drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/slab.h>
  35 #include <linux/delay.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/module.h>
  38 #include <linux/seq_file.h>
  39 #include <linux/ratelimit.h>
  40 #include <trace/events/block.h>
  41 #include "md.h"
  42 #include "raid1.h"
  43 #include "bitmap.h"
  44
  45 #define UNSUPPORTED_MDDEV_FLAGS         \
  46         ((1L << MD_HAS_JOURNAL) |       \
  47          (1L << MD_JOURNAL_CLEAN))
  48
  49 /*
  50  * Number of guaranteed r1bios in case of extreme VM load:
  51  */
  52 #define NR_RAID1_BIOS 256
  53
  54 /* when we get a read error on a read-only array, we redirect to another
  55  * device without failing the first device, or trying to over-write to
  56  * correct the read error.  To keep track of bad blocks on a per-bio
  57  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  58  */
  59 #define IO_BLOCKED ((struct bio *)1)
  60 /* When we successfully write to a known bad-block, we need to remove the
  61  * bad-block marking which must be done from process context.  So we record
  62  * the success by setting devs[n].bio to IO_MADE_GOOD
  63  */
  64 #define IO_MADE_GOOD ((struct bio *)2)
  65
  66 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  67
  68 /* When there are this many requests queue to be written by
  69  * the raid1 thread, we become 'congested' to provide back-pressure
  70  * for writeback.
  71  */
  72 static int max_queued_requests = 1024;
  73
  74 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
  75                           sector_t bi_sector);
  76 static void lower_barrier(struct r1conf *conf);
  77
  78 #define raid1_log(md, fmt, args...)                             \
  79         do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
  80
  81 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  82 {
  83         struct pool_info *pi = data;
  84         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  85
  86         /* allocate a r1bio with room for raid_disks entries in the bios array */
  87         return kzalloc(size, gfp_flags);
  88 }
  89
  90 static void r1bio_pool_free(void *r1_bio, void *data)
  91 {
  92         kfree(r1_bio);
  93 }
  94
  95 #define RESYNC_BLOCK_SIZE (64*1024)
  96 #define RESYNC_DEPTH 32
  97 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  98 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  99 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
 100 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
 101 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
 102 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
 103 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
 104
 105 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
 106 {
 107         struct pool_info *pi = data;
 108         struct r1bio *r1_bio;
 109         struct bio *bio;
 110         int need_pages;
 111         int i, j;
 112
 113         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 114         if (!r1_bio)
 115                 return NULL;
 116
 117         /*
 118          * Allocate bios : 1 for reading, n-1 for writing
 119          */
 120         for (j = pi->raid_disks ; j-- ; ) {
 121                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 122                 if (!bio)
 123                         goto out_free_bio;
 124                 r1_bio->bios[j] = bio;
 125         }
 126         /*
 127          * Allocate RESYNC_PAGES data pages and attach them to
 128          * the first bio.
 129          * If this is a user-requested check/repair, allocate
 130          * RESYNC_PAGES for each bio.
 131          */
 132         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 133                 need_pages = pi->raid_disks;
 134         else
 135                 need_pages = 1;
 136         for (j = 0; j < need_pages; j++) {
 137                 bio = r1_bio->bios[j];
 138                 bio->bi_vcnt = RESYNC_PAGES;
 139
 140                 if (bio_alloc_pages(bio, gfp_flags))
 141                         goto out_free_pages;
 142         }
 143         /* If not user-requests, copy the page pointers to all bios */
 144         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 145                 for (i=0; i<RESYNC_PAGES ; i++)
 146                         for (j=1; j<pi->raid_disks; j++)
 147                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 148                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 149         }
 150
 151         r1_bio->master_bio = NULL;
 152
 153         return r1_bio;
 154
 155 out_free_pages:
 156         while (--j >= 0)
 157                 bio_free_pages(r1_bio->bios[j]);
 158
 159 out_free_bio:
 160         while (++j < pi->raid_disks)
 161                 bio_put(r1_bio->bios[j]);
 162         r1bio_pool_free(r1_bio, data);
 163         return NULL;
 164 }
 165
 166 static void r1buf_pool_free(void *__r1_bio, void *data)
 167 {
 168         struct pool_info *pi = data;
 169         int i,j;
 170         struct r1bio *r1bio = __r1_bio;
 171
 172         for (i = 0; i < RESYNC_PAGES; i++)
 173                 for (j = pi->raid_disks; j-- ;) {
 174                         if (j == 0 ||
 175                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 176                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 177                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 178                 }
 179         for (i=0 ; i < pi->raid_disks; i++)
 180                 bio_put(r1bio->bios[i]);
 181
 182         r1bio_pool_free(r1bio, data);
 183 }
 184
 185 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 186 {
 187         int i;
 188
 189         for (i = 0; i < conf->raid_disks * 2; i++) {
 190                 struct bio **bio = r1_bio->bios + i;
 191                 if (!BIO_SPECIAL(*bio))
 192                         bio_put(*bio);
 193                 *bio = NULL;
 194         }
 195 }
 196
 197 static void free_r1bio(struct r1bio *r1_bio)
 198 {
 199         struct r1conf *conf = r1_bio->mddev->private;
 200
 201         put_all_bios(conf, r1_bio);
 202         mempool_free(r1_bio, conf->r1bio_pool);
 203 }
 204
 205 static void put_buf(struct r1bio *r1_bio)
 206 {
 207         struct r1conf *conf = r1_bio->mddev->private;
 208         int i;
 209
 210         for (i = 0; i < conf->raid_disks * 2; i++) {
 211                 struct bio *bio = r1_bio->bios[i];
 212                 if (bio->bi_end_io)
 213                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 214         }
 215
 216         mempool_free(r1_bio, conf->r1buf_pool);
 217
 218         lower_barrier(conf);
 219 }
 220
 221 static void reschedule_retry(struct r1bio *r1_bio)
 222 {
 223         unsigned long flags;
 224         struct mddev *mddev = r1_bio->mddev;
 225         struct r1conf *conf = mddev->private;
 226
 227         spin_lock_irqsave(&conf->device_lock, flags);
 228         list_add(&r1_bio->retry_list, &conf->retry_list);
 229         conf->nr_queued ++;
 230         spin_unlock_irqrestore(&conf->device_lock, flags);
 231
 232         wake_up(&conf->wait_barrier);
 233         md_wakeup_thread(mddev->thread);
 234 }
 235
 236 /*
 237  * raid_end_bio_io() is called when we have finished servicing a mirrored
 238  * operation and are ready to return a success/failure code to the buffer
 239  * cache layer.
 240  */
 241 static void call_bio_endio(struct r1bio *r1_bio)
 242 {
 243         struct bio *bio = r1_bio->master_bio;
 244         int done;
 245         struct r1conf *conf = r1_bio->mddev->private;
 246         sector_t start_next_window = r1_bio->start_next_window;
 247         sector_t bi_sector = bio->bi_iter.bi_sector;
 248
 249         if (bio->bi_phys_segments) {
 250                 unsigned long flags;
 251                 spin_lock_irqsave(&conf->device_lock, flags);
 252                 bio->bi_phys_segments--;
 253                 done = (bio->bi_phys_segments == 0);
 254                 spin_unlock_irqrestore(&conf->device_lock, flags);
 255                 /*
 256                  * make_request() might be waiting for
 257                  * bi_phys_segments to decrease
 258                  */
 259                 wake_up(&conf->wait_barrier);
 260         } else
 261                 done = 1;
 262
 263         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 264                 bio->bi_error = -EIO;
 265
 266         if (done) {
 267                 bio_endio(bio);
 268                 /*
 269                  * Wake up any possible resync thread that waits for the device
 270                  * to go idle.
 271                  */
 272                 allow_barrier(conf, start_next_window, bi_sector);
 273         }
 274 }
 275
 276 static void raid_end_bio_io(struct r1bio *r1_bio)
 277 {
 278         struct bio *bio = r1_bio->master_bio;
 279
 280         /* if nobody has done the final endio yet, do it now */
 281         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 282                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 283                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 284                          (unsigned long long) bio->bi_iter.bi_sector,
 285                          (unsigned long long) bio_end_sector(bio) - 1);
 286
 287                 call_bio_endio(r1_bio);
 288         }
 289         free_r1bio(r1_bio);
 290 }
 291
 292 /*
 293  * Update disk head position estimator based on IRQ completion info.
 294  */
 295 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 296 {
 297         struct r1conf *conf = r1_bio->mddev->private;
 298
 299         conf->mirrors[disk].head_position =
 300                 r1_bio->sector + (r1_bio->sectors);
 301 }
 302
 303 /*
 304  * Find the disk number which triggered given bio
 305  */
 306 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 307 {
 308         int mirror;
 309         struct r1conf *conf = r1_bio->mddev->private;
 310         int raid_disks = conf->raid_disks;
 311
 312         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 313                 if (r1_bio->bios[mirror] == bio)
 314                         break;
 315
 316         BUG_ON(mirror == raid_disks * 2);
 317         update_head_pos(mirror, r1_bio);
 318
 319         return mirror;
 320 }
 321
 322 static void raid1_end_read_request(struct bio *bio)
 323 {
 324         int uptodate = !bio->bi_error;
 325         struct r1bio *r1_bio = bio->bi_private;
 326         struct r1conf *conf = r1_bio->mddev->private;
 327         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
 328
 329         /*
 330          * this branch is our 'one mirror IO has finished' event handler:
 331          */
 332         update_head_pos(r1_bio->read_disk, r1_bio);
 333
 334         if (uptodate)
 335                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 336         else if (test_bit(FailFast, &rdev->flags) &&
 337                  test_bit(R1BIO_FailFast, &r1_bio->state))
 338                 /* This was a fail-fast read so we definitely
 339                  * want to retry */
 340                 ;
 341         else {
 342                 /* If all other devices have failed, we want to return
 343                  * the error upwards rather than fail the last device.
 344                  * Here we redefine "uptodate" to mean "Don't want to retry"
 345                  */
 346                 unsigned long flags;
 347                 spin_lock_irqsave(&conf->device_lock, flags);
 348                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 349                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 350                      test_bit(In_sync, &rdev->flags)))
 351                         uptodate = 1;
 352                 spin_unlock_irqrestore(&conf->device_lock, flags);
 353         }
 354
 355         if (uptodate) {
 356                 raid_end_bio_io(r1_bio);
 357                 rdev_dec_pending(rdev, conf->mddev);
 358         } else {
 359                 /*
 360                  * oops, read error:
 361                  */
 362                 char b[BDEVNAME_SIZE];
 363                 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
 364                                    mdname(conf->mddev),
 365                                    bdevname(rdev->bdev, b),
 366                                    (unsigned long long)r1_bio->sector);
 367                 set_bit(R1BIO_ReadError, &r1_bio->state);
 368                 reschedule_retry(r1_bio);
 369                 /* don't drop the reference on read_disk yet */
 370         }
 371 }
 372
 373 static void close_write(struct r1bio *r1_bio)
 374 {
 375         /* it really is the end of this request */
 376         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 377                 /* free extra copy of the data pages */
 378                 int i = r1_bio->behind_page_count;
 379                 while (i--)
 380                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 381                 kfree(r1_bio->behind_bvecs);
 382                 r1_bio->behind_bvecs = NULL;
 383         }
 384         /* clear the bitmap if all writes complete successfully */
 385         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 386                         r1_bio->sectors,
 387                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 388                         test_bit(R1BIO_BehindIO, &r1_bio->state));
 389         md_write_end(r1_bio->mddev);
 390 }
 391
 392 static void r1_bio_write_done(struct r1bio *r1_bio)
 393 {
 394         if (!atomic_dec_and_test(&r1_bio->remaining))
 395                 return;
 396
 397         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 398                 reschedule_retry(r1_bio);
 399         else {
 400                 close_write(r1_bio);
 401                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 402                         reschedule_retry(r1_bio);
 403                 else
 404                         raid_end_bio_io(r1_bio);
 405         }
 406 }
 407
 408 static void raid1_end_write_request(struct bio *bio)
 409 {
 410         struct r1bio *r1_bio = bio->bi_private;
 411         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 412         struct r1conf *conf = r1_bio->mddev->private;
 413         struct bio *to_put = NULL;
 414         int mirror = find_bio_disk(r1_bio, bio);
 415         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
 416         bool discard_error;
 417
 418         discard_error = bio->bi_error && bio_op(bio) == REQ_OP_DISCARD;
 419
 420         /*
 421          * 'one mirror IO has finished' event handler:
 422          */
 423         if (bio->bi_error && !discard_error) {
 424                 set_bit(WriteErrorSeen, &rdev->flags);
 425                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
 426                         set_bit(MD_RECOVERY_NEEDED, &
 427                                 conf->mddev->recovery);
 428
 429                 if (test_bit(FailFast, &rdev->flags) &&
 430                     (bio->bi_opf & MD_FAILFAST) &&
 431                     /* We never try FailFast to WriteMostly devices */
 432                     !test_bit(WriteMostly, &rdev->flags)) {
 433                         md_error(r1_bio->mddev, rdev);
 434                         if (!test_bit(Faulty, &rdev->flags))
 435                                 /* This is the only remaining device,
 436                                  * We need to retry the write without
 437                                  * FailFast
 438                                  */
 439                                 set_bit(R1BIO_WriteError, &r1_bio->state);
 440                         else {
 441                                 /* Finished with this branch */
 442                                 r1_bio->bios[mirror] = NULL;
 443                                 to_put = bio;
 444                         }
 445                 } else
 446                         set_bit(R1BIO_WriteError, &r1_bio->state);
 447         } else {
 448                 /*
 449                  * Set R1BIO_Uptodate in our master bio, so that we
 450                  * will return a good error code for to the higher
 451                  * levels even if IO on some other mirrored buffer
 452                  * fails.
 453                  *
 454                  * The 'master' represents the composite IO operation
 455                  * to user-side. So if something waits for IO, then it
 456                  * will wait for the 'master' bio.
