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 <miku@iki.fi>, 2000
  12  *
  13  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  14  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  15  *
  16  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 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 raid1_log(md, fmt, args...)                             \
  50         do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
  51
  52 #include "raid1-10.c"
  53
  54 #define START(node) ((node)->start)
  55 #define LAST(node) ((node)->last)
  56 INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
  57                      START, LAST, static inline, raid1_rb);
  58
  59 static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
  60                                 struct serial_info *si, int idx)
  61 {
  62         unsigned long flags;
  63         int ret = 0;
  64         sector_t lo = r1_bio->sector;
  65         sector_t hi = lo + r1_bio->sectors;
  66         struct serial_in_rdev *serial = &rdev->serial[idx];
  67
  68         spin_lock_irqsave(&serial->serial_lock, flags);
  69         /* collision happened */
  70         if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
  71                 ret = -EBUSY;
  72         else {
  73                 si->start = lo;
  74                 si->last = hi;
  75                 raid1_rb_insert(si, &serial->serial_rb);
  76         }
  77         spin_unlock_irqrestore(&serial->serial_lock, flags);
  78
  79         return ret;
  80 }
  81
  82 static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
  83 {
  84         struct mddev *mddev = rdev->mddev;
  85         struct serial_info *si;
  86         int idx = sector_to_idx(r1_bio->sector);
  87         struct serial_in_rdev *serial = &rdev->serial[idx];
  88
  89         if (WARN_ON(!mddev->serial_info_pool))
  90                 return;
  91         si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
  92         wait_event(serial->serial_io_wait,
  93                    check_and_add_serial(rdev, r1_bio, si, idx) == 0);
  94 }
  95
  96 static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
  97 {
  98         struct serial_info *si;
  99         unsigned long flags;
 100         int found = 0;
 101         struct mddev *mddev = rdev->mddev;
 102         int idx = sector_to_idx(lo);
 103         struct serial_in_rdev *serial = &rdev->serial[idx];
 104
 105         spin_lock_irqsave(&serial->serial_lock, flags);
 106         for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
 107              si; si = raid1_rb_iter_next(si, lo, hi)) {
 108                 if (si->start == lo && si->last == hi) {
 109                         raid1_rb_remove(si, &serial->serial_rb);
 110                         mempool_free(si, mddev->serial_info_pool);
 111                         found = 1;
 112                         break;
 113                 }
 114         }
 115         if (!found)
 116                 WARN(1, "The write IO is not recorded for serialization\n");
 117         spin_unlock_irqrestore(&serial->serial_lock, flags);
 118         wake_up(&serial->serial_io_wait);
 119 }
 120
 121 /*
 122  * for resync bio, r1bio pointer can be retrieved from the per-bio
 123  * 'struct resync_pages'.
 124  */
 125 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
 126 {
 127         return get_resync_pages(bio)->raid_bio;
 128 }
 129
 130 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
 131 {
 132         struct pool_info *pi = data;
 133         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
 134
 135         /* allocate a r1bio with room for raid_disks entries in the bios array */
 136         return kzalloc(size, gfp_flags);
 137 }
 138
 139 #define RESYNC_DEPTH 32
 140 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 141 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 142 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 143 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 144 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 145
 146 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 147 {
 148         struct pool_info *pi = data;
 149         struct r1bio *r1_bio;
 150         struct bio *bio;
 151         int need_pages;
 152         int j;
 153         struct resync_pages *rps;
 154
 155         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 156         if (!r1_bio)
 157                 return NULL;
 158
 159         rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
 160                             gfp_flags);
 161         if (!rps)
 162                 goto out_free_r1bio;
 163
 164         /*
 165          * Allocate bios : 1 for reading, n-1 for writing
 166          */
 167         for (j = pi->raid_disks ; j-- ; ) {
 168                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 169                 if (!bio)
 170                         goto out_free_bio;
 171                 r1_bio->bios[j] = bio;
 172         }
 173         /*
 174          * Allocate RESYNC_PAGES data pages and attach them to
 175          * the first bio.
 176          * If this is a user-requested check/repair, allocate
 177          * RESYNC_PAGES for each bio.
 178          */
 179         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 180                 need_pages = pi->raid_disks;
 181         else
 182                 need_pages = 1;
 183         for (j = 0; j < pi->raid_disks; j++) {
 184                 struct resync_pages *rp = &rps[j];
 185
 186                 bio = r1_bio->bios[j];
 187
 188                 if (j < need_pages) {
 189                         if (resync_alloc_pages(rp, gfp_flags))
 190                                 goto out_free_pages;
 191                 } else {
 192                         memcpy(rp, &rps[0], sizeof(*rp));
 193                         resync_get_all_pages(rp);
 194                 }
 195
 196                 rp->raid_bio = r1_bio;
 197                 bio->bi_private = rp;
 198         }
 199
 200         r1_bio->master_bio = NULL;
 201
 202         return r1_bio;
 203
 204 out_free_pages:
 205         while (--j >= 0)
 206                 resync_free_pages(&rps[j]);
 207
 208 out_free_bio:
 209         while (++j < pi->raid_disks)
 210                 bio_put(r1_bio->bios[j]);
 211         kfree(rps);
 212
 213 out_free_r1bio:
 214         rbio_pool_free(r1_bio, data);
 215         return NULL;
 216 }
 217
 218 static void r1buf_pool_free(void *__r1_bio, void *data)
 219 {
 220         struct pool_info *pi = data;
 221         int i;
 222         struct r1bio *r1bio = __r1_bio;
 223         struct resync_pages *rp = NULL;
 224
 225         for (i = pi->raid_disks; i--; ) {
 226                 rp = get_resync_pages(r1bio->bios[i]);
 227                 resync_free_pages(rp);
 228                 bio_put(r1bio->bios[i]);
 229         }
 230
 231         /* resync pages array stored in the 1st bio's .bi_private */
 232         kfree(rp);
 233
 234         rbio_pool_free(r1bio, data);
 235 }
 236
 237 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 238 {
 239         int i;
 240
 241         for (i = 0; i < conf->raid_disks * 2; i++) {
 242                 struct bio **bio = r1_bio->bios + i;
 243                 if (!BIO_SPECIAL(*bio))
 244                         bio_put(*bio);
 245                 *bio = NULL;
 246         }
 247 }
 248
 249 static void free_r1bio(struct r1bio *r1_bio)
 250 {
 251         struct r1conf *conf = r1_bio->mddev->private;
 252
 253         put_all_bios(conf, r1_bio);
 254         mempool_free(r1_bio, &conf->r1bio_pool);
 255 }
 256
 257 static void put_buf(struct r1bio *r1_bio)
 258 {
 259         struct r1conf *conf = r1_bio->mddev->private;
 260         sector_t sect = r1_bio->sector;
 261         int i;
 262
 263         for (i = 0; i < conf->raid_disks * 2; i++) {
 264                 struct bio *bio = r1_bio->bios[i];
 265                 if (bio->bi_end_io)
 266                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 267         }
 268
 269         mempool_free(r1_bio, &conf->r1buf_pool);
 270
 271         lower_barrier(conf, sect);
 272 }
 273
 274 static void reschedule_retry(struct r1bio *r1_bio)
 275 {
 276         unsigned long flags;
 277         struct mddev *mddev = r1_bio->mddev;
 278         struct r1conf *conf = mddev->private;
 279         int idx;
 280
 281         idx = sector_to_idx(r1_bio->sector);
 282         spin_lock_irqsave(&conf->device_lock, flags);
 283         list_add(&r1_bio->retry_list, &conf->retry_list);
 284         atomic_inc(&conf->nr_queued[idx]);
 285         spin_unlock_irqrestore(&conf->device_lock, flags);
 286
 287         wake_up(&conf->wait_barrier);
 288         md_wakeup_thread(mddev->thread);
 289 }
 290
 291 /*
 292  * raid_end_bio_io() is called when we have finished servicing a mirrored
 293  * operation and are ready to return a success/failure code to the buffer
 294  * cache layer.
 295  */
 296 static void call_bio_endio(struct r1bio *r1_bio)
 297 {
 298         struct bio *bio = r1_bio->master_bio;
 299
 300         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 301                 bio->bi_status = BLK_STS_IOERR;
 302
 303         if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
 304                 bio_end_io_acct(bio, r1_bio->start_time);
 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
 313         /* if nobody has done the final endio yet, do it now */
 314         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 315                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 316                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 317                          (unsigned long long) bio->bi_iter.bi_sector,
 318                          (unsigned long long) bio_end_sector(bio) - 1);
 319
 320                 call_bio_endio(r1_bio);
 321         }
 322         /*
 323          * Wake up any possible resync thread that waits for the device
 324          * to go idle.  All I/Os, even write-behind writes, are done.
 325          */
 326         allow_barrier(conf, r1_bio->sector);
 327
 328         free_r1bio(r1_bio);
 329 }
 330
 331 /*
 332  * Update disk head position estimator based on IRQ completion info.
 333  */
 334 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 335 {
 336         struct r1conf *conf = r1_bio->mddev->private;
 337
 338         conf->mirrors[disk].head_position =
 339                 r1_bio->sector + (r1_bio->sectors);
 340 }
 341
 342 /*
 343  * Find the disk number which triggered given bio
 344  */
 345 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 346 {
 347         int mirror;
 348         struct r1conf *conf = r1_bio->mddev->private;
 349         int raid_disks = conf->raid_disks;
 350
 351         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 352                 if (r1_bio->bios[mirror] == bio)
 353                         break;
 354
 355         BUG_ON(mirror == raid_disks * 2);
 356         update_head_pos(mirror, r1_bio);
 357
 358         return mirror;
 359 }
 360
 361 static void raid1_end_read_request(struct bio *bio)
 362 {
 363         int uptodate = !bio->bi_status;
 364         struct r1bio *r1_bio = bio->bi_private;
 365         struct r1conf *conf = r1_bio->mddev->private;
 366         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 367
 368         /*
 369          * this branch is our 'one mirror IO has finished' event handler:
 370          */
 371         update_head_pos(r1_bio->read_disk, r1_bio);
 372
 373         if (uptodate)
 374                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 375         else if (test_bit(FailFast, &rdev->flags) &&
 376                  test_bit(R1BIO_FailFast, &r1_bio->state))
 377                 /* This was a fail-fast read so we definitely
 378                  * want to retry */
 379                 ;
 380         else {
 381                 /* If all other devices have failed, we want to return
 382                  * the error upwards rather than fail the last device.
 383                  * Here we redefine "uptodate" to mean "Don't want to retry"
 384                  */
 385                 unsigned long flags;
 386                 spin_lock_irqsave(&conf->device_lock, flags);
 387                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 388                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 389                      test_bit(In_sync, &rdev->flags)))
 390                         uptodate = 1;
 391                 spin_unlock_irqrestore(&conf->device_lock, flags);
 392         }
 393
 394         if (uptodate) {
 395                 raid_end_bio_io(r1_bio);
 396                 rdev_dec_pending(rdev, conf->mddev);
 397         } else {
 398                 /*
 399                  * oops, read error:
 400                  */
 401                 char b[BDEVNAME_SIZE];
 402                 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 403                                    mdname(conf->mddev),
 404                                    bdevname(rdev->bdev, b),
 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         /* it really is the end of this request */
 415         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 416                 bio_free_pages(r1_bio->behind_master_bio);
 417                 bio_put(r1_bio->behind_master_bio);
 418                 r1_bio->behind_master_bio = NULL;
 419         }
 420         /* clear the bitmap if all writes complete successfully */
 421         md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 422                            r1_bio->sectors,
 423                            !test_bit(R1BIO_Degraded, &r1_bio->state),
 424                            test_bit(R1BIO_BehindIO, &r1_bio->state));
 425         md_write_end(r1_bio->mddev);
 426 }
 427
 428 static void r1_bio_write_done(struct r1bio *r1_bio)
 429 {
 430         if (!atomic_dec_and_test(&r1_bio->remaining))
 431                 return;
 432
 433         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 434                 reschedule_retry(r1_bio);
 435         else {
 436                 close_write(r1_bio);
 437                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 438                         reschedule_retry(r1_bio);
 439                 else
 440                         raid_end_bio_io(r1_bio);
 441         }
 442 }
 443
 444 static void raid1_end_write_request(struct bio *bio)
 445 {
 446         struct r1bio *r1_bio = bio->bi_private;
 447         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 448         struct r1conf *conf = r1_bio->mddev->private;
 449         struct bio *to_put = NULL;
 450         int mirror = find_bio_disk(r1_bio, bio);
 451         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 452         bool discard_error;
 453         sector_t lo = r1_bio->sector;
 454         sector_t hi = r1_bio->sector + r1_bio->sectors;
 455
 456         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 457
 458         /*
 459          * 'one mirror IO has finished' event handler:
 460          */
 461         if (bio->bi_status && !discard_error) {
 462                 set_bit(WriteErrorSeen, &rdev->flags);
 463                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
 464                         set_bit(MD_RECOVERY_NEEDED, &
 465                                 conf->mddev->recovery);
 466
 467                 if (test_bit(FailFast, &rdev->flags) &&
 468                     (bio->bi_opf & MD_FAILFAST) &&
 469                     /* We never try FailFast to WriteMostly devices */
 470                     !test_bit(WriteMostly, &rdev->flags)) {
 471                         md_error(r1_bio->mddev, rdev);
 472                 }
 473
 474                 /*
 475                  * When the device is faulty, it is not necessary to
 476                  * handle write error.
 477                  */
 478                 if (!test_bit(Faulty, &rdev->flags))
 479                         set_bit(R1BIO_WriteError, &r1_bio->state);
 480                 else {
 481                         /* Fail the request */
 482                         set_bit(R1BIO_Degraded, &r1_bio->state);
 483                         /* Finished with this branch */
 484                         r1_bio->bios[mirror] = NULL;
 485                         to_put = bio;
 486                 }
 487         } else {
 488                 /*
 489                  * Set R1BIO_Uptodate in our master bio, so that we
 490                  * will return a good error code for to the higher
 491                  * levels even if IO on some other mirrored buffer
 492                  * fails.
 493                  *
 494                  * The 'master' represents the composite IO operation
 495                  * to user-side. So if something waits for IO, then it
 496                  * will wait for the 'master' bio.
