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         choose_first = (conf->mddev->recovery_cp < this_sector + sectors);
 544
 545         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 546                 sector_t dist;
 547                 sector_t first_bad;
 548                 int bad_sectors;
 549                 unsigned int pending;
 550                 bool nonrot;
 551
 552                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 553                 if (r1_bio->bios[disk] == IO_BLOCKED
 554                     || rdev == NULL
 555                     || test_bit(Unmerged, &rdev->flags)
 556                     || test_bit(Faulty, &rdev->flags))
 557                         continue;
 558                 if (!test_bit(In_sync, &rdev->flags) &&
 559                     rdev->recovery_offset < this_sector + sectors)
 560                         continue;
 561                 if (test_bit(WriteMostly, &rdev->flags)) {
 562                         /* Don't balance among write-mostly, just
 563                          * use the first as a last resort */
 564                         if (best_disk < 0) {
 565                                 if (is_badblock(rdev, this_sector, sectors,
 566                                                 &first_bad, &bad_sectors)) {
 567                                         if (first_bad < this_sector)
 568                                                 /* Cannot use this */
 569                                                 continue;
 570                                         best_good_sectors = first_bad - this_sector;
 571                                 } else
 572                                         best_good_sectors = sectors;
 573                                 best_disk = disk;
 574                         }
 575                         continue;
 576                 }
 577                 /* This is a reasonable device to use.  It might
 578                  * even be best.
 579                  */
 580                 if (is_badblock(rdev, this_sector, sectors,
 581                                 &first_bad, &bad_sectors)) {
 582                         if (best_dist < MaxSector)
 583                                 /* already have a better device */
 584                                 continue;
 585                         if (first_bad <= this_sector) {
 586                                 /* cannot read here. If this is the 'primary'
 587                                  * device, then we must not read beyond
 588                                  * bad_sectors from another device..
 589                                  */
 590                                 bad_sectors -= (this_sector - first_bad);
 591                                 if (choose_first && sectors > bad_sectors)
 592                                         sectors = bad_sectors;
 593                                 if (best_good_sectors > sectors)
 594                                         best_good_sectors = sectors;
 595
 596                         } else {
 597                                 sector_t good_sectors = first_bad - this_sector;
 598                                 if (good_sectors > best_good_sectors) {
 599                                         best_good_sectors = good_sectors;
 600                                         best_disk = disk;
 601                                 }
 602                                 if (choose_first)
 603                                         break;
 604                         }
 605                         continue;
 606                 } else
 607                         best_good_sectors = sectors;
 608
 609                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 610                 has_nonrot_disk |= nonrot;
 611                 pending = atomic_read(&rdev->nr_pending);
 612                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 613                 if (choose_first) {
 614                         best_disk = disk;
 615                         break;
 616                 }
 617                 /* Don't change to another disk for sequential reads */
 618                 if (conf->mirrors[disk].next_seq_sect == this_sector
 619                     || dist == 0) {
 620                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 621                         struct raid1_info *mirror = &conf->mirrors[disk];
 622
 623                         best_disk = disk;
 624                         /*
 625                          * If buffered sequential IO size exceeds optimal
 626                          * iosize, check if there is idle disk. If yes, choose
 627                          * the idle disk. read_balance could already choose an
 628                          * idle disk before noticing it's a sequential IO in
 629                          * this disk. This doesn't matter because this disk
 630                          * will idle, next time it will be utilized after the
 631                          * first disk has IO size exceeds optimal iosize. In
 632                          * this way, iosize of the first disk will be optimal
 633                          * iosize at least. iosize of the second disk might be
 634                          * small, but not a big deal since when the second disk
 635                          * starts IO, the first disk is likely still busy.
 636                          */
 637                         if (nonrot && opt_iosize > 0 &&
 638                             mirror->seq_start != MaxSector &&
 639                             mirror->next_seq_sect > opt_iosize &&
 640                             mirror->next_seq_sect - opt_iosize >=
 641                             mirror->seq_start) {
 642                                 choose_next_idle = 1;
 643                                 continue;
 644                         }
 645                         break;
 646                 }
 647                 /* If device is idle, use it */
 648                 if (pending == 0) {
 649                         best_disk = disk;
 650                         break;
 651                 }
 652
 653                 if (choose_next_idle)
 654                         continue;
 655
 656                 if (min_pending > pending) {
 657                         min_pending = pending;
 658                         best_pending_disk = disk;
 659                 }
 660
 661                 if (dist < best_dist) {
 662                         best_dist = dist;
 663                         best_dist_disk = disk;
 664                 }
 665         }
 666
 667         /*
 668          * If all disks are rotational, choose the closest disk. If any disk is
 669          * non-rotational, choose the disk with less pending request even the
 670          * disk is rotational, which might/might not be optimal for raids with
 671          * mixed ratation/non-rotational disks depending on workload.
 672          */
 673         if (best_disk == -1) {
 674                 if (has_nonrot_disk)
 675                         best_disk = best_pending_disk;
 676                 else
 677                         best_disk = best_dist_disk;
 678         }
 679
 680         if (best_disk >= 0) {
 681                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 682                 if (!rdev)
 683                         goto retry;
 684                 atomic_inc(&rdev->nr_pending);
 685                 if (test_bit(Faulty, &rdev->flags)) {
 686                         /* cannot risk returning a device that failed
 687                          * before we inc'ed nr_pending
 688                          */
 689                         rdev_dec_pending(rdev, conf->mddev);
 690                         goto retry;
 691                 }
 692                 sectors = best_good_sectors;
 693
 694                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 695                         conf->mirrors[best_disk].seq_start = this_sector;
 696
 697                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 698         }
 699         rcu_read_unlock();
 700         *max_sectors = sectors;
 701
 702         return best_disk;
 703 }
 704
 705 static int raid1_mergeable_bvec(struct request_queue *q,
 706                                 struct bvec_merge_data *bvm,
 707                                 struct bio_vec *biovec)
 708 {
 709         struct mddev *mddev = q->queuedata;
 710         struct r1conf *conf = mddev->private;
 711         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 712         int max = biovec->bv_len;
 713
 714         if (mddev->merge_check_needed) {
 715                 int disk;
 716                 rcu_read_lock();
 717                 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
 718                         struct md_rdev *rdev = rcu_dereference(
 719                                 conf->mirrors[disk].rdev);
 720                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 721                                 struct request_queue *q =
 722                                         bdev_get_queue(rdev->bdev);
 723                                 if (q->merge_bvec_fn) {
 724                                         bvm->bi_sector = sector +
 725                                                 rdev->data_offset;
 726                                         bvm->bi_bdev = rdev->bdev;
 727                                         max = min(max, q->merge_bvec_fn(
 728                                                           q, bvm, biovec));
 729                                 }
 730                         }
 731                 }
 732                 rcu_read_unlock();
 733         }
 734         return max;
 735
 736 }
 737
 738 int md_raid1_congested(struct mddev *mddev, int bits)
 739 {
 740         struct r1conf *conf = mddev->private;
 741         int i, ret = 0;
 742
 743         if ((bits & (1 << BDI_async_congested)) &&
 744             conf->pending_count >= max_queued_requests)
 745                 return 1;
 746
 747         rcu_read_lock();
 748         for (i = 0; i < conf->raid_disks * 2; i++) {
 749                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 750                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 751                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 752
 753                         BUG_ON(!q);
 754
 755                         /* Note the '|| 1' - when read_balance prefers
 756                          * non-congested targets, it can be removed
 757                          */
 758                         if ((bits & (1<<BDI_async_congested)) || 1)
 759                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 760                         else
 761                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 762                 }
 763         }
 764         rcu_read_unlock();
 765         return ret;
 766 }
 767 EXPORT_SYMBOL_GPL(md_raid1_congested);
 768
 769 static int raid1_congested(void *data, int bits)
 770 {
 771         struct mddev *mddev = data;
 772
 773         return mddev_congested(mddev, bits) ||
 774                 md_raid1_congested(mddev, bits);
 775 }
 776
 777 static void flush_pending_writes(struct r1conf *conf)
 778 {
 779         /* Any writes that have been queued but are awaiting
 780          * bitmap updates get flushed here.
 781          */
 782         spin_lock_irq(&conf->device_lock);
 783
 784         if (conf->pending_bio_list.head) {
 785                 struct bio *bio;
 786                 bio = bio_list_get(&conf->pending_bio_list);
 787                 conf->pending_count = 0;
 788                 spin_unlock_irq(&conf->device_lock);
 789                 /* flush any pending bitmap writes to
 790                  * disk before proceeding w/ I/O */
 791                 bitmap_unplug(conf->mddev->bitmap);
 792                 wake_up(&conf->wait_barrier);
 793
 794                 while (bio) { /* submit pending writes */
 795                         struct bio *next = bio->bi_next;
 796                         bio->bi_next = NULL;
 797                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 798                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 799                                 /* Just ignore it */
 800                                 bio_endio(bio, 0);
 801                         else
 802                                 generic_make_request(bio);
 803                         bio = next;
 804                 }
 805         } else
 806                 spin_unlock_irq(&conf->device_lock);
 807 }
 808
 809 /* Barriers....
 810  * Sometimes we need to suspend IO while we do something else,
 811  * either some resync/recovery, or reconfigure the array.
 812  * To do this we raise a 'barrier'.
 813  * The 'barrier' is a counter that can be raised multiple times
 814  * to count how many activities are happening which preclude
 815  * normal IO.
 816  * We can only raise the barrier if there is no pending IO.
 817  * i.e. if nr_pending == 0.
 818  * We choose only to raise the barrier if no-one is waiting for the
 819  * barrier to go down.  This means that as soon as an IO request
 820  * is ready, no other operations which require a barrier will start
 821  * until the IO request has had a chance.
 822  *
 823  * So: regular IO calls 'wait_barrier'.  When that returns there
 824  *    is no backgroup IO happening,  It must arrange to call
 825  *    allow_barrier when it has finished its IO.
 826  * backgroup IO calls must call raise_barrier.  Once that returns
 827  *    there is no normal IO happeing.  It must arrange to call
 828  *    lower_barrier when the particular background IO completes.
 829  */
 830 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
 831 {
 832         spin_lock_irq(&conf->resync_lock);
 833
 834         /* Wait until no block IO is waiting */
 835         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 836                             conf->resync_lock);
 837
 838         /* block any new IO from starting */
 839         conf->barrier++;
 840         conf->next_resync = sector_nr;
 841
 842         /* For these conditions we must wait:
 843          * A: while the array is in frozen state
 844          * B: while barrier >= RESYNC_DEPTH, meaning resync reach
 845          *    the max count which allowed.
 846          * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
 847          *    next resync will reach to the window which normal bios are
 848          *    handling.
 849          * D: while there are any active requests in the current window.
