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  * This program is free software; you can redistribute it and/or modify
  16  * it under the terms of the GNU General Public License as published by
  17  * the Free Software Foundation; either version 2, or (at your option)
  18  * any later version.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * (for example /usr/src/linux/COPYING); if not, write to the Free
  22  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  23  */
  24
  25 #include <linux/raid/raid1.h>
  26
  27 #define MAJOR_NR MD_MAJOR
  28 #define MD_DRIVER
  29 #define MD_PERSONALITY
  30
  31 /*
  32  * Number of guaranteed r1bios in case of extreme VM load:
  33  */
  34 #define NR_RAID1_BIOS 256
  35
  36 static mdk_personality_t raid1_personality;
  37 static spinlock_t retry_list_lock = SPIN_LOCK_UNLOCKED;
  38 static LIST_HEAD(retry_list_head);
  39
  40 static void * r1bio_pool_alloc(int gfp_flags, void *data)
  41 {
  42         mddev_t *mddev = data;
  43         r1bio_t *r1_bio;
  44
  45         /* allocate a r1bio with room for raid_disks entries in the write_bios array */
  46         r1_bio = kmalloc(sizeof(r1bio_t) + sizeof(struct bio*)*mddev->raid_disks,
  47                          gfp_flags);
  48         if (r1_bio)
  49                 memset(r1_bio, 0, sizeof(*r1_bio) + sizeof(struct bio*)*mddev->raid_disks);
  50
  51         return r1_bio;
  52 }
  53
  54 static void r1bio_pool_free(void *r1_bio, void *data)
  55 {
  56         kfree(r1_bio);
  57 }
  58
  59 //#define RESYNC_BLOCK_SIZE (64*1024)
  60 #define RESYNC_BLOCK_SIZE PAGE_SIZE
  61 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  62 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  63 #define RESYNC_WINDOW (2048*1024)
  64
  65 static void * r1buf_pool_alloc(int gfp_flags, void *data)
  66 {
  67         conf_t *conf = data;
  68         struct page *page;
  69         r1bio_t *r1_bio;
  70         struct bio *bio;
  71         int i, j;
  72
  73         r1_bio = r1bio_pool_alloc(gfp_flags, conf->mddev);
  74         if (!r1_bio)
  75                 return NULL;
  76         bio = bio_alloc(gfp_flags, RESYNC_PAGES);
  77         if (!bio)
  78                 goto out_free_r1_bio;
  79
  80         for (i = 0; i < RESYNC_PAGES; i++) {
  81                 page = alloc_page(gfp_flags);
  82                 if (unlikely(!page))
  83                         goto out_free_pages;
  84
  85                 bio->bi_io_vec[i].bv_page = page;
  86                 bio->bi_io_vec[i].bv_len = PAGE_SIZE;
  87                 bio->bi_io_vec[i].bv_offset = 0;
  88         }
  89
  90         /*
  91          * Allocate a single data page for this iovec.
  92          */
  93         bio->bi_vcnt = RESYNC_PAGES;
  94         bio->bi_idx = 0;
  95         bio->bi_size = RESYNC_BLOCK_SIZE;
  96         bio->bi_end_io = NULL;
  97         atomic_set(&bio->bi_cnt, 1);
  98
  99         r1_bio->master_bio = bio;
 100
 101         return r1_bio;
 102
 103 out_free_pages:
 104         for (j = 0; j < i; j++)
 105                 __free_page(bio->bi_io_vec[j].bv_page);
 106         bio_put(bio);
 107 out_free_r1_bio:
 108         r1bio_pool_free(r1_bio, conf->mddev);
 109         return NULL;
 110 }
 111
 112 static void r1buf_pool_free(void *__r1_bio, void *data)
 113 {
 114         int i;
 115         conf_t *conf = data;
 116         r1bio_t *r1bio = __r1_bio;
 117         struct bio *bio = r1bio->master_bio;
 118
 119         if (atomic_read(&bio->bi_cnt) != 1)
 120                 BUG();
 121         for (i = 0; i < RESYNC_PAGES; i++) {
 122                 __free_page(bio->bi_io_vec[i].bv_page);
 123                 bio->bi_io_vec[i].bv_page = NULL;
 124         }
 125         if (atomic_read(&bio->bi_cnt) != 1)
 126                 BUG();
 127         bio_put(bio);
 128         r1bio_pool_free(r1bio, conf->mddev);
 129 }
 130
 131 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 132 {
 133         int i;
 134
 135         if (r1_bio->read_bio) {
 136                 if (atomic_read(&r1_bio->read_bio->bi_cnt) != 1)
 137                         BUG();
 138                 bio_put(r1_bio->read_bio);
 139                 r1_bio->read_bio = NULL;
 140         }
 141         for (i = 0; i < conf->raid_disks; i++) {
 142                 struct bio **bio = r1_bio->write_bios + i;
 143                 if (*bio) {
 144                         if (atomic_read(&(*bio)->bi_cnt) != 1)
 145                                 BUG();
 146                         bio_put(*bio);
 147                 }
 148                 *bio = NULL;
 149         }
 150 }
 151
 152 static inline void free_r1bio(r1bio_t *r1_bio)
 153 {
 154         unsigned long flags;
 155
 156         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 157
 158         /*
 159          * Wake up any possible resync thread that waits for the device
 160          * to go idle.
