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 unplug_slaves(mddev_t *mddev);
  56
  57 static void allow_barrier(conf_t *conf);
  58 static void lower_barrier(conf_t *conf);
  59
  60 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  61 {
  62         struct pool_info *pi = data;
  63         r1bio_t *r1_bio;
  64         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  65
  66         /* allocate a r1bio with room for raid_disks entries in the bios array */
  67         r1_bio = kzalloc(size, gfp_flags);
  68         if (!r1_bio && pi->mddev)
  69                 unplug_slaves(pi->mddev);
  70
  71         return r1_bio;
  72 }
  73
  74 static void r1bio_pool_free(void *r1_bio, void *data)
  75 {
  76         kfree(r1_bio);
  77 }
  78
  79 #define RESYNC_BLOCK_SIZE (64*1024)
  80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  83 #define RESYNC_WINDOW (2048*1024)
  84
  85 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  86 {
  87         struct pool_info *pi = data;
  88         struct page *page;
  89         r1bio_t *r1_bio;
  90         struct bio *bio;
  91         int i, j;
  92
  93         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  94         if (!r1_bio) {
  95                 unplug_slaves(pi->mddev);
  96                 return NULL;
  97         }
  98
  99         /*
 100          * Allocate bios : 1 for reading, n-1 for writing
 101          */
 102         for (j = pi->raid_disks ; j-- ; ) {
 103                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 104                 if (!bio)
 105                         goto out_free_bio;
 106                 r1_bio->bios[j] = bio;
 107         }
 108         /*
 109          * Allocate RESYNC_PAGES data pages and attach them to
 110          * the first bio.
 111          * If this is a user-requested check/repair, allocate
 112          * RESYNC_PAGES for each bio.
 113          */
 114         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 115                 j = pi->raid_disks;
 116         else
 117                 j = 1;
 118         while(j--) {
 119                 bio = r1_bio->bios[j];
 120                 for (i = 0; i < RESYNC_PAGES; i++) {
 121                         page = alloc_page(gfp_flags);
 122                         if (unlikely(!page))
 123                                 goto out_free_pages;
 124
 125                         bio->bi_io_vec[i].bv_page = page;
 126                         bio->bi_vcnt = i+1;
 127                 }
 128         }
 129         /* If not user-requests, copy the page pointers to all bios */
 130         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 131                 for (i=0; i<RESYNC_PAGES ; i++)
 132                         for (j=1; j<pi->raid_disks; j++)
 133                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 134                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 135         }
 136
 137         r1_bio->master_bio = NULL;
 138
 139         return r1_bio;
 140
 141 out_free_pages:
 142         for (j=0 ; j < pi->raid_disks; j++)
 143                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 144                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 145         j = -1;
 146 out_free_bio:
 147         while ( ++j < pi->raid_disks )
 148                 bio_put(r1_bio->bios[j]);
 149         r1bio_pool_free(r1_bio, data);
 150         return NULL;
 151 }
 152
 153 static void r1buf_pool_free(void *__r1_bio, void *data)
 154 {
 155         struct pool_info *pi = data;
 156         int i,j;
 157         r1bio_t *r1bio = __r1_bio;
 158
 159         for (i = 0; i < RESYNC_PAGES; i++)
 160                 for (j = pi->raid_disks; j-- ;) {
 161                         if (j == 0 ||
 162                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 163                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 164                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 165                 }
 166         for (i=0 ; i < pi->raid_disks; i++)
 167                 bio_put(r1bio->bios[i]);
 168
 169         r1bio_pool_free(r1bio, data);
 170 }
 171
 172 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 173 {
 174         int i;
 175
 176         for (i = 0; i < conf->raid_disks; i++) {
 177                 struct bio **bio = r1_bio->bios + i;
 178                 if (*bio && *bio != IO_BLOCKED)
 179                         bio_put(*bio);
 180                 *bio = NULL;
 181         }
 182 }
 183
 184 static void free_r1bio(r1bio_t *r1_bio)
 185 {
 186         conf_t *conf = r1_bio->mddev->private;
 187
 188         /*
 189          * Wake up any possible resync thread that waits for the device
 190          * to go idle.
 191          */
 192         allow_barrier(conf);
 193
 194         put_all_bios(conf, r1_bio);
 195         mempool_free(r1_bio, conf->r1bio_pool);
 196 }
 197
 198 static void put_buf(r1bio_t *r1_bio)
 199 {
 200         conf_t *conf = r1_bio->mddev->private;
 201         int i;
 202
 203         for (i=0; i<conf->raid_disks; i++) {
 204                 struct bio *bio = r1_bio->bios[i];
 205                 if (bio->bi_end_io)
 206                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 207         }
 208
 209         mempool_free(r1_bio, conf->r1buf_pool);
 210
 211         lower_barrier(conf);
 212 }
 213
 214 static void reschedule_retry(r1bio_t *r1_bio)
 215 {
 216         unsigned long flags;
 217         mddev_t *mddev = r1_bio->mddev;
 218         conf_t *conf = mddev->private;
 219
 220         spin_lock_irqsave(&conf->device_lock, flags);
 221         list_add(&r1_bio->retry_list, &conf->retry_list);
 222         conf->nr_queued ++;
 223         spin_unlock_irqrestore(&conf->device_lock, flags);
 224
 225         wake_up(&conf->wait_barrier);
 226         md_wakeup_thread(mddev->thread);
 227 }
 228
 229 /*
 230  * raid_end_bio_io() is called when we have finished servicing a mirrored
 231  * operation and are ready to return a success/failure code to the buffer
 232  * cache layer.
 233  */
 234 static void raid_end_bio_io(r1bio_t *r1_bio)
 235 {
 236         struct bio *bio = r1_bio->master_bio;
 237
 238         /* if nobody has done the final endio yet, do it now */
 239         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 240                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 241                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 242                         (unsigned long long) bio->bi_sector,
 243                         (unsigned long long) bio->bi_sector +
 244                                 (bio->bi_size >> 9) - 1);
 245
 246                 bio_endio(bio,
 247                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 248         }
 249         free_r1bio(r1_bio);
 250 }
 251
 252 /*
 253  * Update disk head position estimator based on IRQ completion info.
 254  */
 255 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 256 {
 257         conf_t *conf = r1_bio->mddev->private;
 258
 259         conf->mirrors[disk].head_position =
 260                 r1_bio->sector + (r1_bio->sectors);
 261 }
 262
 263 static void raid1_end_read_request(struct bio *bio, int error)
 264 {
 265         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 266         r1bio_t *r1_bio = bio->bi_private;
 267         int mirror;
 268         conf_t *conf = r1_bio->mddev->private;
 269
 270         mirror = r1_bio->read_disk;
 271         /*
 272          * this branch is our 'one mirror IO has finished' event handler:
 273          */
 274         update_head_pos(mirror, r1_bio);
 275
 276         if (uptodate)
 277                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 278         else {
 279                 /* If all other devices have failed, we want to return
 280                  * the error upwards rather than fail the last device.
 281                  * Here we redefine "uptodate" to mean "Don't want to retry"
 282                  */
 283                 unsigned long flags;
 284                 spin_lock_irqsave(&conf->device_lock, flags);
 285                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 286                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 287                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 288                         uptodate = 1;
 289                 spin_unlock_irqrestore(&conf->device_lock, flags);
 290         }
 291
 292         if (uptodate)
 293                 raid_end_bio_io(r1_bio);
 294         else {
 295                 /*
 296                  * oops, read error:
 297                  */
 298                 char b[BDEVNAME_SIZE];
 299                 if (printk_ratelimit())
 300                         printk(KERN_ERR "md/raid1:%s: %s: rescheduling sector %llu\n",
 301                                mdname(conf->mddev),
 302                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 303                 reschedule_retry(r1_bio);
 304         }
 305
 306         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 307 }
 308
 309 static void raid1_end_write_request(struct bio *bio, int error)
 310 {
 311         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 312         r1bio_t *r1_bio = bio->bi_private;
 313         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 314         conf_t *conf = r1_bio->mddev->private;
 315         struct bio *to_put = NULL;
 316
 317
 318         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 319                 if (r1_bio->bios[mirror] == bio)
 320                         break;
 321
 322         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
 323                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
 324                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
 325                 r1_bio->mddev->barriers_work = 0;
 326                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
 327         } else {
 328                 /*
 329                  * this branch is our 'one mirror IO has finished' event handler:
 330                  */
 331                 r1_bio->bios[mirror] = NULL;
 332                 to_put = bio;
 333                 if (!uptodate) {
 334                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 335                         /* an I/O failed, we can't clear the bitmap */
 336                         set_bit(R1BIO_Degraded, &r1_bio->state);
 337                 } else
 338                         /*
 339                          * Set R1BIO_Uptodate in our master bio, so that
 340                          * we will return a good error code for to the higher
 341                          * levels even if IO on some other mirrored buffer fails.
 342                          *
 343                          * The 'master' represents the composite IO operation to
 344                          * user-side. So if something waits for IO, then it will
 345                          * wait for the 'master' bio.
 346                          */
 347                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 348
 349                 update_head_pos(mirror, r1_bio);
 350
 351                 if (behind) {
 352                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 353                                 atomic_dec(&r1_bio->behind_remaining);
 354
 355                         /* In behind mode, we ACK the master bio once the I/O has safely
 356                          * reached all non-writemostly disks. Setting the Returned bit
 357                          * ensures that this gets done only once -- we don't ever want to
 358                          * return -EIO here, instead we'll wait */
 359
 360                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 361                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 362                                 /* Maybe we can return now */
 363                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 364                                         struct bio *mbio = r1_bio->master_bio;
 365                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 366                                                (unsigned long long) mbio->bi_sector,
 367                                                (unsigned long long) mbio->bi_sector +
 368                                                (mbio->bi_size >> 9) - 1);
 369                                         bio_endio(mbio, 0);
 370                                 }
 371                         }
 372                 }
 373                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 374         }
 375         /*
 376          *
 377          * Let's see if all mirrored write operations have finished
 378          * already.