 457                  */
 458                 sector_t first_bad;
 459                 int bad_sectors;
 460
 461                 r1_bio->bios[mirror] = NULL;
 462                 to_put = bio;
 463                 /*
 464                  * Do not set R1BIO_Uptodate if the current device is
 465                  * rebuilding or Faulty. This is because we cannot use
 466                  * such device for properly reading the data back (we could
 467                  * potentially use it, if the current write would have felt
 468                  * before rdev->recovery_offset, but for simplicity we don't
 469                  * check this here.
 470                  */
 471                 if (test_bit(In_sync, &rdev->flags) &&
 472                     !test_bit(Faulty, &rdev->flags))
 473                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 474
 475                 /* Maybe we can clear some bad blocks. */
 476                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
 477                                 &first_bad, &bad_sectors) && !discard_error) {
 478                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 479                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 480                 }
 481         }
 482
 483         if (behind) {
 484                 if (test_bit(WriteMostly, &rdev->flags))
 485                         atomic_dec(&r1_bio->behind_remaining);
 486
 487                 /*
 488                  * In behind mode, we ACK the master bio once the I/O
 489                  * has safely reached all non-writemostly
 490                  * disks. Setting the Returned bit ensures that this
 491                  * gets done only once -- we don't ever want to return
 492                  * -EIO here, instead we'll wait
 493                  */
 494                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 495                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 496                         /* Maybe we can return now */
 497                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 498                                 struct bio *mbio = r1_bio->master_bio;
 499                                 pr_debug("raid1: behind end write sectors"
 500                                          " %llu-%llu\n",
 501                                          (unsigned long long) mbio->bi_iter.bi_sector,
 502                                          (unsigned long long) bio_end_sector(mbio) - 1);
 503                                 call_bio_endio(r1_bio);
 504                         }
 505                 }
 506         }
 507         if (r1_bio->bios[mirror] == NULL)
 508                 rdev_dec_pending(rdev, conf->mddev);
 509
 510         /*
 511          * Let's see if all mirrored write operations have finished
 512          * already.
 513          */
 514         r1_bio_write_done(r1_bio);
 515
 516         if (to_put)
 517                 bio_put(to_put);
 518 }
 519
 520 /*
 521  * This routine returns the disk from which the requested read should
 522  * be done. There is a per-array 'next expected sequential IO' sector
 523  * number - if this matches on the next IO then we use the last disk.
 524  * There is also a per-disk 'last know head position' sector that is
 525  * maintained from IRQ contexts, both the normal and the resync IO
 526  * completion handlers update this position correctly. If there is no
 527  * perfect sequential match then we pick the disk whose head is closest.
 528  *
 529  * If there are 2 mirrors in the same 2 devices, performance degrades
 530  * because position is mirror, not device based.
 531  *
 532  * The rdev for the device selected will have nr_pending incremented.
 533  */
 534 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 535 {
 536         const sector_t this_sector = r1_bio->sector;
 537         int sectors;
 538         int best_good_sectors;
 539         int best_disk, best_dist_disk, best_pending_disk;
 540         int has_nonrot_disk;
 541         int disk;
 542         sector_t best_dist;
 543         unsigned int min_pending;
 544         struct md_rdev *rdev;
 545         int choose_first;
 546         int choose_next_idle;
 547
 548         rcu_read_lock();
 549         /*
 550          * Check if we can balance. We can balance on the whole
 551          * device if no resync is going on, or below the resync window.
 552          * We take the first readable disk when above the resync window.
 553          */
 554  retry:
 555         sectors = r1_bio->sectors;
 556         best_disk = -1;
 557         best_dist_disk = -1;
 558         best_dist = MaxSector;
 559         best_pending_disk = -1;
 560         min_pending = UINT_MAX;
 561         best_good_sectors = 0;
 562         has_nonrot_disk = 0;
 563         choose_next_idle = 0;
 564         clear_bit(R1BIO_FailFast, &r1_bio->state);
 565
 566         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 567             (mddev_is_clustered(conf->mddev) &&
 568             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
 569                     this_sector + sectors)))
 570                 choose_first = 1;
 571         else
 572                 choose_first = 0;
 573
 574         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 575                 sector_t dist;
 576                 sector_t first_bad;
 577                 int bad_sectors;
 578                 unsigned int pending;
 579                 bool nonrot;
 580
 581                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 582                 if (r1_bio->bios[disk] == IO_BLOCKED
 583                     || rdev == NULL
 584                     || test_bit(Faulty, &rdev->flags))
 585                         continue;
 586                 if (!test_bit(In_sync, &rdev->flags) &&
 587                     rdev->recovery_offset < this_sector + sectors)
 588                         continue;
 589                 if (test_bit(WriteMostly, &rdev->flags)) {
 590                         /* Don't balance among write-mostly, just
 591                          * use the first as a last resort */
 592                         if (best_dist_disk < 0) {
 593                                 if (is_badblock(rdev, this_sector, sectors,
 594                                                 &first_bad, &bad_sectors)) {
 595                                         if (first_bad <= this_sector)
 596                                                 /* Cannot use this */
 597                                                 continue;
 598                                         best_good_sectors = first_bad - this_sector;
 599                                 } else
 600                                         best_good_sectors = sectors;
 601                                 best_dist_disk = disk;
 602                                 best_pending_disk = disk;
 603                         }
 604                         continue;
 605                 }
 606                 /* This is a reasonable device to use.  It might
 607                  * even be best.
 608                  */
 609                 if (is_badblock(rdev, this_sector, sectors,
 610                                 &first_bad, &bad_sectors)) {
 611                         if (best_dist < MaxSector)
 612                                 /* already have a better device */
 613                                 continue;
 614                         if (first_bad <= this_sector) {
 615                                 /* cannot read here. If this is the 'primary'
 616                                  * device, then we must not read beyond
 617                                  * bad_sectors from another device..
 618                                  */
 619                                 bad_sectors -= (this_sector - first_bad);
 620                                 if (choose_first && sectors > bad_sectors)
 621                                         sectors = bad_sectors;
 622                                 if (best_good_sectors > sectors)
 623                                         best_good_sectors = sectors;
 624
 625                         } else {
 626                                 sector_t good_sectors = first_bad - this_sector;
 627                                 if (good_sectors > best_good_sectors) {
 628                                         best_good_sectors = good_sectors;
 629                                         best_disk = disk;
 630                                 }
 631                                 if (choose_first)
 632                                         break;
 633                         }
 634                         continue;
 635                 } else
 636                         best_good_sectors = sectors;
 637
 638                 if (best_disk >= 0)
 639                         /* At least two disks to choose from so failfast is OK */
 640                         set_bit(R1BIO_FailFast, &r1_bio->state);
 641
 642                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 643                 has_nonrot_disk |= nonrot;
 644                 pending = atomic_read(&rdev->nr_pending);
 645                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 646                 if (choose_first) {
 647                         best_disk = disk;
 648                         break;
 649                 }
 650                 /* Don't change to another disk for sequential reads */
 651                 if (conf->mirrors[disk].next_seq_sect == this_sector
 652                     || dist == 0) {
 653                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 654                         struct raid1_info *mirror = &conf->mirrors[disk];
 655
 656                         best_disk = disk;
 657                         /*
 658                          * If buffered sequential IO size exceeds optimal
 659                          * iosize, check if there is idle disk. If yes, choose
 660                          * the idle disk. read_balance could already choose an
 661                          * idle disk before noticing it's a sequential IO in
 662                          * this disk. This doesn't matter because this disk
 663                          * will idle, next time it will be utilized after the
 664                          * first disk has IO size exceeds optimal iosize. In
 665                          * this way, iosize of the first disk will be optimal
 666                          * iosize at least. iosize of the second disk might be
 667                          * small, but not a big deal since when the second disk
 668                          * starts IO, the first disk is likely still busy.
 669                          */
 670                         if (nonrot && opt_iosize > 0 &&
 671                             mirror->seq_start != MaxSector &&
 672                             mirror->next_seq_sect > opt_iosize &&
 673                             mirror->next_seq_sect - opt_iosize >=
 674                             mirror->seq_start) {
 675                                 choose_next_idle = 1;
 676                                 continue;
 677                         }
 678                         break;
 679                 }
 680
 681                 if (choose_next_idle)
 682                         continue;
 683
 684                 if (min_pending > pending) {
 685                         min_pending = pending;
 686                         best_pending_disk = disk;
 687                 }
 688
 689                 if (dist < best_dist) {
 690                         best_dist = dist;
 691                         best_dist_disk = disk;
 692                 }
 693         }
 694
 695         /*
 696          * If all disks are rotational, choose the closest disk. If any disk is
 697          * non-rotational, choose the disk with less pending request even the
 698          * disk is rotational, which might/might not be optimal for raids with
 699          * mixed ratation/non-rotational disks depending on workload.
 700          */
 701         if (best_disk == -1) {
 702                 if (has_nonrot_disk || min_pending == 0)
 703                         best_disk = best_pending_disk;
 704                 else
 705                         best_disk = best_dist_disk;
 706         }
 707
 708         if (best_disk >= 0) {
 709                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 710                 if (!rdev)
 711                         goto retry;
 712                 atomic_inc(&rdev->nr_pending);
 713                 sectors = best_good_sectors;
 714
 715                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 716                         conf->mirrors[best_disk].seq_start = this_sector;
 717
 718                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 719         }
 720         rcu_read_unlock();
 721         *max_sectors = sectors;
 722
 723         return best_disk;
 724 }
 725
 726 static int raid1_congested(struct mddev *mddev, int bits)
 727 {
 728         struct r1conf *conf = mddev->private;
 729         int i, ret = 0;
 730
 731         if ((bits & (1 << WB_async_congested)) &&
 732             conf->pending_count >= max_queued_requests)
 733                 return 1;
 734
 735         rcu_read_lock();
 736         for (i = 0; i < conf->raid_disks * 2; i++) {
 737                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 738                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 739                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 740
 741                         BUG_ON(!q);
 742
 743                         /* Note the '|| 1' - when read_balance prefers
 744                          * non-congested targets, it can be removed
 745                          */
 746                         if ((bits & (1 << WB_async_congested)) || 1)
 747                                 ret |= bdi_congested(q->backing_dev_info, bits);
 748                         else
 749                                 ret &= bdi_congested(q->backing_dev_info, bits);
 750                 }
 751         }
 752         rcu_read_unlock();
 753         return ret;
 754 }
 755
 756 static void flush_pending_writes(struct r1conf *conf)
 757 {
 758         /* Any writes that have been queued but are awaiting
 759          * bitmap updates get flushed here.
 760          */
 761         spin_lock_irq(&conf->device_lock);
 762
 763         if (conf->pending_bio_list.head) {
 764                 struct bio *bio;
 765                 bio = bio_list_get(&conf->pending_bio_list);
 766                 conf->pending_count = 0;
 767                 spin_unlock_irq(&conf->device_lock);
 768                 /* flush any pending bitmap writes to
 769                  * disk before proceeding w/ I/O */
 770                 bitmap_unplug(conf->mddev->bitmap);
 771                 wake_up(&conf->wait_barrier);
 772
 773                 while (bio) { /* submit pending writes */
 774                         struct bio *next = bio->bi_next;
 775                         struct md_rdev *rdev = (void*)bio->bi_bdev;
 776                         bio->bi_next = NULL;
 777                         bio->bi_bdev = rdev->bdev;
 778                         if (test_bit(Faulty, &rdev->flags)) {
 779                                 bio->bi_error = -EIO;
 780                                 bio_endio(bio);
 781                         } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
 782                                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 783                                 /* Just ignore it */
 784                                 bio_endio(bio);
 785                         else
 786                                 generic_make_request(bio);
 787                         bio = next;
 788                 }
 789         } else
 790                 spin_unlock_irq(&conf->device_lock);
 791 }
 792
 793 /* Barriers....
 794  * Sometimes we need to suspend IO while we do something else,
 795  * either some resync/recovery, or reconfigure the array.
 796  * To do this we raise a 'barrier'.
 797  * The 'barrier' is a counter that can be raised multiple times
 798  * to count how many activities are happening which preclude
 799  * normal IO.
 800  * We can only raise the barrier if there is no pending IO.
 801  * i.e. if nr_pending == 0.
 802  * We choose only to raise the barrier if no-one is waiting for the
 803  * barrier to go down.  This means that as soon as an IO request
 804  * is ready, no other operations which require a barrier will start
 805  * until the IO request has had a chance.
 806  *
 807  * So: regular IO calls 'wait_barrier'.  When that returns there
 808  *    is no backgroup IO happening,  It must arrange to call
 809  *    allow_barrier when it has finished its IO.
 810  * backgroup IO calls must call raise_barrier.  Once that returns
 811  *    there is no normal IO happeing.  It must arrange to call
 812  *    lower_barrier when the particular background IO completes.
 813  */
 814 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
 815 {
 816         spin_lock_irq(&conf->resync_lock);
 817
 818         /* Wait until no block IO is waiting */
 819         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 820                             conf->resync_lock);
 821
 822         /* block any new IO from starting */
 823         conf->barrier++;
 824         conf->next_resync = sector_nr;
 825
 826         /* For these conditions we must wait:
 827          * A: while the array is in frozen state
 828          * B: while barrier >= RESYNC_DEPTH, meaning resync reach
 829          *    the max count which allowed.
 830          * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
 831          *    next resync will reach to the window which normal bios are
 832          *    handling.
 833          * D: while there are any active requests in the current window.
