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 "dm-bio-list.h"
  35 #include <linux/raid/raid1.h>
  36 #include <linux/raid/bitmap.h>
  37
  38 #define DEBUG 0
  39 #if DEBUG
  40 #define PRINTK(x...) printk(x)
  41 #else
  42 #define PRINTK(x...)
  43 #endif
  44
  45 /*
  46  * Number of guaranteed r1bios in case of extreme VM load:
  47  */
  48 #define NR_RAID1_BIOS 256
  49
  50
  51 static void unplug_slaves(mddev_t *mddev);
  52
  53 static void allow_barrier(conf_t *conf);
  54 static void lower_barrier(conf_t *conf);
  55
  56 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  57 {
  58         struct pool_info *pi = data;
  59         r1bio_t *r1_bio;
  60         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  61
  62         /* allocate a r1bio with room for raid_disks entries in the bios array */
  63         r1_bio = kzalloc(size, gfp_flags);
  64         if (!r1_bio)
  65                 unplug_slaves(pi->mddev);
  66
  67         return r1_bio;
  68 }
  69
  70 static void r1bio_pool_free(void *r1_bio, void *data)
  71 {
  72         kfree(r1_bio);
  73 }
  74
  75 #define RESYNC_BLOCK_SIZE (64*1024)
  76 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  77 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  78 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  79 #define RESYNC_WINDOW (2048*1024)
  80
  81 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  82 {
  83         struct pool_info *pi = data;
  84         struct page *page;
  85         r1bio_t *r1_bio;
  86         struct bio *bio;
  87         int i, j;
  88
  89         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  90         if (!r1_bio) {
  91                 unplug_slaves(pi->mddev);
  92                 return NULL;
  93         }
  94
  95         /*
  96          * Allocate bios : 1 for reading, n-1 for writing
  97          */
  98         for (j = pi->raid_disks ; j-- ; ) {
  99                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 100                 if (!bio)
 101                         goto out_free_bio;
 102                 r1_bio->bios[j] = bio;
 103         }
 104         /*
 105          * Allocate RESYNC_PAGES data pages and attach them to
 106          * the first bio.
 107          * If this is a user-requested check/repair, allocate
 108          * RESYNC_PAGES for each bio.
 109          */
 110         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 111                 j = pi->raid_disks;
 112         else
 113                 j = 1;
 114         while(j--) {
 115                 bio = r1_bio->bios[j];
 116                 for (i = 0; i < RESYNC_PAGES; i++) {
 117                         page = alloc_page(gfp_flags);
 118                         if (unlikely(!page))
 119                                 goto out_free_pages;
 120
 121                         bio->bi_io_vec[i].bv_page = page;
 122                 }
 123         }
 124         /* If not user-requests, copy the page pointers to all bios */
 125         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 126                 for (i=0; i<RESYNC_PAGES ; i++)
 127                         for (j=1; j<pi->raid_disks; j++)
 128                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 129                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 130         }
 131
 132         r1_bio->master_bio = NULL;
 133
 134         return r1_bio;
 135
 136 out_free_pages:
 137         for (i=0; i < RESYNC_PAGES ; i++)
 138                 for (j=0 ; j < pi->raid_disks; j++)
 139                         safe_put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 140         j = -1;
 141 out_free_bio:
 142         while ( ++j < pi->raid_disks )
 143                 bio_put(r1_bio->bios[j]);
 144         r1bio_pool_free(r1_bio, data);
 145         return NULL;
 146 }
 147
 148 static void r1buf_pool_free(void *__r1_bio, void *data)
 149 {
 150         struct pool_info *pi = data;
 151         int i,j;
 152         r1bio_t *r1bio = __r1_bio;
 153
 154         for (i = 0; i < RESYNC_PAGES; i++)
 155                 for (j = pi->raid_disks; j-- ;) {
 156                         if (j == 0 ||
 157                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 158                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 159                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 160                 }
 161         for (i=0 ; i < pi->raid_disks; i++)
 162                 bio_put(r1bio->bios[i]);
 163
 164         r1bio_pool_free(r1bio, data);
 165 }
 166
 167 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 168 {
 169         int i;
 170
 171         for (i = 0; i < conf->raid_disks; i++) {
 172                 struct bio **bio = r1_bio->bios + i;
 173                 if (*bio && *bio != IO_BLOCKED)
 174                         bio_put(*bio);
 175                 *bio = NULL;
 176         }
 177 }
 178
 179 static void free_r1bio(r1bio_t *r1_bio)
 180 {
 181         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 182
 183         /*
 184          * Wake up any possible resync thread that waits for the device
 185          * to go idle.
 186          */
 187         allow_barrier(conf);
 188
 189         put_all_bios(conf, r1_bio);
 190         mempool_free(r1_bio, conf->r1bio_pool);
 191 }
 192
 193 static void put_buf(r1bio_t *r1_bio)
 194 {
 195         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 196         int i;
 197
 198         for (i=0; i<conf->raid_disks; i++) {
 199                 struct bio *bio = r1_bio->bios[i];
 200                 if (bio->bi_end_io)
 201                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 202         }
 203
 204         mempool_free(r1_bio, conf->r1buf_pool);
 205
 206         lower_barrier(conf);
 207 }
 208
 209 static void reschedule_retry(r1bio_t *r1_bio)
 210 {
 211         unsigned long flags;
 212         mddev_t *mddev = r1_bio->mddev;
 213         conf_t *conf = mddev_to_conf(mddev);
 214
 215         spin_lock_irqsave(&conf->device_lock, flags);
 216         list_add(&r1_bio->retry_list, &conf->retry_list);
 217         conf->nr_queued ++;
 218         spin_unlock_irqrestore(&conf->device_lock, flags);
 219
 220         wake_up(&conf->wait_barrier);
 221         md_wakeup_thread(mddev->thread);
 222 }
 223
 224 /*
 225  * raid_end_bio_io() is called when we have finished servicing a mirrored
 226  * operation and are ready to return a success/failure code to the buffer
 227  * cache layer.
 228  */
 229 static void raid_end_bio_io(r1bio_t *r1_bio)
 230 {
 231         struct bio *bio = r1_bio->master_bio;
 232
 233         /* if nobody has done the final endio yet, do it now */
 234         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 235                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 236                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 237                         (unsigned long long) bio->bi_sector,
 238                         (unsigned long long) bio->bi_sector +
 239                                 (bio->bi_size >> 9) - 1);
 240
 241                 bio_endio(bio,
 242                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 243         }
 244         free_r1bio(r1_bio);
 245 }
 246
 247 /*
 248  * Update disk head position estimator based on IRQ completion info.
 249  */
 250 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 251 {
 252         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 253
 254         conf->mirrors[disk].head_position =
 255                 r1_bio->sector + (r1_bio->sectors);
 256 }
 257
 258 static void raid1_end_read_request(struct bio *bio, int error)
 259 {
 260         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 261         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 262         int mirror;
 263         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 264
 265         mirror = r1_bio->read_disk;
 266         /*
 267          * this branch is our 'one mirror IO has finished' event handler:
 268          */
 269         update_head_pos(mirror, r1_bio);
 270
 271         if (uptodate)
 272                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 273         else {
 274                 /* If all other devices have failed, we want to return
 275                  * the error upwards rather than fail the last device.
 276                  * Here we redefine "uptodate" to mean "Don't want to retry"
 277                  */
 278                 unsigned long flags;
 279                 spin_lock_irqsave(&conf->device_lock, flags);
 280                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 281                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 282                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 283                         uptodate = 1;
 284                 spin_unlock_irqrestore(&conf->device_lock, flags);
 285         }
 286
 287         if (uptodate)
 288                 raid_end_bio_io(r1_bio);
 289         else {
 290                 /*
 291                  * oops, read error:
 292                  */
 293                 char b[BDEVNAME_SIZE];
 294                 if (printk_ratelimit())
 295                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 296                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 297                 reschedule_retry(r1_bio);
 298         }
 299
 300         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 301 }
 302
 303 static void raid1_end_write_request(struct bio *bio, int error)
 304 {
 305         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 306         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 307         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 308         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 309         struct bio *to_put = NULL;
 310
 311
 312         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 313                 if (r1_bio->bios[mirror] == bio)
 314                         break;
 315
 316         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
 317                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
 318                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
 319                 r1_bio->mddev->barriers_work = 0;
 320                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
 321         } else {
 322                 /*
 323                  * this branch is our 'one mirror IO has finished' event handler:
 324                  */
 325                 r1_bio->bios[mirror] = NULL;
 326                 to_put = bio;
 327                 if (!uptodate) {
 328                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 329                         /* an I/O failed, we can't clear the bitmap */
 330                         set_bit(R1BIO_Degraded, &r1_bio->state);
 331                 } else
 332                         /*
 333                          * Set R1BIO_Uptodate in our master bio, so that
 334                          * we will return a good error code for to the higher
 335                          * levels even if IO on some other mirrored buffer fails.
 336                          *
 337                          * The 'master' represents the composite IO operation to
 338                          * user-side. So if something waits for IO, then it will
 339                          * wait for the 'master' bio.