 379          */
 380         if (atomic_dec_and_test(&r1_bio->remaining)) {
 381                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
 382                         reschedule_retry(r1_bio);
 383                 else {
 384                         /* it really is the end of this request */
 385                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 386                                 /* free extra copy of the data pages */
 387                                 int i = bio->bi_vcnt;
 388                                 while (i--)
 389                                         safe_put_page(bio->bi_io_vec[i].bv_page);
 390                         }
 391                         /* clear the bitmap if all writes complete successfully */
 392                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 393                                         r1_bio->sectors,
 394                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 395                                         behind);
 396                         md_write_end(r1_bio->mddev);
 397                         raid_end_bio_io(r1_bio);
 398                 }
 399         }
 400
 401         if (to_put)
 402                 bio_put(to_put);
 403 }
 404
 405
 406 /*
 407  * This routine returns the disk from which the requested read should
 408  * be done. There is a per-array 'next expected sequential IO' sector
 409  * number - if this matches on the next IO then we use the last disk.
 410  * There is also a per-disk 'last know head position' sector that is
 411  * maintained from IRQ contexts, both the normal and the resync IO
 412  * completion handlers update this position correctly. If there is no
 413  * perfect sequential match then we pick the disk whose head is closest.
 414  *
 415  * If there are 2 mirrors in the same 2 devices, performance degrades
 416  * because position is mirror, not device based.
 417  *
 418  * The rdev for the device selected will have nr_pending incremented.
 419  */
 420 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 421 {
 422         const sector_t this_sector = r1_bio->sector;
 423         int new_disk = conf->last_used, disk = new_disk;
 424         int wonly_disk = -1;
 425         const int sectors = r1_bio->sectors;
 426         sector_t new_distance, current_distance;
 427         mdk_rdev_t *rdev;
 428
 429         rcu_read_lock();
 430         /*
 431          * Check if we can balance. We can balance on the whole
 432          * device if no resync is going on, or below the resync window.
 433          * We take the first readable disk when above the resync window.
 434          */
 435  retry:
 436         if (conf->mddev->recovery_cp < MaxSector &&
 437             (this_sector + sectors >= conf->next_resync)) {
 438                 /* Choose the first operational device, for consistancy */
 439                 new_disk = 0;
 440
 441                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 442                      r1_bio->bios[new_disk] == IO_BLOCKED ||
 443                      !rdev || !test_bit(In_sync, &rdev->flags)
 444                              || test_bit(WriteMostly, &rdev->flags);
 445                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
 446
 447                         if (rdev && test_bit(In_sync, &rdev->flags) &&
 448                                 r1_bio->bios[new_disk] != IO_BLOCKED)
 449                                 wonly_disk = new_disk;
 450
 451                         if (new_disk == conf->raid_disks - 1) {
 452                                 new_disk = wonly_disk;
 453                                 break;
 454                         }
 455                 }
 456                 goto rb_out;
 457         }
 458
 459
 460         /* make sure the disk is operational */
 461         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 462              r1_bio->bios[new_disk] == IO_BLOCKED ||
 463              !rdev || !test_bit(In_sync, &rdev->flags) ||
 464                      test_bit(WriteMostly, &rdev->flags);
 465              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
 466
 467                 if (rdev && test_bit(In_sync, &rdev->flags) &&
 468                     r1_bio->bios[new_disk] != IO_BLOCKED)
 469                         wonly_disk = new_disk;
 470
 471                 if (new_disk <= 0)
 472                         new_disk = conf->raid_disks;
 473                 new_disk--;
 474                 if (new_disk == disk) {
 475                         new_disk = wonly_disk;
 476                         break;
 477                 }
 478         }
 479
 480         if (new_disk < 0)
 481                 goto rb_out;
 482
 483         disk = new_disk;
 484         /* now disk == new_disk == starting point for search */
 485
 486         /*
 487          * Don't change to another disk for sequential reads:
 488          */
 489         if (conf->next_seq_sect == this_sector)
 490                 goto rb_out;
 491         if (this_sector == conf->mirrors[new_disk].head_position)
 492                 goto rb_out;
 493
 494         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 495
 496         /* Find the disk whose head is closest */
 497
 498         do {
 499                 if (disk <= 0)
 500                         disk = conf->raid_disks;
 501                 disk--;
 502
 503                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 504
 505                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
 506                     !test_bit(In_sync, &rdev->flags) ||
 507                     test_bit(WriteMostly, &rdev->flags))
 508                         continue;
 509
 510                 if (!atomic_read(&rdev->nr_pending)) {
 511                         new_disk = disk;
 512                         break;
 513                 }
 514                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 515                 if (new_distance < current_distance) {
 516                         current_distance = new_distance;
 517                         new_disk = disk;
 518                 }
 519         } while (disk != conf->last_used);
 520
 521  rb_out:
 522
 523
 524         if (new_disk >= 0) {
 525                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 526                 if (!rdev)
 527                         goto retry;
 528                 atomic_inc(&rdev->nr_pending);
 529                 if (!test_bit(In_sync, &rdev->flags)) {
 530                         /* cannot risk returning a device that failed
 531                          * before we inc'ed nr_pending
 532                          */
 533                         rdev_dec_pending(rdev, conf->mddev);
 534                         goto retry;
 535                 }
 536                 conf->next_seq_sect = this_sector + sectors;
 537                 conf->last_used = new_disk;
 538         }
 539         rcu_read_unlock();
 540
 541         return new_disk;
 542 }
 543
 544 static void unplug_slaves(mddev_t *mddev)
 545 {
 546         conf_t *conf = mddev->private;
 547         int i;
 548
 549         rcu_read_lock();
 550         for (i=0; i<mddev->raid_disks; i++) {
 551                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 552                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 553                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 554
 555                         atomic_inc(&rdev->nr_pending);
 556                         rcu_read_unlock();
 557
 558                         blk_unplug(r_queue);
 559
 560                         rdev_dec_pending(rdev, mddev);
 561                         rcu_read_lock();
 562                 }
 563         }
 564         rcu_read_unlock();
 565 }
 566
 567 static void raid1_unplug(struct request_queue *q)
 568 {
 569         mddev_t *mddev = q->queuedata;
 570
 571         unplug_slaves(mddev);
 572         md_wakeup_thread(mddev->thread);
 573 }
 574
 575 static int raid1_congested(void *data, int bits)
 576 {
 577         mddev_t *mddev = data;
 578         conf_t *conf = mddev->private;
 579         int i, ret = 0;
 580
 581         if (mddev_congested(mddev, bits))
 582                 return 1;
 583
 584         rcu_read_lock();
 585         for (i = 0; i < mddev->raid_disks; i++) {
 586                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 587                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 588                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 589
 590                         /* Note the '|| 1' - when read_balance prefers
 591                          * non-congested targets, it can be removed
 592                          */
 593                         if ((bits & (1<<BDI_async_congested)) || 1)
 594                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 595                         else
 596                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 597                 }
 598         }
 599         rcu_read_unlock();
 600         return ret;
 601 }
 602
 603
 604 static int flush_pending_writes(conf_t *conf)
 605 {
 606         /* Any writes that have been queued but are awaiting
 607          * bitmap updates get flushed here.
 608          * We return 1 if any requests were actually submitted.
 609          */
 610         int rv = 0;
 611
 612         spin_lock_irq(&conf->device_lock);
 613
 614         if (conf->pending_bio_list.head) {
 615                 struct bio *bio;
 616                 bio = bio_list_get(&conf->pending_bio_list);
 617                 blk_remove_plug(conf->mddev->queue);
 618                 spin_unlock_irq(&conf->device_lock);
 619                 /* flush any pending bitmap writes to
 620                  * disk before proceeding w/ I/O */
 621                 bitmap_unplug(conf->mddev->bitmap);
 622
 623                 while (bio) { /* submit pending writes */
 624                         struct bio *next = bio->bi_next;
 625                         bio->bi_next = NULL;
 626                         generic_make_request(bio);
 627                         bio = next;
 628                 }
 629                 rv = 1;
 630         } else
 631                 spin_unlock_irq(&conf->device_lock);
 632         return rv;
 633 }
 634
 635 /* Barriers....
 636  * Sometimes we need to suspend IO while we do something else,
 637  * either some resync/recovery, or reconfigure the array.
 638  * To do this we raise a 'barrier'.
 639  * The 'barrier' is a counter that can be raised multiple times
 640  * to count how many activities are happening which preclude
 641  * normal IO.
 642  * We can only raise the barrier if there is no pending IO.
 643  * i.e. if nr_pending == 0.
 644  * We choose only to raise the barrier if no-one is waiting for the
 645  * barrier to go down.  This means that as soon as an IO request
 646  * is ready, no other operations which require a barrier will start
 647  * until the IO request has had a chance.
 648  *
 649  * So: regular IO calls 'wait_barrier'.  When that returns there
 650  *    is no backgroup IO happening,  It must arrange to call
 651  *    allow_barrier when it has finished its IO.
 652  * backgroup IO calls must call raise_barrier.  Once that returns
 653  *    there is no normal IO happeing.  It must arrange to call
 654  *    lower_barrier when the particular background IO completes.
