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/seq_file.h>
  38 #include "md.h"
  39 #include "raid1.h"
  40 #include "bitmap.h"
  41
  42 #define DEBUG 0
  43 #if DEBUG
  44 #define PRINTK(x...) printk(x)
  45 #else
  46 #define PRINTK(x...)
  47 #endif
  48
  49 /*
  50  * Number of guaranteed r1bios in case of extreme VM load:
  51  */
  52 #define NR_RAID1_BIOS 256
  53
  54
  55 static void allow_barrier(conf_t *conf);
  56 static void lower_barrier(conf_t *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(r1bio_t, 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         r1bio_t *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         r1bio_t *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(conf_t *conf, r1bio_t *r1_bio)
 164 {
 165         int i;
 166
 167         for (i = 0; i < conf->raid_disks; i++) {
 168                 struct bio **bio = r1_bio->bios + i;
 169                 if (*bio && *bio != IO_BLOCKED)
 170                         bio_put(*bio);
 171                 *bio = NULL;
 172         }
 173 }
 174
 175 static void free_r1bio(r1bio_t *r1_bio)
 176 {
 177         conf_t *conf = r1_bio->mddev->private;
 178
 179         /*
 180          * Wake up any possible resync thread that waits for the device
 181          * to go idle.
 182          */
 183         allow_barrier(conf);
 184
 185         put_all_bios(conf, r1_bio);
 186         mempool_free(r1_bio, conf->r1bio_pool);
 187 }
 188
 189 static void put_buf(r1bio_t *r1_bio)
 190 {
 191         conf_t *conf = r1_bio->mddev->private;
 192         int i;
 193
 194         for (i=0; i<conf->raid_disks; i++) {
 195                 struct bio *bio = r1_bio->bios[i];
 196                 if (bio->bi_end_io)
 197                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 198         }
 199
 200         mempool_free(r1_bio, conf->r1buf_pool);
 201
 202         lower_barrier(conf);
 203 }
 204
 205 static void reschedule_retry(r1bio_t *r1_bio)
 206 {
 207         unsigned long flags;
 208         mddev_t *mddev = r1_bio->mddev;
 209         conf_t *conf = mddev->private;
 210
 211         spin_lock_irqsave(&conf->device_lock, flags);
 212         list_add(&r1_bio->retry_list, &conf->retry_list);
 213         conf->nr_queued ++;
 214         spin_unlock_irqrestore(&conf->device_lock, flags);
 215
 216         wake_up(&conf->wait_barrier);
 217         md_wakeup_thread(mddev->thread);
 218 }
 219
 220 /*
 221  * raid_end_bio_io() is called when we have finished servicing a mirrored
 222  * operation and are ready to return a success/failure code to the buffer
 223  * cache layer.
 224  */
 225 static void raid_end_bio_io(r1bio_t *r1_bio)
 226 {
 227         struct bio *bio = r1_bio->master_bio;
 228
 229         /* if nobody has done the final endio yet, do it now */
 230         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 231                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 232                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 233                         (unsigned long long) bio->bi_sector,
 234                         (unsigned long long) bio->bi_sector +
 235                                 (bio->bi_size >> 9) - 1);
 236
 237                 bio_endio(bio,
 238                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 239         }
 240         free_r1bio(r1_bio);
 241 }
 242
 243 /*
 244  * Update disk head position estimator based on IRQ completion info.
 245  */
 246 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 247 {
 248         conf_t *conf = r1_bio->mddev->private;
 249
 250         conf->mirrors[disk].head_position =
 251                 r1_bio->sector + (r1_bio->sectors);
 252 }
 253
 254 static void raid1_end_read_request(struct bio *bio, int error)
 255 {
 256         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 257         r1bio_t *r1_bio = bio->bi_private;
 258         int mirror;
 259         conf_t *conf = r1_bio->mddev->private;
 260
 261         mirror = r1_bio->read_disk;
 262         /*
 263          * this branch is our 'one mirror IO has finished' event handler:
 264          */
 265         update_head_pos(mirror, r1_bio);
 266
 267         if (uptodate)
 268                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 269         else {
 270                 /* If all other devices have failed, we want to return
 271                  * the error upwards rather than fail the last device.
 272                  * Here we redefine "uptodate" to mean "Don't want to retry"
 273                  */
 274                 unsigned long flags;
 275                 spin_lock_irqsave(&conf->device_lock, flags);
 276                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 277                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 278                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 279                         uptodate = 1;
 280                 spin_unlock_irqrestore(&conf->device_lock, flags);
 281         }
 282
 283         if (uptodate)
 284                 raid_end_bio_io(r1_bio);
 285         else {
 286                 /*
 287                  * oops, read error:
 288                  */
 289                 char b[BDEVNAME_SIZE];
 290                 if (printk_ratelimit())
 291                         printk(KERN_ERR "md/raid1:%s: %s: rescheduling sector %llu\n",
 292                                mdname(conf->mddev),
 293                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 294                 reschedule_retry(r1_bio);
 295         }
 296
 297         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 298 }
 299
 300 static void r1_bio_write_done(r1bio_t *r1_bio)
 301 {
 302         if (atomic_dec_and_test(&r1_bio->remaining))
 303         {
 304                 /* it really is the end of this request */
 305                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 306                         /* free extra copy of the data pages */
 307                         int i = r1_bio->behind_page_count;
 308                         while (i--)
 309                                 safe_put_page(r1_bio->behind_pages[i]);
 310                         kfree(r1_bio->behind_pages);
 311                         r1_bio->behind_pages = NULL;
 312                 }
 313                 /* clear the bitmap if all writes complete successfully */
 314                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 315                                 r1_bio->sectors,
 316                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
 317                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 318                 md_write_end(r1_bio->mddev);
 319                 raid_end_bio_io(r1_bio);
 320         }
 321 }
 322
 323 static void raid1_end_write_request(struct bio *bio, int error)
 324 {
 325         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 326         r1bio_t *r1_bio = bio->bi_private;
 327         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 328         conf_t *conf = r1_bio->mddev->private;
 329         struct bio *to_put = NULL;
 330
 331
 332         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 333                 if (r1_bio->bios[mirror] == bio)
 334                         break;
 335
 336         /*
 337          * 'one mirror IO has finished' event handler:
 338          */
 339         r1_bio->bios[mirror] = NULL;
 340         to_put = bio;
 341         if (!uptodate) {
 342                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 343                 /* an I/O failed, we can't clear the bitmap */
 344                 set_bit(R1BIO_Degraded, &r1_bio->state);
 345         } else
 346                 /*
 347                  * Set R1BIO_Uptodate in our master bio, so that we
 348                  * will return a good error code for to the higher
 349                  * levels even if IO on some other mirrored buffer
 350                  * fails.
 351                  *
 352                  * The 'master' represents the composite IO operation
 353                  * to user-side. So if something waits for IO, then it
 354                  * will wait for the 'master' bio.
 355                  */
 356                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 357
 358         update_head_pos(mirror, r1_bio);
 359
 360         if (behind) {
 361                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 362                         atomic_dec(&r1_bio->behind_remaining);
 363
 364                 /*
 365                  * In behind mode, we ACK the master bio once the I/O
 366                  * has safely reached all non-writemostly
 367                  * disks. Setting the Returned bit ensures that this
 368                  * gets done only once -- we don't ever want to return
 369                  * -EIO here, instead we'll wait
 370                  */
 371                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 372                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 373                         /* Maybe we can return now */
 374                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 375                                 struct bio *mbio = r1_bio->master_bio;
 376                                 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 377                                        (unsigned long long) mbio->bi_sector,
 378                                        (unsigned long long) mbio->bi_sector +
 379                                        (mbio->bi_size >> 9) - 1);
 380                                 bio_endio(mbio, 0);
 381                         }
 382                 }
 383         }
 384         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 385
 386         /*
 387          * Let's see if all mirrored write operations have finished
 388          * already.
 389          */
 390         r1_bio_write_done(r1_bio);
 391
 392         if (to_put)
 393                 bio_put(to_put);
 394 }
 395
 396
 397 /*
 398  * This routine returns the disk from which the requested read should
 399  * be done. There is a per-array 'next expected sequential IO' sector
 400  * number - if this matches on the next IO then we use the last disk.
 401  * There is also a per-disk 'last know head position' sector that is
 402  * maintained from IRQ contexts, both the normal and the resync IO
 403  * completion handlers update this position correctly. If there is no
 404  * perfect sequential match then we pick the disk whose head is closest.
 405  *
 406  * If there are 2 mirrors in the same 2 devices, performance degrades
 407  * because position is mirror, not device based.
 408  *
 409  * The rdev for the device selected will have nr_pending incremented.
 410  */
 411 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 412 {
 413         const sector_t this_sector = r1_bio->sector;
 414         const int sectors = r1_bio->sectors;
 415         int start_disk;
 416         int best_disk;
 417         int i;
 418         sector_t best_dist;
 419         mdk_rdev_t *rdev;
 420         int choose_first;
 421
 422         rcu_read_lock();
 423         /*
 424          * Check if we can balance. We can balance on the whole
 425          * device if no resync is going on, or below the resync window.