 834          */
 835         wait_event_lock_irq(conf->wait_barrier,
 836                             !conf->array_frozen &&
 837                             conf->barrier < RESYNC_DEPTH &&
 838                             conf->current_window_requests == 0 &&
 839                             (conf->start_next_window >=
 840                              conf->next_resync + RESYNC_SECTORS),
 841                             conf->resync_lock);
 842
 843         conf->nr_pending++;
 844         spin_unlock_irq(&conf->resync_lock);
 845 }
 846
 847 static void lower_barrier(struct r1conf *conf)
 848 {
 849         unsigned long flags;
 850         BUG_ON(conf->barrier <= 0);
 851         spin_lock_irqsave(&conf->resync_lock, flags);
 852         conf->barrier--;
 853         conf->nr_pending--;
 854         spin_unlock_irqrestore(&conf->resync_lock, flags);
 855         wake_up(&conf->wait_barrier);
 856 }
 857
 858 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio)
 859 {
 860         bool wait = false;
 861
 862         if (conf->array_frozen || !bio)
 863                 wait = true;
 864         else if (conf->barrier && bio_data_dir(bio) == WRITE) {
 865                 if ((conf->mddev->curr_resync_completed
 866                      >= bio_end_sector(bio)) ||
 867                     (conf->start_next_window + NEXT_NORMALIO_DISTANCE
 868                      <= bio->bi_iter.bi_sector))
 869                         wait = false;
 870                 else
 871                         wait = true;
 872         }
 873
 874         return wait;
 875 }
 876
 877 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
 878 {
 879         sector_t sector = 0;
 880
 881         spin_lock_irq(&conf->resync_lock);
 882         if (need_to_wait_for_sync(conf, bio)) {
 883                 conf->nr_waiting++;
 884                 /* Wait for the barrier to drop.
 885                  * However if there are already pending
 886                  * requests (preventing the barrier from
 887                  * rising completely), and the
 888                  * per-process bio queue isn't empty,
 889                  * then don't wait, as we need to empty
 890                  * that queue to allow conf->start_next_window
 891                  * to increase.
 892                  */
 893                 raid1_log(conf->mddev, "wait barrier");
 894                 wait_event_lock_irq(conf->wait_barrier,
 895                                     !conf->array_frozen &&
 896                                     (!conf->barrier ||
 897                                      ((conf->start_next_window <
 898                                        conf->next_resync + RESYNC_SECTORS) &&
 899                                       current->bio_list &&
 900                                       !bio_list_empty(current->bio_list))),
 901                                     conf->resync_lock);
 902                 conf->nr_waiting--;
 903         }
 904
 905         if (bio && bio_data_dir(bio) == WRITE) {
 906                 if (bio->bi_iter.bi_sector >= conf->next_resync) {
 907                         if (conf->start_next_window == MaxSector)
 908                                 conf->start_next_window =
 909                                         conf->next_resync +
 910                                         NEXT_NORMALIO_DISTANCE;
 911
 912                         if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
 913                             <= bio->bi_iter.bi_sector)
 914                                 conf->next_window_requests++;
 915                         else
 916                                 conf->current_window_requests++;
 917                         sector = conf->start_next_window;
 918                 }
 919         }
 920
 921         conf->nr_pending++;
 922         spin_unlock_irq(&conf->resync_lock);
 923         return sector;
 924 }
 925
 926 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
 927                           sector_t bi_sector)
 928 {
 929         unsigned long flags;
 930
 931         spin_lock_irqsave(&conf->resync_lock, flags);
 932         conf->nr_pending--;
 933         if (start_next_window) {
 934                 if (start_next_window == conf->start_next_window) {
 935                         if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
 936                             <= bi_sector)
 937                                 conf->next_window_requests--;
 938                         else
 939                                 conf->current_window_requests--;
 940                 } else
 941                         conf->current_window_requests--;
 942
 943                 if (!conf->current_window_requests) {
 944                         if (conf->next_window_requests) {
 945                                 conf->current_window_requests =
 946                                         conf->next_window_requests;
 947                                 conf->next_window_requests = 0;
 948                                 conf->start_next_window +=
 949                                         NEXT_NORMALIO_DISTANCE;
 950                         } else
 951                                 conf->start_next_window = MaxSector;
 952                 }
 953         }
 954         spin_unlock_irqrestore(&conf->resync_lock, flags);
 955         wake_up(&conf->wait_barrier);
 956 }
 957
 958 static void freeze_array(struct r1conf *conf, int extra)
 959 {
 960         /* stop syncio and normal IO and wait for everything to
 961          * go quite.
 962          * We wait until nr_pending match nr_queued+extra
 963          * This is called in the context of one normal IO request
 964          * that has failed. Thus any sync request that might be pending
 965          * will be blocked by nr_pending, and we need to wait for
 966          * pending IO requests to complete or be queued for re-try.
 967          * Thus the number queued (nr_queued) plus this request (extra)
 968          * must match the number of pending IOs (nr_pending) before
 969          * we continue.
 970          */
 971         spin_lock_irq(&conf->resync_lock);
 972         conf->array_frozen = 1;
 973         raid1_log(conf->mddev, "wait freeze");
 974         wait_event_lock_irq_cmd(conf->wait_barrier,
 975                                 conf->nr_pending == conf->nr_queued+extra,
 976                                 conf->resync_lock,
 977                                 flush_pending_writes(conf));
 978         spin_unlock_irq(&conf->resync_lock);
 979 }
 980 static void unfreeze_array(struct r1conf *conf)
 981 {
 982         /* reverse the effect of the freeze */
 983         spin_lock_irq(&conf->resync_lock);
 984         conf->array_frozen = 0;
 985         wake_up(&conf->wait_barrier);
 986         spin_unlock_irq(&conf->resync_lock);
 987 }
 988
 989 /* duplicate the data pages for behind I/O
 990  */
 991 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 992 {
 993         int i;
 994         struct bio_vec *bvec;
 995         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 996                                         GFP_NOIO);
 997         if (unlikely(!bvecs))
 998                 return;
 999
1000         bio_for_each_segment_all(bvec, bio, i) {
1001                 bvecs[i] = *bvec;
1002                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
1003                 if (unlikely(!bvecs[i].bv_page))
1004                         goto do_sync_io;
1005                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
1006                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
1007                 kunmap(bvecs[i].bv_page);
1008                 kunmap(bvec->bv_page);
1009         }
1010         r1_bio->behind_bvecs = bvecs;
1011         r1_bio->behind_page_count = bio->bi_vcnt;
1012         set_bit(R1BIO_BehindIO, &r1_bio->state);
1013         return;
1014
1015 do_sync_io:
1016         for (i = 0; i < bio->bi_vcnt; i++)
1017                 if (bvecs[i].bv_page)
1018                         put_page(bvecs[i].bv_page);
1019         kfree(bvecs);
1020         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1021                  bio->bi_iter.bi_size);
1022 }
1023
1024 struct raid1_plug_cb {
1025         struct blk_plug_cb      cb;
1026         struct bio_list         pending;
1027         int                     pending_cnt;
1028 };
1029
1030 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1031 {
1032         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1033                                                   cb);
1034         struct mddev *mddev = plug->cb.data;
1035         struct r1conf *conf = mddev->private;
1036         struct bio *bio;
1037
1038         if (from_schedule || current->bio_list) {
1039                 spin_lock_irq(&conf->device_lock);
1040                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1041                 conf->pending_count += plug->pending_cnt;
1042                 spin_unlock_irq(&conf->device_lock);
1043                 wake_up(&conf->wait_barrier);
1044                 md_wakeup_thread(mddev->thread);
1045                 kfree(plug);
1046                 return;
1047         }
1048
1049         /* we aren't scheduling, so we can do the write-out directly. */
1050         bio = bio_list_get(&plug->pending);
1051         bitmap_unplug(mddev->bitmap);
1052         wake_up(&conf->wait_barrier);
1053
1054         while (bio) { /* submit pending writes */
1055                 struct bio *next = bio->bi_next;
1056                 struct md_rdev *rdev = (void*)bio->bi_bdev;
1057                 bio->bi_next = NULL;
1058                 bio->bi_bdev = rdev->bdev;
1059                 if (test_bit(Faulty, &rdev->flags)) {
1060                         bio->bi_error = -EIO;
1061                         bio_endio(bio);
1062                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1063                                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1064                         /* Just ignore it */
1065                         bio_endio(bio);
1066                 else
1067                         generic_make_request(bio);
1068                 bio = next;
1069         }
1070         kfree(plug);
1071 }
1072
1073 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1074                                  struct r1bio *r1_bio)
1075 {
1076         struct r1conf *conf = mddev->private;
1077         struct raid1_info *mirror;
1078         struct bio *read_bio;
1079         struct bitmap *bitmap = mddev->bitmap;
1080         const int op = bio_op(bio);
1081         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1082         int sectors_handled;
1083         int max_sectors;
1084         int rdisk;
1085
1086         wait_barrier(conf, bio);
1087
1088 read_again:
1089         rdisk = read_balance(conf, r1_bio, &max_sectors);
1090
1091         if (rdisk < 0) {
1092                 /* couldn't find anywhere to read from */
1093                 raid_end_bio_io(r1_bio);
1094                 return;
1095         }
1096         mirror = conf->mirrors + rdisk;
1097
1098         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1099             bitmap) {
1100                 /*
1101                  * Reading from a write-mostly device must take care not to
1102                  * over-take any writes that are 'behind'
1103                  */
1104                 raid1_log(mddev, "wait behind writes");
1105                 wait_event(bitmap->behind_wait,
1106                            atomic_read(&bitmap->behind_writes) == 0);
1107         }
1108         r1_bio->read_disk = rdisk;
1109         r1_bio->start_next_window = 0;
1110
1111         read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1112         bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1113                  max_sectors);
1114
1115         r1_bio->bios[rdisk] = read_bio;
1116
1117         read_bio->bi_iter.bi_sector = r1_bio->sector +
1118                 mirror->rdev->data_offset;
1119         read_bio->bi_bdev = mirror->rdev->bdev;
1120         read_bio->bi_end_io = raid1_end_read_request;
1121         bio_set_op_attrs(read_bio, op, do_sync);
1122         if (test_bit(FailFast, &mirror->rdev->flags) &&
1123             test_bit(R1BIO_FailFast, &r1_bio->state))
1124                 read_bio->bi_opf |= MD_FAILFAST;
1125         read_bio->bi_private = r1_bio;
1126
1127         if (mddev->gendisk)
1128                 trace_block_bio_remap(bdev_get_queue(read_bio->bi_bdev),
1129                                       read_bio, disk_devt(mddev->gendisk),
1130                                       r1_bio->sector);
1131
1132         if (max_sectors < r1_bio->sectors) {
1133                 /*
1134                  * could not read all from this device, so we will need another
1135                  * r1_bio.
1136                  */
1137                 sectors_handled = (r1_bio->sector + max_sectors
1138                                    - bio->bi_iter.bi_sector);
1139                 r1_bio->sectors = max_sectors;
1140                 spin_lock_irq(&conf->device_lock);
1141                 if (bio->bi_phys_segments == 0)
1142                         bio->bi_phys_segments = 2;
1143                 else
1144                         bio->bi_phys_segments++;
1145                 spin_unlock_irq(&conf->device_lock);
1146
1147                 /*
1148                  * Cannot call generic_make_request directly as that will be
1149                  * queued in __make_request and subsequent mempool_alloc might
1150                  * block waiting for it.  So hand bio over to raid1d.
1151                  */
1152                 reschedule_retry(r1_bio);
1153
1154                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1155
1156                 r1_bio->master_bio = bio;
1157                 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1158                 r1_bio->state = 0;
1159                 r1_bio->mddev = mddev;
1160                 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1161                 goto read_again;
1162         } else
1163                 generic_make_request(read_bio);
1164 }
1165
1166 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1167                                 struct r1bio *r1_bio)
1168 {
1169         struct r1conf *conf = mddev->private;
1170         int i, disks;
1171         struct bitmap *bitmap = mddev->bitmap;
1172         unsigned long flags;
1173         struct md_rdev *blocked_rdev;
1174         struct blk_plug_cb *cb;
1175         struct raid1_plug_cb *plug = NULL;
1176         int first_clone;
1177         int sectors_handled;
1178         int max_sectors;
1179         sector_t start_next_window;
1180
1181         /*
1182          * Register the new request and wait if the reconstruction
1183          * thread has put up a bar for new requests.
1184          * Continue immediately if no resync is active currently.
1185          */
1186
1187         md_write_start(mddev, bio); /* wait on superblock update early */
1188
1189         if ((bio_end_sector(bio) > mddev->suspend_lo &&
1190             bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1191             (mddev_is_clustered(mddev) &&
1192              md_cluster_ops->area_resyncing(mddev, WRITE,
1193                      bio->bi_iter.bi_sector, bio_end_sector(bio)))) {
1194
1195                 /*
1196                  * As the suspend_* range is controlled by userspace, we want
1197                  * an interruptible wait.
1198                  */
1199                 DEFINE_WAIT(w);
1200                 for (;;) {
1201                         flush_signals(current);
1202                         prepare_to_wait(&conf->wait_barrier,
1203                                         &w, TASK_INTERRUPTIBLE);
1204                         if (bio_end_sector(bio) <= mddev->suspend_lo ||
1205                             bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1206                             (mddev_is_clustered(mddev) &&
1207                              !md_cluster_ops->area_resyncing(mddev, WRITE,
1208                                      bio->bi_iter.bi_sector,
1209                                      bio_end_sector(bio))))
1210                                 break;
1211                         schedule();
1212                 }
1213                 finish_wait(&conf->wait_barrier, &w);
1214         }
1215         start_next_window = wait_barrier(conf, bio);
1216
1217         if (conf->pending_count >= max_queued_requests) {
1218                 md_wakeup_thread(mddev->thread);
1219                 raid1_log(mddev, "wait queued");
1220                 wait_event(conf->wait_barrier,
1221                            conf->pending_count < max_queued_requests);
1222         }
1223         /* first select target devices under rcu_lock and
1224          * inc refcount on their rdev.  Record them by setting
1225          * bios[x] to bio
1226          * If there are known/acknowledged bad blocks on any device on
1227          * which we have seen a write error, we want to avoid writing those
1228          * blocks.
1229          * This potentially requires several writes to write around
1230          * the bad blocks.  Each set of writes gets it's own r1bio
1231          * with a set of bios attached.
