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 #define RAID_1_10_NAME "raid1"
  53 #include "raid1-10.c"
  54
  55 #define START(node) ((node)->start)
  56 #define LAST(node) ((node)->last)
  57 INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
  58                      START, LAST, static inline, raid1_rb);
  59
  60 static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
  61                                 struct serial_info *si, int idx)
  62 {
  63         unsigned long flags;
  64         int ret = 0;
  65         sector_t lo = r1_bio->sector;
  66         sector_t hi = lo + r1_bio->sectors;
  67         struct serial_in_rdev *serial = &rdev->serial[idx];
  68
  69         spin_lock_irqsave(&serial->serial_lock, flags);
  70         /* collision happened */
  71         if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
  72                 ret = -EBUSY;
  73         else {
  74                 si->start = lo;
  75                 si->last = hi;
  76                 raid1_rb_insert(si, &serial->serial_rb);
  77         }
  78         spin_unlock_irqrestore(&serial->serial_lock, flags);
  79
  80         return ret;
  81 }
  82
  83 static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
  84 {
  85         struct mddev *mddev = rdev->mddev;
  86         struct serial_info *si;
  87         int idx = sector_to_idx(r1_bio->sector);
  88         struct serial_in_rdev *serial = &rdev->serial[idx];
  89
  90         if (WARN_ON(!mddev->serial_info_pool))
  91                 return;
  92         si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
  93         wait_event(serial->serial_io_wait,
  94                    check_and_add_serial(rdev, r1_bio, si, idx) == 0);
  95 }
  96
  97 static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
  98 {
  99         struct serial_info *si;
 100         unsigned long flags;
 101         int found = 0;
 102         struct mddev *mddev = rdev->mddev;
 103         int idx = sector_to_idx(lo);
 104         struct serial_in_rdev *serial = &rdev->serial[idx];
 105
 106         spin_lock_irqsave(&serial->serial_lock, flags);
 107         for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
 108              si; si = raid1_rb_iter_next(si, lo, hi)) {
 109                 if (si->start == lo && si->last == hi) {
 110                         raid1_rb_remove(si, &serial->serial_rb);
 111                         mempool_free(si, mddev->serial_info_pool);
 112                         found = 1;
 113                         break;
 114                 }
 115         }
 116         if (!found)
 117                 WARN(1, "The write IO is not recorded for serialization\n");
 118         spin_unlock_irqrestore(&serial->serial_lock, flags);
 119         wake_up(&serial->serial_io_wait);
 120 }
 121
 122 /*
 123  * for resync bio, r1bio pointer can be retrieved from the per-bio
 124  * 'struct resync_pages'.
 125  */
 126 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
 127 {
 128         return get_resync_pages(bio)->raid_bio;
 129 }
 130
 131 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
 132 {
 133         struct pool_info *pi = data;
 134         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
 135
 136         /* allocate a r1bio with room for raid_disks entries in the bios array */
 137         return kzalloc(size, gfp_flags);
 138 }
 139
 140 #define RESYNC_DEPTH 32
 141 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
 142 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 143 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 144 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 145 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 146
 147 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 148 {
 149         struct pool_info *pi = data;
 150         struct r1bio *r1_bio;
 151         struct bio *bio;
 152         int need_pages;
 153         int j;
 154         struct resync_pages *rps;
 155
 156         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 157         if (!r1_bio)
 158                 return NULL;
 159
 160         rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
 161                             gfp_flags);
 162         if (!rps)
 163                 goto out_free_r1bio;
 164
 165         /*
 166          * Allocate bios : 1 for reading, n-1 for writing
 167          */
 168         for (j = pi->raid_disks ; j-- ; ) {
 169                 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
 170                 if (!bio)
 171                         goto out_free_bio;
 172                 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
 173                 r1_bio->bios[j] = bio;
 174         }
 175         /*
 176          * Allocate RESYNC_PAGES data pages and attach them to
 177          * the first bio.
 178          * If this is a user-requested check/repair, allocate
 179          * RESYNC_PAGES for each bio.
 180          */
 181         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 182                 need_pages = pi->raid_disks;
 183         else
 184                 need_pages = 1;
 185         for (j = 0; j < pi->raid_disks; j++) {
 186                 struct resync_pages *rp = &rps[j];
 187
 188                 bio = r1_bio->bios[j];
 189
 190                 if (j < need_pages) {
 191                         if (resync_alloc_pages(rp, gfp_flags))
 192                                 goto out_free_pages;
 193                 } else {
 194                         memcpy(rp, &rps[0], sizeof(*rp));
 195                         resync_get_all_pages(rp);
 196                 }
 197
 198                 rp->raid_bio = r1_bio;
 199                 bio->bi_private = rp;
 200         }
 201
 202         r1_bio->master_bio = NULL;
 203
 204         return r1_bio;
 205
 206 out_free_pages:
 207         while (--j >= 0)
 208                 resync_free_pages(&rps[j]);
 209
 210 out_free_bio:
 211         while (++j < pi->raid_disks) {
 212                 bio_uninit(r1_bio->bios[j]);
 213                 kfree(r1_bio->bios[j]);
 214         }
 215         kfree(rps);
 216
 217 out_free_r1bio:
 218         rbio_pool_free(r1_bio, data);
 219         return NULL;
 220 }
 221
 222 static void r1buf_pool_free(void *__r1_bio, void *data)
 223 {
 224         struct pool_info *pi = data;
 225         int i;
 226         struct r1bio *r1bio = __r1_bio;
 227         struct resync_pages *rp = NULL;
 228
 229         for (i = pi->raid_disks; i--; ) {
 230                 rp = get_resync_pages(r1bio->bios[i]);
 231                 resync_free_pages(rp);
 232                 bio_uninit(r1bio->bios[i]);
 233                 kfree(r1bio->bios[i]);
 234         }
 235
 236         /* resync pages array stored in the 1st bio's .bi_private */
 237         kfree(rp);
 238
 239         rbio_pool_free(r1bio, data);
 240 }
 241
 242 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 243 {
 244         int i;
 245
 246         for (i = 0; i < conf->raid_disks * 2; i++) {
 247                 struct bio **bio = r1_bio->bios + i;
 248                 if (!BIO_SPECIAL(*bio))
 249                         bio_put(*bio);
 250                 *bio = NULL;
 251         }
 252 }
 253
 254 static void free_r1bio(struct r1bio *r1_bio)
 255 {
 256         struct r1conf *conf = r1_bio->mddev->private;
 257
 258         put_all_bios(conf, r1_bio);
 259         mempool_free(r1_bio, &conf->r1bio_pool);
 260 }
 261
 262 static void put_buf(struct r1bio *r1_bio)
 263 {
 264         struct r1conf *conf = r1_bio->mddev->private;
 265         sector_t sect = r1_bio->sector;
 266         int i;
 267
 268         for (i = 0; i < conf->raid_disks * 2; i++) {
 269                 struct bio *bio = r1_bio->bios[i];
 270                 if (bio->bi_end_io)
 271                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 272         }
 273
 274         mempool_free(r1_bio, &conf->r1buf_pool);
 275
 276         lower_barrier(conf, sect);
 277 }
 278
 279 static void reschedule_retry(struct r1bio *r1_bio)
 280 {
 281         unsigned long flags;
 282         struct mddev *mddev = r1_bio->mddev;
 283         struct r1conf *conf = mddev->private;
 284         int idx;
 285
 286         idx = sector_to_idx(r1_bio->sector);
 287         spin_lock_irqsave(&conf->device_lock, flags);
 288         list_add(&r1_bio->retry_list, &conf->retry_list);
 289         atomic_inc(&conf->nr_queued[idx]);
 290         spin_unlock_irqrestore(&conf->device_lock, flags);
 291
 292         wake_up(&conf->wait_barrier);
 293         md_wakeup_thread(mddev->thread);
 294 }
 295
 296 /*
 297  * raid_end_bio_io() is called when we have finished servicing a mirrored
 298  * operation and are ready to return a success/failure code to the buffer
 299  * cache layer.
 300  */
 301 static void call_bio_endio(struct r1bio *r1_bio)
 302 {
 303         struct bio *bio = r1_bio->master_bio;
 304
 305         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 306                 bio->bi_status = BLK_STS_IOERR;
 307
 308         bio_endio(bio);
 309 }
 310
 311 static void raid_end_bio_io(struct r1bio *r1_bio)
 312 {
 313         struct bio *bio = r1_bio->master_bio;
 314         struct r1conf *conf = r1_bio->mddev->private;
 315         sector_t sector = r1_bio->sector;
 316
 317         /* if nobody has done the final endio yet, do it now */
 318         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 319                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 320                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 321                          (unsigned long long) bio->bi_iter.bi_sector,
 322                          (unsigned long long) bio_end_sector(bio) - 1);
 323
 324                 call_bio_endio(r1_bio);
 325         }
 326
 327         free_r1bio(r1_bio);
 328         /*
 329          * Wake up any possible resync thread that waits for the device
 330          * to go idle.  All I/Os, even write-behind writes, are done.
 331          */
 332         allow_barrier(conf, sector);
 333 }
 334
 335 /*
 336  * Update disk head position estimator based on IRQ completion info.
 337  */
 338 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 339 {
 340         struct r1conf *conf = r1_bio->mddev->private;
 341
 342         conf->mirrors[disk].head_position =
 343                 r1_bio->sector + (r1_bio->sectors);
 344 }
 345
 346 /*
 347  * Find the disk number which triggered given bio
 348  */
 349 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 350 {
 351         int mirror;
 352         struct r1conf *conf = r1_bio->mddev->private;
 353         int raid_disks = conf->raid_disks;
 354
 355         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 356                 if (r1_bio->bios[mirror] == bio)
 357                         break;
 358
 359         BUG_ON(mirror == raid_disks * 2);
 360         update_head_pos(mirror, r1_bio);
 361
 362         return mirror;
 363 }
 364
 365 static void raid1_end_read_request(struct bio *bio)
 366 {
 367         int uptodate = !bio->bi_status;
 368         struct r1bio *r1_bio = bio->bi_private;
 369         struct r1conf *conf = r1_bio->mddev->private;
 370         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 371
 372         /*
 373          * this branch is our 'one mirror IO has finished' event handler:
 374          */
 375         update_head_pos(r1_bio->read_disk, r1_bio);
 376
 377         if (uptodate)
 378                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 379         else if (test_bit(FailFast, &rdev->flags) &&
 380                  test_bit(R1BIO_FailFast, &r1_bio->state))
 381                 /* This was a fail-fast read so we definitely
 382                  * want to retry */
 383                 ;
 384         else {
 385                 /* If all other devices have failed, we want to return
 386                  * the error upwards rather than fail the last device.
 387                  * Here we redefine "uptodate" to mean "Don't want to retry"
 388                  */
 389                 unsigned long flags;
 390                 spin_lock_irqsave(&conf->device_lock, flags);
 391                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 392                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 393                      test_bit(In_sync, &rdev->flags)))
 394                         uptodate = 1;
 395                 spin_unlock_irqrestore(&conf->device_lock, flags);
 396         }
 397
 398         if (uptodate) {
 399                 raid_end_bio_io(r1_bio);
 400                 rdev_dec_pending(rdev, conf->mddev);
 401         } else {
 402                 /*
 403                  * oops, read error:
 404                  */
 405                 pr_err_ratelimited("md/raid1:%s: %pg: rescheduling sector %llu\n",
 406                                    mdname(conf->mddev),
 407                                    rdev->bdev,
 408                                    (unsigned long long)r1_bio->sector);
 409                 set_bit(R1BIO_ReadError, &r1_bio->state);
 410                 reschedule_retry(r1_bio);
 411                 /* don't drop the reference on read_disk yet */
 412         }
 413 }
 414
 415 static void close_write(struct r1bio *r1_bio)
 416 {
 417         /* it really is the end of this request */
 418         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 419                 bio_free_pages(r1_bio->behind_master_bio);
 420                 bio_put(r1_bio->behind_master_bio);
 421                 r1_bio->behind_master_bio = NULL;
 422         }
 423         /* clear the bitmap if all writes complete successfully */
 424         md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 425                            r1_bio->sectors,
 426                            !test_bit(R1BIO_Degraded, &r1_bio->state),
 427                            test_bit(R1BIO_BehindIO, &r1_bio->state));
 428         md_write_end(r1_bio->mddev);
 429 }
 430
 431 static void r1_bio_write_done(struct r1bio *r1_bio)
 432 {
 433         if (!atomic_dec_and_test(&r1_bio->remaining))
 434                 return;
 435
 436         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 437                 reschedule_retry(r1_bio);
 438         else {
 439                 close_write(r1_bio);
 440                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 441                         reschedule_retry(r1_bio);
 442                 else
 443                         raid_end_bio_io(r1_bio);
 444         }
 445 }
 446
 447 static void raid1_end_write_request(struct bio *bio)
 448 {
 449         struct r1bio *r1_bio = bio->bi_private;
 450         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 451         struct r1conf *conf = r1_bio->mddev->private;
 452         struct bio *to_put = NULL;
 453         int mirror = find_bio_disk(r1_bio, bio);
 454         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 455         bool discard_error;
 456         sector_t lo = r1_bio->sector;
 457         sector_t hi = r1_bio->sector + r1_bio->sectors;
 458
 459         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
 460
 461         /*
 462          * 'one mirror IO has finished' event handler:
 463          */
 464         if (bio->bi_status && !discard_error) {
 465                 set_bit(WriteErrorSeen, &rdev->flags);
 466                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
 467                         set_bit(MD_RECOVERY_NEEDED, &
 468                                 conf->mddev->recovery);
 469
 470                 if (test_bit(FailFast, &rdev->flags) &&
 471                     (bio->bi_opf & MD_FAILFAST) &&
 472                     /* We never try FailFast to WriteMostly devices */
 473                     !test_bit(WriteMostly, &rdev->flags)) {
 474                         md_error(r1_bio->mddev, rdev);
 475                 }
 476
 477                 /*
 478                  * When the device is faulty, it is not necessary to
 479                  * handle write error.
