drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/slab.h>
  35 #include <linux/delay.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/seq_file.h>
  38 #include "md.h"
  39 #include "raid1.h"
  40 #include "bitmap.h"
  41
  42 #define DEBUG 0
  43 #if DEBUG
  44 #define PRINTK(x...) printk(x)
  45 #else
  46 #define PRINTK(x...)
  47 #endif
  48
  49 /*
  50  * Number of guaranteed r1bios in case of extreme VM load:
  51  */
  52 #define NR_RAID1_BIOS 256
  53
  54
  55 static void allow_barrier(conf_t *conf);
  56 static void lower_barrier(conf_t *conf);
  57
  58 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  59 {
  60         struct pool_info *pi = data;
  61         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  62
  63         /* allocate a r1bio with room for raid_disks entries in the bios array */
  64         return kzalloc(size, gfp_flags);
  65 }
  66
  67 static void r1bio_pool_free(void *r1_bio, void *data)
  68 {
  69         kfree(r1_bio);
  70 }
  71
  72 #define RESYNC_BLOCK_SIZE (64*1024)
  73 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  74 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  75 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  76 #define RESYNC_WINDOW (2048*1024)
  77
  78 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  79 {
  80         struct pool_info *pi = data;
  81         struct page *page;
  82         r1bio_t *r1_bio;
  83         struct bio *bio;
  84         int i, j;
  85
  86         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  87         if (!r1_bio)
  88                 return NULL;
  89
  90         /*
  91          * Allocate bios : 1 for reading, n-1 for writing
  92          */
  93         for (j = pi->raid_disks ; j-- ; ) {
  94                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
  95                 if (!bio)
  96                         goto out_free_bio;
  97                 r1_bio->bios[j] = bio;
  98         }
  99         /*
 100          * Allocate RESYNC_PAGES data pages and attach them to
 101          * the first bio.
 102          * If this is a user-requested check/repair, allocate
 103          * RESYNC_PAGES for each bio.
 104          */
 105         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 106                 j = pi->raid_disks;
 107         else
 108                 j = 1;
 109         while(j--) {
 110                 bio = r1_bio->bios[j];
 111                 for (i = 0; i < RESYNC_PAGES; i++) {
 112                         page = alloc_page(gfp_flags);
 113                         if (unlikely(!page))
 114                                 goto out_free_pages;
 115
 116                         bio->bi_io_vec[i].bv_page = page;
 117                         bio->bi_vcnt = i+1;
 118                 }
 119         }
 120         /* If not user-requests, copy the page pointers to all bios */
 121         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 122                 for (i=0; i<RESYNC_PAGES ; i++)
 123                         for (j=1; j<pi->raid_disks; j++)
 124                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 125                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 126         }
 127
 128         r1_bio->master_bio = NULL;
 129
 130         return r1_bio;
 131
 132 out_free_pages:
 133         for (j=0 ; j < pi->raid_disks; j++)
 134                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 135                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 136         j = -1;
 137 out_free_bio:
 138         while ( ++j < pi->raid_disks )
 139                 bio_put(r1_bio->bios[j]);
 140         r1bio_pool_free(r1_bio, data);
 141         return NULL;
 142 }
 143
 144 static void r1buf_pool_free(void *__r1_bio, void *data)
 145 {
 146         struct pool_info *pi = data;
 147         int i,j;
 148         r1bio_t *r1bio = __r1_bio;
 149
 150         for (i = 0; i < RESYNC_PAGES; i++)
 151                 for (j = pi->raid_disks; j-- ;) {
 152                         if (j == 0 ||
 153                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 154                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 155                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 156                 }
 157         for (i=0 ; i < pi->raid_disks; i++)
 158                 bio_put(r1bio->bios[i]);
 159
 160         r1bio_pool_free(r1bio, data);
 161 }
 162
 163 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 164 {
 165         int i;
 166
 167         for (i = 0; i < conf->raid_disks; i++) {
 168                 struct bio **bio = r1_bio->bios + i;
 169                 if (*bio && *bio != IO_BLOCKED)
 170                         bio_put(*bio);
 171                 *bio = NULL;
 172         }
 173 }
 174
 175 static void free_r1bio(r1bio_t *r1_bio)
 176 {
 177         conf_t *conf = r1_bio->mddev->private;
 178
 179         /*
 180          * Wake up any possible resync thread that waits for the device
 181          * to go idle.
 182          */
 183         allow_barrier(conf);
 184
 185         put_all_bios(conf, r1_bio);
 186         mempool_free(r1_bio, conf->r1bio_pool);
 187 }
 188
 189 static void put_buf(r1bio_t *r1_bio)
 190 {
 191         conf_t *conf = r1_bio->mddev->private;
 192         int i;
 193
 194         for (i=0; i<conf->raid_disks; i++) {
 195                 struct bio *bio = r1_bio->bios[i];
 196                 if (bio->bi_end_io)
 197                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 198         }
 199
 200         mempool_free(r1_bio, conf->r1buf_pool);
 201
 202         lower_barrier(conf);
 203 }
 204
 205 static void reschedule_retry(r1bio_t *r1_bio)
 206 {
 207         unsigned long flags;
 208         mddev_t *mddev = r1_bio->mddev;
 209         conf_t *conf = mddev->private;
 210
 211         spin_lock_irqsave(&conf->device_lock, flags);
 212         list_add(&r1_bio->retry_list, &conf->retry_list);
 213         conf->nr_queued ++;
 214         spin_unlock_irqrestore(&conf->device_lock, flags);
 215
 216         wake_up(&conf->wait_barrier);
 217         md_wakeup_thread(mddev->thread);
 218 }
 219
 220 /*
 221  * raid_end_bio_io() is called when we have finished servicing a mirrored
 222  * operation and are ready to return a success/failure code to the buffer
 223  * cache layer.
 224  */
 225 static void raid_end_bio_io(r1bio_t *r1_bio)
 226 {
 227         struct bio *bio = r1_bio->master_bio;
 228
 229         /* if nobody has done the final endio yet, do it now */
 230         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 231                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 232                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 233                         (unsigned long long) bio->bi_sector,
 234                         (unsigned long long) bio->bi_sector +
 235                                 (bio->bi_size >> 9) - 1);
 236
 237                 bio_endio(bio,
 238                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 239         }
 240         free_r1bio(r1_bio);
 241 }
 242
 243 /*
 244  * Update disk head position estimator based on IRQ completion info.
 245  */
 246 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 247 {
 248         conf_t *conf = r1_bio->mddev->private;
 249
 250         conf->mirrors[disk].head_position =
 251                 r1_bio->sector + (r1_bio->sectors);
 252 }
 253
 254 static void raid1_end_read_request(struct bio *bio, int error)
 255 {
 256         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 257         r1bio_t *r1_bio = bio->bi_private;
 258         int mirror;
 259         conf_t *conf = r1_bio->mddev->private;
 260
 261         mirror = r1_bio->read_disk;
 262         /*
 263          * this branch is our 'one mirror IO has finished' event handler:
 264          */
 265         update_head_pos(mirror, r1_bio);
 266
 267         if (uptodate)
 268                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 269         else {
 270                 /* If all other devices have failed, we want to return
 271                  * the error upwards rather than fail the last device.
 272                  * Here we redefine "uptodate" to mean "Don't want to retry"
 273                  */
 274                 unsigned long flags;
 275                 spin_lock_irqsave(&conf->device_lock, flags);
 276                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 277                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 278                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 279                         uptodate = 1;
 280                 spin_unlock_irqrestore(&conf->device_lock, flags);
 281         }
 282
 283         if (uptodate)
 284                 raid_end_bio_io(r1_bio);
 285         else {
 286                 /*
 287                  * oops, read error:
 288                  */
 289                 char b[BDEVNAME_SIZE];
 290                 if (printk_ratelimit())
 291                         printk(KERN_ERR "md/raid1:%s: %s: rescheduling sector %llu\n",
 292                                mdname(conf->mddev),
 293                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 294                 reschedule_retry(r1_bio);
 295         }
 296
 297         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 298 }
 299
 300 static void r1_bio_write_done(r1bio_t *r1_bio, int vcnt, struct bio_vec *bv,
 301                               int behind)
 302 {
 303         if (atomic_dec_and_test(&r1_bio->remaining))
 304         {
 305                 /* it really is the end of this request */
 306                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 307                         /* free extra copy of the data pages */
 308                         int i = vcnt;
 309                         while (i--)
 310                                 safe_put_page(bv[i].bv_page);
 311                 }
 312                 /* clear the bitmap if all writes complete successfully */
 313                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 314                                 r1_bio->sectors,
 315                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
 316                                 behind);
 317                 md_write_end(r1_bio->mddev);
 318                 raid_end_bio_io(r1_bio);
 319         }
 320 }
 321
 322 static void raid1_end_write_request(struct bio *bio, int error)
 323 {
 324         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 325         r1bio_t *r1_bio = bio->bi_private;
 326         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 327         conf_t *conf = r1_bio->mddev->private;
 328         struct bio *to_put = NULL;
 329
 330
 331         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 332                 if (r1_bio->bios[mirror] == bio)
 333                         break;
 334
 335         /*
 336          * 'one mirror IO has finished' event handler:
 337          */
 338         r1_bio->bios[mirror] = NULL;
 339         to_put = bio;
 340         if (!uptodate) {
 341                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 342                 /* an I/O failed, we can't clear the bitmap */
 343                 set_bit(R1BIO_Degraded, &r1_bio->state);
 344         } else
 345                 /*
 346                  * Set R1BIO_Uptodate in our master bio, so that we
 347                  * will return a good error code for to the higher
 348                  * levels even if IO on some other mirrored buffer
 349                  * fails.
