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 <linux/ratelimit.h>
  39 #include "md.h"
  40 #include "raid1.h"
  41 #include "bitmap.h"
  42
  43 #define DEBUG 0
  44 #if DEBUG
  45 #define PRINTK(x...) printk(x)
  46 #else
  47 #define PRINTK(x...)
  48 #endif
  49
  50 /*
  51  * Number of guaranteed r1bios in case of extreme VM load:
  52  */
  53 #define NR_RAID1_BIOS 256
  54
  55
  56 static void allow_barrier(conf_t *conf);
  57 static void lower_barrier(conf_t *conf);
  58
  59 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  60 {
  61         struct pool_info *pi = data;
  62         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  63
  64         /* allocate a r1bio with room for raid_disks entries in the bios array */
  65         return kzalloc(size, gfp_flags);
  66 }
  67
  68 static void r1bio_pool_free(void *r1_bio, void *data)
  69 {
  70         kfree(r1_bio);
  71 }
  72
  73 #define RESYNC_BLOCK_SIZE (64*1024)
  74 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  75 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  76 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  77 #define RESYNC_WINDOW (2048*1024)
  78
  79 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  80 {
  81         struct pool_info *pi = data;
  82         struct page *page;
  83         r1bio_t *r1_bio;
  84         struct bio *bio;
  85         int i, j;
  86
  87         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  88         if (!r1_bio)
  89                 return NULL;
  90
  91         /*
  92          * Allocate bios : 1 for reading, n-1 for writing
  93          */
  94         for (j = pi->raid_disks ; j-- ; ) {
  95                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
  96                 if (!bio)
  97                         goto out_free_bio;
  98                 r1_bio->bios[j] = bio;
  99         }
 100         /*
 101          * Allocate RESYNC_PAGES data pages and attach them to
 102          * the first bio.
 103          * If this is a user-requested check/repair, allocate
 104          * RESYNC_PAGES for each bio.
 105          */
 106         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 107                 j = pi->raid_disks;
 108         else
 109                 j = 1;
 110         while(j--) {
 111                 bio = r1_bio->bios[j];
 112                 for (i = 0; i < RESYNC_PAGES; i++) {
 113                         page = alloc_page(gfp_flags);
 114                         if (unlikely(!page))
 115                                 goto out_free_pages;
 116
 117                         bio->bi_io_vec[i].bv_page = page;
 118                         bio->bi_vcnt = i+1;
 119                 }
 120         }
 121         /* If not user-requests, copy the page pointers to all bios */
 122         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 123                 for (i=0; i<RESYNC_PAGES ; i++)
 124                         for (j=1; j<pi->raid_disks; j++)
 125                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 126                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 127         }
 128
 129         r1_bio->master_bio = NULL;
 130
 131         return r1_bio;
 132
 133 out_free_pages:
 134         for (j=0 ; j < pi->raid_disks; j++)
 135                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 136                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 137         j = -1;
 138 out_free_bio:
 139         while ( ++j < pi->raid_disks )
 140                 bio_put(r1_bio->bios[j]);
 141         r1bio_pool_free(r1_bio, data);
 142         return NULL;
 143 }
 144
 145 static void r1buf_pool_free(void *__r1_bio, void *data)
 146 {
 147         struct pool_info *pi = data;
 148         int i,j;
 149         r1bio_t *r1bio = __r1_bio;
 150
 151         for (i = 0; i < RESYNC_PAGES; i++)
 152                 for (j = pi->raid_disks; j-- ;) {
 153                         if (j == 0 ||
 154                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 155                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 156                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 157                 }
 158         for (i=0 ; i < pi->raid_disks; i++)
 159                 bio_put(r1bio->bios[i]);
 160
 161         r1bio_pool_free(r1bio, data);
 162 }
 163
 164 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 165 {
 166         int i;
 167
 168         for (i = 0; i < conf->raid_disks; i++) {
 169                 struct bio **bio = r1_bio->bios + i;
 170                 if (*bio && *bio != IO_BLOCKED)
 171                         bio_put(*bio);
 172                 *bio = NULL;
 173         }
 174 }
 175
 176 static void free_r1bio(r1bio_t *r1_bio)
 177 {
 178         conf_t *conf = r1_bio->mddev->private;
 179
 180         /*
 181          * Wake up any possible resync thread that waits for the device
 182          * to go idle.
 183          */
 184         allow_barrier(conf);
 185
 186         put_all_bios(conf, r1_bio);
 187         mempool_free(r1_bio, conf->r1bio_pool);
 188 }
 189
 190 static void put_buf(r1bio_t *r1_bio)
 191 {
 192         conf_t *conf = r1_bio->mddev->private;
 193         int i;
 194
 195         for (i=0; i<conf->raid_disks; i++) {
 196                 struct bio *bio = r1_bio->bios[i];
 197                 if (bio->bi_end_io)
 198                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 199         }
 200
 201         mempool_free(r1_bio, conf->r1buf_pool);
 202
 203         lower_barrier(conf);
 204 }
 205
 206 static void reschedule_retry(r1bio_t *r1_bio)
 207 {
 208         unsigned long flags;
 209         mddev_t *mddev = r1_bio->mddev;
 210         conf_t *conf = mddev->private;
 211
 212         spin_lock_irqsave(&conf->device_lock, flags);
 213         list_add(&r1_bio->retry_list, &conf->retry_list);
 214         conf->nr_queued ++;
 215         spin_unlock_irqrestore(&conf->device_lock, flags);
 216
 217         wake_up(&conf->wait_barrier);
 218         md_wakeup_thread(mddev->thread);
 219 }
 220
 221 /*
 222  * raid_end_bio_io() is called when we have finished servicing a mirrored
 223  * operation and are ready to return a success/failure code to the buffer
 224  * cache layer.
 225  */
 226 static void raid_end_bio_io(r1bio_t *r1_bio)
 227 {
 228         struct bio *bio = r1_bio->master_bio;
 229
 230         /* if nobody has done the final endio yet, do it now */
 231         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 232                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 233                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 234                         (unsigned long long) bio->bi_sector,
 235                         (unsigned long long) bio->bi_sector +
 236                                 (bio->bi_size >> 9) - 1);
 237
 238                 bio_endio(bio,
 239                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 240         }
 241         free_r1bio(r1_bio);
 242 }
 243
 244 /*
 245  * Update disk head position estimator based on IRQ completion info.
 246  */
 247 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 248 {
 249         conf_t *conf = r1_bio->mddev->private;
 250
 251         conf->mirrors[disk].head_position =
 252                 r1_bio->sector + (r1_bio->sectors);
 253 }
 254
 255 static void raid1_end_read_request(struct bio *bio, int error)
 256 {
 257         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 258         r1bio_t *r1_bio = bio->bi_private;
 259         int mirror;
 260         conf_t *conf = r1_bio->mddev->private;
 261
 262         mirror = r1_bio->read_disk;
 263         /*
 264          * this branch is our 'one mirror IO has finished' event handler:
 265          */
 266         update_head_pos(mirror, r1_bio);
 267
 268         if (uptodate)
 269                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 270         else {
 271                 /* If all other devices have failed, we want to return
 272                  * the error upwards rather than fail the last device.
 273                  * Here we redefine "uptodate" to mean "Don't want to retry"
 274                  */
 275                 unsigned long flags;
 276                 spin_lock_irqsave(&conf->device_lock, flags);
 277                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 278                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 279                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 280                         uptodate = 1;
 281                 spin_unlock_irqrestore(&conf->device_lock, flags);
 282         }
 283
 284         if (uptodate)
 285                 raid_end_bio_io(r1_bio);
 286         else {
 287                 /*
 288                  * oops, read error:
 289                  */
 290                 char b[BDEVNAME_SIZE];
 291                 printk_ratelimited(
 292                         KERN_ERR "md/raid1:%s: %s: "
 293                         "rescheduling sector %llu\n",
 294                         mdname(conf->mddev),
 295                         bdevname(conf->mirrors[mirror].rdev->bdev,
 296                                  b),
 297                         (unsigned long long)r1_bio->sector);
 298                 reschedule_retry(r1_bio);
 299         }
 300
 301         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 302 }
 303
 304 static void r1_bio_write_done(r1bio_t *r1_bio)
 305 {
 306         if (atomic_dec_and_test(&r1_bio->remaining))
 307         {
 308                 /* it really is the end of this request */
 309                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 310                         /* free extra copy of the data pages */
 311                         int i = r1_bio->behind_page_count;
 312                         while (i--)
 313                                 safe_put_page(r1_bio->behind_pages[i]);
 314                         kfree(r1_bio->behind_pages);
 315                         r1_bio->behind_pages = NULL;
 316                 }
 317                 /* clear the bitmap if all writes complete successfully */
 318                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 319                                 r1_bio->sectors,
 320                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
 321                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 322                 md_write_end(r1_bio->mddev);
 323                 raid_end_bio_io(r1_bio);
 324         }
 325 }
 326
 327 static void raid1_end_write_request(struct bio *bio, int error)
 328 {
 329         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 330         r1bio_t *r1_bio = bio->bi_private;
 331         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 332         conf_t *conf = r1_bio->mddev->private;
 333         struct bio *to_put = NULL;
 334
 335
 336         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 337                 if (r1_bio->bios[mirror] == bio)
 338                         break;
 339
 340         /*
 341          * 'one mirror IO has finished' event handler:
 342          */
 343         r1_bio->bios[mirror] = NULL;
 344         to_put = bio;
 345         if (!uptodate) {
 346                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 347                 /* an I/O failed, we can't clear the bitmap */
 348                 set_bit(R1BIO_Degraded, &r1_bio->state);
 349         } else
 350                 /*
 351                  * Set R1BIO_Uptodate in our master bio, so that we
 352                  * will return a good error code for to the higher
 353                  * levels even if IO on some other mirrored buffer
 354                  * fails.