 161          */
 162         spin_lock_irqsave(&conf->resync_lock, flags);
 163         if (!--conf->nr_pending) {
 164                 wake_up(&conf->wait_idle);
 165                 wake_up(&conf->wait_resume);
 166         }
 167         spin_unlock_irqrestore(&conf->resync_lock, flags);
 168
 169         put_all_bios(conf, r1_bio);
 170         mempool_free(r1_bio, conf->r1bio_pool);
 171 }
 172
 173 static inline void put_buf(r1bio_t *r1_bio)
 174 {
 175         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 176         struct bio *bio = r1_bio->master_bio;
 177         unsigned long flags;
 178
 179         /*
 180          * undo any possible partial request fixup magic:
 181          */
 182         if (bio->bi_size != RESYNC_BLOCK_SIZE)
 183                 bio->bi_io_vec[bio->bi_vcnt-1].bv_len = PAGE_SIZE;
 184         put_all_bios(conf, r1_bio);
 185         mempool_free(r1_bio, conf->r1buf_pool);
 186
 187         spin_lock_irqsave(&conf->resync_lock, flags);
 188         if (!conf->barrier)
 189                 BUG();
 190         --conf->barrier;
 191         wake_up(&conf->wait_resume);
 192         wake_up(&conf->wait_idle);
 193
 194         if (!--conf->nr_pending) {
 195                 wake_up(&conf->wait_idle);
 196                 wake_up(&conf->wait_resume);
 197         }
 198         spin_unlock_irqrestore(&conf->resync_lock, flags);
 199 }
 200
 201 static int map(mddev_t *mddev, mdk_rdev_t **rdevp)
 202 {
 203         conf_t *conf = mddev_to_conf(mddev);
 204         int i, disks = conf->raid_disks;
 205
 206         /*
 207          * Later we do read balancing on the read side
 208          * now we use the first available disk.
 209          */
 210
 211         spin_lock_irq(&conf->device_lock);
 212         for (i = 0; i < disks; i++) {
 213                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 214                 if (rdev && rdev->in_sync) {
 215                         *rdevp = rdev;
 216                         atomic_inc(&rdev->nr_pending);
 217                         spin_unlock_irq(&conf->device_lock);
 218                         return 0;
 219                 }
 220         }
 221         spin_unlock_irq(&conf->device_lock);
 222
 223         printk(KERN_ERR "raid1_map(): huh, no more operational devices?\n");
 224         return -1;
 225 }
 226
 227 static void reschedule_retry(r1bio_t *r1_bio)
 228 {
 229         unsigned long flags;
 230         mddev_t *mddev = r1_bio->mddev;
 231
 232         spin_lock_irqsave(&retry_list_lock, flags);
 233         list_add(&r1_bio->retry_list, &retry_list_head);
 234         spin_unlock_irqrestore(&retry_list_lock, flags);
 235
 236         md_wakeup_thread(mddev->thread);
 237 }
 238
 239 /*
 240  * raid_end_bio_io() is called when we have finished servicing a mirrored
 241  * operation and are ready to return a success/failure code to the buffer
 242  * cache layer.
 243  */
 244 static void raid_end_bio_io(r1bio_t *r1_bio)
 245 {
 246         struct bio *bio = r1_bio->master_bio;
 247
 248         bio_endio(bio, bio->bi_size,
 249                 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 250         free_r1bio(r1_bio);
 251 }
 252
 253 /*
 254  * Update disk head position estimator based on IRQ completion info.
 255  */
 256 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 257 {
 258         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 259
 260         conf->mirrors[disk].head_position =
 261                 r1_bio->sector + (r1_bio->master_bio->bi_size >> 9);
 262 }
 263
 264 static int raid1_end_request(struct bio *bio, unsigned int bytes_done, int error)
 265 {
 266         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 267         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 268         int mirror;
 269         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 270
 271         if (bio->bi_size)
 272                 return 1;
 273
 274         if (r1_bio->cmd == READ || r1_bio->cmd == READA)
 275                 mirror = r1_bio->read_disk;
 276         else {
 277                 for (mirror = 0; mirror < conf->raid_disks; mirror++)
 278                         if (r1_bio->write_bios[mirror] == bio)
 279                                 break;
 280         }
 281         /*
 282          * this branch is our 'one mirror IO has finished' event handler:
 283          */
 284         if (!uptodate)
 285                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 286         else
 287                 /*
 288                  * Set R1BIO_Uptodate in our master bio, so that
 289                  * we will return a good error code for to the higher
 290                  * levels even if IO on some other mirrored buffer fails.
 291                  *
 292                  * The 'master' represents the composite IO operation to
 293                  * user-side. So if something waits for IO, then it will
 294                  * wait for the 'master' bio.