 350                  *
 351                  * The 'master' represents the composite IO operation
 352                  * to user-side. So if something waits for IO, then it
 353                  * will wait for the 'master' bio.
 354                  */
 355                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 356
 357         update_head_pos(mirror, r1_bio);
 358
 359         if (behind) {
 360                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 361                         atomic_dec(&r1_bio->behind_remaining);
 362
 363                 /*
 364                  * In behind mode, we ACK the master bio once the I/O
 365                  * has safely reached all non-writemostly
 366                  * disks. Setting the Returned bit ensures that this
 367                  * gets done only once -- we don't ever want to return
 368                  * -EIO here, instead we'll wait
 369                  */
 370                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 371                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 372                         /* Maybe we can return now */
 373                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 374                                 struct bio *mbio = r1_bio->master_bio;
 375                                 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 376                                        (unsigned long long) mbio->bi_sector,
 377                                        (unsigned long long) mbio->bi_sector +
 378                                        (mbio->bi_size >> 9) - 1);
 379                                 bio_endio(mbio, 0);
 380                         }
 381                 }
 382         }
 383         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 384
 385         /*
 386          * Let's see if all mirrored write operations have finished
 387          * already.
 388          */
 389         r1_bio_write_done(r1_bio, bio->bi_vcnt, bio->bi_io_vec, behind);
 390
 391         if (to_put)
 392                 bio_put(to_put);
 393 }
 394
 395
 396 /*
 397  * This routine returns the disk from which the requested read should
 398  * be done. There is a per-array 'next expected sequential IO' sector
 399  * number - if this matches on the next IO then we use the last disk.
 400  * There is also a per-disk 'last know head position' sector that is
 401  * maintained from IRQ contexts, both the normal and the resync IO
 402  * completion handlers update this position correctly. If there is no
 403  * perfect sequential match then we pick the disk whose head is closest.
 404  *
 405  * If there are 2 mirrors in the same 2 devices, performance degrades
 406  * because position is mirror, not device based.
 407  *
 408  * The rdev for the device selected will have nr_pending incremented.
 409  */
 410 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 411 {
 412         const sector_t this_sector = r1_bio->sector;
 413         const int sectors = r1_bio->sectors;
 414         int start_disk;
 415         int best_disk;
 416         int i;
 417         sector_t best_dist;
 418         mdk_rdev_t *rdev;
 419         int choose_first;
 420
 421         rcu_read_lock();
 422         /*
 423          * Check if we can balance. We can balance on the whole
 424          * device if no resync is going on, or below the resync window.
 425          * We take the first readable disk when above the resync window.
 426          */
 427  retry:
 428         best_disk = -1;
 429         best_dist = MaxSector;
 430         if (conf->mddev->recovery_cp < MaxSector &&
 431             (this_sector + sectors >= conf->next_resync)) {
 432                 choose_first = 1;
 433                 start_disk = 0;
 434         } else {
 435                 choose_first = 0;
 436                 start_disk = conf->last_used;
 437         }
 438
 439         for (i = 0 ; i < conf->raid_disks ; i++) {
 440                 sector_t dist;
 441                 int disk = start_disk + i;
 442                 if (disk >= conf->raid_disks)
 443                         disk -= conf->raid_disks;
 444
 445                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 446                 if (r1_bio->bios[disk] == IO_BLOCKED
 447                     || rdev == NULL
 448                     || test_bit(Faulty, &rdev->flags))
 449                         continue;
 450                 if (!test_bit(In_sync, &rdev->flags) &&
 451                     rdev->recovery_offset < this_sector + sectors)
 452                         continue;
 453                 if (test_bit(WriteMostly, &rdev->flags)) {
 454                         /* Don't balance among write-mostly, just
 455                          * use the first as a last resort */
 456                         if (best_disk < 0)
 457                                 best_disk = disk;
 458                         continue;
 459                 }
 460                 /* This is a reasonable device to use.  It might
 461                  * even be best.
 462                  */
 463                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 464                 if (choose_first
 465                     /* Don't change to another disk for sequential reads */
 466                     || conf->next_seq_sect == this_sector
 467                     || dist == 0
 468                     /* If device is idle, use it */
 469                     || atomic_read(&rdev->nr_pending) == 0) {
 470                         best_disk = disk;
 471                         break;
 472                 }
 473                 if (dist < best_dist) {
 474                         best_dist = dist;
 475                         best_disk = disk;
 476                 }
 477         }
 478
 479         if (best_disk >= 0) {
 480                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 481                 if (!rdev)
 482                         goto retry;
 483                 atomic_inc(&rdev->nr_pending);
 484                 if (test_bit(Faulty, &rdev->flags)) {
 485                         /* cannot risk returning a device that failed
 486                          * before we inc'ed nr_pending
 487                          */
 488                         rdev_dec_pending(rdev, conf->mddev);
 489                         goto retry;
 490                 }
 491                 conf->next_seq_sect = this_sector + sectors;
 492                 conf->last_used = best_disk;
 493         }
 494         rcu_read_unlock();
 495
 496         return best_disk;
 497 }
 498
 499 static int raid1_congested(void *data, int bits)
 500 {
 501         mddev_t *mddev = data;
 502         conf_t *conf = mddev->private;
 503         int i, ret = 0;
 504
 505         if (mddev_congested(mddev, bits))
 506                 return 1;
 507
 508         rcu_read_lock();
 509         for (i = 0; i < mddev->raid_disks; i++) {
 510                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 511                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 512                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 513
 514                         /* Note the '|| 1' - when read_balance prefers
 515                          * non-congested targets, it can be removed
 516                          */
 517                         if ((bits & (1<<BDI_async_congested)) || 1)
 518                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 519                         else
 520                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 521                 }
 522         }
 523         rcu_read_unlock();
 524         return ret;
 525 }
 526
 527
 528 static void flush_pending_writes(conf_t *conf)
 529 {
 530         /* Any writes that have been queued but are awaiting
 531          * bitmap updates get flushed here.
 532          */
 533         spin_lock_irq(&conf->device_lock);
 534
 535         if (conf->pending_bio_list.head) {
 536                 struct bio *bio;
 537                 bio = bio_list_get(&conf->pending_bio_list);
 538                 spin_unlock_irq(&conf->device_lock);
 539                 /* flush any pending bitmap writes to
 540                  * disk before proceeding w/ I/O */
 541                 bitmap_unplug(conf->mddev->bitmap);
 542
 543                 while (bio) { /* submit pending writes */
 544                         struct bio *next = bio->bi_next;
 545                         bio->bi_next = NULL;
 546                         generic_make_request(bio);
 547                         bio = next;
 548                 }
 549         } else
 550                 spin_unlock_irq(&conf->device_lock);
 551 }
 552
 553 /* Barriers....
 554  * Sometimes we need to suspend IO while we do something else,
 555  * either some resync/recovery, or reconfigure the array.
 556  * To do this we raise a 'barrier'.
 557  * The 'barrier' is a counter that can be raised multiple times
 558  * to count how many activities are happening which preclude
 559  * normal IO.
 560  * We can only raise the barrier if there is no pending IO.
 561  * i.e. if nr_pending == 0.
 562  * We choose only to raise the barrier if no-one is waiting for the
 563  * barrier to go down.  This means that as soon as an IO request
 564  * is ready, no other operations which require a barrier will start
 565  * until the IO request has had a chance.
 566  *
 567  * So: regular IO calls 'wait_barrier'.  When that returns there
 568  *    is no backgroup IO happening,  It must arrange to call
 569  *    allow_barrier when it has finished its IO.
 570  * backgroup IO calls must call raise_barrier.  Once that returns
 571  *    there is no normal IO happeing.  It must arrange to call
 572  *    lower_barrier when the particular background IO completes.
 573  */
 574 #define RESYNC_DEPTH 32
 575
 576 static void raise_barrier(conf_t *conf)
 577 {
 578         spin_lock_irq(&conf->resync_lock);
 579
 580         /* Wait until no block IO is waiting */
 581         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 582                             conf->resync_lock, );
 583
 584         /* block any new IO from starting */
 585         conf->barrier++;
 586
 587         /* Now wait for all pending IO to complete */
 588         wait_event_lock_irq(conf->wait_barrier,
 589                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 590                             conf->resync_lock, );
 591
 592         spin_unlock_irq(&conf->resync_lock);
 593 }
 594
 595 static void lower_barrier(conf_t *conf)
 596 {
 597         unsigned long flags;
 598         BUG_ON(conf->barrier <= 0);
 599         spin_lock_irqsave(&conf->resync_lock, flags);
 600         conf->barrier--;
 601         spin_unlock_irqrestore(&conf->resync_lock, flags);
 602         wake_up(&conf->wait_barrier);
 603 }
 604
 605 static void wait_barrier(conf_t *conf)
 606 {
 607         spin_lock_irq(&conf->resync_lock);
 608         if (conf->barrier) {
 609                 conf->nr_waiting++;
 610                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 611                                     conf->resync_lock,
 612                                     );
 613                 conf->nr_waiting--;
 614         }
 615         conf->nr_pending++;
 616         spin_unlock_irq(&conf->resync_lock);
 617 }
 618
 619 static void allow_barrier(conf_t *conf)
 620 {
 621         unsigned long flags;
 622         spin_lock_irqsave(&conf->resync_lock, flags);
 623         conf->nr_pending--;
 624         spin_unlock_irqrestore(&conf->resync_lock, flags);
 625         wake_up(&conf->wait_barrier);
 626 }
 627
 628 static void freeze_array(conf_t *conf)
 629 {
 630         /* stop syncio and normal IO and wait for everything to
 631          * go quite.
