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