1232          */
1233
1234         disks = conf->raid_disks * 2;
1235  retry_write:
1236         r1_bio->start_next_window = start_next_window;
1237         blocked_rdev = NULL;
1238         rcu_read_lock();
1239         max_sectors = r1_bio->sectors;
1240         for (i = 0;  i < disks; i++) {
1241                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1242                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1243                         atomic_inc(&rdev->nr_pending);
1244                         blocked_rdev = rdev;
1245                         break;
1246                 }
1247                 r1_bio->bios[i] = NULL;
1248                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1249                         if (i < conf->raid_disks)
1250                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1251                         continue;
1252                 }
1253
1254                 atomic_inc(&rdev->nr_pending);
1255                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1256                         sector_t first_bad;
1257                         int bad_sectors;
1258                         int is_bad;
1259
1260                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1261                                              &first_bad, &bad_sectors);
1262                         if (is_bad < 0) {
1263                                 /* mustn't write here until the bad block is
1264                                  * acknowledged*/
1265                                 set_bit(BlockedBadBlocks, &rdev->flags);
1266                                 blocked_rdev = rdev;
1267                                 break;
1268                         }
1269                         if (is_bad && first_bad <= r1_bio->sector) {
1270                                 /* Cannot write here at all */
1271                                 bad_sectors -= (r1_bio->sector - first_bad);
1272                                 if (bad_sectors < max_sectors)
1273                                         /* mustn't write more than bad_sectors
1274                                          * to other devices yet
1275                                          */
1276                                         max_sectors = bad_sectors;
1277                                 rdev_dec_pending(rdev, mddev);
1278                                 /* We don't set R1BIO_Degraded as that
1279                                  * only applies if the disk is
1280                                  * missing, so it might be re-added,
1281                                  * and we want to know to recover this
1282                                  * chunk.
1283                                  * In this case the device is here,
1284                                  * and the fact that this chunk is not
1285                                  * in-sync is recorded in the bad
1286                                  * block log
1287                                  */
1288                                 continue;
1289                         }
1290                         if (is_bad) {
1291                                 int good_sectors = first_bad - r1_bio->sector;
1292                                 if (good_sectors < max_sectors)
1293                                         max_sectors = good_sectors;
1294                         }
1295                 }
1296                 r1_bio->bios[i] = bio;
1297         }
1298         rcu_read_unlock();
1299
1300         if (unlikely(blocked_rdev)) {
1301                 /* Wait for this device to become unblocked */
1302                 int j;
1303                 sector_t old = start_next_window;
1304
1305                 for (j = 0; j < i; j++)
1306                         if (r1_bio->bios[j])
1307                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1308                 r1_bio->state = 0;
1309                 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector);
1310                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1311                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1312                 start_next_window = wait_barrier(conf, bio);
1313                 /*
1314                  * We must make sure the multi r1bios of bio have
1315                  * the same value of bi_phys_segments
1316                  */
1317                 if (bio->bi_phys_segments && old &&
1318                     old != start_next_window)
1319                         /* Wait for the former r1bio(s) to complete */
1320                         wait_event(conf->wait_barrier,
1321                                    bio->bi_phys_segments == 1);
1322                 goto retry_write;
1323         }
1324
1325         if (max_sectors < r1_bio->sectors) {
1326                 /* We are splitting this write into multiple parts, so
1327                  * we need to prepare for allocating another r1_bio.
1328                  */
1329                 r1_bio->sectors = max_sectors;
1330                 spin_lock_irq(&conf->device_lock);
1331                 if (bio->bi_phys_segments == 0)
1332                         bio->bi_phys_segments = 2;
1333                 else
1334                         bio->bi_phys_segments++;
1335                 spin_unlock_irq(&conf->device_lock);
1336         }
1337         sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1338
1339         atomic_set(&r1_bio->remaining, 1);
1340         atomic_set(&r1_bio->behind_remaining, 0);
1341
1342         first_clone = 1;
1343         for (i = 0; i < disks; i++) {
1344                 struct bio *mbio;
1345                 if (!r1_bio->bios[i])
1346                         continue;
1347
1348                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1349                 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector,
1350                          max_sectors);
1351
1352                 if (first_clone) {
1353                         /* do behind I/O ?
1354                          * Not if there are too many, or cannot
1355                          * allocate memory, or a reader on WriteMostly
1356                          * is waiting for behind writes to flush */
1357                         if (bitmap &&
1358                             (atomic_read(&bitmap->behind_writes)
1359                              < mddev->bitmap_info.max_write_behind) &&
1360                             !waitqueue_active(&bitmap->behind_wait))
1361                                 alloc_behind_pages(mbio, r1_bio);
1362
1363                         bitmap_startwrite(bitmap, r1_bio->sector,
1364                                           r1_bio->sectors,
1365                                           test_bit(R1BIO_BehindIO,
1366                                                    &r1_bio->state));
1367                         first_clone = 0;
1368                 }
1369                 if (r1_bio->behind_bvecs) {
1370                         struct bio_vec *bvec;
1371                         int j;
1372
1373                         /*
1374                          * We trimmed the bio, so _all is legit
1375                          */
1376                         bio_for_each_segment_all(bvec, mbio, j)
1377                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1378                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1379                                 atomic_inc(&r1_bio->behind_remaining);
1380                 }
1381
1382                 r1_bio->bios[i] = mbio;
1383
1384                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1385                                    conf->mirrors[i].rdev->data_offset);
1386                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1387                 mbio->bi_end_io = raid1_end_write_request;
1388                 mbio->bi_opf = bio_op(bio) |
1389                         (bio->bi_opf & (REQ_SYNC | REQ_PREFLUSH | REQ_FUA));
1390                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1391                     !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1392                     conf->raid_disks - mddev->degraded > 1)
1393                         mbio->bi_opf |= MD_FAILFAST;
1394                 mbio->bi_private = r1_bio;
1395
1396                 atomic_inc(&r1_bio->remaining);
1397
1398                 if (mddev->gendisk)
1399                         trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1400                                               mbio, disk_devt(mddev->gendisk),
1401                                               r1_bio->sector);
1402                 /* flush_pending_writes() needs access to the rdev so...*/
1403                 mbio->bi_bdev = (void*)conf->mirrors[i].rdev;
1404
1405                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1406                 if (cb)
1407                         plug = container_of(cb, struct raid1_plug_cb, cb);
1408                 else
1409                         plug = NULL;
1410                 spin_lock_irqsave(&conf->device_lock, flags);
1411                 if (plug) {
1412                         bio_list_add(&plug->pending, mbio);
1413                         plug->pending_cnt++;
1414                 } else {
1415                         bio_list_add(&conf->pending_bio_list, mbio);
1416                         conf->pending_count++;
1417                 }
1418                 spin_unlock_irqrestore(&conf->device_lock, flags);
1419                 if (!plug)
1420                         md_wakeup_thread(mddev->thread);
1421         }
1422         /* Mustn't call r1_bio_write_done before this next test,
1423          * as it could result in the bio being freed.
1424          */
1425         if (sectors_handled < bio_sectors(bio)) {
1426                 r1_bio_write_done(r1_bio);
1427                 /* We need another r1_bio.  It has already been counted
1428                  * in bio->bi_phys_segments
1429                  */
1430                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1431                 r1_bio->master_bio = bio;
1432                 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1433                 r1_bio->state = 0;
1434                 r1_bio->mddev = mddev;
1435                 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1436                 goto retry_write;
1437         }
1438
1439         r1_bio_write_done(r1_bio);
1440
1441         /* In case raid1d snuck in to freeze_array */
1442         wake_up(&conf->wait_barrier);
1443 }
1444
1445 static void raid1_make_request(struct mddev *mddev, struct bio *bio)
1446 {
1447         struct r1conf *conf = mddev->private;
1448         struct r1bio *r1_bio;
1449
1450         /*
1451          * make_request() can abort the operation when read-ahead is being
1452          * used and no empty request is available.
1453          *
1454          */
1455         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1456
1457         r1_bio->master_bio = bio;
1458         r1_bio->sectors = bio_sectors(bio);
1459         r1_bio->state = 0;
1460         r1_bio->mddev = mddev;
1461         r1_bio->sector = bio->bi_iter.bi_sector;
1462
1463         /*
1464          * We might need to issue multiple reads to different devices if there
1465          * are bad blocks around, so we keep track of the number of reads in
1466          * bio->bi_phys_segments.  If this is 0, there is only one r1_bio and
1467          * no locking will be needed when requests complete.  If it is
1468          * non-zero, then it is the number of not-completed requests.
1469          */
1470         bio->bi_phys_segments = 0;
1471         bio_clear_flag(bio, BIO_SEG_VALID);
1472
1473         if (bio_data_dir(bio) == READ)
1474                 raid1_read_request(mddev, bio, r1_bio);
1475         else
1476                 raid1_write_request(mddev, bio, r1_bio);
1477 }
1478
1479 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1480 {
1481         struct r1conf *conf = mddev->private;
1482         int i;
1483
1484         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1485                    conf->raid_disks - mddev->degraded);
1486         rcu_read_lock();
1487         for (i = 0; i < conf->raid_disks; i++) {
1488                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1489                 seq_printf(seq, "%s",
1490                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1491         }
1492         rcu_read_unlock();
1493         seq_printf(seq, "]");
1494 }
1495
1496 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1497 {
1498         char b[BDEVNAME_SIZE];
1499         struct r1conf *conf = mddev->private;
1500         unsigned long flags;
1501
1502         /*
1503          * If it is not operational, then we have already marked it as dead
1504          * else if it is the last working disks, ignore the error, let the
1505          * next level up know.
1506          * else mark the drive as failed
1507          */
1508         spin_lock_irqsave(&conf->device_lock, flags);
1509         if (test_bit(In_sync, &rdev->flags)
1510             && (conf->raid_disks - mddev->degraded) == 1) {
1511                 /*
1512                  * Don't fail the drive, act as though we were just a
1513                  * normal single drive.
1514                  * However don't try a recovery from this drive as
1515                  * it is very likely to fail.
1516                  */
1517                 conf->recovery_disabled = mddev->recovery_disabled;
1518                 spin_unlock_irqrestore(&conf->device_lock, flags);
1519                 return;
1520         }
1521         set_bit(Blocked, &rdev->flags);
1522         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1523                 mddev->degraded++;
1524                 set_bit(Faulty, &rdev->flags);
1525         } else
1526                 set_bit(Faulty, &rdev->flags);
1527         spin_unlock_irqrestore(&conf->device_lock, flags);
1528         /*
1529          * if recovery is running, make sure it aborts.
1530          */
1531         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1532         set_mask_bits(&mddev->sb_flags, 0,
1533                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1534         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1535                 "md/raid1:%s: Operation continuing on %d devices.\n",
1536                 mdname(mddev), bdevname(rdev->bdev, b),
1537                 mdname(mddev), conf->raid_disks - mddev->degraded);
1538 }
1539
1540 static void print_conf(struct r1conf *conf)
1541 {
1542         int i;
1543
1544         pr_debug("RAID1 conf printout:\n");
1545         if (!conf) {
1546                 pr_debug("(!conf)\n");
1547                 return;
1548         }
1549         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1550                  conf->raid_disks);
1551
1552         rcu_read_lock();
1553         for (i = 0; i < conf->raid_disks; i++) {
1554                 char b[BDEVNAME_SIZE];
1555                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1556                 if (rdev)
1557                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1558                                  i, !test_bit(In_sync, &rdev->flags),
1559                                  !test_bit(Faulty, &rdev->flags),
1560                                  bdevname(rdev->bdev,b));
1561         }
1562         rcu_read_unlock();
1563 }
1564
1565 static void close_sync(struct r1conf *conf)
1566 {
1567         wait_barrier(conf, NULL);
1568         allow_barrier(conf, 0, 0);
1569
1570         mempool_destroy(conf->r1buf_pool);
1571         conf->r1buf_pool = NULL;
1572
1573         spin_lock_irq(&conf->resync_lock);
1574         conf->next_resync = MaxSector - 2 * NEXT_NORMALIO_DISTANCE;
1575         conf->start_next_window = MaxSector;
1576         conf->current_window_requests +=
1577                 conf->next_window_requests;
1578         conf->next_window_requests = 0;
1579         spin_unlock_irq(&conf->resync_lock);
1580 }
1581
1582 static int raid1_spare_active(struct mddev *mddev)
1583 {
1584         int i;
1585         struct r1conf *conf = mddev->private;
1586         int count = 0;
1587         unsigned long flags;
1588
1589         /*
1590          * Find all failed disks within the RAID1 configuration
1591          * and mark them readable.
1592          * Called under mddev lock, so rcu protection not needed.
1593          * device_lock used to avoid races with raid1_end_read_request
1594          * which expects 'In_sync' flags and ->degraded to be consistent.
1595          */
1596         spin_lock_irqsave(&conf->device_lock, flags);
1597         for (i = 0; i < conf->raid_disks; i++) {
1598                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1599                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1600                 if (repl
1601                     && !test_bit(Candidate, &repl->flags)
1602                     && repl->recovery_offset == MaxSector
1603                     && !test_bit(Faulty, &repl->flags)
1604                     && !test_and_set_bit(In_sync, &repl->flags)) {
1605                         /* replacement has just become active */
1606                         if (!rdev ||
1607                             !test_and_clear_bit(In_sync, &rdev->flags))
1608                                 count++;
1609                         if (rdev) {
1610                                 /* Replaced device not technically
1611                                  * faulty, but we need to be sure
1612                                  * it gets removed and never re-added
1613                                  */
1614                                 set_bit(Faulty, &rdev->flags);
1615                                 sysfs_notify_dirent_safe(
1616                                         rdev->sysfs_state);
1617                         }
1618                 }
1619                 if (rdev
1620                     && rdev->recovery_offset == MaxSector
1621                     && !test_bit(Faulty, &rdev->flags)
1622                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1623                         count++;
1624                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1625                 }
1626         }
1627         mddev->degraded -= count;
1628         spin_unlock_irqrestore(&conf->device_lock, flags);
1629
1630         print_conf(conf);
1631         return count;
1632 }
1633
1634 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1635 {
1636         struct r1conf *conf = mddev->private;
1637         int err = -EEXIST;
1638         int mirror = 0;
1639         struct raid1_info *p;
1640         int first = 0;
1641         int last = conf->raid_disks - 1;
1642
1643         if (mddev->recovery_disabled == conf->recovery_disabled)
1644                 return -EBUSY;
1645
1646         if (md_integrity_add_rdev(rdev, mddev))
1647                 return -ENXIO;
1648
1649         if (rdev->raid_disk >= 0)
1650                 first = last = rdev->raid_disk;
1651
1652         /*
1653          * find the disk ... but prefer rdev->saved_raid_disk
1654          * if possible.