 497                  */
 498                 sector_t first_bad;
 499                 int bad_sectors;
 500
 501                 r1_bio->bios[mirror] = NULL;
 502                 to_put = bio;
 503                 /*
 504                  * Do not set R1BIO_Uptodate if the current device is
 505                  * rebuilding or Faulty. This is because we cannot use
 506                  * such device for properly reading the data back (we could
 507                  * potentially use it, if the current write would have felt
 508                  * before rdev->recovery_offset, but for simplicity we don't
 509                  * check this here.
 510                  */
 511                 if (test_bit(In_sync, &rdev->flags) &&
 512                     !test_bit(Faulty, &rdev->flags))
 513                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 514
 515                 /* Maybe we can clear some bad blocks. */
 516                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 517                                 &first_bad, &bad_sectors) && !discard_error) {
 518                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 519                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 520                 }
 521         }
 522
 523         if (behind) {
 524                 if (test_bit(CollisionCheck, &rdev->flags))
 525                         remove_serial(rdev, lo, hi);
 526                 if (test_bit(WriteMostly, &rdev->flags))
 527                         atomic_dec(&r1_bio->behind_remaining);
 528
 529                 /*
 530                  * In behind mode, we ACK the master bio once the I/O
 531                  * has safely reached all non-writemostly
 532                  * disks. Setting the Returned bit ensures that this
 533                  * gets done only once -- we don't ever want to return
 534                  * -EIO here, instead we'll wait
 535                  */
 536                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 537                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 538                         /* Maybe we can return now */
 539                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 540                                 struct bio *mbio = r1_bio->master_bio;
 541                                 pr_debug("raid1: behind end write sectors"
 542                                          " %llu-%llu\n",
 543                                          (unsigned long long) mbio->bi_iter.bi_sector,
 544                                          (unsigned long long) bio_end_sector(mbio) - 1);
 545                                 call_bio_endio(r1_bio);
 546                         }
 547                 }
 548         } else if (rdev->mddev->serialize_policy)
 549                 remove_serial(rdev, lo, hi);
 550         if (r1_bio->bios[mirror] == NULL)
 551                 rdev_dec_pending(rdev, conf->mddev);
 552
 553         /*
 554          * Let's see if all mirrored write operations have finished
 555          * already.
 556          */
 557         r1_bio_write_done(r1_bio);
 558
 559         if (to_put)
 560                 bio_put(to_put);
 561 }
 562
 563 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 564                                           sector_t sectors)
 565 {
 566         sector_t len;
 567
 568         WARN_ON(sectors == 0);
 569         /*
 570          * len is the number of sectors from start_sector to end of the
 571          * barrier unit which start_sector belongs to.
 572          */
 573         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 574               start_sector;
 575
 576         if (len > sectors)
 577                 len = sectors;
 578
 579         return len;
 580 }
 581
 582 /*
 583  * This routine returns the disk from which the requested read should
 584  * be done. There is a per-array 'next expected sequential IO' sector
 585  * number - if this matches on the next IO then we use the last disk.
 586  * There is also a per-disk 'last know head position' sector that is
 587  * maintained from IRQ contexts, both the normal and the resync IO
 588  * completion handlers update this position correctly. If there is no
 589  * perfect sequential match then we pick the disk whose head is closest.
 590  *
 591  * If there are 2 mirrors in the same 2 devices, performance degrades
 592  * because position is mirror, not device based.
 593  *
 594  * The rdev for the device selected will have nr_pending incremented.
 595  */
 596 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 597 {
 598         const sector_t this_sector = r1_bio->sector;
 599         int sectors;
 600         int best_good_sectors;
 601         int best_disk, best_dist_disk, best_pending_disk;
 602         int has_nonrot_disk;
 603         int disk;
 604         sector_t best_dist;
 605         unsigned int min_pending;
 606         struct md_rdev *rdev;
 607         int choose_first;
 608         int choose_next_idle;
 609
 610         rcu_read_lock();
 611         /*
 612          * Check if we can balance. We can balance on the whole
 613          * device if no resync is going on, or below the resync window.
 614          * We take the first readable disk when above the resync window.
 615          */
 616  retry:
 617         sectors = r1_bio->sectors;
 618         best_disk = -1;
 619         best_dist_disk = -1;
 620         best_dist = MaxSector;
 621         best_pending_disk = -1;
 622         min_pending = UINT_MAX;
 623         best_good_sectors = 0;
 624         has_nonrot_disk = 0;
 625         choose_next_idle = 0;
 626         clear_bit(R1BIO_FailFast, &r1_bio->state);
 627
 628         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 629             (mddev_is_clustered(conf->mddev) &&
 630             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 631                     this_sector + sectors)))
 632                 choose_first = 1;
 633         else
 634                 choose_first = 0;
 635
 636         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 637                 sector_t dist;
 638                 sector_t first_bad;
 639                 int bad_sectors;
 640                 unsigned int pending;
 641                 bool nonrot;
 642
 643                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 644                 if (r1_bio->bios[disk] == IO_BLOCKED
 645                     || rdev == NULL
 646                     || test_bit(Faulty, &rdev->flags))
 647                         continue;
 648                 if (!test_bit(In_sync, &rdev->flags) &&
 649                     rdev->recovery_offset < this_sector + sectors)
 650                         continue;
 651                 if (test_bit(WriteMostly, &rdev->flags)) {
 652                         /* Don't balance among write-mostly, just
 653                          * use the first as a last resort */
 654                         if (best_dist_disk < 0) {
 655                                 if (is_badblock(rdev, this_sector, sectors,
 656                                                 &first_bad, &bad_sectors)) {
 657                                         if (first_bad <= this_sector)
 658                                                 /* Cannot use this */
 659                                                 continue;
 660                                         best_good_sectors = first_bad - this_sector;
 661                                 } else
 662                                         best_good_sectors = sectors;
 663                                 best_dist_disk = disk;
 664                                 best_pending_disk = disk;
 665                         }
 666                         continue;
 667                 }
 668                 /* This is a reasonable device to use.  It might
 669                  * even be best.
 670                  */
 671                 if (is_badblock(rdev, this_sector, sectors,
 672                                 &first_bad, &bad_sectors)) {
 673                         if (best_dist < MaxSector)
 674                                 /* already have a better device */
 675                                 continue;
 676                         if (first_bad <= this_sector) {
 677                                 /* cannot read here. If this is the 'primary'
 678                                  * device, then we must not read beyond
 679                                  * bad_sectors from another device..
 680                                  */
 681                                 bad_sectors -= (this_sector - first_bad);
 682                                 if (choose_first && sectors > bad_sectors)
 683                                         sectors = bad_sectors;
 684                                 if (best_good_sectors > sectors)
 685                                         best_good_sectors = sectors;
 686
 687                         } else {
 688                                 sector_t good_sectors = first_bad - this_sector;
 689                                 if (good_sectors > best_good_sectors) {
 690                                         best_good_sectors = good_sectors;
 691                                         best_disk = disk;
 692                                 }
 693                                 if (choose_first)
 694                                         break;
 695                         }
 696                         continue;
 697                 } else {
 698                         if ((sectors > best_good_sectors) && (best_disk >= 0))
 699                                 best_disk = -1;
 700                         best_good_sectors = sectors;
 701                 }
 702
 703                 if (best_disk >= 0)
 704                         /* At least two disks to choose from so failfast is OK */
 705                         set_bit(R1BIO_FailFast, &r1_bio->state);
 706
 707                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 708                 has_nonrot_disk |= nonrot;
 709                 pending = atomic_read(&rdev->nr_pending);
 710                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 711                 if (choose_first) {
 712                         best_disk = disk;
 713                         break;
 714                 }
 715                 /* Don't change to another disk for sequential reads */
 716                 if (conf->mirrors[disk].next_seq_sect == this_sector
 717                     || dist == 0) {
 718                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 719                         struct raid1_info *mirror = &conf->mirrors[disk];
 720
 721                         best_disk = disk;
 722                         /*
 723                          * If buffered sequential IO size exceeds optimal
 724                          * iosize, check if there is idle disk. If yes, choose
 725                          * the idle disk. read_balance could already choose an
 726                          * idle disk before noticing it's a sequential IO in
 727                          * this disk. This doesn't matter because this disk
 728                          * will idle, next time it will be utilized after the
 729                          * first disk has IO size exceeds optimal iosize. In
 730                          * this way, iosize of the first disk will be optimal
 731                          * iosize at least. iosize of the second disk might be
 732                          * small, but not a big deal since when the second disk
 733                          * starts IO, the first disk is likely still busy.
 734                          */
 735                         if (nonrot && opt_iosize > 0 &&
 736                             mirror->seq_start != MaxSector &&
 737                             mirror->next_seq_sect > opt_iosize &&
 738                             mirror->next_seq_sect - opt_iosize >=
 739                             mirror->seq_start) {
 740                                 choose_next_idle = 1;
 741                                 continue;
 742                         }
 743                         break;
 744                 }
 745
 746                 if (choose_next_idle)
 747                         continue;
 748
 749                 if (min_pending > pending) {
 750                         min_pending = pending;
 751                         best_pending_disk = disk;
 752                 }
 753
 754                 if (dist < best_dist) {
 755                         best_dist = dist;
 756                         best_dist_disk = disk;
 757                 }
 758         }
 759
 760         /*
 761          * If all disks are rotational, choose the closest disk. If any disk is
 762          * non-rotational, choose the disk with less pending request even the
 763          * disk is rotational, which might/might not be optimal for raids with
 764          * mixed ratation/non-rotational disks depending on workload.
 765          */
 766         if (best_disk == -1) {
 767                 if (has_nonrot_disk || min_pending == 0)
 768                         best_disk = best_pending_disk;
 769                 else
 770                         best_disk = best_dist_disk;
 771         }
 772
 773         if (best_disk >= 0) {
 774                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 775                 if (!rdev)
 776                         goto retry;
 777                 atomic_inc(&rdev->nr_pending);
 778                 sectors = best_good_sectors;
 779
 780                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 781                         conf->mirrors[best_disk].seq_start = this_sector;
 782
 783                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 784         }
 785         rcu_read_unlock();
 786         *max_sectors = sectors;
 787
 788         return best_disk;
 789 }
 790
 791 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 792 {
 793         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 794         md_bitmap_unplug(conf->mddev->bitmap);
 795         wake_up(&conf->wait_barrier);
 796
 797         while (bio) { /* submit pending writes */
 798                 struct bio *next = bio->bi_next;
 799                 struct md_rdev *rdev = (void *)bio->bi_bdev;
 800                 bio->bi_next = NULL;
 801                 bio_set_dev(bio, rdev->bdev);
 802                 if (test_bit(Faulty, &rdev->flags)) {
 803                         bio_io_error(bio);
 804                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 805                                     !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
 806                         /* Just ignore it */
 807                         bio_endio(bio);
 808                 else
 809                         submit_bio_noacct(bio);
 810                 bio = next;
 811                 cond_resched();
 812         }
 813 }
 814
 815 static void flush_pending_writes(struct r1conf *conf)
 816 {
 817         /* Any writes that have been queued but are awaiting
 818          * bitmap updates get flushed here.
 819          */
 820         spin_lock_irq(&conf->device_lock);
 821
 822         if (conf->pending_bio_list.head) {
 823                 struct blk_plug plug;
 824                 struct bio *bio;
 825
 826                 bio = bio_list_get(&conf->pending_bio_list);
 827                 conf->pending_count = 0;
 828                 spin_unlock_irq(&conf->device_lock);
 829
 830                 /*
 831                  * As this is called in a wait_event() loop (see freeze_array),
 832                  * current->state might be TASK_UNINTERRUPTIBLE which will
 833                  * cause a warning when we prepare to wait again.  As it is
 834                  * rare that this path is taken, it is perfectly safe to force
 835                  * us to go around the wait_event() loop again, so the warning
 836                  * is a false-positive.  Silence the warning by resetting
 837                  * thread state
 838                  */
 839                 __set_current_state(TASK_RUNNING);
 840                 blk_start_plug(&plug);
 841                 flush_bio_list(conf, bio);
 842                 blk_finish_plug(&plug);
 843         } else
 844                 spin_unlock_irq(&conf->device_lock);
 845 }
 846
 847 /* Barriers....
 848  * Sometimes we need to suspend IO while we do something else,
 849  * either some resync/recovery, or reconfigure the array.
 850  * To do this we raise a 'barrier'.
 851  * The 'barrier' is a counter that can be raised multiple times
 852  * to count how many activities are happening which preclude
 853  * normal IO.
 854  * We can only raise the barrier if there is no pending IO.
 855  * i.e. if nr_pending == 0.
 856  * We choose only to raise the barrier if no-one is waiting for the
 857  * barrier to go down.  This means that as soon as an IO request
 858  * is ready, no other operations which require a barrier will start
 859  * until the IO request has had a chance.
 860  *
 861  * So: regular IO calls 'wait_barrier'.  When that returns there
 862  *    is no backgroup IO happening,  It must arrange to call
 863  *    allow_barrier when it has finished its IO.
 864  * backgroup IO calls must call raise_barrier.  Once that returns
 865  *    there is no normal IO happeing.  It must arrange to call
 866  *    lower_barrier when the particular background IO completes.
 867  *
 868  * If resync/recovery is interrupted, returns -EINTR;
 869  * Otherwise, returns 0.
 870  */
 871 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 872 {
 873         int idx = sector_to_idx(sector_nr);
 874
 875         spin_lock_irq(&conf->resync_lock);
 876
 877         /* Wait until no block IO is waiting */
 878         wait_event_lock_irq(conf->wait_barrier,
 879                             !atomic_read(&conf->nr_waiting[idx]),
 880                             conf->resync_lock);
 881
 882         /* block any new IO from starting */
 883         atomic_inc(&conf->barrier[idx]);
 884         /*
 885          * In raise_barrier() we firstly increase conf->barrier[idx] then
 886          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 887          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 888          * A memory barrier here to make sure conf->nr_pending[idx] won't
 889          * be fetched before conf->barrier[idx] is increased. Otherwise
 890          * there will be a race between raise_barrier() and _wait_barrier().
 891          */
 892         smp_mb__after_atomic();
 893
 894         /* For these conditions we must wait:
 895          * A: while the array is in frozen state
 896          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 897          *    existing in corresponding I/O barrier bucket.
 898          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 899          *    max resync count which allowed on current I/O barrier bucket.