 655  */
 656 #define RESYNC_DEPTH 32
 657
 658 static void raise_barrier(conf_t *conf)
 659 {
 660         spin_lock_irq(&conf->resync_lock);
 661
 662         /* Wait until no block IO is waiting */
 663         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 664                             conf->resync_lock,
 665                             raid1_unplug(conf->mddev->queue));
 666
 667         /* block any new IO from starting */
 668         conf->barrier++;
 669
 670         /* No wait for all pending IO to complete */
 671         wait_event_lock_irq(conf->wait_barrier,
 672                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 673                             conf->resync_lock,
 674                             raid1_unplug(conf->mddev->queue));
 675
 676         spin_unlock_irq(&conf->resync_lock);
 677 }
 678
 679 static void lower_barrier(conf_t *conf)
 680 {
 681         unsigned long flags;
 682         BUG_ON(conf->barrier <= 0);
 683         spin_lock_irqsave(&conf->resync_lock, flags);
 684         conf->barrier--;
 685         spin_unlock_irqrestore(&conf->resync_lock, flags);
 686         wake_up(&conf->wait_barrier);
 687 }
 688
 689 static void wait_barrier(conf_t *conf)
 690 {
 691         spin_lock_irq(&conf->resync_lock);
 692         if (conf->barrier) {
 693                 conf->nr_waiting++;
 694                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 695                                     conf->resync_lock,
 696                                     raid1_unplug(conf->mddev->queue));
 697                 conf->nr_waiting--;
 698         }
 699         conf->nr_pending++;
 700         spin_unlock_irq(&conf->resync_lock);
 701 }
 702
 703 static void allow_barrier(conf_t *conf)
 704 {
 705         unsigned long flags;
 706         spin_lock_irqsave(&conf->resync_lock, flags);
 707         conf->nr_pending--;
 708         spin_unlock_irqrestore(&conf->resync_lock, flags);
 709         wake_up(&conf->wait_barrier);
 710 }
 711
 712 static void freeze_array(conf_t *conf)
 713 {
 714         /* stop syncio and normal IO and wait for everything to
 715          * go quite.
 716          * We increment barrier and nr_waiting, and then
 717          * wait until nr_pending match nr_queued+1
 718          * This is called in the context of one normal IO request
 719          * that has failed. Thus any sync request that might be pending
 720          * will be blocked by nr_pending, and we need to wait for
 721          * pending IO requests to complete or be queued for re-try.
 722          * Thus the number queued (nr_queued) plus this request (1)
 723          * must match the number of pending IOs (nr_pending) before
 724          * we continue.
 725          */
 726         spin_lock_irq(&conf->resync_lock);
 727         conf->barrier++;
 728         conf->nr_waiting++;
 729         wait_event_lock_irq(conf->wait_barrier,
 730                             conf->nr_pending == conf->nr_queued+1,
 731                             conf->resync_lock,
 732                             ({ flush_pending_writes(conf);
 733                                raid1_unplug(conf->mddev->queue); }));
 734         spin_unlock_irq(&conf->resync_lock);
 735 }
 736 static void unfreeze_array(conf_t *conf)
 737 {
 738         /* reverse the effect of the freeze */
 739         spin_lock_irq(&conf->resync_lock);
 740         conf->barrier--;
 741         conf->nr_waiting--;
 742         wake_up(&conf->wait_barrier);
 743         spin_unlock_irq(&conf->resync_lock);
 744 }
 745
 746
 747 /* duplicate the data pages for behind I/O */
 748 static struct page **alloc_behind_pages(struct bio *bio)
 749 {
 750         int i;
 751         struct bio_vec *bvec;
 752         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 753                                         GFP_NOIO);
 754         if (unlikely(!pages))
 755                 goto do_sync_io;
 756
 757         bio_for_each_segment(bvec, bio, i) {
 758                 pages[i] = alloc_page(GFP_NOIO);
 759                 if (unlikely(!pages[i]))
 760                         goto do_sync_io;
 761                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 762                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 763                 kunmap(pages[i]);
 764                 kunmap(bvec->bv_page);
 765         }
 766
 767         return pages;
 768
 769 do_sync_io:
 770         if (pages)
 771                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 772                         put_page(pages[i]);
 773         kfree(pages);
 774         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 775         return NULL;
 776 }
 777
 778 static int make_request(mddev_t *mddev, struct bio * bio)
 779 {
 780         conf_t *conf = mddev->private;
 781         mirror_info_t *mirror;
 782         r1bio_t *r1_bio;
 783         struct bio *read_bio;
 784         int i, targets = 0, disks;
 785         struct bitmap *bitmap;
 786         unsigned long flags;
 787         struct bio_list bl;
 788         struct page **behind_pages = NULL;
 789         const int rw = bio_data_dir(bio);
 790         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 791         unsigned long do_barriers;
 792         mdk_rdev_t *blocked_rdev;
 793
 794         /*
 795          * Register the new request and wait if the reconstruction
 796          * thread has put up a bar for new requests.
 797          * Continue immediately if no resync is active currently.
 798          * We test barriers_work *after* md_write_start as md_write_start
 799          * may cause the first superblock write, and that will check out
 800          * if barriers work.
 801          */
 802
 803         md_write_start(mddev, bio); /* wait on superblock update early */
 804
 805         if (bio_data_dir(bio) == WRITE &&
 806             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 807             bio->bi_sector < mddev->suspend_hi) {
 808                 /* As the suspend_* range is controlled by
 809                  * userspace, we want an interruptible
 810                  * wait.
 811                  */
 812                 DEFINE_WAIT(w);
 813                 for (;;) {
 814                         flush_signals(current);
 815                         prepare_to_wait(&conf->wait_barrier,
 816                                         &w, TASK_INTERRUPTIBLE);
 817                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 818                             bio->bi_sector >= mddev->suspend_hi)
 819                                 break;
 820                         schedule();
 821                 }
 822                 finish_wait(&conf->wait_barrier, &w);
 823         }
 824         if (unlikely(!mddev->barriers_work &&
 825                      (bio->bi_rw & REQ_HARDBARRIER))) {
 826                 if (rw == WRITE)
 827                         md_write_end(mddev);
 828                 bio_endio(bio, -EOPNOTSUPP);
 829                 return 0;
 830         }
 831
 832         wait_barrier(conf);
 833
 834         bitmap = mddev->bitmap;
 835
 836         /*
 837          * make_request() can abort the operation when READA is being
 838          * used and no empty request is available.
 839          *
 840          */
 841         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 842
 843         r1_bio->master_bio = bio;
 844         r1_bio->sectors = bio->bi_size >> 9;
 845         r1_bio->state = 0;
 846         r1_bio->mddev = mddev;
 847         r1_bio->sector = bio->bi_sector;
 848
 849         if (rw == READ) {
 850                 /*
 851                  * read balancing logic:
 852                  */
 853                 int rdisk = read_balance(conf, r1_bio);
 854
 855                 if (rdisk < 0) {
 856                         /* couldn't find anywhere to read from */
 857                         raid_end_bio_io(r1_bio);
 858                         return 0;
 859                 }
 860                 mirror = conf->mirrors + rdisk;
 861
 862                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 863                     bitmap) {
 864                         /* Reading from a write-mostly device must
 865                          * take care not to over-take any writes
 866                          * that are 'behind'
 867                          */
 868                         wait_event(bitmap->behind_wait,
 869                                    atomic_read(&bitmap->behind_writes) == 0);
 870                 }
 871                 r1_bio->read_disk = rdisk;
 872
 873                 read_bio = bio_clone(bio, GFP_NOIO);
 874
 875                 r1_bio->bios[rdisk] = read_bio;
 876
 877                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 878                 read_bio->bi_bdev = mirror->rdev->bdev;
 879                 read_bio->bi_end_io = raid1_end_read_request;
 880                 read_bio->bi_rw = READ | do_sync;
 881                 read_bio->bi_private = r1_bio;
 882
 883                 generic_make_request(read_bio);
 884                 return 0;
 885         }
 886
 887         /*
 888          * WRITE:
 889          */
 890         /* first select target devices under spinlock and
 891          * inc refcount on their rdev.  Record them by setting
 892          * bios[x] to bio
 893          */
 894         disks = conf->raid_disks;
 895 #if 0
 896         { static int first=1;
 897         if (first) printk("First Write sector %llu disks %d\n",
 898                           (unsigned long long)r1_bio->sector, disks);
 899         first = 0;
 900         }
 901 #endif
 902  retry_write:
 903         blocked_rdev = NULL;
 904         rcu_read_lock();
 905         for (i = 0;  i < disks; i++) {
 906                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 907                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 908                         atomic_inc(&rdev->nr_pending);
 909                         blocked_rdev = rdev;
 910                         break;
 911                 }
 912                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 913                         atomic_inc(&rdev->nr_pending);
 914                         if (test_bit(Faulty, &rdev->flags)) {
 915                                 rdev_dec_pending(rdev, mddev);
 916                                 r1_bio->bios[i] = NULL;
 917                         } else {
 918                                 r1_bio->bios[i] = bio;
 919                                 targets++;
 920                         }
 921                 } else
 922                         r1_bio->bios[i] = NULL;
 923         }
 924         rcu_read_unlock();
 925
 926         if (unlikely(blocked_rdev)) {
 927                 /* Wait for this device to become unblocked */
 928                 int j;
 929
 930                 for (j = 0; j < i; j++)
 931                         if (r1_bio->bios[j])
 932                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 933
 934                 allow_barrier(conf);
 935                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 936                 wait_barrier(conf);
 937                 goto retry_write;
 938         }
 939
 940         BUG_ON(targets == 0); /* we never fail the last device */
 941
 942         if (targets < conf->raid_disks) {
 943                 /* array is degraded, we will not clear the bitmap
 944                  * on I/O completion (see raid1_end_write_request) */
 945                 set_bit(R1BIO_Degraded, &r1_bio->state);
 946         }
 947
 948         /* do behind I/O ?
