drivers/md/raid1.c

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