1655          */
1656         if (rdev->saved_raid_disk >= 0 &&
1657             rdev->saved_raid_disk >= first &&
1658             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1659                 first = last = rdev->saved_raid_disk;
1660
1661         for (mirror = first; mirror <= last; mirror++) {
1662                 p = conf->mirrors+mirror;
1663                 if (!p->rdev) {
1664
1665                         if (mddev->gendisk)
1666                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1667                                                   rdev->data_offset << 9);
1668
1669                         p->head_position = 0;
1670                         rdev->raid_disk = mirror;
1671                         err = 0;
1672                         /* As all devices are equivalent, we don't need a full recovery
1673                          * if this was recently any drive of the array
1674                          */
1675                         if (rdev->saved_raid_disk < 0)
1676                                 conf->fullsync = 1;
1677                         rcu_assign_pointer(p->rdev, rdev);
1678                         break;
1679                 }
1680                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1681                     p[conf->raid_disks].rdev == NULL) {
1682                         /* Add this device as a replacement */
1683                         clear_bit(In_sync, &rdev->flags);
1684                         set_bit(Replacement, &rdev->flags);
1685                         rdev->raid_disk = mirror;
1686                         err = 0;
1687                         conf->fullsync = 1;
1688                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1689                         break;
1690                 }
1691         }
1692         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1693                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1694         print_conf(conf);
1695         return err;
1696 }
1697
1698 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1699 {
1700         struct r1conf *conf = mddev->private;
1701         int err = 0;
1702         int number = rdev->raid_disk;
1703         struct raid1_info *p = conf->mirrors + number;
1704
1705         if (rdev != p->rdev)
1706                 p = conf->mirrors + conf->raid_disks + number;
1707
1708         print_conf(conf);
1709         if (rdev == p->rdev) {
1710                 if (test_bit(In_sync, &rdev->flags) ||
1711                     atomic_read(&rdev->nr_pending)) {
1712                         err = -EBUSY;
1713                         goto abort;
1714                 }
1715                 /* Only remove non-faulty devices if recovery
1716                  * is not possible.
1717                  */
1718                 if (!test_bit(Faulty, &rdev->flags) &&
1719                     mddev->recovery_disabled != conf->recovery_disabled &&
1720                     mddev->degraded < conf->raid_disks) {
1721                         err = -EBUSY;
1722                         goto abort;
1723                 }
1724                 p->rdev = NULL;
1725                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1726                         synchronize_rcu();
1727                         if (atomic_read(&rdev->nr_pending)) {
1728                                 /* lost the race, try later */
1729                                 err = -EBUSY;
1730                                 p->rdev = rdev;
1731                                 goto abort;
1732                         }
1733                 }
1734                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1735                         /* We just removed a device that is being replaced.
1736                          * Move down the replacement.  We drain all IO before
1737                          * doing this to avoid confusion.
1738                          */
1739                         struct md_rdev *repl =
1740                                 conf->mirrors[conf->raid_disks + number].rdev;
1741                         freeze_array(conf, 0);
1742                         clear_bit(Replacement, &repl->flags);
1743                         p->rdev = repl;
1744                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1745                         unfreeze_array(conf);
1746                         clear_bit(WantReplacement, &rdev->flags);
1747                 } else
1748                         clear_bit(WantReplacement, &rdev->flags);
1749                 err = md_integrity_register(mddev);
1750         }
1751 abort:
1752
1753         print_conf(conf);
1754         return err;
1755 }
1756
1757 static void end_sync_read(struct bio *bio)
1758 {
1759         struct r1bio *r1_bio = bio->bi_private;
1760
1761         update_head_pos(r1_bio->read_disk, r1_bio);
1762
1763         /*
1764          * we have read a block, now it needs to be re-written,
1765          * or re-read if the read failed.
1766          * We don't do much here, just schedule handling by raid1d
1767          */
1768         if (!bio->bi_error)
1769                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1770
1771         if (atomic_dec_and_test(&r1_bio->remaining))
1772                 reschedule_retry(r1_bio);
1773 }
1774
1775 static void end_sync_write(struct bio *bio)
1776 {
1777         int uptodate = !bio->bi_error;
1778         struct r1bio *r1_bio = bio->bi_private;
1779         struct mddev *mddev = r1_bio->mddev;
1780         struct r1conf *conf = mddev->private;
1781         sector_t first_bad;
1782         int bad_sectors;
1783         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1784
1785         if (!uptodate) {
1786                 sector_t sync_blocks = 0;
1787                 sector_t s = r1_bio->sector;
1788                 long sectors_to_go = r1_bio->sectors;
1789                 /* make sure these bits doesn't get cleared. */
1790                 do {
1791                         bitmap_end_sync(mddev->bitmap, s,
1792                                         &sync_blocks, 1);
1793                         s += sync_blocks;
1794                         sectors_to_go -= sync_blocks;
1795                 } while (sectors_to_go > 0);
1796                 set_bit(WriteErrorSeen, &rdev->flags);
1797                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1798                         set_bit(MD_RECOVERY_NEEDED, &
1799                                 mddev->recovery);
1800                 set_bit(R1BIO_WriteError, &r1_bio->state);
1801         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1802                                &first_bad, &bad_sectors) &&
1803                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1804                                 r1_bio->sector,
1805                                 r1_bio->sectors,
1806                                 &first_bad, &bad_sectors)
1807                 )
1808                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1809
1810         if (atomic_dec_and_test(&r1_bio->remaining)) {
1811                 int s = r1_bio->sectors;
1812                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1813                     test_bit(R1BIO_WriteError, &r1_bio->state))
1814                         reschedule_retry(r1_bio);
1815                 else {
1816                         put_buf(r1_bio);
1817                         md_done_sync(mddev, s, uptodate);
1818                 }
1819         }
1820 }
1821
1822 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1823                             int sectors, struct page *page, int rw)
1824 {
1825         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1826                 /* success */
1827                 return 1;
1828         if (rw == WRITE) {
1829                 set_bit(WriteErrorSeen, &rdev->flags);
1830                 if (!test_and_set_bit(WantReplacement,
1831                                       &rdev->flags))
1832                         set_bit(MD_RECOVERY_NEEDED, &
1833                                 rdev->mddev->recovery);
1834         }
1835         /* need to record an error - either for the block or the device */
1836         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1837                 md_error(rdev->mddev, rdev);
1838         return 0;
1839 }
1840
1841 static int fix_sync_read_error(struct r1bio *r1_bio)
1842 {
1843         /* Try some synchronous reads of other devices to get
1844          * good data, much like with normal read errors.  Only
1845          * read into the pages we already have so we don't
1846          * need to re-issue the read request.
1847          * We don't need to freeze the array, because being in an
1848          * active sync request, there is no normal IO, and
1849          * no overlapping syncs.
1850          * We don't need to check is_badblock() again as we
1851          * made sure that anything with a bad block in range
1852          * will have bi_end_io clear.
1853          */
1854         struct mddev *mddev = r1_bio->mddev;
1855         struct r1conf *conf = mddev->private;
1856         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1857         sector_t sect = r1_bio->sector;
1858         int sectors = r1_bio->sectors;
1859         int idx = 0;
1860         struct md_rdev *rdev;
1861
1862         rdev = conf->mirrors[r1_bio->read_disk].rdev;
1863         if (test_bit(FailFast, &rdev->flags)) {
1864                 /* Don't try recovering from here - just fail it
1865                  * ... unless it is the last working device of course */
1866                 md_error(mddev, rdev);
1867                 if (test_bit(Faulty, &rdev->flags))
1868                         /* Don't try to read from here, but make sure
1869                          * put_buf does it's thing
1870                          */
1871                         bio->bi_end_io = end_sync_write;
1872         }
1873
1874         while(sectors) {
1875                 int s = sectors;
1876                 int d = r1_bio->read_disk;
1877                 int success = 0;
1878                 int start;
1879
1880                 if (s > (PAGE_SIZE>>9))
1881                         s = PAGE_SIZE >> 9;
1882                 do {
1883                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1884                                 /* No rcu protection needed here devices
1885                                  * can only be removed when no resync is
1886                                  * active, and resync is currently active
1887                                  */
1888                                 rdev = conf->mirrors[d].rdev;
1889                                 if (sync_page_io(rdev, sect, s<<9,
1890                                                  bio->bi_io_vec[idx].bv_page,
1891                                                  REQ_OP_READ, 0, false)) {
1892                                         success = 1;
1893                                         break;
1894                                 }
1895                         }
1896                         d++;
1897                         if (d == conf->raid_disks * 2)
1898                                 d = 0;
1899                 } while (!success && d != r1_bio->read_disk);
1900
1901                 if (!success) {
1902                         char b[BDEVNAME_SIZE];
1903                         int abort = 0;
1904                         /* Cannot read from anywhere, this block is lost.
1905                          * Record a bad block on each device.  If that doesn't
1906                          * work just disable and interrupt the recovery.
1907                          * Don't fail devices as that won't really help.
1908                          */
1909                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1910                                             mdname(mddev),
1911                                             bdevname(bio->bi_bdev, b),
1912                                             (unsigned long long)r1_bio->sector);
1913                         for (d = 0; d < conf->raid_disks * 2; d++) {
1914                                 rdev = conf->mirrors[d].rdev;
1915                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1916                                         continue;
1917                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1918                                         abort = 1;
1919                         }
1920                         if (abort) {
1921                                 conf->recovery_disabled =
1922                                         mddev->recovery_disabled;
1923                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1924                                 md_done_sync(mddev, r1_bio->sectors, 0);
1925                                 put_buf(r1_bio);
1926                                 return 0;
1927                         }
1928                         /* Try next page */
1929                         sectors -= s;
1930                         sect += s;
1931                         idx++;
1932                         continue;
1933                 }
1934
1935                 start = d;
1936                 /* write it back and re-read */
1937                 while (d != r1_bio->read_disk) {
1938                         if (d == 0)
1939                                 d = conf->raid_disks * 2;
1940                         d--;
1941                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1942                                 continue;
1943                         rdev = conf->mirrors[d].rdev;
1944                         if (r1_sync_page_io(rdev, sect, s,
1945                                             bio->bi_io_vec[idx].bv_page,
1946                                             WRITE) == 0) {
1947                                 r1_bio->bios[d]->bi_end_io = NULL;
1948                                 rdev_dec_pending(rdev, mddev);
1949                         }
1950                 }
1951                 d = start;
1952                 while (d != r1_bio->read_disk) {
1953                         if (d == 0)
1954                                 d = conf->raid_disks * 2;
1955                         d--;
1956                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1957                                 continue;
1958                         rdev = conf->mirrors[d].rdev;
1959                         if (r1_sync_page_io(rdev, sect, s,
1960                                             bio->bi_io_vec[idx].bv_page,
1961                                             READ) != 0)
1962                                 atomic_add(s, &rdev->corrected_errors);
1963                 }
1964                 sectors -= s;
1965                 sect += s;
1966                 idx ++;
1967         }
1968         set_bit(R1BIO_Uptodate, &r1_bio->state);
1969         bio->bi_error = 0;
1970         return 1;
1971 }
1972
1973 static void process_checks(struct r1bio *r1_bio)
1974 {
1975         /* We have read all readable devices.  If we haven't
1976          * got the block, then there is no hope left.
1977          * If we have, then we want to do a comparison
1978          * and skip the write if everything is the same.
