drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/slab.h>
  35 #include <linux/delay.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/module.h>
  38 #include <linux/seq_file.h>
  39 #include <linux/ratelimit.h>
  40 #include "md.h"
  41 #include "raid1.h"
  42 #include "bitmap.h"
  43
  44 /*
  45  * Number of guaranteed r1bios in case of extreme VM load:
  46  */
  47 #define NR_RAID1_BIOS 256
  48
  49 /* when we get a read error on a read-only array, we redirect to another
  50  * device without failing the first device, or trying to over-write to
  51  * correct the read error.  To keep track of bad blocks on a per-bio
  52  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
  53  */
  54 #define IO_BLOCKED ((struct bio *)1)
  55 /* When we successfully write to a known bad-block, we need to remove the
  56  * bad-block marking which must be done from process context.  So we record
  57  * the success by setting devs[n].bio to IO_MADE_GOOD
  58  */
  59 #define IO_MADE_GOOD ((struct bio *)2)
  60
  61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
  62
  63 /* When there are this many requests queue to be written by
  64  * the raid1 thread, we become 'congested' to provide back-pressure
  65  * for writeback.
  66  */
  67 static int max_queued_requests = 1024;
  68
  69 static void allow_barrier(struct r1conf *conf);
  70 static void lower_barrier(struct r1conf *conf);
  71
  72 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  73 {
  74         struct pool_info *pi = data;
  75         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
  76
  77         /* allocate a r1bio with room for raid_disks entries in the bios array */
  78         return kzalloc(size, gfp_flags);
  79 }
  80
  81 static void r1bio_pool_free(void *r1_bio, void *data)
  82 {
  83         kfree(r1_bio);
  84 }
  85
  86 #define RESYNC_BLOCK_SIZE (64*1024)
  87 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  88 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  89 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  90 #define RESYNC_WINDOW (2048*1024)
  91
  92 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  93 {
  94         struct pool_info *pi = data;
  95         struct r1bio *r1_bio;
  96         struct bio *bio;
  97         int i, j;
  98
  99         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
 100         if (!r1_bio)
 101                 return NULL;
 102
 103         /*
 104          * Allocate bios : 1 for reading, n-1 for writing
 105          */
 106         for (j = pi->raid_disks ; j-- ; ) {
 107                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
 108                 if (!bio)
 109                         goto out_free_bio;
 110                 r1_bio->bios[j] = bio;
 111         }
 112         /*
 113          * Allocate RESYNC_PAGES data pages and attach them to
 114          * the first bio.
 115          * If this is a user-requested check/repair, allocate
 116          * RESYNC_PAGES for each bio.
 117          */
 118         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 119                 j = pi->raid_disks;
 120         else
 121                 j = 1;
 122         while(j--) {
 123                 bio = r1_bio->bios[j];
 124                 bio->bi_vcnt = RESYNC_PAGES;
 125
 126                 if (bio_alloc_pages(bio, gfp_flags))
 127                         goto out_free_bio;
 128         }
 129         /* If not user-requests, copy the page pointers to all bios */
 130         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 131                 for (i=0; i<RESYNC_PAGES ; i++)
 132                         for (j=1; j<pi->raid_disks; j++)
 133                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 134                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 135         }
 136
 137         r1_bio->master_bio = NULL;
 138
 139         return r1_bio;
 140
 141 out_free_bio:
 142         while (++j < pi->raid_disks)
 143                 bio_put(r1_bio->bios[j]);
 144         r1bio_pool_free(r1_bio, data);
 145         return NULL;
 146 }
 147
 148 static void r1buf_pool_free(void *__r1_bio, void *data)
 149 {
 150         struct pool_info *pi = data;
 151         int i,j;
 152         struct r1bio *r1bio = __r1_bio;
 153
 154         for (i = 0; i < RESYNC_PAGES; i++)
 155                 for (j = pi->raid_disks; j-- ;) {
 156                         if (j == 0 ||
 157                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 158                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 159                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 160                 }
 161         for (i=0 ; i < pi->raid_disks; i++)
 162                 bio_put(r1bio->bios[i]);
 163
 164         r1bio_pool_free(r1bio, data);
 165 }
 166
 167 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
 168 {
 169         int i;
 170
 171         for (i = 0; i < conf->raid_disks * 2; i++) {
 172                 struct bio **bio = r1_bio->bios + i;
 173                 if (!BIO_SPECIAL(*bio))
 174                         bio_put(*bio);
 175                 *bio = NULL;
 176         }
 177 }
 178
 179 static void free_r1bio(struct r1bio *r1_bio)
 180 {
 181         struct r1conf *conf = r1_bio->mddev->private;
 182
 183         put_all_bios(conf, r1_bio);
 184         mempool_free(r1_bio, conf->r1bio_pool);
 185 }
 186
 187 static void put_buf(struct r1bio *r1_bio)
 188 {
 189         struct r1conf *conf = r1_bio->mddev->private;
 190         int i;
 191
 192         for (i = 0; i < conf->raid_disks * 2; i++) {
 193                 struct bio *bio = r1_bio->bios[i];
 194                 if (bio->bi_end_io)
 195                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 196         }
 197
 198         mempool_free(r1_bio, conf->r1buf_pool);
 199
 200         lower_barrier(conf);
 201 }
 202
 203 static void reschedule_retry(struct r1bio *r1_bio)
 204 {
 205         unsigned long flags;
 206         struct mddev *mddev = r1_bio->mddev;
 207         struct r1conf *conf = mddev->private;
 208
 209         spin_lock_irqsave(&conf->device_lock, flags);
 210         list_add(&r1_bio->retry_list, &conf->retry_list);
 211         conf->nr_queued ++;
 212         spin_unlock_irqrestore(&conf->device_lock, flags);
 213
 214         wake_up(&conf->wait_barrier);
 215         md_wakeup_thread(mddev->thread);
 216 }
 217
 218 /*
 219  * raid_end_bio_io() is called when we have finished servicing a mirrored
 220  * operation and are ready to return a success/failure code to the buffer
 221  * cache layer.
 222  */
 223 static void call_bio_endio(struct r1bio *r1_bio)
 224 {
 225         struct bio *bio = r1_bio->master_bio;
 226         int done;
 227         struct r1conf *conf = r1_bio->mddev->private;
 228
 229         if (bio->bi_phys_segments) {
 230                 unsigned long flags;
 231                 spin_lock_irqsave(&conf->device_lock, flags);
 232                 bio->bi_phys_segments--;
 233                 done = (bio->bi_phys_segments == 0);
 234                 spin_unlock_irqrestore(&conf->device_lock, flags);
 235         } else
 236                 done = 1;
 237
 238         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 239                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 240         if (done) {
 241                 bio_endio(bio, 0);
 242                 /*
 243                  * Wake up any possible resync thread that waits for the device
 244                  * to go idle.
 245                  */
 246                 allow_barrier(conf);
 247         }
 248 }
 249
 250 static void raid_end_bio_io(struct r1bio *r1_bio)
 251 {
 252         struct bio *bio = r1_bio->master_bio;
 253
 254         /* if nobody has done the final endio yet, do it now */
 255         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 256                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
 257                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
 258                          (unsigned long long) bio->bi_sector,
 259                          (unsigned long long) bio->bi_sector +
 260                          bio_sectors(bio) - 1);
 261
 262                 call_bio_endio(r1_bio);
 263         }
 264         free_r1bio(r1_bio);
 265 }
 266
 267 /*
 268  * Update disk head position estimator based on IRQ completion info.
 269  */
 270 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
 271 {
 272         struct r1conf *conf = r1_bio->mddev->private;
 273
 274         conf->mirrors[disk].head_position =
 275                 r1_bio->sector + (r1_bio->sectors);
 276 }
 277
 278 /*
 279  * Find the disk number which triggered given bio
 280  */
 281 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
 282 {
 283         int mirror;
 284         struct r1conf *conf = r1_bio->mddev->private;
 285         int raid_disks = conf->raid_disks;
 286
 287         for (mirror = 0; mirror < raid_disks * 2; mirror++)
 288                 if (r1_bio->bios[mirror] == bio)
 289                         break;
 290
 291         BUG_ON(mirror == raid_disks * 2);
 292         update_head_pos(mirror, r1_bio);
 293
 294         return mirror;
 295 }
 296
 297 static void raid1_end_read_request(struct bio *bio, int error)
 298 {
 299         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 300         struct r1bio *r1_bio = bio->bi_private;
 301         int mirror;
 302         struct r1conf *conf = r1_bio->mddev->private;
 303
 304         mirror = r1_bio->read_disk;
 305         /*
 306          * this branch is our 'one mirror IO has finished' event handler:
 307          */
 308         update_head_pos(mirror, r1_bio);
 309
 310         if (uptodate)
 311                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 312         else {
 313                 /* If all other devices have failed, we want to return
 314                  * the error upwards rather than fail the last device.
 315                  * Here we redefine "uptodate" to mean "Don't want to retry"
 316                  */
 317                 unsigned long flags;
 318                 spin_lock_irqsave(&conf->device_lock, flags);
 319                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 320                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 321                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 322                         uptodate = 1;
 323                 spin_unlock_irqrestore(&conf->device_lock, flags);
 324         }
 325
 326         if (uptodate) {
 327                 raid_end_bio_io(r1_bio);
 328                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 329         } else {
 330                 /*
 331                  * oops, read error:
 332                  */
 333                 char b[BDEVNAME_SIZE];
 334                 printk_ratelimited(
 335                         KERN_ERR "md/raid1:%s: %s: "
 336                         "rescheduling sector %llu\n",
 337                         mdname(conf->mddev),
 338                         bdevname(conf->mirrors[mirror].rdev->bdev,
 339                                  b),
 340                         (unsigned long long)r1_bio->sector);
 341                 set_bit(R1BIO_ReadError, &r1_bio->state);
 342                 reschedule_retry(r1_bio);
 343                 /* don't drop the reference on read_disk yet */
 344         }
 345 }
 346
 347 static void close_write(struct r1bio *r1_bio)
 348 {
 349         /* it really is the end of this request */
 350         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 351                 /* free extra copy of the data pages */
 352                 int i = r1_bio->behind_page_count;
 353                 while (i--)
 354                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
 355                 kfree(r1_bio->behind_bvecs);
 356                 r1_bio->behind_bvecs = NULL;
 357         }
 358         /* clear the bitmap if all writes complete successfully */
 359         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 360                         r1_bio->sectors,
 361                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 362                         test_bit(R1BIO_BehindIO, &r1_bio->state));
 363         md_write_end(r1_bio->mddev);
 364 }
 365
 366 static void r1_bio_write_done(struct r1bio *r1_bio)
 367 {
 368         if (!atomic_dec_and_test(&r1_bio->remaining))
 369                 return;
 370
 371         if (test_bit(R1BIO_WriteError, &r1_bio->state))
 372                 reschedule_retry(r1_bio);
 373         else {
 374                 close_write(r1_bio);
 375                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
 376                         reschedule_retry(r1_bio);
 377                 else
 378                         raid_end_bio_io(r1_bio);
 379         }
 380 }
 381
 382 static void raid1_end_write_request(struct bio *bio, int error)
 383 {
 384         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 385         struct r1bio *r1_bio = bio->bi_private;
 386         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 387         struct r1conf *conf = r1_bio->mddev->private;
 388         struct bio *to_put = NULL;
 389
 390         mirror = find_bio_disk(r1_bio, bio);
 391
 392         /*
 393          * 'one mirror IO has finished' event handler:
 394          */
 395         if (!uptodate) {
 396                 set_bit(WriteErrorSeen,
 397                         &conf->mirrors[mirror].rdev->flags);
 398                 if (!test_and_set_bit(WantReplacement,
 399                                       &conf->mirrors[mirror].rdev->flags))
 400                         set_bit(MD_RECOVERY_NEEDED, &
 401                                 conf->mddev->recovery);
 402
 403                 set_bit(R1BIO_WriteError, &r1_bio->state);
 404         } else {
 405                 /*
 406                  * Set R1BIO_Uptodate in our master bio, so that we
 407                  * will return a good error code for to the higher
 408                  * levels even if IO on some other mirrored buffer
 409                  * fails.
 410                  *
 411                  * The 'master' represents the composite IO operation
 412                  * to user-side. So if something waits for IO, then it
 413                  * will wait for the 'master' bio.
 414                  */
 415                 sector_t first_bad;
 416                 int bad_sectors;
 417
 418                 r1_bio->bios[mirror] = NULL;
 419                 to_put = bio;
 420                 /*
 421                  * Do not set R1BIO_Uptodate if the current device is
 422                  * rebuilding or Faulty. This is because we cannot use
 423                  * such device for properly reading the data back (we could
 424                  * potentially use it, if the current write would have felt
 425                  * before rdev->recovery_offset, but for simplicity we don't
 426                  * check this here.
