drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/slab.h>
  35 #include <linux/delay.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/seq_file.h>
  38 #include "md.h"
  39 #include "raid1.h"
  40 #include "bitmap.h"
  41
  42 #define DEBUG 0
  43 #if DEBUG
  44 #define PRINTK(x...) printk(x)
  45 #else
  46 #define PRINTK(x...)
  47 #endif
  48
  49 /*
  50  * Number of guaranteed r1bios in case of extreme VM load:
  51  */
  52 #define NR_RAID1_BIOS 256
  53
  54
  55 static void unplug_slaves(mddev_t *mddev);
  56
  57 static void allow_barrier(conf_t *conf);
  58 static void lower_barrier(conf_t *conf);
  59
  60 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  61 {
  62         struct pool_info *pi = data;
  63         r1bio_t *r1_bio;
  64         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  65
  66         /* allocate a r1bio with room for raid_disks entries in the bios array */
  67         r1_bio = kzalloc(size, gfp_flags);
  68         if (!r1_bio && pi->mddev)
  69                 unplug_slaves(pi->mddev);
  70
  71         return r1_bio;
  72 }
  73
  74 static void r1bio_pool_free(void *r1_bio, void *data)
  75 {
  76         kfree(r1_bio);
  77 }
  78
  79 #define RESYNC_BLOCK_SIZE (64*1024)
  80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  83 #define RESYNC_WINDOW (2048*1024)
  84
  85 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  86 {
  87         struct pool_info *pi = data;
  88         struct page *page;
  89         r1bio_t *r1_bio;
  90         struct bio *bio;
  91         int i, j;
  92
  93         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  94         if (!r1_bio) {
  95                 unplug_slaves(pi->mddev);
  96                 return NULL;
  97         }
  98
  99         /*
 100          * Allocate bios : 1 for reading, n-1 for writing
 101          */
 102         for (j = pi->raid_disks ; j-- ; ) {
 103                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 104                 if (!bio)
 105                         goto out_free_bio;
 106                 r1_bio->bios[j] = bio;
 107         }
 108         /*
 109          * Allocate RESYNC_PAGES data pages and attach them to
 110          * the first bio.
 111          * If this is a user-requested check/repair, allocate
 112          * RESYNC_PAGES for each bio.
 113          */
 114         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 115                 j = pi->raid_disks;
 116         else
 117                 j = 1;
 118         while(j--) {
 119                 bio = r1_bio->bios[j];
 120                 for (i = 0; i < RESYNC_PAGES; i++) {
 121                         page = alloc_page(gfp_flags);
 122                         if (unlikely(!page))
 123                                 goto out_free_pages;
 124
 125                         bio->bi_io_vec[i].bv_page = page;
 126                         bio->bi_vcnt = i+1;
 127                 }
 128         }
 129         /* If not user-requests, copy the page pointers to all bios */
 130         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 131                 for (i=0; i<RESYNC_PAGES ; i++)
 132                         for (j=1; j<pi->raid_disks; j++)
 133                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 134                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 135         }
 136
 137         r1_bio->master_bio = NULL;
 138
 139         return r1_bio;
 140
 141 out_free_pages:
 142         for (j=0 ; j < pi->raid_disks; j++)
 143                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 144                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 145         j = -1;
 146 out_free_bio:
 147         while ( ++j < pi->raid_disks )
 148                 bio_put(r1_bio->bios[j]);
 149         r1bio_pool_free(r1_bio, data);
 150         return NULL;
 151 }
 152
 153 static void r1buf_pool_free(void *__r1_bio, void *data)
 154 {
 155         struct pool_info *pi = data;
 156         int i,j;
 157         r1bio_t *r1bio = __r1_bio;
 158
 159         for (i = 0; i < RESYNC_PAGES; i++)
 160                 for (j = pi->raid_disks; j-- ;) {
 161                         if (j == 0 ||
 162                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 163                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 164                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 165                 }
 166         for (i=0 ; i < pi->raid_disks; i++)
 167                 bio_put(r1bio->bios[i]);
 168
 169         r1bio_pool_free(r1bio, data);
 170 }
 171
 172 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 173 {
 174         int i;
 175
 176         for (i = 0; i < conf->raid_disks; i++) {
 177                 struct bio **bio = r1_bio->bios + i;
 178                 if (*bio && *bio != IO_BLOCKED)
 179                         bio_put(*bio);
 180                 *bio = NULL;
 181         }
 182 }
 183
 184 static void free_r1bio(r1bio_t *r1_bio)
 185 {
 186         conf_t *conf = r1_bio->mddev->private;
 187
 188         /*
 189          * Wake up any possible resync thread that waits for the device
 190          * to go idle.
 191          */
 192         allow_barrier(conf);
 193
 194         put_all_bios(conf, r1_bio);
 195         mempool_free(r1_bio, conf->r1bio_pool);
 196 }
 197
 198 static void put_buf(r1bio_t *r1_bio)
 199 {
 200         conf_t *conf = r1_bio->mddev->private;
 201         int i;
 202
 203         for (i=0; i<conf->raid_disks; i++) {
 204                 struct bio *bio = r1_bio->bios[i];
 205                 if (bio->bi_end_io)
 206                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 207         }
 208
 209         mempool_free(r1_bio, conf->r1buf_pool);
 210
 211         lower_barrier(conf);
 212 }
 213
 214 static void reschedule_retry(r1bio_t *r1_bio)
 215 {
 216         unsigned long flags;
 217         mddev_t *mddev = r1_bio->mddev;
 218         conf_t *conf = mddev->private;
 219
 220         spin_lock_irqsave(&conf->device_lock, flags);
 221         list_add(&r1_bio->retry_list, &conf->retry_list);
 222         conf->nr_queued ++;
 223         spin_unlock_irqrestore(&conf->device_lock, flags);
 224
 225         wake_up(&conf->wait_barrier);
 226         md_wakeup_thread(mddev->thread);
 227 }
 228
 229 /*
 230  * raid_end_bio_io() is called when we have finished servicing a mirrored
 231  * operation and are ready to return a success/failure code to the buffer
 232  * cache layer.
 233  */
 234 static void raid_end_bio_io(r1bio_t *r1_bio)
 235 {
 236         struct bio *bio = r1_bio->master_bio;
 237
 238         /* if nobody has done the final endio yet, do it now */
 239         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 240                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 241                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 242                         (unsigned long long) bio->bi_sector,
 243                         (unsigned long long) bio->bi_sector +
 244                                 (bio->bi_size >> 9) - 1);
 245
 246                 bio_endio(bio,
 247                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 248         }
 249         free_r1bio(r1_bio);
 250 }
 251
 252 /*
 253  * Update disk head position estimator based on IRQ completion info.
 254  */
 255 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 256 {
 257         conf_t *conf = r1_bio->mddev->private;
 258
 259         conf->mirrors[disk].head_position =
 260                 r1_bio->sector + (r1_bio->sectors);
 261 }
 262
 263 static void raid1_end_read_request(struct bio *bio, int error)
 264 {
 265         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 266         r1bio_t *r1_bio = bio->bi_private;
 267         int mirror;
 268         conf_t *conf = r1_bio->mddev->private;
 269
 270         mirror = r1_bio->read_disk;
 271         /*
 272          * this branch is our 'one mirror IO has finished' event handler:
 273          */
 274         update_head_pos(mirror, r1_bio);
 275
 276         if (uptodate)
 277                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 278         else {
 279                 /* If all other devices have failed, we want to return
 280                  * the error upwards rather than fail the last device.
 281                  * Here we redefine "uptodate" to mean "Don't want to retry"
 282                  */
 283                 unsigned long flags;
 284                 spin_lock_irqsave(&conf->device_lock, flags);
 285                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 286                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 287                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 288                         uptodate = 1;
 289                 spin_unlock_irqrestore(&conf->device_lock, flags);
 290         }
 291
 292         if (uptodate)
 293                 raid_end_bio_io(r1_bio);
 294         else {
 295                 /*
 296                  * oops, read error:
 297                  */
 298                 char b[BDEVNAME_SIZE];
 299                 if (printk_ratelimit())
 300                         printk(KERN_ERR "md/raid1:%s: %s: rescheduling sector %llu\n",
 301                                mdname(conf->mddev),
 302                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 303                 reschedule_retry(r1_bio);
 304         }
 305
 306         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 307 }
 308
 309 static void raid1_end_write_request(struct bio *bio, int error)
 310 {
 311         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 312         r1bio_t *r1_bio = bio->bi_private;
 313         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 314         conf_t *conf = r1_bio->mddev->private;
 315         struct bio *to_put = NULL;
 316
 317
 318         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 319                 if (r1_bio->bios[mirror] == bio)
 320                         break;
 321
 322         /*
 323          * '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 we
 334                  * will return a good error code for to the higher
 335                  * levels even if IO on some other mirrored buffer
 336                  * fails.
 337                  *
 338                  * The 'master' represents the composite IO operation
 339                  * to user-side. So if something waits for IO, then it
 340                  * will wait for the 'master' bio.
 341                  */
 342                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 343
 344         update_head_pos(mirror, r1_bio);
 345
 346         if (behind) {
 347                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 348                         atomic_dec(&r1_bio->behind_remaining);
 349
 350                 /*
 351                  * In behind mode, we ACK the master bio once the I/O
 352                  * has safely reached all non-writemostly
 353                  * disks. Setting the Returned bit ensures that this
 354                  * gets done only once -- we don't ever want to return
 355                  * -EIO here, instead we'll wait
 356                  */
 357                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 358                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 359                         /* Maybe we can return now */
 360                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 361                                 struct bio *mbio = r1_bio->master_bio;
 362                                 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 363                                        (unsigned long long) mbio->bi_sector,
 364                                        (unsigned long long) mbio->bi_sector +
 365                                        (mbio->bi_size >> 9) - 1);
 366                                 bio_endio(mbio, 0);
 367                         }
 368                 }
 369         }
 370         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 371
 372         /*
 373          * Let's see if all mirrored write operations have finished
 374          * already.
