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 <[email protected]>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <[email protected]>
  13  * Various fixes by Neil Brown <[email protected]>
  14  *
  15  * Changes by Peter T. Breuer <[email protected]> 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  * This routine returns the disk from which the requested read should
 499  * be done. There is a per-array 'next expected sequential IO' sector
 500  * number - if this matches on the next IO then we use the last disk.
 501  * There is also a per-disk 'last know head position' sector that is
 502  * maintained from IRQ contexts, both the normal and the resync IO
 503  * completion handlers update this position correctly. If there is no
 504  * perfect sequential match then we pick the disk whose head is closest.
 505  *
 506  * If there are 2 mirrors in the same 2 devices, performance degrades
 507  * because position is mirror, not device based.
 508  *
 509  * The rdev for the device selected will have nr_pending incremented.
 510  */
 511 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 512 {
 513         const sector_t this_sector = r1_bio->sector;
 514         int sectors;
 515         int best_good_sectors;
 516         int best_disk, best_dist_disk, best_pending_disk;
 517         int has_nonrot_disk;
 518         int disk;
 519         sector_t best_dist;
 520         unsigned int min_pending;
 521         struct md_rdev *rdev;
 522         int choose_first;
 523         int choose_next_idle;
 524
 525         rcu_read_lock();
 526         /*
 527          * Check if we can balance. We can balance on the whole
 528          * device if no resync is going on, or below the resync window.
 529          * We take the first readable disk when above the resync window.
 530          */
 531  retry:
 532         sectors = r1_bio->sectors;
 533         best_disk = -1;
 534         best_dist_disk = -1;
 535         best_dist = MaxSector;
 536         best_pending_disk = -1;
 537         min_pending = UINT_MAX;
 538         best_good_sectors = 0;
 539         has_nonrot_disk = 0;
 540         choose_next_idle = 0;
 541
 542         choose_first = (conf->mddev->recovery_cp < this_sector + sectors);
 543
 544         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 545                 sector_t dist;
 546                 sector_t first_bad;
 547                 int bad_sectors;
 548                 unsigned int pending;
 549                 bool nonrot;
 550
 551                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 552                 if (r1_bio->bios[disk] == IO_BLOCKED
 553                     || rdev == NULL
 554                     || test_bit(Unmerged, &rdev->flags)
 555                     || test_bit(Faulty, &rdev->flags))
 556                         continue;
 557                 if (!test_bit(In_sync, &rdev->flags) &&
 558                     rdev->recovery_offset < this_sector + sectors)
 559                         continue;
 560                 if (test_bit(WriteMostly, &rdev->flags)) {
 561                         /* Don't balance among write-mostly, just
 562                          * use the first as a last resort */
 563                         if (best_disk < 0) {
 564                                 if (is_badblock(rdev, this_sector, sectors,
 565                                                 &first_bad, &bad_sectors)) {
 566                                         if (first_bad < this_sector)
 567                                                 /* Cannot use this */
 568                                                 continue;
 569                                         best_good_sectors = first_bad - this_sector;
 570                                 } else
 571                                         best_good_sectors = sectors;
 572                                 best_disk = disk;
 573                         }
 574                         continue;
 575                 }
 576                 /* This is a reasonable device to use.  It might
 577                  * even be best.
 578                  */
 579                 if (is_badblock(rdev, this_sector, sectors,
 580                                 &first_bad, &bad_sectors)) {
 581                         if (best_dist < MaxSector)
 582                                 /* already have a better device */
 583                                 continue;
 584                         if (first_bad <= this_sector) {
 585                                 /* cannot read here. If this is the 'primary'
 586                                  * device, then we must not read beyond
 587                                  * bad_sectors from another device..
 588                                  */
 589                                 bad_sectors -= (this_sector - first_bad);
 590                                 if (choose_first && sectors > bad_sectors)
 591                                         sectors = bad_sectors;
 592                                 if (best_good_sectors > sectors)
 593                                         best_good_sectors = sectors;
 594
 595                         } else {
 596                                 sector_t good_sectors = first_bad - this_sector;
 597                                 if (good_sectors > best_good_sectors) {
 598                                         best_good_sectors = good_sectors;
 599                                         best_disk = disk;
 600                                 }
 601                                 if (choose_first)
 602                                         break;
 603                         }
 604                         continue;
 605                 } else
 606                         best_good_sectors = sectors;
 607
 608                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 609                 has_nonrot_disk |= nonrot;
 610                 pending = atomic_read(&rdev->nr_pending);
 611                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 612                 if (choose_first) {
 613                         best_disk = disk;
 614                         break;
 615                 }
 616                 /* Don't change to another disk for sequential reads */
 617                 if (conf->mirrors[disk].next_seq_sect == this_sector
 618                     || dist == 0) {
 619                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 620                         struct raid1_info *mirror = &conf->mirrors[disk];
 621
 622                         best_disk = disk;
 623                         /*
 624                          * If buffered sequential IO size exceeds optimal
 625                          * iosize, check if there is idle disk. If yes, choose
 626                          * the idle disk. read_balance could already choose an
 627                          * idle disk before noticing it's a sequential IO in
 628                          * this disk. This doesn't matter because this disk
 629                          * will idle, next time it will be utilized after the
 630                          * first disk has IO size exceeds optimal iosize. In
 631                          * this way, iosize of the first disk will be optimal
 632                          * iosize at least. iosize of the second disk might be
 633                          * small, but not a big deal since when the second disk
 634                          * starts IO, the first disk is likely still busy.
 635                          */
 636                         if (nonrot && opt_iosize > 0 &&
 637                             mirror->seq_start != MaxSector &&
 638                             mirror->next_seq_sect > opt_iosize &&
 639                             mirror->next_seq_sect - opt_iosize >=
 640                             mirror->seq_start) {
 641                                 choose_next_idle = 1;
 642                                 continue;
 643                         }
 644                         break;
 645                 }
 646                 /* If device is idle, use it */
 647                 if (pending == 0) {
 648                         best_disk = disk;
 649                         break;
 650                 }
 651
 652                 if (choose_next_idle)
 653                         continue;
 654
 655                 if (min_pending > pending) {
 656                         min_pending = pending;
 657                         best_pending_disk = disk;
 658                 }
 659
 660                 if (dist < best_dist) {
 661                         best_dist = dist;
 662                         best_dist_disk = disk;
 663                 }
 664         }
 665
 666         /*
 667          * If all disks are rotational, choose the closest disk. If any disk is
 668          * non-rotational, choose the disk with less pending request even the
 669          * disk is rotational, which might/might not be optimal for raids with
 670          * mixed ratation/non-rotational disks depending on workload.
 671          */
 672         if (best_disk == -1) {
 673                 if (has_nonrot_disk)
 674                         best_disk = best_pending_disk;
 675                 else
 676                         best_disk = best_dist_disk;
 677         }
 678
 679         if (best_disk >= 0) {
 680                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 681                 if (!rdev)
 682                         goto retry;
 683                 atomic_inc(&rdev->nr_pending);
 684                 if (test_bit(Faulty, &rdev->flags)) {
 685                         /* cannot risk returning a device that failed
 686                          * before we inc'ed nr_pending
 687                          */
 688                         rdev_dec_pending(rdev, conf->mddev);
 689                         goto retry;
 690                 }
 691                 sectors = best_good_sectors;
 692
 693                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 694                         conf->mirrors[best_disk].seq_start = this_sector;
 695
 696                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 697         }
 698         rcu_read_unlock();
 699         *max_sectors = sectors;
 700
 701         return best_disk;
 702 }
 703
 704 static int raid1_mergeable_bvec(struct request_queue *q,
 705                                 struct bvec_merge_data *bvm,
 706                                 struct bio_vec *biovec)
 707 {
 708         struct mddev *mddev = q->queuedata;
 709         struct r1conf *conf = mddev->private;
 710         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 711         int max = biovec->bv_len;
 712
 713         if (mddev->merge_check_needed) {
 714                 int disk;
 715                 rcu_read_lock();
 716                 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
 717                         struct md_rdev *rdev = rcu_dereference(
 718                                 conf->mirrors[disk].rdev);
 719                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 720                                 struct request_queue *q =
 721                                         bdev_get_queue(rdev->bdev);
 722                                 if (q->merge_bvec_fn) {
 723                                         bvm->bi_sector = sector +
 724                                                 rdev->data_offset;
 725                                         bvm->bi_bdev = rdev->bdev;
 726                                         max = min(max, q->merge_bvec_fn(
 727                                                           q, bvm, biovec));
 728                                 }
 729                         }
 730                 }
 731                 rcu_read_unlock();
 732         }
 733         return max;
 734
 735 }
 736
 737 int md_raid1_congested(struct mddev *mddev, int bits)
 738 {
 739         struct r1conf *conf = mddev->private;
 740         int i, ret = 0;
 741
 742         if ((bits & (1 << BDI_async_congested)) &&
 743             conf->pending_count >= max_queued_requests)
 744                 return 1;
 745
 746         rcu_read_lock();
 747         for (i = 0; i < conf->raid_disks * 2; i++) {
 748                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 749                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 750                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 751
 752                         BUG_ON(!q);
 753
 754                         /* Note the '|| 1' - when read_balance prefers
 755                          * non-congested targets, it can be removed
 756                          */
 757                         if ((bits & (1<<BDI_async_congested)) || 1)
 758                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 759                         else
 760                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 761                 }
 762         }
 763         rcu_read_unlock();
 764         return ret;
 765 }
 766 EXPORT_SYMBOL_GPL(md_raid1_congested);
 767
 768 static int raid1_congested(void *data, int bits)
 769 {
 770         struct mddev *mddev = data;
 771
 772         return mddev_congested(mddev, bits) ||
 773                 md_raid1_congested(mddev, bits);
 774 }
 775
 776 static void flush_pending_writes(struct r1conf *conf)
 777 {
 778         /* Any writes that have been queued but are awaiting
 779          * bitmap updates get flushed here.
 780          */
 781         spin_lock_irq(&conf->device_lock);
 782
 783         if (conf->pending_bio_list.head) {
 784                 struct bio *bio;
 785                 bio = bio_list_get(&conf->pending_bio_list);
 786                 conf->pending_count = 0;
 787                 spin_unlock_irq(&conf->device_lock);
 788                 /* flush any pending bitmap writes to
 789                  * disk before proceeding w/ I/O */
 790                 bitmap_unplug(conf->mddev->bitmap);
 791                 wake_up(&conf->wait_barrier);
 792
 793                 while (bio) { /* submit pending writes */
 794                         struct bio *next = bio->bi_next;
 795                         bio->bi_next = NULL;
 796                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 797                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 798                                 /* Just ignore it */
 799                                 bio_endio(bio, 0);
 800                         else
 801                                 generic_make_request(bio);
 802                         bio = next;
 803                 }
 804         } else
 805                 spin_unlock_irq(&conf->device_lock);
 806 }
 807
 808 /* Barriers....
 809  * Sometimes we need to suspend IO while we do something else,
 810  * either some resync/recovery, or reconfigure the array.
 811  * To do this we raise a 'barrier'.
 812  * The 'barrier' is a counter that can be raised multiple times
 813  * to count how many activities are happening which preclude
 814  * normal IO.
 815  * We can only raise the barrier if there is no pending IO.
 816  * i.e. if nr_pending == 0.
 817  * We choose only to raise the barrier if no-one is waiting for the
 818  * barrier to go down.  This means that as soon as an IO request
 819  * is ready, no other operations which require a barrier will start
 820  * until the IO request has had a chance.
 821  *
 822  * So: regular IO calls 'wait_barrier'.  When that returns there
 823  *    is no backgroup IO happening,  It must arrange to call
 824  *    allow_barrier when it has finished its IO.
