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 unplug_slaves(mddev_t *mddev);
  41
  42
  43 static void * r1bio_pool_alloc(int gfp_flags, void *data)
  44 {
  45         mddev_t *mddev = data;
  46         r1bio_t *r1_bio;
  47
  48         /* allocate a r1bio with room for raid_disks entries in the bios array */
  49         r1_bio = kmalloc(sizeof(r1bio_t) + sizeof(struct bio*)*mddev->raid_disks,
  50                          gfp_flags);
  51         if (r1_bio)
  52                 memset(r1_bio, 0, sizeof(*r1_bio) + sizeof(struct bio*)*mddev->raid_disks);
  53         else
  54                 unplug_slaves(mddev);
  55
  56         return r1_bio;
  57 }
  58
  59 static void r1bio_pool_free(void *r1_bio, void *data)
  60 {
  61         kfree(r1_bio);
  62 }
  63
  64 #define RESYNC_BLOCK_SIZE (64*1024)
  65 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  66 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  67 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  68 #define RESYNC_WINDOW (2048*1024)
  69
  70 static void * r1buf_pool_alloc(int gfp_flags, void *data)
  71 {
  72         conf_t *conf = data;
  73         struct page *page;
  74         r1bio_t *r1_bio;
  75         struct bio *bio;
  76         int i, j;
  77
  78         r1_bio = r1bio_pool_alloc(gfp_flags, conf->mddev);
  79         if (!r1_bio) {
  80                 unplug_slaves(conf->mddev);
  81                 return NULL;
  82         }
  83
  84         /*
  85          * Allocate bios : 1 for reading, n-1 for writing
  86          */
  87         for (j = conf->raid_disks ; j-- ; ) {
  88                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
  89                 if (!bio)
  90                         goto out_free_bio;
  91                 r1_bio->bios[j] = bio;
  92         }
  93         /*
  94          * Allocate RESYNC_PAGES data pages and attach them to
  95          * the first bio;
  96          */
  97         bio = r1_bio->bios[0];
  98         for (i = 0; i < RESYNC_PAGES; i++) {
  99                 page = alloc_page(gfp_flags);
 100                 if (unlikely(!page))
 101                         goto out_free_pages;
 102
 103                 bio->bi_io_vec[i].bv_page = page;
 104         }
 105
 106         r1_bio->master_bio = bio;
 107
 108         return r1_bio;
 109
 110 out_free_pages:
 111         for ( ; i > 0 ; i--)
 112                 __free_page(bio->bi_io_vec[i-1].bv_page);
 113 out_free_bio:
 114         while ( ++j < conf->raid_disks )
 115                 bio_put(r1_bio->bios[j]);
 116         r1bio_pool_free(r1_bio, conf->mddev);
 117         return NULL;
 118 }
 119
 120 static void r1buf_pool_free(void *__r1_bio, void *data)
 121 {
 122         int i;
 123         conf_t *conf = data;
 124         r1bio_t *r1bio = __r1_bio;
 125         struct bio *bio = r1bio->bios[0];
 126
 127         for (i = 0; i < RESYNC_PAGES; i++) {
 128                 __free_page(bio->bi_io_vec[i].bv_page);
 129                 bio->bi_io_vec[i].bv_page = NULL;
 130         }
 131         for (i=0 ; i < conf->raid_disks; i++)
 132                 bio_put(r1bio->bios[i]);
 133
 134         r1bio_pool_free(r1bio, conf->mddev);
 135 }
 136
 137 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 138 {
 139         int i;
 140
 141         for (i = 0; i < conf->raid_disks; i++) {
 142                 struct bio **bio = r1_bio->bios + i;
 143                 if (*bio)
 144                         bio_put(*bio);
 145                 *bio = NULL;
 146         }
 147 }
 148
 149 static inline void free_r1bio(r1bio_t *r1_bio)
 150 {
 151         unsigned long flags;
 152
 153         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 154
 155         /*
 156          * Wake up any possible resync thread that waits for the device
 157          * to go idle.
 158          */
 159         spin_lock_irqsave(&conf->resync_lock, flags);
 160         if (!--conf->nr_pending) {
 161                 wake_up(&conf->wait_idle);
 162                 wake_up(&conf->wait_resume);
 163         }
 164         spin_unlock_irqrestore(&conf->resync_lock, flags);
 165
 166         put_all_bios(conf, r1_bio);
 167         mempool_free(r1_bio, conf->r1bio_pool);
 168 }
 169
 170 static inline void put_buf(r1bio_t *r1_bio)
 171 {
 172         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 173         unsigned long flags;
 174
 175         mempool_free(r1_bio, conf->r1buf_pool);
 176
 177         spin_lock_irqsave(&conf->resync_lock, flags);
 178         if (!conf->barrier)
 179                 BUG();
 180         --conf->barrier;
 181         wake_up(&conf->wait_resume);
 182         wake_up(&conf->wait_idle);
 183
 184         if (!--conf->nr_pending) {
 185                 wake_up(&conf->wait_idle);
 186                 wake_up(&conf->wait_resume);
 187         }
 188         spin_unlock_irqrestore(&conf->resync_lock, flags);
 189 }
 190
 191 static int map(mddev_t *mddev, mdk_rdev_t **rdevp)
 192 {
 193         conf_t *conf = mddev_to_conf(mddev);
 194         int i, disks = conf->raid_disks;
 195
 196         /*
 197          * Later we do read balancing on the read side
 198          * now we use the first available disk.