 375          */
 376         if (atomic_dec_and_test(&r1_bio->remaining)) {
 377                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 378                         /* free extra copy of the data pages */
 379                         int i = bio->bi_vcnt;
 380                         while (i--)
 381                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 382                 }
 383                 /* clear the bitmap if all writes complete successfully */
 384                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 385                                 r1_bio->sectors,
 386                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
 387                                 behind);
 388                 md_write_end(r1_bio->mddev);
 389                 raid_end_bio_io(r1_bio);
 390         }
 391
 392         if (to_put)
 393                 bio_put(to_put);
 394 }
 395
 396
 397 /*
 398  * This routine returns the disk from which the requested read should
 399  * be done. There is a per-array 'next expected sequential IO' sector
 400  * number - if this matches on the next IO then we use the last disk.
 401  * There is also a per-disk 'last know head position' sector that is
 402  * maintained from IRQ contexts, both the normal and the resync IO
 403  * completion handlers update this position correctly. If there is no
 404  * perfect sequential match then we pick the disk whose head is closest.
 405  *
 406  * If there are 2 mirrors in the same 2 devices, performance degrades
 407  * because position is mirror, not device based.
 408  *
 409  * The rdev for the device selected will have nr_pending incremented.
 410  */
 411 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 412 {
 413         const sector_t this_sector = r1_bio->sector;
 414         int new_disk = conf->last_used, disk = new_disk;
 415         int wonly_disk = -1;
 416         const int sectors = r1_bio->sectors;
 417         sector_t new_distance, current_distance;
 418         mdk_rdev_t *rdev;
 419
 420         rcu_read_lock();
 421         /*
 422          * Check if we can balance. We can balance on the whole
 423          * device if no resync is going on, or below the resync window.
 424          * We take the first readable disk when above the resync window.
 425          */
 426  retry:
 427         if (conf->mddev->recovery_cp < MaxSector &&
 428             (this_sector + sectors >= conf->next_resync)) {
 429                 /* Choose the first operational device, for consistancy */
 430                 new_disk = 0;
 431
 432                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 433                      r1_bio->bios[new_disk] == IO_BLOCKED ||
 434                      !rdev || !test_bit(In_sync, &rdev->flags)
 435                              || test_bit(WriteMostly, &rdev->flags);
 436                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
 437
 438                         if (rdev && test_bit(In_sync, &rdev->flags) &&
 439                                 r1_bio->bios[new_disk] != IO_BLOCKED)
 440                                 wonly_disk = new_disk;
 441
 442                         if (new_disk == conf->raid_disks - 1) {
 443                                 new_disk = wonly_disk;
 444                                 break;
 445                         }
 446                 }
 447                 goto rb_out;
 448         }
 449
 450
 451         /* make sure the disk is operational */
 452         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 453              r1_bio->bios[new_disk] == IO_BLOCKED ||
 454              !rdev || !test_bit(In_sync, &rdev->flags) ||
 455                      test_bit(WriteMostly, &rdev->flags);
 456              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
 457
 458                 if (rdev && test_bit(In_sync, &rdev->flags) &&
 459                     r1_bio->bios[new_disk] != IO_BLOCKED)
 460                         wonly_disk = new_disk;
 461
 462                 if (new_disk <= 0)
 463                         new_disk = conf->raid_disks;
 464                 new_disk--;
 465                 if (new_disk == disk) {
 466                         new_disk = wonly_disk;
 467                         break;
 468                 }
 469         }
 470
 471         if (new_disk < 0)
 472                 goto rb_out;
 473
 474         disk = new_disk;
 475         /* now disk == new_disk == starting point for search */
 476
 477         /*
 478          * Don't change to another disk for sequential reads:
 479          */
 480         if (conf->next_seq_sect == this_sector)
 481                 goto rb_out;
 482         if (this_sector == conf->mirrors[new_disk].head_position)
 483                 goto rb_out;
 484
 485         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 486
 487         /* Find the disk whose head is closest */
 488
 489         do {
 490                 if (disk <= 0)
 491                         disk = conf->raid_disks;
 492                 disk--;
 493
 494                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 495
 496                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
 497                     !test_bit(In_sync, &rdev->flags) ||
 498                     test_bit(WriteMostly, &rdev->flags))
 499                         continue;
 500
 501                 if (!atomic_read(&rdev->nr_pending)) {
 502                         new_disk = disk;
 503                         break;
 504                 }
 505                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 506                 if (new_distance < current_distance) {
 507                         current_distance = new_distance;
 508                         new_disk = disk;
 509                 }
 510         } while (disk != conf->last_used);
 511
 512  rb_out:
 513
 514
 515         if (new_disk >= 0) {
 516                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 517                 if (!rdev)
 518                         goto retry;
 519                 atomic_inc(&rdev->nr_pending);
 520                 if (!test_bit(In_sync, &rdev->flags)) {
 521                         /* cannot risk returning a device that failed
 522                          * before we inc'ed nr_pending
 523                          */
 524                         rdev_dec_pending(rdev, conf->mddev);
 525                         goto retry;
 526                 }
 527                 conf->next_seq_sect = this_sector + sectors;
 528                 conf->last_used = new_disk;
 529         }
 530         rcu_read_unlock();
 531
 532         return new_disk;
 533 }
 534
 535 static void unplug_slaves(mddev_t *mddev)
 536 {
 537         conf_t *conf = mddev->private;
 538         int i;
 539
 540         rcu_read_lock();
 541         for (i=0; i<mddev->raid_disks; i++) {
 542                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 543                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 544                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 545
 546                         atomic_inc(&rdev->nr_pending);
 547                         rcu_read_unlock();
 548
 549                         blk_unplug(r_queue);
 550
 551                         rdev_dec_pending(rdev, mddev);
 552                         rcu_read_lock();
 553                 }
 554         }
 555         rcu_read_unlock();
 556 }
 557
 558 static void raid1_unplug(struct request_queue *q)
 559 {
 560         mddev_t *mddev = q->queuedata;
 561
 562         unplug_slaves(mddev);
 563         md_wakeup_thread(mddev->thread);
 564 }
 565
 566 static int raid1_congested(void *data, int bits)
 567 {
 568         mddev_t *mddev = data;
 569         conf_t *conf = mddev->private;
 570         int i, ret = 0;
 571
 572         if (mddev_congested(mddev, bits))
 573                 return 1;
 574
 575         rcu_read_lock();
 576         for (i = 0; i < mddev->raid_disks; i++) {
 577                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 578                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 579                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 580
 581                         /* Note the '|| 1' - when read_balance prefers
 582                          * non-congested targets, it can be removed
 583                          */
 584                         if ((bits & (1<<BDI_async_congested)) || 1)
 585                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 586                         else
 587                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 588                 }
 589         }
 590         rcu_read_unlock();
 591         return ret;
 592 }
 593
 594
 595 static int flush_pending_writes(conf_t *conf)
 596 {
 597         /* Any writes that have been queued but are awaiting
 598          * bitmap updates get flushed here.
 599          * We return 1 if any requests were actually submitted.
 600          */
 601         int rv = 0;
 602
 603         spin_lock_irq(&conf->device_lock);
 604
 605         if (conf->pending_bio_list.head) {
 606                 struct bio *bio;
 607                 bio = bio_list_get(&conf->pending_bio_list);
 608                 blk_remove_plug(conf->mddev->queue);
 609                 spin_unlock_irq(&conf->device_lock);
 610                 /* flush any pending bitmap writes to
 611                  * disk before proceeding w/ I/O */
 612                 bitmap_unplug(conf->mddev->bitmap);
 613
 614                 while (bio) { /* submit pending writes */
 615                         struct bio *next = bio->bi_next;
 616                         bio->bi_next = NULL;
 617                         generic_make_request(bio);
 618                         bio = next;
 619                 }
 620                 rv = 1;
 621         } else
 622                 spin_unlock_irq(&conf->device_lock);
 623         return rv;
 624 }
 625
 626 /* Barriers....
 627  * Sometimes we need to suspend IO while we do something else,
 628  * either some resync/recovery, or reconfigure the array.
 629  * To do this we raise a 'barrier'.
 630  * The 'barrier' is a counter that can be raised multiple times
 631  * to count how many activities are happening which preclude
 632  * normal IO.
 633  * We can only raise the barrier if there is no pending IO.
 634  * i.e. if nr_pending == 0.
 635  * We choose only to raise the barrier if no-one is waiting for the
 636  * barrier to go down.  This means that as soon as an IO request
 637  * is ready, no other operations which require a barrier will start
 638  * until the IO request has had a chance.
 639  *
 640  * So: regular IO calls 'wait_barrier'.  When that returns there
 641  *    is no backgroup IO happening,  It must arrange to call
 642  *    allow_barrier when it has finished its IO.
 643  * backgroup IO calls must call raise_barrier.  Once that returns
 644  *    there is no normal IO happeing.  It must arrange to call
 645  *    lower_barrier when the particular background IO completes.