 900          */
 901         wait_event_lock_irq(conf->wait_barrier,
 902                             (!conf->array_frozen &&
 903                              !atomic_read(&conf->nr_pending[idx]) &&
 904                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 905                                 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 906                             conf->resync_lock);
 907
 908         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 909                 atomic_dec(&conf->barrier[idx]);
 910                 spin_unlock_irq(&conf->resync_lock);
 911                 wake_up(&conf->wait_barrier);
 912                 return -EINTR;
 913         }
 914
 915         atomic_inc(&conf->nr_sync_pending);
 916         spin_unlock_irq(&conf->resync_lock);
 917
 918         return 0;
 919 }
 920
 921 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 922 {
 923         int idx = sector_to_idx(sector_nr);
 924
 925         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 926
 927         atomic_dec(&conf->barrier[idx]);
 928         atomic_dec(&conf->nr_sync_pending);
 929         wake_up(&conf->wait_barrier);
 930 }
 931
 932 static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
 933 {
 934         bool ret = true;
 935
 936         /*
 937          * We need to increase conf->nr_pending[idx] very early here,
 938          * then raise_barrier() can be blocked when it waits for
 939          * conf->nr_pending[idx] to be 0. Then we can avoid holding
 940          * conf->resync_lock when there is no barrier raised in same
 941          * barrier unit bucket. Also if the array is frozen, I/O
 942          * should be blocked until array is unfrozen.
 943          */
 944         atomic_inc(&conf->nr_pending[idx]);
 945         /*
 946          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 947          * check conf->barrier[idx]. In raise_barrier() we firstly increase
 948          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 949          * barrier is necessary here to make sure conf->barrier[idx] won't be
 950          * fetched before conf->nr_pending[idx] is increased. Otherwise there
 951          * will be a race between _wait_barrier() and raise_barrier().
 952          */
 953         smp_mb__after_atomic();
 954
 955         /*
 956          * Don't worry about checking two atomic_t variables at same time
 957          * here. If during we check conf->barrier[idx], the array is
 958          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 959          * 0, it is safe to return and make the I/O continue. Because the
 960          * array is frozen, all I/O returned here will eventually complete
 961          * or be queued, no race will happen. See code comment in
 962          * frozen_array().
 963          */
 964         if (!READ_ONCE(conf->array_frozen) &&
 965             !atomic_read(&conf->barrier[idx]))
 966                 return ret;
 967
 968         /*
 969          * After holding conf->resync_lock, conf->nr_pending[idx]
 970          * should be decreased before waiting for barrier to drop.
 971          * Otherwise, we may encounter a race condition because
 972          * raise_barrer() might be waiting for conf->nr_pending[idx]
 973          * to be 0 at same time.
 974          */
 975         spin_lock_irq(&conf->resync_lock);
 976         atomic_inc(&conf->nr_waiting[idx]);
 977         atomic_dec(&conf->nr_pending[idx]);
 978         /*
 979          * In case freeze_array() is waiting for
 980          * get_unqueued_pending() == extra
 981          */
 982         wake_up(&conf->wait_barrier);
 983         /* Wait for the barrier in same barrier unit bucket to drop. */
 984
 985         /* Return false when nowait flag is set */
 986         if (nowait) {
 987                 ret = false;
 988         } else {
 989                 wait_event_lock_irq(conf->wait_barrier,
 990                                 !conf->array_frozen &&
 991                                 !atomic_read(&conf->barrier[idx]),
 992                                 conf->resync_lock);
 993                 atomic_inc(&conf->nr_pending[idx]);
 994         }
 995
 996         atomic_dec(&conf->nr_waiting[idx]);
 997         spin_unlock_irq(&conf->resync_lock);
 998         return ret;
 999 }
1000
1001 static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1002 {
1003         int idx = sector_to_idx(sector_nr);
1004         bool ret = true;
1005
1006         /*
1007          * Very similar to _wait_barrier(). The difference is, for read
1008          * I/O we don't need wait for sync I/O, but if the whole array
1009          * is frozen, the read I/O still has to wait until the array is
1010          * unfrozen. Since there is no ordering requirement with
1011          * conf->barrier[idx] here, memory barrier is unnecessary as well.
1012          */
1013         atomic_inc(&conf->nr_pending[idx]);
1014
1015         if (!READ_ONCE(conf->array_frozen))
1016                 return ret;
1017
1018         spin_lock_irq(&conf->resync_lock);
1019         atomic_inc(&conf->nr_waiting[idx]);
1020         atomic_dec(&conf->nr_pending[idx]);
1021         /*
1022          * In case freeze_array() is waiting for
1023          * get_unqueued_pending() == extra
1024          */
1025         wake_up(&conf->wait_barrier);
1026         /* Wait for array to be unfrozen */
1027
1028         /* Return false when nowait flag is set */
1029         if (nowait) {
1030                 /* Return false when nowait flag is set */
1031                 ret = false;
1032         } else {
1033                 wait_event_lock_irq(conf->wait_barrier,
1034                                 !conf->array_frozen,
1035                                 conf->resync_lock);
1036                 atomic_inc(&conf->nr_pending[idx]);
1037         }
1038
1039         atomic_dec(&conf->nr_waiting[idx]);
1040         spin_unlock_irq(&conf->resync_lock);
1041         return ret;
1042 }
1043
1044 static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1045 {
1046         int idx = sector_to_idx(sector_nr);
1047
1048         return _wait_barrier(conf, idx, nowait);
1049 }
1050
1051 static void _allow_barrier(struct r1conf *conf, int idx)
1052 {
1053         atomic_dec(&conf->nr_pending[idx]);
1054         wake_up(&conf->wait_barrier);
1055 }
1056
1057 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1058 {
1059         int idx = sector_to_idx(sector_nr);
1060
1061         _allow_barrier(conf, idx);
1062 }
1063
1064 /* conf->resync_lock should be held */
1065 static int get_unqueued_pending(struct r1conf *conf)
1066 {
1067         int idx, ret;
1068
1069         ret = atomic_read(&conf->nr_sync_pending);
1070         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1071                 ret += atomic_read(&conf->nr_pending[idx]) -
1072                         atomic_read(&conf->nr_queued[idx]);
1073
1074         return ret;
1075 }
1076
1077 static void freeze_array(struct r1conf *conf, int extra)
1078 {
1079         /* Stop sync I/O and normal I/O and wait for everything to
1080          * go quiet.
1081          * This is called in two situations:
1082          * 1) management command handlers (reshape, remove disk, quiesce).
1083          * 2) one normal I/O request failed.
1084
1085          * After array_frozen is set to 1, new sync IO will be blocked at
1086          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1087          * or wait_read_barrier(). The flying I/Os will either complete or be
1088          * queued. When everything goes quite, there are only queued I/Os left.
1089
1090          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1091          * barrier bucket index which this I/O request hits. When all sync and
1092          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1093          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1094          * in handle_read_error(), we may call freeze_array() before trying to
1095          * fix the read error. In this case, the error read I/O is not queued,
1096          * so get_unqueued_pending() == 1.
1097          *
1098          * Therefore before this function returns, we need to wait until
1099          * get_unqueued_pendings(conf) gets equal to extra. For
1100          * normal I/O context, extra is 1, in rested situations extra is 0.
1101          */
1102         spin_lock_irq(&conf->resync_lock);
1103         conf->array_frozen = 1;
1104         raid1_log(conf->mddev, "wait freeze");
1105         wait_event_lock_irq_cmd(
1106                 conf->wait_barrier,
1107                 get_unqueued_pending(conf) == extra,
1108                 conf->resync_lock,
1109                 flush_pending_writes(conf));
1110         spin_unlock_irq(&conf->resync_lock);
1111 }
1112 static void unfreeze_array(struct r1conf *conf)
1113 {
1114         /* reverse the effect of the freeze */
1115         spin_lock_irq(&conf->resync_lock);
1116         conf->array_frozen = 0;
1117         spin_unlock_irq(&conf->resync_lock);
1118         wake_up(&conf->wait_barrier);
1119 }
1120
1121 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1122                                            struct bio *bio)
1123 {
1124         int size = bio->bi_iter.bi_size;
1125         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1126         int i = 0;
1127         struct bio *behind_bio = NULL;
1128
1129         behind_bio = bio_alloc_bioset(GFP_NOIO, vcnt, &r1_bio->mddev->bio_set);
1130         if (!behind_bio)
1131                 return;
1132
1133         /* discard op, we don't support writezero/writesame yet */
1134         if (!bio_has_data(bio)) {
1135                 behind_bio->bi_iter.bi_size = size;
1136                 goto skip_copy;
1137         }
1138
1139         behind_bio->bi_write_hint = bio->bi_write_hint;
1140
1141         while (i < vcnt && size) {
1142                 struct page *page;
1143                 int len = min_t(int, PAGE_SIZE, size);
1144
1145                 page = alloc_page(GFP_NOIO);
1146                 if (unlikely(!page))
1147                         goto free_pages;
1148
1149                 bio_add_page(behind_bio, page, len, 0);
1150
1151                 size -= len;
1152                 i++;
1153         }
1154
1155         bio_copy_data(behind_bio, bio);
1156 skip_copy:
1157         r1_bio->behind_master_bio = behind_bio;
1158         set_bit(R1BIO_BehindIO, &r1_bio->state);
1159
1160         return;
1161
1162 free_pages:
1163         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1164                  bio->bi_iter.bi_size);
1165         bio_free_pages(behind_bio);
1166         bio_put(behind_bio);
1167 }
1168
1169 struct raid1_plug_cb {
1170         struct blk_plug_cb      cb;
1171         struct bio_list         pending;
1172         int                     pending_cnt;
1173 };
1174
1175 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1176 {
1177         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1178                                                   cb);
1179         struct mddev *mddev = plug->cb.data;
1180         struct r1conf *conf = mddev->private;
1181         struct bio *bio;
1182
1183         if (from_schedule || current->bio_list) {
1184                 spin_lock_irq(&conf->device_lock);
1185                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1186                 conf->pending_count += plug->pending_cnt;
1187                 spin_unlock_irq(&conf->device_lock);
1188                 wake_up(&conf->wait_barrier);
1189                 md_wakeup_thread(mddev->thread);
1190                 kfree(plug);
1191                 return;
1192         }
1193
1194         /* we aren't scheduling, so we can do the write-out directly. */
1195         bio = bio_list_get(&plug->pending);
1196         flush_bio_list(conf, bio);
1197         kfree(plug);
1198 }
1199
1200 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1201 {
1202         r1_bio->master_bio = bio;
1203         r1_bio->sectors = bio_sectors(bio);
1204         r1_bio->state = 0;
1205         r1_bio->mddev = mddev;
1206         r1_bio->sector = bio->bi_iter.bi_sector;
1207 }
1208
1209 static inline struct r1bio *
1210 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1211 {
1212         struct r1conf *conf = mddev->private;
1213         struct r1bio *r1_bio;
1214
1215         r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1216         /* Ensure no bio records IO_BLOCKED */
1217         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1218         init_r1bio(r1_bio, mddev, bio);
1219         return r1_bio;
1220 }
1221
1222 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1223                                int max_read_sectors, struct r1bio *r1_bio)
1224 {
1225         struct r1conf *conf = mddev->private;
1226         struct raid1_info *mirror;
1227         struct bio *read_bio;
1228         struct bitmap *bitmap = mddev->bitmap;
1229         const int op = bio_op(bio);
1230         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1231         int max_sectors;
1232         int rdisk;
1233         bool r1bio_existed = !!r1_bio;
1234         char b[BDEVNAME_SIZE];
1235
1236         /*
1237          * If r1_bio is set, we are blocking the raid1d thread
1238          * so there is a tiny risk of deadlock.  So ask for
1239          * emergency memory if needed.
1240          */
1241         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1242
1243         if (r1bio_existed) {
1244                 /* Need to get the block device name carefully */
1245                 struct md_rdev *rdev;
1246                 rcu_read_lock();
1247                 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1248                 if (rdev)
1249                         bdevname(rdev->bdev, b);
1250                 else
1251                         strcpy(b, "???");
1252                 rcu_read_unlock();
1253         }
1254
1255         /*
1256          * Still need barrier for READ in case that whole
1257          * array is frozen.
1258          */
1259         if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1260                                 bio->bi_opf & REQ_NOWAIT)) {
1261                 bio_wouldblock_error(bio);
1262                 return;
1263         }
1264
1265         if (!r1_bio)
1266                 r1_bio = alloc_r1bio(mddev, bio);
1267         else
1268                 init_r1bio(r1_bio, mddev, bio);
1269         r1_bio->sectors = max_read_sectors;
1270
1271         /*
1272          * make_request() can abort the operation when read-ahead is being
1273          * used and no empty request is available.
