1 /* 2 * Copyright (C) 2003 Sistina Software Limited. 3 * Copyright (C) 2005-2008 Red Hat, Inc. All rights reserved. 4 * 5 * This file is released under the GPL. 6 */ 7 8 #include "dm-bio-record.h" 9 10 #include <linux/init.h> 11 #include <linux/mempool.h> 12 #include <linux/module.h> 13 #include <linux/pagemap.h> 14 #include <linux/slab.h> 15 #include <linux/workqueue.h> 16 #include <linux/device-mapper.h> 17 #include <linux/dm-io.h> 18 #include <linux/dm-dirty-log.h> 19 #include <linux/dm-kcopyd.h> 20 #include <linux/dm-region-hash.h> 21 22 #define DM_MSG_PREFIX "raid1" 23 24 #define MAX_RECOVERY 1 /* Maximum number of regions recovered in parallel. */ 25 26 #define DM_RAID1_HANDLE_ERRORS 0x01 27 #define errors_handled(p) ((p)->features & DM_RAID1_HANDLE_ERRORS) 28 29 static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped); 30 31 /*----------------------------------------------------------------- 32 * Mirror set structures. 33 *---------------------------------------------------------------*/ 34 enum dm_raid1_error { 35 DM_RAID1_WRITE_ERROR, 36 DM_RAID1_FLUSH_ERROR, 37 DM_RAID1_SYNC_ERROR, 38 DM_RAID1_READ_ERROR 39 }; 40 41 struct mirror { 42 struct mirror_set *ms; 43 atomic_t error_count; 44 unsigned long error_type; 45 struct dm_dev *dev; 46 sector_t offset; 47 }; 48 49 struct mirror_set { 50 struct dm_target *ti; 51 struct list_head list; 52 53 uint64_t features; 54 55 spinlock_t lock; /* protects the lists */ 56 struct bio_list reads; 57 struct bio_list writes; 58 struct bio_list failures; 59 struct bio_list holds; /* bios are waiting until suspend */ 60 61 struct dm_region_hash *rh; 62 struct dm_kcopyd_client *kcopyd_client; 63 struct dm_io_client *io_client; 64 65 /* recovery */ 66 region_t nr_regions; 67 int in_sync; 68 int log_failure; 69 int leg_failure; 70 atomic_t suspend; 71 72 atomic_t default_mirror; /* Default mirror */ 73 74 struct workqueue_struct *kmirrord_wq; 75 struct work_struct kmirrord_work; 76 struct timer_list timer; 77 unsigned long timer_pending; 78 79 struct work_struct trigger_event; 80 81 unsigned nr_mirrors; 82 struct mirror mirror[0]; 83 }; 84 85 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(raid1_resync_throttle, 86 "A percentage of time allocated for raid resynchronization"); 87 88 static void wakeup_mirrord(void *context) 89 { 90 struct mirror_set *ms = context; 91 92 queue_work(ms->kmirrord_wq, &ms->kmirrord_work); 93 } 94 95 static void delayed_wake_fn(unsigned long data) 96 { 97 struct mirror_set *ms = (struct mirror_set *) data; 98 99 clear_bit(0, &ms->timer_pending); 100 wakeup_mirrord(ms); 101 } 102 103 static void delayed_wake(struct mirror_set *ms) 104 { 105 if (test_and_set_bit(0, &ms->timer_pending)) 106 return; 107 108 ms->timer.expires = jiffies + HZ / 5; 109 ms->timer.data = (unsigned long) ms; 110 ms->timer.function = delayed_wake_fn; 111 add_timer(&ms->timer); 112 } 113 114 static void wakeup_all_recovery_waiters(void *context) 115 { 116 wake_up_all(&_kmirrord_recovery_stopped); 117 } 118 119 static void queue_bio(struct mirror_set *ms, struct bio *bio, int rw) 120 { 121 unsigned long flags; 122 int should_wake = 0; 123 struct bio_list *bl; 124 125 bl = (rw == WRITE) ? &ms->writes : &ms->reads; 126 spin_lock_irqsave(&ms->lock, flags); 127 should_wake = !(bl->head); 128 bio_list_add(bl, bio); 129 spin_unlock_irqrestore(&ms->lock, flags); 130 131 if (should_wake) 132 wakeup_mirrord(ms); 133 } 134 135 static void dispatch_bios(void *context, struct bio_list *bio_list) 136 { 137 struct mirror_set *ms = context; 138 struct bio *bio; 139 140 while ((bio = bio_list_pop(bio_list))) 141 queue_bio(ms, bio, WRITE); 142 } 143 144 struct dm_raid1_bio_record { 145 struct mirror *m; 146 /* if details->bi_bdev == NULL, details were not saved */ 147 struct dm_bio_details details; 148 region_t write_region; 149 }; 150 151 /* 152 * Every mirror should look like this one. 153 */ 154 #define DEFAULT_MIRROR 0 155 156 /* 157 * This is yucky. We squirrel the mirror struct away inside 158 * bi_next for read/write buffers. This is safe since the bh 159 * doesn't get submitted to the lower levels of block layer. 