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                         if (r_queue->unplug_fn)
 553                                 r_queue->unplug_fn(r_queue);
 554
 555                         rdev_dec_pending(rdev, mddev);
 556                         rcu_read_lock();
 557                 }
 558         }
 559         rcu_read_unlock();
 560 }
 561
 562 static void raid1_unplug(struct request_queue *q)
 563 {
 564         mddev_t *mddev = q->queuedata;
 565
 566         unplug_slaves(mddev);
 567         md_wakeup_thread(mddev->thread);
 568 }
 569
 570 static int raid1_congested(void *data, int bits)
 571 {
 572         mddev_t *mddev = data;
 573         conf_t *conf = mddev_to_conf(mddev);
 574         int i, ret = 0;
 575
 576         rcu_read_lock();
 577         for (i = 0; i < mddev->raid_disks; i++) {
 578                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 579                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 580                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 581
 582                         /* Note the '|| 1' - when read_balance prefers
 583                          * non-congested targets, it can be removed
 584                          */
 585                         if ((bits & (1<<BDI_write_congested)) || 1)
 586                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 587                         else
 588                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 589                 }
 590         }
 591         rcu_read_unlock();
 592         return ret;
 593 }
 594
 595
 596 /* Barriers....
 597  * Sometimes we need to suspend IO while we do something else,
 598  * either some resync/recovery, or reconfigure the array.
 599  * To do this we raise a 'barrier'.
 600  * The 'barrier' is a counter that can be raised multiple times
 601  * to count how many activities are happening which preclude
 602  * normal IO.
 603  * We can only raise the barrier if there is no pending IO.
 604  * i.e. if nr_pending == 0.
 605  * We choose only to raise the barrier if no-one is waiting for the
 606  * barrier to go down.  This means that as soon as an IO request
 607  * is ready, no other operations which require a barrier will start
 608  * until the IO request has had a chance.
 609  *
 610  * So: regular IO calls 'wait_barrier'.  When that returns there
 611  *    is no backgroup IO happening,  It must arrange to call
 612  *    allow_barrier when it has finished its IO.
 613  * backgroup IO calls must call raise_barrier.  Once that returns
 614  *    there is no normal IO happeing.  It must arrange to call
 615  *    lower_barrier when the particular background IO completes.
 616  */
 617 #define RESYNC_DEPTH 32
 618
 619 static void raise_barrier(conf_t *conf)
 620 {
 621         spin_lock_irq(&conf->resync_lock);
 622
 623         /* Wait until no block IO is waiting */
 624         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 625                             conf->resync_lock,
 626                             raid1_unplug(conf->mddev->queue));
 627
 628         /* block any new IO from starting */
 629         conf->barrier++;
 630
 631         /* No wait for all pending IO to complete */
 632         wait_event_lock_irq(conf->wait_barrier,
 633                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 634                             conf->resync_lock,
 635                             raid1_unplug(conf->mddev->queue));
 636
 637         spin_unlock_irq(&conf->resync_lock);
 638 }
 639
 640 static void lower_barrier(conf_t *conf)
 641 {
 642         unsigned long flags;
 643         spin_lock_irqsave(&conf->resync_lock, flags);
 644         conf->barrier--;
 645         spin_unlock_irqrestore(&conf->resync_lock, flags);
 646         wake_up(&conf->wait_barrier);
 647 }
 648
 649 static void wait_barrier(conf_t *conf)
 650 {
 651         spin_lock_irq(&conf->resync_lock);
 652         if (conf->barrier) {
 653                 conf->nr_waiting++;
 654                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 655                                     conf->resync_lock,
 656                                     raid1_unplug(conf->mddev->queue));
 657                 conf->nr_waiting--;
 658         }
 659         conf->nr_pending++;
 660         spin_unlock_irq(&conf->resync_lock);
 661 }
 662
 663 static void allow_barrier(conf_t *conf)
 664 {
 665         unsigned long flags;
 666         spin_lock_irqsave(&conf->resync_lock, flags);
 667         conf->nr_pending--;
 668         spin_unlock_irqrestore(&conf->resync_lock, flags);
 669         wake_up(&conf->wait_barrier);
 670 }
 671
 672 static void freeze_array(conf_t *conf)
 673 {
 674         /* stop syncio and normal IO and wait for everything to
 675          * go quite.
 676          * We increment barrier and nr_waiting, and then
 677          * wait until barrier+nr_pending match nr_queued+2
 678          */
 679         spin_lock_irq(&conf->resync_lock);
 680         conf->barrier++;
 681         conf->nr_waiting++;
 682         wait_event_lock_irq(conf->wait_barrier,
 683                             conf->barrier+conf->nr_pending == conf->nr_queued+2,
 684                             conf->resync_lock,
 685                             raid1_unplug(conf->mddev->queue));
 686         spin_unlock_irq(&conf->resync_lock);
 687 }
 688 static void unfreeze_array(conf_t *conf)
 689 {
 690         /* reverse the effect of the freeze */
 691         spin_lock_irq(&conf->resync_lock);
 692         conf->barrier--;
 693         conf->nr_waiting--;
 694         wake_up(&conf->wait_barrier);
 695         spin_unlock_irq(&conf->resync_lock);
 696 }
 697
 698
 699 /* duplicate the data pages for behind I/O */
 700 static struct page **alloc_behind_pages(struct bio *bio)
 701 {
 702         int i;
 703         struct bio_vec *bvec;
 704         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 705                                         GFP_NOIO);
 706         if (unlikely(!pages))
 707                 goto do_sync_io;
 708
 709         bio_for_each_segment(bvec, bio, i) {
 710                 pages[i] = alloc_page(GFP_NOIO);
 711                 if (unlikely(!pages[i]))
 712                         goto do_sync_io;
 713                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 714                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 715                 kunmap(pages[i]);
 716                 kunmap(bvec->bv_page);
 717         }
 718
 719         return pages;
 720
 721 do_sync_io:
 722         if (pages)
 723                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 724                         put_page(pages[i]);
 725         kfree(pages);
 726         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 727         return NULL;
 728 }
 729
 730 static int make_request(struct request_queue *q, struct bio * bio)
 731 {
 732         mddev_t *mddev = q->queuedata;
 733         conf_t *conf = mddev_to_conf(mddev);
 734         mirror_info_t *mirror;
 735         r1bio_t *r1_bio;
 736         struct bio *read_bio;
 737         int i, targets = 0, disks;
 738         mdk_rdev_t *rdev;
 739         struct bitmap *bitmap = mddev->bitmap;
 740         unsigned long flags;
 741         struct bio_list bl;
 742         struct page **behind_pages = NULL;
 743         const int rw = bio_data_dir(bio);
 744         const int do_sync = bio_sync(bio);
 745         int do_barriers;
 746
 747         /*
 748          * Register the new request and wait if the reconstruction
 749          * thread has put up a bar for new requests.
 750          * Continue immediately if no resync is active currently.
 751          * We test barriers_work *after* md_write_start as md_write_start
 752          * may cause the first superblock write, and that will check out
 753          * if barriers work.