 949          * Not if there are too many, or cannot allocate memory,
 950          * or a reader on WriteMostly is waiting for behind writes
 951          * to flush */
 952         if (bitmap &&
 953             (atomic_read(&bitmap->behind_writes)
 954              < mddev->bitmap_info.max_write_behind) &&
 955             !waitqueue_active(&bitmap->behind_wait) &&
 956             (behind_pages = alloc_behind_pages(bio)) != NULL)
 957                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 958
 959         atomic_set(&r1_bio->remaining, 0);
 960         atomic_set(&r1_bio->behind_remaining, 0);
 961
 962         do_barriers = bio->bi_rw & REQ_HARDBARRIER;
 963         if (do_barriers)
 964                 set_bit(R1BIO_Barrier, &r1_bio->state);
 965
 966         bio_list_init(&bl);
 967         for (i = 0; i < disks; i++) {
 968                 struct bio *mbio;
 969                 if (!r1_bio->bios[i])
 970                         continue;
 971
 972                 mbio = bio_clone(bio, GFP_NOIO);
 973                 r1_bio->bios[i] = mbio;
 974
 975                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 976                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 977                 mbio->bi_end_io = raid1_end_write_request;
 978                 mbio->bi_rw = WRITE | do_barriers | do_sync;
 979                 mbio->bi_private = r1_bio;
 980
 981                 if (behind_pages) {
 982                         struct bio_vec *bvec;
 983                         int j;
 984
 985                         /* Yes, I really want the '__' version so that
 986                          * we clear any unused pointer in the io_vec, rather
 987                          * than leave them unchanged.  This is important
 988                          * because when we come to free the pages, we won't
 989                          * know the originial bi_idx, so we just free
 990                          * them all
 991                          */
 992                         __bio_for_each_segment(bvec, mbio, j, 0)
 993                                 bvec->bv_page = behind_pages[j];
 994                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 995                                 atomic_inc(&r1_bio->behind_remaining);
 996                 }
 997
 998                 atomic_inc(&r1_bio->remaining);
 999
1000                 bio_list_add(&bl, mbio);
1001         }
1002         kfree(behind_pages); /* the behind pages are attached to the bios now */
1003
1004         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
1005                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
1006         spin_lock_irqsave(&conf->device_lock, flags);
1007         bio_list_merge(&conf->pending_bio_list, &bl);
1008         bio_list_init(&bl);
1009
1010         blk_plug_device(mddev->queue);
1011         spin_unlock_irqrestore(&conf->device_lock, flags);
1012
1013         /* In case raid1d snuck into freeze_array */
1014         wake_up(&conf->wait_barrier);
1015
1016         if (do_sync)
1017                 md_wakeup_thread(mddev->thread);
1018 #if 0
1019         while ((bio = bio_list_pop(&bl)) != NULL)
1020                 generic_make_request(bio);
1021 #endif
1022
1023         return 0;
1024 }
1025
1026 static void status(struct seq_file *seq, mddev_t *mddev)
1027 {
1028         conf_t *conf = mddev->private;
1029         int i;
1030
1031         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1032                    conf->raid_disks - mddev->degraded);
1033         rcu_read_lock();
1034         for (i = 0; i < conf->raid_disks; i++) {
1035                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1036                 seq_printf(seq, "%s",
1037                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1038         }
1039         rcu_read_unlock();
1040         seq_printf(seq, "]");
1041 }
1042
1043
1044 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1045 {
1046         char b[BDEVNAME_SIZE];
1047         conf_t *conf = mddev->private;
1048
1049         /*
1050          * If it is not operational, then we have already marked it as dead
1051          * else if it is the last working disks, ignore the error, let the
1052          * next level up know.
1053          * else mark the drive as failed
1054          */
1055         if (test_bit(In_sync, &rdev->flags)
1056             && (conf->raid_disks - mddev->degraded) == 1) {
1057                 /*
1058                  * Don't fail the drive, act as though we were just a
1059                  * normal single drive.
1060                  * However don't try a recovery from this drive as
1061                  * it is very likely to fail.
1062                  */
1063                 mddev->recovery_disabled = 1;
1064                 return;
1065         }
1066         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1067                 unsigned long flags;
1068                 spin_lock_irqsave(&conf->device_lock, flags);
1069                 mddev->degraded++;
1070                 set_bit(Faulty, &rdev->flags);
1071                 spin_unlock_irqrestore(&conf->device_lock, flags);
1072                 /*
1073                  * if recovery is running, make sure it aborts.
1074                  */
1075                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1076         } else
1077                 set_bit(Faulty, &rdev->flags);
1078         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1079         printk(KERN_ALERT "md/raid1:%s: Disk failure on %s, disabling device.\n"
1080                KERN_ALERT "md/raid1:%s: Operation continuing on %d devices.\n",
1081                mdname(mddev), bdevname(rdev->bdev, b),
1082                mdname(mddev), conf->raid_disks - mddev->degraded);
1083 }
1084
1085 static void print_conf(conf_t *conf)
1086 {
1087         int i;
1088
1089         printk(KERN_DEBUG "RAID1 conf printout:\n");
1090         if (!conf) {
1091                 printk(KERN_DEBUG "(!conf)\n");
1092                 return;
1093         }
1094         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1095                 conf->raid_disks);
1096
1097         rcu_read_lock();
1098         for (i = 0; i < conf->raid_disks; i++) {
1099                 char b[BDEVNAME_SIZE];
1100                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1101                 if (rdev)
1102                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1103                                i, !test_bit(In_sync, &rdev->flags),
1104                                !test_bit(Faulty, &rdev->flags),
1105                                bdevname(rdev->bdev,b));
1106         }
1107         rcu_read_unlock();
1108 }
1109
1110 static void close_sync(conf_t *conf)
1111 {
1112         wait_barrier(conf);
1113         allow_barrier(conf);
1114
1115         mempool_destroy(conf->r1buf_pool);
1116         conf->r1buf_pool = NULL;
1117 }
1118
1119 static int raid1_spare_active(mddev_t *mddev)
1120 {
1121         int i;
1122         conf_t *conf = mddev->private;
1123         int count = 0;
1124         unsigned long flags;
1125
1126         /*
1127          * Find all failed disks within the RAID1 configuration
1128          * and mark them readable.
1129          * Called under mddev lock, so rcu protection not needed.
1130          */
1131         for (i = 0; i < conf->raid_disks; i++) {
1132                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1133                 if (rdev
1134                     && !test_bit(Faulty, &rdev->flags)
1135                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1136                         count++;
1137                         sysfs_notify_dirent(rdev->sysfs_state);
1138                 }
1139         }
1140         spin_lock_irqsave(&conf->device_lock, flags);
1141         mddev->degraded -= count;
1142         spin_unlock_irqrestore(&conf->device_lock, flags);
1143
1144         print_conf(conf);
1145         return count;
1146 }
1147
1148
1149 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1150 {
1151         conf_t *conf = mddev->private;
1152         int err = -EEXIST;
1153         int mirror = 0;
1154         mirror_info_t *p;
1155         int first = 0;
1156         int last = mddev->raid_disks - 1;
1157
1158         if (rdev->raid_disk >= 0)
1159                 first = last = rdev->raid_disk;
1160
1161         for (mirror = first; mirror <= last; mirror++)
1162                 if ( !(p=conf->mirrors+mirror)->rdev) {
1163
1164                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1165                                           rdev->data_offset << 9);
1166                         /* as we don't honour merge_bvec_fn, we must
1167                          * never risk violating it, so limit
1168                          * ->max_segments to one lying with a single
1169                          * page, as a one page request is never in
1170                          * violation.
1171                          */
1172                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1173                                 blk_queue_max_segments(mddev->queue, 1);
1174                                 blk_queue_segment_boundary(mddev->queue,
1175                                                            PAGE_CACHE_SIZE - 1);
1176                         }
1177
1178                         p->head_position = 0;
1179                         rdev->raid_disk = mirror;
1180                         err = 0;
1181                         /* As all devices are equivalent, we don't need a full recovery
1182                          * if this was recently any drive of the array
1183                          */
1184                         if (rdev->saved_raid_disk < 0)
1185                                 conf->fullsync = 1;
1186                         rcu_assign_pointer(p->rdev, rdev);
1187                         break;
1188                 }
1189         md_integrity_add_rdev(rdev, mddev);
1190         print_conf(conf);
1191         return err;
1192 }
1193
1194 static int raid1_remove_disk(mddev_t *mddev, int number)
1195 {
1196         conf_t *conf = mddev->private;
1197         int err = 0;
1198         mdk_rdev_t *rdev;
1199         mirror_info_t *p = conf->mirrors+ number;
1200
1201         print_conf(conf);
1202         rdev = p->rdev;
1203         if (rdev) {
1204                 if (test_bit(In_sync, &rdev->flags) ||
1205                     atomic_read(&rdev->nr_pending)) {
1206                         err = -EBUSY;
1207                         goto abort;
1208                 }
1209                 /* Only remove non-faulty devices is recovery
1210                  * is not possible.
1211                  */
1212                 if (!test_bit(Faulty, &rdev->flags) &&
1213                     mddev->degraded < conf->raid_disks) {
1214                         err = -EBUSY;
1215                         goto abort;
1216                 }
1217                 p->rdev = NULL;
1218                 synchronize_rcu();
1219                 if (atomic_read(&rdev->nr_pending)) {
1220                         /* lost the race, try later */
1221                         err = -EBUSY;
1222                         p->rdev = rdev;
1223                         goto abort;
1224                 }
1225                 md_integrity_register(mddev);
1226         }
1227 abort:
1228
1229         print_conf(conf);
1230         return err;
1231 }
1232
1233
1234 static void end_sync_read(struct bio *bio, int error)
1235 {
1236         r1bio_t *r1_bio = bio->bi_private;
1237         int i;
1238
1239         for (i=r1_bio->mddev->raid_disks; i--; )
1240                 if (r1_bio->bios[i] == bio)
1241                         break;
1242         BUG_ON(i < 0);
1243         update_head_pos(i, r1_bio);
1244         /*
1245          * we have read a block, now it needs to be re-written,
1246          * or re-read if the read failed.