 355                  *
 356                  * The 'master' represents the composite IO operation
 357                  * to user-side. So if something waits for IO, then it
 358                  * will wait for the 'master' bio.
 359                  */
 360                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 361
 362         update_head_pos(mirror, r1_bio);
 363
 364         if (behind) {
 365                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 366                         atomic_dec(&r1_bio->behind_remaining);
 367
 368                 /*
 369                  * In behind mode, we ACK the master bio once the I/O
 370                  * has safely reached all non-writemostly
 371                  * disks. Setting the Returned bit ensures that this
 372                  * gets done only once -- we don't ever want to return
 373                  * -EIO here, instead we'll wait
 374                  */
 375                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 376                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 377                         /* Maybe we can return now */
 378                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 379                                 struct bio *mbio = r1_bio->master_bio;
 380                                 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 381                                        (unsigned long long) mbio->bi_sector,
 382                                        (unsigned long long) mbio->bi_sector +
 383                                        (mbio->bi_size >> 9) - 1);
 384                                 bio_endio(mbio, 0);
 385                         }
 386                 }
 387         }
 388         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 389
 390         /*
 391          * Let's see if all mirrored write operations have finished
 392          * already.
 393          */
 394         r1_bio_write_done(r1_bio);
 395
 396         if (to_put)
 397                 bio_put(to_put);
 398 }
 399
 400
 401 /*
 402  * This routine returns the disk from which the requested read should
 403  * be done. There is a per-array 'next expected sequential IO' sector
 404  * number - if this matches on the next IO then we use the last disk.
 405  * There is also a per-disk 'last know head position' sector that is
 406  * maintained from IRQ contexts, both the normal and the resync IO
 407  * completion handlers update this position correctly. If there is no
 408  * perfect sequential match then we pick the disk whose head is closest.
 409  *
 410  * If there are 2 mirrors in the same 2 devices, performance degrades
 411  * because position is mirror, not device based.
 412  *
 413  * The rdev for the device selected will have nr_pending incremented.
 414  */
 415 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 416 {
 417         const sector_t this_sector = r1_bio->sector;
 418         const int sectors = r1_bio->sectors;
 419         int start_disk;
 420         int best_disk;
 421         int i;
 422         sector_t best_dist;
 423         mdk_rdev_t *rdev;
 424         int choose_first;
 425
 426         rcu_read_lock();
 427         /*
 428          * Check if we can balance. We can balance on the whole
 429          * device if no resync is going on, or below the resync window.
 430          * We take the first readable disk when above the resync window.
 431          */
 432  retry:
 433         best_disk = -1;
 434         best_dist = MaxSector;
 435         if (conf->mddev->recovery_cp < MaxSector &&
 436             (this_sector + sectors >= conf->next_resync)) {
 437                 choose_first = 1;
 438                 start_disk = 0;
 439         } else {
 440                 choose_first = 0;
 441                 start_disk = conf->last_used;
 442         }
 443
 444         for (i = 0 ; i < conf->raid_disks ; i++) {
 445                 sector_t dist;
 446                 int disk = start_disk + i;
 447                 if (disk >= conf->raid_disks)
 448                         disk -= conf->raid_disks;
 449
 450                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 451                 if (r1_bio->bios[disk] == IO_BLOCKED
 452                     || rdev == NULL
 453                     || test_bit(Faulty, &rdev->flags))
 454                         continue;
 455                 if (!test_bit(In_sync, &rdev->flags) &&
 456                     rdev->recovery_offset < this_sector + sectors)
 457                         continue;
 458                 if (test_bit(WriteMostly, &rdev->flags)) {
 459                         /* Don't balance among write-mostly, just
 460                          * use the first as a last resort */
 461                         if (best_disk < 0)
 462                                 best_disk = disk;
 463                         continue;
 464                 }
 465                 /* This is a reasonable device to use.  It might
 466                  * even be best.
 467                  */
 468                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 469                 if (choose_first
 470                     /* Don't change to another disk for sequential reads */
 471                     || conf->next_seq_sect == this_sector
 472                     || dist == 0
 473                     /* If device is idle, use it */
 474                     || atomic_read(&rdev->nr_pending) == 0) {
 475                         best_disk = disk;
 476                         break;
 477                 }
 478                 if (dist < best_dist) {
 479                         best_dist = dist;
 480                         best_disk = disk;
 481                 }
 482         }
 483
 484         if (best_disk >= 0) {
 485                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 486                 if (!rdev)
 487                         goto retry;
 488                 atomic_inc(&rdev->nr_pending);
 489                 if (test_bit(Faulty, &rdev->flags)) {
 490                         /* cannot risk returning a device that failed
 491                          * before we inc'ed nr_pending
 492                          */
 493                         rdev_dec_pending(rdev, conf->mddev);
 494                         goto retry;
 495                 }
 496                 conf->next_seq_sect = this_sector + sectors;
 497                 conf->last_used = best_disk;
 498         }
 499         rcu_read_unlock();
 500
 501         return best_disk;
 502 }
 503
 504 int md_raid1_congested(mddev_t *mddev, int bits)
 505 {
 506         conf_t *conf = mddev->private;
 507         int i, ret = 0;
 508
 509         rcu_read_lock();
 510         for (i = 0; i < mddev->raid_disks; i++) {
 511                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 512                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 513                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 514
 515                         BUG_ON(!q);
 516
 517                         /* Note the '|| 1' - when read_balance prefers
 518                          * non-congested targets, it can be removed
 519                          */
 520                         if ((bits & (1<<BDI_async_congested)) || 1)
 521                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 522                         else
 523                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 524                 }
 525         }
 526         rcu_read_unlock();
 527         return ret;
 528 }
 529 EXPORT_SYMBOL_GPL(md_raid1_congested);
 530
 531 static int raid1_congested(void *data, int bits)
 532 {
 533         mddev_t *mddev = data;
 534
 535         return mddev_congested(mddev, bits) ||
 536                 md_raid1_congested(mddev, bits);
 537 }
 538
 539 static void flush_pending_writes(conf_t *conf)
 540 {
 541         /* Any writes that have been queued but are awaiting
 542          * bitmap updates get flushed here.
 543          */
 544         spin_lock_irq(&conf->device_lock);
 545
 546         if (conf->pending_bio_list.head) {
 547                 struct bio *bio;
 548                 bio = bio_list_get(&conf->pending_bio_list);
 549                 spin_unlock_irq(&conf->device_lock);
 550                 /* flush any pending bitmap writes to
 551                  * disk before proceeding w/ I/O */
 552                 bitmap_unplug(conf->mddev->bitmap);
 553
 554                 while (bio) { /* submit pending writes */
 555                         struct bio *next = bio->bi_next;
 556                         bio->bi_next = NULL;
 557                         generic_make_request(bio);
 558                         bio = next;
 559                 }
 560         } else
 561                 spin_unlock_irq(&conf->device_lock);
 562 }
 563
 564 /* Barriers....
 565  * Sometimes we need to suspend IO while we do something else,
 566  * either some resync/recovery, or reconfigure the array.
 567  * To do this we raise a 'barrier'.
 568  * The 'barrier' is a counter that can be raised multiple times
 569  * to count how many activities are happening which preclude
 570  * normal IO.
 571  * We can only raise the barrier if there is no pending IO.
 572  * i.e. if nr_pending == 0.
 573  * We choose only to raise the barrier if no-one is waiting for the
 574  * barrier to go down.  This means that as soon as an IO request
 575  * is ready, no other operations which require a barrier will start
 576  * until the IO request has had a chance.
 577  *
 578  * So: regular IO calls 'wait_barrier'.  When that returns there
 579  *    is no backgroup IO happening,  It must arrange to call
 580  *    allow_barrier when it has finished its IO.
 581  * backgroup IO calls must call raise_barrier.  Once that returns
 582  *    there is no normal IO happeing.  It must arrange to call
 583  *    lower_barrier when the particular background IO completes.
 584  */
 585 #define RESYNC_DEPTH 32
 586
 587 static void raise_barrier(conf_t *conf)
 588 {
 589         spin_lock_irq(&conf->resync_lock);
 590
 591         /* Wait until no block IO is waiting */
 592         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 593                             conf->resync_lock, );
 594
 595         /* block any new IO from starting */
 596         conf->barrier++;
 597
 598         /* Now wait for all pending IO to complete */
 599         wait_event_lock_irq(conf->wait_barrier,
 600                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 601                             conf->resync_lock, );
 602
 603         spin_unlock_irq(&conf->resync_lock);
 604 }
 605
 606 static void lower_barrier(conf_t *conf)
 607 {
 608         unsigned long flags;
 609         BUG_ON(conf->barrier <= 0);
 610         spin_lock_irqsave(&conf->resync_lock, flags);
 611         conf->barrier--;
 612         spin_unlock_irqrestore(&conf->resync_lock, flags);
 613         wake_up(&conf->wait_barrier);
 614 }
 615
 616 static void wait_barrier(conf_t *conf)
 617 {
 618         spin_lock_irq(&conf->resync_lock);
 619         if (conf->barrier) {
 620                 conf->nr_waiting++;
 621                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 622                                     conf->resync_lock,
 623                                     );
 624                 conf->nr_waiting--;
 625         }
 626         conf->nr_pending++;
 627         spin_unlock_irq(&conf->resync_lock);
 628 }
 629
 630 static void allow_barrier(conf_t *conf)
 631 {
 632         unsigned long flags;
 633         spin_lock_irqsave(&conf->resync_lock, flags);
 634         conf->nr_pending--;
 635         spin_unlock_irqrestore(&conf->resync_lock, flags);
 636         wake_up(&conf->wait_barrier);
 637 }
 638
 639 static void freeze_array(conf_t *conf)
 640 {
 641         /* stop syncio and normal IO and wait for everything to
 642          * go quite.
