drivers/md/raid1.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * raid1.c : Multiple Devices driver for Linux
   4  *
   5  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   6  *
   7  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   8  *
   9  * RAID-1 management functions.
  10  *
  11  * Better read-balancing code written by Mika Kuoppala <[email protected]>, 2000
  12  *
  13  * Fixes to reconstruction by Jakob Østergaard" <[email protected]>
  14  * Various fixes by Neil Brown <[email protected]>
  15  *
  16  * Changes by Peter T. Breuer <[email protected]> 31/1/2003 to support
  17  * bitmapped intelligence in resync:
  18  *
  19  *      - bitmap marked during normal i/o
  20  *      - bitmap used to skip nondirty blocks during sync
  21  *
  22  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  23  * - persistent bitmap code
  24  */
  25
  26 #include <linux/slab.h>
  27 #include <linux/delay.h>
  28 #include <linux/blkdev.h>
  29 #include <linux/module.h>
  30 #include <linux/seq_file.h>
  31 #include <linux/ratelimit.h>
  32 #include <linux/interval_tree_generic.h>
  33
  34 #include <trace/events/block.h>
  35
  36 #include "md.h"
  37 #include "raid1.h"
  38 #include "md-bitmap.h"
  39
  40 #define UNSUPPORTED_MDDEV_FLAGS         \
  41         ((1L << MD_HAS_JOURNAL) |       \
  42          (1L << MD_JOURNAL_CLEAN) |     \
  43          (1L << MD_HAS_PPL) |           \
  44          (1L << MD_HAS_MULTIPLE_PPLS))
  45
  46 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
  47 static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
  48
  49 #define 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                  * For failfast, this is the only remaining device,
 478                  * We need to retry the write without FailFast.
 479                  */
 480                 if (!test_bit(Faulty, &rdev->flags))
 481                         set_bit(R1BIO_WriteError, &r1_bio->state);
 482                 else {
 483                         /* Fail the request */
 484                         set_bit(R1BIO_Degraded, &r1_bio->state);
 485                         /* Finished with this branch */
 486                         r1_bio->bios[mirror] = NULL;
 487                         to_put = bio;
 488                 }
 489         } else {
 490                 /*
 491                  * Set R1BIO_Uptodate in our master bio, so that we
 492                  * will return a good error code for to the higher
 493                  * levels even if IO on some other mirrored buffer
 494                  * fails.
 495                  *
 496                  * The 'master' represents the composite IO operation
 497                  * to user-side. So if something waits for IO, then it
 498                  * will wait for the 'master' bio.
 499                  */
 500                 sector_t first_bad;
 501                 int bad_sectors;
 502
 503                 r1_bio->bios[mirror] = NULL;
 504                 to_put = bio;
 505                 /*
 506                  * Do not set R1BIO_Uptodate if the current device is
 507                  * rebuilding or Faulty. This is because we cannot use
 508                  * such device for properly reading the data back (we could
 509                  * potentially use it, if the current write would have felt
 510                  * before rdev->recovery_offset, but for simplicity we don't
 511                  * check this here.
 512                  */
 513                 if (test_bit(In_sync, &rdev->flags) &&
 514                     !test_bit(Faulty, &rdev->flags))
 515                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 516
 517                 /* Maybe we can clear some bad blocks. */
 518                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 519                                 &first_bad, &bad_sectors) && !discard_error) {
 520                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 521                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 522                 }
 523         }
 524
 525         if (behind) {
 526                 if (test_bit(CollisionCheck, &rdev->flags))
 527                         remove_serial(rdev, lo, hi);
 528                 if (test_bit(WriteMostly, &rdev->flags))
 529                         atomic_dec(&r1_bio->behind_remaining);
 530
 531                 /*
 532                  * In behind mode, we ACK the master bio once the I/O
 533                  * has safely reached all non-writemostly
 534                  * disks. Setting the Returned bit ensures that this
 535                  * gets done only once -- we don't ever want to return
 536                  * -EIO here, instead we'll wait
 537                  */
 538                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 539                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 540                         /* Maybe we can return now */
 541                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 542                                 struct bio *mbio = r1_bio->master_bio;
 543                                 pr_debug("raid1: behind end write sectors"
 544                                          " %llu-%llu\n",
 545                                          (unsigned long long) mbio->bi_iter.bi_sector,
 546                                          (unsigned long long) bio_end_sector(mbio) - 1);
 547                                 call_bio_endio(r1_bio);
 548                         }
 549                 }
 550         } else if (rdev->mddev->serialize_policy)
 551                 remove_serial(rdev, lo, hi);
 552         if (r1_bio->bios[mirror] == NULL)
 553                 rdev_dec_pending(rdev, conf->mddev);
 554
 555         /*
 556          * Let's see if all mirrored write operations have finished
 557          * already.
 558          */
 559         r1_bio_write_done(r1_bio);
 560
 561         if (to_put)
 562                 bio_put(to_put);
 563 }
 564
 565 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 566                                           sector_t sectors)
 567 {
 568         sector_t len;
 569
 570         WARN_ON(sectors == 0);
 571         /*
 572          * len is the number of sectors from start_sector to end of the
 573          * barrier unit which start_sector belongs to.
 574          */
 575         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 576               start_sector;
 577
 578         if (len > sectors)
 579                 len = sectors;
 580
 581         return len;
 582 }
 583
 584 /*
 585  * This routine returns the disk from which the requested read should
 586  * be done. There is a per-array 'next expected sequential IO' sector
 587  * number - if this matches on the next IO then we use the last disk.
 588  * There is also a per-disk 'last know head position' sector that is
 589  * maintained from IRQ contexts, both the normal and the resync IO
 590  * completion handlers update this position correctly. If there is no
 591  * perfect sequential match then we pick the disk whose head is closest.
 592  *
 593  * If there are 2 mirrors in the same 2 devices, performance degrades
 594  * because position is mirror, not device based.
 595  *
 596  * The rdev for the device selected will have nr_pending incremented.
 597  */
 598 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 599 {
 600         const sector_t this_sector = r1_bio->sector;
 601         int sectors;
 602         int best_good_sectors;
 603         int best_disk, best_dist_disk, best_pending_disk;
 604         int has_nonrot_disk;
 605         int disk;
 606         sector_t best_dist;
 607         unsigned int min_pending;
 608         struct md_rdev *rdev;
 609         int choose_first;
 610         int choose_next_idle;
 611
 612         rcu_read_lock();
 613         /*
 614          * Check if we can balance. We can balance on the whole
 615          * device if no resync is going on, or below the resync window.
 616          * We take the first readable disk when above the resync window.
 617          */
 618  retry:
 619         sectors = r1_bio->sectors;
 620         best_disk = -1;
 621         best_dist_disk = -1;
 622         best_dist = MaxSector;
 623         best_pending_disk = -1;
 624         min_pending = UINT_MAX;
 625         best_good_sectors = 0;
 626         has_nonrot_disk = 0;
 627         choose_next_idle = 0;
 628         clear_bit(R1BIO_FailFast, &r1_bio->state);
 629
 630         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 631             (mddev_is_clustered(conf->mddev) &&
 632             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 633                     this_sector + sectors)))
 634                 choose_first = 1;
 635         else
 636                 choose_first = 0;
 637
 638         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 639                 sector_t dist;
 640                 sector_t first_bad;
 641                 int bad_sectors;
 642                 unsigned int pending;
 643                 bool nonrot;
 644
 645                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 646                 if (r1_bio->bios[disk] == IO_BLOCKED
 647                     || rdev == NULL
 648                     || test_bit(Faulty, &rdev->flags))
 649                         continue;
 650                 if (!test_bit(In_sync, &rdev->flags) &&
 651                     rdev->recovery_offset < this_sector + sectors)
 652                         continue;
 653                 if (test_bit(WriteMostly, &rdev->flags)) {
 654                         /* Don't balance among write-mostly, just
 655                          * use the first as a last resort */
 656                         if (best_dist_disk < 0) {
 657                                 if (is_badblock(rdev, this_sector, sectors,
 658                                                 &first_bad, &bad_sectors)) {
 659                                         if (first_bad <= this_sector)
 660                                                 /* Cannot use this */
 661                                                 continue;
 662                                         best_good_sectors = first_bad - this_sector;
 663                                 } else
 664                                         best_good_sectors = sectors;
 665                                 best_dist_disk = disk;
 666                                 best_pending_disk = disk;
 667                         }
 668                         continue;
 669                 }
 670                 /* This is a reasonable device to use.  It might
 671                  * even be best.
 672                  */
 673                 if (is_badblock(rdev, this_sector, sectors,
 674                                 &first_bad, &bad_sectors)) {
 675                         if (best_dist < MaxSector)
 676                                 /* already have a better device */
 677                                 continue;
 678                         if (first_bad <= this_sector) {
 679                                 /* cannot read here. If this is the 'primary'
 680                                  * device, then we must not read beyond
 681                                  * bad_sectors from another device..
 682                                  */
 683                                 bad_sectors -= (this_sector - first_bad);
 684                                 if (choose_first && sectors > bad_sectors)
 685                                         sectors = bad_sectors;
 686                                 if (best_good_sectors > sectors)
 687                                         best_good_sectors = sectors;
 688
 689                         } else {
 690                                 sector_t good_sectors = first_bad - this_sector;
 691                                 if (good_sectors > best_good_sectors) {
 692                                         best_good_sectors = good_sectors;
 693                                         best_disk = disk;
 694                                 }
 695                                 if (choose_first)
 696                                         break;
 697                         }
 698                         continue;
 699                 } else {
 700                         if ((sectors > best_good_sectors) && (best_disk >= 0))
 701                                 best_disk = -1;
 702                         best_good_sectors = sectors;
 703                 }
 704
 705                 if (best_disk >= 0)
 706                         /* At least two disks to choose from so failfast is OK */
 707                         set_bit(R1BIO_FailFast, &r1_bio->state);
 708
 709                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 710                 has_nonrot_disk |= nonrot;
 711                 pending = atomic_read(&rdev->nr_pending);
 712                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 713                 if (choose_first) {
 714                         best_disk = disk;
 715                         break;
 716                 }
 717                 /* Don't change to another disk for sequential reads */
 718                 if (conf->mirrors[disk].next_seq_sect == this_sector
 719                     || dist == 0) {
 720                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 721                         struct raid1_info *mirror = &conf->mirrors[disk];
 722
 723                         best_disk = disk;
 724                         /*
 725                          * If buffered sequential IO size exceeds optimal
 726                          * iosize, check if there is idle disk. If yes, choose
 727                          * the idle disk. read_balance could already choose an
 728                          * idle disk before noticing it's a sequential IO in
 729                          * this disk. This doesn't matter because this disk
 730                          * will idle, next time it will be utilized after the
 731                          * first disk has IO size exceeds optimal iosize. In
 732                          * this way, iosize of the first disk will be optimal
 733                          * iosize at least. iosize of the second disk might be
 734                          * small, but not a big deal since when the second disk
 735                          * starts IO, the first disk is likely still busy.
 736                          */
 737                         if (nonrot && opt_iosize > 0 &&
 738                             mirror->seq_start != MaxSector &&
 739                             mirror->next_seq_sect > opt_iosize &&
 740                             mirror->next_seq_sect - opt_iosize >=
 741                             mirror->seq_start) {
 742                                 choose_next_idle = 1;
 743                                 continue;
 744                         }
 745                         break;
 746                 }
 747
 748                 if (choose_next_idle)
 749                         continue;
 750
 751                 if (min_pending > pending) {
 752                         min_pending = pending;
 753                         best_pending_disk = disk;
 754                 }
 755
 756                 if (dist < best_dist) {
 757                         best_dist = dist;
 758                         best_dist_disk = disk;
 759                 }
 760         }
 761
 762         /*
 763          * If all disks are rotational, choose the closest disk. If any disk is
 764          * non-rotational, choose the disk with less pending request even the
 765          * disk is rotational, which might/might not be optimal for raids with
 766          * mixed ratation/non-rotational disks depending on workload.