1247          * We don't do much here, just schedule handling by raid1d
1248          */
1249         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1250                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1251
1252         if (atomic_dec_and_test(&r1_bio->remaining))
1253                 reschedule_retry(r1_bio);
1254 }
1255
1256 static void end_sync_write(struct bio *bio, int error)
1257 {
1258         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1259         r1bio_t *r1_bio = bio->bi_private;
1260         mddev_t *mddev = r1_bio->mddev;
1261         conf_t *conf = mddev->private;
1262         int i;
1263         int mirror=0;
1264
1265         for (i = 0; i < conf->raid_disks; i++)
1266                 if (r1_bio->bios[i] == bio) {
1267                         mirror = i;
1268                         break;
1269                 }
1270         if (!uptodate) {
1271                 int sync_blocks = 0;
1272                 sector_t s = r1_bio->sector;
1273                 long sectors_to_go = r1_bio->sectors;
1274                 /* make sure these bits doesn't get cleared. */
1275                 do {
1276                         bitmap_end_sync(mddev->bitmap, s,
1277                                         &sync_blocks, 1);
1278                         s += sync_blocks;
1279                         sectors_to_go -= sync_blocks;
1280                 } while (sectors_to_go > 0);
1281                 md_error(mddev, conf->mirrors[mirror].rdev);
1282         }
1283
1284         update_head_pos(mirror, r1_bio);
1285
1286         if (atomic_dec_and_test(&r1_bio->remaining)) {
1287                 sector_t s = r1_bio->sectors;
1288                 put_buf(r1_bio);
1289                 md_done_sync(mddev, s, uptodate);
1290         }
1291 }
1292
1293 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1294 {
1295         conf_t *conf = mddev->private;
1296         int i;
1297         int disks = conf->raid_disks;
1298         struct bio *bio, *wbio;
1299
1300         bio = r1_bio->bios[r1_bio->read_disk];
1301
1302
1303         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1304                 /* We have read all readable devices.  If we haven't
1305                  * got the block, then there is no hope left.
1306                  * If we have, then we want to do a comparison
1307                  * and skip the write if everything is the same.
1308                  * If any blocks failed to read, then we need to
1309                  * attempt an over-write
1310                  */
1311                 int primary;
1312                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1313                         for (i=0; i<mddev->raid_disks; i++)
1314                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1315                                         md_error(mddev, conf->mirrors[i].rdev);
1316
1317                         md_done_sync(mddev, r1_bio->sectors, 1);
1318                         put_buf(r1_bio);
1319                         return;
1320                 }
1321                 for (primary=0; primary<mddev->raid_disks; primary++)
1322                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1323                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1324                                 r1_bio->bios[primary]->bi_end_io = NULL;
1325                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1326                                 break;
1327                         }
1328                 r1_bio->read_disk = primary;
1329                 for (i=0; i<mddev->raid_disks; i++)
1330                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1331                                 int j;
1332                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1333                                 struct bio *pbio = r1_bio->bios[primary];
1334                                 struct bio *sbio = r1_bio->bios[i];
1335
1336                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1337                                         for (j = vcnt; j-- ; ) {
1338                                                 struct page *p, *s;
1339                                                 p = pbio->bi_io_vec[j].bv_page;
1340                                                 s = sbio->bi_io_vec[j].bv_page;
1341                                                 if (memcmp(page_address(p),
1342                                                            page_address(s),
1343                                                            PAGE_SIZE))
1344                                                         break;
1345                                         }
1346                                 } else
1347                                         j = 0;
1348                                 if (j >= 0)
1349                                         mddev->resync_mismatches += r1_bio->sectors;
1350                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1351                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1352                                         sbio->bi_end_io = NULL;
1353                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1354                                 } else {
1355                                         /* fixup the bio for reuse */
1356                                         int size;
1357                                         sbio->bi_vcnt = vcnt;
1358                                         sbio->bi_size = r1_bio->sectors << 9;
1359                                         sbio->bi_idx = 0;
1360                                         sbio->bi_phys_segments = 0;
1361                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1362                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1363                                         sbio->bi_next = NULL;
1364                                         sbio->bi_sector = r1_bio->sector +
1365                                                 conf->mirrors[i].rdev->data_offset;
1366                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1367                                         size = sbio->bi_size;
1368                                         for (j = 0; j < vcnt ; j++) {
1369                                                 struct bio_vec *bi;
1370                                                 bi = &sbio->bi_io_vec[j];
1371                                                 bi->bv_offset = 0;
1372                                                 if (size > PAGE_SIZE)
1373                                                         bi->bv_len = PAGE_SIZE;
1374                                                 else
1375                                                         bi->bv_len = size;
1376                                                 size -= PAGE_SIZE;
1377                                                 memcpy(page_address(bi->bv_page),
1378                                                        page_address(pbio->bi_io_vec[j].bv_page),
1379                                                        PAGE_SIZE);
1380                                         }
1381
1382                                 }
1383                         }
1384         }
1385         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1386                 /* ouch - failed to read all of that.
1387                  * Try some synchronous reads of other devices to get
1388                  * good data, much like with normal read errors.  Only
1389                  * read into the pages we already have so we don't
1390                  * need to re-issue the read request.
1391                  * We don't need to freeze the array, because being in an
1392                  * active sync request, there is no normal IO, and
1393                  * no overlapping syncs.
1394                  */
1395                 sector_t sect = r1_bio->sector;
1396                 int sectors = r1_bio->sectors;
1397                 int idx = 0;
1398
1399                 while(sectors) {
1400                         int s = sectors;
1401                         int d = r1_bio->read_disk;
1402                         int success = 0;
1403                         mdk_rdev_t *rdev;
1404
1405                         if (s > (PAGE_SIZE>>9))
1406                                 s = PAGE_SIZE >> 9;
1407                         do {
1408                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1409                                         /* No rcu protection needed here devices
1410                                          * can only be removed when no resync is
1411                                          * active, and resync is currently active
1412                                          */
1413                                         rdev = conf->mirrors[d].rdev;
1414                                         if (sync_page_io(rdev->bdev,
1415                                                          sect + rdev->data_offset,
1416                                                          s<<9,
1417                                                          bio->bi_io_vec[idx].bv_page,
1418                                                          READ)) {
1419                                                 success = 1;
1420                                                 break;
1421                                         }
1422                                 }
1423                                 d++;
1424                                 if (d == conf->raid_disks)
1425                                         d = 0;
1426                         } while (!success && d != r1_bio->read_disk);
1427
1428                         if (success) {
1429                                 int start = d;
1430                                 /* write it back and re-read */
1431                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1432                                 while (d != r1_bio->read_disk) {
1433                                         if (d == 0)
1434                                                 d = conf->raid_disks;
1435                                         d--;
1436                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1437                                                 continue;
1438                                         rdev = conf->mirrors[d].rdev;
1439                                         atomic_add(s, &rdev->corrected_errors);
1440                                         if (sync_page_io(rdev->bdev,
1441                                                          sect + rdev->data_offset,
1442                                                          s<<9,
1443                                                          bio->bi_io_vec[idx].bv_page,
1444                                                          WRITE) == 0)
1445                                                 md_error(mddev, rdev);
1446                                 }
1447                                 d = start;
1448                                 while (d != r1_bio->read_disk) {
1449                                         if (d == 0)
1450                                                 d = conf->raid_disks;
1451                                         d--;
1452                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1453                                                 continue;
1454                                         rdev = conf->mirrors[d].rdev;
1455                                         if (sync_page_io(rdev->bdev,
1456                                                          sect + rdev->data_offset,
1457                                                          s<<9,
1458                                                          bio->bi_io_vec[idx].bv_page,
1459                                                          READ) == 0)
1460                                                 md_error(mddev, rdev);
1461                                 }
1462                         } else {
1463                                 char b[BDEVNAME_SIZE];
1464                                 /* Cannot read from anywhere, array is toast */
1465                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1466                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1467                                        " for block %llu\n",
1468                                        mdname(mddev),
1469                                        bdevname(bio->bi_bdev, b),
1470                                        (unsigned long long)r1_bio->sector);
1471                                 md_done_sync(mddev, r1_bio->sectors, 0);
1472                                 put_buf(r1_bio);
1473                                 return;
1474                         }
1475                         sectors -= s;
1476                         sect += s;
1477                         idx ++;
1478                 }
1479         }
1480
1481         /*
1482          * schedule writes
1483          */
1484         atomic_set(&r1_bio->remaining, 1);
1485         for (i = 0; i < disks ; i++) {
1486                 wbio = r1_bio->bios[i];
1487                 if (wbio->bi_end_io == NULL ||
1488                     (wbio->bi_end_io == end_sync_read &&
1489                      (i == r1_bio->read_disk ||
1490                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1491                         continue;
1492
1493                 wbio->bi_rw = WRITE;
1494                 wbio->bi_end_io = end_sync_write;
1495                 atomic_inc(&r1_bio->remaining);
1496                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1497
1498                 generic_make_request(wbio);
1499         }
1500
1501         if (atomic_dec_and_test(&r1_bio->remaining)) {
1502                 /* if we're here, all write(s) have completed, so clean up */
1503                 md_done_sync(mddev, r1_bio->sectors, 1);
1504                 put_buf(r1_bio);
1505         }
1506 }
1507
1508 /*
1509  * This is a kernel thread which:
1510  *
1511  *      1.      Retries failed read operations on working mirrors.
1512  *      2.      Updates the raid superblock when problems encounter.
1513  *      3.      Performs writes following reads for array syncronising.
1514  */
1515
1516 static void fix_read_error(conf_t *conf, int read_disk,
1517                            sector_t sect, int sectors)
1518 {
1519         mddev_t *mddev = conf->mddev;
1520         while(sectors) {
1521                 int s = sectors;
1522                 int d = read_disk;
1523                 int success = 0;
1524                 int start;
1525                 mdk_rdev_t *rdev;
1526
1527                 if (s > (PAGE_SIZE>>9))
1528                         s = PAGE_SIZE >> 9;
1529
1530                 do {
1531                         /* Note: no rcu protection needed here
1532                          * as this is synchronous in the raid1d thread
1533                          * which is the thread that might remove
1534                          * a device.  If raid1d ever becomes multi-threaded....