1979          * If any blocks failed to read, then we need to
1980          * attempt an over-write
1981          */
1982         struct mddev *mddev = r1_bio->mddev;
1983         struct r1conf *conf = mddev->private;
1984         int primary;
1985         int i;
1986         int vcnt;
1987
1988         /* Fix variable parts of all bios */
1989         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1990         for (i = 0; i < conf->raid_disks * 2; i++) {
1991                 int j;
1992                 int size;
1993                 int error;
1994                 struct bio *b = r1_bio->bios[i];
1995                 if (b->bi_end_io != end_sync_read)
1996                         continue;
1997                 /* fixup the bio for reuse, but preserve errno */
1998                 error = b->bi_error;
1999                 bio_reset(b);
2000                 b->bi_error = error;
2001                 b->bi_vcnt = vcnt;
2002                 b->bi_iter.bi_size = r1_bio->sectors << 9;
2003                 b->bi_iter.bi_sector = r1_bio->sector +
2004                         conf->mirrors[i].rdev->data_offset;
2005                 b->bi_bdev = conf->mirrors[i].rdev->bdev;
2006                 b->bi_end_io = end_sync_read;
2007                 b->bi_private = r1_bio;
2008
2009                 size = b->bi_iter.bi_size;
2010                 for (j = 0; j < vcnt ; j++) {
2011                         struct bio_vec *bi;
2012                         bi = &b->bi_io_vec[j];
2013                         bi->bv_offset = 0;
2014                         if (size > PAGE_SIZE)
2015                                 bi->bv_len = PAGE_SIZE;
2016                         else
2017                                 bi->bv_len = size;
2018                         size -= PAGE_SIZE;
2019                 }
2020         }
2021         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2022                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2023                     !r1_bio->bios[primary]->bi_error) {
2024                         r1_bio->bios[primary]->bi_end_io = NULL;
2025                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2026                         break;
2027                 }
2028         r1_bio->read_disk = primary;
2029         for (i = 0; i < conf->raid_disks * 2; i++) {
2030                 int j;
2031                 struct bio *pbio = r1_bio->bios[primary];
2032                 struct bio *sbio = r1_bio->bios[i];
2033                 int error = sbio->bi_error;
2034
2035                 if (sbio->bi_end_io != end_sync_read)
2036                         continue;
2037                 /* Now we can 'fixup' the error value */
2038                 sbio->bi_error = 0;
2039
2040                 if (!error) {
2041                         for (j = vcnt; j-- ; ) {
2042                                 struct page *p, *s;
2043                                 p = pbio->bi_io_vec[j].bv_page;
2044                                 s = sbio->bi_io_vec[j].bv_page;
2045                                 if (memcmp(page_address(p),
2046                                            page_address(s),
2047                                            sbio->bi_io_vec[j].bv_len))
2048                                         break;
2049                         }
2050                 } else
2051                         j = 0;
2052                 if (j >= 0)
2053                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2054                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2055                               && !error)) {
2056                         /* No need to write to this device. */
2057                         sbio->bi_end_io = NULL;
2058                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2059                         continue;
2060                 }
2061
2062                 bio_copy_data(sbio, pbio);
2063         }
2064 }
2065
2066 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2067 {
2068         struct r1conf *conf = mddev->private;
2069         int i;
2070         int disks = conf->raid_disks * 2;
2071         struct bio *bio, *wbio;
2072
2073         bio = r1_bio->bios[r1_bio->read_disk];
2074
2075         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2076                 /* ouch - failed to read all of that. */
2077                 if (!fix_sync_read_error(r1_bio))
2078                         return;
2079
2080         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2081                 process_checks(r1_bio);
2082
2083         /*
2084          * schedule writes
2085          */
2086         atomic_set(&r1_bio->remaining, 1);
2087         for (i = 0; i < disks ; i++) {
2088                 wbio = r1_bio->bios[i];
2089                 if (wbio->bi_end_io == NULL ||
2090                     (wbio->bi_end_io == end_sync_read &&
2091                      (i == r1_bio->read_disk ||
2092                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2093                         continue;
2094
2095                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2096                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2097                         wbio->bi_opf |= MD_FAILFAST;
2098
2099                 wbio->bi_end_io = end_sync_write;
2100                 atomic_inc(&r1_bio->remaining);
2101                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2102
2103                 generic_make_request(wbio);
2104         }
2105
2106         if (atomic_dec_and_test(&r1_bio->remaining)) {
2107                 /* if we're here, all write(s) have completed, so clean up */
2108                 int s = r1_bio->sectors;
2109                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2110                     test_bit(R1BIO_WriteError, &r1_bio->state))
2111                         reschedule_retry(r1_bio);
2112                 else {
2113                         put_buf(r1_bio);
2114                         md_done_sync(mddev, s, 1);
2115                 }
2116         }
2117 }
2118
2119 /*
2120  * This is a kernel thread which:
2121  *
2122  *      1.      Retries failed read operations on working mirrors.
2123  *      2.      Updates the raid superblock when problems encounter.
2124  *      3.      Performs writes following reads for array synchronising.
2125  */
2126
2127 static void fix_read_error(struct r1conf *conf, int read_disk,
2128                            sector_t sect, int sectors)
2129 {
2130         struct mddev *mddev = conf->mddev;
2131         while(sectors) {
2132                 int s = sectors;
2133                 int d = read_disk;
2134                 int success = 0;
2135                 int start;
2136                 struct md_rdev *rdev;
2137
2138                 if (s > (PAGE_SIZE>>9))
2139                         s = PAGE_SIZE >> 9;
2140
2141                 do {
2142                         sector_t first_bad;
2143                         int bad_sectors;
2144
2145                         rcu_read_lock();
2146                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2147                         if (rdev &&
2148                             (test_bit(In_sync, &rdev->flags) ||
2149                              (!test_bit(Faulty, &rdev->flags) &&
2150                               rdev->recovery_offset >= sect + s)) &&
2151                             is_badblock(rdev, sect, s,
2152                                         &first_bad, &bad_sectors) == 0) {
2153                                 atomic_inc(&rdev->nr_pending);
2154                                 rcu_read_unlock();
2155                                 if (sync_page_io(rdev, sect, s<<9,
2156                                          conf->tmppage, REQ_OP_READ, 0, false))
2157                                         success = 1;
2158                                 rdev_dec_pending(rdev, mddev);
2159                                 if (success)
2160                                         break;
2161                         } else
2162                                 rcu_read_unlock();
2163                         d++;
2164                         if (d == conf->raid_disks * 2)
2165                                 d = 0;
2166                 } while (!success && d != read_disk);
2167
2168                 if (!success) {
2169                         /* Cannot read from anywhere - mark it bad */
2170                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2171                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2172                                 md_error(mddev, rdev);
2173                         break;
2174                 }
2175                 /* write it back and re-read */
2176                 start = d;
2177                 while (d != read_disk) {
2178                         if (d==0)
2179                                 d = conf->raid_disks * 2;
2180                         d--;
2181                         rcu_read_lock();
2182                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2183                         if (rdev &&
2184                             !test_bit(Faulty, &rdev->flags)) {
2185                                 atomic_inc(&rdev->nr_pending);
2186                                 rcu_read_unlock();
2187                                 r1_sync_page_io(rdev, sect, s,
2188                                                 conf->tmppage, WRITE);
2189                                 rdev_dec_pending(rdev, mddev);
2190                         } else
2191                                 rcu_read_unlock();
2192                 }
2193                 d = start;
2194                 while (d != read_disk) {
2195                         char b[BDEVNAME_SIZE];
2196                         if (d==0)
2197                                 d = conf->raid_disks * 2;
2198                         d--;
2199                         rcu_read_lock();
2200                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2201                         if (rdev &&
2202                             !test_bit(Faulty, &rdev->flags)) {
2203                                 atomic_inc(&rdev->nr_pending);
2204                                 rcu_read_unlock();
2205                                 if (r1_sync_page_io(rdev, sect, s,
2206                                                     conf->tmppage, READ)) {
2207                                         atomic_add(s, &rdev->corrected_errors);
2208                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2209                                                 mdname(mddev), s,
2210                                                 (unsigned long long)(sect +
2211                                                                      rdev->data_offset),
2212                                                 bdevname(rdev->bdev, b));
2213                                 }
2214                                 rdev_dec_pending(rdev, mddev);
2215                         } else
2216                                 rcu_read_unlock();
2217                 }
2218                 sectors -= s;
2219                 sect += s;
2220         }
2221 }
2222
2223 static int narrow_write_error(struct r1bio *r1_bio, int i)
2224 {
2225         struct mddev *mddev = r1_bio->mddev;
2226         struct r1conf *conf = mddev->private;
2227         struct md_rdev *rdev = conf->mirrors[i].rdev;
2228
2229         /* bio has the data to be written to device 'i' where
2230          * we just recently had a write error.
2231          * We repeatedly clone the bio and trim down to one block,
2232          * then try the write.  Where the write fails we record
2233          * a bad block.
2234          * It is conceivable that the bio doesn't exactly align with
2235          * blocks.  We must handle this somehow.
2236          *
2237          * We currently own a reference on the rdev.
2238          */
2239
2240         int block_sectors;
2241         sector_t sector;
2242         int sectors;
2243         int sect_to_write = r1_bio->sectors;
2244         int ok = 1;
2245
2246         if (rdev->badblocks.shift < 0)
2247                 return 0;
2248
2249         block_sectors = roundup(1 << rdev->badblocks.shift,
2250                                 bdev_logical_block_size(rdev->bdev) >> 9);
2251         sector = r1_bio->sector;
2252         sectors = ((sector + block_sectors)
2253                    & ~(sector_t)(block_sectors - 1))
2254                 - sector;
2255
2256         while (sect_to_write) {
2257                 struct bio *wbio;
2258                 if (sectors > sect_to_write)
2259                         sectors = sect_to_write;
2260                 /* Write at 'sector' for 'sectors'*/
2261
2262                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2263                         unsigned vcnt = r1_bio->behind_page_count;
2264                         struct bio_vec *vec = r1_bio->behind_bvecs;
2265
2266                         while (!vec->bv_page) {
2267                                 vec++;
2268                                 vcnt--;
2269                         }
2270
2271                         wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2272                         memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2273
2274                         wbio->bi_vcnt = vcnt;
2275                 } else {
2276                         wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2277                 }
2278
2279                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2280                 wbio->bi_iter.bi_sector = r1_bio->sector;
2281                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2282
2283                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2284                 wbio->bi_iter.bi_sector += rdev->data_offset;
2285                 wbio->bi_bdev = rdev->bdev;
2286
2287                 if (submit_bio_wait(wbio) < 0)
2288                         /* failure! */
2289                         ok = rdev_set_badblocks(rdev, sector,
2290                                                 sectors, 0)
2291                                 && ok;
2292
2293                 bio_put(wbio);
2294                 sect_to_write -= sectors;
2295                 sector += sectors;
2296                 sectors = block_sectors;
2297         }
2298         return ok;
2299 }
2300
2301 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2302 {
2303         int m;
2304         int s = r1_bio->sectors;
2305         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2306                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2307                 struct bio *bio = r1_bio->bios[m];
2308                 if (bio->bi_end_io == NULL)
2309                         continue;
2310                 if (!bio->bi_error &&
2311                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2312                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2313                 }
2314                 if (bio->bi_error &&
2315                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2316                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2317                                 md_error(conf->mddev, rdev);
2318                 }
2319         }
2320         put_buf(r1_bio);
2321         md_done_sync(conf->mddev, s, 1);
2322 }
2323
2324 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2325 {
2326         int m;
2327         bool fail = false;
2328         for (m = 0; m < conf->raid_disks * 2 ; m++)
2329                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2330                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2331                         rdev_clear_badblocks(rdev,
2332                                              r1_bio->sector,
2333                                              r1_bio->sectors, 0);
2334                         rdev_dec_pending(rdev, conf->mddev);
2335                 } else if (r1_bio->bios[m] != NULL) {
2336                         /* This drive got a write error.  We need to
2337                          * narrow down and record precise write
2338                          * errors.
2339                          */
2340                         fail = true;
2341                         if (!narrow_write_error(r1_bio, m)) {
2342                                 md_error(conf->mddev,
2343                                          conf->mirrors[m].rdev);
2344                                 /* an I/O failed, we can't clear the bitmap */
2345                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2346                         }
2347                         rdev_dec_pending(conf->mirrors[m].rdev,
2348                                          conf->mddev);
2349                 }
2350         if (fail) {
2351                 spin_lock_irq(&conf->device_lock);
2352                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2353                 conf->nr_queued++;
2354                 spin_unlock_irq(&conf->device_lock);
2355                 md_wakeup_thread(conf->mddev->thread);
2356         } else {
2357                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2358                         close_write(r1_bio);
2359                 raid_end_bio_io(r1_bio);
2360         }
2361 }
2362
2363 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2364 {
2365         int disk;
2366         int max_sectors;
2367         struct mddev *mddev = conf->mddev;
2368         struct bio *bio;
2369         char b[BDEVNAME_SIZE];
2370         struct md_rdev *rdev;
2371         dev_t bio_dev;
2372         sector_t bio_sector;
2373
2374         clear_bit(R1BIO_ReadError, &r1_bio->state);
2375         /* we got a read error. Maybe the drive is bad.  Maybe just
2376          * the block and we can fix it.
2377          * We freeze all other IO, and try reading the block from
2378          * other devices.  When we find one, we re-write
2379          * and check it that fixes the read error.