 340                          */
 341                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 342
 343                 update_head_pos(mirror, r1_bio);
 344
 345                 if (behind) {
 346                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 347                                 atomic_dec(&r1_bio->behind_remaining);
 348
 349                         /* In behind mode, we ACK the master bio once the I/O has safely
 350                          * reached all non-writemostly disks. Setting the Returned bit
 351                          * ensures that this gets done only once -- we don't ever want to
 352                          * return -EIO here, instead we'll wait */
 353
 354                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 355                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 356                                 /* Maybe we can return now */
 357                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 358                                         struct bio *mbio = r1_bio->master_bio;
 359                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 360                                                (unsigned long long) mbio->bi_sector,
 361                                                (unsigned long long) mbio->bi_sector +
 362                                                (mbio->bi_size >> 9) - 1);
 363                                         bio_endio(mbio, 0);
 364                                 }
 365                         }
 366                 }
 367                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 368         }
 369         /*
 370          *
 371          * Let's see if all mirrored write operations have finished
 372          * already.
 373          */
 374         if (atomic_dec_and_test(&r1_bio->remaining)) {
 375                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
 376                         reschedule_retry(r1_bio);
 377                 else {
 378                         /* it really is the end of this request */
 379                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 380                                 /* free extra copy of the data pages */
 381                                 int i = bio->bi_vcnt;
 382                                 while (i--)
 383                                         safe_put_page(bio->bi_io_vec[i].bv_page);
 384                         }
 385                         /* clear the bitmap if all writes complete successfully */
 386                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 387                                         r1_bio->sectors,
 388                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 389                                         behind);
 390                         md_write_end(r1_bio->mddev);
 391                         raid_end_bio_io(r1_bio);
 392                 }
 393         }
 394
 395         if (to_put)
 396                 bio_put(to_put);
 397 }
 398
 399
 400 /*
 401  * This routine returns the disk from which the requested read should
 402  * be done. There is a per-array 'next expected sequential IO' sector
 403  * number - if this matches on the next IO then we use the last disk.
 404  * There is also a per-disk 'last know head position' sector that is
 405  * maintained from IRQ contexts, both the normal and the resync IO
 406  * completion handlers update this position correctly. If there is no
 407  * perfect sequential match then we pick the disk whose head is closest.
 408  *
 409  * If there are 2 mirrors in the same 2 devices, performance degrades
 410  * because position is mirror, not device based.
 411  *
 412  * The rdev for the device selected will have nr_pending incremented.
 413  */
 414 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 415 {
 416         const unsigned long this_sector = r1_bio->sector;
 417         int new_disk = conf->last_used, disk = new_disk;
 418         int wonly_disk = -1;
 419         const int sectors = r1_bio->sectors;
 420         sector_t new_distance, current_distance;
 421         mdk_rdev_t *rdev;
 422
 423         rcu_read_lock();
 424         /*
 425          * Check if we can balance. We can balance on the whole
 426          * device if no resync is going on, or below the resync window.
 427          * We take the first readable disk when above the resync window.
 428          */
 429  retry:
 430         if (conf->mddev->recovery_cp < MaxSector &&
 431             (this_sector + sectors >= conf->next_resync)) {
 432                 /* Choose the first operation device, for consistancy */
 433                 new_disk = 0;
 434
 435                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 436                      r1_bio->bios[new_disk] == IO_BLOCKED ||
 437                      !rdev || !test_bit(In_sync, &rdev->flags)
 438                              || test_bit(WriteMostly, &rdev->flags);
 439                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
 440
 441                         if (rdev && test_bit(In_sync, &rdev->flags) &&
 442                                 r1_bio->bios[new_disk] != IO_BLOCKED)
 443                                 wonly_disk = new_disk;
 444
 445                         if (new_disk == conf->raid_disks - 1) {
 446                                 new_disk = wonly_disk;
 447                                 break;
 448                         }
 449                 }
 450                 goto rb_out;
 451         }
 452
 453
 454         /* make sure the disk is operational */
 455         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 456              r1_bio->bios[new_disk] == IO_BLOCKED ||
 457              !rdev || !test_bit(In_sync, &rdev->flags) ||
 458                      test_bit(WriteMostly, &rdev->flags);
 459              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
 460
 461                 if (rdev && test_bit(In_sync, &rdev->flags) &&
 462                     r1_bio->bios[new_disk] != IO_BLOCKED)
 463                         wonly_disk = new_disk;
 464
 465                 if (new_disk <= 0)
 466                         new_disk = conf->raid_disks;
 467                 new_disk--;
 468                 if (new_disk == disk) {
 469                         new_disk = wonly_disk;
 470                         break;
 471                 }
 472         }
 473
 474         if (new_disk < 0)
 475                 goto rb_out;
 476
 477         disk = new_disk;
 478         /* now disk == new_disk == starting point for search */
 479
 480         /*
 481          * Don't change to another disk for sequential reads:
 482          */
 483         if (conf->next_seq_sect == this_sector)
 484                 goto rb_out;
 485         if (this_sector == conf->mirrors[new_disk].head_position)
 486                 goto rb_out;
 487
 488         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 489
 490         /* Find the disk whose head is closest */
 491
 492         do {
 493                 if (disk <= 0)
 494                         disk = conf->raid_disks;
 495                 disk--;
 496
 497                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 498
 499                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
 500                     !test_bit(In_sync, &rdev->flags) ||
 501                     test_bit(WriteMostly, &rdev->flags))
 502                         continue;
 503
 504                 if (!atomic_read(&rdev->nr_pending)) {
 505                         new_disk = disk;
 506                         break;
 507                 }
 508                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 509                 if (new_distance < current_distance) {
 510                         current_distance = new_distance;
 511                         new_disk = disk;
 512                 }
 513         } while (disk != conf->last_used);
 514
 515  rb_out:
 516
 517
 518         if (new_disk >= 0) {
 519                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 520                 if (!rdev)
 521                         goto retry;
 522                 atomic_inc(&rdev->nr_pending);
 523                 if (!test_bit(In_sync, &rdev->flags)) {
 524                         /* cannot risk returning a device that failed
 525                          * before we inc'ed nr_pending
 526                          */
 527                         rdev_dec_pending(rdev, conf->mddev);
 528                         goto retry;
 529                 }
 530                 conf->next_seq_sect = this_sector + sectors;
 531                 conf->last_used = new_disk;
 532         }
 533         rcu_read_unlock();
 534
 535         return new_disk;
 536 }
 537
 538 static void unplug_slaves(mddev_t *mddev)
 539 {
 540         conf_t *conf = mddev_to_conf(mddev);
 541         int i;
 542
 543         rcu_read_lock();
 544         for (i=0; i<mddev->raid_disks; i++) {
 545                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 546                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 547                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 548
 549                         atomic_inc(&rdev->nr_pending);
 550                         rcu_read_unlock();
 551
 552                         blk_unplug(r_queue);
 553
 554                         rdev_dec_pending(rdev, mddev);
 555                         rcu_read_lock();
 556                 }
 557         }
 558         rcu_read_unlock();
 559 }
 560
 561 static void raid1_unplug(struct request_queue *q)
 562 {
 563         mddev_t *mddev = q->queuedata;
 564
 565         unplug_slaves(mddev);
 566         md_wakeup_thread(mddev->thread);
 567 }
 568
 569 static int raid1_congested(void *data, int bits)
 570 {
 571         mddev_t *mddev = data;
 572         conf_t *conf = mddev_to_conf(mddev);
 573         int i, ret = 0;
 574
 575         rcu_read_lock();
 576         for (i = 0; i < mddev->raid_disks; i++) {
 577                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 578                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 579                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 580
 581                         /* Note the '|| 1' - when read_balance prefers
 582                          * non-congested targets, it can be removed
 583                          */
 584                         if ((bits & (1<<BDI_write_congested)) || 1)
 585                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 586                         else
 587                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 588                 }
 589         }
 590         rcu_read_unlock();
 591         return ret;
 592 }
 593
 594
 595 static int flush_pending_writes(conf_t *conf)
 596 {
 597         /* Any writes that have been queued but are awaiting
 598          * bitmap updates get flushed here.
 599          * We return 1 if any requests were actually submitted.
 600          */
 601         int rv = 0;
 602
 603         spin_lock_irq(&conf->device_lock);
 604
 605         if (conf->pending_bio_list.head) {
 606                 struct bio *bio;
 607                 bio = bio_list_get(&conf->pending_bio_list);
 608                 blk_remove_plug(conf->mddev->queue);
 609                 spin_unlock_irq(&conf->device_lock);
 610                 /* flush any pending bitmap writes to
 611                  * disk before proceeding w/ I/O */
 612                 bitmap_unplug(conf->mddev->bitmap);
 613
 614                 while (bio) { /* submit pending writes */
 615                         struct bio *next = bio->bi_next;
 616                         bio->bi_next = NULL;
 617                         generic_make_request(bio);
 618                         bio = next;
 619                 }
 620                 rv = 1;
 621         } else
 622                 spin_unlock_irq(&conf->device_lock);
 623         return rv;
 624 }
 625
 626 /* Barriers....
 627  * Sometimes we need to suspend IO while we do something else,
 628  * either some resync/recovery, or reconfigure the array.
 629  * To do this we raise a 'barrier'.
 630  * The 'barrier' is a counter that can be raised multiple times
 631  * to count how many activities are happening which preclude
 632  * normal IO.
 633  * We can only raise the barrier if there is no pending IO.
 634  * i.e. if nr_pending == 0.
 635  * We choose only to raise the barrier if no-one is waiting for the
 636  * barrier to go down.  This means that as soon as an IO request
 637  * is ready, no other operations which require a barrier will start
 638  * until the IO request has had a chance.