 850          */
 851         wait_event_lock_irq(conf->wait_barrier,
 852                             !conf->array_frozen &&
 853                             conf->barrier < RESYNC_DEPTH &&
 854                             conf->current_window_requests == 0 &&
 855                             (conf->start_next_window >=
 856                              conf->next_resync + RESYNC_SECTORS),
 857                             conf->resync_lock);
 858
 859         conf->nr_pending++;
 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         conf->nr_pending--;
 870         spin_unlock_irqrestore(&conf->resync_lock, flags);
 871         wake_up(&conf->wait_barrier);
 872 }
 873
 874 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio)
 875 {
 876         bool wait = false;
 877
 878         if (conf->array_frozen || !bio)
 879                 wait = true;
 880         else if (conf->barrier && bio_data_dir(bio) == WRITE) {
 881                 if ((conf->mddev->curr_resync_completed
 882                      >= bio_end_sector(bio)) ||
 883                     (conf->next_resync + NEXT_NORMALIO_DISTANCE
 884                      <= bio->bi_iter.bi_sector))
 885                         wait = false;
 886                 else
 887                         wait = true;
 888         }
 889
 890         return wait;
 891 }
 892
 893 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
 894 {
 895         sector_t sector = 0;
 896
 897         spin_lock_irq(&conf->resync_lock);
 898         if (need_to_wait_for_sync(conf, bio)) {
 899                 conf->nr_waiting++;
 900                 /* Wait for the barrier to drop.
 901                  * However if there are already pending
 902                  * requests (preventing the barrier from
 903                  * rising completely), and the
 904                  * pre-process bio queue isn't empty,
 905                  * then don't wait, as we need to empty
 906                  * that queue to get the nr_pending
 907                  * count down.
 908                  */
 909                 wait_event_lock_irq(conf->wait_barrier,
 910                                     !conf->array_frozen &&
 911                                     (!conf->barrier ||
 912                                     ((conf->start_next_window <
 913                                       conf->next_resync + RESYNC_SECTORS) &&
 914                                      current->bio_list &&
 915                                      !bio_list_empty(current->bio_list))),
 916                                     conf->resync_lock);
 917                 conf->nr_waiting--;
 918         }
 919
 920         if (bio && bio_data_dir(bio) == WRITE) {
 921                 if (bio->bi_iter.bi_sector >=
 922                     conf->mddev->curr_resync_completed) {
 923                         if (conf->start_next_window == MaxSector)
 924                                 conf->start_next_window =
 925                                         conf->next_resync +
 926                                         NEXT_NORMALIO_DISTANCE;
 927
 928                         if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
 929                             <= bio->bi_iter.bi_sector)
 930                                 conf->next_window_requests++;
 931                         else
 932                                 conf->current_window_requests++;
 933                         sector = conf->start_next_window;
 934                 }
 935         }
 936
 937         conf->nr_pending++;
 938         spin_unlock_irq(&conf->resync_lock);
 939         return sector;
 940 }
 941
 942 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
 943                           sector_t bi_sector)
 944 {
 945         unsigned long flags;
 946
 947         spin_lock_irqsave(&conf->resync_lock, flags);
 948         conf->nr_pending--;
 949         if (start_next_window) {
 950                 if (start_next_window == conf->start_next_window) {
 951                         if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
 952                             <= bi_sector)
 953                                 conf->next_window_requests--;
 954                         else
 955                                 conf->current_window_requests--;
 956                 } else
 957                         conf->current_window_requests--;
 958
 959                 if (!conf->current_window_requests) {
 960                         if (conf->next_window_requests) {
 961                                 conf->current_window_requests =
 962                                         conf->next_window_requests;
 963                                 conf->next_window_requests = 0;
 964                                 conf->start_next_window +=
 965                                         NEXT_NORMALIO_DISTANCE;
 966                         } else
 967                                 conf->start_next_window = MaxSector;
 968                 }
 969         }
 970         spin_unlock_irqrestore(&conf->resync_lock, flags);
 971         wake_up(&conf->wait_barrier);
 972 }
 973
 974 static void freeze_array(struct r1conf *conf, int extra)
 975 {
 976         /* stop syncio and normal IO and wait for everything to
 977          * go quite.
 978          * We wait until nr_pending match nr_queued+extra
 979          * This is called in the context of one normal IO request
 980          * that has failed. Thus any sync request that might be pending
 981          * will be blocked by nr_pending, and we need to wait for
 982          * pending IO requests to complete or be queued for re-try.
 983          * Thus the number queued (nr_queued) plus this request (extra)
 984          * must match the number of pending IOs (nr_pending) before
 985          * we continue.
 986          */
 987         spin_lock_irq(&conf->resync_lock);
 988         conf->array_frozen = 1;
 989         wait_event_lock_irq_cmd(conf->wait_barrier,
 990                                 conf->nr_pending == conf->nr_queued+extra,
 991                                 conf->resync_lock,
 992                                 flush_pending_writes(conf));
 993         spin_unlock_irq(&conf->resync_lock);
 994 }
 995 static void unfreeze_array(struct r1conf *conf)
 996 {
 997         /* reverse the effect of the freeze */
 998         spin_lock_irq(&conf->resync_lock);
 999         conf->array_frozen = 0;
1000         wake_up(&conf->wait_barrier);
1001         spin_unlock_irq(&conf->resync_lock);
1002 }
1003
1004
1005 /* duplicate the data pages for behind I/O
1006  */
1007 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
1008 {
1009         int i;
1010         struct bio_vec *bvec;
1011         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
1012                                         GFP_NOIO);
1013         if (unlikely(!bvecs))
1014                 return;
1015
1016         bio_for_each_segment_all(bvec, bio, i) {
1017                 bvecs[i] = *bvec;
1018                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
1019                 if (unlikely(!bvecs[i].bv_page))
1020                         goto do_sync_io;
1021                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
1022                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
1023                 kunmap(bvecs[i].bv_page);
1024                 kunmap(bvec->bv_page);
1025         }
1026         r1_bio->behind_bvecs = bvecs;
1027         r1_bio->behind_page_count = bio->bi_vcnt;
1028         set_bit(R1BIO_BehindIO, &r1_bio->state);
1029         return;
1030
1031 do_sync_io:
1032         for (i = 0; i < bio->bi_vcnt; i++)
1033                 if (bvecs[i].bv_page)
1034                         put_page(bvecs[i].bv_page);
1035         kfree(bvecs);
1036         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1037                  bio->bi_iter.bi_size);
1038 }
1039
1040 struct raid1_plug_cb {
1041         struct blk_plug_cb      cb;
1042         struct bio_list         pending;
1043         int                     pending_cnt;
1044 };
1045
1046 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1047 {
1048         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1049                                                   cb);
1050         struct mddev *mddev = plug->cb.data;
1051         struct r1conf *conf = mddev->private;
1052         struct bio *bio;
1053
1054         if (from_schedule || current->bio_list) {
1055                 spin_lock_irq(&conf->device_lock);
1056                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1057                 conf->pending_count += plug->pending_cnt;
1058                 spin_unlock_irq(&conf->device_lock);
1059                 wake_up(&conf->wait_barrier);
1060                 md_wakeup_thread(mddev->thread);
1061                 kfree(plug);
1062                 return;
1063         }
1064
1065         /* we aren't scheduling, so we can do the write-out directly. */
1066         bio = bio_list_get(&plug->pending);
1067         bitmap_unplug(mddev->bitmap);
1068         wake_up(&conf->wait_barrier);
1069
1070         while (bio) { /* submit pending writes */
1071                 struct bio *next = bio->bi_next;
1072                 bio->bi_next = NULL;
1073                 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1074                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1075                         /* Just ignore it */
1076                         bio_endio(bio, 0);
1077                 else
1078                         generic_make_request(bio);
1079                 bio = next;
1080         }
1081         kfree(plug);
1082 }
1083
1084 static void make_request(struct mddev *mddev, struct bio * bio)
1085 {
1086         struct r1conf *conf = mddev->private;
1087         struct raid1_info *mirror;
1088         struct r1bio *r1_bio;
1089         struct bio *read_bio;
1090         int i, disks;
1091         struct bitmap *bitmap;
1092         unsigned long flags;
1093         const int rw = bio_data_dir(bio);
1094         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1095         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1096         const unsigned long do_discard = (bio->bi_rw
1097                                           & (REQ_DISCARD | REQ_SECURE));
1098         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1099         struct md_rdev *blocked_rdev;
1100         struct blk_plug_cb *cb;
1101         struct raid1_plug_cb *plug = NULL;
1102         int first_clone;
1103         int sectors_handled;
1104         int max_sectors;
1105         sector_t start_next_window;
1106
1107         /*
1108          * Register the new request and wait if the reconstruction
1109          * thread has put up a bar for new requests.
1110          * Continue immediately if no resync is active currently.
1111          */
1112
1113         md_write_start(mddev, bio); /* wait on superblock update early */
1114
1115         if (bio_data_dir(bio) == WRITE &&
1116             bio_end_sector(bio) > mddev->suspend_lo &&
1117             bio->bi_iter.bi_sector < mddev->suspend_hi) {
1118                 /* As the suspend_* range is controlled by
1119                  * userspace, we want an interruptible
1120                  * wait.
1121                  */
1122                 DEFINE_WAIT(w);
1123                 for (;;) {
1124                         flush_signals(current);
1125                         prepare_to_wait(&conf->wait_barrier,
1126                                         &w, TASK_INTERRUPTIBLE);
1127                         if (bio_end_sector(bio) <= mddev->suspend_lo ||
1128                             bio->bi_iter.bi_sector >= mddev->suspend_hi)
1129                                 break;
1130                         schedule();
1131                 }
1132                 finish_wait(&conf->wait_barrier, &w);
1133         }
1134
1135         start_next_window = wait_barrier(conf, bio);
1136
1137         bitmap = mddev->bitmap;
1138
1139         /*
1140          * make_request() can abort the operation when READA is being
1141          * used and no empty request is available.
1142          *
1143          */
1144         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1145
1146         r1_bio->master_bio = bio;
1147         r1_bio->sectors = bio_sectors(bio);
1148         r1_bio->state = 0;
1149         r1_bio->mddev = mddev;
1150         r1_bio->sector = bio->bi_iter.bi_sector;
1151
1152         /* We might need to issue multiple reads to different
1153          * devices if there are bad blocks around, so we keep
1154          * track of the number of reads in bio->bi_phys_segments.
1155          * If this is 0, there is only one r1_bio and no locking
1156          * will be needed when requests complete.  If it is
1157          * non-zero, then it is the number of not-completed requests.
1158          */
1159         bio->bi_phys_segments = 0;
1160         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1161
1162         if (rw == READ) {
1163                 /*
1164                  * read balancing logic:
1165                  */
1166                 int rdisk;
1167
1168 read_again:
1169                 rdisk = read_balance(conf, r1_bio, &max_sectors);
1170
1171                 if (rdisk < 0) {
1172                         /* couldn't find anywhere to read from */
1173                         raid_end_bio_io(r1_bio);
1174                         return;
1175                 }
1176                 mirror = conf->mirrors + rdisk;
1177
1178                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1179                     bitmap) {
1180                         /* Reading from a write-mostly device must
1181                          * take care not to over-take any writes
1182                          * that are 'behind'
1183                          */
1184                         wait_event(bitmap->behind_wait,
1185                                    atomic_read(&bitmap->behind_writes) == 0);
1186                 }
1187                 r1_bio->read_disk = rdisk;
1188                 r1_bio->start_next_window = 0;
1189
1190                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1191                 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1192                          max_sectors);
1193
1194                 r1_bio->bios[rdisk] = read_bio;
1195
1196                 read_bio->bi_iter.bi_sector = r1_bio->sector +
1197                         mirror->rdev->data_offset;
1198                 read_bio->bi_bdev = mirror->rdev->bdev;
1199                 read_bio->bi_end_io = raid1_end_read_request;
1200                 read_bio->bi_rw = READ | do_sync;
1201                 read_bio->bi_private = r1_bio;
1202
1203                 if (max_sectors < r1_bio->sectors) {
1204                         /* could not read all from this device, so we will
1205                          * need another r1_bio.