 480                  */
 481                 if (!test_bit(Faulty, &rdev->flags))
 482                         set_bit(R1BIO_WriteError, &r1_bio->state);
 483                 else {
 484                         /* Fail the request */
 485                         set_bit(R1BIO_Degraded, &r1_bio->state);
 486                         /* Finished with this branch */
 487                         r1_bio->bios[mirror] = NULL;
 488                         to_put = bio;
 489                 }
 490         } else {
 491                 /*
 492                  * Set R1BIO_Uptodate in our master bio, so that we
 493                  * will return a good error code for to the higher
 494                  * levels even if IO on some other mirrored buffer
 495                  * fails.
 496                  *
 497                  * The 'master' represents the composite IO operation
 498                  * to user-side. So if something waits for IO, then it
 499                  * will wait for the 'master' bio.
 500                  */
 501                 sector_t first_bad;
 502                 int bad_sectors;
 503
 504                 r1_bio->bios[mirror] = NULL;
 505                 to_put = bio;
 506                 /*
 507                  * Do not set R1BIO_Uptodate if the current device is
 508                  * rebuilding or Faulty. This is because we cannot use
 509                  * such device for properly reading the data back (we could
 510                  * potentially use it, if the current write would have felt
 511                  * before rdev->recovery_offset, but for simplicity we don't
 512                  * check this here.
 513                  */
 514                 if (test_bit(In_sync, &rdev->flags) &&
 515                     !test_bit(Faulty, &rdev->flags))
 516                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 517
 518                 /* Maybe we can clear some bad blocks. */
 519                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 520                                 &first_bad, &bad_sectors) && !discard_error) {
 521                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 522                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 523                 }
 524         }
 525
 526         if (behind) {
 527                 if (test_bit(CollisionCheck, &rdev->flags))
 528                         remove_serial(rdev, lo, hi);
 529                 if (test_bit(WriteMostly, &rdev->flags))
 530                         atomic_dec(&r1_bio->behind_remaining);
 531
 532                 /*
 533                  * In behind mode, we ACK the master bio once the I/O
 534                  * has safely reached all non-writemostly
 535                  * disks. Setting the Returned bit ensures that this
 536                  * gets done only once -- we don't ever want to return
 537                  * -EIO here, instead we'll wait
 538                  */
 539                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 540                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 541                         /* Maybe we can return now */
 542                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 543                                 struct bio *mbio = r1_bio->master_bio;
 544                                 pr_debug("raid1: behind end write sectors"
 545                                          " %llu-%llu\n",
 546                                          (unsigned long long) mbio->bi_iter.bi_sector,
 547                                          (unsigned long long) bio_end_sector(mbio) - 1);
 548                                 call_bio_endio(r1_bio);
 549                         }
 550                 }
 551         } else if (rdev->mddev->serialize_policy)
 552                 remove_serial(rdev, lo, hi);
 553         if (r1_bio->bios[mirror] == NULL)
 554                 rdev_dec_pending(rdev, conf->mddev);
 555
 556         /*
 557          * Let's see if all mirrored write operations have finished
 558          * already.
 559          */
 560         r1_bio_write_done(r1_bio);
 561
 562         if (to_put)
 563                 bio_put(to_put);
 564 }
 565
 566 static sector_t align_to_barrier_unit_end(sector_t start_sector,
 567                                           sector_t sectors)
 568 {
 569         sector_t len;
 570
 571         WARN_ON(sectors == 0);
 572         /*
 573          * len is the number of sectors from start_sector to end of the
 574          * barrier unit which start_sector belongs to.
 575          */
 576         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
 577               start_sector;
 578
 579         if (len > sectors)
 580                 len = sectors;
 581
 582         return len;
 583 }
 584
 585 /*
 586  * This routine returns the disk from which the requested read should
 587  * be done. There is a per-array 'next expected sequential IO' sector
 588  * number - if this matches on the next IO then we use the last disk.
 589  * There is also a per-disk 'last know head position' sector that is
 590  * maintained from IRQ contexts, both the normal and the resync IO
 591  * completion handlers update this position correctly. If there is no
 592  * perfect sequential match then we pick the disk whose head is closest.
 593  *
 594  * If there are 2 mirrors in the same 2 devices, performance degrades
 595  * because position is mirror, not device based.
 596  *
 597  * The rdev for the device selected will have nr_pending incremented.
 598  */
 599 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 600 {
 601         const sector_t this_sector = r1_bio->sector;
 602         int sectors;
 603         int best_good_sectors;
 604         int best_disk, best_dist_disk, best_pending_disk;
 605         int has_nonrot_disk;
 606         int disk;
 607         sector_t best_dist;
 608         unsigned int min_pending;
 609         struct md_rdev *rdev;
 610         int choose_first;
 611         int choose_next_idle;
 612
 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 = 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 = bdev_nonrot(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 = 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         *max_sectors = sectors;
 788
 789         return best_disk;
 790 }
 791
 792 static void wake_up_barrier(struct r1conf *conf)
 793 {
 794         if (wq_has_sleeper(&conf->wait_barrier))
 795                 wake_up(&conf->wait_barrier);
 796 }
 797
 798 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
 799 {
 800         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
 801         raid1_prepare_flush_writes(conf->mddev->bitmap);
 802         wake_up_barrier(conf);
 803
 804         while (bio) { /* submit pending writes */
 805                 struct bio *next = bio->bi_next;
 806
 807                 raid1_submit_write(bio);
 808                 bio = next;
 809                 cond_resched();
 810         }
 811 }
 812
 813 static void flush_pending_writes(struct r1conf *conf)
 814 {
 815         /* Any writes that have been queued but are awaiting
 816          * bitmap updates get flushed here.
 817          */
 818         spin_lock_irq(&conf->device_lock);
 819
 820         if (conf->pending_bio_list.head) {
 821                 struct blk_plug plug;
 822                 struct bio *bio;
 823
 824                 bio = bio_list_get(&conf->pending_bio_list);
 825                 spin_unlock_irq(&conf->device_lock);
 826
 827                 /*
 828                  * As this is called in a wait_event() loop (see freeze_array),
 829                  * current->state might be TASK_UNINTERRUPTIBLE which will
 830                  * cause a warning when we prepare to wait again.  As it is
 831                  * rare that this path is taken, it is perfectly safe to force
 832                  * us to go around the wait_event() loop again, so the warning
 833                  * is a false-positive.  Silence the warning by resetting
 834                  * thread state
 835                  */
 836                 __set_current_state(TASK_RUNNING);
 837                 blk_start_plug(&plug);
 838                 flush_bio_list(conf, bio);
 839                 blk_finish_plug(&plug);
 840         } else
 841                 spin_unlock_irq(&conf->device_lock);
 842 }
 843
 844 /* Barriers....
 845  * Sometimes we need to suspend IO while we do something else,
 846  * either some resync/recovery, or reconfigure the array.
 847  * To do this we raise a 'barrier'.
 848  * The 'barrier' is a counter that can be raised multiple times
 849  * to count how many activities are happening which preclude
 850  * normal IO.
 851  * We can only raise the barrier if there is no pending IO.
 852  * i.e. if nr_pending == 0.
 853  * We choose only to raise the barrier if no-one is waiting for the
 854  * barrier to go down.  This means that as soon as an IO request
 855  * is ready, no other operations which require a barrier will start
 856  * until the IO request has had a chance.
 857  *
 858  * So: regular IO calls 'wait_barrier'.  When that returns there
 859  *    is no backgroup IO happening,  It must arrange to call
 860  *    allow_barrier when it has finished its IO.
 861  * backgroup IO calls must call raise_barrier.  Once that returns
 862  *    there is no normal IO happeing.  It must arrange to call
 863  *    lower_barrier when the particular background IO completes.
 864  *
 865  * If resync/recovery is interrupted, returns -EINTR;
 866  * Otherwise, returns 0.
 867  */
 868 static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
 869 {
 870         int idx = sector_to_idx(sector_nr);
 871
 872         spin_lock_irq(&conf->resync_lock);
 873
 874         /* Wait until no block IO is waiting */
 875         wait_event_lock_irq(conf->wait_barrier,
 876                             !atomic_read(&conf->nr_waiting[idx]),
 877                             conf->resync_lock);
 878
 879         /* block any new IO from starting */
 880         atomic_inc(&conf->barrier[idx]);
 881         /*
 882          * In raise_barrier() we firstly increase conf->barrier[idx] then
 883          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
 884          * increase conf->nr_pending[idx] then check conf->barrier[idx].
 885          * A memory barrier here to make sure conf->nr_pending[idx] won't
 886          * be fetched before conf->barrier[idx] is increased. Otherwise
 887          * there will be a race between raise_barrier() and _wait_barrier().
 888          */
 889         smp_mb__after_atomic();
 890
 891         /* For these conditions we must wait:
 892          * A: while the array is in frozen state
 893          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
 894          *    existing in corresponding I/O barrier bucket.
 895          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
 896          *    max resync count which allowed on current I/O barrier bucket.
 897          */
 898         wait_event_lock_irq(conf->wait_barrier,
 899                             (!conf->array_frozen &&
 900                              !atomic_read(&conf->nr_pending[idx]) &&
 901                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
 902                                 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
 903                             conf->resync_lock);
 904
 905         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
 906                 atomic_dec(&conf->barrier[idx]);
 907                 spin_unlock_irq(&conf->resync_lock);
 908                 wake_up(&conf->wait_barrier);
 909                 return -EINTR;
 910         }
 911
 912         atomic_inc(&conf->nr_sync_pending);
 913         spin_unlock_irq(&conf->resync_lock);
 914
 915         return 0;
 916 }
 917
 918 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
 919 {
 920         int idx = sector_to_idx(sector_nr);
 921
 922         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
 923
 924         atomic_dec(&conf->barrier[idx]);
 925         atomic_dec(&conf->nr_sync_pending);
 926         wake_up(&conf->wait_barrier);
 927 }
 928
 929 static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
 930 {
 931         bool ret = true;
 932
 933         /*
 934          * We need to increase conf->nr_pending[idx] very early here,
 935          * then raise_barrier() can be blocked when it waits for
 936          * conf->nr_pending[idx] to be 0. Then we can avoid holding
 937          * conf->resync_lock when there is no barrier raised in same
 938          * barrier unit bucket. Also if the array is frozen, I/O
 939          * should be blocked until array is unfrozen.
 940          */
 941         atomic_inc(&conf->nr_pending[idx]);
 942         /*
 943          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
 944          * check conf->barrier[idx]. In raise_barrier() we firstly increase
 945          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
 946          * barrier is necessary here to make sure conf->barrier[idx] won't be
 947          * fetched before conf->nr_pending[idx] is increased. Otherwise there
 948          * will be a race between _wait_barrier() and raise_barrier().
 949          */
 950         smp_mb__after_atomic();
 951
 952         /*
 953          * Don't worry about checking two atomic_t variables at same time
 954          * here. If during we check conf->barrier[idx], the array is
 955          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
 956          * 0, it is safe to return and make the I/O continue. Because the
 957          * array is frozen, all I/O returned here will eventually complete
 958          * or be queued, no race will happen. See code comment in
 959          * frozen_array().
 960          */
 961         if (!READ_ONCE(conf->array_frozen) &&
 962             !atomic_read(&conf->barrier[idx]))
 963                 return ret;
 964
 965         /*
 966          * After holding conf->resync_lock, conf->nr_pending[idx]
 967          * should be decreased before waiting for barrier to drop.
 968          * Otherwise, we may encounter a race condition because
 969          * raise_barrer() might be waiting for conf->nr_pending[idx]
 970          * to be 0 at same time.
 971          */
 972         spin_lock_irq(&conf->resync_lock);
 973         atomic_inc(&conf->nr_waiting[idx]);
 974         atomic_dec(&conf->nr_pending[idx]);
 975         /*
 976          * In case freeze_array() is waiting for
 977          * get_unqueued_pending() == extra
 978          */
 979         wake_up_barrier(conf);
 980         /* Wait for the barrier in same barrier unit bucket to drop. */
 981
 982         /* Return false when nowait flag is set */
 983         if (nowait) {
 984                 ret = false;
 985         } else {
 986                 wait_event_lock_irq(conf->wait_barrier,
 987                                 !conf->array_frozen &&
 988                                 !atomic_read(&conf->barrier[idx]),
 989                                 conf->resync_lock);
 990                 atomic_inc(&conf->nr_pending[idx]);
 991         }
 992
 993         atomic_dec(&conf->nr_waiting[idx]);
 994         spin_unlock_irq(&conf->resync_lock);
 995         return ret;
 996 }
 997
 998 static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
 999 {
1000         int idx = sector_to_idx(sector_nr);
1001         bool ret = true;
1002
1003         /*
1004          * Very similar to _wait_barrier(). The difference is, for read
1005          * I/O we don't need wait for sync I/O, but if the whole array
1006          * is frozen, the read I/O still has to wait until the array is
1007          * unfrozen. Since there is no ordering requirement with
1008          * conf->barrier[idx] here, memory barrier is unnecessary as well.