1274          */
1275         rdisk = read_balance(conf, r1_bio, &max_sectors);
1276
1277         if (rdisk < 0) {
1278                 /* couldn't find anywhere to read from */
1279                 if (r1bio_existed) {
1280                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1281                                             mdname(mddev),
1282                                             b,
1283                                             (unsigned long long)r1_bio->sector);
1284                 }
1285                 raid_end_bio_io(r1_bio);
1286                 return;
1287         }
1288         mirror = conf->mirrors + rdisk;
1289
1290         if (r1bio_existed)
1291                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1292                                     mdname(mddev),
1293                                     (unsigned long long)r1_bio->sector,
1294                                     bdevname(mirror->rdev->bdev, b));
1295
1296         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1297             bitmap) {
1298                 /*
1299                  * Reading from a write-mostly device must take care not to
1300                  * over-take any writes that are 'behind'
1301                  */
1302                 raid1_log(mddev, "wait behind writes");
1303                 wait_event(bitmap->behind_wait,
1304                            atomic_read(&bitmap->behind_writes) == 0);
1305         }
1306
1307         if (max_sectors < bio_sectors(bio)) {
1308                 struct bio *split = bio_split(bio, max_sectors,
1309                                               gfp, &conf->bio_split);
1310                 bio_chain(split, bio);
1311                 submit_bio_noacct(bio);
1312                 bio = split;
1313                 r1_bio->master_bio = bio;
1314                 r1_bio->sectors = max_sectors;
1315         }
1316
1317         r1_bio->read_disk = rdisk;
1318
1319         if (!r1bio_existed && blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1320                 r1_bio->start_time = bio_start_io_acct(bio);
1321
1322         read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1323
1324         r1_bio->bios[rdisk] = read_bio;
1325
1326         read_bio->bi_iter.bi_sector = r1_bio->sector +
1327                 mirror->rdev->data_offset;
1328         bio_set_dev(read_bio, mirror->rdev->bdev);
1329         read_bio->bi_end_io = raid1_end_read_request;
1330         bio_set_op_attrs(read_bio, op, do_sync);
1331         if (test_bit(FailFast, &mirror->rdev->flags) &&
1332             test_bit(R1BIO_FailFast, &r1_bio->state))
1333                 read_bio->bi_opf |= MD_FAILFAST;
1334         read_bio->bi_private = r1_bio;
1335
1336         if (mddev->gendisk)
1337                 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1338                                       r1_bio->sector);
1339
1340         submit_bio_noacct(read_bio);
1341 }
1342
1343 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1344                                 int max_write_sectors)
1345 {
1346         struct r1conf *conf = mddev->private;
1347         struct r1bio *r1_bio;
1348         int i, disks;
1349         struct bitmap *bitmap = mddev->bitmap;
1350         unsigned long flags;
1351         struct md_rdev *blocked_rdev;
1352         struct blk_plug_cb *cb;
1353         struct raid1_plug_cb *plug = NULL;
1354         int first_clone;
1355         int max_sectors;
1356         bool write_behind = false;
1357
1358         if (mddev_is_clustered(mddev) &&
1359              md_cluster_ops->area_resyncing(mddev, WRITE,
1360                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1361
1362                 DEFINE_WAIT(w);
1363                 if (bio->bi_opf & REQ_NOWAIT) {
1364                         bio_wouldblock_error(bio);
1365                         return;
1366                 }
1367                 for (;;) {
1368                         prepare_to_wait(&conf->wait_barrier,
1369                                         &w, TASK_IDLE);
1370                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1371                                                         bio->bi_iter.bi_sector,
1372                                                         bio_end_sector(bio)))
1373                                 break;
1374                         schedule();
1375                 }
1376                 finish_wait(&conf->wait_barrier, &w);
1377         }
1378
1379         /*
1380          * Register the new request and wait if the reconstruction
1381          * thread has put up a bar for new requests.
1382          * Continue immediately if no resync is active currently.
1383          */
1384         if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1385                                 bio->bi_opf & REQ_NOWAIT)) {
1386                 bio_wouldblock_error(bio);
1387                 return;
1388         }
1389
1390         r1_bio = alloc_r1bio(mddev, bio);
1391         r1_bio->sectors = max_write_sectors;
1392
1393         /* first select target devices under rcu_lock and
1394          * inc refcount on their rdev.  Record them by setting
1395          * bios[x] to bio
1396          * If there are known/acknowledged bad blocks on any device on
1397          * which we have seen a write error, we want to avoid writing those
1398          * blocks.
1399          * This potentially requires several writes to write around
1400          * the bad blocks.  Each set of writes gets it's own r1bio
1401          * with a set of bios attached.
1402          */
1403
1404         disks = conf->raid_disks * 2;
1405  retry_write:
1406         blocked_rdev = NULL;
1407         rcu_read_lock();
1408         max_sectors = r1_bio->sectors;
1409         for (i = 0;  i < disks; i++) {
1410                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1411
1412                 /*
1413                  * The write-behind io is only attempted on drives marked as
1414                  * write-mostly, which means we could allocate write behind
1415                  * bio later.
1416                  */
1417                 if (rdev && test_bit(WriteMostly, &rdev->flags))
1418                         write_behind = true;
1419
1420                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1421                         atomic_inc(&rdev->nr_pending);
1422                         blocked_rdev = rdev;
1423                         break;
1424                 }
1425                 r1_bio->bios[i] = NULL;
1426                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1427                         if (i < conf->raid_disks)
1428                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1429                         continue;
1430                 }
1431
1432                 atomic_inc(&rdev->nr_pending);
1433                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1434                         sector_t first_bad;
1435                         int bad_sectors;
1436                         int is_bad;
1437
1438                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1439                                              &first_bad, &bad_sectors);
1440                         if (is_bad < 0) {
1441                                 /* mustn't write here until the bad block is
1442                                  * acknowledged*/
1443                                 set_bit(BlockedBadBlocks, &rdev->flags);
1444                                 blocked_rdev = rdev;
1445                                 break;
1446                         }
1447                         if (is_bad && first_bad <= r1_bio->sector) {
1448                                 /* Cannot write here at all */
1449                                 bad_sectors -= (r1_bio->sector - first_bad);
1450                                 if (bad_sectors < max_sectors)
1451                                         /* mustn't write more than bad_sectors
1452                                          * to other devices yet
1453                                          */
1454                                         max_sectors = bad_sectors;
1455                                 rdev_dec_pending(rdev, mddev);
1456                                 /* We don't set R1BIO_Degraded as that
1457                                  * only applies if the disk is
1458                                  * missing, so it might be re-added,
1459                                  * and we want to know to recover this
1460                                  * chunk.
1461                                  * In this case the device is here,
1462                                  * and the fact that this chunk is not
1463                                  * in-sync is recorded in the bad
1464                                  * block log
1465                                  */
1466                                 continue;
1467                         }
1468                         if (is_bad) {
1469                                 int good_sectors = first_bad - r1_bio->sector;
1470                                 if (good_sectors < max_sectors)
1471                                         max_sectors = good_sectors;
1472                         }
1473                 }
1474                 r1_bio->bios[i] = bio;
1475         }
1476         rcu_read_unlock();
1477
1478         if (unlikely(blocked_rdev)) {
1479                 /* Wait for this device to become unblocked */
1480                 int j;
1481
1482                 for (j = 0; j < i; j++)
1483                         if (r1_bio->bios[j])
1484                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1485                 r1_bio->state = 0;
1486                 allow_barrier(conf, bio->bi_iter.bi_sector);
1487
1488                 if (bio->bi_opf & REQ_NOWAIT) {
1489                         bio_wouldblock_error(bio);
1490                         return;
1491                 }
1492                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1493                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1494                 wait_barrier(conf, bio->bi_iter.bi_sector, false);
1495                 goto retry_write;
1496         }
1497
1498         /*
1499          * When using a bitmap, we may call alloc_behind_master_bio below.
1500          * alloc_behind_master_bio allocates a copy of the data payload a page
1501          * at a time and thus needs a new bio that can fit the whole payload
1502          * this bio in page sized chunks.
1503          */
1504         if (write_behind && bitmap)
1505                 max_sectors = min_t(int, max_sectors,
1506                                     BIO_MAX_VECS * (PAGE_SIZE >> 9));
1507         if (max_sectors < bio_sectors(bio)) {
1508                 struct bio *split = bio_split(bio, max_sectors,
1509                                               GFP_NOIO, &conf->bio_split);
1510                 bio_chain(split, bio);
1511                 submit_bio_noacct(bio);
1512                 bio = split;
1513                 r1_bio->master_bio = bio;
1514                 r1_bio->sectors = max_sectors;
1515         }
1516
1517         if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1518                 r1_bio->start_time = bio_start_io_acct(bio);
1519         atomic_set(&r1_bio->remaining, 1);
1520         atomic_set(&r1_bio->behind_remaining, 0);
1521
1522         first_clone = 1;
1523
1524         for (i = 0; i < disks; i++) {
1525                 struct bio *mbio = NULL;
1526                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1527                 if (!r1_bio->bios[i])
1528                         continue;
1529
1530                 if (first_clone) {
1531                         /* do behind I/O ?
1532                          * Not if there are too many, or cannot
1533                          * allocate memory, or a reader on WriteMostly
1534                          * is waiting for behind writes to flush */
1535                         if (bitmap &&
1536                             test_bit(WriteMostly, &rdev->flags) &&
1537                             (atomic_read(&bitmap->behind_writes)
1538                              < mddev->bitmap_info.max_write_behind) &&
1539                             !waitqueue_active(&bitmap->behind_wait)) {
1540                                 alloc_behind_master_bio(r1_bio, bio);
1541                         }
1542
1543                         md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1544                                              test_bit(R1BIO_BehindIO, &r1_bio->state));
1545                         first_clone = 0;
1546                 }
1547
1548                 if (r1_bio->behind_master_bio)
1549                         mbio = bio_clone_fast(r1_bio->behind_master_bio,
1550                                               GFP_NOIO, &mddev->bio_set);
1551                 else
1552                         mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1553
1554                 if (r1_bio->behind_master_bio) {
1555                         if (test_bit(CollisionCheck, &rdev->flags))
1556                                 wait_for_serialization(rdev, r1_bio);
1557                         if (test_bit(WriteMostly, &rdev->flags))
1558                                 atomic_inc(&r1_bio->behind_remaining);
1559                 } else if (mddev->serialize_policy)
1560                         wait_for_serialization(rdev, r1_bio);
1561
1562                 r1_bio->bios[i] = mbio;
1563
1564                 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1565                 bio_set_dev(mbio, rdev->bdev);
1566                 mbio->bi_end_io = raid1_end_write_request;
1567                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1568                 if (test_bit(FailFast, &rdev->flags) &&
1569                     !test_bit(WriteMostly, &rdev->flags) &&
1570                     conf->raid_disks - mddev->degraded > 1)
1571                         mbio->bi_opf |= MD_FAILFAST;
1572                 mbio->bi_private = r1_bio;
1573
1574                 atomic_inc(&r1_bio->remaining);
1575
1576                 if (mddev->gendisk)
1577                         trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1578                                               r1_bio->sector);
1579                 /* flush_pending_writes() needs access to the rdev so...*/
1580                 mbio->bi_bdev = (void *)rdev;
1581
1582                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1583                 if (cb)
1584                         plug = container_of(cb, struct raid1_plug_cb, cb);
1585                 else
1586                         plug = NULL;
1587                 if (plug) {
1588                         bio_list_add(&plug->pending, mbio);
1589                         plug->pending_cnt++;
1590                 } else {
1591                         spin_lock_irqsave(&conf->device_lock, flags);
1592                         bio_list_add(&conf->pending_bio_list, mbio);
1593                         conf->pending_count++;
1594                         spin_unlock_irqrestore(&conf->device_lock, flags);
1595                         md_wakeup_thread(mddev->thread);
1596                 }
1597         }
1598
1599         r1_bio_write_done(r1_bio);
1600
1601         /* In case raid1d snuck in to freeze_array */
1602         wake_up(&conf->wait_barrier);
1603 }
1604
1605 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1606 {
1607         sector_t sectors;
1608
1609         if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1610             && md_flush_request(mddev, bio))
1611                 return true;
1612
1613         /*
1614          * There is a limit to the maximum size, but
1615          * the read/write handler might find a lower limit
1616          * due to bad blocks.  To avoid multiple splits,
1617          * we pass the maximum number of sectors down
1618          * and let the lower level perform the split.
1619          */
1620         sectors = align_to_barrier_unit_end(
1621                 bio->bi_iter.bi_sector, bio_sectors(bio));
1622
1623         if (bio_data_dir(bio) == READ)
1624                 raid1_read_request(mddev, bio, sectors, NULL);
1625         else {
1626                 if (!md_write_start(mddev,bio))
1627                         return false;
1628                 raid1_write_request(mddev, bio, sectors);
1629         }
1630         return true;
1631 }
1632
1633 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1634 {
1635         struct r1conf *conf = mddev->private;
1636         int i;
1637
1638         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1639                    conf->raid_disks - mddev->degraded);
1640         rcu_read_lock();
1641         for (i = 0; i < conf->raid_disks; i++) {
1642                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1643                 seq_printf(seq, "%s",
1644                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1645         }
1646         rcu_read_unlock();
1647         seq_printf(seq, "]");
1648 }
1649
1650 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1651 {
1652         char b[BDEVNAME_SIZE];
1653         struct r1conf *conf = mddev->private;
1654         unsigned long flags;
1655
1656         /*
1657          * If it is not operational, then we have already marked it as dead
1658          * else if it is the last working disks with "fail_last_dev == false",
1659          * ignore the error, let the next level up know.
1660          * else mark the drive as failed
1661          */
1662         spin_lock_irqsave(&conf->device_lock, flags);
1663         if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1664             && (conf->raid_disks - mddev->degraded) == 1) {
1665                 /*
1666                  * Don't fail the drive, act as though we were just a
1667                  * normal single drive.
1668                  * However don't try a recovery from this drive as
1669                  * it is very likely to fail.
1670                  */
1671                 conf->recovery_disabled = mddev->recovery_disabled;
1672                 spin_unlock_irqrestore(&conf->device_lock, flags);
1673                 return;
1674         }
1675         set_bit(Blocked, &rdev->flags);
1676         if (test_and_clear_bit(In_sync, &rdev->flags))
1677                 mddev->degraded++;
1678         set_bit(Faulty, &rdev->flags);
1679         spin_unlock_irqrestore(&conf->device_lock, flags);
1680         /*
1681          * if recovery is running, make sure it aborts.
1682          */
1683         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1684         set_mask_bits(&mddev->sb_flags, 0,
1685                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1686         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1687                 "md/raid1:%s: Operation continuing on %d devices.\n",
1688                 mdname(mddev), bdevname(rdev->bdev, b),
1689                 mdname(mddev), conf->raid_disks - mddev->degraded);
1690 }
1691
1692 static void print_conf(struct r1conf *conf)
1693 {
1694         int i;
1695
1696         pr_debug("RAID1 conf printout:\n");
1697         if (!conf) {
1698                 pr_debug("(!conf)\n");
1699                 return;
1700         }
1701         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1702                  conf->raid_disks);
1703
1704         rcu_read_lock();
1705         for (i = 0; i < conf->raid_disks; i++) {
1706                 char b[BDEVNAME_SIZE];
1707                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1708                 if (rdev)
1709                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1710                                  i, !test_bit(In_sync, &rdev->flags),
1711                                  !test_bit(Faulty, &rdev->flags),
1712                                  bdevname(rdev->bdev,b));
1713         }
1714         rcu_read_unlock();
1715 }
1716
1717 static void close_sync(struct r1conf *conf)
1718 {
1719         int idx;
1720
1721         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1722                 _wait_barrier(conf, idx, false);
1723                 _allow_barrier(conf, idx);
1724         }
1725
1726         mempool_exit(&conf->r1buf_pool);
1727 }
1728
1729 static int raid1_spare_active(struct mddev *mddev)
1730 {
1731         int i;
1732         struct r1conf *conf = mddev->private;
1733         int count = 0;
1734         unsigned long flags;
1735
1736         /*
1737          * Find all failed disks within the RAID1 configuration
1738          * and mark them readable.