 426          * We take the first readable disk when above the resync window.
 427          */
 428  retry:
 429         best_disk = -1;
 430         best_dist = MaxSector;
 431         if (conf->mddev->recovery_cp < MaxSector &&
 432             (this_sector + sectors >= conf->next_resync)) {
 433                 choose_first = 1;
 434                 start_disk = 0;
 435         } else {
 436                 choose_first = 0;
 437                 start_disk = conf->last_used;
 438         }
 439
 440         for (i = 0 ; i < conf->raid_disks ; i++) {
 441                 sector_t dist;
 442                 int disk = start_disk + i;
 443                 if (disk >= conf->raid_disks)
 444                         disk -= conf->raid_disks;
 445
 446                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 447                 if (r1_bio->bios[disk] == IO_BLOCKED
 448                     || rdev == NULL
 449                     || test_bit(Faulty, &rdev->flags))
 450                         continue;
 451                 if (!test_bit(In_sync, &rdev->flags) &&
 452                     rdev->recovery_offset < this_sector + sectors)
 453                         continue;
 454                 if (test_bit(WriteMostly, &rdev->flags)) {
 455                         /* Don't balance among write-mostly, just
 456                          * use the first as a last resort */
 457                         if (best_disk < 0)
 458                                 best_disk = disk;
 459                         continue;
 460                 }
 461                 /* This is a reasonable device to use.  It might
 462                  * even be best.
 463                  */
 464                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 465                 if (choose_first
 466                     /* Don't change to another disk for sequential reads */
 467                     || conf->next_seq_sect == this_sector
 468                     || dist == 0
 469                     /* If device is idle, use it */
 470                     || atomic_read(&rdev->nr_pending) == 0) {
 471                         best_disk = disk;
 472                         break;
 473                 }
 474                 if (dist < best_dist) {
 475                         best_dist = dist;
 476                         best_disk = disk;
 477                 }
 478         }
 479
 480         if (best_disk >= 0) {
 481                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 482                 if (!rdev)
 483                         goto retry;
 484                 atomic_inc(&rdev->nr_pending);
 485                 if (test_bit(Faulty, &rdev->flags)) {
 486                         /* cannot risk returning a device that failed
 487                          * before we inc'ed nr_pending
 488                          */
 489                         rdev_dec_pending(rdev, conf->mddev);
 490                         goto retry;
 491                 }
 492                 conf->next_seq_sect = this_sector + sectors;
 493                 conf->last_used = best_disk;
 494         }
 495         rcu_read_unlock();
 496
 497         return best_disk;
 498 }
 499
 500 int md_raid1_congested(mddev_t *mddev, int bits)
 501 {
 502         conf_t *conf = mddev->private;
 503         int i, ret = 0;
 504
 505         rcu_read_lock();
 506         for (i = 0; i < mddev->raid_disks; i++) {
 507                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 508                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 509                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 510
 511                         BUG_ON(!q);
 512
 513                         /* Note the '|| 1' - when read_balance prefers
 514                          * non-congested targets, it can be removed
 515                          */
 516                         if ((bits & (1<<BDI_async_congested)) || 1)
 517                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 518                         else
 519                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 520                 }
 521         }
 522         rcu_read_unlock();
 523         return ret;
 524 }
 525 EXPORT_SYMBOL_GPL(md_raid1_congested);
 526
 527 static int raid1_congested(void *data, int bits)
 528 {
 529         mddev_t *mddev = data;
 530
 531         return mddev_congested(mddev, bits) ||
 532                 md_raid1_congested(mddev, bits);
 533 }
 534
 535 static void flush_pending_writes(conf_t *conf)
 536 {
 537         /* Any writes that have been queued but are awaiting
 538          * bitmap updates get flushed here.
 539          */
 540         spin_lock_irq(&conf->device_lock);
 541
 542         if (conf->pending_bio_list.head) {
 543                 struct bio *bio;
 544                 bio = bio_list_get(&conf->pending_bio_list);
 545                 spin_unlock_irq(&conf->device_lock);
 546                 /* flush any pending bitmap writes to
 547                  * disk before proceeding w/ I/O */
 548                 bitmap_unplug(conf->mddev->bitmap);
 549
 550                 while (bio) { /* submit pending writes */
 551                         struct bio *next = bio->bi_next;
 552                         bio->bi_next = NULL;
 553                         generic_make_request(bio);
 554                         bio = next;
 555                 }
 556         } else
 557                 spin_unlock_irq(&conf->device_lock);
 558 }
 559
 560 /* Barriers....
 561  * Sometimes we need to suspend IO while we do something else,
 562  * either some resync/recovery, or reconfigure the array.
 563  * To do this we raise a 'barrier'.
 564  * The 'barrier' is a counter that can be raised multiple times
 565  * to count how many activities are happening which preclude
 566  * normal IO.
 567  * We can only raise the barrier if there is no pending IO.
 568  * i.e. if nr_pending == 0.
 569  * We choose only to raise the barrier if no-one is waiting for the
 570  * barrier to go down.  This means that as soon as an IO request
 571  * is ready, no other operations which require a barrier will start
 572  * until the IO request has had a chance.
 573  *
 574  * So: regular IO calls 'wait_barrier'.  When that returns there
 575  *    is no backgroup IO happening,  It must arrange to call
 576  *    allow_barrier when it has finished its IO.
 577  * backgroup IO calls must call raise_barrier.  Once that returns
 578  *    there is no normal IO happeing.  It must arrange to call
 579  *    lower_barrier when the particular background IO completes.
 580  */
 581 #define RESYNC_DEPTH 32
 582
 583 static void raise_barrier(conf_t *conf)
 584 {
 585         spin_lock_irq(&conf->resync_lock);
 586
 587         /* Wait until no block IO is waiting */
 588         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 589                             conf->resync_lock, );
 590
 591         /* block any new IO from starting */
 592         conf->barrier++;
 593
 594         /* Now wait for all pending IO to complete */
 595         wait_event_lock_irq(conf->wait_barrier,
 596                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 597                             conf->resync_lock, );
 598
 599         spin_unlock_irq(&conf->resync_lock);
 600 }
 601
 602 static void lower_barrier(conf_t *conf)
 603 {
 604         unsigned long flags;
 605         BUG_ON(conf->barrier <= 0);
 606         spin_lock_irqsave(&conf->resync_lock, flags);
 607         conf->barrier--;
 608         spin_unlock_irqrestore(&conf->resync_lock, flags);
 609         wake_up(&conf->wait_barrier);
 610 }
 611
 612 static void wait_barrier(conf_t *conf)
 613 {
 614         spin_lock_irq(&conf->resync_lock);
 615         if (conf->barrier) {
 616                 conf->nr_waiting++;
 617                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 618                                     conf->resync_lock,
 619                                     );
 620                 conf->nr_waiting--;
 621         }
 622         conf->nr_pending++;
 623         spin_unlock_irq(&conf->resync_lock);
 624 }
 625
 626 static void allow_barrier(conf_t *conf)
 627 {
 628         unsigned long flags;
 629         spin_lock_irqsave(&conf->resync_lock, flags);
 630         conf->nr_pending--;
 631         spin_unlock_irqrestore(&conf->resync_lock, flags);
 632         wake_up(&conf->wait_barrier);
 633 }
 634
 635 static void freeze_array(conf_t *conf)
 636 {
 637         /* stop syncio and normal IO and wait for everything to
 638          * go quite.
 639          * We increment barrier and nr_waiting, and then
 640          * wait until nr_pending match nr_queued+1
 641          * This is called in the context of one normal IO request
 642          * that has failed. Thus any sync request that might be pending
 643          * will be blocked by nr_pending, and we need to wait for
 644          * pending IO requests to complete or be queued for re-try.
 645          * Thus the number queued (nr_queued) plus this request (1)
 646          * must match the number of pending IOs (nr_pending) before
 647          * we continue.