 427                  */
 428                 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
 429                     !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
 430                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 431
 432                 /* Maybe we can clear some bad blocks. */
 433                 if (is_badblock(conf->mirrors[mirror].rdev,
 434                                 r1_bio->sector, r1_bio->sectors,
 435                                 &first_bad, &bad_sectors)) {
 436                         r1_bio->bios[mirror] = IO_MADE_GOOD;
 437                         set_bit(R1BIO_MadeGood, &r1_bio->state);
 438                 }
 439         }
 440
 441         if (behind) {
 442                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 443                         atomic_dec(&r1_bio->behind_remaining);
 444
 445                 /*
 446                  * In behind mode, we ACK the master bio once the I/O
 447                  * has safely reached all non-writemostly
 448                  * disks. Setting the Returned bit ensures that this
 449                  * gets done only once -- we don't ever want to return
 450                  * -EIO here, instead we'll wait
 451                  */
 452                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 453                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 454                         /* Maybe we can return now */
 455                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 456                                 struct bio *mbio = r1_bio->master_bio;
 457                                 pr_debug("raid1: behind end write sectors"
 458                                          " %llu-%llu\n",
 459                                          (unsigned long long) mbio->bi_sector,
 460                                          (unsigned long long) mbio->bi_sector +
 461                                          bio_sectors(mbio) - 1);
 462                                 call_bio_endio(r1_bio);
 463                         }
 464                 }
 465         }
 466         if (r1_bio->bios[mirror] == NULL)
 467                 rdev_dec_pending(conf->mirrors[mirror].rdev,
 468                                  conf->mddev);
 469
 470         /*
 471          * Let's see if all mirrored write operations have finished
 472          * already.
 473          */
 474         r1_bio_write_done(r1_bio);
 475
 476         if (to_put)
 477                 bio_put(to_put);
 478 }
 479
 480
 481 /*
 482  * This routine returns the disk from which the requested read should
 483  * be done. There is a per-array 'next expected sequential IO' sector
 484  * number - if this matches on the next IO then we use the last disk.
 485  * There is also a per-disk 'last know head position' sector that is
 486  * maintained from IRQ contexts, both the normal and the resync IO
 487  * completion handlers update this position correctly. If there is no
 488  * perfect sequential match then we pick the disk whose head is closest.
 489  *
 490  * If there are 2 mirrors in the same 2 devices, performance degrades
 491  * because position is mirror, not device based.
 492  *
 493  * The rdev for the device selected will have nr_pending incremented.
 494  */
 495 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
 496 {
 497         const sector_t this_sector = r1_bio->sector;
 498         int sectors;
 499         int best_good_sectors;
 500         int best_disk, best_dist_disk, best_pending_disk;
 501         int has_nonrot_disk;
 502         int disk;
 503         sector_t best_dist;
 504         unsigned int min_pending;
 505         struct md_rdev *rdev;
 506         int choose_first;
 507         int choose_next_idle;
 508
 509         rcu_read_lock();
 510         /*
 511          * Check if we can balance. We can balance on the whole
 512          * device if no resync is going on, or below the resync window.
 513          * We take the first readable disk when above the resync window.
 514          */
 515  retry:
 516         sectors = r1_bio->sectors;
 517         best_disk = -1;
 518         best_dist_disk = -1;
 519         best_dist = MaxSector;
 520         best_pending_disk = -1;
 521         min_pending = UINT_MAX;
 522         best_good_sectors = 0;
 523         has_nonrot_disk = 0;
 524         choose_next_idle = 0;
 525
 526         if (conf->mddev->recovery_cp < MaxSector &&
 527             (this_sector + sectors >= conf->next_resync))
 528                 choose_first = 1;
 529         else
 530                 choose_first = 0;
 531
 532         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
 533                 sector_t dist;
 534                 sector_t first_bad;
 535                 int bad_sectors;
 536                 unsigned int pending;
 537                 bool nonrot;
 538
 539                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 540                 if (r1_bio->bios[disk] == IO_BLOCKED
 541                     || rdev == NULL
 542                     || test_bit(Unmerged, &rdev->flags)
 543                     || test_bit(Faulty, &rdev->flags))
 544                         continue;
 545                 if (!test_bit(In_sync, &rdev->flags) &&
 546                     rdev->recovery_offset < this_sector + sectors)
 547                         continue;
 548                 if (test_bit(WriteMostly, &rdev->flags)) {
 549                         /* Don't balance among write-mostly, just
 550                          * use the first as a last resort */
 551                         if (best_disk < 0) {
 552                                 if (is_badblock(rdev, this_sector, sectors,
 553                                                 &first_bad, &bad_sectors)) {
 554                                         if (first_bad < this_sector)
 555                                                 /* Cannot use this */
 556                                                 continue;
 557                                         best_good_sectors = first_bad - this_sector;
 558                                 } else
 559                                         best_good_sectors = sectors;
 560                                 best_disk = disk;
 561                         }
 562                         continue;
 563                 }
 564                 /* This is a reasonable device to use.  It might
 565                  * even be best.
 566                  */
 567                 if (is_badblock(rdev, this_sector, sectors,
 568                                 &first_bad, &bad_sectors)) {
 569                         if (best_dist < MaxSector)
 570                                 /* already have a better device */
 571                                 continue;
 572                         if (first_bad <= this_sector) {
 573                                 /* cannot read here. If this is the 'primary'
 574                                  * device, then we must not read beyond
 575                                  * bad_sectors from another device..
 576                                  */
 577                                 bad_sectors -= (this_sector - first_bad);
 578                                 if (choose_first && sectors > bad_sectors)
 579                                         sectors = bad_sectors;
 580                                 if (best_good_sectors > sectors)
 581                                         best_good_sectors = sectors;
 582
 583                         } else {
 584                                 sector_t good_sectors = first_bad - this_sector;
 585                                 if (good_sectors > best_good_sectors) {
 586                                         best_good_sectors = good_sectors;
 587                                         best_disk = disk;
 588                                 }
 589                                 if (choose_first)
 590                                         break;
 591                         }
 592                         continue;
 593                 } else
 594                         best_good_sectors = sectors;
 595
 596                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
 597                 has_nonrot_disk |= nonrot;
 598                 pending = atomic_read(&rdev->nr_pending);
 599                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 600                 if (choose_first) {
 601                         best_disk = disk;
 602                         break;
 603                 }
 604                 /* Don't change to another disk for sequential reads */
 605                 if (conf->mirrors[disk].next_seq_sect == this_sector
 606                     || dist == 0) {
 607                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
 608                         struct raid1_info *mirror = &conf->mirrors[disk];
 609
 610                         best_disk = disk;
 611                         /*
 612                          * If buffered sequential IO size exceeds optimal
 613                          * iosize, check if there is idle disk. If yes, choose
 614                          * the idle disk. read_balance could already choose an
 615                          * idle disk before noticing it's a sequential IO in
 616                          * this disk. This doesn't matter because this disk
 617                          * will idle, next time it will be utilized after the
 618                          * first disk has IO size exceeds optimal iosize. In
 619                          * this way, iosize of the first disk will be optimal
 620                          * iosize at least. iosize of the second disk might be
 621                          * small, but not a big deal since when the second disk
 622                          * starts IO, the first disk is likely still busy.
 623                          */
 624                         if (nonrot && opt_iosize > 0 &&
 625                             mirror->seq_start != MaxSector &&
 626                             mirror->next_seq_sect > opt_iosize &&
 627                             mirror->next_seq_sect - opt_iosize >=
 628                             mirror->seq_start) {
 629                                 choose_next_idle = 1;
 630                                 continue;
 631                         }
 632                         break;
 633                 }
 634                 /* If device is idle, use it */
 635                 if (pending == 0) {
 636                         best_disk = disk;
 637                         break;
 638                 }
 639
 640                 if (choose_next_idle)
 641                         continue;
 642
 643                 if (min_pending > pending) {
 644                         min_pending = pending;
 645                         best_pending_disk = disk;
 646                 }
 647
 648                 if (dist < best_dist) {
 649                         best_dist = dist;
 650                         best_dist_disk = disk;
 651                 }
 652         }
 653
 654         /*
 655          * If all disks are rotational, choose the closest disk. If any disk is
 656          * non-rotational, choose the disk with less pending request even the
 657          * disk is rotational, which might/might not be optimal for raids with
 658          * mixed ratation/non-rotational disks depending on workload.
 659          */
 660         if (best_disk == -1) {
 661                 if (has_nonrot_disk)
 662                         best_disk = best_pending_disk;
 663                 else
 664                         best_disk = best_dist_disk;
 665         }
 666
 667         if (best_disk >= 0) {
 668                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 669                 if (!rdev)
 670                         goto retry;
 671                 atomic_inc(&rdev->nr_pending);
 672                 if (test_bit(Faulty, &rdev->flags)) {
 673                         /* cannot risk returning a device that failed
 674                          * before we inc'ed nr_pending
 675                          */
 676                         rdev_dec_pending(rdev, conf->mddev);
 677                         goto retry;
 678                 }
 679                 sectors = best_good_sectors;
 680
 681                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
 682                         conf->mirrors[best_disk].seq_start = this_sector;
 683
 684                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
 685         }
 686         rcu_read_unlock();
 687         *max_sectors = sectors;
 688
 689         return best_disk;
 690 }
 691
 692 static int raid1_mergeable_bvec(struct request_queue *q,
 693                                 struct bvec_merge_data *bvm,
 694                                 struct bio_vec *biovec)
 695 {
 696         struct mddev *mddev = q->queuedata;
 697         struct r1conf *conf = mddev->private;
 698         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 699         int max = biovec->bv_len;
 700
 701         if (mddev->merge_check_needed) {
 702                 int disk;
 703                 rcu_read_lock();
 704                 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
 705                         struct md_rdev *rdev = rcu_dereference(
 706                                 conf->mirrors[disk].rdev);
 707                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
 708                                 struct request_queue *q =
 709                                         bdev_get_queue(rdev->bdev);
 710                                 if (q->merge_bvec_fn) {
 711                                         bvm->bi_sector = sector +
 712                                                 rdev->data_offset;
 713                                         bvm->bi_bdev = rdev->bdev;
 714                                         max = min(max, q->merge_bvec_fn(
 715                                                           q, bvm, biovec));
 716                                 }
 717                         }
 718                 }
 719                 rcu_read_unlock();
 720         }
 721         return max;
 722
 723 }
 724
 725 int md_raid1_congested(struct mddev *mddev, int bits)
 726 {
 727         struct r1conf *conf = mddev->private;
 728         int i, ret = 0;
 729
 730         if ((bits & (1 << BDI_async_congested)) &&
 731             conf->pending_count >= max_queued_requests)
 732                 return 1;
 733
 734         rcu_read_lock();
 735         for (i = 0; i < conf->raid_disks * 2; i++) {
 736                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
 737                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 738                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 739
 740                         BUG_ON(!q);
 741
 742                         /* Note the '|| 1' - when read_balance prefers
 743                          * non-congested targets, it can be removed
 744                          */
 745                         if ((bits & (1<<BDI_async_congested)) || 1)
 746                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 747                         else
 748                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 749                 }
 750         }
 751         rcu_read_unlock();
 752         return ret;
 753 }
 754 EXPORT_SYMBOL_GPL(md_raid1_congested);
 755
 756 static int raid1_congested(void *data, int bits)
 757 {
 758         struct mddev *mddev = data;
 759
 760         return mddev_congested(mddev, bits) ||
 761                 md_raid1_congested(mddev, bits);
 762 }
 763
 764 static void flush_pending_writes(struct r1conf *conf)
 765 {
 766         /* Any writes that have been queued but are awaiting
 767          * bitmap updates get flushed here.
 768          */
 769         spin_lock_irq(&conf->device_lock);
 770
 771         if (conf->pending_bio_list.head) {
 772                 struct bio *bio;
 773                 bio = bio_list_get(&conf->pending_bio_list);
 774                 conf->pending_count = 0;
 775                 spin_unlock_irq(&conf->device_lock);
 776                 /* flush any pending bitmap writes to
 777                  * disk before proceeding w/ I/O */
 778                 bitmap_unplug(conf->mddev->bitmap);
 779                 wake_up(&conf->wait_barrier);
 780
 781                 while (bio) { /* submit pending writes */
 782                         struct bio *next = bio->bi_next;
 783                         bio->bi_next = NULL;
 784                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 785                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 786                                 /* Just ignore it */
 787                                 bio_endio(bio, 0);
 788                         else
 789                                 generic_make_request(bio);
 790                         bio = next;
 791                 }
 792         } else
 793                 spin_unlock_irq(&conf->device_lock);
 794 }
 795
 796 /* Barriers....
 797  * Sometimes we need to suspend IO while we do something else,
 798  * either some resync/recovery, or reconfigure the array.
 799  * To do this we raise a 'barrier'.
 800  * The 'barrier' is a counter that can be raised multiple times
 801  * to count how many activities are happening which preclude
 802  * normal IO.
 803  * We can only raise the barrier if there is no pending IO.
 804  * i.e. if nr_pending == 0.