1206                          */
1207
1208                         sectors_handled = (r1_bio->sector + max_sectors
1209                                            - bio->bi_iter.bi_sector);
1210                         r1_bio->sectors = max_sectors;
1211                         spin_lock_irq(&conf->device_lock);
1212                         if (bio->bi_phys_segments == 0)
1213                                 bio->bi_phys_segments = 2;
1214                         else
1215                                 bio->bi_phys_segments++;
1216                         spin_unlock_irq(&conf->device_lock);
1217                         /* Cannot call generic_make_request directly
1218                          * as that will be queued in __make_request
1219                          * and subsequent mempool_alloc might block waiting
1220                          * for it.  So hand bio over to raid1d.
1221                          */
1222                         reschedule_retry(r1_bio);
1223
1224                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1225
1226                         r1_bio->master_bio = bio;
1227                         r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1228                         r1_bio->state = 0;
1229                         r1_bio->mddev = mddev;
1230                         r1_bio->sector = bio->bi_iter.bi_sector +
1231                                 sectors_handled;
1232                         goto read_again;
1233                 } else
1234                         generic_make_request(read_bio);
1235                 return;
1236         }
1237
1238         /*
1239          * WRITE:
1240          */
1241         if (conf->pending_count >= max_queued_requests) {
1242                 md_wakeup_thread(mddev->thread);
1243                 wait_event(conf->wait_barrier,
1244                            conf->pending_count < max_queued_requests);
1245         }
1246         /* first select target devices under rcu_lock and
1247          * inc refcount on their rdev.  Record them by setting
1248          * bios[x] to bio
1249          * If there are known/acknowledged bad blocks on any device on
1250          * which we have seen a write error, we want to avoid writing those
1251          * blocks.
1252          * This potentially requires several writes to write around
1253          * the bad blocks.  Each set of writes gets it's own r1bio
1254          * with a set of bios attached.
1255          */
1256
1257         disks = conf->raid_disks * 2;
1258  retry_write:
1259         r1_bio->start_next_window = start_next_window;
1260         blocked_rdev = NULL;
1261         rcu_read_lock();
1262         max_sectors = r1_bio->sectors;
1263         for (i = 0;  i < disks; i++) {
1264                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1265                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1266                         atomic_inc(&rdev->nr_pending);
1267                         blocked_rdev = rdev;
1268                         break;
1269                 }
1270                 r1_bio->bios[i] = NULL;
1271                 if (!rdev || test_bit(Faulty, &rdev->flags)
1272                     || test_bit(Unmerged, &rdev->flags)) {
1273                         if (i < conf->raid_disks)
1274                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1275                         continue;
1276                 }
1277
1278                 atomic_inc(&rdev->nr_pending);
1279                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1280                         sector_t first_bad;
1281                         int bad_sectors;
1282                         int is_bad;
1283
1284                         is_bad = is_badblock(rdev, r1_bio->sector,
1285                                              max_sectors,
1286                                              &first_bad, &bad_sectors);
1287                         if (is_bad < 0) {
1288                                 /* mustn't write here until the bad block is
1289                                  * acknowledged*/
1290                                 set_bit(BlockedBadBlocks, &rdev->flags);
1291                                 blocked_rdev = rdev;
1292                                 break;
1293                         }
1294                         if (is_bad && first_bad <= r1_bio->sector) {
1295                                 /* Cannot write here at all */
1296                                 bad_sectors -= (r1_bio->sector - first_bad);
1297                                 if (bad_sectors < max_sectors)
1298                                         /* mustn't write more than bad_sectors
1299                                          * to other devices yet
1300                                          */
1301                                         max_sectors = bad_sectors;
1302                                 rdev_dec_pending(rdev, mddev);
1303                                 /* We don't set R1BIO_Degraded as that
1304                                  * only applies if the disk is
1305                                  * missing, so it might be re-added,
1306                                  * and we want to know to recover this
1307                                  * chunk.
1308                                  * In this case the device is here,
1309                                  * and the fact that this chunk is not
1310                                  * in-sync is recorded in the bad
1311                                  * block log
1312                                  */
1313                                 continue;
1314                         }
1315                         if (is_bad) {
1316                                 int good_sectors = first_bad - r1_bio->sector;
1317                                 if (good_sectors < max_sectors)
1318                                         max_sectors = good_sectors;
1319                         }
1320                 }
1321                 r1_bio->bios[i] = bio;
1322         }
1323         rcu_read_unlock();
1324
1325         if (unlikely(blocked_rdev)) {
1326                 /* Wait for this device to become unblocked */
1327                 int j;
1328                 sector_t old = start_next_window;
1329
1330                 for (j = 0; j < i; j++)
1331                         if (r1_bio->bios[j])
1332                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1333                 r1_bio->state = 0;
1334                 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector);
1335                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1336                 start_next_window = wait_barrier(conf, bio);
1337                 /*
1338                  * We must make sure the multi r1bios of bio have
1339                  * the same value of bi_phys_segments
1340                  */
1341                 if (bio->bi_phys_segments && old &&
1342                     old != start_next_window)
1343                         /* Wait for the former r1bio(s) to complete */
1344                         wait_event(conf->wait_barrier,
1345                                    bio->bi_phys_segments == 1);
1346                 goto retry_write;
1347         }
1348
1349         if (max_sectors < r1_bio->sectors) {
1350                 /* We are splitting this write into multiple parts, so
1351                  * we need to prepare for allocating another r1_bio.
1352                  */
1353                 r1_bio->sectors = max_sectors;
1354                 spin_lock_irq(&conf->device_lock);
1355                 if (bio->bi_phys_segments == 0)
1356                         bio->bi_phys_segments = 2;
1357                 else
1358                         bio->bi_phys_segments++;
1359                 spin_unlock_irq(&conf->device_lock);
1360         }
1361         sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1362
1363         atomic_set(&r1_bio->remaining, 1);
1364         atomic_set(&r1_bio->behind_remaining, 0);
1365
1366         first_clone = 1;
1367         for (i = 0; i < disks; i++) {
1368                 struct bio *mbio;
1369                 if (!r1_bio->bios[i])
1370                         continue;
1371
1372                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1373                 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors);
1374
1375                 if (first_clone) {
1376                         /* do behind I/O ?
1377                          * Not if there are too many, or cannot
1378                          * allocate memory, or a reader on WriteMostly
1379                          * is waiting for behind writes to flush */
1380                         if (bitmap &&
1381                             (atomic_read(&bitmap->behind_writes)
1382                              < mddev->bitmap_info.max_write_behind) &&
1383                             !waitqueue_active(&bitmap->behind_wait))
1384                                 alloc_behind_pages(mbio, r1_bio);
1385
1386                         bitmap_startwrite(bitmap, r1_bio->sector,
1387                                           r1_bio->sectors,
1388                                           test_bit(R1BIO_BehindIO,
1389                                                    &r1_bio->state));
1390                         first_clone = 0;
1391                 }
1392                 if (r1_bio->behind_bvecs) {
1393                         struct bio_vec *bvec;
1394                         int j;
1395
1396                         /*
1397                          * We trimmed the bio, so _all is legit
1398                          */
1399                         bio_for_each_segment_all(bvec, mbio, j)
1400                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1401                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1402                                 atomic_inc(&r1_bio->behind_remaining);
1403                 }
1404
1405                 r1_bio->bios[i] = mbio;
1406
1407                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1408                                    conf->mirrors[i].rdev->data_offset);
1409                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1410                 mbio->bi_end_io = raid1_end_write_request;
1411                 mbio->bi_rw =
1412                         WRITE | do_flush_fua | do_sync | do_discard | do_same;
1413                 mbio->bi_private = r1_bio;
1414
1415                 atomic_inc(&r1_bio->remaining);
1416
1417                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1418                 if (cb)
1419                         plug = container_of(cb, struct raid1_plug_cb, cb);
1420                 else
1421                         plug = NULL;
1422                 spin_lock_irqsave(&conf->device_lock, flags);
1423                 if (plug) {
1424                         bio_list_add(&plug->pending, mbio);
1425                         plug->pending_cnt++;
1426                 } else {
1427                         bio_list_add(&conf->pending_bio_list, mbio);
1428                         conf->pending_count++;
1429                 }
1430                 spin_unlock_irqrestore(&conf->device_lock, flags);
1431                 if (!plug)
1432                         md_wakeup_thread(mddev->thread);
1433         }
1434         /* Mustn't call r1_bio_write_done before this next test,
1435          * as it could result in the bio being freed.
1436          */
1437         if (sectors_handled < bio_sectors(bio)) {
1438                 r1_bio_write_done(r1_bio);
1439                 /* We need another r1_bio.  It has already been counted
1440                  * in bio->bi_phys_segments
1441                  */
1442                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1443                 r1_bio->master_bio = bio;
1444                 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1445                 r1_bio->state = 0;
1446                 r1_bio->mddev = mddev;
1447                 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1448                 goto retry_write;
1449         }
1450
1451         r1_bio_write_done(r1_bio);
1452
1453         /* In case raid1d snuck in to freeze_array */
1454         wake_up(&conf->wait_barrier);
1455 }
1456
1457 static void status(struct seq_file *seq, struct mddev *mddev)
1458 {
1459         struct r1conf *conf = mddev->private;
1460         int i;
1461
1462         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1463                    conf->raid_disks - mddev->degraded);
1464         rcu_read_lock();
1465         for (i = 0; i < conf->raid_disks; i++) {
1466                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1467                 seq_printf(seq, "%s",
1468                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1469         }
1470         rcu_read_unlock();
1471         seq_printf(seq, "]");
1472 }
1473
1474
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         } else
1505                 set_bit(Faulty, &rdev->flags);
1506         /*
1507          * if recovery is running, make sure it aborts.
1508          */
1509         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
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         spin_lock_irq(&conf->resync_lock);
1552         conf->next_resync = 0;
1553         conf->start_next_window = MaxSector;
1554         conf->current_window_requests +=
1555                 conf->next_window_requests;
1556         conf->next_window_requests = 0;
1557         spin_unlock_irq(&conf->resync_lock);
1558 }
1559
1560 static int raid1_spare_active(struct mddev *mddev)
1561 {
1562         int i;
1563         struct r1conf *conf = mddev->private;
1564         int count = 0;
1565         unsigned long flags;
1566
1567         /*
1568          * Find all failed disks within the RAID1 configuration
1569          * and mark them readable.
1570          * Called under mddev lock, so rcu protection not needed.