 643          * We increment barrier and nr_waiting, and then
 644          * wait until nr_pending match nr_queued+1
 645          * This is called in the context of one normal IO request
 646          * that has failed. Thus any sync request that might be pending
 647          * will be blocked by nr_pending, and we need to wait for
 648          * pending IO requests to complete or be queued for re-try.
 649          * Thus the number queued (nr_queued) plus this request (1)
 650          * must match the number of pending IOs (nr_pending) before
 651          * we continue.
 652          */
 653         spin_lock_irq(&conf->resync_lock);
 654         conf->barrier++;
 655         conf->nr_waiting++;
 656         wait_event_lock_irq(conf->wait_barrier,
 657                             conf->nr_pending == conf->nr_queued+1,
 658                             conf->resync_lock,
 659                             flush_pending_writes(conf));
 660         spin_unlock_irq(&conf->resync_lock);
 661 }
 662 static void unfreeze_array(conf_t *conf)
 663 {
 664         /* reverse the effect of the freeze */
 665         spin_lock_irq(&conf->resync_lock);
 666         conf->barrier--;
 667         conf->nr_waiting--;
 668         wake_up(&conf->wait_barrier);
 669         spin_unlock_irq(&conf->resync_lock);
 670 }
 671
 672
 673 /* duplicate the data pages for behind I/O
 674  */
 675 static void alloc_behind_pages(struct bio *bio, r1bio_t *r1_bio)
 676 {
 677         int i;
 678         struct bio_vec *bvec;
 679         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page*),
 680                                         GFP_NOIO);
 681         if (unlikely(!pages))
 682                 return;
 683
 684         bio_for_each_segment(bvec, bio, i) {
 685                 pages[i] = alloc_page(GFP_NOIO);
 686                 if (unlikely(!pages[i]))
 687                         goto do_sync_io;
 688                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 689                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 690                 kunmap(pages[i]);
 691                 kunmap(bvec->bv_page);
 692         }
 693         r1_bio->behind_pages = pages;
 694         r1_bio->behind_page_count = bio->bi_vcnt;
 695         set_bit(R1BIO_BehindIO, &r1_bio->state);
 696         return;
 697
 698 do_sync_io:
 699         for (i = 0; i < bio->bi_vcnt; i++)
 700                 if (pages[i])
 701                         put_page(pages[i]);
 702         kfree(pages);
 703         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 704 }
 705
 706 static int make_request(mddev_t *mddev, struct bio * bio)
 707 {
 708         conf_t *conf = mddev->private;
 709         mirror_info_t *mirror;
 710         r1bio_t *r1_bio;
 711         struct bio *read_bio;
 712         int i, targets = 0, disks;
 713         struct bitmap *bitmap;
 714         unsigned long flags;
 715         const int rw = bio_data_dir(bio);
 716         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 717         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 718         mdk_rdev_t *blocked_rdev;
 719         int plugged;
 720
 721         /*
 722          * Register the new request and wait if the reconstruction
 723          * thread has put up a bar for new requests.
 724          * Continue immediately if no resync is active currently.
 725          */
 726
 727         md_write_start(mddev, bio); /* wait on superblock update early */
 728
 729         if (bio_data_dir(bio) == WRITE &&
 730             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 731             bio->bi_sector < mddev->suspend_hi) {
 732                 /* As the suspend_* range is controlled by
 733                  * userspace, we want an interruptible
 734                  * wait.
 735                  */
 736                 DEFINE_WAIT(w);
 737                 for (;;) {
 738                         flush_signals(current);
 739                         prepare_to_wait(&conf->wait_barrier,
 740                                         &w, TASK_INTERRUPTIBLE);
 741                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 742                             bio->bi_sector >= mddev->suspend_hi)
 743                                 break;
 744                         schedule();
 745                 }
 746                 finish_wait(&conf->wait_barrier, &w);
 747         }
 748
 749         wait_barrier(conf);
 750
 751         bitmap = mddev->bitmap;
 752
 753         /*
 754          * make_request() can abort the operation when READA is being
 755          * used and no empty request is available.
 756          *
 757          */
 758         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 759
 760         r1_bio->master_bio = bio;
 761         r1_bio->sectors = bio->bi_size >> 9;
 762         r1_bio->state = 0;
 763         r1_bio->mddev = mddev;
 764         r1_bio->sector = bio->bi_sector;
 765
 766         if (rw == READ) {
 767                 /*
 768                  * read balancing logic:
 769                  */
 770                 int rdisk = read_balance(conf, r1_bio);
 771
 772                 if (rdisk < 0) {
 773                         /* couldn't find anywhere to read from */
 774                         raid_end_bio_io(r1_bio);
 775                         return 0;
 776                 }
 777                 mirror = conf->mirrors + rdisk;
 778
 779                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 780                     bitmap) {
 781                         /* Reading from a write-mostly device must
 782                          * take care not to over-take any writes
 783                          * that are 'behind'
 784                          */
 785                         wait_event(bitmap->behind_wait,
 786                                    atomic_read(&bitmap->behind_writes) == 0);
 787                 }
 788                 r1_bio->read_disk = rdisk;
 789
 790                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 791
 792                 r1_bio->bios[rdisk] = read_bio;
 793
 794                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 795                 read_bio->bi_bdev = mirror->rdev->bdev;
 796                 read_bio->bi_end_io = raid1_end_read_request;
 797                 read_bio->bi_rw = READ | do_sync;
 798                 read_bio->bi_private = r1_bio;
 799
 800                 generic_make_request(read_bio);
 801                 return 0;
 802         }
 803
 804         /*
 805          * WRITE:
 806          */
 807         /* first select target devices under spinlock and
 808          * inc refcount on their rdev.  Record them by setting
 809          * bios[x] to bio
 810          */
 811         plugged = mddev_check_plugged(mddev);
 812
 813         disks = conf->raid_disks;
 814  retry_write:
 815         blocked_rdev = NULL;
 816         rcu_read_lock();
 817         for (i = 0;  i < disks; i++) {
 818                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 819                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 820                         atomic_inc(&rdev->nr_pending);
 821                         blocked_rdev = rdev;
 822                         break;
 823                 }
 824                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 825                         atomic_inc(&rdev->nr_pending);
 826                         if (test_bit(Faulty, &rdev->flags)) {
 827                                 rdev_dec_pending(rdev, mddev);
 828                                 r1_bio->bios[i] = NULL;
 829                         } else {
 830                                 r1_bio->bios[i] = bio;
 831                                 targets++;
 832                         }
 833                 } else
 834                         r1_bio->bios[i] = NULL;
 835         }
 836         rcu_read_unlock();
 837
 838         if (unlikely(blocked_rdev)) {
 839                 /* Wait for this device to become unblocked */
 840                 int j;
 841
 842                 for (j = 0; j < i; j++)
 843                         if (r1_bio->bios[j])
 844                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 845
 846                 allow_barrier(conf);
 847                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 848                 wait_barrier(conf);
 849                 goto retry_write;
 850         }
 851
 852         BUG_ON(targets == 0); /* we never fail the last device */
 853
 854         if (targets < conf->raid_disks) {
 855                 /* array is degraded, we will not clear the bitmap
 856                  * on I/O completion (see raid1_end_write_request) */
 857                 set_bit(R1BIO_Degraded, &r1_bio->state);
 858         }
 859
 860         /* do behind I/O ?
