 0;
1912         info->seq_start = MaxSector;
1913         WRITE_ONCE(info->rdev, rdev);
1914
1915         return true;
1916 }
1917
1918 static bool raid1_remove_conf(struct r1conf *conf, int disk)
1919 {
1920         struct raid1_info *info = conf->mirrors + disk;
1921         struct md_rdev *rdev = info->rdev;
1922
1923         if (!rdev || test_bit(In_sync, &rdev->flags) ||
1924             atomic_read(&rdev->nr_pending))
1925                 return false;
1926
1927         /* Only remove non-faulty devices if recovery is not possible. */
1928         if (!test_bit(Faulty, &rdev->flags) &&
1929             rdev->mddev->recovery_disabled != conf->recovery_disabled &&
1930             rdev->mddev->degraded < conf->raid_disks)
1931                 return false;
1932
1933         if (test_and_clear_bit(Nonrot, &rdev->flags))
1934                 WRITE_ONCE(conf->nonrot_disks, conf->nonrot_disks - 1);
1935
1936         WRITE_ONCE(info->rdev, NULL);
1937         return true;
1938 }
1939
1940 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1941 {
1942         struct r1conf *conf = mddev->private;
1943         int err = -EEXIST;
1944         int mirror = 0, repl_slot = -1;
1945         struct raid1_info *p;
1946         int first = 0;
1947         int last = conf->raid_disks - 1;
1948
1949         if (mddev->recovery_disabled == conf->recovery_disabled)
1950                 return -EBUSY;
1951
1952         if (rdev->raid_disk >= 0)
1953                 first = last = rdev->raid_disk;
1954
1955         /*
1956          * find the disk ... but prefer rdev->saved_raid_disk
1957          * if possible.
1958          */
1959         if (rdev->saved_raid_disk >= 0 &&
1960             rdev->saved_raid_disk >= first &&
1961             rdev->saved_raid_disk < conf->raid_disks &&
1962             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1963                 first = last = rdev->saved_raid_disk;
1964
1965         for (mirror = first; mirror <= last; mirror++) {
1966                 p = conf->mirrors + mirror;
1967                 if (!p->rdev) {
1968                         err = mddev_stack_new_rdev(mddev, rdev);
1969                         if (err)
1970                                 return err;
1971
1972                         raid1_add_conf(conf, rdev, mirror, false);
1973                         /* As all devices are equivalent, we don't need a full recovery
1974                          * if this was recently any drive of the array
1975                          */
1976                         if (rdev->saved_raid_disk < 0)
1977                                 conf->fullsync = 1;
1978                         break;
1979                 }
1980                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1981                     p[conf->raid_disks].rdev == NULL && repl_slot < 0)
1982                         repl_slot = mirror;
1983         }
1984
1985         if (err && repl_slot >= 0) {
1986                 /* Add this device as a replacement */
1987                 clear_bit(In_sync, &rdev->flags);
1988                 set_bit(Replacement, &rdev->flags);
1989                 raid1_add_conf(conf, rdev, repl_slot, true);
1990                 err = 0;
1991                 conf->fullsync = 1;
1992         }
1993
1994         print_conf(conf);
1995         return err;
1996 }
1997
1998 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1999 {
2000         struct r1conf *conf = mddev->private;
2001         int err = 0;
2002         int number = rdev->raid_disk;
2003         struct raid1_info *p = conf->mirrors + number;
2004
2005         if (unlikely(number >= conf->raid_disks))
2006                 goto abort;
2007
2008         if (rdev != p->rdev) {
2009                 number += conf->raid_disks;
2010                 p = conf->mirrors + number;
2011         }
2012
2013         print_conf(conf);
2014         if (rdev == p->rdev) {
2015                 if (!raid1_remove_conf(conf, number)) {
2016                         err = -EBUSY;
2017                         goto abort;
2018                 }
2019
2020                 if (number < conf->raid_disks &&
2021                     conf->mirrors[conf->raid_disks + number].rdev) {
2022                         /* We just removed a device that is being replaced.
2023                          * Move down the replacement.  We drain all IO before
2024                          * doing this to avoid confusion.
2025                          */
2026                         struct md_rdev *repl =
2027                                 conf->mirrors[conf->raid_disks + number].rdev;
2028                         freeze_array(conf, 0);
2029                         if (atomic_read(&repl->nr_pending)) {
2030                                 /* It means that some queued IO of retry_list
2031                                  * hold repl. Thus, we cannot set replacement
2032                                  * as NULL, avoiding rdev NULL pointer
2033                                  * dereference in sync_request_write and
2034                                  * handle_write_finished.
2035                                  */
2036                                 err = -EBUSY;
2037                                 unfreeze_array(conf);
2038                                 goto abort;
2039                         }
2040                         clear_bit(Replacement, &repl->flags);
2041                         WRITE_ONCE(p->rdev, repl);
2042                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
2043                         unfreeze_array(conf);
2044                 }
2045
2046                 clear_bit(WantReplacement, &rdev->flags);
2047                 err = md_integrity_register(mddev);
2048         }
2049 abort:
2050
2051         print_conf(conf);
2052         return err;
2053 }
2054
2055 static void end_sync_read(struct bio *bio)
2056 {
2057         struct r1bio *r1_bio = get_resync_r1bio(bio);
2058
2059         update_head_pos(r1_bio->read_disk, r1_bio);
2060
2061         /*
2062          * we have read a block, now it needs to be re-written,
2063          * or re-read if the read failed.
2064          * We don't do much here, just schedule handling by raid1d
2065          */
2066         if (!bio->bi_status)
2067                 set_bit(R1BIO_Uptodate, &r1_bio->state);
2068
2069         if (atomic_dec_and_test(&r1_bio->remaining))
2070                 reschedule_retry(r1_bio);
2071 }
2072
2073 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
2074 {
2075         sector_t sync_blocks = 0;
2076         sector_t s = r1_bio->sector;
2077         long sectors_to_go = r1_bio->sectors;
2078
2079         /* make sure these bits don't get cleared. */
2080         do {
2081                 mddev->bitmap_ops->end_sync(mddev, s, &sync_blocks);
2082                 s += sync_blocks;
2083                 sectors_to_go -= sync_blocks;
2084         } while (sectors_to_go > 0);
2085 }
2086
2087 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
2088 {
2089         if (atomic_dec_and_test(&r1_bio->remaining)) {
2090                 struct mddev *mddev = r1_bio->mddev;
2091                 int s = r1_bio->sectors;
2092
2093                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2094                     test_bit(R1BIO_WriteError, &r1_bio->state))
2095                         reschedule_retry(r1_bio);
2096                 else {
2097                         put_buf(r1_bio);
2098                         md_done_sync(mddev, s, uptodate);
2099                 }
2100         }
2101 }
2102
2103 static void end_sync_write(struct bio *bio)
2104 {
2105         int uptodate = !bio->bi_status;
2106         struct r1bio *r1_bio = get_resync_r1bio(bio);
2107         struct mddev *mddev = r1_bio->mddev;
2108         struct r1conf *conf = mddev->private;
2109         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
2110
2111         if (!uptodate) {
2112                 abort_sync_write(mddev, r1_bio);
2113                 set_bit(WriteErrorSeen, &rdev->flags);
2114                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2115                         set_bit(MD_RECOVERY_NEEDED, &
2116                                 mddev->recovery);
2117                 set_bit(R1BIO_WriteError, &r1_bio->state);
2118         } else if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors) &&
2119                    !rdev_has_badblock(conf->mirrors[r1_bio->read_disk].rdev,
2120                                       r1_bio->sector, r1_bio->sectors)) {
2121                 set_bit(R1BIO_MadeGood, &r1_bio->state);
2122         }
2123
2124         put_sync_write_buf(r1_bio, uptodate);
2125 }
2126
2127 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
2128                            int sectors, struct page *page, blk_opf_t rw)
2129 {
2130         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2131                 /* success */
2132                 return 1;
2133         if (rw == REQ_OP_WRITE) {
2134                 set_bit(WriteErrorSeen, &rdev->flags);
2135                 if (!test_and_set_bit(WantReplacement,
2136                                       &rdev->flags))
2137                         set_bit(MD_RECOVERY_NEEDED, &
2138                                 rdev->mddev->recovery);
2139         }
2140         /* need to record an error - either for the block or the device */
2141         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2142                 md_error(rdev->mddev, rdev);
2143         return 0;
2144 }
2145
2146 static int fix_sync_read_error(struct r1bio *r1_bio)
2147 {
2148         /* Try some synchronous reads of other devices to get
2149          * good data, much like with normal read errors.  Only
2150          * read into the pages we already have so we don't
2151          * need to re-issue the read request.