 648          */
 649         spin_lock_irq(&conf->resync_lock);
 650         conf->barrier++;
 651         conf->nr_waiting++;
 652         wait_event_lock_irq(conf->wait_barrier,
 653                             conf->nr_pending == conf->nr_queued+1,
 654                             conf->resync_lock,
 655                             flush_pending_writes(conf));
 656         spin_unlock_irq(&conf->resync_lock);
 657 }
 658 static void unfreeze_array(conf_t *conf)
 659 {
 660         /* reverse the effect of the freeze */
 661         spin_lock_irq(&conf->resync_lock);
 662         conf->barrier--;
 663         conf->nr_waiting--;
 664         wake_up(&conf->wait_barrier);
 665         spin_unlock_irq(&conf->resync_lock);
 666 }
 667
 668
 669 /* duplicate the data pages for behind I/O
 670  */
 671 static void alloc_behind_pages(struct bio *bio, r1bio_t *r1_bio)
 672 {
 673         int i;
 674         struct bio_vec *bvec;
 675         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page*),
 676                                         GFP_NOIO);
 677         if (unlikely(!pages))
 678                 return;
 679
 680         bio_for_each_segment(bvec, bio, i) {
 681                 pages[i] = alloc_page(GFP_NOIO);
 682                 if (unlikely(!pages[i]))
 683                         goto do_sync_io;
 684                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 685                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 686                 kunmap(pages[i]);
 687                 kunmap(bvec->bv_page);
 688         }
 689         r1_bio->behind_pages = pages;
 690         r1_bio->behind_page_count = bio->bi_vcnt;
 691         set_bit(R1BIO_BehindIO, &r1_bio->state);
 692         return;
 693
 694 do_sync_io:
 695         for (i = 0; i < bio->bi_vcnt; i++)
 696                 if (pages[i])
 697                         put_page(pages[i]);
 698         kfree(pages);
 699         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 700 }
 701
 702 static int make_request(mddev_t *mddev, struct bio * bio)
 703 {
 704         conf_t *conf = mddev->private;
 705         mirror_info_t *mirror;
 706         r1bio_t *r1_bio;
 707         struct bio *read_bio;
 708         int i, targets = 0, disks;
 709         struct bitmap *bitmap;
 710         unsigned long flags;
 711         const int rw = bio_data_dir(bio);
 712         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 713         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 714         mdk_rdev_t *blocked_rdev;
 715         int plugged;
 716
 717         /*
 718          * Register the new request and wait if the reconstruction
 719          * thread has put up a bar for new requests.
 720          * Continue immediately if no resync is active currently.
 721          */
 722
 723         md_write_start(mddev, bio); /* wait on superblock update early */
 724
 725         if (bio_data_dir(bio) == WRITE &&
 726             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 727             bio->bi_sector < mddev->suspend_hi) {
 728                 /* As the suspend_* range is controlled by
 729                  * userspace, we want an interruptible
 730                  * wait.
 731                  */
 732                 DEFINE_WAIT(w);
 733                 for (;;) {
 734                         flush_signals(current);
 735                         prepare_to_wait(&conf->wait_barrier,
 736                                         &w, TASK_INTERRUPTIBLE);
 737                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 738                             bio->bi_sector >= mddev->suspend_hi)
 739                                 break;
 740                         schedule();
 741                 }
 742                 finish_wait(&conf->wait_barrier, &w);
 743         }
 744
 745         wait_barrier(conf);
 746
 747         bitmap = mddev->bitmap;
 748
 749         /*
 750          * make_request() can abort the operation when READA is being
 751          * used and no empty request is available.
 752          *
 753          */
 754         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 755
 756         r1_bio->master_bio = bio;
 757         r1_bio->sectors = bio->bi_size >> 9;
 758         r1_bio->state = 0;
 759         r1_bio->mddev = mddev;
 760         r1_bio->sector = bio->bi_sector;
 761
 762         if (rw == READ) {
 763                 /*
 764                  * read balancing logic:
 765                  */
 766                 int rdisk = read_balance(conf, r1_bio);
 767
 768                 if (rdisk < 0) {
 769                         /* couldn't find anywhere to read from */
 770                         raid_end_bio_io(r1_bio);
 771                         return 0;
 772                 }
 773                 mirror = conf->mirrors + rdisk;
 774
 775                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 776                     bitmap) {
 777                         /* Reading from a write-mostly device must
 778                          * take care not to over-take any writes
 779                          * that are 'behind'
 780                          */
 781                         wait_event(bitmap->behind_wait,
 782                                    atomic_read(&bitmap->behind_writes) == 0);
 783                 }
 784                 r1_bio->read_disk = rdisk;
 785
 786                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 787
 788                 r1_bio->bios[rdisk] = read_bio;
 789
 790                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 791                 read_bio->bi_bdev = mirror->rdev->bdev;
 792                 read_bio->bi_end_io = raid1_end_read_request;
 793                 read_bio->bi_rw = READ | do_sync;
 794                 read_bio->bi_private = r1_bio;
 795
 796                 generic_make_request(read_bio);
 797                 return 0;
 798         }
 799
 800         /*
 801          * WRITE:
 802          */
 803         /* first select target devices under spinlock and
 804          * inc refcount on their rdev.  Record them by setting
 805          * bios[x] to bio
 806          */
 807         plugged = mddev_check_plugged(mddev);
 808
 809         disks = conf->raid_disks;
 810  retry_write:
 811         blocked_rdev = NULL;
 812         rcu_read_lock();
 813         for (i = 0;  i < disks; i++) {
 814                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 815                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 816                         atomic_inc(&rdev->nr_pending);
 817                         blocked_rdev = rdev;
 818                         break;
 819                 }
 820                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 821                         atomic_inc(&rdev->nr_pending);
 822                         if (test_bit(Faulty, &rdev->flags)) {
 823                                 rdev_dec_pending(rdev, mddev);
 824                                 r1_bio->bios[i] = NULL;
 825                         } else {
 826                                 r1_bio->bios[i] = bio;
 827                                 targets++;
 828                         }
 829                 } else
 830                         r1_bio->bios[i] = NULL;
 831         }
 832         rcu_read_unlock();
 833
 834         if (unlikely(blocked_rdev)) {
 835                 /* Wait for this device to become unblocked */
 836                 int j;
 837
 838                 for (j = 0; j < i; j++)
 839                         if (r1_bio->bios[j])
 840                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 841
 842                 allow_barrier(conf);
 843                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 844                 wait_barrier(conf);
 845                 goto retry_write;
 846         }
 847
 848         BUG_ON(targets == 0); /* we never fail the last device */
 849
 850         if (targets < conf->raid_disks) {
 851                 /* array is degraded, we will not clear the bitmap
 852                  * on I/O completion (see raid1_end_write_request) */
 853                 set_bit(R1BIO_Degraded, &r1_bio->state);
 854         }
 855
 856         /* do behind I/O ?
 857          * Not if there are too many, or cannot allocate memory,
 858          * or a reader on WriteMostly is waiting for behind writes
 859          * to flush */
 860         if (bitmap &&
 861             (atomic_read(&bitmap->behind_writes)
 862              < mddev->bitmap_info.max_write_behind) &&
 863             !waitqueue_active(&bitmap->behind_wait))
 864                 alloc_behind_pages(bio, r1_bio);
 865
 866         atomic_set(&r1_bio->remaining, 1);
 867         atomic_set(&r1_bio->behind_remaining, 0);
 868
 869         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 870                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 871         for (i = 0; i < disks; i++) {
 872                 struct bio *mbio;
 873                 if (!r1_bio->bios[i])
 874                         continue;
 875
 876                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 877                 r1_bio->bios[i] = mbio;
 878
 879                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 880                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 881                 mbio->bi_end_io = raid1_end_write_request;
 882                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
 883                 mbio->bi_private = r1_bio;
 884
 885                 if (r1_bio->behind_pages) {
 886                         struct bio_vec *bvec;
 887                         int j;
 888
 889                         /* Yes, I really want the '__' version so that
 890                          * we clear any unused pointer in the io_vec, rather
 891                          * than leave them unchanged.  This is important
 892                          * because when we come to free the pages, we won't
 893                          * know the original bi_idx, so we just free
 894                          * them all
 895                          */
 896                         __bio_for_each_segment(bvec, mbio, j, 0)
 897                                 bvec->bv_page = r1_bio->behind_pages[j];
 898                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 899                                 atomic_inc(&r1_bio->behind_remaining);
 900                 }
 901
 902                 atomic_inc(&r1_bio->remaining);
 903                 spin_lock_irqsave(&conf->device_lock, flags);
 904                 bio_list_add(&conf->pending_bio_list, mbio);
 905                 spin_unlock_irqrestore(&conf->device_lock, flags);
 906         }
 907         r1_bio_write_done(r1_bio);
 908
 909         /* In case raid1d snuck in to freeze_array */
 910         wake_up(&conf->wait_barrier);
 911
 912         if (do_sync || !bitmap || !plugged)
 913                 md_wakeup_thread(mddev->thread);
 914
 915         return 0;
 916 }
 917
 918 static void status(struct seq_file *seq, mddev_t *mddev)
 919 {
 920         conf_t *conf = mddev->private;
 921         int i;
 922
 923         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 924                    conf->raid_disks - mddev->degraded);
 925         rcu_read_lock();
 926         for (i = 0; i < conf->raid_disks; i++) {
 927                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 928                 seq_printf(seq, "%s",
 929                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
 930         }
 931         rcu_read_unlock();
 932         seq_printf(seq, "]");
 933 }
 934
 935
 936 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 937 {
 938         char b[BDEVNAME_SIZE];
 939         conf_t *conf = mddev->private;
 940
 941         /*
 942          * If it is not operational, then we have already marked it as dead
 943          * else if it is the last working disks, ignore the error, let the
 944          * next level up know.
 945          * else mark the drive as failed
 946          */
 947         if (test_bit(In_sync, &rdev->flags)
 948             && (conf->raid_disks - mddev->degraded) == 1) {
 949                 /*
 950                  * Don't fail the drive, act as though we were just a
 951                  * normal single drive.