 825  * backgroup IO calls must call raise_barrier.  Once that returns
 826  *    there is no normal IO happeing.  It must arrange to call
 827  *    lower_barrier when the particular background IO completes.
 828  */
 829 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
 830 {
 831         spin_lock_irq(&conf->resync_lock);
 832
 833         /* Wait until no block IO is waiting */
 834         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 835                             conf->resync_lock);
 836
 837         /* block any new IO from starting */
 838         conf->barrier++;
 839         conf->next_resync = sector_nr;
 840
 841         /* For these conditions we must wait:
 842          * A: while the array is in frozen state
 843          * B: while barrier >= RESYNC_DEPTH, meaning resync reach
 844          *    the max count which allowed.
 845          * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
 846          *    next resync will reach to the window which normal bios are
 847          *    handling.
 848          * D: while there are any active requests in the current window.
 849          */
 850         wait_event_lock_irq(conf->wait_barrier,
 851                             !conf->array_frozen &&
 852                             conf->barrier < RESYNC_DEPTH &&
 853                             conf->current_window_requests == 0 &&
 854                             (conf->start_next_window >=
 855                              conf->next_resync + RESYNC_SECTORS),
 856                             conf->resync_lock);
 857
 858         conf->nr_pending++;
 859         spin_unlock_irq(&conf->resync_lock);
 860 }
 861
 862 static void lower_barrier(struct r1conf *conf)
 863 {
 864         unsigned long flags;
 865         BUG_ON(conf->barrier <= 0);
 866         spin_lock_irqsave(&conf->resync_lock, flags);
 867         conf->barrier--;
 868         conf->nr_pending--;
 869         spin_unlock_irqrestore(&conf->resync_lock, flags);
 870         wake_up(&conf->wait_barrier);
 871 }
 872
 873 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio)
 874 {
 875         bool wait = false;
 876
 877         if (conf->array_frozen || !bio)
 878                 wait = true;
 879         else if (conf->barrier && bio_data_dir(bio) == WRITE) {
 880                 if ((conf->mddev->curr_resync_completed
 881                      >= bio_end_sector(bio)) ||
 882                     (conf->next_resync + NEXT_NORMALIO_DISTANCE
 883                      <= bio->bi_iter.bi_sector))
 884                         wait = false;
 885                 else
 886                         wait = true;
 887         }
 888
 889         return wait;
 890 }
 891
 892 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
 893 {
 894         sector_t sector = 0;
 895
 896         spin_lock_irq(&conf->resync_lock);
 897         if (need_to_wait_for_sync(conf, bio)) {
 898                 conf->nr_waiting++;
 899                 /* Wait for the barrier to drop.
 900                  * However if there are already pending
 901                  * requests (preventing the barrier from
 902                  * rising completely), and the
 903                  * per-process bio queue isn't empty,
 904                  * then don't wait, as we need to empty
 905                  * that queue to allow conf->start_next_window
 906                  * to increase.
 907                  */
 908                 wait_event_lock_irq(conf->wait_barrier,
 909                                     !conf->array_frozen &&
 910                                     (!conf->barrier ||
 911                                      ((conf->start_next_window <
 912                                        conf->next_resync + RESYNC_SECTORS) &&
 913                                       current->bio_list &&
 914                                       !bio_list_empty(current->bio_list))),
 915                                     conf->resync_lock);
 916                 conf->nr_waiting--;
 917         }
 918
 919         if (bio && bio_data_dir(bio) == WRITE) {
 920                 if (bio->bi_iter.bi_sector >=
 921                     conf->mddev->curr_resync_completed) {
 922                         if (conf->start_next_window == MaxSector)
 923                                 conf->start_next_window =
 924                                         conf->next_resync +
 925                                         NEXT_NORMALIO_DISTANCE;
 926
 927                         if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
 928                             <= bio->bi_iter.bi_sector)
 929                                 conf->next_window_requests++;
 930                         else
 931                                 conf->current_window_requests++;
 932                         sector = conf->start_next_window;
 933                 }
 934         }
 935
 936         conf->nr_pending++;
 937         spin_unlock_irq(&conf->resync_lock);
 938         return sector;
 939 }
 940
 941 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
 942                           sector_t bi_sector)
 943 {
 944         unsigned long flags;
 945
 946         spin_lock_irqsave(&conf->resync_lock, flags);
 947         conf->nr_pending--;
 948         if (start_next_window) {
 949                 if (start_next_window == conf->start_next_window) {
 950                         if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
 951                             <= bi_sector)
 952                                 conf->next_window_requests--;
 953                         else
 954                                 conf->current_window_requests--;
 955                 } else
 956                         conf->current_window_requests--;
 957
 958                 if (!conf->current_window_requests) {
 959                         if (conf->next_window_requests) {
 960                                 conf->current_window_requests =
 961                                         conf->next_window_requests;
 962                                 conf->next_window_requests = 0;
 963                                 conf->start_next_window +=
 964                                         NEXT_NORMALIO_DISTANCE;
 965                         } else
 966                                 conf->start_next_window = MaxSector;
 967                 }
 968         }
 969         spin_unlock_irqrestore(&conf->resync_lock, flags);
 970         wake_up(&conf->wait_barrier);
 971 }
 972
 973 static void freeze_array(struct r1conf *conf, int extra)
 974 {
 975         /* stop syncio and normal IO and wait for everything to
 976          * go quite.
 977          * We wait until nr_pending match nr_queued+extra
 978          * This is called in the context of one normal IO request
 979          * that has failed. Thus any sync request that might be pending
 980          * will be blocked by nr_pending, and we need to wait for
 981          * pending IO requests to complete or be queued for re-try.
 982          * Thus the number queued (nr_queued) plus this request (extra)
 983          * must match the number of pending IOs (nr_pending) before
 984          * we continue.
 985          */
 986         spin_lock_irq(&conf->resync_lock);
 987         conf->array_frozen = 1;
 988         wait_event_lock_irq_cmd(conf->wait_barrier,
 989                                 conf->nr_pending == conf->nr_queued+extra,
 990                                 conf->resync_lock,
 991                                 flush_pending_writes(conf));
 992         spin_unlock_irq(&conf->resync_lock);
 993 }
 994 static void unfreeze_array(struct r1conf *conf)
 995 {
 996         /* reverse the effect of the freeze */
 997         spin_lock_irq(&conf->resync_lock);
 998         conf->array_frozen = 0;
 999         wake_up(&conf->wait_barrier);
1000         spin_unlock_irq(&conf->resync_lock);
1001 }
1002
1003 /* duplicate the data pages for behind I/O
1004  */
1005 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
1006 {
1007         int i;
1008         struct bio_vec *bvec;
1009         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
1010                                         GFP_NOIO);
1011         if (unlikely(!bvecs))
1012                 return;
1013
1014         bio_for_each_segment_all(bvec, bio, i) {
1015                 bvecs[i] = *bvec;
1016                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
1017                 if (unlikely(!bvecs[i].bv_page))
1018                         goto do_sync_io;
1019                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
1020                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
1021                 kunmap(bvecs[i].bv_page);
1022                 kunmap(bvec->bv_page);
1023         }
1024         r1_bio->behind_bvecs = bvecs;
1025         r1_bio->behind_page_count = bio->bi_vcnt;
1026         set_bit(R1BIO_BehindIO, &r1_bio->state);
1027         return;
1028
1029 do_sync_io:
1030         for (i = 0; i < bio->bi_vcnt; i++)
1031                 if (bvecs[i].bv_page)
1032                         put_page(bvecs[i].bv_page);
1033         kfree(bvecs);
1034         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1035                  bio->bi_iter.bi_size);
1036 }
1037
1038 struct raid1_plug_cb {
1039         struct blk_plug_cb      cb;
1040         struct bio_list         pending;
1041         int                     pending_cnt;
1042 };
1043
1044 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1045 {
1046         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1047                                                   cb);
1048         struct mddev *mddev = plug->cb.data;
1049         struct r1conf *conf = mddev->private;
1050         struct bio *bio;
1051
1052         if (from_schedule || current->bio_list) {
1053                 spin_lock_irq(&conf->device_lock);
1054                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1055                 conf->pending_count += plug->pending_cnt;
1056                 spin_unlock_irq(&conf->device_lock);
1057                 wake_up(&conf->wait_barrier);
1058                 md_wakeup_thread(mddev->thread);
1059                 kfree(plug);
1060                 return;
1061         }
1062
1063         /* we aren't scheduling, so we can do the write-out directly. */
1064         bio = bio_list_get(&plug->pending);
1065         bitmap_unplug(mddev->bitmap);
1066         wake_up(&conf->wait_barrier);
1067
1068         while (bio) { /* submit pending writes */
1069                 struct bio *next = bio->bi_next;
1070                 bio->bi_next = NULL;
1071                 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1072                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1073                         /* Just ignore it */
1074                         bio_endio(bio, 0);
1075                 else
1076                         generic_make_request(bio);
1077                 bio = next;
1078         }
1079         kfree(plug);
1080 }
1081
1082 static void make_request(struct mddev *mddev, struct bio * bio)
1083 {
1084         struct r1conf *conf = mddev->private;
1085         struct raid1_info *mirror;
1086         struct r1bio *r1_bio;
1087         struct bio *read_bio;
1088         int i, disks;
1089         struct bitmap *bitmap;
1090         unsigned long flags;
1091         const int rw = bio_data_dir(bio);
1092         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1093         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1094         const unsigned long do_discard = (bio->bi_rw
1095                                           & (REQ_DISCARD | REQ_SECURE));
1096         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1097         struct md_rdev *blocked_rdev;
1098         struct blk_plug_cb *cb;
1099         struct raid1_plug_cb *plug = NULL;
1100         int first_clone;
1101         int sectors_handled;
1102         int max_sectors;
1103         sector_t start_next_window;
1104
1105         /*
1106          * Register the new request and wait if the reconstruction
1107          * thread has put up a bar for new requests.
1108          * Continue immediately if no resync is active currently.
1109          */
1110
1111         md_write_start(mddev, bio); /* wait on superblock update early */
1112
1113         if (bio_data_dir(bio) == WRITE &&
1114             bio_end_sector(bio) > mddev->suspend_lo &&
1115             bio->bi_iter.bi_sector < mddev->suspend_hi) {
1116                 /* As the suspend_* range is controlled by
1117                  * userspace, we want an interruptible
1118                  * wait.
1119                  */
1120                 DEFINE_WAIT(w);
1121                 for (;;) {
1122                         flush_signals(current);
1123                         prepare_to_wait(&conf->wait_barrier,
1124                                         &w, TASK_INTERRUPTIBLE);
1125                         if (bio_end_sector(bio) <= mddev->suspend_lo ||
1126                             bio->bi_iter.bi_sector >= mddev->suspend_hi)
1127                                 break;
1128                         schedule();
1129                 }
1130                 finish_wait(&conf->wait_barrier, &w);
1131         }
1132
1133         start_next_window = wait_barrier(conf, bio);
1134
1135         bitmap = mddev->bitmap;
1136
1137         /*
1138          * make_request() can abort the operation when READA is being
1139          * used and no empty request is available.
1140          *
1141          */
1142         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1143
1144         r1_bio->master_bio = bio;
1145         r1_bio->sectors = bio_sectors(bio);
1146         r1_bio->state = 0;
1147         r1_bio->mddev = mddev;
1148         r1_bio->sector = bio->bi_iter.bi_sector;
1149
1150         /* We might need to issue multiple reads to different
1151          * devices if there are bad blocks around, so we keep
1152          * track of the number of reads in bio->bi_phys_segments.
1153          * If this is 0, there is only one r1_bio and no locking
1154          * will be needed when requests complete.  If it is
1155          * non-zero, then it is the number of not-completed requests.