1535                          */
1536                         rdev = conf->mirrors[d].rdev;
1537                         if (rdev &&
1538                             test_bit(In_sync, &rdev->flags) &&
1539                             sync_page_io(rdev->bdev,
1540                                          sect + rdev->data_offset,
1541                                          s<<9,
1542                                          conf->tmppage, READ))
1543                                 success = 1;
1544                         else {
1545                                 d++;
1546                                 if (d == conf->raid_disks)
1547                                         d = 0;
1548                         }
1549                 } while (!success && d != read_disk);
1550
1551                 if (!success) {
1552                         /* Cannot read from anywhere -- bye bye array */
1553                         md_error(mddev, conf->mirrors[read_disk].rdev);
1554                         break;
1555                 }
1556                 /* write it back and re-read */
1557                 start = d;
1558                 while (d != read_disk) {
1559                         if (d==0)
1560                                 d = conf->raid_disks;
1561                         d--;
1562                         rdev = conf->mirrors[d].rdev;
1563                         if (rdev &&
1564                             test_bit(In_sync, &rdev->flags)) {
1565                                 if (sync_page_io(rdev->bdev,
1566                                                  sect + rdev->data_offset,
1567                                                  s<<9, conf->tmppage, WRITE)
1568                                     == 0)
1569                                         /* Well, this device is dead */
1570                                         md_error(mddev, rdev);
1571                         }
1572                 }
1573                 d = start;
1574                 while (d != read_disk) {
1575                         char b[BDEVNAME_SIZE];
1576                         if (d==0)
1577                                 d = conf->raid_disks;
1578                         d--;
1579                         rdev = conf->mirrors[d].rdev;
1580                         if (rdev &&
1581                             test_bit(In_sync, &rdev->flags)) {
1582                                 if (sync_page_io(rdev->bdev,
1583                                                  sect + rdev->data_offset,
1584                                                  s<<9, conf->tmppage, READ)
1585                                     == 0)
1586                                         /* Well, this device is dead */
1587                                         md_error(mddev, rdev);
1588                                 else {
1589                                         atomic_add(s, &rdev->corrected_errors);
1590                                         printk(KERN_INFO
1591                                                "md/raid1:%s: read error corrected "
1592                                                "(%d sectors at %llu on %s)\n",
1593                                                mdname(mddev), s,
1594                                                (unsigned long long)(sect +
1595                                                    rdev->data_offset),
1596                                                bdevname(rdev->bdev, b));
1597                                 }
1598                         }
1599                 }
1600                 sectors -= s;
1601                 sect += s;
1602         }
1603 }
1604
1605 static void raid1d(mddev_t *mddev)
1606 {
1607         r1bio_t *r1_bio;
1608         struct bio *bio;
1609         unsigned long flags;
1610         conf_t *conf = mddev->private;
1611         struct list_head *head = &conf->retry_list;
1612         int unplug=0;
1613         mdk_rdev_t *rdev;
1614
1615         md_check_recovery(mddev);
1616
1617         for (;;) {
1618                 char b[BDEVNAME_SIZE];
1619
1620                 unplug += flush_pending_writes(conf);
1621
1622                 spin_lock_irqsave(&conf->device_lock, flags);
1623                 if (list_empty(head)) {
1624                         spin_unlock_irqrestore(&conf->device_lock, flags);
1625                         break;
1626                 }
1627                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1628                 list_del(head->prev);
1629                 conf->nr_queued--;
1630                 spin_unlock_irqrestore(&conf->device_lock, flags);
1631
1632                 mddev = r1_bio->mddev;
1633                 conf = mddev->private;
1634                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1635                         sync_request_write(mddev, r1_bio);
1636                         unplug = 1;
1637                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1638                         /* some requests in the r1bio were REQ_HARDBARRIER
1639                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1640                          * Better resubmit without the barrier.
1641                          * We know which devices to resubmit for, because
1642                          * all others have had their bios[] entry cleared.
1643                          * We already have a nr_pending reference on these rdevs.
1644                          */
1645                         int i;
1646                         const unsigned long do_sync = (r1_bio->master_bio->bi_rw & REQ_SYNC);
1647                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1648                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1649                         for (i=0; i < conf->raid_disks; i++)
1650                                 if (r1_bio->bios[i])
1651                                         atomic_inc(&r1_bio->remaining);
1652                         for (i=0; i < conf->raid_disks; i++)
1653                                 if (r1_bio->bios[i]) {
1654                                         struct bio_vec *bvec;
1655                                         int j;
1656
1657                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1658                                         /* copy pages from the failed bio, as
1659                                          * this might be a write-behind device */
1660                                         __bio_for_each_segment(bvec, bio, j, 0)
1661                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1662                                         bio_put(r1_bio->bios[i]);
1663                                         bio->bi_sector = r1_bio->sector +
1664                                                 conf->mirrors[i].rdev->data_offset;
1665                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1666                                         bio->bi_end_io = raid1_end_write_request;
1667                                         bio->bi_rw = WRITE | do_sync;
1668                                         bio->bi_private = r1_bio;
1669                                         r1_bio->bios[i] = bio;
1670                                         generic_make_request(bio);
1671                                 }
1672                 } else {
1673                         int disk;
1674
1675                         /* we got a read error. Maybe the drive is bad.  Maybe just
1676                          * the block and we can fix it.
1677                          * We freeze all other IO, and try reading the block from
1678                          * other devices.  When we find one, we re-write
1679                          * and check it that fixes the read error.
1680                          * This is all done synchronously while the array is
1681                          * frozen
1682                          */
1683                         if (mddev->ro == 0) {
1684                                 freeze_array(conf);
1685                                 fix_read_error(conf, r1_bio->read_disk,
1686                                                r1_bio->sector,
1687                                                r1_bio->sectors);
1688                                 unfreeze_array(conf);
1689                         } else
1690                                 md_error(mddev,
1691                                          conf->mirrors[r1_bio->read_disk].rdev);
1692
1693                         bio = r1_bio->bios[r1_bio->read_disk];
1694                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1695                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1696                                        " read error for block %llu\n",
1697                                        mdname(mddev),
1698                                        bdevname(bio->bi_bdev,b),
1699                                        (unsigned long long)r1_bio->sector);
1700                                 raid_end_bio_io(r1_bio);
1701                         } else {
1702                                 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1703                                 r1_bio->bios[r1_bio->read_disk] =
1704                                         mddev->ro ? IO_BLOCKED : NULL;
1705                                 r1_bio->read_disk = disk;
1706                                 bio_put(bio);
1707                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1708                                 r1_bio->bios[r1_bio->read_disk] = bio;
1709                                 rdev = conf->mirrors[disk].rdev;
1710                                 if (printk_ratelimit())
1711                                         printk(KERN_ERR "md/raid1:%s: redirecting sector %llu to"
1712                                                " other mirror: %s\n",
1713                                                mdname(mddev),
1714                                                (unsigned long long)r1_bio->sector,
1715                                                bdevname(rdev->bdev,b));
1716                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1717                                 bio->bi_bdev = rdev->bdev;
1718                                 bio->bi_end_io = raid1_end_read_request;
1719                                 bio->bi_rw = READ | do_sync;
1720                                 bio->bi_private = r1_bio;
1721                                 unplug = 1;
1722                                 generic_make_request(bio);
1723                         }
1724                 }
1725                 cond_resched();
1726         }
1727         if (unplug)
1728                 unplug_slaves(mddev);
1729 }
1730
1731
1732 static int init_resync(conf_t *conf)
1733 {
1734         int buffs;
1735
1736         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1737         BUG_ON(conf->r1buf_pool);
1738         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1739                                           conf->poolinfo);
1740         if (!conf->r1buf_pool)
1741                 return -ENOMEM;
1742         conf->next_resync = 0;
1743         return 0;
1744 }
1745
1746 /*
1747  * perform a "sync" on one "block"
1748  *
1749  * We need to make sure that no normal I/O request - particularly write
1750  * requests - conflict with active sync requests.
1751  *
1752  * This is achieved by tracking pending requests and a 'barrier' concept
1753  * that can be installed to exclude normal IO requests.
1754  */
1755
1756 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1757 {
1758         conf_t *conf = mddev->private;
1759         r1bio_t *r1_bio;
1760         struct bio *bio;
1761         sector_t max_sector, nr_sectors;
1762         int disk = -1;
1763         int i;
1764         int wonly = -1;
1765         int write_targets = 0, read_targets = 0;
1766         int sync_blocks;
1767         int still_degraded = 0;
1768
1769         if (!conf->r1buf_pool)
1770                 if (init_resync(conf))
1771                         return 0;
1772
1773         max_sector = mddev->dev_sectors;
1774         if (sector_nr >= max_sector) {
1775                 /* If we aborted, we need to abort the
1776                  * sync on the 'current' bitmap chunk (there will
1777                  * only be one in raid1 resync.
1778                  * We can find the current addess in mddev->curr_resync
1779                  */
1780                 if (mddev->curr_resync < max_sector) /* aborted */
1781                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1782                                                 &sync_blocks, 1);
1783                 else /* completed sync */
1784                         conf->fullsync = 0;
1785
1786                 bitmap_close_sync(mddev->bitmap);
1787                 close_sync(conf);
1788                 return 0;
1789         }
1790
1791         if (mddev->bitmap == NULL &&
1792             mddev->recovery_cp == MaxSector &&
1793             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1794             conf->fullsync == 0) {
1795                 *skipped = 1;
1796                 return max_sector - sector_nr;
1797         }
1798         /* before building a request, check if we can skip these blocks..
1799          * This call the bitmap_start_sync doesn't actually record anything
1800          */
1801         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1802             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1803                 /* We can skip this block, and probably several more */
1804                 *skipped = 1;
1805                 return sync_blocks;
1806         }
1807         /*
1808          * If there is non-resync activity waiting for a turn,
1809          * and resync is going fast enough,
1810          * then let it though before starting on this new sync request.
