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