kernel/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 "md.h"
  41 #include "raid1.h"
  42 #include "bitmap.h"
  43
  44 /*
  45  * Number of guaranteed r1bios in case of extreme VM load:
  46  */
  47 #define NR_RAID1_BIOS 256
  48
  49 /* when we get a read error on a read-only array, we redirect to another
  50  * device without failing the first device, or trying to over-write to
  51  * correct the read error.  To keep track of bad blocks on a per-bio
  52  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  53  */
  54 #define IO_BLOCKED ((struct bio *)1)
  55 /* When we successfully write to a known bad-block, we need to remove the
  56  * bad-block marking which must be done from process context.  So we record
  57  * the success by setting devs[n].bio to IO_MADE_GOOD
  58  */
  59 #define IO_MADE_GOOD ((struct bio *)2)
  60
  61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  62
  63 /* When there are this many requests queue to be written by
  64  * the raid1 thread, we become 'congested' to provide back-pressure
  65  * for writeback.
  66  */
  67 static int max_queued_requests = 1024;
  68
  69 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
  70                           sector_t bi_sector);
  71 static void lower_barrier(struct r1conf *conf);
  72
  73 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  74 {
  75         struct pool_info *pi = data;
  76         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  77
  78         /* allocate a r1bio with room for raid_disks entries in the bios array */
  79         return kzalloc(size, gfp_flags);
  80 }
  81
  82 static void r1bio_pool_free(void *r1_bio, void *data)
  83 {
  84         kfree(r1_bio);
  85 }
  86
  87 #define RESYNC_BLOCK_SIZE (64*1024)
  88 #define RESYNC_DEPTH 32
  89 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  90 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  91 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
  92 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
  93 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
  94
  95 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  96 {
  97         struct pool_info *pi = data;
  98         struct r1bio *r1_bio;
  99         struct bio *bio;
 100         int need_pages;
 101         int i, j;
 102
 103         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 104         if (!r1_bio)
 105                 return NULL;
 106
 107         /*
 108          * Allocate bios : 1 for reading, n-1 for writing
 109          */
 110         for (j = pi->raid_disks ; j-- ; ) {
 111                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 112                 if (!bio)
 113                         goto out_free_bio;
 114                 r1_bio->bios[j] = bio;
 115         }
 116         /*
 117          * Allocate RESYNC_PAGES data pages and attach them to
 118          * the first bio.
 119          * If this is a user-requested check/repair, allocate
 120          * RESYNC_PAGES for each bio.
 121          */
 122         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 123                 need_pages = pi->raid_disks;
 124         else
 125                 need_pages = 1;
 126         for (j = 0; j < need_pages; j++) {
 127                 bio = r1_bio->bios[j];
 128                 bio->bi_vcnt = RESYNC_PAGES;
 129
 130                 if (bio_alloc_pages(bio, gfp_flags))
 131                         goto out_free_pages;
 132         }
 133         /* If not user-requests, copy the page pointers to all bios */
 134         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 135                 for (i=0; i<RESYNC_PAGES ; i++)
 136                         for (j=1; j<pi->raid_disks; j++)
 137                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 138                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 139         }
 140
 141         r1_bio->master_bio = NULL;
 142
 143         return r1_bio;
 144
 145 out_free_pages:
 146         while (--j >= 0) {
 147                 struct bio_vec *bv;
 148
 149                 bio_for_each_segment_all(bv, r1_bio->bios[j], i)
 150                         __free_page(bv->bv_page);
 151         }
 152
 153 out_free_bio:
 154         while (++j < pi->raid_disks)
 155                 bio_put(r1_bio->bios[j]);
 156         r1bio_pool_free(r1_bio, data);
 157         return NULL;
 158 }
 159
 160 static void r1buf_pool_free(void *__r1_bio, void *data)
 161 {
 162         struct pool_info *pi = data;
 163         int i,j;
 164         struct r1bio *r1bio = __r1_bio;
 165
 166         for (i = 0; i < RESYNC_PAGES; i++)
 167                 for (j = pi->raid_disks; j-- ;) {
 168                         if (j == 0 ||
 169                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 170                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 171                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 172                 }
 173         for (i=0 ; i < pi->raid_disks; i++)
 174                 bio_put(r1bio->bios[i]);
 175
 176         r1bio_pool_free(r1bio, data);
 177 }
 178
 179 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 180 {
 181         int i;
 182
 183         for (i = 0; i < conf->raid_disks * 2; i++) {
 184                 struct bio **bio = r1_bio->bios + i;
 185                 if (!BIO_SPECIAL(*bio))
 186                         bio_put(*bio);
 187                 *bio = NULL;
 188         }
 189 }
 190
 191 static void free_r1bio(struct r1bio *r1_bio)
 192 {
 193         struct r1conf *conf = r1_bio->mddev->private;
 194
 195         put_all_bios(conf, r1_bio);
 196         mempool_free(r1_bio, conf->r1bio_pool);
 197 }
 198
 199 static void put_buf(struct r1bio *r1_bio)
 200 {
 201         struct r1conf *conf = r1_bio->mddev->private;
 202         int i;
 203
 204         for (i = 0; i < conf->raid_disks * 2; i++) {
 205                 struct bio *bio = r1_bio->bios[i];
 206                 if (bio->bi_end_io)
 207                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 208         }
 209
 210         mempool_free(r1_bio, conf->r1buf_pool);
 211
 212         lower_barrier(conf);
 213 }
 214
 215 static void reschedule_retry(struct r1bio *r1_bio)
 216 {
 217         unsigned long flags;
 218         struct mddev *mddev = r1_bio->mddev;
 219         struct r1conf *conf = mddev->private;
 220
 221         spin_lock_irqsave(&conf->device_lock, flags);
 222         list_add(&r1_bio->retry_list, &conf->retry_list);
 223         conf->nr_queued ++;
 224         spin_unlock_irqrestore(&conf->device_lock, flags);
 225
 226         wake_up(&conf->wait_barrier);
 227         md_wakeup_thread(mddev->thread);
 228 }
 229
 230 /*
 231  * raid_end_bio_io() is called when we have finished servicing a mirrored
 232  * operation and are ready to return a success/failure code to the buffer
 233  * cache layer.
 234  */
 235 static void call_bio_endio(struct r1bio *r1_bio)
 236 {
 237         struct bio *bio = r1_bio->master_bio;
 238         int done;
 239         struct r1conf *conf = r1_bio->mddev->private;
 240         sector_t start_next_window = r1_bio->start_next_window;
 241         sector_t bi_sector = bio->bi_iter.bi_sector;
 242
 243         if (bio->bi_phys_segments) {
 244                 unsigned long flags;
 245                 spin_lock_irqsave(&conf->device_lock, flags);
 246                 bio->bi_phys_segments--;
 247                 done = (bio->bi_phys_segments == 0);
 248                 spin_unlock_irqrestore(&conf->device_lock, flags);
 249                 /*
 250                  * make_request() might be waiting for
 251                  * bi_phys_segments to decrease
 252                  */
 253                 wake_up(&conf->wait_barrier);
 254         } else
 255                 done = 1;
 256
 257         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 258                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 259         if (done) {
 260                 bio_endio(bio, 0);
 261                 /*
 262                  * Wake up any possible resync thread that waits for the device
 263                  * to go idle.
 264                  */
 265                 allow_barrier(conf, start_next_window, bi_sector);
 266         }
 267 }
 268
 269 static void raid_end_bio_io(struct r1bio *r1_bio)
 270 {
 271         struct bio *bio = r1_bio->master_bio;
 272
 273         /* if nobody has done the final endio yet, do it now */
 274         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 275                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 276                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 277                          (unsigned long long) bio->bi_iter.bi_sector,
 278                          (unsigned long long) bio_end_sector(bio) - 1);
 279
 280                 call_bio_endio(r1_bio);
 281         }
 282         free_r1bio(r1_bio);
 283 }
 284
 285 /*
 286  * Update disk head position estimator based on IRQ completion info.
 287  */
 288 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 289 {
 290         struct r1conf *conf = r1_bio->mddev->private;
 291
 292         conf->mirrors[disk].head_position =
 293                 r1_bio->sector + (r1_bio->sectors);
 294 }
 295
 296 /*
 297  * Find the disk number which triggered given bio
 298  */
 299 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 300 {
 301         int mirror;
 302         struct r1conf *conf = r1_bio->mddev->private;
 303         int raid_disks = conf->raid_disks;
 304
 305         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 306                 if (r1_bio->bios[mirror] == bio)
 307                         break;
 308
 309         BUG_ON(mirror == raid_disks * 2);
 310         update_head_pos(mirror, r1_bio);
 311
 312         return mirror;
 313 }
 314
 315 static void raid1_end_read_request(struct bio *bio, int error)
 316 {
 317         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 318         struct r1bio *r1_bio = bio->bi_private;
 319         int mirror;
 320         struct r1conf *conf = r1_bio->mddev->private;
 321
 322         mirror = r1_bio->read_disk;
 323         /*
 324          * this branch is our 'one mirror IO has finished' event handler:
 325          */
 326         update_head_pos(mirror, r1_bio);
 327
 328         if (uptodate)
 329                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 330         else {
 331                 /* If all other devices have failed, we want to return
 332                  * the error upwards rather than fail the last device.
 333                  * Here we redefine "uptodate" to mean "Don't want to retry"
 334                  */
 335                 unsigned long flags;
 336                 spin_lock_irqsave(&conf->device_lock, flags);
 337                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 338                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 339                      test_bit(In_sync, &conf->mirrors[mirror].rdev->flags)))
 340                         uptodate = 1;
 341                 spin_unlock_irqrestore(&conf->device_lock, flags);
 342         }
 343
 344         if (uptodate) {
 345                 raid_end_bio_io(r1_bio);
 346                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 347         } else {
 348                 /*
 349                  * oops, read error:
 350                  */
 351                 char b[BDEVNAME_SIZE];
 352                 printk_ratelimited(
 353                         KERN_ERR "md/raid1:%s: %s: "
 354                         "rescheduling sector %llu\n",
 355                         mdname(conf->mddev),
 356                         bdevname(conf->mirrors[mirror].rdev->bdev,
 357                                  b),
 358                         (unsigned long long)r1_bio->sector);
 359                 set_bit(R1BIO_ReadError, &r1_bio->state);
 360                 reschedule_retry(r1_bio);
 361                 /* don't drop the reference on read_disk yet */
 362         }
 363 }
 364
 365 static void close_write(struct r1bio *r1_bio)
 366 {
 367         /* it really is the end of this request */
 368         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 369                 /* free extra copy of the data pages */
 370                 int i = r1_bio->behind_page_count;
 371                 while (i--)
 372                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 373                 kfree(r1_bio->behind_bvecs);
 374                 r1_bio->behind_bvecs = NULL;
 375         }
 376         /* clear the bitmap if all writes complete successfully */
 377         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 378                         r1_bio->sectors,
 379                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 380                         test_bit(R1BIO_BehindIO, &r1_bio->state));
 381         md_write_end(r1_bio->mddev);
 382 }
 383
 384 static void r1_bio_write_done(struct r1bio *r1_bio)
 385 {
 386         if (!atomic_dec_and_test(&r1_bio->remaining))
 387                 return;
 388
 389         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 390                 reschedule_retry(r1_bio);
 391         else {
 392                 close_write(r1_bio);
 393                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 394                         reschedule_retry(r1_bio);
 395                 else
 396                         raid_end_bio_io(r1_bio);
 397         }
 398 }
 399
 400 static void raid1_end_write_request(struct bio *bio, int error)
 401 {
 402         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 403         struct r1bio *r1_bio = bio->bi_private;
 404         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 405         struct r1conf *conf = r1_bio->mddev->private;
 406         struct bio *to_put = NULL;
 407
 408         mirror = find_bio_disk(r1_bio, bio);
 409
 410         /*
 411          * 'one mirror IO has finished' event handler:
 412          */
 413         if (!uptodate) {
 414                 set_bit(WriteErrorSeen,
 415                         &conf->mirrors[mirror].rdev->flags);
 416                 if (!test_and_set_bit(WantReplacement,
 417                                       &conf->mirrors[mirror].rdev->flags))
 418                         set_bit(MD_RECOVERY_NEEDED, &
 419                                 conf->mddev->recovery);
 420
 421                 set_bit(R1BIO_WriteError, &r1_bio->state);
 422         } else {
 423                 /*
 424                  * Set R1BIO_Uptodate in our master bio, so that we
 425                  * will return a good error code for to the higher
 426                  * levels even if IO on some other mirrored buffer
 427                  * fails.
 428                  *
 429                  * The 'master' represents the composite IO operation
 430                  * to user-side. So if something waits for IO, then it
 431                  * will wait for the 'master' bio.
 432                  */
 433                 sector_t first_bad;
 434                 int bad_sectors;
 435
 436                 r1_bio->bios[mirror] = NULL;
 437                 to_put = bio;
 438                 /*
 439                  * Do not set R1BIO_Uptodate if the current device is
 440                  * rebuilding or Faulty. This is because we cannot use
 441                  * such device for properly reading the data back (we could
 442                  * potentially use it, if the current write would have felt
 443                  * before rdev->recovery_offset, but for simplicity we don't
 444                  * check this here.