 639  *
 640  * So: regular IO calls 'wait_barrier'.  When that returns there
 641  *    is no backgroup IO happening,  It must arrange to call
 642  *    allow_barrier when it has finished its IO.
 643  * backgroup IO calls must call raise_barrier.  Once that returns
 644  *    there is no normal IO happeing.  It must arrange to call
 645  *    lower_barrier when the particular background IO completes.
 646  */
 647 #define RESYNC_DEPTH 32
 648
 649 static void raise_barrier(conf_t *conf)
 650 {
 651         spin_lock_irq(&conf->resync_lock);
 652
 653         /* Wait until no block IO is waiting */
 654         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 655                             conf->resync_lock,
 656                             raid1_unplug(conf->mddev->queue));
 657
 658         /* block any new IO from starting */
 659         conf->barrier++;
 660
 661         /* No wait for all pending IO to complete */
 662         wait_event_lock_irq(conf->wait_barrier,
 663                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 664                             conf->resync_lock,
 665                             raid1_unplug(conf->mddev->queue));
 666
 667         spin_unlock_irq(&conf->resync_lock);
 668 }
 669
 670 static void lower_barrier(conf_t *conf)
 671 {
 672         unsigned long flags;
 673         spin_lock_irqsave(&conf->resync_lock, flags);
 674         conf->barrier--;
 675         spin_unlock_irqrestore(&conf->resync_lock, flags);
 676         wake_up(&conf->wait_barrier);
 677 }
 678
 679 static void wait_barrier(conf_t *conf)
 680 {
 681         spin_lock_irq(&conf->resync_lock);
 682         if (conf->barrier) {
 683                 conf->nr_waiting++;
 684                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 685                                     conf->resync_lock,
 686                                     raid1_unplug(conf->mddev->queue));
 687                 conf->nr_waiting--;
 688         }
 689         conf->nr_pending++;
 690         spin_unlock_irq(&conf->resync_lock);
 691 }
 692
 693 static void allow_barrier(conf_t *conf)
 694 {
 695         unsigned long flags;
 696         spin_lock_irqsave(&conf->resync_lock, flags);
 697         conf->nr_pending--;
 698         spin_unlock_irqrestore(&conf->resync_lock, flags);
 699         wake_up(&conf->wait_barrier);
 700 }
 701
 702 static void freeze_array(conf_t *conf)
 703 {
 704         /* stop syncio and normal IO and wait for everything to
 705          * go quite.
 706          * We increment barrier and nr_waiting, and then
 707          * wait until nr_pending match nr_queued+1
 708          * This is called in the context of one normal IO request
 709          * that has failed. Thus any sync request that might be pending
 710          * will be blocked by nr_pending, and we need to wait for
 711          * pending IO requests to complete or be queued for re-try.
 712          * Thus the number queued (nr_queued) plus this request (1)
 713          * must match the number of pending IOs (nr_pending) before
 714          * we continue.
 715          */
 716         spin_lock_irq(&conf->resync_lock);
 717         conf->barrier++;
 718         conf->nr_waiting++;
 719         wait_event_lock_irq(conf->wait_barrier,
 720                             conf->nr_pending == conf->nr_queued+1,
 721                             conf->resync_lock,
 722                             ({ flush_pending_writes(conf);
 723                                raid1_unplug(conf->mddev->queue); }));
 724         spin_unlock_irq(&conf->resync_lock);
 725 }
 726 static void unfreeze_array(conf_t *conf)
 727 {
 728         /* reverse the effect of the freeze */
 729         spin_lock_irq(&conf->resync_lock);
 730         conf->barrier--;
 731         conf->nr_waiting--;
 732         wake_up(&conf->wait_barrier);
 733         spin_unlock_irq(&conf->resync_lock);
 734 }
 735
 736
 737 /* duplicate the data pages for behind I/O */
 738 static struct page **alloc_behind_pages(struct bio *bio)
 739 {
 740         int i;
 741         struct bio_vec *bvec;
 742         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 743                                         GFP_NOIO);
 744         if (unlikely(!pages))
 745                 goto do_sync_io;
 746
 747         bio_for_each_segment(bvec, bio, i) {
 748                 pages[i] = alloc_page(GFP_NOIO);
 749                 if (unlikely(!pages[i]))
 750                         goto do_sync_io;
 751                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 752                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 753                 kunmap(pages[i]);
 754                 kunmap(bvec->bv_page);
 755         }
 756
 757         return pages;
 758
 759 do_sync_io:
 760         if (pages)
 761                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 762                         put_page(pages[i]);
 763         kfree(pages);
 764         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 765         return NULL;
 766 }
 767
 768 static int make_request(struct request_queue *q, struct bio * bio)
 769 {
 770         mddev_t *mddev = q->queuedata;
 771         conf_t *conf = mddev_to_conf(mddev);
 772         mirror_info_t *mirror;
 773         r1bio_t *r1_bio;
 774         struct bio *read_bio;
 775         int i, targets = 0, disks;
 776         struct bitmap *bitmap;
 777         unsigned long flags;
 778         struct bio_list bl;
 779         struct page **behind_pages = NULL;
 780         const int rw = bio_data_dir(bio);
 781         const int do_sync = bio_sync(bio);
 782         int do_barriers;
 783         mdk_rdev_t *blocked_rdev;
 784
 785         /*
 786          * Register the new request and wait if the reconstruction
 787          * thread has put up a bar for new requests.
 788          * Continue immediately if no resync is active currently.
 789          * We test barriers_work *after* md_write_start as md_write_start
 790          * may cause the first superblock write, and that will check out
 791          * if barriers work.
 792          */
 793
 794         md_write_start(mddev, bio); /* wait on superblock update early */
 795
 796         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
 797                 if (rw == WRITE)
 798                         md_write_end(mddev);
 799                 bio_endio(bio, -EOPNOTSUPP);
 800                 return 0;
 801         }
 802
 803         wait_barrier(conf);
 804
 805         bitmap = mddev->bitmap;
 806
 807         disk_stat_inc(mddev->gendisk, ios[rw]);
 808         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 809
 810         /*
 811          * make_request() can abort the operation when READA is being
 812          * used and no empty request is available.
 813          *
 814          */
 815         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 816
 817         r1_bio->master_bio = bio;
 818         r1_bio->sectors = bio->bi_size >> 9;
 819         r1_bio->state = 0;
 820         r1_bio->mddev = mddev;
 821         r1_bio->sector = bio->bi_sector;
 822
 823         if (rw == READ) {
 824                 /*
 825                  * read balancing logic:
 826                  */
 827                 int rdisk = read_balance(conf, r1_bio);
 828
 829                 if (rdisk < 0) {
 830                         /* couldn't find anywhere to read from */
 831                         raid_end_bio_io(r1_bio);
 832                         return 0;
 833                 }
 834                 mirror = conf->mirrors + rdisk;
 835
 836                 r1_bio->read_disk = rdisk;
 837
 838                 read_bio = bio_clone(bio, GFP_NOIO);
 839
 840                 r1_bio->bios[rdisk] = read_bio;
 841
 842                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 843                 read_bio->bi_bdev = mirror->rdev->bdev;
 844                 read_bio->bi_end_io = raid1_end_read_request;
 845                 read_bio->bi_rw = READ | do_sync;
 846                 read_bio->bi_private = r1_bio;
 847
 848                 generic_make_request(read_bio);
 849                 return 0;
 850         }
 851
 852         /*
 853          * WRITE:
 854          */
 855         /* first select target devices under spinlock and
 856          * inc refcount on their rdev.  Record them by setting
 857          * bios[x] to bio
 858          */
 859         disks = conf->raid_disks;
 860 #if 0
 861         { static int first=1;
 862         if (first) printk("First Write sector %llu disks %d\n",
 863                           (unsigned long long)r1_bio->sector, disks);
 864         first = 0;
 865         }
 866 #endif
 867  retry_write:
 868         blocked_rdev = NULL;
 869         rcu_read_lock();
 870         for (i = 0;  i < disks; i++) {
 871                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 872                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 873                         atomic_inc(&rdev->nr_pending);
 874                         blocked_rdev = rdev;
 875                         break;
 876                 }
 877                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 878                         atomic_inc(&rdev->nr_pending);
 879                         if (test_bit(Faulty, &rdev->flags)) {
 880                                 rdev_dec_pending(rdev, mddev);
 881                                 r1_bio->bios[i] = NULL;
 882                         } else
 883                                 r1_bio->bios[i] = bio;
 884                         targets++;
 885                 } else
 886                         r1_bio->bios[i] = NULL;
 887         }
 888         rcu_read_unlock();
 889
 890         if (unlikely(blocked_rdev)) {
 891                 /* Wait for this device to become unblocked */
 892                 int j;
 893
 894                 for (j = 0; j < i; j++)
 895                         if (r1_bio->bios[j])
 896                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 897
 898                 allow_barrier(conf);
 899                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 900                 wait_barrier(conf);
 901                 goto retry_write;
 902         }
 903
 904         BUG_ON(targets == 0); /* we never fail the last device */
 905
 906         if (targets < conf->raid_disks) {
 907                 /* array is degraded, we will not clear the bitmap
 908                  * on I/O completion (see raid1_end_write_request) */
 909                 set_bit(R1BIO_Degraded, &r1_bio->state);
 910         }
 911
 912         /* do behind I/O ? */
 913         if (bitmap &&
 914             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
 915             (behind_pages = alloc_behind_pages(bio)) != NULL)