 632          * We increment barrier and nr_waiting, and then
 633          * wait until nr_pending match nr_queued+1
 634          * This is called in the context of one normal IO request
 635          * that has failed. Thus any sync request that might be pending
 636          * will be blocked by nr_pending, and we need to wait for
 637          * pending IO requests to complete or be queued for re-try.
 638          * Thus the number queued (nr_queued) plus this request (1)
 639          * must match the number of pending IOs (nr_pending) before
 640          * we continue.
 641          */
 642         spin_lock_irq(&conf->resync_lock);
 643         conf->barrier++;
 644         conf->nr_waiting++;
 645         wait_event_lock_irq(conf->wait_barrier,
 646                             conf->nr_pending == conf->nr_queued+1,
 647                             conf->resync_lock,
 648                             flush_pending_writes(conf));
 649         spin_unlock_irq(&conf->resync_lock);
 650 }
 651 static void unfreeze_array(conf_t *conf)
 652 {
 653         /* reverse the effect of the freeze */
 654         spin_lock_irq(&conf->resync_lock);
 655         conf->barrier--;
 656         conf->nr_waiting--;
 657         wake_up(&conf->wait_barrier);
 658         spin_unlock_irq(&conf->resync_lock);
 659 }
 660
 661
 662 /* duplicate the data pages for behind I/O
 663  * We return a list of bio_vec rather than just page pointers
 664  * as it makes freeing easier
 665  */
 666 static struct bio_vec *alloc_behind_pages(struct bio *bio)
 667 {
 668         int i;
 669         struct bio_vec *bvec;
 670         struct bio_vec *pages = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
 671                                         GFP_NOIO);
 672         if (unlikely(!pages))
 673                 goto do_sync_io;
 674
 675         bio_for_each_segment(bvec, bio, i) {
 676                 pages[i].bv_page = alloc_page(GFP_NOIO);
 677                 if (unlikely(!pages[i].bv_page))
 678                         goto do_sync_io;
 679                 memcpy(kmap(pages[i].bv_page) + bvec->bv_offset,
 680                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 681                 kunmap(pages[i].bv_page);
 682                 kunmap(bvec->bv_page);
 683         }
 684
 685         return pages;
 686
 687 do_sync_io:
 688         if (pages)
 689                 for (i = 0; i < bio->bi_vcnt && pages[i].bv_page; i++)
 690                         put_page(pages[i].bv_page);
 691         kfree(pages);
 692         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 693         return NULL;
 694 }
 695
 696 static int make_request(mddev_t *mddev, struct bio * bio)
 697 {
 698         conf_t *conf = mddev->private;
 699         mirror_info_t *mirror;
 700         r1bio_t *r1_bio;
 701         struct bio *read_bio;
 702         int i, targets = 0, disks;
 703         struct bitmap *bitmap;
 704         unsigned long flags;
 705         struct bio_vec *behind_pages = NULL;
 706         const int rw = bio_data_dir(bio);
 707         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 708         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 709         mdk_rdev_t *blocked_rdev;
 710         int plugged;
 711
 712         /*
 713          * Register the new request and wait if the reconstruction
 714          * thread has put up a bar for new requests.
 715          * Continue immediately if no resync is active currently.
 716          */
 717
 718         md_write_start(mddev, bio); /* wait on superblock update early */
 719
 720         if (bio_data_dir(bio) == WRITE &&
 721             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 722             bio->bi_sector < mddev->suspend_hi) {
 723                 /* As the suspend_* range is controlled by
 724                  * userspace, we want an interruptible
 725                  * wait.
 726                  */
 727                 DEFINE_WAIT(w);
 728                 for (;;) {
 729                         flush_signals(current);
 730                         prepare_to_wait(&conf->wait_barrier,
 731                                         &w, TASK_INTERRUPTIBLE);
 732                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 733                             bio->bi_sector >= mddev->suspend_hi)
 734                                 break;
 735                         schedule();
 736                 }
 737                 finish_wait(&conf->wait_barrier, &w);
 738         }
 739
 740         wait_barrier(conf);
 741
 742         bitmap = mddev->bitmap;
 743
 744         /*
 745          * make_request() can abort the operation when READA is being
 746          * used and no empty request is available.
 747          *
 748          */
 749         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 750
 751         r1_bio->master_bio = bio;
 752         r1_bio->sectors = bio->bi_size >> 9;
 753         r1_bio->state = 0;
 754         r1_bio->mddev = mddev;
 755         r1_bio->sector = bio->bi_sector;
 756
 757         if (rw == READ) {
 758                 /*
 759                  * read balancing logic:
 760                  */
 761                 int rdisk = read_balance(conf, r1_bio);
 762
 763                 if (rdisk < 0) {
 764                         /* couldn't find anywhere to read from */
 765                         raid_end_bio_io(r1_bio);
 766                         return 0;
 767                 }
 768                 mirror = conf->mirrors + rdisk;
 769
 770                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 771                     bitmap) {
 772                         /* Reading from a write-mostly device must
 773                          * take care not to over-take any writes
 774                          * that are 'behind'
 775                          */
 776                         wait_event(bitmap->behind_wait,
 777                                    atomic_read(&bitmap->behind_writes) == 0);
 778                 }
 779                 r1_bio->read_disk = rdisk;
 780
 781                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 782
 783                 r1_bio->bios[rdisk] = read_bio;
 784
 785                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 786                 read_bio->bi_bdev = mirror->rdev->bdev;
 787                 read_bio->bi_end_io = raid1_end_read_request;
 788                 read_bio->bi_rw = READ | do_sync;
 789                 read_bio->bi_private = r1_bio;
 790
 791                 generic_make_request(read_bio);
 792                 return 0;
 793         }
 794
 795         /*
 796          * WRITE:
 797          */
 798         /* first select target devices under spinlock and
 799          * inc refcount on their rdev.  Record them by setting
 800          * bios[x] to bio
 801          */
 802         plugged = mddev_check_plugged(mddev);
 803
 804         disks = conf->raid_disks;
 805  retry_write:
 806         blocked_rdev = NULL;
 807         rcu_read_lock();
 808         for (i = 0;  i < disks; i++) {
 809                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 810                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 811                         atomic_inc(&rdev->nr_pending);
 812                         blocked_rdev = rdev;
 813                         break;
 814                 }
 815                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 816                         atomic_inc(&rdev->nr_pending);
 817                         if (test_bit(Faulty, &rdev->flags)) {
 818                                 rdev_dec_pending(rdev, mddev);
 819                                 r1_bio->bios[i] = NULL;
 820                         } else {
 821                                 r1_bio->bios[i] = bio;
 822                                 targets++;
 823                         }
 824                 } else
 825                         r1_bio->bios[i] = NULL;
 826         }
 827         rcu_read_unlock();
 828
 829         if (unlikely(blocked_rdev)) {
 830                 /* Wait for this device to become unblocked */
 831                 int j;
 832
 833                 for (j = 0; j < i; j++)
 834                         if (r1_bio->bios[j])
 835                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 836
 837                 allow_barrier(conf);
 838                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 839                 wait_barrier(conf);
 840                 goto retry_write;
 841         }
 842
 843         BUG_ON(targets == 0); /* we never fail the last device */
 844
 845         if (targets < conf->raid_disks) {
 846                 /* array is degraded, we will not clear the bitmap
 847                  * on I/O completion (see raid1_end_write_request) */
 848                 set_bit(R1BIO_Degraded, &r1_bio->state);
 849         }
 850
 851         /* do behind I/O ?