 445                  */
 446                 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
 447                     !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
 448                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 449
 450                 /* Maybe we can clear some bad blocks. */
 451                 if (is_badblock(conf->mirrors[mirror].rdev,
 452                                 r1_bio->sector, r1_bio->sectors,
 453                                 &first_bad, &bad_sectors)) {
 454                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 455                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 456                 }
 457         }
 458
 459         if (behind) {
 460                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 461                         atomic_dec(&r1_bio->behind_remaining);
 462
 463                 /*
 464                  * In behind mode, we ACK the master bio once the I/O
 465                  * has safely reached all non-writemostly
 466                  * disks. Setting the Returned bit ensures that this
 467                  * gets done only once -- we don't ever want to return
 468                  * -EIO here, instead we'll wait
 469                  */
 470                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 471                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 472                         /* Maybe we can return now */
 473                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 474                                 struct bio *mbio = r1_bio->master_bio;
 475                                 pr_debug("raid1: behind end write sectors"
 476                                          " %llu-%llu\n",
 477                                          (unsigned long long) mbio->bi_iter.bi_sector,
 478                                          (unsigned long long) bio_end_sector(mbio) - 1);
 479                                 call_bio_endio(r1_bio);
 480                         }
 481                 }
 482         }
 483         if (r1_bio->bios[mirror] == NULL)
 484                 rdev_dec_pending(conf->mirrors[mirror].rdev,
 485                                  conf->mddev);
 486
 487         /*
 488          * Let's see if all mirrored write operations have finished
 489          * already.
 490          */
 491         r1_bio_write_done(r1_bio);
 492
 493         if (to_put)
 494                 bio_put(to_put);
 495 }
 496
 497 /*
 498  * This routine returns the disk from which the requested read should
 499  * be done. There is a per-array 'next expected sequential IO' sector
 500  * number - if this matches on the next IO then we use the last disk.
 501  * There is also a per-disk 'last know head position' sector that is
 502  * maintained from IRQ contexts, both the normal and the resync IO
 503  * completion handlers update this position correctly. If there is no
 504  * perfect sequential match then we pick the disk whose head is closest.
 505  *
 506  * If there are 2 mirrors in the same 2 devices, performance degrades
 507  * because position is mirror, not device based.
 508  *
 509  * The rdev for the device selected will have nr_pending incremented.
 510  */
 511 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 512 {
 513         const sector_t this_sector = r1_bio->sector;
 514         int sectors;
 515         int best_good_sectors;
 516         int best_disk, best_dist_disk, best_pending_disk;
 517         int has_nonrot_disk;
 518         int disk;
 519         sector_t best_dist;
 520         unsigned int min_pending;
 521         struct md_rdev *rdev;
 522         int choose_first;
 523         int choose_next_idle;
 524
 525         rcu_read_lock();
 526         /*
 527          * Check if we can balance. We can balance on the whole
 528          * device if no resync is going on, or below the resync window.
 529          * We take the first readable disk when above the resync window.
 530          */
 531  retry:
 532         sectors = r1_bio->sectors;
 533         best_disk = -1;
 534         best_dist_disk = -1;
 535         best_dist = MaxSector;
 536         best_pending_disk = -1;
 537         min_pending = UINT_MAX;
 538         best_good_sectors = 0;
 539         has_nonrot_disk = 0;
 540         choose_next_idle = 0;
 541
 542         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
 543             (mddev_is_clustered(conf->mddev) &&
 544             md_cluster_ops->area_resyncing(conf->mddev, this_sector,
 545                     this_sector + sectors)))
 546                 choose_first = 1;
 547         else
 548                 choose_first = 0;
 549
 550         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 551                 sector_t dist;
 552                 sector_t first_bad;
 553                 int bad_sectors;
 554                 unsigned int pending;
 555                 bool nonrot;
 556
 557                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 558                 if (r1_bio->bios[disk] == IO_BLOCKED
 559                     || rdev == NULL
 560                     || test_bit(Unmerged, &rdev->flags)
 561                     || test_bit(Faulty, &rdev->flags))
 562                         continue;
 563                 if (!test_bit(In_sync, &rdev->flags) &&
 564                     rdev->recovery_offset < this_sector + sectors)
 565                         continue;
 566                 if (test_bit(WriteMostly, &rdev->flags)) {
 567                         /* Don't balance among write-mostly, just
 568                          * use the first as a last resort */
 569                         if (best_dist_disk < 0) {
 570                                 if (is_badblock(rdev, this_sector, sectors,
 571                                                 &first_bad, &bad_sectors)) {
 572                                         if (first_bad < this_sector)
 573                                                 /* Cannot use this */
 574                                                 continue;
 575                                         best_good_sectors = first_bad - this_sector;
 576                                 } else
 577                                         best_good_sectors = sectors;
 578                                 best_dist_disk = disk;
 579                                 best_pending_disk = disk;
 580                         }
 581                         continue;
 582                 }
 583                 /* This is a reasonable device to use.  It might
 584                  * even be best.
 585                  */
 586                 if (is_badblock(rdev, this_sector, sectors,
 587                                 &first_bad, &bad_sectors)) {
 588                         if (best_dist < MaxSector)
 589                                 /* already have a better device */
 590                                 continue;
 591                         if (first_bad <= this_sector) {
 592                                 /* cannot read here. If this is the 'primary'
 593                                  * device, then we must not read beyond
 594                                  * bad_sectors from another device..
 595                                  */
 596                                 bad_sectors -= (this_sector - first_bad);
 597                                 if (choose_first && sectors > bad_sectors)
 598                                         sectors = bad_sectors;
 599                                 if (best_good_sectors > sectors)
 600                                         best_good_sectors = sectors;
 601
 602                         } else {
 603                                 sector_t good_sectors = first_bad - this_sector;
 604                                 if (good_sectors > best_good_sectors) {
 605                                         best_good_sectors = good_sectors;
 606                                         best_disk = disk;
 607                                 }
 608                                 if (choose_first)
 609                                         break;
 610                         }
 611                         continue;
 612                 } else
 613                         best_good_sectors = sectors;
 614
 615                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 616                 has_nonrot_disk |= nonrot;
 617                 pending = atomic_read(&rdev->nr_pending);
 618                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 619                 if (choose_first) {
 620                         best_disk = disk;
 621                         break;
 622                 }
 623                 /* Don't change to another disk for sequential reads */
 624                 if (conf->mirrors[disk].next_seq_sect == this_sector
 625                     || dist == 0) {
 626                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 627                         struct raid1_info *mirror = &conf->mirrors[disk];
 628
 629                         best_disk = disk;
 630                         /*
 631                          * If buffered sequential IO size exceeds optimal
 632                          * iosize, check if there is idle disk. If yes, choose
 633                          * the idle disk. read_balance could already choose an
 634                          * idle disk before noticing it's a sequential IO in
 635                          * this disk. This doesn't matter because this disk
 636                          * will idle, next time it will be utilized after the
 637                          * first disk has IO size exceeds optimal iosize. In
 638                          * this way, iosize of the first disk will be optimal
 639                          * iosize at least. iosize of the second disk might be
 640                          * small, but not a big deal since when the second disk
 641                          * starts IO, the first disk is likely still busy.
 642                          */
 643                         if (nonrot && opt_iosize > 0 &&
 644                             mirror->seq_start != MaxSector &&
 645                             mirror->next_seq_sect > opt_iosize &&
 646                             mirror->next_seq_sect - opt_iosize >=
 647                             mirror->seq_start) {
 648                                 choose_next_idle = 1;
 649                                 continue;
 650                         }
 651                         break;
 652                 }
 653                 /* If device is idle, use it */
 654                 if (pending == 0) {
 655                         best_disk = disk;
 656                         break;
 657                 }
 658
 659                 if (choose_next_idle)
 660                         continue;
 661
 662                 if (min_pending > pending) {
 663                         min_pending = pending;
 664                         best_pending_disk = disk;
 665                 }
 666
 667                 if (dist < best_dist) {
 668                         best_dist = dist;
 669                         best_dist_disk = disk;
 670                 }
 671         }
 672
 673         /*
 674          * If all disks are rotational, choose the closest disk. If any disk is
 675          * non-rotational, choose the disk with less pending request even the
 676          * disk is rotational, which might/might not be optimal for raids with
 677          * mixed ratation/non-rotational disks depending on workload.
 678          */
 679         if (best_disk == -1) {
 680                 if (has_nonrot_disk)
 681                         best_disk = best_pending_disk;
 682                 else
 683                         best_disk = best_dist_disk;
 684         }
 685
 686         if (best_disk >= 0) {
 687                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 688                 if (!rdev)
 689                         goto retry;
 690                 atomic_inc(&rdev->nr_pending);
 691                 if (test_bit(Faulty, &rdev->flags)) {
 692                         /* cannot risk returning a device that failed
 693                          * before we inc'ed nr_pending
 694                          */
 695                         rdev_dec_pending(rdev, conf->mddev);
 696                         goto retry;
 697                 }
 698                 sectors = best_good_sectors;
 699
 700                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 701                         conf->mirrors[best_disk].seq_start = this_sector;
 702
 703                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 704         }
 705         rcu_read_unlock();
 706         *max_sectors = sectors;
 707
 708         return best_disk;
 709 }
 710
 711 static int raid1_mergeable_bvec(struct mddev *mddev,
 712                                 struct bvec_merge_data *bvm,
 713                                 struct bio_vec *biovec)
 714 {
 715         struct r1conf *conf = mddev->private;
 716         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 717         int max = biovec->bv_len;
 718
 719         if (mddev->merge_check_needed) {
 720                 int disk;
 721                 rcu_read_lock();
 722                 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
 723                         struct md_rdev *rdev = rcu_dereference(
 724                                 conf->mirrors[disk].rdev);
 725                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 726                                 struct request_queue *q =
 727                                         bdev_get_queue(rdev->bdev);
 728                                 if (q->merge_bvec_fn) {
 729                                         bvm->bi_sector = sector +
 730                                                 rdev->data_offset;
 731                                         bvm->bi_bdev = rdev->bdev;
 732                                         max = min(max, q->merge_bvec_fn(
 733                                                           q, bvm, biovec));
 734                                 }
 735                         }
 736                 }
 737                 rcu_read_unlock();
 738         }
 739         return max;
 740
 741 }
 742
 743 static int raid1_congested(struct mddev *mddev, int bits)
 744 {
 745         struct r1conf *conf = mddev->private;
 746         int i, ret = 0;
 747
 748         if ((bits & (1 << BDI_async_congested)) &&
 749             conf->pending_count >= max_queued_requests)
 750                 return 1;
 751
 752         rcu_read_lock();
 753         for (i = 0; i < conf->raid_disks * 2; i++) {
 754                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 755                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 756                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 757
 758                         BUG_ON(!q);
 759
 760                         /* Note the '|| 1' - when read_balance prefers
 761                          * non-congested targets, it can be removed
 762                          */
 763                         if ((bits & (1<<BDI_async_congested)) || 1)
 764                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 765                         else
 766                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 767                 }
 768         }
 769         rcu_read_unlock();
 770         return ret;
 771 }
 772
 773 static void flush_pending_writes(struct r1conf *conf)
 774 {
 775         /* Any writes that have been queued but are awaiting
 776          * bitmap updates get flushed here.
 777          */
 778         spin_lock_irq(&conf->device_lock);
 779
 780         if (conf->pending_bio_list.head) {
 781                 struct bio *bio;
 782                 bio = bio_list_get(&conf->pending_bio_list);
 783                 conf->pending_count = 0;
 784                 spin_unlock_irq(&conf->device_lock);
 785                 /* flush any pending bitmap writes to
 786                  * disk before proceeding w/ I/O */
 787                 bitmap_unplug(conf->mddev->bitmap);
 788                 wake_up(&conf->wait_barrier);
 789
 790                 while (bio) { /* submit pending writes */
 791                         struct bio *next = bio->bi_next;
 792                         bio->bi_next = NULL;
 793                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 794                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 795                                 /* Just ignore it */
 796                                 bio_endio(bio, 0);
 797                         else
 798                                 generic_make_request(bio);
 799                         bio = next;
 800                 }
 801         } else
 802                 spin_unlock_irq(&conf->device_lock);
 803 }
 804
 805 /* Barriers....
 806  * Sometimes we need to suspend IO while we do something else,
 807  * either some resync/recovery, or reconfigure the array.
 808  * To do this we raise a 'barrier'.
 809  * The 'barrier' is a counter that can be raised multiple times
 810  * to count how many activities are happening which preclude
 811  * normal IO.
 812  * We can only raise the barrier if there is no pending IO.
 813  * i.e. if nr_pending == 0.
 814  * We choose only to raise the barrier if no-one is waiting for the
 815  * barrier to go down.  This means that as soon as an IO request
 816  * is ready, no other operations which require a barrier will start
 817  * until the IO request has had a chance.
 818  *
 819  * So: regular IO calls 'wait_barrier'.  When that returns there
 820  *    is no backgroup IO happening,  It must arrange to call
 821  *    allow_barrier when it has finished its IO.
 822  * backgroup IO calls must call raise_barrier.  Once that returns
 823  *    there is no normal IO happeing.  It must arrange to call
 824  *    lower_barrier when the particular background IO completes.
 825  */
 826 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
 827 {
 828         spin_lock_irq(&conf->resync_lock);
 829
 830         /* Wait until no block IO is waiting */
 831         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 832                             conf->resync_lock);
 833
 834         /* block any new IO from starting */
 835         conf->barrier++;
 836         conf->next_resync = sector_nr;
 837
 838         /* For these conditions we must wait:
 839          * A: while the array is in frozen state
 840          * B: while barrier >= RESYNC_DEPTH, meaning resync reach
 841          *    the max count which allowed.