 754          */
 755
 756         md_write_start(mddev, bio); /* wait on superblock update early */
 757
 758         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
 759                 if (rw == WRITE)
 760                         md_write_end(mddev);
 761                 bio_endio(bio, -EOPNOTSUPP);
 762                 return 0;
 763         }
 764
 765         wait_barrier(conf);
 766
 767         disk_stat_inc(mddev->gendisk, ios[rw]);
 768         disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
 769
 770         /*
 771          * make_request() can abort the operation when READA is being
 772          * used and no empty request is available.
 773          *
 774          */
 775         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 776
 777         r1_bio->master_bio = bio;
 778         r1_bio->sectors = bio->bi_size >> 9;
 779         r1_bio->state = 0;
 780         r1_bio->mddev = mddev;
 781         r1_bio->sector = bio->bi_sector;
 782
 783         if (rw == READ) {
 784                 /*
 785                  * read balancing logic:
 786                  */
 787                 int rdisk = read_balance(conf, r1_bio);
 788
 789                 if (rdisk < 0) {
 790                         /* couldn't find anywhere to read from */
 791                         raid_end_bio_io(r1_bio);
 792                         return 0;
 793                 }
 794                 mirror = conf->mirrors + rdisk;
 795
 796                 r1_bio->read_disk = rdisk;
 797
 798                 read_bio = bio_clone(bio, GFP_NOIO);
 799
 800                 r1_bio->bios[rdisk] = read_bio;
 801
 802                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 803                 read_bio->bi_bdev = mirror->rdev->bdev;
 804                 read_bio->bi_end_io = raid1_end_read_request;
 805                 read_bio->bi_rw = READ | do_sync;
 806                 read_bio->bi_private = r1_bio;
 807
 808                 generic_make_request(read_bio);
 809                 return 0;
 810         }
 811
 812         /*
 813          * WRITE:
 814          */
 815         /* first select target devices under spinlock and
 816          * inc refcount on their rdev.  Record them by setting
 817          * bios[x] to bio
 818          */
 819         disks = conf->raid_disks;
 820 #if 0
 821         { static int first=1;
 822         if (first) printk("First Write sector %llu disks %d\n",
 823                           (unsigned long long)r1_bio->sector, disks);
 824         first = 0;
 825         }
 826 #endif
 827         rcu_read_lock();
 828         for (i = 0;  i < disks; i++) {
 829                 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
 830                     !test_bit(Faulty, &rdev->flags)) {
 831                         atomic_inc(&rdev->nr_pending);
 832                         if (test_bit(Faulty, &rdev->flags)) {
 833                                 rdev_dec_pending(rdev, mddev);
 834                                 r1_bio->bios[i] = NULL;
 835                         } else
 836                                 r1_bio->bios[i] = bio;
 837                         targets++;
 838                 } else
 839                         r1_bio->bios[i] = NULL;
 840         }
 841         rcu_read_unlock();
 842
 843         BUG_ON(targets == 0); /* we never fail the last device */
 844
 845         if (targets < conf->raid_disks) {
 846                 /* array is degraded, we will not clear the bitmap
 847                  * on I/O completion (see raid1_end_write_request) */
 848                 set_bit(R1BIO_Degraded, &r1_bio->state);
 849         }
 850
 851         /* do behind I/O ? */
 852         if (bitmap &&
 853             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
 854             (behind_pages = alloc_behind_pages(bio)) != NULL)
 855                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 856
 857         atomic_set(&r1_bio->remaining, 0);
 858         atomic_set(&r1_bio->behind_remaining, 0);
 859
 860         do_barriers = bio_barrier(bio);
 861         if (do_barriers)
 862                 set_bit(R1BIO_Barrier, &r1_bio->state);
 863
 864         bio_list_init(&bl);
 865         for (i = 0; i < disks; i++) {
 866                 struct bio *mbio;
 867                 if (!r1_bio->bios[i])
 868                         continue;
 869
 870                 mbio = bio_clone(bio, GFP_NOIO);
 871                 r1_bio->bios[i] = mbio;
 872
 873                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 874                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 875                 mbio->bi_end_io = raid1_end_write_request;
 876                 mbio->bi_rw = WRITE | do_barriers | do_sync;
 877                 mbio->bi_private = r1_bio;
 878
 879                 if (behind_pages) {
 880                         struct bio_vec *bvec;
 881                         int j;
 882
 883                         /* Yes, I really want the '__' version so that
 884                          * we clear any unused pointer in the io_vec, rather
 885                          * than leave them unchanged.  This is important
 886                          * because when we come to free the pages, we won't
 887                          * know the originial bi_idx, so we just free
 888                          * them all
 889                          */
 890                         __bio_for_each_segment(bvec, mbio, j, 0)
 891                                 bvec->bv_page = behind_pages[j];
 892                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 893                                 atomic_inc(&r1_bio->behind_remaining);
 894                 }
 895
 896                 atomic_inc(&r1_bio->remaining);
 897
 898                 bio_list_add(&bl, mbio);
 899         }
 900         kfree(behind_pages); /* the behind pages are attached to the bios now */
 901
 902         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 903                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 904         spin_lock_irqsave(&conf->device_lock, flags);
 905         bio_list_merge(&conf->pending_bio_list, &bl);
 906         bio_list_init(&bl);
 907
 908         blk_plug_device(mddev->queue);
 909         spin_unlock_irqrestore(&conf->device_lock, flags);
 910
 911         if (do_sync)
 912                 md_wakeup_thread(mddev->thread);
 913 #if 0
 914         while ((bio = bio_list_pop(&bl)) != NULL)
 915                 generic_make_request(bio);
 916 #endif
 917
 918         return 0;
 919 }
 920
 921 static void status(struct seq_file *seq, mddev_t *mddev)
 922 {
 923         conf_t *conf = mddev_to_conf(mddev);
 924         int i;
 925
 926         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 927                    conf->raid_disks - mddev->degraded);
 928         rcu_read_lock();
 929         for (i = 0; i < conf->raid_disks; i++) {
 930                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 931                 seq_printf(seq, "%s",
 932                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
 933         }
 934         rcu_read_unlock();
 935         seq_printf(seq, "]");
 936 }
 937
 938
 939 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 940 {
 941         char b[BDEVNAME_SIZE];
 942         conf_t *conf = mddev_to_conf(mddev);
 943
 944         /*
 945          * If it is not operational, then we have already marked it as dead
 946          * else if it is the last working disks, ignore the error, let the
 947          * next level up know.
 948          * else mark the drive as failed
 949          */
 950         if (test_bit(In_sync, &rdev->flags)
 951             && (conf->raid_disks - mddev->degraded) == 1)
 952                 /*
 953                  * Don't fail the drive, act as though we were just a
 954                  * normal single drive
 955                  */
 956                 return;
 957         if (test_and_clear_bit(In_sync, &rdev->flags)) {
 958                 unsigned long flags;
 959                 spin_lock_irqsave(&conf->device_lock, flags);
 960                 mddev->degraded++;
 961                 set_bit(Faulty, &rdev->flags);
 962                 spin_unlock_irqrestore(&conf->device_lock, flags);
 963                 /*
 964                  * if recovery is running, make sure it aborts.