 852          * Not if there are too many, or cannot allocate memory,
 853          * or a reader on WriteMostly is waiting for behind writes
 854          * to flush */
 855         if (bitmap &&
 856             (atomic_read(&bitmap->behind_writes)
 857              < mddev->bitmap_info.max_write_behind) &&
 858             !waitqueue_active(&bitmap->behind_wait) &&
 859             (behind_pages = alloc_behind_pages(bio)) != NULL)
 860                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 861
 862         atomic_set(&r1_bio->remaining, 1);
 863         atomic_set(&r1_bio->behind_remaining, 0);
 864
 865         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 866                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 867         for (i = 0; i < disks; i++) {
 868                 struct bio *mbio;
 869                 if (!r1_bio->bios[i])
 870                         continue;
 871
 872                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 873                 r1_bio->bios[i] = mbio;
 874
 875                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 876                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 877                 mbio->bi_end_io = raid1_end_write_request;
 878                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
 879                 mbio->bi_private = r1_bio;
 880
 881                 if (behind_pages) {
 882                         struct bio_vec *bvec;
 883                         int j;
 884
 885                         /* Yes, I really want the '__' version so that
 886                          * we clear any unused pointer in the io_vec, rather
 887                          * than leave them unchanged.  This is important
 888                          * because when we come to free the pages, we won't
 889                          * know the original bi_idx, so we just free
 890                          * them all
 891                          */
 892                         __bio_for_each_segment(bvec, mbio, j, 0)
 893                                 bvec->bv_page = behind_pages[j].bv_page;
 894                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 895                                 atomic_inc(&r1_bio->behind_remaining);
 896                 }
 897
 898                 atomic_inc(&r1_bio->remaining);
 899                 spin_lock_irqsave(&conf->device_lock, flags);
 900                 bio_list_add(&conf->pending_bio_list, mbio);
 901                 spin_unlock_irqrestore(&conf->device_lock, flags);
 902         }
 903         r1_bio_write_done(r1_bio, bio->bi_vcnt, behind_pages, behind_pages != NULL);
 904         kfree(behind_pages); /* the behind pages are attached to the bios now */
 905
 906         /* In case raid1d snuck in to freeze_array */
 907         wake_up(&conf->wait_barrier);
 908
 909         if (do_sync || !bitmap || !plugged)
 910                 md_wakeup_thread(mddev->thread);
 911
 912         return 0;
 913 }
 914
 915 static void status(struct seq_file *seq, mddev_t *mddev)
 916 {
 917         conf_t *conf = mddev->private;
 918         int i;
 919
 920         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 921                    conf->raid_disks - mddev->degraded);
 922         rcu_read_lock();
 923         for (i = 0; i < conf->raid_disks; i++) {
 924                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 925                 seq_printf(seq, "%s",
 926                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
 927         }
 928         rcu_read_unlock();
 929         seq_printf(seq, "]");
 930 }
 931
 932
 933 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 934 {
 935         char b[BDEVNAME_SIZE];
 936         conf_t *conf = mddev->private;
 937
 938         /*
 939          * If it is not operational, then we have already marked it as dead
 940          * else if it is the last working disks, ignore the error, let the
 941          * next level up know.
 942          * else mark the drive as failed
 943          */
 944         if (test_bit(In_sync, &rdev->flags)
 945             && (conf->raid_disks - mddev->degraded) == 1) {
 946                 /*
 947                  * Don't fail the drive, act as though we were just a
 948                  * normal single drive.
 949                  * However don't try a recovery from this drive as
 950                  * it is very likely to fail.
 951                  */
 952                 mddev->recovery_disabled = 1;
 953                 return;
 954         }
 955         if (test_and_clear_bit(In_sync, &rdev->flags)) {
 956                 unsigned long flags;
 957                 spin_lock_irqsave(&conf->device_lock, flags);
 958                 mddev->degraded++;
 959                 set_bit(Faulty, &rdev->flags);
 960                 spin_unlock_irqrestore(&conf->device_lock, flags);
 961                 /*
 962                  * if recovery is running, make sure it aborts.
 963                  */
 964                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
 965         } else
 966                 set_bit(Faulty, &rdev->flags);
 967         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 968         printk(KERN_ALERT
 969                "md/raid1:%s: Disk failure on %s, disabling device.\n"
 970                "md/raid1:%s: Operation continuing on %d devices.\n",
 971                mdname(mddev), bdevname(rdev->bdev, b),
 972                mdname(mddev), conf->raid_disks - mddev->degraded);
 973 }
 974
 975 static void print_conf(conf_t *conf)
 976 {
 977         int i;
 978
 979         printk(KERN_DEBUG "RAID1 conf printout:\n");
 980         if (!conf) {
 981                 printk(KERN_DEBUG "(!conf)\n");
 982                 return;
 983         }
 984         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
 985                 conf->raid_disks);
 986
 987         rcu_read_lock();
 988         for (i = 0; i < conf->raid_disks; i++) {
 989                 char b[BDEVNAME_SIZE];
 990                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 991                 if (rdev)
 992                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
 993                                i, !test_bit(In_sync, &rdev->flags),
 994                                !test_bit(Faulty, &rdev->flags),
 995                                bdevname(rdev->bdev,b));
 996         }
 997         rcu_read_unlock();
 998 }
 999
1000 static void close_sync(conf_t *conf)
1001 {
1002         wait_barrier(conf);
1003         allow_barrier(conf);
1004
1005         mempool_destroy(conf->r1buf_pool);
1006         conf->r1buf_pool = NULL;
1007 }
1008
1009 static int raid1_spare_active(mddev_t *mddev)
1010 {
1011         int i;
1012         conf_t *conf = mddev->private;
1013         int count = 0;
1014         unsigned long flags;
1015
1016         /*
1017          * Find all failed disks within the RAID1 configuration
1018          * and mark them readable.
1019          * Called under mddev lock, so rcu protection not needed.
1020          */
1021         for (i = 0; i < conf->raid_disks; i++) {
1022                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1023                 if (rdev
1024                     && !test_bit(Faulty, &rdev->flags)
1025                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1026                         count++;
1027                         sysfs_notify_dirent(rdev->sysfs_state);
1028                 }
1029         }
1030         spin_lock_irqsave(&conf->device_lock, flags);
1031         mddev->degraded -= count;
1032         spin_unlock_irqrestore(&conf->device_lock, flags);
1033
1034         print_conf(conf);
1035         return count;
1036 }
1037
1038
1039 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1040 {
1041         conf_t *conf = mddev->private;
1042         int err = -EEXIST;
1043         int mirror = 0;
1044         mirror_info_t *p;
1045         int first = 0;
1046         int last = mddev->raid_disks - 1;
1047
1048         if (rdev->raid_disk >= 0)
1049                 first = last = rdev->raid_disk;
1050
1051         for (mirror = first; mirror <= last; mirror++)
1052                 if ( !(p=conf->mirrors+mirror)->rdev) {
1053
1054                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1055                                           rdev->data_offset << 9);
1056                         /* as we don't honour merge_bvec_fn, we must
1057                          * never risk violating it, so limit
1058                          * ->max_segments to one lying with a single
1059                          * page, as a one page request is never in
1060                          * violation.
1061                          */
1062                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1063                                 blk_queue_max_segments(mddev->queue, 1);
1064                                 blk_queue_segment_boundary(mddev->queue,
1065                                                            PAGE_CACHE_SIZE - 1);
1066                         }
1067
1068                         p->head_position = 0;
1069                         rdev->raid_disk = mirror;
1070                         err = 0;
1071                         /* As all devices are equivalent, we don't need a full recovery
1072                          * if this was recently any drive of the array
1073                          */
1074                         if (rdev->saved_raid_disk < 0)
1075                                 conf->fullsync = 1;
1076                         rcu_assign_pointer(p->rdev, rdev);
1077                         break;
1078                 }
1079         md_integrity_add_rdev(rdev, mddev);
1080         print_conf(conf);
1081         return err;
1082 }
1083
1084 static int raid1_remove_disk(mddev_t *mddev, int number)
1085 {
1086         conf_t *conf = mddev->private;
1087         int err = 0;
1088         mdk_rdev_t *rdev;
1089         mirror_info_t *p = conf->mirrors+ number;
1090
1091         print_conf(conf);
1092         rdev = p->rdev;
1093         if (rdev) {
1094                 if (test_bit(In_sync, &rdev->flags) ||
1095                     atomic_read(&rdev->nr_pending)) {
1096                         err = -EBUSY;
1097                         goto abort;
1098                 }
1099                 /* Only remove non-faulty devices if recovery
1100                  * is not possible.
1101                  */
1102                 if (!test_bit(Faulty, &rdev->flags) &&
1103                     !mddev->recovery_disabled &&
1104                     mddev->degraded < conf->raid_disks) {
1105                         err = -EBUSY;
1106                         goto abort;
1107                 }
1108                 p->rdev = NULL;
1109                 synchronize_rcu();
1110                 if (atomic_read(&rdev->nr_pending)) {
1111                         /* lost the race, try later */
1112                         err = -EBUSY;
1113                         p->rdev = rdev;
1114                         goto abort;
1115                 }
1116                 err = md_integrity_register(mddev);
1117         }
1118 abort:
1119
1120         print_conf(conf);
1121         return err;
1122 }
1123
1124
1125 static void end_sync_read(struct bio *bio, int error)
1126 {
1127         r1bio_t *r1_bio = bio->bi_private;
1128         int i;
1129
1130         for (i=r1_bio->mddev->raid_disks; i--; )
1131                 if (r1_bio->bios[i] == bio)
1132                         break;
1133         BUG_ON(i < 0);
1134         update_head_pos(i, r1_bio);
1135         /*
1136          * we have read a block, now it needs to be re-written,
1137          * or re-read if the read failed.
1138          * We don't do much here, just schedule handling by raid1d
1139          */
1140         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1141                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1142
1143         if (atomic_dec_and_test(&r1_bio->remaining))
1144                 reschedule_retry(r1_bio);
1145 }
1146
1147 static void end_sync_write(struct bio *bio, int error)
1148 {
1149         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1150         r1bio_t *r1_bio = bio->bi_private;
1151         mddev_t *mddev = r1_bio->mddev;
1152         conf_t *conf = mddev->private;
1153         int i;
1154         int mirror=0;
1155
1156         for (i = 0; i < conf->raid_disks; i++)
1157                 if (r1_bio->bios[i] == bio) {
1158                         mirror = i;
1159                         break;
1160                 }
1161         if (!uptodate) {
1162                 sector_t sync_blocks = 0;
1163                 sector_t s = r1_bio->sector;
1164                 long sectors_to_go = r1_bio->sectors;
1165                 /* make sure these bits doesn't get cleared. */
1166                 do {
1167                         bitmap_end_sync(mddev->bitmap, s,
1168                                         &sync_blocks, 1);
1169                         s += sync_blocks;
1170                         sectors_to_go -= sync_blocks;
1171                 } while (sectors_to_go > 0);
1172                 md_error(mddev, conf->mirrors[mirror].rdev);
1173         }
1174
1175         update_head_pos(mirror, r1_bio);
1176
1177         if (atomic_dec_and_test(&r1_bio->remaining)) {
1178                 sector_t s = r1_bio->sectors;
1179                 put_buf(r1_bio);
1180                 md_done_sync(mddev, s, uptodate);
1181         }
1182 }
1183
1184 static int fix_sync_read_error(r1bio_t *r1_bio)
1185 {
1186         /* Try some synchronous reads of other devices to get
1187          * good data, much like with normal read errors.  Only
1188          * read into the pages we already have so we don't
1189          * need to re-issue the read request.