2380          * This is all done synchronously while the array is
2381          * frozen
2382          */
2383
2384         bio = r1_bio->bios[r1_bio->read_disk];
2385         bdevname(bio->bi_bdev, b);
2386         bio_dev = bio->bi_bdev->bd_dev;
2387         bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2388         bio_put(bio);
2389         r1_bio->bios[r1_bio->read_disk] = NULL;
2390
2391         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2392         if (mddev->ro == 0
2393             && !test_bit(FailFast, &rdev->flags)) {
2394                 freeze_array(conf, 1);
2395                 fix_read_error(conf, r1_bio->read_disk,
2396                                r1_bio->sector, r1_bio->sectors);
2397                 unfreeze_array(conf);
2398         } else {
2399                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2400         }
2401
2402         rdev_dec_pending(rdev, conf->mddev);
2403
2404 read_more:
2405         disk = read_balance(conf, r1_bio, &max_sectors);
2406         if (disk == -1) {
2407                 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2408                                     mdname(mddev), b, (unsigned long long)r1_bio->sector);
2409                 raid_end_bio_io(r1_bio);
2410         } else {
2411                 const unsigned long do_sync
2412                         = r1_bio->master_bio->bi_opf & REQ_SYNC;
2413                 r1_bio->read_disk = disk;
2414                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2415                 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2416                          max_sectors);
2417                 r1_bio->bios[r1_bio->read_disk] = bio;
2418                 rdev = conf->mirrors[disk].rdev;
2419                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
2420                                     mdname(mddev),
2421                                     (unsigned long long)r1_bio->sector,
2422                                     bdevname(rdev->bdev, b));
2423                 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2424                 bio->bi_bdev = rdev->bdev;
2425                 bio->bi_end_io = raid1_end_read_request;
2426                 bio_set_op_attrs(bio, REQ_OP_READ, do_sync);
2427                 if (test_bit(FailFast, &rdev->flags) &&
2428                     test_bit(R1BIO_FailFast, &r1_bio->state))
2429                         bio->bi_opf |= MD_FAILFAST;
2430                 bio->bi_private = r1_bio;
2431                 if (max_sectors < r1_bio->sectors) {
2432                         /* Drat - have to split this up more */
2433                         struct bio *mbio = r1_bio->master_bio;
2434                         int sectors_handled = (r1_bio->sector + max_sectors
2435                                                - mbio->bi_iter.bi_sector);
2436                         r1_bio->sectors = max_sectors;
2437                         spin_lock_irq(&conf->device_lock);
2438                         if (mbio->bi_phys_segments == 0)
2439                                 mbio->bi_phys_segments = 2;
2440                         else
2441                                 mbio->bi_phys_segments++;
2442                         spin_unlock_irq(&conf->device_lock);
2443                         trace_block_bio_remap(bdev_get_queue(bio->bi_bdev),
2444                                               bio, bio_dev, bio_sector);
2445                         generic_make_request(bio);
2446                         bio = NULL;
2447
2448                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2449
2450                         r1_bio->master_bio = mbio;
2451                         r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2452                         r1_bio->state = 0;
2453                         set_bit(R1BIO_ReadError, &r1_bio->state);
2454                         r1_bio->mddev = mddev;
2455                         r1_bio->sector = mbio->bi_iter.bi_sector +
2456                                 sectors_handled;
2457
2458                         goto read_more;
2459                 } else {
2460                         trace_block_bio_remap(bdev_get_queue(bio->bi_bdev),
2461                                               bio, bio_dev, bio_sector);
2462                         generic_make_request(bio);
2463                 }
2464         }
2465 }
2466
2467 static void raid1d(struct md_thread *thread)
2468 {
2469         struct mddev *mddev = thread->mddev;
2470         struct r1bio *r1_bio;
2471         unsigned long flags;
2472         struct r1conf *conf = mddev->private;
2473         struct list_head *head = &conf->retry_list;
2474         struct blk_plug plug;
2475
2476         md_check_recovery(mddev);
2477
2478         if (!list_empty_careful(&conf->bio_end_io_list) &&
2479             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2480                 LIST_HEAD(tmp);
2481                 spin_lock_irqsave(&conf->device_lock, flags);
2482                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2483                         while (!list_empty(&conf->bio_end_io_list)) {
2484                                 list_move(conf->bio_end_io_list.prev, &tmp);
2485                                 conf->nr_queued--;
2486                         }
2487                 }
2488                 spin_unlock_irqrestore(&conf->device_lock, flags);
2489                 while (!list_empty(&tmp)) {
2490                         r1_bio = list_first_entry(&tmp, struct r1bio,
2491                                                   retry_list);
2492                         list_del(&r1_bio->retry_list);
2493                         if (mddev->degraded)
2494                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2495                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2496                                 close_write(r1_bio);
2497                         raid_end_bio_io(r1_bio);
2498                 }
2499         }
2500
2501         blk_start_plug(&plug);
2502         for (;;) {
2503
2504                 flush_pending_writes(conf);
2505
2506                 spin_lock_irqsave(&conf->device_lock, flags);
2507                 if (list_empty(head)) {
2508                         spin_unlock_irqrestore(&conf->device_lock, flags);
2509                         break;
2510                 }
2511                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2512                 list_del(head->prev);
2513                 conf->nr_queued--;
2514                 spin_unlock_irqrestore(&conf->device_lock, flags);
2515
2516                 mddev = r1_bio->mddev;
2517                 conf = mddev->private;
2518                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2519                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2520                             test_bit(R1BIO_WriteError, &r1_bio->state))
2521                                 handle_sync_write_finished(conf, r1_bio);
2522                         else
2523                                 sync_request_write(mddev, r1_bio);
2524                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2525                            test_bit(R1BIO_WriteError, &r1_bio->state))
2526                         handle_write_finished(conf, r1_bio);
2527                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2528                         handle_read_error(conf, r1_bio);
2529                 else
2530                         /* just a partial read to be scheduled from separate
2531                          * context
2532                          */
2533                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2534
2535                 cond_resched();
2536                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2537                         md_check_recovery(mddev);
2538         }
2539         blk_finish_plug(&plug);
2540 }
2541
2542 static int init_resync(struct r1conf *conf)
2543 {
2544         int buffs;
2545
2546         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2547         BUG_ON(conf->r1buf_pool);
2548         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2549                                           conf->poolinfo);
2550         if (!conf->r1buf_pool)
2551                 return -ENOMEM;
2552         conf->next_resync = 0;
2553         return 0;
2554 }
2555
2556 /*
2557  * perform a "sync" on one "block"
2558  *
2559  * We need to make sure that no normal I/O request - particularly write
2560  * requests - conflict with active sync requests.
2561  *
2562  * This is achieved by tracking pending requests and a 'barrier' concept
2563  * that can be installed to exclude normal IO requests.
2564  */
2565
2566 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2567                                    int *skipped)
2568 {
2569         struct r1conf *conf = mddev->private;
2570         struct r1bio *r1_bio;
2571         struct bio *bio;
2572         sector_t max_sector, nr_sectors;
2573         int disk = -1;
2574         int i;
2575         int wonly = -1;
2576         int write_targets = 0, read_targets = 0;
2577         sector_t sync_blocks;
2578         int still_degraded = 0;
2579         int good_sectors = RESYNC_SECTORS;
2580         int min_bad = 0; /* number of sectors that are bad in all devices */
2581
2582         if (!conf->r1buf_pool)
2583                 if (init_resync(conf))
2584                         return 0;
2585
2586         max_sector = mddev->dev_sectors;
2587         if (sector_nr >= max_sector) {
2588                 /* If we aborted, we need to abort the
2589                  * sync on the 'current' bitmap chunk (there will
2590                  * only be one in raid1 resync.
2591                  * We can find the current addess in mddev->curr_resync
2592                  */
2593                 if (mddev->curr_resync < max_sector) /* aborted */
2594                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2595                                                 &sync_blocks, 1);
2596                 else /* completed sync */
2597                         conf->fullsync = 0;
2598
2599                 bitmap_close_sync(mddev->bitmap);
2600                 close_sync(conf);
2601
2602                 if (mddev_is_clustered(mddev)) {
2603                         conf->cluster_sync_low = 0;
2604                         conf->cluster_sync_high = 0;
2605                 }
2606                 return 0;
2607         }
2608
2609         if (mddev->bitmap == NULL &&
2610             mddev->recovery_cp == MaxSector &&
2611             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2612             conf->fullsync == 0) {
2613                 *skipped = 1;
2614                 return max_sector - sector_nr;
2615         }
2616         /* before building a request, check if we can skip these blocks..
2617          * This call the bitmap_start_sync doesn't actually record anything
2618          */
2619         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2620             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2621                 /* We can skip this block, and probably several more */
2622                 *skipped = 1;
2623                 return sync_blocks;
2624         }
2625
2626         /*
2627          * If there is non-resync activity waiting for a turn, then let it
2628          * though before starting on this new sync request.
2629          */
2630         if (conf->nr_waiting)
2631                 schedule_timeout_uninterruptible(1);
2632
2633         /* we are incrementing sector_nr below. To be safe, we check against
2634          * sector_nr + two times RESYNC_SECTORS
2635          */
2636
2637         bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2638                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2639         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2640
2641         raise_barrier(conf, sector_nr);
2642
2643         rcu_read_lock();
2644         /*
2645          * If we get a correctably read error during resync or recovery,
2646          * we might want to read from a different device.  So we
2647          * flag all drives that could conceivably be read from for READ,
2648          * and any others (which will be non-In_sync devices) for WRITE.
2649          * If a read fails, we try reading from something else for which READ
2650          * is OK.
2651          */
2652
2653         r1_bio->mddev = mddev;
2654         r1_bio->sector = sector_nr;
2655         r1_bio->state = 0;
2656         set_bit(R1BIO_IsSync, &r1_bio->state);
2657
2658         for (i = 0; i < conf->raid_disks * 2; i++) {
2659                 struct md_rdev *rdev;
2660                 bio = r1_bio->bios[i];
2661                 bio_reset(bio);
2662
2663                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2664                 if (rdev == NULL ||
2665                     test_bit(Faulty, &rdev->flags)) {
2666                         if (i < conf->raid_disks)
2667                                 still_degraded = 1;
2668                 } else if (!test_bit(In_sync, &rdev->flags)) {
2669                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2670                         bio->bi_end_io = end_sync_write;
2671                         write_targets ++;
2672                 } else {
2673                         /* may need to read from here */
2674                         sector_t first_bad = MaxSector;
2675                         int bad_sectors;
2676
2677                         if (is_badblock(rdev, sector_nr, good_sectors,
2678                                         &first_bad, &bad_sectors)) {
2679                                 if (first_bad > sector_nr)
2680                                         good_sectors = first_bad - sector_nr;
2681                                 else {
2682                                         bad_sectors -= (sector_nr - first_bad);
2683                                         if (min_bad == 0 ||
2684                                             min_bad > bad_sectors)
2685                                                 min_bad = bad_sectors;
2686                                 }
2687                         }
2688                         if (sector_nr < first_bad) {
2689                                 if (test_bit(WriteMostly, &rdev->flags)) {
2690                                         if (wonly < 0)
2691                                                 wonly = i;
2692                                 } else {
2693                                         if (disk < 0)
2694                                                 disk = i;
2695                                 }
2696                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2697                                 bio->bi_end_io = end_sync_read;
2698                                 read_targets++;
2699                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2700                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2701                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2702                                 /*
2703                                  * The device is suitable for reading (InSync),
2704                                  * but has bad block(s) here. Let's try to correct them,
2705                                  * if we are doing resync or repair. Otherwise, leave
2706                                  * this device alone for this sync request.
2707                                  */
2708                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2709                                 bio->bi_end_io = end_sync_write;
2710                                 write_targets++;
2711                         }
2712                 }
2713                 if (bio->bi_end_io) {
2714                         atomic_inc(&rdev->nr_pending);
2715                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2716                         bio->bi_bdev = rdev->bdev;
2717                         bio->bi_private = r1_bio;
2718                         if (test_bit(FailFast, &rdev->flags))
2719                                 bio->bi_opf |= MD_FAILFAST;
2720                 }
2721         }
2722         rcu_read_unlock();
2723         if (disk < 0)
2724                 disk = wonly;
2725         r1_bio->read_disk = disk;
2726
2727         if (read_targets == 0 && min_bad > 0) {
2728                 /* These sectors are bad on all InSync devices, so we
2729                  * need to mark them bad on all write targets
2730                  */
2731                 int ok = 1;
2732                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2733                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2734                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2735                                 ok = rdev_set_badblocks(rdev, sector_nr,
2736                                                         min_bad, 0
2737                                         ) && ok;
2738                         }
2739                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2740                 *skipped = 1;
2741                 put_buf(r1_bio);
2742
2743                 if (!ok) {
2744                         /* Cannot record the badblocks, so need to
2745                          * abort the resync.
2746                          * If there are multiple read targets, could just
2747                          * fail the really bad ones ???