 965                  */
 966                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 967         } else
 968                 set_bit(Faulty, &rdev->flags);
 969         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 970         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
 971                 "       Operation continuing on %d devices\n",
 972                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
 973 }
 974
 975 static void print_conf(conf_t *conf)
 976 {
 977         int i;
 978
 979         printk("RAID1 conf printout:\n");
 980         if (!conf) {
 981                 printk("(!conf)\n");
 982                 return;
 983         }
 984         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
 985                 conf->raid_disks);
 986
 987         rcu_read_lock();
 988         for (i = 0; i < conf->raid_disks; i++) {
 989                 char b[BDEVNAME_SIZE];
 990                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 991                 if (rdev)
 992                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 993                                i, !test_bit(In_sync, &rdev->flags),
 994                                !test_bit(Faulty, &rdev->flags),
 995                                bdevname(rdev->bdev,b));
 996         }
 997         rcu_read_unlock();
 998 }
 999
1000 static void close_sync(conf_t *conf)
1001 {
1002         wait_barrier(conf);
1003         allow_barrier(conf);
1004
1005         mempool_destroy(conf->r1buf_pool);
1006         conf->r1buf_pool = NULL;
1007 }
1008
1009 static int raid1_spare_active(mddev_t *mddev)
1010 {
1011         int i;
1012         conf_t *conf = mddev->private;
1013
1014         /*
1015          * Find all failed disks within the RAID1 configuration
1016          * and mark them readable.
1017          * Called under mddev lock, so rcu protection not needed.
1018          */
1019         for (i = 0; i < conf->raid_disks; i++) {
1020                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1021                 if (rdev
1022                     && !test_bit(Faulty, &rdev->flags)
1023                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1024                         unsigned long flags;
1025                         spin_lock_irqsave(&conf->device_lock, flags);
1026                         mddev->degraded--;
1027                         spin_unlock_irqrestore(&conf->device_lock, flags);
1028                 }
1029         }
1030
1031         print_conf(conf);
1032         return 0;
1033 }
1034
1035
1036 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1037 {
1038         conf_t *conf = mddev->private;
1039         int found = 0;
1040         int mirror = 0;
1041         mirror_info_t *p;
1042
1043         for (mirror=0; mirror < mddev->raid_disks; mirror++)
1044                 if ( !(p=conf->mirrors+mirror)->rdev) {
1045
1046                         blk_queue_stack_limits(mddev->queue,
1047                                                rdev->bdev->bd_disk->queue);
1048                         /* as we don't honour merge_bvec_fn, we must never risk
1049                          * violating it, so limit ->max_sector to one PAGE, as
1050                          * a one page request is never in violation.
1051                          */
1052                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1053                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
1054                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1055
1056                         p->head_position = 0;
1057                         rdev->raid_disk = mirror;
1058                         found = 1;
1059                         /* As all devices are equivalent, we don't need a full recovery
1060                          * if this was recently any drive of the array
1061                          */
1062                         if (rdev->saved_raid_disk < 0)
1063                                 conf->fullsync = 1;
1064                         rcu_assign_pointer(p->rdev, rdev);
1065                         break;
1066                 }
1067
1068         print_conf(conf);
1069         return found;
1070 }
1071
1072 static int raid1_remove_disk(mddev_t *mddev, int number)
1073 {
1074         conf_t *conf = mddev->private;
1075         int err = 0;
1076         mdk_rdev_t *rdev;
1077         mirror_info_t *p = conf->mirrors+ number;
1078
1079         print_conf(conf);
1080         rdev = p->rdev;
1081         if (rdev) {
1082                 if (test_bit(In_sync, &rdev->flags) ||
1083                     atomic_read(&rdev->nr_pending)) {
1084                         err = -EBUSY;
1085                         goto abort;
1086                 }
1087                 p->rdev = NULL;
1088                 synchronize_rcu();
1089                 if (atomic_read(&rdev->nr_pending)) {
1090                         /* lost the race, try later */
1091                         err = -EBUSY;
1092                         p->rdev = rdev;
1093                 }
1094         }
1095 abort:
1096
1097         print_conf(conf);
1098         return err;
1099 }
1100
1101
1102 static void end_sync_read(struct bio *bio, int error)
1103 {
1104         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1105         int i;
1106
1107         for (i=r1_bio->mddev->raid_disks; i--; )
1108                 if (r1_bio->bios[i] == bio)
1109                         break;
1110         BUG_ON(i < 0);
1111         update_head_pos(i, r1_bio);
1112         /*
1113          * we have read a block, now it needs to be re-written,
1114          * or re-read if the read failed.
1115          * We don't do much here, just schedule handling by raid1d
1116          */
1117         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1118                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1119
1120         if (atomic_dec_and_test(&r1_bio->remaining))
1121                 reschedule_retry(r1_bio);
1122 }
1123
1124 static void end_sync_write(struct bio *bio, int error)
1125 {
1126         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1127         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1128         mddev_t *mddev = r1_bio->mddev;
1129         conf_t *conf = mddev_to_conf(mddev);
1130         int i;
1131         int mirror=0;
1132
1133         for (i = 0; i < conf->raid_disks; i++)
1134                 if (r1_bio->bios[i] == bio) {
1135                         mirror = i;
1136                         break;
1137                 }
1138         if (!uptodate) {
1139                 int sync_blocks = 0;
1140                 sector_t s = r1_bio->sector;
1141                 long sectors_to_go = r1_bio->sectors;
1142                 /* make sure these bits doesn't get cleared. */
1143                 do {
1144                         bitmap_end_sync(mddev->bitmap, s,
1145                                         &sync_blocks, 1);
1146                         s += sync_blocks;
1147                         sectors_to_go -= sync_blocks;
1148                 } while (sectors_to_go > 0);
1149                 md_error(mddev, conf->mirrors[mirror].rdev);
1150         }
1151
1152         update_head_pos(mirror, r1_bio);
1153
1154         if (atomic_dec_and_test(&r1_bio->remaining)) {
1155                 md_done_sync(mddev, r1_bio->sectors, uptodate);
1156                 put_buf(r1_bio);
1157         }
1158 }
1159
1160 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1161 {
1162         conf_t *conf = mddev_to_conf(mddev);
1163         int i;
1164         int disks = conf->raid_disks;
1165         struct bio *bio, *wbio;
1166
1167         bio = r1_bio->bios[r1_bio->read_disk];
1168
1169
1170         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1171                 /* We have read all readable devices.  If we haven't
1172                  * got the block, then there is no hope left.
1173                  * If we have, then we want to do a comparison
1174                  * and skip the write if everything is the same.