1811          */
1812         if (!go_faster && conf->nr_waiting)
1813                 msleep_interruptible(1000);
1814
1815         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1816         raise_barrier(conf);
1817
1818         conf->next_resync = sector_nr;
1819
1820         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1821         rcu_read_lock();
1822         /*
1823          * If we get a correctably read error during resync or recovery,
1824          * we might want to read from a different device.  So we
1825          * flag all drives that could conceivably be read from for READ,
1826          * and any others (which will be non-In_sync devices) for WRITE.
1827          * If a read fails, we try reading from something else for which READ
1828          * is OK.
1829          */
1830
1831         r1_bio->mddev = mddev;
1832         r1_bio->sector = sector_nr;
1833         r1_bio->state = 0;
1834         set_bit(R1BIO_IsSync, &r1_bio->state);
1835
1836         for (i=0; i < conf->raid_disks; i++) {
1837                 mdk_rdev_t *rdev;
1838                 bio = r1_bio->bios[i];
1839
1840                 /* take from bio_init */
1841                 bio->bi_next = NULL;
1842                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1843                 bio->bi_flags |= 1 << BIO_UPTODATE;
1844                 bio->bi_comp_cpu = -1;
1845                 bio->bi_rw = READ;
1846                 bio->bi_vcnt = 0;
1847                 bio->bi_idx = 0;
1848                 bio->bi_phys_segments = 0;
1849                 bio->bi_size = 0;
1850                 bio->bi_end_io = NULL;
1851                 bio->bi_private = NULL;
1852
1853                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1854                 if (rdev == NULL ||
1855                            test_bit(Faulty, &rdev->flags)) {
1856                         still_degraded = 1;
1857                         continue;
1858                 } else if (!test_bit(In_sync, &rdev->flags)) {
1859                         bio->bi_rw = WRITE;
1860                         bio->bi_end_io = end_sync_write;
1861                         write_targets ++;
1862                 } else {
1863                         /* may need to read from here */
1864                         bio->bi_rw = READ;
1865                         bio->bi_end_io = end_sync_read;
1866                         if (test_bit(WriteMostly, &rdev->flags)) {
1867                                 if (wonly < 0)
1868                                         wonly = i;
1869                         } else {
1870                                 if (disk < 0)
1871                                         disk = i;
1872                         }
1873                         read_targets++;
1874                 }
1875                 atomic_inc(&rdev->nr_pending);
1876                 bio->bi_sector = sector_nr + rdev->data_offset;
1877                 bio->bi_bdev = rdev->bdev;
1878                 bio->bi_private = r1_bio;
1879         }
1880         rcu_read_unlock();
1881         if (disk < 0)
1882                 disk = wonly;
1883         r1_bio->read_disk = disk;
1884
1885         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1886                 /* extra read targets are also write targets */
1887                 write_targets += read_targets-1;
1888
1889         if (write_targets == 0 || read_targets == 0) {
1890                 /* There is nowhere to write, so all non-sync
1891                  * drives must be failed - so we are finished
1892                  */
1893                 sector_t rv = max_sector - sector_nr;
1894                 *skipped = 1;
1895                 put_buf(r1_bio);
1896                 return rv;
1897         }
1898
1899         if (max_sector > mddev->resync_max)
1900                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1901         nr_sectors = 0;
1902         sync_blocks = 0;
1903         do {
1904                 struct page *page;
1905                 int len = PAGE_SIZE;
1906                 if (sector_nr + (len>>9) > max_sector)
1907                         len = (max_sector - sector_nr) << 9;
1908                 if (len == 0)
1909                         break;
1910                 if (sync_blocks == 0) {
1911                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1912                                                &sync_blocks, still_degraded) &&
1913                             !conf->fullsync &&
1914                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1915                                 break;
1916                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1917                         if ((len >> 9) > sync_blocks)
1918                                 len = sync_blocks<<9;
1919                 }
1920
1921                 for (i=0 ; i < conf->raid_disks; i++) {
1922                         bio = r1_bio->bios[i];
1923                         if (bio->bi_end_io) {
1924                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1925                                 if (bio_add_page(bio, page, len, 0) == 0) {
1926                                         /* stop here */
1927                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1928                                         while (i > 0) {
1929                                                 i--;
1930                                                 bio = r1_bio->bios[i];
1931                                                 if (bio->bi_end_io==NULL)
1932                                                         continue;
1933                                                 /* remove last page from this bio */
1934                                                 bio->bi_vcnt--;
1935                                                 bio->bi_size -= len;
1936                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1937                                         }
1938                                         goto bio_full;
1939                                 }
1940                         }
1941                 }
1942                 nr_sectors += len>>9;
1943                 sector_nr += len>>9;
1944                 sync_blocks -= (len>>9);
1945         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1946  bio_full:
1947         r1_bio->sectors = nr_sectors;
1948
1949         /* For a user-requested sync, we read all readable devices and do a
1950          * compare
1951          */
1952         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1953                 atomic_set(&r1_bio->remaining, read_targets);
1954                 for (i=0; i<conf->raid_disks; i++) {
1955                         bio = r1_bio->bios[i];
1956                         if (bio->bi_end_io == end_sync_read) {
1957                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1958                                 generic_make_request(bio);
1959                         }
1960                 }
1961         } else {
1962                 atomic_set(&r1_bio->remaining, 1);
1963                 bio = r1_bio->bios[r1_bio->read_disk];
1964                 md_sync_acct(bio->bi_bdev, nr_sectors);
1965                 generic_make_request(bio);
1966
1967         }
1968         return nr_sectors;
1969 }
1970
1971 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1972 {
1973         if (sectors)
1974                 return sectors;
1975
1976         return mddev->dev_sectors;
1977 }
1978
1979 static conf_t *setup_conf(mddev_t *mddev)
1980 {
1981         conf_t *conf;
1982         int i;
1983         mirror_info_t *disk;
1984         mdk_rdev_t *rdev;
1985         int err = -ENOMEM;
1986
1987         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1988         if (!conf)
1989                 goto abort;
1990
1991         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1992                                  GFP_KERNEL);
1993         if (!conf->mirrors)
1994                 goto abort;
1995
1996         conf->tmppage = alloc_page(GFP_KERNEL);
1997         if (!conf->tmppage)
1998                 goto abort;
1999
2000         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2001         if (!conf->poolinfo)
2002                 goto abort;
2003         conf->poolinfo->raid_disks = mddev->raid_disks;
2004         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2005                                           r1bio_pool_free,
2006                                           conf->poolinfo);
2007         if (!conf->r1bio_pool)
2008                 goto abort;
2009
2010         conf->poolinfo->mddev = mddev;
2011
2012         spin_lock_init(&conf->device_lock);
2013         list_for_each_entry(rdev, &mddev->disks, same_set) {
2014                 int disk_idx = rdev->raid_disk;
2015                 if (disk_idx >= mddev->raid_disks
2016                     || disk_idx < 0)
2017                         continue;
2018                 disk = conf->mirrors + disk_idx;
2019
2020                 disk->rdev = rdev;
2021
2022                 disk->head_position = 0;
2023         }
2024         conf->raid_disks = mddev->raid_disks;
2025         conf->mddev = mddev;
2026         INIT_LIST_HEAD(&conf->retry_list);
2027
2028         spin_lock_init(&conf->resync_lock);
2029         init_waitqueue_head(&conf->wait_barrier);
2030
2031         bio_list_init(&conf->pending_bio_list);
2032         bio_list_init(&conf->flushing_bio_list);
2033
2034         conf->last_used = -1;
2035         for (i = 0; i < conf->raid_disks; i++) {
2036
2037                 disk = conf->mirrors + i;
2038
2039                 if (!disk->rdev ||
2040                     !test_bit(In_sync, &disk->rdev->flags)) {
2041                         disk->head_position = 0;
2042                         if (disk->rdev)
2043                                 conf->fullsync = 1;
2044                 } else if (conf->last_used < 0)
2045                         /*
2046                          * The first working device is used as a
2047                          * starting point to read balancing.
2048                          */
2049                         conf->last_used = i;
2050         }
2051
2052         err = -EIO;
2053         if (conf->last_used < 0) {
2054                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2055                        mdname(mddev));
2056                 goto abort;
2057         }
2058         err = -ENOMEM;
2059         conf->thread = md_register_thread(raid1d, mddev, NULL);
2060         if (!conf->thread) {
2061                 printk(KERN_ERR
2062                        "md/raid1:%s: couldn't allocate thread\n",
2063                        mdname(mddev));
2064                 goto abort;
2065         }
2066
2067         return conf;
2068
2069  abort:
2070         if (conf) {
2071                 if (conf->r1bio_pool)
2072                         mempool_destroy(conf->r1bio_pool);
2073                 kfree(conf->mirrors);
2074                 safe_put_page(conf->tmppage);
2075                 kfree(conf->poolinfo);
2076                 kfree(conf);
2077         }
2078         return ERR_PTR(err);
2079 }
2080
2081 static int run(mddev_t *mddev)
2082 {
2083         conf_t *conf;
2084         int i;
2085         mdk_rdev_t *rdev;
2086
2087         if (mddev->level != 1) {
2088                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2089                        mdname(mddev), mddev->level);
2090                 return -EIO;
2091         }
2092         if (mddev->reshape_position != MaxSector) {
2093                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2094                        mdname(mddev));
2095                 return -EIO;
2096         }
2097         /*
2098          * copy the already verified devices into our private RAID1
2099          * bookkeeping area. [whatever we allocate in run(),
2100          * should be freed in stop()]
2101          */
2102         if (mddev->private == NULL)
2103                 conf = setup_conf(mddev);
2104         else
2105                 conf = mddev->private;
2106
2107         if (IS_ERR(conf))
2108                 return PTR_ERR(conf);
2109
2110         mddev->queue->queue_lock = &conf->device_lock;
2111         list_for_each_entry(rdev, &mddev->disks, same_set) {
2112                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2113                                   rdev->data_offset << 9);
2114                 /* as we don't honour merge_bvec_fn, we must never risk
2115                  * violating it, so limit ->max_segments to 1 lying within
2116                  * a single page, as a one page request is never in violation.
