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