 767          */
 768         if (best_disk == -1) {
 769                 if (has_nonrot_disk || min_pending == 0)
 770                         best_disk = best_pending_disk;
 771                 else
 772                         best_disk = best_dist_disk;
 773         }
 774
 775         if (best_disk >= 0) {
 776                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 777                 if (!rdev)
 778                         goto retry;
 779                 atomic_inc(&rdev->nr_pending);
 780                 sectors = best_good_sectors;
 781
 782                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 783                         conf->mirrors[best_disk].seq_start = this_sector;
 784
 785                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 786         }
 787         rcu_read_unlock();
 788         *max_sectors = sectors;
 789
 790         return best_disk;
 791 }
 792
 793 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 794 {
 795         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 796         md_bitmap_unplug(conf->mddev->bitmap);
 797         wake_up(&conf->wait_barrier);
 798
 799         while (bio) { /* submit pending writes */
 800                 struct bio *next = bio->bi_next;
 801                 struct md_rdev *rdev = (void *)bio->bi_bdev;
 802                 bio->bi_next = NULL;
 803                 bio_set_dev(bio, rdev->bdev);
 804                 if (test_bit(Faulty, &rdev->flags)) {
 805                         bio_io_error(bio);
 806                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 807                                     !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
 808                         /* Just ignore it */
 809                         bio_endio(bio);
 810                 else
 811                         submit_bio_noacct(bio);
 812                 bio = next;
 813                 cond_resched();
 814         }
 815 }
 816
 817 static void flush_pending_writes(struct r1conf *conf)
 818 {
 819         /* Any writes that have been queued but are awaiting
 820          * bitmap updates get flushed here.
 821          */
 822         spin_lock_irq(&conf->device_lock);
 823
 824         if (conf->pending_bio_list.head) {
 825                 struct blk_plug plug;
 826                 struct bio *bio;
 827
 828                 bio = bio_list_get(&conf->pending_bio_list);
 829                 conf->pending_count = 0;
 830                 spin_unlock_irq(&conf->device_lock);
 831
 832                 /*
 833                  * As this is called in a wait_event() loop (see freeze_array),
 834                  * current->state might be TASK_UNINTERRUPTIBLE which will
 835                  * cause a warning when we prepare to wait again.  As it is
 836                  * rare that this path is taken, it is perfectly safe to force
 837                  * us to go around the wait_event() loop again, so the warning
 838                  * is a false-positive.  Silence the warning by resetting
 839                  * thread state
 840                  */
 841                 __set_current_state(TASK_RUNNING);
 842                 blk_start_plug(&plug);
 843                 flush_bio_list(conf, bio);
 844                 blk_finish_plug(&plug);
 845         } else
 846                 spin_unlock_irq(&conf->device_lock);
 847 }
 848
 849 /* Barriers....
 850  * Sometimes we need to suspend IO while we do something else,
 851  * either some resync/recovery, or reconfigure the array.
 852  * To do this we raise a 'barrier'.
 853  * The 'barrier' is a counter that can be raised multiple times
 854  * to count how many activities are happening which preclude
 855  * normal IO.
 856  * We can only raise the barrier if there is no pending IO.
 857  * i.e. if nr_pending == 0.
 858  * We choose only to raise the barrier if no-one is waiting for the
 859  * barrier to go down.  This means that as soon as an IO request
 860  * is ready, no other operations which require a barrier will start
 861  * until the IO request has had a chance.
 862  *
 863  * So: regular IO calls 'wait_barrier'.  When that returns there
 864  *    is no backgroup IO happening,  It must arrange to call
 865  *    allow_barrier when it has finished its IO.
 866  * backgroup IO calls must call raise_barrier.  Once that returns
 867  *    there is no normal IO happeing.  It must arrange to call
 868  *    lower_barrier when the particular background IO completes.
 869  *
 870  * If resync/recovery is interrupted, returns -EINTR;
 871  * Otherwise, returns 0.
 872  */
 873 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 874 {
 875         int idx = sector_to_idx(sector_nr);
 876
 877         spin_lock_irq(&conf->resync_lock);
 878
 879         /* Wait until no block IO is waiting */
 880         wait_event_lock_irq(conf->wait_barrier,
 881                             !atomic_read(&conf->nr_waiting[idx]),
 882                             conf->resync_lock);
 883
 884         /* block any new IO from starting */
 885         atomic_inc(&conf->barrier[idx]);
 886         /*
 887          * In raise_barrier() we firstly increase conf->barrier[idx] then
 888          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 889          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 890          * A memory barrier here to make sure conf->nr_pending[idx] won't
 891          * be fetched before conf->barrier[idx] is increased. Otherwise
 892          * there will be a race between raise_barrier() and _wait_barrier().
 893          */
 894         smp_mb__after_atomic();
 895
 896         /* For these conditions we must wait:
 897          * A: while the array is in frozen state
 898          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 899          *    existing in corresponding I/O barrier bucket.
 900          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 901          *    max resync count which allowed on current I/O barrier bucket.
 902          */
 903         wait_event_lock_irq(conf->wait_barrier,
 904                             (!conf->array_frozen &&
 905                              !atomic_read(&conf->nr_pending[idx]) &&
 906                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 907                                 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 908                             conf->resync_lock);
 909
 910         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 911                 atomic_dec(&conf->barrier[idx]);
 912                 spin_unlock_irq(&conf->resync_lock);
 913                 wake_up(&conf->wait_barrier);
 914                 return -EINTR;
 915         }
 916
 917         atomic_inc(&conf->nr_sync_pending);
 918         spin_unlock_irq(&conf->resync_lock);
 919
 920         return 0;
 921 }
 922
 923 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 924 {
 925         int idx = sector_to_idx(sector_nr);
 926
 927         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 928
 929         atomic_dec(&conf->barrier[idx]);
 930         atomic_dec(&conf->nr_sync_pending);
 931         wake_up(&conf->wait_barrier);
 932 }
 933
 934 static void _wait_barrier(struct r1conf *conf, int idx)
 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;
 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         wait_event_lock_irq(conf->wait_barrier,
 985                             !conf->array_frozen &&
 986                              !atomic_read(&conf->barrier[idx]),
 987                             conf->resync_lock);
 988         atomic_inc(&conf->nr_pending[idx]);
 989         atomic_dec(&conf->nr_waiting[idx]);
 990         spin_unlock_irq(&conf->resync_lock);
 991 }
 992
 993 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
 994 {
 995         int idx = sector_to_idx(sector_nr);
 996
 997         /*
 998          * Very similar to _wait_barrier(). The difference is, for read
 999          * I/O we don't need wait for sync I/O, but if the whole array
1000          * is frozen, the read I/O still has to wait until the array is
1001          * unfrozen. Since there is no ordering requirement with
1002          * conf->barrier[idx] here, memory barrier is unnecessary as well.
1003          */
1004         atomic_inc(&conf->nr_pending[idx]);
1005
1006         if (!READ_ONCE(conf->array_frozen))
1007                 return;
1008
1009         spin_lock_irq(&conf->resync_lock);
1010         atomic_inc(&conf->nr_waiting[idx]);
1011         atomic_dec(&conf->nr_pending[idx]);
1012         /*
1013          * In case freeze_array() is waiting for
1014          * get_unqueued_pending() == extra
1015          */
1016         wake_up(&conf->wait_barrier);
1017         /* Wait for array to be unfrozen */
1018         wait_event_lock_irq(conf->wait_barrier,
1019                             !conf->array_frozen,
1020                             conf->resync_lock);
1021         atomic_inc(&conf->nr_pending[idx]);
1022         atomic_dec(&conf->nr_waiting[idx]);
1023         spin_unlock_irq(&conf->resync_lock);
1024 }
1025
1026 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1027 {
1028         int idx = sector_to_idx(sector_nr);
1029
1030         _wait_barrier(conf, idx);
1031 }
1032
1033 static void _allow_barrier(struct r1conf *conf, int idx)
1034 {
1035         atomic_dec(&conf->nr_pending[idx]);
1036         wake_up(&conf->wait_barrier);
1037 }
1038
1039 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1040 {
1041         int idx = sector_to_idx(sector_nr);
1042
1043         _allow_barrier(conf, idx);
1044 }
1045
1046 /* conf->resync_lock should be held */
1047 static int get_unqueued_pending(struct r1conf *conf)
1048 {
1049         int idx, ret;
1050
1051         ret = atomic_read(&conf->nr_sync_pending);
1052         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1053                 ret += atomic_read(&conf->nr_pending[idx]) -
1054                         atomic_read(&conf->nr_queued[idx]);
1055
1056         return ret;
1057 }
1058
1059 static void freeze_array(struct r1conf *conf, int extra)
1060 {
1061         /* Stop sync I/O and normal I/O and wait for everything to
1062          * go quiet.
1063          * This is called in two situations:
1064          * 1) management command handlers (reshape, remove disk, quiesce).
1065          * 2) one normal I/O request failed.
1066
1067          * After array_frozen is set to 1, new sync IO will be blocked at
1068          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1069          * or wait_read_barrier(). The flying I/Os will either complete or be
1070          * queued. When everything goes quite, there are only queued I/Os left.
1071
1072          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1073          * barrier bucket index which this I/O request hits. When all sync and
1074          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1075          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1076          * in handle_read_error(), we may call freeze_array() before trying to
1077          * fix the read error. In this case, the error read I/O is not queued,
1078          * so get_unqueued_pending() == 1.
1079          *
1080          * Therefore before this function returns, we need to wait until
1081          * get_unqueued_pendings(conf) gets equal to extra. For
1082          * normal I/O context, extra is 1, in rested situations extra is 0.
1083          */
1084         spin_lock_irq(&conf->resync_lock);
1085         conf->array_frozen = 1;
1086         raid1_log(conf->mddev, "wait freeze");
1087         wait_event_lock_irq_cmd(
1088                 conf->wait_barrier,
1089                 get_unqueued_pending(conf) == extra,
1090                 conf->resync_lock,
1091                 flush_pending_writes(conf));
1092         spin_unlock_irq(&conf->resync_lock);
1093 }
1094 static void unfreeze_array(struct r1conf *conf)
1095 {
1096         /* reverse the effect of the freeze */
1097         spin_lock_irq(&conf->resync_lock);
1098         conf->array_frozen = 0;
1099         spin_unlock_irq(&conf->resync_lock);
1100         wake_up(&conf->wait_barrier);
1101 }
1102
1103 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1104                                            struct bio *bio)
1105 {
1106         int size = bio->bi_iter.bi_size;
1107         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1108         int i = 0;
1109         struct bio *behind_bio = NULL;
1110
1111         behind_bio = bio_alloc_bioset(GFP_NOIO, vcnt, &r1_bio->mddev->bio_set);
1112         if (!behind_bio)
1113                 return;
1114
1115         /* discard op, we don't support writezero/writesame yet */
1116         if (!bio_has_data(bio)) {
1117                 behind_bio->bi_iter.bi_size = size;
1118                 goto skip_copy;
1119         }
1120
1121         behind_bio->bi_write_hint = bio->bi_write_hint;
1122
1123         while (i < vcnt && size) {
1124                 struct page *page;
1125                 int len = min_t(int, PAGE_SIZE, size);
1126
1127                 page = alloc_page(GFP_NOIO);
1128                 if (unlikely(!page))
1129                         goto free_pages;
1130
1131                 bio_add_page(behind_bio, page, len, 0);
1132
1133                 size -= len;
1134                 i++;
1135         }
1136
1137         bio_copy_data(behind_bio, bio);
1138 skip_copy:
1139         r1_bio->behind_master_bio = behind_bio;
1140         set_bit(R1BIO_BehindIO, &r1_bio->state);
1141
1142         return;
1143
1144 free_pages:
1145         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1146                  bio->bi_iter.bi_size);
1147         bio_free_pages(behind_bio);
1148         bio_put(behind_bio);
1149 }
1150
1151 struct raid1_plug_cb {
1152         struct blk_plug_cb      cb;
1153         struct bio_list         pending;
1154         int                     pending_cnt;
1155 };
1156
1157 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1158 {
1159         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1160                                                   cb);
1161         struct mddev *mddev = plug->cb.data;
1162         struct r1conf *conf = mddev->private;
1163         struct bio *bio;
1164
1165         if (from_schedule || current->bio_list) {
1166                 spin_lock_irq(&conf->device_lock);
1167                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1168                 conf->pending_count += plug->pending_cnt;
1169                 spin_unlock_irq(&conf->device_lock);
1170                 wake_up(&conf->wait_barrier);
1171                 md_wakeup_thread(mddev->thread);
1172                 kfree(plug);
1173                 return;
1174         }
1175
1176         /* we aren't scheduling, so we can do the write-out directly. */
1177         bio = bio_list_get(&plug->pending);
1178         flush_bio_list(conf, bio);
1179         kfree(plug);
1180 }
1181
1182 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1183 {
1184         r1_bio->master_bio = bio;
1185         r1_bio->sectors = bio_sectors(bio);
1186         r1_bio->state = 0;
1187         r1_bio->mddev = mddev;
1188         r1_bio->sector = bio->bi_iter.bi_sector;
1189 }
1190
1191 static inline struct r1bio *
1192 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1193 {
1194         struct r1conf *conf = mddev->private;
1195         struct r1bio *r1_bio;
1196
1197         r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1198         /* Ensure no bio records IO_BLOCKED */
1199         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1200         init_r1bio(r1_bio, mddev, bio);
1201         return r1_bio;
1202 }
1203
1204 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1205                                int max_read_sectors, struct r1bio *r1_bio)
1206 {
1207         struct r1conf *conf = mddev->private;
1208         struct raid1_info *mirror;
1209         struct bio *read_bio;
1210         struct bitmap *bitmap = mddev->bitmap;
1211         const int op = bio_op(bio);
1212         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1213         int max_sectors;
1214         int rdisk;
1215         bool r1bio_existed = !!r1_bio;
1216         char b[BDEVNAME_SIZE];
1217
1218         /*
1219          * If r1_bio is set, we are blocking the raid1d thread
1220          * so there is a tiny risk of deadlock.  So ask for
1221          * emergency memory if needed.