 805  * We choose only to raise the barrier if no-one is waiting for the
 806  * barrier to go down.  This means that as soon as an IO request
 807  * is ready, no other operations which require a barrier will start
 808  * until the IO request has had a chance.
 809  *
 810  * So: regular IO calls 'wait_barrier'.  When that returns there
 811  *    is no backgroup IO happening,  It must arrange to call
 812  *    allow_barrier when it has finished its IO.
 813  * backgroup IO calls must call raise_barrier.  Once that returns
 814  *    there is no normal IO happeing.  It must arrange to call
 815  *    lower_barrier when the particular background IO completes.
 816  */
 817 #define RESYNC_DEPTH 32
 818
 819 static void raise_barrier(struct r1conf *conf)
 820 {
 821         spin_lock_irq(&conf->resync_lock);
 822
 823         /* Wait until no block IO is waiting */
 824         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 825                             conf->resync_lock);
 826
 827         /* block any new IO from starting */
 828         conf->barrier++;
 829
 830         /* Now wait for all pending IO to complete */
 831         wait_event_lock_irq(conf->wait_barrier,
 832                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 833                             conf->resync_lock);
 834
 835         spin_unlock_irq(&conf->resync_lock);
 836 }
 837
 838 static void lower_barrier(struct r1conf *conf)
 839 {
 840         unsigned long flags;
 841         BUG_ON(conf->barrier <= 0);
 842         spin_lock_irqsave(&conf->resync_lock, flags);
 843         conf->barrier--;
 844         spin_unlock_irqrestore(&conf->resync_lock, flags);
 845         wake_up(&conf->wait_barrier);
 846 }
 847
 848 static void wait_barrier(struct r1conf *conf)
 849 {
 850         spin_lock_irq(&conf->resync_lock);
 851         if (conf->barrier) {
 852                 conf->nr_waiting++;
 853                 /* Wait for the barrier to drop.
 854                  * However if there are already pending
 855                  * requests (preventing the barrier from
 856                  * rising completely), and the
 857                  * pre-process bio queue isn't empty,
 858                  * then don't wait, as we need to empty
 859                  * that queue to get the nr_pending
 860                  * count down.
 861                  */
 862                 wait_event_lock_irq(conf->wait_barrier,
 863                                     !conf->barrier ||
 864                                     (conf->nr_pending &&
 865                                      current->bio_list &&
 866                                      !bio_list_empty(current->bio_list)),
 867                                     conf->resync_lock);
 868                 conf->nr_waiting--;
 869         }
 870         conf->nr_pending++;
 871         spin_unlock_irq(&conf->resync_lock);
 872 }
 873
 874 static void allow_barrier(struct r1conf *conf)
 875 {
 876         unsigned long flags;
 877         spin_lock_irqsave(&conf->resync_lock, flags);
 878         conf->nr_pending--;
 879         spin_unlock_irqrestore(&conf->resync_lock, flags);
 880         wake_up(&conf->wait_barrier);
 881 }
 882
 883 static void freeze_array(struct r1conf *conf, int extra)
 884 {
 885         /* stop syncio and normal IO and wait for everything to
 886          * go quite.
 887          * We increment barrier and nr_waiting, and then
 888          * wait until nr_pending match nr_queued+extra
 889          * This is called in the context of one normal IO request
 890          * that has failed. Thus any sync request that might be pending
 891          * will be blocked by nr_pending, and we need to wait for
 892          * pending IO requests to complete or be queued for re-try.
 893          * Thus the number queued (nr_queued) plus this request (extra)
 894          * must match the number of pending IOs (nr_pending) before
 895          * we continue.
 896          */
 897         spin_lock_irq(&conf->resync_lock);
 898         conf->barrier++;
 899         conf->nr_waiting++;
 900         wait_event_lock_irq_cmd(conf->wait_barrier,
 901                                 conf->nr_pending == conf->nr_queued+extra,
 902                                 conf->resync_lock,
 903                                 flush_pending_writes(conf));
 904         spin_unlock_irq(&conf->resync_lock);
 905 }
 906 static void unfreeze_array(struct r1conf *conf)
 907 {
 908         /* reverse the effect of the freeze */
 909         spin_lock_irq(&conf->resync_lock);
 910         conf->barrier--;
 911         conf->nr_waiting--;
 912         wake_up(&conf->wait_barrier);
 913         spin_unlock_irq(&conf->resync_lock);
 914 }
 915
 916
 917 /* duplicate the data pages for behind I/O
 918  */
 919 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
 920 {
 921         int i;
 922         struct bio_vec *bvec;
 923         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 924                                         GFP_NOIO);
 925         if (unlikely(!bvecs))
 926                 return;
 927
 928         bio_for_each_segment_all(bvec, bio, i) {
 929                 bvecs[i] = *bvec;
 930                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
 931                 if (unlikely(!bvecs[i].bv_page))
 932                         goto do_sync_io;
 933                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
 934                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 935                 kunmap(bvecs[i].bv_page);
 936                 kunmap(bvec->bv_page);
 937         }
 938         r1_bio->behind_bvecs = bvecs;
 939         r1_bio->behind_page_count = bio->bi_vcnt;
 940         set_bit(R1BIO_BehindIO, &r1_bio->state);
 941         return;
 942
 943 do_sync_io:
 944         for (i = 0; i < bio->bi_vcnt; i++)
 945                 if (bvecs[i].bv_page)
 946                         put_page(bvecs[i].bv_page);
 947         kfree(bvecs);
 948         pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 949 }
 950
 951 struct raid1_plug_cb {
 952         struct blk_plug_cb      cb;
 953         struct bio_list         pending;
 954         int                     pending_cnt;
 955 };
 956
 957 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
 958 {
 959         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
 960                                                   cb);
 961         struct mddev *mddev = plug->cb.data;
 962         struct r1conf *conf = mddev->private;
 963         struct bio *bio;
 964
 965         if (from_schedule || current->bio_list) {
 966                 spin_lock_irq(&conf->device_lock);
 967                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
 968                 conf->pending_count += plug->pending_cnt;
 969                 spin_unlock_irq(&conf->device_lock);
 970                 wake_up(&conf->wait_barrier);
 971                 md_wakeup_thread(mddev->thread);
 972                 kfree(plug);
 973                 return;
 974         }
 975
 976         /* we aren't scheduling, so we can do the write-out directly. */
 977         bio = bio_list_get(&plug->pending);
 978         bitmap_unplug(mddev->bitmap);
 979         wake_up(&conf->wait_barrier);
 980
 981         while (bio) { /* submit pending writes */
 982                 struct bio *next = bio->bi_next;
 983                 bio->bi_next = NULL;
 984                 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
 985                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
 986                         /* Just ignore it */
 987                         bio_endio(bio, 0);
 988                 else
 989                         generic_make_request(bio);
 990                 bio = next;
 991         }
 992         kfree(plug);
 993 }
 994
 995 static void make_request(struct mddev *mddev, struct bio * bio)
 996 {
 997         struct r1conf *conf = mddev->private;
 998         struct raid1_info *mirror;
 999         struct r1bio *r1_bio;
1000         struct bio *read_bio;
1001         int i, disks;
1002         struct bitmap *bitmap;
1003         unsigned long flags;
1004         const int rw = bio_data_dir(bio);
1005         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1006         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1007         const unsigned long do_discard = (bio->bi_rw
1008                                           & (REQ_DISCARD | REQ_SECURE));
1009         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1010         struct md_rdev *blocked_rdev;
1011         struct blk_plug_cb *cb;
1012         struct raid1_plug_cb *plug = NULL;
1013         int first_clone;
1014         int sectors_handled;
1015         int max_sectors;
1016
1017         /*
1018          * Register the new request and wait if the reconstruction
1019          * thread has put up a bar for new requests.
1020          * Continue immediately if no resync is active currently.
1021          */
1022
1023         md_write_start(mddev, bio); /* wait on superblock update early */
1024
1025         if (bio_data_dir(bio) == WRITE &&
1026             bio_end_sector(bio) > mddev->suspend_lo &&
1027             bio->bi_sector < mddev->suspend_hi) {
1028                 /* As the suspend_* range is controlled by
1029                  * userspace, we want an interruptible
1030                  * wait.
1031                  */
1032                 DEFINE_WAIT(w);
1033                 for (;;) {
1034                         flush_signals(current);
1035                         prepare_to_wait(&conf->wait_barrier,
1036                                         &w, TASK_INTERRUPTIBLE);
1037                         if (bio_end_sector(bio) <= mddev->suspend_lo ||
1038                             bio->bi_sector >= mddev->suspend_hi)
1039                                 break;
1040                         schedule();
1041                 }
1042                 finish_wait(&conf->wait_barrier, &w);
1043         }
1044
1045         wait_barrier(conf);
1046
1047         bitmap = mddev->bitmap;
1048
1049         /*
1050          * make_request() can abort the operation when READA is being
1051          * used and no empty request is available.
1052          *
1053          */
1054         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1055
1056         r1_bio->master_bio = bio;
1057         r1_bio->sectors = bio_sectors(bio);
1058         r1_bio->state = 0;
1059         r1_bio->mddev = mddev;
1060         r1_bio->sector = bio->bi_sector;
1061
1062         /* We might need to issue multiple reads to different
1063          * devices if there are bad blocks around, so we keep
1064          * track of the number of reads in bio->bi_phys_segments.
1065          * If this is 0, there is only one r1_bio and no locking
1066          * will be needed when requests complete.  If it is
1067          * non-zero, then it is the number of not-completed requests.
1068          */
1069         bio->bi_phys_segments = 0;
1070         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1071
1072         if (rw == READ) {
1073                 /*
1074                  * read balancing logic:
1075                  */
1076                 int rdisk;
1077
1078 read_again:
1079                 rdisk = read_balance(conf, r1_bio, &max_sectors);
1080
1081                 if (rdisk < 0) {
1082                         /* couldn't find anywhere to read from */
1083                         raid_end_bio_io(r1_bio);
1084                         return;
1085                 }
1086                 mirror = conf->mirrors + rdisk;
1087
1088                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1089                     bitmap) {
1090                         /* Reading from a write-mostly device must
1091                          * take care not to over-take any writes
1092                          * that are 'behind'
1093                          */
1094                         wait_event(bitmap->behind_wait,
1095                                    atomic_read(&bitmap->behind_writes) == 0);
1096                 }
1097                 r1_bio->read_disk = rdisk;
1098
1099                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1100                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1101                             max_sectors);
1102
1103                 r1_bio->bios[rdisk] = read_bio;
1104
1105                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1106                 read_bio->bi_bdev = mirror->rdev->bdev;
1107                 read_bio->bi_end_io = raid1_end_read_request;
1108                 read_bio->bi_rw = READ | do_sync;
1109                 read_bio->bi_private = r1_bio;
1110
1111                 if (max_sectors < r1_bio->sectors) {
1112                         /* could not read all from this device, so we will
1113                          * need another r1_bio.
1114                          */
1115
1116                         sectors_handled = (r1_bio->sector + max_sectors
1117                                            - bio->bi_sector);
1118                         r1_bio->sectors = max_sectors;
1119                         spin_lock_irq(&conf->device_lock);
1120                         if (bio->bi_phys_segments == 0)
1121                                 bio->bi_phys_segments = 2;
1122                         else
1123                                 bio->bi_phys_segments++;
1124                         spin_unlock_irq(&conf->device_lock);
1125                         /* Cannot call generic_make_request directly
1126                          * as that will be queued in __make_request
1127                          * and subsequent mempool_alloc might block waiting
1128                          * for it.  So hand bio over to raid1d.
1129                          */
1130                         reschedule_retry(r1_bio);
1131
1132                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1133
1134                         r1_bio->master_bio = bio;
1135                         r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1136                         r1_bio->state = 0;
1137                         r1_bio->mddev = mddev;
1138                         r1_bio->sector = bio->bi_sector + sectors_handled;
1139                         goto read_again;
1140                 } else
1141                         generic_make_request(read_bio);
1142                 return;
1143         }
1144
1145         /*
1146          * WRITE:
1147          */
1148         if (conf->pending_count >= max_queued_requests) {
1149                 md_wakeup_thread(mddev->thread);
1150                 wait_event(conf->wait_barrier,
1151                            conf->pending_count < max_queued_requests);
1152         }
1153         /* first select target devices under rcu_lock and
1154          * inc refcount on their rdev.  Record them by setting
1155          * bios[x] to bio
1156          * If there are known/acknowledged bad blocks on any device on
1157          * which we have seen a write error, we want to avoid writing those
1158          * blocks.
1159          * This potentially requires several writes to write around
1160          * the bad blocks.  Each set of writes gets it's own r1bio
1161          * with a set of bios attached.