 199          */
 200
 201         spin_lock_irq(&conf->device_lock);
 202         for (i = 0; i < disks; i++) {
 203                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 204                 if (rdev && rdev->in_sync) {
 205                         *rdevp = rdev;
 206                         atomic_inc(&rdev->nr_pending);
 207                         spin_unlock_irq(&conf->device_lock);
 208                         return 0;
 209                 }
 210         }
 211         spin_unlock_irq(&conf->device_lock);
 212
 213         printk(KERN_ERR "raid1_map(): huh, no more operational devices?\n");
 214         return -1;
 215 }
 216
 217 static void reschedule_retry(r1bio_t *r1_bio)
 218 {
 219         unsigned long flags;
 220         mddev_t *mddev = r1_bio->mddev;
 221
 222         spin_lock_irqsave(&retry_list_lock, flags);
 223         list_add(&r1_bio->retry_list, &retry_list_head);
 224         spin_unlock_irqrestore(&retry_list_lock, flags);
 225
 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         bio_endio(bio, bio->bi_size,
 239                 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 240         free_r1bio(r1_bio);
 241 }
 242
 243 /*
 244  * Update disk head position estimator based on IRQ completion info.
 245  */
 246 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 247 {
 248         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 249
 250         conf->mirrors[disk].head_position =
 251                 r1_bio->sector + (r1_bio->sectors);
 252 }
 253
 254 static int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
 255 {
 256         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 257         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 258         int mirror;
 259         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 260
 261         if (bio->bi_size)
 262                 return 1;
 263
 264         mirror = r1_bio->read_disk;
 265         /*
 266          * this branch is our 'one mirror IO has finished' event handler:
 267          */
 268         if (!uptodate)
 269                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 270         else
 271                 /*
 272                  * Set R1BIO_Uptodate in our master bio, so that
 273                  * we will return a good error code for to the higher
 274                  * levels even if IO on some other mirrored buffer fails.
 275                  *
 276                  * The 'master' represents the composite IO operation to
 277                  * user-side. So if something waits for IO, then it will
 278                  * wait for the 'master' bio.
 279                  */
 280                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 281
 282         update_head_pos(mirror, r1_bio);
 283
 284         /*
 285          * we have only one bio on the read side
 286          */
 287         if (uptodate)
 288                 raid_end_bio_io(r1_bio);
 289         else {
 290                 /*
 291                  * oops, read error:
 292                  */
 293                 char b[BDEVNAME_SIZE];
 294                 printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 295                        bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 296                 reschedule_retry(r1_bio);
 297         }
 298
 299         atomic_dec(&conf->mirrors[mirror].rdev->nr_pending);
 300         return 0;
 301 }
 302
 303 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
 304 {
 305         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 306         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 307         int mirror;
 308         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 309
 310         if (bio->bi_size)
 311                 return 1;
 312
 313         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 314                 if (r1_bio->bios[mirror] == bio)
 315                         break;
 316
 317         /*
 318          * this branch is our 'one mirror IO has finished' event handler:
 319          */
 320         if (!uptodate)
 321                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 322         else
 323                 /*
 324                  * Set R1BIO_Uptodate in our master bio, so that
 325                  * we will return a good error code for to the higher
 326                  * levels even if IO on some other mirrored buffer fails.
 327                  *
 328                  * The 'master' represents the composite IO operation to
 329                  * user-side. So if something waits for IO, then it will
 330                  * wait for the 'master' bio.
 331                  */
 332                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 333
 334         update_head_pos(mirror, r1_bio);
 335
 336         /*
 337          *
 338          * Let's see if all mirrored write operations have finished
 339          * already.
 340          */
 341         if (atomic_dec_and_test(&r1_bio->remaining)) {
 342                 md_write_end(r1_bio->mddev);
 343                 raid_end_bio_io(r1_bio);
 344         }
 345
 346         atomic_dec(&conf->mirrors[mirror].rdev->nr_pending);
 347         return 0;
 348 }
 349
 350
 351 /*
 352  * This routine returns the disk from which the requested read should
 353  * be done. There is a per-array 'next expected sequential IO' sector
 354  * number - if this matches on the next IO then we use the last disk.
 355  * There is also a per-disk 'last know head position' sector that is
 356  * maintained from IRQ contexts, both the normal and the resync IO
 357  * completion handlers update this position correctly. If there is no
 358  * perfect sequential match then we pick the disk whose head is closest.
 359  *
 360  * If there are 2 mirrors in the same 2 devices, performance degrades
 361  * because position is mirror, not device based.
 362  *
 363  * The rdev for the device selected will have nr_pending incremented.
 364  */
 365 static int read_balance(conf_t *conf, struct bio *bio, r1bio_t *r1_bio)
 366 {
 367         const unsigned long this_sector = r1_bio->sector;
 368         int new_disk = conf->last_used, disk = new_disk;
 369         const int sectors = bio->bi_size >> 9;
 370         sector_t new_distance, current_distance;
 371
 372         spin_lock_irq(&conf->device_lock);
 373         /*
 374          * Check if it if we can balance. We can balance on the whole
 375          * device if no resync is going on, or below the resync window.
