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(Faulty, &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 /*
 499  * This routine returns the disk from which the requested read should
 500  * be done. There is a per-array 'next expected sequential IO' sector
 501  * number - if this matches on the next IO then we use the last disk.
 502  * There is also a per-disk 'last know head position' sector that is
 503  * maintained from IRQ contexts, both the normal and the resync IO
 504  * completion handlers update this position correctly. If there is no
 505  * perfect sequential match then we pick the disk whose head is closest.
 506  *
 507  * If there are 2 mirrors in the same 2 devices, performance degrades
 508  * because position is mirror, not device based.
 509  *
 510  * The rdev for the device selected will have nr_pending incremented.
 511  */
 512 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 513 {
 514         const sector_t this_sector = r1_bio->sector;
 515         int sectors;
 516         int best_good_sectors;
 517         int best_disk, best_dist_disk, best_pending_disk;
 518         int has_nonrot_disk;
 519         int disk;
 520         sector_t best_dist;
 521         unsigned int min_pending;
 522         struct md_rdev *rdev;
 523         int choose_first;
 524         int choose_next_idle;
 525
 526         rcu_read_lock();
 527         /*
 528          * Check if we can balance. We can balance on the whole
 529          * device if no resync is going on, or below the resync window.
 530          * We take the first readable disk when above the resync window.
 531          */
 532  retry:
 533         sectors = r1_bio->sectors;
 534         best_disk = -1;
 535         best_dist_disk = -1;
 536         best_dist = MaxSector;
 537         best_pending_disk = -1;
 538         min_pending = UINT_MAX;
 539         best_good_sectors = 0;
 540         has_nonrot_disk = 0;
 541         choose_next_idle = 0;
 542
 543         if (conf->mddev->recovery_cp < MaxSector &&
 544             (this_sector + sectors >= conf->next_resync))
 545                 choose_first = 1;
 546         else
 547                 choose_first = 0;
 548
 549         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 550                 sector_t dist;
 551                 sector_t first_bad;
 552                 int bad_sectors;
 553                 unsigned int pending;
 554                 bool nonrot;
 555
 556                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 557                 if (r1_bio->bios[disk] == IO_BLOCKED
 558                     || rdev == NULL
 559                     || test_bit(Unmerged, &rdev->flags)
 560                     || test_bit(Faulty, &rdev->flags))
 561                         continue;
 562                 if (!test_bit(In_sync, &rdev->flags) &&
 563                     rdev->recovery_offset < this_sector + sectors)
 564                         continue;
 565                 if (test_bit(WriteMostly, &rdev->flags)) {
 566                         /* Don't balance among write-mostly, just
 567                          * use the first as a last resort */
 568                         if (best_disk < 0) {
 569                                 if (is_badblock(rdev, this_sector, sectors,
 570                                                 &first_bad, &bad_sectors)) {
 571                                         if (first_bad < this_sector)
 572                                                 /* Cannot use this */
 573                                                 continue;
 574                                         best_good_sectors = first_bad - this_sector;
 575                                 } else
 576                                         best_good_sectors = sectors;
 577                                 best_disk = disk;
 578                         }
 579                         continue;
 580                 }
 581                 /* This is a reasonable device to use.  It might
 582                  * even be best.
 583                  */
 584                 if (is_badblock(rdev, this_sector, sectors,
 585                                 &first_bad, &bad_sectors)) {
 586                         if (best_dist < MaxSector)
 587                                 /* already have a better device */
 588                                 continue;
 589                         if (first_bad <= this_sector) {
 590                                 /* cannot read here. If this is the 'primary'
 591                                  * device, then we must not read beyond
 592                                  * bad_sectors from another device..
 593                                  */
 594                                 bad_sectors -= (this_sector - first_bad);
 595                                 if (choose_first && sectors > bad_sectors)
 596                                         sectors = bad_sectors;
 597                                 if (best_good_sectors > sectors)
 598                                         best_good_sectors = sectors;
 599
 600                         } else {
 601                                 sector_t good_sectors = first_bad - this_sector;
 602                                 if (good_sectors > best_good_sectors) {
 603                                         best_good_sectors = good_sectors;
 604                                         best_disk = disk;
 605                                 }
 606                                 if (choose_first)
 607                                         break;
 608                         }
 609                         continue;
 610                 } else
 611                         best_good_sectors = sectors;
 612
 613                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 614                 has_nonrot_disk |= nonrot;
 615                 pending = atomic_read(&rdev->nr_pending);
 616                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 617                 if (choose_first) {
 618                         best_disk = disk;
 619                         break;
 620                 }
 621                 /* Don't change to another disk for sequential reads */
 622                 if (conf->mirrors[disk].next_seq_sect == this_sector
 623                     || dist == 0) {
 624                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 625                         struct raid1_info *mirror = &conf->mirrors[disk];
 626
 627                         best_disk = disk;
 628                         /*
 629                          * If buffered sequential IO size exceeds optimal
 630                          * iosize, check if there is idle disk. If yes, choose
 631                          * the idle disk. read_balance could already choose an
 632                          * idle disk before noticing it's a sequential IO in
 633                          * this disk. This doesn't matter because this disk
 634                          * will idle, next time it will be utilized after the
 635                          * first disk has IO size exceeds optimal iosize. In
 636                          * this way, iosize of the first disk will be optimal
 637                          * iosize at least. iosize of the second disk might be
 638                          * small, but not a big deal since when the second disk
 639                          * starts IO, the first disk is likely still busy.
 640                          */
 641                         if (nonrot && opt_iosize > 0 &&
 642                             mirror->seq_start != MaxSector &&
 643                             mirror->next_seq_sect > opt_iosize &&
 644                             mirror->next_seq_sect - opt_iosize >=
 645                             mirror->seq_start) {
 646                                 choose_next_idle = 1;
 647                                 continue;
 648                         }
 649                         break;
 650                 }
 651                 /* If device is idle, use it */
 652                 if (pending == 0) {
 653                         best_disk = disk;
 654                         break;
 655                 }
 656
 657                 if (choose_next_idle)
 658                         continue;
 659
 660                 if (min_pending > pending) {
 661                         min_pending = pending;
 662                         best_pending_disk = disk;
 663                 }
 664
 665                 if (dist < best_dist) {
 666                         best_dist = dist;
 667                         best_dist_disk = disk;
 668                 }
 669         }
 670
 671         /*
 672          * If all disks are rotational, choose the closest disk. If any disk is
 673          * non-rotational, choose the disk with less pending request even the
 674          * disk is rotational, which might/might not be optimal for raids with
 675          * mixed ratation/non-rotational disks depending on workload.
 676          */
 677         if (best_disk == -1) {
 678                 if (has_nonrot_disk)
 679                         best_disk = best_pending_disk;
 680                 else
 681                         best_disk = best_dist_disk;
 682         }
 683
 684         if (best_disk >= 0) {
 685                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 686                 if (!rdev)
 687                         goto retry;
 688                 atomic_inc(&rdev->nr_pending);
 689                 if (test_bit(Faulty, &rdev->flags)) {
 690                         /* cannot risk returning a device that failed
 691                          * before we inc'ed nr_pending
 692                          */
 693                         rdev_dec_pending(rdev, conf->mddev);
 694                         goto retry;
 695                 }
 696                 sectors = best_good_sectors;
 697
 698                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 699                         conf->mirrors[best_disk].seq_start = this_sector;
 700
 701                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 702         }
 703         rcu_read_unlock();
 704         *max_sectors = sectors;
 705
 706         return best_disk;
 707 }
 708
 709 static int raid1_mergeable_bvec(struct request_queue *q,
 710                                 struct bvec_merge_data *bvm,
 711                                 struct bio_vec *biovec)
 712 {
 713         struct mddev *mddev = q->queuedata;
 714         struct r1conf *conf = mddev->private;
 715         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 716         int max = biovec->bv_len;
 717
 718         if (mddev->merge_check_needed) {
 719                 int disk;
 720                 rcu_read_lock();
 721                 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
 722                         struct md_rdev *rdev = rcu_dereference(
 723                                 conf->mirrors[disk].rdev);
 724                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 725                                 struct request_queue *q =
 726                                         bdev_get_queue(rdev->bdev);
 727                                 if (q->merge_bvec_fn) {
 728                                         bvm->bi_sector = sector +
 729                                                 rdev->data_offset;
 730                                         bvm->bi_bdev = rdev->bdev;
 731                                         max = min(max, q->merge_bvec_fn(
 732                                                           q, bvm, biovec));
 733                                 }
 734                         }
 735                 }
 736                 rcu_read_unlock();
 737         }
 738         return max;
 739
 740 }
 741
 742 int md_raid1_congested(struct mddev *mddev, int bits)
 743 {
 744         struct r1conf *conf = mddev->private;
 745         int i, ret = 0;
 746
 747         if ((bits & (1 << BDI_async_congested)) &&
 748             conf->pending_count >= max_queued_requests)
 749                 return 1;
 750
 751         rcu_read_lock();
 752         for (i = 0; i < conf->raid_disks * 2; i++) {
 753                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 754                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 755                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 756
 757                         BUG_ON(!q);
 758
 759                         /* Note the '|| 1' - when read_balance prefers
 760                          * non-congested targets, it can be removed
 761                          */
 762                         if ((bits & (1<<BDI_async_congested)) || 1)
 763                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 764                         else
 765                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 766                 }
 767         }
 768         rcu_read_unlock();
 769         return ret;
 770 }
 771 EXPORT_SYMBOL_GPL(md_raid1_congested);
 772
 773 static int raid1_congested(void *data, int bits)
 774 {
 775         struct mddev *mddev = data;
 776
 777         return mddev_congested(mddev, bits) ||
 778                 md_raid1_congested(mddev, bits);
 779 }
 780
 781 static void flush_pending_writes(struct r1conf *conf)
 782 {
 783         /* Any writes that have been queued but are awaiting
 784          * bitmap updates get flushed here.