1175                  * If any blocks failed to read, then we need to
1176                  * attempt an over-write
1177                  */
1178                 int primary;
1179                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1180                         for (i=0; i<mddev->raid_disks; i++)
1181                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1182                                         md_error(mddev, conf->mirrors[i].rdev);
1183
1184                         md_done_sync(mddev, r1_bio->sectors, 1);
1185                         put_buf(r1_bio);
1186                         return;
1187                 }
1188                 for (primary=0; primary<mddev->raid_disks; primary++)
1189                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1190                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1191                                 r1_bio->bios[primary]->bi_end_io = NULL;
1192                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1193                                 break;
1194                         }
1195                 r1_bio->read_disk = primary;
1196                 for (i=0; i<mddev->raid_disks; i++)
1197                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1198                                 int j;
1199                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1200                                 struct bio *pbio = r1_bio->bios[primary];
1201                                 struct bio *sbio = r1_bio->bios[i];
1202
1203                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1204                                         for (j = vcnt; j-- ; ) {
1205                                                 struct page *p, *s;
1206                                                 p = pbio->bi_io_vec[j].bv_page;
1207                                                 s = sbio->bi_io_vec[j].bv_page;
1208                                                 if (memcmp(page_address(p),
1209                                                            page_address(s),
1210                                                            PAGE_SIZE))
1211                                                         break;
1212                                         }
1213                                 } else
1214                                         j = 0;
1215                                 if (j >= 0)
1216                                         mddev->resync_mismatches += r1_bio->sectors;
1217                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1218                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1219                                         sbio->bi_end_io = NULL;
1220                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1221                                 } else {
1222                                         /* fixup the bio for reuse */
1223                                         sbio->bi_vcnt = vcnt;
1224                                         sbio->bi_size = r1_bio->sectors << 9;
1225                                         sbio->bi_idx = 0;
1226                                         sbio->bi_phys_segments = 0;
1227                                         sbio->bi_hw_segments = 0;
1228                                         sbio->bi_hw_front_size = 0;
1229                                         sbio->bi_hw_back_size = 0;
1230                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1231                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1232                                         sbio->bi_next = NULL;
1233                                         sbio->bi_sector = r1_bio->sector +
1234                                                 conf->mirrors[i].rdev->data_offset;
1235                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1236                                         for (j = 0; j < vcnt ; j++)
1237                                                 memcpy(page_address(sbio->bi_io_vec[j].bv_page),
1238                                                        page_address(pbio->bi_io_vec[j].bv_page),
1239                                                        PAGE_SIZE);
1240
1241                                 }
1242                         }
1243         }
1244         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1245                 /* ouch - failed to read all of that.
1246                  * Try some synchronous reads of other devices to get
1247                  * good data, much like with normal read errors.  Only
1248                  * read into the pages we already have so we don't
1249                  * need to re-issue the read request.
1250                  * We don't need to freeze the array, because being in an
1251                  * active sync request, there is no normal IO, and
1252                  * no overlapping syncs.
1253                  */
1254                 sector_t sect = r1_bio->sector;
1255                 int sectors = r1_bio->sectors;
1256                 int idx = 0;
1257
1258                 while(sectors) {
1259                         int s = sectors;
1260                         int d = r1_bio->read_disk;
1261                         int success = 0;
1262                         mdk_rdev_t *rdev;
1263
1264                         if (s > (PAGE_SIZE>>9))
1265                                 s = PAGE_SIZE >> 9;
1266                         do {
1267                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1268                                         /* No rcu protection needed here devices
1269                                          * can only be removed when no resync is
1270                                          * active, and resync is currently active
1271                                          */
1272                                         rdev = conf->mirrors[d].rdev;
1273                                         if (sync_page_io(rdev->bdev,
1274                                                          sect + rdev->data_offset,
1275                                                          s<<9,
1276                                                          bio->bi_io_vec[idx].bv_page,
1277                                                          READ)) {
1278                                                 success = 1;
1279                                                 break;
1280                                         }
1281                                 }
1282                                 d++;
1283                                 if (d == conf->raid_disks)
1284                                         d = 0;
1285                         } while (!success && d != r1_bio->read_disk);
1286
1287                         if (success) {
1288                                 int start = d;
1289                                 /* write it back and re-read */
1290                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1291                                 while (d != r1_bio->read_disk) {
1292                                         if (d == 0)
1293                                                 d = conf->raid_disks;
1294                                         d--;
1295                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1296                                                 continue;
1297                                         rdev = conf->mirrors[d].rdev;
1298                                         atomic_add(s, &rdev->corrected_errors);
1299                                         if (sync_page_io(rdev->bdev,
1300                                                          sect + rdev->data_offset,
1301                                                          s<<9,
1302                                                          bio->bi_io_vec[idx].bv_page,
1303                                                          WRITE) == 0)
1304                                                 md_error(mddev, rdev);
1305                                 }
1306                                 d = start;
1307                                 while (d != r1_bio->read_disk) {
1308                                         if (d == 0)
1309                                                 d = conf->raid_disks;
1310                                         d--;
1311                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1312                                                 continue;
1313                                         rdev = conf->mirrors[d].rdev;
1314                                         if (sync_page_io(rdev->bdev,
1315                                                          sect + rdev->data_offset,
1316                                                          s<<9,
1317                                                          bio->bi_io_vec[idx].bv_page,
1318                                                          READ) == 0)
1319                                                 md_error(mddev, rdev);
1320                                 }
1321                         } else {
1322                                 char b[BDEVNAME_SIZE];
1323                                 /* Cannot read from anywhere, array is toast */
1324                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1325                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1326                                        " for block %llu\n",
1327                                        bdevname(bio->bi_bdev,b),
1328                                        (unsigned long long)r1_bio->sector);
1329                                 md_done_sync(mddev, r1_bio->sectors, 0);
1330                                 put_buf(r1_bio);
1331                                 return;
1332                         }
1333                         sectors -= s;
1334                         sect += s;
1335                         idx ++;
1336                 }
1337         }
1338
1339         /*
1340          * schedule writes
1341          */
1342         atomic_set(&r1_bio->remaining, 1);
1343         for (i = 0; i < disks ; i++) {
1344                 wbio = r1_bio->bios[i];
1345                 if (wbio->bi_end_io == NULL ||
1346                     (wbio->bi_end_io == end_sync_read &&
1347                      (i == r1_bio->read_disk ||
1348                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1349                         continue;
1350
1351                 wbio->bi_rw = WRITE;
1352                 wbio->bi_end_io = end_sync_write;
1353                 atomic_inc(&r1_bio->remaining);
1354                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1355
1356                 generic_make_request(wbio);
1357         }
1358
1359         if (atomic_dec_and_test(&r1_bio->remaining)) {
1360                 /* if we're here, all write(s) have completed, so clean up */
1361                 md_done_sync(mddev, r1_bio->sectors, 1);
1362                 put_buf(r1_bio);
1363         }
1364 }
1365
1366 /*
1367  * This is a kernel thread which:
1368  *
1369  *      1.      Retries failed read operations on working mirrors.
1370  *      2.      Updates the raid superblock when problems encounter.
1371  *      3.      Performs writes following reads for array syncronising.
1372  */
1373
1374 static void fix_read_error(conf_t *conf, int read_disk,
1375                            sector_t sect, int sectors)
1376 {
1377         mddev_t *mddev = conf->mddev;
1378         while(sectors) {
1379                 int s = sectors;
1380                 int d = read_disk;
1381                 int success = 0;
1382                 int start;
1383                 mdk_rdev_t *rdev;
1384
1385                 if (s > (PAGE_SIZE>>9))
1386                         s = PAGE_SIZE >> 9;
1387
1388                 do {
1389                         /* Note: no rcu protection needed here
1390                          * as this is synchronous in the raid1d thread
1391                          * which is the thread that might remove
1392                          * a device.  If raid1d ever becomes multi-threaded....