 646  */
 647 #define RESYNC_DEPTH 32
 648
 649 static void raise_barrier(conf_t *conf)
 650 {
 651         spin_lock_irq(&conf->resync_lock);
 652
 653         /* Wait until no block IO is waiting */
 654         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 655                             conf->resync_lock,
 656                             raid1_unplug(conf->mddev->queue));
 657
 658         /* block any new IO from starting */
 659         conf->barrier++;
 660
 661         /* No wait for all pending IO to complete */
 662         wait_event_lock_irq(conf->wait_barrier,
 663                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 664                             conf->resync_lock,
 665                             raid1_unplug(conf->mddev->queue));
 666
 667         spin_unlock_irq(&conf->resync_lock);
 668 }
 669
 670 static void lower_barrier(conf_t *conf)
 671 {
 672         unsigned long flags;
 673         BUG_ON(conf->barrier <= 0);
 674         spin_lock_irqsave(&conf->resync_lock, flags);
 675         conf->barrier--;
 676         spin_unlock_irqrestore(&conf->resync_lock, flags);
 677         wake_up(&conf->wait_barrier);
 678 }
 679
 680 static void wait_barrier(conf_t *conf)
 681 {
 682         spin_lock_irq(&conf->resync_lock);
 683         if (conf->barrier) {
 684                 conf->nr_waiting++;
 685                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 686                                     conf->resync_lock,
 687                                     raid1_unplug(conf->mddev->queue));
 688                 conf->nr_waiting--;
 689         }
 690         conf->nr_pending++;
 691         spin_unlock_irq(&conf->resync_lock);
 692 }
 693
 694 static void allow_barrier(conf_t *conf)
 695 {
 696         unsigned long flags;
 697         spin_lock_irqsave(&conf->resync_lock, flags);
 698         conf->nr_pending--;
 699         spin_unlock_irqrestore(&conf->resync_lock, flags);
 700         wake_up(&conf->wait_barrier);
 701 }
 702
 703 static void freeze_array(conf_t *conf)
 704 {
 705         /* stop syncio and normal IO and wait for everything to
 706          * go quite.
 707          * We increment barrier and nr_waiting, and then
 708          * wait until nr_pending match nr_queued+1
 709          * This is called in the context of one normal IO request
 710          * that has failed. Thus any sync request that might be pending
 711          * will be blocked by nr_pending, and we need to wait for
 712          * pending IO requests to complete or be queued for re-try.
 713          * Thus the number queued (nr_queued) plus this request (1)
 714          * must match the number of pending IOs (nr_pending) before
 715          * we continue.
 716          */
 717         spin_lock_irq(&conf->resync_lock);
 718         conf->barrier++;
 719         conf->nr_waiting++;
 720         wait_event_lock_irq(conf->wait_barrier,
 721                             conf->nr_pending == conf->nr_queued+1,
 722                             conf->resync_lock,
 723                             ({ flush_pending_writes(conf);
 724                                raid1_unplug(conf->mddev->queue); }));
 725         spin_unlock_irq(&conf->resync_lock);
 726 }
 727 static void unfreeze_array(conf_t *conf)
 728 {
 729         /* reverse the effect of the freeze */
 730         spin_lock_irq(&conf->resync_lock);
 731         conf->barrier--;
 732         conf->nr_waiting--;
 733         wake_up(&conf->wait_barrier);
 734         spin_unlock_irq(&conf->resync_lock);
 735 }
 736
 737
 738 /* duplicate the data pages for behind I/O */
 739 static struct page **alloc_behind_pages(struct bio *bio)
 740 {
 741         int i;
 742         struct bio_vec *bvec;
 743         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 744                                         GFP_NOIO);
 745         if (unlikely(!pages))
 746                 goto do_sync_io;
 747
 748         bio_for_each_segment(bvec, bio, i) {
 749                 pages[i] = alloc_page(GFP_NOIO);
 750                 if (unlikely(!pages[i]))
 751                         goto do_sync_io;
 752                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 753                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 754                 kunmap(pages[i]);
 755                 kunmap(bvec->bv_page);
 756         }
 757
 758         return pages;
 759
 760 do_sync_io:
 761         if (pages)
 762                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 763                         put_page(pages[i]);
 764         kfree(pages);
 765         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 766         return NULL;
 767 }
 768
 769 static int make_request(mddev_t *mddev, struct bio * bio)
 770 {
 771         conf_t *conf = mddev->private;
 772         mirror_info_t *mirror;
 773         r1bio_t *r1_bio;
 774         struct bio *read_bio;
 775         int i, targets = 0, disks;
 776         struct bitmap *bitmap;
 777         unsigned long flags;
 778         struct bio_list bl;
 779         struct page **behind_pages = NULL;
 780         const int rw = bio_data_dir(bio);
 781         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 782         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 783         mdk_rdev_t *blocked_rdev;
 784
 785         /*
 786          * Register the new request and wait if the reconstruction
 787          * thread has put up a bar for new requests.
 788          * Continue immediately if no resync is active currently.
 789          */
 790
 791         md_write_start(mddev, bio); /* wait on superblock update early */
 792
 793         if (bio_data_dir(bio) == WRITE &&
 794             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 795             bio->bi_sector < mddev->suspend_hi) {
 796                 /* As the suspend_* range is controlled by
 797                  * userspace, we want an interruptible
 798                  * wait.
 799                  */
 800                 DEFINE_WAIT(w);
 801                 for (;;) {
 802                         flush_signals(current);
 803                         prepare_to_wait(&conf->wait_barrier,
 804                                         &w, TASK_INTERRUPTIBLE);
 805                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 806                             bio->bi_sector >= mddev->suspend_hi)
 807                                 break;
 808                         schedule();
 809                 }
 810                 finish_wait(&conf->wait_barrier, &w);
 811         }
 812
 813         wait_barrier(conf);
 814
 815         bitmap = mddev->bitmap;
 816
 817         /*
 818          * make_request() can abort the operation when READA is being
 819          * used and no empty request is available.
 820          *
 821          */
 822         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 823
 824         r1_bio->master_bio = bio;
 825         r1_bio->sectors = bio->bi_size >> 9;
 826         r1_bio->state = 0;
 827         r1_bio->mddev = mddev;
 828         r1_bio->sector = bio->bi_sector;
 829
 830         if (rw == READ) {
 831                 /*
 832                  * read balancing logic:
 833                  */
 834                 int rdisk = read_balance(conf, r1_bio);
 835
 836                 if (rdisk < 0) {
 837                         /* couldn't find anywhere to read from */
 838                         raid_end_bio_io(r1_bio);
 839                         return 0;
 840                 }
 841                 mirror = conf->mirrors + rdisk;
 842
 843                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 844                     bitmap) {
 845                         /* Reading from a write-mostly device must
 846                          * take care not to over-take any writes
 847                          * that are 'behind'
 848                          */
 849                         wait_event(bitmap->behind_wait,
 850                                    atomic_read(&bitmap->behind_writes) == 0);
 851                 }
 852                 r1_bio->read_disk = rdisk;
 853
 854                 read_bio = bio_clone(bio, GFP_NOIO);
 855
 856                 r1_bio->bios[rdisk] = read_bio;
 857
 858                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 859                 read_bio->bi_bdev = mirror->rdev->bdev;
 860                 read_bio->bi_end_io = raid1_end_read_request;
 861                 read_bio->bi_rw = READ | do_sync;
 862                 read_bio->bi_private = r1_bio;
 863
 864                 generic_make_request(read_bio);
 865                 return 0;
 866         }
 867
 868         /*
 869          * WRITE:
 870          */
 871         /* first select target devices under spinlock and
 872          * inc refcount on their rdev.  Record them by setting
 873          * bios[x] to bio
 874          */
 875         disks = conf->raid_disks;
 876 #if 0
 877         { static int first=1;
 878         if (first) printk("First Write sector %llu disks %d\n",
 879                           (unsigned long long)r1_bio->sector, disks);
 880         first = 0;
 881         }
 882 #endif
 883  retry_write:
 884         blocked_rdev = NULL;
 885         rcu_read_lock();
 886         for (i = 0;  i < disks; i++) {
 887                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 888                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 889                         atomic_inc(&rdev->nr_pending);
 890                         blocked_rdev = rdev;
 891                         break;
 892                 }
 893                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 894                         atomic_inc(&rdev->nr_pending);
 895                         if (test_bit(Faulty, &rdev->flags)) {
 896                                 rdev_dec_pending(rdev, mddev);
 897                                 r1_bio->bios[i] = NULL;
 898                         } else {
 899                                 r1_bio->bios[i] = bio;
 900                                 targets++;
 901                         }
 902                 } else
 903                         r1_bio->bios[i] = NULL;
 904         }
 905         rcu_read_unlock();
 906
 907         if (unlikely(blocked_rdev)) {
 908                 /* Wait for this device to become unblocked */
 909                 int j;
 910
 911                 for (j = 0; j < i; j++)
 912                         if (r1_bio->bios[j])
 913                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 914
 915                 allow_barrier(conf);
 916                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 917                 wait_barrier(conf);
 918                 goto retry_write;
 919         }
 920
 921         BUG_ON(targets == 0); /* we never fail the last device */
 922
 923         if (targets < conf->raid_disks) {
 924                 /* array is degraded, we will not clear the bitmap
 925                  * on I/O completion (see raid1_end_write_request) */
 926                 set_bit(R1BIO_Degraded, &r1_bio->state);
 927         }
 928
 929         /* do behind I/O ?
