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