1393                          */
1394                         rdev = conf->mirrors[d].rdev;
1395                         if (rdev &&
1396                             test_bit(In_sync, &rdev->flags) &&
1397                             sync_page_io(rdev->bdev,
1398                                          sect + rdev->data_offset,
1399                                          s<<9,
1400                                          conf->tmppage, READ))
1401                                 success = 1;
1402                         else {
1403                                 d++;
1404                                 if (d == conf->raid_disks)
1405                                         d = 0;
1406                         }
1407                 } while (!success && d != read_disk);
1408
1409                 if (!success) {
1410                         /* Cannot read from anywhere -- bye bye array */
1411                         md_error(mddev, conf->mirrors[read_disk].rdev);
1412                         break;
1413                 }
1414                 /* write it back and re-read */
1415                 start = d;
1416                 while (d != read_disk) {
1417                         if (d==0)
1418                                 d = conf->raid_disks;
1419                         d--;
1420                         rdev = conf->mirrors[d].rdev;
1421                         if (rdev &&
1422                             test_bit(In_sync, &rdev->flags)) {
1423                                 if (sync_page_io(rdev->bdev,
1424                                                  sect + rdev->data_offset,
1425                                                  s<<9, conf->tmppage, WRITE)
1426                                     == 0)
1427                                         /* Well, this device is dead */
1428                                         md_error(mddev, rdev);
1429                         }
1430                 }
1431                 d = start;
1432                 while (d != read_disk) {
1433                         char b[BDEVNAME_SIZE];
1434                         if (d==0)
1435                                 d = conf->raid_disks;
1436                         d--;
1437                         rdev = conf->mirrors[d].rdev;
1438                         if (rdev &&
1439                             test_bit(In_sync, &rdev->flags)) {
1440                                 if (sync_page_io(rdev->bdev,
1441                                                  sect + rdev->data_offset,
1442                                                  s<<9, conf->tmppage, READ)
1443                                     == 0)
1444                                         /* Well, this device is dead */
1445                                         md_error(mddev, rdev);
1446                                 else {
1447                                         atomic_add(s, &rdev->corrected_errors);
1448                                         printk(KERN_INFO
1449                                                "raid1:%s: read error corrected "
1450                                                "(%d sectors at %llu on %s)\n",
1451                                                mdname(mddev), s,
1452                                                (unsigned long long)(sect +
1453                                                    rdev->data_offset),
1454                                                bdevname(rdev->bdev, b));
1455                                 }
1456                         }
1457                 }
1458                 sectors -= s;
1459                 sect += s;
1460         }
1461 }
1462
1463 static void raid1d(mddev_t *mddev)
1464 {
1465         r1bio_t *r1_bio;
1466         struct bio *bio;
1467         unsigned long flags;
1468         conf_t *conf = mddev_to_conf(mddev);
1469         struct list_head *head = &conf->retry_list;
1470         int unplug=0;
1471         mdk_rdev_t *rdev;
1472
1473         md_check_recovery(mddev);
1474
1475         for (;;) {
1476                 char b[BDEVNAME_SIZE];
1477                 spin_lock_irqsave(&conf->device_lock, flags);
1478
1479                 if (conf->pending_bio_list.head) {
1480                         bio = bio_list_get(&conf->pending_bio_list);
1481                         blk_remove_plug(mddev->queue);
1482                         spin_unlock_irqrestore(&conf->device_lock, flags);
1483                         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1484                         bitmap_unplug(mddev->bitmap);
1485
1486                         while (bio) { /* submit pending writes */
1487                                 struct bio *next = bio->bi_next;
1488                                 bio->bi_next = NULL;
1489                                 generic_make_request(bio);
1490                                 bio = next;
1491                         }
1492                         unplug = 1;
1493
1494                         continue;
1495                 }
1496
1497                 if (list_empty(head))
1498                         break;
1499                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1500                 list_del(head->prev);
1501                 conf->nr_queued--;
1502                 spin_unlock_irqrestore(&conf->device_lock, flags);
1503
1504                 mddev = r1_bio->mddev;
1505                 conf = mddev_to_conf(mddev);
1506                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1507                         sync_request_write(mddev, r1_bio);
1508                         unplug = 1;
1509                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1510                         /* some requests in the r1bio were BIO_RW_BARRIER
1511                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1512                          * Better resubmit without the barrier.
1513                          * We know which devices to resubmit for, because
1514                          * all others have had their bios[] entry cleared.
1515                          * We already have a nr_pending reference on these rdevs.
1516                          */
1517                         int i;
1518                         const int do_sync = bio_sync(r1_bio->master_bio);
1519                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1520                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1521                         for (i=0; i < conf->raid_disks; i++)
1522                                 if (r1_bio->bios[i])
1523                                         atomic_inc(&r1_bio->remaining);
1524                         for (i=0; i < conf->raid_disks; i++)
1525                                 if (r1_bio->bios[i]) {
1526                                         struct bio_vec *bvec;
1527                                         int j;
1528
1529                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1530                                         /* copy pages from the failed bio, as
1531                                          * this might be a write-behind device */
1532                                         __bio_for_each_segment(bvec, bio, j, 0)
1533                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1534                                         bio_put(r1_bio->bios[i]);
1535                                         bio->bi_sector = r1_bio->sector +
1536                                                 conf->mirrors[i].rdev->data_offset;
1537                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1538                                         bio->bi_end_io = raid1_end_write_request;
1539                                         bio->bi_rw = WRITE | do_sync;
1540                                         bio->bi_private = r1_bio;
1541                                         r1_bio->bios[i] = bio;
1542                                         generic_make_request(bio);
1543                                 }
1544                 } else {
1545                         int disk;
1546
1547                         /* we got a read error. Maybe the drive is bad.  Maybe just
1548                          * the block and we can fix it.
1549                          * We freeze all other IO, and try reading the block from
1550                          * other devices.  When we find one, we re-write
1551                          * and check it that fixes the read error.