 930          * Not if there are too many, or cannot allocate memory,
 931          * or a reader on WriteMostly is waiting for behind writes
 932          * to flush */
 933         if (bitmap &&
 934             (atomic_read(&bitmap->behind_writes)
 935              < mddev->bitmap_info.max_write_behind) &&
 936             !waitqueue_active(&bitmap->behind_wait) &&
 937             (behind_pages = alloc_behind_pages(bio)) != NULL)
 938                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 939
 940         atomic_set(&r1_bio->remaining, 0);
 941         atomic_set(&r1_bio->behind_remaining, 0);
 942
 943         bio_list_init(&bl);
 944         for (i = 0; i < disks; i++) {
 945                 struct bio *mbio;
 946                 if (!r1_bio->bios[i])
 947                         continue;
 948
 949                 mbio = bio_clone(bio, GFP_NOIO);
 950                 r1_bio->bios[i] = mbio;
 951
 952                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 953                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 954                 mbio->bi_end_io = raid1_end_write_request;
 955                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
 956                 mbio->bi_private = r1_bio;
 957
 958                 if (behind_pages) {
 959                         struct bio_vec *bvec;
 960                         int j;
 961
 962                         /* Yes, I really want the '__' version so that
 963                          * we clear any unused pointer in the io_vec, rather
 964                          * than leave them unchanged.  This is important
 965                          * because when we come to free the pages, we won't
 966                          * know the originial bi_idx, so we just free
 967                          * them all
 968                          */
 969                         __bio_for_each_segment(bvec, mbio, j, 0)
 970                                 bvec->bv_page = behind_pages[j];
 971                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 972                                 atomic_inc(&r1_bio->behind_remaining);
 973                 }
 974
 975                 atomic_inc(&r1_bio->remaining);
 976
 977                 bio_list_add(&bl, mbio);
 978         }
 979         kfree(behind_pages); /* the behind pages are attached to the bios now */
 980
 981         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 982                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 983         spin_lock_irqsave(&conf->device_lock, flags);
 984         bio_list_merge(&conf->pending_bio_list, &bl);
 985         bio_list_init(&bl);
 986
 987         blk_plug_device(mddev->queue);
 988         spin_unlock_irqrestore(&conf->device_lock, flags);
 989
 990         /* In case raid1d snuck into freeze_array */
 991         wake_up(&conf->wait_barrier);
 992
 993         if (do_sync)
 994                 md_wakeup_thread(mddev->thread);
 995 #if 0
 996         while ((bio = bio_list_pop(&bl)) != NULL)
 997                 generic_make_request(bio);
 998 #endif
 999
1000         return 0;
1001 }
1002
1003 static void status(struct seq_file *seq, mddev_t *mddev)
1004 {
1005         conf_t *conf = mddev->private;
1006         int i;
1007
1008         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1009                    conf->raid_disks - mddev->degraded);
1010         rcu_read_lock();
1011         for (i = 0; i < conf->raid_disks; i++) {
1012                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1013                 seq_printf(seq, "%s",
1014                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1015         }
1016         rcu_read_unlock();
1017         seq_printf(seq, "]");
1018 }
1019
1020
1021 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1022 {
1023         char b[BDEVNAME_SIZE];
1024         conf_t *conf = mddev->private;
1025
1026         /*
1027          * If it is not operational, then we have already marked it as dead
1028          * else if it is the last working disks, ignore the error, let the
1029          * next level up know.
1030          * else mark the drive as failed
1031          */
1032         if (test_bit(In_sync, &rdev->flags)
1033             && (conf->raid_disks - mddev->degraded) == 1) {
1034                 /*
1035                  * Don't fail the drive, act as though we were just a
1036                  * normal single drive.
1037                  * However don't try a recovery from this drive as
1038                  * it is very likely to fail.
1039                  */
1040                 mddev->recovery_disabled = 1;
1041                 return;
1042         }
1043         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1044                 unsigned long flags;
1045                 spin_lock_irqsave(&conf->device_lock, flags);
1046                 mddev->degraded++;
1047                 set_bit(Faulty, &rdev->flags);
1048                 spin_unlock_irqrestore(&conf->device_lock, flags);
1049                 /*
1050                  * if recovery is running, make sure it aborts.
1051                  */
1052                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1053         } else
1054                 set_bit(Faulty, &rdev->flags);
1055         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1056         printk(KERN_ALERT "md/raid1:%s: Disk failure on %s, disabling device.\n"
1057                KERN_ALERT "md/raid1:%s: Operation continuing on %d devices.\n",
1058                mdname(mddev), bdevname(rdev->bdev, b),
1059                mdname(mddev), conf->raid_disks - mddev->degraded);
1060 }
1061
1062 static void print_conf(conf_t *conf)
1063 {
1064         int i;
1065
1066         printk(KERN_DEBUG "RAID1 conf printout:\n");
1067         if (!conf) {
1068                 printk(KERN_DEBUG "(!conf)\n");
1069                 return;
1070         }
1071         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1072                 conf->raid_disks);
1073
1074         rcu_read_lock();
1075         for (i = 0; i < conf->raid_disks; i++) {
1076                 char b[BDEVNAME_SIZE];
1077                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1078                 if (rdev)
1079                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1080                                i, !test_bit(In_sync, &rdev->flags),
1081                                !test_bit(Faulty, &rdev->flags),
1082                                bdevname(rdev->bdev,b));
1083         }
1084         rcu_read_unlock();
1085 }
1086
1087 static void close_sync(conf_t *conf)
1088 {
1089         wait_barrier(conf);
1090         allow_barrier(conf);
1091
1092         mempool_destroy(conf->r1buf_pool);
1093         conf->r1buf_pool = NULL;
1094 }
1095
1096 static int raid1_spare_active(mddev_t *mddev)
1097 {
1098         int i;
1099         conf_t *conf = mddev->private;
1100         int count = 0;
1101         unsigned long flags;
1102
1103         /*
1104          * Find all failed disks within the RAID1 configuration
1105          * and mark them readable.
1106          * Called under mddev lock, so rcu protection not needed.
1107          */
1108         for (i = 0; i < conf->raid_disks; i++) {
1109                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1110                 if (rdev
1111                     && !test_bit(Faulty, &rdev->flags)
1112                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1113                         count++;
1114                         sysfs_notify_dirent(rdev->sysfs_state);
1115                 }
1116         }
1117         spin_lock_irqsave(&conf->device_lock, flags);
1118         mddev->degraded -= count;
1119         spin_unlock_irqrestore(&conf->device_lock, flags);
1120
1121         print_conf(conf);
1122         return count;
1123 }
1124
1125
1126 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1127 {
1128         conf_t *conf = mddev->private;
1129         int err = -EEXIST;
1130         int mirror = 0;
1131         mirror_info_t *p;
1132         int first = 0;
1133         int last = mddev->raid_disks - 1;
1134
1135         if (rdev->raid_disk >= 0)
1136                 first = last = rdev->raid_disk;
1137
1138         for (mirror = first; mirror <= last; mirror++)
1139                 if ( !(p=conf->mirrors+mirror)->rdev) {
1140
1141                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1142                                           rdev->data_offset << 9);
1143                         /* as we don't honour merge_bvec_fn, we must
1144                          * never risk violating it, so limit
1145                          * ->max_segments to one lying with a single
1146                          * page, as a one page request is never in
1147                          * violation.
1148                          */
1149                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1150                                 blk_queue_max_segments(mddev->queue, 1);
1151                                 blk_queue_segment_boundary(mddev->queue,
1152                                                            PAGE_CACHE_SIZE - 1);
1153                         }
1154
1155                         p->head_position = 0;
1156                         rdev->raid_disk = mirror;
1157                         err = 0;
1158                         /* As all devices are equivalent, we don't need a full recovery
1159                          * if this was recently any drive of the array
1160                          */
1161                         if (rdev->saved_raid_disk < 0)
1162                                 conf->fullsync = 1;
1163                         rcu_assign_pointer(p->rdev, rdev);
1164                         break;
1165                 }
1166         md_integrity_add_rdev(rdev, mddev);
1167         print_conf(conf);
1168         return err;
1169 }
1170
1171 static int raid1_remove_disk(mddev_t *mddev, int number)
1172 {
1173         conf_t *conf = mddev->private;
1174         int err = 0;
1175         mdk_rdev_t *rdev;
1176         mirror_info_t *p = conf->mirrors+ number;
1177
1178         print_conf(conf);
1179         rdev = p->rdev;
1180         if (rdev) {
1181                 if (test_bit(In_sync, &rdev->flags) ||
1182                     atomic_read(&rdev->nr_pending)) {
1183                         err = -EBUSY;
1184                         goto abort;
1185                 }
1186                 /* Only remove non-faulty devices is recovery
1187                  * is not possible.
1188                  */
1189                 if (!test_bit(Faulty, &rdev->flags) &&
1190                     mddev->degraded < conf->raid_disks) {
1191                         err = -EBUSY;
1192                         goto abort;
1193                 }
1194                 p->rdev = NULL;
1195                 synchronize_rcu();
1196                 if (atomic_read(&rdev->nr_pending)) {
1197                         /* lost the race, try later */
1198                         err = -EBUSY;
1199                         p->rdev = rdev;
1200                         goto abort;
1201                 }
1202                 md_integrity_register(mddev);
1203         }
1204 abort:
1205
1206         print_conf(conf);
1207         return err;
1208 }
1209
1210
1211 static void end_sync_read(struct bio *bio, int error)
1212 {
1213         r1bio_t *r1_bio = bio->bi_private;
1214         int i;
1215
1216         for (i=r1_bio->mddev->raid_disks; i--; )
1217                 if (r1_bio->bios[i] == bio)
1218                         break;
1219         BUG_ON(i < 0);
1220         update_head_pos(i, r1_bio);
1221         /*
1222          * we have read a block, now it needs to be re-written,
1223          * or re-read if the read failed.