1739          * Called under mddev lock, so rcu protection not needed.
1740          * device_lock used to avoid races with raid1_end_read_request
1741          * which expects 'In_sync' flags and ->degraded to be consistent.
1742          */
1743         spin_lock_irqsave(&conf->device_lock, flags);
1744         for (i = 0; i < conf->raid_disks; i++) {
1745                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1746                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1747                 if (repl
1748                     && !test_bit(Candidate, &repl->flags)
1749                     && repl->recovery_offset == MaxSector
1750                     && !test_bit(Faulty, &repl->flags)
1751                     && !test_and_set_bit(In_sync, &repl->flags)) {
1752                         /* replacement has just become active */
1753                         if (!rdev ||
1754                             !test_and_clear_bit(In_sync, &rdev->flags))
1755                                 count++;
1756                         if (rdev) {
1757                                 /* Replaced device not technically
1758                                  * faulty, but we need to be sure
1759                                  * it gets removed and never re-added
1760                                  */
1761                                 set_bit(Faulty, &rdev->flags);
1762                                 sysfs_notify_dirent_safe(
1763                                         rdev->sysfs_state);
1764                         }
1765                 }
1766                 if (rdev
1767                     && rdev->recovery_offset == MaxSector
1768                     && !test_bit(Faulty, &rdev->flags)
1769                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1770                         count++;
1771                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1772                 }
1773         }
1774         mddev->degraded -= count;
1775         spin_unlock_irqrestore(&conf->device_lock, flags);
1776
1777         print_conf(conf);
1778         return count;
1779 }
1780
1781 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1782 {
1783         struct r1conf *conf = mddev->private;
1784         int err = -EEXIST;
1785         int mirror = 0;
1786         struct raid1_info *p;
1787         int first = 0;
1788         int last = conf->raid_disks - 1;
1789
1790         if (mddev->recovery_disabled == conf->recovery_disabled)
1791                 return -EBUSY;
1792
1793         if (md_integrity_add_rdev(rdev, mddev))
1794                 return -ENXIO;
1795
1796         if (rdev->raid_disk >= 0)
1797                 first = last = rdev->raid_disk;
1798
1799         /*
1800          * find the disk ... but prefer rdev->saved_raid_disk
1801          * if possible.
1802          */
1803         if (rdev->saved_raid_disk >= 0 &&
1804             rdev->saved_raid_disk >= first &&
1805             rdev->saved_raid_disk < conf->raid_disks &&
1806             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1807                 first = last = rdev->saved_raid_disk;
1808
1809         for (mirror = first; mirror <= last; mirror++) {
1810                 p = conf->mirrors + mirror;
1811                 if (!p->rdev) {
1812                         if (mddev->gendisk)
1813                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1814                                                   rdev->data_offset << 9);
1815
1816                         p->head_position = 0;
1817                         rdev->raid_disk = mirror;
1818                         err = 0;
1819                         /* As all devices are equivalent, we don't need a full recovery
1820                          * if this was recently any drive of the array
1821                          */
1822                         if (rdev->saved_raid_disk < 0)
1823                                 conf->fullsync = 1;
1824                         rcu_assign_pointer(p->rdev, rdev);
1825                         break;
1826                 }
1827                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1828                     p[conf->raid_disks].rdev == NULL) {
1829                         /* Add this device as a replacement */
1830                         clear_bit(In_sync, &rdev->flags);
1831                         set_bit(Replacement, &rdev->flags);
1832                         rdev->raid_disk = mirror;
1833                         err = 0;
1834                         conf->fullsync = 1;
1835                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1836                         break;
1837                 }
1838         }
1839         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1840                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1841         print_conf(conf);
1842         return err;
1843 }
1844
1845 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1846 {
1847         struct r1conf *conf = mddev->private;
1848         int err = 0;
1849         int number = rdev->raid_disk;
1850         struct raid1_info *p = conf->mirrors + number;
1851
1852         if (rdev != p->rdev)
1853                 p = conf->mirrors + conf->raid_disks + number;
1854
1855         print_conf(conf);
1856         if (rdev == p->rdev) {
1857                 if (test_bit(In_sync, &rdev->flags) ||
1858                     atomic_read(&rdev->nr_pending)) {
1859                         err = -EBUSY;
1860                         goto abort;
1861                 }
1862                 /* Only remove non-faulty devices if recovery
1863                  * is not possible.
1864                  */
1865                 if (!test_bit(Faulty, &rdev->flags) &&
1866                     mddev->recovery_disabled != conf->recovery_disabled &&
1867                     mddev->degraded < conf->raid_disks) {
1868                         err = -EBUSY;
1869                         goto abort;
1870                 }
1871                 p->rdev = NULL;
1872                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1873                         synchronize_rcu();
1874                         if (atomic_read(&rdev->nr_pending)) {
1875                                 /* lost the race, try later */
1876                                 err = -EBUSY;
1877                                 p->rdev = rdev;
1878                                 goto abort;
1879                         }
1880                 }
1881                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1882                         /* We just removed a device that is being replaced.
1883                          * Move down the replacement.  We drain all IO before
1884                          * doing this to avoid confusion.
1885                          */
1886                         struct md_rdev *repl =
1887                                 conf->mirrors[conf->raid_disks + number].rdev;
1888                         freeze_array(conf, 0);
1889                         if (atomic_read(&repl->nr_pending)) {
1890                                 /* It means that some queued IO of retry_list
1891                                  * hold repl. Thus, we cannot set replacement
1892                                  * as NULL, avoiding rdev NULL pointer
1893                                  * dereference in sync_request_write and
1894                                  * handle_write_finished.
1895                                  */
1896                                 err = -EBUSY;
1897                                 unfreeze_array(conf);
1898                                 goto abort;
1899                         }
1900                         clear_bit(Replacement, &repl->flags);
1901                         p->rdev = repl;
1902                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1903                         unfreeze_array(conf);
1904                 }
1905
1906                 clear_bit(WantReplacement, &rdev->flags);
1907                 err = md_integrity_register(mddev);
1908         }
1909 abort:
1910
1911         print_conf(conf);
1912         return err;
1913 }
1914
1915 static void end_sync_read(struct bio *bio)
1916 {
1917         struct r1bio *r1_bio = get_resync_r1bio(bio);
1918
1919         update_head_pos(r1_bio->read_disk, r1_bio);
1920
1921         /*
1922          * we have read a block, now it needs to be re-written,
1923          * or re-read if the read failed.
1924          * We don't do much here, just schedule handling by raid1d
1925          */
1926         if (!bio->bi_status)
1927                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1928
1929         if (atomic_dec_and_test(&r1_bio->remaining))
1930                 reschedule_retry(r1_bio);
1931 }
1932
1933 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1934 {
1935         sector_t sync_blocks = 0;
1936         sector_t s = r1_bio->sector;
1937         long sectors_to_go = r1_bio->sectors;
1938
1939         /* make sure these bits don't get cleared. */
1940         do {
1941                 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1942                 s += sync_blocks;
1943                 sectors_to_go -= sync_blocks;
1944         } while (sectors_to_go > 0);
1945 }
1946
1947 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1948 {
1949         if (atomic_dec_and_test(&r1_bio->remaining)) {
1950                 struct mddev *mddev = r1_bio->mddev;
1951                 int s = r1_bio->sectors;
1952
1953                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1954                     test_bit(R1BIO_WriteError, &r1_bio->state))
1955                         reschedule_retry(r1_bio);
1956                 else {
1957                         put_buf(r1_bio);
1958                         md_done_sync(mddev, s, uptodate);
1959                 }
1960         }
1961 }
1962
1963 static void end_sync_write(struct bio *bio)
1964 {
1965         int uptodate = !bio->bi_status;
1966         struct r1bio *r1_bio = get_resync_r1bio(bio);
1967         struct mddev *mddev = r1_bio->mddev;
1968         struct r1conf *conf = mddev->private;
1969         sector_t first_bad;
1970         int bad_sectors;
1971         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1972
1973         if (!uptodate) {
1974                 abort_sync_write(mddev, r1_bio);
1975                 set_bit(WriteErrorSeen, &rdev->flags);
1976                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1977                         set_bit(MD_RECOVERY_NEEDED, &
1978                                 mddev->recovery);
1979                 set_bit(R1BIO_WriteError, &r1_bio->state);
1980         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1981                                &first_bad, &bad_sectors) &&
1982                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1983                                 r1_bio->sector,
1984                                 r1_bio->sectors,
1985                                 &first_bad, &bad_sectors)
1986                 )
1987                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1988
1989         put_sync_write_buf(r1_bio, uptodate);
1990 }
1991
1992 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1993                             int sectors, struct page *page, int rw)
1994 {
1995         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1996                 /* success */
1997                 return 1;
1998         if (rw == WRITE) {
1999                 set_bit(WriteErrorSeen, &rdev->flags);
2000                 if (!test_and_set_bit(WantReplacement,
2001                                       &rdev->flags))
2002                         set_bit(MD_RECOVERY_NEEDED, &
2003                                 rdev->mddev->recovery);
2004         }
2005         /* need to record an error - either for the block or the device */
2006         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2007                 md_error(rdev->mddev, rdev);
2008         return 0;
2009 }
2010
2011 static int fix_sync_read_error(struct r1bio *r1_bio)
2012 {
2013         /* Try some synchronous reads of other devices to get
2014          * good data, much like with normal read errors.  Only
2015          * read into the pages we already have so we don't
2016          * need to re-issue the read request.
2017          * We don't need to freeze the array, because being in an
2018          * active sync request, there is no normal IO, and
2019          * no overlapping syncs.
2020          * We don't need to check is_badblock() again as we
2021          * made sure that anything with a bad block in range
2022          * will have bi_end_io clear.
2023          */
2024         struct mddev *mddev = r1_bio->mddev;
2025         struct r1conf *conf = mddev->private;
2026         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2027         struct page **pages = get_resync_pages(bio)->pages;
2028         sector_t sect = r1_bio->sector;
2029         int sectors = r1_bio->sectors;
2030         int idx = 0;
2031         struct md_rdev *rdev;
2032
2033         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2034         if (test_bit(FailFast, &rdev->flags)) {
2035                 /* Don't try recovering from here - just fail it
2036                  * ... unless it is the last working device of course */
2037                 md_error(mddev, rdev);
2038                 if (test_bit(Faulty, &rdev->flags))
2039                         /* Don't try to read from here, but make sure
2040                          * put_buf does it's thing
2041                          */
2042                         bio->bi_end_io = end_sync_write;
2043         }
2044
2045         while(sectors) {
2046                 int s = sectors;
2047                 int d = r1_bio->read_disk;
2048                 int success = 0;
2049                 int start;
2050
2051                 if (s > (PAGE_SIZE>>9))
2052                         s = PAGE_SIZE >> 9;
2053                 do {
2054                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2055                                 /* No rcu protection needed here devices
2056                                  * can only be removed when no resync is
2057                                  * active, and resync is currently active
2058                                  */
2059                                 rdev = conf->mirrors[d].rdev;
2060                                 if (sync_page_io(rdev, sect, s<<9,
2061                                                  pages[idx],
2062                                                  REQ_OP_READ, 0, false)) {
2063                                         success = 1;
2064                                         break;
2065                                 }
2066                         }
2067                         d++;
2068                         if (d == conf->raid_disks * 2)
2069                                 d = 0;
2070                 } while (!success && d != r1_bio->read_disk);
2071
2072                 if (!success) {
2073                         char b[BDEVNAME_SIZE];
2074                         int abort = 0;
2075                         /* Cannot read from anywhere, this block is lost.
2076                          * Record a bad block on each device.  If that doesn't
2077                          * work just disable and interrupt the recovery.
2078                          * Don't fail devices as that won't really help.
2079                          */
2080                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2081                                             mdname(mddev), bio_devname(bio, b),
2082                                             (unsigned long long)r1_bio->sector);
2083                         for (d = 0; d < conf->raid_disks * 2; d++) {
2084                                 rdev = conf->mirrors[d].rdev;
2085                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2086                                         continue;
2087                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2088                                         abort = 1;
2089                         }
2090                         if (abort) {
2091                                 conf->recovery_disabled =
2092                                         mddev->recovery_disabled;
2093                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2094                                 md_done_sync(mddev, r1_bio->sectors, 0);
2095                                 put_buf(r1_bio);
2096                                 return 0;
2097                         }
2098                         /* Try next page */
2099                         sectors -= s;
2100                         sect += s;
2101                         idx++;
2102                         continue;
2103                 }
2104
2105                 start = d;
2106                 /* write it back and re-read */
2107                 while (d != r1_bio->read_disk) {
2108                         if (d == 0)
2109                                 d = conf->raid_disks * 2;
2110                         d--;
2111                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2112                                 continue;
2113                         rdev = conf->mirrors[d].rdev;
2114                         if (r1_sync_page_io(rdev, sect, s,
2115                                             pages[idx],
2116                                             WRITE) == 0) {
2117                                 r1_bio->bios[d]->bi_end_io = NULL;
2118                                 rdev_dec_pending(rdev, mddev);
2119                         }
2120                 }
2121                 d = start;
2122                 while (d != r1_bio->read_disk) {
2123                         if (d == 0)
2124                                 d = conf->raid_disks * 2;
2125                         d--;
2126                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2127                                 continue;
2128                         rdev = conf->mirrors[d].rdev;
2129                         if (r1_sync_page_io(rdev, sect, s,
2130                                             pages[idx],
2131                                             READ) != 0)
2132                                 atomic_add(s, &rdev->corrected_errors);
2133                 }
2134                 sectors -= s;
2135                 sect += s;
2136                 idx ++;
2137         }
2138         set_bit(R1BIO_Uptodate, &r1_bio->state);
2139         bio->bi_status = 0;
2140         return 1;
2141 }
2142
2143 static void process_checks(struct r1bio *r1_bio)
2144 {
2145         /* We have read all readable devices.  If we haven't
2146          * got the block, then there is no hope left.