 952                  * However don't try a recovery from this drive as
 953                  * it is very likely to fail.
 954                  */
 955                 mddev->recovery_disabled = 1;
 956                 return;
 957         }
 958         if (test_and_clear_bit(In_sync, &rdev->flags)) {
 959                 unsigned long flags;
 960                 spin_lock_irqsave(&conf->device_lock, flags);
 961                 mddev->degraded++;
 962                 set_bit(Faulty, &rdev->flags);
 963                 spin_unlock_irqrestore(&conf->device_lock, flags);
 964                 /*
 965                  * if recovery is running, make sure it aborts.
 966                  */
 967                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
 968         } else
 969                 set_bit(Faulty, &rdev->flags);
 970         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 971         printk(KERN_ALERT
 972                "md/raid1:%s: Disk failure on %s, disabling device.\n"
 973                "md/raid1:%s: Operation continuing on %d devices.\n",
 974                mdname(mddev), bdevname(rdev->bdev, b),
 975                mdname(mddev), conf->raid_disks - mddev->degraded);
 976 }
 977
 978 static void print_conf(conf_t *conf)
 979 {
 980         int i;
 981
 982         printk(KERN_DEBUG "RAID1 conf printout:\n");
 983         if (!conf) {
 984                 printk(KERN_DEBUG "(!conf)\n");
 985                 return;
 986         }
 987         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
 988                 conf->raid_disks);
 989
 990         rcu_read_lock();
 991         for (i = 0; i < conf->raid_disks; i++) {
 992                 char b[BDEVNAME_SIZE];
 993                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 994                 if (rdev)
 995                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
 996                                i, !test_bit(In_sync, &rdev->flags),
 997                                !test_bit(Faulty, &rdev->flags),
 998                                bdevname(rdev->bdev,b));
 999         }
1000         rcu_read_unlock();
1001 }
1002
1003 static void close_sync(conf_t *conf)
1004 {
1005         wait_barrier(conf);
1006         allow_barrier(conf);
1007
1008         mempool_destroy(conf->r1buf_pool);
1009         conf->r1buf_pool = NULL;
1010 }
1011
1012 static int raid1_spare_active(mddev_t *mddev)
1013 {
1014         int i;
1015         conf_t *conf = mddev->private;
1016         int count = 0;
1017         unsigned long flags;
1018
1019         /*
1020          * Find all failed disks within the RAID1 configuration
1021          * and mark them readable.
1022          * Called under mddev lock, so rcu protection not needed.
1023          */
1024         for (i = 0; i < conf->raid_disks; i++) {
1025                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1026                 if (rdev
1027                     && !test_bit(Faulty, &rdev->flags)
1028                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1029                         count++;
1030                         sysfs_notify_dirent(rdev->sysfs_state);
1031                 }
1032         }
1033         spin_lock_irqsave(&conf->device_lock, flags);
1034         mddev->degraded -= count;
1035         spin_unlock_irqrestore(&conf->device_lock, flags);
1036
1037         print_conf(conf);
1038         return count;
1039 }
1040
1041
1042 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1043 {
1044         conf_t *conf = mddev->private;
1045         int err = -EEXIST;
1046         int mirror = 0;
1047         mirror_info_t *p;
1048         int first = 0;
1049         int last = mddev->raid_disks - 1;
1050
1051         if (rdev->raid_disk >= 0)
1052                 first = last = rdev->raid_disk;
1053
1054         for (mirror = first; mirror <= last; mirror++)
1055                 if ( !(p=conf->mirrors+mirror)->rdev) {
1056
1057                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1058                                           rdev->data_offset << 9);
1059                         /* as we don't honour merge_bvec_fn, we must
1060                          * never risk violating it, so limit
1061                          * ->max_segments to one lying with a single
1062                          * page, as a one page request is never in
1063                          * violation.
1064                          */
1065                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1066                                 blk_queue_max_segments(mddev->queue, 1);
1067                                 blk_queue_segment_boundary(mddev->queue,
1068                                                            PAGE_CACHE_SIZE - 1);
1069                         }
1070
1071                         p->head_position = 0;
1072                         rdev->raid_disk = mirror;
1073                         err = 0;
1074                         /* As all devices are equivalent, we don't need a full recovery
1075                          * if this was recently any drive of the array
1076                          */
1077                         if (rdev->saved_raid_disk < 0)
1078                                 conf->fullsync = 1;
1079                         rcu_assign_pointer(p->rdev, rdev);
1080                         break;
1081                 }
1082         md_integrity_add_rdev(rdev, mddev);
1083         print_conf(conf);
1084         return err;
1085 }
1086
1087 static int raid1_remove_disk(mddev_t *mddev, int number)
1088 {
1089         conf_t *conf = mddev->private;
1090         int err = 0;
1091         mdk_rdev_t *rdev;
1092         mirror_info_t *p = conf->mirrors+ number;
1093
1094         print_conf(conf);
1095         rdev = p->rdev;
1096         if (rdev) {
1097                 if (test_bit(In_sync, &rdev->flags) ||
1098                     atomic_read(&rdev->nr_pending)) {
1099                         err = -EBUSY;
1100                         goto abort;
1101                 }
1102                 /* Only remove non-faulty devices if recovery
1103                  * is not possible.
1104                  */
1105                 if (!test_bit(Faulty, &rdev->flags) &&
1106                     !mddev->recovery_disabled &&
1107                     mddev->degraded < conf->raid_disks) {
1108                         err = -EBUSY;
1109                         goto abort;
1110                 }
1111                 p->rdev = NULL;
1112                 synchronize_rcu();
1113                 if (atomic_read(&rdev->nr_pending)) {
1114                         /* lost the race, try later */
1115                         err = -EBUSY;
1116                         p->rdev = rdev;
1117                         goto abort;
1118                 }
1119                 err = md_integrity_register(mddev);
1120         }
1121 abort:
1122
1123         print_conf(conf);
1124         return err;
1125 }
1126
1127
1128 static void end_sync_read(struct bio *bio, int error)
1129 {
1130         r1bio_t *r1_bio = bio->bi_private;
1131         int i;
1132
1133         for (i=r1_bio->mddev->raid_disks; i--; )
1134                 if (r1_bio->bios[i] == bio)
1135                         break;
1136         BUG_ON(i < 0);
1137         update_head_pos(i, r1_bio);
1138         /*
1139          * we have read a block, now it needs to be re-written,
1140          * or re-read if the read failed.
1141          * We don't do much here, just schedule handling by raid1d
1142          */
1143         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1144                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1145
1146         if (atomic_dec_and_test(&r1_bio->remaining))
1147                 reschedule_retry(r1_bio);
1148 }
1149
1150 static void end_sync_write(struct bio *bio, int error)
1151 {
1152         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1153         r1bio_t *r1_bio = bio->bi_private;
1154         mddev_t *mddev = r1_bio->mddev;
1155         conf_t *conf = mddev->private;
1156         int i;
1157         int mirror=0;
1158
1159         for (i = 0; i < conf->raid_disks; i++)
1160                 if (r1_bio->bios[i] == bio) {
1161                         mirror = i;
1162                         break;
1163                 }
1164         if (!uptodate) {
1165                 sector_t sync_blocks = 0;
1166                 sector_t s = r1_bio->sector;
1167                 long sectors_to_go = r1_bio->sectors;
1168                 /* make sure these bits doesn't get cleared. */
1169                 do {
1170                         bitmap_end_sync(mddev->bitmap, s,
1171                                         &sync_blocks, 1);
1172                         s += sync_blocks;
1173                         sectors_to_go -= sync_blocks;
1174                 } while (sectors_to_go > 0);
1175                 md_error(mddev, conf->mirrors[mirror].rdev);
1176         }
1177
1178         update_head_pos(mirror, r1_bio);
1179
1180         if (atomic_dec_and_test(&r1_bio->remaining)) {
1181                 sector_t s = r1_bio->sectors;
1182                 put_buf(r1_bio);
1183                 md_done_sync(mddev, s, uptodate);
1184         }
1185 }
1186
1187 static int fix_sync_read_error(r1bio_t *r1_bio)
1188 {
1189         /* Try some synchronous reads of other devices to get
1190          * good data, much like with normal read errors.  Only
1191          * read into the pages we already have so we don't
1192          * need to re-issue the read request.
1193          * We don't need to freeze the array, because being in an
1194          * active sync request, there is no normal IO, and
1195          * no overlapping syncs.