 861          * Not if there are too many, or cannot allocate memory,
 862          * or a reader on WriteMostly is waiting for behind writes
 863          * to flush */
 864         if (bitmap &&
 865             (atomic_read(&bitmap->behind_writes)
 866              < mddev->bitmap_info.max_write_behind) &&
 867             !waitqueue_active(&bitmap->behind_wait))
 868                 alloc_behind_pages(bio, r1_bio);
 869
 870         atomic_set(&r1_bio->remaining, 1);
 871         atomic_set(&r1_bio->behind_remaining, 0);
 872
 873         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 874                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 875         for (i = 0; i < disks; i++) {
 876                 struct bio *mbio;
 877                 if (!r1_bio->bios[i])
 878                         continue;
 879
 880                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 881                 r1_bio->bios[i] = mbio;
 882
 883                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 884                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 885                 mbio->bi_end_io = raid1_end_write_request;
 886                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
 887                 mbio->bi_private = r1_bio;
 888
 889                 if (r1_bio->behind_pages) {
 890                         struct bio_vec *bvec;
 891                         int j;
 892
 893                         /* Yes, I really want the '__' version so that
 894                          * we clear any unused pointer in the io_vec, rather
 895                          * than leave them unchanged.  This is important
 896                          * because when we come to free the pages, we won't
 897                          * know the original bi_idx, so we just free
 898                          * them all
 899                          */
 900                         __bio_for_each_segment(bvec, mbio, j, 0)
 901                                 bvec->bv_page = r1_bio->behind_pages[j];
 902                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 903                                 atomic_inc(&r1_bio->behind_remaining);
 904                 }
 905
 906                 atomic_inc(&r1_bio->remaining);
 907                 spin_lock_irqsave(&conf->device_lock, flags);
 908                 bio_list_add(&conf->pending_bio_list, mbio);
 909                 spin_unlock_irqrestore(&conf->device_lock, flags);
 910         }
 911         r1_bio_write_done(r1_bio);
 912
 913         /* In case raid1d snuck in to freeze_array */
 914         wake_up(&conf->wait_barrier);
 915
 916         if (do_sync || !bitmap || !plugged)
 917                 md_wakeup_thread(mddev->thread);
 918
 919         return 0;
 920 }
 921
 922 static void status(struct seq_file *seq, mddev_t *mddev)
 923 {
 924         conf_t *conf = mddev->private;
 925         int i;
 926
 927         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 928                    conf->raid_disks - mddev->degraded);
 929         rcu_read_lock();
 930         for (i = 0; i < conf->raid_disks; i++) {
 931                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 932                 seq_printf(seq, "%s",
 933                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
 934         }
 935         rcu_read_unlock();
 936         seq_printf(seq, "]");
 937 }
 938
 939
 940 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 941 {
 942         char b[BDEVNAME_SIZE];
 943         conf_t *conf = mddev->private;
 944
 945         /*
 946          * If it is not operational, then we have already marked it as dead
 947          * else if it is the last working disks, ignore the error, let the
 948          * next level up know.
 949          * else mark the drive as failed
 950          */
 951         if (test_bit(In_sync, &rdev->flags)
 952             && (conf->raid_disks - mddev->degraded) == 1) {
 953                 /*
 954                  * Don't fail the drive, act as though we were just a
 955                  * normal single drive.
 956                  * However don't try a recovery from this drive as
 957                  * it is very likely to fail.
 958                  */
 959                 mddev->recovery_disabled = 1;
 960                 return;
 961         }
 962         if (test_and_clear_bit(In_sync, &rdev->flags)) {
 963                 unsigned long flags;
 964                 spin_lock_irqsave(&conf->device_lock, flags);
 965                 mddev->degraded++;
 966                 set_bit(Faulty, &rdev->flags);
 967                 spin_unlock_irqrestore(&conf->device_lock, flags);
 968                 /*
 969                  * if recovery is running, make sure it aborts.
 970                  */
 971                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
 972         } else
 973                 set_bit(Faulty, &rdev->flags);
 974         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 975         printk(KERN_ALERT
 976                "md/raid1:%s: Disk failure on %s, disabling device.\n"
 977                "md/raid1:%s: Operation continuing on %d devices.\n",
 978                mdname(mddev), bdevname(rdev->bdev, b),
 979                mdname(mddev), conf->raid_disks - mddev->degraded);
 980 }
 981
 982 static void print_conf(conf_t *conf)
 983 {
 984         int i;
 985
 986         printk(KERN_DEBUG "RAID1 conf printout:\n");
 987         if (!conf) {
 988                 printk(KERN_DEBUG "(!conf)\n");
 989                 return;
 990         }
 991         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
 992                 conf->raid_disks);
 993
 994         rcu_read_lock();
 995         for (i = 0; i < conf->raid_disks; i++) {
 996                 char b[BDEVNAME_SIZE];
 997                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 998                 if (rdev)
 999                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1000                                i, !test_bit(In_sync, &rdev->flags),
1001                                !test_bit(Faulty, &rdev->flags),
1002                                bdevname(rdev->bdev,b));
1003         }
1004         rcu_read_unlock();
1005 }
1006
1007 static void close_sync(conf_t *conf)
1008 {
1009         wait_barrier(conf);
1010         allow_barrier(conf);
1011
1012         mempool_destroy(conf->r1buf_pool);
1013         conf->r1buf_pool = NULL;
1014 }
1015
1016 static int raid1_spare_active(mddev_t *mddev)
1017 {
1018         int i;
1019         conf_t *conf = mddev->private;
1020         int count = 0;
1021         unsigned long flags;
1022
1023         /*
1024          * Find all failed disks within the RAID1 configuration
1025          * and mark them readable.
1026          * Called under mddev lock, so rcu protection not needed.
1027          */
1028         for (i = 0; i < conf->raid_disks; i++) {
1029                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1030                 if (rdev
1031                     && !test_bit(Faulty, &rdev->flags)
1032                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1033                         count++;
1034                         sysfs_notify_dirent(rdev->sysfs_state);
1035                 }
1036         }
1037         spin_lock_irqsave(&conf->device_lock, flags);
1038         mddev->degraded -= count;
1039         spin_unlock_irqrestore(&conf->device_lock, flags);
1040
1041         print_conf(conf);
1042         return count;
1043 }
1044
1045
1046 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1047 {
1048         conf_t *conf = mddev->private;
1049         int err = -EEXIST;
1050         int mirror = 0;
1051         mirror_info_t *p;
1052         int first = 0;
1053         int last = mddev->raid_disks - 1;
1054
1055         if (rdev->raid_disk >= 0)
1056                 first = last = rdev->raid_disk;
1057
1058         for (mirror = first; mirror <= last; mirror++)
1059                 if ( !(p=conf->mirrors+mirror)->rdev) {
1060
1061                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1062                                           rdev->data_offset << 9);
1063                         /* as we don't honour merge_bvec_fn, we must
1064                          * never risk violating it, so limit
1065                          * ->max_segments to one lying with a single
1066                          * page, as a one page request is never in
1067                          * violation.
1068                          */
1069                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1070                                 blk_queue_max_segments(mddev->queue, 1);
1071                                 blk_queue_segment_boundary(mddev->queue,
1072                                                            PAGE_CACHE_SIZE - 1);
1073                         }
1074
1075                         p->head_position = 0;
1076                         rdev->raid_disk = mirror;
1077                         err = 0;
1078                         /* As all devices are equivalent, we don't need a full recovery
1079                          * if this was recently any drive of the array
1080                          */
1081                         if (rdev->saved_raid_disk < 0)
1082                                 conf->fullsync = 1;
1083                         rcu_assign_pointer(p->rdev, rdev);
1084                         break;
1085                 }
1086         md_integrity_add_rdev(rdev, mddev);
1087         print_conf(conf);
1088         return err;
1089 }
1090
1091 static int raid1_remove_disk(mddev_t *mddev, int number)
1092 {
1093         conf_t *conf = mddev->private;
1094         int err = 0;
1095         mdk_rdev_t *rdev;
1096         mirror_info_t *p = conf->mirrors+ number;
1097
1098         print_conf(conf);
1099         rdev = p->rdev;
1100         if (rdev) {
1101                 if (test_bit(In_sync, &rdev->flags) ||
1102                     atomic_read(&rdev->nr_pending)) {
1103                         err = -EBUSY;
1104                         goto abort;
1105                 }
1106                 /* Only remove non-faulty devices if recovery
1107                  * is not possible.
1108                  */
1109                 if (!test_bit(Faulty, &rdev->flags) &&
1110                     !mddev->recovery_disabled &&
1111                     mddev->degraded < conf->raid_disks) {
1112                         err = -EBUSY;
1113                         goto abort;
1114                 }
1115                 p->rdev = NULL;
1116                 synchronize_rcu();
1117                 if (atomic_read(&rdev->nr_pending)) {
1118                         /* lost the race, try later */
1119                         err = -EBUSY;
1120                         p->rdev = rdev;
1121                         goto abort;
1122                 }
1123                 err = md_integrity_register(mddev);
1124         }
1125 abort:
1126
1127         print_conf(conf);
1128         return err;
1129 }
1130
1131
1132 static void end_sync_read(struct bio *bio, int error)
1133 {
1134         r1bio_t *r1_bio = bio->bi_private;
1135         int i;
1136
1137         for (i=r1_bio->mddev->raid_disks; i--; )
1138                 if (r1_bio->bios[i] == bio)
1139                         break;
1140         BUG_ON(i < 0);
1141         update_head_pos(i, r1_bio);
1142         /*
1143          * we have read a block, now it needs to be re-written,
1144          * or re-read if the read failed.
1145          * We don't do much here, just schedule handling by raid1d
1146          */
1147         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1148                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1149
1150         if (atomic_dec_and_test(&r1_bio->remaining))
1151                 reschedule_retry(r1_bio);
1152 }
1153
1154 static void end_sync_write(struct bio *bio, int error)
1155 {
1156         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1157         r1bio_t *r1_bio = bio->bi_private;
1158         mddev_t *mddev = r1_bio->mddev;
1159         conf_t *conf = mddev->private;
1160         int i;
1161         int mirror=0;
1162
1163         for (i = 0; i < conf->raid_disks; i++)
1164                 if (r1_bio->bios[i] == bio) {
1165                         mirror = i;
1166                         break;
1167                 }
1168         if (!uptodate) {
1169                 sector_t sync_blocks = 0;
1170                 sector_t s = r1_bio->sector;
1171                 long sectors_to_go = r1_bio->sectors;
1172                 /* make sure these bits doesn't get cleared. */
1173                 do {
1174                         bitmap_end_sync(mddev->bitmap, s,
1175                                         &sync_blocks, 1);
1176                         s += sync_blocks;
1177                         sectors_to_go -= sync_blocks;
1178                 } while (sectors_to_go > 0);
1179                 md_error(mddev, conf->mirrors[mirror].rdev);
1180         }
1181
1182         update_head_pos(mirror, r1_bio);
1183
1184         if (atomic_dec_and_test(&r1_bio->remaining)) {
1185                 sector_t s = r1_bio->sectors;
1186                 put_buf(r1_bio);
1187                 md_done_sync(mddev, s, uptodate);
1188         }
1189 }
1190
1191 static int fix_sync_read_error(r1bio_t *r1_bio)
1192 {
1193         /* Try some synchronous reads of other devices to get
1194          * good data, much like with normal read errors.  Only
1195          * read into the pages we already have so we don't
1196          * need to re-issue the read request.