 785          */
 786         spin_lock_irq(&conf->device_lock);
 787
 788         if (conf->pending_bio_list.head) {
 789                 struct bio *bio;
 790                 bio = bio_list_get(&conf->pending_bio_list);
 791                 conf->pending_count = 0;
 792                 spin_unlock_irq(&conf->device_lock);
 793                 /* flush any pending bitmap writes to
 794                  * disk before proceeding w/ I/O */
 795                 bitmap_unplug(conf->mddev->bitmap);
 796                 wake_up(&conf->wait_barrier);
 797
 798                 while (bio) { /* submit pending writes */
 799                         struct bio *next = bio->bi_next;
 800                         bio->bi_next = NULL;
 801                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 802                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 803                                 /* Just ignore it */
 804                                 bio_endio(bio, 0);
 805                         else
 806                                 generic_make_request(bio);
 807                         bio = next;
 808                 }
 809         } else
 810                 spin_unlock_irq(&conf->device_lock);
 811 }
 812
 813 /* Barriers....
 814  * Sometimes we need to suspend IO while we do something else,
 815  * either some resync/recovery, or reconfigure the array.
 816  * To do this we raise a 'barrier'.
 817  * The 'barrier' is a counter that can be raised multiple times
 818  * to count how many activities are happening which preclude
 819  * normal IO.
 820  * We can only raise the barrier if there is no pending IO.
 821  * i.e. if nr_pending == 0.
 822  * We choose only to raise the barrier if no-one is waiting for the
 823  * barrier to go down.  This means that as soon as an IO request
 824  * is ready, no other operations which require a barrier will start
 825  * until the IO request has had a chance.
 826  *
 827  * So: regular IO calls 'wait_barrier'.  When that returns there
 828  *    is no backgroup IO happening,  It must arrange to call
 829  *    allow_barrier when it has finished its IO.
 830  * backgroup IO calls must call raise_barrier.  Once that returns
 831  *    there is no normal IO happeing.  It must arrange to call
 832  *    lower_barrier when the particular background IO completes.
 833  */
 834 static void raise_barrier(struct r1conf *conf)
 835 {
 836         spin_lock_irq(&conf->resync_lock);
 837
 838         /* Wait until no block IO is waiting */
 839         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 840                             conf->resync_lock);
 841
 842         /* block any new IO from starting */
 843         conf->barrier++;
 844
 845         /* For these conditions we must wait:
 846          * A: while the array is in frozen state
 847          * B: while barrier >= RESYNC_DEPTH, meaning resync reach
 848          *    the max count which allowed.
 849          * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
 850          *    next resync will reach to the window which normal bios are
 851          *    handling.
 852          */
 853         wait_event_lock_irq(conf->wait_barrier,
 854                             !conf->array_frozen &&
 855                             conf->barrier < RESYNC_DEPTH &&
 856                             (conf->start_next_window >=
 857                              conf->next_resync + RESYNC_SECTORS),
 858                             conf->resync_lock);
 859
 860         spin_unlock_irq(&conf->resync_lock);
 861 }
 862
 863 static void lower_barrier(struct r1conf *conf)
 864 {
 865         unsigned long flags;
 866         BUG_ON(conf->barrier <= 0);
 867         spin_lock_irqsave(&conf->resync_lock, flags);
 868         conf->barrier--;
 869         spin_unlock_irqrestore(&conf->resync_lock, flags);
 870         wake_up(&conf->wait_barrier);
 871 }
 872
 873 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio)
 874 {
 875         bool wait = false;
 876
 877         if (conf->array_frozen || !bio)
 878                 wait = true;
 879         else if (conf->barrier && bio_data_dir(bio) == WRITE) {
 880                 if (conf->next_resync < RESYNC_WINDOW_SECTORS)
 881                         wait = true;
 882                 else if ((conf->next_resync - RESYNC_WINDOW_SECTORS
 883                                 >= bio_end_sector(bio)) ||
 884                          (conf->next_resync + NEXT_NORMALIO_DISTANCE
 885                                 <= bio->bi_iter.bi_sector))
 886                         wait = false;
 887                 else
 888                         wait = true;
 889         }
 890
 891         return wait;
 892 }
 893
 894 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
 895 {
 896         sector_t sector = 0;
 897
 898         spin_lock_irq(&conf->resync_lock);
 899         if (need_to_wait_for_sync(conf, bio)) {
 900                 conf->nr_waiting++;
 901                 /* Wait for the barrier to drop.
 902                  * However if there are already pending
 903                  * requests (preventing the barrier from
 904                  * rising completely), and the
 905                  * pre-process bio queue isn't empty,
 906                  * then don't wait, as we need to empty
 907                  * that queue to get the nr_pending
 908                  * count down.
 909                  */
 910                 wait_event_lock_irq(conf->wait_barrier,
 911                                     !conf->array_frozen &&
 912                                     (!conf->barrier ||
 913                                     ((conf->start_next_window <
 914                                       conf->next_resync + RESYNC_SECTORS) &&
 915                                      current->bio_list &&
 916                                      !bio_list_empty(current->bio_list))),
 917                                     conf->resync_lock);
 918                 conf->nr_waiting--;
 919         }
 920
 921         if (bio && bio_data_dir(bio) == WRITE) {
 922                 if (conf->next_resync + NEXT_NORMALIO_DISTANCE
 923                     <= bio->bi_iter.bi_sector) {
 924                         if (conf->start_next_window == MaxSector)
 925                                 conf->start_next_window =
 926                                         conf->next_resync +
 927                                         NEXT_NORMALIO_DISTANCE;
 928
 929                         if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
 930                             <= bio->bi_iter.bi_sector)
 931                                 conf->next_window_requests++;
 932                         else
 933                                 conf->current_window_requests++;
 934                         sector = conf->start_next_window;
 935                 }
 936         }
 937
 938         conf->nr_pending++;
 939         spin_unlock_irq(&conf->resync_lock);
 940         return sector;
 941 }
 942
 943 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
 944                           sector_t bi_sector)
 945 {
 946         unsigned long flags;
 947
 948         spin_lock_irqsave(&conf->resync_lock, flags);
 949         conf->nr_pending--;
 950         if (start_next_window) {
 951                 if (start_next_window == conf->start_next_window) {
 952                         if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
 953                             <= bi_sector)
 954                                 conf->next_window_requests--;
 955                         else
 956                                 conf->current_window_requests--;
 957                 } else
 958                         conf->current_window_requests--;
 959
 960                 if (!conf->current_window_requests) {
 961                         if (conf->next_window_requests) {
 962                                 conf->current_window_requests =
 963                                         conf->next_window_requests;
 964                                 conf->next_window_requests = 0;
 965                                 conf->start_next_window +=
 966                                         NEXT_NORMALIO_DISTANCE;
 967                         } else
 968                                 conf->start_next_window = MaxSector;
 969                 }
 970         }
 971         spin_unlock_irqrestore(&conf->resync_lock, flags);
 972         wake_up(&conf->wait_barrier);
 973 }
 974
 975 static void freeze_array(struct r1conf *conf, int extra)
 976 {
 977         /* stop syncio and normal IO and wait for everything to
 978          * go quite.
 979          * We wait until nr_pending match nr_queued+extra
 980          * This is called in the context of one normal IO request
 981          * that has failed. Thus any sync request that might be pending
 982          * will be blocked by nr_pending, and we need to wait for
 983          * pending IO requests to complete or be queued for re-try.
 984          * Thus the number queued (nr_queued) plus this request (extra)
 985          * must match the number of pending IOs (nr_pending) before
 986          * we continue.
 987          */
 988         spin_lock_irq(&conf->resync_lock);
 989         conf->array_frozen = 1;
 990         wait_event_lock_irq_cmd(conf->wait_barrier,
 991                                 conf->nr_pending == conf->nr_queued+extra,
 992                                 conf->resync_lock,
 993                                 flush_pending_writes(conf));
 994         spin_unlock_irq(&conf->resync_lock);
 995 }
 996 static void unfreeze_array(struct r1conf *conf)
 997 {
 998         /* reverse the effect of the freeze */
 999         spin_lock_irq(&conf->resync_lock);
1000         conf->array_frozen = 0;
1001         wake_up(&conf->wait_barrier);
1002         spin_unlock_irq(&conf->resync_lock);
1003 }
1004
1005
1006 /* duplicate the data pages for behind I/O
1007  */
1008 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
1009 {
1010         int i;
1011         struct bio_vec *bvec;
1012         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
1013                                         GFP_NOIO);
1014         if (unlikely(!bvecs))
1015                 return;
1016
1017         bio_for_each_segment_all(bvec, bio, i) {
1018                 bvecs[i] = *bvec;
1019                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
1020                 if (unlikely(!bvecs[i].bv_page))
1021                         goto do_sync_io;
1022                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
1023                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
1024                 kunmap(bvecs[i].bv_page);
1025                 kunmap(bvec->bv_page);
1026         }
1027         r1_bio->behind_bvecs = bvecs;
1028         r1_bio->behind_page_count = bio->bi_vcnt;
1029         set_bit(R1BIO_BehindIO, &r1_bio->state);
1030         return;
1031
1032 do_sync_io:
1033         for (i = 0; i < bio->bi_vcnt; i++)
1034                 if (bvecs[i].bv_page)
1035                         put_page(bvecs[i].bv_page);
1036         kfree(bvecs);
1037         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1038                  bio->bi_iter.bi_size);
1039 }
1040
1041 struct raid1_plug_cb {
1042         struct blk_plug_cb      cb;
1043         struct bio_list         pending;
1044         int                     pending_cnt;
1045 };
1046
1047 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1048 {
1049         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1050                                                   cb);
1051         struct mddev *mddev = plug->cb.data;
1052         struct r1conf *conf = mddev->private;
1053         struct bio *bio;
1054
1055         if (from_schedule || current->bio_list) {
1056                 spin_lock_irq(&conf->device_lock);
1057                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1058                 conf->pending_count += plug->pending_cnt;
1059                 spin_unlock_irq(&conf->device_lock);
1060                 wake_up(&conf->wait_barrier);
1061                 md_wakeup_thread(mddev->thread);
1062                 kfree(plug);
1063                 return;
1064         }
1065
1066         /* we aren't scheduling, so we can do the write-out directly. */
1067         bio = bio_list_get(&plug->pending);
1068         bitmap_unplug(mddev->bitmap);
1069         wake_up(&conf->wait_barrier);
1070
1071         while (bio) { /* submit pending writes */
1072                 struct bio *next = bio->bi_next;
1073                 bio->bi_next = NULL;
1074                 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1075                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1076                         /* Just ignore it */
1077                         bio_endio(bio, 0);
1078                 else
1079                         generic_make_request(bio);
1080                 bio = next;
1081         }
1082         kfree(plug);
1083 }
1084
1085 static void make_request(struct mddev *mddev, struct bio * bio)
1086 {
1087         struct r1conf *conf = mddev->private;
1088         struct raid1_info *mirror;
1089         struct r1bio *r1_bio;
1090         struct bio *read_bio;
1091         int i, disks;
1092         struct bitmap *bitmap;
1093         unsigned long flags;
1094         const int rw = bio_data_dir(bio);
1095         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1096         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1097         const unsigned long do_discard = (bio->bi_rw
1098                                           & (REQ_DISCARD | REQ_SECURE));
1099         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1100         struct md_rdev *blocked_rdev;
1101         struct blk_plug_cb *cb;
1102         struct raid1_plug_cb *plug = NULL;
1103         int first_clone;
1104         int sectors_handled;
1105         int max_sectors;
1106         sector_t start_next_window;
1107
1108         /*
1109          * Register the new request and wait if the reconstruction
1110          * thread has put up a bar for new requests.