1224          * We don't do much here, just schedule handling by raid1d
1225          */
1226         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1227                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1228
1229         if (atomic_dec_and_test(&r1_bio->remaining))
1230                 reschedule_retry(r1_bio);
1231 }
1232
1233 static void end_sync_write(struct bio *bio, int error)
1234 {
1235         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1236         r1bio_t *r1_bio = bio->bi_private;
1237         mddev_t *mddev = r1_bio->mddev;
1238         conf_t *conf = mddev->private;
1239         int i;
1240         int mirror=0;
1241
1242         for (i = 0; i < conf->raid_disks; i++)
1243                 if (r1_bio->bios[i] == bio) {
1244                         mirror = i;
1245                         break;
1246                 }
1247         if (!uptodate) {
1248                 int sync_blocks = 0;
1249                 sector_t s = r1_bio->sector;
1250                 long sectors_to_go = r1_bio->sectors;
1251                 /* make sure these bits doesn't get cleared. */
1252                 do {
1253                         bitmap_end_sync(mddev->bitmap, s,
1254                                         &sync_blocks, 1);
1255                         s += sync_blocks;
1256                         sectors_to_go -= sync_blocks;
1257                 } while (sectors_to_go > 0);
1258                 md_error(mddev, conf->mirrors[mirror].rdev);
1259         }
1260
1261         update_head_pos(mirror, r1_bio);
1262
1263         if (atomic_dec_and_test(&r1_bio->remaining)) {
1264                 sector_t s = r1_bio->sectors;
1265                 put_buf(r1_bio);
1266                 md_done_sync(mddev, s, uptodate);
1267         }
1268 }
1269
1270 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1271 {
1272         conf_t *conf = mddev->private;
1273         int i;
1274         int disks = conf->raid_disks;
1275         struct bio *bio, *wbio;
1276
1277         bio = r1_bio->bios[r1_bio->read_disk];
1278
1279
1280         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1281                 /* We have read all readable devices.  If we haven't
1282                  * got the block, then there is no hope left.
1283                  * If we have, then we want to do a comparison
1284                  * and skip the write if everything is the same.
1285                  * If any blocks failed to read, then we need to
1286                  * attempt an over-write
1287                  */
1288                 int primary;
1289                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1290                         for (i=0; i<mddev->raid_disks; i++)
1291                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1292                                         md_error(mddev, conf->mirrors[i].rdev);
1293
1294                         md_done_sync(mddev, r1_bio->sectors, 1);
1295                         put_buf(r1_bio);
1296                         return;
1297                 }
1298                 for (primary=0; primary<mddev->raid_disks; primary++)
1299                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1300                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1301                                 r1_bio->bios[primary]->bi_end_io = NULL;
1302                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1303                                 break;
1304                         }
1305                 r1_bio->read_disk = primary;
1306                 for (i=0; i<mddev->raid_disks; i++)
1307                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1308                                 int j;
1309                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1310                                 struct bio *pbio = r1_bio->bios[primary];
1311                                 struct bio *sbio = r1_bio->bios[i];
1312
1313                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1314                                         for (j = vcnt; j-- ; ) {
1315                                                 struct page *p, *s;
1316                                                 p = pbio->bi_io_vec[j].bv_page;
1317                                                 s = sbio->bi_io_vec[j].bv_page;
1318                                                 if (memcmp(page_address(p),
1319                                                            page_address(s),
1320                                                            PAGE_SIZE))
1321                                                         break;
1322                                         }
1323                                 } else
1324                                         j = 0;
1325                                 if (j >= 0)
1326                                         mddev->resync_mismatches += r1_bio->sectors;
1327                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1328                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1329                                         sbio->bi_end_io = NULL;
1330                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1331                                 } else {
1332                                         /* fixup the bio for reuse */
1333                                         int size;
1334                                         sbio->bi_vcnt = vcnt;
1335                                         sbio->bi_size = r1_bio->sectors << 9;
1336                                         sbio->bi_idx = 0;
1337                                         sbio->bi_phys_segments = 0;
1338                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1339                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1340                                         sbio->bi_next = NULL;
1341                                         sbio->bi_sector = r1_bio->sector +
1342                                                 conf->mirrors[i].rdev->data_offset;
1343                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1344                                         size = sbio->bi_size;
1345                                         for (j = 0; j < vcnt ; j++) {
1346                                                 struct bio_vec *bi;
1347                                                 bi = &sbio->bi_io_vec[j];
1348                                                 bi->bv_offset = 0;
1349                                                 if (size > PAGE_SIZE)
1350                                                         bi->bv_len = PAGE_SIZE;
1351                                                 else
1352                                                         bi->bv_len = size;
1353                                                 size -= PAGE_SIZE;
1354                                                 memcpy(page_address(bi->bv_page),
1355                                                        page_address(pbio->bi_io_vec[j].bv_page),
1356                                                        PAGE_SIZE);
1357                                         }
1358
1359                                 }
1360                         }
1361         }
1362         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1363                 /* ouch - failed to read all of that.
1364                  * Try some synchronous reads of other devices to get
1365                  * good data, much like with normal read errors.  Only
1366                  * read into the pages we already have so we don't
1367                  * need to re-issue the read request.
1368                  * We don't need to freeze the array, because being in an
1369                  * active sync request, there is no normal IO, and
1370                  * no overlapping syncs.
1371                  */
1372                 sector_t sect = r1_bio->sector;
1373                 int sectors = r1_bio->sectors;
1374                 int idx = 0;
1375
1376                 while(sectors) {
1377                         int s = sectors;
1378                         int d = r1_bio->read_disk;
1379                         int success = 0;
1380                         mdk_rdev_t *rdev;
1381
1382                         if (s > (PAGE_SIZE>>9))
1383                                 s = PAGE_SIZE >> 9;
1384                         do {
1385                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1386                                         /* No rcu protection needed here devices
1387                                          * can only be removed when no resync is
1388                                          * active, and resync is currently active
1389                                          */
1390                                         rdev = conf->mirrors[d].rdev;
1391                                         if (sync_page_io(rdev->bdev,
1392                                                          sect + rdev->data_offset,
1393                                                          s<<9,
1394                                                          bio->bi_io_vec[idx].bv_page,
1395                                                          READ)) {
1396                                                 success = 1;
1397                                                 break;
1398                                         }
1399                                 }
1400                                 d++;
1401                                 if (d == conf->raid_disks)
1402                                         d = 0;
1403                         } while (!success && d != r1_bio->read_disk);
1404
1405                         if (success) {
1406                                 int start = d;
1407                                 /* write it back and re-read */
1408                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1409                                 while (d != r1_bio->read_disk) {
1410                                         if (d == 0)
1411                                                 d = conf->raid_disks;
1412                                         d--;
1413                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1414                                                 continue;
1415                                         rdev = conf->mirrors[d].rdev;
1416                                         atomic_add(s, &rdev->corrected_errors);
1417                                         if (sync_page_io(rdev->bdev,
1418                                                          sect + rdev->data_offset,
1419                                                          s<<9,
1420                                                          bio->bi_io_vec[idx].bv_page,
1421                                                          WRITE) == 0)
1422                                                 md_error(mddev, rdev);
1423                                 }
1424                                 d = start;
1425                                 while (d != r1_bio->read_disk) {
1426                                         if (d == 0)
1427                                                 d = conf->raid_disks;
1428                                         d--;
1429                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1430                                                 continue;
1431                                         rdev = conf->mirrors[d].rdev;
1432                                         if (sync_page_io(rdev->bdev,
1433                                                          sect + rdev->data_offset,
1434                                                          s<<9,
1435                                                          bio->bi_io_vec[idx].bv_page,
1436                                                          READ) == 0)
1437                                                 md_error(mddev, rdev);
1438                                 }
1439                         } else {
1440                                 char b[BDEVNAME_SIZE];
1441                                 /* Cannot read from anywhere, array is toast */
1442                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1443                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1444                                        " for block %llu\n",
1445                                        mdname(mddev),
1446                                        bdevname(bio->bi_bdev, b),
1447                                        (unsigned long long)r1_bio->sector);
1448                                 md_done_sync(mddev, r1_bio->sectors, 0);
1449                                 put_buf(r1_bio);
1450                                 return;
1451                         }
1452                         sectors -= s;
1453                         sect += s;
1454                         idx ++;
1455                 }
1456         }
1457
1458         /*
1459          * schedule writes
1460          */
1461         atomic_set(&r1_bio->remaining, 1);
1462         for (i = 0; i < disks ; i++) {
1463                 wbio = r1_bio->bios[i];
1464                 if (wbio->bi_end_io == NULL ||
1465                     (wbio->bi_end_io == end_sync_read &&
1466                      (i == r1_bio->read_disk ||
1467                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1468                         continue;
1469
1470                 wbio->bi_rw = WRITE;
1471                 wbio->bi_end_io = end_sync_write;
1472                 atomic_inc(&r1_bio->remaining);
1473                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1474
1475                 generic_make_request(wbio);
1476         }
1477
1478         if (atomic_dec_and_test(&r1_bio->remaining)) {
1479                 /* if we're here, all write(s) have completed, so clean up */
1480                 md_done_sync(mddev, r1_bio->sectors, 1);
1481                 put_buf(r1_bio);
1482         }
1483 }
1484
1485 /*
1486  * This is a kernel thread which:
1487  *
1488  *      1.      Retries failed read operations on working mirrors.
1489  *      2.      Updates the raid superblock when problems encounter.
1490  *      3.      Performs writes following reads for array syncronising.