1009          */
1010         atomic_inc(&conf->nr_pending[idx]);
1011
1012         if (!READ_ONCE(conf->array_frozen))
1013                 return ret;
1014
1015         spin_lock_irq(&conf->resync_lock);
1016         atomic_inc(&conf->nr_waiting[idx]);
1017         atomic_dec(&conf->nr_pending[idx]);
1018         /*
1019          * In case freeze_array() is waiting for
1020          * get_unqueued_pending() == extra
1021          */
1022         wake_up_barrier(conf);
1023         /* Wait for array to be unfrozen */
1024
1025         /* Return false when nowait flag is set */
1026         if (nowait) {
1027                 /* Return false when nowait flag is set */
1028                 ret = false;
1029         } else {
1030                 wait_event_lock_irq(conf->wait_barrier,
1031                                 !conf->array_frozen,
1032                                 conf->resync_lock);
1033                 atomic_inc(&conf->nr_pending[idx]);
1034         }
1035
1036         atomic_dec(&conf->nr_waiting[idx]);
1037         spin_unlock_irq(&conf->resync_lock);
1038         return ret;
1039 }
1040
1041 static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1042 {
1043         int idx = sector_to_idx(sector_nr);
1044
1045         return _wait_barrier(conf, idx, nowait);
1046 }
1047
1048 static void _allow_barrier(struct r1conf *conf, int idx)
1049 {
1050         atomic_dec(&conf->nr_pending[idx]);
1051         wake_up_barrier(conf);
1052 }
1053
1054 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1055 {
1056         int idx = sector_to_idx(sector_nr);
1057
1058         _allow_barrier(conf, idx);
1059 }
1060
1061 /* conf->resync_lock should be held */
1062 static int get_unqueued_pending(struct r1conf *conf)
1063 {
1064         int idx, ret;
1065
1066         ret = atomic_read(&conf->nr_sync_pending);
1067         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1068                 ret += atomic_read(&conf->nr_pending[idx]) -
1069                         atomic_read(&conf->nr_queued[idx]);
1070
1071         return ret;
1072 }
1073
1074 static void freeze_array(struct r1conf *conf, int extra)
1075 {
1076         /* Stop sync I/O and normal I/O and wait for everything to
1077          * go quiet.
1078          * This is called in two situations:
1079          * 1) management command handlers (reshape, remove disk, quiesce).
1080          * 2) one normal I/O request failed.
1081
1082          * After array_frozen is set to 1, new sync IO will be blocked at
1083          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1084          * or wait_read_barrier(). The flying I/Os will either complete or be
1085          * queued. When everything goes quite, there are only queued I/Os left.
1086
1087          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1088          * barrier bucket index which this I/O request hits. When all sync and
1089          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1090          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1091          * in handle_read_error(), we may call freeze_array() before trying to
1092          * fix the read error. In this case, the error read I/O is not queued,
1093          * so get_unqueued_pending() == 1.
1094          *
1095          * Therefore before this function returns, we need to wait until
1096          * get_unqueued_pendings(conf) gets equal to extra. For
1097          * normal I/O context, extra is 1, in rested situations extra is 0.
1098          */
1099         spin_lock_irq(&conf->resync_lock);
1100         conf->array_frozen = 1;
1101         raid1_log(conf->mddev, "wait freeze");
1102         wait_event_lock_irq_cmd(
1103                 conf->wait_barrier,
1104                 get_unqueued_pending(conf) == extra,
1105                 conf->resync_lock,
1106                 flush_pending_writes(conf));
1107         spin_unlock_irq(&conf->resync_lock);
1108 }
1109 static void unfreeze_array(struct r1conf *conf)
1110 {
1111         /* reverse the effect of the freeze */
1112         spin_lock_irq(&conf->resync_lock);
1113         conf->array_frozen = 0;
1114         spin_unlock_irq(&conf->resync_lock);
1115         wake_up(&conf->wait_barrier);
1116 }
1117
1118 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1119                                            struct bio *bio)
1120 {
1121         int size = bio->bi_iter.bi_size;
1122         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1123         int i = 0;
1124         struct bio *behind_bio = NULL;
1125
1126         behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1127                                       &r1_bio->mddev->bio_set);
1128
1129         /* discard op, we don't support writezero/writesame yet */
1130         if (!bio_has_data(bio)) {
1131                 behind_bio->bi_iter.bi_size = size;
1132                 goto skip_copy;
1133         }
1134
1135         while (i < vcnt && size) {
1136                 struct page *page;
1137                 int len = min_t(int, PAGE_SIZE, size);
1138
1139                 page = alloc_page(GFP_NOIO);
1140                 if (unlikely(!page))
1141                         goto free_pages;
1142
1143                 if (!bio_add_page(behind_bio, page, len, 0)) {
1144                         put_page(page);
1145                         goto free_pages;
1146                 }
1147
1148                 size -= len;
1149                 i++;
1150         }
1151
1152         bio_copy_data(behind_bio, bio);
1153 skip_copy:
1154         r1_bio->behind_master_bio = behind_bio;
1155         set_bit(R1BIO_BehindIO, &r1_bio->state);
1156
1157         return;
1158
1159 free_pages:
1160         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1161                  bio->bi_iter.bi_size);
1162         bio_free_pages(behind_bio);
1163         bio_put(behind_bio);
1164 }
1165
1166 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1167 {
1168         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1169                                                   cb);
1170         struct mddev *mddev = plug->cb.data;
1171         struct r1conf *conf = mddev->private;
1172         struct bio *bio;
1173
1174         if (from_schedule) {
1175                 spin_lock_irq(&conf->device_lock);
1176                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1177                 spin_unlock_irq(&conf->device_lock);
1178                 wake_up_barrier(conf);
1179                 md_wakeup_thread(mddev->thread);
1180                 kfree(plug);
1181                 return;
1182         }
1183
1184         /* we aren't scheduling, so we can do the write-out directly. */
1185         bio = bio_list_get(&plug->pending);
1186         flush_bio_list(conf, bio);
1187         kfree(plug);
1188 }
1189
1190 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1191 {
1192         r1_bio->master_bio = bio;
1193         r1_bio->sectors = bio_sectors(bio);
1194         r1_bio->state = 0;
1195         r1_bio->mddev = mddev;
1196         r1_bio->sector = bio->bi_iter.bi_sector;
1197 }
1198
1199 static inline struct r1bio *
1200 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1201 {
1202         struct r1conf *conf = mddev->private;
1203         struct r1bio *r1_bio;
1204
1205         r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1206         /* Ensure no bio records IO_BLOCKED */
1207         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1208         init_r1bio(r1_bio, mddev, bio);
1209         return r1_bio;
1210 }
1211
1212 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1213                                int max_read_sectors, struct r1bio *r1_bio)
1214 {
1215         struct r1conf *conf = mddev->private;
1216         struct raid1_info *mirror;
1217         struct bio *read_bio;
1218         struct bitmap *bitmap = mddev->bitmap;
1219         const enum req_op op = bio_op(bio);
1220         const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1221         int max_sectors;
1222         int rdisk;
1223         bool r1bio_existed = !!r1_bio;
1224         char b[BDEVNAME_SIZE];
1225
1226         /*
1227          * If r1_bio is set, we are blocking the raid1d thread
1228          * so there is a tiny risk of deadlock.  So ask for
1229          * emergency memory if needed.
1230          */
1231         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1232
1233         if (r1bio_existed) {
1234                 /* Need to get the block device name carefully */
1235                 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
1236
1237                 if (rdev)
1238                         snprintf(b, sizeof(b), "%pg", rdev->bdev);
1239                 else
1240                         strcpy(b, "???");
1241         }
1242
1243         /*
1244          * Still need barrier for READ in case that whole
1245          * array is frozen.
1246          */
1247         if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1248                                 bio->bi_opf & REQ_NOWAIT)) {
1249                 bio_wouldblock_error(bio);
1250                 return;
1251         }
1252
1253         if (!r1_bio)
1254                 r1_bio = alloc_r1bio(mddev, bio);
1255         else
1256                 init_r1bio(r1_bio, mddev, bio);
1257         r1_bio->sectors = max_read_sectors;
1258
1259         /*
1260          * make_request() can abort the operation when read-ahead is being
1261          * used and no empty request is available.
1262          */
1263         rdisk = read_balance(conf, r1_bio, &max_sectors);
1264
1265         if (rdisk < 0) {
1266                 /* couldn't find anywhere to read from */
1267                 if (r1bio_existed) {
1268                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1269                                             mdname(mddev),
1270                                             b,
1271                                             (unsigned long long)r1_bio->sector);
1272                 }
1273                 raid_end_bio_io(r1_bio);
1274                 return;
1275         }
1276         mirror = conf->mirrors + rdisk;
1277
1278         if (r1bio_existed)
1279                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %pg\n",
1280                                     mdname(mddev),
1281                                     (unsigned long long)r1_bio->sector,
1282                                     mirror->rdev->bdev);
1283
1284         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1285             bitmap) {
1286                 /*
1287                  * Reading from a write-mostly device must take care not to
1288                  * over-take any writes that are 'behind'
1289                  */
1290                 raid1_log(mddev, "wait behind writes");
1291                 wait_event(bitmap->behind_wait,
1292                            atomic_read(&bitmap->behind_writes) == 0);
1293         }
1294
1295         if (max_sectors < bio_sectors(bio)) {
1296                 struct bio *split = bio_split(bio, max_sectors,
1297                                               gfp, &conf->bio_split);
1298                 bio_chain(split, bio);
1299                 submit_bio_noacct(bio);
1300                 bio = split;
1301                 r1_bio->master_bio = bio;
1302                 r1_bio->sectors = max_sectors;
1303         }
1304
1305         r1_bio->read_disk = rdisk;
1306         if (!r1bio_existed) {
1307                 md_account_bio(mddev, &bio);
1308                 r1_bio->master_bio = bio;
1309         }
1310         read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1311                                    &mddev->bio_set);
1312
1313         r1_bio->bios[rdisk] = read_bio;
1314
1315         read_bio->bi_iter.bi_sector = r1_bio->sector +
1316                 mirror->rdev->data_offset;
1317         read_bio->bi_end_io = raid1_end_read_request;
1318         read_bio->bi_opf = op | do_sync;
1319         if (test_bit(FailFast, &mirror->rdev->flags) &&
1320             test_bit(R1BIO_FailFast, &r1_bio->state))
1321                 read_bio->bi_opf |= MD_FAILFAST;
1322         read_bio->bi_private = r1_bio;
1323
1324         if (mddev->gendisk)
1325                 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1326                                       r1_bio->sector);
1327
1328         submit_bio_noacct(read_bio);
1329 }
1330
1331 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1332                                 int max_write_sectors)
1333 {
1334         struct r1conf *conf = mddev->private;
1335         struct r1bio *r1_bio;
1336         int i, disks;
1337         struct bitmap *bitmap = mddev->bitmap;
1338         unsigned long flags;
1339         struct md_rdev *blocked_rdev;
1340         int first_clone;
1341         int max_sectors;
1342         bool write_behind = false;
1343         bool is_discard = (bio_op(bio) == REQ_OP_DISCARD);
1344
1345         if (mddev_is_clustered(mddev) &&
1346              md_cluster_ops->area_resyncing(mddev, WRITE,
1347                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1348
1349                 DEFINE_WAIT(w);
1350                 if (bio->bi_opf & REQ_NOWAIT) {
1351                         bio_wouldblock_error(bio);
1352                         return;
1353                 }
1354                 for (;;) {
1355                         prepare_to_wait(&conf->wait_barrier,
1356                                         &w, TASK_IDLE);
1357                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1358                                                         bio->bi_iter.bi_sector,
1359                                                         bio_end_sector(bio)))
1360                                 break;
1361                         schedule();
1362                 }
1363                 finish_wait(&conf->wait_barrier, &w);
1364         }
1365
1366         /*
1367          * Register the new request and wait if the reconstruction
1368          * thread has put up a bar for new requests.
1369          * Continue immediately if no resync is active currently.
1370          */
1371         if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1372                                 bio->bi_opf & REQ_NOWAIT)) {
1373                 bio_wouldblock_error(bio);
1374                 return;
1375         }
1376
1377  retry_write:
1378         r1_bio = alloc_r1bio(mddev, bio);
1379         r1_bio->sectors = max_write_sectors;
1380
1381         /* first select target devices under rcu_lock and
1382          * inc refcount on their rdev.  Record them by setting
1383          * bios[x] to bio
1384          * If there are known/acknowledged bad blocks on any device on
1385          * which we have seen a write error, we want to avoid writing those
1386          * blocks.
1387          * This potentially requires several writes to write around
1388          * the bad blocks.  Each set of writes gets it's own r1bio
1389          * with a set of bios attached.
1390          */
1391
1392         disks = conf->raid_disks * 2;
1393         blocked_rdev = NULL;
1394         max_sectors = r1_bio->sectors;
1395         for (i = 0;  i < disks; i++) {
1396                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1397
1398                 /*
1399                  * The write-behind io is only attempted on drives marked as
1400                  * write-mostly, which means we could allocate write behind
1401                  * bio later.
