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