1197          * We don't need to freeze the array, because being in an
1198          * active sync request, there is no normal IO, and
1199          * no overlapping syncs.
1200          */
1201         mddev_t *mddev = r1_bio->mddev;
1202         conf_t *conf = mddev->private;
1203         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1204         sector_t sect = r1_bio->sector;
1205         int sectors = r1_bio->sectors;
1206         int idx = 0;
1207
1208         while(sectors) {
1209                 int s = sectors;
1210                 int d = r1_bio->read_disk;
1211                 int success = 0;
1212                 mdk_rdev_t *rdev;
1213                 int start;
1214
1215                 if (s > (PAGE_SIZE>>9))
1216                         s = PAGE_SIZE >> 9;
1217                 do {
1218                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1219                                 /* No rcu protection needed here devices
1220                                  * can only be removed when no resync is
1221                                  * active, and resync is currently active
1222                                  */
1223                                 rdev = conf->mirrors[d].rdev;
1224                                 if (sync_page_io(rdev,
1225                                                  sect,
1226                                                  s<<9,
1227                                                  bio->bi_io_vec[idx].bv_page,
1228                                                  READ, false)) {
1229                                         success = 1;
1230                                         break;
1231                                 }
1232                         }
1233                         d++;
1234                         if (d == conf->raid_disks)
1235                                 d = 0;
1236                 } while (!success && d != r1_bio->read_disk);
1237
1238                 if (!success) {
1239                         char b[BDEVNAME_SIZE];
1240                         /* Cannot read from anywhere, array is toast */
1241                         md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1242                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1243                                " for block %llu\n",
1244                                mdname(mddev),
1245                                bdevname(bio->bi_bdev, b),
1246                                (unsigned long long)r1_bio->sector);
1247                         md_done_sync(mddev, r1_bio->sectors, 0);
1248                         put_buf(r1_bio);
1249                         return 0;
1250                 }
1251
1252                 start = d;
1253                 /* write it back and re-read */
1254                 while (d != r1_bio->read_disk) {
1255                         if (d == 0)
1256                                 d = conf->raid_disks;
1257                         d--;
1258                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1259                                 continue;
1260                         rdev = conf->mirrors[d].rdev;
1261                         if (sync_page_io(rdev,
1262                                          sect,
1263                                          s<<9,
1264                                          bio->bi_io_vec[idx].bv_page,
1265                                          WRITE, false) == 0) {
1266                                 r1_bio->bios[d]->bi_end_io = NULL;
1267                                 rdev_dec_pending(rdev, mddev);
1268                                 md_error(mddev, rdev);
1269                         } else
1270                                 atomic_add(s, &rdev->corrected_errors);
1271                 }
1272                 d = start;
1273                 while (d != r1_bio->read_disk) {
1274                         if (d == 0)
1275                                 d = conf->raid_disks;
1276                         d--;
1277                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1278                                 continue;
1279                         rdev = conf->mirrors[d].rdev;
1280                         if (sync_page_io(rdev,
1281                                          sect,
1282                                          s<<9,
1283                                          bio->bi_io_vec[idx].bv_page,
1284                                          READ, false) == 0)
1285                                 md_error(mddev, rdev);
1286                 }
1287                 sectors -= s;
1288                 sect += s;
1289                 idx ++;
1290         }
1291         set_bit(R1BIO_Uptodate, &r1_bio->state);
1292         set_bit(BIO_UPTODATE, &bio->bi_flags);
1293         return 1;
1294 }
1295
1296 static int process_checks(r1bio_t *r1_bio)
1297 {
1298         /* We have read all readable devices.  If we haven't
1299          * got the block, then there is no hope left.
1300          * If we have, then we want to do a comparison
1301          * and skip the write if everything is the same.
1302          * If any blocks failed to read, then we need to
1303          * attempt an over-write
1304          */
1305         mddev_t *mddev = r1_bio->mddev;
1306         conf_t *conf = mddev->private;
1307         int primary;
1308         int i;
1309
1310         for (primary = 0; primary < conf->raid_disks; primary++)
1311                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1312                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1313                         r1_bio->bios[primary]->bi_end_io = NULL;
1314                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1315                         break;
1316                 }
1317         r1_bio->read_disk = primary;
1318         for (i = 0; i < conf->raid_disks; i++) {
1319                 int j;
1320                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1321                 struct bio *pbio = r1_bio->bios[primary];
1322                 struct bio *sbio = r1_bio->bios[i];
1323                 int size;
1324
1325                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1326                         continue;
1327
1328                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1329                         for (j = vcnt; j-- ; ) {
1330                                 struct page *p, *s;
1331                                 p = pbio->bi_io_vec[j].bv_page;
1332                                 s = sbio->bi_io_vec[j].bv_page;
1333                                 if (memcmp(page_address(p),
1334                                            page_address(s),
1335                                            PAGE_SIZE))
1336                                         break;
1337                         }
1338                 } else
1339                         j = 0;
1340                 if (j >= 0)
1341                         mddev->resync_mismatches += r1_bio->sectors;
1342                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1343                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1344                         /* No need to write to this device. */
1345                         sbio->bi_end_io = NULL;
1346                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1347                         continue;
1348                 }
1349                 /* fixup the bio for reuse */
1350                 sbio->bi_vcnt = vcnt;
1351                 sbio->bi_size = r1_bio->sectors << 9;
1352                 sbio->bi_idx = 0;
1353                 sbio->bi_phys_segments = 0;
1354                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1355                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1356                 sbio->bi_next = NULL;
1357                 sbio->bi_sector = r1_bio->sector +
1358                         conf->mirrors[i].rdev->data_offset;
1359                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1360                 size = sbio->bi_size;
1361                 for (j = 0; j < vcnt ; j++) {
1362                         struct bio_vec *bi;
1363                         bi = &sbio->bi_io_vec[j];
1364                         bi->bv_offset = 0;
1365                         if (size > PAGE_SIZE)
1366                                 bi->bv_len = PAGE_SIZE;
1367                         else
1368                                 bi->bv_len = size;
1369                         size -= PAGE_SIZE;
1370                         memcpy(page_address(bi->bv_page),
1371                                page_address(pbio->bi_io_vec[j].bv_page),
1372                                PAGE_SIZE);
1373                 }
1374         }
1375         return 0;
1376 }
1377
1378 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1379 {
1380         conf_t *conf = mddev->private;
1381         int i;
1382         int disks = conf->raid_disks;
1383         struct bio *bio, *wbio;
1384
1385         bio = r1_bio->bios[r1_bio->read_disk];
1386
1387         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1388                 /* ouch - failed to read all of that. */
1389                 if (!fix_sync_read_error(r1_bio))
1390                         return;
1391
1392         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1393                 if (process_checks(r1_bio) < 0)
1394                         return;
1395         /*
1396          * schedule writes
1397          */
1398         atomic_set(&r1_bio->remaining, 1);
1399         for (i = 0; i < disks ; i++) {
1400                 wbio = r1_bio->bios[i];
1401                 if (wbio->bi_end_io == NULL ||
1402                     (wbio->bi_end_io == end_sync_read &&
1403                      (i == r1_bio->read_disk ||
1404                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1405                         continue;
1406
1407                 wbio->bi_rw = WRITE;
1408                 wbio->bi_end_io = end_sync_write;
1409                 atomic_inc(&r1_bio->remaining);
1410                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1411
1412                 generic_make_request(wbio);
1413         }
1414
1415         if (atomic_dec_and_test(&r1_bio->remaining)) {
1416                 /* if we're here, all write(s) have completed, so clean up */
1417                 md_done_sync(mddev, r1_bio->sectors, 1);
1418                 put_buf(r1_bio);
1419         }
1420 }
1421
1422 /*
1423  * This is a kernel thread which:
1424  *
1425  *      1.      Retries failed read operations on working mirrors.
1426  *      2.      Updates the raid superblock when problems encounter.
1427  *      3.      Performs writes following reads for array syncronising.
1428  */
1429
1430 static void fix_read_error(conf_t *conf, int read_disk,
1431                            sector_t sect, int sectors)
1432 {
1433         mddev_t *mddev = conf->mddev;
1434         while(sectors) {
1435                 int s = sectors;
1436                 int d = read_disk;
1437                 int success = 0;
1438                 int start;
1439                 mdk_rdev_t *rdev;
1440
1441                 if (s > (PAGE_SIZE>>9))
1442                         s = PAGE_SIZE >> 9;
1443
1444                 do {
1445                         /* Note: no rcu protection needed here
1446                          * as this is synchronous in the raid1d thread
1447                          * which is the thread that might remove
1448                          * a device.  If raid1d ever becomes multi-threaded....