 295                  */
 296                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 297
 298         update_head_pos(mirror, r1_bio);
 299         if ((r1_bio->cmd == READ) || (r1_bio->cmd == READA)) {
 300                 if (!r1_bio->read_bio)
 301                         BUG();
 302                 /*
 303                  * we have only one bio on the read side
 304                  */
 305                 if (uptodate)
 306                         raid_end_bio_io(r1_bio);
 307                 else {
 308                         /*
 309                          * oops, read error:
 310                          */
 311                         char b[BDEVNAME_SIZE];
 312                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 313                                 bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 314                         reschedule_retry(r1_bio);
 315                 }
 316         } else {
 317
 318                 if (r1_bio->read_bio)
 319                         BUG();
 320                 /*
 321                  * WRITE:
 322                  *
 323                  * Let's see if all mirrored write operations have finished
 324                  * already.
 325                  */
 326                 if (atomic_dec_and_test(&r1_bio->remaining)) {
 327                         md_write_end(r1_bio->mddev);
 328                         raid_end_bio_io(r1_bio);
 329                 }
 330         }
 331         atomic_dec(&conf->mirrors[mirror].rdev->nr_pending);
 332         return 0;
 333 }
 334
 335 /*
 336  * This routine returns the disk from which the requested read should
 337  * be done. There is a per-array 'next expected sequential IO' sector
 338  * number - if this matches on the next IO then we use the last disk.
 339  * There is also a per-disk 'last know head position' sector that is
 340  * maintained from IRQ contexts, both the normal and the resync IO
 341  * completion handlers update this position correctly. If there is no
 342  * perfect sequential match then we pick the disk whose head is closest.
 343  *
 344  * If there are 2 mirrors in the same 2 devices, performance degrades
 345  * because position is mirror, not device based.
 346  *
 347  * The rdev for the device selected will have nr_pending incremented.
 348  */
 349 static int read_balance(conf_t *conf, struct bio *bio, r1bio_t *r1_bio)
 350 {
 351         const unsigned long this_sector = r1_bio->sector;
 352         int new_disk = conf->last_used, disk = new_disk;
 353         const int sectors = bio->bi_size >> 9;
 354         sector_t new_distance, current_distance;
 355
 356         spin_lock_irq(&conf->device_lock);
 357         /*
 358          * Check if it if we can balance. We can balance on the whole
 359          * device if no resync is going on, or below the resync window.
 360          * We take the first readable disk when above the resync window.
 361          */
 362         if (!conf->mddev->in_sync && (this_sector + sectors >= conf->next_resync)) {
 363                 /* make sure that disk is operational */
 364                 new_disk = 0;
 365
 366                 while (!conf->mirrors[new_disk].rdev ||
 367                        !conf->mirrors[new_disk].rdev->in_sync) {
 368                         new_disk++;
 369                         if (new_disk == conf->raid_disks) {
 370                                 new_disk = 0;
 371                                 break;
 372                         }
 373                 }
 374                 goto rb_out;
 375         }
 376
 377
 378         /* make sure the disk is operational */
 379         while (!conf->mirrors[new_disk].rdev ||
 380                !conf->mirrors[new_disk].rdev->in_sync) {
 381                 if (new_disk <= 0)
 382                         new_disk = conf->raid_disks;
 383                 new_disk--;
 384                 if (new_disk == disk) {
 385                         new_disk = conf->last_used;
 386                         goto rb_out;
 387                 }
 388         }
 389         disk = new_disk;
 390         /* now disk == new_disk == starting point for search */
 391
 392         /*
 393          * Don't change to another disk for sequential reads:
 394          */
 395         if (conf->next_seq_sect == this_sector)
 396                 goto rb_out;
 397         if (this_sector == conf->mirrors[new_disk].head_position)
 398                 goto rb_out;
 399
 400         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 401
 402         /* Find the disk whose head is closest */
 403
 404         do {
 405                 if (disk <= 0)
 406                         disk = conf->raid_disks;
 407                 disk--;
 408
 409                 if (!conf->mirrors[disk].rdev ||
 410                     !conf->mirrors[disk].rdev->in_sync)
 411                         continue;
 412
 413                 if (!atomic_read(&conf->mirrors[disk].rdev->nr_pending)) {
 414                         new_disk = disk;
 415                         break;
 416                 }
 417                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 418                 if (new_distance < current_distance) {
 419                         current_distance = new_distance;
 420                         new_disk = disk;
 421                 }
 422         } while (disk != conf->last_used);
 423
 424 rb_out:
 425         r1_bio->read_disk = new_disk;
 426         conf->next_seq_sect = this_sector + sectors;
 427
 428         conf->last_used = new_disk;
 429
 430         if (conf->mirrors[new_disk].rdev)
 431                 atomic_inc(&conf->mirrors[new_disk].rdev->nr_pending);
 432         spin_unlock_irq(&conf->device_lock);
 433
 434         return new_disk;
 435 }
 436
 437 /*
 438  * Throttle resync depth, so that we can both get proper overlapping of
 439  * requests, but are still able to handle normal requests quickly.