2152          * We don't need to freeze the array, because being in an
2153          * active sync request, there is no normal IO, and
2154          * no overlapping syncs.
2155          * We don't need to check is_badblock() again as we
2156          * made sure that anything with a bad block in range
2157          * will have bi_end_io clear.
2158          */
2159         struct mddev *mddev = r1_bio->mddev;
2160         struct r1conf *conf = mddev->private;
2161         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2162         struct page **pages = get_resync_pages(bio)->pages;
2163         sector_t sect = r1_bio->sector;
2164         int sectors = r1_bio->sectors;
2165         int idx = 0;
2166         struct md_rdev *rdev;
2167
2168         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2169         if (test_bit(FailFast, &rdev->flags)) {
2170                 /* Don't try recovering from here - just fail it
2171                  * ... unless it is the last working device of course */
2172                 md_error(mddev, rdev);
2173                 if (test_bit(Faulty, &rdev->flags))
2174                         /* Don't try to read from here, but make sure
2175                          * put_buf does it's thing
2176                          */
2177                         bio->bi_end_io = end_sync_write;
2178         }
2179
2180         while(sectors) {
2181                 int s = sectors;
2182                 int d = r1_bio->read_disk;
2183                 int success = 0;
2184                 int start;
2185
2186                 if (s > (PAGE_SIZE>>9))
2187                         s = PAGE_SIZE >> 9;
2188                 do {
2189                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2190                                 /* No rcu protection needed here devices
2191                                  * can only be removed when no resync is
2192                                  * active, and resync is currently active
2193                                  */
2194                                 rdev = conf->mirrors[d].rdev;
2195                                 if (sync_page_io(rdev, sect, s<<9,
2196                                                  pages[idx],
2197                                                  REQ_OP_READ, false)) {
2198                                         success = 1;
2199                                         break;
2200                                 }
2201                         }
2202                         d++;
2203                         if (d == conf->raid_disks * 2)
2204                                 d = 0;
2205                 } while (!success && d != r1_bio->read_disk);
2206
2207                 if (!success) {
2208                         int abort = 0;
2209                         /* Cannot read from anywhere, this block is lost.
2210                          * Record a bad block on each device.  If that doesn't
2211                          * work just disable and interrupt the recovery.
2212                          * Don't fail devices as that won't really help.
2213                          */
2214                         pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2215                                             mdname(mddev), bio->bi_bdev,
2216                                             (unsigned long long)r1_bio->sector);
2217                         for (d = 0; d < conf->raid_disks * 2; d++) {
2218                                 rdev = conf->mirrors[d].rdev;
2219                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2220                                         continue;
2221                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2222                                         abort = 1;
2223                         }
2224                         if (abort) {
2225                                 conf->recovery_disabled =
2226                                         mddev->recovery_disabled;
2227                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2228                                 md_done_sync(mddev, r1_bio->sectors, 0);
2229                                 put_buf(r1_bio);
2230                                 return 0;
2231                         }
2232                         /* Try next page */
2233                         sectors -= s;
2234                         sect += s;
2235                         idx++;
2236                         continue;
2237                 }
2238
2239                 start = d;
2240                 /* write it back and re-read */
2241                 while (d != r1_bio->read_disk) {
2242                         if (d == 0)
2243                                 d = conf->raid_disks * 2;
2244                         d--;
2245                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2246                                 continue;
2247                         rdev = conf->mirrors[d].rdev;
2248                         if (r1_sync_page_io(rdev, sect, s,
2249                                             pages[idx],
2250                                             REQ_OP_WRITE) == 0) {
2251                                 r1_bio->bios[d]->bi_end_io = NULL;
2252                                 rdev_dec_pending(rdev, mddev);
2253                         }
2254                 }
2255                 d = start;
2256                 while (d != r1_bio->read_disk) {
2257                         if (d == 0)
2258                                 d = conf->raid_disks * 2;
2259                         d--;
2260                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2261                                 continue;
2262                         rdev = conf->mirrors[d].rdev;
2263                         if (r1_sync_page_io(rdev, sect, s,
2264                                             pages[idx],
2265                                             REQ_OP_READ) != 0)
2266                                 atomic_add(s, &rdev->corrected_errors);
2267                 }
2268                 sectors -= s;
2269                 sect += s;
2270                 idx ++;
2271         }
2272         set_bit(R1BIO_Uptodate, &r1_bio->state);
2273         bio->bi_status = 0;
2274         return 1;
2275 }
2276
2277 static void process_checks(struct r1bio *r1_bio)
2278 {
2279         /* We have read all readable devices.  If we haven't
2280          * got the block, then there is no hope left.
2281          * If we have, then we want to do a comparison
2282          * and skip the write if everything is the same.