1571          */
1572         for (i = 0; i < conf->raid_disks; i++) {
1573                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1574                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1575                 if (repl
1576                     && repl->recovery_offset == MaxSector
1577                     && !test_bit(Faulty, &repl->flags)
1578                     && !test_and_set_bit(In_sync, &repl->flags)) {
1579                         /* replacement has just become active */
1580                         if (!rdev ||
1581                             !test_and_clear_bit(In_sync, &rdev->flags))
1582                                 count++;
1583                         if (rdev) {
1584                                 /* Replaced device not technically
1585                                  * faulty, but we need to be sure
1586                                  * it gets removed and never re-added
1587                                  */
1588                                 set_bit(Faulty, &rdev->flags);
1589                                 sysfs_notify_dirent_safe(
1590                                         rdev->sysfs_state);
1591                         }
1592                 }
1593                 if (rdev
1594                     && rdev->recovery_offset == MaxSector
1595                     && !test_bit(Faulty, &rdev->flags)
1596                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1597                         count++;
1598                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1599                 }
1600         }
1601         spin_lock_irqsave(&conf->device_lock, flags);
1602         mddev->degraded -= count;
1603         spin_unlock_irqrestore(&conf->device_lock, flags);
1604
1605         print_conf(conf);
1606         return count;
1607 }
1608
1609
1610 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1611 {
1612         struct r1conf *conf = mddev->private;
1613         int err = -EEXIST;
1614         int mirror = 0;
1615         struct raid1_info *p;
1616         int first = 0;
1617         int last = conf->raid_disks - 1;
1618         struct request_queue *q = bdev_get_queue(rdev->bdev);
1619
1620         if (mddev->recovery_disabled == conf->recovery_disabled)
1621                 return -EBUSY;
1622
1623         if (rdev->raid_disk >= 0)
1624                 first = last = rdev->raid_disk;
1625
1626         if (q->merge_bvec_fn) {
1627                 set_bit(Unmerged, &rdev->flags);
1628                 mddev->merge_check_needed = 1;
1629         }
1630
1631         for (mirror = first; mirror <= last; mirror++) {
1632                 p = conf->mirrors+mirror;
1633                 if (!p->rdev) {
1634
1635                         if (mddev->gendisk)
1636                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1637                                                   rdev->data_offset << 9);
1638
1639                         p->head_position = 0;
1640                         rdev->raid_disk = mirror;
1641                         err = 0;
1642                         /* As all devices are equivalent, we don't need a full recovery
1643                          * if this was recently any drive of the array
1644                          */
1645                         if (rdev->saved_raid_disk < 0)
1646                                 conf->fullsync = 1;
1647                         rcu_assign_pointer(p->rdev, rdev);
1648                         break;
1649                 }
1650                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1651                     p[conf->raid_disks].rdev == NULL) {
1652                         /* Add this device as a replacement */
1653                         clear_bit(In_sync, &rdev->flags);
1654                         set_bit(Replacement, &rdev->flags);
1655                         rdev->raid_disk = mirror;
1656                         err = 0;
1657                         conf->fullsync = 1;
1658                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1659                         break;
1660                 }
1661         }
1662         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1663                 /* Some requests might not have seen this new
1664                  * merge_bvec_fn.  We must wait for them to complete
1665                  * before merging the device fully.
1666                  * First we make sure any code which has tested
1667                  * our function has submitted the request, then
1668                  * we wait for all outstanding requests to complete.
1669                  */
1670                 synchronize_sched();
1671                 freeze_array(conf, 0);
1672                 unfreeze_array(conf);
1673                 clear_bit(Unmerged, &rdev->flags);
1674         }
1675         md_integrity_add_rdev(rdev, mddev);
1676         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1677                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1678         print_conf(conf);
1679         return err;
1680 }
1681
1682 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1683 {
1684         struct r1conf *conf = mddev->private;
1685         int err = 0;
1686         int number = rdev->raid_disk;
1687         struct raid1_info *p = conf->mirrors + number;
1688
1689         if (rdev != p->rdev)
1690                 p = conf->mirrors + conf->raid_disks + number;
1691
1692         print_conf(conf);
1693         if (rdev == p->rdev) {
1694                 if (test_bit(In_sync, &rdev->flags) ||
1695                     atomic_read(&rdev->nr_pending)) {
1696                         err = -EBUSY;
1697                         goto abort;
1698                 }
1699                 /* Only remove non-faulty devices if recovery
1700                  * is not possible.
1701                  */
1702                 if (!test_bit(Faulty, &rdev->flags) &&
1703                     mddev->recovery_disabled != conf->recovery_disabled &&
1704                     mddev->degraded < conf->raid_disks) {
1705                         err = -EBUSY;
1706                         goto abort;
1707                 }
1708                 p->rdev = NULL;
1709                 synchronize_rcu();
1710                 if (atomic_read(&rdev->nr_pending)) {
1711                         /* lost the race, try later */
1712                         err = -EBUSY;
1713                         p->rdev = rdev;
1714                         goto abort;
1715                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1716                         /* We just removed a device that is being replaced.
1717                          * Move down the replacement.  We drain all IO before
1718                          * doing this to avoid confusion.
1719                          */
1720                         struct md_rdev *repl =
1721                                 conf->mirrors[conf->raid_disks + number].rdev;
1722                         freeze_array(conf, 0);
1723                         clear_bit(Replacement, &repl->flags);
1724                         p->rdev = repl;
1725                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1726                         unfreeze_array(conf);
1727                         clear_bit(WantReplacement, &rdev->flags);
1728                 } else
1729                         clear_bit(WantReplacement, &rdev->flags);
1730                 err = md_integrity_register(mddev);
1731         }
1732 abort:
1733
1734         print_conf(conf);
1735         return err;
1736 }
1737
1738
1739 static void end_sync_read(struct bio *bio, int error)
1740 {
1741         struct r1bio *r1_bio = bio->bi_private;
1742
1743         update_head_pos(r1_bio->read_disk, r1_bio);
1744
1745         /*
1746          * we have read a block, now it needs to be re-written,
1747          * or re-read if the read failed.
1748          * We don't do much here, just schedule handling by raid1d
1749          */
1750         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1751                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1752
1753         if (atomic_dec_and_test(&r1_bio->remaining))
1754                 reschedule_retry(r1_bio);
1755 }
1756
1757 static void end_sync_write(struct bio *bio, int error)
1758 {
1759         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1760         struct r1bio *r1_bio = bio->bi_private;
1761         struct mddev *mddev = r1_bio->mddev;
1762         struct r1conf *conf = mddev->private;
1763         int mirror=0;
1764         sector_t first_bad;
1765         int bad_sectors;
1766
1767         mirror = find_bio_disk(r1_bio, bio);
1768
1769         if (!uptodate) {
1770                 sector_t sync_blocks = 0;
1771                 sector_t s = r1_bio->sector;
1772                 long sectors_to_go = r1_bio->sectors;
1773                 /* make sure these bits doesn't get cleared. */
1774                 do {
1775                         bitmap_end_sync(mddev->bitmap, s,
1776                                         &sync_blocks, 1);
1777                         s += sync_blocks;
1778                         sectors_to_go -= sync_blocks;
1779                 } while (sectors_to_go > 0);
1780                 set_bit(WriteErrorSeen,
1781                         &conf->mirrors[mirror].rdev->flags);
1782                 if (!test_and_set_bit(WantReplacement,
1783                                       &conf->mirrors[mirror].rdev->flags))
1784                         set_bit(MD_RECOVERY_NEEDED, &
1785                                 mddev->recovery);
1786                 set_bit(R1BIO_WriteError, &r1_bio->state);
1787         } else if (is_badblock(conf->mirrors[mirror].rdev,
1788                                r1_bio->sector,
1789                                r1_bio->sectors,
1790                                &first_bad, &bad_sectors) &&
1791                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1792                                 r1_bio->sector,
1793                                 r1_bio->sectors,
1794                                 &first_bad, &bad_sectors)
1795                 )
1796                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1797
1798         if (atomic_dec_and_test(&r1_bio->remaining)) {
1799                 int s = r1_bio->sectors;
1800                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1801                     test_bit(R1BIO_WriteError, &r1_bio->state))
1802                         reschedule_retry(r1_bio);
1803                 else {
1804                         put_buf(r1_bio);
1805                         md_done_sync(mddev, s, uptodate);
1806                 }
1807         }
1808 }
1809
1810 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1811                             int sectors, struct page *page, int rw)
1812 {
1813         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1814                 /* success */
1815                 return 1;
1816         if (rw == WRITE) {
1817                 set_bit(WriteErrorSeen, &rdev->flags);
1818                 if (!test_and_set_bit(WantReplacement,
1819                                       &rdev->flags))
1820                         set_bit(MD_RECOVERY_NEEDED, &
1821                                 rdev->mddev->recovery);
1822         }
1823         /* need to record an error - either for the block or the device */
1824         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1825                 md_error(rdev->mddev, rdev);
1826         return 0;
1827 }
1828
1829 static int fix_sync_read_error(struct r1bio *r1_bio)
1830 {
1831         /* Try some synchronous reads of other devices to get
1832          * good data, much like with normal read errors.  Only
1833          * read into the pages we already have so we don't
1834          * need to re-issue the read request.
1835          * We don't need to freeze the array, because being in an
1836          * active sync request, there is no normal IO, and
1837          * no overlapping syncs.
1838          * We don't need to check is_badblock() again as we
1839          * made sure that anything with a bad block in range
1840          * will have bi_end_io clear.
1841          */
1842         struct mddev *mddev = r1_bio->mddev;
1843         struct r1conf *conf = mddev->private;
1844         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1845         sector_t sect = r1_bio->sector;
1846         int sectors = r1_bio->sectors;
1847         int idx = 0;
1848
1849         while(sectors) {
1850                 int s = sectors;
1851                 int d = r1_bio->read_disk;
1852                 int success = 0;
1853                 struct md_rdev *rdev;
1854                 int start;
1855
1856                 if (s > (PAGE_SIZE>>9))
1857                         s = PAGE_SIZE >> 9;
1858                 do {
1859                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1860                                 /* No rcu protection needed here devices
1861                                  * can only be removed when no resync is
1862                                  * active, and resync is currently active
1863                                  */
1864                                 rdev = conf->mirrors[d].rdev;
1865                                 if (sync_page_io(rdev, sect, s<<9,
1866                                                  bio->bi_io_vec[idx].bv_page,
1867                                                  READ, false)) {
1868                                         success = 1;
1869                                         break;
1870                                 }
1871                         }
1872                         d++;
1873                         if (d == conf->raid_disks * 2)
1874                                 d = 0;
1875                 } while (!success && d != r1_bio->read_disk);
1876
1877                 if (!success) {
1878                         char b[BDEVNAME_SIZE];
1879                         int abort = 0;
1880                         /* Cannot read from anywhere, this block is lost.
1881                          * Record a bad block on each device.  If that doesn't
1882                          * work just disable and interrupt the recovery.
1883                          * Don't fail devices as that won't really help.