1196          */
1197         mddev_t *mddev = r1_bio->mddev;
1198         conf_t *conf = mddev->private;
1199         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1200         sector_t sect = r1_bio->sector;
1201         int sectors = r1_bio->sectors;
1202         int idx = 0;
1203
1204         while(sectors) {
1205                 int s = sectors;
1206                 int d = r1_bio->read_disk;
1207                 int success = 0;
1208                 mdk_rdev_t *rdev;
1209                 int start;
1210
1211                 if (s > (PAGE_SIZE>>9))
1212                         s = PAGE_SIZE >> 9;
1213                 do {
1214                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1215                                 /* No rcu protection needed here devices
1216                                  * can only be removed when no resync is
1217                                  * active, and resync is currently active
1218                                  */
1219                                 rdev = conf->mirrors[d].rdev;
1220                                 if (sync_page_io(rdev,
1221                                                  sect,
1222                                                  s<<9,
1223                                                  bio->bi_io_vec[idx].bv_page,
1224                                                  READ, false)) {
1225                                         success = 1;
1226                                         break;
1227                                 }
1228                         }
1229                         d++;
1230                         if (d == conf->raid_disks)
1231                                 d = 0;
1232                 } while (!success && d != r1_bio->read_disk);
1233
1234                 if (!success) {
1235                         char b[BDEVNAME_SIZE];
1236                         /* Cannot read from anywhere, array is toast */
1237                         md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1238                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1239                                " for block %llu\n",
1240                                mdname(mddev),
1241                                bdevname(bio->bi_bdev, b),
1242                                (unsigned long long)r1_bio->sector);
1243                         md_done_sync(mddev, r1_bio->sectors, 0);
1244                         put_buf(r1_bio);
1245                         return 0;
1246                 }
1247
1248                 start = d;
1249                 /* write it back and re-read */
1250                 while (d != r1_bio->read_disk) {
1251                         if (d == 0)
1252                                 d = conf->raid_disks;
1253                         d--;
1254                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1255                                 continue;
1256                         rdev = conf->mirrors[d].rdev;
1257                         if (sync_page_io(rdev,
1258                                          sect,
1259                                          s<<9,
1260                                          bio->bi_io_vec[idx].bv_page,
1261                                          WRITE, false) == 0) {
1262                                 r1_bio->bios[d]->bi_end_io = NULL;
1263                                 rdev_dec_pending(rdev, mddev);
1264                                 md_error(mddev, rdev);
1265                         } else
1266                                 atomic_add(s, &rdev->corrected_errors);
1267                 }
1268                 d = start;
1269                 while (d != r1_bio->read_disk) {
1270                         if (d == 0)
1271                                 d = conf->raid_disks;
1272                         d--;
1273                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1274                                 continue;
1275                         rdev = conf->mirrors[d].rdev;
1276                         if (sync_page_io(rdev,
1277                                          sect,
1278                                          s<<9,
1279                                          bio->bi_io_vec[idx].bv_page,
1280                                          READ, false) == 0)
1281                                 md_error(mddev, rdev);
1282                 }
1283                 sectors -= s;
1284                 sect += s;
1285                 idx ++;
1286         }
1287         set_bit(R1BIO_Uptodate, &r1_bio->state);
1288         set_bit(BIO_UPTODATE, &bio->bi_flags);
1289         return 1;
1290 }
1291
1292 static int process_checks(r1bio_t *r1_bio)
1293 {
1294         /* We have read all readable devices.  If we haven't
1295          * got the block, then there is no hope left.
1296          * If we have, then we want to do a comparison
1297          * and skip the write if everything is the same.
1298          * If any blocks failed to read, then we need to
1299          * attempt an over-write
1300          */
1301         mddev_t *mddev = r1_bio->mddev;
1302         conf_t *conf = mddev->private;
1303         int primary;
1304         int i;
1305
1306         for (primary = 0; primary < conf->raid_disks; primary++)
1307                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1308                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1309                         r1_bio->bios[primary]->bi_end_io = NULL;
1310                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1311                         break;
1312                 }
1313         r1_bio->read_disk = primary;
1314         for (i = 0; i < conf->raid_disks; i++) {
1315                 int j;
1316                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1317                 struct bio *pbio = r1_bio->bios[primary];
1318                 struct bio *sbio = r1_bio->bios[i];
1319                 int size;
1320
1321                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1322                         continue;
1323
1324                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1325                         for (j = vcnt; j-- ; ) {
1326                                 struct page *p, *s;
1327                                 p = pbio->bi_io_vec[j].bv_page;
1328                                 s = sbio->bi_io_vec[j].bv_page;
1329                                 if (memcmp(page_address(p),
1330                                            page_address(s),
1331                                            PAGE_SIZE))
1332                                         break;
1333                         }
1334                 } else
1335                         j = 0;
1336                 if (j >= 0)
1337                         mddev->resync_mismatches += r1_bio->sectors;
1338                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1339                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1340                         /* No need to write to this device. */
1341                         sbio->bi_end_io = NULL;
1342                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1343                         continue;
1344                 }
1345                 /* fixup the bio for reuse */
1346                 sbio->bi_vcnt = vcnt;
1347                 sbio->bi_size = r1_bio->sectors << 9;
1348                 sbio->bi_idx = 0;
1349                 sbio->bi_phys_segments = 0;
1350                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1351                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1352                 sbio->bi_next = NULL;
1353                 sbio->bi_sector = r1_bio->sector +
1354                         conf->mirrors[i].rdev->data_offset;
1355                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1356                 size = sbio->bi_size;
1357                 for (j = 0; j < vcnt ; j++) {
1358                         struct bio_vec *bi;
1359                         bi = &sbio->bi_io_vec[j];
1360                         bi->bv_offset = 0;
1361                         if (size > PAGE_SIZE)
1362                                 bi->bv_len = PAGE_SIZE;
1363                         else
1364                                 bi->bv_len = size;
1365                         size -= PAGE_SIZE;
1366                         memcpy(page_address(bi->bv_page),
1367                                page_address(pbio->bi_io_vec[j].bv_page),
1368                                PAGE_SIZE);
1369                 }
1370         }
1371         return 0;
1372 }
1373
1374 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1375 {
1376         conf_t *conf = mddev->private;
1377         int i;
1378         int disks = conf->raid_disks;
1379         struct bio *bio, *wbio;
1380
1381         bio = r1_bio->bios[r1_bio->read_disk];
1382
1383         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1384                 /* ouch - failed to read all of that. */
1385                 if (!fix_sync_read_error(r1_bio))
1386                         return;
1387
1388         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1389                 if (process_checks(r1_bio) < 0)
1390                         return;
1391         /*
1392          * schedule writes
1393          */
1394         atomic_set(&r1_bio->remaining, 1);
1395         for (i = 0; i < disks ; i++) {
1396                 wbio = r1_bio->bios[i];
1397                 if (wbio->bi_end_io == NULL ||
1398                     (wbio->bi_end_io == end_sync_read &&
1399                      (i == r1_bio->read_disk ||
1400                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1401                         continue;
1402
1403                 wbio->bi_rw = WRITE;
1404                 wbio->bi_end_io = end_sync_write;
1405                 atomic_inc(&r1_bio->remaining);
1406                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1407
1408                 generic_make_request(wbio);
1409         }
1410
1411         if (atomic_dec_and_test(&r1_bio->remaining)) {
1412                 /* if we're here, all write(s) have completed, so clean up */
1413                 md_done_sync(mddev, r1_bio->sectors, 1);
1414                 put_buf(r1_bio);
1415         }
1416 }
1417
1418 /*
1419  * This is a kernel thread which:
1420  *
1421  *      1.      Retries failed read operations on working mirrors.
1422  *      2.      Updates the raid superblock when problems encounter.
1423  *      3.      Performs writes following reads for array syncronising.
1424  */
1425
1426 static void fix_read_error(conf_t *conf, int read_disk,
1427                            sector_t sect, int sectors)
1428 {
1429         mddev_t *mddev = conf->mddev;
1430         while(sectors) {
1431                 int s = sectors;
1432                 int d = read_disk;
1433                 int success = 0;
1434                 int start;
1435                 mdk_rdev_t *rdev;
1436
1437                 if (s > (PAGE_SIZE>>9))
1438                         s = PAGE_SIZE >> 9;
1439
1440                 do {
1441                         /* Note: no rcu protection needed here
1442                          * as this is synchronous in the raid1d thread
1443                          * which is the thread that might remove
1444                          * a device.  If raid1d ever becomes multi-threaded....