2283          * If any blocks failed to read, then we need to
2284          * attempt an over-write
2285          */
2286         struct mddev *mddev = r1_bio->mddev;
2287         struct r1conf *conf = mddev->private;
2288         int primary;
2289         int i;
2290         int vcnt;
2291
2292         /* Fix variable parts of all bios */
2293         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2294         for (i = 0; i < conf->raid_disks * 2; i++) {
2295                 blk_status_t status;
2296                 struct bio *b = r1_bio->bios[i];
2297                 struct resync_pages *rp = get_resync_pages(b);
2298                 if (b->bi_end_io != end_sync_read)
2299                         continue;
2300                 /* fixup the bio for reuse, but preserve errno */
2301                 status = b->bi_status;
2302                 bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2303                 b->bi_status = status;
2304                 b->bi_iter.bi_sector = r1_bio->sector +
2305                         conf->mirrors[i].rdev->data_offset;
2306                 b->bi_end_io = end_sync_read;
2307                 rp->raid_bio = r1_bio;
2308                 b->bi_private = rp;
2309
2310                 /* initialize bvec table again */
2311                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2312         }
2313         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2314                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2315                     !r1_bio->bios[primary]->bi_status) {
2316                         r1_bio->bios[primary]->bi_end_io = NULL;
2317                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2318                         break;
2319                 }
2320         r1_bio->read_disk = primary;
2321         for (i = 0; i < conf->raid_disks * 2; i++) {
2322                 int j = 0;
2323                 struct bio *pbio = r1_bio->bios[primary];
2324                 struct bio *sbio = r1_bio->bios[i];
2325                 blk_status_t status = sbio->bi_status;
2326                 struct page **ppages = get_resync_pages(pbio)->pages;
2327                 struct page **spages = get_resync_pages(sbio)->pages;
2328                 struct bio_vec *bi;
2329                 int page_len[RESYNC_PAGES] = { 0 };
2330                 struct bvec_iter_all iter_all;
2331
2332                 if (sbio->bi_end_io != end_sync_read)
2333                         continue;
2334                 /* Now we can 'fixup' the error value */
2335                 sbio->bi_status = 0;
2336
2337                 bio_for_each_segment_all(bi, sbio, iter_all)
2338                         page_len[j++] = bi->bv_len;
2339
2340                 if (!status) {
2341                         for (j = vcnt; j-- ; ) {
2342                                 if (memcmp(page_address(ppages[j]),
2343                                            page_address(spages[j]),
2344                                            page_len[j]))
2345                                         break;
2346                         }
2347                 } else
2348                         j = 0;
2349                 if (j >= 0)
2350                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2351                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2352                               && !status)) {
2353                         /* No need to write to this device. */
2354                         sbio->bi_end_io = NULL;
2355                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2356                         continue;
2357                 }
2358
2359                 bio_copy_data(sbio, pbio);
2360         }
2361 }
2362
2363 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2364 {
2365         struct r1conf *conf = mddev->private;
2366         int i;
2367         int disks = conf->raid_disks * 2;
2368         struct bio *wbio;
2369
2370         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2371                 /* ouch - failed to read all of that. */
2372                 if (!fix_sync_read_error(r1_bio))
2373                         return;
2374
2375         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2376                 process_checks(r1_bio);
2377
2378         /*
2379          * schedule writes
2380          */
2381         atomic_set(&r1_bio->remaining, 1);
2382         for (i = 0; i < disks ; i++) {
2383                 wbio = r1_bio->bios[i];
2384                 if (wbio->bi_end_io == NULL ||
2385                     (wbio->bi_end_io == end_sync_read &&
2386                      (i == r1_bio->read_disk ||
2387                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2388                         continue;
2389                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2390                         abort_sync_write(mddev, r1_bio);
2391                         continue;
2392                 }
2393
2394                 wbio->bi_opf = REQ_OP_WRITE;
2395                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2396                         wbio->bi_opf |= MD_FAILFAST;
2397
2398                 wbio->bi_end_io = end_sync_write;
2399                 atomic_inc(&r1_bio->remaining);
2400                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2401
2402                 submit_bio_noacct(wbio);
2403         }
2404
2405         put_sync_write_buf(r1_bio, 1);
2406 }
2407
2408 /*
2409  * This is a kernel thread which:
2410  *
2411  *      1.      Retries failed read operations on working mirrors.
2412  *      2.      Updates the raid superblock when problems encounter.
2413  *      3.      Performs writes following reads for array synchronising.
2414  */
2415
2416 static void fix_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2417 {
2418         sector_t sect = r1_bio->sector;
2419         int sectors = r1_bio->sectors;
2420         int read_disk = r1_bio->read_disk;
2421         struct mddev *mddev = conf->mddev;
2422         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2423
2424         if (exceed_read_errors(mddev, rdev)) {
2425                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2426                 return;
2427         }
2428
2429         while(sectors) {
2430                 int s = sectors;
2431                 int d = read_disk;
2432                 int success = 0;
2433                 int start;
2434
2435                 if (s > (PAGE_SIZE>>9))
2436                         s = PAGE_SIZE >> 9;
2437
2438                 do {
2439                         rdev = conf->mirrors[d].rdev;
2440                         if (rdev &&
2441                             (test_bit(In_sync, &rdev->flags) ||
2442                              (!test_bit(Faulty, &rdev->flags) &&
2443                               rdev->recovery_offset >= sect + s)) &&
2444                             rdev_has_badblock(rdev, sect, s) == 0) {
2445                                 atomic_inc(&rdev->nr_pending);
2446                                 if (sync_page_io(rdev, sect, s<<9,
2447                                          conf->tmppage, REQ_OP_READ, false))
2448                                         success = 1;
2449                                 rdev_dec_pending(rdev, mddev);
2450                                 if (success)
2451                                         break;
2452                         }
2453
2454                         d++;
2455                         if (d == conf->raid_disks * 2)
2456                                 d = 0;
2457                 } while (d != read_disk);
2458
2459                 if (!success) {
2460                         /* Cannot read from anywhere - mark it bad */
2461                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2462                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2463                                 md_error(mddev, rdev);
2464                         break;
2465                 }
2466                 /* write it back and re-read */
2467                 start = d;
2468                 while (d != read_disk) {
2469                         if (d==0)
2470                                 d = conf->raid_disks * 2;
2471                         d--;
2472                         rdev = conf->mirrors[d].rdev;
2473                         if (rdev &&
2474                             !test_bit(Faulty, &rdev->flags)) {
2475                                 atomic_inc(&rdev->nr_pending);
2476                                 r1_sync_page_io(rdev, sect, s,
2477                                                 conf->tmppage, REQ_OP_WRITE);
2478                                 rdev_dec_pending(rdev, mddev);
2479                         }
2480                 }
2481                 d = start;
2482                 while (d != read_disk) {
2483                         if (d==0)
2484                                 d = conf->raid_disks * 2;
2485                         d--;
2486                         rdev = conf->mirrors[d].rdev;
2487                         if (rdev &&
2488                             !test_bit(Faulty, &rdev->flags)) {
2489                                 atomic_inc(&rdev->nr_pending);
2490                                 if (r1_sync_page_io(rdev, sect, s,
2491                                                 conf->tmppage, REQ_OP_READ)) {
2492                                         atomic_add(s, &rdev->corrected_errors);
2493                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %pg)\n",
2494                                                 mdname(mddev), s,
2495                                                 (unsigned long long)(sect +
2496                                                                      rdev->data_offset),
2497                                                 rdev->bdev);
2498                                 }
2499                                 rdev_dec_pending(rdev, mddev);
2500                         }
2501                 }
2502                 sectors -= s;
2503                 sect += s;
2504         }
2505 }
2506
2507 static int narrow_write_error(struct r1bio *r1_bio, int i)
2508 {
2509         struct mddev *mddev = r1_bio->mddev;
2510         struct r1conf *conf = mddev->private;
2511         struct md_rdev *rdev = conf->mirrors[i].rdev;
2512
2513         /* bio has the data to be written to device 'i' where
2514          * we just recently had a write error.