 376          * We take the first readable disk when above the resync window.
 377          */
 378         if (!conf->mddev->in_sync && (this_sector + sectors >= conf->next_resync)) {
 379                 /* make sure that disk is operational */
 380                 new_disk = 0;
 381
 382                 while (!conf->mirrors[new_disk].rdev ||
 383                        !conf->mirrors[new_disk].rdev->in_sync) {
 384                         new_disk++;
 385                         if (new_disk == conf->raid_disks) {
 386                                 new_disk = 0;
 387                                 break;
 388                         }
 389                 }
 390                 goto rb_out;
 391         }
 392
 393
 394         /* make sure the disk is operational */
 395         while (!conf->mirrors[new_disk].rdev ||
 396                !conf->mirrors[new_disk].rdev->in_sync) {
 397                 if (new_disk <= 0)
 398                         new_disk = conf->raid_disks;
 399                 new_disk--;
 400                 if (new_disk == disk) {
 401                         new_disk = conf->last_used;
 402                         goto rb_out;
 403                 }
 404         }
 405         disk = new_disk;
 406         /* now disk == new_disk == starting point for search */
 407
 408         /*
 409          * Don't change to another disk for sequential reads:
 410          */
 411         if (conf->next_seq_sect == this_sector)
 412                 goto rb_out;
 413         if (this_sector == conf->mirrors[new_disk].head_position)
 414                 goto rb_out;
 415
 416         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 417
 418         /* Find the disk whose head is closest */
 419
 420         do {
 421                 if (disk <= 0)
 422                         disk = conf->raid_disks;
 423                 disk--;
 424
 425                 if (!conf->mirrors[disk].rdev ||
 426                     !conf->mirrors[disk].rdev->in_sync)
 427                         continue;
 428
 429                 if (!atomic_read(&conf->mirrors[disk].rdev->nr_pending)) {
 430                         new_disk = disk;
 431                         break;
 432                 }
 433                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 434                 if (new_distance < current_distance) {
 435                         current_distance = new_distance;
 436                         new_disk = disk;
 437                 }
 438         } while (disk != conf->last_used);
 439
 440 rb_out:
 441         r1_bio->read_disk = new_disk;
 442         conf->next_seq_sect = this_sector + sectors;
 443
 444         conf->last_used = new_disk;
 445
 446         if (conf->mirrors[new_disk].rdev)
 447                 atomic_inc(&conf->mirrors[new_disk].rdev->nr_pending);
 448         spin_unlock_irq(&conf->device_lock);
 449
 450         return new_disk;
 451 }
 452
 453 static void unplug_slaves(mddev_t *mddev)
 454 {
 455         conf_t *conf = mddev_to_conf(mddev);
 456         int i;
 457         unsigned long flags;
 458
 459         spin_lock_irqsave(&conf->device_lock, flags);
 460         for (i=0; i<mddev->raid_disks; i++) {
 461                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
 462                 if (rdev && !rdev->faulty) {
 463                         request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
 464
 465                         if (r_queue->unplug_fn)
 466                                 r_queue->unplug_fn(r_queue);
 467                 }
 468         }
 469         spin_unlock_irqrestore(&conf->device_lock, flags);
 470 }
 471 static void raid1_unplug(request_queue_t *q)
 472 {
 473         unplug_slaves(q->queuedata);
 474 }
 475
 476 /*
 477  * Throttle resync depth, so that we can both get proper overlapping of
 478  * requests, but are still able to handle normal requests quickly.
 479  */
 480 #define RESYNC_DEPTH 32
 481
 482 static void device_barrier(conf_t *conf, sector_t sect)
 483 {
 484         spin_lock_irq(&conf->resync_lock);
 485         wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume),
 486                             conf->resync_lock, unplug_slaves(conf->mddev));
 487
 488         if (!conf->barrier++) {
 489                 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
 490                                     conf->resync_lock, unplug_slaves(conf->mddev));
 491                 if (conf->nr_pending)
 492                         BUG();
 493         }
 494         wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH,
 495                             conf->resync_lock, unplug_slaves(conf->mddev));
 496         conf->next_resync = sect;
 497         spin_unlock_irq(&conf->resync_lock);
 498 }
 499
 500 static int make_request(request_queue_t *q, struct bio * bio)
 501 {
 502         mddev_t *mddev = q->queuedata;
 503         conf_t *conf = mddev_to_conf(mddev);
 504         mirror_info_t *mirror;
 505         r1bio_t *r1_bio;
 506         struct bio *read_bio;
 507         int i, disks = conf->raid_disks;
 508
 509         /*
 510          * Register the new request and wait if the reconstruction
 511          * thread has put up a bar for new requests.
 512          * Continue immediately if no resync is active currently.
 513          */
 514         spin_lock_irq(&conf->resync_lock);
 515         wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock, );
 516         conf->nr_pending++;
 517         spin_unlock_irq(&conf->resync_lock);
 518
 519         if (bio_data_dir(bio)==WRITE) {
 520                 disk_stat_inc(mddev->gendisk, writes);
 521                 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
 522         } else {
 523                 disk_stat_inc(mddev->gendisk, reads);
 524                 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
 525         }
 526
 527         /*
 528          * make_request() can abort the operation when READA is being
 529          * used and no empty request is available.