 440  */
 441 #define RESYNC_DEPTH 32
 442
 443 static void device_barrier(conf_t *conf, sector_t sect)
 444 {
 445         spin_lock_irq(&conf->resync_lock);
 446         wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume), conf->resync_lock);
 447
 448         if (!conf->barrier++) {
 449                 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending, conf->resync_lock);
 450                 if (conf->nr_pending)
 451                         BUG();
 452         }
 453         wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH, conf->resync_lock);
 454         conf->next_resync = sect;
 455         spin_unlock_irq(&conf->resync_lock);
 456 }
 457
 458 static int make_request(request_queue_t *q, struct bio * bio)
 459 {
 460         mddev_t *mddev = q->queuedata;
 461         conf_t *conf = mddev_to_conf(mddev);
 462         mirror_info_t *mirror;
 463         r1bio_t *r1_bio;
 464         struct bio *read_bio;
 465         int i, disks = conf->raid_disks;
 466
 467         /*
 468          * Register the new request and wait if the reconstruction
 469          * thread has put up a bar for new requests.
 470          * Continue immediately if no resync is active currently.
 471          */
 472         spin_lock_irq(&conf->resync_lock);
 473         wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock);
 474         conf->nr_pending++;
 475         spin_unlock_irq(&conf->resync_lock);
 476
 477         if (bio_data_dir(bio)==WRITE) {
 478                 disk_stat_inc(mddev->gendisk, writes);
 479                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
 480         } else {
 481                 disk_stat_inc(mddev->gendisk, reads);
 482                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
 483         }
 484
 485         /*
 486          * make_request() can abort the operation when READA is being
 487          * used and no empty request is available.
 488          *
 489          */
 490         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 491
 492         r1_bio->master_bio = bio;
 493
 494         r1_bio->mddev = mddev;
 495         r1_bio->sector = bio->bi_sector;
 496         r1_bio->cmd = bio_data_dir(bio);
 497
 498         if (r1_bio->cmd == READ) {
 499                 /*
 500                  * read balancing logic:
 501                  */
 502                 mirror = conf->mirrors + read_balance(conf, bio, r1_bio);
 503
 504                 read_bio = bio_clone(bio, GFP_NOIO);
 505                 if (r1_bio->read_bio)
 506                         BUG();
 507                 r1_bio->read_bio = read_bio;
 508
 509                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 510                 read_bio->bi_bdev = mirror->rdev->bdev;
 511                 read_bio->bi_end_io = raid1_end_request;
 512                 read_bio->bi_rw = r1_bio->cmd;
 513                 read_bio->bi_private = r1_bio;
 514
 515                 generic_make_request(read_bio);
 516                 return 0;
 517         }
 518
 519         /*
 520          * WRITE:
 521          */
 522         /* first select target devices under spinlock and
 523          * inc refcount on their rdev.  Record them by setting
 524          * write_bios[x] to bio
 525          */
 526         spin_lock_irq(&conf->device_lock);
 527         for (i = 0;  i < disks; i++) {
 528                 if (conf->mirrors[i].rdev &&
 529                     !conf->mirrors[i].rdev->faulty) {
 530                         atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 531                         r1_bio->write_bios[i] = bio;
 532                 } else
 533                         r1_bio->write_bios[i] = NULL;
 534         }
 535         spin_unlock_irq(&conf->device_lock);
 536
 537         atomic_set(&r1_bio->remaining, 1);
 538         md_write_start(mddev);
 539         for (i = 0; i < disks; i++) {
 540                 struct bio *mbio;
 541                 if (!r1_bio->write_bios[i])
 542                         continue;
 543
 544                 mbio = bio_clone(bio, GFP_NOIO);
 545                 r1_bio->write_bios[i] = mbio;
 546
 547                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 548                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 549                 mbio->bi_end_io = raid1_end_request;
 550                 mbio->bi_rw = r1_bio->cmd;
 551                 mbio->bi_private = r1_bio;
 552
 553                 atomic_inc(&r1_bio->remaining);
 554                 generic_make_request(mbio);
 555         }
 556
 557         if (atomic_dec_and_test(&r1_bio->remaining)) {
 558                 md_write_end(mddev);
 559                 raid_end_bio_io(r1_bio);
 560         }
 561
 562         return 0;
 563 }
 564
 565 static void status(struct seq_file *seq, mddev_t *mddev)
 566 {
 567         conf_t *conf = mddev_to_conf(mddev);
 568         int i;
 569
 570         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 571                                                 conf->working_disks);
 572         for (i = 0; i < conf->raid_disks; i++)
 573                 seq_printf(seq, "%s",
 574                               conf->mirrors[i].rdev &&
 575                               conf->mirrors[i].rdev->in_sync ? "U" : "_");
 576         seq_printf(seq, "]");
 577 }
 578
 579
 580 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 581 {
 582         char b[BDEVNAME_SIZE];
 583         conf_t *conf = mddev_to_conf(mddev);
 584
 585         /*
 586          * If it is not operational, then we have already marked it as dead
 587          * else if it is the last working disks, ignore the error, let the
 588          * next level up know.
 589          * else mark the drive as failed
 590          */
 591         if (rdev->in_sync
 592             && conf->working_disks == 1)
 593                 /*
 594                  * Don't fail the drive, act as though we were just a
 595                  * normal single drive
 596                  */
 597                 return;
 598         if (rdev->in_sync) {
 599                 mddev->degraded++;
 600                 conf->working_disks--;
 601                 /*
 602                  * if recovery is running, make sure it aborts.