1111          * Continue immediately if no resync is active currently.
1112          */
1113
1114         md_write_start(mddev, bio); /* wait on superblock update early */
1115
1116         if (bio_data_dir(bio) == WRITE &&
1117             bio_end_sector(bio) > mddev->suspend_lo &&
1118             bio->bi_iter.bi_sector < mddev->suspend_hi) {
1119                 /* As the suspend_* range is controlled by
1120                  * userspace, we want an interruptible
1121                  * wait.
1122                  */
1123                 DEFINE_WAIT(w);
1124                 for (;;) {
1125                         flush_signals(current);
1126                         prepare_to_wait(&conf->wait_barrier,
1127                                         &w, TASK_INTERRUPTIBLE);
1128                         if (bio_end_sector(bio) <= mddev->suspend_lo ||
1129                             bio->bi_iter.bi_sector >= mddev->suspend_hi)
1130                                 break;
1131                         schedule();
1132                 }
1133                 finish_wait(&conf->wait_barrier, &w);
1134         }
1135
1136         start_next_window = wait_barrier(conf, bio);
1137
1138         bitmap = mddev->bitmap;
1139
1140         /*
1141          * make_request() can abort the operation when READA is being
1142          * used and no empty request is available.
1143          *
1144          */
1145         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1146
1147         r1_bio->master_bio = bio;
1148         r1_bio->sectors = bio_sectors(bio);
1149         r1_bio->state = 0;
1150         r1_bio->mddev = mddev;
1151         r1_bio->sector = bio->bi_iter.bi_sector;
1152
1153         /* We might need to issue multiple reads to different
1154          * devices if there are bad blocks around, so we keep
1155          * track of the number of reads in bio->bi_phys_segments.
1156          * If this is 0, there is only one r1_bio and no locking
1157          * will be needed when requests complete.  If it is
1158          * non-zero, then it is the number of not-completed requests.
1159          */
1160         bio->bi_phys_segments = 0;
1161         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1162
1163         if (rw == READ) {
1164                 /*
1165                  * read balancing logic:
1166                  */
1167                 int rdisk;
1168
1169 read_again:
1170                 rdisk = read_balance(conf, r1_bio, &max_sectors);
1171
1172                 if (rdisk < 0) {
1173                         /* couldn't find anywhere to read from */
1174                         raid_end_bio_io(r1_bio);
1175                         return;
1176                 }
1177                 mirror = conf->mirrors + rdisk;
1178
1179                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1180                     bitmap) {
1181                         /* Reading from a write-mostly device must
1182                          * take care not to over-take any writes
1183                          * that are 'behind'
1184                          */
1185                         wait_event(bitmap->behind_wait,
1186                                    atomic_read(&bitmap->behind_writes) == 0);
1187                 }
1188                 r1_bio->read_disk = rdisk;
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
1475 static void error(struct mddev *mddev, struct md_rdev *rdev)
1476 {
1477         char b[BDEVNAME_SIZE];
1478         struct r1conf *conf = mddev->private;
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         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1499                 unsigned long flags;
1500                 spin_lock_irqsave(&conf->device_lock, flags);
1501                 mddev->degraded++;
1502                 set_bit(Faulty, &rdev->flags);
1503                 spin_unlock_irqrestore(&conf->device_lock, flags);
1504                 /*
1505                  * if recovery is running, make sure it aborts.
1506                  */
1507                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1508         } else
1509                 set_bit(Faulty, &rdev->flags);
1510         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1511         printk(KERN_ALERT
1512                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1513                "md/raid1:%s: Operation continuing on %d devices.\n",
1514                mdname(mddev), bdevname(rdev->bdev, b),
1515                mdname(mddev), conf->raid_disks - mddev->degraded);
1516 }
1517
1518 static void print_conf(struct r1conf *conf)
1519 {
1520         int i;
1521
1522         printk(KERN_DEBUG "RAID1 conf printout:\n");
1523         if (!conf) {
1524                 printk(KERN_DEBUG "(!conf)\n");
1525                 return;
1526         }
1527         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1528                 conf->raid_disks);
1529
1530         rcu_read_lock();
1531         for (i = 0; i < conf->raid_disks; i++) {
1532                 char b[BDEVNAME_SIZE];
1533                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1534                 if (rdev)
1535                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1536                                i, !test_bit(In_sync, &rdev->flags),
1537                                !test_bit(Faulty, &rdev->flags),
1538                                bdevname(rdev->bdev,b));
1539         }
1540         rcu_read_unlock();
1541 }
1542
1543 static void close_sync(struct r1conf *conf)
1544 {
1545         wait_barrier(conf, NULL);
1546         allow_barrier(conf, 0, 0);
1547
1548         mempool_destroy(conf->r1buf_pool);
1549         conf->r1buf_pool = NULL;
1550
1551         conf->next_resync = 0;
1552         conf->start_next_window = MaxSector;
1553 }
1554
1555 static int raid1_spare_active(struct mddev *mddev)
1556 {
1557         int i;
1558         struct r1conf *conf = mddev->private;
1559         int count = 0;
1560         unsigned long flags;
1561
1562         /*
1563          * Find all failed disks within the RAID1 configuration
1564          * and mark them readable.
1565          * Called under mddev lock, so rcu protection not needed.
1566          */
1567         for (i = 0; i < conf->raid_disks; i++) {
1568                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1569                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1570                 if (repl
1571                     && repl->recovery_offset == MaxSector
1572                     && !test_bit(Faulty, &repl->flags)
1573                     && !test_and_set_bit(In_sync, &repl->flags)) {
1574                         /* replacement has just become active */
1575                         if (!rdev ||
1576                             !test_and_clear_bit(In_sync, &rdev->flags))
1577                                 count++;
1578                         if (rdev) {
1579                                 /* Replaced device not technically
1580                                  * faulty, but we need to be sure
1581                                  * it gets removed and never re-added
1582                                  */
1583                                 set_bit(Faulty, &rdev->flags);
1584                                 sysfs_notify_dirent_safe(
1585                                         rdev->sysfs_state);
1586                         }
1587                 }
1588                 if (rdev
1589                     && rdev->recovery_offset == MaxSector
1590                     && !test_bit(Faulty, &rdev->flags)
1591                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1592                         count++;
1593                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1594                 }
1595         }
1596         spin_lock_irqsave(&conf->device_lock, flags);
1597         mddev->degraded -= count;
1598         spin_unlock_irqrestore(&conf->device_lock, flags);
1599
1600         print_conf(conf);
1601         return count;
1602 }
1603
1604
1605 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1606 {
1607         struct r1conf *conf = mddev->private;
1608         int err = -EEXIST;
1609         int mirror = 0;
1610         struct raid1_info *p;
1611         int first = 0;
1612         int last = conf->raid_disks - 1;
1613         struct request_queue *q = bdev_get_queue(rdev->bdev);
1614
1615         if (mddev->recovery_disabled == conf->recovery_disabled)
1616                 return -EBUSY;
1617
1618         if (rdev->raid_disk >= 0)
1619                 first = last = rdev->raid_disk;
1620
1621         if (q->merge_bvec_fn) {
1622                 set_bit(Unmerged, &rdev->flags);
1623                 mddev->merge_check_needed = 1;
1624         }
1625
1626         for (mirror = first; mirror <= last; mirror++) {
1627                 p = conf->mirrors+mirror;
1628                 if (!p->rdev) {
1629
1630                         if (mddev->gendisk)
1631                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1632                                                   rdev->data_offset << 9);
1633
1634                         p->head_position = 0;
1635                         rdev->raid_disk = mirror;
1636                         err = 0;
1637                         /* As all devices are equivalent, we don't need a full recovery
1638                          * if this was recently any drive of the array
1639                          */
1640                         if (rdev->saved_raid_disk < 0)
1641                                 conf->fullsync = 1;
1642                         rcu_assign_pointer(p->rdev, rdev);
1643                         break;
1644                 }
1645                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1646                     p[conf->raid_disks].rdev == NULL) {
1647                         /* Add this device as a replacement */
1648                         clear_bit(In_sync, &rdev->flags);
1649                         set_bit(Replacement, &rdev->flags);
1650                         rdev->raid_disk = mirror;
1651                         err = 0;
1652                         conf->fullsync = 1;
1653                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1654                         break;
1655                 }
1656         }
1657         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1658                 /* Some requests might not have seen this new
1659                  * merge_bvec_fn.  We must wait for them to complete
1660                  * before merging the device fully.
1661                  * First we make sure any code which has tested
1662                  * our function has submitted the request, then
1663                  * we wait for all outstanding requests to complete.
1664                  */
1665                 synchronize_sched();
1666                 freeze_array(conf, 0);
1667                 unfreeze_array(conf);
1668                 clear_bit(Unmerged, &rdev->flags);
1669         }
1670         md_integrity_add_rdev(rdev, mddev);
1671         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1672                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1673         print_conf(conf);
1674         return err;
1675 }
1676
1677 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1678 {
1679         struct r1conf *conf = mddev->private;
1680         int err = 0;
1681         int number = rdev->raid_disk;
1682         struct raid1_info *p = conf->mirrors + number;
1683
1684         if (rdev != p->rdev)
1685                 p = conf->mirrors + conf->raid_disks + number;
1686
1687         print_conf(conf);
1688         if (rdev == p->rdev) {
1689                 if (test_bit(In_sync, &rdev->flags) ||
1690                     atomic_read(&rdev->nr_pending)) {
1691                         err = -EBUSY;
1692                         goto abort;
1693                 }
1694                 /* Only remove non-faulty devices if recovery
1695                  * is not possible.