1402                  */
1403                 if (!is_discard && rdev && test_bit(WriteMostly, &rdev->flags))
1404                         write_behind = true;
1405
1406                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1407                         atomic_inc(&rdev->nr_pending);
1408                         blocked_rdev = rdev;
1409                         break;
1410                 }
1411                 r1_bio->bios[i] = NULL;
1412                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1413                         if (i < conf->raid_disks)
1414                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1415                         continue;
1416                 }
1417
1418                 atomic_inc(&rdev->nr_pending);
1419                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1420                         sector_t first_bad;
1421                         int bad_sectors;
1422                         int is_bad;
1423
1424                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1425                                              &first_bad, &bad_sectors);
1426                         if (is_bad < 0) {
1427                                 /* mustn't write here until the bad block is
1428                                  * acknowledged*/
1429                                 set_bit(BlockedBadBlocks, &rdev->flags);
1430                                 blocked_rdev = rdev;
1431                                 break;
1432                         }
1433                         if (is_bad && first_bad <= r1_bio->sector) {
1434                                 /* Cannot write here at all */
1435                                 bad_sectors -= (r1_bio->sector - first_bad);
1436                                 if (bad_sectors < max_sectors)
1437                                         /* mustn't write more than bad_sectors
1438                                          * to other devices yet
1439                                          */
1440                                         max_sectors = bad_sectors;
1441                                 rdev_dec_pending(rdev, mddev);
1442                                 /* We don't set R1BIO_Degraded as that
1443                                  * only applies if the disk is
1444                                  * missing, so it might be re-added,
1445                                  * and we want to know to recover this
1446                                  * chunk.
1447                                  * In this case the device is here,
1448                                  * and the fact that this chunk is not
1449                                  * in-sync is recorded in the bad
1450                                  * block log
1451                                  */
1452                                 continue;
1453                         }
1454                         if (is_bad) {
1455                                 int good_sectors = first_bad - r1_bio->sector;
1456                                 if (good_sectors < max_sectors)
1457                                         max_sectors = good_sectors;
1458                         }
1459                 }
1460                 r1_bio->bios[i] = bio;
1461         }
1462
1463         if (unlikely(blocked_rdev)) {
1464                 /* Wait for this device to become unblocked */
1465                 int j;
1466
1467                 for (j = 0; j < i; j++)
1468                         if (r1_bio->bios[j])
1469                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1470                 free_r1bio(r1_bio);
1471                 allow_barrier(conf, bio->bi_iter.bi_sector);
1472
1473                 if (bio->bi_opf & REQ_NOWAIT) {
1474                         bio_wouldblock_error(bio);
1475                         return;
1476                 }
1477                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1478                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1479                 wait_barrier(conf, bio->bi_iter.bi_sector, false);
1480                 goto retry_write;
1481         }
1482
1483         /*
1484          * When using a bitmap, we may call alloc_behind_master_bio below.
1485          * alloc_behind_master_bio allocates a copy of the data payload a page
1486          * at a time and thus needs a new bio that can fit the whole payload
1487          * this bio in page sized chunks.
1488          */
1489         if (write_behind && bitmap)
1490                 max_sectors = min_t(int, max_sectors,
1491                                     BIO_MAX_VECS * (PAGE_SIZE >> 9));
1492         if (max_sectors < bio_sectors(bio)) {
1493                 struct bio *split = bio_split(bio, max_sectors,
1494                                               GFP_NOIO, &conf->bio_split);
1495                 bio_chain(split, bio);
1496                 submit_bio_noacct(bio);
1497                 bio = split;
1498                 r1_bio->master_bio = bio;
1499                 r1_bio->sectors = max_sectors;
1500         }
1501
1502         md_account_bio(mddev, &bio);
1503         r1_bio->master_bio = bio;
1504         atomic_set(&r1_bio->remaining, 1);
1505         atomic_set(&r1_bio->behind_remaining, 0);
1506
1507         first_clone = 1;
1508
1509         for (i = 0; i < disks; i++) {
1510                 struct bio *mbio = NULL;
1511                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1512                 if (!r1_bio->bios[i])
1513                         continue;
1514
1515                 if (first_clone) {
1516                         /* do behind I/O ?
1517                          * Not if there are too many, or cannot
1518                          * allocate memory, or a reader on WriteMostly
1519                          * is waiting for behind writes to flush */
1520                         if (bitmap && write_behind &&
1521                             (atomic_read(&bitmap->behind_writes)
1522                              < mddev->bitmap_info.max_write_behind) &&
1523                             !waitqueue_active(&bitmap->behind_wait)) {
1524                                 alloc_behind_master_bio(r1_bio, bio);
1525                         }
1526
1527                         md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1528                                              test_bit(R1BIO_BehindIO, &r1_bio->state));
1529                         first_clone = 0;
1530                 }
1531
1532                 if (r1_bio->behind_master_bio) {
1533                         mbio = bio_alloc_clone(rdev->bdev,
1534                                                r1_bio->behind_master_bio,
1535                                                GFP_NOIO, &mddev->bio_set);
1536                         if (test_bit(CollisionCheck, &rdev->flags))
1537                                 wait_for_serialization(rdev, r1_bio);
1538                         if (test_bit(WriteMostly, &rdev->flags))
1539                                 atomic_inc(&r1_bio->behind_remaining);
1540                 } else {
1541                         mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1542                                                &mddev->bio_set);
1543
1544                         if (mddev->serialize_policy)
1545                                 wait_for_serialization(rdev, r1_bio);
1546                 }
1547
1548                 r1_bio->bios[i] = mbio;
1549
1550                 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1551                 mbio->bi_end_io = raid1_end_write_request;
1552                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1553                 if (test_bit(FailFast, &rdev->flags) &&
1554                     !test_bit(WriteMostly, &rdev->flags) &&
1555                     conf->raid_disks - mddev->degraded > 1)
1556                         mbio->bi_opf |= MD_FAILFAST;
1557                 mbio->bi_private = r1_bio;
1558
1559                 atomic_inc(&r1_bio->remaining);
1560
1561                 if (mddev->gendisk)
1562                         trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1563                                               r1_bio->sector);
1564                 /* flush_pending_writes() needs access to the rdev so...*/
1565                 mbio->bi_bdev = (void *)rdev;
1566                 if (!raid1_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) {
1567                         spin_lock_irqsave(&conf->device_lock, flags);
1568                         bio_list_add(&conf->pending_bio_list, mbio);
1569                         spin_unlock_irqrestore(&conf->device_lock, flags);
1570                         md_wakeup_thread(mddev->thread);
1571                 }
1572         }
1573
1574         r1_bio_write_done(r1_bio);
1575
1576         /* In case raid1d snuck in to freeze_array */
1577         wake_up_barrier(conf);
1578 }
1579
1580 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1581 {
1582         sector_t sectors;
1583
1584         if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1585             && md_flush_request(mddev, bio))
1586                 return true;
1587
1588         /*
1589          * There is a limit to the maximum size, but
1590          * the read/write handler might find a lower limit
1591          * due to bad blocks.  To avoid multiple splits,
1592          * we pass the maximum number of sectors down
1593          * and let the lower level perform the split.
1594          */
1595         sectors = align_to_barrier_unit_end(
1596                 bio->bi_iter.bi_sector, bio_sectors(bio));
1597
1598         if (bio_data_dir(bio) == READ)
1599                 raid1_read_request(mddev, bio, sectors, NULL);
1600         else {
1601                 if (!md_write_start(mddev,bio))
1602                         return false;
1603                 raid1_write_request(mddev, bio, sectors);
1604         }
1605         return true;
1606 }
1607
1608 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1609 {
1610         struct r1conf *conf = mddev->private;
1611         int i;
1612
1613         lockdep_assert_held(&mddev->lock);
1614
1615         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1616                    conf->raid_disks - mddev->degraded);
1617         for (i = 0; i < conf->raid_disks; i++) {
1618                 struct md_rdev *rdev = READ_ONCE(conf->mirrors[i].rdev);
1619
1620                 seq_printf(seq, "%s",
1621                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1622         }
1623         seq_printf(seq, "]");
1624 }
1625
1626 /**
1627  * raid1_error() - RAID1 error handler.
1628  * @mddev: affected md device.
1629  * @rdev: member device to fail.
1630  *
1631  * The routine acknowledges &rdev failure and determines new @mddev state.
1632  * If it failed, then:
1633  *      - &MD_BROKEN flag is set in &mddev->flags.
1634  *      - recovery is disabled.
1635  * Otherwise, it must be degraded:
1636  *      - recovery is interrupted.
1637  *      - &mddev->degraded is bumped.
1638  *
1639  * @rdev is marked as &Faulty excluding case when array is failed and
1640  * &mddev->fail_last_dev is off.
1641  */
1642 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1643 {
1644         struct r1conf *conf = mddev->private;
1645         unsigned long flags;
1646
1647         spin_lock_irqsave(&conf->device_lock, flags);
1648
1649         if (test_bit(In_sync, &rdev->flags) &&
1650             (conf->raid_disks - mddev->degraded) == 1) {
1651                 set_bit(MD_BROKEN, &mddev->flags);
1652
1653                 if (!mddev->fail_last_dev) {
1654                         conf->recovery_disabled = mddev->recovery_disabled;
1655                         spin_unlock_irqrestore(&conf->device_lock, flags);
1656                         return;
1657                 }
1658         }
1659         set_bit(Blocked, &rdev->flags);
1660         if (test_and_clear_bit(In_sync, &rdev->flags))
1661                 mddev->degraded++;
1662         set_bit(Faulty, &rdev->flags);
1663         spin_unlock_irqrestore(&conf->device_lock, flags);
1664         /*
1665          * if recovery is running, make sure it aborts.
1666          */
1667         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1668         set_mask_bits(&mddev->sb_flags, 0,
1669                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1670         pr_crit("md/raid1:%s: Disk failure on %pg, disabling device.\n"
1671                 "md/raid1:%s: Operation continuing on %d devices.\n",
1672                 mdname(mddev), rdev->bdev,
1673                 mdname(mddev), conf->raid_disks - mddev->degraded);
1674 }
1675
1676 static void print_conf(struct r1conf *conf)
1677 {
1678         int i;
1679
1680         pr_debug("RAID1 conf printout:\n");
1681         if (!conf) {
1682                 pr_debug("(!conf)\n");
1683                 return;
1684         }
1685         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1686                  conf->raid_disks);
1687
1688         lockdep_assert_held(&conf->mddev->reconfig_mutex);
1689         for (i = 0; i < conf->raid_disks; i++) {
1690                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1691                 if (rdev)
1692                         pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
1693                                  i, !test_bit(In_sync, &rdev->flags),
1694                                  !test_bit(Faulty, &rdev->flags),
1695                                  rdev->bdev);
1696         }
1697 }
1698
1699 static void close_sync(struct r1conf *conf)
1700 {
1701         int idx;
1702
1703         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1704                 _wait_barrier(conf, idx, false);
1705                 _allow_barrier(conf, idx);
1706         }
1707
1708         mempool_exit(&conf->r1buf_pool);
1709 }
1710
1711 static int raid1_spare_active(struct mddev *mddev)
1712 {
1713         int i;
1714         struct r1conf *conf = mddev->private;
1715         int count = 0;
1716         unsigned long flags;
1717
1718         /*
1719          * Find all failed disks within the RAID1 configuration
1720          * and mark them readable.
1721          * Called under mddev lock, so rcu protection not needed.
1722          * device_lock used to avoid races with raid1_end_read_request
1723          * which expects 'In_sync' flags and ->degraded to be consistent.
1724          */
1725         spin_lock_irqsave(&conf->device_lock, flags);
1726         for (i = 0; i < conf->raid_disks; i++) {
1727                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1728                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1729                 if (repl
1730                     && !test_bit(Candidate, &repl->flags)
1731                     && repl->recovery_offset == MaxSector
1732                     && !test_bit(Faulty, &repl->flags)
1733                     && !test_and_set_bit(In_sync, &repl->flags)) {
1734                         /* replacement has just become active */
1735                         if (!rdev ||
1736                             !test_and_clear_bit(In_sync, &rdev->flags))
1737                                 count++;
1738                         if (rdev) {
1739                                 /* Replaced device not technically
1740                                  * faulty, but we need to be sure
1741                                  * it gets removed and never re-added
1742                                  */
1743                                 set_bit(Faulty, &rdev->flags);
1744                                 sysfs_notify_dirent_safe(
1745                                         rdev->sysfs_state);
1746                         }
1747                 }
1748                 if (rdev
1749                     && rdev->recovery_offset == MaxSector
1750                     && !test_bit(Faulty, &rdev->flags)
1751                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1752                         count++;
1753                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1754                 }
1755         }
1756         mddev->degraded -= count;
1757         spin_unlock_irqrestore(&conf->device_lock, flags);
1758
1759         print_conf(conf);
1760         return count;
1761 }
1762
1763 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1764 {
1765         struct r1conf *conf = mddev->private;
1766         int err = -EEXIST;
1767         int mirror = 0, repl_slot = -1;
1768         struct raid1_info *p;
1769         int first = 0;
1770         int last = conf->raid_disks - 1;
1771
1772         if (mddev->recovery_disabled == conf->recovery_disabled)
1773                 return -EBUSY;
1774
1775         if (md_integrity_add_rdev(rdev, mddev))
1776                 return -ENXIO;
1777
1778         if (rdev->raid_disk >= 0)
1779                 first = last = rdev->raid_disk;
1780
1781         /*
1782          * find the disk ... but prefer rdev->saved_raid_disk
1783          * if possible.