2147          * If we have, then we want to do a comparison
2148          * and skip the write if everything is the same.
2149          * If any blocks failed to read, then we need to
2150          * attempt an over-write
2151          */
2152         struct mddev *mddev = r1_bio->mddev;
2153         struct r1conf *conf = mddev->private;
2154         int primary;
2155         int i;
2156         int vcnt;
2157
2158         /* Fix variable parts of all bios */
2159         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2160         for (i = 0; i < conf->raid_disks * 2; i++) {
2161                 blk_status_t status;
2162                 struct bio *b = r1_bio->bios[i];
2163                 struct resync_pages *rp = get_resync_pages(b);
2164                 if (b->bi_end_io != end_sync_read)
2165                         continue;
2166                 /* fixup the bio for reuse, but preserve errno */
2167                 status = b->bi_status;
2168                 bio_reset(b);
2169                 b->bi_status = status;
2170                 b->bi_iter.bi_sector = r1_bio->sector +
2171                         conf->mirrors[i].rdev->data_offset;
2172                 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2173                 b->bi_end_io = end_sync_read;
2174                 rp->raid_bio = r1_bio;
2175                 b->bi_private = rp;
2176
2177                 /* initialize bvec table again */
2178                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2179         }
2180         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2181                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2182                     !r1_bio->bios[primary]->bi_status) {
2183                         r1_bio->bios[primary]->bi_end_io = NULL;
2184                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2185                         break;
2186                 }
2187         r1_bio->read_disk = primary;
2188         for (i = 0; i < conf->raid_disks * 2; i++) {
2189                 int j = 0;
2190                 struct bio *pbio = r1_bio->bios[primary];
2191                 struct bio *sbio = r1_bio->bios[i];
2192                 blk_status_t status = sbio->bi_status;
2193                 struct page **ppages = get_resync_pages(pbio)->pages;
2194                 struct page **spages = get_resync_pages(sbio)->pages;
2195                 struct bio_vec *bi;
2196                 int page_len[RESYNC_PAGES] = { 0 };
2197                 struct bvec_iter_all iter_all;
2198
2199                 if (sbio->bi_end_io != end_sync_read)
2200                         continue;
2201                 /* Now we can 'fixup' the error value */
2202                 sbio->bi_status = 0;
2203
2204                 bio_for_each_segment_all(bi, sbio, iter_all)
2205                         page_len[j++] = bi->bv_len;
2206
2207                 if (!status) {
2208                         for (j = vcnt; j-- ; ) {
2209                                 if (memcmp(page_address(ppages[j]),
2210                                            page_address(spages[j]),
2211                                            page_len[j]))
2212                                         break;
2213                         }
2214                 } else
2215                         j = 0;
2216                 if (j >= 0)
2217                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2218                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2219                               && !status)) {
2220                         /* No need to write to this device. */
2221                         sbio->bi_end_io = NULL;
2222                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2223                         continue;
2224                 }
2225
2226                 bio_copy_data(sbio, pbio);
2227         }
2228 }
2229
2230 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2231 {
2232         struct r1conf *conf = mddev->private;
2233         int i;
2234         int disks = conf->raid_disks * 2;
2235         struct bio *wbio;
2236
2237         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2238                 /* ouch - failed to read all of that. */
2239                 if (!fix_sync_read_error(r1_bio))
2240                         return;
2241
2242         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2243                 process_checks(r1_bio);
2244
2245         /*
2246          * schedule writes
2247          */
2248         atomic_set(&r1_bio->remaining, 1);
2249         for (i = 0; i < disks ; i++) {
2250                 wbio = r1_bio->bios[i];
2251                 if (wbio->bi_end_io == NULL ||
2252                     (wbio->bi_end_io == end_sync_read &&
2253                      (i == r1_bio->read_disk ||
2254                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2255                         continue;
2256                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2257                         abort_sync_write(mddev, r1_bio);
2258                         continue;
2259                 }
2260
2261                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2262                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2263                         wbio->bi_opf |= MD_FAILFAST;
2264
2265                 wbio->bi_end_io = end_sync_write;
2266                 atomic_inc(&r1_bio->remaining);
2267                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2268
2269                 submit_bio_noacct(wbio);
2270         }
2271
2272         put_sync_write_buf(r1_bio, 1);
2273 }
2274
2275 /*
2276  * This is a kernel thread which:
2277  *
2278  *      1.      Retries failed read operations on working mirrors.
2279  *      2.      Updates the raid superblock when problems encounter.
2280  *      3.      Performs writes following reads for array synchronising.
2281  */
2282
2283 static void fix_read_error(struct r1conf *conf, int read_disk,
2284                            sector_t sect, int sectors)
2285 {
2286         struct mddev *mddev = conf->mddev;
2287         while(sectors) {
2288                 int s = sectors;
2289                 int d = read_disk;
2290                 int success = 0;
2291                 int start;
2292                 struct md_rdev *rdev;
2293
2294                 if (s > (PAGE_SIZE>>9))
2295                         s = PAGE_SIZE >> 9;
2296
2297                 do {
2298                         sector_t first_bad;
2299                         int bad_sectors;
2300
2301                         rcu_read_lock();
2302                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2303                         if (rdev &&
2304                             (test_bit(In_sync, &rdev->flags) ||
2305                              (!test_bit(Faulty, &rdev->flags) &&
2306                               rdev->recovery_offset >= sect + s)) &&
2307                             is_badblock(rdev, sect, s,
2308                                         &first_bad, &bad_sectors) == 0) {
2309                                 atomic_inc(&rdev->nr_pending);
2310                                 rcu_read_unlock();
2311                                 if (sync_page_io(rdev, sect, s<<9,
2312                                          conf->tmppage, REQ_OP_READ, 0, false))
2313                                         success = 1;
2314                                 rdev_dec_pending(rdev, mddev);
2315                                 if (success)
2316                                         break;
2317                         } else
2318                                 rcu_read_unlock();
2319                         d++;
2320                         if (d == conf->raid_disks * 2)
2321                                 d = 0;
2322                 } while (!success && d != read_disk);
2323
2324                 if (!success) {
2325                         /* Cannot read from anywhere - mark it bad */
2326                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2327                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2328                                 md_error(mddev, rdev);
2329                         break;
2330                 }
2331                 /* write it back and re-read */
2332                 start = d;
2333                 while (d != read_disk) {
2334                         if (d==0)
2335                                 d = conf->raid_disks * 2;
2336                         d--;
2337                         rcu_read_lock();
2338                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2339                         if (rdev &&
2340                             !test_bit(Faulty, &rdev->flags)) {
2341                                 atomic_inc(&rdev->nr_pending);
2342                                 rcu_read_unlock();
2343                                 r1_sync_page_io(rdev, sect, s,
2344                                                 conf->tmppage, WRITE);
2345                                 rdev_dec_pending(rdev, mddev);
2346                         } else
2347                                 rcu_read_unlock();
2348                 }
2349                 d = start;
2350                 while (d != read_disk) {
2351                         char b[BDEVNAME_SIZE];
2352                         if (d==0)
2353                                 d = conf->raid_disks * 2;
2354                         d--;
2355                         rcu_read_lock();
2356                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2357                         if (rdev &&
2358                             !test_bit(Faulty, &rdev->flags)) {
2359                                 atomic_inc(&rdev->nr_pending);
2360                                 rcu_read_unlock();
2361                                 if (r1_sync_page_io(rdev, sect, s,
2362                                                     conf->tmppage, READ)) {
2363                                         atomic_add(s, &rdev->corrected_errors);
2364                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2365                                                 mdname(mddev), s,
2366                                                 (unsigned long long)(sect +
2367                                                                      rdev->data_offset),
2368                                                 bdevname(rdev->bdev, b));
2369                                 }
2370                                 rdev_dec_pending(rdev, mddev);
2371                         } else
2372                                 rcu_read_unlock();
2373                 }
2374                 sectors -= s;
2375                 sect += s;
2376         }
2377 }
2378
2379 static int narrow_write_error(struct r1bio *r1_bio, int i)
2380 {
2381         struct mddev *mddev = r1_bio->mddev;
2382         struct r1conf *conf = mddev->private;
2383         struct md_rdev *rdev = conf->mirrors[i].rdev;
2384
2385         /* bio has the data to be written to device 'i' where
2386          * we just recently had a write error.
2387          * We repeatedly clone the bio and trim down to one block,
2388          * then try the write.  Where the write fails we record
2389          * a bad block.
2390          * It is conceivable that the bio doesn't exactly align with
2391          * blocks.  We must handle this somehow.
2392          *
2393          * We currently own a reference on the rdev.
2394          */
2395
2396         int block_sectors;
2397         sector_t sector;
2398         int sectors;
2399         int sect_to_write = r1_bio->sectors;
2400         int ok = 1;
2401
2402         if (rdev->badblocks.shift < 0)
2403                 return 0;
2404
2405         block_sectors = roundup(1 << rdev->badblocks.shift,
2406                                 bdev_logical_block_size(rdev->bdev) >> 9);
2407         sector = r1_bio->sector;
2408         sectors = ((sector + block_sectors)
2409                    & ~(sector_t)(block_sectors - 1))
2410                 - sector;
2411
2412         while (sect_to_write) {
2413                 struct bio *wbio;
2414                 if (sectors > sect_to_write)
2415                         sectors = sect_to_write;
2416                 /* Write at 'sector' for 'sectors'*/
2417
2418                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2419                         wbio = bio_clone_fast(r1_bio->behind_master_bio,
2420                                               GFP_NOIO,
2421                                               &mddev->bio_set);
2422                 } else {
2423                         wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2424                                               &mddev->bio_set);
2425                 }
2426
2427                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2428                 wbio->bi_iter.bi_sector = r1_bio->sector;
2429                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2430
2431                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2432                 wbio->bi_iter.bi_sector += rdev->data_offset;
2433                 bio_set_dev(wbio, rdev->bdev);
2434
2435                 if (submit_bio_wait(wbio) < 0)
2436                         /* failure! */
2437                         ok = rdev_set_badblocks(rdev, sector,
2438                                                 sectors, 0)
2439                                 && ok;
2440
2441                 bio_put(wbio);
2442                 sect_to_write -= sectors;
2443                 sector += sectors;
2444                 sectors = block_sectors;
2445         }
2446         return ok;
2447 }
2448
2449 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2450 {
2451         int m;
2452         int s = r1_bio->sectors;
2453         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2454                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2455                 struct bio *bio = r1_bio->bios[m];
2456                 if (bio->bi_end_io == NULL)
2457                         continue;
2458                 if (!bio->bi_status &&
2459                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2460                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2461                 }
2462                 if (bio->bi_status &&
2463                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2464                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2465                                 md_error(conf->mddev, rdev);
2466                 }
2467         }
2468         put_buf(r1_bio);
2469         md_done_sync(conf->mddev, s, 1);
2470 }
2471
2472 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2473 {
2474         int m, idx;
2475         bool fail = false;
2476
2477         for (m = 0; m < conf->raid_disks * 2 ; m++)
2478                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2479                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2480                         rdev_clear_badblocks(rdev,
2481                                              r1_bio->sector,
2482                                              r1_bio->sectors, 0);
2483                         rdev_dec_pending(rdev, conf->mddev);
2484                 } else if (r1_bio->bios[m] != NULL) {
2485                         /* This drive got a write error.  We need to
2486                          * narrow down and record precise write
2487                          * errors.
2488                          */
2489                         fail = true;
2490                         if (!narrow_write_error(r1_bio, m)) {
2491                                 md_error(conf->mddev,
2492                                          conf->mirrors[m].rdev);
2493                                 /* an I/O failed, we can't clear the bitmap */
2494                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2495                         }
2496                         rdev_dec_pending(conf->mirrors[m].rdev,
2497                                          conf->mddev);
2498                 }
2499         if (fail) {
2500                 spin_lock_irq(&conf->device_lock);
2501                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2502                 idx = sector_to_idx(r1_bio->sector);
2503                 atomic_inc(&conf->nr_queued[idx]);
2504                 spin_unlock_irq(&conf->device_lock);
2505                 /*
2506                  * In case freeze_array() is waiting for condition
2507                  * get_unqueued_pending() == extra to be true.
2508                  */
2509                 wake_up(&conf->wait_barrier);
2510                 md_wakeup_thread(conf->mddev->thread);
2511         } else {
2512                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2513                         close_write(r1_bio);
2514                 raid_end_bio_io(r1_bio);
2515         }
2516 }
2517
2518 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2519 {
2520         struct mddev *mddev = conf->mddev;
2521         struct bio *bio;
2522         struct md_rdev *rdev;
2523
2524         clear_bit(R1BIO_ReadError, &r1_bio->state);
2525         /* we got a read error. Maybe the drive is bad.  Maybe just
2526          * the block and we can fix it.
2527          * We freeze all other IO, and try reading the block from
2528          * other devices.  When we find one, we re-write
2529          * and check it that fixes the read error.