 842          * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
 843          *    next resync will reach to the window which normal bios are
 844          *    handling.
 845          * D: while there are any active requests in the current window.
 846          */
 847         wait_event_lock_irq(conf->wait_barrier,
 848                             !conf->array_frozen &&
 849                             conf->barrier < RESYNC_DEPTH &&
 850                             conf->current_window_requests == 0 &&
 851                             (conf->start_next_window >=
 852                              conf->next_resync + RESYNC_SECTORS),
 853                             conf->resync_lock);
 854
 855         conf->nr_pending++;
 856         spin_unlock_irq(&conf->resync_lock);
 857 }
 858
 859 static void lower_barrier(struct r1conf *conf)
 860 {
 861         unsigned long flags;
 862         BUG_ON(conf->barrier <= 0);
 863         spin_lock_irqsave(&conf->resync_lock, flags);
 864         conf->barrier--;
 865         conf->nr_pending--;
 866         spin_unlock_irqrestore(&conf->resync_lock, flags);
 867         wake_up(&conf->wait_barrier);
 868 }
 869
 870 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio)
 871 {
 872         bool wait = false;
 873
 874         if (conf->array_frozen || !bio)
 875                 wait = true;
 876         else if (conf->barrier && bio_data_dir(bio) == WRITE) {
 877                 if ((conf->mddev->curr_resync_completed
 878                      >= bio_end_sector(bio)) ||
 879                     (conf->next_resync + NEXT_NORMALIO_DISTANCE
 880                      <= bio->bi_iter.bi_sector))
 881                         wait = false;
 882                 else
 883                         wait = true;
 884         }
 885
 886         return wait;
 887 }
 888
 889 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
 890 {
 891         sector_t sector = 0;
 892
 893         spin_lock_irq(&conf->resync_lock);
 894         if (need_to_wait_for_sync(conf, bio)) {
 895                 conf->nr_waiting++;
 896                 /* Wait for the barrier to drop.
 897                  * However if there are already pending
 898                  * requests (preventing the barrier from
 899                  * rising completely), and the
 900                  * per-process bio queue isn't empty,
 901                  * then don't wait, as we need to empty
 902                  * that queue to allow conf->start_next_window
 903                  * to increase.
 904                  */
 905                 wait_event_lock_irq(conf->wait_barrier,
 906                                     !conf->array_frozen &&
 907                                     (!conf->barrier ||
 908                                      ((conf->start_next_window <
 909                                        conf->next_resync + RESYNC_SECTORS) &&
 910                                       current->bio_list &&
 911                                       !bio_list_empty(current->bio_list))),
 912                                     conf->resync_lock);
 913                 conf->nr_waiting--;
 914         }
 915
 916         if (bio && bio_data_dir(bio) == WRITE) {
 917                 if (bio->bi_iter.bi_sector >=
 918                     conf->mddev->curr_resync_completed) {
 919                         if (conf->start_next_window == MaxSector)
 920                                 conf->start_next_window =
 921                                         conf->next_resync +
 922                                         NEXT_NORMALIO_DISTANCE;
 923
 924                         if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
 925                             <= bio->bi_iter.bi_sector)
 926                                 conf->next_window_requests++;
 927                         else
 928                                 conf->current_window_requests++;
 929                         sector = conf->start_next_window;
 930                 }
 931         }
 932
 933         conf->nr_pending++;
 934         spin_unlock_irq(&conf->resync_lock);
 935         return sector;
 936 }
 937
 938 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
 939                           sector_t bi_sector)
 940 {
 941         unsigned long flags;
 942
 943         spin_lock_irqsave(&conf->resync_lock, flags);
 944         conf->nr_pending--;
 945         if (start_next_window) {
 946                 if (start_next_window == conf->start_next_window) {
 947                         if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
 948                             <= bi_sector)
 949                                 conf->next_window_requests--;
 950                         else
 951                                 conf->current_window_requests--;
 952                 } else
 953                         conf->current_window_requests--;
 954
 955                 if (!conf->current_window_requests) {
 956                         if (conf->next_window_requests) {
 957                                 conf->current_window_requests =
 958                                         conf->next_window_requests;
 959                                 conf->next_window_requests = 0;
 960                                 conf->start_next_window +=
 961                                         NEXT_NORMALIO_DISTANCE;
 962                         } else
 963                                 conf->start_next_window = MaxSector;
 964                 }
 965         }
 966         spin_unlock_irqrestore(&conf->resync_lock, flags);
 967         wake_up(&conf->wait_barrier);
 968 }
 969
 970 static void freeze_array(struct r1conf *conf, int extra)
 971 {
 972         /* stop syncio and normal IO and wait for everything to
 973          * go quite.
 974          * We wait until nr_pending match nr_queued+extra
 975          * This is called in the context of one normal IO request
 976          * that has failed. Thus any sync request that might be pending
 977          * will be blocked by nr_pending, and we need to wait for
 978          * pending IO requests to complete or be queued for re-try.
 979          * Thus the number queued (nr_queued) plus this request (extra)
 980          * must match the number of pending IOs (nr_pending) before
 981          * we continue.
 982          */
 983         spin_lock_irq(&conf->resync_lock);
 984         conf->array_frozen = 1;
 985         wait_event_lock_irq_cmd(conf->wait_barrier,
 986                                 conf->nr_pending == conf->nr_queued+extra,
 987                                 conf->resync_lock,
 988                                 flush_pending_writes(conf));
 989         spin_unlock_irq(&conf->resync_lock);
 990 }
 991 static void unfreeze_array(struct r1conf *conf)
 992 {
 993         /* reverse the effect of the freeze */
 994         spin_lock_irq(&conf->resync_lock);
 995         conf->array_frozen = 0;
 996         wake_up(&conf->wait_barrier);
 997         spin_unlock_irq(&conf->resync_lock);
 998 }
 999
1000 /* duplicate the data pages for behind I/O
1001  */
1002 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
1003 {
1004         int i;
1005         struct bio_vec *bvec;
1006         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
1007                                         GFP_NOIO);
1008         if (unlikely(!bvecs))
1009                 return;
1010
1011         bio_for_each_segment_all(bvec, bio, i) {
1012                 bvecs[i] = *bvec;
1013                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
1014                 if (unlikely(!bvecs[i].bv_page))
1015                         goto do_sync_io;
1016                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
1017                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
1018                 kunmap(bvecs[i].bv_page);
1019                 kunmap(bvec->bv_page);
1020         }
1021         r1_bio->behind_bvecs = bvecs;
1022         r1_bio->behind_page_count = bio->bi_vcnt;
1023         set_bit(R1BIO_BehindIO, &r1_bio->state);
1024         return;
1025
1026 do_sync_io:
1027         for (i = 0; i < bio->bi_vcnt; i++)
1028                 if (bvecs[i].bv_page)
1029                         put_page(bvecs[i].bv_page);
1030         kfree(bvecs);
1031         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1032                  bio->bi_iter.bi_size);
1033 }
1034
1035 struct raid1_plug_cb {
1036         struct blk_plug_cb      cb;
1037         struct bio_list         pending;
1038         int                     pending_cnt;
1039 };
1040
1041 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1042 {
1043         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1044                                                   cb);
1045         struct mddev *mddev = plug->cb.data;
1046         struct r1conf *conf = mddev->private;
1047         struct bio *bio;
1048
1049         if (from_schedule || current->bio_list) {
1050                 spin_lock_irq(&conf->device_lock);
1051                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1052                 conf->pending_count += plug->pending_cnt;
1053                 spin_unlock_irq(&conf->device_lock);
1054                 wake_up(&conf->wait_barrier);
1055                 md_wakeup_thread(mddev->thread);
1056                 kfree(plug);
1057                 return;
1058         }
1059
1060         /* we aren't scheduling, so we can do the write-out directly. */
1061         bio = bio_list_get(&plug->pending);
1062         bitmap_unplug(mddev->bitmap);
1063         wake_up(&conf->wait_barrier);
1064
1065         while (bio) { /* submit pending writes */
1066                 struct bio *next = bio->bi_next;
1067                 bio->bi_next = NULL;
1068                 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1069                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1070                         /* Just ignore it */
1071                         bio_endio(bio, 0);
1072                 else
1073                         generic_make_request(bio);
1074                 bio = next;
1075         }
1076         kfree(plug);
1077 }
1078
1079 static void make_request(struct mddev *mddev, struct bio * bio)
1080 {
1081         struct r1conf *conf = mddev->private;
1082         struct raid1_info *mirror;
1083         struct r1bio *r1_bio;
1084         struct bio *read_bio;
1085         int i, disks;
1086         struct bitmap *bitmap;
1087         unsigned long flags;
1088         const int rw = bio_data_dir(bio);
1089         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1090         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1091         const unsigned long do_discard = (bio->bi_rw
1092                                           & (REQ_DISCARD | REQ_SECURE));
1093         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1094         struct md_rdev *blocked_rdev;
1095         struct blk_plug_cb *cb;
1096         struct raid1_plug_cb *plug = NULL;
1097         int first_clone;
1098         int sectors_handled;
1099         int max_sectors;
1100         sector_t start_next_window;
1101
1102         /*
1103          * Register the new request and wait if the reconstruction
1104          * thread has put up a bar for new requests.
1105          * Continue immediately if no resync is active currently.
1106          */
1107
1108         md_write_start(mddev, bio); /* wait on superblock update early */
1109
1110         if (bio_data_dir(bio) == WRITE &&
1111             ((bio_end_sector(bio) > mddev->suspend_lo &&
1112             bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1113             (mddev_is_clustered(mddev) &&
1114              md_cluster_ops->area_resyncing(mddev, bio->bi_iter.bi_sector, bio_end_sector(bio))))) {
1115                 /* As the suspend_* range is controlled by
1116                  * userspace, we want an interruptible
1117                  * wait.
1118                  */
1119                 DEFINE_WAIT(w);
1120                 for (;;) {
1121                         flush_signals(current);
1122                         prepare_to_wait(&conf->wait_barrier,
1123                                         &w, TASK_INTERRUPTIBLE);
1124                         if (bio_end_sector(bio) <= mddev->suspend_lo ||
1125                             bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1126                             (mddev_is_clustered(mddev) &&
1127                              !md_cluster_ops->area_resyncing(mddev,
1128                                      bio->bi_iter.bi_sector, bio_end_sector(bio))))
1129                                 break;
1130                         schedule();
1131                 }
1132                 finish_wait(&conf->wait_barrier, &w);
1133         }
1134
1135         start_next_window = wait_barrier(conf, bio);
1136
1137         bitmap = mddev->bitmap;
1138
1139         /*
1140          * make_request() can abort the operation when READA is being
1141          * used and no empty request is available.
1142          *
1143          */
1144         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1145
1146         r1_bio->master_bio = bio;
1147         r1_bio->sectors = bio_sectors(bio);
1148         r1_bio->state = 0;
1149         r1_bio->mddev = mddev;
1150         r1_bio->sector = bio->bi_iter.bi_sector;
1151
1152         /* We might need to issue multiple reads to different
1153          * devices if there are bad blocks around, so we keep
1154          * track of the number of reads in bio->bi_phys_segments.
1155          * If this is 0, there is only one r1_bio and no locking
1156          * will be needed when requests complete.  If it is
1157          * non-zero, then it is the number of not-completed requests.
1158          */
1159         bio->bi_phys_segments = 0;
1160         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1161
1162         if (rw == READ) {
1163                 /*
1164                  * read balancing logic:
1165                  */
1166                 int rdisk;
1167
1168 read_again:
1169                 rdisk = read_balance(conf, r1_bio, &max_sectors);
1170
1171                 if (rdisk < 0) {
1172                         /* couldn't find anywhere to read from */
1173                         raid_end_bio_io(r1_bio);
1174                         return;
1175                 }
1176                 mirror = conf->mirrors + rdisk;
1177
1178                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1179                     bitmap) {
1180                         /* Reading from a write-mostly device must
1181                          * take care not to over-take any writes
1182                          * that are 'behind'
1183                          */
1184                         wait_event(bitmap->behind_wait,
1185                                    atomic_read(&bitmap->behind_writes) == 0);
1186                 }
1187                 r1_bio->read_disk = rdisk;
1188                 r1_bio->start_next_window = 0;
1189
1190                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1191                 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1192                          max_sectors);
1193
1194                 r1_bio->bios[rdisk] = read_bio;
1195
1196                 read_bio->bi_iter.bi_sector = r1_bio->sector +
1197                         mirror->rdev->data_offset;
1198                 read_bio->bi_bdev = mirror->rdev->bdev;
1199                 read_bio->bi_end_io = raid1_end_read_request;
1200                 read_bio->bi_rw = READ | do_sync;
1201                 read_bio->bi_private = r1_bio;
1202
1203                 if (max_sectors < r1_bio->sectors) {
1204                         /* could not read all from this device, so we will
1205                          * need another r1_bio.
1206                          */
1207
1208                         sectors_handled = (r1_bio->sector + max_sectors
1209                                            - bio->bi_iter.bi_sector);
1210                         r1_bio->sectors = max_sectors;
1211                         spin_lock_irq(&conf->device_lock);
1212                         if (bio->bi_phys_segments == 0)
1213                                 bio->bi_phys_segments = 2;
1214                         else
1215                                 bio->bi_phys_segments++;
1216                         spin_unlock_irq(&conf->device_lock);
1217                         /* Cannot call generic_make_request directly
1218                          * as that will be queued in __make_request
1219                          * and subsequent mempool_alloc might block waiting
1220                          * for it.  So hand bio over to raid1d.