1784          */
1785         if (rdev->saved_raid_disk >= 0 &&
1786             rdev->saved_raid_disk >= first &&
1787             rdev->saved_raid_disk < conf->raid_disks &&
1788             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1789                 first = last = rdev->saved_raid_disk;
1790
1791         for (mirror = first; mirror <= last; mirror++) {
1792                 p = conf->mirrors + mirror;
1793                 if (!p->rdev) {
1794                         if (mddev->gendisk)
1795                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1796                                                   rdev->data_offset << 9);
1797
1798                         p->head_position = 0;
1799                         rdev->raid_disk = mirror;
1800                         err = 0;
1801                         /* As all devices are equivalent, we don't need a full recovery
1802                          * if this was recently any drive of the array
1803                          */
1804                         if (rdev->saved_raid_disk < 0)
1805                                 conf->fullsync = 1;
1806                         WRITE_ONCE(p->rdev, rdev);
1807                         break;
1808                 }
1809                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1810                     p[conf->raid_disks].rdev == NULL && repl_slot < 0)
1811                         repl_slot = mirror;
1812         }
1813
1814         if (err && repl_slot >= 0) {
1815                 /* Add this device as a replacement */
1816                 p = conf->mirrors + repl_slot;
1817                 clear_bit(In_sync, &rdev->flags);
1818                 set_bit(Replacement, &rdev->flags);
1819                 rdev->raid_disk = repl_slot;
1820                 err = 0;
1821                 conf->fullsync = 1;
1822                 WRITE_ONCE(p[conf->raid_disks].rdev, rdev);
1823         }
1824
1825         print_conf(conf);
1826         return err;
1827 }
1828
1829 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1830 {
1831         struct r1conf *conf = mddev->private;
1832         int err = 0;
1833         int number = rdev->raid_disk;
1834         struct raid1_info *p = conf->mirrors + number;
1835
1836         if (unlikely(number >= conf->raid_disks))
1837                 goto abort;
1838
1839         if (rdev != p->rdev)
1840                 p = conf->mirrors + conf->raid_disks + number;
1841
1842         print_conf(conf);
1843         if (rdev == p->rdev) {
1844                 if (test_bit(In_sync, &rdev->flags) ||
1845                     atomic_read(&rdev->nr_pending)) {
1846                         err = -EBUSY;
1847                         goto abort;
1848                 }
1849                 /* Only remove non-faulty devices if recovery
1850                  * is not possible.
1851                  */
1852                 if (!test_bit(Faulty, &rdev->flags) &&
1853                     mddev->recovery_disabled != conf->recovery_disabled &&
1854                     mddev->degraded < conf->raid_disks) {
1855                         err = -EBUSY;
1856                         goto abort;
1857                 }
1858                 WRITE_ONCE(p->rdev, NULL);
1859                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1860                         /* We just removed a device that is being replaced.
1861                          * Move down the replacement.  We drain all IO before
1862                          * doing this to avoid confusion.
1863                          */
1864                         struct md_rdev *repl =
1865                                 conf->mirrors[conf->raid_disks + number].rdev;
1866                         freeze_array(conf, 0);
1867                         if (atomic_read(&repl->nr_pending)) {
1868                                 /* It means that some queued IO of retry_list
1869                                  * hold repl. Thus, we cannot set replacement
1870                                  * as NULL, avoiding rdev NULL pointer
1871                                  * dereference in sync_request_write and
1872                                  * handle_write_finished.
1873                                  */
1874                                 err = -EBUSY;
1875                                 unfreeze_array(conf);
1876                                 goto abort;
1877                         }
1878                         clear_bit(Replacement, &repl->flags);
1879                         WRITE_ONCE(p->rdev, repl);
1880                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1881                         unfreeze_array(conf);
1882                 }
1883
1884                 clear_bit(WantReplacement, &rdev->flags);
1885                 err = md_integrity_register(mddev);
1886         }
1887 abort:
1888
1889         print_conf(conf);
1890         return err;
1891 }
1892
1893 static void end_sync_read(struct bio *bio)
1894 {
1895         struct r1bio *r1_bio = get_resync_r1bio(bio);
1896
1897         update_head_pos(r1_bio->read_disk, r1_bio);
1898
1899         /*
1900          * we have read a block, now it needs to be re-written,
1901          * or re-read if the read failed.
1902          * We don't do much here, just schedule handling by raid1d
1903          */
1904         if (!bio->bi_status)
1905                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1906
1907         if (atomic_dec_and_test(&r1_bio->remaining))
1908                 reschedule_retry(r1_bio);
1909 }
1910
1911 static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1912 {
1913         sector_t sync_blocks = 0;
1914         sector_t s = r1_bio->sector;
1915         long sectors_to_go = r1_bio->sectors;
1916
1917         /* make sure these bits don't get cleared. */
1918         do {
1919                 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1920                 s += sync_blocks;
1921                 sectors_to_go -= sync_blocks;
1922         } while (sectors_to_go > 0);
1923 }
1924
1925 static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1926 {
1927         if (atomic_dec_and_test(&r1_bio->remaining)) {
1928                 struct mddev *mddev = r1_bio->mddev;
1929                 int s = r1_bio->sectors;
1930
1931                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1932                     test_bit(R1BIO_WriteError, &r1_bio->state))
1933                         reschedule_retry(r1_bio);
1934                 else {
1935                         put_buf(r1_bio);
1936                         md_done_sync(mddev, s, uptodate);
1937                 }
1938         }
1939 }
1940
1941 static void end_sync_write(struct bio *bio)
1942 {
1943         int uptodate = !bio->bi_status;
1944         struct r1bio *r1_bio = get_resync_r1bio(bio);
1945         struct mddev *mddev = r1_bio->mddev;
1946         struct r1conf *conf = mddev->private;
1947         sector_t first_bad;
1948         int bad_sectors;
1949         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1950
1951         if (!uptodate) {
1952                 abort_sync_write(mddev, r1_bio);
1953                 set_bit(WriteErrorSeen, &rdev->flags);
1954                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1955                         set_bit(MD_RECOVERY_NEEDED, &
1956                                 mddev->recovery);
1957                 set_bit(R1BIO_WriteError, &r1_bio->state);
1958         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1959                                &first_bad, &bad_sectors) &&
1960                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1961                                 r1_bio->sector,
1962                                 r1_bio->sectors,
1963                                 &first_bad, &bad_sectors)
1964                 )
1965                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1966
1967         put_sync_write_buf(r1_bio, uptodate);
1968 }
1969
1970 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1971                            int sectors, struct page *page, blk_opf_t rw)
1972 {
1973         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1974                 /* success */
1975                 return 1;
1976         if (rw == REQ_OP_WRITE) {
1977                 set_bit(WriteErrorSeen, &rdev->flags);
1978                 if (!test_and_set_bit(WantReplacement,
1979                                       &rdev->flags))
1980                         set_bit(MD_RECOVERY_NEEDED, &
1981                                 rdev->mddev->recovery);
1982         }
1983         /* need to record an error - either for the block or the device */
1984         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1985                 md_error(rdev->mddev, rdev);
1986         return 0;
1987 }
1988
1989 static int fix_sync_read_error(struct r1bio *r1_bio)
1990 {
1991         /* Try some synchronous reads of other devices to get
1992          * good data, much like with normal read errors.  Only
1993          * read into the pages we already have so we don't
1994          * need to re-issue the read request.
1995          * We don't need to freeze the array, because being in an
1996          * active sync request, there is no normal IO, and
1997          * no overlapping syncs.
1998          * We don't need to check is_badblock() again as we
1999          * made sure that anything with a bad block in range
2000          * will have bi_end_io clear.
2001          */
2002         struct mddev *mddev = r1_bio->mddev;
2003         struct r1conf *conf = mddev->private;
2004         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2005         struct page **pages = get_resync_pages(bio)->pages;
2006         sector_t sect = r1_bio->sector;
2007         int sectors = r1_bio->sectors;
2008         int idx = 0;
2009         struct md_rdev *rdev;
2010
2011         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2012         if (test_bit(FailFast, &rdev->flags)) {
2013                 /* Don't try recovering from here - just fail it
2014                  * ... unless it is the last working device of course */
2015                 md_error(mddev, rdev);
2016                 if (test_bit(Faulty, &rdev->flags))
2017                         /* Don't try to read from here, but make sure
2018                          * put_buf does it's thing
2019                          */
2020                         bio->bi_end_io = end_sync_write;
2021         }
2022
2023         while(sectors) {
2024                 int s = sectors;
2025                 int d = r1_bio->read_disk;
2026                 int success = 0;
2027                 int start;
2028
2029                 if (s > (PAGE_SIZE>>9))
2030                         s = PAGE_SIZE >> 9;
2031                 do {
2032                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2033                                 /* No rcu protection needed here devices
2034                                  * can only be removed when no resync is
2035                                  * active, and resync is currently active
2036                                  */
2037                                 rdev = conf->mirrors[d].rdev;
2038                                 if (sync_page_io(rdev, sect, s<<9,
2039                                                  pages[idx],
2040                                                  REQ_OP_READ, false)) {
2041                                         success = 1;
2042                                         break;
2043                                 }
2044                         }
2045                         d++;
2046                         if (d == conf->raid_disks * 2)
2047                                 d = 0;
2048                 } while (!success && d != r1_bio->read_disk);
2049
2050                 if (!success) {
2051                         int abort = 0;
2052                         /* Cannot read from anywhere, this block is lost.
2053                          * Record a bad block on each device.  If that doesn't
2054                          * work just disable and interrupt the recovery.
2055                          * Don't fail devices as that won't really help.
2056                          */
2057                         pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2058                                             mdname(mddev), bio->bi_bdev,
2059                                             (unsigned long long)r1_bio->sector);
2060                         for (d = 0; d < conf->raid_disks * 2; d++) {
2061                                 rdev = conf->mirrors[d].rdev;
2062                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2063                                         continue;
2064                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2065                                         abort = 1;
2066                         }
2067                         if (abort) {
2068                                 conf->recovery_disabled =
2069                                         mddev->recovery_disabled;
2070                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2071                                 md_done_sync(mddev, r1_bio->sectors, 0);
2072                                 put_buf(r1_bio);
2073                                 return 0;
2074                         }
2075                         /* Try next page */
2076                         sectors -= s;
2077                         sect += s;
2078                         idx++;
2079                         continue;
2080                 }
2081
2082                 start = d;
2083                 /* write it back and re-read */
2084                 while (d != r1_bio->read_disk) {
2085                         if (d == 0)
2086                                 d = conf->raid_disks * 2;
2087                         d--;
2088                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2089                                 continue;
2090                         rdev = conf->mirrors[d].rdev;
2091                         if (r1_sync_page_io(rdev, sect, s,
2092                                             pages[idx],
2093                                             REQ_OP_WRITE) == 0) {
2094                                 r1_bio->bios[d]->bi_end_io = NULL;
2095                                 rdev_dec_pending(rdev, mddev);
2096                         }
2097                 }
2098                 d = start;
2099                 while (d != r1_bio->read_disk) {
2100                         if (d == 0)
2101                                 d = conf->raid_disks * 2;
2102                         d--;
2103                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2104                                 continue;
2105                         rdev = conf->mirrors[d].rdev;
2106                         if (r1_sync_page_io(rdev, sect, s,
2107                                             pages[idx],
2108                                             REQ_OP_READ) != 0)
2109                                 atomic_add(s, &rdev->corrected_errors);
2110                 }
2111                 sectors -= s;
2112                 sect += s;
2113                 idx ++;
2114         }
2115         set_bit(R1BIO_Uptodate, &r1_bio->state);
2116         bio->bi_status = 0;
2117         return 1;
2118 }
2119
2120 static void process_checks(struct r1bio *r1_bio)
2121 {
2122         /* We have read all readable devices.  If we haven't
2123          * got the block, then there is no hope left.
2124          * If we have, then we want to do a comparison
2125          * and skip the write if everything is the same.