2530          * This is all done synchronously while the array is
2531          * frozen
2532          */
2533
2534         bio = r1_bio->bios[r1_bio->read_disk];
2535         bio_put(bio);
2536         r1_bio->bios[r1_bio->read_disk] = NULL;
2537
2538         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2539         if (mddev->ro == 0
2540             && !test_bit(FailFast, &rdev->flags)) {
2541                 freeze_array(conf, 1);
2542                 fix_read_error(conf, r1_bio->read_disk,
2543                                r1_bio->sector, r1_bio->sectors);
2544                 unfreeze_array(conf);
2545         } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2546                 md_error(mddev, rdev);
2547         } else {
2548                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2549         }
2550
2551         rdev_dec_pending(rdev, conf->mddev);
2552         allow_barrier(conf, r1_bio->sector);
2553         bio = r1_bio->master_bio;
2554
2555         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2556         r1_bio->state = 0;
2557         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2558 }
2559
2560 static void raid1d(struct md_thread *thread)
2561 {
2562         struct mddev *mddev = thread->mddev;
2563         struct r1bio *r1_bio;
2564         unsigned long flags;
2565         struct r1conf *conf = mddev->private;
2566         struct list_head *head = &conf->retry_list;
2567         struct blk_plug plug;
2568         int idx;
2569
2570         md_check_recovery(mddev);
2571
2572         if (!list_empty_careful(&conf->bio_end_io_list) &&
2573             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2574                 LIST_HEAD(tmp);
2575                 spin_lock_irqsave(&conf->device_lock, flags);
2576                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2577                         list_splice_init(&conf->bio_end_io_list, &tmp);
2578                 spin_unlock_irqrestore(&conf->device_lock, flags);
2579                 while (!list_empty(&tmp)) {
2580                         r1_bio = list_first_entry(&tmp, struct r1bio,
2581                                                   retry_list);
2582                         list_del(&r1_bio->retry_list);
2583                         idx = sector_to_idx(r1_bio->sector);
2584                         atomic_dec(&conf->nr_queued[idx]);
2585                         if (mddev->degraded)
2586                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2587                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2588                                 close_write(r1_bio);
2589                         raid_end_bio_io(r1_bio);
2590                 }
2591         }
2592
2593         blk_start_plug(&plug);
2594         for (;;) {
2595
2596                 flush_pending_writes(conf);
2597
2598                 spin_lock_irqsave(&conf->device_lock, flags);
2599                 if (list_empty(head)) {
2600                         spin_unlock_irqrestore(&conf->device_lock, flags);
2601                         break;
2602                 }
2603                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2604                 list_del(head->prev);
2605                 idx = sector_to_idx(r1_bio->sector);
2606                 atomic_dec(&conf->nr_queued[idx]);
2607                 spin_unlock_irqrestore(&conf->device_lock, flags);
2608
2609                 mddev = r1_bio->mddev;
2610                 conf = mddev->private;
2611                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2612                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2613                             test_bit(R1BIO_WriteError, &r1_bio->state))
2614                                 handle_sync_write_finished(conf, r1_bio);
2615                         else
2616                                 sync_request_write(mddev, r1_bio);
2617                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2618                            test_bit(R1BIO_WriteError, &r1_bio->state))
2619                         handle_write_finished(conf, r1_bio);
2620                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2621                         handle_read_error(conf, r1_bio);
2622                 else
2623                         WARN_ON_ONCE(1);
2624
2625                 cond_resched();
2626                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2627                         md_check_recovery(mddev);
2628         }
2629         blk_finish_plug(&plug);
2630 }
2631
2632 static int init_resync(struct r1conf *conf)
2633 {
2634         int buffs;
2635
2636         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2637         BUG_ON(mempool_initialized(&conf->r1buf_pool));
2638
2639         return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2640                             r1buf_pool_free, conf->poolinfo);
2641 }
2642
2643 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2644 {
2645         struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2646         struct resync_pages *rps;
2647         struct bio *bio;
2648         int i;
2649
2650         for (i = conf->poolinfo->raid_disks; i--; ) {
2651                 bio = r1bio->bios[i];
2652                 rps = bio->bi_private;
2653                 bio_reset(bio);
2654                 bio->bi_private = rps;
2655         }
2656         r1bio->master_bio = NULL;
2657         return r1bio;
2658 }
2659
2660 /*
2661  * perform a "sync" on one "block"
2662  *
2663  * We need to make sure that no normal I/O request - particularly write
2664  * requests - conflict with active sync requests.
2665  *
2666  * This is achieved by tracking pending requests and a 'barrier' concept
2667  * that can be installed to exclude normal IO requests.
2668  */
2669
2670 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2671                                    int *skipped)
2672 {
2673         struct r1conf *conf = mddev->private;
2674         struct r1bio *r1_bio;
2675         struct bio *bio;
2676         sector_t max_sector, nr_sectors;
2677         int disk = -1;
2678         int i;
2679         int wonly = -1;
2680         int write_targets = 0, read_targets = 0;
2681         sector_t sync_blocks;
2682         int still_degraded = 0;
2683         int good_sectors = RESYNC_SECTORS;
2684         int min_bad = 0; /* number of sectors that are bad in all devices */
2685         int idx = sector_to_idx(sector_nr);
2686         int page_idx = 0;
2687
2688         if (!mempool_initialized(&conf->r1buf_pool))
2689                 if (init_resync(conf))
2690                         return 0;
2691
2692         max_sector = mddev->dev_sectors;
2693         if (sector_nr >= max_sector) {
2694                 /* If we aborted, we need to abort the
2695                  * sync on the 'current' bitmap chunk (there will
2696                  * only be one in raid1 resync.
2697                  * We can find the current addess in mddev->curr_resync
2698                  */
2699                 if (mddev->curr_resync < max_sector) /* aborted */
2700                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2701                                            &sync_blocks, 1);
2702                 else /* completed sync */
2703                         conf->fullsync = 0;
2704
2705                 md_bitmap_close_sync(mddev->bitmap);
2706                 close_sync(conf);
2707
2708                 if (mddev_is_clustered(mddev)) {
2709                         conf->cluster_sync_low = 0;
2710                         conf->cluster_sync_high = 0;
2711                 }
2712                 return 0;
2713         }
2714
2715         if (mddev->bitmap == NULL &&
2716             mddev->recovery_cp == MaxSector &&
2717             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2718             conf->fullsync == 0) {
2719                 *skipped = 1;
2720                 return max_sector - sector_nr;
2721         }
2722         /* before building a request, check if we can skip these blocks..
2723          * This call the bitmap_start_sync doesn't actually record anything
2724          */
2725         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2726             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2727                 /* We can skip this block, and probably several more */
2728                 *skipped = 1;
2729                 return sync_blocks;
2730         }
2731
2732         /*
2733          * If there is non-resync activity waiting for a turn, then let it
2734          * though before starting on this new sync request.
2735          */
2736         if (atomic_read(&conf->nr_waiting[idx]))
2737                 schedule_timeout_uninterruptible(1);
2738
2739         /* we are incrementing sector_nr below. To be safe, we check against
2740          * sector_nr + two times RESYNC_SECTORS
2741          */
2742
2743         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2744                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2745
2746
2747         if (raise_barrier(conf, sector_nr))
2748                 return 0;
2749
2750         r1_bio = raid1_alloc_init_r1buf(conf);
2751
2752         rcu_read_lock();
2753         /*
2754          * If we get a correctably read error during resync or recovery,
2755          * we might want to read from a different device.  So we
2756          * flag all drives that could conceivably be read from for READ,
2757          * and any others (which will be non-In_sync devices) for WRITE.
2758          * If a read fails, we try reading from something else for which READ
2759          * is OK.
2760          */
2761
2762         r1_bio->mddev = mddev;
2763         r1_bio->sector = sector_nr;
2764         r1_bio->state = 0;
2765         set_bit(R1BIO_IsSync, &r1_bio->state);
2766         /* make sure good_sectors won't go across barrier unit boundary */
2767         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2768
2769         for (i = 0; i < conf->raid_disks * 2; i++) {
2770                 struct md_rdev *rdev;
2771                 bio = r1_bio->bios[i];
2772
2773                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2774                 if (rdev == NULL ||
2775                     test_bit(Faulty, &rdev->flags)) {
2776                         if (i < conf->raid_disks)
2777                                 still_degraded = 1;
2778                 } else if (!test_bit(In_sync, &rdev->flags)) {
2779                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2780                         bio->bi_end_io = end_sync_write;
2781                         write_targets ++;
2782                 } else {
2783                         /* may need to read from here */
2784                         sector_t first_bad = MaxSector;
2785                         int bad_sectors;
2786
2787                         if (is_badblock(rdev, sector_nr, good_sectors,
2788                                         &first_bad, &bad_sectors)) {
2789                                 if (first_bad > sector_nr)
2790                                         good_sectors = first_bad - sector_nr;
2791                                 else {
2792                                         bad_sectors -= (sector_nr - first_bad);
2793                                         if (min_bad == 0 ||
2794                                             min_bad > bad_sectors)
2795                                                 min_bad = bad_sectors;
2796                                 }
2797                         }
2798                         if (sector_nr < first_bad) {
2799                                 if (test_bit(WriteMostly, &rdev->flags)) {
2800                                         if (wonly < 0)
2801                                                 wonly = i;
2802                                 } else {
2803                                         if (disk < 0)
2804                                                 disk = i;
2805                                 }
2806                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2807                                 bio->bi_end_io = end_sync_read;
2808                                 read_targets++;
2809                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2810                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2811                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2812                                 /*
2813                                  * The device is suitable for reading (InSync),
2814                                  * but has bad block(s) here. Let's try to correct them,
2815                                  * if we are doing resync or repair. Otherwise, leave
2816                                  * this device alone for this sync request.
2817                                  */
2818                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2819                                 bio->bi_end_io = end_sync_write;
2820                                 write_targets++;
2821                         }
2822                 }
2823                 if (rdev && bio->bi_end_io) {
2824                         atomic_inc(&rdev->nr_pending);
2825                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2826                         bio_set_dev(bio, rdev->bdev);
2827                         if (test_bit(FailFast, &rdev->flags))
2828                                 bio->bi_opf |= MD_FAILFAST;
2829                 }
2830         }
2831         rcu_read_unlock();
2832         if (disk < 0)
2833                 disk = wonly;
2834         r1_bio->read_disk = disk;
2835
2836         if (read_targets == 0 && min_bad > 0) {
2837                 /* These sectors are bad on all InSync devices, so we
2838                  * need to mark them bad on all write targets
2839                  */
2840                 int ok = 1;
2841                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2842                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2843                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2844                                 ok = rdev_set_badblocks(rdev, sector_nr,
2845                                                         min_bad, 0
2846                                         ) && ok;
2847                         }
2848                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2849                 *skipped = 1;
2850                 put_buf(r1_bio);
2851
2852                 if (!ok) {
2853                         /* Cannot record the badblocks, so need to
2854                          * abort the resync.
2855                          * If there are multiple read targets, could just
2856                          * fail the really bad ones ???
2857                          */
2858                         conf->recovery_disabled = mddev->recovery_disabled;
2859                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2860                         return 0;
2861                 } else
2862                         return min_bad;
2863
2864         }
2865         if (min_bad > 0 && min_bad < good_sectors) {
2866                 /* only resync enough to reach the next bad->good
2867                  * transition */
2868                 good_sectors = min_bad;
2869         }
2870
2871         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2872                 /* extra read targets are also write targets */
2873                 write_targets += read_targets-1;
2874
2875         if (write_targets == 0 || read_targets == 0) {
2876                 /* There is nowhere to write, so all non-sync
2877                  * drives must be failed - so we are finished
2878                  */
2879                 sector_t rv;
2880                 if (min_bad > 0)
2881                         max_sector = sector_nr + min_bad;
2882                 rv = max_sector - sector_nr;
2883                 *skipped = 1;
2884                 put_buf(r1_bio);
2885                 return rv;
2886         }
2887
2888         if (max_sector > mddev->resync_max)
2889                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2890         if (max_sector > sector_nr + good_sectors)
2891                 max_sector = sector_nr + good_sectors;
2892         nr_sectors = 0;
2893         sync_blocks = 0;
2894         do {
2895                 struct page *page;
2896                 int len = PAGE_SIZE;
2897                 if (sector_nr + (len>>9) > max_sector)
2898                         len = (max_sector - sector_nr) << 9;
2899                 if (len == 0)
2900                         break;
2901                 if (sync_blocks == 0) {
2902                         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2903                                                   &sync_blocks, still_degraded) &&
2904                             !conf->fullsync &&
2905                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2906                                 break;
2907                         if ((len >> 9) > sync_blocks)
2908                                 len = sync_blocks<<9;
2909                 }
2910
2911                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2912                         struct resync_pages *rp;
2913
2914                         bio = r1_bio->bios[i];
2915                         rp = get_resync_pages(bio);
2916                         if (bio->bi_end_io) {
2917                                 page = resync_fetch_page(rp, page_idx);
2918
2919                                 /*
2920                                  * won't fail because the vec table is big
2921                                  * enough to hold all these pages
2922                                  */
2923                                 bio_add_page(bio, page, len, 0);
2924                         }
2925                 }
2926                 nr_sectors += len>>9;
2927                 sector_nr += len>>9;
2928                 sync_blocks -= (len>>9);
2929         } while (++page_idx < RESYNC_PAGES);
2930
2931         r1_bio->sectors = nr_sectors;
2932
2933         if (mddev_is_clustered(mddev) &&
2934                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2935                 conf->cluster_sync_low = mddev->curr_resync_completed;
2936                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2937                 /* Send resync message */
2938                 md_cluster_ops->resync_info_update(mddev,
2939                                 conf->cluster_sync_low,
2940                                 conf->cluster_sync_high);
2941         }
2942
2943         /* For a user-requested sync, we read all readable devices and do a
2944          * compare
2945          */
2946         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2947                 atomic_set(&r1_bio->remaining, read_targets);
2948                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2949                         bio = r1_bio->bios[i];