2515          * We repeatedly clone the bio and trim down to one block,
2516          * then try the write.  Where the write fails we record
2517          * a bad block.
2518          * It is conceivable that the bio doesn't exactly align with
2519          * blocks.  We must handle this somehow.
2520          *
2521          * We currently own a reference on the rdev.
2522          */
2523
2524         int block_sectors;
2525         sector_t sector;
2526         int sectors;
2527         int sect_to_write = r1_bio->sectors;
2528         int ok = 1;
2529
2530         if (rdev->badblocks.shift < 0)
2531                 return 0;
2532
2533         block_sectors = roundup(1 << rdev->badblocks.shift,
2534                                 bdev_logical_block_size(rdev->bdev) >> 9);
2535         sector = r1_bio->sector;
2536         sectors = ((sector + block_sectors)
2537                    & ~(sector_t)(block_sectors - 1))
2538                 - sector;
2539
2540         while (sect_to_write) {
2541                 struct bio *wbio;
2542                 if (sectors > sect_to_write)
2543                         sectors = sect_to_write;
2544                 /* Write at 'sector' for 'sectors'*/
2545
2546                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2547                         wbio = bio_alloc_clone(rdev->bdev,
2548                                                r1_bio->behind_master_bio,
2549                                                GFP_NOIO, &mddev->bio_set);
2550                 } else {
2551                         wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2552                                                GFP_NOIO, &mddev->bio_set);
2553                 }
2554
2555                 wbio->bi_opf = REQ_OP_WRITE;
2556                 wbio->bi_iter.bi_sector = r1_bio->sector;
2557                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2558
2559                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2560                 wbio->bi_iter.bi_sector += rdev->data_offset;
2561
2562                 if (submit_bio_wait(wbio) < 0)
2563                         /* failure! */
2564                         ok = rdev_set_badblocks(rdev, sector,
2565                                                 sectors, 0)
2566                                 && ok;
2567
2568                 bio_put(wbio);
2569                 sect_to_write -= sectors;
2570                 sector += sectors;
2571                 sectors = block_sectors;
2572         }
2573         return ok;
2574 }
2575
2576 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2577 {
2578         int m;
2579         int s = r1_bio->sectors;
2580         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2581                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2582                 struct bio *bio = r1_bio->bios[m];
2583                 if (bio->bi_end_io == NULL)
2584                         continue;
2585                 if (!bio->bi_status &&
2586                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2587                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2588                 }
2589                 if (bio->bi_status &&
2590                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2591                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2592                                 md_error(conf->mddev, rdev);
2593                 }
2594         }
2595         put_buf(r1_bio);
2596         md_done_sync(conf->mddev, s, 1);
2597 }
2598
2599 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2600 {
2601         int m, idx;
2602         bool fail = false;
2603
2604         for (m = 0; m < conf->raid_disks * 2 ; m++)
2605                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2606                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2607                         rdev_clear_badblocks(rdev,
2608                                              r1_bio->sector,
2609                                              r1_bio->sectors, 0);
2610                         rdev_dec_pending(rdev, conf->mddev);
2611                 } else if (r1_bio->bios[m] != NULL) {
2612                         /* This drive got a write error.  We need to
2613                          * narrow down and record precise write
2614                          * errors.
2615                          */
2616                         fail = true;
2617                         if (!narrow_write_error(r1_bio, m)) {
2618                                 md_error(conf->mddev,
2619                                          conf->mirrors[m].rdev);
2620                                 /* an I/O failed, we can't clear the bitmap */
2621                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2622                         }
2623                         rdev_dec_pending(conf->mirrors[m].rdev,
2624                                          conf->mddev);
2625                 }
2626         if (fail) {
2627                 spin_lock_irq(&conf->device_lock);
2628                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2629                 idx = sector_to_idx(r1_bio->sector);
2630                 atomic_inc(&conf->nr_queued[idx]);
2631                 spin_unlock_irq(&conf->device_lock);
2632                 /*
2633                  * In case freeze_array() is waiting for condition
2634                  * get_unqueued_pending() == extra to be true.
2635                  */
2636                 wake_up(&conf->wait_barrier);
2637                 md_wakeup_thread(conf->mddev->thread);
2638         } else {
2639                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2640                         close_write(r1_bio);
2641                 raid_end_bio_io(r1_bio);
2642         }
2643 }
2644
2645 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2646 {
2647         struct mddev *mddev = conf->mddev;
2648         struct bio *bio;
2649         struct md_rdev *rdev;
2650         sector_t sector;
2651
2652         clear_bit(R1BIO_ReadError, &r1_bio->state);
2653         /* we got a read error. Maybe the drive is bad.  Maybe just
2654          * the block and we can fix it.
2655          * We freeze all other IO, and try reading the block from
2656          * other devices.  When we find one, we re-write
2657          * and check it that fixes the read error.
2658          * This is all done synchronously while the array is
2659          * frozen
2660          */
2661
2662         bio = r1_bio->bios[r1_bio->read_disk];
2663         bio_put(bio);
2664         r1_bio->bios[r1_bio->read_disk] = NULL;
2665
2666         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2667         if (mddev->ro == 0
2668             && !test_bit(FailFast, &rdev->flags)) {
2669                 freeze_array(conf, 1);
2670                 fix_read_error(conf, r1_bio);
2671                 unfreeze_array(conf);
2672         } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2673                 md_error(mddev, rdev);
2674         } else {
2675                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2676         }
2677
2678         rdev_dec_pending(rdev, conf->mddev);
2679         sector = r1_bio->sector;
2680         bio = r1_bio->master_bio;
2681
2682         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2683         r1_bio->state = 0;
2684         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2685         allow_barrier(conf, sector);
2686 }
2687
2688 static void raid1d(struct md_thread *thread)
2689 {
2690         struct mddev *mddev = thread->mddev;
2691         struct r1bio *r1_bio;
2692         unsigned long flags;
2693         struct r1conf *conf = mddev->private;
2694         struct list_head *head = &conf->retry_list;
2695         struct blk_plug plug;
2696         int idx;
2697
2698         md_check_recovery(mddev);
2699
2700         if (!list_empty_careful(&conf->bio_end_io_list) &&
2701             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2702                 LIST_HEAD(tmp);
2703                 spin_lock_irqsave(&conf->device_lock, flags);
2704                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2705                         list_splice_init(&conf->bio_end_io_list, &tmp);
2706                 spin_unlock_irqrestore(&conf->device_lock, flags);
2707                 while (!list_empty(&tmp)) {
2708                         r1_bio = list_first_entry(&tmp, struct r1bio,
2709                                                   retry_list);
2710                         list_del(&r1_bio->retry_list);
2711                         idx = sector_to_idx(r1_bio->sector);
2712                         atomic_dec(&conf->nr_queued[idx]);
2713                         if (mddev->degraded)
2714                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2715                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2716                                 close_write(r1_bio);
2717                         raid_end_bio_io(r1_bio);
2718                 }
2719         }
2720
2721         blk_start_plug(&plug);
2722         for (;;) {
2723
2724                 flush_pending_writes(conf);
2725
2726                 spin_lock_irqsave(&conf->device_lock, flags);
2727                 if (list_empty(head)) {
2728                         spin_unlock_irqrestore(&conf->device_lock, flags);
2729                         break;
2730                 }
2731                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2732                 list_del(head->prev);
2733                 idx = sector_to_idx(r1_bio->sector);
2734                 atomic_dec(&conf->nr_queued[idx]);
2735                 spin_unlock_irqrestore(&conf->device_lock, flags);
2736
2737                 mddev = r1_bio->mddev;
2738                 conf = mddev->private;
2739                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2740                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2741                             test_bit(R1BIO_WriteError, &r1_bio->state))
2742                                 handle_sync_write_finished(conf, r1_bio);
2743                         else
2744                                 sync_request_write(mddev, r1_bio);
2745                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2746                            test_bit(R1BIO_WriteError, &r1_bio->state))
2747                         handle_write_finished(conf, r1_bio);
2748                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2749                         handle_read_error(conf, r1_bio);
2750                 else
2751                         WARN_ON_ONCE(1);
2752
2753                 cond_resched();
2754                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2755                         md_check_recovery(mddev);
2756         }
2757         blk_finish_plug(&plug);
2758 }
2759
2760 static int init_resync(struct r1conf *conf)
2761 {
2762         int buffs;
2763
2764         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2765         BUG_ON(mempool_initialized(&conf->r1buf_pool));
2766
2767         return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2768                             r1buf_pool_free, conf->poolinfo);
2769 }
2770
2771 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2772 {
2773         struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2774         struct resync_pages *rps;
2775         struct bio *bio;
2776         int i;
2777
2778         for (i = conf->poolinfo->raid_disks; i--; ) {
2779                 bio = r1bio->bios[i];
2780                 rps = bio->bi_private;
2781                 bio_reset(bio, NULL, 0);
2782                 bio->bi_private = rps;
2783         }
2784         r1bio->master_bio = NULL;
2785         return r1bio;
2786 }
2787
2788 /*
2789  * perform a "sync" on one "block"
2790  *
2791  * We need to make sure that no normal I/O request - particularly write
2792  * requests - conflict with active sync requests.