1884                          */
1885                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1886                                " for block %llu\n",
1887                                mdname(mddev),
1888                                bdevname(bio->bi_bdev, b),
1889                                (unsigned long long)r1_bio->sector);
1890                         for (d = 0; d < conf->raid_disks * 2; d++) {
1891                                 rdev = conf->mirrors[d].rdev;
1892                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1893                                         continue;
1894                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1895                                         abort = 1;
1896                         }
1897                         if (abort) {
1898                                 conf->recovery_disabled =
1899                                         mddev->recovery_disabled;
1900                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1901                                 md_done_sync(mddev, r1_bio->sectors, 0);
1902                                 put_buf(r1_bio);
1903                                 return 0;
1904                         }
1905                         /* Try next page */
1906                         sectors -= s;
1907                         sect += s;
1908                         idx++;
1909                         continue;
1910                 }
1911
1912                 start = d;
1913                 /* write it back and re-read */
1914                 while (d != r1_bio->read_disk) {
1915                         if (d == 0)
1916                                 d = conf->raid_disks * 2;
1917                         d--;
1918                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1919                                 continue;
1920                         rdev = conf->mirrors[d].rdev;
1921                         if (r1_sync_page_io(rdev, sect, s,
1922                                             bio->bi_io_vec[idx].bv_page,
1923                                             WRITE) == 0) {
1924                                 r1_bio->bios[d]->bi_end_io = NULL;
1925                                 rdev_dec_pending(rdev, mddev);
1926                         }
1927                 }
1928                 d = start;
1929                 while (d != r1_bio->read_disk) {
1930                         if (d == 0)
1931                                 d = conf->raid_disks * 2;
1932                         d--;
1933                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1934                                 continue;
1935                         rdev = conf->mirrors[d].rdev;
1936                         if (r1_sync_page_io(rdev, sect, s,
1937                                             bio->bi_io_vec[idx].bv_page,
1938                                             READ) != 0)
1939                                 atomic_add(s, &rdev->corrected_errors);
1940                 }
1941                 sectors -= s;
1942                 sect += s;
1943                 idx ++;
1944         }
1945         set_bit(R1BIO_Uptodate, &r1_bio->state);
1946         set_bit(BIO_UPTODATE, &bio->bi_flags);
1947         return 1;
1948 }
1949
1950 static int process_checks(struct r1bio *r1_bio)
1951 {
1952         /* We have read all readable devices.  If we haven't
1953          * got the block, then there is no hope left.
1954          * If we have, then we want to do a comparison
1955          * and skip the write if everything is the same.
1956          * If any blocks failed to read, then we need to
1957          * attempt an over-write
1958          */
1959         struct mddev *mddev = r1_bio->mddev;
1960         struct r1conf *conf = mddev->private;
1961         int primary;
1962         int i;
1963         int vcnt;
1964
1965         /* Fix variable parts of all bios */
1966         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1967         for (i = 0; i < conf->raid_disks * 2; i++) {
1968                 int j;
1969                 int size;
1970                 int uptodate;
1971                 struct bio *b = r1_bio->bios[i];
1972                 if (b->bi_end_io != end_sync_read)
1973                         continue;
1974                 /* fixup the bio for reuse, but preserve BIO_UPTODATE */
1975                 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags);
1976                 bio_reset(b);
1977                 if (!uptodate)
1978                         clear_bit(BIO_UPTODATE, &b->bi_flags);
1979                 b->bi_vcnt = vcnt;
1980                 b->bi_iter.bi_size = r1_bio->sectors << 9;
1981                 b->bi_iter.bi_sector = r1_bio->sector +
1982                         conf->mirrors[i].rdev->data_offset;
1983                 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1984                 b->bi_end_io = end_sync_read;
1985                 b->bi_private = r1_bio;
1986
1987                 size = b->bi_iter.bi_size;
1988                 for (j = 0; j < vcnt ; j++) {
1989                         struct bio_vec *bi;
1990                         bi = &b->bi_io_vec[j];
1991                         bi->bv_offset = 0;
1992                         if (size > PAGE_SIZE)
1993                                 bi->bv_len = PAGE_SIZE;
1994                         else
1995                                 bi->bv_len = size;
1996                         size -= PAGE_SIZE;
1997                 }
1998         }
1999         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2000                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2001                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
2002                         r1_bio->bios[primary]->bi_end_io = NULL;
2003                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2004                         break;
2005                 }
2006         r1_bio->read_disk = primary;
2007         for (i = 0; i < conf->raid_disks * 2; i++) {
2008                 int j;
2009                 struct bio *pbio = r1_bio->bios[primary];
2010                 struct bio *sbio = r1_bio->bios[i];
2011                 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags);
2012
2013                 if (sbio->bi_end_io != end_sync_read)
2014                         continue;
2015                 /* Now we can 'fixup' the BIO_UPTODATE flag */
2016                 set_bit(BIO_UPTODATE, &sbio->bi_flags);
2017
2018                 if (uptodate) {
2019                         for (j = vcnt; j-- ; ) {
2020                                 struct page *p, *s;
2021                                 p = pbio->bi_io_vec[j].bv_page;
2022                                 s = sbio->bi_io_vec[j].bv_page;
2023                                 if (memcmp(page_address(p),
2024                                            page_address(s),
2025                                            sbio->bi_io_vec[j].bv_len))
2026                                         break;
2027                         }
2028                 } else
2029                         j = 0;
2030                 if (j >= 0)
2031                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2032                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2033                               && uptodate)) {
2034                         /* No need to write to this device. */
2035                         sbio->bi_end_io = NULL;
2036                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2037                         continue;
2038                 }
2039
2040                 bio_copy_data(sbio, pbio);
2041         }
2042         return 0;
2043 }
2044
2045 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2046 {
2047         struct r1conf *conf = mddev->private;
2048         int i;
2049         int disks = conf->raid_disks * 2;
2050         struct bio *bio, *wbio;
2051
2052         bio = r1_bio->bios[r1_bio->read_disk];
2053
2054         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2055                 /* ouch - failed to read all of that. */
2056                 if (!fix_sync_read_error(r1_bio))
2057                         return;
2058
2059         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2060                 if (process_checks(r1_bio) < 0)
2061                         return;
2062         /*
2063          * schedule writes
2064          */
2065         atomic_set(&r1_bio->remaining, 1);
2066         for (i = 0; i < disks ; i++) {
2067                 wbio = r1_bio->bios[i];
2068                 if (wbio->bi_end_io == NULL ||
2069                     (wbio->bi_end_io == end_sync_read &&
2070                      (i == r1_bio->read_disk ||
2071                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2072                         continue;
2073
2074                 wbio->bi_rw = WRITE;
2075                 wbio->bi_end_io = end_sync_write;
2076                 atomic_inc(&r1_bio->remaining);
2077                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2078
2079                 generic_make_request(wbio);
2080         }
2081
2082         if (atomic_dec_and_test(&r1_bio->remaining)) {
2083                 /* if we're here, all write(s) have completed, so clean up */
2084                 int s = r1_bio->sectors;
2085                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2086                     test_bit(R1BIO_WriteError, &r1_bio->state))
2087                         reschedule_retry(r1_bio);
2088                 else {
2089                         put_buf(r1_bio);
2090                         md_done_sync(mddev, s, 1);
2091                 }
2092         }
2093 }
2094
2095 /*
2096  * This is a kernel thread which:
2097  *
2098  *      1.      Retries failed read operations on working mirrors.
2099  *      2.      Updates the raid superblock when problems encounter.
2100  *      3.      Performs writes following reads for array synchronising.
2101  */
2102
2103 static void fix_read_error(struct r1conf *conf, int read_disk,
2104                            sector_t sect, int sectors)
2105 {
2106         struct mddev *mddev = conf->mddev;
2107         while(sectors) {
2108                 int s = sectors;
2109                 int d = read_disk;
2110                 int success = 0;
2111                 int start;
2112                 struct md_rdev *rdev;
2113
2114                 if (s > (PAGE_SIZE>>9))
2115                         s = PAGE_SIZE >> 9;
2116
2117                 do {
2118                         /* Note: no rcu protection needed here
2119                          * as this is synchronous in the raid1d thread
2120                          * which is the thread that might remove
2121                          * a device.  If raid1d ever becomes multi-threaded....
2122                          */
2123                         sector_t first_bad;
2124                         int bad_sectors;
2125
2126                         rdev = conf->mirrors[d].rdev;
2127                         if (rdev &&
2128                             (test_bit(In_sync, &rdev->flags) ||
2129                              (!test_bit(Faulty, &rdev->flags) &&
2130                               rdev->recovery_offset >= sect + s)) &&
2131                             is_badblock(rdev, sect, s,
2132                                         &first_bad, &bad_sectors) == 0 &&
2133                             sync_page_io(rdev, sect, s<<9,
2134                                          conf->tmppage, READ, false))
2135                                 success = 1;
2136                         else {
2137                                 d++;
2138                                 if (d == conf->raid_disks * 2)
2139                                         d = 0;
2140                         }
2141                 } while (!success && d != read_disk);
2142
2143                 if (!success) {
2144                         /* Cannot read from anywhere - mark it bad */
2145                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2146                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2147                                 md_error(mddev, rdev);
2148                         break;
2149                 }
2150                 /* write it back and re-read */
2151                 start = d;
2152                 while (d != read_disk) {
2153                         if (d==0)
2154                                 d = conf->raid_disks * 2;
2155                         d--;
2156                         rdev = conf->mirrors[d].rdev;
2157                         if (rdev &&
2158                             !test_bit(Faulty, &rdev->flags))
2159                                 r1_sync_page_io(rdev, sect, s,
2160                                                 conf->tmppage, WRITE);
2161                 }
2162                 d = start;
2163                 while (d != read_disk) {
2164                         char b[BDEVNAME_SIZE];
2165                         if (d==0)
2166                                 d = conf->raid_disks * 2;
2167                         d--;
2168                         rdev = conf->mirrors[d].rdev;
2169                         if (rdev &&
2170                             !test_bit(Faulty, &rdev->flags)) {
2171                                 if (r1_sync_page_io(rdev, sect, s,
2172                                                     conf->tmppage, READ)) {
2173                                         atomic_add(s, &rdev->corrected_errors);
2174                                         printk(KERN_INFO
2175                                                "md/raid1:%s: read error corrected "
2176                                                "(%d sectors at %llu on %s)\n",
2177                                                mdname(mddev), s,
2178                                                (unsigned long long)(sect +
2179                                                    rdev->data_offset),
2180                                                bdevname(rdev->bdev, b));
2181                                 }
2182                         }
2183                 }
2184                 sectors -= s;
2185                 sect += s;
2186         }
2187 }
2188
2189 static int narrow_write_error(struct r1bio *r1_bio, int i)
2190 {
2191         struct mddev *mddev = r1_bio->mddev;
2192         struct r1conf *conf = mddev->private;
2193         struct md_rdev *rdev = conf->mirrors[i].rdev;
2194
2195         /* bio has the data to be written to device 'i' where
2196          * we just recently had a write error.
2197          * We repeatedly clone the bio and trim down to one block,
2198          * then try the write.  Where the write fails we record
2199          * a bad block.
2200          * It is conceivable that the bio doesn't exactly align with
2201          * blocks.  We must handle this somehow.
2202          *
2203          * We currently own a reference on the rdev.