1156          */
1157         bio->bi_phys_segments = 0;
1158         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1159
1160         if (rw == READ) {
1161                 /*
1162                  * read balancing logic:
1163                  */
1164                 int rdisk;
1165
1166 read_again:
1167                 rdisk = read_balance(conf, r1_bio, &max_sectors);
1168
1169                 if (rdisk < 0) {
1170                         /* couldn't find anywhere to read from */
1171                         raid_end_bio_io(r1_bio);
1172                         return;
1173                 }
1174                 mirror = conf->mirrors + rdisk;
1175
1176                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1177                     bitmap) {
1178                         /* Reading from a write-mostly device must
1179                          * take care not to over-take any writes
1180                          * that are 'behind'
1181                          */
1182                         wait_event(bitmap->behind_wait,
1183                                    atomic_read(&bitmap->behind_writes) == 0);
1184                 }
1185                 r1_bio->read_disk = rdisk;
1186                 r1_bio->start_next_window = 0;
1187
1188                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1189                 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1190                          max_sectors);
1191
1192                 r1_bio->bios[rdisk] = read_bio;
1193
1194                 read_bio->bi_iter.bi_sector = r1_bio->sector +
1195                         mirror->rdev->data_offset;
1196                 read_bio->bi_bdev = mirror->rdev->bdev;
1197                 read_bio->bi_end_io = raid1_end_read_request;
1198                 read_bio->bi_rw = READ | do_sync;
1199                 read_bio->bi_private = r1_bio;
1200
1201                 if (max_sectors < r1_bio->sectors) {
1202                         /* could not read all from this device, so we will
1203                          * need another r1_bio.
1204                          */
1205
1206                         sectors_handled = (r1_bio->sector + max_sectors
1207                                            - bio->bi_iter.bi_sector);
1208                         r1_bio->sectors = max_sectors;
1209                         spin_lock_irq(&conf->device_lock);
1210                         if (bio->bi_phys_segments == 0)
1211                                 bio->bi_phys_segments = 2;
1212                         else
1213                                 bio->bi_phys_segments++;
1214                         spin_unlock_irq(&conf->device_lock);
1215                         /* Cannot call generic_make_request directly
1216                          * as that will be queued in __make_request
1217                          * and subsequent mempool_alloc might block waiting
1218                          * for it.  So hand bio over to raid1d.
1219                          */
1220                         reschedule_retry(r1_bio);
1221
1222                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1223
1224                         r1_bio->master_bio = bio;
1225                         r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1226                         r1_bio->state = 0;
1227                         r1_bio->mddev = mddev;
1228                         r1_bio->sector = bio->bi_iter.bi_sector +
1229                                 sectors_handled;
1230                         goto read_again;
1231                 } else
1232                         generic_make_request(read_bio);
1233                 return;
1234         }
1235
1236         /*
1237          * WRITE:
1238          */
1239         if (conf->pending_count >= max_queued_requests) {
1240                 md_wakeup_thread(mddev->thread);
1241                 wait_event(conf->wait_barrier,
1242                            conf->pending_count < max_queued_requests);
1243         }
1244         /* first select target devices under rcu_lock and
1245          * inc refcount on their rdev.  Record them by setting
1246          * bios[x] to bio
1247          * If there are known/acknowledged bad blocks on any device on
1248          * which we have seen a write error, we want to avoid writing those
1249          * blocks.
1250          * This potentially requires several writes to write around
1251          * the bad blocks.  Each set of writes gets it's own r1bio
1252          * with a set of bios attached.
1253          */
1254
1255         disks = conf->raid_disks * 2;
1256  retry_write:
1257         r1_bio->start_next_window = start_next_window;
1258         blocked_rdev = NULL;
1259         rcu_read_lock();
1260         max_sectors = r1_bio->sectors;
1261         for (i = 0;  i < disks; i++) {
1262                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1263                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1264                         atomic_inc(&rdev->nr_pending);
1265                         blocked_rdev = rdev;
1266                         break;
1267                 }
1268                 r1_bio->bios[i] = NULL;
1269                 if (!rdev || test_bit(Faulty, &rdev->flags)
1270                     || test_bit(Unmerged, &rdev->flags)) {
1271                         if (i < conf->raid_disks)
1272                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1273                         continue;
1274                 }
1275
1276                 atomic_inc(&rdev->nr_pending);
1277                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1278                         sector_t first_bad;
1279                         int bad_sectors;
1280                         int is_bad;
1281
1282                         is_bad = is_badblock(rdev, r1_bio->sector,
1283                                              max_sectors,
1284                                              &first_bad, &bad_sectors);
1285                         if (is_bad < 0) {
1286                                 /* mustn't write here until the bad block is
1287                                  * acknowledged*/
1288                                 set_bit(BlockedBadBlocks, &rdev->flags);
1289                                 blocked_rdev = rdev;
1290                                 break;
1291                         }
1292                         if (is_bad && first_bad <= r1_bio->sector) {
1293                                 /* Cannot write here at all */
1294                                 bad_sectors -= (r1_bio->sector - first_bad);
1295                                 if (bad_sectors < max_sectors)
1296                                         /* mustn't write more than bad_sectors
1297                                          * to other devices yet
1298                                          */
1299                                         max_sectors = bad_sectors;
1300                                 rdev_dec_pending(rdev, mddev);
1301                                 /* We don't set R1BIO_Degraded as that
1302                                  * only applies if the disk is
1303                                  * missing, so it might be re-added,
1304                                  * and we want to know to recover this
1305                                  * chunk.
1306                                  * In this case the device is here,
1307                                  * and the fact that this chunk is not
1308                                  * in-sync is recorded in the bad
1309                                  * block log
1310                                  */
1311                                 continue;
1312                         }
1313                         if (is_bad) {
1314                                 int good_sectors = first_bad - r1_bio->sector;
1315                                 if (good_sectors < max_sectors)
1316                                         max_sectors = good_sectors;
1317                         }
1318                 }
1319                 r1_bio->bios[i] = bio;
1320         }
1321         rcu_read_unlock();
1322
1323         if (unlikely(blocked_rdev)) {
1324                 /* Wait for this device to become unblocked */
1325                 int j;
1326                 sector_t old = start_next_window;
1327
1328                 for (j = 0; j < i; j++)
1329                         if (r1_bio->bios[j])
1330                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1331                 r1_bio->state = 0;
1332                 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector);
1333                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1334                 start_next_window = wait_barrier(conf, bio);
1335                 /*
1336                  * We must make sure the multi r1bios of bio have
1337                  * the same value of bi_phys_segments
1338                  */
1339                 if (bio->bi_phys_segments && old &&
1340                     old != start_next_window)
1341                         /* Wait for the former r1bio(s) to complete */
1342                         wait_event(conf->wait_barrier,
1343                                    bio->bi_phys_segments == 1);
1344                 goto retry_write;
1345         }
1346
1347         if (max_sectors < r1_bio->sectors) {
1348                 /* We are splitting this write into multiple parts, so
1349                  * we need to prepare for allocating another r1_bio.
1350                  */
1351                 r1_bio->sectors = max_sectors;
1352                 spin_lock_irq(&conf->device_lock);
1353                 if (bio->bi_phys_segments == 0)
1354                         bio->bi_phys_segments = 2;
1355                 else
1356                         bio->bi_phys_segments++;
1357                 spin_unlock_irq(&conf->device_lock);
1358         }
1359         sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1360
1361         atomic_set(&r1_bio->remaining, 1);
1362         atomic_set(&r1_bio->behind_remaining, 0);
1363
1364         first_clone = 1;
1365         for (i = 0; i < disks; i++) {
1366                 struct bio *mbio;
1367                 if (!r1_bio->bios[i])
1368                         continue;
1369
1370                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1371                 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors);
1372
1373                 if (first_clone) {
1374                         /* do behind I/O ?
1375                          * Not if there are too many, or cannot
1376                          * allocate memory, or a reader on WriteMostly
1377                          * is waiting for behind writes to flush */
1378                         if (bitmap &&
1379                             (atomic_read(&bitmap->behind_writes)
1380                              < mddev->bitmap_info.max_write_behind) &&
1381                             !waitqueue_active(&bitmap->behind_wait))
1382                                 alloc_behind_pages(mbio, r1_bio);
1383
1384                         bitmap_startwrite(bitmap, r1_bio->sector,
1385                                           r1_bio->sectors,
1386                                           test_bit(R1BIO_BehindIO,
1387                                                    &r1_bio->state));
1388                         first_clone = 0;
1389                 }
1390                 if (r1_bio->behind_bvecs) {
1391                         struct bio_vec *bvec;
1392                         int j;
1393
1394                         /*
1395                          * We trimmed the bio, so _all is legit
1396                          */
1397                         bio_for_each_segment_all(bvec, mbio, j)
1398                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1399                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1400                                 atomic_inc(&r1_bio->behind_remaining);
1401                 }
1402
1403                 r1_bio->bios[i] = mbio;
1404
1405                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1406                                    conf->mirrors[i].rdev->data_offset);
1407                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1408                 mbio->bi_end_io = raid1_end_write_request;
1409                 mbio->bi_rw =
1410                         WRITE | do_flush_fua | do_sync | do_discard | do_same;
1411                 mbio->bi_private = r1_bio;
1412
1413                 atomic_inc(&r1_bio->remaining);
1414
1415                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1416                 if (cb)
1417                         plug = container_of(cb, struct raid1_plug_cb, cb);
1418                 else
1419                         plug = NULL;
1420                 spin_lock_irqsave(&conf->device_lock, flags);
1421                 if (plug) {
1422                         bio_list_add(&plug->pending, mbio);
1423                         plug->pending_cnt++;
1424                 } else {
1425                         bio_list_add(&conf->pending_bio_list, mbio);
1426                         conf->pending_count++;
1427                 }
1428                 spin_unlock_irqrestore(&conf->device_lock, flags);
1429                 if (!plug)
1430                         md_wakeup_thread(mddev->thread);
1431         }
1432         /* Mustn't call r1_bio_write_done before this next test,
1433          * as it could result in the bio being freed.
1434          */
1435         if (sectors_handled < bio_sectors(bio)) {
1436                 r1_bio_write_done(r1_bio);
1437                 /* We need another r1_bio.  It has already been counted
1438                  * in bio->bi_phys_segments
1439                  */
1440                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1441                 r1_bio->master_bio = bio;
1442                 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1443                 r1_bio->state = 0;
1444                 r1_bio->mddev = mddev;
1445                 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1446                 goto retry_write;
1447         }
1448
1449         r1_bio_write_done(r1_bio);
1450
1451         /* In case raid1d snuck in to freeze_array */
1452         wake_up(&conf->wait_barrier);
1453 }
1454
1455 static void status(struct seq_file *seq, struct mddev *mddev)
1456 {
1457         struct r1conf *conf = mddev->private;
1458         int i;
1459
1460         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1461                    conf->raid_disks - mddev->degraded);
1462         rcu_read_lock();
1463         for (i = 0; i < conf->raid_disks; i++) {
1464                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1465                 seq_printf(seq, "%s",
1466                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1467         }
1468         rcu_read_unlock();
1469         seq_printf(seq, "]");
1470 }
1471
1472 static void error(struct mddev *mddev, struct md_rdev *rdev)
1473 {
1474         char b[BDEVNAME_SIZE];
1475         struct r1conf *conf = mddev->private;
1476
1477         /*
1478          * If it is not operational, then we have already marked it as dead
1479          * else if it is the last working disks, ignore the error, let the
1480          * next level up know.
1481          * else mark the drive as failed
1482          */
1483         if (test_bit(In_sync, &rdev->flags)
1484             && (conf->raid_disks - mddev->degraded) == 1) {
1485                 /*
1486                  * Don't fail the drive, act as though we were just a
1487                  * normal single drive.