2793  *
2794  * This is achieved by tracking pending requests and a 'barrier' concept
2795  * that can be installed to exclude normal IO requests.
2796  */
2797
2798 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2799                                    sector_t max_sector, int *skipped)
2800 {
2801         struct r1conf *conf = mddev->private;
2802         struct r1bio *r1_bio;
2803         struct bio *bio;
2804         sector_t nr_sectors;
2805         int disk = -1;
2806         int i;
2807         int wonly = -1;
2808         int write_targets = 0, read_targets = 0;
2809         sector_t sync_blocks;
2810         bool still_degraded = false;
2811         int good_sectors = RESYNC_SECTORS;
2812         int min_bad = 0; /* number of sectors that are bad in all devices */
2813         int idx = sector_to_idx(sector_nr);
2814         int page_idx = 0;
2815
2816         if (!mempool_initialized(&conf->r1buf_pool))
2817                 if (init_resync(conf))
2818                         return 0;
2819
2820         if (sector_nr >= max_sector) {
2821                 /* If we aborted, we need to abort the
2822                  * sync on the 'current' bitmap chunk (there will
2823                  * only be one in raid1 resync.
2824                  * We can find the current addess in mddev->curr_resync
2825                  */
2826                 if (mddev->curr_resync < max_sector) /* aborted */
2827                         mddev->bitmap_ops->end_sync(mddev, mddev->curr_resync,
2828                                                     &sync_blocks);
2829                 else /* completed sync */
2830                         conf->fullsync = 0;
2831
2832                 mddev->bitmap_ops->close_sync(mddev);
2833                 close_sync(conf);
2834
2835                 if (mddev_is_clustered(mddev)) {
2836                         conf->cluster_sync_low = 0;
2837                         conf->cluster_sync_high = 0;
2838                 }
2839                 return 0;
2840         }
2841
2842         if (mddev->bitmap == NULL &&
2843             mddev->recovery_cp == MaxSector &&
2844             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2845             conf->fullsync == 0) {
2846                 *skipped = 1;
2847                 return max_sector - sector_nr;
2848         }
2849         /* before building a request, check if we can skip these blocks..
2850          * This call the bitmap_start_sync doesn't actually record anything
2851          */
2852         if (!mddev->bitmap_ops->start_sync(mddev, sector_nr, &sync_blocks, true) &&
2853             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2854                 /* We can skip this block, and probably several more */
2855                 *skipped = 1;
2856                 return sync_blocks;
2857         }
2858
2859         /*
2860          * If there is non-resync activity waiting for a turn, then let it
2861          * though before starting on this new sync request.
2862          */
2863         if (atomic_read(&conf->nr_waiting[idx]))
2864                 schedule_timeout_uninterruptible(1);
2865
2866         /* we are incrementing sector_nr below. To be safe, we check against
2867          * sector_nr + two times RESYNC_SECTORS
2868          */
2869
2870         mddev->bitmap_ops->cond_end_sync(mddev, sector_nr,
2871                 mddev_is_clustered(mddev) &&
2872                 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2873
2874         if (raise_barrier(conf, sector_nr))
2875                 return 0;
2876
2877         r1_bio = raid1_alloc_init_r1buf(conf);
2878
2879         /*
2880          * If we get a correctably read error during resync or recovery,
2881          * we might want to read from a different device.  So we
2882          * flag all drives that could conceivably be read from for READ,
2883          * and any others (which will be non-In_sync devices) for WRITE.
2884          * If a read fails, we try reading from something else for which READ
2885          * is OK.
2886          */
2887
2888         r1_bio->mddev = mddev;
2889         r1_bio->sector = sector_nr;
2890         r1_bio->state = 0;
2891         set_bit(R1BIO_IsSync, &r1_bio->state);
2892         /* make sure good_sectors won't go across barrier unit boundary */
2893         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2894
2895         for (i = 0; i < conf->raid_disks * 2; i++) {
2896                 struct md_rdev *rdev;
2897                 bio = r1_bio->bios[i];
2898
2899                 rdev = conf->mirrors[i].rdev;
2900                 if (rdev == NULL ||
2901                     test_bit(Faulty, &rdev->flags)) {
2902                         if (i < conf->raid_disks)
2903                                 still_degraded = true;
2904                 } else if (!test_bit(In_sync, &rdev->flags)) {
2905                         bio->bi_opf = REQ_OP_WRITE;
2906                         bio->bi_end_io = end_sync_write;
2907                         write_targets ++;
2908                 } else {
2909                         /* may need to read from here */
2910                         sector_t first_bad = MaxSector;
2911                         int bad_sectors;
2912
2913                         if (is_badblock(rdev, sector_nr, good_sectors,
2914                                         &first_bad, &bad_sectors)) {
2915                                 if (first_bad > sector_nr)
2916                                         good_sectors = first_bad - sector_nr;
2917                                 else {
2918                                         bad_sectors -= (sector_nr - first_bad);
2919                                         if (min_bad == 0 ||
2920                                             min_bad > bad_sectors)
2921                                                 min_bad = bad_sectors;
2922                                 }
2923                         }
2924                         if (sector_nr < first_bad) {
2925                                 if (test_bit(WriteMostly, &rdev->flags)) {
2926                                         if (wonly < 0)
2927                                                 wonly = i;
2928                                 } else {
2929                                         if (disk < 0)
2930                                                 disk = i;
2931                                 }
2932                                 bio->bi_opf = REQ_OP_READ;
2933                                 bio->bi_end_io = end_sync_read;
2934                                 read_targets++;
2935                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2936                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2937                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2938                                 /*
2939                                  * The device is suitable for reading (InSync),
2940                                  * but has bad block(s) here. Let's try to correct them,
2941                                  * if we are doing resync or repair. Otherwise, leave
2942                                  * this device alone for this sync request.