2204          */
2205
2206         int block_sectors;
2207         sector_t sector;
2208         int sectors;
2209         int sect_to_write = r1_bio->sectors;
2210         int ok = 1;
2211
2212         if (rdev->badblocks.shift < 0)
2213                 return 0;
2214
2215         block_sectors = 1 << rdev->badblocks.shift;
2216         sector = r1_bio->sector;
2217         sectors = ((sector + block_sectors)
2218                    & ~(sector_t)(block_sectors - 1))
2219                 - sector;
2220
2221         while (sect_to_write) {
2222                 struct bio *wbio;
2223                 if (sectors > sect_to_write)
2224                         sectors = sect_to_write;
2225                 /* Write at 'sector' for 'sectors'*/
2226
2227                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2228                         unsigned vcnt = r1_bio->behind_page_count;
2229                         struct bio_vec *vec = r1_bio->behind_bvecs;
2230
2231                         while (!vec->bv_page) {
2232                                 vec++;
2233                                 vcnt--;
2234                         }
2235
2236                         wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2237                         memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2238
2239                         wbio->bi_vcnt = vcnt;
2240                 } else {
2241                         wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2242                 }
2243
2244                 wbio->bi_rw = WRITE;
2245                 wbio->bi_iter.bi_sector = r1_bio->sector;
2246                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2247
2248                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2249                 wbio->bi_iter.bi_sector += rdev->data_offset;
2250                 wbio->bi_bdev = rdev->bdev;
2251                 if (submit_bio_wait(WRITE, wbio) == 0)
2252                         /* failure! */
2253                         ok = rdev_set_badblocks(rdev, sector,
2254                                                 sectors, 0)
2255                                 && ok;
2256
2257                 bio_put(wbio);
2258                 sect_to_write -= sectors;
2259                 sector += sectors;
2260                 sectors = block_sectors;
2261         }
2262         return ok;
2263 }
2264
2265 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2266 {
2267         int m;
2268         int s = r1_bio->sectors;
2269         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2270                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2271                 struct bio *bio = r1_bio->bios[m];
2272                 if (bio->bi_end_io == NULL)
2273                         continue;
2274                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2275                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2276                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2277                 }
2278                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2279                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2280                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2281                                 md_error(conf->mddev, rdev);
2282                 }
2283         }
2284         put_buf(r1_bio);
2285         md_done_sync(conf->mddev, s, 1);
2286 }
2287
2288 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2289 {
2290         int m;
2291         for (m = 0; m < conf->raid_disks * 2 ; m++)
2292                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2293                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2294                         rdev_clear_badblocks(rdev,
2295                                              r1_bio->sector,
2296                                              r1_bio->sectors, 0);
2297                         rdev_dec_pending(rdev, conf->mddev);
2298                 } else if (r1_bio->bios[m] != NULL) {
2299                         /* This drive got a write error.  We need to
2300                          * narrow down and record precise write
2301                          * errors.
2302                          */
2303                         if (!narrow_write_error(r1_bio, m)) {
2304                                 md_error(conf->mddev,
2305                                          conf->mirrors[m].rdev);
2306                                 /* an I/O failed, we can't clear the bitmap */
2307                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2308                         }
2309                         rdev_dec_pending(conf->mirrors[m].rdev,
2310                                          conf->mddev);
2311                 }
2312         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2313                 close_write(r1_bio);
2314         raid_end_bio_io(r1_bio);
2315 }
2316
2317 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2318 {
2319         int disk;
2320         int max_sectors;
2321         struct mddev *mddev = conf->mddev;
2322         struct bio *bio;
2323         char b[BDEVNAME_SIZE];
2324         struct md_rdev *rdev;
2325
2326         clear_bit(R1BIO_ReadError, &r1_bio->state);
2327         /* we got a read error. Maybe the drive is bad.  Maybe just
2328          * the block and we can fix it.
2329          * We freeze all other IO, and try reading the block from
2330          * other devices.  When we find one, we re-write
2331          * and check it that fixes the read error.
2332          * This is all done synchronously while the array is
2333          * frozen
2334          */
2335         if (mddev->ro == 0) {
2336                 freeze_array(conf, 1);
2337                 fix_read_error(conf, r1_bio->read_disk,
2338                                r1_bio->sector, r1_bio->sectors);
2339                 unfreeze_array(conf);
2340         } else
2341                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2342         rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2343
2344         bio = r1_bio->bios[r1_bio->read_disk];
2345         bdevname(bio->bi_bdev, b);
2346 read_more:
2347         disk = read_balance(conf, r1_bio, &max_sectors);
2348         if (disk == -1) {
2349                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2350                        " read error for block %llu\n",
2351                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2352                 raid_end_bio_io(r1_bio);
2353         } else {
2354                 const unsigned long do_sync
2355                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2356                 if (bio) {
2357                         r1_bio->bios[r1_bio->read_disk] =
2358                                 mddev->ro ? IO_BLOCKED : NULL;
2359                         bio_put(bio);
2360                 }
2361                 r1_bio->read_disk = disk;
2362                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2363                 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2364                          max_sectors);
2365                 r1_bio->bios[r1_bio->read_disk] = bio;
2366                 rdev = conf->mirrors[disk].rdev;
2367                 printk_ratelimited(KERN_ERR
2368                                    "md/raid1:%s: redirecting sector %llu"
2369                                    " to other mirror: %s\n",
2370                                    mdname(mddev),
2371                                    (unsigned long long)r1_bio->sector,
2372                                    bdevname(rdev->bdev, b));
2373                 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2374                 bio->bi_bdev = rdev->bdev;
2375                 bio->bi_end_io = raid1_end_read_request;
2376                 bio->bi_rw = READ | do_sync;
2377                 bio->bi_private = r1_bio;
2378                 if (max_sectors < r1_bio->sectors) {
2379                         /* Drat - have to split this up more */
2380                         struct bio *mbio = r1_bio->master_bio;
2381                         int sectors_handled = (r1_bio->sector + max_sectors
2382                                                - mbio->bi_iter.bi_sector);
2383                         r1_bio->sectors = max_sectors;
2384                         spin_lock_irq(&conf->device_lock);
2385                         if (mbio->bi_phys_segments == 0)
2386                                 mbio->bi_phys_segments = 2;
2387                         else
2388                                 mbio->bi_phys_segments++;
2389                         spin_unlock_irq(&conf->device_lock);
2390                         generic_make_request(bio);
2391                         bio = NULL;
2392
2393                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2394
2395                         r1_bio->master_bio = mbio;
2396                         r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2397                         r1_bio->state = 0;
2398                         set_bit(R1BIO_ReadError, &r1_bio->state);
2399                         r1_bio->mddev = mddev;
2400                         r1_bio->sector = mbio->bi_iter.bi_sector +
2401                                 sectors_handled;
2402
2403                         goto read_more;
2404                 } else
2405                         generic_make_request(bio);
2406         }
2407 }
2408
2409 static void raid1d(struct md_thread *thread)
2410 {
2411         struct mddev *mddev = thread->mddev;
2412         struct r1bio *r1_bio;
2413         unsigned long flags;
2414         struct r1conf *conf = mddev->private;
2415         struct list_head *head = &conf->retry_list;
2416         struct blk_plug plug;
2417
2418         md_check_recovery(mddev);
2419
2420         blk_start_plug(&plug);
2421         for (;;) {
2422
2423                 flush_pending_writes(conf);
2424
2425                 spin_lock_irqsave(&conf->device_lock, flags);
2426                 if (list_empty(head)) {
2427                         spin_unlock_irqrestore(&conf->device_lock, flags);
2428                         break;
2429                 }
2430                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2431                 list_del(head->prev);
2432                 conf->nr_queued--;
2433                 spin_unlock_irqrestore(&conf->device_lock, flags);
2434
2435                 mddev = r1_bio->mddev;
2436                 conf = mddev->private;
2437                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2438                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2439                             test_bit(R1BIO_WriteError, &r1_bio->state))
2440                                 handle_sync_write_finished(conf, r1_bio);
2441                         else
2442                                 sync_request_write(mddev, r1_bio);
2443                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2444                            test_bit(R1BIO_WriteError, &r1_bio->state))
2445                         handle_write_finished(conf, r1_bio);
2446                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2447                         handle_read_error(conf, r1_bio);
2448                 else
2449                         /* just a partial read to be scheduled from separate
2450                          * context
2451                          */
2452                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2453
2454                 cond_resched();
2455                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2456                         md_check_recovery(mddev);
2457         }
2458         blk_finish_plug(&plug);
2459 }
2460
2461
2462 static int init_resync(struct r1conf *conf)
2463 {
2464         int buffs;
2465
2466         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2467         BUG_ON(conf->r1buf_pool);
2468         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2469                                           conf->poolinfo);
2470         if (!conf->r1buf_pool)
2471                 return -ENOMEM;
2472         conf->next_resync = 0;
2473         return 0;
2474 }
2475
2476 /*
2477  * perform a "sync" on one "block"
2478  *
2479  * We need to make sure that no normal I/O request - particularly write
2480  * requests - conflict with active sync requests.
2481  *
2482  * This is achieved by tracking pending requests and a 'barrier' concept
2483  * that can be installed to exclude normal IO requests.
2484  */
2485
2486 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2487 {
2488         struct r1conf *conf = mddev->private;
2489         struct r1bio *r1_bio;
2490         struct bio *bio;
2491         sector_t max_sector, nr_sectors;
2492         int disk = -1;
2493         int i;
2494         int wonly = -1;
2495         int write_targets = 0, read_targets = 0;
2496         sector_t sync_blocks;
2497         int still_degraded = 0;
2498         int good_sectors = RESYNC_SECTORS;
2499         int min_bad = 0; /* number of sectors that are bad in all devices */
2500
2501         if (!conf->r1buf_pool)
2502                 if (init_resync(conf))
2503                         return 0;
2504
2505         max_sector = mddev->dev_sectors;
2506         if (sector_nr >= max_sector) {
2507                 /* If we aborted, we need to abort the
2508                  * sync on the 'current' bitmap chunk (there will
2509                  * only be one in raid1 resync.
2510                  * We can find the current addess in mddev->curr_resync
2511                  */
2512                 if (mddev->curr_resync < max_sector) /* aborted */
2513                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2514                                                 &sync_blocks, 1);
2515                 else /* completed sync */
2516                         conf->fullsync = 0;
2517
2518                 bitmap_close_sync(mddev->bitmap);
2519                 close_sync(conf);
2520                 return 0;
2521         }
2522
2523         if (mddev->bitmap == NULL &&
2524             mddev->recovery_cp == MaxSector &&
2525             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2526             conf->fullsync == 0) {
2527                 *skipped = 1;
2528                 return max_sector - sector_nr;
2529         }
2530         /* before building a request, check if we can skip these blocks..
2531          * This call the bitmap_start_sync doesn't actually record anything
2532          */
2533         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2534             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2535                 /* We can skip this block, and probably several more */
2536                 *skipped = 1;
2537                 return sync_blocks;
2538         }
2539         /*
2540          * If there is non-resync activity waiting for a turn,
2541          * and resync is going fast enough,
2542          * then let it though before starting on this new sync request.