1162          */
1163
1164         disks = conf->raid_disks * 2;
1165  retry_write:
1166         blocked_rdev = NULL;
1167         rcu_read_lock();
1168         max_sectors = r1_bio->sectors;
1169         for (i = 0;  i < disks; i++) {
1170                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1171                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1172                         atomic_inc(&rdev->nr_pending);
1173                         blocked_rdev = rdev;
1174                         break;
1175                 }
1176                 r1_bio->bios[i] = NULL;
1177                 if (!rdev || test_bit(Faulty, &rdev->flags)
1178                     || test_bit(Unmerged, &rdev->flags)) {
1179                         if (i < conf->raid_disks)
1180                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1181                         continue;
1182                 }
1183
1184                 atomic_inc(&rdev->nr_pending);
1185                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1186                         sector_t first_bad;
1187                         int bad_sectors;
1188                         int is_bad;
1189
1190                         is_bad = is_badblock(rdev, r1_bio->sector,
1191                                              max_sectors,
1192                                              &first_bad, &bad_sectors);
1193                         if (is_bad < 0) {
1194                                 /* mustn't write here until the bad block is
1195                                  * acknowledged*/
1196                                 set_bit(BlockedBadBlocks, &rdev->flags);
1197                                 blocked_rdev = rdev;
1198                                 break;
1199                         }
1200                         if (is_bad && first_bad <= r1_bio->sector) {
1201                                 /* Cannot write here at all */
1202                                 bad_sectors -= (r1_bio->sector - first_bad);
1203                                 if (bad_sectors < max_sectors)
1204                                         /* mustn't write more than bad_sectors
1205                                          * to other devices yet
1206                                          */
1207                                         max_sectors = bad_sectors;
1208                                 rdev_dec_pending(rdev, mddev);
1209                                 /* We don't set R1BIO_Degraded as that
1210                                  * only applies if the disk is
1211                                  * missing, so it might be re-added,
1212                                  * and we want to know to recover this
1213                                  * chunk.
1214                                  * In this case the device is here,
1215                                  * and the fact that this chunk is not
1216                                  * in-sync is recorded in the bad
1217                                  * block log
1218                                  */
1219                                 continue;
1220                         }
1221                         if (is_bad) {
1222                                 int good_sectors = first_bad - r1_bio->sector;
1223                                 if (good_sectors < max_sectors)
1224                                         max_sectors = good_sectors;
1225                         }
1226                 }
1227                 r1_bio->bios[i] = bio;
1228         }
1229         rcu_read_unlock();
1230
1231         if (unlikely(blocked_rdev)) {
1232                 /* Wait for this device to become unblocked */
1233                 int j;
1234
1235                 for (j = 0; j < i; j++)
1236                         if (r1_bio->bios[j])
1237                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1238                 r1_bio->state = 0;
1239                 allow_barrier(conf);
1240                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1241                 wait_barrier(conf);
1242                 goto retry_write;
1243         }
1244
1245         if (max_sectors < r1_bio->sectors) {
1246                 /* We are splitting this write into multiple parts, so
1247                  * we need to prepare for allocating another r1_bio.
1248                  */
1249                 r1_bio->sectors = max_sectors;
1250                 spin_lock_irq(&conf->device_lock);
1251                 if (bio->bi_phys_segments == 0)
1252                         bio->bi_phys_segments = 2;
1253                 else
1254                         bio->bi_phys_segments++;
1255                 spin_unlock_irq(&conf->device_lock);
1256         }
1257         sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1258
1259         atomic_set(&r1_bio->remaining, 1);
1260         atomic_set(&r1_bio->behind_remaining, 0);
1261
1262         first_clone = 1;
1263         for (i = 0; i < disks; i++) {
1264                 struct bio *mbio;
1265                 if (!r1_bio->bios[i])
1266                         continue;
1267
1268                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1269                 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1270
1271                 if (first_clone) {
1272                         /* do behind I/O ?
1273                          * Not if there are too many, or cannot
1274                          * allocate memory, or a reader on WriteMostly
1275                          * is waiting for behind writes to flush */
1276                         if (bitmap &&
1277                             (atomic_read(&bitmap->behind_writes)
1278                              < mddev->bitmap_info.max_write_behind) &&
1279                             !waitqueue_active(&bitmap->behind_wait))
1280                                 alloc_behind_pages(mbio, r1_bio);
1281
1282                         bitmap_startwrite(bitmap, r1_bio->sector,
1283                                           r1_bio->sectors,
1284                                           test_bit(R1BIO_BehindIO,
1285                                                    &r1_bio->state));
1286                         first_clone = 0;
1287                 }
1288                 if (r1_bio->behind_bvecs) {
1289                         struct bio_vec *bvec;
1290                         int j;
1291
1292                         /*
1293                          * We trimmed the bio, so _all is legit
1294                          */
1295                         bio_for_each_segment_all(bvec, mbio, j)
1296                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1297                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1298                                 atomic_inc(&r1_bio->behind_remaining);
1299                 }
1300
1301                 r1_bio->bios[i] = mbio;
1302
1303                 mbio->bi_sector = (r1_bio->sector +
1304                                    conf->mirrors[i].rdev->data_offset);
1305                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1306                 mbio->bi_end_io = raid1_end_write_request;
1307                 mbio->bi_rw =
1308                         WRITE | do_flush_fua | do_sync | do_discard | do_same;
1309                 mbio->bi_private = r1_bio;
1310
1311                 atomic_inc(&r1_bio->remaining);
1312
1313                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1314                 if (cb)
1315                         plug = container_of(cb, struct raid1_plug_cb, cb);
1316                 else
1317                         plug = NULL;
1318                 spin_lock_irqsave(&conf->device_lock, flags);
1319                 if (plug) {
1320                         bio_list_add(&plug->pending, mbio);
1321                         plug->pending_cnt++;
1322                 } else {
1323                         bio_list_add(&conf->pending_bio_list, mbio);
1324                         conf->pending_count++;
1325                 }
1326                 spin_unlock_irqrestore(&conf->device_lock, flags);
1327                 if (!plug)
1328                         md_wakeup_thread(mddev->thread);
1329         }
1330         /* Mustn't call r1_bio_write_done before this next test,
1331          * as it could result in the bio being freed.
1332          */
1333         if (sectors_handled < bio_sectors(bio)) {
1334                 r1_bio_write_done(r1_bio);
1335                 /* We need another r1_bio.  It has already been counted
1336                  * in bio->bi_phys_segments
1337                  */
1338                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1339                 r1_bio->master_bio = bio;
1340                 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1341                 r1_bio->state = 0;
1342                 r1_bio->mddev = mddev;
1343                 r1_bio->sector = bio->bi_sector + sectors_handled;
1344                 goto retry_write;
1345         }
1346
1347         r1_bio_write_done(r1_bio);
1348
1349         /* In case raid1d snuck in to freeze_array */
1350         wake_up(&conf->wait_barrier);
1351 }
1352
1353 static void status(struct seq_file *seq, struct mddev *mddev)
1354 {
1355         struct r1conf *conf = mddev->private;
1356         int i;
1357
1358         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1359                    conf->raid_disks - mddev->degraded);
1360         rcu_read_lock();
1361         for (i = 0; i < conf->raid_disks; i++) {
1362                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1363                 seq_printf(seq, "%s",
1364                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1365         }
1366         rcu_read_unlock();
1367         seq_printf(seq, "]");
1368 }
1369
1370
1371 static void error(struct mddev *mddev, struct md_rdev *rdev)
1372 {
1373         char b[BDEVNAME_SIZE];
1374         struct r1conf *conf = mddev->private;
1375
1376         /*
1377          * If it is not operational, then we have already marked it as dead
1378          * else if it is the last working disks, ignore the error, let the
1379          * next level up know.
1380          * else mark the drive as failed
1381          */
1382         if (test_bit(In_sync, &rdev->flags)
1383             && (conf->raid_disks - mddev->degraded) == 1) {
1384                 /*
1385                  * Don't fail the drive, act as though we were just a
1386                  * normal single drive.
1387                  * However don't try a recovery from this drive as
1388                  * it is very likely to fail.
1389                  */
1390                 conf->recovery_disabled = mddev->recovery_disabled;
1391                 return;
1392         }
1393         set_bit(Blocked, &rdev->flags);
1394         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1395                 unsigned long flags;
1396                 spin_lock_irqsave(&conf->device_lock, flags);
1397                 mddev->degraded++;
1398                 set_bit(Faulty, &rdev->flags);
1399                 spin_unlock_irqrestore(&conf->device_lock, flags);
1400                 /*
1401                  * if recovery is running, make sure it aborts.
1402                  */
1403                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1404         } else
1405                 set_bit(Faulty, &rdev->flags);
1406         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1407         printk(KERN_ALERT
1408                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1409                "md/raid1:%s: Operation continuing on %d devices.\n",
1410                mdname(mddev), bdevname(rdev->bdev, b),
1411                mdname(mddev), conf->raid_disks - mddev->degraded);
1412 }
1413
1414 static void print_conf(struct r1conf *conf)
1415 {
1416         int i;
1417
1418         printk(KERN_DEBUG "RAID1 conf printout:\n");
1419         if (!conf) {
1420                 printk(KERN_DEBUG "(!conf)\n");
1421                 return;
1422         }
1423         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1424                 conf->raid_disks);
1425
1426         rcu_read_lock();
1427         for (i = 0; i < conf->raid_disks; i++) {
1428                 char b[BDEVNAME_SIZE];
1429                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1430                 if (rdev)
1431                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1432                                i, !test_bit(In_sync, &rdev->flags),
1433                                !test_bit(Faulty, &rdev->flags),
1434                                bdevname(rdev->bdev,b));
1435         }
1436         rcu_read_unlock();
1437 }
1438
1439 static void close_sync(struct r1conf *conf)
1440 {
1441         wait_barrier(conf);
1442         allow_barrier(conf);
1443
1444         mempool_destroy(conf->r1buf_pool);
1445         conf->r1buf_pool = NULL;
1446 }
1447
1448 static int raid1_spare_active(struct mddev *mddev)
1449 {
1450         int i;
1451         struct r1conf *conf = mddev->private;
1452         int count = 0;
1453         unsigned long flags;
1454
1455         /*
1456          * Find all failed disks within the RAID1 configuration
1457          * and mark them readable.
1458          * Called under mddev lock, so rcu protection not needed.
1459          */
1460         for (i = 0; i < conf->raid_disks; i++) {
1461                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1462                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1463                 if (repl
1464                     && repl->recovery_offset == MaxSector
1465                     && !test_bit(Faulty, &repl->flags)
1466                     && !test_and_set_bit(In_sync, &repl->flags)) {
1467                         /* replacement has just become active */
1468                         if (!rdev ||
1469                             !test_and_clear_bit(In_sync, &rdev->flags))
1470                                 count++;
1471                         if (rdev) {
1472                                 /* Replaced device not technically
1473                                  * faulty, but we need to be sure
1474                                  * it gets removed and never re-added
1475                                  */
1476                                 set_bit(Faulty, &rdev->flags);
1477                                 sysfs_notify_dirent_safe(
1478                                         rdev->sysfs_state);
1479                         }
1480                 }
1481                 if (rdev
1482                     && !test_bit(Faulty, &rdev->flags)
1483                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1484                         count++;
1485                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1486                 }
1487         }
1488         spin_lock_irqsave(&conf->device_lock, flags);
1489         mddev->degraded -= count;
1490         spin_unlock_irqrestore(&conf->device_lock, flags);
1491
1492         print_conf(conf);
1493         return count;
1494 }
1495
1496
1497 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1498 {
1499         struct r1conf *conf = mddev->private;
1500         int err = -EEXIST;
1501         int mirror = 0;
1502         struct raid1_info *p;
1503         int first = 0;
1504         int last = conf->raid_disks - 1;
1505         struct request_queue *q = bdev_get_queue(rdev->bdev);
1506
1507         if (mddev->recovery_disabled == conf->recovery_disabled)
1508                 return -EBUSY;
1509
1510         if (rdev->raid_disk >= 0)
1511                 first = last = rdev->raid_disk;
1512
1513         if (q->merge_bvec_fn) {
1514                 set_bit(Unmerged, &rdev->flags);
1515                 mddev->merge_check_needed = 1;
1516         }
1517
1518         for (mirror = first; mirror <= last; mirror++) {
1519                 p = conf->mirrors+mirror;
1520                 if (!p->rdev) {
1521
1522                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1523                                           rdev->data_offset << 9);
1524
1525                         p->head_position = 0;
1526                         rdev->raid_disk = mirror;
1527                         err = 0;
1528                         /* As all devices are equivalent, we don't need a full recovery
1529                          * if this was recently any drive of the array
1530                          */
1531                         if (rdev->saved_raid_disk < 0)
1532                                 conf->fullsync = 1;
1533                         rcu_assign_pointer(p->rdev, rdev);
1534                         break;
1535                 }
1536                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1537                     p[conf->raid_disks].rdev == NULL) {
1538                         /* Add this device as a replacement */
1539                         clear_bit(In_sync, &rdev->flags);
1540                         set_bit(Replacement, &rdev->flags);
1541                         rdev->raid_disk = mirror;
1542                         err = 0;
1543                         conf->fullsync = 1;
1544                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1545                         break;
1546                 }
1547         }
1548         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1549                 /* Some requests might not have seen this new
1550                  * merge_bvec_fn.  We must wait for them to complete
1551                  * before merging the device fully.