1552                          * This is all done synchronously while the array is
1553                          * frozen
1554                          */
1555                         if (mddev->ro == 0) {
1556                                 freeze_array(conf);
1557                                 fix_read_error(conf, r1_bio->read_disk,
1558                                                r1_bio->sector,
1559                                                r1_bio->sectors);
1560                                 unfreeze_array(conf);
1561                         }
1562
1563                         bio = r1_bio->bios[r1_bio->read_disk];
1564                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1565                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1566                                        " read error for block %llu\n",
1567                                        bdevname(bio->bi_bdev,b),
1568                                        (unsigned long long)r1_bio->sector);
1569                                 raid_end_bio_io(r1_bio);
1570                         } else {
1571                                 const int do_sync = bio_sync(r1_bio->master_bio);
1572                                 r1_bio->bios[r1_bio->read_disk] =
1573                                         mddev->ro ? IO_BLOCKED : NULL;
1574                                 r1_bio->read_disk = disk;
1575                                 bio_put(bio);
1576                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1577                                 r1_bio->bios[r1_bio->read_disk] = bio;
1578                                 rdev = conf->mirrors[disk].rdev;
1579                                 if (printk_ratelimit())
1580                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1581                                                " another mirror\n",
1582                                                bdevname(rdev->bdev,b),
1583                                                (unsigned long long)r1_bio->sector);
1584                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1585                                 bio->bi_bdev = rdev->bdev;
1586                                 bio->bi_end_io = raid1_end_read_request;
1587                                 bio->bi_rw = READ | do_sync;
1588                                 bio->bi_private = r1_bio;
1589                                 unplug = 1;
1590                                 generic_make_request(bio);
1591                         }
1592                 }
1593         }
1594         spin_unlock_irqrestore(&conf->device_lock, flags);
1595         if (unplug)
1596                 unplug_slaves(mddev);
1597 }
1598
1599
1600 static int init_resync(conf_t *conf)
1601 {
1602         int buffs;
1603
1604         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1605         BUG_ON(conf->r1buf_pool);
1606         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1607                                           conf->poolinfo);
1608         if (!conf->r1buf_pool)
1609                 return -ENOMEM;
1610         conf->next_resync = 0;
1611         return 0;
1612 }
1613
1614 /*
1615  * perform a "sync" on one "block"
1616  *
1617  * We need to make sure that no normal I/O request - particularly write
1618  * requests - conflict with active sync requests.
1619  *
1620  * This is achieved by tracking pending requests and a 'barrier' concept
1621  * that can be installed to exclude normal IO requests.
1622  */
1623
1624 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1625 {
1626         conf_t *conf = mddev_to_conf(mddev);
1627         r1bio_t *r1_bio;
1628         struct bio *bio;
1629         sector_t max_sector, nr_sectors;
1630         int disk = -1;
1631         int i;
1632         int wonly = -1;
1633         int write_targets = 0, read_targets = 0;
1634         int sync_blocks;
1635         int still_degraded = 0;
1636
1637         if (!conf->r1buf_pool)
1638         {
1639 /*
1640                 printk("sync start - bitmap %p\n", mddev->bitmap);
1641 */
1642                 if (init_resync(conf))
1643                         return 0;
1644         }
1645
1646         max_sector = mddev->size << 1;
1647         if (sector_nr >= max_sector) {
1648                 /* If we aborted, we need to abort the
1649                  * sync on the 'current' bitmap chunk (there will
1650                  * only be one in raid1 resync.
1651                  * We can find the current addess in mddev->curr_resync
1652                  */
1653                 if (mddev->curr_resync < max_sector) /* aborted */
1654                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1655                                                 &sync_blocks, 1);
1656                 else /* completed sync */
1657                         conf->fullsync = 0;
1658
1659                 bitmap_close_sync(mddev->bitmap);
1660                 close_sync(conf);
1661                 return 0;
1662         }
1663
1664         if (mddev->bitmap == NULL &&
1665             mddev->recovery_cp == MaxSector &&
1666             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1667             conf->fullsync == 0) {
1668                 *skipped = 1;
1669                 return max_sector - sector_nr;
1670         }
1671         /* before building a request, check if we can skip these blocks..
1672          * This call the bitmap_start_sync doesn't actually record anything
1673          */
1674         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1675             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1676                 /* We can skip this block, and probably several more */
1677                 *skipped = 1;
1678                 return sync_blocks;
1679         }
1680         /*
1681          * If there is non-resync activity waiting for a turn,
1682          * and resync is going fast enough,
1683          * then let it though before starting on this new sync request.
1684          */
1685         if (!go_faster && conf->nr_waiting)
1686                 msleep_interruptible(1000);
1687
1688         raise_barrier(conf);
1689
1690         conf->next_resync = sector_nr;
1691
1692         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1693         rcu_read_lock();
1694         /*
1695          * If we get a correctably read error during resync or recovery,
1696          * we might want to read from a different device.  So we
1697          * flag all drives that could conceivably be read from for READ,
1698          * and any others (which will be non-In_sync devices) for WRITE.
1699          * If a read fails, we try reading from something else for which READ
1700          * is OK.
1701          */
1702
1703         r1_bio->mddev = mddev;
1704         r1_bio->sector = sector_nr;
1705         r1_bio->state = 0;
1706         set_bit(R1BIO_IsSync, &r1_bio->state);
1707
1708         for (i=0; i < conf->raid_disks; i++) {
1709                 mdk_rdev_t *rdev;
1710                 bio = r1_bio->bios[i];
1711
1712                 /* take from bio_init */
1713                 bio->bi_next = NULL;
1714                 bio->bi_flags |= 1 << BIO_UPTODATE;
1715                 bio->bi_rw = READ;
1716                 bio->bi_vcnt = 0;
1717                 bio->bi_idx = 0;
1718                 bio->bi_phys_segments = 0;
1719                 bio->bi_hw_segments = 0;
1720                 bio->bi_size = 0;
1721                 bio->bi_end_io = NULL;
1722                 bio->bi_private = NULL;
1723
1724                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1725                 if (rdev == NULL ||
1726                            test_bit(Faulty, &rdev->flags)) {
1727                         still_degraded = 1;
1728                         continue;
1729                 } else if (!test_bit(In_sync, &rdev->flags)) {
1730                         bio->bi_rw = WRITE;
1731                         bio->bi_end_io = end_sync_write;
1732                         write_targets ++;
1733                 } else {
1734                         /* may need to read from here */
1735                         bio->bi_rw = READ;
1736                         bio->bi_end_io = end_sync_read;
1737                         if (test_bit(WriteMostly, &rdev->flags)) {
1738                                 if (wonly < 0)
1739                                         wonly = i;
1740                         } else {
1741                                 if (disk < 0)
1742                                         disk = i;
1743                         }
1744                         read_targets++;
1745                 }
1746                 atomic_inc(&rdev->nr_pending);
1747                 bio->bi_sector = sector_nr + rdev->data_offset;
1748                 bio->bi_bdev = rdev->bdev;
1749                 bio->bi_private = r1_bio;
1750         }
1751         rcu_read_unlock();
1752         if (disk < 0)
1753                 disk = wonly;
1754         r1_bio->read_disk = disk;
1755
1756         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1757                 /* extra read targets are also write targets */
1758                 write_targets += read_targets-1;
1759
1760         if (write_targets == 0 || read_targets == 0) {
1761                 /* There is nowhere to write, so all non-sync
1762                  * drives must be failed - so we are finished
1763                  */
1764                 sector_t rv = max_sector - sector_nr;
1765                 *skipped = 1;
1766                 put_buf(r1_bio);
1767                 return rv;
1768         }
1769
1770         nr_sectors = 0;
1771         sync_blocks = 0;
1772         do {
1773                 struct page *page;
1774                 int len = PAGE_SIZE;
1775                 if (sector_nr + (len>>9) > max_sector)
1776                         len = (max_sector - sector_nr) << 9;
1777                 if (len == 0)
1778                         break;
1779                 if (sync_blocks == 0) {
1780                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1781                                                &sync_blocks, still_degraded) &&