2543          */
2544         if (!go_faster && conf->nr_waiting)
2545                 msleep_interruptible(1000);
2546
2547         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2548         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2549
2550         raise_barrier(conf, sector_nr);
2551
2552         rcu_read_lock();
2553         /*
2554          * If we get a correctably read error during resync or recovery,
2555          * we might want to read from a different device.  So we
2556          * flag all drives that could conceivably be read from for READ,
2557          * and any others (which will be non-In_sync devices) for WRITE.
2558          * If a read fails, we try reading from something else for which READ
2559          * is OK.
2560          */
2561
2562         r1_bio->mddev = mddev;
2563         r1_bio->sector = sector_nr;
2564         r1_bio->state = 0;
2565         set_bit(R1BIO_IsSync, &r1_bio->state);
2566
2567         for (i = 0; i < conf->raid_disks * 2; i++) {
2568                 struct md_rdev *rdev;
2569                 bio = r1_bio->bios[i];
2570                 bio_reset(bio);
2571
2572                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2573                 if (rdev == NULL ||
2574                     test_bit(Faulty, &rdev->flags)) {
2575                         if (i < conf->raid_disks)
2576                                 still_degraded = 1;
2577                 } else if (!test_bit(In_sync, &rdev->flags)) {
2578                         bio->bi_rw = WRITE;
2579                         bio->bi_end_io = end_sync_write;
2580                         write_targets ++;
2581                 } else {
2582                         /* may need to read from here */
2583                         sector_t first_bad = MaxSector;
2584                         int bad_sectors;
2585
2586                         if (is_badblock(rdev, sector_nr, good_sectors,
2587                                         &first_bad, &bad_sectors)) {
2588                                 if (first_bad > sector_nr)
2589                                         good_sectors = first_bad - sector_nr;
2590                                 else {
2591                                         bad_sectors -= (sector_nr - first_bad);
2592                                         if (min_bad == 0 ||
2593                                             min_bad > bad_sectors)
2594                                                 min_bad = bad_sectors;
2595                                 }
2596                         }
2597                         if (sector_nr < first_bad) {
2598                                 if (test_bit(WriteMostly, &rdev->flags)) {
2599                                         if (wonly < 0)
2600                                                 wonly = i;
2601                                 } else {
2602                                         if (disk < 0)
2603                                                 disk = i;
2604                                 }
2605                                 bio->bi_rw = READ;
2606                                 bio->bi_end_io = end_sync_read;
2607                                 read_targets++;
2608                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2609                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2610                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2611                                 /*
2612                                  * The device is suitable for reading (InSync),
2613                                  * but has bad block(s) here. Let's try to correct them,
2614                                  * if we are doing resync or repair. Otherwise, leave
2615                                  * this device alone for this sync request.
2616                                  */
2617                                 bio->bi_rw = WRITE;
2618                                 bio->bi_end_io = end_sync_write;
2619                                 write_targets++;
2620                         }
2621                 }
2622                 if (bio->bi_end_io) {
2623                         atomic_inc(&rdev->nr_pending);
2624                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2625                         bio->bi_bdev = rdev->bdev;
2626                         bio->bi_private = r1_bio;
2627                 }
2628         }
2629         rcu_read_unlock();
2630         if (disk < 0)
2631                 disk = wonly;
2632         r1_bio->read_disk = disk;
2633
2634         if (read_targets == 0 && min_bad > 0) {
2635                 /* These sectors are bad on all InSync devices, so we
2636                  * need to mark them bad on all write targets
2637                  */
2638                 int ok = 1;
2639                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2640                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2641                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2642                                 ok = rdev_set_badblocks(rdev, sector_nr,
2643                                                         min_bad, 0
2644                                         ) && ok;
2645                         }
2646                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2647                 *skipped = 1;
2648                 put_buf(r1_bio);
2649
2650                 if (!ok) {
2651                         /* Cannot record the badblocks, so need to
2652                          * abort the resync.
2653                          * If there are multiple read targets, could just
2654                          * fail the really bad ones ???
2655                          */
2656                         conf->recovery_disabled = mddev->recovery_disabled;
2657                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2658                         return 0;
2659                 } else
2660                         return min_bad;
2661
2662         }
2663         if (min_bad > 0 && min_bad < good_sectors) {
2664                 /* only resync enough to reach the next bad->good
2665                  * transition */
2666                 good_sectors = min_bad;
2667         }
2668
2669         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2670                 /* extra read targets are also write targets */
2671                 write_targets += read_targets-1;
2672
2673         if (write_targets == 0 || read_targets == 0) {
2674                 /* There is nowhere to write, so all non-sync
2675                  * drives must be failed - so we are finished
2676                  */
2677                 sector_t rv;
2678                 if (min_bad > 0)
2679                         max_sector = sector_nr + min_bad;
2680                 rv = max_sector - sector_nr;
2681                 *skipped = 1;
2682                 put_buf(r1_bio);
2683                 return rv;
2684         }
2685
2686         if (max_sector > mddev->resync_max)
2687                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2688         if (max_sector > sector_nr + good_sectors)
2689                 max_sector = sector_nr + good_sectors;
2690         nr_sectors = 0;
2691         sync_blocks = 0;
2692         do {
2693                 struct page *page;
2694                 int len = PAGE_SIZE;
2695                 if (sector_nr + (len>>9) > max_sector)
2696                         len = (max_sector - sector_nr) << 9;
2697                 if (len == 0)
2698                         break;
2699                 if (sync_blocks == 0) {
2700                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2701                                                &sync_blocks, still_degraded) &&
2702                             !conf->fullsync &&
2703                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2704                                 break;
2705                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2706                         if ((len >> 9) > sync_blocks)
2707                                 len = sync_blocks<<9;
2708                 }
2709
2710                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2711                         bio = r1_bio->bios[i];
2712                         if (bio->bi_end_io) {
2713                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2714                                 if (bio_add_page(bio, page, len, 0) == 0) {
2715                                         /* stop here */
2716                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2717                                         while (i > 0) {
2718                                                 i--;
2719                                                 bio = r1_bio->bios[i];
2720                                                 if (bio->bi_end_io==NULL)
2721                                                         continue;
2722                                                 /* remove last page from this bio */
2723                                                 bio->bi_vcnt--;
2724                                                 bio->bi_iter.bi_size -= len;
2725                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2726                                         }
2727                                         goto bio_full;
2728                                 }
2729                         }
2730                 }
2731                 nr_sectors += len>>9;
2732                 sector_nr += len>>9;
2733                 sync_blocks -= (len>>9);
2734         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2735  bio_full:
2736         r1_bio->sectors = nr_sectors;
2737
2738         /* For a user-requested sync, we read all readable devices and do a
2739          * compare
2740          */
2741         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2742                 atomic_set(&r1_bio->remaining, read_targets);
2743                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2744                         bio = r1_bio->bios[i];
2745                         if (bio->bi_end_io == end_sync_read) {
2746                                 read_targets--;
2747                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2748                                 generic_make_request(bio);
2749                         }
2750                 }
2751         } else {
2752                 atomic_set(&r1_bio->remaining, 1);
2753                 bio = r1_bio->bios[r1_bio->read_disk];
2754                 md_sync_acct(bio->bi_bdev, nr_sectors);
2755                 generic_make_request(bio);
2756
2757         }
2758         return nr_sectors;
2759 }
2760
2761 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2762 {
2763         if (sectors)
2764                 return sectors;
2765
2766         return mddev->dev_sectors;
2767 }
2768
2769 static struct r1conf *setup_conf(struct mddev *mddev)
2770 {
2771         struct r1conf *conf;
2772         int i;
2773         struct raid1_info *disk;
2774         struct md_rdev *rdev;
2775         int err = -ENOMEM;
2776
2777         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2778         if (!conf)
2779                 goto abort;
2780
2781         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2782                                 * mddev->raid_disks * 2,
2783                                  GFP_KERNEL);
2784         if (!conf->mirrors)
2785                 goto abort;
2786
2787         conf->tmppage = alloc_page(GFP_KERNEL);
2788         if (!conf->tmppage)
2789                 goto abort;
2790
2791         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2792         if (!conf->poolinfo)
2793                 goto abort;
2794         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2795         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2796                                           r1bio_pool_free,
2797                                           conf->poolinfo);
2798         if (!conf->r1bio_pool)
2799                 goto abort;
2800
2801         conf->poolinfo->mddev = mddev;
2802
2803         err = -EINVAL;
2804         spin_lock_init(&conf->device_lock);
2805         rdev_for_each(rdev, mddev) {
2806                 struct request_queue *q;
2807                 int disk_idx = rdev->raid_disk;
2808                 if (disk_idx >= mddev->raid_disks
2809                     || disk_idx < 0)
2810                         continue;
2811                 if (test_bit(Replacement, &rdev->flags))
2812                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
2813                 else
2814                         disk = conf->mirrors + disk_idx;
2815
2816                 if (disk->rdev)
2817                         goto abort;
2818                 disk->rdev = rdev;
2819                 q = bdev_get_queue(rdev->bdev);
2820                 if (q->merge_bvec_fn)
2821                         mddev->merge_check_needed = 1;
2822
2823                 disk->head_position = 0;
2824                 disk->seq_start = MaxSector;
2825         }
2826         conf->raid_disks = mddev->raid_disks;
2827         conf->mddev = mddev;
2828         INIT_LIST_HEAD(&conf->retry_list);
2829
2830         spin_lock_init(&conf->resync_lock);
2831         init_waitqueue_head(&conf->wait_barrier);
2832
2833         bio_list_init(&conf->pending_bio_list);
2834         conf->pending_count = 0;
2835         conf->recovery_disabled = mddev->recovery_disabled - 1;
2836
2837         conf->start_next_window = MaxSector;
2838         conf->current_window_requests = conf->next_window_requests = 0;
2839
2840         err = -EIO;
2841         for (i = 0; i < conf->raid_disks * 2; i++) {
2842
2843                 disk = conf->mirrors + i;
2844
2845                 if (i < conf->raid_disks &&
2846                     disk[conf->raid_disks].rdev) {
2847                         /* This slot has a replacement. */
2848                         if (!disk->rdev) {
2849                                 /* No original, just make the replacement
2850                                  * a recovering spare
2851                                  */
2852                                 disk->rdev =
2853                                         disk[conf->raid_disks].rdev;