2943                                  */
2944                                 bio->bi_opf = REQ_OP_WRITE;
2945                                 bio->bi_end_io = end_sync_write;
2946                                 write_targets++;
2947                         }
2948                 }
2949                 if (rdev && bio->bi_end_io) {
2950                         atomic_inc(&rdev->nr_pending);
2951                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2952                         bio_set_dev(bio, rdev->bdev);
2953                         if (test_bit(FailFast, &rdev->flags))
2954                                 bio->bi_opf |= MD_FAILFAST;
2955                 }
2956         }
2957         if (disk < 0)
2958                 disk = wonly;
2959         r1_bio->read_disk = disk;
2960
2961         if (read_targets == 0 && min_bad > 0) {
2962                 /* These sectors are bad on all InSync devices, so we
2963                  * need to mark them bad on all write targets
2964                  */
2965                 int ok = 1;
2966                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2967                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2968                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2969                                 ok = rdev_set_badblocks(rdev, sector_nr,
2970                                                         min_bad, 0
2971                                         ) && ok;
2972                         }
2973                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2974                 *skipped = 1;
2975                 put_buf(r1_bio);
2976
2977                 if (!ok) {
2978                         /* Cannot record the badblocks, so need to
2979                          * abort the resync.
2980                          * If there are multiple read targets, could just
2981                          * fail the really bad ones ???
2982                          */
2983                         conf->recovery_disabled = mddev->recovery_disabled;
2984                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2985                         return 0;
2986                 } else
2987                         return min_bad;
2988
2989         }
2990         if (min_bad > 0 && min_bad < good_sectors) {
2991                 /* only resync enough to reach the next bad->good
2992                  * transition */
2993                 good_sectors = min_bad;
2994         }
2995
2996         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2997                 /* extra read targets are also write targets */
2998                 write_targets += read_targets-1;
2999
3000         if (write_targets == 0 || read_targets == 0) {
3001                 /* There is nowhere to write, so all non-sync
3002                  * drives must be failed - so we are finished
3003                  */
3004                 sector_t rv;
3005                 if (min_bad > 0)
3006                         max_sector = sector_nr + min_bad;
3007                 rv = max_sector - sector_nr;
3008                 *skipped = 1;
3009                 put_buf(r1_bio);
3010                 return rv;
3011         }
3012
3013         if (max_sector > mddev->resync_max)
3014                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3015         if (max_sector > sector_nr + good_sectors)
3016                 max_sector = sector_nr + good_sectors;
3017         nr_sectors = 0;
3018         sync_blocks = 0;
3019         do {
3020                 struct page *page;
3021                 int len = PAGE_SIZE;
3022                 if (sector_nr + (len>>9) > max_sector)
3023                         len = (max_sector - sector_nr) << 9;
3024                 if (len == 0)
3025                         break;
3026                 if (sync_blocks == 0) {
3027                         if (!mddev->bitmap_ops->start_sync(mddev, sector_nr,
3028                                                 &sync_blocks, still_degraded) &&
3029                             !conf->fullsync &&
3030                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
3031                                 break;
3032                         if ((len >> 9) > sync_blocks)
3033                                 len = sync_blocks<<9;
3034                 }
3035
3036                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
3037                         struct resync_pages *rp;
3038
3039                         bio = r1_bio->bios[i];
3040                         rp = get_resync_pages(bio);
3041                         if (bio->bi_end_io) {
3042                                 page = resync_fetch_page(rp, page_idx);
3043
3044                                 /*
3045                                  * won't fail because the vec table is big
3046                                  * enough to hold all these pages
3047                                  */
3048                                 __bio_add_page(bio, page, len, 0);
3049                         }
3050                 }
3051                 nr_sectors += len>>9;
3052                 sector_nr += len>>9;
3053                 sync_blocks -= (len>>9);
3054         } while (++page_idx < RESYNC_PAGES);
3055
3056         r1_bio->sectors = nr_sectors;
3057
3058         if (mddev_is_clustered(mddev) &&
3059                         conf->cluster_sync_high < sector_nr + nr_sectors) {
3060                 conf->cluster_sync_low = mddev->curr_resync_completed;
3061                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
3062                 /* Send resync message */
3063                 md_cluster_ops->resync_info_update(mddev,
3064                                 conf->cluster_sync_low,
3065                                 conf->cluster_sync_high);
3066         }
3067
3068         /* For a user-requested sync, we read all readable devices and do a
3069          * compare
3070          */
3071         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
3072                 atomic_set(&r1_bio->remaining, read_targets);
3073                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
3074                         bio = r1_bio->bios[i];
3075                         if (bio->bi_end_io == end_sync_read) {
3076                                 read_targets--;
3077                                 md_sync_acct_bio(bio, nr_sectors);
3078                                 if (read_targets == 1)
3079                                         bio->bi_opf &= ~MD_FAILFAST;
3080                                 submit_bio_noacct(bio);
3081                         }
3082                 }
3083         } else {
3084                 atomic_set(&r1_bio->remaining, 1);
3085                 bio = r1_bio->bios[r1_bio->read_disk];
3086                 md_sync_acct_bio(bio, nr_sectors);
3087                 if (read_targets == 1)
3088                         bio->bi_opf &= ~MD_FAILFAST;
3089                 submit_bio_noacct(bio);
3090         }
3091         return nr_sectors;
3092 }
3093
3094 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3095 {
3096         if (sectors)
3097                 return sectors;
3098
3099         return mddev->dev_sectors;
3100 }
3101
3102 static struct r1conf *setup_conf(struct mddev *mddev)
3103 {
3104         struct r1conf *conf;
3105         int i;
3106         struct raid1_info *disk;
3107         struct md_rdev *rdev;
3108         int err = -ENOMEM;
3109
3110         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
3111         if (!conf)
3112                 goto abort;
3113
3114         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
3115                                    sizeof(atomic_t), GFP_KERNEL);
3116         if (!conf->nr_pending)
3117                 goto abort;
3118
3119         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
3120                                    sizeof(atomic_t), GFP_KERNEL);
3121         if (!conf->nr_waiting)
3122                 goto abort;
3123
3124         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
3125                                   sizeof(atomic_t), GFP_KERNEL);
3126         if (!conf->nr_queued)
3127                 goto abort;
3128
3129         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