2126          * If any blocks failed to read, then we need to
2127          * attempt an over-write
2128          */
2129         struct mddev *mddev = r1_bio->mddev;
2130         struct r1conf *conf = mddev->private;
2131         int primary;
2132         int i;
2133         int vcnt;
2134
2135         /* Fix variable parts of all bios */
2136         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2137         for (i = 0; i < conf->raid_disks * 2; i++) {
2138                 blk_status_t status;
2139                 struct bio *b = r1_bio->bios[i];
2140                 struct resync_pages *rp = get_resync_pages(b);
2141                 if (b->bi_end_io != end_sync_read)
2142                         continue;
2143                 /* fixup the bio for reuse, but preserve errno */
2144                 status = b->bi_status;
2145                 bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2146                 b->bi_status = status;
2147                 b->bi_iter.bi_sector = r1_bio->sector +
2148                         conf->mirrors[i].rdev->data_offset;
2149                 b->bi_end_io = end_sync_read;
2150                 rp->raid_bio = r1_bio;
2151                 b->bi_private = rp;
2152
2153                 /* initialize bvec table again */
2154                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2155         }
2156         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2157                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2158                     !r1_bio->bios[primary]->bi_status) {
2159                         r1_bio->bios[primary]->bi_end_io = NULL;
2160                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2161                         break;
2162                 }
2163         r1_bio->read_disk = primary;
2164         for (i = 0; i < conf->raid_disks * 2; i++) {
2165                 int j = 0;
2166                 struct bio *pbio = r1_bio->bios[primary];
2167                 struct bio *sbio = r1_bio->bios[i];
2168                 blk_status_t status = sbio->bi_status;
2169                 struct page **ppages = get_resync_pages(pbio)->pages;
2170                 struct page **spages = get_resync_pages(sbio)->pages;
2171                 struct bio_vec *bi;
2172                 int page_len[RESYNC_PAGES] = { 0 };
2173                 struct bvec_iter_all iter_all;
2174
2175                 if (sbio->bi_end_io != end_sync_read)
2176                         continue;
2177                 /* Now we can 'fixup' the error value */
2178                 sbio->bi_status = 0;
2179
2180                 bio_for_each_segment_all(bi, sbio, iter_all)
2181                         page_len[j++] = bi->bv_len;
2182
2183                 if (!status) {
2184                         for (j = vcnt; j-- ; ) {
2185                                 if (memcmp(page_address(ppages[j]),
2186                                            page_address(spages[j]),
2187                                            page_len[j]))
2188                                         break;
2189                         }
2190                 } else
2191                         j = 0;
2192                 if (j >= 0)
2193                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2194                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2195                               && !status)) {
2196                         /* No need to write to this device. */
2197                         sbio->bi_end_io = NULL;
2198                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2199                         continue;
2200                 }
2201
2202                 bio_copy_data(sbio, pbio);
2203         }
2204 }
2205
2206 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2207 {
2208         struct r1conf *conf = mddev->private;
2209         int i;
2210         int disks = conf->raid_disks * 2;
2211         struct bio *wbio;
2212
2213         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2214                 /* ouch - failed to read all of that. */
2215                 if (!fix_sync_read_error(r1_bio))
2216                         return;
2217
2218         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2219                 process_checks(r1_bio);
2220
2221         /*
2222          * schedule writes
2223          */
2224         atomic_set(&r1_bio->remaining, 1);
2225         for (i = 0; i < disks ; i++) {
2226                 wbio = r1_bio->bios[i];
2227                 if (wbio->bi_end_io == NULL ||
2228                     (wbio->bi_end_io == end_sync_read &&
2229                      (i == r1_bio->read_disk ||
2230                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2231                         continue;
2232                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2233                         abort_sync_write(mddev, r1_bio);
2234                         continue;
2235                 }
2236
2237                 wbio->bi_opf = REQ_OP_WRITE;
2238                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2239                         wbio->bi_opf |= MD_FAILFAST;
2240
2241                 wbio->bi_end_io = end_sync_write;
2242                 atomic_inc(&r1_bio->remaining);
2243                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2244
2245                 submit_bio_noacct(wbio);
2246         }
2247
2248         put_sync_write_buf(r1_bio, 1);
2249 }
2250
2251 /*
2252  * This is a kernel thread which:
2253  *
2254  *      1.      Retries failed read operations on working mirrors.
2255  *      2.      Updates the raid superblock when problems encounter.
2256  *      3.      Performs writes following reads for array synchronising.
2257  */
2258
2259 static void fix_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2260 {
2261         sector_t sect = r1_bio->sector;
2262         int sectors = r1_bio->sectors;
2263         int read_disk = r1_bio->read_disk;
2264         struct mddev *mddev = conf->mddev;
2265         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2266
2267         if (exceed_read_errors(mddev, rdev)) {
2268                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2269                 return;
2270         }
2271
2272         while(sectors) {
2273                 int s = sectors;
2274                 int d = read_disk;
2275                 int success = 0;
2276                 int start;
2277
2278                 if (s > (PAGE_SIZE>>9))
2279                         s = PAGE_SIZE >> 9;
2280
2281                 do {
2282                         sector_t first_bad;
2283                         int bad_sectors;
2284
2285                         rdev = conf->mirrors[d].rdev;
2286                         if (rdev &&
2287                             (test_bit(In_sync, &rdev->flags) ||
2288                              (!test_bit(Faulty, &rdev->flags) &&
2289                               rdev->recovery_offset >= sect + s)) &&
2290                             is_badblock(rdev, sect, s,
2291                                         &first_bad, &bad_sectors) == 0) {
2292                                 atomic_inc(&rdev->nr_pending);
2293                                 if (sync_page_io(rdev, sect, s<<9,
2294                                          conf->tmppage, REQ_OP_READ, false))
2295                                         success = 1;
2296                                 rdev_dec_pending(rdev, mddev);
2297                                 if (success)
2298                                         break;
2299                         }
2300
2301                         d++;
2302                         if (d == conf->raid_disks * 2)
2303                                 d = 0;
2304                 } while (d != read_disk);
2305
2306                 if (!success) {
2307                         /* Cannot read from anywhere - mark it bad */
2308                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2309                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2310                                 md_error(mddev, rdev);
2311                         break;
2312                 }
2313                 /* write it back and re-read */
2314                 start = d;
2315                 while (d != read_disk) {
2316                         if (d==0)
2317                                 d = conf->raid_disks * 2;
2318                         d--;
2319                         rdev = conf->mirrors[d].rdev;
2320                         if (rdev &&
2321                             !test_bit(Faulty, &rdev->flags)) {
2322                                 atomic_inc(&rdev->nr_pending);
2323                                 r1_sync_page_io(rdev, sect, s,
2324                                                 conf->tmppage, REQ_OP_WRITE);
2325                                 rdev_dec_pending(rdev, mddev);
2326                         }
2327                 }
2328                 d = start;
2329                 while (d != read_disk) {
2330                         if (d==0)
2331                                 d = conf->raid_disks * 2;
2332                         d--;
2333                         rdev = conf->mirrors[d].rdev;
2334                         if (rdev &&
2335                             !test_bit(Faulty, &rdev->flags)) {
2336                                 atomic_inc(&rdev->nr_pending);
2337                                 if (r1_sync_page_io(rdev, sect, s,
2338                                                 conf->tmppage, REQ_OP_READ)) {
2339                                         atomic_add(s, &rdev->corrected_errors);
2340                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %pg)\n",
2341                                                 mdname(mddev), s,
2342                                                 (unsigned long long)(sect +
2343                                                                      rdev->data_offset),
2344                                                 rdev->bdev);
2345                                 }
2346                                 rdev_dec_pending(rdev, mddev);
2347                         }
2348                 }
2349                 sectors -= s;
2350                 sect += s;
2351         }
2352 }
2353
2354 static int narrow_write_error(struct r1bio *r1_bio, int i)
2355 {
2356         struct mddev *mddev = r1_bio->mddev;
2357         struct r1conf *conf = mddev->private;
2358         struct md_rdev *rdev = conf->mirrors[i].rdev;
2359
2360         /* bio has the data to be written to device 'i' where
2361          * we just recently had a write error.
2362          * We repeatedly clone the bio and trim down to one block,
2363          * then try the write.  Where the write fails we record
2364          * a bad block.
2365          * It is conceivable that the bio doesn't exactly align with
2366          * blocks.  We must handle this somehow.
2367          *
2368          * We currently own a reference on the rdev.
2369          */
2370
2371         int block_sectors;
2372         sector_t sector;
2373         int sectors;
2374         int sect_to_write = r1_bio->sectors;
2375         int ok = 1;
2376
2377         if (rdev->badblocks.shift < 0)
2378                 return 0;
2379
2380         block_sectors = roundup(1 << rdev->badblocks.shift,
2381                                 bdev_logical_block_size(rdev->bdev) >> 9);
2382         sector = r1_bio->sector;
2383         sectors = ((sector + block_sectors)
2384                    & ~(sector_t)(block_sectors - 1))
2385                 - sector;
2386
2387         while (sect_to_write) {
2388                 struct bio *wbio;
2389                 if (sectors > sect_to_write)
2390                         sectors = sect_to_write;
2391                 /* Write at 'sector' for 'sectors'*/
2392
2393                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2394                         wbio = bio_alloc_clone(rdev->bdev,
2395                                                r1_bio->behind_master_bio,
2396                                                GFP_NOIO, &mddev->bio_set);
2397                 } else {
2398                         wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2399                                                GFP_NOIO, &mddev->bio_set);
2400                 }
2401
2402                 wbio->bi_opf = REQ_OP_WRITE;
2403                 wbio->bi_iter.bi_sector = r1_bio->sector;
2404                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2405
2406                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2407                 wbio->bi_iter.bi_sector += rdev->data_offset;
2408
2409                 if (submit_bio_wait(wbio) < 0)
2410                         /* failure! */
2411                         ok = rdev_set_badblocks(rdev, sector,
2412                                                 sectors, 0)
2413                                 && ok;
2414
2415                 bio_put(wbio);
2416                 sect_to_write -= sectors;
2417                 sector += sectors;
2418                 sectors = block_sectors;
2419         }
2420         return ok;
2421 }
2422
2423 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2424 {
2425         int m;
2426         int s = r1_bio->sectors;
2427         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2428                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2429                 struct bio *bio = r1_bio->bios[m];
2430                 if (bio->bi_end_io == NULL)
2431                         continue;
2432                 if (!bio->bi_status &&
2433                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2434                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2435                 }
2436                 if (bio->bi_status &&
2437                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2438                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2439                                 md_error(conf->mddev, rdev);
2440                 }
2441         }
2442         put_buf(r1_bio);
2443         md_done_sync(conf->mddev, s, 1);
2444 }
2445
2446 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2447 {
2448         int m, idx;
2449         bool fail = false;
2450
2451         for (m = 0; m < conf->raid_disks * 2 ; m++)
2452                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2453                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2454                         rdev_clear_badblocks(rdev,
2455                                              r1_bio->sector,
2456                                              r1_bio->sectors, 0);
2457                         rdev_dec_pending(rdev, conf->mddev);
2458                 } else if (r1_bio->bios[m] != NULL) {
2459                         /* This drive got a write error.  We need to
2460                          * narrow down and record precise write
2461                          * errors.
2462                          */
2463                         fail = true;
2464                         if (!narrow_write_error(r1_bio, m)) {
2465                                 md_error(conf->mddev,
2466                                          conf->mirrors[m].rdev);
2467                                 /* an I/O failed, we can't clear the bitmap */
2468                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2469                         }
2470                         rdev_dec_pending(conf->mirrors[m].rdev,
2471                                          conf->mddev);
2472                 }
2473         if (fail) {
2474                 spin_lock_irq(&conf->device_lock);
2475                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2476                 idx = sector_to_idx(r1_bio->sector);
2477                 atomic_inc(&conf->nr_queued[idx]);
2478                 spin_unlock_irq(&conf->device_lock);
2479                 /*
2480                  * In case freeze_array() is waiting for condition
2481                  * get_unqueued_pending() == extra to be true.
2482                  */
2483                 wake_up(&conf->wait_barrier);
2484                 md_wakeup_thread(conf->mddev->thread);
2485         } else {
2486                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2487                         close_write(r1_bio);
2488                 raid_end_bio_io(r1_bio);
2489         }
2490 }
2491
2492 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2493 {
2494         struct mddev *mddev = conf->mddev;
2495         struct bio *bio;
2496         struct md_rdev *rdev;
2497         sector_t sector;
2498
2499         clear_bit(R1BIO_ReadError, &r1_bio->state);
2500         /* we got a read error. Maybe the drive is bad.  Maybe just
2501          * the block and we can fix it.
2502          * We freeze all other IO, and try reading the block from
2503          * other devices.  When we find one, we re-write
2504          * and check it that fixes the read error.
2505          * This is all done synchronously while the array is
2506          * frozen
2507          */
2508
2509         bio = r1_bio->bios[r1_bio->read_disk];
2510         bio_put(bio);
2511         r1_bio->bios[r1_bio->read_disk] = NULL;
2512
2513         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2514         if (mddev->ro == 0
2515             && !test_bit(FailFast, &rdev->flags)) {
2516                 freeze_array(conf, 1);
2517                 fix_read_error(conf, r1_bio);
2518                 unfreeze_array(conf);
2519         } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2520                 md_error(mddev, rdev);
2521         } else {
2522                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2523         }
2524
2525         rdev_dec_pending(rdev, conf->mddev);
2526         sector = r1_bio->sector;
2527         bio = r1_bio->master_bio;
2528
2529         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2530         r1_bio->state = 0;
2531         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2532         allow_barrier(conf, sector);
2533 }
2534
2535 static void raid1d(struct md_thread *thread)
2536 {
2537         struct mddev *mddev = thread->mddev;
2538         struct r1bio *r1_bio;
2539         unsigned long flags;
2540         struct r1conf *conf = mddev->private;
2541         struct list_head *head = &conf->retry_list;
2542         struct blk_plug plug;
2543         int idx;
2544
2545         md_check_recovery(mddev);
2546
2547         if (!list_empty_careful(&conf->bio_end_io_list) &&
2548             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2549                 LIST_HEAD(tmp);
2550                 spin_lock_irqsave(&conf->device_lock, flags);
2551                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2552                         list_splice_init(&conf->bio_end_io_list, &tmp);
2553                 spin_unlock_irqrestore(&conf->device_lock, flags);
2554                 while (!list_empty(&tmp)) {
2555                         r1_bio = list_first_entry(&tmp, struct r1bio,
2556                                                   retry_list);
2557                         list_del(&r1_bio->retry_list);
2558                         idx = sector_to_idx(r1_bio->sector);
2559                         atomic_dec(&conf->nr_queued[idx]);
2560                         if (mddev->degraded)
2561                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2562                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2563                                 close_write(r1_bio);
2564                         raid_end_bio_io(r1_bio);
2565                 }
2566         }
2567
2568         blk_start_plug(&plug);
2569         for (;;) {
2570
2571                 flush_pending_writes(conf);
2572
2573                 spin_lock_irqsave(&conf->device_lock, flags);
2574                 if (list_empty(head)) {
2575                         spin_unlock_irqrestore(&conf->device_lock, flags);
2576                         break;
2577                 }
2578                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2579                 list_del(head->prev);
2580                 idx = sector_to_idx(r1_bio->sector);
2581                 atomic_dec(&conf->nr_queued[idx]);
2582                 spin_unlock_irqrestore(&conf->device_lock, flags);
2583
2584                 mddev = r1_bio->mddev;
2585                 conf = mddev->private;
2586                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2587                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2588                             test_bit(R1BIO_WriteError, &r1_bio->state))
2589                                 handle_sync_write_finished(conf, r1_bio);
2590                         else
2591                                 sync_request_write(mddev, r1_bio);
2592                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2593                            test_bit(R1BIO_WriteError, &r1_bio->state))
2594                         handle_write_finished(conf, r1_bio);
2595                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2596                         handle_read_error(conf, r1_bio);
2597                 else
2598                         WARN_ON_ONCE(1);
2599
2600                 cond_resched();
2601                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2602                         md_check_recovery(mddev);
2603         }
2604         blk_finish_plug(&plug);
2605 }
2606
2607 static int init_resync(struct r1conf *conf)
2608 {
2609         int buffs;
2610
2611         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2612         BUG_ON(mempool_initialized(&conf->r1buf_pool));
2613
2614         return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2615                             r1buf_pool_free, conf->poolinfo);
2616 }
2617
2618 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2619 {
2620         struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2621         struct resync_pages *rps;
2622         struct bio *bio;
2623         int i;
2624
2625         for (i = conf->poolinfo->raid_disks; i--; ) {
2626                 bio = r1bio->bios[i];
2627                 rps = bio->bi_private;
2628                 bio_reset(bio, NULL, 0);
2629                 bio->bi_private = rps;
2630         }
2631         r1bio->master_bio = NULL;
2632         return r1bio;
2633 }
2634
2635 /*
2636  * perform a "sync" on one "block"
2637  *
2638  * We need to make sure that no normal I/O request - particularly write
2639  * requests - conflict with active sync requests.