1491  */
1492
1493 static void fix_read_error(conf_t *conf, int read_disk,
1494                            sector_t sect, int sectors)
1495 {
1496         mddev_t *mddev = conf->mddev;
1497         while(sectors) {
1498                 int s = sectors;
1499                 int d = read_disk;
1500                 int success = 0;
1501                 int start;
1502                 mdk_rdev_t *rdev;
1503
1504                 if (s > (PAGE_SIZE>>9))
1505                         s = PAGE_SIZE >> 9;
1506
1507                 do {
1508                         /* Note: no rcu protection needed here
1509                          * as this is synchronous in the raid1d thread
1510                          * which is the thread that might remove
1511                          * a device.  If raid1d ever becomes multi-threaded....
1512                          */
1513                         rdev = conf->mirrors[d].rdev;
1514                         if (rdev &&
1515                             test_bit(In_sync, &rdev->flags) &&
1516                             sync_page_io(rdev->bdev,
1517                                          sect + rdev->data_offset,
1518                                          s<<9,
1519                                          conf->tmppage, READ))
1520                                 success = 1;
1521                         else {
1522                                 d++;
1523                                 if (d == conf->raid_disks)
1524                                         d = 0;
1525                         }
1526                 } while (!success && d != read_disk);
1527
1528                 if (!success) {
1529                         /* Cannot read from anywhere -- bye bye array */
1530                         md_error(mddev, conf->mirrors[read_disk].rdev);
1531                         break;
1532                 }
1533                 /* write it back and re-read */
1534                 start = d;
1535                 while (d != read_disk) {
1536                         if (d==0)
1537                                 d = conf->raid_disks;
1538                         d--;
1539                         rdev = conf->mirrors[d].rdev;
1540                         if (rdev &&
1541                             test_bit(In_sync, &rdev->flags)) {
1542                                 if (sync_page_io(rdev->bdev,
1543                                                  sect + rdev->data_offset,
1544                                                  s<<9, conf->tmppage, WRITE)
1545                                     == 0)
1546                                         /* Well, this device is dead */
1547                                         md_error(mddev, rdev);
1548                         }
1549                 }
1550                 d = start;
1551                 while (d != read_disk) {
1552                         char b[BDEVNAME_SIZE];
1553                         if (d==0)
1554                                 d = conf->raid_disks;
1555                         d--;
1556                         rdev = conf->mirrors[d].rdev;
1557                         if (rdev &&
1558                             test_bit(In_sync, &rdev->flags)) {
1559                                 if (sync_page_io(rdev->bdev,
1560                                                  sect + rdev->data_offset,
1561                                                  s<<9, conf->tmppage, READ)
1562                                     == 0)
1563                                         /* Well, this device is dead */
1564                                         md_error(mddev, rdev);
1565                                 else {
1566                                         atomic_add(s, &rdev->corrected_errors);
1567                                         printk(KERN_INFO
1568                                                "md/raid1:%s: read error corrected "
1569                                                "(%d sectors at %llu on %s)\n",
1570                                                mdname(mddev), s,
1571                                                (unsigned long long)(sect +
1572                                                    rdev->data_offset),
1573                                                bdevname(rdev->bdev, b));
1574                                 }
1575                         }
1576                 }
1577                 sectors -= s;
1578                 sect += s;
1579         }
1580 }
1581
1582 static void raid1d(mddev_t *mddev)
1583 {
1584         r1bio_t *r1_bio;
1585         struct bio *bio;
1586         unsigned long flags;
1587         conf_t *conf = mddev->private;
1588         struct list_head *head = &conf->retry_list;
1589         int unplug=0;
1590         mdk_rdev_t *rdev;
1591
1592         md_check_recovery(mddev);
1593
1594         for (;;) {
1595                 char b[BDEVNAME_SIZE];
1596
1597                 unplug += flush_pending_writes(conf);
1598
1599                 spin_lock_irqsave(&conf->device_lock, flags);
1600                 if (list_empty(head)) {
1601                         spin_unlock_irqrestore(&conf->device_lock, flags);
1602                         break;
1603                 }
1604                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1605                 list_del(head->prev);
1606                 conf->nr_queued--;
1607                 spin_unlock_irqrestore(&conf->device_lock, flags);
1608
1609                 mddev = r1_bio->mddev;
1610                 conf = mddev->private;
1611                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1612                         sync_request_write(mddev, r1_bio);
1613                         unplug = 1;
1614                 } else {
1615                         int disk;
1616
1617                         /* we got a read error. Maybe the drive is bad.  Maybe just
1618                          * the block and we can fix it.
1619                          * We freeze all other IO, and try reading the block from
1620                          * other devices.  When we find one, we re-write
1621                          * and check it that fixes the read error.
1622                          * This is all done synchronously while the array is
1623                          * frozen
1624                          */
1625                         if (mddev->ro == 0) {
1626                                 freeze_array(conf);
1627                                 fix_read_error(conf, r1_bio->read_disk,
1628                                                r1_bio->sector,
1629                                                r1_bio->sectors);
1630                                 unfreeze_array(conf);
1631                         } else
1632                                 md_error(mddev,
1633                                          conf->mirrors[r1_bio->read_disk].rdev);
1634
1635                         bio = r1_bio->bios[r1_bio->read_disk];
1636                         if ((disk=read_balance(conf, r1_bio)) == -1) {
1637                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1638                                        " read error for block %llu\n",
1639                                        mdname(mddev),
1640                                        bdevname(bio->bi_bdev,b),
1641                                        (unsigned long long)r1_bio->sector);
1642                                 raid_end_bio_io(r1_bio);
1643                         } else {
1644                                 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1645                                 r1_bio->bios[r1_bio->read_disk] =
1646                                         mddev->ro ? IO_BLOCKED : NULL;
1647                                 r1_bio->read_disk = disk;
1648                                 bio_put(bio);
1649                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1650                                 r1_bio->bios[r1_bio->read_disk] = bio;
1651                                 rdev = conf->mirrors[disk].rdev;
1652                                 if (printk_ratelimit())
1653                                         printk(KERN_ERR "md/raid1:%s: redirecting sector %llu to"
1654                                                " other mirror: %s\n",
1655                                                mdname(mddev),
1656                                                (unsigned long long)r1_bio->sector,
1657                                                bdevname(rdev->bdev,b));
1658                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1659                                 bio->bi_bdev = rdev->bdev;
1660                                 bio->bi_end_io = raid1_end_read_request;
1661                                 bio->bi_rw = READ | do_sync;
1662                                 bio->bi_private = r1_bio;
1663                                 unplug = 1;
1664                                 generic_make_request(bio);
1665                         }
1666                 }
1667                 cond_resched();
1668         }
1669         if (unplug)
1670                 unplug_slaves(mddev);
1671 }
1672
1673
1674 static int init_resync(conf_t *conf)
1675 {
1676         int buffs;
1677
1678         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1679         BUG_ON(conf->r1buf_pool);
1680         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1681                                           conf->poolinfo);
1682         if (!conf->r1buf_pool)
1683                 return -ENOMEM;
1684         conf->next_resync = 0;
1685         return 0;
1686 }
1687
1688 /*
1689  * perform a "sync" on one "block"
1690  *
1691  * We need to make sure that no normal I/O request - particularly write
1692  * requests - conflict with active sync requests.
1693  *
1694  * This is achieved by tracking pending requests and a 'barrier' concept
1695  * that can be installed to exclude normal IO requests.
1696  */
1697
1698 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1699 {
1700         conf_t *conf = mddev->private;
1701         r1bio_t *r1_bio;
1702         struct bio *bio;
1703         sector_t max_sector, nr_sectors;
1704         int disk = -1;
1705         int i;
1706         int wonly = -1;
1707         int write_targets = 0, read_targets = 0;
1708         int sync_blocks;
1709         int still_degraded = 0;
1710
1711         if (!conf->r1buf_pool)
1712                 if (init_resync(conf))
1713                         return 0;
1714
1715         max_sector = mddev->dev_sectors;
1716         if (sector_nr >= max_sector) {
1717                 /* If we aborted, we need to abort the
1718                  * sync on the 'current' bitmap chunk (there will
1719                  * only be one in raid1 resync.
1720                  * We can find the current addess in mddev->curr_resync
1721                  */
1722                 if (mddev->curr_resync < max_sector) /* aborted */
1723                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1724                                                 &sync_blocks, 1);
1725                 else /* completed sync */
1726                         conf->fullsync = 0;
1727
1728                 bitmap_close_sync(mddev->bitmap);
1729                 close_sync(conf);
1730                 return 0;
1731         }
1732
1733         if (mddev->bitmap == NULL &&
1734             mddev->recovery_cp == MaxSector &&
1735             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1736             conf->fullsync == 0) {
1737                 *skipped = 1;
1738                 return max_sector - sector_nr;
1739         }
1740         /* before building a request, check if we can skip these blocks..
1741          * This call the bitmap_start_sync doesn't actually record anything
1742          */
1743         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1744             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1745                 /* We can skip this block, and probably several more */
1746                 *skipped = 1;
1747                 return sync_blocks;
1748         }
1749         /*
1750          * If there is non-resync activity waiting for a turn,
1751          * and resync is going fast enough,
1752          * then let it though before starting on this new sync request.
1753          */
1754         if (!go_faster && conf->nr_waiting)
1755                 msleep_interruptible(1000);
1756
1757         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1758         raise_barrier(conf);
1759
1760         conf->next_resync = sector_nr;
1761
1762         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1763         rcu_read_lock();
1764         /*
1765          * If we get a correctably read error during resync or recovery,
1766          * we might want to read from a different device.  So we
1767          * flag all drives that could conceivably be read from for READ,
1768          * and any others (which will be non-In_sync devices) for WRITE.