1696                  */
1697                 if (!test_bit(Faulty, &rdev->flags) &&
1698                     mddev->recovery_disabled != conf->recovery_disabled &&
1699                     mddev->degraded < conf->raid_disks) {
1700                         err = -EBUSY;
1701                         goto abort;
1702                 }
1703                 p->rdev = NULL;
1704                 synchronize_rcu();
1705                 if (atomic_read(&rdev->nr_pending)) {
1706                         /* lost the race, try later */
1707                         err = -EBUSY;
1708                         p->rdev = rdev;
1709                         goto abort;
1710                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1711                         /* We just removed a device that is being replaced.
1712                          * Move down the replacement.  We drain all IO before
1713                          * doing this to avoid confusion.
1714                          */
1715                         struct md_rdev *repl =
1716                                 conf->mirrors[conf->raid_disks + number].rdev;
1717                         freeze_array(conf, 0);
1718                         clear_bit(Replacement, &repl->flags);
1719                         p->rdev = repl;
1720                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1721                         unfreeze_array(conf);
1722                         clear_bit(WantReplacement, &rdev->flags);
1723                 } else
1724                         clear_bit(WantReplacement, &rdev->flags);
1725                 err = md_integrity_register(mddev);
1726         }
1727 abort:
1728
1729         print_conf(conf);
1730         return err;
1731 }
1732
1733
1734 static void end_sync_read(struct bio *bio, int error)
1735 {
1736         struct r1bio *r1_bio = bio->bi_private;
1737
1738         update_head_pos(r1_bio->read_disk, r1_bio);
1739
1740         /*
1741          * we have read a block, now it needs to be re-written,
1742          * or re-read if the read failed.
1743          * We don't do much here, just schedule handling by raid1d
1744          */
1745         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1746                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1747
1748         if (atomic_dec_and_test(&r1_bio->remaining))
1749                 reschedule_retry(r1_bio);
1750 }
1751
1752 static void end_sync_write(struct bio *bio, int error)
1753 {
1754         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1755         struct r1bio *r1_bio = bio->bi_private;
1756         struct mddev *mddev = r1_bio->mddev;
1757         struct r1conf *conf = mddev->private;
1758         int mirror=0;
1759         sector_t first_bad;
1760         int bad_sectors;
1761
1762         mirror = find_bio_disk(r1_bio, bio);
1763
1764         if (!uptodate) {
1765                 sector_t sync_blocks = 0;
1766                 sector_t s = r1_bio->sector;
1767                 long sectors_to_go = r1_bio->sectors;
1768                 /* make sure these bits doesn't get cleared. */
1769                 do {
1770                         bitmap_end_sync(mddev->bitmap, s,
1771                                         &sync_blocks, 1);
1772                         s += sync_blocks;
1773                         sectors_to_go -= sync_blocks;
1774                 } while (sectors_to_go > 0);
1775                 set_bit(WriteErrorSeen,
1776                         &conf->mirrors[mirror].rdev->flags);
1777                 if (!test_and_set_bit(WantReplacement,
1778                                       &conf->mirrors[mirror].rdev->flags))
1779                         set_bit(MD_RECOVERY_NEEDED, &
1780                                 mddev->recovery);
1781                 set_bit(R1BIO_WriteError, &r1_bio->state);
1782         } else if (is_badblock(conf->mirrors[mirror].rdev,
1783                                r1_bio->sector,
1784                                r1_bio->sectors,
1785                                &first_bad, &bad_sectors) &&
1786                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1787                                 r1_bio->sector,
1788                                 r1_bio->sectors,
1789                                 &first_bad, &bad_sectors)
1790                 )
1791                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1792
1793         if (atomic_dec_and_test(&r1_bio->remaining)) {
1794                 int s = r1_bio->sectors;
1795                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1796                     test_bit(R1BIO_WriteError, &r1_bio->state))
1797                         reschedule_retry(r1_bio);
1798                 else {
1799                         put_buf(r1_bio);
1800                         md_done_sync(mddev, s, uptodate);
1801                 }
1802         }
1803 }
1804
1805 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1806                             int sectors, struct page *page, int rw)
1807 {
1808         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1809                 /* success */
1810                 return 1;
1811         if (rw == WRITE) {
1812                 set_bit(WriteErrorSeen, &rdev->flags);
1813                 if (!test_and_set_bit(WantReplacement,
1814                                       &rdev->flags))
1815                         set_bit(MD_RECOVERY_NEEDED, &
1816                                 rdev->mddev->recovery);
1817         }
1818         /* need to record an error - either for the block or the device */
1819         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1820                 md_error(rdev->mddev, rdev);
1821         return 0;
1822 }
1823
1824 static int fix_sync_read_error(struct r1bio *r1_bio)
1825 {
1826         /* Try some synchronous reads of other devices to get
1827          * good data, much like with normal read errors.  Only
1828          * read into the pages we already have so we don't
1829          * need to re-issue the read request.
1830          * We don't need to freeze the array, because being in an
1831          * active sync request, there is no normal IO, and
1832          * no overlapping syncs.
1833          * We don't need to check is_badblock() again as we
1834          * made sure that anything with a bad block in range
1835          * will have bi_end_io clear.
1836          */
1837         struct mddev *mddev = r1_bio->mddev;
1838         struct r1conf *conf = mddev->private;
1839         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1840         sector_t sect = r1_bio->sector;
1841         int sectors = r1_bio->sectors;
1842         int idx = 0;
1843
1844         while(sectors) {
1845                 int s = sectors;
1846                 int d = r1_bio->read_disk;
1847                 int success = 0;
1848                 struct md_rdev *rdev;
1849                 int start;
1850
1851                 if (s > (PAGE_SIZE>>9))
1852                         s = PAGE_SIZE >> 9;
1853                 do {
1854                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1855                                 /* No rcu protection needed here devices
1856                                  * can only be removed when no resync is
1857                                  * active, and resync is currently active
1858                                  */
1859                                 rdev = conf->mirrors[d].rdev;
1860                                 if (sync_page_io(rdev, sect, s<<9,
1861                                                  bio->bi_io_vec[idx].bv_page,
1862                                                  READ, false)) {
1863                                         success = 1;
1864                                         break;
1865                                 }
1866                         }
1867                         d++;
1868                         if (d == conf->raid_disks * 2)
1869                                 d = 0;
1870                 } while (!success && d != r1_bio->read_disk);
1871
1872                 if (!success) {
1873                         char b[BDEVNAME_SIZE];
1874                         int abort = 0;
1875                         /* Cannot read from anywhere, this block is lost.
1876                          * Record a bad block on each device.  If that doesn't
1877                          * work just disable and interrupt the recovery.
1878                          * Don't fail devices as that won't really help.
1879                          */
1880                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1881                                " for block %llu\n",
1882                                mdname(mddev),
1883                                bdevname(bio->bi_bdev, b),
1884                                (unsigned long long)r1_bio->sector);
1885                         for (d = 0; d < conf->raid_disks * 2; d++) {
1886                                 rdev = conf->mirrors[d].rdev;
1887                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1888                                         continue;
1889                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1890                                         abort = 1;
1891                         }
1892                         if (abort) {
1893                                 conf->recovery_disabled =
1894                                         mddev->recovery_disabled;
1895                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1896                                 md_done_sync(mddev, r1_bio->sectors, 0);
1897                                 put_buf(r1_bio);
1898                                 return 0;
1899                         }
1900                         /* Try next page */
1901                         sectors -= s;
1902                         sect += s;
1903                         idx++;
1904                         continue;
1905                 }
1906
1907                 start = d;
1908                 /* write it back and re-read */
1909                 while (d != r1_bio->read_disk) {
1910                         if (d == 0)
1911                                 d = conf->raid_disks * 2;
1912                         d--;
1913                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1914                                 continue;
1915                         rdev = conf->mirrors[d].rdev;
1916                         if (r1_sync_page_io(rdev, sect, s,
1917                                             bio->bi_io_vec[idx].bv_page,
1918                                             WRITE) == 0) {
1919                                 r1_bio->bios[d]->bi_end_io = NULL;
1920                                 rdev_dec_pending(rdev, mddev);
1921                         }
1922                 }
1923                 d = start;
1924                 while (d != r1_bio->read_disk) {
1925                         if (d == 0)
1926                                 d = conf->raid_disks * 2;
1927                         d--;
1928                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1929                                 continue;
1930                         rdev = conf->mirrors[d].rdev;
1931                         if (r1_sync_page_io(rdev, sect, s,
1932                                             bio->bi_io_vec[idx].bv_page,
1933                                             READ) != 0)
1934                                 atomic_add(s, &rdev->corrected_errors);
1935                 }
1936                 sectors -= s;
1937                 sect += s;
1938                 idx ++;
1939         }
1940         set_bit(R1BIO_Uptodate, &r1_bio->state);
1941         set_bit(BIO_UPTODATE, &bio->bi_flags);
1942         return 1;
1943 }
1944
1945 static int process_checks(struct r1bio *r1_bio)
1946 {
1947         /* We have read all readable devices.  If we haven't
1948          * got the block, then there is no hope left.
1949          * If we have, then we want to do a comparison
1950          * and skip the write if everything is the same.