2640  *
2641  * This is achieved by tracking pending requests and a 'barrier' concept
2642  * that can be installed to exclude normal IO requests.
2643  */
2644
2645 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2646                                    int *skipped)
2647 {
2648         struct r1conf *conf = mddev->private;
2649         struct r1bio *r1_bio;
2650         struct bio *bio;
2651         sector_t max_sector, nr_sectors;
2652         int disk = -1;
2653         int i;
2654         int wonly = -1;
2655         int write_targets = 0, read_targets = 0;
2656         sector_t sync_blocks;
2657         int still_degraded = 0;
2658         int good_sectors = RESYNC_SECTORS;
2659         int min_bad = 0; /* number of sectors that are bad in all devices */
2660         int idx = sector_to_idx(sector_nr);
2661         int page_idx = 0;
2662
2663         if (!mempool_initialized(&conf->r1buf_pool))
2664                 if (init_resync(conf))
2665                         return 0;
2666
2667         max_sector = mddev->dev_sectors;
2668         if (sector_nr >= max_sector) {
2669                 /* If we aborted, we need to abort the
2670                  * sync on the 'current' bitmap chunk (there will
2671                  * only be one in raid1 resync.
2672                  * We can find the current addess in mddev->curr_resync
2673                  */
2674                 if (mddev->curr_resync < max_sector) /* aborted */
2675                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2676                                            &sync_blocks, 1);
2677                 else /* completed sync */
2678                         conf->fullsync = 0;
2679
2680                 md_bitmap_close_sync(mddev->bitmap);
2681                 close_sync(conf);
2682
2683                 if (mddev_is_clustered(mddev)) {
2684                         conf->cluster_sync_low = 0;
2685                         conf->cluster_sync_high = 0;
2686                 }
2687                 return 0;
2688         }
2689
2690         if (mddev->bitmap == NULL &&
2691             mddev->recovery_cp == MaxSector &&
2692             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2693             conf->fullsync == 0) {
2694                 *skipped = 1;
2695                 return max_sector - sector_nr;
2696         }
2697         /* before building a request, check if we can skip these blocks..
2698          * This call the bitmap_start_sync doesn't actually record anything
2699          */
2700         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2701             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2702                 /* We can skip this block, and probably several more */
2703                 *skipped = 1;
2704                 return sync_blocks;
2705         }
2706
2707         /*
2708          * If there is non-resync activity waiting for a turn, then let it
2709          * though before starting on this new sync request.
2710          */
2711         if (atomic_read(&conf->nr_waiting[idx]))
2712                 schedule_timeout_uninterruptible(1);
2713
2714         /* we are incrementing sector_nr below. To be safe, we check against
2715          * sector_nr + two times RESYNC_SECTORS
2716          */
2717
2718         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2719                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2720
2721
2722         if (raise_barrier(conf, sector_nr))
2723                 return 0;
2724
2725         r1_bio = raid1_alloc_init_r1buf(conf);
2726
2727         /*
2728          * If we get a correctably read error during resync or recovery,
2729          * we might want to read from a different device.  So we
2730          * flag all drives that could conceivably be read from for READ,
2731          * and any others (which will be non-In_sync devices) for WRITE.
2732          * If a read fails, we try reading from something else for which READ
2733          * is OK.
2734          */
2735
2736         r1_bio->mddev = mddev;
2737         r1_bio->sector = sector_nr;
2738         r1_bio->state = 0;
2739         set_bit(R1BIO_IsSync, &r1_bio->state);
2740         /* make sure good_sectors won't go across barrier unit boundary */
2741         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2742
2743         for (i = 0; i < conf->raid_disks * 2; i++) {
2744                 struct md_rdev *rdev;
2745                 bio = r1_bio->bios[i];
2746
2747                 rdev = conf->mirrors[i].rdev;
2748                 if (rdev == NULL ||
2749                     test_bit(Faulty, &rdev->flags)) {
2750                         if (i < conf->raid_disks)
2751                                 still_degraded = 1;
2752                 } else if (!test_bit(In_sync, &rdev->flags)) {
2753                         bio->bi_opf = REQ_OP_WRITE;
2754                         bio->bi_end_io = end_sync_write;
2755                         write_targets ++;
2756                 } else {
2757                         /* may need to read from here */
2758                         sector_t first_bad = MaxSector;
2759                         int bad_sectors;
2760
2761                         if (is_badblock(rdev, sector_nr, good_sectors,
2762                                         &first_bad, &bad_sectors)) {
2763                                 if (first_bad > sector_nr)
2764                                         good_sectors = first_bad - sector_nr;
2765                                 else {
2766                                         bad_sectors -= (sector_nr - first_bad);
2767                                         if (min_bad == 0 ||
2768                                             min_bad > bad_sectors)
2769                                                 min_bad = bad_sectors;
2770                                 }
2771                         }
2772                         if (sector_nr < first_bad) {
2773                                 if (test_bit(WriteMostly, &rdev->flags)) {
2774                                         if (wonly < 0)
2775                                                 wonly = i;
2776                                 } else {
2777                                         if (disk < 0)
2778                                                 disk = i;
2779                                 }
2780                                 bio->bi_opf = REQ_OP_READ;
2781                                 bio->bi_end_io = end_sync_read;
2782                                 read_targets++;
2783                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2784                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2785                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2786                                 /*
2787                                  * The device is suitable for reading (InSync),
2788                                  * but has bad block(s) here. Let's try to correct them,
2789                                  * if we are doing resync or repair. Otherwise, leave
2790                                  * this device alone for this sync request.
2791                                  */
2792                                 bio->bi_opf = REQ_OP_WRITE;
2793                                 bio->bi_end_io = end_sync_write;
2794                                 write_targets++;
2795                         }
2796                 }
2797                 if (rdev && bio->bi_end_io) {
2798                         atomic_inc(&rdev->nr_pending);
2799                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2800                         bio_set_dev(bio, rdev->bdev);
2801                         if (test_bit(FailFast, &rdev->flags))
2802                                 bio->bi_opf |= MD_FAILFAST;
2803                 }
2804         }
2805         if (disk < 0)
2806                 disk = wonly;
2807         r1_bio->read_disk = disk;
2808
2809         if (read_targets == 0 && min_bad > 0) {
2810                 /* These sectors are bad on all InSync devices, so we
2811                  * need to mark them bad on all write targets
2812                  */
2813                 int ok = 1;
2814                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2815                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2816                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2817                                 ok = rdev_set_badblocks(rdev, sector_nr,
2818                                                         min_bad, 0
2819                                         ) && ok;
2820                         }
2821                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2822                 *skipped = 1;
2823                 put_buf(r1_bio);
2824
2825                 if (!ok) {
2826                         /* Cannot record the badblocks, so need to
2827                          * abort the resync.
2828                          * If there are multiple read targets, could just
2829                          * fail the really bad ones ???
2830                          */
2831                         conf->recovery_disabled = mddev->recovery_disabled;
2832                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2833                         return 0;
2834                 } else
2835                         return min_bad;
2836
2837         }
2838         if (min_bad > 0 && min_bad < good_sectors) {
2839                 /* only resync enough to reach the next bad->good
2840                  * transition */
2841                 good_sectors = min_bad;
2842         }
2843
2844         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2845                 /* extra read targets are also write targets */
2846                 write_targets += read_targets-1;
2847
2848         if (write_targets == 0 || read_targets == 0) {
2849                 /* There is nowhere to write, so all non-sync
2850                  * drives must be failed - so we are finished
2851                  */
2852                 sector_t rv;
2853                 if (min_bad > 0)
2854                         max_sector = sector_nr + min_bad;
2855                 rv = max_sector - sector_nr;
2856                 *skipped = 1;
2857                 put_buf(r1_bio);
2858                 return rv;
2859         }
2860
2861         if (max_sector > mddev->resync_max)
2862                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2863         if (max_sector > sector_nr + good_sectors)
2864                 max_sector = sector_nr + good_sectors;
2865         nr_sectors = 0;
2866         sync_blocks = 0;
2867         do {
2868                 struct page *page;
2869                 int len = PAGE_SIZE;
2870                 if (sector_nr + (len>>9) > max_sector)
2871                         len = (max_sector - sector_nr) << 9;
2872                 if (len == 0)
2873                         break;
2874                 if (sync_blocks == 0) {
2875                         if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2876                                                   &sync_blocks, still_degraded) &&
2877                             !conf->fullsync &&
2878                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2879                                 break;
2880                         if ((len >> 9) > sync_blocks)
2881                                 len = sync_blocks<<9;
2882                 }
2883
2884                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2885                         struct resync_pages *rp;
2886
2887                         bio = r1_bio->bios[i];
2888                         rp = get_resync_pages(bio);
2889                         if (bio->bi_end_io) {
2890                                 page = resync_fetch_page(rp, page_idx);
2891
2892                                 /*
2893                                  * won't fail because the vec table is big
2894                                  * enough to hold all these pages
2895                                  */
2896                                 __bio_add_page(bio, page, len, 0);
2897                         }
2898                 }
2899                 nr_sectors += len>>9;
2900                 sector_nr += len>>9;
2901                 sync_blocks -= (len>>9);
2902         } while (++page_idx < RESYNC_PAGES);
2903
2904         r1_bio->sectors = nr_sectors;
2905
2906         if (mddev_is_clustered(mddev) &&
2907                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2908                 conf->cluster_sync_low = mddev->curr_resync_completed;
2909                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2910                 /* Send resync message */
2911                 md_cluster_ops->resync_info_update(mddev,
2912                                 conf->cluster_sync_low,
2913                                 conf->cluster_sync_high);
2914         }
2915
2916         /* For a user-requested sync, we read all readable devices and do a
2917          * compare
2918          */
2919         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2920                 atomic_set(&r1_bio->remaining, read_targets);
2921                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2922                         bio = r1_bio->bios[i];
2923                         if (bio->bi_end_io == end_sync_read) {
2924                                 read_targets--;
2925                                 md_sync_acct_bio(bio, nr_sectors);
2926                                 if (read_targets == 1)
2927                                         bio->bi_opf &= ~MD_FAILFAST;
2928                                 submit_bio_noacct(bio);
2929                         }
2930                 }
2931         } else {
2932                 atomic_set(&r1_bio->remaining, 1);
2933                 bio = r1_bio->bios[r1_bio->read_disk];
2934                 md_sync_acct_bio(bio, nr_sectors);
2935                 if (read_targets == 1)
2936                         bio->bi_opf &= ~MD_FAILFAST;
2937                 submit_bio_noacct(bio);
2938         }
2939         return nr_sectors;
2940 }
2941
2942 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2943 {
2944         if (sectors)
2945                 return sectors;
2946
2947         return mddev->dev_sectors;
2948 }
2949
2950 static struct r1conf *setup_conf(struct mddev *mddev)
2951 {
2952         struct r1conf *conf;
2953         int i;
2954         struct raid1_info *disk;
2955         struct md_rdev *rdev;
2956         int err = -ENOMEM;
2957
2958         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2959         if (!conf)
2960                 goto abort;
2961
2962         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2963                                    sizeof(atomic_t), GFP_KERNEL);
2964         if (!conf->nr_pending)
2965                 goto abort;
2966
2967         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2968                                    sizeof(atomic_t), GFP_KERNEL);
2969         if (!conf->nr_waiting)
2970                 goto abort;
2971
2972         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2973                                   sizeof(atomic_t), GFP_KERNEL);
2974         if (!conf->nr_queued)
2975                 goto abort;
2976
2977         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2978                                 sizeof(atomic_t), GFP_KERNEL);
2979         if (!conf->barrier)
2980                 goto abort;
2981
2982         conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2983                                             mddev->raid_disks, 2),
2984                                 GFP_KERNEL);
2985         if (!conf->mirrors)
2986                 goto abort;
2987
2988         conf->tmppage = alloc_page(GFP_KERNEL);