1445                          */
1446                         rdev = conf->mirrors[d].rdev;
1447                         if (rdev &&
1448                             test_bit(In_sync, &rdev->flags) &&
1449                             sync_page_io(rdev, sect, s<<9,
1450                                          conf->tmppage, READ, false))
1451                                 success = 1;
1452                         else {
1453                                 d++;
1454                                 if (d == conf->raid_disks)
1455                                         d = 0;
1456                         }
1457                 } while (!success && d != read_disk);
1458
1459                 if (!success) {
1460                         /* Cannot read from anywhere -- bye bye array */
1461                         md_error(mddev, conf->mirrors[read_disk].rdev);
1462                         break;
1463                 }
1464                 /* write it back and re-read */
1465                 start = d;
1466                 while (d != read_disk) {
1467                         if (d==0)
1468                                 d = conf->raid_disks;
1469                         d--;
1470                         rdev = conf->mirrors[d].rdev;
1471                         if (rdev &&
1472                             test_bit(In_sync, &rdev->flags)) {
1473                                 if (sync_page_io(rdev, sect, s<<9,
1474                                                  conf->tmppage, WRITE, false)
1475                                     == 0)
1476                                         /* Well, this device is dead */
1477                                         md_error(mddev, rdev);
1478                         }
1479                 }
1480                 d = start;
1481                 while (d != read_disk) {
1482                         char b[BDEVNAME_SIZE];
1483                         if (d==0)
1484                                 d = conf->raid_disks;
1485                         d--;
1486                         rdev = conf->mirrors[d].rdev;
1487                         if (rdev &&
1488                             test_bit(In_sync, &rdev->flags)) {
1489                                 if (sync_page_io(rdev, sect, s<<9,
1490                                                  conf->tmppage, READ, false)
1491                                     == 0)
1492                                         /* Well, this device is dead */
1493                                         md_error(mddev, rdev);
1494                                 else {
1495                                         atomic_add(s, &rdev->corrected_errors);
1496                                         printk(KERN_INFO
1497                                                "md/raid1:%s: read error corrected "
1498                                                "(%d sectors at %llu on %s)\n",
1499                                                mdname(mddev), s,
1500                                                (unsigned long long)(sect +
1501                                                    rdev->data_offset),
1502                                                bdevname(rdev->bdev, b));
1503                                 }
1504                         }
1505                 }
1506                 sectors -= s;
1507                 sect += s;
1508         }
1509 }
1510
1511 static void raid1d(mddev_t *mddev)
1512 {
1513         r1bio_t *r1_bio;
1514         struct bio *bio;
1515         unsigned long flags;
1516         conf_t *conf = mddev->private;
1517         struct list_head *head = &conf->retry_list;
1518         mdk_rdev_t *rdev;
1519         struct blk_plug plug;
1520
1521         md_check_recovery(mddev);
1522
1523         blk_start_plug(&plug);
1524         for (;;) {
1525                 char b[BDEVNAME_SIZE];
1526
1527                 if (atomic_read(&mddev->plug_cnt) == 0)
1528                         flush_pending_writes(conf);
1529
1530                 spin_lock_irqsave(&conf->device_lock, flags);
1531                 if (list_empty(head)) {
1532                         spin_unlock_irqrestore(&conf->device_lock, flags);
1533                         break;
1534                 }
1535                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1536                 list_del(head->prev);
1537                 conf->nr_queued--;
1538                 spin_unlock_irqrestore(&conf->device_lock, flags);
1539
1540                 mddev = r1_bio->mddev;
1541                 conf = mddev->private;
1542                 if (test_bit(R1BIO_IsSync, &r1_bio->state))
1543                         sync_request_write(mddev, r1_bio);
1544                 else {
1545                         int disk;
1546
1547                         /* we got a read error. Maybe the drive is bad.  Maybe just
1548                          * the block and we can fix it.
1549                          * We freeze all other IO, and try reading the block from
1550                          * other devices.  When we find one, we re-write
1551                          * and check it that fixes the read error.
1552                          * This is all done synchronously while the array is
1553                          * frozen
1554                          */
1555                         if (mddev->ro == 0) {
1556                                 freeze_array(conf);
1557                                 fix_read_error(conf, r1_bio->read_disk,
1558                                                r1_bio->sector,
1559                                                r1_bio->sectors);
1560                                 unfreeze_array(conf);
1561                         } else
1562                                 md_error(mddev,
1563                                          conf->mirrors[r1_bio->read_disk].rdev);
1564
1565                         bio = r1_bio->bios[r1_bio->read_disk];
1566                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1567                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1568                                        " read error for block %llu\n",
1569                                        mdname(mddev),
1570                                        bdevname(bio->bi_bdev,b),
1571                                        (unsigned long long)r1_bio->sector);
1572                                 raid_end_bio_io(r1_bio);
1573                         } else {
1574                                 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1575                                 r1_bio->bios[r1_bio->read_disk] =
1576                                         mddev->ro ? IO_BLOCKED : NULL;
1577                                 r1_bio->read_disk = disk;
1578                                 bio_put(bio);
1579                                 bio = bio_clone_mddev(r1_bio->master_bio,
1580                                                       GFP_NOIO, mddev);
1581                                 r1_bio->bios[r1_bio->read_disk] = bio;
1582                                 rdev = conf->mirrors[disk].rdev;
1583                                 if (printk_ratelimit())
1584                                         printk(KERN_ERR "md/raid1:%s: redirecting sector %llu to"
1585                                                " other mirror: %s\n",
1586                                                mdname(mddev),
1587                                                (unsigned long long)r1_bio->sector,
1588                                                bdevname(rdev->bdev,b));
1589                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1590                                 bio->bi_bdev = rdev->bdev;
1591                                 bio->bi_end_io = raid1_end_read_request;
1592                                 bio->bi_rw = READ | do_sync;
1593                                 bio->bi_private = r1_bio;
1594                                 generic_make_request(bio);
1595                         }
1596                 }
1597                 cond_resched();
1598         }
1599         blk_finish_plug(&plug);
1600 }
1601
1602
1603 static int init_resync(conf_t *conf)
1604 {
1605         int buffs;
1606
1607         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1608         BUG_ON(conf->r1buf_pool);
1609         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1610                                           conf->poolinfo);
1611         if (!conf->r1buf_pool)
1612                 return -ENOMEM;
1613         conf->next_resync = 0;
1614         return 0;
1615 }
1616
1617 /*
1618  * perform a "sync" on one "block"
1619  *
1620  * We need to make sure that no normal I/O request - particularly write
1621  * requests - conflict with active sync requests.
1622  *
1623  * This is achieved by tracking pending requests and a 'barrier' concept
1624  * that can be installed to exclude normal IO requests.
1625  */
1626
1627 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1628 {
1629         conf_t *conf = mddev->private;
1630         r1bio_t *r1_bio;
1631         struct bio *bio;
1632         sector_t max_sector, nr_sectors;
1633         int disk = -1;
1634         int i;
1635         int wonly = -1;
1636         int write_targets = 0, read_targets = 0;
1637         sector_t sync_blocks;
1638         int still_degraded = 0;
1639
1640         if (!conf->r1buf_pool)
1641                 if (init_resync(conf))
1642                         return 0;
1643
1644         max_sector = mddev->dev_sectors;
1645         if (sector_nr >= max_sector) {
1646                 /* If we aborted, we need to abort the
1647                  * sync on the 'current' bitmap chunk (there will
1648                  * only be one in raid1 resync.
1649                  * We can find the current addess in mddev->curr_resync
1650                  */
1651                 if (mddev->curr_resync < max_sector) /* aborted */
1652                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1653                                                 &sync_blocks, 1);
1654                 else /* completed sync */
1655                         conf->fullsync = 0;
1656
1657                 bitmap_close_sync(mddev->bitmap);
1658                 close_sync(conf);
1659                 return 0;
1660         }
1661
1662         if (mddev->bitmap == NULL &&
1663             mddev->recovery_cp == MaxSector &&
1664             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1665             conf->fullsync == 0) {
1666                 *skipped = 1;
1667                 return max_sector - sector_nr;
1668         }
1669         /* before building a request, check if we can skip these blocks..
1670          * This call the bitmap_start_sync doesn't actually record anything
1671          */
1672         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1673             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1674                 /* We can skip this block, and probably several more */
1675                 *skipped = 1;
1676                 return sync_blocks;
1677         }
1678         /*
1679          * If there is non-resync activity waiting for a turn,
1680          * and resync is going fast enough,
1681          * then let it though before starting on this new sync request.
1682          */
1683         if (!go_faster && conf->nr_waiting)
1684                 msleep_interruptible(1000);
1685
1686         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1687         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1688         raise_barrier(conf);
1689
1690         conf->next_resync = sector_nr;
1691
1692         rcu_read_lock();
1693         /*
1694          * If we get a correctably read error during resync or recovery,
1695          * we might want to read from a different device.  So we
1696          * flag all drives that could conceivably be read from for READ,
1697          * and any others (which will be non-In_sync devices) for WRITE.
1698          * If a read fails, we try reading from something else for which READ
1699          * is OK.