2117                  */
2118                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2119                         blk_queue_max_segments(mddev->queue, 1);
2120                         blk_queue_segment_boundary(mddev->queue,
2121                                                    PAGE_CACHE_SIZE - 1);
2122                 }
2123         }
2124
2125         mddev->degraded = 0;
2126         for (i=0; i < conf->raid_disks; i++)
2127                 if (conf->mirrors[i].rdev == NULL ||
2128                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2129                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2130                         mddev->degraded++;
2131
2132         if (conf->raid_disks - mddev->degraded == 1)
2133                 mddev->recovery_cp = MaxSector;
2134
2135         if (mddev->recovery_cp != MaxSector)
2136                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2137                        " -- starting background reconstruction\n",
2138                        mdname(mddev));
2139         printk(KERN_INFO
2140                 "md/raid1:%s: active with %d out of %d mirrors\n",
2141                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2142                 mddev->raid_disks);
2143
2144         /*
2145          * Ok, everything is just fine now
2146          */
2147         mddev->thread = conf->thread;
2148         conf->thread = NULL;
2149         mddev->private = conf;
2150
2151         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2152
2153         mddev->queue->unplug_fn = raid1_unplug;
2154         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2155         mddev->queue->backing_dev_info.congested_data = mddev;
2156         md_integrity_register(mddev);
2157         return 0;
2158 }
2159
2160 static int stop(mddev_t *mddev)
2161 {
2162         conf_t *conf = mddev->private;
2163         struct bitmap *bitmap = mddev->bitmap;
2164
2165         /* wait for behind writes to complete */
2166         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2167                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2168                        mdname(mddev));
2169                 /* need to kick something here to make sure I/O goes? */
2170                 wait_event(bitmap->behind_wait,
2171                            atomic_read(&bitmap->behind_writes) == 0);
2172         }
2173
2174         raise_barrier(conf);
2175         lower_barrier(conf);
2176
2177         md_unregister_thread(mddev->thread);
2178         mddev->thread = NULL;
2179         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2180         if (conf->r1bio_pool)
2181                 mempool_destroy(conf->r1bio_pool);
2182         kfree(conf->mirrors);
2183         kfree(conf->poolinfo);
2184         kfree(conf);
2185         mddev->private = NULL;
2186         return 0;
2187 }
2188
2189 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2190 {
2191         /* no resync is happening, and there is enough space
2192          * on all devices, so we can resize.
2193          * We need to make sure resync covers any new space.
2194          * If the array is shrinking we should possibly wait until
2195          * any io in the removed space completes, but it hardly seems
2196          * worth it.
2197          */
2198         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2199         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2200                 return -EINVAL;
2201         set_capacity(mddev->gendisk, mddev->array_sectors);
2202         revalidate_disk(mddev->gendisk);
2203         if (sectors > mddev->dev_sectors &&
2204             mddev->recovery_cp == MaxSector) {
2205                 mddev->recovery_cp = mddev->dev_sectors;
2206                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2207         }
2208         mddev->dev_sectors = sectors;
2209         mddev->resync_max_sectors = sectors;
2210         return 0;
2211 }
2212
2213 static int raid1_reshape(mddev_t *mddev)
2214 {
2215         /* We need to:
2216          * 1/ resize the r1bio_pool
2217          * 2/ resize conf->mirrors
2218          *
2219          * We allocate a new r1bio_pool if we can.
2220          * Then raise a device barrier and wait until all IO stops.
2221          * Then resize conf->mirrors and swap in the new r1bio pool.
2222          *
2223          * At the same time, we "pack" the devices so that all the missing
2224          * devices have the higher raid_disk numbers.
2225          */
2226         mempool_t *newpool, *oldpool;
2227         struct pool_info *newpoolinfo;
2228         mirror_info_t *newmirrors;
2229         conf_t *conf = mddev->private;
2230         int cnt, raid_disks;
2231         unsigned long flags;
2232         int d, d2, err;
2233
2234         /* Cannot change chunk_size, layout, or level */
2235         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2236             mddev->layout != mddev->new_layout ||
2237             mddev->level != mddev->new_level) {
2238                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2239                 mddev->new_layout = mddev->layout;
2240                 mddev->new_level = mddev->level;
2241                 return -EINVAL;
2242         }
2243
2244         err = md_allow_write(mddev);
2245         if (err)
2246                 return err;
2247
2248         raid_disks = mddev->raid_disks + mddev->delta_disks;
2249
2250         if (raid_disks < conf->raid_disks) {
2251                 cnt=0;
2252                 for (d= 0; d < conf->raid_disks; d++)
2253                         if (conf->mirrors[d].rdev)
2254                                 cnt++;
2255                 if (cnt > raid_disks)
2256                         return -EBUSY;
2257         }
2258
2259         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2260         if (!newpoolinfo)
2261                 return -ENOMEM;
2262         newpoolinfo->mddev = mddev;
2263         newpoolinfo->raid_disks = raid_disks;
2264
2265         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2266                                  r1bio_pool_free, newpoolinfo);
2267         if (!newpool) {
2268                 kfree(newpoolinfo);
2269                 return -ENOMEM;
2270         }
2271         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2272         if (!newmirrors) {
2273                 kfree(newpoolinfo);
2274                 mempool_destroy(newpool);
2275                 return -ENOMEM;
2276         }
2277
2278         raise_barrier(conf);
2279
2280         /* ok, everything is stopped */
2281         oldpool = conf->r1bio_pool;
2282         conf->r1bio_pool = newpool;
2283
2284         for (d = d2 = 0; d < conf->raid_disks; d++) {
2285                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2286                 if (rdev && rdev->raid_disk != d2) {
2287                         char nm[20];
2288                         sprintf(nm, "rd%d", rdev->raid_disk);
2289                         sysfs_remove_link(&mddev->kobj, nm);
2290                         rdev->raid_disk = d2;
2291                         sprintf(nm, "rd%d", rdev->raid_disk);
2292                         sysfs_remove_link(&mddev->kobj, nm);
2293                         if (sysfs_create_link(&mddev->kobj,
2294                                               &rdev->kobj, nm))
2295                                 printk(KERN_WARNING
2296                                        "md/raid1:%s: cannot register "
2297                                        "%s\n",
2298                                        mdname(mddev), nm);
2299                 }
2300                 if (rdev)
2301                         newmirrors[d2++].rdev = rdev;
2302         }
2303         kfree(conf->mirrors);
2304         conf->mirrors = newmirrors;
2305         kfree(conf->poolinfo);
2306         conf->poolinfo = newpoolinfo;
2307
2308         spin_lock_irqsave(&conf->device_lock, flags);
2309         mddev->degraded += (raid_disks - conf->raid_disks);
2310         spin_unlock_irqrestore(&conf->device_lock, flags);
2311         conf->raid_disks = mddev->raid_disks = raid_disks;
2312         mddev->delta_disks = 0;
2313
2314         conf->last_used = 0; /* just make sure it is in-range */
2315         lower_barrier(conf);
2316
2317         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2318         md_wakeup_thread(mddev->thread);
2319
2320         mempool_destroy(oldpool);
2321         return 0;
2322 }
2323
2324 static void raid1_quiesce(mddev_t *mddev, int state)
2325 {
2326         conf_t *conf = mddev->private;
2327
2328         switch(state) {
2329         case 2: /* wake for suspend */
2330                 wake_up(&conf->wait_barrier);
2331                 break;
2332         case 1:
2333                 raise_barrier(conf);
2334                 break;
2335         case 0:
2336                 lower_barrier(conf);
2337                 break;
2338         }
2339 }
2340
2341 static void *raid1_takeover(mddev_t *mddev)
2342 {
2343         /* raid1 can take over:
2344          *  raid5 with 2 devices, any layout or chunk size
2345          */
2346         if (mddev->level == 5 && mddev->raid_disks == 2) {
2347                 conf_t *conf;
2348                 mddev->new_level = 1;
2349                 mddev->new_layout = 0;
2350                 mddev->new_chunk_sectors = 0;
2351                 conf = setup_conf(mddev);
2352                 if (!IS_ERR(conf))
2353                         conf->barrier = 1;
2354                 return conf;
2355         }
2356         return ERR_PTR(-EINVAL);
2357 }
2358
2359 static struct mdk_personality raid1_personality =
2360 {
2361         .name           = "raid1",
2362         .level          = 1,
2363         .owner          = THIS_MODULE,
2364         .make_request   = make_request,
2365         .run            = run,
2366         .stop           = stop,
2367         .status         = status,
2368         .error_handler  = error,
2369         .hot_add_disk   = raid1_add_disk,
2370         .hot_remove_disk= raid1_remove_disk,
2371         .spare_active   = raid1_spare_active,
2372         .sync_request   = sync_request,
2373         .resize         = raid1_resize,
2374         .size           = raid1_size,
2375         .check_reshape  = raid1_reshape,
2376         .quiesce        = raid1_quiesce,
2377         .takeover       = raid1_takeover,
2378 };
2379
2380 static int __init raid_init(void)
2381 {
2382         return register_md_personality(&raid1_personality);
2383 }
2384
2385 static void raid_exit(void)
2386 {
2387         unregister_md_personality(&raid1_personality);
2388 }
2389
2390 module_init(raid_init);
2391 module_exit(raid_exit);
2392 MODULE_LICENSE("GPL");
2393 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2394 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2395 MODULE_ALIAS("md-raid1");
2396 MODULE_ALIAS("md-level-1");