1449                          */
1450                         rdev = conf->mirrors[d].rdev;
1451                         if (rdev &&
1452                             test_bit(In_sync, &rdev->flags) &&
1453                             sync_page_io(rdev, sect, s<<9,
1454                                          conf->tmppage, READ, false))
1455                                 success = 1;
1456                         else {
1457                                 d++;
1458                                 if (d == conf->raid_disks)
1459                                         d = 0;
1460                         }
1461                 } while (!success && d != read_disk);
1462
1463                 if (!success) {
1464                         /* Cannot read from anywhere -- bye bye array */
1465                         md_error(mddev, conf->mirrors[read_disk].rdev);
1466                         break;
1467                 }
1468                 /* write it back and re-read */
1469                 start = d;
1470                 while (d != read_disk) {
1471                         if (d==0)
1472                                 d = conf->raid_disks;
1473                         d--;
1474                         rdev = conf->mirrors[d].rdev;
1475                         if (rdev &&
1476                             test_bit(In_sync, &rdev->flags)) {
1477                                 if (sync_page_io(rdev, sect, s<<9,
1478                                                  conf->tmppage, WRITE, false)
1479                                     == 0)
1480                                         /* Well, this device is dead */
1481                                         md_error(mddev, rdev);
1482                         }
1483                 }
1484                 d = start;
1485                 while (d != read_disk) {
1486                         char b[BDEVNAME_SIZE];
1487                         if (d==0)
1488                                 d = conf->raid_disks;
1489                         d--;
1490                         rdev = conf->mirrors[d].rdev;
1491                         if (rdev &&
1492                             test_bit(In_sync, &rdev->flags)) {
1493                                 if (sync_page_io(rdev, sect, s<<9,
1494                                                  conf->tmppage, READ, false)
1495                                     == 0)
1496                                         /* Well, this device is dead */
1497                                         md_error(mddev, rdev);
1498                                 else {
1499                                         atomic_add(s, &rdev->corrected_errors);
1500                                         printk(KERN_INFO
1501                                                "md/raid1:%s: read error corrected "
1502                                                "(%d sectors at %llu on %s)\n",
1503                                                mdname(mddev), s,
1504                                                (unsigned long long)(sect +
1505                                                    rdev->data_offset),
1506                                                bdevname(rdev->bdev, b));
1507                                 }
1508                         }
1509                 }
1510                 sectors -= s;
1511                 sect += s;
1512         }
1513 }
1514
1515 static void raid1d(mddev_t *mddev)
1516 {
1517         r1bio_t *r1_bio;
1518         struct bio *bio;
1519         unsigned long flags;
1520         conf_t *conf = mddev->private;
1521         struct list_head *head = &conf->retry_list;
1522         mdk_rdev_t *rdev;
1523         struct blk_plug plug;
1524
1525         md_check_recovery(mddev);
1526
1527         blk_start_plug(&plug);
1528         for (;;) {
1529                 char b[BDEVNAME_SIZE];
1530
1531                 if (atomic_read(&mddev->plug_cnt) == 0)
1532                         flush_pending_writes(conf);
1533
1534                 spin_lock_irqsave(&conf->device_lock, flags);
1535                 if (list_empty(head)) {
1536                         spin_unlock_irqrestore(&conf->device_lock, flags);
1537                         break;
1538                 }
1539                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1540                 list_del(head->prev);
1541                 conf->nr_queued--;
1542                 spin_unlock_irqrestore(&conf->device_lock, flags);
1543
1544                 mddev = r1_bio->mddev;
1545                 conf = mddev->private;
1546                 if (test_bit(R1BIO_IsSync, &r1_bio->state))
1547                         sync_request_write(mddev, r1_bio);
1548                 else {
1549                         int disk;
1550
1551                         /* we got a read error. Maybe the drive is bad.  Maybe just
1552                          * the block and we can fix it.
1553                          * We freeze all other IO, and try reading the block from
1554                          * other devices.  When we find one, we re-write
1555                          * and check it that fixes the read error.
1556                          * This is all done synchronously while the array is
1557                          * frozen
1558                          */
1559                         if (mddev->ro == 0) {
1560                                 freeze_array(conf);
1561                                 fix_read_error(conf, r1_bio->read_disk,
1562                                                r1_bio->sector,
1563                                                r1_bio->sectors);
1564                                 unfreeze_array(conf);
1565                         } else
1566                                 md_error(mddev,
1567                                          conf->mirrors[r1_bio->read_disk].rdev);
1568
1569                         bio = r1_bio->bios[r1_bio->read_disk];
1570                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1571                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1572                                        " read error for block %llu\n",
1573                                        mdname(mddev),
1574                                        bdevname(bio->bi_bdev,b),
1575                                        (unsigned long long)r1_bio->sector);
1576                                 raid_end_bio_io(r1_bio);
1577                         } else {
1578                                 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1579                                 r1_bio->bios[r1_bio->read_disk] =
1580                                         mddev->ro ? IO_BLOCKED : NULL;
1581                                 r1_bio->read_disk = disk;
1582                                 bio_put(bio);
1583                                 bio = bio_clone_mddev(r1_bio->master_bio,
1584                                                       GFP_NOIO, mddev);
1585                                 r1_bio->bios[r1_bio->read_disk] = bio;
1586                                 rdev = conf->mirrors[disk].rdev;
1587                                 printk_ratelimited(
1588                                         KERN_ERR
1589                                         "md/raid1:%s: redirecting sector %llu"
1590                                         " to other mirror: %s\n",
1591                                         mdname(mddev),
1592                                         (unsigned long long)r1_bio->sector,
1593                                         bdevname(rdev->bdev, b));
1594                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1595                                 bio->bi_bdev = rdev->bdev;
1596                                 bio->bi_end_io = raid1_end_read_request;
1597                                 bio->bi_rw = READ | do_sync;
1598                                 bio->bi_private = r1_bio;
1599                                 generic_make_request(bio);
1600                         }
1601                 }
1602                 cond_resched();
1603         }
1604         blk_finish_plug(&plug);
1605 }
1606
1607
1608 static int init_resync(conf_t *conf)
1609 {
1610         int buffs;
1611
1612         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1613         BUG_ON(conf->r1buf_pool);
1614         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1615                                           conf->poolinfo);
1616         if (!conf->r1buf_pool)
1617                 return -ENOMEM;
1618         conf->next_resync = 0;
1619         return 0;
1620 }
1621
1622 /*
1623  * perform a "sync" on one "block"
1624  *
1625  * We need to make sure that no normal I/O request - particularly write
1626  * requests - conflict with active sync requests.
1627  *
1628  * This is achieved by tracking pending requests and a 'barrier' concept
1629  * that can be installed to exclude normal IO requests.
1630  */
1631
1632 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1633 {
1634         conf_t *conf = mddev->private;
1635         r1bio_t *r1_bio;
1636         struct bio *bio;
1637         sector_t max_sector, nr_sectors;
1638         int disk = -1;
1639         int i;
1640         int wonly = -1;
1641         int write_targets = 0, read_targets = 0;
1642         sector_t sync_blocks;
1643         int still_degraded = 0;
1644
1645         if (!conf->r1buf_pool)
1646                 if (init_resync(conf))
1647                         return 0;
1648
1649         max_sector = mddev->dev_sectors;
1650         if (sector_nr >= max_sector) {
1651                 /* If we aborted, we need to abort the
1652                  * sync on the 'current' bitmap chunk (there will
1653                  * only be one in raid1 resync.
1654                  * We can find the current addess in mddev->curr_resync
1655                  */
1656                 if (mddev->curr_resync < max_sector) /* aborted */
1657                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1658                                                 &sync_blocks, 1);
1659                 else /* completed sync */
1660                         conf->fullsync = 0;
1661
1662                 bitmap_close_sync(mddev->bitmap);
1663                 close_sync(conf);
1664                 return 0;
1665         }
1666
1667         if (mddev->bitmap == NULL &&
1668             mddev->recovery_cp == MaxSector &&
1669             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1670             conf->fullsync == 0) {
1671                 *skipped = 1;
1672                 return max_sector - sector_nr;
1673         }
1674         /* before building a request, check if we can skip these blocks..
1675          * This call the bitmap_start_sync doesn't actually record anything
1676          */
1677         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1678             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1679                 /* We can skip this block, and probably several more */
1680                 *skipped = 1;
1681                 return sync_blocks;
1682         }
1683         /*
1684          * If there is non-resync activity waiting for a turn,
1685          * and resync is going fast enough,
1686          * then let it though before starting on this new sync request.
1687          */
1688         if (!go_faster && conf->nr_waiting)
1689                 msleep_interruptible(1000);
1690
1691         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1692         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1693         raise_barrier(conf);
1694
1695         conf->next_resync = sector_nr;
1696
1697         rcu_read_lock();
1698         /*
1699          * If we get a correctably read error during resync or recovery,
1700          * we might want to read from a different device.  So we
1701          * flag all drives that could conceivably be read from for READ,
1702          * and any others (which will be non-In_sync devices) for WRITE.
1703          * If a read fails, we try reading from something else for which READ
1704          * is OK.