 603                  */
 604                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 605         }
 606         rdev->in_sync = 0;
 607         rdev->faulty = 1;
 608         mddev->sb_dirty = 1;
 609         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
 610                 "       Operation continuing on %d devices\n",
 611                 bdevname(rdev->bdev,b), conf->working_disks);
 612 }
 613
 614 static void print_conf(conf_t *conf)
 615 {
 616         int i;
 617         mirror_info_t *tmp;
 618
 619         printk("RAID1 conf printout:\n");
 620         if (!conf) {
 621                 printk("(!conf)\n");
 622                 return;
 623         }
 624         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 625                 conf->raid_disks);
 626
 627         for (i = 0; i < conf->raid_disks; i++) {
 628                 char b[BDEVNAME_SIZE];
 629                 tmp = conf->mirrors + i;
 630                 if (tmp->rdev)
 631                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 632                                 i, !tmp->rdev->in_sync, !tmp->rdev->faulty,
 633                                 bdevname(tmp->rdev->bdev,b));
 634         }
 635 }
 636
 637 static void close_sync(conf_t *conf)
 638 {
 639         spin_lock_irq(&conf->resync_lock);
 640         wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock);
 641         spin_unlock_irq(&conf->resync_lock);
 642
 643         if (conf->barrier) BUG();
 644         if (waitqueue_active(&conf->wait_idle)) BUG();
 645
 646         mempool_destroy(conf->r1buf_pool);
 647         conf->r1buf_pool = NULL;
 648 }
 649
 650 static int raid1_spare_active(mddev_t *mddev)
 651 {
 652         int i;
 653         conf_t *conf = mddev->private;
 654         mirror_info_t *tmp;
 655
 656         spin_lock_irq(&conf->device_lock);
 657         /*
 658          * Find all failed disks within the RAID1 configuration
 659          * and mark them readable
 660          */
 661         for (i = 0; i < conf->raid_disks; i++) {
 662                 tmp = conf->mirrors + i;
 663                 if (tmp->rdev
 664                     && !tmp->rdev->faulty
 665                     && !tmp->rdev->in_sync) {
 666                         conf->working_disks++;
 667                         mddev->degraded--;
 668                         tmp->rdev->in_sync = 1;
 669                 }
 670         }
 671         spin_unlock_irq(&conf->device_lock);
 672
 673         print_conf(conf);
 674         return 0;
 675 }
 676
 677
 678 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 679 {
 680         conf_t *conf = mddev->private;
 681         int found = 0;
 682         int mirror;
 683         mirror_info_t *p;
 684
 685         spin_lock_irq(&conf->device_lock);
 686         for (mirror=0; mirror < mddev->raid_disks; mirror++)
 687                 if ( !(p=conf->mirrors+mirror)->rdev) {
 688                         p->rdev = rdev;
 689
 690                         blk_queue_stack_limits(mddev->queue,
 691                                                rdev->bdev->bd_disk->queue);
 692                         /* as we don't honour merge_bvec_fn, we must never risk
 693                          * violating it, so limit ->max_sector to one PAGE, as
 694                          * a one page request is never in violation.
 695                          */
 696                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 697                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 698                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
 699
 700                         p->head_position = 0;
 701                         rdev->raid_disk = mirror;
 702                         found = 1;
 703                         break;
 704                 }
 705         spin_unlock_irq(&conf->device_lock);
 706
 707         print_conf(conf);
 708         return found;
 709 }
 710
 711 static int raid1_remove_disk(mddev_t *mddev, int number)
 712 {
 713         conf_t *conf = mddev->private;
 714         int err = 1;
 715         mirror_info_t *p = conf->mirrors+ number;
 716
 717         print_conf(conf);
 718         spin_lock_irq(&conf->device_lock);
 719         if (p->rdev) {
 720                 if (p->rdev->in_sync ||
 721                     atomic_read(&p->rdev->nr_pending)) {
 722                         err = -EBUSY;
 723                         goto abort;
 724                 }
 725                 p->rdev = NULL;
 726                 err = 0;
 727         }
 728         if (err)
 729                 MD_BUG();
 730 abort:
 731         spin_unlock_irq(&conf->device_lock);
 732
 733         print_conf(conf);
 734         return err;
 735 }
 736
 737
 738 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
 739 {
 740         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 741         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 742         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 743
 744         if (bio->bi_size)
 745                 return 1;
 746
 747         if (r1_bio->read_bio != bio)
 748                 BUG();
 749         update_head_pos(r1_bio->read_disk, r1_bio);
 750         /*
 751          * we have read a block, now it needs to be re-written,
 752          * or re-read if the read failed.