1190          * We don't need to freeze the array, because being in an
1191          * active sync request, there is no normal IO, and
1192          * no overlapping syncs.
1193          */
1194         mddev_t *mddev = r1_bio->mddev;
1195         conf_t *conf = mddev->private;
1196         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1197         sector_t sect = r1_bio->sector;
1198         int sectors = r1_bio->sectors;
1199         int idx = 0;
1200
1201         while(sectors) {
1202                 int s = sectors;
1203                 int d = r1_bio->read_disk;
1204                 int success = 0;
1205                 mdk_rdev_t *rdev;
1206                 int start;
1207
1208                 if (s > (PAGE_SIZE>>9))
1209                         s = PAGE_SIZE >> 9;
1210                 do {
1211                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1212                                 /* No rcu protection needed here devices
1213                                  * can only be removed when no resync is
1214                                  * active, and resync is currently active
1215                                  */
1216                                 rdev = conf->mirrors[d].rdev;
1217                                 if (sync_page_io(rdev,
1218                                                  sect,
1219                                                  s<<9,
1220                                                  bio->bi_io_vec[idx].bv_page,
1221                                                  READ, false)) {
1222                                         success = 1;
1223                                         break;
1224                                 }
1225                         }
1226                         d++;
1227                         if (d == conf->raid_disks)
1228                                 d = 0;
1229                 } while (!success && d != r1_bio->read_disk);
1230
1231                 if (!success) {
1232                         char b[BDEVNAME_SIZE];
1233                         /* Cannot read from anywhere, array is toast */
1234                         md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1235                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1236                                " for block %llu\n",
1237                                mdname(mddev),
1238                                bdevname(bio->bi_bdev, b),
1239                                (unsigned long long)r1_bio->sector);
1240                         md_done_sync(mddev, r1_bio->sectors, 0);
1241                         put_buf(r1_bio);
1242                         return 0;
1243                 }
1244
1245                 start = d;
1246                 /* write it back and re-read */
1247                 while (d != r1_bio->read_disk) {
1248                         if (d == 0)
1249                                 d = conf->raid_disks;
1250                         d--;
1251                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1252                                 continue;
1253                         rdev = conf->mirrors[d].rdev;
1254                         if (sync_page_io(rdev,
1255                                          sect,
1256                                          s<<9,
1257                                          bio->bi_io_vec[idx].bv_page,
1258                                          WRITE, false) == 0) {
1259                                 r1_bio->bios[d]->bi_end_io = NULL;
1260                                 rdev_dec_pending(rdev, mddev);
1261                                 md_error(mddev, rdev);
1262                         } else
1263                                 atomic_add(s, &rdev->corrected_errors);
1264                 }
1265                 d = start;
1266                 while (d != r1_bio->read_disk) {
1267                         if (d == 0)
1268                                 d = conf->raid_disks;
1269                         d--;
1270                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1271                                 continue;
1272                         rdev = conf->mirrors[d].rdev;
1273                         if (sync_page_io(rdev,
1274                                          sect,
1275                                          s<<9,
1276                                          bio->bi_io_vec[idx].bv_page,
1277                                          READ, false) == 0)
1278                                 md_error(mddev, rdev);
1279                 }
1280                 sectors -= s;
1281                 sect += s;
1282                 idx ++;
1283         }
1284         set_bit(R1BIO_Uptodate, &r1_bio->state);
1285         return 1;
1286 }
1287
1288 static int process_checks(r1bio_t *r1_bio)
1289 {
1290         /* We have read all readable devices.  If we haven't
1291          * got the block, then there is no hope left.
1292          * If we have, then we want to do a comparison
1293          * and skip the write if everything is the same.
1294          * If any blocks failed to read, then we need to
1295          * attempt an over-write
1296          */
1297         mddev_t *mddev = r1_bio->mddev;
1298         conf_t *conf = mddev->private;
1299         int primary;
1300         int i;
1301
1302         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1303                 for (i=0; i < conf->raid_disks; i++)
1304                         if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1305                                 md_error(mddev, conf->mirrors[i].rdev);
1306
1307                 md_done_sync(mddev, r1_bio->sectors, 1);
1308                 put_buf(r1_bio);
1309                 return -1;
1310         }
1311         for (primary = 0; primary < conf->raid_disks; primary++)
1312                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1313                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1314                         r1_bio->bios[primary]->bi_end_io = NULL;
1315                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1316                         break;
1317                 }
1318         r1_bio->read_disk = primary;
1319         for (i = 0; i < conf->raid_disks; i++) {
1320                 int j;
1321                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1322                 struct bio *pbio = r1_bio->bios[primary];
1323                 struct bio *sbio = r1_bio->bios[i];
1324                 int size;
1325
1326                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1327                         continue;
1328
1329                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1330                         for (j = vcnt; j-- ; ) {
1331                                 struct page *p, *s;
1332                                 p = pbio->bi_io_vec[j].bv_page;
1333                                 s = sbio->bi_io_vec[j].bv_page;
1334                                 if (memcmp(page_address(p),
1335                                            page_address(s),
1336                                            PAGE_SIZE))
1337                                         break;
1338                         }
1339                 } else
1340                         j = 0;
1341                 if (j >= 0)
1342                         mddev->resync_mismatches += r1_bio->sectors;
1343                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1344                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1345                         /* No need to write to this device. */
1346                         sbio->bi_end_io = NULL;
1347                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1348                         continue;
1349                 }
1350                 /* fixup the bio for reuse */
1351                 sbio->bi_vcnt = vcnt;
1352                 sbio->bi_size = r1_bio->sectors << 9;
1353                 sbio->bi_idx = 0;
1354                 sbio->bi_phys_segments = 0;
1355                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1356                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1357                 sbio->bi_next = NULL;
1358                 sbio->bi_sector = r1_bio->sector +
1359                         conf->mirrors[i].rdev->data_offset;
1360                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1361                 size = sbio->bi_size;
1362                 for (j = 0; j < vcnt ; j++) {
1363                         struct bio_vec *bi;
1364                         bi = &sbio->bi_io_vec[j];
1365                         bi->bv_offset = 0;
1366                         if (size > PAGE_SIZE)
1367                                 bi->bv_len = PAGE_SIZE;
1368                         else
1369                                 bi->bv_len = size;
1370                         size -= PAGE_SIZE;
1371                         memcpy(page_address(bi->bv_page),
1372                                page_address(pbio->bi_io_vec[j].bv_page),
1373                                PAGE_SIZE);
1374                 }
1375         }
1376         return 0;
1377 }
1378
1379 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1380 {
1381         conf_t *conf = mddev->private;
1382         int i;
1383         int disks = conf->raid_disks;
1384         struct bio *bio, *wbio;
1385
1386         bio = r1_bio->bios[r1_bio->read_disk];
1387
1388
1389         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1390                 if (process_checks(r1_bio) < 0)
1391                         return;
1392
1393         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1394                 /* ouch - failed to read all of that. */
1395                 if (!fix_sync_read_error(r1_bio))
1396                         return;
1397         /*
1398          * schedule writes
1399          */
1400         atomic_set(&r1_bio->remaining, 1);
1401         for (i = 0; i < disks ; i++) {
1402                 wbio = r1_bio->bios[i];
1403                 if (wbio->bi_end_io == NULL ||
1404                     (wbio->bi_end_io == end_sync_read &&
1405                      (i == r1_bio->read_disk ||
1406                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1407                         continue;
1408
1409                 wbio->bi_rw = WRITE;
1410                 wbio->bi_end_io = end_sync_write;
1411                 atomic_inc(&r1_bio->remaining);
1412                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1413
1414                 generic_make_request(wbio);
1415         }
1416
1417         if (atomic_dec_and_test(&r1_bio->remaining)) {
1418                 /* if we're here, all write(s) have completed, so clean up */
1419                 md_done_sync(mddev, r1_bio->sectors, 1);
1420                 put_buf(r1_bio);
1421         }
1422 }
1423
1424 /*
1425  * This is a kernel thread which:
1426  *
1427  *      1.      Retries failed read operations on working mirrors.
1428  *      2.      Updates the raid superblock when problems encounter.
1429  *      3.      Performs writes following reads for array syncronising.
1430  */
1431
1432 static void fix_read_error(conf_t *conf, int read_disk,
1433                            sector_t sect, int sectors)
1434 {
1435         mddev_t *mddev = conf->mddev;
1436         while(sectors) {
1437                 int s = sectors;
1438                 int d = read_disk;
1439                 int success = 0;
1440                 int start;
1441                 mdk_rdev_t *rdev;
1442
1443                 if (s > (PAGE_SIZE>>9))
1444                         s = PAGE_SIZE >> 9;
1445
1446                 do {
1447                         /* Note: no rcu protection needed here
1448                          * as this is synchronous in the raid1d thread
1449                          * which is the thread that might remove
1450                          * a device.  If raid1d ever becomes multi-threaded....