1769          * If a read fails, we try reading from something else for which READ
1770          * is OK.
1771          */
1772
1773         r1_bio->mddev = mddev;
1774         r1_bio->sector = sector_nr;
1775         r1_bio->state = 0;
1776         set_bit(R1BIO_IsSync, &r1_bio->state);
1777
1778         for (i=0; i < conf->raid_disks; i++) {
1779                 mdk_rdev_t *rdev;
1780                 bio = r1_bio->bios[i];
1781
1782                 /* take from bio_init */
1783                 bio->bi_next = NULL;
1784                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1785                 bio->bi_flags |= 1 << BIO_UPTODATE;
1786                 bio->bi_comp_cpu = -1;
1787                 bio->bi_rw = READ;
1788                 bio->bi_vcnt = 0;
1789                 bio->bi_idx = 0;
1790                 bio->bi_phys_segments = 0;
1791                 bio->bi_size = 0;
1792                 bio->bi_end_io = NULL;
1793                 bio->bi_private = NULL;
1794
1795                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1796                 if (rdev == NULL ||
1797                            test_bit(Faulty, &rdev->flags)) {
1798                         still_degraded = 1;
1799                         continue;
1800                 } else if (!test_bit(In_sync, &rdev->flags)) {
1801                         bio->bi_rw = WRITE;
1802                         bio->bi_end_io = end_sync_write;
1803                         write_targets ++;
1804                 } else {
1805                         /* may need to read from here */
1806                         bio->bi_rw = READ;
1807                         bio->bi_end_io = end_sync_read;
1808                         if (test_bit(WriteMostly, &rdev->flags)) {
1809                                 if (wonly < 0)
1810                                         wonly = i;
1811                         } else {
1812                                 if (disk < 0)
1813                                         disk = i;
1814                         }
1815                         read_targets++;
1816                 }
1817                 atomic_inc(&rdev->nr_pending);
1818                 bio->bi_sector = sector_nr + rdev->data_offset;
1819                 bio->bi_bdev = rdev->bdev;
1820                 bio->bi_private = r1_bio;
1821         }
1822         rcu_read_unlock();
1823         if (disk < 0)
1824                 disk = wonly;
1825         r1_bio->read_disk = disk;
1826
1827         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1828                 /* extra read targets are also write targets */
1829                 write_targets += read_targets-1;
1830
1831         if (write_targets == 0 || read_targets == 0) {
1832                 /* There is nowhere to write, so all non-sync
1833                  * drives must be failed - so we are finished
1834                  */
1835                 sector_t rv = max_sector - sector_nr;
1836                 *skipped = 1;
1837                 put_buf(r1_bio);
1838                 return rv;
1839         }
1840
1841         if (max_sector > mddev->resync_max)
1842                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1843         nr_sectors = 0;
1844         sync_blocks = 0;
1845         do {
1846                 struct page *page;
1847                 int len = PAGE_SIZE;
1848                 if (sector_nr + (len>>9) > max_sector)
1849                         len = (max_sector - sector_nr) << 9;
1850                 if (len == 0)
1851                         break;
1852                 if (sync_blocks == 0) {
1853                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1854                                                &sync_blocks, still_degraded) &&
1855                             !conf->fullsync &&
1856                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1857                                 break;
1858                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1859                         if ((len >> 9) > sync_blocks)
1860                                 len = sync_blocks<<9;
1861                 }
1862
1863                 for (i=0 ; i < conf->raid_disks; i++) {
1864                         bio = r1_bio->bios[i];
1865                         if (bio->bi_end_io) {
1866                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1867                                 if (bio_add_page(bio, page, len, 0) == 0) {
1868                                         /* stop here */
1869                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1870                                         while (i > 0) {
1871                                                 i--;
1872                                                 bio = r1_bio->bios[i];
1873                                                 if (bio->bi_end_io==NULL)
1874                                                         continue;
1875                                                 /* remove last page from this bio */
1876                                                 bio->bi_vcnt--;
1877                                                 bio->bi_size -= len;
1878                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1879                                         }
1880                                         goto bio_full;
1881                                 }
1882                         }
1883                 }
1884                 nr_sectors += len>>9;
1885                 sector_nr += len>>9;
1886                 sync_blocks -= (len>>9);
1887         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1888  bio_full:
1889         r1_bio->sectors = nr_sectors;
1890
1891         /* For a user-requested sync, we read all readable devices and do a
1892          * compare
1893          */
1894         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1895                 atomic_set(&r1_bio->remaining, read_targets);
1896                 for (i=0; i<conf->raid_disks; i++) {
1897                         bio = r1_bio->bios[i];
1898                         if (bio->bi_end_io == end_sync_read) {
1899                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1900                                 generic_make_request(bio);
1901                         }
1902                 }
1903         } else {
1904                 atomic_set(&r1_bio->remaining, 1);
1905                 bio = r1_bio->bios[r1_bio->read_disk];
1906                 md_sync_acct(bio->bi_bdev, nr_sectors);
1907                 generic_make_request(bio);
1908
1909         }
1910         return nr_sectors;
1911 }
1912
1913 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1914 {
1915         if (sectors)
1916                 return sectors;
1917
1918         return mddev->dev_sectors;
1919 }
1920
1921 static conf_t *setup_conf(mddev_t *mddev)
1922 {
1923         conf_t *conf;
1924         int i;
1925         mirror_info_t *disk;
1926         mdk_rdev_t *rdev;
1927         int err = -ENOMEM;
1928
1929         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1930         if (!conf)
1931                 goto abort;
1932
1933         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1934                                  GFP_KERNEL);
1935         if (!conf->mirrors)
1936                 goto abort;
1937
1938         conf->tmppage = alloc_page(GFP_KERNEL);
1939         if (!conf->tmppage)
1940                 goto abort;
1941
1942         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1943         if (!conf->poolinfo)
1944                 goto abort;
1945         conf->poolinfo->raid_disks = mddev->raid_disks;
1946         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1947                                           r1bio_pool_free,
1948                                           conf->poolinfo);
1949         if (!conf->r1bio_pool)
1950                 goto abort;
1951
1952         conf->poolinfo->mddev = mddev;
1953
1954         spin_lock_init(&conf->device_lock);
1955         list_for_each_entry(rdev, &mddev->disks, same_set) {
1956                 int disk_idx = rdev->raid_disk;
1957                 if (disk_idx >= mddev->raid_disks
1958                     || disk_idx < 0)
1959                         continue;
1960                 disk = conf->mirrors + disk_idx;
1961
1962                 disk->rdev = rdev;
1963
1964                 disk->head_position = 0;
1965         }
1966         conf->raid_disks = mddev->raid_disks;
1967         conf->mddev = mddev;
1968         INIT_LIST_HEAD(&conf->retry_list);
1969
1970         spin_lock_init(&conf->resync_lock);
1971         init_waitqueue_head(&conf->wait_barrier);
1972
1973         bio_list_init(&conf->pending_bio_list);
1974         bio_list_init(&conf->flushing_bio_list);
1975
1976         conf->last_used = -1;
1977         for (i = 0; i < conf->raid_disks; i++) {
1978
1979                 disk = conf->mirrors + i;
1980
1981                 if (!disk->rdev ||
1982                     !test_bit(In_sync, &disk->rdev->flags)) {
1983                         disk->head_position = 0;
1984                         if (disk->rdev)
1985                                 conf->fullsync = 1;
1986                 } else if (conf->last_used < 0)
1987                         /*
1988                          * The first working device is used as a
1989                          * starting point to read balancing.
1990                          */
1991                         conf->last_used = i;
1992         }
1993
1994         err = -EIO;
1995         if (conf->last_used < 0) {
1996                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
1997                        mdname(mddev));
1998                 goto abort;
1999         }
2000         err = -ENOMEM;
2001         conf->thread = md_register_thread(raid1d, mddev, NULL);
2002         if (!conf->thread) {
2003                 printk(KERN_ERR
2004                        "md/raid1:%s: couldn't allocate thread\n",
2005                        mdname(mddev));
2006                 goto abort;
2007         }
2008
2009         return conf;
2010
2011  abort:
2012         if (conf) {
2013                 if (conf->r1bio_pool)
2014                         mempool_destroy(conf->r1bio_pool);
2015                 kfree(conf->mirrors);
2016                 safe_put_page(conf->tmppage);
2017                 kfree(conf->poolinfo);
2018                 kfree(conf);
2019         }
2020         return ERR_PTR(err);
2021 }
2022
2023 static int run(mddev_t *mddev)
2024 {
2025         conf_t *conf;
2026         int i;
2027         mdk_rdev_t *rdev;
2028
2029         if (mddev->level != 1) {
2030                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2031                        mdname(mddev), mddev->level);
2032                 return -EIO;
2033         }
2034         if (mddev->reshape_position != MaxSector) {
2035                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2036                        mdname(mddev));
2037                 return -EIO;
2038         }
2039         /*
2040          * copy the already verified devices into our private RAID1
2041          * bookkeeping area. [whatever we allocate in run(),
2042          * should be freed in stop()]
2043          */
2044         if (mddev->private == NULL)
2045                 conf = setup_conf(mddev);
2046         else
2047                 conf = mddev->private;
2048
2049         if (IS_ERR(conf))
2050                 return PTR_ERR(conf);
2051
2052         mddev->queue->queue_lock = &conf->device_lock;
2053         list_for_each_entry(rdev, &mddev->disks, same_set) {
2054                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2055                                   rdev->data_offset << 9);
2056                 /* as we don't honour merge_bvec_fn, we must never risk
2057                  * violating it, so limit ->max_segments to 1 lying within
2058                  * a single page, as a one page request is never in violation.