 530          *
 531          */
 532         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 533
 534         r1_bio->master_bio = bio;
 535         r1_bio->sectors = bio->bi_size >> 9;
 536
 537         r1_bio->mddev = mddev;
 538         r1_bio->sector = bio->bi_sector;
 539
 540         if (bio_data_dir(bio) == READ) {
 541                 /*
 542                  * read balancing logic:
 543                  */
 544                 mirror = conf->mirrors + read_balance(conf, bio, r1_bio);
 545
 546                 read_bio = bio_clone(bio, GFP_NOIO);
 547
 548                 r1_bio->bios[r1_bio->read_disk] = read_bio;
 549
 550                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 551                 read_bio->bi_bdev = mirror->rdev->bdev;
 552                 read_bio->bi_end_io = raid1_end_read_request;
 553                 read_bio->bi_rw = READ;
 554                 read_bio->bi_private = r1_bio;
 555
 556                 generic_make_request(read_bio);
 557                 return 0;
 558         }
 559
 560         /*
 561          * WRITE:
 562          */
 563         /* first select target devices under spinlock and
 564          * inc refcount on their rdev.  Record them by setting
 565          * bios[x] to bio
 566          */
 567         spin_lock_irq(&conf->device_lock);
 568         for (i = 0;  i < disks; i++) {
 569                 if (conf->mirrors[i].rdev &&
 570                     !conf->mirrors[i].rdev->faulty) {
 571                         atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 572                         r1_bio->bios[i] = bio;
 573                 } else
 574                         r1_bio->bios[i] = NULL;
 575         }
 576         spin_unlock_irq(&conf->device_lock);
 577
 578         atomic_set(&r1_bio->remaining, 1);
 579         md_write_start(mddev);
 580         for (i = 0; i < disks; i++) {
 581                 struct bio *mbio;
 582                 if (!r1_bio->bios[i])
 583                         continue;
 584
 585                 mbio = bio_clone(bio, GFP_NOIO);
 586                 r1_bio->bios[i] = mbio;
 587
 588                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 589                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 590                 mbio->bi_end_io = raid1_end_write_request;
 591                 mbio->bi_rw = WRITE;
 592                 mbio->bi_private = r1_bio;
 593
 594                 atomic_inc(&r1_bio->remaining);
 595                 generic_make_request(mbio);
 596         }
 597
 598         if (atomic_dec_and_test(&r1_bio->remaining)) {
 599                 md_write_end(mddev);
 600                 raid_end_bio_io(r1_bio);
 601         }
 602
 603         return 0;
 604 }
 605
 606 static void status(struct seq_file *seq, mddev_t *mddev)
 607 {
 608         conf_t *conf = mddev_to_conf(mddev);
 609         int i;
 610
 611         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 612                                                 conf->working_disks);
 613         for (i = 0; i < conf->raid_disks; i++)
 614                 seq_printf(seq, "%s",
 615                               conf->mirrors[i].rdev &&
 616                               conf->mirrors[i].rdev->in_sync ? "U" : "_");
 617         seq_printf(seq, "]");
 618 }
 619
 620
 621 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
 622 {
 623         char b[BDEVNAME_SIZE];
 624         conf_t *conf = mddev_to_conf(mddev);
 625
 626         /*
 627          * If it is not operational, then we have already marked it as dead
 628          * else if it is the last working disks, ignore the error, let the
 629          * next level up know.
 630          * else mark the drive as failed
 631          */
 632         if (rdev->in_sync
 633             && conf->working_disks == 1)
 634                 /*
 635                  * Don't fail the drive, act as though we were just a
 636                  * normal single drive
 637                  */
 638                 return;
 639         if (rdev->in_sync) {
 640                 mddev->degraded++;
 641                 conf->working_disks--;
 642                 /*
 643                  * if recovery is running, make sure it aborts.