1222          */
1223         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1224
1225         if (r1bio_existed) {
1226                 /* Need to get the block device name carefully */
1227                 struct md_rdev *rdev;
1228                 rcu_read_lock();
1229                 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1230                 if (rdev)
1231                         bdevname(rdev->bdev, b);
1232                 else
1233                         strcpy(b, "???");
1234                 rcu_read_unlock();
1235         }
1236
1237         /*
1238          * Still need barrier for READ in case that whole
1239          * array is frozen.
1240          */
1241         wait_read_barrier(conf, bio->bi_iter.bi_sector);
1242
1243         if (!r1_bio)
1244                 r1_bio = alloc_r1bio(mddev, bio);
1245         else
1246                 init_r1bio(r1_bio, mddev, bio);
1247         r1_bio->sectors = max_read_sectors;
1248
1249         /*
1250          * make_request() can abort the operation when read-ahead is being
1251          * used and no empty request is available.
1252          */
1253         rdisk = read_balance(conf, r1_bio, &max_sectors);
1254
1255         if (rdisk < 0) {
1256                 /* couldn't find anywhere to read from */
1257                 if (r1bio_existed) {
1258                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1259                                             mdname(mddev),
1260                                             b,
1261                                             (unsigned long long)r1_bio->sector);
1262                 }
1263                 raid_end_bio_io(r1_bio);
1264                 return;
1265         }
1266         mirror = conf->mirrors + rdisk;
1267
1268         if (r1bio_existed)
1269                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1270                                     mdname(mddev),
1271                                     (unsigned long long)r1_bio->sector,
1272                                     bdevname(mirror->rdev->bdev, b));
1273
1274         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1275             bitmap) {
1276                 /*
1277                  * Reading from a write-mostly device must take care not to
1278                  * over-take any writes that are 'behind'
1279                  */
1280                 raid1_log(mddev, "wait behind writes");
1281                 wait_event(bitmap->behind_wait,
1282                            atomic_read(&bitmap->behind_writes) == 0);
1283         }
1284
1285         if (max_sectors < bio_sectors(bio)) {
1286                 struct bio *split = bio_split(bio, max_sectors,
1287                                               gfp, &conf->bio_split);
1288                 bio_chain(split, bio);
1289                 submit_bio_noacct(bio);
1290                 bio = split;
1291                 r1_bio->master_bio = bio;
1292                 r1_bio->sectors = max_sectors;
1293         }
1294
1295         r1_bio->read_disk = rdisk;
1296
1297         if (!r1bio_existed && blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1298                 r1_bio->start_time = bio_start_io_acct(bio);
1299
1300         read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1301
1302         r1_bio->bios[rdisk] = read_bio;
1303
1304         read_bio->bi_iter.bi_sector = r1_bio->sector +
1305                 mirror->rdev->data_offset;
1306         bio_set_dev(read_bio, mirror->rdev->bdev);
1307         read_bio->bi_end_io = raid1_end_read_request;
1308         bio_set_op_attrs(read_bio, op, do_sync);
1309         if (test_bit(FailFast, &mirror->rdev->flags) &&
1310             test_bit(R1BIO_FailFast, &r1_bio->state))
1311                 read_bio->bi_opf |= MD_FAILFAST;
1312         read_bio->bi_private = r1_bio;
1313
1314         if (mddev->gendisk)
1315                 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1316                                       r1_bio->sector);
1317
1318         submit_bio_noacct(read_bio);
1319 }
1320
1321 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1322                                 int max_write_sectors)
1323 {
1324         struct r1conf *conf = mddev->private;
1325         struct r1bio *r1_bio;
1326         int i, disks;
1327         struct bitmap *bitmap = mddev->bitmap;
1328         unsigned long flags;
1329         struct md_rdev *blocked_rdev;
1330         struct blk_plug_cb *cb;
1331         struct raid1_plug_cb *plug = NULL;
1332         int first_clone;
1333         int max_sectors;
1334
1335         if (mddev_is_clustered(mddev) &&
1336              md_cluster_ops->area_resyncing(mddev, WRITE,
1337                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1338
1339                 DEFINE_WAIT(w);
1340                 for (;;) {
1341                         prepare_to_wait(&conf->wait_barrier,
1342                                         &w, TASK_IDLE);
1343                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1344                                                         bio->bi_iter.bi_sector,
1345                                                         bio_end_sector(bio)))
1346                                 break;
1347                         schedule();
1348                 }
1349                 finish_wait(&conf->wait_barrier, &w);
1350         }
1351
1352         /*
1353          * Register the new request and wait if the reconstruction
1354          * thread has put up a bar for new requests.
1355          * Continue immediately if no resync is active currently.
1356          */
1357         wait_barrier(conf, bio->bi_iter.bi_sector);
1358
1359         r1_bio = alloc_r1bio(mddev, bio);
1360         r1_bio->sectors = max_write_sectors;
1361
1362         if (conf->pending_count >= max_queued_requests) {
1363                 md_wakeup_thread(mddev->thread);
1364                 raid1_log(mddev, "wait queued");
1365                 wait_event(conf->wait_barrier,
1366                            conf->pending_count < max_queued_requests);
1367         }
1368         /* first select target devices under rcu_lock and
1369          * inc refcount on their rdev.  Record them by setting
1370          * bios[x] to bio
1371          * If there are known/acknowledged bad blocks on any device on
1372          * which we have seen a write error, we want to avoid writing those
1373          * blocks.
1374          * This potentially requires several writes to write around
1375          * the bad blocks.  Each set of writes gets it's own r1bio
1376          * with a set of bios attached.
1377          */
1378
1379         disks = conf->raid_disks * 2;
1380  retry_write:
1381         blocked_rdev = NULL;
1382         rcu_read_lock();
1383         max_sectors = r1_bio->sectors;
1384         for (i = 0;  i < disks; i++) {
1385                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1386                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1387                         atomic_inc(&rdev->nr_pending);
1388                         blocked_rdev = rdev;
1389                         break;
1390                 }
1391                 r1_bio->bios[i] = NULL;
1392                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1393                         if (i < conf->raid_disks)
1394                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1395                         continue;
1396                 }
1397
1398                 atomic_inc(&rdev->nr_pending);
1399                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1400                         sector_t first_bad;
1401                         int bad_sectors;
1402                         int is_bad;
1403
1404                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1405                                              &first_bad, &bad_sectors);
1406                         if (is_bad < 0) {
1407                                 /* mustn't write here until the bad block is
1408                                  * acknowledged*/
1409                                 set_bit(BlockedBadBlocks, &rdev->flags);
1410                                 blocked_rdev = rdev;
1411                                 break;
1412                         }
1413                         if (is_bad && first_bad <= r1_bio->sector) {
1414                                 /* Cannot write here at all */
1415                                 bad_sectors -= (r1_bio->sector - first_bad);
1416                                 if (bad_sectors < max_sectors)
1417                                         /* mustn't write more than bad_sectors
1418                                          * to other devices yet
1419                                          */
1420                                         max_sectors = bad_sectors;
1421                                 rdev_dec_pending(rdev, mddev);
1422                                 /* We don't set R1BIO_Degraded as that
1423                                  * only applies if the disk is
1424                                  * missing, so it might be re-added,
1425                                  * and we want to know to recover this
1426                                  * chunk.
1427                                  * In this case the device is here,
1428                                  * and the fact that this chunk is not
1429                                  * in-sync is recorded in the bad
1430                                  * block log
1431                                  */
1432                                 continue;
1433                         }
1434                         if (is_bad) {
1435                                 int good_sectors = first_bad - r1_bio->sector;
1436                                 if (good_sectors < max_sectors)
1437                                         max_sectors = good_sectors;
1438                         }
1439                 }
1440                 r1_bio->bios[i] = bio;
1441         }
1442         rcu_read_unlock();
1443
1444         if (unlikely(blocked_rdev)) {
1445                 /* Wait for this device to become unblocked */
1446                 int j;
1447
1448                 for (j = 0; j < i; j++)
1449                         if (r1_bio->bios[j])
1450                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1451                 r1_bio->state = 0;
1452                 allow_barrier(conf, bio->bi_iter.bi_sector);
1453                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1454                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1455                 wait_barrier(conf, bio->bi_iter.bi_sector);
1456                 goto retry_write;
1457         }
1458
1459         if (max_sectors < bio_sectors(bio)) {
1460                 struct bio *split = bio_split(bio, max_sectors,
1461                                               GFP_NOIO, &conf->bio_split);
1462                 bio_chain(split, bio);
1463                 submit_bio_noacct(bio);
1464                 bio = split;
1465                 r1_bio->master_bio = bio;
1466                 r1_bio->sectors = max_sectors;
1467         }
1468
1469         if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1470                 r1_bio->start_time = bio_start_io_acct(bio);
1471         atomic_set(&r1_bio->remaining, 1);
1472         atomic_set(&r1_bio->behind_remaining, 0);
1473
1474         first_clone = 1;
1475
1476         for (i = 0; i < disks; i++) {
1477                 struct bio *mbio = NULL;
1478                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1479                 if (!r1_bio->bios[i])
1480                         continue;
1481
1482                 if (first_clone) {
1483                         /* do behind I/O ?
1484                          * Not if there are too many, or cannot
1485                          * allocate memory, or a reader on WriteMostly
1486                          * is waiting for behind writes to flush */
1487                         if (bitmap &&
1488                             (atomic_read(&bitmap->behind_writes)
1489                              < mddev->bitmap_info.max_write_behind) &&
1490                             !waitqueue_active(&bitmap->behind_wait)) {
1491                                 alloc_behind_master_bio(r1_bio, bio);
1492                         }
1493
1494                         md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1495                                              test_bit(R1BIO_BehindIO, &r1_bio->state));
1496                         first_clone = 0;
1497                 }
1498
1499                 if (r1_bio->behind_master_bio)
1500                         mbio = bio_clone_fast(r1_bio->behind_master_bio,
1501                                               GFP_NOIO, &mddev->bio_set);
1502                 else
1503                         mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1504
1505                 if (r1_bio->behind_master_bio) {
1506                         if (test_bit(CollisionCheck, &rdev->flags))
1507                                 wait_for_serialization(rdev, r1_bio);
1508                         if (test_bit(WriteMostly, &rdev->flags))
1509                                 atomic_inc(&r1_bio->behind_remaining);
1510                 } else if (mddev->serialize_policy)
1511                         wait_for_serialization(rdev, r1_bio);
1512
1513                 r1_bio->bios[i] = mbio;
1514
1515                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1516                                    conf->mirrors[i].rdev->data_offset);
1517                 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1518                 mbio->bi_end_io = raid1_end_write_request;
1519                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1520                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1521                     !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1522                     conf->raid_disks - mddev->degraded > 1)
1523                         mbio->bi_opf |= MD_FAILFAST;
1524                 mbio->bi_private = r1_bio;
1525
1526                 atomic_inc(&r1_bio->remaining);
1527
1528                 if (mddev->gendisk)
1529                         trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1530                                               r1_bio->sector);
1531                 /* flush_pending_writes() needs access to the rdev so...*/
1532                 mbio->bi_bdev = (void *)conf->mirrors[i].rdev;
1533
1534                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1535                 if (cb)
1536                         plug = container_of(cb, struct raid1_plug_cb, cb);
1537                 else
1538                         plug = NULL;
1539                 if (plug) {
1540                         bio_list_add(&plug->pending, mbio);
1541                         plug->pending_cnt++;
1542                 } else {
1543                         spin_lock_irqsave(&conf->device_lock, flags);
1544                         bio_list_add(&conf->pending_bio_list, mbio);
1545                         conf->pending_count++;
1546                         spin_unlock_irqrestore(&conf->device_lock, flags);
1547                         md_wakeup_thread(mddev->thread);
1548                 }
1549         }
1550
1551         r1_bio_write_done(r1_bio);
1552
1553         /* In case raid1d snuck in to freeze_array */
1554         wake_up(&conf->wait_barrier);
1555 }
1556
1557 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1558 {
1559         sector_t sectors;
1560
1561         if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1562             && md_flush_request(mddev, bio))
1563                 return true;
1564
1565         /*
1566          * There is a limit to the maximum size, but
1567          * the read/write handler might find a lower limit
1568          * due to bad blocks.  To avoid multiple splits,
1569          * we pass the maximum number of sectors down
1570          * and let the lower level perform the split.
