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