2854                                 disk[conf->raid_disks].rdev = NULL;
2855                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2856                                 /* Original is not in_sync - bad */
2857                                 goto abort;
2858                 }
2859
2860                 if (!disk->rdev ||
2861                     !test_bit(In_sync, &disk->rdev->flags)) {
2862                         disk->head_position = 0;
2863                         if (disk->rdev &&
2864                             (disk->rdev->saved_raid_disk < 0))
2865                                 conf->fullsync = 1;
2866                 }
2867         }
2868
2869         err = -ENOMEM;
2870         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2871         if (!conf->thread) {
2872                 printk(KERN_ERR
2873                        "md/raid1:%s: couldn't allocate thread\n",
2874                        mdname(mddev));
2875                 goto abort;
2876         }
2877
2878         return conf;
2879
2880  abort:
2881         if (conf) {
2882                 if (conf->r1bio_pool)
2883                         mempool_destroy(conf->r1bio_pool);
2884                 kfree(conf->mirrors);
2885                 safe_put_page(conf->tmppage);
2886                 kfree(conf->poolinfo);
2887                 kfree(conf);
2888         }
2889         return ERR_PTR(err);
2890 }
2891
2892 static int stop(struct mddev *mddev);
2893 static int run(struct mddev *mddev)
2894 {
2895         struct r1conf *conf;
2896         int i;
2897         struct md_rdev *rdev;
2898         int ret;
2899         bool discard_supported = false;
2900
2901         if (mddev->level != 1) {
2902                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2903                        mdname(mddev), mddev->level);
2904                 return -EIO;
2905         }
2906         if (mddev->reshape_position != MaxSector) {
2907                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2908                        mdname(mddev));
2909                 return -EIO;
2910         }
2911         /*
2912          * copy the already verified devices into our private RAID1
2913          * bookkeeping area. [whatever we allocate in run(),
2914          * should be freed in stop()]
2915          */
2916         if (mddev->private == NULL)
2917                 conf = setup_conf(mddev);
2918         else
2919                 conf = mddev->private;
2920
2921         if (IS_ERR(conf))
2922                 return PTR_ERR(conf);
2923
2924         if (mddev->queue)
2925                 blk_queue_max_write_same_sectors(mddev->queue, 0);
2926
2927         rdev_for_each(rdev, mddev) {
2928                 if (!mddev->gendisk)
2929                         continue;
2930                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2931                                   rdev->data_offset << 9);
2932                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2933                         discard_supported = true;
2934         }
2935
2936         mddev->degraded = 0;
2937         for (i=0; i < conf->raid_disks; i++)
2938                 if (conf->mirrors[i].rdev == NULL ||
2939                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2940                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2941                         mddev->degraded++;
2942
2943         if (conf->raid_disks - mddev->degraded == 1)
2944                 mddev->recovery_cp = MaxSector;
2945
2946         if (mddev->recovery_cp != MaxSector)
2947                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2948                        " -- starting background reconstruction\n",
2949                        mdname(mddev));
2950         printk(KERN_INFO
2951                 "md/raid1:%s: active with %d out of %d mirrors\n",
2952                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2953                 mddev->raid_disks);
2954
2955         /*
2956          * Ok, everything is just fine now
2957          */
2958         mddev->thread = conf->thread;
2959         conf->thread = NULL;
2960         mddev->private = conf;
2961
2962         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2963
2964         if (mddev->queue) {
2965                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2966                 mddev->queue->backing_dev_info.congested_data = mddev;
2967                 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2968
2969                 if (discard_supported)
2970                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2971                                                 mddev->queue);
2972                 else
2973                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2974                                                   mddev->queue);
2975         }
2976
2977         ret =  md_integrity_register(mddev);
2978         if (ret)
2979                 stop(mddev);
2980         return ret;
2981 }
2982
2983 static int stop(struct mddev *mddev)
2984 {
2985         struct r1conf *conf = mddev->private;
2986         struct bitmap *bitmap = mddev->bitmap;
2987
2988         /* wait for behind writes to complete */
2989         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2990                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2991                        mdname(mddev));
2992                 /* need to kick something here to make sure I/O goes? */
2993                 wait_event(bitmap->behind_wait,
2994                            atomic_read(&bitmap->behind_writes) == 0);
2995         }
2996
2997         freeze_array(conf, 0);
2998         unfreeze_array(conf);
2999
3000         md_unregister_thread(&mddev->thread);
3001         if (conf->r1bio_pool)
3002                 mempool_destroy(conf->r1bio_pool);
3003         kfree(conf->mirrors);
3004         safe_put_page(conf->tmppage);
3005         kfree(conf->poolinfo);
3006         kfree(conf);
3007         mddev->private = NULL;
3008         return 0;
3009 }
3010
3011 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3012 {
3013         /* no resync is happening, and there is enough space
3014          * on all devices, so we can resize.
3015          * We need to make sure resync covers any new space.
3016          * If the array is shrinking we should possibly wait until
3017          * any io in the removed space completes, but it hardly seems
3018          * worth it.
3019          */
3020         sector_t newsize = raid1_size(mddev, sectors, 0);
3021         if (mddev->external_size &&
3022             mddev->array_sectors > newsize)
3023                 return -EINVAL;
3024         if (mddev->bitmap) {
3025                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3026                 if (ret)
3027                         return ret;
3028         }
3029         md_set_array_sectors(mddev, newsize);
3030         set_capacity(mddev->gendisk, mddev->array_sectors);
3031         revalidate_disk(mddev->gendisk);
3032         if (sectors > mddev->dev_sectors &&
3033             mddev->recovery_cp > mddev->dev_sectors) {
3034                 mddev->recovery_cp = mddev->dev_sectors;
3035                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3036         }
3037         mddev->dev_sectors = sectors;
3038         mddev->resync_max_sectors = sectors;
3039         return 0;
3040 }
3041
3042 static int raid1_reshape(struct mddev *mddev)
3043 {
3044         /* We need to:
3045          * 1/ resize the r1bio_pool
3046          * 2/ resize conf->mirrors
3047          *
3048          * We allocate a new r1bio_pool if we can.
3049          * Then raise a device barrier and wait until all IO stops.
3050          * Then resize conf->mirrors and swap in the new r1bio pool.
3051          *
3052          * At the same time, we "pack" the devices so that all the missing
3053          * devices have the higher raid_disk numbers.
3054          */
3055         mempool_t *newpool, *oldpool;
3056         struct pool_info *newpoolinfo;
3057         struct raid1_info *newmirrors;
3058         struct r1conf *conf = mddev->private;
3059         int cnt, raid_disks;
3060         unsigned long flags;
3061         int d, d2, err;
3062
3063         /* Cannot change chunk_size, layout, or level */
3064         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3065             mddev->layout != mddev->new_layout ||
3066             mddev->level != mddev->new_level) {
3067                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3068                 mddev->new_layout = mddev->layout;
3069                 mddev->new_level = mddev->level;
3070                 return -EINVAL;
3071         }
3072
3073         err = md_allow_write(mddev);
3074         if (err)
3075                 return err;
3076
3077         raid_disks = mddev->raid_disks + mddev->delta_disks;
3078
3079         if (raid_disks < conf->raid_disks) {
3080                 cnt=0;
3081                 for (d= 0; d < conf->raid_disks; d++)
3082                         if (conf->mirrors[d].rdev)
3083                                 cnt++;
3084                 if (cnt > raid_disks)
3085                         return -EBUSY;
3086         }
3087
3088         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3089         if (!newpoolinfo)
3090                 return -ENOMEM;
3091         newpoolinfo->mddev = mddev;
3092         newpoolinfo->raid_disks = raid_disks * 2;
3093
3094         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3095                                  r1bio_pool_free, newpoolinfo);
3096         if (!newpool) {
3097                 kfree(newpoolinfo);
3098                 return -ENOMEM;
3099         }
3100         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3101                              GFP_KERNEL);
3102         if (!newmirrors) {
3103                 kfree(newpoolinfo);
3104                 mempool_destroy(newpool);
3105                 return -ENOMEM;
3106         }
3107
3108         freeze_array(conf, 0);
3109
3110         /* ok, everything is stopped */
3111         oldpool = conf->r1bio_pool;
3112         conf->r1bio_pool = newpool;
3113
3114         for (d = d2 = 0; d < conf->raid_disks; d++) {
3115                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3116                 if (rdev && rdev->raid_disk != d2) {
3117                         sysfs_unlink_rdev(mddev, rdev);
3118                         rdev->raid_disk = d2;
3119                         sysfs_unlink_rdev(mddev, rdev);
3120                         if (sysfs_link_rdev(mddev, rdev))
3121                                 printk(KERN_WARNING
3122                                        "md/raid1:%s: cannot register rd%d\n",
3123                                        mdname(mddev), rdev->raid_disk);
3124                 }
3125                 if (rdev)
3126                         newmirrors[d2++].rdev = rdev;
3127         }
3128         kfree(conf->mirrors);
3129         conf->mirrors = newmirrors;
3130         kfree(conf->poolinfo);
3131         conf->poolinfo = newpoolinfo;
3132
3133         spin_lock_irqsave(&conf->device_lock, flags);
3134         mddev->degraded += (raid_disks - conf->raid_disks);
3135         spin_unlock_irqrestore(&conf->device_lock, flags);
3136         conf->raid_disks = mddev->raid_disks = raid_disks;
3137         mddev->delta_disks = 0;
3138
3139         unfreeze_array(conf);
3140
3141         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3142         md_wakeup_thread(mddev->thread);
3143
3144         mempool_destroy(oldpool);
3145         return 0;
3146 }
3147
3148 static void raid1_quiesce(struct mddev *mddev, int state)
3149 {
3150         struct r1conf *conf = mddev->private;
3151
3152         switch(state) {
3153         case 2: /* wake for suspend */
3154                 wake_up(&conf->wait_barrier);
3155                 break;
3156         case 1:
3157                 freeze_array(conf, 0);
3158                 break;
3159         case 0:
3160                 unfreeze_array(conf);
3161                 break;
3162         }
3163 }
3164
3165 static void *raid1_takeover(struct mddev *mddev)
3166 {
3167         /* raid1 can take over:
3168          *  raid5 with 2 devices, any layout or chunk size
3169          */
3170         if (mddev->level == 5 && mddev->raid_disks == 2) {
3171                 struct r1conf *conf;
3172                 mddev->new_level = 1;
3173                 mddev->new_layout = 0;
3174                 mddev->new_chunk_sectors = 0;
3175                 conf = setup_conf(mddev);
3176                 if (!IS_ERR(conf))
3177                         /* Array must appear to be quiesced */
3178                         conf->array_frozen = 1;
3179                 return conf;
3180         }
3181         return ERR_PTR(-EINVAL);
3182 }
3183
3184 static struct md_personality raid1_personality =
3185 {
3186         .name           = "raid1",
3187         .level          = 1,
3188         .owner          = THIS_MODULE,
3189         .make_request   = make_request,
3190         .run            = run,
3191         .stop           = stop,
3192         .status         = status,
3193         .error_handler  = error,
3194         .hot_add_disk   = raid1_add_disk,
3195         .hot_remove_disk= raid1_remove_disk,
3196         .spare_active   = raid1_spare_active,
3197         .sync_request   = sync_request,
3198         .resize         = raid1_resize,
3199         .size           = raid1_size,
3200         .check_reshape  = raid1_reshape,
3201         .quiesce        = raid1_quiesce,
3202         .takeover       = raid1_takeover,
3203 };
3204
3205 static int __init raid_init(void)
3206 {
3207         return register_md_personality(&raid1_personality);
3208 }
3209
3210 static void raid_exit(void)
3211 {
3212         unregister_md_personality(&raid1_personality);
3213 }
3214
3215 module_init(raid_init);
3216 module_exit(raid_exit);
3217 MODULE_LICENSE("GPL");
3218 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3219 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3220 MODULE_ALIAS("md-raid1");
3221 MODULE_ALIAS("md-level-1");
3222
3223 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);