1221                          */
1222                         reschedule_retry(r1_bio);
1223
1224                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1225
1226                         r1_bio->master_bio = bio;
1227                         r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1228                         r1_bio->state = 0;
1229                         r1_bio->mddev = mddev;
1230                         r1_bio->sector = bio->bi_iter.bi_sector +
1231                                 sectors_handled;
1232                         goto read_again;
1233                 } else
1234                         generic_make_request(read_bio);
1235                 return;
1236         }
1237
1238         /*
1239          * WRITE:
1240          */
1241         if (conf->pending_count >= max_queued_requests) {
1242                 md_wakeup_thread(mddev->thread);
1243                 wait_event(conf->wait_barrier,
1244                            conf->pending_count < max_queued_requests);
1245         }
1246         /* first select target devices under rcu_lock and
1247          * inc refcount on their rdev.  Record them by setting
1248          * bios[x] to bio
1249          * If there are known/acknowledged bad blocks on any device on
1250          * which we have seen a write error, we want to avoid writing those
1251          * blocks.
1252          * This potentially requires several writes to write around
1253          * the bad blocks.  Each set of writes gets it's own r1bio
1254          * with a set of bios attached.
1255          */
1256
1257         disks = conf->raid_disks * 2;
1258  retry_write:
1259         r1_bio->start_next_window = start_next_window;
1260         blocked_rdev = NULL;
1261         rcu_read_lock();
1262         max_sectors = r1_bio->sectors;
1263         for (i = 0;  i < disks; i++) {
1264                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1265                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1266                         atomic_inc(&rdev->nr_pending);
1267                         blocked_rdev = rdev;
1268                         break;
1269                 }
1270                 r1_bio->bios[i] = NULL;
1271                 if (!rdev || test_bit(Faulty, &rdev->flags)
1272                     || test_bit(Unmerged, &rdev->flags)) {
1273                         if (i < conf->raid_disks)
1274                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1275                         continue;
1276                 }
1277
1278                 atomic_inc(&rdev->nr_pending);
1279                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1280                         sector_t first_bad;
1281                         int bad_sectors;
1282                         int is_bad;
1283
1284                         is_bad = is_badblock(rdev, r1_bio->sector,
1285                                              max_sectors,
1286                                              &first_bad, &bad_sectors);
1287                         if (is_bad < 0) {
1288                                 /* mustn't write here until the bad block is
1289                                  * acknowledged*/
1290                                 set_bit(BlockedBadBlocks, &rdev->flags);
1291                                 blocked_rdev = rdev;
1292                                 break;
1293                         }
1294                         if (is_bad && first_bad <= r1_bio->sector) {
1295                                 /* Cannot write here at all */
1296                                 bad_sectors -= (r1_bio->sector - first_bad);
1297                                 if (bad_sectors < max_sectors)
1298                                         /* mustn't write more than bad_sectors
1299                                          * to other devices yet
1300                                          */
1301                                         max_sectors = bad_sectors;
1302                                 rdev_dec_pending(rdev, mddev);
1303                                 /* We don't set R1BIO_Degraded as that
1304                                  * only applies if the disk is
1305                                  * missing, so it might be re-added,
1306                                  * and we want to know to recover this
1307                                  * chunk.
1308                                  * In this case the device is here,
1309                                  * and the fact that this chunk is not
1310                                  * in-sync is recorded in the bad
1311                                  * block log
1312                                  */
1313                                 continue;
1314                         }
1315                         if (is_bad) {
1316                                 int good_sectors = first_bad - r1_bio->sector;
1317                                 if (good_sectors < max_sectors)
1318                                         max_sectors = good_sectors;
1319                         }
1320                 }
1321                 r1_bio->bios[i] = bio;
1322         }
1323         rcu_read_unlock();
1324
1325         if (unlikely(blocked_rdev)) {
1326                 /* Wait for this device to become unblocked */
1327                 int j;
1328                 sector_t old = start_next_window;
1329
1330                 for (j = 0; j < i; j++)
1331                         if (r1_bio->bios[j])
1332                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1333                 r1_bio->state = 0;
1334                 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector);
1335                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1336                 start_next_window = wait_barrier(conf, bio);
1337                 /*
1338                  * We must make sure the multi r1bios of bio have
1339                  * the same value of bi_phys_segments
1340                  */
1341                 if (bio->bi_phys_segments && old &&
1342                     old != start_next_window)
1343                         /* Wait for the former r1bio(s) to complete */
1344                         wait_event(conf->wait_barrier,
1345                                    bio->bi_phys_segments == 1);
1346                 goto retry_write;
1347         }
1348
1349         if (max_sectors < r1_bio->sectors) {
1350                 /* We are splitting this write into multiple parts, so
1351                  * we need to prepare for allocating another r1_bio.
1352                  */
1353                 r1_bio->sectors = max_sectors;
1354                 spin_lock_irq(&conf->device_lock);
1355                 if (bio->bi_phys_segments == 0)
1356                         bio->bi_phys_segments = 2;
1357                 else
1358                         bio->bi_phys_segments++;
1359                 spin_unlock_irq(&conf->device_lock);
1360         }
1361         sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1362
1363         atomic_set(&r1_bio->remaining, 1);
1364         atomic_set(&r1_bio->behind_remaining, 0);
1365
1366         first_clone = 1;
1367         for (i = 0; i < disks; i++) {
1368                 struct bio *mbio;
1369                 if (!r1_bio->bios[i])
1370                         continue;
1371
1372                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1373                 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors);
1374
1375                 if (first_clone) {
1376                         /* do behind I/O ?
1377                          * Not if there are too many, or cannot
1378                          * allocate memory, or a reader on WriteMostly
1379                          * is waiting for behind writes to flush */
1380                         if (bitmap &&
1381                             (atomic_read(&bitmap->behind_writes)
1382                              < mddev->bitmap_info.max_write_behind) &&
1383                             !waitqueue_active(&bitmap->behind_wait))
1384                                 alloc_behind_pages(mbio, r1_bio);
1385
1386                         bitmap_startwrite(bitmap, r1_bio->sector,
1387                                           r1_bio->sectors,
1388                                           test_bit(R1BIO_BehindIO,
1389                                                    &r1_bio->state));
1390                         first_clone = 0;
1391                 }
1392                 if (r1_bio->behind_bvecs) {
1393                         struct bio_vec *bvec;
1394                         int j;
1395
1396                         /*
1397                          * We trimmed the bio, so _all is legit
1398                          */
1399                         bio_for_each_segment_all(bvec, mbio, j)
1400                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1401                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1402                                 atomic_inc(&r1_bio->behind_remaining);
1403                 }
1404
1405                 r1_bio->bios[i] = mbio;
1406
1407                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1408                                    conf->mirrors[i].rdev->data_offset);
1409                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1410                 mbio->bi_end_io = raid1_end_write_request;
1411                 mbio->bi_rw =
1412                         WRITE | do_flush_fua | do_sync | do_discard | do_same;
1413                 mbio->bi_private = r1_bio;
1414
1415                 atomic_inc(&r1_bio->remaining);
1416
1417                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1418                 if (cb)
1419                         plug = container_of(cb, struct raid1_plug_cb, cb);
1420                 else
1421                         plug = NULL;
1422                 spin_lock_irqsave(&conf->device_lock, flags);
1423                 if (plug) {
1424                         bio_list_add(&plug->pending, mbio);
1425                         plug->pending_cnt++;
1426                 } else {
1427                         bio_list_add(&conf->pending_bio_list, mbio);
1428                         conf->pending_count++;
1429                 }
1430                 spin_unlock_irqrestore(&conf->device_lock, flags);
1431                 if (!plug)
1432                         md_wakeup_thread(mddev->thread);
1433         }
1434         /* Mustn't call r1_bio_write_done before this next test,
1435          * as it could result in the bio being freed.
1436          */
1437         if (sectors_handled < bio_sectors(bio)) {
1438                 r1_bio_write_done(r1_bio);
1439                 /* We need another r1_bio.  It has already been counted
1440                  * in bio->bi_phys_segments
1441                  */
1442                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1443                 r1_bio->master_bio = bio;
1444                 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1445                 r1_bio->state = 0;
1446                 r1_bio->mddev = mddev;
1447                 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1448                 goto retry_write;
1449         }
1450
1451         r1_bio_write_done(r1_bio);
1452
1453         /* In case raid1d snuck in to freeze_array */
1454         wake_up(&conf->wait_barrier);
1455 }
1456
1457 static void status(struct seq_file *seq, struct mddev *mddev)
1458 {
1459         struct r1conf *conf = mddev->private;
1460         int i;
1461
1462         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1463                    conf->raid_disks - mddev->degraded);
1464         rcu_read_lock();
1465         for (i = 0; i < conf->raid_disks; i++) {
1466                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1467                 seq_printf(seq, "%s",
1468                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1469         }
1470         rcu_read_unlock();
1471         seq_printf(seq, "]");
1472 }
1473
1474 static void error(struct mddev *mddev, struct md_rdev *rdev)
1475 {
1476         char b[BDEVNAME_SIZE];
1477         struct r1conf *conf = mddev->private;
1478         unsigned long flags;
1479
1480         /*
1481          * If it is not operational, then we have already marked it as dead
1482          * else if it is the last working disks, ignore the error, let the
1483          * next level up know.
1484          * else mark the drive as failed
1485          */
1486         if (test_bit(In_sync, &rdev->flags)
1487             && (conf->raid_disks - mddev->degraded) == 1) {
1488                 /*
1489                  * Don't fail the drive, act as though we were just a
1490                  * normal single drive.
1491                  * However don't try a recovery from this drive as
1492                  * it is very likely to fail.
1493                  */
1494                 conf->recovery_disabled = mddev->recovery_disabled;
1495                 return;
1496         }
1497         set_bit(Blocked, &rdev->flags);
1498         spin_lock_irqsave(&conf->device_lock, flags);
1499         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1500                 mddev->degraded++;
1501                 set_bit(Faulty, &rdev->flags);
1502         } else
1503                 set_bit(Faulty, &rdev->flags);
1504         spin_unlock_irqrestore(&conf->device_lock, flags);
1505         /*
1506          * if recovery is running, make sure it aborts.
1507          */
1508         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1509         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1510         printk(KERN_ALERT
1511                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1512                "md/raid1:%s: Operation continuing on %d devices.\n",
1513                mdname(mddev), bdevname(rdev->bdev, b),
1514                mdname(mddev), conf->raid_disks - mddev->degraded);
1515 }
1516
1517 static void print_conf(struct r1conf *conf)
1518 {
1519         int i;
1520
1521         printk(KERN_DEBUG "RAID1 conf printout:\n");
1522         if (!conf) {
1523                 printk(KERN_DEBUG "(!conf)\n");
1524                 return;
1525         }
1526         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1527                 conf->raid_disks);
1528
1529         rcu_read_lock();
1530         for (i = 0; i < conf->raid_disks; i++) {
1531                 char b[BDEVNAME_SIZE];
1532                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1533                 if (rdev)
1534                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1535                                i, !test_bit(In_sync, &rdev->flags),
1536                                !test_bit(Faulty, &rdev->flags),
1537                                bdevname(rdev->bdev,b));
1538         }
1539         rcu_read_unlock();
1540 }
1541
1542 static void close_sync(struct r1conf *conf)
1543 {
1544         wait_barrier(conf, NULL);
1545         allow_barrier(conf, 0, 0);
1546
1547         mempool_destroy(conf->r1buf_pool);
1548         conf->r1buf_pool = NULL;
1549
1550         spin_lock_irq(&conf->resync_lock);
1551         conf->next_resync = 0;
1552         conf->start_next_window = MaxSector;
1553         conf->current_window_requests +=
1554                 conf->next_window_requests;
1555         conf->next_window_requests = 0;
1556         spin_unlock_irq(&conf->resync_lock);
1557 }
1558
1559 static int raid1_spare_active(struct mddev *mddev)
1560 {
1561         int i;
1562         struct r1conf *conf = mddev->private;
1563         int count = 0;
1564         unsigned long flags;
1565
1566         /*
1567          * Find all failed disks within the RAID1 configuration
1568          * and mark them readable.
1569          * Called under mddev lock, so rcu protection not needed.
1570          * device_lock used to avoid races with raid1_end_read_request
1571          * which expects 'In_sync' flags and ->degraded to be consistent.