1552                  * First we make sure any code which has tested
1553                  * our function has submitted the request, then
1554                  * we wait for all outstanding requests to complete.
1555                  */
1556                 synchronize_sched();
1557                 freeze_array(conf, 0);
1558                 unfreeze_array(conf);
1559                 clear_bit(Unmerged, &rdev->flags);
1560         }
1561         md_integrity_add_rdev(rdev, mddev);
1562         if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
1563                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1564         print_conf(conf);
1565         return err;
1566 }
1567
1568 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1569 {
1570         struct r1conf *conf = mddev->private;
1571         int err = 0;
1572         int number = rdev->raid_disk;
1573         struct raid1_info *p = conf->mirrors + number;
1574
1575         if (rdev != p->rdev)
1576                 p = conf->mirrors + conf->raid_disks + number;
1577
1578         print_conf(conf);
1579         if (rdev == p->rdev) {
1580                 if (test_bit(In_sync, &rdev->flags) ||
1581                     atomic_read(&rdev->nr_pending)) {
1582                         err = -EBUSY;
1583                         goto abort;
1584                 }
1585                 /* Only remove non-faulty devices if recovery
1586                  * is not possible.
1587                  */
1588                 if (!test_bit(Faulty, &rdev->flags) &&
1589                     mddev->recovery_disabled != conf->recovery_disabled &&
1590                     mddev->degraded < conf->raid_disks) {
1591                         err = -EBUSY;
1592                         goto abort;
1593                 }
1594                 p->rdev = NULL;
1595                 synchronize_rcu();
1596                 if (atomic_read(&rdev->nr_pending)) {
1597                         /* lost the race, try later */
1598                         err = -EBUSY;
1599                         p->rdev = rdev;
1600                         goto abort;
1601                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1602                         /* We just removed a device that is being replaced.
1603                          * Move down the replacement.  We drain all IO before
1604                          * doing this to avoid confusion.
1605                          */
1606                         struct md_rdev *repl =
1607                                 conf->mirrors[conf->raid_disks + number].rdev;
1608                         freeze_array(conf, 0);
1609                         clear_bit(Replacement, &repl->flags);
1610                         p->rdev = repl;
1611                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1612                         unfreeze_array(conf);
1613                         clear_bit(WantReplacement, &rdev->flags);
1614                 } else
1615                         clear_bit(WantReplacement, &rdev->flags);
1616                 err = md_integrity_register(mddev);
1617         }
1618 abort:
1619
1620         print_conf(conf);
1621         return err;
1622 }
1623
1624
1625 static void end_sync_read(struct bio *bio, int error)
1626 {
1627         struct r1bio *r1_bio = bio->bi_private;
1628
1629         update_head_pos(r1_bio->read_disk, r1_bio);
1630
1631         /*
1632          * we have read a block, now it needs to be re-written,
1633          * or re-read if the read failed.
1634          * We don't do much here, just schedule handling by raid1d
1635          */
1636         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1637                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1638
1639         if (atomic_dec_and_test(&r1_bio->remaining))
1640                 reschedule_retry(r1_bio);
1641 }
1642
1643 static void end_sync_write(struct bio *bio, int error)
1644 {
1645         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1646         struct r1bio *r1_bio = bio->bi_private;
1647         struct mddev *mddev = r1_bio->mddev;
1648         struct r1conf *conf = mddev->private;
1649         int mirror=0;
1650         sector_t first_bad;
1651         int bad_sectors;
1652
1653         mirror = find_bio_disk(r1_bio, bio);
1654
1655         if (!uptodate) {
1656                 sector_t sync_blocks = 0;
1657                 sector_t s = r1_bio->sector;
1658                 long sectors_to_go = r1_bio->sectors;
1659                 /* make sure these bits doesn't get cleared. */
1660                 do {
1661                         bitmap_end_sync(mddev->bitmap, s,
1662                                         &sync_blocks, 1);
1663                         s += sync_blocks;
1664                         sectors_to_go -= sync_blocks;
1665                 } while (sectors_to_go > 0);
1666                 set_bit(WriteErrorSeen,
1667                         &conf->mirrors[mirror].rdev->flags);
1668                 if (!test_and_set_bit(WantReplacement,
1669                                       &conf->mirrors[mirror].rdev->flags))
1670                         set_bit(MD_RECOVERY_NEEDED, &
1671                                 mddev->recovery);
1672                 set_bit(R1BIO_WriteError, &r1_bio->state);
1673         } else if (is_badblock(conf->mirrors[mirror].rdev,
1674                                r1_bio->sector,
1675                                r1_bio->sectors,
1676                                &first_bad, &bad_sectors) &&
1677                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1678                                 r1_bio->sector,
1679                                 r1_bio->sectors,
1680                                 &first_bad, &bad_sectors)
1681                 )
1682                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1683
1684         if (atomic_dec_and_test(&r1_bio->remaining)) {
1685                 int s = r1_bio->sectors;
1686                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1687                     test_bit(R1BIO_WriteError, &r1_bio->state))
1688                         reschedule_retry(r1_bio);
1689                 else {
1690                         put_buf(r1_bio);
1691                         md_done_sync(mddev, s, uptodate);
1692                 }
1693         }
1694 }
1695
1696 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1697                             int sectors, struct page *page, int rw)
1698 {
1699         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1700                 /* success */
1701                 return 1;
1702         if (rw == WRITE) {
1703                 set_bit(WriteErrorSeen, &rdev->flags);
1704                 if (!test_and_set_bit(WantReplacement,
1705                                       &rdev->flags))
1706                         set_bit(MD_RECOVERY_NEEDED, &
1707                                 rdev->mddev->recovery);
1708         }
1709         /* need to record an error - either for the block or the device */
1710         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1711                 md_error(rdev->mddev, rdev);
1712         return 0;
1713 }
1714
1715 static int fix_sync_read_error(struct r1bio *r1_bio)
1716 {
1717         /* Try some synchronous reads of other devices to get
1718          * good data, much like with normal read errors.  Only
1719          * read into the pages we already have so we don't
1720          * need to re-issue the read request.
1721          * We don't need to freeze the array, because being in an
1722          * active sync request, there is no normal IO, and
1723          * no overlapping syncs.
1724          * We don't need to check is_badblock() again as we
1725          * made sure that anything with a bad block in range
1726          * will have bi_end_io clear.
1727          */
1728         struct mddev *mddev = r1_bio->mddev;
1729         struct r1conf *conf = mddev->private;
1730         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1731         sector_t sect = r1_bio->sector;
1732         int sectors = r1_bio->sectors;
1733         int idx = 0;
1734
1735         while(sectors) {
1736                 int s = sectors;
1737                 int d = r1_bio->read_disk;
1738                 int success = 0;
1739                 struct md_rdev *rdev;
1740                 int start;
1741
1742                 if (s > (PAGE_SIZE>>9))
1743                         s = PAGE_SIZE >> 9;
1744                 do {
1745                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1746                                 /* No rcu protection needed here devices
1747                                  * can only be removed when no resync is
1748                                  * active, and resync is currently active
1749                                  */
1750                                 rdev = conf->mirrors[d].rdev;
1751                                 if (sync_page_io(rdev, sect, s<<9,
1752                                                  bio->bi_io_vec[idx].bv_page,
1753                                                  READ, false)) {
1754                                         success = 1;
1755                                         break;
1756                                 }
1757                         }
1758                         d++;
1759                         if (d == conf->raid_disks * 2)
1760                                 d = 0;
1761                 } while (!success && d != r1_bio->read_disk);
1762
1763                 if (!success) {
1764                         char b[BDEVNAME_SIZE];
1765                         int abort = 0;
1766                         /* Cannot read from anywhere, this block is lost.
1767                          * Record a bad block on each device.  If that doesn't
1768                          * work just disable and interrupt the recovery.
1769                          * Don't fail devices as that won't really help.
1770                          */
1771                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1772                                " for block %llu\n",
1773                                mdname(mddev),
1774                                bdevname(bio->bi_bdev, b),
1775                                (unsigned long long)r1_bio->sector);
1776                         for (d = 0; d < conf->raid_disks * 2; d++) {
1777                                 rdev = conf->mirrors[d].rdev;
1778                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1779                                         continue;
1780                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1781                                         abort = 1;
1782                         }
1783                         if (abort) {
1784                                 conf->recovery_disabled =
1785                                         mddev->recovery_disabled;
1786                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1787                                 md_done_sync(mddev, r1_bio->sectors, 0);
1788                                 put_buf(r1_bio);
1789                                 return 0;
1790                         }
1791                         /* Try next page */
1792                         sectors -= s;
1793                         sect += s;
1794                         idx++;
1795                         continue;
1796                 }
1797
1798                 start = d;
1799                 /* write it back and re-read */
1800                 while (d != r1_bio->read_disk) {
1801                         if (d == 0)
1802                                 d = conf->raid_disks * 2;
1803                         d--;
1804                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1805                                 continue;
1806                         rdev = conf->mirrors[d].rdev;
1807                         if (r1_sync_page_io(rdev, sect, s,
1808                                             bio->bi_io_vec[idx].bv_page,
1809                                             WRITE) == 0) {
1810                                 r1_bio->bios[d]->bi_end_io = NULL;
1811                                 rdev_dec_pending(rdev, mddev);
1812                         }
1813                 }
1814                 d = start;
1815                 while (d != r1_bio->read_disk) {
1816                         if (d == 0)
1817                                 d = conf->raid_disks * 2;
1818                         d--;
1819                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1820                                 continue;
1821                         rdev = conf->mirrors[d].rdev;
1822                         if (r1_sync_page_io(rdev, sect, s,
1823                                             bio->bi_io_vec[idx].bv_page,
1824                                             READ) != 0)
1825                                 atomic_add(s, &rdev->corrected_errors);
1826                 }
1827                 sectors -= s;
1828                 sect += s;
1829                 idx ++;
1830         }
1831         set_bit(R1BIO_Uptodate, &r1_bio->state);
1832         set_bit(BIO_UPTODATE, &bio->bi_flags);
1833         return 1;
1834 }
1835
1836 static int process_checks(struct r1bio *r1_bio)
1837 {
1838         /* We have read all readable devices.  If we haven't
1839          * got the block, then there is no hope left.
1840          * If we have, then we want to do a comparison
1841          * and skip the write if everything is the same.
1842          * If any blocks failed to read, then we need to
1843          * attempt an over-write
1844          */
1845         struct mddev *mddev = r1_bio->mddev;
1846         struct r1conf *conf = mddev->private;
1847         int primary;
1848         int i;
1849         int vcnt;
1850
1851         for (primary = 0; primary < conf->raid_disks * 2; primary++)
1852                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1853                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1854                         r1_bio->bios[primary]->bi_end_io = NULL;
1855                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1856                         break;
1857                 }
1858         r1_bio->read_disk = primary;
1859         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1860         for (i = 0; i < conf->raid_disks * 2; i++) {
1861                 int j;
1862                 struct bio *pbio = r1_bio->bios[primary];
1863                 struct bio *sbio = r1_bio->bios[i];
1864                 int size;
1865
1866                 if (sbio->bi_end_io != end_sync_read)
1867                         continue;
1868
1869                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1870                         for (j = vcnt; j-- ; ) {
1871                                 struct page *p, *s;
1872                                 p = pbio->bi_io_vec[j].bv_page;
1873                                 s = sbio->bi_io_vec[j].bv_page;
1874                                 if (memcmp(page_address(p),
1875                                            page_address(s),
1876                                            sbio->bi_io_vec[j].bv_len))
1877                                         break;
1878                         }
1879                 } else
1880                         j = 0;
1881                 if (j >= 0)
1882                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1883                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1884                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1885                         /* No need to write to this device. */
1886                         sbio->bi_end_io = NULL;
1887                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1888                         continue;
1889                 }
1890                 /* fixup the bio for reuse */
1891                 bio_reset(sbio);
1892                 sbio->bi_vcnt = vcnt;
1893                 sbio->bi_size = r1_bio->sectors << 9;
1894                 sbio->bi_sector = r1_bio->sector +
1895                         conf->mirrors[i].rdev->data_offset;
1896                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1897                 sbio->bi_end_io = end_sync_read;
1898                 sbio->bi_private = r1_bio;
1899
1900                 size = sbio->bi_size;
1901                 for (j = 0; j < vcnt ; j++) {
1902                         struct bio_vec *bi;
1903                         bi = &sbio->bi_io_vec[j];
1904                         bi->bv_offset = 0;
1905                         if (size > PAGE_SIZE)
1906                                 bi->bv_len = PAGE_SIZE;
1907                         else
1908                                 bi->bv_len = size;
1909                         size -= PAGE_SIZE;
1910                 }
1911
1912                 bio_copy_data(sbio, pbio);
1913         }
1914         return 0;
1915 }
1916
1917 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1918 {
1919         struct r1conf *conf = mddev->private;
1920         int i;
1921         int disks = conf->raid_disks * 2;
1922         struct bio *bio, *wbio;
1923
1924         bio = r1_bio->bios[r1_bio->read_disk];
1925
1926         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1927                 /* ouch - failed to read all of that. */
1928                 if (!fix_sync_read_error(r1_bio))
1929                         return;
1930
1931         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1932                 if (process_checks(r1_bio) < 0)
1933                         return;
1934         /*
1935          * schedule writes
1936          */
1937         atomic_set(&r1_bio->remaining, 1);
1938         for (i = 0; i < disks ; i++) {
1939                 wbio = r1_bio->bios[i];
1940                 if (wbio->bi_end_io == NULL ||
1941                     (wbio->bi_end_io == end_sync_read &&
1942                      (i == r1_bio->read_disk ||
1943                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1944                         continue;
1945
1946                 wbio->bi_rw = WRITE;
1947                 wbio->bi_end_io = end_sync_write;
1948                 atomic_inc(&r1_bio->remaining);
1949                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
1950
1951                 generic_make_request(wbio);
1952         }
1953
1954         if (atomic_dec_and_test(&r1_bio->remaining)) {
1955                 /* if we're here, all write(s) have completed, so clean up */
1956                 int s = r1_bio->sectors;
1957                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1958                     test_bit(R1BIO_WriteError, &r1_bio->state))
1959                         reschedule_retry(r1_bio);
1960                 else {
1961                         put_buf(r1_bio);
1962                         md_done_sync(mddev, s, 1);
1963                 }
1964         }
1965 }
1966
1967 /*
1968  * This is a kernel thread which:
1969  *
1970  *      1.      Retries failed read operations on working mirrors.