 753          * We don't do much here, just schedule handling by raid1d
 754          */
 755         if (!uptodate)
 756                 md_error(r1_bio->mddev,
 757                          conf->mirrors[r1_bio->read_disk].rdev);
 758         else
 759                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 760         atomic_dec(&conf->mirrors[r1_bio->read_disk].rdev->nr_pending);
 761         reschedule_retry(r1_bio);
 762         return 0;
 763 }
 764
 765 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
 766 {
 767         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 768         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 769         mddev_t *mddev = r1_bio->mddev;
 770         conf_t *conf = mddev_to_conf(mddev);
 771         int i;
 772         int mirror=0;
 773
 774         if (bio->bi_size)
 775                 return 1;
 776
 777         for (i = 0; i < conf->raid_disks; i++)
 778                 if (r1_bio->write_bios[i] == bio) {
 779                         mirror = i;
 780                         break;
 781                 }
 782         if (!uptodate)
 783                 md_error(mddev, conf->mirrors[mirror].rdev);
 784         update_head_pos(mirror, r1_bio);
 785
 786         if (atomic_dec_and_test(&r1_bio->remaining)) {
 787                 md_done_sync(mddev, r1_bio->master_bio->bi_size >> 9, uptodate);
 788                 put_buf(r1_bio);
 789         }
 790         atomic_dec(&conf->mirrors[mirror].rdev->nr_pending);
 791         return 0;
 792 }
 793
 794 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
 795 {
 796         conf_t *conf = mddev_to_conf(mddev);
 797         int i;
 798         int disks = conf->raid_disks;
 799         struct bio *bio, *mbio;
 800
 801         bio = r1_bio->master_bio;
 802
 803         /*
 804          * have to allocate lots of bio structures and
 805          * schedule writes
 806          */
 807         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 808                 /*
 809                  * There is no point trying a read-for-reconstruct as
 810                  * reconstruct is about to be aborted
 811                  */
 812                 char b[BDEVNAME_SIZE];
 813                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
 814                         " for block %llu\n",
 815                         bdevname(bio->bi_bdev,b),
 816                         (unsigned long long)r1_bio->sector);
 817                 md_done_sync(mddev, r1_bio->master_bio->bi_size >> 9, 0);
 818                 put_buf(r1_bio);
 819                 return;
 820         }
 821
 822         spin_lock_irq(&conf->device_lock);
 823         for (i = 0; i < disks ; i++) {
 824                 r1_bio->write_bios[i] = NULL;
 825                 if (!conf->mirrors[i].rdev ||
 826                     conf->mirrors[i].rdev->faulty)
 827                         continue;
 828                 if (conf->mirrors[i].rdev->bdev == bio->bi_bdev)
 829                         /*
 830                          * we read from here, no need to write
 831                          */
 832                         continue;
 833                 if (conf->mirrors[i].rdev->in_sync &&
 834                         r1_bio->sector + (bio->bi_size>>9) <= mddev->recovery_cp)
 835                         /*
 836                          * don't need to write this we are just rebuilding
 837                          */
 838                         continue;
 839                 atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 840                 r1_bio->write_bios[i] = bio;
 841         }
 842         spin_unlock_irq(&conf->device_lock);
 843
 844         atomic_set(&r1_bio->remaining, 1);
 845         for (i = disks; i-- ; ) {
 846                 if (!r1_bio->write_bios[i])
 847                         continue;
 848                 mbio = bio_clone(bio, GFP_NOIO);
 849                 r1_bio->write_bios[i] = mbio;
 850                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 851                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 852                 mbio->bi_end_io = end_sync_write;
 853                 mbio->bi_rw = WRITE;
 854                 mbio->bi_private = r1_bio;
 855
 856                 atomic_inc(&r1_bio->remaining);
 857                 md_sync_acct(conf->mirrors[i].rdev, mbio->bi_size >> 9);
 858                 generic_make_request(mbio);
 859         }
 860
 861         if (atomic_dec_and_test(&r1_bio->remaining)) {
 862                 md_done_sync(mddev, r1_bio->master_bio->bi_size >> 9, 1);
 863                 put_buf(r1_bio);
 864         }
 865 }
 866
 867 /*
 868  * This is a kernel thread which:
 869  *
 870  *      1.      Retries failed read operations on working mirrors.
 871  *      2.      Updates the raid superblock when problems encounter.
 872  *      3.      Performs writes following reads for array syncronising.
 873  */
 874
 875 static void raid1d(mddev_t *mddev)
 876 {
 877         struct list_head *head = &retry_list_head;
 878         r1bio_t *r1_bio;
 879         struct bio *bio;
 880         unsigned long flags;
 881         conf_t *conf = mddev_to_conf(mddev);
 882         mdk_rdev_t *rdev;
 883
 884         md_check_recovery(mddev);
 885         md_handle_safemode(mddev);
 886
 887         for (;;) {
 888                 char b[BDEVNAME_SIZE];
 889                 spin_lock_irqsave(&retry_list_lock, flags);
 890                 if (list_empty(head))
 891                         break;
 892                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
 893                 list_del(head->prev);
 894                 spin_unlock_irqrestore(&retry_list_lock, flags);
 895
 896                 mddev = r1_bio->mddev;
 897                 conf = mddev_to_conf(mddev);
 898                 bio = r1_bio->master_bio;
 899                 switch(r1_bio->cmd) {
 900                 case SPECIAL:
 901                         sync_request_write(mddev, r1_bio);
 902                         break;
 903                 case READ:
 904                 case READA:
 905                         if (map(mddev, &rdev) == -1) {
 906                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
 907                                 " read error for block %llu\n",
 908                                 bdevname(bio->bi_bdev,b),
 909                                 (unsigned long long)r1_bio->sector);
 910                                 raid_end_bio_io(r1_bio);
 911                                 break;
 912                         }
 913                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
 914                                 " another mirror\n",
 915                                 bdevname(rdev->bdev,b),
 916                                 (unsigned long long)r1_bio->sector);
 917                         bio->bi_bdev = rdev->bdev;
 918                         bio->bi_sector = r1_bio->sector + rdev->data_offset;
 919                         bio->bi_rw = r1_bio->cmd;
 920
 921                         generic_make_request(bio);
 922                         break;
 923                 }
 924         }
 925         spin_unlock_irqrestore(&retry_list_lock, flags);
 926 }
 927
 928
 929 static int init_resync(conf_t *conf)
 930 {
 931         int buffs;
 932
 933         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
 934         if (conf->r1buf_pool)
 935                 BUG();
 936         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, conf);
 937         if (!conf->r1buf_pool)
 938                 return -ENOMEM;
 939         conf->next_resync = 0;
 940         return 0;
 941 }
 942
 943 /*
 944  * perform a "sync" on one "block"
 945  *
 946  * We need to make sure that no normal I/O request - particularly write
 947  * requests - conflict with active sync requests.