1572          */
1573         spin_lock_irqsave(&conf->device_lock, flags);
1574         for (i = 0; i < conf->raid_disks; i++) {
1575                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1576                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1577                 if (repl
1578                     && !test_bit(Candidate, &repl->flags)
1579                     && repl->recovery_offset == MaxSector
1580                     && !test_bit(Faulty, &repl->flags)
1581                     && !test_and_set_bit(In_sync, &repl->flags)) {
1582                         /* replacement has just become active */
1583                         if (!rdev ||
1584                             !test_and_clear_bit(In_sync, &rdev->flags))
1585                                 count++;
1586                         if (rdev) {
1587                                 /* Replaced device not technically
1588                                  * faulty, but we need to be sure
1589                                  * it gets removed and never re-added
1590                                  */
1591                                 set_bit(Faulty, &rdev->flags);
1592                                 sysfs_notify_dirent_safe(
1593                                         rdev->sysfs_state);
1594                         }
1595                 }
1596                 if (rdev
1597                     && rdev->recovery_offset == MaxSector
1598                     && !test_bit(Faulty, &rdev->flags)
1599                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1600                         count++;
1601                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1602                 }
1603         }
1604         mddev->degraded -= count;
1605         spin_unlock_irqrestore(&conf->device_lock, flags);
1606
1607         print_conf(conf);
1608         return count;
1609 }
1610
1611 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1612 {
1613         struct r1conf *conf = mddev->private;
1614         int err = -EEXIST;
1615         int mirror = 0;
1616         struct raid1_info *p;
1617         int first = 0;
1618         int last = conf->raid_disks - 1;
1619         struct request_queue *q = bdev_get_queue(rdev->bdev);
1620
1621         if (mddev->recovery_disabled == conf->recovery_disabled)
1622                 return -EBUSY;
1623
1624         if (rdev->raid_disk >= 0)
1625                 first = last = rdev->raid_disk;
1626
1627         if (q->merge_bvec_fn) {
1628                 set_bit(Unmerged, &rdev->flags);
1629                 mddev->merge_check_needed = 1;
1630         }
1631
1632         for (mirror = first; mirror <= last; mirror++) {
1633                 p = conf->mirrors+mirror;
1634                 if (!p->rdev) {
1635
1636                         if (mddev->gendisk)
1637                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1638                                                   rdev->data_offset << 9);
1639
1640                         p->head_position = 0;
1641                         rdev->raid_disk = mirror;
1642                         err = 0;
1643                         /* As all devices are equivalent, we don't need a full recovery
1644                          * if this was recently any drive of the array
1645                          */
1646                         if (rdev->saved_raid_disk < 0)
1647                                 conf->fullsync = 1;
1648                         rcu_assign_pointer(p->rdev, rdev);
1649                         break;
1650                 }
1651                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1652                     p[conf->raid_disks].rdev == NULL) {
1653                         /* Add this device as a replacement */
1654                         clear_bit(In_sync, &rdev->flags);
1655                         set_bit(Replacement, &rdev->flags);
1656                         rdev->raid_disk = mirror;
1657                         err = 0;
1658                         conf->fullsync = 1;
1659                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1660                         break;
1661                 }
1662         }
1663         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1664                 /* Some requests might not have seen this new
1665                  * merge_bvec_fn.  We must wait for them to complete
1666                  * before merging the device fully.
1667                  * First we make sure any code which has tested
1668                  * our function has submitted the request, then
1669                  * we wait for all outstanding requests to complete.
1670                  */
1671                 synchronize_sched();
1672                 freeze_array(conf, 0);
1673                 unfreeze_array(conf);
1674                 clear_bit(Unmerged, &rdev->flags);
1675         }
1676         md_integrity_add_rdev(rdev, mddev);
1677         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1678                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1679         print_conf(conf);
1680         return err;
1681 }
1682
1683 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1684 {
1685         struct r1conf *conf = mddev->private;
1686         int err = 0;
1687         int number = rdev->raid_disk;
1688         struct raid1_info *p = conf->mirrors + number;
1689
1690         if (rdev != p->rdev)
1691                 p = conf->mirrors + conf->raid_disks + number;
1692
1693         print_conf(conf);
1694         if (rdev == p->rdev) {
1695                 if (test_bit(In_sync, &rdev->flags) ||
1696                     atomic_read(&rdev->nr_pending)) {
1697                         err = -EBUSY;
1698                         goto abort;
1699                 }
1700                 /* Only remove non-faulty devices if recovery
1701                  * is not possible.
1702                  */
1703                 if (!test_bit(Faulty, &rdev->flags) &&
1704                     mddev->recovery_disabled != conf->recovery_disabled &&
1705                     mddev->degraded < conf->raid_disks) {
1706                         err = -EBUSY;
1707                         goto abort;
1708                 }
1709                 p->rdev = NULL;
1710                 synchronize_rcu();
1711                 if (atomic_read(&rdev->nr_pending)) {
1712                         /* lost the race, try later */
1713                         err = -EBUSY;
1714                         p->rdev = rdev;
1715                         goto abort;
1716                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1717                         /* We just removed a device that is being replaced.
1718                          * Move down the replacement.  We drain all IO before
1719                          * doing this to avoid confusion.
1720                          */
1721                         struct md_rdev *repl =
1722                                 conf->mirrors[conf->raid_disks + number].rdev;
1723                         freeze_array(conf, 0);
1724                         clear_bit(Replacement, &repl->flags);
1725                         p->rdev = repl;
1726                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1727                         unfreeze_array(conf);
1728                         clear_bit(WantReplacement, &rdev->flags);
1729                 } else
1730                         clear_bit(WantReplacement, &rdev->flags);
1731                 err = md_integrity_register(mddev);
1732         }
1733 abort:
1734
1735         print_conf(conf);
1736         return err;
1737 }
1738
1739 static void end_sync_read(struct bio *bio, int error)
1740 {
1741         struct r1bio *r1_bio = bio->bi_private;
1742
1743         update_head_pos(r1_bio->read_disk, r1_bio);
1744
1745         /*
1746          * we have read a block, now it needs to be re-written,
1747          * or re-read if the read failed.
1748          * We don't do much here, just schedule handling by raid1d
1749          */
1750         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1751                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1752
1753         if (atomic_dec_and_test(&r1_bio->remaining))
1754                 reschedule_retry(r1_bio);
1755 }
1756
1757 static void end_sync_write(struct bio *bio, int error)
1758 {
1759         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1760         struct r1bio *r1_bio = bio->bi_private;
1761         struct mddev *mddev = r1_bio->mddev;
1762         struct r1conf *conf = mddev->private;
1763         int mirror=0;
1764         sector_t first_bad;
1765         int bad_sectors;
1766
1767         mirror = find_bio_disk(r1_bio, bio);
1768
1769         if (!uptodate) {
1770                 sector_t sync_blocks = 0;
1771                 sector_t s = r1_bio->sector;
1772                 long sectors_to_go = r1_bio->sectors;
1773                 /* make sure these bits doesn't get cleared. */
1774                 do {
1775                         bitmap_end_sync(mddev->bitmap, s,
1776                                         &sync_blocks, 1);
1777                         s += sync_blocks;
1778                         sectors_to_go -= sync_blocks;
1779                 } while (sectors_to_go > 0);
1780                 set_bit(WriteErrorSeen,
1781                         &conf->mirrors[mirror].rdev->flags);
1782                 if (!test_and_set_bit(WantReplacement,
1783                                       &conf->mirrors[mirror].rdev->flags))
1784                         set_bit(MD_RECOVERY_NEEDED, &
1785                                 mddev->recovery);
1786                 set_bit(R1BIO_WriteError, &r1_bio->state);
1787         } else if (is_badblock(conf->mirrors[mirror].rdev,
1788                                r1_bio->sector,
1789                                r1_bio->sectors,
1790                                &first_bad, &bad_sectors) &&
1791                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1792                                 r1_bio->sector,
1793                                 r1_bio->sectors,
1794                                 &first_bad, &bad_sectors)
1795                 )
1796                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1797
1798         if (atomic_dec_and_test(&r1_bio->remaining)) {
1799                 int s = r1_bio->sectors;
1800                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1801                     test_bit(R1BIO_WriteError, &r1_bio->state))
1802                         reschedule_retry(r1_bio);
1803                 else {
1804                         put_buf(r1_bio);
1805                         md_done_sync(mddev, s, uptodate);
1806                 }
1807         }
1808 }
1809
1810 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1811                             int sectors, struct page *page, int rw)
1812 {
1813         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1814                 /* success */
1815                 return 1;
1816         if (rw == WRITE) {
1817                 set_bit(WriteErrorSeen, &rdev->flags);
1818                 if (!test_and_set_bit(WantReplacement,
1819                                       &rdev->flags))
1820                         set_bit(MD_RECOVERY_NEEDED, &
1821                                 rdev->mddev->recovery);
1822         }
1823         /* need to record an error - either for the block or the device */
1824         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1825                 md_error(rdev->mddev, rdev);
1826         return 0;
1827 }
1828
1829 static int fix_sync_read_error(struct r1bio *r1_bio)
1830 {
1831         /* Try some synchronous reads of other devices to get
1832          * good data, much like with normal read errors.  Only
1833          * read into the pages we already have so we don't
1834          * need to re-issue the read request.
1835          * We don't need to freeze the array, because being in an
1836          * active sync request, there is no normal IO, and
1837          * no overlapping syncs.
1838          * We don't need to check is_badblock() again as we
1839          * made sure that anything with a bad block in range
1840          * will have bi_end_io clear.
1841          */
1842         struct mddev *mddev = r1_bio->mddev;
1843         struct r1conf *conf = mddev->private;
1844         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1845         sector_t sect = r1_bio->sector;
1846         int sectors = r1_bio->sectors;
1847         int idx = 0;
1848
1849         while(sectors) {
1850                 int s = sectors;
1851                 int d = r1_bio->read_disk;
1852                 int success = 0;
1853                 struct md_rdev *rdev;
1854                 int start;
1855
1856                 if (s > (PAGE_SIZE>>9))
1857                         s = PAGE_SIZE >> 9;
1858                 do {
1859                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1860                                 /* No rcu protection needed here devices
1861                                  * can only be removed when no resync is
1862                                  * active, and resync is currently active
1863                                  */
1864                                 rdev = conf->mirrors[d].rdev;
1865                                 if (sync_page_io(rdev, sect, s<<9,
1866                                                  bio->bi_io_vec[idx].bv_page,
1867                                                  READ, false)) {
1868                                         success = 1;
1869                                         break;
1870                                 }
1871                         }
1872                         d++;
1873                         if (d == conf->raid_disks * 2)
1874                                 d = 0;
1875                 } while (!success && d != r1_bio->read_disk);
1876
1877                 if (!success) {
1878                         char b[BDEVNAME_SIZE];
1879                         int abort = 0;
1880                         /* Cannot read from anywhere, this block is lost.
1881                          * Record a bad block on each device.  If that doesn't
1882                          * work just disable and interrupt the recovery.
1883                          * Don't fail devices as that won't really help.
1884                          */
1885                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1886                                " for block %llu\n",
1887                                mdname(mddev),
1888                                bdevname(bio->bi_bdev, b),
1889                                (unsigned long long)r1_bio->sector);
1890                         for (d = 0; d < conf->raid_disks * 2; d++) {
1891                                 rdev = conf->mirrors[d].rdev;
1892                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1893                                         continue;
1894                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1895                                         abort = 1;
1896                         }
1897                         if (abort) {
1898                                 conf->recovery_disabled =
1899                                         mddev->recovery_disabled;
1900                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1901                                 md_done_sync(mddev, r1_bio->sectors, 0);
1902                                 put_buf(r1_bio);
1903                                 return 0;
1904                         }
1905                         /* Try next page */
1906                         sectors -= s;
1907                         sect += s;
1908                         idx++;
1909                         continue;
1910                 }
1911
1912                 start = d;
1913                 /* write it back and re-read */
1914                 while (d != r1_bio->read_disk) {
1915                         if (d == 0)
1916                                 d = conf->raid_disks * 2;
1917                         d--;
1918                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1919                                 continue;
1920                         rdev = conf->mirrors[d].rdev;
1921                         if (r1_sync_page_io(rdev, sect, s,
1922                                             bio->bi_io_vec[idx].bv_page,
1923                                             WRITE) == 0) {
1924                                 r1_bio->bios[d]->bi_end_io = NULL;
1925                                 rdev_dec_pending(rdev, mddev);
1926                         }
1927                 }
1928                 d = start;
1929                 while (d != r1_bio->read_disk) {
1930                         if (d == 0)
1931                                 d = conf->raid_disks * 2;
1932                         d--;
1933                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1934                                 continue;
1935                         rdev = conf->mirrors[d].rdev;
1936                         if (r1_sync_page_io(rdev, sect, s,
1937                                             bio->bi_io_vec[idx].bv_page,
1938                                             READ) != 0)
1939                                 atomic_add(s, &rdev->corrected_errors);
1940                 }
1941                 sectors -= s;
1942                 sect += s;
1943                 idx ++;
1944         }
1945         set_bit(R1BIO_Uptodate, &r1_bio->state);
1946         set_bit(BIO_UPTODATE, &bio->bi_flags);
1947         return 1;
1948 }
1949
1950 static void process_checks(struct r1bio *r1_bio)
1951 {
1952         /* We have read all readable devices.  If we haven't
1953          * got the block, then there is no hope left.
1954          * If we have, then we want to do a comparison
1955          * and skip the write if everything is the same.