1571          */
1572         sectors = align_to_barrier_unit_end(
1573                 bio->bi_iter.bi_sector, bio_sectors(bio));
1574
1575         if (bio_data_dir(bio) == READ)
1576                 raid1_read_request(mddev, bio, sectors, NULL);
1577         else {
1578                 if (!md_write_start(mddev,bio))
1579                         return false;
1580                 raid1_write_request(mddev, bio, sectors);
1581         }
1582         return true;
1583 }
1584
1585 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1586 {
1587         struct r1conf *conf = mddev->private;
1588         int i;
1589
1590         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1591                    conf->raid_disks - mddev->degraded);
1592         rcu_read_lock();
1593         for (i = 0; i < conf->raid_disks; i++) {
1594                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1595                 seq_printf(seq, "%s",
1596                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1597         }
1598         rcu_read_unlock();
1599         seq_printf(seq, "]");
1600 }
1601
1602 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1603 {
1604         char b[BDEVNAME_SIZE];
1605         struct r1conf *conf = mddev->private;
1606         unsigned long flags;
1607
1608         /*
1609          * If it is not operational, then we have already marked it as dead
1610          * else if it is the last working disks with "fail_last_dev == false",
1611          * ignore the error, let the next level up know.
1612          * else mark the drive as failed
1613          */
1614         spin_lock_irqsave(&conf->device_lock, flags);
1615         if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1616             && (conf->raid_disks - mddev->degraded) == 1) {
1617                 /*
1618                  * Don't fail the drive, act as though we were just a
1619                  * normal single drive.
1620                  * However don't try a recovery from this drive as
1621                  * it is very likely to fail.
1622                  */
1623                 conf->recovery_disabled = mddev->recovery_disabled;
1624                 spin_unlock_irqrestore(&conf->device_lock, flags);
1625                 return;
1626         }
1627         set_bit(Blocked, &rdev->flags);
1628         if (test_and_clear_bit(In_sync, &rdev->flags))
1629                 mddev->degraded++;
1630         set_bit(Faulty, &rdev->flags);
1631         spin_unlock_irqrestore(&conf->device_lock, flags);
1632         /*
1633          * if recovery is running, make sure it aborts.
1634          */
1635         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1636         set_mask_bits(&mddev->sb_flags, 0,
1637                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1638         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1639                 "md/raid1:%s: Operation continuing on %d devices.\n",
1640                 mdname(mddev), bdevname(rdev->bdev, b),
1641                 mdname(mddev), conf->raid_disks - mddev->degraded);
1642 }
1643
1644 static void print_conf(struct r1conf *conf)
1645 {
1646         int i;
1647
1648         pr_debug("RAID1 conf printout:\n");
1649         if (!conf) {
1650                 pr_debug("(!conf)\n");
1651                 return;
1652         }
1653         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1654                  conf->raid_disks);
1655
1656         rcu_read_lock();
1657         for (i = 0; i < conf->raid_disks; i++) {
1658                 char b[BDEVNAME_SIZE];
1659                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1660                 if (rdev)
1661                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1662                                  i, !test_bit(In_sync, &rdev->flags),
1663                                  !test_bit(Faulty, &rdev->flags),
1664                                  bdevname(rdev->bdev,b));
1665         }
1666         rcu_read_unlock();
1667 }
1668
1669 static void close_sync(struct r1conf *conf)
1670 {
1671         int idx;
1672
1673         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1674                 _wait_barrier(conf, idx);
1675                 _allow_barrier(conf, idx);
1676         }
1677
1678         mempool_exit(&conf->r1buf_pool);
1679 }
1680
1681 static int raid1_spare_active(struct mddev *mddev)
1682 {
1683         int i;
1684         struct r1conf *conf = mddev->private;
1685         int count = 0;
1686         unsigned long flags;
1687
1688         /*
1689          * Find all failed disks within the RAID1 configuration
1690          * and mark them readable.
1691          * Called under mddev lock, so rcu protection not needed.
1692          * device_lock used to avoid races with raid1_end_read_request
1693          * which expects 'In_sync' flags and ->degraded to be consistent.
1694          */
1695         spin_lock_irqsave(&conf->device_lock, flags);
1696         for (i = 0; i < conf->raid_disks; i++) {
1697                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1698                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1699                 if (repl
1700                     && !test_bit(Candidate, &repl->flags)
1701                     && repl->recovery_offset == MaxSector
1702                     && !test_bit(Faulty, &repl->flags)
1703                     && !test_and_set_bit(In_sync, &repl->flags)) {
1704                         /* replacement has just become active */
1705                         if (!rdev ||
1706                             !test_and_clear_bit(In_sync, &rdev->flags))
1707                                 count++;
1708                         if (rdev) {
1709                                 /* Replaced device not technically
1710                                  * faulty, but we need to be sure
1711                                  * it gets removed and never re-added
1712                                  */
1713                                 set_bit(Faulty, &rdev->flags);
1714                                 sysfs_notify_dirent_safe(
1715                                         rdev->sysfs_state);
1716                         }
1717                 }
1718                 if (rdev
1719                     && rdev->recovery_offset == MaxSector
1720                     && !test_bit(Faulty, &rdev->flags)
1721                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1722                         count++;
1723                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1724                 }
1725         }
1726         mddev->degraded -= count;
1727         spin_unlock_irqrestore(&conf->device_lock, flags);
1728
1729         print_conf(conf);
1730         return count;
1731 }
1732
1733 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1734 {
1735         struct r1conf *conf = mddev->private;
1736         int err = -EEXIST;
1737         int mirror = 0;
1738         struct raid1_info *p;
1739         int first = 0;
1740         int last = conf->raid_disks - 1;
1741
1742         if (mddev->recovery_disabled == conf->recovery_disabled)
1743                 return -EBUSY;
1744
1745         if (md_integrity_add_rdev(rdev, mddev))
1746                 return -ENXIO;
1747
1748         if (rdev->raid_disk >= 0)
1749                 first = last = rdev->raid_disk;
1750
1751         /*
1752          * find the disk ... but prefer rdev->saved_raid_disk
1753          * if possible.
1754          */
1755         if (rdev->saved_raid_disk >= 0 &&
1756             rdev->saved_raid_disk >= first &&
1757             rdev->saved_raid_disk < conf->raid_disks &&
1758             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1759                 first = last = rdev->saved_raid_disk;
1760
1761         for (mirror = first; mirror <= last; mirror++) {
1762                 p = conf->mirrors + mirror;
1763                 if (!p->rdev) {
1764                         if (mddev->gendisk)
1765                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1766                                                   rdev->data_offset << 9);
1767
1768                         p->head_position = 0;
1769                         rdev->raid_disk = mirror;
1770                         err = 0;
1771                         /* As all devices are equivalent, we don't need a full recovery
1772                          * if this was recently any drive of the array
1773                          */
1774                         if (rdev->saved_raid_disk < 0)
1775                                 conf->fullsync = 1;
1776                         rcu_assign_pointer(p->rdev, rdev);
1777                         break;
1778                 }
1779                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1780                     p[conf->raid_disks].rdev == NULL) {
1781                         /* Add this device as a replacement */
1782                         clear_bit(In_sync, &rdev->flags);
1783                         set_bit(Replacement, &rdev->flags);
1784                         rdev->raid_disk = mirror;
1785                         err = 0;
1786                         conf->fullsync = 1;
1787                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1788                         break;
1789                 }
1790         }
1791         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1792                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1793         print_conf(conf);
1794         return err;
1795 }
1796
1797 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1798 {
1799         struct r1conf *conf = mddev->private;
1800         int err = 0;
1801         int number = rdev->raid_disk;
1802         struct raid1_info *p = conf->mirrors + number;
1803
1804         if (rdev != p->rdev)
1805                 p = conf->mirrors + conf->raid_disks + number;
1806
1807         print_conf(conf);
1808         if (rdev == p->rdev) {
1809                 if (test_bit(In_sync, &rdev->flags) ||
1810                     atomic_read(&rdev->nr_pending)) {
1811                         err = -EBUSY;
1812                         goto abort;
1813                 }
1814                 /* Only remove non-faulty devices if recovery
1815                  * is not possible.
1816                  */
1817                 if (!test_bit(Faulty, &rdev->flags) &&
1818                     mddev->recovery_disabled != conf->recovery_disabled &&
1819                     mddev->degraded < conf->raid_disks) {
1820                         err = -EBUSY;
1821                         goto abort;
1822                 }
1823                 p->rdev = NULL;
1824                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1825                         synchronize_rcu();
1826                         if (atomic_read(&rdev->nr_pending)) {
1827                                 /* lost the race, try later */
1828                                 err = -EBUSY;
1829                                 p->rdev = rdev;
1830                                 goto abort;
1831                         }
1832                 }
1833                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1834                         /* We just removed a device that is being replaced.
1835                          * Move down the replacement.  We drain all IO before
1836                          * doing this to avoid confusion.
1837                          */
1838                         struct md_rdev *repl =
1839                                 conf->mirrors[conf->raid_disks + number].rdev;
1840                         freeze_array(conf, 0);
1841                         if (atomic_read(&repl->nr_pending)) {
1842                                 /* It means that some queued IO of retry_list
1843                                  * hold repl. Thus, we cannot set replacement
1844                                  * as NULL, avoiding rdev NULL pointer
1845                                  * dereference in sync_request_write and
1846                                  * handle_write_finished.
1847                                  */
1848                                 err = -EBUSY;
1849                                 unfreeze_array(conf);
1850                                 goto abort;
1851                         }
1852                         clear_bit(Replacement, &repl->flags);
1853                         p->rdev = repl;
1854                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1855                         unfreeze_array(conf);
1856                 }
1857
1858                 clear_bit(WantReplacement, &rdev->flags);
1859                 err = md_integrity_register(mddev);
1860         }
1861 abort:
1862
1863         print_conf(conf);
1864         return err;
1865 }
1866
1867 static void end_sync_read(struct bio *bio)
1868 {
1869         struct r1bio *r1_bio = get_resync_r1bio(bio);
1870
1871         update_head_pos(r1_bio->read_disk, r1_bio);
1872
1873         /*
1874          * we have read a block, now it needs to be re-written,
1875          * or re-read if the read failed.