1488                  * However don't try a recovery from this drive as
1489                  * it is very likely to fail.
1490                  */
1491                 conf->recovery_disabled = mddev->recovery_disabled;
1492                 return;
1493         }
1494         set_bit(Blocked, &rdev->flags);
1495         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1496                 unsigned long flags;
1497                 spin_lock_irqsave(&conf->device_lock, flags);
1498                 mddev->degraded++;
1499                 set_bit(Faulty, &rdev->flags);
1500                 spin_unlock_irqrestore(&conf->device_lock, flags);
1501         } else
1502                 set_bit(Faulty, &rdev->flags);
1503         /*
1504          * if recovery is running, make sure it aborts.
1505          */
1506         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1507         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1508         printk(KERN_ALERT
1509                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1510                "md/raid1:%s: Operation continuing on %d devices.\n",
1511                mdname(mddev), bdevname(rdev->bdev, b),
1512                mdname(mddev), conf->raid_disks - mddev->degraded);
1513 }
1514
1515 static void print_conf(struct r1conf *conf)
1516 {
1517         int i;
1518
1519         printk(KERN_DEBUG "RAID1 conf printout:\n");
1520         if (!conf) {
1521                 printk(KERN_DEBUG "(!conf)\n");
1522                 return;
1523         }
1524         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1525                 conf->raid_disks);
1526
1527         rcu_read_lock();
1528         for (i = 0; i < conf->raid_disks; i++) {
1529                 char b[BDEVNAME_SIZE];
1530                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1531                 if (rdev)
1532                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1533                                i, !test_bit(In_sync, &rdev->flags),
1534                                !test_bit(Faulty, &rdev->flags),
1535                                bdevname(rdev->bdev,b));
1536         }
1537         rcu_read_unlock();
1538 }
1539
1540 static void close_sync(struct r1conf *conf)
1541 {
1542         wait_barrier(conf, NULL);
1543         allow_barrier(conf, 0, 0);
1544
1545         mempool_destroy(conf->r1buf_pool);
1546         conf->r1buf_pool = NULL;
1547
1548         spin_lock_irq(&conf->resync_lock);
1549         conf->next_resync = 0;
1550         conf->start_next_window = MaxSector;
1551         conf->current_window_requests +=
1552                 conf->next_window_requests;
1553         conf->next_window_requests = 0;
1554         spin_unlock_irq(&conf->resync_lock);
1555 }
1556
1557 static int raid1_spare_active(struct mddev *mddev)
1558 {
1559         int i;
1560         struct r1conf *conf = mddev->private;
1561         int count = 0;
1562         unsigned long flags;
1563
1564         /*
1565          * Find all failed disks within the RAID1 configuration
1566          * and mark them readable.
1567          * Called under mddev lock, so rcu protection not needed.
1568          */
1569         for (i = 0; i < conf->raid_disks; i++) {
1570                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1571                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1572                 if (repl
1573                     && repl->recovery_offset == MaxSector
1574                     && !test_bit(Faulty, &repl->flags)
1575                     && !test_and_set_bit(In_sync, &repl->flags)) {
1576                         /* replacement has just become active */
1577                         if (!rdev ||
1578                             !test_and_clear_bit(In_sync, &rdev->flags))
1579                                 count++;
1580                         if (rdev) {
1581                                 /* Replaced device not technically
1582                                  * faulty, but we need to be sure
1583                                  * it gets removed and never re-added
1584                                  */
1585                                 set_bit(Faulty, &rdev->flags);
1586                                 sysfs_notify_dirent_safe(
1587                                         rdev->sysfs_state);
1588                         }
1589                 }
1590                 if (rdev
1591                     && rdev->recovery_offset == MaxSector
1592                     && !test_bit(Faulty, &rdev->flags)
1593                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1594                         count++;
1595                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1596                 }
1597         }
1598         spin_lock_irqsave(&conf->device_lock, flags);
1599         mddev->degraded -= count;
1600         spin_unlock_irqrestore(&conf->device_lock, flags);
1601
1602         print_conf(conf);
1603         return count;
1604 }
1605
1606 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1607 {
1608         struct r1conf *conf = mddev->private;
1609         int err = -EEXIST;
1610         int mirror = 0;
1611         struct raid1_info *p;
1612         int first = 0;
1613         int last = conf->raid_disks - 1;
1614         struct request_queue *q = bdev_get_queue(rdev->bdev);
1615
1616         if (mddev->recovery_disabled == conf->recovery_disabled)
1617                 return -EBUSY;
1618
1619         if (rdev->raid_disk >= 0)
1620                 first = last = rdev->raid_disk;
1621
1622         if (q->merge_bvec_fn) {
1623                 set_bit(Unmerged, &rdev->flags);
1624                 mddev->merge_check_needed = 1;
1625         }
1626
1627         for (mirror = first; mirror <= last; mirror++) {
1628                 p = conf->mirrors+mirror;
1629                 if (!p->rdev) {
1630
1631                         if (mddev->gendisk)
1632                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1633                                                   rdev->data_offset << 9);
1634
1635                         p->head_position = 0;
1636                         rdev->raid_disk = mirror;
1637                         err = 0;
1638                         /* As all devices are equivalent, we don't need a full recovery
1639                          * if this was recently any drive of the array
1640                          */
1641                         if (rdev->saved_raid_disk < 0)
1642                                 conf->fullsync = 1;
1643                         rcu_assign_pointer(p->rdev, rdev);
1644                         break;
1645                 }
1646                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1647                     p[conf->raid_disks].rdev == NULL) {
1648                         /* Add this device as a replacement */
1649                         clear_bit(In_sync, &rdev->flags);
1650                         set_bit(Replacement, &rdev->flags);
1651                         rdev->raid_disk = mirror;
1652                         err = 0;
1653                         conf->fullsync = 1;
1654                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1655                         break;
1656                 }
1657         }
1658         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1659                 /* Some requests might not have seen this new
1660                  * merge_bvec_fn.  We must wait for them to complete
1661                  * before merging the device fully.
1662                  * First we make sure any code which has tested
1663                  * our function has submitted the request, then
1664                  * we wait for all outstanding requests to complete.
1665                  */
1666                 synchronize_sched();
1667                 freeze_array(conf, 0);
1668                 unfreeze_array(conf);
1669                 clear_bit(Unmerged, &rdev->flags);
1670         }
1671         md_integrity_add_rdev(rdev, mddev);
1672         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1673                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1674         print_conf(conf);
1675         return err;
1676 }
1677
1678 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1679 {
1680         struct r1conf *conf = mddev->private;
1681         int err = 0;
1682         int number = rdev->raid_disk;
1683         struct raid1_info *p = conf->mirrors + number;
1684
1685         if (rdev != p->rdev)
1686                 p = conf->mirrors + conf->raid_disks + number;
1687
1688         print_conf(conf);
1689         if (rdev == p->rdev) {
1690                 if (test_bit(In_sync, &rdev->flags) ||
1691                     atomic_read(&rdev->nr_pending)) {
1692                         err = -EBUSY;
1693                         goto abort;
1694                 }
1695                 /* Only remove non-faulty devices if recovery
1696                  * is not possible.
1697                  */
1698                 if (!test_bit(Faulty, &rdev->flags) &&
1699                     mddev->recovery_disabled != conf->recovery_disabled &&
1700                     mddev->degraded < conf->raid_disks) {
1701                         err = -EBUSY;
1702                         goto abort;
1703                 }
1704                 p->rdev = NULL;
1705                 synchronize_rcu();
1706                 if (atomic_read(&rdev->nr_pending)) {
1707                         /* lost the race, try later */
1708                         err = -EBUSY;
1709                         p->rdev = rdev;
1710                         goto abort;
1711                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1712                         /* We just removed a device that is being replaced.
1713                          * Move down the replacement.  We drain all IO before
1714                          * doing this to avoid confusion.
1715                          */
1716                         struct md_rdev *repl =
1717                                 conf->mirrors[conf->raid_disks + number].rdev;
1718                         freeze_array(conf, 0);
1719                         clear_bit(Replacement, &repl->flags);
1720                         p->rdev = repl;
1721                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1722                         unfreeze_array(conf);
1723                         clear_bit(WantReplacement, &rdev->flags);
1724                 } else
1725                         clear_bit(WantReplacement, &rdev->flags);
1726                 err = md_integrity_register(mddev);
1727         }
1728 abort:
1729
1730         print_conf(conf);
1731         return err;
1732 }
1733
1734 static void end_sync_read(struct bio *bio, int error)
1735 {
1736         struct r1bio *r1_bio = bio->bi_private;
1737
1738         update_head_pos(r1_bio->read_disk, r1_bio);
1739
1740         /*
1741          * we have read a block, now it needs to be re-written,
1742          * or re-read if the read failed.
1743          * We don't do much here, just schedule handling by raid1d
1744          */
1745         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1746                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1747
1748         if (atomic_dec_and_test(&r1_bio->remaining))
1749                 reschedule_retry(r1_bio);
1750 }
1751
1752 static void end_sync_write(struct bio *bio, int error)
1753 {
1754         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1755         struct r1bio *r1_bio = bio->bi_private;
1756         struct mddev *mddev = r1_bio->mddev;
1757         struct r1conf *conf = mddev->private;
1758         int mirror=0;
1759         sector_t first_bad;
1760         int bad_sectors;
1761
1762         mirror = find_bio_disk(r1_bio, bio);
1763
1764         if (!uptodate) {
1765                 sector_t sync_blocks = 0;
1766                 sector_t s = r1_bio->sector;
1767                 long sectors_to_go = r1_bio->sectors;
1768                 /* make sure these bits doesn't get cleared. */
1769                 do {
1770                         bitmap_end_sync(mddev->bitmap, s,
1771                                         &sync_blocks, 1);
1772                         s += sync_blocks;
1773                         sectors_to_go -= sync_blocks;
1774                 } while (sectors_to_go > 0);
1775                 set_bit(WriteErrorSeen,
1776                         &conf->mirrors[mirror].rdev->flags);
1777                 if (!test_and_set_bit(WantReplacement,
1778                                       &conf->mirrors[mirror].rdev->flags))
1779                         set_bit(MD_RECOVERY_NEEDED, &
1780                                 mddev->recovery);
1781                 set_bit(R1BIO_WriteError, &r1_bio->state);
1782         } else if (is_badblock(conf->mirrors[mirror].rdev,
1783                                r1_bio->sector,
1784                                r1_bio->sectors,
1785                                &first_bad, &bad_sectors) &&
1786                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1787                                 r1_bio->sector,
1788                                 r1_bio->sectors,
1789                                 &first_bad, &bad_sectors)
1790                 )
1791                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1792
1793         if (atomic_dec_and_test(&r1_bio->remaining)) {
1794                 int s = r1_bio->sectors;
1795                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1796                     test_bit(R1BIO_WriteError, &r1_bio->state))
1797                         reschedule_retry(r1_bio);
1798                 else {
1799                         put_buf(r1_bio);
1800                         md_done_sync(mddev, s, uptodate);
1801                 }
1802         }
1803 }
1804
1805 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1806                             int sectors, struct page *page, int rw)
1807 {
1808         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1809                 /* success */
1810                 return 1;
1811         if (rw == WRITE) {
1812                 set_bit(WriteErrorSeen, &rdev->flags);
1813                 if (!test_and_set_bit(WantReplacement,
1814                                       &rdev->flags))
1815                         set_bit(MD_RECOVERY_NEEDED, &
1816                                 rdev->mddev->recovery);
1817         }
1818         /* need to record an error - either for the block or the device */
1819         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1820                 md_error(rdev->mddev, rdev);
1821         return 0;
1822 }
1823
1824 static int fix_sync_read_error(struct r1bio *r1_bio)
1825 {
1826         /* Try some synchronous reads of other devices to get
1827          * good data, much like with normal read errors.  Only
1828          * read into the pages we already have so we don't
1829          * need to re-issue the read request.