1705          */
1706
1707         r1_bio->mddev = mddev;
1708         r1_bio->sector = sector_nr;
1709         r1_bio->state = 0;
1710         set_bit(R1BIO_IsSync, &r1_bio->state);
1711
1712         for (i=0; i < conf->raid_disks; i++) {
1713                 mdk_rdev_t *rdev;
1714                 bio = r1_bio->bios[i];
1715
1716                 /* take from bio_init */
1717                 bio->bi_next = NULL;
1718                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1719                 bio->bi_flags |= 1 << BIO_UPTODATE;
1720                 bio->bi_comp_cpu = -1;
1721                 bio->bi_rw = READ;
1722                 bio->bi_vcnt = 0;
1723                 bio->bi_idx = 0;
1724                 bio->bi_phys_segments = 0;
1725                 bio->bi_size = 0;
1726                 bio->bi_end_io = NULL;
1727                 bio->bi_private = NULL;
1728
1729                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1730                 if (rdev == NULL ||
1731                            test_bit(Faulty, &rdev->flags)) {
1732                         still_degraded = 1;
1733                         continue;
1734                 } else if (!test_bit(In_sync, &rdev->flags)) {
1735                         bio->bi_rw = WRITE;
1736                         bio->bi_end_io = end_sync_write;
1737                         write_targets ++;
1738                 } else {
1739                         /* may need to read from here */
1740                         bio->bi_rw = READ;
1741                         bio->bi_end_io = end_sync_read;
1742                         if (test_bit(WriteMostly, &rdev->flags)) {
1743                                 if (wonly < 0)
1744                                         wonly = i;
1745                         } else {
1746                                 if (disk < 0)
1747                                         disk = i;
1748                         }
1749                         read_targets++;
1750                 }
1751                 atomic_inc(&rdev->nr_pending);
1752                 bio->bi_sector = sector_nr + rdev->data_offset;
1753                 bio->bi_bdev = rdev->bdev;
1754                 bio->bi_private = r1_bio;
1755         }
1756         rcu_read_unlock();
1757         if (disk < 0)
1758                 disk = wonly;
1759         r1_bio->read_disk = disk;
1760
1761         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1762                 /* extra read targets are also write targets */
1763                 write_targets += read_targets-1;
1764
1765         if (write_targets == 0 || read_targets == 0) {
1766                 /* There is nowhere to write, so all non-sync
1767                  * drives must be failed - so we are finished
1768                  */
1769                 sector_t rv = max_sector - sector_nr;
1770                 *skipped = 1;
1771                 put_buf(r1_bio);
1772                 return rv;
1773         }
1774
1775         if (max_sector > mddev->resync_max)
1776                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1777         nr_sectors = 0;
1778         sync_blocks = 0;
1779         do {
1780                 struct page *page;
1781                 int len = PAGE_SIZE;
1782                 if (sector_nr + (len>>9) > max_sector)
1783                         len = (max_sector - sector_nr) << 9;
1784                 if (len == 0)
1785                         break;
1786                 if (sync_blocks == 0) {
1787                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1788                                                &sync_blocks, still_degraded) &&
1789                             !conf->fullsync &&
1790                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1791                                 break;
1792                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1793                         if ((len >> 9) > sync_blocks)
1794                                 len = sync_blocks<<9;
1795                 }
1796
1797                 for (i=0 ; i < conf->raid_disks; i++) {
1798                         bio = r1_bio->bios[i];
1799                         if (bio->bi_end_io) {
1800                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1801                                 if (bio_add_page(bio, page, len, 0) == 0) {
1802                                         /* stop here */
1803                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1804                                         while (i > 0) {
1805                                                 i--;
1806                                                 bio = r1_bio->bios[i];
1807                                                 if (bio->bi_end_io==NULL)
1808                                                         continue;
1809                                                 /* remove last page from this bio */
1810                                                 bio->bi_vcnt--;
1811                                                 bio->bi_size -= len;
1812                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1813                                         }
1814                                         goto bio_full;
1815                                 }
1816                         }
1817                 }
1818                 nr_sectors += len>>9;
1819                 sector_nr += len>>9;
1820                 sync_blocks -= (len>>9);
1821         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1822  bio_full:
1823         r1_bio->sectors = nr_sectors;
1824
1825         /* For a user-requested sync, we read all readable devices and do a
1826          * compare
1827          */
1828         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1829                 atomic_set(&r1_bio->remaining, read_targets);
1830                 for (i=0; i<conf->raid_disks; i++) {
1831                         bio = r1_bio->bios[i];
1832                         if (bio->bi_end_io == end_sync_read) {
1833                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1834                                 generic_make_request(bio);
1835                         }
1836                 }
1837         } else {
1838                 atomic_set(&r1_bio->remaining, 1);
1839                 bio = r1_bio->bios[r1_bio->read_disk];
1840                 md_sync_acct(bio->bi_bdev, nr_sectors);
1841                 generic_make_request(bio);
1842
1843         }
1844         return nr_sectors;
1845 }
1846
1847 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1848 {
1849         if (sectors)
1850                 return sectors;
1851
1852         return mddev->dev_sectors;
1853 }
1854
1855 static conf_t *setup_conf(mddev_t *mddev)
1856 {
1857         conf_t *conf;
1858         int i;
1859         mirror_info_t *disk;
1860         mdk_rdev_t *rdev;
1861         int err = -ENOMEM;
1862
1863         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1864         if (!conf)
1865                 goto abort;
1866
1867         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1868                                  GFP_KERNEL);
1869         if (!conf->mirrors)
1870                 goto abort;
1871
1872         conf->tmppage = alloc_page(GFP_KERNEL);
1873         if (!conf->tmppage)
1874                 goto abort;
1875
1876         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1877         if (!conf->poolinfo)
1878                 goto abort;
1879         conf->poolinfo->raid_disks = mddev->raid_disks;
1880         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1881                                           r1bio_pool_free,
1882                                           conf->poolinfo);
1883         if (!conf->r1bio_pool)
1884                 goto abort;
1885
1886         conf->poolinfo->mddev = mddev;
1887
1888         spin_lock_init(&conf->device_lock);
1889         list_for_each_entry(rdev, &mddev->disks, same_set) {
1890                 int disk_idx = rdev->raid_disk;
1891                 if (disk_idx >= mddev->raid_disks
1892                     || disk_idx < 0)
1893                         continue;
1894                 disk = conf->mirrors + disk_idx;
1895
1896                 disk->rdev = rdev;
1897
1898                 disk->head_position = 0;
1899         }
1900         conf->raid_disks = mddev->raid_disks;
1901         conf->mddev = mddev;
1902         INIT_LIST_HEAD(&conf->retry_list);
1903
1904         spin_lock_init(&conf->resync_lock);
1905         init_waitqueue_head(&conf->wait_barrier);
1906
1907         bio_list_init(&conf->pending_bio_list);
1908
1909         conf->last_used = -1;
1910         for (i = 0; i < conf->raid_disks; i++) {
1911
1912                 disk = conf->mirrors + i;
1913
1914                 if (!disk->rdev ||
1915                     !test_bit(In_sync, &disk->rdev->flags)) {
1916                         disk->head_position = 0;
1917                         if (disk->rdev)
1918                                 conf->fullsync = 1;
1919                 } else if (conf->last_used < 0)
1920                         /*
1921                          * The first working device is used as a
1922                          * starting point to read balancing.
1923                          */
1924                         conf->last_used = i;
1925         }
1926
1927         err = -EIO;
1928         if (conf->last_used < 0) {
1929                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
1930                        mdname(mddev));
1931                 goto abort;
1932         }
1933         err = -ENOMEM;
1934         conf->thread = md_register_thread(raid1d, mddev, NULL);
1935         if (!conf->thread) {
1936                 printk(KERN_ERR
1937                        "md/raid1:%s: couldn't allocate thread\n",
1938                        mdname(mddev));
1939                 goto abort;
1940         }
1941
1942         return conf;
1943
1944  abort:
1945         if (conf) {
1946                 if (conf->r1bio_pool)
1947                         mempool_destroy(conf->r1bio_pool);
1948                 kfree(conf->mirrors);
1949                 safe_put_page(conf->tmppage);
1950                 kfree(conf->poolinfo);
1951                 kfree(conf);
1952         }
1953         return ERR_PTR(err);
1954 }
1955
1956 static int run(mddev_t *mddev)
1957 {
1958         conf_t *conf;
1959         int i;
1960         mdk_rdev_t *rdev;
1961
1962         if (mddev->level != 1) {
1963                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
1964                        mdname(mddev), mddev->level);
1965                 return -EIO;
1966         }
1967         if (mddev->reshape_position != MaxSector) {
1968                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
1969                        mdname(mddev));
1970                 return -EIO;
1971         }
1972         /*
1973          * copy the already verified devices into our private RAID1
1974          * bookkeeping area. [whatever we allocate in run(),
1975          * should be freed in stop()]
1976          */
1977         if (mddev->private == NULL)
1978                 conf = setup_conf(mddev);
1979         else
1980                 conf = mddev->private;
1981
1982         if (IS_ERR(conf))
1983                 return PTR_ERR(conf);
1984
1985         list_for_each_entry(rdev, &mddev->disks, same_set) {
1986                 if (!mddev->gendisk)
1987                         continue;
1988                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1989                                   rdev->data_offset << 9);
1990                 /* as we don't honour merge_bvec_fn, we must never risk
1991                  * violating it, so limit ->max_segments to 1 lying within
1992                  * a single page, as a one page request is never in violation.