1782                             !conf->fullsync &&
1783                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1784                                 break;
1785                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1786                         if (len > (sync_blocks<<9))
1787                                 len = sync_blocks<<9;
1788                 }
1789
1790                 for (i=0 ; i < conf->raid_disks; i++) {
1791                         bio = r1_bio->bios[i];
1792                         if (bio->bi_end_io) {
1793                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1794                                 if (bio_add_page(bio, page, len, 0) == 0) {
1795                                         /* stop here */
1796                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1797                                         while (i > 0) {
1798                                                 i--;
1799                                                 bio = r1_bio->bios[i];
1800                                                 if (bio->bi_end_io==NULL)
1801                                                         continue;
1802                                                 /* remove last page from this bio */
1803                                                 bio->bi_vcnt--;
1804                                                 bio->bi_size -= len;
1805                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1806                                         }
1807                                         goto bio_full;
1808                                 }
1809                         }
1810                 }
1811                 nr_sectors += len>>9;
1812                 sector_nr += len>>9;
1813                 sync_blocks -= (len>>9);
1814         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1815  bio_full:
1816         r1_bio->sectors = nr_sectors;
1817
1818         /* For a user-requested sync, we read all readable devices and do a
1819          * compare
1820          */
1821         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1822                 atomic_set(&r1_bio->remaining, read_targets);
1823                 for (i=0; i<conf->raid_disks; i++) {
1824                         bio = r1_bio->bios[i];
1825                         if (bio->bi_end_io == end_sync_read) {
1826                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1827                                 generic_make_request(bio);
1828                         }
1829                 }
1830         } else {
1831                 atomic_set(&r1_bio->remaining, 1);
1832                 bio = r1_bio->bios[r1_bio->read_disk];
1833                 md_sync_acct(bio->bi_bdev, nr_sectors);
1834                 generic_make_request(bio);
1835
1836         }
1837         return nr_sectors;
1838 }
1839
1840 static int run(mddev_t *mddev)
1841 {
1842         conf_t *conf;
1843         int i, j, disk_idx;
1844         mirror_info_t *disk;
1845         mdk_rdev_t *rdev;
1846         struct list_head *tmp;
1847
1848         if (mddev->level != 1) {
1849                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1850                        mdname(mddev), mddev->level);
1851                 goto out;
1852         }
1853         if (mddev->reshape_position != MaxSector) {
1854                 printk("raid1: %s: reshape_position set but not supported\n",
1855                        mdname(mddev));
1856                 goto out;
1857         }
1858         /*
1859          * copy the already verified devices into our private RAID1
1860          * bookkeeping area. [whatever we allocate in run(),
1861          * should be freed in stop()]
1862          */
1863         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1864         mddev->private = conf;
1865         if (!conf)
1866                 goto out_no_mem;
1867
1868         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1869                                  GFP_KERNEL);
1870         if (!conf->mirrors)
1871                 goto out_no_mem;
1872
1873         conf->tmppage = alloc_page(GFP_KERNEL);
1874         if (!conf->tmppage)
1875                 goto out_no_mem;
1876
1877         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1878         if (!conf->poolinfo)
1879                 goto out_no_mem;
1880         conf->poolinfo->mddev = mddev;
1881         conf->poolinfo->raid_disks = mddev->raid_disks;
1882         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1883                                           r1bio_pool_free,
1884                                           conf->poolinfo);
1885         if (!conf->r1bio_pool)
1886                 goto out_no_mem;
1887
1888         ITERATE_RDEV(mddev, rdev, tmp) {
1889                 disk_idx = rdev->raid_disk;
1890                 if (disk_idx >= mddev->raid_disks
1891                     || disk_idx < 0)
1892                         continue;
1893                 disk = conf->mirrors + disk_idx;
1894
1895                 disk->rdev = rdev;
1896
1897                 blk_queue_stack_limits(mddev->queue,
1898                                        rdev->bdev->bd_disk->queue);
1899                 /* as we don't honour merge_bvec_fn, we must never risk
1900                  * violating it, so limit ->max_sector to one PAGE, as
1901                  * a one page request is never in violation.
1902                  */
1903                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1904                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1905                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1906
1907                 disk->head_position = 0;
1908         }
1909         conf->raid_disks = mddev->raid_disks;
1910         conf->mddev = mddev;
1911         spin_lock_init(&conf->device_lock);
1912         INIT_LIST_HEAD(&conf->retry_list);
1913
1914         spin_lock_init(&conf->resync_lock);
1915         init_waitqueue_head(&conf->wait_barrier);
1916
1917         bio_list_init(&conf->pending_bio_list);
1918         bio_list_init(&conf->flushing_bio_list);
1919
1920
1921         mddev->degraded = 0;
1922         for (i = 0; i < conf->raid_disks; i++) {
1923
1924                 disk = conf->mirrors + i;
1925
1926                 if (!disk->rdev ||
1927                     !test_bit(In_sync, &disk->rdev->flags)) {
1928                         disk->head_position = 0;
1929                         mddev->degraded++;
1930                         if (disk->rdev)
1931                                 conf->fullsync = 1;
1932                 }
1933         }
1934         if (mddev->degraded == conf->raid_disks) {
1935                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1936                         mdname(mddev));
1937                 goto out_free_conf;
1938         }
1939         if (conf->raid_disks - mddev->degraded == 1)
1940                 mddev->recovery_cp = MaxSector;
1941
1942         /*
1943          * find the first working one and use it as a starting point
1944          * to read balancing.
1945          */
1946         for (j = 0; j < conf->raid_disks &&
1947                      (!conf->mirrors[j].rdev ||
1948                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
1949                 /* nothing */;
1950         conf->last_used = j;
1951
1952
1953         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1954         if (!mddev->thread) {
1955                 printk(KERN_ERR
1956                        "raid1: couldn't allocate thread for %s\n",
1957                        mdname(mddev));
1958                 goto out_free_conf;
1959         }
1960
1961         printk(KERN_INFO
1962                 "raid1: raid set %s active with %d out of %d mirrors\n",
1963                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1964                 mddev->raid_disks);
1965         /*
1966          * Ok, everything is just fine now
1967          */
1968         mddev->array_size = mddev->size;
1969
1970         mddev->queue->unplug_fn = raid1_unplug;
1971         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
1972         mddev->queue->backing_dev_info.congested_data = mddev;
1973
1974         return 0;
1975
1976 out_no_mem:
1977         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1978                mdname(mddev));
1979
1980 out_free_conf:
1981         if (conf) {
1982                 if (conf->r1bio_pool)
1983                         mempool_destroy(conf->r1bio_pool);
1984                 kfree(conf->mirrors);
1985                 safe_put_page(conf->tmppage);
1986                 kfree(conf->poolinfo);
1987                 kfree(conf);
1988                 mddev->private = NULL;
1989         }
1990 out:
1991         return -EIO;
1992 }
1993
1994 static int stop(mddev_t *mddev)
1995 {
1996         conf_t *conf = mddev_to_conf(mddev);
1997         struct bitmap *bitmap = mddev->bitmap;
1998         int behind_wait = 0;
1999
2000         /* wait for behind writes to complete */
2001         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2002                 behind_wait++;
2003                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2004                 set_current_state(TASK_UNINTERRUPTIBLE);
2005                 schedule_timeout(HZ); /* wait a second */
2006                 /* need to kick something here to make sure I/O goes? */
2007         }
2008
2009         md_unregister_thread(mddev->thread);
2010         mddev->thread = NULL;
2011         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2012         if (conf->r1bio_pool)
2013                 mempool_destroy(conf->r1bio_pool);
2014         kfree(conf->mirrors);
2015         kfree(conf->poolinfo);
2016         kfree(conf);
2017         mddev->private = NULL;
2018         return 0;
2019 }
2020
2021 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2022 {
2023         /* no resync is happening, and there is enough space
2024          * on all devices, so we can resize.