1830          * We don't need to freeze the array, because being in an
1831          * active sync request, there is no normal IO, and
1832          * no overlapping syncs.
1833          * We don't need to check is_badblock() again as we
1834          * made sure that anything with a bad block in range
1835          * will have bi_end_io clear.
1836          */
1837         struct mddev *mddev = r1_bio->mddev;
1838         struct r1conf *conf = mddev->private;
1839         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1840         sector_t sect = r1_bio->sector;
1841         int sectors = r1_bio->sectors;
1842         int idx = 0;
1843
1844         while(sectors) {
1845                 int s = sectors;
1846                 int d = r1_bio->read_disk;
1847                 int success = 0;
1848                 struct md_rdev *rdev;
1849                 int start;
1850
1851                 if (s > (PAGE_SIZE>>9))
1852                         s = PAGE_SIZE >> 9;
1853                 do {
1854                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1855                                 /* No rcu protection needed here devices
1856                                  * can only be removed when no resync is
1857                                  * active, and resync is currently active
1858                                  */
1859                                 rdev = conf->mirrors[d].rdev;
1860                                 if (sync_page_io(rdev, sect, s<<9,
1861                                                  bio->bi_io_vec[idx].bv_page,
1862                                                  READ, false)) {
1863                                         success = 1;
1864                                         break;
1865                                 }
1866                         }
1867                         d++;
1868                         if (d == conf->raid_disks * 2)
1869                                 d = 0;
1870                 } while (!success && d != r1_bio->read_disk);
1871
1872                 if (!success) {
1873                         char b[BDEVNAME_SIZE];
1874                         int abort = 0;
1875                         /* Cannot read from anywhere, this block is lost.
1876                          * Record a bad block on each device.  If that doesn't
1877                          * work just disable and interrupt the recovery.
1878                          * Don't fail devices as that won't really help.
1879                          */
1880                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1881                                " for block %llu\n",
1882                                mdname(mddev),
1883                                bdevname(bio->bi_bdev, b),
1884                                (unsigned long long)r1_bio->sector);
1885                         for (d = 0; d < conf->raid_disks * 2; d++) {
1886                                 rdev = conf->mirrors[d].rdev;
1887                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1888                                         continue;
1889                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1890                                         abort = 1;
1891                         }
1892                         if (abort) {
1893                                 conf->recovery_disabled =
1894                                         mddev->recovery_disabled;
1895                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1896                                 md_done_sync(mddev, r1_bio->sectors, 0);
1897                                 put_buf(r1_bio);
1898                                 return 0;
1899                         }
1900                         /* Try next page */
1901                         sectors -= s;
1902                         sect += s;
1903                         idx++;
1904                         continue;
1905                 }
1906
1907                 start = d;
1908                 /* write it back and re-read */
1909                 while (d != r1_bio->read_disk) {
1910                         if (d == 0)
1911                                 d = conf->raid_disks * 2;
1912                         d--;
1913                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1914                                 continue;
1915                         rdev = conf->mirrors[d].rdev;
1916                         if (r1_sync_page_io(rdev, sect, s,
1917                                             bio->bi_io_vec[idx].bv_page,
1918                                             WRITE) == 0) {
1919                                 r1_bio->bios[d]->bi_end_io = NULL;
1920                                 rdev_dec_pending(rdev, mddev);
1921                         }
1922                 }
1923                 d = start;
1924                 while (d != r1_bio->read_disk) {
1925                         if (d == 0)
1926                                 d = conf->raid_disks * 2;
1927                         d--;
1928                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1929                                 continue;
1930                         rdev = conf->mirrors[d].rdev;
1931                         if (r1_sync_page_io(rdev, sect, s,
1932                                             bio->bi_io_vec[idx].bv_page,
1933                                             READ) != 0)
1934                                 atomic_add(s, &rdev->corrected_errors);
1935                 }
1936                 sectors -= s;
1937                 sect += s;
1938                 idx ++;
1939         }
1940         set_bit(R1BIO_Uptodate, &r1_bio->state);
1941         set_bit(BIO_UPTODATE, &bio->bi_flags);
1942         return 1;
1943 }
1944
1945 static void process_checks(struct r1bio *r1_bio)
1946 {
1947         /* We have read all readable devices.  If we haven't
1948          * got the block, then there is no hope left.
1949          * If we have, then we want to do a comparison
1950          * and skip the write if everything is the same.
1951          * If any blocks failed to read, then we need to
1952          * attempt an over-write
1953          */
1954         struct mddev *mddev = r1_bio->mddev;
1955         struct r1conf *conf = mddev->private;
1956         int primary;
1957         int i;
1958         int vcnt;
1959
1960         /* Fix variable parts of all bios */
1961         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1962         for (i = 0; i < conf->raid_disks * 2; i++) {
1963                 int j;
1964                 int size;
1965                 int uptodate;
1966                 struct bio *b = r1_bio->bios[i];
1967                 if (b->bi_end_io != end_sync_read)
1968                         continue;
1969                 /* fixup the bio for reuse, but preserve BIO_UPTODATE */
1970                 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags);
1971                 bio_reset(b);
1972                 if (!uptodate)
1973                         clear_bit(BIO_UPTODATE, &b->bi_flags);
1974                 b->bi_vcnt = vcnt;
1975                 b->bi_iter.bi_size = r1_bio->sectors << 9;
1976                 b->bi_iter.bi_sector = r1_bio->sector +
1977                         conf->mirrors[i].rdev->data_offset;
1978                 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1979                 b->bi_end_io = end_sync_read;
1980                 b->bi_private = r1_bio;
1981
1982                 size = b->bi_iter.bi_size;
1983                 for (j = 0; j < vcnt ; j++) {
1984                         struct bio_vec *bi;
1985                         bi = &b->bi_io_vec[j];
1986                         bi->bv_offset = 0;
1987                         if (size > PAGE_SIZE)
1988                                 bi->bv_len = PAGE_SIZE;
1989                         else
1990                                 bi->bv_len = size;
1991                         size -= PAGE_SIZE;
1992                 }
1993         }
1994         for (primary = 0; primary < conf->raid_disks * 2; primary++)
1995                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1996                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1997                         r1_bio->bios[primary]->bi_end_io = NULL;
1998                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1999                         break;
2000                 }
2001         r1_bio->read_disk = primary;
2002         for (i = 0; i < conf->raid_disks * 2; i++) {
2003                 int j;
2004                 struct bio *pbio = r1_bio->bios[primary];
2005                 struct bio *sbio = r1_bio->bios[i];
2006                 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags);
2007
2008                 if (sbio->bi_end_io != end_sync_read)
2009                         continue;
2010                 /* Now we can 'fixup' the BIO_UPTODATE flag */
2011                 set_bit(BIO_UPTODATE, &sbio->bi_flags);
2012
2013                 if (uptodate) {
2014                         for (j = vcnt; j-- ; ) {
2015                                 struct page *p, *s;
2016                                 p = pbio->bi_io_vec[j].bv_page;
2017                                 s = sbio->bi_io_vec[j].bv_page;
2018                                 if (memcmp(page_address(p),
2019                                            page_address(s),
2020                                            sbio->bi_io_vec[j].bv_len))
2021                                         break;
2022                         }
2023                 } else
2024                         j = 0;
2025                 if (j >= 0)
2026                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2027                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2028                               && uptodate)) {
2029                         /* No need to write to this device. */
2030                         sbio->bi_end_io = NULL;
2031                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2032                         continue;
2033                 }
2034
2035                 bio_copy_data(sbio, pbio);
2036         }
2037 }
2038
2039 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2040 {
2041         struct r1conf *conf = mddev->private;
2042         int i;
2043         int disks = conf->raid_disks * 2;
2044         struct bio *bio, *wbio;
2045
2046         bio = r1_bio->bios[r1_bio->read_disk];
2047
2048         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2049                 /* ouch - failed to read all of that. */
2050                 if (!fix_sync_read_error(r1_bio))
2051                         return;
2052
2053         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2054                 process_checks(r1_bio);
2055
2056         /*
2057          * schedule writes
2058          */
2059         atomic_set(&r1_bio->remaining, 1);
2060         for (i = 0; i < disks ; i++) {
2061                 wbio = r1_bio->bios[i];
2062                 if (wbio->bi_end_io == NULL ||
2063                     (wbio->bi_end_io == end_sync_read &&
2064                      (i == r1_bio->read_disk ||
2065                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2066                         continue;
2067
2068                 wbio->bi_rw = WRITE;
2069                 wbio->bi_end_io = end_sync_write;
2070                 atomic_inc(&r1_bio->remaining);
2071                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2072
2073                 generic_make_request(wbio);
2074         }
2075
2076         if (atomic_dec_and_test(&r1_bio->remaining)) {
2077                 /* if we're here, all write(s) have completed, so clean up */
2078                 int s = r1_bio->sectors;
2079                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2080                     test_bit(R1BIO_WriteError, &r1_bio->state))
2081                         reschedule_retry(r1_bio);
2082                 else {
2083                         put_buf(r1_bio);
2084                         md_done_sync(mddev, s, 1);
2085                 }
2086         }
2087 }
2088
2089 /*
2090  * This is a kernel thread which:
2091  *
2092  *      1.      Retries failed read operations on working mirrors.
2093  *      2.      Updates the raid superblock when problems encounter.
2094  *      3.      Performs writes following reads for array synchronising.
2095  */
2096
2097 static void fix_read_error(struct r1conf *conf, int read_disk,
2098                            sector_t sect, int sectors)
2099 {
2100         struct mddev *mddev = conf->mddev;
2101         while(sectors) {
2102                 int s = sectors;
2103                 int d = read_disk;
2104                 int success = 0;
2105                 int start;
2106                 struct md_rdev *rdev;
2107
2108                 if (s > (PAGE_SIZE>>9))
2109                         s = PAGE_SIZE >> 9;
2110
2111                 do {
2112                         /* Note: no rcu protection needed here
2113                          * as this is synchronous in the raid1d thread
2114                          * which is the thread that might remove
2115                          * a device.  If raid1d ever becomes multi-threaded....
2116                          */
2117                         sector_t first_bad;
2118                         int bad_sectors;
2119
2120                         rdev = conf->mirrors[d].rdev;
2121                         if (rdev &&
2122                             (test_bit(In_sync, &rdev->flags) ||
2123                              (!test_bit(Faulty, &rdev->flags) &&
2124                               rdev->recovery_offset >= sect + s)) &&
2125                             is_badblock(rdev, sect, s,
2126                                         &first_bad, &bad_sectors) == 0 &&
2127                             sync_page_io(rdev, sect, s<<9,
2128                                          conf->tmppage, READ, false))
2129                                 success = 1;
2130                         else {
2131                                 d++;
2132                                 if (d == conf->raid_disks * 2)
2133                                         d = 0;
2134                         }
2135                 } while (!success && d != read_disk);
2136
2137                 if (!success) {
2138                         /* Cannot read from anywhere - mark it bad */
2139                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2140                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2141                                 md_error(mddev, rdev);
2142                         break;
2143                 }
2144                 /* write it back and re-read */
2145                 start = d;
2146                 while (d != read_disk) {
2147                         if (d==0)
2148                                 d = conf->raid_disks * 2;
2149                         d--;
2150                         rdev = conf->mirrors[d].rdev;
2151                         if (rdev &&
2152                             !test_bit(Faulty, &rdev->flags))
2153                                 r1_sync_page_io(rdev, sect, s,
2154                                                 conf->tmppage, WRITE);
2155                 }
2156                 d = start;
2157                 while (d != read_disk) {
2158                         char b[BDEVNAME_SIZE];
2159                         if (d==0)
2160                                 d = conf->raid_disks * 2;
2161                         d--;
2162                         rdev = conf->mirrors[d].rdev;
2163                         if (rdev &&
2164                             !test_bit(Faulty, &rdev->flags)) {
2165                                 if (r1_sync_page_io(rdev, sect, s,
2166                                                     conf->tmppage, READ)) {
2167                                         atomic_add(s, &rdev->corrected_errors);
2168                                         printk(KERN_INFO
2169                                                "md/raid1:%s: read error corrected "
2170                                                "(%d sectors at %llu on %s)\n",
2171                                                mdname(mddev), s,
2172                                                (unsigned long long)(sect +
2173                                                    rdev->data_offset),
2174                                                bdevname(rdev->bdev, b));
2175                                 }
2176                         }
2177                 }
2178                 sectors -= s;
2179                 sect += s;
2180         }
2181 }
2182
2183 static int narrow_write_error(struct r1bio *r1_bio, int i)
2184 {
2185         struct mddev *mddev = r1_bio->mddev;
2186         struct r1conf *conf = mddev->private;
2187         struct md_rdev *rdev = conf->mirrors[i].rdev;
2188
2189         /* bio has the data to be written to device 'i' where
2190          * we just recently had a write error.