1951          * If any blocks failed to read, then we need to
1952          * attempt an over-write
1953          */
1954         struct mddev *mddev = r1_bio->mddev;
1955         struct r1conf *conf = mddev->private;
1956         int primary;
1957         int i;
1958         int vcnt;
1959
1960         /* Fix variable parts of all bios */
1961         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1962         for (i = 0; i < conf->raid_disks * 2; i++) {
1963                 int j;
1964                 int size;
1965                 int uptodate;
1966                 struct bio *b = r1_bio->bios[i];
1967                 if (b->bi_end_io != end_sync_read)
1968                         continue;
1969                 /* fixup the bio for reuse, but preserve BIO_UPTODATE */
1970                 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags);
1971                 bio_reset(b);
1972                 if (!uptodate)
1973                         clear_bit(BIO_UPTODATE, &b->bi_flags);
1974                 b->bi_vcnt = vcnt;
1975                 b->bi_iter.bi_size = r1_bio->sectors << 9;
1976                 b->bi_iter.bi_sector = r1_bio->sector +
1977                         conf->mirrors[i].rdev->data_offset;
1978                 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1979                 b->bi_end_io = end_sync_read;
1980                 b->bi_private = r1_bio;
1981
1982                 size = b->bi_iter.bi_size;
1983                 for (j = 0; j < vcnt ; j++) {
1984                         struct bio_vec *bi;
1985                         bi = &b->bi_io_vec[j];
1986                         bi->bv_offset = 0;
1987                         if (size > PAGE_SIZE)
1988                                 bi->bv_len = PAGE_SIZE;
1989                         else
1990                                 bi->bv_len = size;
1991                         size -= PAGE_SIZE;
1992                 }
1993         }
1994         for (primary = 0; primary < conf->raid_disks * 2; primary++)
1995                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1996                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1997                         r1_bio->bios[primary]->bi_end_io = NULL;
1998                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1999                         break;
2000                 }
2001         r1_bio->read_disk = primary;
2002         for (i = 0; i < conf->raid_disks * 2; i++) {
2003                 int j;
2004                 struct bio *pbio = r1_bio->bios[primary];
2005                 struct bio *sbio = r1_bio->bios[i];
2006                 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags);
2007
2008                 if (sbio->bi_end_io != end_sync_read)
2009                         continue;
2010                 /* Now we can 'fixup' the BIO_UPTODATE flag */
2011                 set_bit(BIO_UPTODATE, &sbio->bi_flags);
2012
2013                 if (uptodate) {
2014                         for (j = vcnt; j-- ; ) {
2015                                 struct page *p, *s;
2016                                 p = pbio->bi_io_vec[j].bv_page;
2017                                 s = sbio->bi_io_vec[j].bv_page;
2018                                 if (memcmp(page_address(p),
2019                                            page_address(s),
2020                                            sbio->bi_io_vec[j].bv_len))
2021                                         break;
2022                         }
2023                 } else
2024                         j = 0;
2025                 if (j >= 0)
2026                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2027                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2028                               && uptodate)) {
2029                         /* No need to write to this device. */
2030                         sbio->bi_end_io = NULL;
2031                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2032                         continue;
2033                 }
2034
2035                 bio_copy_data(sbio, pbio);
2036         }
2037         return 0;
2038 }
2039
2040 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2041 {
2042         struct r1conf *conf = mddev->private;
2043         int i;
2044         int disks = conf->raid_disks * 2;
2045         struct bio *bio, *wbio;
2046
2047         bio = r1_bio->bios[r1_bio->read_disk];
2048
2049         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2050                 /* ouch - failed to read all of that. */
2051                 if (!fix_sync_read_error(r1_bio))
2052                         return;
2053
2054         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2055                 if (process_checks(r1_bio) < 0)
2056                         return;
2057         /*
2058          * schedule writes
2059          */
2060         atomic_set(&r1_bio->remaining, 1);
2061         for (i = 0; i < disks ; i++) {
2062                 wbio = r1_bio->bios[i];
2063                 if (wbio->bi_end_io == NULL ||
2064                     (wbio->bi_end_io == end_sync_read &&
2065                      (i == r1_bio->read_disk ||
2066                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2067                         continue;
2068
2069                 wbio->bi_rw = WRITE;
2070                 wbio->bi_end_io = end_sync_write;
2071                 atomic_inc(&r1_bio->remaining);
2072                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2073
2074                 generic_make_request(wbio);
2075         }
2076
2077         if (atomic_dec_and_test(&r1_bio->remaining)) {
2078                 /* if we're here, all write(s) have completed, so clean up */
2079                 int s = r1_bio->sectors;
2080                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2081                     test_bit(R1BIO_WriteError, &r1_bio->state))
2082                         reschedule_retry(r1_bio);
2083                 else {
2084                         put_buf(r1_bio);
2085                         md_done_sync(mddev, s, 1);
2086                 }
2087         }
2088 }
2089
2090 /*
2091  * This is a kernel thread which:
2092  *
2093  *      1.      Retries failed read operations on working mirrors.
2094  *      2.      Updates the raid superblock when problems encounter.
2095  *      3.      Performs writes following reads for array synchronising.
2096  */
2097
2098 static void fix_read_error(struct r1conf *conf, int read_disk,
2099                            sector_t sect, int sectors)
2100 {
2101         struct mddev *mddev = conf->mddev;
2102         while(sectors) {
2103                 int s = sectors;
2104                 int d = read_disk;
2105                 int success = 0;
2106                 int start;
2107                 struct md_rdev *rdev;
2108
2109                 if (s > (PAGE_SIZE>>9))
2110                         s = PAGE_SIZE >> 9;
2111
2112                 do {
2113                         /* Note: no rcu protection needed here
2114                          * as this is synchronous in the raid1d thread
2115                          * which is the thread that might remove
2116                          * a device.  If raid1d ever becomes multi-threaded....
2117                          */
2118                         sector_t first_bad;
2119                         int bad_sectors;
2120
2121                         rdev = conf->mirrors[d].rdev;
2122                         if (rdev &&
2123                             (test_bit(In_sync, &rdev->flags) ||
2124                              (!test_bit(Faulty, &rdev->flags) &&
2125                               rdev->recovery_offset >= sect + s)) &&
2126                             is_badblock(rdev, sect, s,
2127                                         &first_bad, &bad_sectors) == 0 &&
2128                             sync_page_io(rdev, sect, s<<9,
2129                                          conf->tmppage, READ, false))
2130                                 success = 1;
2131                         else {
2132                                 d++;
2133                                 if (d == conf->raid_disks * 2)
2134                                         d = 0;
2135                         }
2136                 } while (!success && d != read_disk);
2137
2138                 if (!success) {
2139                         /* Cannot read from anywhere - mark it bad */
2140                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2141                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2142                                 md_error(mddev, rdev);
2143                         break;
2144                 }
2145                 /* write it back and re-read */
2146                 start = d;
2147                 while (d != read_disk) {
2148                         if (d==0)
2149                                 d = conf->raid_disks * 2;
2150                         d--;
2151                         rdev = conf->mirrors[d].rdev;
2152                         if (rdev &&
2153                             test_bit(In_sync, &rdev->flags))
2154                                 r1_sync_page_io(rdev, sect, s,
2155                                                 conf->tmppage, WRITE);
2156                 }
2157                 d = start;
2158                 while (d != read_disk) {
2159                         char b[BDEVNAME_SIZE];
2160                         if (d==0)
2161                                 d = conf->raid_disks * 2;
2162                         d--;
2163                         rdev = conf->mirrors[d].rdev;
2164                         if (rdev &&
2165                             test_bit(In_sync, &rdev->flags)) {
2166                                 if (r1_sync_page_io(rdev, sect, s,
2167                                                     conf->tmppage, READ)) {
2168                                         atomic_add(s, &rdev->corrected_errors);
2169                                         printk(KERN_INFO
2170                                                "md/raid1:%s: read error corrected "
2171                                                "(%d sectors at %llu on %s)\n",
2172                                                mdname(mddev), s,
2173                                                (unsigned long long)(sect +
2174                                                    rdev->data_offset),
2175                                                bdevname(rdev->bdev, b));
2176                                 }
2177                         }
2178                 }
2179                 sectors -= s;
2180                 sect += s;
2181         }
2182 }
2183
2184 static int narrow_write_error(struct r1bio *r1_bio, int i)
2185 {
2186         struct mddev *mddev = r1_bio->mddev;
2187         struct r1conf *conf = mddev->private;
2188         struct md_rdev *rdev = conf->mirrors[i].rdev;
2189
2190         /* bio has the data to be written to device 'i' where
2191          * we just recently had a write error.
2192          * We repeatedly clone the bio and trim down to one block,
2193          * then try the write.  Where the write fails we record
2194          * a bad block.
2195          * It is conceivable that the bio doesn't exactly align with
2196          * blocks.  We must handle this somehow.
2197          *
2198          * We currently own a reference on the rdev.
2199          */
2200
2201         int block_sectors;
2202         sector_t sector;
2203         int sectors;
2204         int sect_to_write = r1_bio->sectors;
2205         int ok = 1;
2206
2207         if (rdev->badblocks.shift < 0)
2208                 return 0;
2209
2210         block_sectors = 1 << rdev->badblocks.shift;
2211         sector = r1_bio->sector;
2212         sectors = ((sector + block_sectors)
2213                    & ~(sector_t)(block_sectors - 1))
2214                 - sector;
2215
2216         while (sect_to_write) {
2217                 struct bio *wbio;
2218                 if (sectors > sect_to_write)
2219                         sectors = sect_to_write;
2220                 /* Write at 'sector' for 'sectors'*/
2221
2222                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2223                         unsigned vcnt = r1_bio->behind_page_count;
2224                         struct bio_vec *vec = r1_bio->behind_bvecs;
2225
2226                         while (!vec->bv_page) {
2227                                 vec++;
2228                                 vcnt--;
2229                         }
2230
2231                         wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2232                         memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2233
2234                         wbio->bi_vcnt = vcnt;
2235                 } else {
2236                         wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2237                 }
2238
2239                 wbio->bi_rw = WRITE;
2240                 wbio->bi_iter.bi_sector = r1_bio->sector;
2241                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2242
2243                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2244                 wbio->bi_iter.bi_sector += rdev->data_offset;
2245                 wbio->bi_bdev = rdev->bdev;
2246                 if (submit_bio_wait(WRITE, wbio) == 0)
2247                         /* failure! */
2248                         ok = rdev_set_badblocks(rdev, sector,
2249                                                 sectors, 0)
2250                                 && ok;
2251
2252                 bio_put(wbio);
2253                 sect_to_write -= sectors;
2254                 sector += sectors;
2255                 sectors = block_sectors;
2256         }
2257         return ok;
2258 }
2259
2260 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2261 {
2262         int m;
2263         int s = r1_bio->sectors;
2264         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2265                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2266                 struct bio *bio = r1_bio->bios[m];
2267                 if (bio->bi_end_io == NULL)
2268                         continue;
2269                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2270                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2271                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2272                 }
2273                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2274                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2275                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2276                                 md_error(conf->mddev, rdev);
2277                 }
2278         }
2279         put_buf(r1_bio);
2280         md_done_sync(conf->mddev, s, 1);
2281 }
2282
2283 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2284 {
2285         int m;
2286         for (m = 0; m < conf->raid_disks * 2 ; m++)
2287                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2288                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2289                         rdev_clear_badblocks(rdev,
2290                                              r1_bio->sector,
2291                                              r1_bio->sectors, 0);
2292                         rdev_dec_pending(rdev, conf->mddev);
2293                 } else if (r1_bio->bios[m] != NULL) {
2294                         /* This drive got a write error.  We need to
2295                          * narrow down and record precise write
2296                          * errors.