1451                          */
1452                         rdev = conf->mirrors[d].rdev;
1453                         if (rdev &&
1454                             test_bit(In_sync, &rdev->flags) &&
1455                             sync_page_io(rdev, sect, s<<9,
1456                                          conf->tmppage, READ, false))
1457                                 success = 1;
1458                         else {
1459                                 d++;
1460                                 if (d == conf->raid_disks)
1461                                         d = 0;
1462                         }
1463                 } while (!success && d != read_disk);
1464
1465                 if (!success) {
1466                         /* Cannot read from anywhere -- bye bye array */
1467                         md_error(mddev, conf->mirrors[read_disk].rdev);
1468                         break;
1469                 }
1470                 /* write it back and re-read */
1471                 start = d;
1472                 while (d != read_disk) {
1473                         if (d==0)
1474                                 d = conf->raid_disks;
1475                         d--;
1476                         rdev = conf->mirrors[d].rdev;
1477                         if (rdev &&
1478                             test_bit(In_sync, &rdev->flags)) {
1479                                 if (sync_page_io(rdev, sect, s<<9,
1480                                                  conf->tmppage, WRITE, false)
1481                                     == 0)
1482                                         /* Well, this device is dead */
1483                                         md_error(mddev, rdev);
1484                         }
1485                 }
1486                 d = start;
1487                 while (d != read_disk) {
1488                         char b[BDEVNAME_SIZE];
1489                         if (d==0)
1490                                 d = conf->raid_disks;
1491                         d--;
1492                         rdev = conf->mirrors[d].rdev;
1493                         if (rdev &&
1494                             test_bit(In_sync, &rdev->flags)) {
1495                                 if (sync_page_io(rdev, sect, s<<9,
1496                                                  conf->tmppage, READ, false)
1497                                     == 0)
1498                                         /* Well, this device is dead */
1499                                         md_error(mddev, rdev);
1500                                 else {
1501                                         atomic_add(s, &rdev->corrected_errors);
1502                                         printk(KERN_INFO
1503                                                "md/raid1:%s: read error corrected "
1504                                                "(%d sectors at %llu on %s)\n",
1505                                                mdname(mddev), s,
1506                                                (unsigned long long)(sect +
1507                                                    rdev->data_offset),
1508                                                bdevname(rdev->bdev, b));
1509                                 }
1510                         }
1511                 }
1512                 sectors -= s;
1513                 sect += s;
1514         }
1515 }
1516
1517 static void raid1d(mddev_t *mddev)
1518 {
1519         r1bio_t *r1_bio;
1520         struct bio *bio;
1521         unsigned long flags;
1522         conf_t *conf = mddev->private;
1523         struct list_head *head = &conf->retry_list;
1524         mdk_rdev_t *rdev;
1525         struct blk_plug plug;
1526
1527         md_check_recovery(mddev);
1528
1529         blk_start_plug(&plug);
1530         for (;;) {
1531                 char b[BDEVNAME_SIZE];
1532
1533                 if (atomic_read(&mddev->plug_cnt) == 0)
1534                         flush_pending_writes(conf);
1535
1536                 spin_lock_irqsave(&conf->device_lock, flags);
1537                 if (list_empty(head)) {
1538                         spin_unlock_irqrestore(&conf->device_lock, flags);
1539                         break;
1540                 }
1541                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1542                 list_del(head->prev);
1543                 conf->nr_queued--;
1544                 spin_unlock_irqrestore(&conf->device_lock, flags);
1545
1546                 mddev = r1_bio->mddev;
1547                 conf = mddev->private;
1548                 if (test_bit(R1BIO_IsSync, &r1_bio->state))
1549                         sync_request_write(mddev, r1_bio);
1550                 else {
1551                         int disk;
1552
1553                         /* we got a read error. Maybe the drive is bad.  Maybe just
1554                          * the block and we can fix it.
1555                          * We freeze all other IO, and try reading the block from
1556                          * other devices.  When we find one, we re-write
1557                          * and check it that fixes the read error.
1558                          * This is all done synchronously while the array is
1559                          * frozen
1560                          */
1561                         if (mddev->ro == 0) {
1562                                 freeze_array(conf);
1563                                 fix_read_error(conf, r1_bio->read_disk,
1564                                                r1_bio->sector,
1565                                                r1_bio->sectors);
1566                                 unfreeze_array(conf);
1567                         } else
1568                                 md_error(mddev,
1569                                          conf->mirrors[r1_bio->read_disk].rdev);
1570
1571                         bio = r1_bio->bios[r1_bio->read_disk];
1572                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1573                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1574                                        " read error for block %llu\n",
1575                                        mdname(mddev),
1576                                        bdevname(bio->bi_bdev,b),
1577                                        (unsigned long long)r1_bio->sector);
1578                                 raid_end_bio_io(r1_bio);
1579                         } else {
1580                                 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1581                                 r1_bio->bios[r1_bio->read_disk] =
1582                                         mddev->ro ? IO_BLOCKED : NULL;
1583                                 r1_bio->read_disk = disk;
1584                                 bio_put(bio);
1585                                 bio = bio_clone_mddev(r1_bio->master_bio,
1586                                                       GFP_NOIO, mddev);
1587                                 r1_bio->bios[r1_bio->read_disk] = bio;
1588                                 rdev = conf->mirrors[disk].rdev;
1589                                 if (printk_ratelimit())
1590                                         printk(KERN_ERR "md/raid1:%s: redirecting sector %llu to"
1591                                                " other mirror: %s\n",
1592                                                mdname(mddev),
1593                                                (unsigned long long)r1_bio->sector,
1594                                                bdevname(rdev->bdev,b));
1595                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1596                                 bio->bi_bdev = rdev->bdev;
1597                                 bio->bi_end_io = raid1_end_read_request;
1598                                 bio->bi_rw = READ | do_sync;
1599                                 bio->bi_private = r1_bio;
1600                                 generic_make_request(bio);
1601                         }
1602                 }
1603                 cond_resched();
1604         }
1605         blk_finish_plug(&plug);
1606 }
1607
1608
1609 static int init_resync(conf_t *conf)
1610 {
1611         int buffs;
1612
1613         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1614         BUG_ON(conf->r1buf_pool);
1615         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1616                                           conf->poolinfo);
1617         if (!conf->r1buf_pool)
1618                 return -ENOMEM;
1619         conf->next_resync = 0;
1620         return 0;
1621 }
1622
1623 /*
1624  * perform a "sync" on one "block"
1625  *
1626  * We need to make sure that no normal I/O request - particularly write
1627  * requests - conflict with active sync requests.
1628  *
1629  * This is achieved by tracking pending requests and a 'barrier' concept
1630  * that can be installed to exclude normal IO requests.
1631  */
1632
1633 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1634 {
1635         conf_t *conf = mddev->private;
1636         r1bio_t *r1_bio;
1637         struct bio *bio;
1638         sector_t max_sector, nr_sectors;
1639         int disk = -1;
1640         int i;
1641         int wonly = -1;
1642         int write_targets = 0, read_targets = 0;
1643         sector_t sync_blocks;
1644         int still_degraded = 0;
1645
1646         if (!conf->r1buf_pool)
1647                 if (init_resync(conf))
1648                         return 0;
1649
1650         max_sector = mddev->dev_sectors;
1651         if (sector_nr >= max_sector) {
1652                 /* If we aborted, we need to abort the
1653                  * sync on the 'current' bitmap chunk (there will
1654                  * only be one in raid1 resync.
1655                  * We can find the current addess in mddev->curr_resync
1656                  */
1657                 if (mddev->curr_resync < max_sector) /* aborted */
1658                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1659                                                 &sync_blocks, 1);
1660                 else /* completed sync */
1661                         conf->fullsync = 0;
1662
1663                 bitmap_close_sync(mddev->bitmap);
1664                 close_sync(conf);
1665                 return 0;
1666         }
1667
1668         if (mddev->bitmap == NULL &&
1669             mddev->recovery_cp == MaxSector &&
1670             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1671             conf->fullsync == 0) {
1672                 *skipped = 1;
1673                 return max_sector - sector_nr;
1674         }
1675         /* before building a request, check if we can skip these blocks..
1676          * This call the bitmap_start_sync doesn't actually record anything
1677          */
1678         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1679             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1680                 /* We can skip this block, and probably several more */
1681                 *skipped = 1;
1682                 return sync_blocks;
1683         }
1684         /*
1685          * If there is non-resync activity waiting for a turn,
1686          * and resync is going fast enough,
1687          * then let it though before starting on this new sync request.
1688          */
1689         if (!go_faster && conf->nr_waiting)
1690                 msleep_interruptible(1000);
1691
1692         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1693         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1694         raise_barrier(conf);
1695
1696         conf->next_resync = sector_nr;
1697
1698         rcu_read_lock();
1699         /*
1700          * If we get a correctably read error during resync or recovery,
1701          * we might want to read from a different device.  So we
1702          * flag all drives that could conceivably be read from for READ,
1703          * and any others (which will be non-In_sync devices) for WRITE.
1704          * If a read fails, we try reading from something else for which READ
1705          * is OK.