2748                          */
2749                         conf->recovery_disabled = mddev->recovery_disabled;
2750                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2751                         return 0;
2752                 } else
2753                         return min_bad;
2754
2755         }
2756         if (min_bad > 0 && min_bad < good_sectors) {
2757                 /* only resync enough to reach the next bad->good
2758                  * transition */
2759                 good_sectors = min_bad;
2760         }
2761
2762         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2763                 /* extra read targets are also write targets */
2764                 write_targets += read_targets-1;
2765
2766         if (write_targets == 0 || read_targets == 0) {
2767                 /* There is nowhere to write, so all non-sync
2768                  * drives must be failed - so we are finished
2769                  */
2770                 sector_t rv;
2771                 if (min_bad > 0)
2772                         max_sector = sector_nr + min_bad;
2773                 rv = max_sector - sector_nr;
2774                 *skipped = 1;
2775                 put_buf(r1_bio);
2776                 return rv;
2777         }
2778
2779         if (max_sector > mddev->resync_max)
2780                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2781         if (max_sector > sector_nr + good_sectors)
2782                 max_sector = sector_nr + good_sectors;
2783         nr_sectors = 0;
2784         sync_blocks = 0;
2785         do {
2786                 struct page *page;
2787                 int len = PAGE_SIZE;
2788                 if (sector_nr + (len>>9) > max_sector)
2789                         len = (max_sector - sector_nr) << 9;
2790                 if (len == 0)
2791                         break;
2792                 if (sync_blocks == 0) {
2793                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2794                                                &sync_blocks, still_degraded) &&
2795                             !conf->fullsync &&
2796                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2797                                 break;
2798                         if ((len >> 9) > sync_blocks)
2799                                 len = sync_blocks<<9;
2800                 }
2801
2802                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2803                         bio = r1_bio->bios[i];
2804                         if (bio->bi_end_io) {
2805                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2806                                 if (bio_add_page(bio, page, len, 0) == 0) {
2807                                         /* stop here */
2808                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2809                                         while (i > 0) {
2810                                                 i--;
2811                                                 bio = r1_bio->bios[i];
2812                                                 if (bio->bi_end_io==NULL)
2813                                                         continue;
2814                                                 /* remove last page from this bio */
2815                                                 bio->bi_vcnt--;
2816                                                 bio->bi_iter.bi_size -= len;
2817                                                 bio_clear_flag(bio, BIO_SEG_VALID);
2818                                         }
2819                                         goto bio_full;
2820                                 }
2821                         }
2822                 }
2823                 nr_sectors += len>>9;
2824                 sector_nr += len>>9;
2825                 sync_blocks -= (len>>9);
2826         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2827  bio_full:
2828         r1_bio->sectors = nr_sectors;
2829
2830         if (mddev_is_clustered(mddev) &&
2831                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2832                 conf->cluster_sync_low = mddev->curr_resync_completed;
2833                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2834                 /* Send resync message */
2835                 md_cluster_ops->resync_info_update(mddev,
2836                                 conf->cluster_sync_low,
2837                                 conf->cluster_sync_high);
2838         }
2839
2840         /* For a user-requested sync, we read all readable devices and do a
2841          * compare
2842          */
2843         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2844                 atomic_set(&r1_bio->remaining, read_targets);
2845                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2846                         bio = r1_bio->bios[i];
2847                         if (bio->bi_end_io == end_sync_read) {
2848                                 read_targets--;
2849                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2850                                 if (read_targets == 1)
2851                                         bio->bi_opf &= ~MD_FAILFAST;
2852                                 generic_make_request(bio);
2853                         }
2854                 }
2855         } else {
2856                 atomic_set(&r1_bio->remaining, 1);
2857                 bio = r1_bio->bios[r1_bio->read_disk];
2858                 md_sync_acct(bio->bi_bdev, nr_sectors);
2859                 if (read_targets == 1)
2860                         bio->bi_opf &= ~MD_FAILFAST;
2861                 generic_make_request(bio);
2862
2863         }
2864         return nr_sectors;
2865 }
2866
2867 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2868 {
2869         if (sectors)
2870                 return sectors;
2871
2872         return mddev->dev_sectors;
2873 }
2874
2875 static struct r1conf *setup_conf(struct mddev *mddev)
2876 {
2877         struct r1conf *conf;
2878         int i;
2879         struct raid1_info *disk;
2880         struct md_rdev *rdev;
2881         int err = -ENOMEM;
2882
2883         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2884         if (!conf)
2885                 goto abort;
2886
2887         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2888                                 * mddev->raid_disks * 2,
2889                                  GFP_KERNEL);
2890         if (!conf->mirrors)
2891                 goto abort;
2892
2893         conf->tmppage = alloc_page(GFP_KERNEL);
2894         if (!conf->tmppage)
2895                 goto abort;
2896
2897         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2898         if (!conf->poolinfo)
2899                 goto abort;
2900         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2901         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2902                                           r1bio_pool_free,
2903                                           conf->poolinfo);
2904         if (!conf->r1bio_pool)
2905                 goto abort;
2906
2907         conf->poolinfo->mddev = mddev;
2908
2909         err = -EINVAL;
2910         spin_lock_init(&conf->device_lock);
2911         rdev_for_each(rdev, mddev) {
2912                 struct request_queue *q;
2913                 int disk_idx = rdev->raid_disk;
2914                 if (disk_idx >= mddev->raid_disks
2915                     || disk_idx < 0)
2916                         continue;
2917                 if (test_bit(Replacement, &rdev->flags))
2918                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
2919                 else
2920                         disk = conf->mirrors + disk_idx;
2921
2922                 if (disk->rdev)
2923                         goto abort;
2924                 disk->rdev = rdev;
2925                 q = bdev_get_queue(rdev->bdev);
2926
2927                 disk->head_position = 0;
2928                 disk->seq_start = MaxSector;
2929         }
2930         conf->raid_disks = mddev->raid_disks;
2931         conf->mddev = mddev;
2932         INIT_LIST_HEAD(&conf->retry_list);
2933         INIT_LIST_HEAD(&conf->bio_end_io_list);
2934
2935         spin_lock_init(&conf->resync_lock);
2936         init_waitqueue_head(&conf->wait_barrier);
2937
2938         bio_list_init(&conf->pending_bio_list);
2939         conf->pending_count = 0;
2940         conf->recovery_disabled = mddev->recovery_disabled - 1;
2941
2942         conf->start_next_window = MaxSector;
2943         conf->current_window_requests = conf->next_window_requests = 0;
2944
2945         err = -EIO;
2946         for (i = 0; i < conf->raid_disks * 2; i++) {
2947
2948                 disk = conf->mirrors + i;
2949
2950                 if (i < conf->raid_disks &&
2951                     disk[conf->raid_disks].rdev) {
2952                         /* This slot has a replacement. */
2953                         if (!disk->rdev) {
2954                                 /* No original, just make the replacement
2955                                  * a recovering spare
2956                                  */
2957                                 disk->rdev =
2958                                         disk[conf->raid_disks].rdev;
2959                                 disk[conf->raid_disks].rdev = NULL;
2960                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2961                                 /* Original is not in_sync - bad */
2962                                 goto abort;
2963                 }
2964
2965                 if (!disk->rdev ||
2966                     !test_bit(In_sync, &disk->rdev->flags)) {
2967                         disk->head_position = 0;
2968                         if (disk->rdev &&
2969                             (disk->rdev->saved_raid_disk < 0))
2970                                 conf->fullsync = 1;
2971                 }
2972         }
2973
2974         err = -ENOMEM;
2975         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2976         if (!conf->thread)
2977                 goto abort;
2978
2979         return conf;
2980
2981  abort:
2982         if (conf) {
2983                 mempool_destroy(conf->r1bio_pool);
2984                 kfree(conf->mirrors);
2985                 safe_put_page(conf->tmppage);
2986                 kfree(conf->poolinfo);
2987                 kfree(conf);
2988         }
2989         return ERR_PTR(err);
2990 }
2991
2992 static void raid1_free(struct mddev *mddev, void *priv);
2993 static int raid1_run(struct mddev *mddev)
2994 {
2995         struct r1conf *conf;
2996         int i;
2997         struct md_rdev *rdev;
2998         int ret;
2999         bool discard_supported = false;
3000
3001         if (mddev->level != 1) {
3002                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3003                         mdname(mddev), mddev->level);
3004                 return -EIO;
3005         }
3006         if (mddev->reshape_position != MaxSector) {
3007                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3008                         mdname(mddev));
3009                 return -EIO;
3010         }
3011         /*
3012          * copy the already verified devices into our private RAID1
3013          * bookkeeping area. [whatever we allocate in run(),
3014          * should be freed in raid1_free()]
3015          */
3016         if (mddev->private == NULL)
3017                 conf = setup_conf(mddev);
3018         else
3019                 conf = mddev->private;
3020
3021         if (IS_ERR(conf))
3022                 return PTR_ERR(conf);
3023
3024         if (mddev->queue)
3025                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3026
3027         rdev_for_each(rdev, mddev) {
3028                 if (!mddev->gendisk)
3029                         continue;
3030                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3031                                   rdev->data_offset << 9);
3032                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3033                         discard_supported = true;
3034         }
3035
3036         mddev->degraded = 0;
3037         for (i=0; i < conf->raid_disks; i++)
3038                 if (conf->mirrors[i].rdev == NULL ||
3039                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3040                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3041                         mddev->degraded++;
3042
3043         if (conf->raid_disks - mddev->degraded == 1)
3044                 mddev->recovery_cp = MaxSector;
3045
3046         if (mddev->recovery_cp != MaxSector)
3047                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3048                         mdname(mddev));
3049         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3050                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3051                 mddev->raid_disks);
3052
3053         /*
3054          * Ok, everything is just fine now
3055          */
3056         mddev->thread = conf->thread;
3057         conf->thread = NULL;
3058         mddev->private = conf;
3059         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3060
3061         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3062
3063         if (mddev->queue) {
3064                 if (discard_supported)
3065                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3066                                                 mddev->queue);
3067                 else
3068                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3069                                                   mddev->queue);
3070         }
3071
3072         ret =  md_integrity_register(mddev);
3073         if (ret) {
3074                 md_unregister_thread(&mddev->thread);
3075                 raid1_free(mddev, conf);
3076         }
3077         return ret;
3078 }
3079
3080 static void raid1_free(struct mddev *mddev, void *priv)
3081 {
3082         struct r1conf *conf = priv;
3083
3084         mempool_destroy(conf->r1bio_pool);
3085         kfree(conf->mirrors);
3086         safe_put_page(conf->tmppage);
3087         kfree(conf->poolinfo);
3088         kfree(conf);
3089 }
3090
3091 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3092 {
3093         /* no resync is happening, and there is enough space
3094          * on all devices, so we can resize.
3095          * We need to make sure resync covers any new space.
3096          * If the array is shrinking we should possibly wait until
3097          * any io in the removed space completes, but it hardly seems
3098          * worth it.
3099          */
3100         sector_t newsize = raid1_size(mddev, sectors, 0);
3101         if (mddev->external_size &&
3102             mddev->array_sectors > newsize)
3103                 return -EINVAL;
3104         if (mddev->bitmap) {
3105                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3106                 if (ret)
3107                         return ret;
3108         }
3109         md_set_array_sectors(mddev, newsize);
3110         set_capacity(mddev->gendisk, mddev->array_sectors);
3111         revalidate_disk(mddev->gendisk);
3112         if (sectors > mddev->dev_sectors &&
3113             mddev->recovery_cp > mddev->dev_sectors) {
3114                 mddev->recovery_cp = mddev->dev_sectors;
3115                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3116         }
3117         mddev->dev_sectors = sectors;
3118         mddev->resync_max_sectors = sectors;
3119         return 0;
3120 }
3121
3122 static int raid1_reshape(struct mddev *mddev)
3123 {
3124         /* We need to:
3125          * 1/ resize the r1bio_pool
3126          * 2/ resize conf->mirrors
3127          *
3128          * We allocate a new r1bio_pool if we can.
3129          * Then raise a device barrier and wait until all IO stops.
3130          * Then resize conf->mirrors and swap in the new r1bio pool.
3131          *
3132          * At the same time, we "pack" the devices so that all the missing
3133          * devices have the higher raid_disk numbers.
3134          */
3135         mempool_t *newpool, *oldpool;
3136         struct pool_info *newpoolinfo;
3137         struct raid1_info *newmirrors;
3138         struct r1conf *conf = mddev->private;
3139         int cnt, raid_disks;
3140         unsigned long flags;
3141         int d, d2, err;
3142
3143         /* Cannot change chunk_size, layout, or level */
3144         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3145             mddev->layout != mddev->new_layout ||
3146             mddev->level != mddev->new_level) {
3147                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3148                 mddev->new_layout = mddev->layout;
3149                 mddev->new_level = mddev->level;
3150                 return -EINVAL;
3151         }
3152
3153         if (!mddev_is_clustered(mddev)) {
3154                 err = md_allow_write(mddev);
3155                 if (err)
3156                         return err;
3157         }
3158
3159         raid_disks = mddev->raid_disks + mddev->delta_disks;
3160
3161         if (raid_disks < conf->raid_disks) {
3162                 cnt=0;
3163                 for (d= 0; d < conf->raid_disks; d++)
3164                         if (conf->mirrors[d].rdev)
3165                                 cnt++;
3166                 if (cnt > raid_disks)
3167                         return -EBUSY;
3168         }
3169
3170         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3171         if (!newpoolinfo)
3172                 return -ENOMEM;
3173         newpoolinfo->mddev = mddev;
3174         newpoolinfo->raid_disks = raid_disks * 2;
3175
3176         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3177                                  r1bio_pool_free, newpoolinfo);
3178         if (!newpool) {
3179                 kfree(newpoolinfo);
3180                 return -ENOMEM;
3181         }
3182         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3183                              GFP_KERNEL);
3184         if (!newmirrors) {
3185                 kfree(newpoolinfo);
3186                 mempool_destroy(newpool);
3187                 return -ENOMEM;
3188         }
3189
3190         freeze_array(conf, 0);
3191
3192         /* ok, everything is stopped */
3193         oldpool = conf->r1bio_pool;
3194         conf->r1bio_pool = newpool;
3195
3196         for (d = d2 = 0; d < conf->raid_disks; d++) {
3197                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3198                 if (rdev && rdev->raid_disk != d2) {
3199                         sysfs_unlink_rdev(mddev, rdev);
3200                         rdev->raid_disk = d2;
3201                         sysfs_unlink_rdev(mddev, rdev);
3202                         if (sysfs_link_rdev(mddev, rdev))
3203                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3204                                         mdname(mddev), rdev->raid_disk);
3205                 }
3206                 if (rdev)
3207                         newmirrors[d2++].rdev = rdev;
3208         }
3209         kfree(conf->mirrors);
3210         conf->mirrors = newmirrors;
3211         kfree(conf->poolinfo);
3212         conf->poolinfo = newpoolinfo;
3213
3214         spin_lock_irqsave(&conf->device_lock, flags);
3215         mddev->degraded += (raid_disks - conf->raid_disks);
3216         spin_unlock_irqrestore(&conf->device_lock, flags);
3217         conf->raid_disks = mddev->raid_disks = raid_disks;
3218         mddev->delta_disks = 0;
3219
3220         unfreeze_array(conf);
3221
3222         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3223         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3224         md_wakeup_thread(mddev->thread);
3225
3226         mempool_destroy(oldpool);
3227         return 0;
3228 }
3229
3230 static void raid1_quiesce(struct mddev *mddev, int state)
3231 {
3232         struct r1conf *conf = mddev->private;
3233
3234         switch(state) {
3235         case 2: /* wake for suspend */
3236                 wake_up(&conf->wait_barrier);
3237                 break;
3238         case 1:
3239                 freeze_array(conf, 0);
3240                 break;
3241         case 0:
3242                 unfreeze_array(conf);
3243                 break;
3244         }
3245 }
3246
3247 static void *raid1_takeover(struct mddev *mddev)
3248 {
3249         /* raid1 can take over:
3250          *  raid5 with 2 devices, any layout or chunk size
3251          */
3252         if (mddev->level == 5 && mddev->raid_disks == 2) {
3253                 struct r1conf *conf;
3254                 mddev->new_level = 1;
3255                 mddev->new_layout = 0;
3256                 mddev->new_chunk_sectors = 0;
3257                 conf = setup_conf(mddev);
3258                 if (!IS_ERR(conf)) {
3259                         /* Array must appear to be quiesced */
3260                         conf->array_frozen = 1;
3261                         mddev_clear_unsupported_flags(mddev,
3262                                 UNSUPPORTED_MDDEV_FLAGS);
3263                 }
3264                 return conf;
3265         }
3266         return ERR_PTR(-EINVAL);
3267 }
3268
3269 static struct md_personality raid1_personality =
3270 {
3271         .name           = "raid1",
3272         .level          = 1,
3273         .owner          = THIS_MODULE,
3274         .make_request   = raid1_make_request,
3275         .run            = raid1_run,
3276         .free           = raid1_free,
3277         .status         = raid1_status,
3278         .error_handler  = raid1_error,
3279         .hot_add_disk   = raid1_add_disk,
3280         .hot_remove_disk= raid1_remove_disk,
3281         .spare_active   = raid1_spare_active,
3282         .sync_request   = raid1_sync_request,
3283         .resize         = raid1_resize,
3284         .size           = raid1_size,
3285         .check_reshape  = raid1_reshape,
3286         .quiesce        = raid1_quiesce,
3287         .takeover       = raid1_takeover,
3288         .congested      = raid1_congested,
3289 };
3290
3291 static int __init raid_init(void)
3292 {
3293         return register_md_personality(&raid1_personality);
3294 }
3295
3296 static void raid_exit(void)
3297 {
3298         unregister_md_personality(&raid1_personality);
3299 }
3300
3301 module_init(raid_init);
3302 module_exit(raid_exit);
3303 MODULE_LICENSE("GPL");
3304 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3305 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3306 MODULE_ALIAS("md-raid1");
3307 MODULE_ALIAS("md-level-1");
3308
3309 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);