2297                          */
2298                         if (!narrow_write_error(r1_bio, m)) {
2299                                 md_error(conf->mddev,
2300                                          conf->mirrors[m].rdev);
2301                                 /* an I/O failed, we can't clear the bitmap */
2302                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2303                         }
2304                         rdev_dec_pending(conf->mirrors[m].rdev,
2305                                          conf->mddev);
2306                 }
2307         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2308                 close_write(r1_bio);
2309         raid_end_bio_io(r1_bio);
2310 }
2311
2312 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2313 {
2314         int disk;
2315         int max_sectors;
2316         struct mddev *mddev = conf->mddev;
2317         struct bio *bio;
2318         char b[BDEVNAME_SIZE];
2319         struct md_rdev *rdev;
2320
2321         clear_bit(R1BIO_ReadError, &r1_bio->state);
2322         /* we got a read error. Maybe the drive is bad.  Maybe just
2323          * the block and we can fix it.
2324          * We freeze all other IO, and try reading the block from
2325          * other devices.  When we find one, we re-write
2326          * and check it that fixes the read error.
2327          * This is all done synchronously while the array is
2328          * frozen
2329          */
2330         if (mddev->ro == 0) {
2331                 freeze_array(conf, 1);
2332                 fix_read_error(conf, r1_bio->read_disk,
2333                                r1_bio->sector, r1_bio->sectors);
2334                 unfreeze_array(conf);
2335         } else
2336                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2337         rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2338
2339         bio = r1_bio->bios[r1_bio->read_disk];
2340         bdevname(bio->bi_bdev, b);
2341 read_more:
2342         disk = read_balance(conf, r1_bio, &max_sectors);
2343         if (disk == -1) {
2344                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2345                        " read error for block %llu\n",
2346                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2347                 raid_end_bio_io(r1_bio);
2348         } else {
2349                 const unsigned long do_sync
2350                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2351                 if (bio) {
2352                         r1_bio->bios[r1_bio->read_disk] =
2353                                 mddev->ro ? IO_BLOCKED : NULL;
2354                         bio_put(bio);
2355                 }
2356                 r1_bio->read_disk = disk;
2357                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2358                 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2359                          max_sectors);
2360                 r1_bio->bios[r1_bio->read_disk] = bio;
2361                 rdev = conf->mirrors[disk].rdev;
2362                 printk_ratelimited(KERN_ERR
2363                                    "md/raid1:%s: redirecting sector %llu"
2364                                    " to other mirror: %s\n",
2365                                    mdname(mddev),
2366                                    (unsigned long long)r1_bio->sector,
2367                                    bdevname(rdev->bdev, b));
2368                 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2369                 bio->bi_bdev = rdev->bdev;
2370                 bio->bi_end_io = raid1_end_read_request;
2371                 bio->bi_rw = READ | do_sync;
2372                 bio->bi_private = r1_bio;
2373                 if (max_sectors < r1_bio->sectors) {
2374                         /* Drat - have to split this up more */
2375                         struct bio *mbio = r1_bio->master_bio;
2376                         int sectors_handled = (r1_bio->sector + max_sectors
2377                                                - mbio->bi_iter.bi_sector);
2378                         r1_bio->sectors = max_sectors;
2379                         spin_lock_irq(&conf->device_lock);
2380                         if (mbio->bi_phys_segments == 0)
2381                                 mbio->bi_phys_segments = 2;
2382                         else
2383                                 mbio->bi_phys_segments++;
2384                         spin_unlock_irq(&conf->device_lock);
2385                         generic_make_request(bio);
2386                         bio = NULL;
2387
2388                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2389
2390                         r1_bio->master_bio = mbio;
2391                         r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2392                         r1_bio->state = 0;
2393                         set_bit(R1BIO_ReadError, &r1_bio->state);
2394                         r1_bio->mddev = mddev;
2395                         r1_bio->sector = mbio->bi_iter.bi_sector +
2396                                 sectors_handled;
2397
2398                         goto read_more;
2399                 } else
2400                         generic_make_request(bio);
2401         }
2402 }
2403
2404 static void raid1d(struct md_thread *thread)
2405 {
2406         struct mddev *mddev = thread->mddev;
2407         struct r1bio *r1_bio;
2408         unsigned long flags;
2409         struct r1conf *conf = mddev->private;
2410         struct list_head *head = &conf->retry_list;
2411         struct blk_plug plug;
2412
2413         md_check_recovery(mddev);
2414
2415         blk_start_plug(&plug);
2416         for (;;) {
2417
2418                 flush_pending_writes(conf);
2419
2420                 spin_lock_irqsave(&conf->device_lock, flags);
2421                 if (list_empty(head)) {
2422                         spin_unlock_irqrestore(&conf->device_lock, flags);
2423                         break;
2424                 }
2425                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2426                 list_del(head->prev);
2427                 conf->nr_queued--;
2428                 spin_unlock_irqrestore(&conf->device_lock, flags);
2429
2430                 mddev = r1_bio->mddev;
2431                 conf = mddev->private;
2432                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2433                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2434                             test_bit(R1BIO_WriteError, &r1_bio->state))
2435                                 handle_sync_write_finished(conf, r1_bio);
2436                         else
2437                                 sync_request_write(mddev, r1_bio);
2438                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2439                            test_bit(R1BIO_WriteError, &r1_bio->state))
2440                         handle_write_finished(conf, r1_bio);
2441                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2442                         handle_read_error(conf, r1_bio);
2443                 else
2444                         /* just a partial read to be scheduled from separate
2445                          * context
2446                          */
2447                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2448
2449                 cond_resched();
2450                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2451                         md_check_recovery(mddev);
2452         }
2453         blk_finish_plug(&plug);
2454 }
2455
2456
2457 static int init_resync(struct r1conf *conf)
2458 {
2459         int buffs;
2460
2461         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2462         BUG_ON(conf->r1buf_pool);
2463         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2464                                           conf->poolinfo);
2465         if (!conf->r1buf_pool)
2466                 return -ENOMEM;
2467         conf->next_resync = 0;
2468         return 0;
2469 }
2470
2471 /*
2472  * perform a "sync" on one "block"
2473  *
2474  * We need to make sure that no normal I/O request - particularly write
2475  * requests - conflict with active sync requests.
2476  *
2477  * This is achieved by tracking pending requests and a 'barrier' concept
2478  * that can be installed to exclude normal IO requests.
2479  */
2480
2481 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2482 {
2483         struct r1conf *conf = mddev->private;
2484         struct r1bio *r1_bio;
2485         struct bio *bio;
2486         sector_t max_sector, nr_sectors;
2487         int disk = -1;
2488         int i;
2489         int wonly = -1;
2490         int write_targets = 0, read_targets = 0;
2491         sector_t sync_blocks;
2492         int still_degraded = 0;
2493         int good_sectors = RESYNC_SECTORS;
2494         int min_bad = 0; /* number of sectors that are bad in all devices */
2495
2496         if (!conf->r1buf_pool)
2497                 if (init_resync(conf))
2498                         return 0;
2499
2500         max_sector = mddev->dev_sectors;
2501         if (sector_nr >= max_sector) {
2502                 /* If we aborted, we need to abort the
2503                  * sync on the 'current' bitmap chunk (there will
2504                  * only be one in raid1 resync.
2505                  * We can find the current addess in mddev->curr_resync
2506                  */
2507                 if (mddev->curr_resync < max_sector) /* aborted */
2508                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2509                                                 &sync_blocks, 1);
2510                 else /* completed sync */
2511                         conf->fullsync = 0;
2512
2513                 bitmap_close_sync(mddev->bitmap);
2514                 close_sync(conf);
2515                 return 0;
2516         }
2517
2518         if (mddev->bitmap == NULL &&
2519             mddev->recovery_cp == MaxSector &&
2520             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2521             conf->fullsync == 0) {
2522                 *skipped = 1;
2523                 return max_sector - sector_nr;
2524         }
2525         /* before building a request, check if we can skip these blocks..
2526          * This call the bitmap_start_sync doesn't actually record anything
2527          */
2528         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2529             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2530                 /* We can skip this block, and probably several more */
2531                 *skipped = 1;
2532                 return sync_blocks;
2533         }
2534         /*
2535          * If there is non-resync activity waiting for a turn,
2536          * and resync is going fast enough,
2537          * then let it though before starting on this new sync request.
2538          */
2539         if (!go_faster && conf->nr_waiting)
2540                 msleep_interruptible(1000);
2541
2542         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2543         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2544         raise_barrier(conf);
2545
2546         conf->next_resync = sector_nr;
2547
2548         rcu_read_lock();
2549         /*
2550          * If we get a correctably read error during resync or recovery,
2551          * we might want to read from a different device.  So we
2552          * flag all drives that could conceivably be read from for READ,
2553          * and any others (which will be non-In_sync devices) for WRITE.
2554          * If a read fails, we try reading from something else for which READ
2555          * is OK.
2556          */
2557
2558         r1_bio->mddev = mddev;
2559         r1_bio->sector = sector_nr;
2560         r1_bio->state = 0;
2561         set_bit(R1BIO_IsSync, &r1_bio->state);
2562
2563         for (i = 0; i < conf->raid_disks * 2; i++) {
2564                 struct md_rdev *rdev;
2565                 bio = r1_bio->bios[i];
2566                 bio_reset(bio);
2567
2568                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2569                 if (rdev == NULL ||
2570                     test_bit(Faulty, &rdev->flags)) {
2571                         if (i < conf->raid_disks)
2572                                 still_degraded = 1;
2573                 } else if (!test_bit(In_sync, &rdev->flags)) {
2574                         bio->bi_rw = WRITE;
2575                         bio->bi_end_io = end_sync_write;
2576                         write_targets ++;
2577                 } else {
2578                         /* may need to read from here */
2579                         sector_t first_bad = MaxSector;
2580                         int bad_sectors;
2581
2582                         if (is_badblock(rdev, sector_nr, good_sectors,
2583                                         &first_bad, &bad_sectors)) {
2584                                 if (first_bad > sector_nr)
2585                                         good_sectors = first_bad - sector_nr;
2586                                 else {
2587                                         bad_sectors -= (sector_nr - first_bad);
2588                                         if (min_bad == 0 ||
2589                                             min_bad > bad_sectors)
2590                                                 min_bad = bad_sectors;
2591                                 }
2592                         }
2593                         if (sector_nr < first_bad) {
2594                                 if (test_bit(WriteMostly, &rdev->flags)) {
2595                                         if (wonly < 0)
2596                                                 wonly = i;
2597                                 } else {
2598                                         if (disk < 0)
2599                                                 disk = i;
2600                                 }
2601                                 bio->bi_rw = READ;
2602                                 bio->bi_end_io = end_sync_read;
2603                                 read_targets++;
2604                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2605                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2606                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2607                                 /*
2608                                  * The device is suitable for reading (InSync),
2609                                  * but has bad block(s) here. Let's try to correct them,
2610                                  * if we are doing resync or repair. Otherwise, leave
2611                                  * this device alone for this sync request.