 916                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 917
 918         atomic_set(&r1_bio->remaining, 0);
 919         atomic_set(&r1_bio->behind_remaining, 0);
 920
 921         do_barriers = bio_barrier(bio);
 922         if (do_barriers)
 923                 set_bit(R1BIO_Barrier, &r1_bio->state);
 924
 925         bio_list_init(&bl);
 926         for (i = 0; i < disks; i++) {
 927                 struct bio *mbio;
 928                 if (!r1_bio->bios[i])
 929                         continue;
 930
 931                 mbio = bio_clone(bio, GFP_NOIO);
 932                 r1_bio->bios[i] = mbio;
 933
 934                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 935                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 936                 mbio->bi_end_io = raid1_end_write_request;
 937                 mbio->bi_rw = WRITE | do_barriers | do_sync;
 938                 mbio->bi_private = r1_bio;
 939
 940                 if (behind_pages) {
 941                         struct bio_vec *bvec;
 942                         int j;
 943
 944                         /* Yes, I really want the '__' version so that
 945                          * we clear any unused pointer in the io_vec, rather
 946                          * than leave them unchanged.  This is important
 947                          * because when we come to free the pages, we won't
 948                          * know the originial bi_idx, so we just free
 949                          * them all
 950                          */
 951                         __bio_for_each_segment(bvec, mbio, j, 0)
 952                                 bvec->bv_page = behind_pages[j];
 953                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 954                                 atomic_inc(&r1_bio->behind_remaining);
 955                 }
 956
 957                 atomic_inc(&r1_bio->remaining);
 958
 959                 bio_list_add(&bl, mbio);
 960         }
 961         kfree(behind_pages); /* the behind pages are attached to the bios now */
 962
 963         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 964                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 965         spin_lock_irqsave(&conf->device_lock, flags);
 966         bio_list_merge(&conf->pending_bio_list, &bl);
 967         bio_list_init(&bl);
 968
 969         blk_plug_device(mddev->queue);
 970         spin_unlock_irqrestore(&conf->device_lock, flags);
 971
 972         /* In case raid1d snuck into freeze_array */
 973         wake_up(&conf->wait_barrier);
 974
 975         if (do_sync)
 976                 md_wakeup_thread(mddev->thread);
 977 #if 0
 978         while ((bio = bio_list_pop(&bl)) != NULL)
 979                 generic_make_request(bio);
 980 #endif
 981
 982         return 0;
 983 }
 984
 985 static void status(struct seq_file *seq, mddev_t *mddev)
 986 {
 987         conf_t *conf = mddev_to_conf(mddev);
 988         int i;
 989
 990         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 991                    conf->raid_disks - mddev->degraded);
 992         rcu_read_lock();
 993         for (i = 0; i < conf->raid_disks; i++) {
 994                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 995                 seq_printf(seq, "%s",
 996                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
 997         }
 998         rcu_read_unlock();
 999         seq_printf(seq, "]");
1000 }
1001
1002
1003 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1004 {
1005         char b[BDEVNAME_SIZE];
1006         conf_t *conf = mddev_to_conf(mddev);
1007
1008         /*
1009          * If it is not operational, then we have already marked it as dead
1010          * else if it is the last working disks, ignore the error, let the
1011          * next level up know.
1012          * else mark the drive as failed
1013          */
1014         if (test_bit(In_sync, &rdev->flags)
1015             && (conf->raid_disks - mddev->degraded) == 1)
1016                 /*
1017                  * Don't fail the drive, act as though we were just a
1018                  * normal single drive
1019                  */
1020                 return;
1021         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1022                 unsigned long flags;
1023                 spin_lock_irqsave(&conf->device_lock, flags);
1024                 mddev->degraded++;
1025                 set_bit(Faulty, &rdev->flags);
1026                 spin_unlock_irqrestore(&conf->device_lock, flags);
1027                 /*
1028                  * if recovery is running, make sure it aborts.
1029                  */
1030                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1031         } else
1032                 set_bit(Faulty, &rdev->flags);
1033         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1034         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1035                 "raid1: Operation continuing on %d devices.\n",
1036                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1037 }
1038
1039 static void print_conf(conf_t *conf)
1040 {
1041         int i;
1042
1043         printk("RAID1 conf printout:\n");
1044         if (!conf) {
1045                 printk("(!conf)\n");
1046                 return;
1047         }
1048         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1049                 conf->raid_disks);
1050
1051         rcu_read_lock();
1052         for (i = 0; i < conf->raid_disks; i++) {
1053                 char b[BDEVNAME_SIZE];
1054                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1055                 if (rdev)
1056                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1057                                i, !test_bit(In_sync, &rdev->flags),
1058                                !test_bit(Faulty, &rdev->flags),
1059                                bdevname(rdev->bdev,b));
1060         }
1061         rcu_read_unlock();
1062 }
1063
1064 static void close_sync(conf_t *conf)
1065 {
1066         wait_barrier(conf);
1067         allow_barrier(conf);
1068
1069         mempool_destroy(conf->r1buf_pool);
1070         conf->r1buf_pool = NULL;
1071 }
1072
1073 static int raid1_spare_active(mddev_t *mddev)
1074 {
1075         int i;
1076         conf_t *conf = mddev->private;
1077
1078         /*
1079          * Find all failed disks within the RAID1 configuration
1080          * and mark them readable.
1081          * Called under mddev lock, so rcu protection not needed.
1082          */
1083         for (i = 0; i < conf->raid_disks; i++) {
1084                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1085                 if (rdev
1086                     && !test_bit(Faulty, &rdev->flags)
1087                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1088                         unsigned long flags;
1089                         spin_lock_irqsave(&conf->device_lock, flags);
1090                         mddev->degraded--;
1091                         spin_unlock_irqrestore(&conf->device_lock, flags);
1092                 }
1093         }
1094
1095         print_conf(conf);
1096         return 0;
1097 }
1098
1099
1100 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1101 {
1102         conf_t *conf = mddev->private;
1103         int err = -EEXIST;
1104         int mirror = 0;
1105         mirror_info_t *p;
1106         int first = 0;
1107         int last = mddev->raid_disks - 1;
1108
1109         if (rdev->raid_disk >= 0)
1110                 first = last = rdev->raid_disk;
1111
1112         for (mirror = first; mirror <= last; mirror++)
1113                 if ( !(p=conf->mirrors+mirror)->rdev) {
1114
1115                         blk_queue_stack_limits(mddev->queue,
1116                                                rdev->bdev->bd_disk->queue);
1117                         /* as we don't honour merge_bvec_fn, we must never risk
1118                          * violating it, so limit ->max_sector to one PAGE, as
1119                          * a one page request is never in violation.
1120                          */
1121                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1122                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1123                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1124
1125                         p->head_position = 0;
1126                         rdev->raid_disk = mirror;
1127                         err = 0;
1128                         /* As all devices are equivalent, we don't need a full recovery
1129                          * if this was recently any drive of the array
1130                          */
1131                         if (rdev->saved_raid_disk < 0)
1132                                 conf->fullsync = 1;
1133                         rcu_assign_pointer(p->rdev, rdev);
1134                         break;
1135                 }
1136
1137         print_conf(conf);
1138         return err;
1139 }
1140
1141 static int raid1_remove_disk(mddev_t *mddev, int number)
1142 {
1143         conf_t *conf = mddev->private;
1144         int err = 0;
1145         mdk_rdev_t *rdev;
1146         mirror_info_t *p = conf->mirrors+ number;
1147
1148         print_conf(conf);
1149         rdev = p->rdev;
1150         if (rdev) {
1151                 if (test_bit(In_sync, &rdev->flags) ||
1152                     atomic_read(&rdev->nr_pending)) {
1153                         err = -EBUSY;
1154                         goto abort;
1155                 }
1156                 /* Only remove non-faulty devices is recovery
1157                  * is not possible.
1158                  */
1159                 if (!test_bit(Faulty, &rdev->flags) &&
1160                     mddev->degraded < conf->raid_disks) {
1161                         err = -EBUSY;
1162                         goto abort;
1163                 }
1164                 p->rdev = NULL;
1165                 synchronize_rcu();
1166                 if (atomic_read(&rdev->nr_pending)) {
1167                         /* lost the race, try later */
1168                         err = -EBUSY;
1169                         p->rdev = rdev;
1170                 }
1171         }
1172 abort:
1173
1174         print_conf(conf);
1175         return err;
1176 }
1177
1178
1179 static void end_sync_read(struct bio *bio, int error)
1180 {
1181         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1182         int i;
1183
1184         for (i=r1_bio->mddev->raid_disks; i--; )
1185                 if (r1_bio->bios[i] == bio)
1186                         break;
1187         BUG_ON(i < 0);
1188         update_head_pos(i, r1_bio);
1189         /*
1190          * we have read a block, now it needs to be re-written,
1191          * or re-read if the read failed.