 948  *
 949  * This is achieved by tracking pending requests and a 'barrier' concept
 950  * that can be installed to exclude normal IO requests.
 951  */
 952
 953 static int sync_request(mddev_t *mddev, sector_t sector_nr, int go_faster)
 954 {
 955         conf_t *conf = mddev_to_conf(mddev);
 956         mirror_info_t *mirror;
 957         r1bio_t *r1_bio;
 958         struct bio *read_bio, *bio;
 959         sector_t max_sector, nr_sectors;
 960         int disk, partial;
 961
 962         if (!conf->r1buf_pool)
 963                 if (init_resync(conf))
 964                         return -ENOMEM;
 965
 966         max_sector = mddev->size << 1;
 967         if (sector_nr >= max_sector) {
 968                 close_sync(conf);
 969                 return 0;
 970         }
 971
 972         /*
 973          * If there is non-resync activity waiting for us then
 974          * put in a delay to throttle resync.
 975          */
 976         if (!go_faster && waitqueue_active(&conf->wait_resume))
 977                 schedule_timeout(HZ);
 978         device_barrier(conf, sector_nr + RESYNC_SECTORS);
 979
 980         /*
 981          * If reconstructing, and >1 working disc,
 982          * could dedicate one to rebuild and others to
 983          * service read requests ..
 984          */
 985         disk = conf->last_used;
 986         /* make sure disk is operational */
 987         spin_lock_irq(&conf->device_lock);
 988         while (conf->mirrors[disk].rdev == NULL ||
 989                !conf->mirrors[disk].rdev->in_sync) {
 990                 if (disk <= 0)
 991                         disk = conf->raid_disks;
 992                 disk--;
 993                 if (disk == conf->last_used)
 994                         break;
 995         }
 996         conf->last_used = disk;
 997         atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 998         spin_unlock_irq(&conf->device_lock);
 999
1000         mirror = conf->mirrors + disk;
1001
1002         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1003
1004         spin_lock_irq(&conf->resync_lock);
1005         conf->nr_pending++;
1006         spin_unlock_irq(&conf->resync_lock);
1007
1008         r1_bio->mddev = mddev;
1009         r1_bio->sector = sector_nr;
1010         r1_bio->cmd = SPECIAL;
1011         r1_bio->read_disk = disk;
1012
1013         bio = r1_bio->master_bio;
1014         nr_sectors = RESYNC_BLOCK_SIZE >> 9;
1015         if (max_sector - sector_nr < nr_sectors)
1016                 nr_sectors = max_sector - sector_nr;
1017         bio->bi_size = nr_sectors << 9;
1018         bio->bi_vcnt = (bio->bi_size + PAGE_SIZE-1) / PAGE_SIZE;
1019         /*
1020          * Is there a partial page at the end of the request?
1021          */
1022         partial = bio->bi_size % PAGE_SIZE;
1023         if (partial)
1024                 bio->bi_io_vec[bio->bi_vcnt-1].bv_len = partial;
1025
1026
1027         read_bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1028
1029         read_bio->bi_sector = sector_nr + mirror->rdev->data_offset;
1030         read_bio->bi_bdev = mirror->rdev->bdev;
1031         read_bio->bi_end_io = end_sync_read;
1032         read_bio->bi_rw = READ;
1033         read_bio->bi_private = r1_bio;
1034
1035         if (r1_bio->read_bio)
1036                 BUG();
1037         r1_bio->read_bio = read_bio;
1038
1039         md_sync_acct(mirror->rdev, nr_sectors);
1040
1041         generic_make_request(read_bio);
1042
1043         return nr_sectors;
1044 }
1045
1046 static int run(mddev_t *mddev)
1047 {
1048         conf_t *conf;
1049         int i, j, disk_idx;
1050         mirror_info_t *disk;
1051         mdk_rdev_t *rdev;
1052         struct list_head *tmp;
1053
1054         if (mddev->level != 1) {
1055                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1056                        mdname(mddev), mddev->level);
1057                 goto out;
1058         }
1059         /*
1060          * copy the already verified devices into our private RAID1
1061          * bookkeeping area. [whatever we allocate in run(),
1062          * should be freed in stop()]
1063          */
1064         conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1065         mddev->private = conf;
1066         if (!conf) {
1067                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1068                         mdname(mddev));
1069                 goto out;
1070         }
1071         memset(conf, 0, sizeof(*conf));
1072         conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1073                                  GFP_KERNEL);
1074         if (!conf->mirrors) {
1075                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1076                        mdname(mddev));
1077                 goto out_free_conf;
1078         }
1079         memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1080
1081         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1082                                                 r1bio_pool_free, mddev);
1083         if (!conf->r1bio_pool) {
1084                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1085                         mdname(mddev));
1086                 goto out_free_conf;
1087         }
1088
1089
1090         ITERATE_RDEV(mddev, rdev, tmp) {
1091                 disk_idx = rdev->raid_disk;
1092                 if (disk_idx >= mddev->raid_disks
1093                     || disk_idx < 0)
1094                         continue;
1095                 disk = conf->mirrors + disk_idx;
1096
1097                 disk->rdev = rdev;
1098
1099                 blk_queue_stack_limits(mddev->queue,
1100                                        rdev->bdev->bd_disk->queue);
1101                 /* as we don't honour merge_bvec_fn, we must never risk
1102                  * violating it, so limit ->max_sector to one PAGE, as
1103                  * a one page request is never in violation.