 644                  */
 645                 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
 646         }
 647         rdev->in_sync = 0;
 648         rdev->faulty = 1;
 649         mddev->sb_dirty = 1;
 650         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
 651                 "       Operation continuing on %d devices\n",
 652                 bdevname(rdev->bdev,b), conf->working_disks);
 653 }
 654
 655 static void print_conf(conf_t *conf)
 656 {
 657         int i;
 658         mirror_info_t *tmp;
 659
 660         printk("RAID1 conf printout:\n");
 661         if (!conf) {
 662                 printk("(!conf)\n");
 663                 return;
 664         }
 665         printk(" --- wd:%d rd:%d\n", conf->working_disks,
 666                 conf->raid_disks);
 667
 668         for (i = 0; i < conf->raid_disks; i++) {
 669                 char b[BDEVNAME_SIZE];
 670                 tmp = conf->mirrors + i;
 671                 if (tmp->rdev)
 672                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
 673                                 i, !tmp->rdev->in_sync, !tmp->rdev->faulty,
 674                                 bdevname(tmp->rdev->bdev,b));
 675         }
 676 }
 677
 678 static void close_sync(conf_t *conf)
 679 {
 680         spin_lock_irq(&conf->resync_lock);
 681         wait_event_lock_irq(conf->wait_resume, !conf->barrier,
 682                             conf->resync_lock,  unplug_slaves(conf->mddev));
 683         spin_unlock_irq(&conf->resync_lock);
 684
 685         if (conf->barrier) BUG();
 686         if (waitqueue_active(&conf->wait_idle)) BUG();
 687
 688         mempool_destroy(conf->r1buf_pool);
 689         conf->r1buf_pool = NULL;
 690 }
 691
 692 static int raid1_spare_active(mddev_t *mddev)
 693 {
 694         int i;
 695         conf_t *conf = mddev->private;
 696         mirror_info_t *tmp;
 697
 698         spin_lock_irq(&conf->device_lock);
 699         /*
 700          * Find all failed disks within the RAID1 configuration
 701          * and mark them readable
 702          */
 703         for (i = 0; i < conf->raid_disks; i++) {
 704                 tmp = conf->mirrors + i;
 705                 if (tmp->rdev
 706                     && !tmp->rdev->faulty
 707                     && !tmp->rdev->in_sync) {
 708                         conf->working_disks++;
 709                         mddev->degraded--;
 710                         tmp->rdev->in_sync = 1;
 711                 }
 712         }
 713         spin_unlock_irq(&conf->device_lock);
 714
 715         print_conf(conf);
 716         return 0;
 717 }
 718
 719
 720 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
 721 {
 722         conf_t *conf = mddev->private;
 723         int found = 0;
 724         int mirror;
 725         mirror_info_t *p;
 726
 727         spin_lock_irq(&conf->device_lock);
 728         for (mirror=0; mirror < mddev->raid_disks; mirror++)
 729                 if ( !(p=conf->mirrors+mirror)->rdev) {
 730                         p->rdev = rdev;
 731
 732                         blk_queue_stack_limits(mddev->queue,
 733                                                rdev->bdev->bd_disk->queue);
 734                         /* as we don't honour merge_bvec_fn, we must never risk
 735                          * violating it, so limit ->max_sector to one PAGE, as
 736                          * a one page request is never in violation.
 737                          */
 738                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
 739                             mddev->queue->max_sectors > (PAGE_SIZE>>9))
 740                                 mddev->queue->max_sectors = (PAGE_SIZE>>9);
 741
 742                         p->head_position = 0;
 743                         rdev->raid_disk = mirror;
 744                         found = 1;
 745                         break;
 746                 }
 747         spin_unlock_irq(&conf->device_lock);
 748
 749         print_conf(conf);
 750         return found;
 751 }
 752
 753 static int raid1_remove_disk(mddev_t *mddev, int number)
 754 {
 755         conf_t *conf = mddev->private;
 756         int err = 1;
 757         mirror_info_t *p = conf->mirrors+ number;
 758
 759         print_conf(conf);
 760         spin_lock_irq(&conf->device_lock);
 761         if (p->rdev) {
 762                 if (p->rdev->in_sync ||
 763                     atomic_read(&p->rdev->nr_pending)) {
 764                         err = -EBUSY;
 765                         goto abort;
 766                 }
 767                 p->rdev = NULL;
 768                 err = 0;
 769         }
 770         if (err)
 771                 MD_BUG();
 772 abort:
 773         spin_unlock_irq(&conf->device_lock);
 774
 775         print_conf(conf);
 776         return err;
 777 }
 778
 779
 780 static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
 781 {
 782         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 783         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 784         conf_t *conf = mddev_to_conf(r1_bio->mddev);
 785
 786         if (bio->bi_size)
 787                 return 1;
 788
 789         if (r1_bio->bios[r1_bio->read_disk] != bio)
 790                 BUG();
 791         update_head_pos(r1_bio->read_disk, r1_bio);
 792         /*
 793          * we have read a block, now it needs to be re-written,
 794          * or re-read if the read failed.