1956          * If any blocks failed to read, then we need to
1957          * attempt an over-write
1958          */
1959         struct mddev *mddev = r1_bio->mddev;
1960         struct r1conf *conf = mddev->private;
1961         int primary;
1962         int i;
1963         int vcnt;
1964
1965         /* Fix variable parts of all bios */
1966         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1967         for (i = 0; i < conf->raid_disks * 2; i++) {
1968                 int j;
1969                 int size;
1970                 int uptodate;
1971                 struct bio *b = r1_bio->bios[i];
1972                 if (b->bi_end_io != end_sync_read)
1973                         continue;
1974                 /* fixup the bio for reuse, but preserve BIO_UPTODATE */
1975                 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags);
1976                 bio_reset(b);
1977                 if (!uptodate)
1978                         clear_bit(BIO_UPTODATE, &b->bi_flags);
1979                 b->bi_vcnt = vcnt;
1980                 b->bi_iter.bi_size = r1_bio->sectors << 9;
1981                 b->bi_iter.bi_sector = r1_bio->sector +
1982                         conf->mirrors[i].rdev->data_offset;
1983                 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1984                 b->bi_end_io = end_sync_read;
1985                 b->bi_private = r1_bio;
1986
1987                 size = b->bi_iter.bi_size;
1988                 for (j = 0; j < vcnt ; j++) {
1989                         struct bio_vec *bi;
1990                         bi = &b->bi_io_vec[j];
1991                         bi->bv_offset = 0;
1992                         if (size > PAGE_SIZE)
1993                                 bi->bv_len = PAGE_SIZE;
1994                         else
1995                                 bi->bv_len = size;
1996                         size -= PAGE_SIZE;
1997                 }
1998         }
1999         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2000                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2001                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
2002                         r1_bio->bios[primary]->bi_end_io = NULL;
2003                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2004                         break;
2005                 }
2006         r1_bio->read_disk = primary;
2007         for (i = 0; i < conf->raid_disks * 2; i++) {
2008                 int j;
2009                 struct bio *pbio = r1_bio->bios[primary];
2010                 struct bio *sbio = r1_bio->bios[i];
2011                 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags);
2012
2013                 if (sbio->bi_end_io != end_sync_read)
2014                         continue;
2015                 /* Now we can 'fixup' the BIO_UPTODATE flag */
2016                 set_bit(BIO_UPTODATE, &sbio->bi_flags);
2017
2018                 if (uptodate) {
2019                         for (j = vcnt; j-- ; ) {
2020                                 struct page *p, *s;
2021                                 p = pbio->bi_io_vec[j].bv_page;
2022                                 s = sbio->bi_io_vec[j].bv_page;
2023                                 if (memcmp(page_address(p),
2024                                            page_address(s),
2025                                            sbio->bi_io_vec[j].bv_len))
2026                                         break;
2027                         }
2028                 } else
2029                         j = 0;
2030                 if (j >= 0)
2031                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2032                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2033                               && uptodate)) {
2034                         /* No need to write to this device. */
2035                         sbio->bi_end_io = NULL;
2036                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2037                         continue;
2038                 }
2039
2040                 bio_copy_data(sbio, pbio);
2041         }
2042 }
2043
2044 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2045 {
2046         struct r1conf *conf = mddev->private;
2047         int i;
2048         int disks = conf->raid_disks * 2;
2049         struct bio *bio, *wbio;
2050
2051         bio = r1_bio->bios[r1_bio->read_disk];
2052
2053         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2054                 /* ouch - failed to read all of that. */
2055                 if (!fix_sync_read_error(r1_bio))
2056                         return;
2057
2058         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2059                 process_checks(r1_bio);
2060
2061         /*
2062          * schedule writes
2063          */
2064         atomic_set(&r1_bio->remaining, 1);
2065         for (i = 0; i < disks ; i++) {
2066                 wbio = r1_bio->bios[i];
2067                 if (wbio->bi_end_io == NULL ||
2068                     (wbio->bi_end_io == end_sync_read &&
2069                      (i == r1_bio->read_disk ||
2070                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2071                         continue;
2072
2073                 wbio->bi_rw = WRITE;
2074                 wbio->bi_end_io = end_sync_write;
2075                 atomic_inc(&r1_bio->remaining);
2076                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2077
2078                 generic_make_request(wbio);
2079         }
2080
2081         if (atomic_dec_and_test(&r1_bio->remaining)) {
2082                 /* if we're here, all write(s) have completed, so clean up */
2083                 int s = r1_bio->sectors;
2084                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2085                     test_bit(R1BIO_WriteError, &r1_bio->state))
2086                         reschedule_retry(r1_bio);
2087                 else {
2088                         put_buf(r1_bio);
2089                         md_done_sync(mddev, s, 1);
2090                 }
2091         }
2092 }
2093
2094 /*
2095  * This is a kernel thread which:
2096  *
2097  *      1.      Retries failed read operations on working mirrors.
2098  *      2.      Updates the raid superblock when problems encounter.
2099  *      3.      Performs writes following reads for array synchronising.
2100  */
2101
2102 static void fix_read_error(struct r1conf *conf, int read_disk,
2103                            sector_t sect, int sectors)
2104 {
2105         struct mddev *mddev = conf->mddev;
2106         while(sectors) {
2107                 int s = sectors;
2108                 int d = read_disk;
2109                 int success = 0;
2110                 int start;
2111                 struct md_rdev *rdev;
2112
2113                 if (s > (PAGE_SIZE>>9))
2114                         s = PAGE_SIZE >> 9;
2115
2116                 do {
2117                         /* Note: no rcu protection needed here
2118                          * as this is synchronous in the raid1d thread
2119                          * which is the thread that might remove
2120                          * a device.  If raid1d ever becomes multi-threaded....
2121                          */
2122                         sector_t first_bad;
2123                         int bad_sectors;
2124
2125                         rdev = conf->mirrors[d].rdev;
2126                         if (rdev &&
2127                             (test_bit(In_sync, &rdev->flags) ||
2128                              (!test_bit(Faulty, &rdev->flags) &&
2129                               rdev->recovery_offset >= sect + s)) &&
2130                             is_badblock(rdev, sect, s,
2131                                         &first_bad, &bad_sectors) == 0 &&
2132                             sync_page_io(rdev, sect, s<<9,
2133                                          conf->tmppage, READ, false))
2134                                 success = 1;
2135                         else {
2136                                 d++;
2137                                 if (d == conf->raid_disks * 2)
2138                                         d = 0;
2139                         }
2140                 } while (!success && d != read_disk);
2141
2142                 if (!success) {
2143                         /* Cannot read from anywhere - mark it bad */
2144                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2145                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2146                                 md_error(mddev, rdev);
2147                         break;
2148                 }
2149                 /* write it back and re-read */
2150                 start = d;
2151                 while (d != read_disk) {
2152                         if (d==0)
2153                                 d = conf->raid_disks * 2;
2154                         d--;
2155                         rdev = conf->mirrors[d].rdev;
2156                         if (rdev &&
2157                             !test_bit(Faulty, &rdev->flags))
2158                                 r1_sync_page_io(rdev, sect, s,
2159                                                 conf->tmppage, WRITE);
2160                 }
2161                 d = start;
2162                 while (d != read_disk) {
2163                         char b[BDEVNAME_SIZE];
2164                         if (d==0)
2165                                 d = conf->raid_disks * 2;
2166                         d--;
2167                         rdev = conf->mirrors[d].rdev;
2168                         if (rdev &&
2169                             !test_bit(Faulty, &rdev->flags)) {
2170                                 if (r1_sync_page_io(rdev, sect, s,
2171                                                     conf->tmppage, READ)) {
2172                                         atomic_add(s, &rdev->corrected_errors);
2173                                         printk(KERN_INFO
2174                                                "md/raid1:%s: read error corrected "
2175                                                "(%d sectors at %llu on %s)\n",
2176                                                mdname(mddev), s,
2177                                                (unsigned long long)(sect +
2178                                                    rdev->data_offset),
2179                                                bdevname(rdev->bdev, b));
2180                                 }
2181                         }
2182                 }
2183                 sectors -= s;
2184                 sect += s;
2185         }
2186 }
2187
2188 static int narrow_write_error(struct r1bio *r1_bio, int i)
2189 {
2190         struct mddev *mddev = r1_bio->mddev;
2191         struct r1conf *conf = mddev->private;
2192         struct md_rdev *rdev = conf->mirrors[i].rdev;
2193
2194         /* bio has the data to be written to device 'i' where
2195          * we just recently had a write error.
2196          * We repeatedly clone the bio and trim down to one block,
2197          * then try the write.  Where the write fails we record
2198          * a bad block.
2199          * It is conceivable that the bio doesn't exactly align with
2200          * blocks.  We must handle this somehow.
2201          *
2202          * We currently own a reference on the rdev.
2203          */
2204
2205         int block_sectors;
2206         sector_t sector;
2207         int sectors;
2208         int sect_to_write = r1_bio->sectors;
2209         int ok = 1;
2210
2211         if (rdev->badblocks.shift < 0)
2212                 return 0;
2213
2214         block_sectors = roundup(1 << rdev->badblocks.shift,
2215                                 bdev_logical_block_size(rdev->bdev) >> 9);
2216         sector = r1_bio->sector;
2217         sectors = ((sector + block_sectors)
2218                    & ~(sector_t)(block_sectors - 1))
2219                 - sector;
2220
2221         while (sect_to_write) {
2222                 struct bio *wbio;
2223                 if (sectors > sect_to_write)
2224                         sectors = sect_to_write;
2225                 /* Write at 'sector' for 'sectors'*/
2226
2227                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2228                         unsigned vcnt = r1_bio->behind_page_count;
2229                         struct bio_vec *vec = r1_bio->behind_bvecs;
2230
2231                         while (!vec->bv_page) {
2232                                 vec++;
2233                                 vcnt--;
2234                         }
2235
2236                         wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2237                         memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2238
2239                         wbio->bi_vcnt = vcnt;
2240                 } else {
2241                         wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2242                 }
2243
2244                 wbio->bi_rw = WRITE;
2245                 wbio->bi_iter.bi_sector = r1_bio->sector;
2246                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2247
2248                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2249                 wbio->bi_iter.bi_sector += rdev->data_offset;
2250                 wbio->bi_bdev = rdev->bdev;
2251                 if (submit_bio_wait(WRITE, wbio) == 0)
2252                         /* failure! */
2253                         ok = rdev_set_badblocks(rdev, sector,
2254                                                 sectors, 0)
2255                                 && ok;
2256
2257                 bio_put(wbio);
2258                 sect_to_write -= sectors;
2259                 sector += sectors;
2260                 sectors = block_sectors;
2261         }
2262         return ok;
2263 }
2264
2265 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2266 {
2267         int m;
2268         int s = r1_bio->sectors;
2269         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2270                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2271                 struct bio *bio = r1_bio->bios[m];
2272                 if (bio->bi_end_io == NULL)
2273                         continue;
2274                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2275                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2276                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2277                 }
2278                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2279                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2280                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2281                                 md_error(conf->mddev, rdev);
2282                 }
2283         }
2284         put_buf(r1_bio);
2285         md_done_sync(conf->mddev, s, 1);
2286 }
2287
2288 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2289 {
2290         int m;
2291         for (m = 0; m < conf->raid_disks * 2 ; m++)
2292                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2293                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2294                         rdev_clear_badblocks(rdev,
2295                                              r1_bio->sector,
2296                                              r1_bio->sectors, 0);
2297                         rdev_dec_pending(rdev, conf->mddev);
2298                 } else if (r1_bio->bios[m] != NULL) {
2299                         /* This drive got a write error.  We need to
2300                          * narrow down and record precise write
2301                          * errors.
2302                          */
2303                         if (!narrow_write_error(r1_bio, m)) {
2304                                 md_error(conf->mddev,
2305                                          conf->mirrors[m].rdev);
2306                                 /* an I/O failed, we can't clear the bitmap */
2307                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2308                         }
2309                         rdev_dec_pending(conf->mirrors[m].rdev,
2310                                          conf->mddev);
2311                 }
2312         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2313                 close_write(r1_bio);
2314         raid_end_bio_io(r1_bio);
2315 }
2316
2317 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2318 {
2319         int disk;
2320         int max_sectors;
2321         struct mddev *mddev = conf->mddev;
2322         struct bio *bio;
2323         char b[BDEVNAME_SIZE];
2324         struct md_rdev *rdev;
2325
2326         clear_bit(R1BIO_ReadError, &r1_bio->state);
2327         /* we got a read error. Maybe the drive is bad.  Maybe just
2328          * the block and we can fix it.
2329          * We freeze all other IO, and try reading the block from
2330          * other devices.  When we find one, we re-write
2331          * and check it that fixes the read error.