3130                                 sizeof(atomic_t), GFP_KERNEL);
3131         if (!conf->barrier)
3132                 goto abort;
3133
3134         conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3135                                             mddev->raid_disks, 2),
3136                                 GFP_KERNEL);
3137         if (!conf->mirrors)
3138                 goto abort;
3139
3140         conf->tmppage = alloc_page(GFP_KERNEL);
3141         if (!conf->tmppage)
3142                 goto abort;
3143
3144         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
3145         if (!conf->poolinfo)
3146                 goto abort;
3147         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
3148         err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
3149                            rbio_pool_free, conf->poolinfo);
3150         if (err)
3151                 goto abort;
3152
3153         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3154         if (err)
3155                 goto abort;
3156
3157         conf->poolinfo->mddev = mddev;
3158
3159         err = -EINVAL;
3160         spin_lock_init(&conf->device_lock);
3161         conf->raid_disks = mddev->raid_disks;
3162         rdev_for_each(rdev, mddev) {
3163                 int disk_idx = rdev->raid_disk;
3164
3165                 if (disk_idx >= conf->raid_disks || disk_idx < 0)
3166                         continue;
3167
3168                 if (!raid1_add_conf(conf, rdev, disk_idx,
3169                                     test_bit(Replacement, &rdev->flags)))
3170                         goto abort;
3171         }
3172         conf->mddev = mddev;
3173         INIT_LIST_HEAD(&conf->retry_list);
3174         INIT_LIST_HEAD(&conf->bio_end_io_list);
3175
3176         spin_lock_init(&conf->resync_lock);
3177         init_waitqueue_head(&conf->wait_barrier);
3178
3179         bio_list_init(&conf->pending_bio_list);
3180         conf->recovery_disabled = mddev->recovery_disabled - 1;
3181
3182         err = -EIO;
3183         for (i = 0; i < conf->raid_disks * 2; i++) {
3184
3185                 disk = conf->mirrors + i;
3186
3187                 if (i < conf->raid_disks &&
3188                     disk[conf->raid_disks].rdev) {
3189                         /* This slot has a replacement. */
3190                         if (!disk->rdev) {
3191                                 /* No original, just make the replacement
3192                                  * a recovering spare
3193                                  */
3194                                 disk->rdev =
3195                                         disk[conf->raid_disks].rdev;
3196                                 disk[conf->raid_disks].rdev = NULL;
3197                         } else if (!test_bit(In_sync, &disk->rdev->flags))
3198                                 /* Original is not in_sync - bad */
3199                                 goto abort;
3200                 }
3201
3202                 if (!disk->rdev ||
3203                     !test_bit(In_sync, &disk->rdev->flags)) {
3204                         disk->head_position = 0;
3205                         if (disk->rdev &&
3206                             (disk->rdev->saved_raid_disk < 0))
3207                                 conf->fullsync = 1;
3208                 }
3209         }
3210
3211         err = -ENOMEM;
3212         rcu_assign_pointer(conf->thread,
3213                            md_register_thread(raid1d, mddev, "raid1"));
3214         if (!conf->thread)
3215                 goto abort;
3216
3217         return conf;
3218
3219  abort:
3220         if (conf) {
3221                 mempool_exit(&conf->r1bio_pool);
3222                 kfree(conf->mirrors);
3223                 safe_put_page(conf->tmppage);
3224                 kfree(conf->poolinfo);
3225                 kfree(conf->nr_pending);
3226                 kfree(conf->nr_waiting);
3227                 kfree(conf->nr_queued);
3228                 kfree(conf->barrier);
3229                 bioset_exit(&conf->bio_split);
3230                 kfree(conf);
3231         }
3232         return ERR_PTR(err);
3233 }
3234
3235 static int raid1_set_limits(struct mddev *mddev)
3236 {
3237         struct queue_limits lim;
3238         int err;
3239
3240         md_init_stacking_limits(&lim);
3241         lim.max_write_zeroes_sectors = 0;
3242         lim.features |= BLK_FEAT_ATOMIC_WRITES_STACKED;
3243         err = mddev_stack_rdev_limits(mddev, &lim, MDDEV_STACK_INTEGRITY);
3244         if (err) {
3245                 queue_limits_cancel_update(mddev->gendisk->queue);
3246                 return err;
3247         }
3248         return queue_limits_set(mddev->gendisk->queue, &lim);
3249 }
3250
3251 static int raid1_run(struct mddev *mddev)
3252 {
3253         struct r1conf *conf;
3254         int i;
3255         int ret;
3256
3257         if (mddev->level != 1) {
3258                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3259                         mdname(mddev), mddev->level);
3260                 return -EIO;
3261         }
3262         if (mddev->reshape_position != MaxSector) {
3263                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3264                         mdname(mddev));
3265                 return -EIO;
3266         }
3267
3268         /*
3269          * copy the already verified devices into our private RAID1
3270          * bookkeeping area. [whatever we allocate in run(),
3271          * should be freed in raid1_free()]
3272          */
3273         if (mddev->private == NULL)
3274                 conf = setup_conf(mddev);
3275         else
3276                 conf = mddev->private;
3277
3278         if (IS_ERR(conf))
3279                 return PTR_ERR(conf);
3280
3281         if (!mddev_is_dm(mddev)) {
3282                 ret = raid1_set_limits(mddev);
3283                 if (ret)
3284                         return ret;
3285         }
3286
3287         mddev->degraded = 0;
3288         for (i = 0; i < conf->raid_disks; i++)
3289                 if (conf->mirrors[i].rdev == NULL ||
3290                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3291                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3292                         mddev->degraded++;
3293         /*
3294          * RAID1 needs at least one disk in active
3295          */
3296         if (conf->raid_disks - mddev->degraded < 1) {
3297                 md_unregister_thread(mddev, &conf->thread);
3298                 return -EINVAL;
3299         }
3300
3301         if (conf->raid_disks - mddev->degraded == 1)
3302                 mddev->recovery_cp = MaxSector;
3303
3304         if (mddev->recovery_cp != MaxSector)
3305                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3306                         mdname(mddev));
3307         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3308                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3309                 mddev->raid_disks);
3310
3311         /*
3312          * Ok, everything is just fine now
3313          */
3314         rcu_assign_pointer(mddev->thread, conf->thread);
3315         rcu_assign_pointer(conf->thread, NULL);
3316         mddev->private = conf;
3317         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3318
3319         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3320
3321         ret = md_integrity_register(mddev);
3322         if (ret)
3323                 md_unregister_thread(mddev, &mddev->thread);
3324         return ret;
3325 }
3326
3327 static void raid1_free(struct mddev *mddev, void *priv)
3328 {
3329         struct r1conf *conf = priv;
3330
3331         mempool_exit(&conf->r1bio_pool);
3332         kfree(conf->mirrors);
3333         safe_put_page(conf->tmppage);
3334         kfree(conf->poolinfo);
3335         kfree(conf->nr_pending);
3336         kfree(conf->nr_waiting);
3337         kfree(conf->nr_queued);
3338         kfree(conf->barrier);
3339         bioset_exit(&conf->bio_split);
3340         kfree(conf);
3341 }
3342
3343 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3344 {
3345         /* no resync is happening, and there is enough space
3346          * on all devices, so we can resize.
3347          * We need to make sure resync covers any new space.