1971  *      2.      Updates the raid superblock when problems encounter.
1972  *      3.      Performs writes following reads for array synchronising.
1973  */
1974
1975 static void fix_read_error(struct r1conf *conf, int read_disk,
1976                            sector_t sect, int sectors)
1977 {
1978         struct mddev *mddev = conf->mddev;
1979         while(sectors) {
1980                 int s = sectors;
1981                 int d = read_disk;
1982                 int success = 0;
1983                 int start;
1984                 struct md_rdev *rdev;
1985
1986                 if (s > (PAGE_SIZE>>9))
1987                         s = PAGE_SIZE >> 9;
1988
1989                 do {
1990                         /* Note: no rcu protection needed here
1991                          * as this is synchronous in the raid1d thread
1992                          * which is the thread that might remove
1993                          * a device.  If raid1d ever becomes multi-threaded....
1994                          */
1995                         sector_t first_bad;
1996                         int bad_sectors;
1997
1998                         rdev = conf->mirrors[d].rdev;
1999                         if (rdev &&
2000                             (test_bit(In_sync, &rdev->flags) ||
2001                              (!test_bit(Faulty, &rdev->flags) &&
2002                               rdev->recovery_offset >= sect + s)) &&
2003                             is_badblock(rdev, sect, s,
2004                                         &first_bad, &bad_sectors) == 0 &&
2005                             sync_page_io(rdev, sect, s<<9,
2006                                          conf->tmppage, READ, false))
2007                                 success = 1;
2008                         else {
2009                                 d++;
2010                                 if (d == conf->raid_disks * 2)
2011                                         d = 0;
2012                         }
2013                 } while (!success && d != read_disk);
2014
2015                 if (!success) {
2016                         /* Cannot read from anywhere - mark it bad */
2017                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2018                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2019                                 md_error(mddev, rdev);
2020                         break;
2021                 }
2022                 /* write it back and re-read */
2023                 start = d;
2024                 while (d != read_disk) {
2025                         if (d==0)
2026                                 d = conf->raid_disks * 2;
2027                         d--;
2028                         rdev = conf->mirrors[d].rdev;
2029                         if (rdev &&
2030                             test_bit(In_sync, &rdev->flags))
2031                                 r1_sync_page_io(rdev, sect, s,
2032                                                 conf->tmppage, WRITE);
2033                 }
2034                 d = start;
2035                 while (d != read_disk) {
2036                         char b[BDEVNAME_SIZE];
2037                         if (d==0)
2038                                 d = conf->raid_disks * 2;
2039                         d--;
2040                         rdev = conf->mirrors[d].rdev;
2041                         if (rdev &&
2042                             test_bit(In_sync, &rdev->flags)) {
2043                                 if (r1_sync_page_io(rdev, sect, s,
2044                                                     conf->tmppage, READ)) {
2045                                         atomic_add(s, &rdev->corrected_errors);
2046                                         printk(KERN_INFO
2047                                                "md/raid1:%s: read error corrected "
2048                                                "(%d sectors at %llu on %s)\n",
2049                                                mdname(mddev), s,
2050                                                (unsigned long long)(sect +
2051                                                    rdev->data_offset),
2052                                                bdevname(rdev->bdev, b));
2053                                 }
2054                         }
2055                 }
2056                 sectors -= s;
2057                 sect += s;
2058         }
2059 }
2060
2061 static int narrow_write_error(struct r1bio *r1_bio, int i)
2062 {
2063         struct mddev *mddev = r1_bio->mddev;
2064         struct r1conf *conf = mddev->private;
2065         struct md_rdev *rdev = conf->mirrors[i].rdev;
2066
2067         /* bio has the data to be written to device 'i' where
2068          * we just recently had a write error.
2069          * We repeatedly clone the bio and trim down to one block,
2070          * then try the write.  Where the write fails we record
2071          * a bad block.
2072          * It is conceivable that the bio doesn't exactly align with
2073          * blocks.  We must handle this somehow.
2074          *
2075          * We currently own a reference on the rdev.
2076          */
2077
2078         int block_sectors;
2079         sector_t sector;
2080         int sectors;
2081         int sect_to_write = r1_bio->sectors;
2082         int ok = 1;
2083
2084         if (rdev->badblocks.shift < 0)
2085                 return 0;
2086
2087         block_sectors = 1 << rdev->badblocks.shift;
2088         sector = r1_bio->sector;
2089         sectors = ((sector + block_sectors)
2090                    & ~(sector_t)(block_sectors - 1))
2091                 - sector;
2092
2093         while (sect_to_write) {
2094                 struct bio *wbio;
2095                 if (sectors > sect_to_write)
2096                         sectors = sect_to_write;
2097                 /* Write at 'sector' for 'sectors'*/
2098
2099                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2100                         unsigned vcnt = r1_bio->behind_page_count;
2101                         struct bio_vec *vec = r1_bio->behind_bvecs;
2102
2103                         while (!vec->bv_page) {
2104                                 vec++;
2105                                 vcnt--;
2106                         }
2107
2108                         wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2109                         memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2110
2111                         wbio->bi_vcnt = vcnt;
2112                 } else {
2113                         wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2114                 }
2115
2116                 wbio->bi_rw = WRITE;
2117                 wbio->bi_sector = r1_bio->sector;
2118                 wbio->bi_size = r1_bio->sectors << 9;
2119
2120                 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2121                 wbio->bi_sector += rdev->data_offset;
2122                 wbio->bi_bdev = rdev->bdev;
2123                 if (submit_bio_wait(WRITE, wbio) == 0)
2124                         /* failure! */
2125                         ok = rdev_set_badblocks(rdev, sector,
2126                                                 sectors, 0)
2127                                 && ok;
2128
2129                 bio_put(wbio);
2130                 sect_to_write -= sectors;
2131                 sector += sectors;
2132                 sectors = block_sectors;
2133         }
2134         return ok;
2135 }
2136
2137 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2138 {
2139         int m;
2140         int s = r1_bio->sectors;
2141         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2142                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2143                 struct bio *bio = r1_bio->bios[m];
2144                 if (bio->bi_end_io == NULL)
2145                         continue;
2146                 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2147                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2148                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2149                 }
2150                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2151                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2152                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2153                                 md_error(conf->mddev, rdev);
2154                 }
2155         }
2156         put_buf(r1_bio);
2157         md_done_sync(conf->mddev, s, 1);
2158 }
2159
2160 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2161 {
2162         int m;
2163         for (m = 0; m < conf->raid_disks * 2 ; m++)
2164                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2165                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2166                         rdev_clear_badblocks(rdev,
2167                                              r1_bio->sector,
2168                                              r1_bio->sectors, 0);
2169                         rdev_dec_pending(rdev, conf->mddev);
2170                 } else if (r1_bio->bios[m] != NULL) {
2171                         /* This drive got a write error.  We need to
2172                          * narrow down and record precise write
2173                          * errors.
2174                          */
2175                         if (!narrow_write_error(r1_bio, m)) {
2176                                 md_error(conf->mddev,
2177                                          conf->mirrors[m].rdev);
2178                                 /* an I/O failed, we can't clear the bitmap */
2179                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2180                         }
2181                         rdev_dec_pending(conf->mirrors[m].rdev,
2182                                          conf->mddev);
2183                 }
2184         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2185                 close_write(r1_bio);
2186         raid_end_bio_io(r1_bio);
2187 }
2188
2189 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2190 {
2191         int disk;
2192         int max_sectors;
2193         struct mddev *mddev = conf->mddev;
2194         struct bio *bio;
2195         char b[BDEVNAME_SIZE];
2196         struct md_rdev *rdev;
2197
2198         clear_bit(R1BIO_ReadError, &r1_bio->state);
2199         /* we got a read error. Maybe the drive is bad.  Maybe just
2200          * the block and we can fix it.
2201          * We freeze all other IO, and try reading the block from
2202          * other devices.  When we find one, we re-write
2203          * and check it that fixes the read error.
2204          * This is all done synchronously while the array is
2205          * frozen
2206          */
2207         if (mddev->ro == 0) {
2208                 freeze_array(conf, 1);
2209                 fix_read_error(conf, r1_bio->read_disk,
2210                                r1_bio->sector, r1_bio->sectors);
2211                 unfreeze_array(conf);
2212         } else
2213                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2214         rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2215
2216         bio = r1_bio->bios[r1_bio->read_disk];
2217         bdevname(bio->bi_bdev, b);
2218 read_more:
2219         disk = read_balance(conf, r1_bio, &max_sectors);
2220         if (disk == -1) {
2221                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2222                        " read error for block %llu\n",
2223                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2224                 raid_end_bio_io(r1_bio);
2225         } else {
2226                 const unsigned long do_sync
2227                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2228                 if (bio) {
2229                         r1_bio->bios[r1_bio->read_disk] =
2230                                 mddev->ro ? IO_BLOCKED : NULL;
2231                         bio_put(bio);
2232                 }
2233                 r1_bio->read_disk = disk;
2234                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2235                 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2236                 r1_bio->bios[r1_bio->read_disk] = bio;
2237                 rdev = conf->mirrors[disk].rdev;
2238                 printk_ratelimited(KERN_ERR
2239                                    "md/raid1:%s: redirecting sector %llu"
2240                                    " to other mirror: %s\n",
2241                                    mdname(mddev),
2242                                    (unsigned long long)r1_bio->sector,
2243                                    bdevname(rdev->bdev, b));
2244                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2245                 bio->bi_bdev = rdev->bdev;
2246                 bio->bi_end_io = raid1_end_read_request;
2247                 bio->bi_rw = READ | do_sync;
2248                 bio->bi_private = r1_bio;
2249                 if (max_sectors < r1_bio->sectors) {
2250                         /* Drat - have to split this up more */
2251                         struct bio *mbio = r1_bio->master_bio;
2252                         int sectors_handled = (r1_bio->sector + max_sectors
2253                                                - mbio->bi_sector);
2254                         r1_bio->sectors = max_sectors;
2255                         spin_lock_irq(&conf->device_lock);
2256                         if (mbio->bi_phys_segments == 0)
2257                                 mbio->bi_phys_segments = 2;
2258                         else
2259                                 mbio->bi_phys_segments++;
2260                         spin_unlock_irq(&conf->device_lock);
2261                         generic_make_request(bio);
2262                         bio = NULL;
2263
2264                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2265
2266                         r1_bio->master_bio = mbio;
2267                         r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2268                         r1_bio->state = 0;
2269                         set_bit(R1BIO_ReadError, &r1_bio->state);
2270                         r1_bio->mddev = mddev;
2271                         r1_bio->sector = mbio->bi_sector + sectors_handled;
2272
2273                         goto read_more;
2274                 } else
2275                         generic_make_request(bio);
2276         }
2277 }
2278
2279 static void raid1d(struct md_thread *thread)
2280 {
2281         struct mddev *mddev = thread->mddev;
2282         struct r1bio *r1_bio;
2283         unsigned long flags;
2284         struct r1conf *conf = mddev->private;
2285         struct list_head *head = &conf->retry_list;
2286         struct blk_plug plug;
2287
2288         md_check_recovery(mddev);
2289
2290         blk_start_plug(&plug);
2291         for (;;) {
2292
2293                 flush_pending_writes(conf);
2294
2295                 spin_lock_irqsave(&conf->device_lock, flags);
2296                 if (list_empty(head)) {
2297                         spin_unlock_irqrestore(&conf->device_lock, flags);
2298                         break;
2299                 }
2300                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2301                 list_del(head->prev);
2302                 conf->nr_queued--;
2303                 spin_unlock_irqrestore(&conf->device_lock, flags);
2304
2305                 mddev = r1_bio->mddev;
2306                 conf = mddev->private;
2307                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2308                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2309                             test_bit(R1BIO_WriteError, &r1_bio->state))
2310                                 handle_sync_write_finished(conf, r1_bio);
2311                         else
2312                                 sync_request_write(mddev, r1_bio);
2313                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2314                            test_bit(R1BIO_WriteError, &r1_bio->state))
2315                         handle_write_finished(conf, r1_bio);
2316                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2317                         handle_read_error(conf, r1_bio);
2318                 else
2319                         /* just a partial read to be scheduled from separate
2320                          * context
2321                          */
2322                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2323
2324                 cond_resched();
2325                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2326                         md_check_recovery(mddev);
2327         }
2328         blk_finish_plug(&plug);
2329 }
2330
2331
2332 static int init_resync(struct r1conf *conf)
2333 {
2334         int buffs;
2335
2336         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2337         BUG_ON(conf->r1buf_pool);
2338         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2339                                           conf->poolinfo);
2340         if (!conf->r1buf_pool)
2341                 return -ENOMEM;
2342         conf->next_resync = 0;
2343         return 0;
2344 }
2345
2346 /*
2347  * perform a "sync" on one "block"
2348  *
2349  * We need to make sure that no normal I/O request - particularly write
2350  * requests - conflict with active sync requests.