2950                         if (bio->bi_end_io == end_sync_read) {
2951                                 read_targets--;
2952                                 md_sync_acct_bio(bio, nr_sectors);
2953                                 if (read_targets == 1)
2954                                         bio->bi_opf &= ~MD_FAILFAST;
2955                                 submit_bio_noacct(bio);
2956                         }
2957                 }
2958         } else {
2959                 atomic_set(&r1_bio->remaining, 1);
2960                 bio = r1_bio->bios[r1_bio->read_disk];
2961                 md_sync_acct_bio(bio, nr_sectors);
2962                 if (read_targets == 1)
2963                         bio->bi_opf &= ~MD_FAILFAST;
2964                 submit_bio_noacct(bio);
2965         }
2966         return nr_sectors;
2967 }
2968
2969 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2970 {
2971         if (sectors)
2972                 return sectors;
2973
2974         return mddev->dev_sectors;
2975 }
2976
2977 static struct r1conf *setup_conf(struct mddev *mddev)
2978 {
2979         struct r1conf *conf;
2980         int i;
2981         struct raid1_info *disk;
2982         struct md_rdev *rdev;
2983         int err = -ENOMEM;
2984
2985         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2986         if (!conf)
2987                 goto abort;
2988
2989         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2990                                    sizeof(atomic_t), GFP_KERNEL);
2991         if (!conf->nr_pending)
2992                 goto abort;
2993
2994         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2995                                    sizeof(atomic_t), GFP_KERNEL);
2996         if (!conf->nr_waiting)
2997                 goto abort;
2998
2999         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
3000                                   sizeof(atomic_t), GFP_KERNEL);
3001         if (!conf->nr_queued)
3002                 goto abort;
3003
3004         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
3005                                 sizeof(atomic_t), GFP_KERNEL);
3006         if (!conf->barrier)
3007                 goto abort;
3008
3009         conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3010                                             mddev->raid_disks, 2),
3011                                 GFP_KERNEL);
3012         if (!conf->mirrors)
3013                 goto abort;
3014
3015         conf->tmppage = alloc_page(GFP_KERNEL);
3016         if (!conf->tmppage)
3017                 goto abort;
3018
3019         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
3020         if (!conf->poolinfo)
3021                 goto abort;
3022         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
3023         err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
3024                            rbio_pool_free, conf->poolinfo);
3025         if (err)
3026                 goto abort;
3027
3028         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3029         if (err)
3030                 goto abort;
3031
3032         conf->poolinfo->mddev = mddev;
3033
3034         err = -EINVAL;
3035         spin_lock_init(&conf->device_lock);
3036         rdev_for_each(rdev, mddev) {
3037                 int disk_idx = rdev->raid_disk;
3038                 if (disk_idx >= mddev->raid_disks
3039                     || disk_idx < 0)
3040                         continue;
3041                 if (test_bit(Replacement, &rdev->flags))
3042                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
3043                 else
3044                         disk = conf->mirrors + disk_idx;
3045
3046                 if (disk->rdev)
3047                         goto abort;
3048                 disk->rdev = rdev;
3049                 disk->head_position = 0;
3050                 disk->seq_start = MaxSector;
3051         }
3052         conf->raid_disks = mddev->raid_disks;
3053         conf->mddev = mddev;
3054         INIT_LIST_HEAD(&conf->retry_list);
3055         INIT_LIST_HEAD(&conf->bio_end_io_list);
3056
3057         spin_lock_init(&conf->resync_lock);
3058         init_waitqueue_head(&conf->wait_barrier);
3059
3060         bio_list_init(&conf->pending_bio_list);
3061         conf->pending_count = 0;
3062         conf->recovery_disabled = mddev->recovery_disabled - 1;
3063
3064         err = -EIO;
3065         for (i = 0; i < conf->raid_disks * 2; i++) {
3066
3067                 disk = conf->mirrors + i;
3068
3069                 if (i < conf->raid_disks &&
3070                     disk[conf->raid_disks].rdev) {
3071                         /* This slot has a replacement. */
3072                         if (!disk->rdev) {
3073                                 /* No original, just make the replacement
3074                                  * a recovering spare
3075                                  */
3076                                 disk->rdev =
3077                                         disk[conf->raid_disks].rdev;
3078                                 disk[conf->raid_disks].rdev = NULL;
3079                         } else if (!test_bit(In_sync, &disk->rdev->flags))
3080                                 /* Original is not in_sync - bad */
3081                                 goto abort;
3082                 }
3083
3084                 if (!disk->rdev ||
3085                     !test_bit(In_sync, &disk->rdev->flags)) {
3086                         disk->head_position = 0;
3087                         if (disk->rdev &&
3088                             (disk->rdev->saved_raid_disk < 0))
3089                                 conf->fullsync = 1;
3090                 }
3091         }
3092
3093         err = -ENOMEM;
3094         conf->thread = md_register_thread(raid1d, mddev, "raid1");
3095         if (!conf->thread)
3096                 goto abort;
3097
3098         return conf;
3099
3100  abort:
3101         if (conf) {
3102                 mempool_exit(&conf->r1bio_pool);
3103                 kfree(conf->mirrors);
3104                 safe_put_page(conf->tmppage);
3105                 kfree(conf->poolinfo);
3106                 kfree(conf->nr_pending);
3107                 kfree(conf->nr_waiting);
3108                 kfree(conf->nr_queued);
3109                 kfree(conf->barrier);
3110                 bioset_exit(&conf->bio_split);
3111                 kfree(conf);
3112         }
3113         return ERR_PTR(err);
3114 }
3115
3116 static void raid1_free(struct mddev *mddev, void *priv);
3117 static int raid1_run(struct mddev *mddev)
3118 {
3119         struct r1conf *conf;
3120         int i;
3121         struct md_rdev *rdev;
3122         int ret;
3123         bool discard_supported = false;
3124
3125         if (mddev->level != 1) {
3126                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3127                         mdname(mddev), mddev->level);
3128                 return -EIO;
3129         }
3130         if (mddev->reshape_position != MaxSector) {
3131                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3132                         mdname(mddev));
3133                 return -EIO;
3134         }
3135         if (mddev_init_writes_pending(mddev) < 0)
3136                 return -ENOMEM;
3137         /*
3138          * copy the already verified devices into our private RAID1
3139          * bookkeeping area. [whatever we allocate in run(),
3140          * should be freed in raid1_free()]
3141          */
3142         if (mddev->private == NULL)
3143                 conf = setup_conf(mddev);
3144         else
3145                 conf = mddev->private;
3146
3147         if (IS_ERR(conf))
3148                 return PTR_ERR(conf);
3149
3150         if (mddev->queue) {
3151                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3152                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3153         }
3154
3155         rdev_for_each(rdev, mddev) {
3156                 if (!mddev->gendisk)
3157                         continue;
3158                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3159                                   rdev->data_offset << 9);
3160                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3161                         discard_supported = true;
3162         }
3163
3164         mddev->degraded = 0;
3165         for (i = 0; i < conf->raid_disks; i++)
3166                 if (conf->mirrors[i].rdev == NULL ||
3167                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3168                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3169                         mddev->degraded++;
3170         /*
3171          * RAID1 needs at least one disk in active
3172          */
3173         if (conf->raid_disks - mddev->degraded < 1) {
3174                 ret = -EINVAL;
3175                 goto abort;
3176         }
3177
3178         if (conf->raid_disks - mddev->degraded == 1)
3179                 mddev->recovery_cp = MaxSector;
3180
3181         if (mddev->recovery_cp != MaxSector)
3182                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3183                         mdname(mddev));
3184         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3185                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3186                 mddev->raid_disks);
3187
3188         /*
3189          * Ok, everything is just fine now
3190          */
3191         mddev->thread = conf->thread;
3192         conf->thread = NULL;
3193         mddev->private = conf;
3194         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3195
3196         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3197
3198         if (mddev->queue) {
3199                 if (discard_supported)
3200                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3201                                                 mddev->queue);
3202                 else
3203                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3204                                                   mddev->queue);
3205         }
3206
3207         ret = md_integrity_register(mddev);
3208         if (ret) {
3209                 md_unregister_thread(&mddev->thread);
3210                 goto abort;
3211         }
3212         return 0;
3213
3214 abort:
3215         raid1_free(mddev, conf);
3216         return ret;
3217 }
3218
3219 static void raid1_free(struct mddev *mddev, void *priv)
3220 {
3221         struct r1conf *conf = priv;
3222
3223         mempool_exit(&conf->r1bio_pool);
3224         kfree(conf->mirrors);
3225         safe_put_page(conf->tmppage);
3226         kfree(conf->poolinfo);
3227         kfree(conf->nr_pending);
3228         kfree(conf->nr_waiting);
3229         kfree(conf->nr_queued);
3230         kfree(conf->barrier);
3231         bioset_exit(&conf->bio_split);
3232         kfree(conf);
3233 }
3234
3235 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3236 {
3237         /* no resync is happening, and there is enough space
3238          * on all devices, so we can resize.
3239          * We need to make sure resync covers any new space.
3240          * If the array is shrinking we should possibly wait until
3241          * any io in the removed space completes, but it hardly seems
3242          * worth it.
3243          */
3244         sector_t newsize = raid1_size(mddev, sectors, 0);
3245         if (mddev->external_size &&
3246             mddev->array_sectors > newsize)
3247                 return -EINVAL;
3248         if (mddev->bitmap) {
3249                 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3250                 if (ret)
3251                         return ret;
3252         }
3253         md_set_array_sectors(mddev, newsize);
3254         if (sectors > mddev->dev_sectors &&
3255             mddev->recovery_cp > mddev->dev_sectors) {
3256                 mddev->recovery_cp = mddev->dev_sectors;
3257                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3258         }
3259         mddev->dev_sectors = sectors;
3260         mddev->resync_max_sectors = sectors;
3261         return 0;
3262 }
3263
3264 static int raid1_reshape(struct mddev *mddev)
3265 {
3266         /* We need to:
3267          * 1/ resize the r1bio_pool
3268          * 2/ resize conf->mirrors
3269          *
3270          * We allocate a new r1bio_pool if we can.
3271          * Then raise a device barrier and wait until all IO stops.
3272          * Then resize conf->mirrors and swap in the new r1bio pool.
3273          *
3274          * At the same time, we "pack" the devices so that all the missing
3275          * devices have the higher raid_disk numbers.
3276          */
3277         mempool_t newpool, oldpool;
3278         struct pool_info *newpoolinfo;
3279         struct raid1_info *newmirrors;
3280         struct r1conf *conf = mddev->private;
3281         int cnt, raid_disks;
3282         unsigned long flags;
3283         int d, d2;
3284         int ret;
3285
3286         memset(&newpool, 0, sizeof(newpool));
3287         memset(&oldpool, 0, sizeof(oldpool));
3288
3289         /* Cannot change chunk_size, layout, or level */
3290         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3291             mddev->layout != mddev->new_layout ||
3292             mddev->level != mddev->new_level) {
3293                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3294                 mddev->new_layout = mddev->layout;
3295                 mddev->new_level = mddev->level;
3296                 return -EINVAL;
3297         }
3298
3299         if (!mddev_is_clustered(mddev))
3300                 md_allow_write(mddev);
3301
3302         raid_disks = mddev->raid_disks + mddev->delta_disks;
3303
3304         if (raid_disks < conf->raid_disks) {
3305                 cnt=0;
3306                 for (d= 0; d < conf->raid_disks; d++)
3307                         if (conf->mirrors[d].rdev)
3308                                 cnt++;
3309                 if (cnt > raid_disks)
3310                         return -EBUSY;
3311         }
3312
3313         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3314         if (!newpoolinfo)
3315                 return -ENOMEM;
3316         newpoolinfo->mddev = mddev;
3317         newpoolinfo->raid_disks = raid_disks * 2;
3318
3319         ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3320                            rbio_pool_free, newpoolinfo);
3321         if (ret) {
3322                 kfree(newpoolinfo);
3323                 return ret;
3324         }
3325         newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3326                                          raid_disks, 2),
3327                              GFP_KERNEL);
3328         if (!newmirrors) {
3329                 kfree(newpoolinfo);
3330                 mempool_exit(&newpool);
3331                 return -ENOMEM;
3332         }
3333
3334         freeze_array(conf, 0);
3335
3336         /* ok, everything is stopped */
3337         oldpool = conf->r1bio_pool;
3338         conf->r1bio_pool = newpool;
3339
3340         for (d = d2 = 0; d < conf->raid_disks; d++) {
3341                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3342                 if (rdev && rdev->raid_disk != d2) {
3343                         sysfs_unlink_rdev(mddev, rdev);
3344                         rdev->raid_disk = d2;
3345                         sysfs_unlink_rdev(mddev, rdev);
3346                         if (sysfs_link_rdev(mddev, rdev))
3347                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3348                                         mdname(mddev), rdev->raid_disk);
3349                 }
3350                 if (rdev)
3351                         newmirrors[d2++].rdev = rdev;
3352         }
3353         kfree(conf->mirrors);
3354         conf->mirrors = newmirrors;
3355         kfree(conf->poolinfo);
3356         conf->poolinfo = newpoolinfo;
3357
3358         spin_lock_irqsave(&conf->device_lock, flags);
3359         mddev->degraded += (raid_disks - conf->raid_disks);
3360         spin_unlock_irqrestore(&conf->device_lock, flags);
3361         conf->raid_disks = mddev->raid_disks = raid_disks;
3362         mddev->delta_disks = 0;
3363
3364         unfreeze_array(conf);
3365
3366         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3367         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3368         md_wakeup_thread(mddev->thread);
3369
3370         mempool_exit(&oldpool);
3371         return 0;
3372 }
3373
3374 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3375 {
3376         struct r1conf *conf = mddev->private;
3377
3378         if (quiesce)
3379                 freeze_array(conf, 0);
3380         else
3381                 unfreeze_array(conf);
3382 }
3383
3384 static void *raid1_takeover(struct mddev *mddev)
3385 {
3386         /* raid1 can take over:
3387          *  raid5 with 2 devices, any layout or chunk size
3388          */
3389         if (mddev->level == 5 && mddev->raid_disks == 2) {
3390                 struct r1conf *conf;
3391                 mddev->new_level = 1;
3392                 mddev->new_layout = 0;
3393                 mddev->new_chunk_sectors = 0;
3394                 conf = setup_conf(mddev);
3395                 if (!IS_ERR(conf)) {
3396                         /* Array must appear to be quiesced */
3397                         conf->array_frozen = 1;
3398                         mddev_clear_unsupported_flags(mddev,
3399                                 UNSUPPORTED_MDDEV_FLAGS);
3400                 }
3401                 return conf;
3402         }
3403         return ERR_PTR(-EINVAL);
3404 }
3405
3406 static struct md_personality raid1_personality =
3407 {
3408         .name           = "raid1",
3409         .level          = 1,
3410         .owner          = THIS_MODULE,
3411         .make_request   = raid1_make_request,
3412         .run            = raid1_run,
3413         .free           = raid1_free,
3414         .status         = raid1_status,
3415         .error_handler  = raid1_error,
3416         .hot_add_disk   = raid1_add_disk,
3417         .hot_remove_disk= raid1_remove_disk,
3418         .spare_active   = raid1_spare_active,
3419         .sync_request   = raid1_sync_request,
3420         .resize         = raid1_resize,
3421         .size           = raid1_size,
3422         .check_reshape  = raid1_reshape,
3423         .quiesce        = raid1_quiesce,
3424         .takeover       = raid1_takeover,
3425 };
3426
3427 static int __init raid_init(void)
3428 {
3429         return register_md_personality(&raid1_personality);
3430 }
3431
3432 static void raid_exit(void)
3433 {
3434         unregister_md_personality(&raid1_personality);
3435 }
3436
3437 module_init(raid_init);
3438 module_exit(raid_exit);
3439 MODULE_LICENSE("GPL");
3440 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3441 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3442 MODULE_ALIAS("md-raid1");
3443 MODULE_ALIAS("md-level-1");