3348          * If the array is shrinking we should possibly wait until
3349          * any io in the removed space completes, but it hardly seems
3350          * worth it.
3351          */
3352         sector_t newsize = raid1_size(mddev, sectors, 0);
3353         int ret;
3354
3355         if (mddev->external_size &&
3356             mddev->array_sectors > newsize)
3357                 return -EINVAL;
3358
3359         ret = mddev->bitmap_ops->resize(mddev, newsize, 0, false);
3360         if (ret)
3361                 return ret;
3362
3363         md_set_array_sectors(mddev, newsize);
3364         if (sectors > mddev->dev_sectors &&
3365             mddev->recovery_cp > mddev->dev_sectors) {
3366                 mddev->recovery_cp = mddev->dev_sectors;
3367                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3368         }
3369         mddev->dev_sectors = sectors;
3370         mddev->resync_max_sectors = sectors;
3371         return 0;
3372 }
3373
3374 static int raid1_reshape(struct mddev *mddev)
3375 {
3376         /* We need to:
3377          * 1/ resize the r1bio_pool
3378          * 2/ resize conf->mirrors
3379          *
3380          * We allocate a new r1bio_pool if we can.
3381          * Then raise a device barrier and wait until all IO stops.
3382          * Then resize conf->mirrors and swap in the new r1bio pool.
3383          *
3384          * At the same time, we "pack" the devices so that all the missing
3385          * devices have the higher raid_disk numbers.
3386          */
3387         mempool_t newpool, oldpool;
3388         struct pool_info *newpoolinfo;
3389         struct raid1_info *newmirrors;
3390         struct r1conf *conf = mddev->private;
3391         int cnt, raid_disks;
3392         unsigned long flags;
3393         int d, d2;
3394         int ret;
3395
3396         memset(&newpool, 0, sizeof(newpool));
3397         memset(&oldpool, 0, sizeof(oldpool));
3398
3399         /* Cannot change chunk_size, layout, or level */
3400         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3401             mddev->layout != mddev->new_layout ||
3402             mddev->level != mddev->new_level) {
3403                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3404                 mddev->new_layout = mddev->layout;
3405                 mddev->new_level = mddev->level;
3406                 return -EINVAL;
3407         }
3408
3409         if (!mddev_is_clustered(mddev))
3410                 md_allow_write(mddev);
3411
3412         raid_disks = mddev->raid_disks + mddev->delta_disks;
3413
3414         if (raid_disks < conf->raid_disks) {
3415                 cnt=0;
3416                 for (d= 0; d < conf->raid_disks; d++)
3417                         if (conf->mirrors[d].rdev)
3418                                 cnt++;
3419                 if (cnt > raid_disks)
3420                         return -EBUSY;
3421         }
3422
3423         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3424         if (!newpoolinfo)
3425                 return -ENOMEM;
3426         newpoolinfo->mddev = mddev;
3427         newpoolinfo->raid_disks = raid_disks * 2;
3428
3429         ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3430                            rbio_pool_free, newpoolinfo);
3431         if (ret) {
3432                 kfree(newpoolinfo);
3433                 return ret;
3434         }
3435         newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3436                                          raid_disks, 2),
3437                              GFP_KERNEL);
3438         if (!newmirrors) {
3439                 kfree(newpoolinfo);
3440                 mempool_exit(&newpool);
3441                 return -ENOMEM;
3442         }
3443
3444         freeze_array(conf, 0);
3445
3446         /* ok, everything is stopped */
3447         oldpool = conf->r1bio_pool;
3448         conf->r1bio_pool = newpool;
3449
3450         for (d = d2 = 0; d < conf->raid_disks; d++) {
3451                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3452                 if (rdev && rdev->raid_disk != d2) {
3453                         sysfs_unlink_rdev(mddev, rdev);
3454                         rdev->raid_disk = d2;
3455                         sysfs_unlink_rdev(mddev, rdev);
3456                         if (sysfs_link_rdev(mddev, rdev))
3457                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3458                                         mdname(mddev), rdev->raid_disk);
3459                 }
3460                 if (rdev)
3461                         newmirrors[d2++].rdev = rdev;
3462         }
3463         kfree(conf->mirrors);
3464         conf->mirrors = newmirrors;
3465         kfree(conf->poolinfo);
3466         conf->poolinfo = newpoolinfo;
3467
3468         spin_lock_irqsave(&conf->device_lock, flags);
3469         mddev->degraded += (raid_disks - conf->raid_disks);
3470         spin_unlock_irqrestore(&conf->device_lock, flags);
3471         conf->raid_disks = mddev->raid_disks = raid_disks;
3472         mddev->delta_disks = 0;
3473
3474         unfreeze_array(conf);
3475
3476         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3477         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3478         md_wakeup_thread(mddev->thread);
3479
3480         mempool_exit(&oldpool);
3481         return 0;
3482 }
3483
3484 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3485 {
3486         struct r1conf *conf = mddev->private;
3487
3488         if (quiesce)
3489                 freeze_array(conf, 0);
3490         else
3491                 unfreeze_array(conf);
3492 }
3493
3494 static void *raid1_takeover(struct mddev *mddev)
3495 {
3496         /* raid1 can take over:
3497          *  raid5 with 2 devices, any layout or chunk size
3498          */
3499         if (mddev->level == 5 && mddev->raid_disks == 2) {
3500                 struct r1conf *conf;
3501                 mddev->new_level = 1;
3502                 mddev->new_layout = 0;
3503                 mddev->new_chunk_sectors = 0;
3504                 conf = setup_conf(mddev);
3505                 if (!IS_ERR(conf)) {
3506                         /* Array must appear to be quiesced */
3507                         conf->array_frozen = 1;
3508                         mddev_clear_unsupported_flags(mddev,
3509                                 UNSUPPORTED_MDDEV_FLAGS);
3510                 }
3511                 return conf;
3512         }
3513         return ERR_PTR(-EINVAL);
3514 }
3515
3516 static struct md_personality raid1_personality =
3517 {
3518         .name           = "raid1",
3519         .level          = 1,
3520         .owner          = THIS_MODULE,
3521         .make_request   = raid1_make_request,
3522         .run            = raid1_run,
3523         .free           = raid1_free,
3524         .status         = raid1_status,
3525         .error_handler  = raid1_error,
3526         .hot_add_disk   = raid1_add_disk,
3527         .hot_remove_disk= raid1_remove_disk,
3528         .spare_active   = raid1_spare_active,
3529         .sync_request   = raid1_sync_request,
3530         .resize         = raid1_resize,
3531         .size           = raid1_size,
3532         .check_reshape  = raid1_reshape,
3533         .quiesce        = raid1_quiesce,
3534         .takeover       = raid1_takeover,
3535 };
3536
3537 static int __init raid_init(void)
3538 {
3539         return register_md_personality(&raid1_personality);
3540 }
3541
3542 static void raid_exit(void)
3543 {
3544         unregister_md_personality(&raid1_personality);
3545 }
3546
3547 module_init(raid_init);
3548 module_exit(raid_exit);
3549 MODULE_LICENSE("GPL");
3550 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3551 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3552 MODULE_ALIAS("md-raid1");
3553 MODULE_ALIAS("md-level-1");