1993                  */
1994                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1995                         blk_queue_max_segments(mddev->queue, 1);
1996                         blk_queue_segment_boundary(mddev->queue,
1997                                                    PAGE_CACHE_SIZE - 1);
1998                 }
1999         }
2000
2001         mddev->degraded = 0;
2002         for (i=0; i < conf->raid_disks; i++)
2003                 if (conf->mirrors[i].rdev == NULL ||
2004                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2005                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2006                         mddev->degraded++;
2007
2008         if (conf->raid_disks - mddev->degraded == 1)
2009                 mddev->recovery_cp = MaxSector;
2010
2011         if (mddev->recovery_cp != MaxSector)
2012                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2013                        " -- starting background reconstruction\n",
2014                        mdname(mddev));
2015         printk(KERN_INFO
2016                 "md/raid1:%s: active with %d out of %d mirrors\n",
2017                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2018                 mddev->raid_disks);
2019
2020         /*
2021          * Ok, everything is just fine now
2022          */
2023         mddev->thread = conf->thread;
2024         conf->thread = NULL;
2025         mddev->private = conf;
2026
2027         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2028
2029         if (mddev->queue) {
2030                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2031                 mddev->queue->backing_dev_info.congested_data = mddev;
2032         }
2033         return md_integrity_register(mddev);
2034 }
2035
2036 static int stop(mddev_t *mddev)
2037 {
2038         conf_t *conf = mddev->private;
2039         struct bitmap *bitmap = mddev->bitmap;
2040
2041         /* wait for behind writes to complete */
2042         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2043                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2044                        mdname(mddev));
2045                 /* need to kick something here to make sure I/O goes? */
2046                 wait_event(bitmap->behind_wait,
2047                            atomic_read(&bitmap->behind_writes) == 0);
2048         }
2049
2050         raise_barrier(conf);
2051         lower_barrier(conf);
2052
2053         md_unregister_thread(mddev->thread);
2054         mddev->thread = NULL;
2055         if (conf->r1bio_pool)
2056                 mempool_destroy(conf->r1bio_pool);
2057         kfree(conf->mirrors);
2058         kfree(conf->poolinfo);
2059         kfree(conf);
2060         mddev->private = NULL;
2061         return 0;
2062 }
2063
2064 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2065 {
2066         /* no resync is happening, and there is enough space
2067          * on all devices, so we can resize.
2068          * We need to make sure resync covers any new space.
2069          * If the array is shrinking we should possibly wait until
2070          * any io in the removed space completes, but it hardly seems
2071          * worth it.
2072          */
2073         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2074         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2075                 return -EINVAL;
2076         set_capacity(mddev->gendisk, mddev->array_sectors);
2077         revalidate_disk(mddev->gendisk);
2078         if (sectors > mddev->dev_sectors &&
2079             mddev->recovery_cp > mddev->dev_sectors) {
2080                 mddev->recovery_cp = mddev->dev_sectors;
2081                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2082         }
2083         mddev->dev_sectors = sectors;
2084         mddev->resync_max_sectors = sectors;
2085         return 0;
2086 }
2087
2088 static int raid1_reshape(mddev_t *mddev)
2089 {
2090         /* We need to:
2091          * 1/ resize the r1bio_pool
2092          * 2/ resize conf->mirrors
2093          *
2094          * We allocate a new r1bio_pool if we can.
2095          * Then raise a device barrier and wait until all IO stops.
2096          * Then resize conf->mirrors and swap in the new r1bio pool.
2097          *
2098          * At the same time, we "pack" the devices so that all the missing
2099          * devices have the higher raid_disk numbers.
2100          */
2101         mempool_t *newpool, *oldpool;
2102         struct pool_info *newpoolinfo;
2103         mirror_info_t *newmirrors;
2104         conf_t *conf = mddev->private;
2105         int cnt, raid_disks;
2106         unsigned long flags;
2107         int d, d2, err;
2108
2109         /* Cannot change chunk_size, layout, or level */
2110         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2111             mddev->layout != mddev->new_layout ||
2112             mddev->level != mddev->new_level) {
2113                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2114                 mddev->new_layout = mddev->layout;
2115                 mddev->new_level = mddev->level;
2116                 return -EINVAL;
2117         }
2118
2119         err = md_allow_write(mddev);
2120         if (err)
2121                 return err;
2122
2123         raid_disks = mddev->raid_disks + mddev->delta_disks;
2124
2125         if (raid_disks < conf->raid_disks) {
2126                 cnt=0;
2127                 for (d= 0; d < conf->raid_disks; d++)
2128                         if (conf->mirrors[d].rdev)
2129                                 cnt++;
2130                 if (cnt > raid_disks)
2131                         return -EBUSY;
2132         }
2133
2134         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2135         if (!newpoolinfo)
2136                 return -ENOMEM;
2137         newpoolinfo->mddev = mddev;
2138         newpoolinfo->raid_disks = raid_disks;
2139
2140         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2141                                  r1bio_pool_free, newpoolinfo);
2142         if (!newpool) {
2143                 kfree(newpoolinfo);
2144                 return -ENOMEM;
2145         }
2146         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2147         if (!newmirrors) {
2148                 kfree(newpoolinfo);
2149                 mempool_destroy(newpool);
2150                 return -ENOMEM;
2151         }
2152
2153         raise_barrier(conf);
2154
2155         /* ok, everything is stopped */
2156         oldpool = conf->r1bio_pool;
2157         conf->r1bio_pool = newpool;
2158
2159         for (d = d2 = 0; d < conf->raid_disks; d++) {
2160                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2161                 if (rdev && rdev->raid_disk != d2) {
2162                         char nm[20];
2163                         sprintf(nm, "rd%d", rdev->raid_disk);
2164                         sysfs_remove_link(&mddev->kobj, nm);
2165                         rdev->raid_disk = d2;
2166                         sprintf(nm, "rd%d", rdev->raid_disk);
2167                         sysfs_remove_link(&mddev->kobj, nm);
2168                         if (sysfs_create_link(&mddev->kobj,
2169                                               &rdev->kobj, nm))
2170                                 printk(KERN_WARNING
2171                                        "md/raid1:%s: cannot register "
2172                                        "%s\n",
2173                                        mdname(mddev), nm);
2174                 }
2175                 if (rdev)
2176                         newmirrors[d2++].rdev = rdev;
2177         }
2178         kfree(conf->mirrors);
2179         conf->mirrors = newmirrors;
2180         kfree(conf->poolinfo);
2181         conf->poolinfo = newpoolinfo;
2182
2183         spin_lock_irqsave(&conf->device_lock, flags);
2184         mddev->degraded += (raid_disks - conf->raid_disks);
2185         spin_unlock_irqrestore(&conf->device_lock, flags);
2186         conf->raid_disks = mddev->raid_disks = raid_disks;
2187         mddev->delta_disks = 0;
2188
2189         conf->last_used = 0; /* just make sure it is in-range */
2190         lower_barrier(conf);
2191
2192         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2193         md_wakeup_thread(mddev->thread);
2194
2195         mempool_destroy(oldpool);
2196         return 0;
2197 }
2198
2199 static void raid1_quiesce(mddev_t *mddev, int state)
2200 {
2201         conf_t *conf = mddev->private;
2202
2203         switch(state) {
2204         case 2: /* wake for suspend */
2205                 wake_up(&conf->wait_barrier);
2206                 break;
2207         case 1:
2208                 raise_barrier(conf);
2209                 break;
2210         case 0:
2211                 lower_barrier(conf);
2212                 break;
2213         }
2214 }
2215
2216 static void *raid1_takeover(mddev_t *mddev)
2217 {
2218         /* raid1 can take over:
2219          *  raid5 with 2 devices, any layout or chunk size
2220          */
2221         if (mddev->level == 5 && mddev->raid_disks == 2) {
2222                 conf_t *conf;
2223                 mddev->new_level = 1;
2224                 mddev->new_layout = 0;
2225                 mddev->new_chunk_sectors = 0;
2226                 conf = setup_conf(mddev);
2227                 if (!IS_ERR(conf))
2228                         conf->barrier = 1;
2229                 return conf;
2230         }
2231         return ERR_PTR(-EINVAL);
2232 }
2233
2234 static struct mdk_personality raid1_personality =
2235 {
2236         .name           = "raid1",
2237         .level          = 1,
2238         .owner          = THIS_MODULE,
2239         .make_request   = make_request,
2240         .run            = run,
2241         .stop           = stop,
2242         .status         = status,
2243         .error_handler  = error,
2244         .hot_add_disk   = raid1_add_disk,
2245         .hot_remove_disk= raid1_remove_disk,
2246         .spare_active   = raid1_spare_active,
2247         .sync_request   = sync_request,
2248         .resize         = raid1_resize,
2249         .size           = raid1_size,
2250         .check_reshape  = raid1_reshape,
2251         .quiesce        = raid1_quiesce,
2252         .takeover       = raid1_takeover,
2253 };
2254
2255 static int __init raid_init(void)
2256 {
2257         return register_md_personality(&raid1_personality);
2258 }
2259
2260 static void raid_exit(void)
2261 {
2262         unregister_md_personality(&raid1_personality);
2263 }
2264
2265 module_init(raid_init);
2266 module_exit(raid_exit);
2267 MODULE_LICENSE("GPL");
2268 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2269 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2270 MODULE_ALIAS("md-raid1");
2271 MODULE_ALIAS("md-level-1");