2612                                  */
2613                                 bio->bi_rw = WRITE;
2614                                 bio->bi_end_io = end_sync_write;
2615                                 write_targets++;
2616                         }
2617                 }
2618                 if (bio->bi_end_io) {
2619                         atomic_inc(&rdev->nr_pending);
2620                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2621                         bio->bi_bdev = rdev->bdev;
2622                         bio->bi_private = r1_bio;
2623                 }
2624         }
2625         rcu_read_unlock();
2626         if (disk < 0)
2627                 disk = wonly;
2628         r1_bio->read_disk = disk;
2629
2630         if (read_targets == 0 && min_bad > 0) {
2631                 /* These sectors are bad on all InSync devices, so we
2632                  * need to mark them bad on all write targets
2633                  */
2634                 int ok = 1;
2635                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2636                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2637                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2638                                 ok = rdev_set_badblocks(rdev, sector_nr,
2639                                                         min_bad, 0
2640                                         ) && ok;
2641                         }
2642                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2643                 *skipped = 1;
2644                 put_buf(r1_bio);
2645
2646                 if (!ok) {
2647                         /* Cannot record the badblocks, so need to
2648                          * abort the resync.
2649                          * If there are multiple read targets, could just
2650                          * fail the really bad ones ???
2651                          */
2652                         conf->recovery_disabled = mddev->recovery_disabled;
2653                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2654                         return 0;
2655                 } else
2656                         return min_bad;
2657
2658         }
2659         if (min_bad > 0 && min_bad < good_sectors) {
2660                 /* only resync enough to reach the next bad->good
2661                  * transition */
2662                 good_sectors = min_bad;
2663         }
2664
2665         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2666                 /* extra read targets are also write targets */
2667                 write_targets += read_targets-1;
2668
2669         if (write_targets == 0 || read_targets == 0) {
2670                 /* There is nowhere to write, so all non-sync
2671                  * drives must be failed - so we are finished
2672                  */
2673                 sector_t rv;
2674                 if (min_bad > 0)
2675                         max_sector = sector_nr + min_bad;
2676                 rv = max_sector - sector_nr;
2677                 *skipped = 1;
2678                 put_buf(r1_bio);
2679                 return rv;
2680         }
2681
2682         if (max_sector > mddev->resync_max)
2683                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2684         if (max_sector > sector_nr + good_sectors)
2685                 max_sector = sector_nr + good_sectors;
2686         nr_sectors = 0;
2687         sync_blocks = 0;
2688         do {
2689                 struct page *page;
2690                 int len = PAGE_SIZE;
2691                 if (sector_nr + (len>>9) > max_sector)
2692                         len = (max_sector - sector_nr) << 9;
2693                 if (len == 0)
2694                         break;
2695                 if (sync_blocks == 0) {
2696                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2697                                                &sync_blocks, still_degraded) &&
2698                             !conf->fullsync &&
2699                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2700                                 break;
2701                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2702                         if ((len >> 9) > sync_blocks)
2703                                 len = sync_blocks<<9;
2704                 }
2705
2706                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2707                         bio = r1_bio->bios[i];
2708                         if (bio->bi_end_io) {
2709                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2710                                 if (bio_add_page(bio, page, len, 0) == 0) {
2711                                         /* stop here */
2712                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2713                                         while (i > 0) {
2714                                                 i--;
2715                                                 bio = r1_bio->bios[i];
2716                                                 if (bio->bi_end_io==NULL)
2717                                                         continue;
2718                                                 /* remove last page from this bio */
2719                                                 bio->bi_vcnt--;
2720                                                 bio->bi_iter.bi_size -= len;
2721                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2722                                         }
2723                                         goto bio_full;
2724                                 }
2725                         }
2726                 }
2727                 nr_sectors += len>>9;
2728                 sector_nr += len>>9;
2729                 sync_blocks -= (len>>9);
2730         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2731  bio_full:
2732         r1_bio->sectors = nr_sectors;
2733
2734         /* For a user-requested sync, we read all readable devices and do a
2735          * compare
2736          */
2737         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2738                 atomic_set(&r1_bio->remaining, read_targets);
2739                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2740                         bio = r1_bio->bios[i];
2741                         if (bio->bi_end_io == end_sync_read) {
2742                                 read_targets--;
2743                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2744                                 generic_make_request(bio);
2745                         }
2746                 }
2747         } else {
2748                 atomic_set(&r1_bio->remaining, 1);
2749                 bio = r1_bio->bios[r1_bio->read_disk];
2750                 md_sync_acct(bio->bi_bdev, nr_sectors);
2751                 generic_make_request(bio);
2752
2753         }
2754         return nr_sectors;
2755 }
2756
2757 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2758 {
2759         if (sectors)
2760                 return sectors;
2761
2762         return mddev->dev_sectors;
2763 }
2764
2765 static struct r1conf *setup_conf(struct mddev *mddev)
2766 {
2767         struct r1conf *conf;
2768         int i;
2769         struct raid1_info *disk;
2770         struct md_rdev *rdev;
2771         int err = -ENOMEM;
2772
2773         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2774         if (!conf)
2775                 goto abort;
2776
2777         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2778                                 * mddev->raid_disks * 2,
2779                                  GFP_KERNEL);
2780         if (!conf->mirrors)
2781                 goto abort;
2782
2783         conf->tmppage = alloc_page(GFP_KERNEL);
2784         if (!conf->tmppage)
2785                 goto abort;
2786
2787         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2788         if (!conf->poolinfo)
2789                 goto abort;
2790         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2791         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2792                                           r1bio_pool_free,
2793                                           conf->poolinfo);
2794         if (!conf->r1bio_pool)
2795                 goto abort;
2796
2797         conf->poolinfo->mddev = mddev;
2798
2799         err = -EINVAL;
2800         spin_lock_init(&conf->device_lock);
2801         rdev_for_each(rdev, mddev) {
2802                 struct request_queue *q;
2803                 int disk_idx = rdev->raid_disk;
2804                 if (disk_idx >= mddev->raid_disks
2805                     || disk_idx < 0)
2806                         continue;
2807                 if (test_bit(Replacement, &rdev->flags))
2808                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
2809                 else
2810                         disk = conf->mirrors + disk_idx;
2811
2812                 if (disk->rdev)
2813                         goto abort;
2814                 disk->rdev = rdev;
2815                 q = bdev_get_queue(rdev->bdev);
2816                 if (q->merge_bvec_fn)
2817                         mddev->merge_check_needed = 1;
2818
2819                 disk->head_position = 0;
2820                 disk->seq_start = MaxSector;
2821         }
2822         conf->raid_disks = mddev->raid_disks;
2823         conf->mddev = mddev;
2824         INIT_LIST_HEAD(&conf->retry_list);
2825
2826         spin_lock_init(&conf->resync_lock);
2827         init_waitqueue_head(&conf->wait_barrier);
2828
2829         bio_list_init(&conf->pending_bio_list);
2830         conf->pending_count = 0;
2831         conf->recovery_disabled = mddev->recovery_disabled - 1;
2832
2833         conf->start_next_window = MaxSector;
2834         conf->current_window_requests = conf->next_window_requests = 0;
2835
2836         err = -EIO;
2837         for (i = 0; i < conf->raid_disks * 2; i++) {
2838
2839                 disk = conf->mirrors + i;
2840
2841                 if (i < conf->raid_disks &&
2842                     disk[conf->raid_disks].rdev) {
2843                         /* This slot has a replacement. */
2844                         if (!disk->rdev) {
2845                                 /* No original, just make the replacement
2846                                  * a recovering spare
2847                                  */
2848                                 disk->rdev =
2849                                         disk[conf->raid_disks].rdev;
2850                                 disk[conf->raid_disks].rdev = NULL;
2851                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2852                                 /* Original is not in_sync - bad */
2853                                 goto abort;
2854                 }
2855
2856                 if (!disk->rdev ||
2857                     !test_bit(In_sync, &disk->rdev->flags)) {
2858                         disk->head_position = 0;
2859                         if (disk->rdev &&
2860                             (disk->rdev->saved_raid_disk < 0))
2861                                 conf->fullsync = 1;
2862                 }
2863         }
2864
2865         err = -ENOMEM;
2866         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2867         if (!conf->thread) {
2868                 printk(KERN_ERR
2869                        "md/raid1:%s: couldn't allocate thread\n",
2870                        mdname(mddev));
2871                 goto abort;
2872         }
2873
2874         return conf;
2875
2876  abort:
2877         if (conf) {
2878                 if (conf->r1bio_pool)
2879                         mempool_destroy(conf->r1bio_pool);
2880                 kfree(conf->mirrors);
2881                 safe_put_page(conf->tmppage);
2882                 kfree(conf->poolinfo);
2883                 kfree(conf);
2884         }
2885         return ERR_PTR(err);
2886 }
2887
2888 static int stop(struct mddev *mddev);
2889 static int run(struct mddev *mddev)
2890 {
2891         struct r1conf *conf;
2892         int i;
2893         struct md_rdev *rdev;
2894         int ret;
2895         bool discard_supported = false;
2896
2897         if (mddev->level != 1) {
2898                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2899                        mdname(mddev), mddev->level);
2900                 return -EIO;