 795          * We don't do much here, just schedule handling by raid1d
 796          */
 797         if (!uptodate)
 798                 md_error(r1_bio->mddev,
 799                          conf->mirrors[r1_bio->read_disk].rdev);
 800         else
 801                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 802         atomic_dec(&conf->mirrors[r1_bio->read_disk].rdev->nr_pending);
 803         reschedule_retry(r1_bio);
 804         return 0;
 805 }
 806
 807 static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
 808 {
 809         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 810         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 811         mddev_t *mddev = r1_bio->mddev;
 812         conf_t *conf = mddev_to_conf(mddev);
 813         int i;
 814         int mirror=0;
 815
 816         if (bio->bi_size)
 817                 return 1;
 818
 819         for (i = 0; i < conf->raid_disks; i++)
 820                 if (r1_bio->bios[i] == bio) {
 821                         mirror = i;
 822                         break;
 823                 }
 824         if (!uptodate)
 825                 md_error(mddev, conf->mirrors[mirror].rdev);
 826         update_head_pos(mirror, r1_bio);
 827
 828         if (atomic_dec_and_test(&r1_bio->remaining)) {
 829                 md_done_sync(mddev, r1_bio->sectors, uptodate);
 830                 put_buf(r1_bio);
 831         }
 832         atomic_dec(&conf->mirrors[mirror].rdev->nr_pending);
 833         return 0;
 834 }
 835
 836 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
 837 {
 838         conf_t *conf = mddev_to_conf(mddev);
 839         int i;
 840         int disks = conf->raid_disks;
 841         struct bio *bio, *wbio;
 842
 843         bio = r1_bio->bios[r1_bio->read_disk];
 844
 845         /*
 846          * schedule writes
 847          */
 848         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 849                 /*
 850                  * There is no point trying a read-for-reconstruct as
 851                  * reconstruct is about to be aborted
 852                  */
 853                 char b[BDEVNAME_SIZE];
 854                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
 855                         " for block %llu\n",
 856                         bdevname(bio->bi_bdev,b),
 857                         (unsigned long long)r1_bio->sector);
 858                 md_done_sync(mddev, r1_bio->sectors, 0);
 859                 put_buf(r1_bio);
 860                 return;
 861         }
 862
 863         atomic_set(&r1_bio->remaining, 1);
 864         for (i = 0; i < disks ; i++) {
 865                 wbio = r1_bio->bios[i];
 866                 if (wbio->bi_end_io != end_sync_write)
 867                         continue;
 868
 869                 atomic_inc(&conf->mirrors[i].rdev->nr_pending);
 870                 atomic_inc(&r1_bio->remaining);
 871                 md_sync_acct(conf->mirrors[i].rdev, wbio->bi_size >> 9);
 872                 generic_make_request(wbio);
 873         }
 874
 875         if (atomic_dec_and_test(&r1_bio->remaining)) {
 876                 md_done_sync(mddev, r1_bio->sectors, 1);
 877                 put_buf(r1_bio);
 878         }
 879 }
 880
 881 /*
 882  * This is a kernel thread which:
 883  *
 884  *      1.      Retries failed read operations on working mirrors.
 885  *      2.      Updates the raid superblock when problems encounter.
 886  *      3.      Performs writes following reads for array syncronising.
 887  */
 888
 889 static void raid1d(mddev_t *mddev)
 890 {
 891         struct list_head *head = &retry_list_head;
 892         r1bio_t *r1_bio;
 893         struct bio *bio;
 894         unsigned long flags;
 895         conf_t *conf = mddev_to_conf(mddev);
 896         int unplug=0;
 897         mdk_rdev_t *rdev;
 898
 899         md_check_recovery(mddev);
 900         md_handle_safemode(mddev);
 901
 902         for (;;) {
 903                 char b[BDEVNAME_SIZE];
 904                 spin_lock_irqsave(&retry_list_lock, flags);
 905                 if (list_empty(head))
 906                         break;
 907                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
 908                 list_del(head->prev);
 909                 spin_unlock_irqrestore(&retry_list_lock, flags);
 910
 911                 mddev = r1_bio->mddev;
 912                 conf = mddev_to_conf(mddev);
 913                 bio = r1_bio->master_bio;
 914                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
 915                         sync_request_write(mddev, r1_bio);
 916                         unplug = 1;
 917                 } else {
 918                         if (map(mddev, &rdev) == -1) {
 919                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
 920                                        " read error for block %llu\n",
 921                                        bdevname(bio->bi_bdev,b),
 922                                        (unsigned long long)r1_bio->sector);
 923                                 raid_end_bio_io(r1_bio);
 924                         } else {
 925                                 printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
 926                                        " another mirror\n",
 927                                        bdevname(rdev->bdev,b),
 928                                        (unsigned long long)r1_bio->sector);
 929                                 bio->bi_bdev = rdev->bdev;
 930                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
 931                                 bio->bi_rw = READ;
 932                                 unplug = 1;
 933                                 generic_make_request(bio);
 934                         }
 935                 }
 936         }
 937         spin_unlock_irqrestore(&retry_list_lock, flags);
 938         if (unplug)
 939                 unplug_slaves(mddev);
 940 }
 941
 942
 943 static int init_resync(conf_t *conf)
 944 {
 945         int buffs;
 946
 947         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
 948         if (conf->r1buf_pool)
 949                 BUG();
 950         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, conf);
 951         if (!conf->r1buf_pool)
 952                 return -ENOMEM;
 953         conf->next_resync = 0;
 954         return 0;
 955 }
 956
 957 /*
 958  * perform a "sync" on one "block"
 959  *
 960  * We need to make sure that no normal I/O request - particularly write
 961  * requests - conflict with active sync requests.
 962  *
 963  * This is achieved by tracking pending requests and a 'barrier' concept
 964  * that can be installed to exclude normal IO requests.
 965  */
 966
 967 static int sync_request(mddev_t *mddev, sector_t sector_nr, int go_faster)
 968 {
 969         conf_t *conf = mddev_to_conf(mddev);
 970         mirror_info_t *mirror;
 971         r1bio_t *r1_bio;
 972         struct bio *bio;
 973         sector_t max_sector, nr_sectors;
 974         int disk;
 975         int i;
 976
 977         if (!conf->r1buf_pool)
 978                 if (init_resync(conf))
 979                         return -ENOMEM;
 980
 981         max_sector = mddev->size << 1;
 982         if (sector_nr >= max_sector) {
 983                 close_sync(conf);
 984                 return 0;
 985         }
 986
 987         /*
 988          * If there is non-resync activity waiting for us then
 989          * put in a delay to throttle resync.