1192          * We don't do much here, just schedule handling by raid1d
1193          */
1194         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1195                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1196
1197         if (atomic_dec_and_test(&r1_bio->remaining))
1198                 reschedule_retry(r1_bio);
1199 }
1200
1201 static void end_sync_write(struct bio *bio, int error)
1202 {
1203         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1204         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1205         mddev_t *mddev = r1_bio->mddev;
1206         conf_t *conf = mddev_to_conf(mddev);
1207         int i;
1208         int mirror=0;
1209
1210         for (i = 0; i < conf->raid_disks; i++)
1211                 if (r1_bio->bios[i] == bio) {
1212                         mirror = i;
1213                         break;
1214                 }
1215         if (!uptodate) {
1216                 int sync_blocks = 0;
1217                 sector_t s = r1_bio->sector;
1218                 long sectors_to_go = r1_bio->sectors;
1219                 /* make sure these bits doesn't get cleared. */
1220                 do {
1221                         bitmap_end_sync(mddev->bitmap, s,
1222                                         &sync_blocks, 1);
1223                         s += sync_blocks;
1224                         sectors_to_go -= sync_blocks;
1225                 } while (sectors_to_go > 0);
1226                 md_error(mddev, conf->mirrors[mirror].rdev);
1227         }
1228
1229         update_head_pos(mirror, r1_bio);
1230
1231         if (atomic_dec_and_test(&r1_bio->remaining)) {
1232                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1233                 put_buf(r1_bio);
1234         }
1235 }
1236
1237 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1238 {
1239         conf_t *conf = mddev_to_conf(mddev);
1240         int i;
1241         int disks = conf->raid_disks;
1242         struct bio *bio, *wbio;
1243
1244         bio = r1_bio->bios[r1_bio->read_disk];
1245
1246
1247         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1248                 /* We have read all readable devices.  If we haven't
1249                  * got the block, then there is no hope left.
1250                  * If we have, then we want to do a comparison
1251                  * and skip the write if everything is the same.
1252                  * If any blocks failed to read, then we need to
1253                  * attempt an over-write
1254                  */
1255                 int primary;
1256                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1257                         for (i=0; i<mddev->raid_disks; i++)
1258                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1259                                         md_error(mddev, conf->mirrors[i].rdev);
1260
1261                         md_done_sync(mddev, r1_bio->sectors, 1);
1262                         put_buf(r1_bio);
1263                         return;
1264                 }
1265                 for (primary=0; primary<mddev->raid_disks; primary++)
1266                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1267                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1268                                 r1_bio->bios[primary]->bi_end_io = NULL;
1269                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1270                                 break;
1271                         }
1272                 r1_bio->read_disk = primary;
1273                 for (i=0; i<mddev->raid_disks; i++)
1274                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1275                                 int j;
1276                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1277                                 struct bio *pbio = r1_bio->bios[primary];
1278                                 struct bio *sbio = r1_bio->bios[i];
1279
1280                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1281                                         for (j = vcnt; j-- ; ) {
1282                                                 struct page *p, *s;
1283                                                 p = pbio->bi_io_vec[j].bv_page;
1284                                                 s = sbio->bi_io_vec[j].bv_page;
1285                                                 if (memcmp(page_address(p),
1286                                                            page_address(s),
1287                                                            PAGE_SIZE))
1288                                                         break;
1289                                         }
1290                                 } else
1291                                         j = 0;
1292                                 if (j >= 0)
1293                                         mddev->resync_mismatches += r1_bio->sectors;
1294                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1295                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1296                                         sbio->bi_end_io = NULL;
1297                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1298                                 } else {
1299                                         /* fixup the bio for reuse */
1300                                         int size;
1301                                         sbio->bi_vcnt = vcnt;
1302                                         sbio->bi_size = r1_bio->sectors << 9;
1303                                         sbio->bi_idx = 0;
1304                                         sbio->bi_phys_segments = 0;
1305                                         sbio->bi_hw_segments = 0;
1306                                         sbio->bi_hw_front_size = 0;
1307                                         sbio->bi_hw_back_size = 0;
1308                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1309                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1310                                         sbio->bi_next = NULL;
1311                                         sbio->bi_sector = r1_bio->sector +
1312                                                 conf->mirrors[i].rdev->data_offset;
1313                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1314                                         size = sbio->bi_size;
1315                                         for (j = 0; j < vcnt ; j++) {
1316                                                 struct bio_vec *bi;
1317                                                 bi = &sbio->bi_io_vec[j];
1318                                                 bi->bv_offset = 0;
1319                                                 if (size > PAGE_SIZE)
1320                                                         bi->bv_len = PAGE_SIZE;
1321                                                 else
1322                                                         bi->bv_len = size;
1323                                                 size -= PAGE_SIZE;
1324                                                 memcpy(page_address(bi->bv_page),
1325                                                        page_address(pbio->bi_io_vec[j].bv_page),
1326                                                        PAGE_SIZE);
1327                                         }
1328
1329                                 }
1330                         }
1331         }
1332         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1333                 /* ouch - failed to read all of that.
1334                  * Try some synchronous reads of other devices to get
1335                  * good data, much like with normal read errors.  Only
1336                  * read into the pages we already have so we don't
1337                  * need to re-issue the read request.
1338                  * We don't need to freeze the array, because being in an
1339                  * active sync request, there is no normal IO, and
1340                  * no overlapping syncs.
1341                  */
1342                 sector_t sect = r1_bio->sector;
1343                 int sectors = r1_bio->sectors;
1344                 int idx = 0;
1345
1346                 while(sectors) {
1347                         int s = sectors;
1348                         int d = r1_bio->read_disk;
1349                         int success = 0;
1350                         mdk_rdev_t *rdev;
1351
1352                         if (s > (PAGE_SIZE>>9))
1353                                 s = PAGE_SIZE >> 9;
1354                         do {
1355                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1356                                         /* No rcu protection needed here devices
1357                                          * can only be removed when no resync is
1358                                          * active, and resync is currently active
1359                                          */
1360                                         rdev = conf->mirrors[d].rdev;
1361                                         if (sync_page_io(rdev->bdev,
1362                                                          sect + rdev->data_offset,
1363                                                          s<<9,
1364                                                          bio->bi_io_vec[idx].bv_page,
1365                                                          READ)) {
1366                                                 success = 1;
1367                                                 break;
1368                                         }
1369                                 }
1370                                 d++;
1371                                 if (d == conf->raid_disks)
1372                                         d = 0;
1373                         } while (!success && d != r1_bio->read_disk);
1374
1375                         if (success) {
1376                                 int start = d;
1377                                 /* write it back and re-read */
1378                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1379                                 while (d != r1_bio->read_disk) {
1380                                         if (d == 0)
1381                                                 d = conf->raid_disks;
1382                                         d--;
1383                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1384                                                 continue;
1385                                         rdev = conf->mirrors[d].rdev;
1386                                         atomic_add(s, &rdev->corrected_errors);
1387                                         if (sync_page_io(rdev->bdev,
1388                                                          sect + rdev->data_offset,
1389                                                          s<<9,
1390                                                          bio->bi_io_vec[idx].bv_page,
1391                                                          WRITE) == 0)
1392                                                 md_error(mddev, rdev);
1393                                 }
1394                                 d = start;
1395                                 while (d != r1_bio->read_disk) {
1396                                         if (d == 0)
1397                                                 d = conf->raid_disks;
1398                                         d--;
1399                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1400                                                 continue;
1401                                         rdev = conf->mirrors[d].rdev;
1402                                         if (sync_page_io(rdev->bdev,
1403                                                          sect + rdev->data_offset,
1404                                                          s<<9,
1405                                                          bio->bi_io_vec[idx].bv_page,
1406                                                          READ) == 0)
1407                                                 md_error(mddev, rdev);
1408                                 }
1409                         } else {
1410                                 char b[BDEVNAME_SIZE];
1411                                 /* Cannot read from anywhere, array is toast */
1412                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1413                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1414                                        " for block %llu\n",
1415                                        bdevname(bio->bi_bdev,b),
1416                                        (unsigned long long)r1_bio->sector);
1417                                 md_done_sync(mddev, r1_bio->sectors, 0);
1418                                 put_buf(r1_bio);
1419                                 return;
1420                         }
1421                         sectors -= s;
1422                         sect += s;
1423                         idx ++;
1424                 }
1425         }
1426
1427         /*
1428          * schedule writes
1429          */
1430         atomic_set(&r1_bio->remaining, 1);
1431         for (i = 0; i < disks ; i++) {
1432                 wbio = r1_bio->bios[i];
1433                 if (wbio->bi_end_io == NULL ||
1434                     (wbio->bi_end_io == end_sync_read &&
1435                      (i == r1_bio->read_disk ||
1436                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1437                         continue;
1438
1439                 wbio->bi_rw = WRITE;
1440                 wbio->bi_end_io = end_sync_write;
1441                 atomic_inc(&r1_bio->remaining);
1442                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1443
1444                 generic_make_request(wbio);
1445         }
1446
1447         if (atomic_dec_and_test(&r1_bio->remaining)) {
1448                 /* if we're here, all write(s) have completed, so clean up */
1449                 md_done_sync(mddev, r1_bio->sectors, 1);
1450                 put_buf(r1_bio);
1451         }
1452 }
1453
1454 /*
1455  * This is a kernel thread which:
1456  *
1457  *      1.      Retries failed read operations on working mirrors.
1458  *      2.      Updates the raid superblock when problems encounter.
1459  *      3.      Performs writes following reads for array syncronising.