2989         if (!conf->tmppage)
2990                 goto abort;
2991
2992         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2993         if (!conf->poolinfo)
2994                 goto abort;
2995         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2996         err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2997                            rbio_pool_free, conf->poolinfo);
2998         if (err)
2999                 goto abort;
3000
3001         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3002         if (err)
3003                 goto abort;
3004
3005         conf->poolinfo->mddev = mddev;
3006
3007         err = -EINVAL;
3008         spin_lock_init(&conf->device_lock);
3009         rdev_for_each(rdev, mddev) {
3010                 int disk_idx = rdev->raid_disk;
3011                 if (disk_idx >= mddev->raid_disks
3012                     || disk_idx < 0)
3013                         continue;
3014                 if (test_bit(Replacement, &rdev->flags))
3015                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
3016                 else
3017                         disk = conf->mirrors + disk_idx;
3018
3019                 if (disk->rdev)
3020                         goto abort;
3021                 disk->rdev = rdev;
3022                 disk->head_position = 0;
3023                 disk->seq_start = MaxSector;
3024         }
3025         conf->raid_disks = mddev->raid_disks;
3026         conf->mddev = mddev;
3027         INIT_LIST_HEAD(&conf->retry_list);
3028         INIT_LIST_HEAD(&conf->bio_end_io_list);
3029
3030         spin_lock_init(&conf->resync_lock);
3031         init_waitqueue_head(&conf->wait_barrier);
3032
3033         bio_list_init(&conf->pending_bio_list);
3034         conf->recovery_disabled = mddev->recovery_disabled - 1;
3035
3036         err = -EIO;
3037         for (i = 0; i < conf->raid_disks * 2; i++) {
3038
3039                 disk = conf->mirrors + i;
3040
3041                 if (i < conf->raid_disks &&
3042                     disk[conf->raid_disks].rdev) {
3043                         /* This slot has a replacement. */
3044                         if (!disk->rdev) {
3045                                 /* No original, just make the replacement
3046                                  * a recovering spare
3047                                  */
3048                                 disk->rdev =
3049                                         disk[conf->raid_disks].rdev;
3050                                 disk[conf->raid_disks].rdev = NULL;
3051                         } else if (!test_bit(In_sync, &disk->rdev->flags))
3052                                 /* Original is not in_sync - bad */
3053                                 goto abort;
3054                 }
3055
3056                 if (!disk->rdev ||
3057                     !test_bit(In_sync, &disk->rdev->flags)) {
3058                         disk->head_position = 0;
3059                         if (disk->rdev &&
3060                             (disk->rdev->saved_raid_disk < 0))
3061                                 conf->fullsync = 1;
3062                 }
3063         }
3064
3065         err = -ENOMEM;
3066         rcu_assign_pointer(conf->thread,
3067                            md_register_thread(raid1d, mddev, "raid1"));
3068         if (!conf->thread)
3069                 goto abort;
3070
3071         return conf;
3072
3073  abort:
3074         if (conf) {
3075                 mempool_exit(&conf->r1bio_pool);
3076                 kfree(conf->mirrors);
3077                 safe_put_page(conf->tmppage);
3078                 kfree(conf->poolinfo);
3079                 kfree(conf->nr_pending);
3080                 kfree(conf->nr_waiting);
3081                 kfree(conf->nr_queued);
3082                 kfree(conf->barrier);
3083                 bioset_exit(&conf->bio_split);
3084                 kfree(conf);
3085         }
3086         return ERR_PTR(err);
3087 }
3088
3089 static void raid1_free(struct mddev *mddev, void *priv);
3090 static int raid1_run(struct mddev *mddev)
3091 {
3092         struct r1conf *conf;
3093         int i;
3094         struct md_rdev *rdev;
3095         int ret;
3096
3097         if (mddev->level != 1) {
3098                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3099                         mdname(mddev), mddev->level);
3100                 return -EIO;
3101         }
3102         if (mddev->reshape_position != MaxSector) {
3103                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3104                         mdname(mddev));
3105                 return -EIO;
3106         }
3107
3108         /*
3109          * copy the already verified devices into our private RAID1
3110          * bookkeeping area. [whatever we allocate in run(),
3111          * should be freed in raid1_free()]
3112          */
3113         if (mddev->private == NULL)
3114                 conf = setup_conf(mddev);
3115         else
3116                 conf = mddev->private;
3117
3118         if (IS_ERR(conf))
3119                 return PTR_ERR(conf);
3120
3121         if (mddev->queue)
3122                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3123
3124         rdev_for_each(rdev, mddev) {
3125                 if (!mddev->gendisk)
3126                         continue;
3127                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3128                                   rdev->data_offset << 9);
3129         }
3130
3131         mddev->degraded = 0;
3132         for (i = 0; i < conf->raid_disks; i++)
3133                 if (conf->mirrors[i].rdev == NULL ||
3134                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3135                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3136                         mddev->degraded++;
3137         /*
3138          * RAID1 needs at least one disk in active
3139          */
3140         if (conf->raid_disks - mddev->degraded < 1) {
3141                 md_unregister_thread(mddev, &conf->thread);
3142                 ret = -EINVAL;
3143                 goto abort;
3144         }
3145
3146         if (conf->raid_disks - mddev->degraded == 1)
3147                 mddev->recovery_cp = MaxSector;
3148
3149         if (mddev->recovery_cp != MaxSector)
3150                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3151                         mdname(mddev));
3152         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3153                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3154                 mddev->raid_disks);
3155
3156         /*
3157          * Ok, everything is just fine now
3158          */
3159         rcu_assign_pointer(mddev->thread, conf->thread);
3160         rcu_assign_pointer(conf->thread, NULL);
3161         mddev->private = conf;
3162         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3163
3164         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3165
3166         ret = md_integrity_register(mddev);
3167         if (ret) {
3168                 md_unregister_thread(mddev, &mddev->thread);
3169                 goto abort;
3170         }
3171         return 0;
3172
3173 abort:
3174         raid1_free(mddev, conf);
3175         return ret;
3176 }
3177
3178 static void raid1_free(struct mddev *mddev, void *priv)
3179 {
3180         struct r1conf *conf = priv;
3181
3182         mempool_exit(&conf->r1bio_pool);
3183         kfree(conf->mirrors);
3184         safe_put_page(conf->tmppage);
3185         kfree(conf->poolinfo);
3186         kfree(conf->nr_pending);
3187         kfree(conf->nr_waiting);
3188         kfree(conf->nr_queued);
3189         kfree(conf->barrier);
3190         bioset_exit(&conf->bio_split);
3191         kfree(conf);
3192 }
3193
3194 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3195 {
3196         /* no resync is happening, and there is enough space
3197          * on all devices, so we can resize.
3198          * We need to make sure resync covers any new space.
3199          * If the array is shrinking we should possibly wait until
3200          * any io in the removed space completes, but it hardly seems
3201          * worth it.
3202          */
3203         sector_t newsize = raid1_size(mddev, sectors, 0);
3204         if (mddev->external_size &&
3205             mddev->array_sectors > newsize)
3206                 return -EINVAL;
3207         if (mddev->bitmap) {
3208                 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3209                 if (ret)
3210                         return ret;
3211         }
3212         md_set_array_sectors(mddev, newsize);
3213         if (sectors > mddev->dev_sectors &&
3214             mddev->recovery_cp > mddev->dev_sectors) {
3215                 mddev->recovery_cp = mddev->dev_sectors;
3216                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3217         }
3218         mddev->dev_sectors = sectors;
3219         mddev->resync_max_sectors = sectors;
3220         return 0;
3221 }
3222
3223 static int raid1_reshape(struct mddev *mddev)
3224 {
3225         /* We need to:
3226          * 1/ resize the r1bio_pool
3227          * 2/ resize conf->mirrors
3228          *
3229          * We allocate a new r1bio_pool if we can.
3230          * Then raise a device barrier and wait until all IO stops.
3231          * Then resize conf->mirrors and swap in the new r1bio pool.
3232          *
3233          * At the same time, we "pack" the devices so that all the missing
3234          * devices have the higher raid_disk numbers.
3235          */
3236         mempool_t newpool, oldpool;
3237         struct pool_info *newpoolinfo;
3238         struct raid1_info *newmirrors;
3239         struct r1conf *conf = mddev->private;
3240         int cnt, raid_disks;
3241         unsigned long flags;
3242         int d, d2;
3243         int ret;
3244
3245         memset(&newpool, 0, sizeof(newpool));
3246         memset(&oldpool, 0, sizeof(oldpool));
3247
3248         /* Cannot change chunk_size, layout, or level */
3249         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3250             mddev->layout != mddev->new_layout ||
3251             mddev->level != mddev->new_level) {
3252                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3253                 mddev->new_layout = mddev->layout;
3254                 mddev->new_level = mddev->level;
3255                 return -EINVAL;
3256         }
3257
3258         if (!mddev_is_clustered(mddev))
3259                 md_allow_write(mddev);
3260
3261         raid_disks = mddev->raid_disks + mddev->delta_disks;
3262
3263         if (raid_disks < conf->raid_disks) {
3264                 cnt=0;
3265                 for (d= 0; d < conf->raid_disks; d++)
3266                         if (conf->mirrors[d].rdev)
3267                                 cnt++;
3268                 if (cnt > raid_disks)
3269                         return -EBUSY;
3270         }
3271
3272         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3273         if (!newpoolinfo)
3274                 return -ENOMEM;
3275         newpoolinfo->mddev = mddev;
3276         newpoolinfo->raid_disks = raid_disks * 2;
3277
3278         ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3279                            rbio_pool_free, newpoolinfo);
3280         if (ret) {
3281                 kfree(newpoolinfo);
3282                 return ret;
3283         }
3284         newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3285                                          raid_disks, 2),
3286                              GFP_KERNEL);
3287         if (!newmirrors) {
3288                 kfree(newpoolinfo);
3289                 mempool_exit(&newpool);
3290                 return -ENOMEM;
3291         }
3292
3293         freeze_array(conf, 0);
3294
3295         /* ok, everything is stopped */
3296         oldpool = conf->r1bio_pool;
3297         conf->r1bio_pool = newpool;
3298
3299         for (d = d2 = 0; d < conf->raid_disks; d++) {
3300                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3301                 if (rdev && rdev->raid_disk != d2) {
3302                         sysfs_unlink_rdev(mddev, rdev);
3303                         rdev->raid_disk = d2;
3304                         sysfs_unlink_rdev(mddev, rdev);
3305                         if (sysfs_link_rdev(mddev, rdev))
3306                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3307                                         mdname(mddev), rdev->raid_disk);
3308                 }
3309                 if (rdev)
3310                         newmirrors[d2++].rdev = rdev;
3311         }
3312         kfree(conf->mirrors);
3313         conf->mirrors = newmirrors;
3314         kfree(conf->poolinfo);
3315         conf->poolinfo = newpoolinfo;
3316
3317         spin_lock_irqsave(&conf->device_lock, flags);
3318         mddev->degraded += (raid_disks - conf->raid_disks);
3319         spin_unlock_irqrestore(&conf->device_lock, flags);
3320         conf->raid_disks = mddev->raid_disks = raid_disks;
3321         mddev->delta_disks = 0;
3322
3323         unfreeze_array(conf);
3324
3325         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3326         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3327         md_wakeup_thread(mddev->thread);
3328
3329         mempool_exit(&oldpool);
3330         return 0;
3331 }
3332
3333 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3334 {
3335         struct r1conf *conf = mddev->private;
3336
3337         if (quiesce)
3338                 freeze_array(conf, 0);
3339         else
3340                 unfreeze_array(conf);
3341 }
3342
3343 static void *raid1_takeover(struct mddev *mddev)
3344 {
3345         /* raid1 can take over:
3346          *  raid5 with 2 devices, any layout or chunk size
3347          */
3348         if (mddev->level == 5 && mddev->raid_disks == 2) {
3349                 struct r1conf *conf;
3350                 mddev->new_level = 1;
3351                 mddev->new_layout = 0;
3352                 mddev->new_chunk_sectors = 0;
3353                 conf = setup_conf(mddev);
3354                 if (!IS_ERR(conf)) {
3355                         /* Array must appear to be quiesced */
3356                         conf->array_frozen = 1;
3357                         mddev_clear_unsupported_flags(mddev,
3358                                 UNSUPPORTED_MDDEV_FLAGS);
3359                 }
3360                 return conf;
3361         }
3362         return ERR_PTR(-EINVAL);
3363 }
3364
3365 static struct md_personality raid1_personality =
3366 {
3367         .name           = "raid1",
3368         .level          = 1,
3369         .owner          = THIS_MODULE,
3370         .make_request   = raid1_make_request,
3371         .run            = raid1_run,
3372         .free           = raid1_free,
3373         .status         = raid1_status,
3374         .error_handler  = raid1_error,
3375         .hot_add_disk   = raid1_add_disk,
3376         .hot_remove_disk= raid1_remove_disk,
3377         .spare_active   = raid1_spare_active,
3378         .sync_request   = raid1_sync_request,
3379         .resize         = raid1_resize,
3380         .size           = raid1_size,
3381         .check_reshape  = raid1_reshape,
3382         .quiesce        = raid1_quiesce,
3383         .takeover       = raid1_takeover,
3384 };
3385
3386 static int __init raid_init(void)
3387 {
3388         return register_md_personality(&raid1_personality);
3389 }
3390
3391 static void raid_exit(void)
3392 {
3393         unregister_md_personality(&raid1_personality);
3394 }
3395
3396 module_init(raid_init);
3397 module_exit(raid_exit);
3398 MODULE_LICENSE("GPL");
3399 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3400 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3401 MODULE_ALIAS("md-raid1");
3402 MODULE_ALIAS("md-level-1");