 990          */
 991         if (!go_faster && waitqueue_active(&conf->wait_resume))
 992                 schedule_timeout(HZ);
 993         device_barrier(conf, sector_nr + RESYNC_SECTORS);
 994
 995         /*
 996          * If reconstructing, and >1 working disc,
 997          * could dedicate one to rebuild and others to
 998          * service read requests ..
 999          */
1000         disk = conf->last_used;
1001         /* make sure disk is operational */
1002         spin_lock_irq(&conf->device_lock);
1003         while (conf->mirrors[disk].rdev == NULL ||
1004                !conf->mirrors[disk].rdev->in_sync) {
1005                 if (disk <= 0)
1006                         disk = conf->raid_disks;
1007                 disk--;
1008                 if (disk == conf->last_used)
1009                         break;
1010         }
1011         conf->last_used = disk;
1012         atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
1013         spin_unlock_irq(&conf->device_lock);
1014
1015         mirror = conf->mirrors + disk;
1016
1017         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1018
1019         spin_lock_irq(&conf->resync_lock);
1020         conf->nr_pending++;
1021         spin_unlock_irq(&conf->resync_lock);
1022
1023         r1_bio->mddev = mddev;
1024         r1_bio->sector = sector_nr;
1025         set_bit(R1BIO_IsSync, &r1_bio->state);
1026         r1_bio->read_disk = disk;
1027
1028         for (i=0; i < conf->raid_disks; i++) {
1029                 bio = r1_bio->bios[i];
1030
1031                 /* take from bio_init */
1032                 bio->bi_next = NULL;
1033                 bio->bi_flags |= 1 << BIO_UPTODATE;
1034                 bio->bi_rw = 0;
1035                 bio->bi_vcnt = 0;
1036                 bio->bi_idx = 0;
1037                 bio->bi_phys_segments = 0;
1038                 bio->bi_hw_segments = 0;
1039                 bio->bi_size = 0;
1040                 bio->bi_end_io = NULL;
1041                 bio->bi_private = NULL;
1042
1043                 if (i == disk) {
1044                         bio->bi_rw = READ;
1045                         bio->bi_end_io = end_sync_read;
1046                 } else if (conf->mirrors[i].rdev &&
1047                            !conf->mirrors[i].rdev->faulty &&
1048                            (!conf->mirrors[i].rdev->in_sync ||
1049                             sector_nr + RESYNC_SECTORS > mddev->recovery_cp)) {
1050                         bio->bi_rw = WRITE;
1051                         bio->bi_end_io = end_sync_write;
1052                 } else
1053                         continue;
1054                 bio->bi_sector = sector_nr + conf->mirrors[i].rdev->data_offset;
1055                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1056                 bio->bi_private = r1_bio;
1057         }
1058         nr_sectors = 0;
1059         do {
1060                 struct page *page;
1061                 int len = PAGE_SIZE;
1062                 if (sector_nr + (len>>9) > max_sector)
1063                         len = (max_sector - sector_nr) << 9;
1064                 if (len == 0)
1065                         break;
1066                 for (i=0 ; i < conf->raid_disks; i++) {
1067                         bio = r1_bio->bios[i];
1068                         if (bio->bi_end_io) {
1069                                 page = r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page;
1070                                 if (bio_add_page(bio, page, len, 0) == 0) {
1071                                         /* stop here */
1072                                         r1_bio->bios[0]->bi_io_vec[bio->bi_vcnt].bv_page = page;
1073                                         while (i > 0) {
1074                                                 i--;
1075                                                 bio = r1_bio->bios[i];
1076                                                 if (bio->bi_end_io==NULL) continue;
1077                                                 /* remove last page from this bio */
1078                                                 bio->bi_vcnt--;
1079                                                 bio->bi_size -= len;
1080                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1081                                         }
1082                                         goto bio_full;
1083                                 }
1084                         }
1085                 }
1086                 nr_sectors += len>>9;
1087                 sector_nr += len>>9;
1088         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1089  bio_full:
1090         bio = r1_bio->bios[disk];
1091         r1_bio->sectors = nr_sectors;
1092
1093         md_sync_acct(mirror->rdev, nr_sectors);
1094
1095         generic_make_request(bio);
1096
1097         return nr_sectors;
1098 }
1099
1100 static int run(mddev_t *mddev)
1101 {
1102         conf_t *conf;
1103         int i, j, disk_idx;
1104         mirror_info_t *disk;
1105         mdk_rdev_t *rdev;
1106         struct list_head *tmp;
1107
1108         if (mddev->level != 1) {
1109                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1110                        mdname(mddev), mddev->level);
1111                 goto out;
1112         }
1113         /*
1114          * copy the already verified devices into our private RAID1
1115          * bookkeeping area. [whatever we allocate in run(),
1116          * should be freed in stop()]
1117          */
1118         conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1119         mddev->private = conf;
1120         if (!conf) {
1121                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1122                         mdname(mddev));
1123                 goto out;
1124         }
1125         memset(conf, 0, sizeof(*conf));
1126         conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1127                                  GFP_KERNEL);
1128         if (!conf->mirrors) {
1129                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1130                        mdname(mddev));
1131                 goto out_free_conf;
1132         }
1133         memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1134
1135         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1136                                                 r1bio_pool_free, mddev);
1137         if (!conf->r1bio_pool) {
1138                 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
1139                         mdname(mddev));
1140                 goto out_free_conf;
1141         }
1142         mddev->queue->unplug_fn = raid1_unplug;
1143
1144
1145         ITERATE_RDEV(mddev, rdev, tmp) {
1146                 disk_idx = rdev->raid_disk;
1147                 if (disk_idx >= mddev->raid_disks
1148                     || disk_idx < 0)
1149                         continue;
1150                 disk = conf->mirrors + disk_idx;
1151
1152                 disk->rdev = rdev;
1153
1154                 blk_queue_stack_limits(mddev->queue,
1155                                        rdev->bdev->bd_disk->queue);
1156                 /* as we don't honour merge_bvec_fn, we must never risk
1157                  * violating it, so limit ->max_sector to one PAGE, as
1158                  * a one page request is never in violation.