2191          * We repeatedly clone the bio and trim down to one block,
2192          * then try the write.  Where the write fails we record
2193          * a bad block.
2194          * It is conceivable that the bio doesn't exactly align with
2195          * blocks.  We must handle this somehow.
2196          *
2197          * We currently own a reference on the rdev.
2198          */
2199
2200         int block_sectors;
2201         sector_t sector;
2202         int sectors;
2203         int sect_to_write = r1_bio->sectors;
2204         int ok = 1;
2205
2206         if (rdev->badblocks.shift < 0)
2207                 return 0;
2208
2209         block_sectors = 1 << rdev->badblocks.shift;
2210         sector = r1_bio->sector;
2211         sectors = ((sector + block_sectors)
2212                    & ~(sector_t)(block_sectors - 1))
2213                 - sector;
2214
2215         while (sect_to_write) {
2216                 struct bio *wbio;
2217                 if (sectors > sect_to_write)
2218                         sectors = sect_to_write;
2219                 /* Write at 'sector' for 'sectors'*/
2220
2221                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2222                         unsigned vcnt = r1_bio->behind_page_count;
2223                         struct bio_vec *vec = r1_bio->behind_bvecs;
2224
2225                         while (!vec->bv_page) {
2226                                 vec++;
2227                                 vcnt--;
2228                         }
2229
2230                         wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2231                         memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2232
2233                         wbio->bi_vcnt = vcnt;
2234                 } else {
2235                         wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2236                 }
2237
2238                 wbio->bi_rw = WRITE;
2239                 wbio->bi_iter.bi_sector = r1_bio->sector;
2240                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2241
2242                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2243                 wbio->bi_iter.bi_sector += rdev->data_offset;
2244                 wbio->bi_bdev = rdev->bdev;
2245                 if (submit_bio_wait(WRITE, wbio) == 0)
2246                         /* failure! */
2247                         ok = rdev_set_badblocks(rdev, sector,
2248                                                 sectors, 0)
2249                                 && ok;
2250
2251                 bio_put(wbio);
2252                 sect_to_write -= sectors;
2253                 sector += sectors;
2254                 sectors = block_sectors;
2255         }
2256         return ok;
2257 }
2258
2259 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2260 {
2261         int m;
2262         int s = r1_bio->sectors;
2263         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2264                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2265                 struct bio *bio = r1_bio->bios[m];
2266                 if (bio->bi_end_io == NULL)
2267                         continue;
2268                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2269                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2270                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2271                 }
2272                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2273                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2274                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2275                                 md_error(conf->mddev, rdev);
2276                 }
2277         }
2278         put_buf(r1_bio);
2279         md_done_sync(conf->mddev, s, 1);
2280 }
2281
2282 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2283 {
2284         int m;
2285         for (m = 0; m < conf->raid_disks * 2 ; m++)
2286                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2287                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2288                         rdev_clear_badblocks(rdev,
2289                                              r1_bio->sector,
2290                                              r1_bio->sectors, 0);
2291                         rdev_dec_pending(rdev, conf->mddev);
2292                 } else if (r1_bio->bios[m] != NULL) {
2293                         /* This drive got a write error.  We need to
2294                          * narrow down and record precise write
2295                          * errors.
2296                          */
2297                         if (!narrow_write_error(r1_bio, m)) {
2298                                 md_error(conf->mddev,
2299                                          conf->mirrors[m].rdev);
2300                                 /* an I/O failed, we can't clear the bitmap */
2301                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2302                         }
2303                         rdev_dec_pending(conf->mirrors[m].rdev,
2304                                          conf->mddev);
2305                 }
2306         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2307                 close_write(r1_bio);
2308         raid_end_bio_io(r1_bio);
2309 }
2310
2311 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2312 {
2313         int disk;
2314         int max_sectors;
2315         struct mddev *mddev = conf->mddev;
2316         struct bio *bio;
2317         char b[BDEVNAME_SIZE];
2318         struct md_rdev *rdev;
2319
2320         clear_bit(R1BIO_ReadError, &r1_bio->state);
2321         /* we got a read error. Maybe the drive is bad.  Maybe just
2322          * the block and we can fix it.
2323          * We freeze all other IO, and try reading the block from
2324          * other devices.  When we find one, we re-write
2325          * and check it that fixes the read error.
2326          * This is all done synchronously while the array is
2327          * frozen
2328          */
2329         if (mddev->ro == 0) {
2330                 freeze_array(conf, 1);
2331                 fix_read_error(conf, r1_bio->read_disk,
2332                                r1_bio->sector, r1_bio->sectors);
2333                 unfreeze_array(conf);
2334         } else
2335                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2336         rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2337
2338         bio = r1_bio->bios[r1_bio->read_disk];
2339         bdevname(bio->bi_bdev, b);
2340 read_more:
2341         disk = read_balance(conf, r1_bio, &max_sectors);
2342         if (disk == -1) {
2343                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2344                        " read error for block %llu\n",
2345                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2346                 raid_end_bio_io(r1_bio);
2347         } else {
2348                 const unsigned long do_sync
2349                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2350                 if (bio) {
2351                         r1_bio->bios[r1_bio->read_disk] =
2352                                 mddev->ro ? IO_BLOCKED : NULL;
2353                         bio_put(bio);
2354                 }
2355                 r1_bio->read_disk = disk;
2356                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2357                 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2358                          max_sectors);
2359                 r1_bio->bios[r1_bio->read_disk] = bio;
2360                 rdev = conf->mirrors[disk].rdev;
2361                 printk_ratelimited(KERN_ERR
2362                                    "md/raid1:%s: redirecting sector %llu"
2363                                    " to other mirror: %s\n",
2364                                    mdname(mddev),
2365                                    (unsigned long long)r1_bio->sector,
2366                                    bdevname(rdev->bdev, b));
2367                 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2368                 bio->bi_bdev = rdev->bdev;
2369                 bio->bi_end_io = raid1_end_read_request;
2370                 bio->bi_rw = READ | do_sync;
2371                 bio->bi_private = r1_bio;
2372                 if (max_sectors < r1_bio->sectors) {
2373                         /* Drat - have to split this up more */
2374                         struct bio *mbio = r1_bio->master_bio;
2375                         int sectors_handled = (r1_bio->sector + max_sectors
2376                                                - mbio->bi_iter.bi_sector);
2377                         r1_bio->sectors = max_sectors;
2378                         spin_lock_irq(&conf->device_lock);
2379                         if (mbio->bi_phys_segments == 0)
2380                                 mbio->bi_phys_segments = 2;
2381                         else
2382                                 mbio->bi_phys_segments++;
2383                         spin_unlock_irq(&conf->device_lock);
2384                         generic_make_request(bio);
2385                         bio = NULL;
2386
2387                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2388
2389                         r1_bio->master_bio = mbio;
2390                         r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2391                         r1_bio->state = 0;
2392                         set_bit(R1BIO_ReadError, &r1_bio->state);
2393                         r1_bio->mddev = mddev;
2394                         r1_bio->sector = mbio->bi_iter.bi_sector +
2395                                 sectors_handled;
2396
2397                         goto read_more;
2398                 } else
2399                         generic_make_request(bio);
2400         }
2401 }
2402
2403 static void raid1d(struct md_thread *thread)
2404 {
2405         struct mddev *mddev = thread->mddev;
2406         struct r1bio *r1_bio;
2407         unsigned long flags;
2408         struct r1conf *conf = mddev->private;
2409         struct list_head *head = &conf->retry_list;
2410         struct blk_plug plug;
2411
2412         md_check_recovery(mddev);
2413
2414         blk_start_plug(&plug);
2415         for (;;) {
2416
2417                 flush_pending_writes(conf);
2418
2419                 spin_lock_irqsave(&conf->device_lock, flags);
2420                 if (list_empty(head)) {
2421                         spin_unlock_irqrestore(&conf->device_lock, flags);
2422                         break;
2423                 }
2424                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2425                 list_del(head->prev);
2426                 conf->nr_queued--;
2427                 spin_unlock_irqrestore(&conf->device_lock, flags);
2428
2429                 mddev = r1_bio->mddev;
2430                 conf = mddev->private;
2431                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2432                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2433                             test_bit(R1BIO_WriteError, &r1_bio->state))
2434                                 handle_sync_write_finished(conf, r1_bio);
2435                         else
2436                                 sync_request_write(mddev, r1_bio);
2437                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2438                            test_bit(R1BIO_WriteError, &r1_bio->state))
2439                         handle_write_finished(conf, r1_bio);
2440                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2441                         handle_read_error(conf, r1_bio);
2442                 else
2443                         /* just a partial read to be scheduled from separate
2444                          * context
2445                          */
2446                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2447
2448                 cond_resched();
2449                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2450                         md_check_recovery(mddev);
2451         }
2452         blk_finish_plug(&plug);
2453 }
2454
2455 static int init_resync(struct r1conf *conf)
2456 {
2457         int buffs;
2458
2459         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2460         BUG_ON(conf->r1buf_pool);
2461         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2462                                           conf->poolinfo);
2463         if (!conf->r1buf_pool)
2464                 return -ENOMEM;
2465         conf->next_resync = 0;
2466         return 0;
2467 }
2468
2469 /*
2470  * perform a "sync" on one "block"
2471  *
2472  * We need to make sure that no normal I/O request - particularly write
2473  * requests - conflict with active sync requests.
2474  *
2475  * This is achieved by tracking pending requests and a 'barrier' concept
2476  * that can be installed to exclude normal IO requests.
2477  */
2478
2479 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2480 {
2481         struct r1conf *conf = mddev->private;
2482         struct r1bio *r1_bio;
2483         struct bio *bio;
2484         sector_t max_sector, nr_sectors;
2485         int disk = -1;
2486         int i;
2487         int wonly = -1;
2488         int write_targets = 0, read_targets = 0;
2489         sector_t sync_blocks;
2490         int still_degraded = 0;
2491         int good_sectors = RESYNC_SECTORS;
2492         int min_bad = 0; /* number of sectors that are bad in all devices */
2493
2494         if (!conf->r1buf_pool)
2495                 if (init_resync(conf))
2496                         return 0;
2497
2498         max_sector = mddev->dev_sectors;
2499         if (sector_nr >= max_sector) {
2500                 /* If we aborted, we need to abort the
2501                  * sync on the 'current' bitmap chunk (there will
2502                  * only be one in raid1 resync.