1876          * We don't do much here, just schedule handling by raid1d
1877          */
1878         if (!bio->bi_status)
1879                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1880
1881         if (atomic_dec_and_test(&r1_bio->remaining))
1882                 reschedule_retry(r1_bio);
1883 }
1884
1885 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1886 {
1887         sector_t sync_blocks = 0;
1888         sector_t s = r1_bio->sector;
1889         long sectors_to_go = r1_bio->sectors;
1890
1891         /* make sure these bits don't get cleared. */
1892         do {
1893                 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1894                 s += sync_blocks;
1895                 sectors_to_go -= sync_blocks;
1896         } while (sectors_to_go > 0);
1897 }
1898
1899 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1900 {
1901         if (atomic_dec_and_test(&r1_bio->remaining)) {
1902                 struct mddev *mddev = r1_bio->mddev;
1903                 int s = r1_bio->sectors;
1904
1905                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1906                     test_bit(R1BIO_WriteError, &r1_bio->state))
1907                         reschedule_retry(r1_bio);
1908                 else {
1909                         put_buf(r1_bio);
1910                         md_done_sync(mddev, s, uptodate);
1911                 }
1912         }
1913 }
1914
1915 static void end_sync_write(struct bio *bio)
1916 {
1917         int uptodate = !bio->bi_status;
1918         struct r1bio *r1_bio = get_resync_r1bio(bio);
1919         struct mddev *mddev = r1_bio->mddev;
1920         struct r1conf *conf = mddev->private;
1921         sector_t first_bad;
1922         int bad_sectors;
1923         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1924
1925         if (!uptodate) {
1926                 abort_sync_write(mddev, r1_bio);
1927                 set_bit(WriteErrorSeen, &rdev->flags);
1928                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1929                         set_bit(MD_RECOVERY_NEEDED, &
1930                                 mddev->recovery);
1931                 set_bit(R1BIO_WriteError, &r1_bio->state);
1932         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1933                                &first_bad, &bad_sectors) &&
1934                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1935                                 r1_bio->sector,
1936                                 r1_bio->sectors,
1937                                 &first_bad, &bad_sectors)
1938                 )
1939                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1940
1941         put_sync_write_buf(r1_bio, uptodate);
1942 }
1943
1944 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1945                             int sectors, struct page *page, int rw)
1946 {
1947         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1948                 /* success */
1949                 return 1;
1950         if (rw == WRITE) {
1951                 set_bit(WriteErrorSeen, &rdev->flags);
1952                 if (!test_and_set_bit(WantReplacement,
1953                                       &rdev->flags))
1954                         set_bit(MD_RECOVERY_NEEDED, &
1955                                 rdev->mddev->recovery);
1956         }
1957         /* need to record an error - either for the block or the device */
1958         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1959                 md_error(rdev->mddev, rdev);
1960         return 0;
1961 }
1962
1963 static int fix_sync_read_error(struct r1bio *r1_bio)
1964 {
1965         /* Try some synchronous reads of other devices to get
1966          * good data, much like with normal read errors.  Only
1967          * read into the pages we already have so we don't
1968          * need to re-issue the read request.
1969          * We don't need to freeze the array, because being in an
1970          * active sync request, there is no normal IO, and
1971          * no overlapping syncs.
1972          * We don't need to check is_badblock() again as we
1973          * made sure that anything with a bad block in range
1974          * will have bi_end_io clear.
1975          */
1976         struct mddev *mddev = r1_bio->mddev;
1977         struct r1conf *conf = mddev->private;
1978         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1979         struct page **pages = get_resync_pages(bio)->pages;
1980         sector_t sect = r1_bio->sector;
1981         int sectors = r1_bio->sectors;
1982         int idx = 0;
1983         struct md_rdev *rdev;
1984
1985         rdev = conf->mirrors[r1_bio->read_disk].rdev;
1986         if (test_bit(FailFast, &rdev->flags)) {
1987                 /* Don't try recovering from here - just fail it
1988                  * ... unless it is the last working device of course */
1989                 md_error(mddev, rdev);
1990                 if (test_bit(Faulty, &rdev->flags))
1991                         /* Don't try to read from here, but make sure
1992                          * put_buf does it's thing
1993                          */
1994                         bio->bi_end_io = end_sync_write;
1995         }
1996
1997         while(sectors) {
1998                 int s = sectors;
1999                 int d = r1_bio->read_disk;
2000                 int success = 0;
2001                 int start;
2002
2003                 if (s > (PAGE_SIZE>>9))
2004                         s = PAGE_SIZE >> 9;
2005                 do {
2006                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2007                                 /* No rcu protection needed here devices
2008                                  * can only be removed when no resync is
2009                                  * active, and resync is currently active
2010                                  */
2011                                 rdev = conf->mirrors[d].rdev;
2012                                 if (sync_page_io(rdev, sect, s<<9,
2013                                                  pages[idx],
2014                                                  REQ_OP_READ, 0, false)) {
2015                                         success = 1;
2016                                         break;
2017                                 }
2018                         }
2019                         d++;
2020                         if (d == conf->raid_disks * 2)
2021                                 d = 0;
2022                 } while (!success && d != r1_bio->read_disk);
2023
2024                 if (!success) {
2025                         char b[BDEVNAME_SIZE];
2026                         int abort = 0;
2027                         /* Cannot read from anywhere, this block is lost.
2028                          * Record a bad block on each device.  If that doesn't
2029                          * work just disable and interrupt the recovery.
2030                          * Don't fail devices as that won't really help.
2031                          */
2032                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2033                                             mdname(mddev), bio_devname(bio, b),
2034                                             (unsigned long long)r1_bio->sector);
2035                         for (d = 0; d < conf->raid_disks * 2; d++) {
2036                                 rdev = conf->mirrors[d].rdev;
2037                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2038                                         continue;
2039                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2040                                         abort = 1;
2041                         }
2042                         if (abort) {
2043                                 conf->recovery_disabled =
2044                                         mddev->recovery_disabled;
2045                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2046                                 md_done_sync(mddev, r1_bio->sectors, 0);
2047                                 put_buf(r1_bio);
2048                                 return 0;
2049                         }
2050                         /* Try next page */
2051                         sectors -= s;
2052                         sect += s;
2053                         idx++;
2054                         continue;
2055                 }
2056
2057                 start = d;
2058                 /* write it back and re-read */
2059                 while (d != r1_bio->read_disk) {
2060                         if (d == 0)
2061                                 d = conf->raid_disks * 2;
2062                         d--;
2063                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2064                                 continue;
2065                         rdev = conf->mirrors[d].rdev;
2066                         if (r1_sync_page_io(rdev, sect, s,
2067                                             pages[idx],
2068                                             WRITE) == 0) {
2069                                 r1_bio->bios[d]->bi_end_io = NULL;
2070                                 rdev_dec_pending(rdev, mddev);
2071                         }
2072                 }
2073                 d = start;
2074                 while (d != r1_bio->read_disk) {
2075                         if (d == 0)
2076                                 d = conf->raid_disks * 2;
2077                         d--;
2078                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2079                                 continue;
2080                         rdev = conf->mirrors[d].rdev;
2081                         if (r1_sync_page_io(rdev, sect, s,
2082                                             pages[idx],
2083                                             READ) != 0)
2084                                 atomic_add(s, &rdev->corrected_errors);
2085                 }
2086                 sectors -= s;
2087                 sect += s;
2088                 idx ++;
2089         }
2090         set_bit(R1BIO_Uptodate, &r1_bio->state);
2091         bio->bi_status = 0;
2092         return 1;
2093 }
2094
2095 static void process_checks(struct r1bio *r1_bio)
2096 {
2097         /* We have read all readable devices.  If we haven't
2098          * got the block, then there is no hope left.
2099          * If we have, then we want to do a comparison
2100          * and skip the write if everything is the same.
2101          * If any blocks failed to read, then we need to
2102          * attempt an over-write
2103          */
2104         struct mddev *mddev = r1_bio->mddev;
2105         struct r1conf *conf = mddev->private;
2106         int primary;
2107         int i;
2108         int vcnt;
2109
2110         /* Fix variable parts of all bios */
2111         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2112         for (i = 0; i < conf->raid_disks * 2; i++) {
2113                 blk_status_t status;
2114                 struct bio *b = r1_bio->bios[i];
2115                 struct resync_pages *rp = get_resync_pages(b);
2116                 if (b->bi_end_io != end_sync_read)
2117                         continue;
2118                 /* fixup the bio for reuse, but preserve errno */
2119                 status = b->bi_status;
2120                 bio_reset(b);
2121                 b->bi_status = status;
2122                 b->bi_iter.bi_sector = r1_bio->sector +
2123                         conf->mirrors[i].rdev->data_offset;
2124                 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2125                 b->bi_end_io = end_sync_read;
2126                 rp->raid_bio = r1_bio;
2127                 b->bi_private = rp;
2128
2129                 /* initialize bvec table again */
2130                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2131         }
2132         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2133                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2134                     !r1_bio->bios[primary]->bi_status) {
2135                         r1_bio->bios[primary]->bi_end_io = NULL;
2136                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2137                         break;
2138                 }
2139         r1_bio->read_disk = primary;
2140         for (i = 0; i < conf->raid_disks * 2; i++) {
2141                 int j = 0;
2142                 struct bio *pbio = r1_bio->bios[primary];
2143                 struct bio *sbio = r1_bio->bios[i];
2144                 blk_status_t status = sbio->bi_status;
2145                 struct page **ppages = get_resync_pages(pbio)->pages;
2146                 struct page **spages = get_resync_pages(sbio)->pages;
2147                 struct bio_vec *bi;
2148                 int page_len[RESYNC_PAGES] = { 0 };
2149                 struct bvec_iter_all iter_all;
2150
2151                 if (sbio->bi_end_io != end_sync_read)
2152                         continue;
2153                 /* Now we can 'fixup' the error value */
2154                 sbio->bi_status = 0;
2155
2156                 bio_for_each_segment_all(bi, sbio, iter_all)
2157                         page_len[j++] = bi->bv_len;
2158
2159                 if (!status) {
2160                         for (j = vcnt; j-- ; ) {
2161                                 if (memcmp(page_address(ppages[j]),
2162                                            page_address(spages[j]),
2163                                            page_len[j]))
2164                                         break;
2165                         }
2166                 } else
2167                         j = 0;
2168                 if (j >= 0)
2169                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2170                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2171                               && !status)) {
2172                         /* No need to write to this device. */
2173                         sbio->bi_end_io = NULL;
2174                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2175                         continue;
2176                 }
2177
2178                 bio_copy_data(sbio, pbio);
2179         }
2180 }
2181
2182 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2183 {
2184         struct r1conf *conf = mddev->private;
2185         int i;
2186         int disks = conf->raid_disks * 2;
2187         struct bio *wbio;
2188
2189         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2190                 /* ouch - failed to read all of that. */
2191                 if (!fix_sync_read_error(r1_bio))
2192                         return;
2193
2194         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2195                 process_checks(r1_bio);
2196
2197         /*
2198          * schedule writes
2199          */
2200         atomic_set(&r1_bio->remaining, 1);
2201         for (i = 0; i < disks ; i++) {
2202                 wbio = r1_bio->bios[i];
2203                 if (wbio->bi_end_io == NULL ||
2204                     (wbio->bi_end_io == end_sync_read &&
2205                      (i == r1_bio->read_disk ||
2206                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2207                         continue;
2208                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2209                         abort_sync_write(mddev, r1_bio);
2210                         continue;
2211                 }
2212
2213                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2214                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2215                         wbio->bi_opf |= MD_FAILFAST;
2216
2217                 wbio->bi_end_io = end_sync_write;
2218                 atomic_inc(&r1_bio->remaining);
2219                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2220
2221                 submit_bio_noacct(wbio);
2222         }
2223
2224         put_sync_write_buf(r1_bio, 1);
2225 }
2226
2227 /*
2228  * This is a kernel thread which:
2229  *
2230  *      1.      Retries failed read operations on working mirrors.
2231  *      2.      Updates the raid superblock when problems encounter.
2232  *      3.      Performs writes following reads for array synchronising.