1706          */
1707
1708         r1_bio->mddev = mddev;
1709         r1_bio->sector = sector_nr;
1710         r1_bio->state = 0;
1711         set_bit(R1BIO_IsSync, &r1_bio->state);
1712
1713         for (i=0; i < conf->raid_disks; i++) {
1714                 mdk_rdev_t *rdev;
1715                 bio = r1_bio->bios[i];
1716
1717                 /* take from bio_init */
1718                 bio->bi_next = NULL;
1719                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1720                 bio->bi_flags |= 1 << BIO_UPTODATE;
1721                 bio->bi_comp_cpu = -1;
1722                 bio->bi_rw = READ;
1723                 bio->bi_vcnt = 0;
1724                 bio->bi_idx = 0;
1725                 bio->bi_phys_segments = 0;
1726                 bio->bi_size = 0;
1727                 bio->bi_end_io = NULL;
1728                 bio->bi_private = NULL;
1729
1730                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1731                 if (rdev == NULL ||
1732                            test_bit(Faulty, &rdev->flags)) {
1733                         still_degraded = 1;
1734                         continue;
1735                 } else if (!test_bit(In_sync, &rdev->flags)) {
1736                         bio->bi_rw = WRITE;
1737                         bio->bi_end_io = end_sync_write;
1738                         write_targets ++;
1739                 } else {
1740                         /* may need to read from here */
1741                         bio->bi_rw = READ;
1742                         bio->bi_end_io = end_sync_read;
1743                         if (test_bit(WriteMostly, &rdev->flags)) {
1744                                 if (wonly < 0)
1745                                         wonly = i;
1746                         } else {
1747                                 if (disk < 0)
1748                                         disk = i;
1749                         }
1750                         read_targets++;
1751                 }
1752                 atomic_inc(&rdev->nr_pending);
1753                 bio->bi_sector = sector_nr + rdev->data_offset;
1754                 bio->bi_bdev = rdev->bdev;
1755                 bio->bi_private = r1_bio;
1756         }
1757         rcu_read_unlock();
1758         if (disk < 0)
1759                 disk = wonly;
1760         r1_bio->read_disk = disk;
1761
1762         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1763                 /* extra read targets are also write targets */
1764                 write_targets += read_targets-1;
1765
1766         if (write_targets == 0 || read_targets == 0) {
1767                 /* There is nowhere to write, so all non-sync
1768                  * drives must be failed - so we are finished
1769                  */
1770                 sector_t rv = max_sector - sector_nr;
1771                 *skipped = 1;
1772                 put_buf(r1_bio);
1773                 return rv;
1774         }
1775
1776         if (max_sector > mddev->resync_max)
1777                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1778         nr_sectors = 0;
1779         sync_blocks = 0;
1780         do {
1781                 struct page *page;
1782                 int len = PAGE_SIZE;
1783                 if (sector_nr + (len>>9) > max_sector)
1784                         len = (max_sector - sector_nr) << 9;
1785                 if (len == 0)
1786                         break;
1787                 if (sync_blocks == 0) {
1788                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1789                                                &sync_blocks, still_degraded) &&
1790                             !conf->fullsync &&
1791                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1792                                 break;
1793                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1794                         if ((len >> 9) > sync_blocks)
1795                                 len = sync_blocks<<9;
1796                 }
1797
1798                 for (i=0 ; i < conf->raid_disks; i++) {
1799                         bio = r1_bio->bios[i];
1800                         if (bio->bi_end_io) {
1801                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1802                                 if (bio_add_page(bio, page, len, 0) == 0) {
1803                                         /* stop here */
1804                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1805                                         while (i > 0) {
1806                                                 i--;
1807                                                 bio = r1_bio->bios[i];
1808                                                 if (bio->bi_end_io==NULL)
1809                                                         continue;
1810                                                 /* remove last page from this bio */
1811                                                 bio->bi_vcnt--;
1812                                                 bio->bi_size -= len;
1813                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1814                                         }
1815                                         goto bio_full;
1816                                 }
1817                         }
1818                 }
1819                 nr_sectors += len>>9;
1820                 sector_nr += len>>9;
1821                 sync_blocks -= (len>>9);
1822         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1823  bio_full:
1824         r1_bio->sectors = nr_sectors;
1825
1826         /* For a user-requested sync, we read all readable devices and do a
1827          * compare
1828          */
1829         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1830                 atomic_set(&r1_bio->remaining, read_targets);
1831                 for (i=0; i<conf->raid_disks; i++) {
1832                         bio = r1_bio->bios[i];
1833                         if (bio->bi_end_io == end_sync_read) {
1834                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1835                                 generic_make_request(bio);
1836                         }
1837                 }
1838         } else {
1839                 atomic_set(&r1_bio->remaining, 1);
1840                 bio = r1_bio->bios[r1_bio->read_disk];
1841                 md_sync_acct(bio->bi_bdev, nr_sectors);
1842                 generic_make_request(bio);
1843
1844         }
1845         return nr_sectors;
1846 }
1847
1848 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1849 {
1850         if (sectors)
1851                 return sectors;
1852
1853         return mddev->dev_sectors;
1854 }
1855
1856 static conf_t *setup_conf(mddev_t *mddev)
1857 {
1858         conf_t *conf;
1859         int i;
1860         mirror_info_t *disk;
1861         mdk_rdev_t *rdev;
1862         int err = -ENOMEM;
1863
1864         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1865         if (!conf)
1866                 goto abort;
1867
1868         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1869                                  GFP_KERNEL);
1870         if (!conf->mirrors)
1871                 goto abort;
1872
1873         conf->tmppage = alloc_page(GFP_KERNEL);
1874         if (!conf->tmppage)
1875                 goto abort;
1876
1877         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1878         if (!conf->poolinfo)
1879                 goto abort;
1880         conf->poolinfo->raid_disks = mddev->raid_disks;
1881         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1882                                           r1bio_pool_free,
1883                                           conf->poolinfo);
1884         if (!conf->r1bio_pool)
1885                 goto abort;
1886
1887         conf->poolinfo->mddev = mddev;
1888
1889         spin_lock_init(&conf->device_lock);
1890         list_for_each_entry(rdev, &mddev->disks, same_set) {
1891                 int disk_idx = rdev->raid_disk;
1892                 if (disk_idx >= mddev->raid_disks
1893                     || disk_idx < 0)
1894                         continue;
1895                 disk = conf->mirrors + disk_idx;
1896
1897                 disk->rdev = rdev;
1898
1899                 disk->head_position = 0;
1900         }
1901         conf->raid_disks = mddev->raid_disks;
1902         conf->mddev = mddev;
1903         INIT_LIST_HEAD(&conf->retry_list);
1904
1905         spin_lock_init(&conf->resync_lock);
1906         init_waitqueue_head(&conf->wait_barrier);
1907
1908         bio_list_init(&conf->pending_bio_list);
1909
1910         conf->last_used = -1;
1911         for (i = 0; i < conf->raid_disks; i++) {
1912
1913                 disk = conf->mirrors + i;
1914
1915                 if (!disk->rdev ||
1916                     !test_bit(In_sync, &disk->rdev->flags)) {
1917                         disk->head_position = 0;
1918                         if (disk->rdev)
1919                                 conf->fullsync = 1;
1920                 } else if (conf->last_used < 0)
1921                         /*
1922                          * The first working device is used as a
1923                          * starting point to read balancing.
1924                          */
1925                         conf->last_used = i;
1926         }
1927
1928         err = -EIO;
1929         if (conf->last_used < 0) {
1930                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
1931                        mdname(mddev));
1932                 goto abort;
1933         }
1934         err = -ENOMEM;
1935         conf->thread = md_register_thread(raid1d, mddev, NULL);
1936         if (!conf->thread) {
1937                 printk(KERN_ERR
1938                        "md/raid1:%s: couldn't allocate thread\n",
1939                        mdname(mddev));
1940                 goto abort;
1941         }
1942
1943         return conf;
1944
1945  abort:
1946         if (conf) {
1947                 if (conf->r1bio_pool)
1948                         mempool_destroy(conf->r1bio_pool);
1949                 kfree(conf->mirrors);
1950                 safe_put_page(conf->tmppage);
1951                 kfree(conf->poolinfo);
1952                 kfree(conf);
1953         }
1954         return ERR_PTR(err);
1955 }
1956
1957 static int run(mddev_t *mddev)
1958 {
1959         conf_t *conf;
1960         int i;
1961         mdk_rdev_t *rdev;
1962
1963         if (mddev->level != 1) {
1964                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
1965                        mdname(mddev), mddev->level);
1966                 return -EIO;
1967         }
1968         if (mddev->reshape_position != MaxSector) {
1969                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
1970                        mdname(mddev));
1971                 return -EIO;
1972         }
1973         /*
1974          * copy the already verified devices into our private RAID1
1975          * bookkeeping area. [whatever we allocate in run(),
1976          * should be freed in stop()]
1977          */
1978         if (mddev->private == NULL)
1979                 conf = setup_conf(mddev);
1980         else
1981                 conf = mddev->private;
1982
1983         if (IS_ERR(conf))
1984                 return PTR_ERR(conf);
1985
1986         list_for_each_entry(rdev, &mddev->disks, same_set) {
1987                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1988                                   rdev->data_offset << 9);
1989                 /* as we don't honour merge_bvec_fn, we must never risk
1990                  * violating it, so limit ->max_segments to 1 lying within
1991                  * a single page, as a one page request is never in violation.