2025          * We need to make sure resync covers any new space.
2026          * If the array is shrinking we should possibly wait until
2027          * any io in the removed space completes, but it hardly seems
2028          * worth it.
2029          */
2030         mddev->array_size = sectors>>1;
2031         set_capacity(mddev->gendisk, mddev->array_size << 1);
2032         mddev->changed = 1;
2033         if (mddev->array_size > mddev->size && mddev->recovery_cp == MaxSector) {
2034                 mddev->recovery_cp = mddev->size << 1;
2035                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2036         }
2037         mddev->size = mddev->array_size;
2038         mddev->resync_max_sectors = sectors;
2039         return 0;
2040 }
2041
2042 static int raid1_reshape(mddev_t *mddev)
2043 {
2044         /* We need to:
2045          * 1/ resize the r1bio_pool
2046          * 2/ resize conf->mirrors
2047          *
2048          * We allocate a new r1bio_pool if we can.
2049          * Then raise a device barrier and wait until all IO stops.
2050          * Then resize conf->mirrors and swap in the new r1bio pool.
2051          *
2052          * At the same time, we "pack" the devices so that all the missing
2053          * devices have the higher raid_disk numbers.
2054          */
2055         mempool_t *newpool, *oldpool;
2056         struct pool_info *newpoolinfo;
2057         mirror_info_t *newmirrors;
2058         conf_t *conf = mddev_to_conf(mddev);
2059         int cnt, raid_disks;
2060         unsigned long flags;
2061         int d, d2;
2062
2063         /* Cannot change chunk_size, layout, or level */
2064         if (mddev->chunk_size != mddev->new_chunk ||
2065             mddev->layout != mddev->new_layout ||
2066             mddev->level != mddev->new_level) {
2067                 mddev->new_chunk = mddev->chunk_size;
2068                 mddev->new_layout = mddev->layout;
2069                 mddev->new_level = mddev->level;
2070                 return -EINVAL;
2071         }
2072
2073         md_allow_write(mddev);
2074
2075         raid_disks = mddev->raid_disks + mddev->delta_disks;
2076
2077         if (raid_disks < conf->raid_disks) {
2078                 cnt=0;
2079                 for (d= 0; d < conf->raid_disks; d++)
2080                         if (conf->mirrors[d].rdev)
2081                                 cnt++;
2082                 if (cnt > raid_disks)
2083                         return -EBUSY;
2084         }
2085
2086         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2087         if (!newpoolinfo)
2088                 return -ENOMEM;
2089         newpoolinfo->mddev = mddev;
2090         newpoolinfo->raid_disks = raid_disks;
2091
2092         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2093                                  r1bio_pool_free, newpoolinfo);
2094         if (!newpool) {
2095                 kfree(newpoolinfo);
2096                 return -ENOMEM;
2097         }
2098         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2099         if (!newmirrors) {
2100                 kfree(newpoolinfo);
2101                 mempool_destroy(newpool);
2102                 return -ENOMEM;
2103         }
2104
2105         raise_barrier(conf);
2106
2107         /* ok, everything is stopped */
2108         oldpool = conf->r1bio_pool;
2109         conf->r1bio_pool = newpool;
2110
2111         for (d = d2 = 0; d < conf->raid_disks; d++) {
2112                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2113                 if (rdev && rdev->raid_disk != d2) {
2114                         char nm[20];
2115                         sprintf(nm, "rd%d", rdev->raid_disk);
2116                         sysfs_remove_link(&mddev->kobj, nm);
2117                         rdev->raid_disk = d2;
2118                         sprintf(nm, "rd%d", rdev->raid_disk);
2119                         sysfs_remove_link(&mddev->kobj, nm);
2120                         if (sysfs_create_link(&mddev->kobj,
2121                                               &rdev->kobj, nm))
2122                                 printk(KERN_WARNING
2123                                        "md/raid1: cannot register "
2124                                        "%s for %s\n",
2125                                        nm, mdname(mddev));
2126                 }
2127                 if (rdev)
2128                         newmirrors[d2++].rdev = rdev;
2129         }
2130         kfree(conf->mirrors);
2131         conf->mirrors = newmirrors;
2132         kfree(conf->poolinfo);
2133         conf->poolinfo = newpoolinfo;
2134
2135         spin_lock_irqsave(&conf->device_lock, flags);
2136         mddev->degraded += (raid_disks - conf->raid_disks);
2137         spin_unlock_irqrestore(&conf->device_lock, flags);
2138         conf->raid_disks = mddev->raid_disks = raid_disks;
2139         mddev->delta_disks = 0;
2140
2141         conf->last_used = 0; /* just make sure it is in-range */
2142         lower_barrier(conf);
2143
2144         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2145         md_wakeup_thread(mddev->thread);
2146
2147         mempool_destroy(oldpool);
2148         return 0;
2149 }
2150
2151 static void raid1_quiesce(mddev_t *mddev, int state)
2152 {
2153         conf_t *conf = mddev_to_conf(mddev);
2154
2155         switch(state) {
2156         case 1:
2157                 raise_barrier(conf);
2158                 break;
2159         case 0:
2160                 lower_barrier(conf);
2161                 break;
2162         }
2163 }
2164
2165
2166 static struct mdk_personality raid1_personality =
2167 {
2168         .name           = "raid1",
2169         .level          = 1,
2170         .owner          = THIS_MODULE,
2171         .make_request   = make_request,
2172         .run            = run,
2173         .stop           = stop,
2174         .status         = status,
2175         .error_handler  = error,
2176         .hot_add_disk   = raid1_add_disk,
2177         .hot_remove_disk= raid1_remove_disk,
2178         .spare_active   = raid1_spare_active,
2179         .sync_request   = sync_request,
2180         .resize         = raid1_resize,
2181         .check_reshape  = raid1_reshape,
2182         .quiesce        = raid1_quiesce,
2183 };
2184
2185 static int __init raid_init(void)
2186 {
2187         return register_md_personality(&raid1_personality);
2188 }
2189
2190 static void raid_exit(void)
2191 {
2192         unregister_md_personality(&raid1_personality);
2193 }
2194
2195 module_init(raid_init);
2196 module_exit(raid_exit);
2197 MODULE_LICENSE("GPL");
2198 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2199 MODULE_ALIAS("md-raid1");
2200 MODULE_ALIAS("md-level-1");