2901         }
2902         if (mddev->reshape_position != MaxSector) {
2903                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2904                        mdname(mddev));
2905                 return -EIO;
2906         }
2907         /*
2908          * copy the already verified devices into our private RAID1
2909          * bookkeeping area. [whatever we allocate in run(),
2910          * should be freed in stop()]
2911          */
2912         if (mddev->private == NULL)
2913                 conf = setup_conf(mddev);
2914         else
2915                 conf = mddev->private;
2916
2917         if (IS_ERR(conf))
2918                 return PTR_ERR(conf);
2919
2920         if (mddev->queue)
2921                 blk_queue_max_write_same_sectors(mddev->queue, 0);
2922
2923         rdev_for_each(rdev, mddev) {
2924                 if (!mddev->gendisk)
2925                         continue;
2926                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2927                                   rdev->data_offset << 9);
2928                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2929                         discard_supported = true;
2930         }
2931
2932         mddev->degraded = 0;
2933         for (i=0; i < conf->raid_disks; i++)
2934                 if (conf->mirrors[i].rdev == NULL ||
2935                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2936                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2937                         mddev->degraded++;
2938
2939         if (conf->raid_disks - mddev->degraded == 1)
2940                 mddev->recovery_cp = MaxSector;
2941
2942         if (mddev->recovery_cp != MaxSector)
2943                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2944                        " -- starting background reconstruction\n",
2945                        mdname(mddev));
2946         printk(KERN_INFO
2947                 "md/raid1:%s: active with %d out of %d mirrors\n",
2948                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2949                 mddev->raid_disks);
2950
2951         /*
2952          * Ok, everything is just fine now
2953          */
2954         mddev->thread = conf->thread;
2955         conf->thread = NULL;
2956         mddev->private = conf;
2957
2958         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2959
2960         if (mddev->queue) {
2961                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2962                 mddev->queue->backing_dev_info.congested_data = mddev;
2963                 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2964
2965                 if (discard_supported)
2966                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2967                                                 mddev->queue);
2968                 else
2969                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2970                                                   mddev->queue);
2971         }
2972
2973         ret =  md_integrity_register(mddev);
2974         if (ret)
2975                 stop(mddev);
2976         return ret;
2977 }
2978
2979 static int stop(struct mddev *mddev)
2980 {
2981         struct r1conf *conf = mddev->private;
2982         struct bitmap *bitmap = mddev->bitmap;
2983
2984         /* wait for behind writes to complete */
2985         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2986                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2987                        mdname(mddev));
2988                 /* need to kick something here to make sure I/O goes? */
2989                 wait_event(bitmap->behind_wait,
2990                            atomic_read(&bitmap->behind_writes) == 0);
2991         }
2992
2993         freeze_array(conf, 0);
2994         unfreeze_array(conf);
2995
2996         md_unregister_thread(&mddev->thread);
2997         if (conf->r1bio_pool)
2998                 mempool_destroy(conf->r1bio_pool);
2999         kfree(conf->mirrors);
3000         safe_put_page(conf->tmppage);
3001         kfree(conf->poolinfo);
3002         kfree(conf);
3003         mddev->private = NULL;
3004         return 0;
3005 }
3006
3007 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3008 {
3009         /* no resync is happening, and there is enough space
3010          * on all devices, so we can resize.
3011          * We need to make sure resync covers any new space.
3012          * If the array is shrinking we should possibly wait until
3013          * any io in the removed space completes, but it hardly seems
3014          * worth it.
3015          */
3016         sector_t newsize = raid1_size(mddev, sectors, 0);
3017         if (mddev->external_size &&
3018             mddev->array_sectors > newsize)
3019                 return -EINVAL;
3020         if (mddev->bitmap) {
3021                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3022                 if (ret)
3023                         return ret;
3024         }
3025         md_set_array_sectors(mddev, newsize);
3026         set_capacity(mddev->gendisk, mddev->array_sectors);
3027         revalidate_disk(mddev->gendisk);
3028         if (sectors > mddev->dev_sectors &&
3029             mddev->recovery_cp > mddev->dev_sectors) {
3030                 mddev->recovery_cp = mddev->dev_sectors;
3031                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3032         }
3033         mddev->dev_sectors = sectors;
3034         mddev->resync_max_sectors = sectors;
3035         return 0;
3036 }
3037
3038 static int raid1_reshape(struct mddev *mddev)
3039 {
3040         /* We need to:
3041          * 1/ resize the r1bio_pool
3042          * 2/ resize conf->mirrors
3043          *
3044          * We allocate a new r1bio_pool if we can.
3045          * Then raise a device barrier and wait until all IO stops.
3046          * Then resize conf->mirrors and swap in the new r1bio pool.
3047          *
3048          * At the same time, we "pack" the devices so that all the missing
3049          * devices have the higher raid_disk numbers.
3050          */
3051         mempool_t *newpool, *oldpool;
3052         struct pool_info *newpoolinfo;
3053         struct raid1_info *newmirrors;
3054         struct r1conf *conf = mddev->private;
3055         int cnt, raid_disks;
3056         unsigned long flags;
3057         int d, d2, err;
3058
3059         /* Cannot change chunk_size, layout, or level */
3060         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3061             mddev->layout != mddev->new_layout ||
3062             mddev->level != mddev->new_level) {
3063                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3064                 mddev->new_layout = mddev->layout;
3065                 mddev->new_level = mddev->level;
3066                 return -EINVAL;
3067         }
3068
3069         err = md_allow_write(mddev);
3070         if (err)
3071                 return err;
3072
3073         raid_disks = mddev->raid_disks + mddev->delta_disks;
3074
3075         if (raid_disks < conf->raid_disks) {
3076                 cnt=0;
3077                 for (d= 0; d < conf->raid_disks; d++)
3078                         if (conf->mirrors[d].rdev)
3079                                 cnt++;
3080                 if (cnt > raid_disks)
3081                         return -EBUSY;
3082         }
3083
3084         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3085         if (!newpoolinfo)
3086                 return -ENOMEM;
3087         newpoolinfo->mddev = mddev;
3088         newpoolinfo->raid_disks = raid_disks * 2;
3089
3090         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3091                                  r1bio_pool_free, newpoolinfo);
3092         if (!newpool) {
3093                 kfree(newpoolinfo);
3094                 return -ENOMEM;
3095         }
3096         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3097                              GFP_KERNEL);
3098         if (!newmirrors) {
3099                 kfree(newpoolinfo);
3100                 mempool_destroy(newpool);
3101                 return -ENOMEM;
3102         }
3103
3104         freeze_array(conf, 0);
3105
3106         /* ok, everything is stopped */
3107         oldpool = conf->r1bio_pool;
3108         conf->r1bio_pool = newpool;
3109
3110         for (d = d2 = 0; d < conf->raid_disks; d++) {
3111                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3112                 if (rdev && rdev->raid_disk != d2) {
3113                         sysfs_unlink_rdev(mddev, rdev);
3114                         rdev->raid_disk = d2;
3115                         sysfs_unlink_rdev(mddev, rdev);
3116                         if (sysfs_link_rdev(mddev, rdev))
3117                                 printk(KERN_WARNING
3118                                        "md/raid1:%s: cannot register rd%d\n",
3119                                        mdname(mddev), rdev->raid_disk);
3120                 }
3121                 if (rdev)
3122                         newmirrors[d2++].rdev = rdev;
3123         }
3124         kfree(conf->mirrors);
3125         conf->mirrors = newmirrors;
3126         kfree(conf->poolinfo);
3127         conf->poolinfo = newpoolinfo;
3128
3129         spin_lock_irqsave(&conf->device_lock, flags);
3130         mddev->degraded += (raid_disks - conf->raid_disks);
3131         spin_unlock_irqrestore(&conf->device_lock, flags);
3132         conf->raid_disks = mddev->raid_disks = raid_disks;
3133         mddev->delta_disks = 0;
3134
3135         unfreeze_array(conf);
3136
3137         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3138         md_wakeup_thread(mddev->thread);
3139
3140         mempool_destroy(oldpool);
3141         return 0;
3142 }
3143
3144 static void raid1_quiesce(struct mddev *mddev, int state)
3145 {
3146         struct r1conf *conf = mddev->private;
3147
3148         switch(state) {
3149         case 2: /* wake for suspend */
3150                 wake_up(&conf->wait_barrier);
3151                 break;
3152         case 1:
3153                 freeze_array(conf, 0);
3154                 break;
3155         case 0:
3156                 unfreeze_array(conf);
3157                 break;
3158         }
3159 }
3160
3161 static void *raid1_takeover(struct mddev *mddev)
3162 {
3163         /* raid1 can take over:
3164          *  raid5 with 2 devices, any layout or chunk size
3165          */
3166         if (mddev->level == 5 && mddev->raid_disks == 2) {
3167                 struct r1conf *conf;
3168                 mddev->new_level = 1;
3169                 mddev->new_layout = 0;
3170                 mddev->new_chunk_sectors = 0;
3171                 conf = setup_conf(mddev);
3172                 if (!IS_ERR(conf))
3173                         /* Array must appear to be quiesced */
3174                         conf->array_frozen = 1;
3175                 return conf;
3176         }
3177         return ERR_PTR(-EINVAL);
3178 }
3179
3180 static struct md_personality raid1_personality =
3181 {
3182         .name           = "raid1",
3183         .level          = 1,
3184         .owner          = THIS_MODULE,
3185         .make_request   = make_request,
3186         .run            = run,
3187         .stop           = stop,
3188         .status         = status,
3189         .error_handler  = error,
3190         .hot_add_disk   = raid1_add_disk,
3191         .hot_remove_disk= raid1_remove_disk,
3192         .spare_active   = raid1_spare_active,
3193         .sync_request   = sync_request,
3194         .resize         = raid1_resize,
3195         .size           = raid1_size,
3196         .check_reshape  = raid1_reshape,
3197         .quiesce        = raid1_quiesce,
3198         .takeover       = raid1_takeover,
3199 };
3200
3201 static int __init raid_init(void)
3202 {
3203         return register_md_personality(&raid1_personality);
3204 }
3205
3206 static void raid_exit(void)
3207 {
3208         unregister_md_personality(&raid1_personality);
3209 }
3210
3211 module_init(raid_init);
3212 module_exit(raid_exit);
3213 MODULE_LICENSE("GPL");
3214 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3215 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3216 MODULE_ALIAS("md-raid1");
3217 MODULE_ALIAS("md-level-1");
3218
3219 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);