2332          * This is all done synchronously while the array is
2333          * frozen
2334          */
2335         if (mddev->ro == 0) {
2336                 freeze_array(conf, 1);
2337                 fix_read_error(conf, r1_bio->read_disk,
2338                                r1_bio->sector, r1_bio->sectors);
2339                 unfreeze_array(conf);
2340         } else
2341                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2342         rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2343
2344         bio = r1_bio->bios[r1_bio->read_disk];
2345         bdevname(bio->bi_bdev, b);
2346 read_more:
2347         disk = read_balance(conf, r1_bio, &max_sectors);
2348         if (disk == -1) {
2349                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2350                        " read error for block %llu\n",
2351                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2352                 raid_end_bio_io(r1_bio);
2353         } else {
2354                 const unsigned long do_sync
2355                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2356                 if (bio) {
2357                         r1_bio->bios[r1_bio->read_disk] =
2358                                 mddev->ro ? IO_BLOCKED : NULL;
2359                         bio_put(bio);
2360                 }
2361                 r1_bio->read_disk = disk;
2362                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2363                 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2364                          max_sectors);
2365                 r1_bio->bios[r1_bio->read_disk] = bio;
2366                 rdev = conf->mirrors[disk].rdev;
2367                 printk_ratelimited(KERN_ERR
2368                                    "md/raid1:%s: redirecting sector %llu"
2369                                    " to other mirror: %s\n",
2370                                    mdname(mddev),
2371                                    (unsigned long long)r1_bio->sector,
2372                                    bdevname(rdev->bdev, b));
2373                 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2374                 bio->bi_bdev = rdev->bdev;
2375                 bio->bi_end_io = raid1_end_read_request;
2376                 bio->bi_rw = READ | do_sync;
2377                 bio->bi_private = r1_bio;
2378                 if (max_sectors < r1_bio->sectors) {
2379                         /* Drat - have to split this up more */
2380                         struct bio *mbio = r1_bio->master_bio;
2381                         int sectors_handled = (r1_bio->sector + max_sectors
2382                                                - mbio->bi_iter.bi_sector);
2383                         r1_bio->sectors = max_sectors;
2384                         spin_lock_irq(&conf->device_lock);
2385                         if (mbio->bi_phys_segments == 0)
2386                                 mbio->bi_phys_segments = 2;
2387                         else
2388                                 mbio->bi_phys_segments++;
2389                         spin_unlock_irq(&conf->device_lock);
2390                         generic_make_request(bio);
2391                         bio = NULL;
2392
2393                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2394
2395                         r1_bio->master_bio = mbio;
2396                         r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2397                         r1_bio->state = 0;
2398                         set_bit(R1BIO_ReadError, &r1_bio->state);
2399                         r1_bio->mddev = mddev;
2400                         r1_bio->sector = mbio->bi_iter.bi_sector +
2401                                 sectors_handled;
2402
2403                         goto read_more;
2404                 } else
2405                         generic_make_request(bio);
2406         }
2407 }
2408
2409 static void raid1d(struct md_thread *thread)
2410 {
2411         struct mddev *mddev = thread->mddev;
2412         struct r1bio *r1_bio;
2413         unsigned long flags;
2414         struct r1conf *conf = mddev->private;
2415         struct list_head *head = &conf->retry_list;
2416         struct blk_plug plug;
2417
2418         md_check_recovery(mddev);
2419
2420         blk_start_plug(&plug);
2421         for (;;) {
2422
2423                 flush_pending_writes(conf);
2424
2425                 spin_lock_irqsave(&conf->device_lock, flags);
2426                 if (list_empty(head)) {
2427                         spin_unlock_irqrestore(&conf->device_lock, flags);
2428                         break;
2429                 }
2430                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2431                 list_del(head->prev);
2432                 conf->nr_queued--;
2433                 spin_unlock_irqrestore(&conf->device_lock, flags);
2434
2435                 mddev = r1_bio->mddev;
2436                 conf = mddev->private;
2437                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2438                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2439                             test_bit(R1BIO_WriteError, &r1_bio->state))
2440                                 handle_sync_write_finished(conf, r1_bio);
2441                         else
2442                                 sync_request_write(mddev, r1_bio);
2443                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2444                            test_bit(R1BIO_WriteError, &r1_bio->state))
2445                         handle_write_finished(conf, r1_bio);
2446                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2447                         handle_read_error(conf, r1_bio);
2448                 else
2449                         /* just a partial read to be scheduled from separate
2450                          * context
2451                          */
2452                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2453
2454                 cond_resched();
2455                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2456                         md_check_recovery(mddev);
2457         }
2458         blk_finish_plug(&plug);
2459 }
2460
2461 static int init_resync(struct r1conf *conf)
2462 {
2463         int buffs;
2464
2465         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2466         BUG_ON(conf->r1buf_pool);
2467         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2468                                           conf->poolinfo);
2469         if (!conf->r1buf_pool)
2470                 return -ENOMEM;
2471         conf->next_resync = 0;
2472         return 0;
2473 }
2474
2475 /*
2476  * perform a "sync" on one "block"
2477  *
2478  * We need to make sure that no normal I/O request - particularly write
2479  * requests - conflict with active sync requests.
2480  *
2481  * This is achieved by tracking pending requests and a 'barrier' concept
2482  * that can be installed to exclude normal IO requests.
2483  */
2484
2485 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
2486 {
2487         struct r1conf *conf = mddev->private;
2488         struct r1bio *r1_bio;
2489         struct bio *bio;
2490         sector_t max_sector, nr_sectors;
2491         int disk = -1;
2492         int i;
2493         int wonly = -1;
2494         int write_targets = 0, read_targets = 0;
2495         sector_t sync_blocks;
2496         int still_degraded = 0;
2497         int good_sectors = RESYNC_SECTORS;
2498         int min_bad = 0; /* number of sectors that are bad in all devices */
2499
2500         if (!conf->r1buf_pool)
2501                 if (init_resync(conf))
2502                         return 0;
2503
2504         max_sector = mddev->dev_sectors;
2505         if (sector_nr >= max_sector) {
2506                 /* If we aborted, we need to abort the
2507                  * sync on the 'current' bitmap chunk (there will
2508                  * only be one in raid1 resync.
2509                  * We can find the current addess in mddev->curr_resync
2510                  */
2511                 if (mddev->curr_resync < max_sector) /* aborted */
2512                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2513                                                 &sync_blocks, 1);
2514                 else /* completed sync */
2515                         conf->fullsync = 0;
2516
2517                 bitmap_close_sync(mddev->bitmap);
2518                 close_sync(conf);
2519                 return 0;
2520         }
2521
2522         if (mddev->bitmap == NULL &&
2523             mddev->recovery_cp == MaxSector &&
2524             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2525             conf->fullsync == 0) {
2526                 *skipped = 1;
2527                 return max_sector - sector_nr;
2528         }
2529         /* before building a request, check if we can skip these blocks..
2530          * This call the bitmap_start_sync doesn't actually record anything
2531          */
2532         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2533             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2534                 /* We can skip this block, and probably several more */
2535                 *skipped = 1;
2536                 return sync_blocks;
2537         }
2538
2539         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2540         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2541
2542         raise_barrier(conf, sector_nr);
2543
2544         rcu_read_lock();
2545         /*
2546          * If we get a correctably read error during resync or recovery,
2547          * we might want to read from a different device.  So we
2548          * flag all drives that could conceivably be read from for READ,
2549          * and any others (which will be non-In_sync devices) for WRITE.
2550          * If a read fails, we try reading from something else for which READ
2551          * is OK.
2552          */
2553
2554         r1_bio->mddev = mddev;
2555         r1_bio->sector = sector_nr;
2556         r1_bio->state = 0;
2557         set_bit(R1BIO_IsSync, &r1_bio->state);
2558
2559         for (i = 0; i < conf->raid_disks * 2; i++) {
2560                 struct md_rdev *rdev;
2561                 bio = r1_bio->bios[i];
2562                 bio_reset(bio);
2563
2564                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2565                 if (rdev == NULL ||
2566                     test_bit(Faulty, &rdev->flags)) {
2567                         if (i < conf->raid_disks)
2568                                 still_degraded = 1;
2569                 } else if (!test_bit(In_sync, &rdev->flags)) {
2570                         bio->bi_rw = WRITE;
2571                         bio->bi_end_io = end_sync_write;
2572                         write_targets ++;
2573                 } else {
2574                         /* may need to read from here */
2575                         sector_t first_bad = MaxSector;
2576                         int bad_sectors;
2577
2578                         if (is_badblock(rdev, sector_nr, good_sectors,
2579                                         &first_bad, &bad_sectors)) {
2580                                 if (first_bad > sector_nr)
2581                                         good_sectors = first_bad - sector_nr;
2582                                 else {
2583                                         bad_sectors -= (sector_nr - first_bad);
2584                                         if (min_bad == 0 ||
2585                                             min_bad > bad_sectors)
2586                                                 min_bad = bad_sectors;
2587                                 }
2588                         }
2589                         if (sector_nr < first_bad) {
2590                                 if (test_bit(WriteMostly, &rdev->flags)) {
2591                                         if (wonly < 0)
2592                                                 wonly = i;
2593                                 } else {
2594                                         if (disk < 0)
2595                                                 disk = i;
2596                                 }
2597                                 bio->bi_rw = READ;
2598                                 bio->bi_end_io = end_sync_read;
2599                                 read_targets++;
2600                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2601                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2602                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2603                                 /*
2604                                  * The device is suitable for reading (InSync),
2605                                  * but has bad block(s) here. Let's try to correct them,
2606                                  * if we are doing resync or repair. Otherwise, leave
2607                                  * this device alone for this sync request.
2608                                  */
2609                                 bio->bi_rw = WRITE;
2610                                 bio->bi_end_io = end_sync_write;
2611                                 write_targets++;
2612                         }
2613                 }
2614                 if (bio->bi_end_io) {
2615                         atomic_inc(&rdev->nr_pending);
2616                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2617                         bio->bi_bdev = rdev->bdev;
2618                         bio->bi_private = r1_bio;
2619                 }
2620         }
2621         rcu_read_unlock();
2622         if (disk < 0)
2623                 disk = wonly;
2624         r1_bio->read_disk = disk;
2625
2626         if (read_targets == 0 && min_bad > 0) {
2627                 /* These sectors are bad on all InSync devices, so we
2628                  * need to mark them bad on all write targets
2629                  */
2630                 int ok = 1;
2631                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2632                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2633                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2634                                 ok = rdev_set_badblocks(rdev, sector_nr,
2635                                                         min_bad, 0
2636                                         ) && ok;
2637                         }
2638                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2639                 *skipped = 1;
2640                 put_buf(r1_bio);
2641
2642                 if (!ok) {
2643                         /* Cannot record the badblocks, so need to
2644                          * abort the resync.
2645                          * If there are multiple read targets, could just
2646                          * fail the really bad ones ???
2647                          */
2648                         conf->recovery_disabled = mddev->recovery_disabled;
2649                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2650                         return 0;
2651                 } else
2652                         return min_bad;
2653
2654         }
2655         if (min_bad > 0 && min_bad < good_sectors) {
2656                 /* only resync enough to reach the next bad->good
2657                  * transition */
2658                 good_sectors = min_bad;
2659         }
2660
2661         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2662                 /* extra read targets are also write targets */
2663                 write_targets += read_targets-1;
2664
2665         if (write_targets == 0 || read_targets == 0) {
2666                 /* There is nowhere to write, so all non-sync
2667                  * drives must be failed - so we are finished
2668                  */
2669                 sector_t rv;
2670                 if (min_bad > 0)
2671                         max_sector = sector_nr + min_bad;
2672                 rv = max_sector - sector_nr;
2673                 *skipped = 1;
2674                 put_buf(r1_bio);
2675                 return rv;
2676         }
2677
2678         if (max_sector > mddev->resync_max)
2679                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2680         if (max_sector > sector_nr + good_sectors)
2681                 max_sector = sector_nr + good_sectors;
2682         nr_sectors = 0;
2683         sync_blocks = 0;
2684         do {
2685                 struct page *page;
2686                 int len = PAGE_SIZE;
2687                 if (sector_nr + (len>>9) > max_sector)
2688                         len = (max_sector - sector_nr) << 9;
2689                 if (len == 0)
2690                         break;
2691                 if (sync_blocks == 0) {
2692                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2693                                                &sync_blocks, still_degraded) &&
2694                             !conf->fullsync &&
2695                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2696                                 break;
2697                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2698                         if ((len >> 9) > sync_blocks)
2699                                 len = sync_blocks<<9;
2700                 }
2701
2702                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2703                         bio = r1_bio->bios[i];
2704                         if (bio->bi_end_io) {
2705                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2706                                 if (bio_add_page(bio, page, len, 0) == 0) {
2707                                         /* stop here */
2708                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2709                                         while (i > 0) {
2710                                                 i--;
2711                                                 bio = r1_bio->bios[i];
2712                                                 if (bio->bi_end_io==NULL)
2713                                                         continue;
2714                                                 /* remove last page from this bio */
2715                                                 bio->bi_vcnt--;
2716                                                 bio->bi_iter.bi_size -= len;
2717                                                 __clear_bit(BIO_SEG_VALID, &bio->bi_flags);
2718                                         }