1700          */
1701
1702         r1_bio->mddev = mddev;
1703         r1_bio->sector = sector_nr;
1704         r1_bio->state = 0;
1705         set_bit(R1BIO_IsSync, &r1_bio->state);
1706
1707         for (i=0; i < conf->raid_disks; i++) {
1708                 mdk_rdev_t *rdev;
1709                 bio = r1_bio->bios[i];
1710
1711                 /* take from bio_init */
1712                 bio->bi_next = NULL;
1713                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1714                 bio->bi_flags |= 1 << BIO_UPTODATE;
1715                 bio->bi_comp_cpu = -1;
1716                 bio->bi_rw = READ;
1717                 bio->bi_vcnt = 0;
1718                 bio->bi_idx = 0;
1719                 bio->bi_phys_segments = 0;
1720                 bio->bi_size = 0;
1721                 bio->bi_end_io = NULL;
1722                 bio->bi_private = NULL;
1723
1724                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1725                 if (rdev == NULL ||
1726                            test_bit(Faulty, &rdev->flags)) {
1727                         still_degraded = 1;
1728                         continue;
1729                 } else if (!test_bit(In_sync, &rdev->flags)) {
1730                         bio->bi_rw = WRITE;
1731                         bio->bi_end_io = end_sync_write;
1732                         write_targets ++;
1733                 } else {
1734                         /* may need to read from here */
1735                         bio->bi_rw = READ;
1736                         bio->bi_end_io = end_sync_read;
1737                         if (test_bit(WriteMostly, &rdev->flags)) {
1738                                 if (wonly < 0)
1739                                         wonly = i;
1740                         } else {
1741                                 if (disk < 0)
1742                                         disk = i;
1743                         }
1744                         read_targets++;
1745                 }
1746                 atomic_inc(&rdev->nr_pending);
1747                 bio->bi_sector = sector_nr + rdev->data_offset;
1748                 bio->bi_bdev = rdev->bdev;
1749                 bio->bi_private = r1_bio;
1750         }
1751         rcu_read_unlock();
1752         if (disk < 0)
1753                 disk = wonly;
1754         r1_bio->read_disk = disk;
1755
1756         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1757                 /* extra read targets are also write targets */
1758                 write_targets += read_targets-1;
1759
1760         if (write_targets == 0 || read_targets == 0) {
1761                 /* There is nowhere to write, so all non-sync
1762                  * drives must be failed - so we are finished
1763                  */
1764                 sector_t rv = max_sector - sector_nr;
1765                 *skipped = 1;
1766                 put_buf(r1_bio);
1767                 return rv;
1768         }
1769
1770         if (max_sector > mddev->resync_max)
1771                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1772         nr_sectors = 0;
1773         sync_blocks = 0;
1774         do {
1775                 struct page *page;
1776                 int len = PAGE_SIZE;
1777                 if (sector_nr + (len>>9) > max_sector)
1778                         len = (max_sector - sector_nr) << 9;
1779                 if (len == 0)
1780                         break;
1781                 if (sync_blocks == 0) {
1782                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1783                                                &sync_blocks, still_degraded) &&
1784                             !conf->fullsync &&
1785                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1786                                 break;
1787                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1788                         if ((len >> 9) > sync_blocks)
1789                                 len = sync_blocks<<9;
1790                 }
1791
1792                 for (i=0 ; i < conf->raid_disks; i++) {
1793                         bio = r1_bio->bios[i];
1794                         if (bio->bi_end_io) {
1795                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1796                                 if (bio_add_page(bio, page, len, 0) == 0) {
1797                                         /* stop here */
1798                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1799                                         while (i > 0) {
1800                                                 i--;
1801                                                 bio = r1_bio->bios[i];
1802                                                 if (bio->bi_end_io==NULL)
1803                                                         continue;
1804                                                 /* remove last page from this bio */
1805                                                 bio->bi_vcnt--;
1806                                                 bio->bi_size -= len;
1807                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1808                                         }
1809                                         goto bio_full;
1810                                 }
1811                         }
1812                 }
1813                 nr_sectors += len>>9;
1814                 sector_nr += len>>9;
1815                 sync_blocks -= (len>>9);
1816         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1817  bio_full:
1818         r1_bio->sectors = nr_sectors;
1819
1820         /* For a user-requested sync, we read all readable devices and do a
1821          * compare
1822          */
1823         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1824                 atomic_set(&r1_bio->remaining, read_targets);
1825                 for (i=0; i<conf->raid_disks; i++) {
1826                         bio = r1_bio->bios[i];
1827                         if (bio->bi_end_io == end_sync_read) {
1828                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1829                                 generic_make_request(bio);
1830                         }
1831                 }
1832         } else {
1833                 atomic_set(&r1_bio->remaining, 1);
1834                 bio = r1_bio->bios[r1_bio->read_disk];
1835                 md_sync_acct(bio->bi_bdev, nr_sectors);
1836                 generic_make_request(bio);
1837
1838         }
1839         return nr_sectors;
1840 }
1841
1842 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1843 {
1844         if (sectors)
1845                 return sectors;
1846
1847         return mddev->dev_sectors;
1848 }
1849
1850 static conf_t *setup_conf(mddev_t *mddev)
1851 {
1852         conf_t *conf;
1853         int i;
1854         mirror_info_t *disk;
1855         mdk_rdev_t *rdev;
1856         int err = -ENOMEM;
1857
1858         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1859         if (!conf)
1860                 goto abort;
1861
1862         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1863                                  GFP_KERNEL);
1864         if (!conf->mirrors)
1865                 goto abort;
1866
1867         conf->tmppage = alloc_page(GFP_KERNEL);
1868         if (!conf->tmppage)
1869                 goto abort;
1870
1871         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1872         if (!conf->poolinfo)
1873                 goto abort;
1874         conf->poolinfo->raid_disks = mddev->raid_disks;
1875         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1876                                           r1bio_pool_free,
1877                                           conf->poolinfo);
1878         if (!conf->r1bio_pool)
1879                 goto abort;
1880
1881         conf->poolinfo->mddev = mddev;
1882
1883         spin_lock_init(&conf->device_lock);
1884         list_for_each_entry(rdev, &mddev->disks, same_set) {
1885                 int disk_idx = rdev->raid_disk;
1886                 if (disk_idx >= mddev->raid_disks
1887                     || disk_idx < 0)
1888                         continue;
1889                 disk = conf->mirrors + disk_idx;
1890
1891                 disk->rdev = rdev;
1892
1893                 disk->head_position = 0;
1894         }
1895         conf->raid_disks = mddev->raid_disks;
1896         conf->mddev = mddev;
1897         INIT_LIST_HEAD(&conf->retry_list);
1898
1899         spin_lock_init(&conf->resync_lock);
1900         init_waitqueue_head(&conf->wait_barrier);
1901
1902         bio_list_init(&conf->pending_bio_list);
1903
1904         conf->last_used = -1;
1905         for (i = 0; i < conf->raid_disks; i++) {
1906
1907                 disk = conf->mirrors + i;
1908
1909                 if (!disk->rdev ||
1910                     !test_bit(In_sync, &disk->rdev->flags)) {
1911                         disk->head_position = 0;
1912                         if (disk->rdev)
1913                                 conf->fullsync = 1;
1914                 } else if (conf->last_used < 0)
1915                         /*
1916                          * The first working device is used as a
1917                          * starting point to read balancing.
1918                          */
1919                         conf->last_used = i;
1920         }
1921
1922         err = -EIO;
1923         if (conf->last_used < 0) {
1924                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
1925                        mdname(mddev));
1926                 goto abort;
1927         }
1928         err = -ENOMEM;
1929         conf->thread = md_register_thread(raid1d, mddev, NULL);
1930         if (!conf->thread) {
1931                 printk(KERN_ERR
1932                        "md/raid1:%s: couldn't allocate thread\n",
1933                        mdname(mddev));
1934                 goto abort;
1935         }
1936
1937         return conf;
1938
1939  abort:
1940         if (conf) {
1941                 if (conf->r1bio_pool)
1942                         mempool_destroy(conf->r1bio_pool);
1943                 kfree(conf->mirrors);
1944                 safe_put_page(conf->tmppage);
1945                 kfree(conf->poolinfo);
1946                 kfree(conf);
1947         }
1948         return ERR_PTR(err);
1949 }
1950
1951 static int run(mddev_t *mddev)
1952 {
1953         conf_t *conf;
1954         int i;
1955         mdk_rdev_t *rdev;
1956
1957         if (mddev->level != 1) {
1958                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
1959                        mdname(mddev), mddev->level);
1960                 return -EIO;
1961         }
1962         if (mddev->reshape_position != MaxSector) {
1963                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
1964                        mdname(mddev));
1965                 return -EIO;
1966         }
1967         /*
1968          * copy the already verified devices into our private RAID1
1969          * bookkeeping area. [whatever we allocate in run(),
1970          * should be freed in stop()]
1971          */
1972         if (mddev->private == NULL)
1973                 conf = setup_conf(mddev);
1974         else
1975                 conf = mddev->private;
1976
1977         if (IS_ERR(conf))
1978                 return PTR_ERR(conf);
1979
1980         list_for_each_entry(rdev, &mddev->disks, same_set) {
1981                 if (!mddev->gendisk)
1982                         continue;
1983                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1984                                   rdev->data_offset << 9);
1985                 /* as we don't honour merge_bvec_fn, we must never risk
1986                  * violating it, so limit ->max_segments to 1 lying within
1987                  * a single page, as a one page request is never in violation.