1460  */
1461
1462 static void fix_read_error(conf_t *conf, int read_disk,
1463                            sector_t sect, int sectors)
1464 {
1465         mddev_t *mddev = conf->mddev;
1466         while(sectors) {
1467                 int s = sectors;
1468                 int d = read_disk;
1469                 int success = 0;
1470                 int start;
1471                 mdk_rdev_t *rdev;
1472
1473                 if (s > (PAGE_SIZE>>9))
1474                         s = PAGE_SIZE >> 9;
1475
1476                 do {
1477                         /* Note: no rcu protection needed here
1478                          * as this is synchronous in the raid1d thread
1479                          * which is the thread that might remove
1480                          * a device.  If raid1d ever becomes multi-threaded....
1481                          */
1482                         rdev = conf->mirrors[d].rdev;
1483                         if (rdev &&
1484                             test_bit(In_sync, &rdev->flags) &&
1485                             sync_page_io(rdev->bdev,
1486                                          sect + rdev->data_offset,
1487                                          s<<9,
1488                                          conf->tmppage, READ))
1489                                 success = 1;
1490                         else {
1491                                 d++;
1492                                 if (d == conf->raid_disks)
1493                                         d = 0;
1494                         }
1495                 } while (!success && d != read_disk);
1496
1497                 if (!success) {
1498                         /* Cannot read from anywhere -- bye bye array */
1499                         md_error(mddev, conf->mirrors[read_disk].rdev);
1500                         break;
1501                 }
1502                 /* write it back and re-read */
1503                 start = d;
1504                 while (d != read_disk) {
1505                         if (d==0)
1506                                 d = conf->raid_disks;
1507                         d--;
1508                         rdev = conf->mirrors[d].rdev;
1509                         if (rdev &&
1510                             test_bit(In_sync, &rdev->flags)) {
1511                                 if (sync_page_io(rdev->bdev,
1512                                                  sect + rdev->data_offset,
1513                                                  s<<9, conf->tmppage, WRITE)
1514                                     == 0)
1515                                         /* Well, this device is dead */
1516                                         md_error(mddev, rdev);
1517                         }
1518                 }
1519                 d = start;
1520                 while (d != read_disk) {
1521                         char b[BDEVNAME_SIZE];
1522                         if (d==0)
1523                                 d = conf->raid_disks;
1524                         d--;
1525                         rdev = conf->mirrors[d].rdev;
1526                         if (rdev &&
1527                             test_bit(In_sync, &rdev->flags)) {
1528                                 if (sync_page_io(rdev->bdev,
1529                                                  sect + rdev->data_offset,
1530                                                  s<<9, conf->tmppage, READ)
1531                                     == 0)
1532                                         /* Well, this device is dead */
1533                                         md_error(mddev, rdev);
1534                                 else {
1535                                         atomic_add(s, &rdev->corrected_errors);
1536                                         printk(KERN_INFO
1537                                                "raid1:%s: read error corrected "
1538                                                "(%d sectors at %llu on %s)\n",
1539                                                mdname(mddev), s,
1540                                                (unsigned long long)(sect +
1541                                                    rdev->data_offset),
1542                                                bdevname(rdev->bdev, b));
1543                                 }
1544                         }
1545                 }
1546                 sectors -= s;
1547                 sect += s;
1548         }
1549 }
1550
1551 static void raid1d(mddev_t *mddev)
1552 {
1553         r1bio_t *r1_bio;
1554         struct bio *bio;
1555         unsigned long flags;
1556         conf_t *conf = mddev_to_conf(mddev);
1557         struct list_head *head = &conf->retry_list;
1558         int unplug=0;
1559         mdk_rdev_t *rdev;
1560
1561         md_check_recovery(mddev);
1562
1563         for (;;) {
1564                 char b[BDEVNAME_SIZE];
1565
1566                 unplug += flush_pending_writes(conf);
1567
1568                 spin_lock_irqsave(&conf->device_lock, flags);
1569                 if (list_empty(head)) {
1570                         spin_unlock_irqrestore(&conf->device_lock, flags);
1571                         break;
1572                 }
1573                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1574                 list_del(head->prev);
1575                 conf->nr_queued--;
1576                 spin_unlock_irqrestore(&conf->device_lock, flags);
1577
1578                 mddev = r1_bio->mddev;
1579                 conf = mddev_to_conf(mddev);
1580                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1581                         sync_request_write(mddev, r1_bio);
1582                         unplug = 1;
1583                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1584                         /* some requests in the r1bio were BIO_RW_BARRIER
1585                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1586                          * Better resubmit without the barrier.
1587                          * We know which devices to resubmit for, because
1588                          * all others have had their bios[] entry cleared.
1589                          * We already have a nr_pending reference on these rdevs.
1590                          */
1591                         int i;
1592                         const int do_sync = bio_sync(r1_bio->master_bio);
1593                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1594                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1595                         for (i=0; i < conf->raid_disks; i++)
1596                                 if (r1_bio->bios[i])
1597                                         atomic_inc(&r1_bio->remaining);
1598                         for (i=0; i < conf->raid_disks; i++)
1599                                 if (r1_bio->bios[i]) {
1600                                         struct bio_vec *bvec;
1601                                         int j;
1602
1603                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1604                                         /* copy pages from the failed bio, as
1605                                          * this might be a write-behind device */
1606                                         __bio_for_each_segment(bvec, bio, j, 0)
1607                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1608                                         bio_put(r1_bio->bios[i]);
1609                                         bio->bi_sector = r1_bio->sector +
1610                                                 conf->mirrors[i].rdev->data_offset;
1611                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1612                                         bio->bi_end_io = raid1_end_write_request;
1613                                         bio->bi_rw = WRITE | do_sync;
1614                                         bio->bi_private = r1_bio;
1615                                         r1_bio->bios[i] = bio;
1616                                         generic_make_request(bio);
1617                                 }
1618                 } else {
1619                         int disk;
1620
1621                         /* we got a read error. Maybe the drive is bad.  Maybe just
1622                          * the block and we can fix it.
1623                          * We freeze all other IO, and try reading the block from
1624                          * other devices.  When we find one, we re-write
1625                          * and check it that fixes the read error.
1626                          * This is all done synchronously while the array is
1627                          * frozen
1628                          */
1629                         if (mddev->ro == 0) {
1630                                 freeze_array(conf);
1631                                 fix_read_error(conf, r1_bio->read_disk,
1632                                                r1_bio->sector,
1633                                                r1_bio->sectors);
1634                                 unfreeze_array(conf);
1635                         }
1636
1637                         bio = r1_bio->bios[r1_bio->read_disk];
1638                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1639                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1640                                        " read error for block %llu\n",
1641                                        bdevname(bio->bi_bdev,b),
1642                                        (unsigned long long)r1_bio->sector);
1643                                 raid_end_bio_io(r1_bio);
1644                         } else {
1645                                 const int do_sync = bio_sync(r1_bio->master_bio);
1646                                 r1_bio->bios[r1_bio->read_disk] =
1647                                         mddev->ro ? IO_BLOCKED : NULL;
1648                                 r1_bio->read_disk = disk;
1649                                 bio_put(bio);
1650                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1651                                 r1_bio->bios[r1_bio->read_disk] = bio;
1652                                 rdev = conf->mirrors[disk].rdev;
1653                                 if (printk_ratelimit())
1654                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1655                                                " another mirror\n",
1656                                                bdevname(rdev->bdev,b),
1657                                                (unsigned long long)r1_bio->sector);
1658                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1659                                 bio->bi_bdev = rdev->bdev;
1660                                 bio->bi_end_io = raid1_end_read_request;
1661                                 bio->bi_rw = READ | do_sync;
1662                                 bio->bi_private = r1_bio;
1663                                 unplug = 1;
1664                                 generic_make_request(bio);
1665                         }
1666                 }
1667         }
1668         if (unplug)
1669                 unplug_slaves(mddev);
1670 }
1671
1672
1673 static int init_resync(conf_t *conf)
1674 {
1675         int buffs;
1676
1677         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1678         BUG_ON(conf->r1buf_pool);
1679         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1680                                           conf->poolinfo);
1681         if (!conf->r1buf_pool)
1682                 return -ENOMEM;
1683         conf->next_resync = 0;
1684         return 0;
1685 }
1686
1687 /*
1688  * perform a "sync" on one "block"
1689  *
1690  * We need to make sure that no normal I/O request - particularly write
1691  * requests - conflict with active sync requests.
1692  *
1693  * This is achieved by tracking pending requests and a 'barrier' concept
1694  * that can be installed to exclude normal IO requests.
1695  */
1696
1697 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1698 {
1699         conf_t *conf = mddev_to_conf(mddev);
1700         r1bio_t *r1_bio;
1701         struct bio *bio;
1702         sector_t max_sector, nr_sectors;
1703         int disk = -1;
1704         int i;
1705         int wonly = -1;
1706         int write_targets = 0, read_targets = 0;
1707         int sync_blocks;
1708         int still_degraded = 0;
1709
1710         if (!conf->r1buf_pool)
1711         {
1712 /*
1713                 printk("sync start - bitmap %p\n", mddev->bitmap);
1714 */
1715                 if (init_resync(conf))
1716                         return 0;
1717         }
1718
1719         max_sector = mddev->size << 1;
1720         if (sector_nr >= max_sector) {
1721                 /* If we aborted, we need to abort the
1722                  * sync on the 'current' bitmap chunk (there will
1723                  * only be one in raid1 resync.