2351  *
2352  * This is achieved by tracking pending requests and a 'barrier' concept
2353  * that can be installed to exclude normal IO requests.
2354  */
2355
2356 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2357 {
2358         struct r1conf *conf = mddev->private;
2359         struct r1bio *r1_bio;
2360         struct bio *bio;
2361         sector_t max_sector, nr_sectors;
2362         int disk = -1;
2363         int i;
2364         int wonly = -1;
2365         int write_targets = 0, read_targets = 0;
2366         sector_t sync_blocks;
2367         int still_degraded = 0;
2368         int good_sectors = RESYNC_SECTORS;
2369         int min_bad = 0; /* number of sectors that are bad in all devices */
2370
2371         if (!conf->r1buf_pool)
2372                 if (init_resync(conf))
2373                         return 0;
2374
2375         max_sector = mddev->dev_sectors;
2376         if (sector_nr >= max_sector) {
2377                 /* If we aborted, we need to abort the
2378                  * sync on the 'current' bitmap chunk (there will
2379                  * only be one in raid1 resync.
2380                  * We can find the current addess in mddev->curr_resync
2381                  */
2382                 if (mddev->curr_resync < max_sector) /* aborted */
2383                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2384                                                 &sync_blocks, 1);
2385                 else /* completed sync */
2386                         conf->fullsync = 0;
2387
2388                 bitmap_close_sync(mddev->bitmap);
2389                 close_sync(conf);
2390                 return 0;
2391         }
2392
2393         if (mddev->bitmap == NULL &&
2394             mddev->recovery_cp == MaxSector &&
2395             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2396             conf->fullsync == 0) {
2397                 *skipped = 1;
2398                 return max_sector - sector_nr;
2399         }
2400         /* before building a request, check if we can skip these blocks..
2401          * This call the bitmap_start_sync doesn't actually record anything
2402          */
2403         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2404             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2405                 /* We can skip this block, and probably several more */
2406                 *skipped = 1;
2407                 return sync_blocks;
2408         }
2409         /*
2410          * If there is non-resync activity waiting for a turn,
2411          * and resync is going fast enough,
2412          * then let it though before starting on this new sync request.
2413          */
2414         if (!go_faster && conf->nr_waiting)
2415                 msleep_interruptible(1000);
2416
2417         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2418         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2419         raise_barrier(conf);
2420
2421         conf->next_resync = sector_nr;
2422
2423         rcu_read_lock();
2424         /*
2425          * If we get a correctably read error during resync or recovery,
2426          * we might want to read from a different device.  So we
2427          * flag all drives that could conceivably be read from for READ,
2428          * and any others (which will be non-In_sync devices) for WRITE.
2429          * If a read fails, we try reading from something else for which READ
2430          * is OK.
2431          */
2432
2433         r1_bio->mddev = mddev;
2434         r1_bio->sector = sector_nr;
2435         r1_bio->state = 0;
2436         set_bit(R1BIO_IsSync, &r1_bio->state);
2437
2438         for (i = 0; i < conf->raid_disks * 2; i++) {
2439                 struct md_rdev *rdev;
2440                 bio = r1_bio->bios[i];
2441                 bio_reset(bio);
2442
2443                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2444                 if (rdev == NULL ||
2445                     test_bit(Faulty, &rdev->flags)) {
2446                         if (i < conf->raid_disks)
2447                                 still_degraded = 1;
2448                 } else if (!test_bit(In_sync, &rdev->flags)) {
2449                         bio->bi_rw = WRITE;
2450                         bio->bi_end_io = end_sync_write;
2451                         write_targets ++;
2452                 } else {
2453                         /* may need to read from here */
2454                         sector_t first_bad = MaxSector;
2455                         int bad_sectors;
2456
2457                         if (is_badblock(rdev, sector_nr, good_sectors,
2458                                         &first_bad, &bad_sectors)) {
2459                                 if (first_bad > sector_nr)
2460                                         good_sectors = first_bad - sector_nr;
2461                                 else {
2462                                         bad_sectors -= (sector_nr - first_bad);
2463                                         if (min_bad == 0 ||
2464                                             min_bad > bad_sectors)
2465                                                 min_bad = bad_sectors;
2466                                 }
2467                         }
2468                         if (sector_nr < first_bad) {
2469                                 if (test_bit(WriteMostly, &rdev->flags)) {
2470                                         if (wonly < 0)
2471                                                 wonly = i;
2472                                 } else {
2473                                         if (disk < 0)
2474                                                 disk = i;
2475                                 }
2476                                 bio->bi_rw = READ;
2477                                 bio->bi_end_io = end_sync_read;
2478                                 read_targets++;
2479                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2480                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2481                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2482                                 /*
2483                                  * The device is suitable for reading (InSync),
2484                                  * but has bad block(s) here. Let's try to correct them,
2485                                  * if we are doing resync or repair. Otherwise, leave
2486                                  * this device alone for this sync request.
2487                                  */
2488                                 bio->bi_rw = WRITE;
2489                                 bio->bi_end_io = end_sync_write;
2490                                 write_targets++;
2491                         }
2492                 }
2493                 if (bio->bi_end_io) {
2494                         atomic_inc(&rdev->nr_pending);
2495                         bio->bi_sector = sector_nr + rdev->data_offset;
2496                         bio->bi_bdev = rdev->bdev;
2497                         bio->bi_private = r1_bio;
2498                 }
2499         }
2500         rcu_read_unlock();
2501         if (disk < 0)
2502                 disk = wonly;
2503         r1_bio->read_disk = disk;
2504
2505         if (read_targets == 0 && min_bad > 0) {
2506                 /* These sectors are bad on all InSync devices, so we
2507                  * need to mark them bad on all write targets
2508                  */
2509                 int ok = 1;
2510                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2511                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2512                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2513                                 ok = rdev_set_badblocks(rdev, sector_nr,
2514                                                         min_bad, 0
2515                                         ) && ok;
2516                         }
2517                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2518                 *skipped = 1;
2519                 put_buf(r1_bio);
2520
2521                 if (!ok) {
2522                         /* Cannot record the badblocks, so need to
2523                          * abort the resync.
2524                          * If there are multiple read targets, could just
2525                          * fail the really bad ones ???
2526                          */
2527                         conf->recovery_disabled = mddev->recovery_disabled;
2528                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2529                         return 0;
2530                 } else
2531                         return min_bad;
2532
2533         }
2534         if (min_bad > 0 && min_bad < good_sectors) {
2535                 /* only resync enough to reach the next bad->good
2536                  * transition */
2537                 good_sectors = min_bad;
2538         }
2539
2540         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2541                 /* extra read targets are also write targets */
2542                 write_targets += read_targets-1;
2543
2544         if (write_targets == 0 || read_targets == 0) {
2545                 /* There is nowhere to write, so all non-sync
2546                  * drives must be failed - so we are finished
2547                  */
2548                 sector_t rv;
2549                 if (min_bad > 0)
2550                         max_sector = sector_nr + min_bad;
2551                 rv = max_sector - sector_nr;
2552                 *skipped = 1;
2553                 put_buf(r1_bio);
2554                 return rv;
2555         }
2556
2557         if (max_sector > mddev->resync_max)
2558                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2559         if (max_sector > sector_nr + good_sectors)
2560                 max_sector = sector_nr + good_sectors;
2561         nr_sectors = 0;
2562         sync_blocks = 0;
2563         do {
2564                 struct page *page;
2565                 int len = PAGE_SIZE;
2566                 if (sector_nr + (len>>9) > max_sector)
2567                         len = (max_sector - sector_nr) << 9;
2568                 if (len == 0)
2569                         break;
2570                 if (sync_blocks == 0) {
2571                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2572                                                &sync_blocks, still_degraded) &&
2573                             !conf->fullsync &&
2574                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2575                                 break;
2576                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2577                         if ((len >> 9) > sync_blocks)
2578                                 len = sync_blocks<<9;
2579                 }
2580
2581                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2582                         bio = r1_bio->bios[i];
2583                         if (bio->bi_end_io) {
2584                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2585                                 if (bio_add_page(bio, page, len, 0) == 0) {
2586                                         /* stop here */
2587                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2588                                         while (i > 0) {
2589                                                 i--;
2590                                                 bio = r1_bio->bios[i];
2591                                                 if (bio->bi_end_io==NULL)
2592                                                         continue;
2593                                                 /* remove last page from this bio */
2594                                                 bio->bi_vcnt--;
2595                                                 bio->bi_size -= len;
2596                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2597                                         }
2598                                         goto bio_full;
2599                                 }
2600                         }
2601                 }
2602                 nr_sectors += len>>9;
2603                 sector_nr += len>>9;
2604                 sync_blocks -= (len>>9);
2605         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2606  bio_full:
2607         r1_bio->sectors = nr_sectors;
2608
2609         /* For a user-requested sync, we read all readable devices and do a
2610          * compare
2611          */
2612         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2613                 atomic_set(&r1_bio->remaining, read_targets);
2614                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2615                         bio = r1_bio->bios[i];
2616                         if (bio->bi_end_io == end_sync_read) {
2617                                 read_targets--;
2618                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2619                                 generic_make_request(bio);
2620                         }
2621                 }
2622         } else {
2623                 atomic_set(&r1_bio->remaining, 1);
2624                 bio = r1_bio->bios[r1_bio->read_disk];
2625                 md_sync_acct(bio->bi_bdev, nr_sectors);
2626                 generic_make_request(bio);
2627
2628         }
2629         return nr_sectors;
2630 }
2631
2632 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2633 {
2634         if (sectors)
2635                 return sectors;
2636
2637         return mddev->dev_sectors;
2638 }
2639
2640 static struct r1conf *setup_conf(struct mddev *mddev)
2641 {
2642         struct r1conf *conf;
2643         int i;
2644         struct raid1_info *disk;
2645         struct md_rdev *rdev;
2646         int err = -ENOMEM;
2647
2648         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2649         if (!conf)
2650                 goto abort;
2651
2652         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2653                                 * mddev->raid_disks * 2,
2654                                  GFP_KERNEL);
2655         if (!conf->mirrors)
2656                 goto abort;
2657
2658         conf->tmppage = alloc_page(GFP_KERNEL);
2659         if (!conf->tmppage)
2660                 goto abort;
2661
2662         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2663         if (!conf->poolinfo)
2664                 goto abort;
2665         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2666         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2667                                           r1bio_pool_free,
2668                                           conf->poolinfo);
2669         if (!conf->r1bio_pool)
2670                 goto abort;
2671
2672         conf->poolinfo->mddev = mddev;
2673
2674         err = -EINVAL;
2675         spin_lock_init(&conf->device_lock);
2676         rdev_for_each(rdev, mddev) {
2677                 struct request_queue *q;
2678                 int disk_idx = rdev->raid_disk;
2679                 if (disk_idx >= mddev->raid_disks
2680                     || disk_idx < 0)
2681                         continue;
2682                 if (test_bit(Replacement, &rdev->flags))
2683                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
2684                 else
2685                         disk = conf->mirrors + disk_idx;
2686
2687                 if (disk->rdev)
2688                         goto abort;