1159                  */
1160                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1161                     mddev->queue->max_sectors > (PAGE_SIZE>>9))
1162                         mddev->queue->max_sectors = (PAGE_SIZE>>9);
1163
1164                 disk->head_position = 0;
1165                 if (!rdev->faulty && rdev->in_sync)
1166                         conf->working_disks++;
1167         }
1168         conf->raid_disks = mddev->raid_disks;
1169         conf->mddev = mddev;
1170         conf->device_lock = SPIN_LOCK_UNLOCKED;
1171         if (conf->working_disks == 1)
1172                 mddev->recovery_cp = MaxSector;
1173
1174         conf->resync_lock = SPIN_LOCK_UNLOCKED;
1175         init_waitqueue_head(&conf->wait_idle);
1176         init_waitqueue_head(&conf->wait_resume);
1177
1178         if (!conf->working_disks) {
1179                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
1180                         mdname(mddev));
1181                 goto out_free_conf;
1182         }
1183
1184         mddev->degraded = 0;
1185         for (i = 0; i < conf->raid_disks; i++) {
1186
1187                 disk = conf->mirrors + i;
1188
1189                 if (!disk->rdev) {
1190                         disk->head_position = 0;
1191                         mddev->degraded++;
1192                 }
1193         }
1194
1195         /*
1196          * find the first working one and use it as a starting point
1197          * to read balancing.
1198          */
1199         for (j = 0; j < conf->raid_disks &&
1200                      (!conf->mirrors[j].rdev ||
1201                       !conf->mirrors[j].rdev->in_sync) ; j++)
1202                 /* nothing */;
1203         conf->last_used = j;
1204
1205
1206
1207         {
1208                 mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
1209                 if (!mddev->thread) {
1210                         printk(KERN_ERR
1211                                 "raid1: couldn't allocate thread for %s\n",
1212                                 mdname(mddev));
1213                         goto out_free_conf;
1214                 }
1215         }
1216         printk(KERN_INFO
1217                 "raid1: raid set %s active with %d out of %d mirrors\n",
1218                 mdname(mddev), mddev->raid_disks - mddev->degraded,
1219                 mddev->raid_disks);
1220         /*
1221          * Ok, everything is just fine now
1222          */
1223         mddev->array_size = mddev->size;
1224
1225         return 0;
1226
1227 out_free_conf:
1228         if (conf->r1bio_pool)
1229                 mempool_destroy(conf->r1bio_pool);
1230         if (conf->mirrors)
1231                 kfree(conf->mirrors);
1232         kfree(conf);
1233         mddev->private = NULL;
1234 out:
1235         return -EIO;
1236 }
1237
1238 static int stop(mddev_t *mddev)
1239 {
1240         conf_t *conf = mddev_to_conf(mddev);
1241
1242         md_unregister_thread(mddev->thread);
1243         mddev->thread = NULL;
1244         if (conf->r1bio_pool)
1245                 mempool_destroy(conf->r1bio_pool);
1246         if (conf->mirrors)
1247                 kfree(conf->mirrors);
1248         kfree(conf);
1249         mddev->private = NULL;
1250         return 0;
1251 }
1252
1253 static mdk_personality_t raid1_personality =
1254 {
1255         .name           = "raid1",
1256         .owner          = THIS_MODULE,
1257         .make_request   = make_request,
1258         .run            = run,
1259         .stop           = stop,
1260         .status         = status,
1261         .error_handler  = error,
1262         .hot_add_disk   = raid1_add_disk,
1263         .hot_remove_disk= raid1_remove_disk,
1264         .spare_active   = raid1_spare_active,
1265         .sync_request   = sync_request,
1266 };
1267
1268 static int __init raid_init(void)
1269 {
1270         return register_md_personality(RAID1, &raid1_personality);
1271 }
1272
1273 static void raid_exit(void)
1274 {
1275         unregister_md_personality(RAID1);
1276 }
1277
1278 module_init(raid_init);
1279 module_exit(raid_exit);
1280 MODULE_LICENSE("GPL");
1281 MODULE_ALIAS("md-personality-3"); /* RAID1 */