1724                  * We can find the current addess in mddev->curr_resync
1725                  */
1726                 if (mddev->curr_resync < max_sector) /* aborted */
1727                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1728                                                 &sync_blocks, 1);
1729                 else /* completed sync */
1730                         conf->fullsync = 0;
1731
1732                 bitmap_close_sync(mddev->bitmap);
1733                 close_sync(conf);
1734                 return 0;
1735         }
1736
1737         if (mddev->bitmap == NULL &&
1738             mddev->recovery_cp == MaxSector &&
1739             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1740             conf->fullsync == 0) {
1741                 *skipped = 1;
1742                 return max_sector - sector_nr;
1743         }
1744         /* before building a request, check if we can skip these blocks..
1745          * This call the bitmap_start_sync doesn't actually record anything
1746          */
1747         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1748             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1749                 /* We can skip this block, and probably several more */
1750                 *skipped = 1;
1751                 return sync_blocks;
1752         }
1753         /*
1754          * If there is non-resync activity waiting for a turn,
1755          * and resync is going fast enough,
1756          * then let it though before starting on this new sync request.
1757          */
1758         if (!go_faster && conf->nr_waiting)
1759                 msleep_interruptible(1000);
1760
1761         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1762         raise_barrier(conf);
1763
1764         conf->next_resync = sector_nr;
1765
1766         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1767         rcu_read_lock();
1768         /*
1769          * If we get a correctably read error during resync or recovery,
1770          * we might want to read from a different device.  So we
1771          * flag all drives that could conceivably be read from for READ,
1772          * and any others (which will be non-In_sync devices) for WRITE.
1773          * If a read fails, we try reading from something else for which READ
1774          * is OK.
1775          */
1776
1777         r1_bio->mddev = mddev;
1778         r1_bio->sector = sector_nr;
1779         r1_bio->state = 0;
1780         set_bit(R1BIO_IsSync, &r1_bio->state);
1781
1782         for (i=0; i < conf->raid_disks; i++) {
1783                 mdk_rdev_t *rdev;
1784                 bio = r1_bio->bios[i];
1785
1786                 /* take from bio_init */
1787                 bio->bi_next = NULL;
1788                 bio->bi_flags |= 1 << BIO_UPTODATE;
1789                 bio->bi_rw = READ;
1790                 bio->bi_vcnt = 0;
1791                 bio->bi_idx = 0;
1792                 bio->bi_phys_segments = 0;
1793                 bio->bi_hw_segments = 0;
1794                 bio->bi_size = 0;
1795                 bio->bi_end_io = NULL;
1796                 bio->bi_private = NULL;
1797
1798                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1799                 if (rdev == NULL ||
1800                            test_bit(Faulty, &rdev->flags)) {
1801                         still_degraded = 1;
1802                         continue;
1803                 } else if (!test_bit(In_sync, &rdev->flags)) {
1804                         bio->bi_rw = WRITE;
1805                         bio->bi_end_io = end_sync_write;
1806                         write_targets ++;
1807                 } else {
1808                         /* may need to read from here */
1809                         bio->bi_rw = READ;
1810                         bio->bi_end_io = end_sync_read;
1811                         if (test_bit(WriteMostly, &rdev->flags)) {
1812                                 if (wonly < 0)
1813                                         wonly = i;
1814                         } else {
1815                                 if (disk < 0)
1816                                         disk = i;
1817                         }
1818                         read_targets++;
1819                 }
1820                 atomic_inc(&rdev->nr_pending);
1821                 bio->bi_sector = sector_nr + rdev->data_offset;
1822                 bio->bi_bdev = rdev->bdev;
1823                 bio->bi_private = r1_bio;
1824         }
1825         rcu_read_unlock();
1826         if (disk < 0)
1827                 disk = wonly;
1828         r1_bio->read_disk = disk;
1829
1830         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1831                 /* extra read targets are also write targets */
1832                 write_targets += read_targets-1;
1833
1834         if (write_targets == 0 || read_targets == 0) {
1835                 /* There is nowhere to write, so all non-sync
1836                  * drives must be failed - so we are finished
1837                  */
1838                 sector_t rv = max_sector - sector_nr;
1839                 *skipped = 1;
1840                 put_buf(r1_bio);
1841                 return rv;
1842         }
1843
1844         if (max_sector > mddev->resync_max)
1845                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1846         nr_sectors = 0;
1847         sync_blocks = 0;
1848         do {
1849                 struct page *page;
1850                 int len = PAGE_SIZE;
1851                 if (sector_nr + (len>>9) > max_sector)
1852                         len = (max_sector - sector_nr) << 9;
1853                 if (len == 0)
1854                         break;
1855                 if (sync_blocks == 0) {
1856                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1857                                                &sync_blocks, still_degraded) &&
1858                             !conf->fullsync &&
1859                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1860                                 break;
1861                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1862                         if (len > (sync_blocks<<9))
1863                                 len = sync_blocks<<9;
1864                 }
1865
1866                 for (i=0 ; i < conf->raid_disks; i++) {
1867                         bio = r1_bio->bios[i];
1868                         if (bio->bi_end_io) {
1869                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1870                                 if (bio_add_page(bio, page, len, 0) == 0) {
1871                                         /* stop here */
1872                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1873                                         while (i > 0) {
1874                                                 i--;
1875                                                 bio = r1_bio->bios[i];
1876                                                 if (bio->bi_end_io==NULL)
1877                                                         continue;
1878                                                 /* remove last page from this bio */
1879                                                 bio->bi_vcnt--;
1880                                                 bio->bi_size -= len;
1881                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1882                                         }
1883                                         goto bio_full;
1884                                 }
1885                         }
1886                 }
1887                 nr_sectors += len>>9;
1888                 sector_nr += len>>9;
1889                 sync_blocks -= (len>>9);
1890         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1891  bio_full:
1892         r1_bio->sectors = nr_sectors;
1893
1894         /* For a user-requested sync, we read all readable devices and do a
1895          * compare
1896          */
1897         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1898                 atomic_set(&r1_bio->remaining, read_targets);
1899                 for (i=0; i<conf->raid_disks; i++) {
1900                         bio = r1_bio->bios[i];
1901                         if (bio->bi_end_io == end_sync_read) {
1902                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1903                                 generic_make_request(bio);
1904                         }
1905                 }
1906         } else {
1907                 atomic_set(&r1_bio->remaining, 1);
1908                 bio = r1_bio->bios[r1_bio->read_disk];
1909                 md_sync_acct(bio->bi_bdev, nr_sectors);
1910                 generic_make_request(bio);
1911
1912         }
1913         return nr_sectors;
1914 }
1915
1916 static int run(mddev_t *mddev)
1917 {
1918         conf_t *conf;
1919         int i, j, disk_idx;
1920         mirror_info_t *disk;
1921         mdk_rdev_t *rdev;
1922         struct list_head *tmp;
1923
1924         if (mddev->level != 1) {
1925                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1926                        mdname(mddev), mddev->level);
1927                 goto out;
1928         }
1929         if (mddev->reshape_position != MaxSector) {
1930                 printk("raid1: %s: reshape_position set but not supported\n",
1931                        mdname(mddev));
1932                 goto out;
1933         }
1934         /*
1935          * copy the already verified devices into our private RAID1
1936          * bookkeeping area. [whatever we allocate in run(),
1937          * should be freed in stop()]
1938          */
1939         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1940         mddev->private = conf;
1941         if (!conf)
1942                 goto out_no_mem;
1943
1944         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1945                                  GFP_KERNEL);
1946         if (!conf->mirrors)
1947                 goto out_no_mem;
1948
1949         conf->tmppage = alloc_page(GFP_KERNEL);
1950         if (!conf->tmppage)
1951                 goto out_no_mem;
1952
1953         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1954         if (!conf->poolinfo)
1955                 goto out_no_mem;
1956         conf->poolinfo->mddev = mddev;
1957         conf->poolinfo->raid_disks = mddev->raid_disks;
1958         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1959                                           r1bio_pool_free,
1960                                           conf->poolinfo);
1961         if (!conf->r1bio_pool)
1962                 goto out_no_mem;
1963
1964         spin_lock_init(&conf->device_lock);
1965         mddev->queue->queue_lock = &conf->device_lock;
1966
1967         rdev_for_each(rdev, tmp, mddev) {
1968                 disk_idx = rdev->raid_disk;
1969                 if (disk_idx >= mddev->raid_disks
1970                     || disk_idx < 0)
1971                         continue;
1972                 disk = conf->mirrors + disk_idx;
1973
1974                 disk->rdev = rdev;
1975
1976                 blk_queue_stack_limits(mddev->queue,
1977                                        rdev->bdev->bd_disk->queue);
1978                 /* as we don't honour merge_bvec_fn, we must never risk
1979                  * violating it, so limit ->max_sector to one PAGE, as
1980                  * a one page request is never in violation.