2233  */
2234
2235 static void fix_read_error(struct r1conf *conf, int read_disk,
2236                            sector_t sect, int sectors)
2237 {
2238         struct mddev *mddev = conf->mddev;
2239         while(sectors) {
2240                 int s = sectors;
2241                 int d = read_disk;
2242                 int success = 0;
2243                 int start;
2244                 struct md_rdev *rdev;
2245
2246                 if (s > (PAGE_SIZE>>9))
2247                         s = PAGE_SIZE >> 9;
2248
2249                 do {
2250                         sector_t first_bad;
2251                         int bad_sectors;
2252
2253                         rcu_read_lock();
2254                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2255                         if (rdev &&
2256                             (test_bit(In_sync, &rdev->flags) ||
2257                              (!test_bit(Faulty, &rdev->flags) &&
2258                               rdev->recovery_offset >= sect + s)) &&
2259                             is_badblock(rdev, sect, s,
2260                                         &first_bad, &bad_sectors) == 0) {
2261                                 atomic_inc(&rdev->nr_pending);
2262                                 rcu_read_unlock();
2263                                 if (sync_page_io(rdev, sect, s<<9,
2264                                          conf->tmppage, REQ_OP_READ, 0, false))
2265                                         success = 1;
2266                                 rdev_dec_pending(rdev, mddev);
2267                                 if (success)
2268                                         break;
2269                         } else
2270                                 rcu_read_unlock();
2271                         d++;
2272                         if (d == conf->raid_disks * 2)
2273                                 d = 0;
2274                 } while (!success && d != read_disk);
2275
2276                 if (!success) {
2277                         /* Cannot read from anywhere - mark it bad */
2278                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2279                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2280                                 md_error(mddev, rdev);
2281                         break;
2282                 }
2283                 /* write it back and re-read */
2284                 start = d;
2285                 while (d != read_disk) {
2286                         if (d==0)
2287                                 d = conf->raid_disks * 2;
2288                         d--;
2289                         rcu_read_lock();
2290                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2291                         if (rdev &&
2292                             !test_bit(Faulty, &rdev->flags)) {
2293                                 atomic_inc(&rdev->nr_pending);
2294                                 rcu_read_unlock();
2295                                 r1_sync_page_io(rdev, sect, s,
2296                                                 conf->tmppage, WRITE);
2297                                 rdev_dec_pending(rdev, mddev);
2298                         } else
2299                                 rcu_read_unlock();
2300                 }
2301                 d = start;
2302                 while (d != read_disk) {
2303                         char b[BDEVNAME_SIZE];
2304                         if (d==0)
2305                                 d = conf->raid_disks * 2;
2306                         d--;
2307                         rcu_read_lock();
2308                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2309                         if (rdev &&
2310                             !test_bit(Faulty, &rdev->flags)) {
2311                                 atomic_inc(&rdev->nr_pending);
2312                                 rcu_read_unlock();
2313                                 if (r1_sync_page_io(rdev, sect, s,
2314                                                     conf->tmppage, READ)) {
2315                                         atomic_add(s, &rdev->corrected_errors);
2316                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2317                                                 mdname(mddev), s,
2318                                                 (unsigned long long)(sect +
2319                                                                      rdev->data_offset),
2320                                                 bdevname(rdev->bdev, b));
2321                                 }
2322                                 rdev_dec_pending(rdev, mddev);
2323                         } else
2324                                 rcu_read_unlock();
2325                 }
2326                 sectors -= s;
2327                 sect += s;
2328         }
2329 }
2330
2331 static int narrow_write_error(struct r1bio *r1_bio, int i)
2332 {
2333         struct mddev *mddev = r1_bio->mddev;
2334         struct r1conf *conf = mddev->private;
2335         struct md_rdev *rdev = conf->mirrors[i].rdev;
2336
2337         /* bio has the data to be written to device 'i' where
2338          * we just recently had a write error.
2339          * We repeatedly clone the bio and trim down to one block,
2340          * then try the write.  Where the write fails we record
2341          * a bad block.
2342          * It is conceivable that the bio doesn't exactly align with
2343          * blocks.  We must handle this somehow.
2344          *
2345          * We currently own a reference on the rdev.
2346          */
2347
2348         int block_sectors;
2349         sector_t sector;
2350         int sectors;
2351         int sect_to_write = r1_bio->sectors;
2352         int ok = 1;
2353
2354         if (rdev->badblocks.shift < 0)
2355                 return 0;
2356
2357         block_sectors = roundup(1 << rdev->badblocks.shift,
2358                                 bdev_logical_block_size(rdev->bdev) >> 9);
2359         sector = r1_bio->sector;
2360         sectors = ((sector + block_sectors)
2361                    & ~(sector_t)(block_sectors - 1))
2362                 - sector;
2363
2364         while (sect_to_write) {
2365                 struct bio *wbio;
2366                 if (sectors > sect_to_write)
2367                         sectors = sect_to_write;
2368                 /* Write at 'sector' for 'sectors'*/
2369
2370                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2371                         wbio = bio_clone_fast(r1_bio->behind_master_bio,
2372                                               GFP_NOIO,
2373                                               &mddev->bio_set);
2374                 } else {
2375                         wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2376                                               &mddev->bio_set);
2377                 }
2378
2379                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2380                 wbio->bi_iter.bi_sector = r1_bio->sector;
2381                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2382
2383                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2384                 wbio->bi_iter.bi_sector += rdev->data_offset;
2385                 bio_set_dev(wbio, rdev->bdev);
2386
2387                 if (submit_bio_wait(wbio) < 0)
2388                         /* failure! */
2389                         ok = rdev_set_badblocks(rdev, sector,
2390                                                 sectors, 0)
2391                                 && ok;
2392
2393                 bio_put(wbio);
2394                 sect_to_write -= sectors;
2395                 sector += sectors;
2396                 sectors = block_sectors;
2397         }
2398         return ok;
2399 }
2400
2401 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2402 {
2403         int m;
2404         int s = r1_bio->sectors;
2405         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2406                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2407                 struct bio *bio = r1_bio->bios[m];
2408                 if (bio->bi_end_io == NULL)
2409                         continue;
2410                 if (!bio->bi_status &&
2411                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2412                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2413                 }
2414                 if (bio->bi_status &&
2415                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2416                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2417                                 md_error(conf->mddev, rdev);
2418                 }
2419         }
2420         put_buf(r1_bio);
2421         md_done_sync(conf->mddev, s, 1);
2422 }
2423
2424 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2425 {
2426         int m, idx;
2427         bool fail = false;
2428
2429         for (m = 0; m < conf->raid_disks * 2 ; m++)
2430                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2431                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2432                         rdev_clear_badblocks(rdev,
2433                                              r1_bio->sector,
2434                                              r1_bio->sectors, 0);
2435                         rdev_dec_pending(rdev, conf->mddev);
2436                 } else if (r1_bio->bios[m] != NULL) {
2437                         /* This drive got a write error.  We need to
2438                          * narrow down and record precise write
2439                          * errors.
2440                          */
2441                         fail = true;
2442                         if (!narrow_write_error(r1_bio, m)) {
2443                                 md_error(conf->mddev,
2444                                          conf->mirrors[m].rdev);
2445                                 /* an I/O failed, we can't clear the bitmap */
2446                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2447                         }
2448                         rdev_dec_pending(conf->mirrors[m].rdev,
2449                                          conf->mddev);
2450                 }
2451         if (fail) {
2452                 spin_lock_irq(&conf->device_lock);
2453                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2454                 idx = sector_to_idx(r1_bio->sector);
2455                 atomic_inc(&conf->nr_queued[idx]);
2456                 spin_unlock_irq(&conf->device_lock);
2457                 /*
2458                  * In case freeze_array() is waiting for condition
2459                  * get_unqueued_pending() == extra to be true.
2460                  */
2461                 wake_up(&conf->wait_barrier);
2462                 md_wakeup_thread(conf->mddev->thread);
2463         } else {
2464                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2465                         close_write(r1_bio);
2466                 raid_end_bio_io(r1_bio);
2467         }
2468 }
2469
2470 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2471 {
2472         struct mddev *mddev = conf->mddev;
2473         struct bio *bio;
2474         struct md_rdev *rdev;
2475
2476         clear_bit(R1BIO_ReadError, &r1_bio->state);
2477         /* we got a read error. Maybe the drive is bad.  Maybe just
2478          * the block and we can fix it.
2479          * We freeze all other IO, and try reading the block from
2480          * other devices.  When we find one, we re-write
2481          * and check it that fixes the read error.
2482          * This is all done synchronously while the array is
2483          * frozen
2484          */
2485
2486         bio = r1_bio->bios[r1_bio->read_disk];
2487         bio_put(bio);
2488         r1_bio->bios[r1_bio->read_disk] = NULL;
2489
2490         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2491         if (mddev->ro == 0
2492             && !test_bit(FailFast, &rdev->flags)) {
2493                 freeze_array(conf, 1);
2494                 fix_read_error(conf, r1_bio->read_disk,
2495                                r1_bio->sector, r1_bio->sectors);
2496                 unfreeze_array(conf);
2497         } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2498                 md_error(mddev, rdev);
2499         } else {
2500                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2501         }
2502
2503         rdev_dec_pending(rdev, conf->mddev);
2504         allow_barrier(conf, r1_bio->sector);
2505         bio = r1_bio->master_bio;
2506
2507         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2508         r1_bio->state = 0;
2509         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2510 }
2511
2512 static void raid1d(struct md_thread *thread)
2513 {
2514         struct mddev *mddev = thread->mddev;
2515         struct r1bio *r1_bio;
2516         unsigned long flags;
2517         struct r1conf *conf = mddev->private;
2518         struct list_head *head = &conf->retry_list;
2519         struct blk_plug plug;
2520         int idx;
2521
2522         md_check_recovery(mddev);
2523
2524         if (!list_empty_careful(&conf->bio_end_io_list) &&
2525             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2526                 LIST_HEAD(tmp);
2527                 spin_lock_irqsave(&conf->device_lock, flags);
2528                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2529                         list_splice_init(&conf->bio_end_io_list, &tmp);
2530                 spin_unlock_irqrestore(&conf->device_lock, flags);
2531                 while (!list_empty(&tmp)) {
2532                         r1_bio = list_first_entry(&tmp, struct r1bio,
2533                                                   retry_list);
2534                         list_del(&r1_bio->retry_list);
2535                         idx = sector_to_idx(r1_bio->sector);
2536                         atomic_dec(&conf->nr_queued[idx]);
2537                         if (mddev->degraded)
2538                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2539                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2540                                 close_write(r1_bio);
2541                         raid_end_bio_io(r1_bio);
2542                 }
2543         }
2544
2545         blk_start_plug(&plug);
2546         for (;;) {
2547
2548                 flush_pending_writes(conf);
2549
2550                 spin_lock_irqsave(&conf->device_lock, flags);
2551                 if (list_empty(head)) {
2552                         spin_unlock_irqrestore(&conf->device_lock, flags);
2553                         break;
2554                 }
2555                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2556                 list_del(head->prev);
2557                 idx = sector_to_idx(r1_bio->sector);
2558                 atomic_dec(&conf->nr_queued[idx]);
2559                 spin_unlock_irqrestore(&conf->device_lock, flags);
2560
2561                 mddev = r1_bio->mddev;
2562                 conf = mddev->private;
2563                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2564                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2565                             test_bit(R1BIO_WriteError, &r1_bio->state))
2566                                 handle_sync_write_finished(conf, r1_bio);
2567                         else
2568                                 sync_request_write(mddev, r1_bio);
2569                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2570                            test_bit(R1BIO_WriteError, &r1_bio->state))
2571                         handle_write_finished(conf, r1_bio);
2572                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2573                         handle_read_error(conf, r1_bio);
2574                 else
2575                         WARN_ON_ONCE(1);
2576
2577                 cond_resched();
2578                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2579                         md_check_recovery(mddev);
2580         }
2581         blk_finish_plug(&plug);
2582 }
2583
2584 static int init_resync(struct r1conf *conf)
2585 {
2586         int buffs;
2587
2588         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2589         BUG_ON(mempool_initialized(&conf->r1buf_pool));
2590
2591         return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2592                             r1buf_pool_free, conf->poolinfo);
2593 }
2594
2595 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2596 {
2597         struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2598         struct resync_pages *rps;
2599         struct bio *bio;
2600         int i;
2601
2602         for (i = conf->poolinfo->raid_disks; i--; ) {
2603                 bio = r1bio->bios[i];
2604                 rps = bio->bi_private;
2605                 bio_reset(bio);
2606                 bio->bi_private = rps;
2607         }
2608         r1bio->master_bio = NULL;
2609         return r1bio;
2610 }
2611
2612 /*
2613  * perform a "sync" on one "block"
2614  *
2615  * We need to make sure that no normal I/O request - particularly write
2616  * requests - conflict with active sync requests.
2617  *
2618  * This is achieved by tracking pending requests and a 'barrier' concept
2619  * that can be installed to exclude normal IO requests.
2620  */
2621
2622 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2623                                    int *skipped)
2624 {
2625         struct r1conf *conf = mddev->private;
2626         struct r1bio *r1_bio;
2627         struct bio *bio;
2628         sector_t max_sector, nr_sectors;
2629         int disk = -1;
2630         int i;
2631         int wonly = -1;
2632         int write_targets = 0, read_targets = 0;
2633         sector_t sync_blocks;
2634         int still_degraded = 0;
2635         int good_sectors = RESYNC_SECTORS;
2636         int min_bad = 0; /* number of sectors that are bad in all devices */
2637         int idx = sector_to_idx(sector_nr);
2638         int page_idx = 0;
2639
2640         if (!mempool_initialized(&conf->r1buf_pool))
2641                 if (init_resync(conf))
2642                         return 0;
2643
2644         max_sector = mddev->dev_sectors;
2645         if (sector_nr >= max_sector) {
2646                 /* If we aborted, we need to abort the
2647                  * sync on the 'current' bitmap chunk (there will
2648                  * only be one in raid1 resync.