2503                  * We can find the current addess in mddev->curr_resync
2504                  */
2505                 if (mddev->curr_resync < max_sector) /* aborted */
2506                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2507                                                 &sync_blocks, 1);
2508                 else /* completed sync */
2509                         conf->fullsync = 0;
2510
2511                 bitmap_close_sync(mddev->bitmap);
2512                 close_sync(conf);
2513                 return 0;
2514         }
2515
2516         if (mddev->bitmap == NULL &&
2517             mddev->recovery_cp == MaxSector &&
2518             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2519             conf->fullsync == 0) {
2520                 *skipped = 1;
2521                 return max_sector - sector_nr;
2522         }
2523         /* before building a request, check if we can skip these blocks..
2524          * This call the bitmap_start_sync doesn't actually record anything
2525          */
2526         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2527             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2528                 /* We can skip this block, and probably several more */
2529                 *skipped = 1;
2530                 return sync_blocks;
2531         }
2532         /*
2533          * If there is non-resync activity waiting for a turn,
2534          * and resync is going fast enough,
2535          * then let it though before starting on this new sync request.
2536          */
2537         if (!go_faster && conf->nr_waiting)
2538                 msleep_interruptible(1000);
2539
2540         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2541         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2542
2543         raise_barrier(conf, sector_nr);
2544
2545         rcu_read_lock();
2546         /*
2547          * If we get a correctably read error during resync or recovery,
2548          * we might want to read from a different device.  So we
2549          * flag all drives that could conceivably be read from for READ,
2550          * and any others (which will be non-In_sync devices) for WRITE.
2551          * If a read fails, we try reading from something else for which READ
2552          * is OK.
2553          */
2554
2555         r1_bio->mddev = mddev;
2556         r1_bio->sector = sector_nr;
2557         r1_bio->state = 0;
2558         set_bit(R1BIO_IsSync, &r1_bio->state);
2559
2560         for (i = 0; i < conf->raid_disks * 2; i++) {
2561                 struct md_rdev *rdev;
2562                 bio = r1_bio->bios[i];
2563                 bio_reset(bio);
2564
2565                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2566                 if (rdev == NULL ||
2567                     test_bit(Faulty, &rdev->flags)) {
2568                         if (i < conf->raid_disks)
2569                                 still_degraded = 1;
2570                 } else if (!test_bit(In_sync, &rdev->flags)) {
2571                         bio->bi_rw = WRITE;
2572                         bio->bi_end_io = end_sync_write;
2573                         write_targets ++;
2574                 } else {
2575                         /* may need to read from here */
2576                         sector_t first_bad = MaxSector;
2577                         int bad_sectors;
2578
2579                         if (is_badblock(rdev, sector_nr, good_sectors,
2580                                         &first_bad, &bad_sectors)) {
2581                                 if (first_bad > sector_nr)
2582                                         good_sectors = first_bad - sector_nr;
2583                                 else {
2584                                         bad_sectors -= (sector_nr - first_bad);
2585                                         if (min_bad == 0 ||
2586                                             min_bad > bad_sectors)
2587                                                 min_bad = bad_sectors;
2588                                 }
2589                         }
2590                         if (sector_nr < first_bad) {
2591                                 if (test_bit(WriteMostly, &rdev->flags)) {
2592                                         if (wonly < 0)
2593                                                 wonly = i;
2594                                 } else {
2595                                         if (disk < 0)
2596                                                 disk = i;
2597                                 }
2598                                 bio->bi_rw = READ;
2599                                 bio->bi_end_io = end_sync_read;
2600                                 read_targets++;
2601                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2602                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2603                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2604                                 /*
2605                                  * The device is suitable for reading (InSync),
2606                                  * but has bad block(s) here. Let's try to correct them,
2607                                  * if we are doing resync or repair. Otherwise, leave
2608                                  * this device alone for this sync request.
2609                                  */
2610                                 bio->bi_rw = WRITE;
2611                                 bio->bi_end_io = end_sync_write;
2612                                 write_targets++;
2613                         }
2614                 }
2615                 if (bio->bi_end_io) {
2616                         atomic_inc(&rdev->nr_pending);
2617                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2618                         bio->bi_bdev = rdev->bdev;
2619                         bio->bi_private = r1_bio;
2620                 }
2621         }
2622         rcu_read_unlock();
2623         if (disk < 0)
2624                 disk = wonly;
2625         r1_bio->read_disk = disk;
2626
2627         if (read_targets == 0 && min_bad > 0) {
2628                 /* These sectors are bad on all InSync devices, so we
2629                  * need to mark them bad on all write targets
2630                  */
2631                 int ok = 1;
2632                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2633                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2634                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2635                                 ok = rdev_set_badblocks(rdev, sector_nr,
2636                                                         min_bad, 0
2637                                         ) && ok;
2638                         }
2639                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2640                 *skipped = 1;
2641                 put_buf(r1_bio);
2642
2643                 if (!ok) {
2644                         /* Cannot record the badblocks, so need to
2645                          * abort the resync.
2646                          * If there are multiple read targets, could just
2647                          * fail the really bad ones ???
2648                          */
2649                         conf->recovery_disabled = mddev->recovery_disabled;
2650                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2651                         return 0;
2652                 } else
2653                         return min_bad;
2654
2655         }
2656         if (min_bad > 0 && min_bad < good_sectors) {
2657                 /* only resync enough to reach the next bad->good
2658                  * transition */
2659                 good_sectors = min_bad;
2660         }
2661
2662         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2663                 /* extra read targets are also write targets */
2664                 write_targets += read_targets-1;
2665
2666         if (write_targets == 0 || read_targets == 0) {
2667                 /* There is nowhere to write, so all non-sync
2668                  * drives must be failed - so we are finished
2669                  */
2670                 sector_t rv;
2671                 if (min_bad > 0)
2672                         max_sector = sector_nr + min_bad;
2673                 rv = max_sector - sector_nr;
2674                 *skipped = 1;
2675                 put_buf(r1_bio);
2676                 return rv;
2677         }
2678
2679         if (max_sector > mddev->resync_max)
2680                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2681         if (max_sector > sector_nr + good_sectors)
2682                 max_sector = sector_nr + good_sectors;
2683         nr_sectors = 0;
2684         sync_blocks = 0;
2685         do {
2686                 struct page *page;
2687                 int len = PAGE_SIZE;
2688                 if (sector_nr + (len>>9) > max_sector)
2689                         len = (max_sector - sector_nr) << 9;
2690                 if (len == 0)
2691                         break;
2692                 if (sync_blocks == 0) {
2693                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2694                                                &sync_blocks, still_degraded) &&
2695                             !conf->fullsync &&
2696                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2697                                 break;
2698                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2699                         if ((len >> 9) > sync_blocks)
2700                                 len = sync_blocks<<9;
2701                 }
2702
2703                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2704                         bio = r1_bio->bios[i];
2705                         if (bio->bi_end_io) {
2706                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2707                                 if (bio_add_page(bio, page, len, 0) == 0) {
2708                                         /* stop here */
2709                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2710                                         while (i > 0) {
2711                                                 i--;
2712                                                 bio = r1_bio->bios[i];
2713                                                 if (bio->bi_end_io==NULL)
2714                                                         continue;
2715                                                 /* remove last page from this bio */
2716                                                 bio->bi_vcnt--;
2717                                                 bio->bi_iter.bi_size -= len;
2718                                                 __clear_bit(BIO_SEG_VALID, &bio->bi_flags);
2719                                         }
2720                                         goto bio_full;
2721                                 }
2722                         }
2723                 }
2724                 nr_sectors += len>>9;
2725                 sector_nr += len>>9;
2726                 sync_blocks -= (len>>9);
2727         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2728  bio_full:
2729         r1_bio->sectors = nr_sectors;
2730
2731         /* For a user-requested sync, we read all readable devices and do a
2732          * compare
2733          */
2734         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2735                 atomic_set(&r1_bio->remaining, read_targets);
2736                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2737                         bio = r1_bio->bios[i];
2738                         if (bio->bi_end_io == end_sync_read) {
2739                                 read_targets--;
2740                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2741                                 generic_make_request(bio);
2742                         }
2743                 }
2744         } else {
2745                 atomic_set(&r1_bio->remaining, 1);
2746                 bio = r1_bio->bios[r1_bio->read_disk];
2747                 md_sync_acct(bio->bi_bdev, nr_sectors);
2748                 generic_make_request(bio);
2749
2750         }
2751         return nr_sectors;
2752 }
2753
2754 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2755 {
2756         if (sectors)
2757                 return sectors;
2758
2759         return mddev->dev_sectors;
2760 }
2761
2762 static struct r1conf *setup_conf(struct mddev *mddev)
2763 {
2764         struct r1conf *conf;
2765         int i;
2766         struct raid1_info *disk;
2767         struct md_rdev *rdev;
2768         int err = -ENOMEM;
2769
2770         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2771         if (!conf)
2772                 goto abort;
2773
2774         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2775                                 * mddev->raid_disks * 2,
2776                                  GFP_KERNEL);
2777         if (!conf->mirrors)
2778                 goto abort;
2779
2780         conf->tmppage = alloc_page(GFP_KERNEL);
2781         if (!conf->tmppage)
2782                 goto abort;
2783
2784         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2785         if (!conf->poolinfo)
2786                 goto abort;
2787         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2788         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2789                                           r1bio_pool_free,
2790                                           conf->poolinfo);
2791         if (!conf->r1bio_pool)
2792                 goto abort;
2793
2794         conf->poolinfo->mddev = mddev;
2795
2796         err = -EINVAL;
2797         spin_lock_init(&conf->device_lock);
2798         rdev_for_each(rdev, mddev) {
2799                 struct request_queue *q;
2800                 int disk_idx = rdev->raid_disk;
2801                 if (disk_idx >= mddev->raid_disks
2802                     || disk_idx < 0)
2803                         continue;
2804                 if (test_bit(Replacement, &rdev->flags))
2805                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
2806                 else
2807                         disk = conf->mirrors + disk_idx;
2808
2809                 if (disk->rdev)
2810                         goto abort;
2811                 disk->rdev = rdev;
2812                 q = bdev_get_queue(rdev->bdev);
2813                 if (q->merge_bvec_fn)
2814                         mddev->merge_check_needed = 1;
2815
2816                 disk->head_position = 0;
2817                 disk->seq_start = MaxSector;
2818         }
2819         conf->raid_disks = mddev->raid_disks;
2820         conf->mddev = mddev;
2821         INIT_LIST_HEAD(&conf->retry_list);
2822
2823         spin_lock_init(&conf->resync_lock);
2824         init_waitqueue_head(&conf->wait_barrier);
2825
2826         bio_list_init(&conf->pending_bio_list);
2827         conf->pending_count = 0;
2828         conf->recovery_disabled = mddev->recovery_disabled - 1;