2059                  */
2060                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2061                         blk_queue_max_segments(mddev->queue, 1);
2062                         blk_queue_segment_boundary(mddev->queue,
2063                                                    PAGE_CACHE_SIZE - 1);
2064                 }
2065         }
2066
2067         mddev->degraded = 0;
2068         for (i=0; i < conf->raid_disks; i++)
2069                 if (conf->mirrors[i].rdev == NULL ||
2070                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2071                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2072                         mddev->degraded++;
2073
2074         if (conf->raid_disks - mddev->degraded == 1)
2075                 mddev->recovery_cp = MaxSector;
2076
2077         if (mddev->recovery_cp != MaxSector)
2078                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2079                        " -- starting background reconstruction\n",
2080                        mdname(mddev));
2081         printk(KERN_INFO
2082                 "md/raid1:%s: active with %d out of %d mirrors\n",
2083                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2084                 mddev->raid_disks);
2085
2086         /*
2087          * Ok, everything is just fine now
2088          */
2089         mddev->thread = conf->thread;
2090         conf->thread = NULL;
2091         mddev->private = conf;
2092
2093         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2094
2095         mddev->queue->unplug_fn = raid1_unplug;
2096         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2097         mddev->queue->backing_dev_info.congested_data = mddev;
2098         md_integrity_register(mddev);
2099         return 0;
2100 }
2101
2102 static int stop(mddev_t *mddev)
2103 {
2104         conf_t *conf = mddev->private;
2105         struct bitmap *bitmap = mddev->bitmap;
2106
2107         /* wait for behind writes to complete */
2108         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2109                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2110                        mdname(mddev));
2111                 /* need to kick something here to make sure I/O goes? */
2112                 wait_event(bitmap->behind_wait,
2113                            atomic_read(&bitmap->behind_writes) == 0);
2114         }
2115
2116         raise_barrier(conf);
2117         lower_barrier(conf);
2118
2119         md_unregister_thread(mddev->thread);
2120         mddev->thread = NULL;
2121         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2122         if (conf->r1bio_pool)
2123                 mempool_destroy(conf->r1bio_pool);
2124         kfree(conf->mirrors);
2125         kfree(conf->poolinfo);
2126         kfree(conf);
2127         mddev->private = NULL;
2128         return 0;
2129 }
2130
2131 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2132 {
2133         /* no resync is happening, and there is enough space
2134          * on all devices, so we can resize.
2135          * We need to make sure resync covers any new space.
2136          * If the array is shrinking we should possibly wait until
2137          * any io in the removed space completes, but it hardly seems
2138          * worth it.
2139          */
2140         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2141         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2142                 return -EINVAL;
2143         set_capacity(mddev->gendisk, mddev->array_sectors);
2144         revalidate_disk(mddev->gendisk);
2145         if (sectors > mddev->dev_sectors &&
2146             mddev->recovery_cp == MaxSector) {
2147                 mddev->recovery_cp = mddev->dev_sectors;
2148                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2149         }
2150         mddev->dev_sectors = sectors;
2151         mddev->resync_max_sectors = sectors;
2152         return 0;
2153 }
2154
2155 static int raid1_reshape(mddev_t *mddev)
2156 {
2157         /* We need to:
2158          * 1/ resize the r1bio_pool
2159          * 2/ resize conf->mirrors
2160          *
2161          * We allocate a new r1bio_pool if we can.
2162          * Then raise a device barrier and wait until all IO stops.
2163          * Then resize conf->mirrors and swap in the new r1bio pool.
2164          *
2165          * At the same time, we "pack" the devices so that all the missing
2166          * devices have the higher raid_disk numbers.
2167          */
2168         mempool_t *newpool, *oldpool;
2169         struct pool_info *newpoolinfo;
2170         mirror_info_t *newmirrors;
2171         conf_t *conf = mddev->private;
2172         int cnt, raid_disks;
2173         unsigned long flags;
2174         int d, d2, err;
2175
2176         /* Cannot change chunk_size, layout, or level */
2177         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2178             mddev->layout != mddev->new_layout ||
2179             mddev->level != mddev->new_level) {
2180                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2181                 mddev->new_layout = mddev->layout;
2182                 mddev->new_level = mddev->level;
2183                 return -EINVAL;
2184         }
2185
2186         err = md_allow_write(mddev);
2187         if (err)
2188                 return err;
2189
2190         raid_disks = mddev->raid_disks + mddev->delta_disks;
2191
2192         if (raid_disks < conf->raid_disks) {
2193                 cnt=0;
2194                 for (d= 0; d < conf->raid_disks; d++)
2195                         if (conf->mirrors[d].rdev)
2196                                 cnt++;
2197                 if (cnt > raid_disks)
2198                         return -EBUSY;
2199         }
2200
2201         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2202         if (!newpoolinfo)
2203                 return -ENOMEM;
2204         newpoolinfo->mddev = mddev;
2205         newpoolinfo->raid_disks = raid_disks;
2206
2207         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2208                                  r1bio_pool_free, newpoolinfo);
2209         if (!newpool) {
2210                 kfree(newpoolinfo);
2211                 return -ENOMEM;
2212         }
2213         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2214         if (!newmirrors) {
2215                 kfree(newpoolinfo);
2216                 mempool_destroy(newpool);
2217                 return -ENOMEM;
2218         }
2219
2220         raise_barrier(conf);
2221
2222         /* ok, everything is stopped */
2223         oldpool = conf->r1bio_pool;
2224         conf->r1bio_pool = newpool;
2225
2226         for (d = d2 = 0; d < conf->raid_disks; d++) {
2227                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2228                 if (rdev && rdev->raid_disk != d2) {
2229                         char nm[20];
2230                         sprintf(nm, "rd%d", rdev->raid_disk);
2231                         sysfs_remove_link(&mddev->kobj, nm);
2232                         rdev->raid_disk = d2;
2233                         sprintf(nm, "rd%d", rdev->raid_disk);
2234                         sysfs_remove_link(&mddev->kobj, nm);
2235                         if (sysfs_create_link(&mddev->kobj,
2236                                               &rdev->kobj, nm))
2237                                 printk(KERN_WARNING
2238                                        "md/raid1:%s: cannot register "
2239                                        "%s\n",
2240                                        mdname(mddev), nm);
2241                 }
2242                 if (rdev)
2243                         newmirrors[d2++].rdev = rdev;
2244         }
2245         kfree(conf->mirrors);
2246         conf->mirrors = newmirrors;
2247         kfree(conf->poolinfo);
2248         conf->poolinfo = newpoolinfo;
2249
2250         spin_lock_irqsave(&conf->device_lock, flags);
2251         mddev->degraded += (raid_disks - conf->raid_disks);
2252         spin_unlock_irqrestore(&conf->device_lock, flags);
2253         conf->raid_disks = mddev->raid_disks = raid_disks;
2254         mddev->delta_disks = 0;
2255
2256         conf->last_used = 0; /* just make sure it is in-range */
2257         lower_barrier(conf);
2258
2259         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2260         md_wakeup_thread(mddev->thread);
2261
2262         mempool_destroy(oldpool);
2263         return 0;
2264 }
2265
2266 static void raid1_quiesce(mddev_t *mddev, int state)
2267 {
2268         conf_t *conf = mddev->private;
2269
2270         switch(state) {
2271         case 2: /* wake for suspend */
2272                 wake_up(&conf->wait_barrier);
2273                 break;
2274         case 1:
2275                 raise_barrier(conf);
2276                 break;
2277         case 0:
2278                 lower_barrier(conf);
2279                 break;
2280         }
2281 }
2282
2283 static void *raid1_takeover(mddev_t *mddev)
2284 {
2285         /* raid1 can take over:
2286          *  raid5 with 2 devices, any layout or chunk size
2287          */
2288         if (mddev->level == 5 && mddev->raid_disks == 2) {
2289                 conf_t *conf;
2290                 mddev->new_level = 1;
2291                 mddev->new_layout = 0;
2292                 mddev->new_chunk_sectors = 0;
2293                 conf = setup_conf(mddev);
2294                 if (!IS_ERR(conf))
2295                         conf->barrier = 1;
2296                 return conf;
2297         }
2298         return ERR_PTR(-EINVAL);
2299 }
2300
2301 static struct mdk_personality raid1_personality =
2302 {
2303         .name           = "raid1",
2304         .level          = 1,
2305         .owner          = THIS_MODULE,
2306         .make_request   = make_request,
2307         .run            = run,
2308         .stop           = stop,
2309         .status         = status,
2310         .error_handler  = error,
2311         .hot_add_disk   = raid1_add_disk,
2312         .hot_remove_disk= raid1_remove_disk,
2313         .spare_active   = raid1_spare_active,
2314         .sync_request   = sync_request,
2315         .resize         = raid1_resize,
2316         .size           = raid1_size,
2317         .check_reshape  = raid1_reshape,
2318         .quiesce        = raid1_quiesce,
2319         .takeover       = raid1_takeover,
2320 };
2321
2322 static int __init raid_init(void)
2323 {
2324         return register_md_personality(&raid1_personality);
2325 }
2326
2327 static void raid_exit(void)
2328 {
2329         unregister_md_personality(&raid1_personality);
2330 }
2331
2332 module_init(raid_init);
2333 module_exit(raid_exit);
2334 MODULE_LICENSE("GPL");
2335 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2336 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2337 MODULE_ALIAS("md-raid1");
2338 MODULE_ALIAS("md-level-1");