2649                  * We can find the current addess in mddev->curr_resync
2650                  */
2651                 if (mddev->curr_resync < max_sector) /* aborted */
2652                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2653                                            &sync_blocks, 1);
2654                 else /* completed sync */
2655                         conf->fullsync = 0;
2656
2657                 md_bitmap_close_sync(mddev->bitmap);
2658                 close_sync(conf);
2659
2660                 if (mddev_is_clustered(mddev)) {
2661                         conf->cluster_sync_low = 0;
2662                         conf->cluster_sync_high = 0;
2663                 }
2664                 return 0;
2665         }
2666
2667         if (mddev->bitmap == NULL &&
2668             mddev->recovery_cp == MaxSector &&
2669             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2670             conf->fullsync == 0) {
2671                 *skipped = 1;
2672                 return max_sector - sector_nr;
2673         }
2674         /* before building a request, check if we can skip these blocks..
2675          * This call the bitmap_start_sync doesn't actually record anything
2676          */
2677         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2678             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2679                 /* We can skip this block, and probably several more */
2680                 *skipped = 1;
2681                 return sync_blocks;
2682         }
2683
2684         /*
2685          * If there is non-resync activity waiting for a turn, then let it
2686          * though before starting on this new sync request.
2687          */
2688         if (atomic_read(&conf->nr_waiting[idx]))
2689                 schedule_timeout_uninterruptible(1);
2690
2691         /* we are incrementing sector_nr below. To be safe, we check against
2692          * sector_nr + two times RESYNC_SECTORS
2693          */
2694
2695         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2696                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2697
2698
2699         if (raise_barrier(conf, sector_nr))
2700                 return 0;
2701
2702         r1_bio = raid1_alloc_init_r1buf(conf);
2703
2704         rcu_read_lock();
2705         /*
2706          * If we get a correctably read error during resync or recovery,
2707          * we might want to read from a different device.  So we
2708          * flag all drives that could conceivably be read from for READ,
2709          * and any others (which will be non-In_sync devices) for WRITE.
2710          * If a read fails, we try reading from something else for which READ
2711          * is OK.
2712          */
2713
2714         r1_bio->mddev = mddev;
2715         r1_bio->sector = sector_nr;
2716         r1_bio->state = 0;
2717         set_bit(R1BIO_IsSync, &r1_bio->state);
2718         /* make sure good_sectors won't go across barrier unit boundary */
2719         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2720
2721         for (i = 0; i < conf->raid_disks * 2; i++) {
2722                 struct md_rdev *rdev;
2723                 bio = r1_bio->bios[i];
2724
2725                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2726                 if (rdev == NULL ||
2727                     test_bit(Faulty, &rdev->flags)) {
2728                         if (i < conf->raid_disks)
2729                                 still_degraded = 1;
2730                 } else if (!test_bit(In_sync, &rdev->flags)) {
2731                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2732                         bio->bi_end_io = end_sync_write;
2733                         write_targets ++;
2734                 } else {
2735                         /* may need to read from here */
2736                         sector_t first_bad = MaxSector;
2737                         int bad_sectors;
2738
2739                         if (is_badblock(rdev, sector_nr, good_sectors,
2740                                         &first_bad, &bad_sectors)) {
2741                                 if (first_bad > sector_nr)
2742                                         good_sectors = first_bad - sector_nr;
2743                                 else {
2744                                         bad_sectors -= (sector_nr - first_bad);
2745                                         if (min_bad == 0 ||
2746                                             min_bad > bad_sectors)
2747                                                 min_bad = bad_sectors;
2748                                 }
2749                         }
2750                         if (sector_nr < first_bad) {
2751                                 if (test_bit(WriteMostly, &rdev->flags)) {
2752                                         if (wonly < 0)
2753                                                 wonly = i;
2754                                 } else {
2755                                         if (disk < 0)
2756                                                 disk = i;
2757                                 }
2758                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2759                                 bio->bi_end_io = end_sync_read;
2760                                 read_targets++;
2761                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2762                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2763                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2764                                 /*
2765                                  * The device is suitable for reading (InSync),
2766                                  * but has bad block(s) here. Let's try to correct them,
2767                                  * if we are doing resync or repair. Otherwise, leave
2768                                  * this device alone for this sync request.
2769                                  */
2770                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2771                                 bio->bi_end_io = end_sync_write;
2772                                 write_targets++;
2773                         }
2774                 }
2775                 if (rdev && bio->bi_end_io) {
2776                         atomic_inc(&rdev->nr_pending);
2777                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2778                         bio_set_dev(bio, rdev->bdev);
2779                         if (test_bit(FailFast, &rdev->flags))
2780                                 bio->bi_opf |= MD_FAILFAST;
2781                 }
2782         }
2783         rcu_read_unlock();
2784         if (disk < 0)
2785                 disk = wonly;
2786         r1_bio->read_disk = disk;
2787
2788         if (read_targets == 0 && min_bad > 0) {
2789                 /* These sectors are bad on all InSync devices, so we
2790                  * need to mark them bad on all write targets
2791                  */
2792                 int ok = 1;
2793                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2794                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2795                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2796                                 ok = rdev_set_badblocks(rdev, sector_nr,
2797                                                         min_bad, 0
2798                                         ) && ok;
2799                         }
2800                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2801                 *skipped = 1;
2802                 put_buf(r1_bio);
2803
2804                 if (!ok) {
2805                         /* Cannot record the badblocks, so need to
2806                          * abort the resync.
2807                          * If there are multiple read targets, could just
2808                          * fail the really bad ones ???
2809                          */
2810                         conf->recovery_disabled = mddev->recovery_disabled;
2811                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2812                         return 0;
2813                 } else
2814                         return min_bad;
2815
2816         }
2817         if (min_bad > 0 && min_bad < good_sectors) {
2818                 /* only resync enough to reach the next bad->good
2819                  * transition */
2820                 good_sectors = min_bad;
2821         }
2822
2823         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2824                 /* extra read targets are also write targets */
2825                 write_targets += read_targets-1;
2826
2827         if (write_targets == 0 || read_targets == 0) {
2828                 /* There is nowhere to write, so all non-sync
2829                  * drives must be failed - so we are finished
2830                  */
2831                 sector_t rv;
2832                 if (min_bad > 0)
2833                         max_sector = sector_nr + min_bad;
2834                 rv = max_sector - sector_nr;
2835                 *skipped = 1;
2836                 put_buf(r1_bio);
2837                 return rv;
2838         }
2839
2840         if (max_sector > mddev->resync_max)
2841                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2842         if (max_sector > sector_nr + good_sectors)
2843                 max_sector = sector_nr + good_sectors;
2844         nr_sectors = 0;
2845         sync_blocks = 0;
2846         do {
2847                 struct page *page;
2848                 int len = PAGE_SIZE;
2849                 if (sector_nr + (len>>9) > max_sector)
2850                         len = (max_sector - sector_nr) << 9;
2851                 if (len == 0)
2852                         break;
2853                 if (sync_blocks == 0) {
2854                         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2855                                                   &sync_blocks, still_degraded) &&
2856                             !conf->fullsync &&
2857                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2858                                 break;
2859                         if ((len >> 9) > sync_blocks)
2860                                 len = sync_blocks<<9;
2861                 }
2862
2863                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2864                         struct resync_pages *rp;
2865
2866                         bio = r1_bio->bios[i];
2867                         rp = get_resync_pages(bio);
2868                         if (bio->bi_end_io) {
2869                                 page = resync_fetch_page(rp, page_idx);
2870
2871                                 /*
2872                                  * won't fail because the vec table is big
2873                                  * enough to hold all these pages
2874                                  */
2875                                 bio_add_page(bio, page, len, 0);
2876                         }
2877                 }
2878                 nr_sectors += len>>9;
2879                 sector_nr += len>>9;
2880                 sync_blocks -= (len>>9);
2881         } while (++page_idx < RESYNC_PAGES);
2882
2883         r1_bio->sectors = nr_sectors;
2884
2885         if (mddev_is_clustered(mddev) &&
2886                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2887                 conf->cluster_sync_low = mddev->curr_resync_completed;
2888                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2889                 /* Send resync message */
2890                 md_cluster_ops->resync_info_update(mddev,
2891                                 conf->cluster_sync_low,
2892                                 conf->cluster_sync_high);
2893         }
2894
2895         /* For a user-requested sync, we read all readable devices and do a
2896          * compare
2897          */
2898         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2899                 atomic_set(&r1_bio->remaining, read_targets);
2900                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2901                         bio = r1_bio->bios[i];
2902                         if (bio->bi_end_io == end_sync_read) {
2903                                 read_targets--;
2904                                 md_sync_acct_bio(bio, nr_sectors);
2905                                 if (read_targets == 1)
2906                                         bio->bi_opf &= ~MD_FAILFAST;
2907                                 submit_bio_noacct(bio);
2908                         }
2909                 }
2910         } else {
2911                 atomic_set(&r1_bio->remaining, 1);
2912                 bio = r1_bio->bios[r1_bio->read_disk];
2913                 md_sync_acct_bio(bio, nr_sectors);
2914                 if (read_targets == 1)
2915                         bio->bi_opf &= ~MD_FAILFAST;
2916                 submit_bio_noacct(bio);
2917         }
2918         return nr_sectors;
2919 }
2920
2921 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2922 {
2923         if (sectors)
2924                 return sectors;
2925
2926         return mddev->dev_sectors;
2927 }
2928
2929 static struct r1conf *setup_conf(struct mddev *mddev)
2930 {
2931         struct r1conf *conf;
2932         int i;
2933         struct raid1_info *disk;
2934         struct md_rdev *rdev;
2935         int err = -ENOMEM;
2936
2937         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2938         if (!conf)
2939                 goto abort;
2940
2941         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2942                                    sizeof(atomic_t), GFP_KERNEL);
2943         if (!conf->nr_pending)
2944                 goto abort;
2945
2946         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2947                                    sizeof(atomic_t), GFP_KERNEL);
2948         if (!conf->nr_waiting)
2949                 goto abort;
2950
2951         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2952                                   sizeof(atomic_t), GFP_KERNEL);
2953         if (!conf->nr_queued)
2954                 goto abort;
2955
2956         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2957                                 sizeof(atomic_t), GFP_KERNEL);
2958         if (!conf->barrier)
2959                 goto abort;
2960
2961         conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2962                                             mddev->raid_disks, 2),
2963                                 GFP_KERNEL);
2964         if (!conf->mirrors)
2965                 goto abort;
2966
2967         conf->tmppage = alloc_page(GFP_KERNEL);
2968         if (!conf->tmppage)
2969                 goto abort;
2970
2971         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2972         if (!conf->poolinfo)
2973                 goto abort;
2974         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2975         err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2976                            rbio_pool_free, conf->poolinfo);
2977         if (err)
2978                 goto abort;
2979
2980         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2981         if (err)
2982                 goto abort;
2983
2984         conf->poolinfo->mddev = mddev;
2985
2986         err = -EINVAL;
2987         spin_lock_init(&conf->device_lock);
2988         rdev_for_each(rdev, mddev) {
2989                 int disk_idx = rdev->raid_disk;
2990                 if (disk_idx >= mddev->raid_disks
2991                     || disk_idx < 0)
2992                         continue;