1988                  */
1989                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1990                         blk_queue_max_segments(mddev->queue, 1);
1991                         blk_queue_segment_boundary(mddev->queue,
1992                                                    PAGE_CACHE_SIZE - 1);
1993                 }
1994         }
1995
1996         mddev->degraded = 0;
1997         for (i=0; i < conf->raid_disks; i++)
1998                 if (conf->mirrors[i].rdev == NULL ||
1999                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2000                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2001                         mddev->degraded++;
2002
2003         if (conf->raid_disks - mddev->degraded == 1)
2004                 mddev->recovery_cp = MaxSector;
2005
2006         if (mddev->recovery_cp != MaxSector)
2007                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2008                        " -- starting background reconstruction\n",
2009                        mdname(mddev));
2010         printk(KERN_INFO
2011                 "md/raid1:%s: active with %d out of %d mirrors\n",
2012                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2013                 mddev->raid_disks);
2014
2015         /*
2016          * Ok, everything is just fine now
2017          */
2018         mddev->thread = conf->thread;
2019         conf->thread = NULL;
2020         mddev->private = conf;
2021
2022         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2023
2024         if (mddev->queue) {
2025                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2026                 mddev->queue->backing_dev_info.congested_data = mddev;
2027         }
2028         return md_integrity_register(mddev);
2029 }
2030
2031 static int stop(mddev_t *mddev)
2032 {
2033         conf_t *conf = mddev->private;
2034         struct bitmap *bitmap = mddev->bitmap;
2035
2036         /* wait for behind writes to complete */
2037         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2038                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2039                        mdname(mddev));
2040                 /* need to kick something here to make sure I/O goes? */
2041                 wait_event(bitmap->behind_wait,
2042                            atomic_read(&bitmap->behind_writes) == 0);
2043         }
2044
2045         raise_barrier(conf);
2046         lower_barrier(conf);
2047
2048         md_unregister_thread(mddev->thread);
2049         mddev->thread = NULL;
2050         if (conf->r1bio_pool)
2051                 mempool_destroy(conf->r1bio_pool);
2052         kfree(conf->mirrors);
2053         kfree(conf->poolinfo);
2054         kfree(conf);
2055         mddev->private = NULL;
2056         return 0;
2057 }
2058
2059 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2060 {
2061         /* no resync is happening, and there is enough space
2062          * on all devices, so we can resize.
2063          * We need to make sure resync covers any new space.
2064          * If the array is shrinking we should possibly wait until
2065          * any io in the removed space completes, but it hardly seems
2066          * worth it.
2067          */
2068         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2069         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2070                 return -EINVAL;
2071         set_capacity(mddev->gendisk, mddev->array_sectors);
2072         revalidate_disk(mddev->gendisk);
2073         if (sectors > mddev->dev_sectors &&
2074             mddev->recovery_cp > mddev->dev_sectors) {
2075                 mddev->recovery_cp = mddev->dev_sectors;
2076                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2077         }
2078         mddev->dev_sectors = sectors;
2079         mddev->resync_max_sectors = sectors;
2080         return 0;
2081 }
2082
2083 static int raid1_reshape(mddev_t *mddev)
2084 {
2085         /* We need to:
2086          * 1/ resize the r1bio_pool
2087          * 2/ resize conf->mirrors
2088          *
2089          * We allocate a new r1bio_pool if we can.
2090          * Then raise a device barrier and wait until all IO stops.
2091          * Then resize conf->mirrors and swap in the new r1bio pool.
2092          *
2093          * At the same time, we "pack" the devices so that all the missing
2094          * devices have the higher raid_disk numbers.
2095          */
2096         mempool_t *newpool, *oldpool;
2097         struct pool_info *newpoolinfo;
2098         mirror_info_t *newmirrors;
2099         conf_t *conf = mddev->private;
2100         int cnt, raid_disks;
2101         unsigned long flags;
2102         int d, d2, err;
2103
2104         /* Cannot change chunk_size, layout, or level */
2105         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2106             mddev->layout != mddev->new_layout ||
2107             mddev->level != mddev->new_level) {
2108                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2109                 mddev->new_layout = mddev->layout;
2110                 mddev->new_level = mddev->level;
2111                 return -EINVAL;
2112         }
2113
2114         err = md_allow_write(mddev);
2115         if (err)
2116                 return err;
2117
2118         raid_disks = mddev->raid_disks + mddev->delta_disks;
2119
2120         if (raid_disks < conf->raid_disks) {
2121                 cnt=0;
2122                 for (d= 0; d < conf->raid_disks; d++)
2123                         if (conf->mirrors[d].rdev)
2124                                 cnt++;
2125                 if (cnt > raid_disks)
2126                         return -EBUSY;
2127         }
2128
2129         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2130         if (!newpoolinfo)
2131                 return -ENOMEM;
2132         newpoolinfo->mddev = mddev;
2133         newpoolinfo->raid_disks = raid_disks;
2134
2135         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2136                                  r1bio_pool_free, newpoolinfo);
2137         if (!newpool) {
2138                 kfree(newpoolinfo);
2139                 return -ENOMEM;
2140         }
2141         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2142         if (!newmirrors) {
2143                 kfree(newpoolinfo);
2144                 mempool_destroy(newpool);
2145                 return -ENOMEM;
2146         }
2147
2148         raise_barrier(conf);
2149
2150         /* ok, everything is stopped */
2151         oldpool = conf->r1bio_pool;
2152         conf->r1bio_pool = newpool;
2153
2154         for (d = d2 = 0; d < conf->raid_disks; d++) {
2155                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2156                 if (rdev && rdev->raid_disk != d2) {
2157                         char nm[20];
2158                         sprintf(nm, "rd%d", rdev->raid_disk);
2159                         sysfs_remove_link(&mddev->kobj, nm);
2160                         rdev->raid_disk = d2;
2161                         sprintf(nm, "rd%d", rdev->raid_disk);
2162                         sysfs_remove_link(&mddev->kobj, nm);
2163                         if (sysfs_create_link(&mddev->kobj,
2164                                               &rdev->kobj, nm))
2165                                 printk(KERN_WARNING
2166                                        "md/raid1:%s: cannot register "
2167                                        "%s\n",
2168                                        mdname(mddev), nm);
2169                 }
2170                 if (rdev)
2171                         newmirrors[d2++].rdev = rdev;
2172         }
2173         kfree(conf->mirrors);
2174         conf->mirrors = newmirrors;
2175         kfree(conf->poolinfo);
2176         conf->poolinfo = newpoolinfo;
2177
2178         spin_lock_irqsave(&conf->device_lock, flags);
2179         mddev->degraded += (raid_disks - conf->raid_disks);
2180         spin_unlock_irqrestore(&conf->device_lock, flags);
2181         conf->raid_disks = mddev->raid_disks = raid_disks;
2182         mddev->delta_disks = 0;
2183
2184         conf->last_used = 0; /* just make sure it is in-range */
2185         lower_barrier(conf);
2186
2187         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2188         md_wakeup_thread(mddev->thread);
2189
2190         mempool_destroy(oldpool);
2191         return 0;
2192 }
2193
2194 static void raid1_quiesce(mddev_t *mddev, int state)
2195 {
2196         conf_t *conf = mddev->private;
2197
2198         switch(state) {
2199         case 2: /* wake for suspend */
2200                 wake_up(&conf->wait_barrier);
2201                 break;
2202         case 1:
2203                 raise_barrier(conf);
2204                 break;
2205         case 0:
2206                 lower_barrier(conf);
2207                 break;
2208         }
2209 }
2210
2211 static void *raid1_takeover(mddev_t *mddev)
2212 {
2213         /* raid1 can take over:
2214          *  raid5 with 2 devices, any layout or chunk size
2215          */
2216         if (mddev->level == 5 && mddev->raid_disks == 2) {
2217                 conf_t *conf;
2218                 mddev->new_level = 1;
2219                 mddev->new_layout = 0;
2220                 mddev->new_chunk_sectors = 0;
2221                 conf = setup_conf(mddev);
2222                 if (!IS_ERR(conf))
2223                         conf->barrier = 1;
2224                 return conf;
2225         }
2226         return ERR_PTR(-EINVAL);
2227 }
2228
2229 static struct mdk_personality raid1_personality =
2230 {
2231         .name           = "raid1",
2232         .level          = 1,
2233         .owner          = THIS_MODULE,
2234         .make_request   = make_request,
2235         .run            = run,
2236         .stop           = stop,
2237         .status         = status,
2238         .error_handler  = error,
2239         .hot_add_disk   = raid1_add_disk,
2240         .hot_remove_disk= raid1_remove_disk,
2241         .spare_active   = raid1_spare_active,
2242         .sync_request   = sync_request,
2243         .resize         = raid1_resize,
2244         .size           = raid1_size,
2245         .check_reshape  = raid1_reshape,
2246         .quiesce        = raid1_quiesce,
2247         .takeover       = raid1_takeover,
2248 };
2249
2250 static int __init raid_init(void)
2251 {
2252         return register_md_personality(&raid1_personality);
2253 }
2254
2255 static void raid_exit(void)
2256 {
2257         unregister_md_personality(&raid1_personality);
2258 }
2259
2260 module_init(raid_init);
2261 module_exit(raid_exit);
2262 MODULE_LICENSE("GPL");
2263 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2264 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2265 MODULE_ALIAS("md-raid1");
2266 MODULE_ALIAS("md-level-1");