1992                  */
1993                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1994                         blk_queue_max_segments(mddev->queue, 1);
1995                         blk_queue_segment_boundary(mddev->queue,
1996                                                    PAGE_CACHE_SIZE - 1);
1997                 }
1998         }
1999
2000         mddev->degraded = 0;
2001         for (i=0; i < conf->raid_disks; i++)
2002                 if (conf->mirrors[i].rdev == NULL ||
2003                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2004                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2005                         mddev->degraded++;
2006
2007         if (conf->raid_disks - mddev->degraded == 1)
2008                 mddev->recovery_cp = MaxSector;
2009
2010         if (mddev->recovery_cp != MaxSector)
2011                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2012                        " -- starting background reconstruction\n",
2013                        mdname(mddev));
2014         printk(KERN_INFO
2015                 "md/raid1:%s: active with %d out of %d mirrors\n",
2016                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2017                 mddev->raid_disks);
2018
2019         /*
2020          * Ok, everything is just fine now
2021          */
2022         mddev->thread = conf->thread;
2023         conf->thread = NULL;
2024         mddev->private = conf;
2025
2026         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2027
2028         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2029         mddev->queue->backing_dev_info.congested_data = mddev;
2030         return md_integrity_register(mddev);
2031 }
2032
2033 static int stop(mddev_t *mddev)
2034 {
2035         conf_t *conf = mddev->private;
2036         struct bitmap *bitmap = mddev->bitmap;
2037
2038         /* wait for behind writes to complete */
2039         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2040                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2041                        mdname(mddev));
2042                 /* need to kick something here to make sure I/O goes? */
2043                 wait_event(bitmap->behind_wait,
2044                            atomic_read(&bitmap->behind_writes) == 0);
2045         }
2046
2047         raise_barrier(conf);
2048         lower_barrier(conf);
2049
2050         md_unregister_thread(mddev->thread);
2051         mddev->thread = NULL;
2052         if (conf->r1bio_pool)
2053                 mempool_destroy(conf->r1bio_pool);
2054         kfree(conf->mirrors);
2055         kfree(conf->poolinfo);
2056         kfree(conf);
2057         mddev->private = NULL;
2058         return 0;
2059 }
2060
2061 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2062 {
2063         /* no resync is happening, and there is enough space
2064          * on all devices, so we can resize.
2065          * We need to make sure resync covers any new space.
2066          * If the array is shrinking we should possibly wait until
2067          * any io in the removed space completes, but it hardly seems
2068          * worth it.
2069          */
2070         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2071         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2072                 return -EINVAL;
2073         set_capacity(mddev->gendisk, mddev->array_sectors);
2074         revalidate_disk(mddev->gendisk);
2075         if (sectors > mddev->dev_sectors &&
2076             mddev->recovery_cp == MaxSector) {
2077                 mddev->recovery_cp = mddev->dev_sectors;
2078                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2079         }
2080         mddev->dev_sectors = sectors;
2081         mddev->resync_max_sectors = sectors;
2082         return 0;
2083 }
2084
2085 static int raid1_reshape(mddev_t *mddev)
2086 {
2087         /* We need to:
2088          * 1/ resize the r1bio_pool
2089          * 2/ resize conf->mirrors
2090          *
2091          * We allocate a new r1bio_pool if we can.
2092          * Then raise a device barrier and wait until all IO stops.
2093          * Then resize conf->mirrors and swap in the new r1bio pool.
2094          *
2095          * At the same time, we "pack" the devices so that all the missing
2096          * devices have the higher raid_disk numbers.
2097          */
2098         mempool_t *newpool, *oldpool;
2099         struct pool_info *newpoolinfo;
2100         mirror_info_t *newmirrors;
2101         conf_t *conf = mddev->private;
2102         int cnt, raid_disks;
2103         unsigned long flags;
2104         int d, d2, err;
2105
2106         /* Cannot change chunk_size, layout, or level */
2107         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2108             mddev->layout != mddev->new_layout ||
2109             mddev->level != mddev->new_level) {
2110                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2111                 mddev->new_layout = mddev->layout;
2112                 mddev->new_level = mddev->level;
2113                 return -EINVAL;
2114         }
2115
2116         err = md_allow_write(mddev);
2117         if (err)
2118                 return err;
2119
2120         raid_disks = mddev->raid_disks + mddev->delta_disks;
2121
2122         if (raid_disks < conf->raid_disks) {
2123                 cnt=0;
2124                 for (d= 0; d < conf->raid_disks; d++)
2125                         if (conf->mirrors[d].rdev)
2126                                 cnt++;
2127                 if (cnt > raid_disks)
2128                         return -EBUSY;
2129         }
2130
2131         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2132         if (!newpoolinfo)
2133                 return -ENOMEM;
2134         newpoolinfo->mddev = mddev;
2135         newpoolinfo->raid_disks = raid_disks;
2136
2137         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2138                                  r1bio_pool_free, newpoolinfo);
2139         if (!newpool) {
2140                 kfree(newpoolinfo);
2141                 return -ENOMEM;
2142         }
2143         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2144         if (!newmirrors) {
2145                 kfree(newpoolinfo);
2146                 mempool_destroy(newpool);
2147                 return -ENOMEM;
2148         }
2149
2150         raise_barrier(conf);
2151
2152         /* ok, everything is stopped */
2153         oldpool = conf->r1bio_pool;
2154         conf->r1bio_pool = newpool;
2155
2156         for (d = d2 = 0; d < conf->raid_disks; d++) {
2157                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2158                 if (rdev && rdev->raid_disk != d2) {
2159                         char nm[20];
2160                         sprintf(nm, "rd%d", rdev->raid_disk);
2161                         sysfs_remove_link(&mddev->kobj, nm);
2162                         rdev->raid_disk = d2;
2163                         sprintf(nm, "rd%d", rdev->raid_disk);
2164                         sysfs_remove_link(&mddev->kobj, nm);
2165                         if (sysfs_create_link(&mddev->kobj,
2166                                               &rdev->kobj, nm))
2167                                 printk(KERN_WARNING
2168                                        "md/raid1:%s: cannot register "
2169                                        "%s\n",
2170                                        mdname(mddev), nm);
2171                 }
2172                 if (rdev)
2173                         newmirrors[d2++].rdev = rdev;
2174         }
2175         kfree(conf->mirrors);
2176         conf->mirrors = newmirrors;
2177         kfree(conf->poolinfo);
2178         conf->poolinfo = newpoolinfo;
2179
2180         spin_lock_irqsave(&conf->device_lock, flags);
2181         mddev->degraded += (raid_disks - conf->raid_disks);
2182         spin_unlock_irqrestore(&conf->device_lock, flags);
2183         conf->raid_disks = mddev->raid_disks = raid_disks;
2184         mddev->delta_disks = 0;
2185
2186         conf->last_used = 0; /* just make sure it is in-range */
2187         lower_barrier(conf);
2188
2189         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2190         md_wakeup_thread(mddev->thread);
2191
2192         mempool_destroy(oldpool);
2193         return 0;
2194 }
2195
2196 static void raid1_quiesce(mddev_t *mddev, int state)
2197 {
2198         conf_t *conf = mddev->private;
2199
2200         switch(state) {
2201         case 2: /* wake for suspend */
2202                 wake_up(&conf->wait_barrier);
2203                 break;
2204         case 1:
2205                 raise_barrier(conf);
2206                 break;
2207         case 0:
2208                 lower_barrier(conf);
2209                 break;
2210         }
2211 }
2212
2213 static void *raid1_takeover(mddev_t *mddev)
2214 {
2215         /* raid1 can take over:
2216          *  raid5 with 2 devices, any layout or chunk size
2217          */
2218         if (mddev->level == 5 && mddev->raid_disks == 2) {
2219                 conf_t *conf;
2220                 mddev->new_level = 1;
2221                 mddev->new_layout = 0;
2222                 mddev->new_chunk_sectors = 0;
2223                 conf = setup_conf(mddev);
2224                 if (!IS_ERR(conf))
2225                         conf->barrier = 1;
2226                 return conf;
2227         }
2228         return ERR_PTR(-EINVAL);
2229 }
2230
2231 static struct mdk_personality raid1_personality =
2232 {
2233         .name           = "raid1",
2234         .level          = 1,
2235         .owner          = THIS_MODULE,
2236         .make_request   = make_request,
2237         .run            = run,
2238         .stop           = stop,
2239         .status         = status,
2240         .error_handler  = error,
2241         .hot_add_disk   = raid1_add_disk,
2242         .hot_remove_disk= raid1_remove_disk,
2243         .spare_active   = raid1_spare_active,
2244         .sync_request   = sync_request,
2245         .resize         = raid1_resize,
2246         .size           = raid1_size,
2247         .check_reshape  = raid1_reshape,
2248         .quiesce        = raid1_quiesce,
2249         .takeover       = raid1_takeover,
2250 };
2251
2252 static int __init raid_init(void)
2253 {
2254         return register_md_personality(&raid1_personality);
2255 }
2256
2257 static void raid_exit(void)
2258 {
2259         unregister_md_personality(&raid1_personality);
2260 }
2261
2262 module_init(raid_init);
2263 module_exit(raid_exit);
2264 MODULE_LICENSE("GPL");
2265 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2266 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2267 MODULE_ALIAS("md-raid1");
2268 MODULE_ALIAS("md-level-1");