2993                 if (test_bit(Replacement, &rdev->flags))
2994                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
2995                 else
2996                         disk = conf->mirrors + disk_idx;
2997
2998                 if (disk->rdev)
2999                         goto abort;
3000                 disk->rdev = rdev;
3001                 disk->head_position = 0;
3002                 disk->seq_start = MaxSector;
3003         }
3004         conf->raid_disks = mddev->raid_disks;
3005         conf->mddev = mddev;
3006         INIT_LIST_HEAD(&conf->retry_list);
3007         INIT_LIST_HEAD(&conf->bio_end_io_list);
3008
3009         spin_lock_init(&conf->resync_lock);
3010         init_waitqueue_head(&conf->wait_barrier);
3011
3012         bio_list_init(&conf->pending_bio_list);
3013         conf->pending_count = 0;
3014         conf->recovery_disabled = mddev->recovery_disabled - 1;
3015
3016         err = -EIO;
3017         for (i = 0; i < conf->raid_disks * 2; i++) {
3018
3019                 disk = conf->mirrors + i;
3020
3021                 if (i < conf->raid_disks &&
3022                     disk[conf->raid_disks].rdev) {
3023                         /* This slot has a replacement. */
3024                         if (!disk->rdev) {
3025                                 /* No original, just make the replacement
3026                                  * a recovering spare
3027                                  */
3028                                 disk->rdev =
3029                                         disk[conf->raid_disks].rdev;
3030                                 disk[conf->raid_disks].rdev = NULL;
3031                         } else if (!test_bit(In_sync, &disk->rdev->flags))
3032                                 /* Original is not in_sync - bad */
3033                                 goto abort;
3034                 }
3035
3036                 if (!disk->rdev ||
3037                     !test_bit(In_sync, &disk->rdev->flags)) {
3038                         disk->head_position = 0;
3039                         if (disk->rdev &&
3040                             (disk->rdev->saved_raid_disk < 0))
3041                                 conf->fullsync = 1;
3042                 }
3043         }
3044
3045         err = -ENOMEM;
3046         conf->thread = md_register_thread(raid1d, mddev, "raid1");
3047         if (!conf->thread)
3048                 goto abort;
3049
3050         return conf;
3051
3052  abort:
3053         if (conf) {
3054                 mempool_exit(&conf->r1bio_pool);
3055                 kfree(conf->mirrors);
3056                 safe_put_page(conf->tmppage);
3057                 kfree(conf->poolinfo);
3058                 kfree(conf->nr_pending);
3059                 kfree(conf->nr_waiting);
3060                 kfree(conf->nr_queued);
3061                 kfree(conf->barrier);
3062                 bioset_exit(&conf->bio_split);
3063                 kfree(conf);
3064         }
3065         return ERR_PTR(err);
3066 }
3067
3068 static void raid1_free(struct mddev *mddev, void *priv);
3069 static int raid1_run(struct mddev *mddev)
3070 {
3071         struct r1conf *conf;
3072         int i;
3073         struct md_rdev *rdev;
3074         int ret;
3075         bool discard_supported = false;
3076
3077         if (mddev->level != 1) {
3078                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3079                         mdname(mddev), mddev->level);
3080                 return -EIO;
3081         }
3082         if (mddev->reshape_position != MaxSector) {
3083                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3084                         mdname(mddev));
3085                 return -EIO;
3086         }
3087         if (mddev_init_writes_pending(mddev) < 0)
3088                 return -ENOMEM;
3089         /*
3090          * copy the already verified devices into our private RAID1
3091          * bookkeeping area. [whatever we allocate in run(),
3092          * should be freed in raid1_free()]
3093          */
3094         if (mddev->private == NULL)
3095                 conf = setup_conf(mddev);
3096         else
3097                 conf = mddev->private;
3098
3099         if (IS_ERR(conf))
3100                 return PTR_ERR(conf);
3101
3102         if (mddev->queue) {
3103                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3104                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3105         }
3106
3107         rdev_for_each(rdev, mddev) {
3108                 if (!mddev->gendisk)
3109                         continue;
3110                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3111                                   rdev->data_offset << 9);
3112                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3113                         discard_supported = true;
3114         }
3115
3116         mddev->degraded = 0;
3117         for (i = 0; i < conf->raid_disks; i++)
3118                 if (conf->mirrors[i].rdev == NULL ||
3119                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3120                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3121                         mddev->degraded++;
3122         /*
3123          * RAID1 needs at least one disk in active
3124          */
3125         if (conf->raid_disks - mddev->degraded < 1) {
3126                 ret = -EINVAL;
3127                 goto abort;
3128         }
3129
3130         if (conf->raid_disks - mddev->degraded == 1)
3131                 mddev->recovery_cp = MaxSector;
3132
3133         if (mddev->recovery_cp != MaxSector)
3134                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3135                         mdname(mddev));
3136         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3137                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3138                 mddev->raid_disks);
3139
3140         /*
3141          * Ok, everything is just fine now
3142          */
3143         mddev->thread = conf->thread;
3144         conf->thread = NULL;
3145         mddev->private = conf;
3146         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3147
3148         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3149
3150         if (mddev->queue) {
3151                 if (discard_supported)
3152                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3153                                                 mddev->queue);
3154                 else
3155                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3156                                                   mddev->queue);
3157         }
3158
3159         ret = md_integrity_register(mddev);
3160         if (ret) {
3161                 md_unregister_thread(&mddev->thread);
3162                 goto abort;
3163         }
3164         return 0;
3165
3166 abort:
3167         raid1_free(mddev, conf);
3168         return ret;
3169 }
3170
3171 static void raid1_free(struct mddev *mddev, void *priv)
3172 {
3173         struct r1conf *conf = priv;
3174
3175         mempool_exit(&conf->r1bio_pool);
3176         kfree(conf->mirrors);
3177         safe_put_page(conf->tmppage);
3178         kfree(conf->poolinfo);
3179         kfree(conf->nr_pending);
3180         kfree(conf->nr_waiting);
3181         kfree(conf->nr_queued);
3182         kfree(conf->barrier);
3183         bioset_exit(&conf->bio_split);
3184         kfree(conf);
3185 }
3186
3187 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3188 {
3189         /* no resync is happening, and there is enough space
3190          * on all devices, so we can resize.
3191          * We need to make sure resync covers any new space.
3192          * If the array is shrinking we should possibly wait until
3193          * any io in the removed space completes, but it hardly seems
3194          * worth it.
3195          */
3196         sector_t newsize = raid1_size(mddev, sectors, 0);
3197         if (mddev->external_size &&
3198             mddev->array_sectors > newsize)
3199                 return -EINVAL;
3200         if (mddev->bitmap) {
3201                 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3202                 if (ret)
3203                         return ret;
3204         }
3205         md_set_array_sectors(mddev, newsize);
3206         if (sectors > mddev->dev_sectors &&
3207             mddev->recovery_cp > mddev->dev_sectors) {
3208                 mddev->recovery_cp = mddev->dev_sectors;
3209                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3210         }
3211         mddev->dev_sectors = sectors;
3212         mddev->resync_max_sectors = sectors;
3213         return 0;
3214 }
3215
3216 static int raid1_reshape(struct mddev *mddev)
3217 {
3218         /* We need to:
3219          * 1/ resize the r1bio_pool
3220          * 2/ resize conf->mirrors
3221          *
3222          * We allocate a new r1bio_pool if we can.
3223          * Then raise a device barrier and wait until all IO stops.
3224          * Then resize conf->mirrors and swap in the new r1bio pool.
3225          *
3226          * At the same time, we "pack" the devices so that all the missing
3227          * devices have the higher raid_disk numbers.
3228          */
3229         mempool_t newpool, oldpool;
3230         struct pool_info *newpoolinfo;
3231         struct raid1_info *newmirrors;
3232         struct r1conf *conf = mddev->private;
3233         int cnt, raid_disks;
3234         unsigned long flags;
3235         int d, d2;
3236         int ret;
3237
3238         memset(&newpool, 0, sizeof(newpool));
3239         memset(&oldpool, 0, sizeof(oldpool));
3240
3241         /* Cannot change chunk_size, layout, or level */
3242         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3243             mddev->layout != mddev->new_layout ||
3244             mddev->level != mddev->new_level) {
3245                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3246                 mddev->new_layout = mddev->layout;
3247                 mddev->new_level = mddev->level;
3248                 return -EINVAL;
3249         }
3250
3251         if (!mddev_is_clustered(mddev))
3252                 md_allow_write(mddev);
3253
3254         raid_disks = mddev->raid_disks + mddev->delta_disks;
3255
3256         if (raid_disks < conf->raid_disks) {
3257                 cnt=0;
3258                 for (d= 0; d < conf->raid_disks; d++)
3259                         if (conf->mirrors[d].rdev)
3260                                 cnt++;
3261                 if (cnt > raid_disks)
3262                         return -EBUSY;
3263         }
3264
3265         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3266         if (!newpoolinfo)
3267                 return -ENOMEM;
3268         newpoolinfo->mddev = mddev;
3269         newpoolinfo->raid_disks = raid_disks * 2;
3270
3271         ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3272                            rbio_pool_free, newpoolinfo);
3273         if (ret) {
3274                 kfree(newpoolinfo);
3275                 return ret;
3276         }
3277         newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3278                                          raid_disks, 2),
3279                              GFP_KERNEL);
3280         if (!newmirrors) {
3281                 kfree(newpoolinfo);
3282                 mempool_exit(&newpool);
3283                 return -ENOMEM;
3284         }
3285
3286         freeze_array(conf, 0);
3287
3288         /* ok, everything is stopped */
3289         oldpool = conf->r1bio_pool;
3290         conf->r1bio_pool = newpool;
3291
3292         for (d = d2 = 0; d < conf->raid_disks; d++) {
3293                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3294                 if (rdev && rdev->raid_disk != d2) {
3295                         sysfs_unlink_rdev(mddev, rdev);
3296                         rdev->raid_disk = d2;
3297                         sysfs_unlink_rdev(mddev, rdev);
3298                         if (sysfs_link_rdev(mddev, rdev))
3299                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3300                                         mdname(mddev), rdev->raid_disk);
3301                 }
3302                 if (rdev)
3303                         newmirrors[d2++].rdev = rdev;
3304         }
3305         kfree(conf->mirrors);
3306         conf->mirrors = newmirrors;
3307         kfree(conf->poolinfo);
3308         conf->poolinfo = newpoolinfo;
3309
3310         spin_lock_irqsave(&conf->device_lock, flags);
3311         mddev->degraded += (raid_disks - conf->raid_disks);
3312         spin_unlock_irqrestore(&conf->device_lock, flags);
3313         conf->raid_disks = mddev->raid_disks = raid_disks;
3314         mddev->delta_disks = 0;
3315
3316         unfreeze_array(conf);
3317
3318         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3319         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3320         md_wakeup_thread(mddev->thread);
3321
3322         mempool_exit(&oldpool);
3323         return 0;
3324 }
3325
3326 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3327 {
3328         struct r1conf *conf = mddev->private;
3329
3330         if (quiesce)
3331                 freeze_array(conf, 0);
3332         else
3333                 unfreeze_array(conf);
3334 }
3335
3336 static void *raid1_takeover(struct mddev *mddev)
3337 {
3338         /* raid1 can take over:
3339          *  raid5 with 2 devices, any layout or chunk size
3340          */
3341         if (mddev->level == 5 && mddev->raid_disks == 2) {
3342                 struct r1conf *conf;
3343                 mddev->new_level = 1;
3344                 mddev->new_layout = 0;
3345                 mddev->new_chunk_sectors = 0;
3346                 conf = setup_conf(mddev);
3347                 if (!IS_ERR(conf)) {
3348                         /* Array must appear to be quiesced */
3349                         conf->array_frozen = 1;
3350                         mddev_clear_unsupported_flags(mddev,
3351                                 UNSUPPORTED_MDDEV_FLAGS);
3352                 }
3353                 return conf;
3354         }
3355         return ERR_PTR(-EINVAL);
3356 }
3357
3358 static struct md_personality raid1_personality =
3359 {
3360         .name           = "raid1",
3361         .level          = 1,
3362         .owner          = THIS_MODULE,
3363         .make_request   = raid1_make_request,
3364         .run            = raid1_run,
3365         .free           = raid1_free,
3366         .status         = raid1_status,
3367         .error_handler  = raid1_error,
3368         .hot_add_disk   = raid1_add_disk,
3369         .hot_remove_disk= raid1_remove_disk,
3370         .spare_active   = raid1_spare_active,
3371         .sync_request   = raid1_sync_request,
3372         .resize         = raid1_resize,
3373         .size           = raid1_size,
3374         .check_reshape  = raid1_reshape,
3375         .quiesce        = raid1_quiesce,
3376         .takeover       = raid1_takeover,
3377 };
3378
3379 static int __init raid_init(void)
3380 {
3381         return register_md_personality(&raid1_personality);
3382 }
3383
3384 static void raid_exit(void)
3385 {
3386         unregister_md_personality(&raid1_personality);
3387 }
3388
3389 module_init(raid_init);
3390 module_exit(raid_exit);
3391 MODULE_LICENSE("GPL");
3392 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3393 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3394 MODULE_ALIAS("md-raid1");
3395 MODULE_ALIAS("md-level-1");
3396
3397 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);