1981                  */
1982                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1983                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1984                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1985
1986                 disk->head_position = 0;
1987         }
1988         conf->raid_disks = mddev->raid_disks;
1989         conf->mddev = mddev;
1990         INIT_LIST_HEAD(&conf->retry_list);
1991
1992         spin_lock_init(&conf->resync_lock);
1993         init_waitqueue_head(&conf->wait_barrier);
1994
1995         bio_list_init(&conf->pending_bio_list);
1996         bio_list_init(&conf->flushing_bio_list);
1997
1998
1999         mddev->degraded = 0;
2000         for (i = 0; i < conf->raid_disks; i++) {
2001
2002                 disk = conf->mirrors + i;
2003
2004                 if (!disk->rdev ||
2005                     !test_bit(In_sync, &disk->rdev->flags)) {
2006                         disk->head_position = 0;
2007                         mddev->degraded++;
2008                         if (disk->rdev)
2009                                 conf->fullsync = 1;
2010                 }
2011         }
2012         if (mddev->degraded == conf->raid_disks) {
2013                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2014                         mdname(mddev));
2015                 goto out_free_conf;
2016         }
2017         if (conf->raid_disks - mddev->degraded == 1)
2018                 mddev->recovery_cp = MaxSector;
2019
2020         /*
2021          * find the first working one and use it as a starting point
2022          * to read balancing.
2023          */
2024         for (j = 0; j < conf->raid_disks &&
2025                      (!conf->mirrors[j].rdev ||
2026                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
2027                 /* nothing */;
2028         conf->last_used = j;
2029
2030
2031         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
2032         if (!mddev->thread) {
2033                 printk(KERN_ERR
2034                        "raid1: couldn't allocate thread for %s\n",
2035                        mdname(mddev));
2036                 goto out_free_conf;
2037         }
2038
2039         printk(KERN_INFO
2040                 "raid1: raid set %s active with %d out of %d mirrors\n",
2041                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2042                 mddev->raid_disks);
2043         /*
2044          * Ok, everything is just fine now
2045          */
2046         mddev->array_size = mddev->size;
2047
2048         mddev->queue->unplug_fn = raid1_unplug;
2049         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2050         mddev->queue->backing_dev_info.congested_data = mddev;
2051
2052         return 0;
2053
2054 out_no_mem:
2055         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
2056                mdname(mddev));
2057
2058 out_free_conf:
2059         if (conf) {
2060                 if (conf->r1bio_pool)
2061                         mempool_destroy(conf->r1bio_pool);
2062                 kfree(conf->mirrors);
2063                 safe_put_page(conf->tmppage);
2064                 kfree(conf->poolinfo);
2065                 kfree(conf);
2066                 mddev->private = NULL;
2067         }
2068 out:
2069         return -EIO;
2070 }
2071
2072 static int stop(mddev_t *mddev)
2073 {
2074         conf_t *conf = mddev_to_conf(mddev);
2075         struct bitmap *bitmap = mddev->bitmap;
2076         int behind_wait = 0;
2077
2078         /* wait for behind writes to complete */
2079         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2080                 behind_wait++;
2081                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2082                 set_current_state(TASK_UNINTERRUPTIBLE);
2083                 schedule_timeout(HZ); /* wait a second */
2084                 /* need to kick something here to make sure I/O goes? */
2085         }
2086
2087         md_unregister_thread(mddev->thread);
2088         mddev->thread = NULL;
2089         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2090         if (conf->r1bio_pool)
2091                 mempool_destroy(conf->r1bio_pool);
2092         kfree(conf->mirrors);
2093         kfree(conf->poolinfo);
2094         kfree(conf);
2095         mddev->private = NULL;
2096         return 0;
2097 }
2098
2099 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2100 {
2101         /* no resync is happening, and there is enough space
2102          * on all devices, so we can resize.
2103          * We need to make sure resync covers any new space.
2104          * If the array is shrinking we should possibly wait until
2105          * any io in the removed space completes, but it hardly seems
2106          * worth it.
2107          */
2108         mddev->array_size = sectors>>1;
2109         set_capacity(mddev->gendisk, mddev->array_size << 1);
2110         mddev->changed = 1;
2111         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
2112                 mddev->recovery_cp = mddev->size << 1;
2113                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2114         }
2115         mddev->size = mddev->array_size;
2116         mddev->resync_max_sectors = sectors;
2117         return 0;
2118 }
2119
2120 static int raid1_reshape(mddev_t *mddev)
2121 {
2122         /* We need to:
2123          * 1/ resize the r1bio_pool
2124          * 2/ resize conf->mirrors
2125          *
2126          * We allocate a new r1bio_pool if we can.
2127          * Then raise a device barrier and wait until all IO stops.
2128          * Then resize conf->mirrors and swap in the new r1bio pool.
2129          *
2130          * At the same time, we "pack" the devices so that all the missing
2131          * devices have the higher raid_disk numbers.
2132          */
2133         mempool_t *newpool, *oldpool;
2134         struct pool_info *newpoolinfo;
2135         mirror_info_t *newmirrors;
2136         conf_t *conf = mddev_to_conf(mddev);
2137         int cnt, raid_disks;
2138         unsigned long flags;
2139         int d, d2, err;
2140
2141         /* Cannot change chunk_size, layout, or level */
2142         if (mddev->chunk_size != mddev->new_chunk ||
2143             mddev->layout != mddev->new_layout ||
2144             mddev->level != mddev->new_level) {
2145                 mddev->new_chunk = mddev->chunk_size;
2146                 mddev->new_layout = mddev->layout;
2147                 mddev->new_level = mddev->level;
2148                 return -EINVAL;
2149         }
2150
2151         err = md_allow_write(mddev);
2152         if (err)
2153                 return err;
2154
2155         raid_disks = mddev->raid_disks + mddev->delta_disks;
2156
2157         if (raid_disks < conf->raid_disks) {
2158                 cnt=0;
2159                 for (d= 0; d < conf->raid_disks; d++)
2160                         if (conf->mirrors[d].rdev)
2161                                 cnt++;
2162                 if (cnt > raid_disks)
2163                         return -EBUSY;
2164         }
2165
2166         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2167         if (!newpoolinfo)
2168                 return -ENOMEM;
2169         newpoolinfo->mddev = mddev;
2170         newpoolinfo->raid_disks = raid_disks;
2171
2172         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2173                                  r1bio_pool_free, newpoolinfo);
2174         if (!newpool) {
2175                 kfree(newpoolinfo);
2176                 return -ENOMEM;
2177         }
2178         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2179         if (!newmirrors) {
2180                 kfree(newpoolinfo);
2181                 mempool_destroy(newpool);
2182                 return -ENOMEM;
2183         }
2184
2185         raise_barrier(conf);
2186
2187         /* ok, everything is stopped */
2188         oldpool = conf->r1bio_pool;
2189         conf->r1bio_pool = newpool;
2190
2191         for (d = d2 = 0; d < conf->raid_disks; d++) {
2192                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2193                 if (rdev && rdev->raid_disk != d2) {
2194                         char nm[20];
2195                         sprintf(nm, "rd%d", rdev->raid_disk);
2196                         sysfs_remove_link(&mddev->kobj, nm);
2197                         rdev->raid_disk = d2;
2198                         sprintf(nm, "rd%d", rdev->raid_disk);
2199                         sysfs_remove_link(&mddev->kobj, nm);
2200                         if (sysfs_create_link(&mddev->kobj,
2201                                               &rdev->kobj, nm))
2202                                 printk(KERN_WARNING
2203                                        "md/raid1: cannot register "
2204                                        "%s for %s\n",
2205                                        nm, mdname(mddev));
2206                 }
2207                 if (rdev)
2208                         newmirrors[d2++].rdev = rdev;
2209         }
2210         kfree(conf->mirrors);
2211         conf->mirrors = newmirrors;
2212         kfree(conf->poolinfo);
2213         conf->poolinfo = newpoolinfo;
2214
2215         spin_lock_irqsave(&conf->device_lock, flags);
2216         mddev->degraded += (raid_disks - conf->raid_disks);
2217         spin_unlock_irqrestore(&conf->device_lock, flags);
2218         conf->raid_disks = mddev->raid_disks = raid_disks;
2219         mddev->delta_disks = 0;
2220
2221         conf->last_used = 0; /* just make sure it is in-range */
2222         lower_barrier(conf);
2223
2224         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2225         md_wakeup_thread(mddev->thread);
2226
2227         mempool_destroy(oldpool);
2228         return 0;
2229 }
2230
2231 static void raid1_quiesce(mddev_t *mddev, int state)
2232 {
2233         conf_t *conf = mddev_to_conf(mddev);
2234
2235         switch(state) {
2236         case 1:
2237                 raise_barrier(conf);
2238                 break;
2239         case 0:
2240                 lower_barrier(conf);
2241                 break;
2242         }
2243 }
2244
2245
2246 static struct mdk_personality raid1_personality =
2247 {
2248         .name           = "raid1",
2249         .level          = 1,
2250         .owner          = THIS_MODULE,
2251         .make_request   = make_request,
2252         .run            = run,
2253         .stop           = stop,
2254         .status         = status,
2255         .error_handler  = error,
2256         .hot_add_disk   = raid1_add_disk,
2257         .hot_remove_disk= raid1_remove_disk,
2258         .spare_active   = raid1_spare_active,
2259         .sync_request   = sync_request,
2260         .resize         = raid1_resize,
2261         .check_reshape  = raid1_reshape,
2262         .quiesce        = raid1_quiesce,
2263 };
2264
2265 static int __init raid_init(void)
2266 {
2267         return register_md_personality(&raid1_personality);
2268 }
2269
2270 static void raid_exit(void)
2271 {
2272         unregister_md_personality(&raid1_personality);
2273 }
2274
2275 module_init(raid_init);
2276 module_exit(raid_exit);
2277 MODULE_LICENSE("GPL");
2278 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2279 MODULE_ALIAS("md-raid1");
2280 MODULE_ALIAS("md-level-1");