2829
2830         conf->start_next_window = MaxSector;
2831         conf->current_window_requests = conf->next_window_requests = 0;
2832
2833         err = -EIO;
2834         for (i = 0; i < conf->raid_disks * 2; i++) {
2835
2836                 disk = conf->mirrors + i;
2837
2838                 if (i < conf->raid_disks &&
2839                     disk[conf->raid_disks].rdev) {
2840                         /* This slot has a replacement. */
2841                         if (!disk->rdev) {
2842                                 /* No original, just make the replacement
2843                                  * a recovering spare
2844                                  */
2845                                 disk->rdev =
2846                                         disk[conf->raid_disks].rdev;
2847                                 disk[conf->raid_disks].rdev = NULL;
2848                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2849                                 /* Original is not in_sync - bad */
2850                                 goto abort;
2851                 }
2852
2853                 if (!disk->rdev ||
2854                     !test_bit(In_sync, &disk->rdev->flags)) {
2855                         disk->head_position = 0;
2856                         if (disk->rdev &&
2857                             (disk->rdev->saved_raid_disk < 0))
2858                                 conf->fullsync = 1;
2859                 }
2860         }
2861
2862         err = -ENOMEM;
2863         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2864         if (!conf->thread) {
2865                 printk(KERN_ERR
2866                        "md/raid1:%s: couldn't allocate thread\n",
2867                        mdname(mddev));
2868                 goto abort;
2869         }
2870
2871         return conf;
2872
2873  abort:
2874         if (conf) {
2875                 if (conf->r1bio_pool)
2876                         mempool_destroy(conf->r1bio_pool);
2877                 kfree(conf->mirrors);
2878                 safe_put_page(conf->tmppage);
2879                 kfree(conf->poolinfo);
2880                 kfree(conf);
2881         }
2882         return ERR_PTR(err);
2883 }
2884
2885 static int stop(struct mddev *mddev);
2886 static int run(struct mddev *mddev)
2887 {
2888         struct r1conf *conf;
2889         int i;
2890         struct md_rdev *rdev;
2891         int ret;
2892         bool discard_supported = false;
2893
2894         if (mddev->level != 1) {
2895                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2896                        mdname(mddev), mddev->level);
2897                 return -EIO;
2898         }
2899         if (mddev->reshape_position != MaxSector) {
2900                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2901                        mdname(mddev));
2902                 return -EIO;
2903         }
2904         /*
2905          * copy the already verified devices into our private RAID1
2906          * bookkeeping area. [whatever we allocate in run(),
2907          * should be freed in stop()]
2908          */
2909         if (mddev->private == NULL)
2910                 conf = setup_conf(mddev);
2911         else
2912                 conf = mddev->private;
2913
2914         if (IS_ERR(conf))
2915                 return PTR_ERR(conf);
2916
2917         if (mddev->queue)
2918                 blk_queue_max_write_same_sectors(mddev->queue, 0);
2919
2920         rdev_for_each(rdev, mddev) {
2921                 if (!mddev->gendisk)
2922                         continue;
2923                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2924                                   rdev->data_offset << 9);
2925                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2926                         discard_supported = true;
2927         }
2928
2929         mddev->degraded = 0;
2930         for (i=0; i < conf->raid_disks; i++)
2931                 if (conf->mirrors[i].rdev == NULL ||
2932                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2933                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2934                         mddev->degraded++;
2935
2936         if (conf->raid_disks - mddev->degraded == 1)
2937                 mddev->recovery_cp = MaxSector;
2938
2939         if (mddev->recovery_cp != MaxSector)
2940                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2941                        " -- starting background reconstruction\n",
2942                        mdname(mddev));
2943         printk(KERN_INFO
2944                 "md/raid1:%s: active with %d out of %d mirrors\n",
2945                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2946                 mddev->raid_disks);
2947
2948         /*
2949          * Ok, everything is just fine now
2950          */
2951         mddev->thread = conf->thread;
2952         conf->thread = NULL;
2953         mddev->private = conf;
2954
2955         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2956
2957         if (mddev->queue) {
2958                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2959                 mddev->queue->backing_dev_info.congested_data = mddev;
2960                 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2961
2962                 if (discard_supported)
2963                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2964                                                 mddev->queue);
2965                 else
2966                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2967                                                   mddev->queue);
2968         }
2969
2970         ret =  md_integrity_register(mddev);
2971         if (ret)
2972                 stop(mddev);
2973         return ret;
2974 }
2975
2976 static int stop(struct mddev *mddev)
2977 {
2978         struct r1conf *conf = mddev->private;
2979         struct bitmap *bitmap = mddev->bitmap;
2980
2981         /* wait for behind writes to complete */
2982         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2983                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2984                        mdname(mddev));
2985                 /* need to kick something here to make sure I/O goes? */
2986                 wait_event(bitmap->behind_wait,
2987                            atomic_read(&bitmap->behind_writes) == 0);
2988         }
2989
2990         freeze_array(conf, 0);
2991         unfreeze_array(conf);
2992
2993         md_unregister_thread(&mddev->thread);
2994         if (conf->r1bio_pool)
2995                 mempool_destroy(conf->r1bio_pool);
2996         kfree(conf->mirrors);
2997         safe_put_page(conf->tmppage);
2998         kfree(conf->poolinfo);
2999         kfree(conf);
3000         mddev->private = NULL;
3001         return 0;
3002 }
3003
3004 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3005 {
3006         /* no resync is happening, and there is enough space
3007          * on all devices, so we can resize.
3008          * We need to make sure resync covers any new space.
3009          * If the array is shrinking we should possibly wait until
3010          * any io in the removed space completes, but it hardly seems
3011          * worth it.
3012          */
3013         sector_t newsize = raid1_size(mddev, sectors, 0);
3014         if (mddev->external_size &&
3015             mddev->array_sectors > newsize)
3016                 return -EINVAL;
3017         if (mddev->bitmap) {
3018                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3019                 if (ret)
3020                         return ret;
3021         }
3022         md_set_array_sectors(mddev, newsize);
3023         set_capacity(mddev->gendisk, mddev->array_sectors);
3024         revalidate_disk(mddev->gendisk);
3025         if (sectors > mddev->dev_sectors &&
3026             mddev->recovery_cp > mddev->dev_sectors) {
3027                 mddev->recovery_cp = mddev->dev_sectors;
3028                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3029         }
3030         mddev->dev_sectors = sectors;
3031         mddev->resync_max_sectors = sectors;
3032         return 0;
3033 }
3034
3035 static int raid1_reshape(struct mddev *mddev)
3036 {
3037         /* We need to:
3038          * 1/ resize the r1bio_pool
3039          * 2/ resize conf->mirrors
3040          *
3041          * We allocate a new r1bio_pool if we can.
3042          * Then raise a device barrier and wait until all IO stops.
3043          * Then resize conf->mirrors and swap in the new r1bio pool.
3044          *
3045          * At the same time, we "pack" the devices so that all the missing
3046          * devices have the higher raid_disk numbers.
3047          */
3048         mempool_t *newpool, *oldpool;
3049         struct pool_info *newpoolinfo;
3050         struct raid1_info *newmirrors;
3051         struct r1conf *conf = mddev->private;
3052         int cnt, raid_disks;
3053         unsigned long flags;
3054         int d, d2, err;
3055
3056         /* Cannot change chunk_size, layout, or level */
3057         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3058             mddev->layout != mddev->new_layout ||
3059             mddev->level != mddev->new_level) {
3060                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3061                 mddev->new_layout = mddev->layout;
3062                 mddev->new_level = mddev->level;
3063                 return -EINVAL;
3064         }
3065
3066         err = md_allow_write(mddev);
3067         if (err)
3068                 return err;
3069
3070         raid_disks = mddev->raid_disks + mddev->delta_disks;
3071
3072         if (raid_disks < conf->raid_disks) {
3073                 cnt=0;
3074                 for (d= 0; d < conf->raid_disks; d++)
3075                         if (conf->mirrors[d].rdev)
3076                                 cnt++;
3077                 if (cnt > raid_disks)
3078                         return -EBUSY;
3079         }
3080
3081         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3082         if (!newpoolinfo)
3083                 return -ENOMEM;
3084         newpoolinfo->mddev = mddev;
3085         newpoolinfo->raid_disks = raid_disks * 2;
3086
3087         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3088                                  r1bio_pool_free, newpoolinfo);
3089         if (!newpool) {
3090                 kfree(newpoolinfo);
3091                 return -ENOMEM;
3092         }
3093         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3094                              GFP_KERNEL);
3095         if (!newmirrors) {
3096                 kfree(newpoolinfo);
3097                 mempool_destroy(newpool);
3098                 return -ENOMEM;
3099         }
3100
3101         freeze_array(conf, 0);
3102
3103         /* ok, everything is stopped */
3104         oldpool = conf->r1bio_pool;
3105         conf->r1bio_pool = newpool;
3106
3107         for (d = d2 = 0; d < conf->raid_disks; d++) {
3108                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3109                 if (rdev && rdev->raid_disk != d2) {
3110                         sysfs_unlink_rdev(mddev, rdev);
3111                         rdev->raid_disk = d2;
3112                         sysfs_unlink_rdev(mddev, rdev);
3113                         if (sysfs_link_rdev(mddev, rdev))
3114                                 printk(KERN_WARNING
3115                                        "md/raid1:%s: cannot register rd%d\n",
3116                                        mdname(mddev), rdev->raid_disk);
3117                 }
3118                 if (rdev)
3119                         newmirrors[d2++].rdev = rdev;
3120         }
3121         kfree(conf->mirrors);
3122         conf->mirrors = newmirrors;
3123         kfree(conf->poolinfo);
3124         conf->poolinfo = newpoolinfo;
3125
3126         spin_lock_irqsave(&conf->device_lock, flags);
3127         mddev->degraded += (raid_disks - conf->raid_disks);
3128         spin_unlock_irqrestore(&conf->device_lock, flags);
3129         conf->raid_disks = mddev->raid_disks = raid_disks;
3130         mddev->delta_disks = 0;
3131
3132         unfreeze_array(conf);
3133
3134         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3135         md_wakeup_thread(mddev->thread);
3136
3137         mempool_destroy(oldpool);
3138         return 0;
3139 }
3140
3141 static void raid1_quiesce(struct mddev *mddev, int state)
3142 {
3143         struct r1conf *conf = mddev->private;
3144
3145         switch(state) {
3146         case 2: /* wake for suspend */
3147                 wake_up(&conf->wait_barrier);
3148                 break;
3149         case 1:
3150                 freeze_array(conf, 0);
3151                 break;
3152         case 0:
3153                 unfreeze_array(conf);
3154                 break;
3155         }
3156 }
3157
3158 static void *raid1_takeover(struct mddev *mddev)
3159 {
3160         /* raid1 can take over:
3161          *  raid5 with 2 devices, any layout or chunk size
3162          */
3163         if (mddev->level == 5 && mddev->raid_disks == 2) {
3164                 struct r1conf *conf;
3165                 mddev->new_level = 1;
3166                 mddev->new_layout = 0;
3167                 mddev->new_chunk_sectors = 0;
3168                 conf = setup_conf(mddev);
3169                 if (!IS_ERR(conf))
3170                         /* Array must appear to be quiesced */
3171                         conf->array_frozen = 1;
3172                 return conf;
3173         }
3174         return ERR_PTR(-EINVAL);
3175 }
3176
3177 static struct md_personality raid1_personality =
3178 {
3179         .name           = "raid1",
3180         .level          = 1,
3181         .owner          = THIS_MODULE,
3182         .make_request   = make_request,
3183         .run            = run,
3184         .stop           = stop,
3185         .status         = status,
3186         .error_handler  = error,
3187         .hot_add_disk   = raid1_add_disk,
3188         .hot_remove_disk= raid1_remove_disk,
3189         .spare_active   = raid1_spare_active,
3190         .sync_request   = sync_request,
3191         .resize         = raid1_resize,
3192         .size           = raid1_size,
3193         .check_reshape  = raid1_reshape,
3194         .quiesce        = raid1_quiesce,
3195         .takeover       = raid1_takeover,
3196 };
3197
3198 static int __init raid_init(void)
3199 {
3200         return register_md_personality(&raid1_personality);
3201 }
3202
3203 static void raid_exit(void)
3204 {
3205         unregister_md_personality(&raid1_personality);
3206 }
3207
3208 module_init(raid_init);
3209 module_exit(raid_exit);
3210 MODULE_LICENSE("GPL");
3211 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3212 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3213 MODULE_ALIAS("md-raid1");
3214 MODULE_ALIAS("md-level-1");
3215
3216 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);