1104                  */
1105                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1106                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1107                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
1108
1109                 disk->head_position = 0;
1110                 if (!rdev->faulty && rdev->in_sync)
1111                         conf->working_disks++;
1112         }
1113         conf->raid_disks = mddev->raid_disks;
1114         conf->mddev = mddev;
1115         conf->device_lock = SPIN_LOCK_UNLOCKED;
1116         if (conf->working_disks == 1)
1117                 mddev->recovery_cp = MaxSector;
1118
1119         conf->resync_lock = SPIN_LOCK_UNLOCKED;
1120         init_waitqueue_head(&conf->wait_idle);
1121         init_waitqueue_head(&conf->wait_resume);
1122
1123         if (!conf->working_disks) {
1124                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1125                         mdname(mddev));
1126                 goto out_free_conf;
1127         }
1128
1129         mddev->degraded = 0;
1130         for (i = 0; i < conf->raid_disks; i++) {
1131
1132                 disk = conf->mirrors + i;
1133
1134                 if (!disk->rdev) {
1135                         disk->head_position = 0;
1136                         mddev->degraded++;
1137                 }
1138         }
1139
1140         /*
1141          * find the first working one and use it as a starting point
1142          * to read balancing.
1143          */
1144         for (j = 0; j < conf->raid_disks &&
1145                      (!conf->mirrors[j].rdev ||
1146                       !conf->mirrors[j].rdev->in_sync) ; j++)
1147                 /* nothing */;
1148         conf->last_used = j;
1149
1150
1151
1152         {
1153                 mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1154                 if (!mddev->thread) {
1155                         printk(KERN_ERR
1156                                 "raid1: couldn't allocate thread for %s\n",
1157                                 mdname(mddev));
1158                         goto out_free_conf;
1159                 }
1160         }
1161         printk(KERN_INFO
1162                 "raid1: raid set %s active with %d out of %d mirrors\n",
1163                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1164                 mddev->raid_disks);
1165         /*
1166          * Ok, everything is just fine now
1167          */
1168         mddev->array_size = mddev->size;
1169
1170         return 0;
1171
1172 out_free_conf:
1173         if (conf->r1bio_pool)
1174                 mempool_destroy(conf->r1bio_pool);
1175         if (conf->mirrors)
1176                 kfree(conf->mirrors);
1177         kfree(conf);
1178         mddev->private = NULL;
1179 out:
1180         return -EIO;
1181 }
1182
1183 static int stop(mddev_t *mddev)
1184 {
1185         conf_t *conf = mddev_to_conf(mddev);
1186
1187         md_unregister_thread(mddev->thread);
1188         mddev->thread = NULL;
1189         if (conf->r1bio_pool)
1190                 mempool_destroy(conf->r1bio_pool);
1191         if (conf->mirrors)
1192                 kfree(conf->mirrors);
1193         kfree(conf);
1194         mddev->private = NULL;
1195         return 0;
1196 }
1197
1198 static mdk_personality_t raid1_personality =
1199 {
1200         .name           = "raid1",
1201         .owner          = THIS_MODULE,
1202         .make_request   = make_request,
1203         .run            = run,
1204         .stop           = stop,
1205         .status         = status,
1206         .error_handler  = error,
1207         .hot_add_disk   = raid1_add_disk,
1208         .hot_remove_disk= raid1_remove_disk,
1209         .spare_active   = raid1_spare_active,
1210         .sync_request   = sync_request,
1211 };
1212
1213 static int __init raid_init(void)
1214 {
1215         return register_md_personality(RAID1, &raid1_personality);
1216 }
1217
1218 static void raid_exit(void)
1219 {
1220         unregister_md_personality(RAID1);
1221 }
1222
1223 module_init(raid_init);
1224 module_exit(raid_exit);
1225 MODULE_LICENSE("GPL");
1226 MODULE_ALIAS("md-personality-3"); /* RAID1 */