2719                                         goto bio_full;
2720                                 }
2721                         }
2722                 }
2723                 nr_sectors += len>>9;
2724                 sector_nr += len>>9;
2725                 sync_blocks -= (len>>9);
2726         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2727  bio_full:
2728         r1_bio->sectors = nr_sectors;
2729
2730         /* For a user-requested sync, we read all readable devices and do a
2731          * compare
2732          */
2733         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2734                 atomic_set(&r1_bio->remaining, read_targets);
2735                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2736                         bio = r1_bio->bios[i];
2737                         if (bio->bi_end_io == end_sync_read) {
2738                                 read_targets--;
2739                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2740                                 generic_make_request(bio);
2741                         }
2742                 }
2743         } else {
2744                 atomic_set(&r1_bio->remaining, 1);
2745                 bio = r1_bio->bios[r1_bio->read_disk];
2746                 md_sync_acct(bio->bi_bdev, nr_sectors);
2747                 generic_make_request(bio);
2748
2749         }
2750         return nr_sectors;
2751 }
2752
2753 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2754 {
2755         if (sectors)
2756                 return sectors;
2757
2758         return mddev->dev_sectors;
2759 }
2760
2761 static struct r1conf *setup_conf(struct mddev *mddev)
2762 {
2763         struct r1conf *conf;
2764         int i;
2765         struct raid1_info *disk;
2766         struct md_rdev *rdev;
2767         int err = -ENOMEM;
2768
2769         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2770         if (!conf)
2771                 goto abort;
2772
2773         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2774                                 * mddev->raid_disks * 2,
2775                                  GFP_KERNEL);
2776         if (!conf->mirrors)
2777                 goto abort;
2778
2779         conf->tmppage = alloc_page(GFP_KERNEL);
2780         if (!conf->tmppage)
2781                 goto abort;
2782
2783         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2784         if (!conf->poolinfo)
2785                 goto abort;
2786         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2787         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2788                                           r1bio_pool_free,
2789                                           conf->poolinfo);
2790         if (!conf->r1bio_pool)
2791                 goto abort;
2792
2793         conf->poolinfo->mddev = mddev;
2794
2795         err = -EINVAL;
2796         spin_lock_init(&conf->device_lock);
2797         rdev_for_each(rdev, mddev) {
2798                 struct request_queue *q;
2799                 int disk_idx = rdev->raid_disk;
2800                 if (disk_idx >= mddev->raid_disks
2801                     || disk_idx < 0)
2802                         continue;
2803                 if (test_bit(Replacement, &rdev->flags))
2804                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
2805                 else
2806                         disk = conf->mirrors + disk_idx;
2807
2808                 if (disk->rdev)
2809                         goto abort;
2810                 disk->rdev = rdev;
2811                 q = bdev_get_queue(rdev->bdev);
2812                 if (q->merge_bvec_fn)
2813                         mddev->merge_check_needed = 1;
2814
2815                 disk->head_position = 0;
2816                 disk->seq_start = MaxSector;
2817         }
2818         conf->raid_disks = mddev->raid_disks;
2819         conf->mddev = mddev;
2820         INIT_LIST_HEAD(&conf->retry_list);
2821
2822         spin_lock_init(&conf->resync_lock);
2823         init_waitqueue_head(&conf->wait_barrier);
2824
2825         bio_list_init(&conf->pending_bio_list);
2826         conf->pending_count = 0;
2827         conf->recovery_disabled = mddev->recovery_disabled - 1;
2828
2829         conf->start_next_window = MaxSector;
2830         conf->current_window_requests = conf->next_window_requests = 0;
2831
2832         err = -EIO;
2833         for (i = 0; i < conf->raid_disks * 2; i++) {
2834
2835                 disk = conf->mirrors + i;
2836
2837                 if (i < conf->raid_disks &&
2838                     disk[conf->raid_disks].rdev) {
2839                         /* This slot has a replacement. */
2840                         if (!disk->rdev) {
2841                                 /* No original, just make the replacement
2842                                  * a recovering spare
2843                                  */
2844                                 disk->rdev =
2845                                         disk[conf->raid_disks].rdev;
2846                                 disk[conf->raid_disks].rdev = NULL;
2847                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2848                                 /* Original is not in_sync - bad */
2849                                 goto abort;
2850                 }
2851
2852                 if (!disk->rdev ||
2853                     !test_bit(In_sync, &disk->rdev->flags)) {
2854                         disk->head_position = 0;
2855                         if (disk->rdev &&
2856                             (disk->rdev->saved_raid_disk < 0))
2857                                 conf->fullsync = 1;
2858                 }
2859         }
2860
2861         err = -ENOMEM;
2862         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2863         if (!conf->thread) {
2864                 printk(KERN_ERR
2865                        "md/raid1:%s: couldn't allocate thread\n",
2866                        mdname(mddev));
2867                 goto abort;
2868         }
2869
2870         return conf;
2871
2872  abort:
2873         if (conf) {
2874                 if (conf->r1bio_pool)
2875                         mempool_destroy(conf->r1bio_pool);
2876                 kfree(conf->mirrors);
2877                 safe_put_page(conf->tmppage);
2878                 kfree(conf->poolinfo);
2879                 kfree(conf);
2880         }
2881         return ERR_PTR(err);
2882 }
2883
2884 static void raid1_free(struct mddev *mddev, void *priv);
2885 static int run(struct mddev *mddev)
2886 {
2887         struct r1conf *conf;
2888         int i;
2889         struct md_rdev *rdev;
2890         int ret;
2891         bool discard_supported = false;
2892
2893         if (mddev->level != 1) {
2894                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2895                        mdname(mddev), mddev->level);
2896                 return -EIO;
2897         }
2898         if (mddev->reshape_position != MaxSector) {
2899                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2900                        mdname(mddev));
2901                 return -EIO;
2902         }
2903         /*
2904          * copy the already verified devices into our private RAID1
2905          * bookkeeping area. [whatever we allocate in run(),
2906          * should be freed in raid1_free()]
2907          */
2908         if (mddev->private == NULL)
2909                 conf = setup_conf(mddev);
2910         else
2911                 conf = mddev->private;
2912
2913         if (IS_ERR(conf))
2914                 return PTR_ERR(conf);
2915
2916         if (mddev->queue)
2917                 blk_queue_max_write_same_sectors(mddev->queue, 0);
2918
2919         rdev_for_each(rdev, mddev) {
2920                 if (!mddev->gendisk)
2921                         continue;
2922                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2923                                   rdev->data_offset << 9);
2924                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2925                         discard_supported = true;
2926         }
2927
2928         mddev->degraded = 0;
2929         for (i=0; i < conf->raid_disks; i++)
2930                 if (conf->mirrors[i].rdev == NULL ||
2931                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2932                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2933                         mddev->degraded++;
2934
2935         if (conf->raid_disks - mddev->degraded == 1)
2936                 mddev->recovery_cp = MaxSector;
2937
2938         if (mddev->recovery_cp != MaxSector)
2939                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2940                        " -- starting background reconstruction\n",
2941                        mdname(mddev));
2942         printk(KERN_INFO
2943                 "md/raid1:%s: active with %d out of %d mirrors\n",
2944                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2945                 mddev->raid_disks);
2946
2947         /*
2948          * Ok, everything is just fine now
2949          */
2950         mddev->thread = conf->thread;
2951         conf->thread = NULL;
2952         mddev->private = conf;
2953
2954         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2955
2956         if (mddev->queue) {
2957                 if (discard_supported)
2958                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2959                                                 mddev->queue);
2960                 else
2961                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2962                                                   mddev->queue);
2963         }
2964
2965         ret =  md_integrity_register(mddev);
2966         if (ret) {
2967                 md_unregister_thread(&mddev->thread);
2968                 raid1_free(mddev, conf);
2969         }
2970         return ret;
2971 }
2972
2973 static void raid1_free(struct mddev *mddev, void *priv)
2974 {
2975         struct r1conf *conf = priv;
2976
2977         if (conf->r1bio_pool)
2978                 mempool_destroy(conf->r1bio_pool);
2979         kfree(conf->mirrors);
2980         safe_put_page(conf->tmppage);
2981         kfree(conf->poolinfo);
2982         kfree(conf);
2983 }
2984
2985 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2986 {
2987         /* no resync is happening, and there is enough space
2988          * on all devices, so we can resize.
2989          * We need to make sure resync covers any new space.
2990          * If the array is shrinking we should possibly wait until
2991          * any io in the removed space completes, but it hardly seems
2992          * worth it.
2993          */
2994         sector_t newsize = raid1_size(mddev, sectors, 0);
2995         if (mddev->external_size &&
2996             mddev->array_sectors > newsize)
2997                 return -EINVAL;
2998         if (mddev->bitmap) {
2999                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3000                 if (ret)
3001                         return ret;
3002         }
3003         md_set_array_sectors(mddev, newsize);
3004         set_capacity(mddev->gendisk, mddev->array_sectors);
3005         revalidate_disk(mddev->gendisk);
3006         if (sectors > mddev->dev_sectors &&
3007             mddev->recovery_cp > mddev->dev_sectors) {
3008                 mddev->recovery_cp = mddev->dev_sectors;
3009                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3010         }
3011         mddev->dev_sectors = sectors;
3012         mddev->resync_max_sectors = sectors;
3013         return 0;
3014 }
3015
3016 static int raid1_reshape(struct mddev *mddev)
3017 {
3018         /* We need to:
3019          * 1/ resize the r1bio_pool
3020          * 2/ resize conf->mirrors
3021          *
3022          * We allocate a new r1bio_pool if we can.
3023          * Then raise a device barrier and wait until all IO stops.
3024          * Then resize conf->mirrors and swap in the new r1bio pool.
3025          *
3026          * At the same time, we "pack" the devices so that all the missing
3027          * devices have the higher raid_disk numbers.
3028          */
3029         mempool_t *newpool, *oldpool;
3030         struct pool_info *newpoolinfo;
3031         struct raid1_info *newmirrors;
3032         struct r1conf *conf = mddev->private;
3033         int cnt, raid_disks;
3034         unsigned long flags;
3035         int d, d2, err;
3036
3037         /* Cannot change chunk_size, layout, or level */
3038         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3039             mddev->layout != mddev->new_layout ||
3040             mddev->level != mddev->new_level) {
3041                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3042                 mddev->new_layout = mddev->layout;
3043                 mddev->new_level = mddev->level;
3044                 return -EINVAL;
3045         }
3046
3047         err = md_allow_write(mddev);
3048         if (err)
3049                 return err;
3050
3051         raid_disks = mddev->raid_disks + mddev->delta_disks;
3052
3053         if (raid_disks < conf->raid_disks) {
3054                 cnt=0;
3055                 for (d= 0; d < conf->raid_disks; d++)
3056                         if (conf->mirrors[d].rdev)
3057                                 cnt++;
3058                 if (cnt > raid_disks)
3059                         return -EBUSY;
3060         }
3061
3062         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3063         if (!newpoolinfo)
3064                 return -ENOMEM;
3065         newpoolinfo->mddev = mddev;
3066         newpoolinfo->raid_disks = raid_disks * 2;
3067
3068         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3069                                  r1bio_pool_free, newpoolinfo);
3070         if (!newpool) {
3071                 kfree(newpoolinfo);
3072                 return -ENOMEM;
3073         }
3074         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3075                              GFP_KERNEL);
3076         if (!newmirrors) {
3077                 kfree(newpoolinfo);
3078                 mempool_destroy(newpool);
3079                 return -ENOMEM;
3080         }
3081
3082         freeze_array(conf, 0);
3083
3084         /* ok, everything is stopped */
3085         oldpool = conf->r1bio_pool;
3086         conf->r1bio_pool = newpool;
3087
3088         for (d = d2 = 0; d < conf->raid_disks; d++) {
3089                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3090                 if (rdev && rdev->raid_disk != d2) {
3091                         sysfs_unlink_rdev(mddev, rdev);
3092                         rdev->raid_disk = d2;
3093                         sysfs_unlink_rdev(mddev, rdev);
3094                         if (sysfs_link_rdev(mddev, rdev))
3095                                 printk(KERN_WARNING
3096                                        "md/raid1:%s: cannot register rd%d\n",
3097                                        mdname(mddev), rdev->raid_disk);
3098                 }
3099                 if (rdev)
3100                         newmirrors[d2++].rdev = rdev;
3101         }
3102         kfree(conf->mirrors);
3103         conf->mirrors = newmirrors;
3104         kfree(conf->poolinfo);
3105         conf->poolinfo = newpoolinfo;
3106
3107         spin_lock_irqsave(&conf->device_lock, flags);
3108         mddev->degraded += (raid_disks - conf->raid_disks);
3109         spin_unlock_irqrestore(&conf->device_lock, flags);
3110         conf->raid_disks = mddev->raid_disks = raid_disks;
3111         mddev->delta_disks = 0;
3112
3113         unfreeze_array(conf);
3114
3115         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3116         md_wakeup_thread(mddev->thread);
3117
3118         mempool_destroy(oldpool);
3119         return 0;
3120 }
3121
3122 static void raid1_quiesce(struct mddev *mddev, int state)
3123 {
3124         struct r1conf *conf = mddev->private;
3125
3126         switch(state) {
3127         case 2: /* wake for suspend */
3128                 wake_up(&conf->wait_barrier);
3129                 break;
3130         case 1:
3131                 freeze_array(conf, 0);
3132                 break;
3133         case 0:
3134                 unfreeze_array(conf);
3135                 break;
3136         }
3137 }
3138
3139 static void *raid1_takeover(struct mddev *mddev)
3140 {
3141         /* raid1 can take over:
3142          *  raid5 with 2 devices, any layout or chunk size
3143          */
3144         if (mddev->level == 5 && mddev->raid_disks == 2) {
3145                 struct r1conf *conf;
3146                 mddev->new_level = 1;
3147                 mddev->new_layout = 0;
3148                 mddev->new_chunk_sectors = 0;
3149                 conf = setup_conf(mddev);
3150                 if (!IS_ERR(conf))
3151                         /* Array must appear to be quiesced */
3152                         conf->array_frozen = 1;
3153                 return conf;
3154         }
3155         return ERR_PTR(-EINVAL);
3156 }
3157
3158 static struct md_personality raid1_personality =
3159 {
3160         .name           = "raid1",
3161         .level          = 1,
3162         .owner          = THIS_MODULE,
3163         .make_request   = make_request,
3164         .run            = run,
3165         .free           = raid1_free,
3166         .status         = status,
3167         .error_handler  = error,
3168         .hot_add_disk   = raid1_add_disk,
3169         .hot_remove_disk= raid1_remove_disk,
3170         .spare_active   = raid1_spare_active,
3171         .sync_request   = sync_request,
3172         .resize         = raid1_resize,
3173         .size           = raid1_size,
3174         .check_reshape  = raid1_reshape,
3175         .quiesce        = raid1_quiesce,
3176         .takeover       = raid1_takeover,
3177         .congested      = raid1_congested,
3178         .mergeable_bvec = raid1_mergeable_bvec,
3179 };
3180
3181 static int __init raid_init(void)
3182 {
3183         return register_md_personality(&raid1_personality);
3184 }
3185
3186 static void raid_exit(void)
3187 {
3188         unregister_md_personality(&raid1_personality);
3189 }
3190
3191 module_init(raid_init);
3192 module_exit(raid_exit);
3193 MODULE_LICENSE("GPL");
3194 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3195 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3196 MODULE_ALIAS("md-raid1");
3197 MODULE_ALIAS("md-level-1");
3198
3199 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);