2689                 disk->rdev = rdev;
2690                 q = bdev_get_queue(rdev->bdev);
2691                 if (q->merge_bvec_fn)
2692                         mddev->merge_check_needed = 1;
2693
2694                 disk->head_position = 0;
2695                 disk->seq_start = MaxSector;
2696         }
2697         conf->raid_disks = mddev->raid_disks;
2698         conf->mddev = mddev;
2699         INIT_LIST_HEAD(&conf->retry_list);
2700
2701         spin_lock_init(&conf->resync_lock);
2702         init_waitqueue_head(&conf->wait_barrier);
2703
2704         bio_list_init(&conf->pending_bio_list);
2705         conf->pending_count = 0;
2706         conf->recovery_disabled = mddev->recovery_disabled - 1;
2707
2708         err = -EIO;
2709         for (i = 0; i < conf->raid_disks * 2; i++) {
2710
2711                 disk = conf->mirrors + i;
2712
2713                 if (i < conf->raid_disks &&
2714                     disk[conf->raid_disks].rdev) {
2715                         /* This slot has a replacement. */
2716                         if (!disk->rdev) {
2717                                 /* No original, just make the replacement
2718                                  * a recovering spare
2719                                  */
2720                                 disk->rdev =
2721                                         disk[conf->raid_disks].rdev;
2722                                 disk[conf->raid_disks].rdev = NULL;
2723                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2724                                 /* Original is not in_sync - bad */
2725                                 goto abort;
2726                 }
2727
2728                 if (!disk->rdev ||
2729                     !test_bit(In_sync, &disk->rdev->flags)) {
2730                         disk->head_position = 0;
2731                         if (disk->rdev &&
2732                             (disk->rdev->saved_raid_disk < 0))
2733                                 conf->fullsync = 1;
2734                 }
2735         }
2736
2737         err = -ENOMEM;
2738         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2739         if (!conf->thread) {
2740                 printk(KERN_ERR
2741                        "md/raid1:%s: couldn't allocate thread\n",
2742                        mdname(mddev));
2743                 goto abort;
2744         }
2745
2746         return conf;
2747
2748  abort:
2749         if (conf) {
2750                 if (conf->r1bio_pool)
2751                         mempool_destroy(conf->r1bio_pool);
2752                 kfree(conf->mirrors);
2753                 safe_put_page(conf->tmppage);
2754                 kfree(conf->poolinfo);
2755                 kfree(conf);
2756         }
2757         return ERR_PTR(err);
2758 }
2759
2760 static int stop(struct mddev *mddev);
2761 static int run(struct mddev *mddev)
2762 {
2763         struct r1conf *conf;
2764         int i;
2765         struct md_rdev *rdev;
2766         int ret;
2767         bool discard_supported = false;
2768
2769         if (mddev->level != 1) {
2770                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2771                        mdname(mddev), mddev->level);
2772                 return -EIO;
2773         }
2774         if (mddev->reshape_position != MaxSector) {
2775                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2776                        mdname(mddev));
2777                 return -EIO;
2778         }
2779         /*
2780          * copy the already verified devices into our private RAID1
2781          * bookkeeping area. [whatever we allocate in run(),
2782          * should be freed in stop()]
2783          */
2784         if (mddev->private == NULL)
2785                 conf = setup_conf(mddev);
2786         else
2787                 conf = mddev->private;
2788
2789         if (IS_ERR(conf))
2790                 return PTR_ERR(conf);
2791
2792         if (mddev->queue)
2793                 blk_queue_max_write_same_sectors(mddev->queue, 0);
2794
2795         rdev_for_each(rdev, mddev) {
2796                 if (!mddev->gendisk)
2797                         continue;
2798                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2799                                   rdev->data_offset << 9);
2800                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2801                         discard_supported = true;
2802         }
2803
2804         mddev->degraded = 0;
2805         for (i=0; i < conf->raid_disks; i++)
2806                 if (conf->mirrors[i].rdev == NULL ||
2807                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2808                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2809                         mddev->degraded++;
2810
2811         if (conf->raid_disks - mddev->degraded == 1)
2812                 mddev->recovery_cp = MaxSector;
2813
2814         if (mddev->recovery_cp != MaxSector)
2815                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2816                        " -- starting background reconstruction\n",
2817                        mdname(mddev));
2818         printk(KERN_INFO
2819                 "md/raid1:%s: active with %d out of %d mirrors\n",
2820                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2821                 mddev->raid_disks);
2822
2823         /*
2824          * Ok, everything is just fine now
2825          */
2826         mddev->thread = conf->thread;
2827         conf->thread = NULL;
2828         mddev->private = conf;
2829
2830         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2831
2832         if (mddev->queue) {
2833                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2834                 mddev->queue->backing_dev_info.congested_data = mddev;
2835                 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2836
2837                 if (discard_supported)
2838                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2839                                                 mddev->queue);
2840                 else
2841                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2842                                                   mddev->queue);
2843         }
2844
2845         ret =  md_integrity_register(mddev);
2846         if (ret)
2847                 stop(mddev);
2848         return ret;
2849 }
2850
2851 static int stop(struct mddev *mddev)
2852 {
2853         struct r1conf *conf = mddev->private;
2854         struct bitmap *bitmap = mddev->bitmap;
2855
2856         /* wait for behind writes to complete */
2857         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2858                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2859                        mdname(mddev));
2860                 /* need to kick something here to make sure I/O goes? */
2861                 wait_event(bitmap->behind_wait,
2862                            atomic_read(&bitmap->behind_writes) == 0);
2863         }
2864
2865         raise_barrier(conf);
2866         lower_barrier(conf);
2867
2868         md_unregister_thread(&mddev->thread);
2869         if (conf->r1bio_pool)
2870                 mempool_destroy(conf->r1bio_pool);
2871         kfree(conf->mirrors);
2872         safe_put_page(conf->tmppage);
2873         kfree(conf->poolinfo);
2874         kfree(conf);
2875         mddev->private = NULL;
2876         return 0;
2877 }
2878
2879 static int raid1_resize(struct mddev *mddev, sector_t sectors)
2880 {
2881         /* no resync is happening, and there is enough space
2882          * on all devices, so we can resize.
2883          * We need to make sure resync covers any new space.
2884          * If the array is shrinking we should possibly wait until
2885          * any io in the removed space completes, but it hardly seems
2886          * worth it.
2887          */
2888         sector_t newsize = raid1_size(mddev, sectors, 0);
2889         if (mddev->external_size &&
2890             mddev->array_sectors > newsize)
2891                 return -EINVAL;
2892         if (mddev->bitmap) {
2893                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2894                 if (ret)
2895                         return ret;
2896         }
2897         md_set_array_sectors(mddev, newsize);
2898         set_capacity(mddev->gendisk, mddev->array_sectors);
2899         revalidate_disk(mddev->gendisk);
2900         if (sectors > mddev->dev_sectors &&
2901             mddev->recovery_cp > mddev->dev_sectors) {
2902                 mddev->recovery_cp = mddev->dev_sectors;
2903                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2904         }
2905         mddev->dev_sectors = sectors;
2906         mddev->resync_max_sectors = sectors;
2907         return 0;
2908 }
2909
2910 static int raid1_reshape(struct mddev *mddev)
2911 {
2912         /* We need to:
2913          * 1/ resize the r1bio_pool
2914          * 2/ resize conf->mirrors
2915          *
2916          * We allocate a new r1bio_pool if we can.
2917          * Then raise a device barrier and wait until all IO stops.
2918          * Then resize conf->mirrors and swap in the new r1bio pool.
2919          *
2920          * At the same time, we "pack" the devices so that all the missing
2921          * devices have the higher raid_disk numbers.
2922          */
2923         mempool_t *newpool, *oldpool;
2924         struct pool_info *newpoolinfo;
2925         struct raid1_info *newmirrors;
2926         struct r1conf *conf = mddev->private;
2927         int cnt, raid_disks;
2928         unsigned long flags;
2929         int d, d2, err;
2930
2931         /* Cannot change chunk_size, layout, or level */
2932         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2933             mddev->layout != mddev->new_layout ||
2934             mddev->level != mddev->new_level) {
2935                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2936                 mddev->new_layout = mddev->layout;
2937                 mddev->new_level = mddev->level;
2938                 return -EINVAL;
2939         }
2940
2941         err = md_allow_write(mddev);
2942         if (err)
2943                 return err;
2944
2945         raid_disks = mddev->raid_disks + mddev->delta_disks;
2946
2947         if (raid_disks < conf->raid_disks) {
2948                 cnt=0;
2949                 for (d= 0; d < conf->raid_disks; d++)
2950                         if (conf->mirrors[d].rdev)
2951                                 cnt++;
2952                 if (cnt > raid_disks)
2953                         return -EBUSY;
2954         }
2955
2956         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2957         if (!newpoolinfo)
2958                 return -ENOMEM;
2959         newpoolinfo->mddev = mddev;
2960         newpoolinfo->raid_disks = raid_disks * 2;
2961
2962         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2963                                  r1bio_pool_free, newpoolinfo);
2964         if (!newpool) {
2965                 kfree(newpoolinfo);
2966                 return -ENOMEM;
2967         }
2968         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
2969                              GFP_KERNEL);
2970         if (!newmirrors) {
2971                 kfree(newpoolinfo);
2972                 mempool_destroy(newpool);
2973                 return -ENOMEM;
2974         }
2975
2976         freeze_array(conf, 0);
2977
2978         /* ok, everything is stopped */
2979         oldpool = conf->r1bio_pool;
2980         conf->r1bio_pool = newpool;
2981
2982         for (d = d2 = 0; d < conf->raid_disks; d++) {
2983                 struct md_rdev *rdev = conf->mirrors[d].rdev;
2984                 if (rdev && rdev->raid_disk != d2) {
2985                         sysfs_unlink_rdev(mddev, rdev);
2986                         rdev->raid_disk = d2;
2987                         sysfs_unlink_rdev(mddev, rdev);
2988                         if (sysfs_link_rdev(mddev, rdev))
2989                                 printk(KERN_WARNING
2990                                        "md/raid1:%s: cannot register rd%d\n",
2991                                        mdname(mddev), rdev->raid_disk);
2992                 }
2993                 if (rdev)
2994                         newmirrors[d2++].rdev = rdev;
2995         }
2996         kfree(conf->mirrors);
2997         conf->mirrors = newmirrors;
2998         kfree(conf->poolinfo);
2999         conf->poolinfo = newpoolinfo;
3000
3001         spin_lock_irqsave(&conf->device_lock, flags);
3002         mddev->degraded += (raid_disks - conf->raid_disks);
3003         spin_unlock_irqrestore(&conf->device_lock, flags);
3004         conf->raid_disks = mddev->raid_disks = raid_disks;
3005         mddev->delta_disks = 0;
3006
3007         unfreeze_array(conf);
3008
3009         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3010         md_wakeup_thread(mddev->thread);
3011
3012         mempool_destroy(oldpool);
3013         return 0;
3014 }
3015
3016 static void raid1_quiesce(struct mddev *mddev, int state)
3017 {
3018         struct r1conf *conf = mddev->private;
3019
3020         switch(state) {
3021         case 2: /* wake for suspend */
3022                 wake_up(&conf->wait_barrier);
3023                 break;
3024         case 1:
3025                 raise_barrier(conf);
3026                 break;
3027         case 0:
3028                 lower_barrier(conf);
3029                 break;
3030         }
3031 }
3032
3033 static void *raid1_takeover(struct mddev *mddev)
3034 {
3035         /* raid1 can take over:
3036          *  raid5 with 2 devices, any layout or chunk size
3037          */
3038         if (mddev->level == 5 && mddev->raid_disks == 2) {
3039                 struct r1conf *conf;
3040                 mddev->new_level = 1;
3041                 mddev->new_layout = 0;
3042                 mddev->new_chunk_sectors = 0;
3043                 conf = setup_conf(mddev);
3044                 if (!IS_ERR(conf))
3045                         conf->barrier = 1;
3046                 return conf;
3047         }
3048         return ERR_PTR(-EINVAL);
3049 }
3050
3051 static struct md_personality raid1_personality =
3052 {
3053         .name           = "raid1",
3054         .level          = 1,
3055         .owner          = THIS_MODULE,
3056         .make_request   = make_request,
3057         .run            = run,
3058         .stop           = stop,
3059         .status         = status,
3060         .error_handler  = error,
3061         .hot_add_disk   = raid1_add_disk,
3062         .hot_remove_disk= raid1_remove_disk,
3063         .spare_active   = raid1_spare_active,
3064         .sync_request   = sync_request,
3065         .resize         = raid1_resize,
3066         .size           = raid1_size,
3067         .check_reshape  = raid1_reshape,
3068         .quiesce        = raid1_quiesce,
3069         .takeover       = raid1_takeover,
3070 };
3071
3072 static int __init raid_init(void)
3073 {
3074         return register_md_personality(&raid1_personality);
3075 }
3076
3077 static void raid_exit(void)
3078 {
3079         unregister_md_personality(&raid1_personality);
3080 }
3081
3082 module_init(raid_init);
3083 module_exit(raid_exit);
3084 MODULE_LICENSE("GPL");
3085 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3086 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3087 MODULE_ALIAS("md-raid1");
3088 MODULE_ALIAS("md-level-1");
3089
3090 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);