160 */ 161 static struct mirror *bio_get_m(struct bio *bio) 162 { 163 return (struct mirror *) bio->bi_next; 164 } 165 166 static void bio_set_m(struct bio *bio, struct mirror *m) 167 { 168 bio->bi_next = (struct bio *) m; 169 } 170 171 static struct mirror *get_default_mirror(struct mirror_set *ms) 172 { 173 return &ms->mirror[atomic_read(&ms->default_mirror)]; 174 } 175 176 static void set_default_mirror(struct mirror *m) 177 { 178 struct mirror_set *ms = m->ms; 179 struct mirror *m0 = &(ms->mirror[0]); 180 181 atomic_set(&ms->default_mirror, m - m0); 182 } 183 184 static struct mirror *get_valid_mirror(struct mirror_set *ms) 185 { 186 struct mirror *m; 187 188 for (m = ms->mirror; m < ms->mirror + ms->nr_mirrors; m++) 189 if (!atomic_read(&m->error_count)) 190 return m; 191 192 return NULL; 193 } 194 195 /* fail_mirror 196 * @m: mirror device to fail 197 * @error_type: one of the enum's, DM_RAID1_*_ERROR 198 * 199 * If errors are being handled, record the type of 200 * error encountered for this device. If this type 201 * of error has already been recorded, we can return; 202 * otherwise, we must signal userspace by triggering 203 * an event. Additionally, if the device is the 204 * primary device, we must choose a new primary, but 205 * only if the mirror is in-sync. 206 * 207 * This function must not block. 208 */ 209 static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type) 210 { 211 struct mirror_set *ms = m->ms; 212 struct mirror *new; 213 214 ms->leg_failure = 1; 215 216 /* 217 * error_count is used for nothing more than a 218 * simple way to tell if a device has encountered 219 * errors. 220 */ 221 atomic_inc(&m->error_count); 222 223 if (test_and_set_bit(error_type, &m->error_type)) 224 return; 225 226 if (!errors_handled(ms)) 227 return; 228 229 if (m != get_default_mirror(ms)) 230 goto out; 231 232 if (!ms->in_sync) { 233 /* 234 * Better to issue requests to same failing device 235 * than to risk returning corrupt data. 236 */ 237 DMERR("Primary mirror (%s) failed while out-of-sync: " 238 "Reads may fail.", m->dev->name); 239 goto out; 240 } 241 242 new = get_valid_mirror(ms); 243 if (new) 244 set_default_mirror(new); 245 else 246 DMWARN("All sides of mirror have failed."); 247 248 out: 249 schedule_work(&ms->trigger_event); 250 } 251 252 static int mirror_flush(struct dm_target *ti) 253 { 254 struct mirror_set *ms = ti->private; 255 unsigned long error_bits; 256 257 unsigned int i; 258 struct dm_io_region io[ms->nr_mirrors]; 259 struct mirror *m; 260 struct dm_io_request io_req = { 261 .bi_rw = WRITE_FLUSH, 262 .mem.type = DM_IO_KMEM, 263 .mem.ptr.addr = NULL, 264 .client = ms->io_client, 265 }; 266 267 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++) { 268 io[i].bdev = m->dev->bdev; 269 io[i].sector = 0; 270 io[i].count = 0; 271 } 272 273 error_bits = -1; 274 dm_io(&io_req, ms->nr_mirrors, io, &error_bits); 275 if (unlikely(error_bits != 0)) { 276 for (i = 0; i < ms->nr_mirrors; i++) 277 if (test_bit(i, &error_bits)) 278 fail_mirror(ms->mirror + i, 279 DM_RAID1_FLUSH_ERROR); 280 return -EIO; 281 } 282 283 return 0; 284 } 285 286 /*----------------------------------------------------------------- 287 * Recovery. 288 * 289 * When a mirror is first activated we may find that some regions 290 * are in the no-sync state. We have to recover these by 291 * recopying from the default mirror to all the others. 292 *---------------------------------------------------------------*/ 293 static void recovery_complete(int read_err, unsigned long write_err, 294 void *context) 295 { 296 struct dm_region *reg = context; 297 struct mirror_set *ms = dm_rh_region_context(reg); 298 int m, bit = 0; 299 300 if (read_err) { 301 /* Read error means the failure of default mirror. */ 302 DMERR_LIMIT("Unable to read primary mirror during recovery"); 303 fail_mirror(get_default_mirror(ms), DM_RAID1_SYNC_ERROR); 304 } 305 306 if (write_err) { 307 DMERR_LIMIT("Write error during recovery (error = 0x%lx)", 308 write_err); 309 /* 310 * Bits correspond to devices (excluding default mirror). 311 * The default mirror cannot change during recovery. 312 */ 313 for (m = 0; m < ms->nr_mirrors; m++) { 314 if (&ms->mirror[m] == get_default_mirror(ms)) 315 continue; 316 if (test_bit(bit, &write_err)) 317 fail_mirror(ms->mirror + m, 318 DM_RAID1_SYNC_ERROR); 319 bit++; 320 } 321 } 322 323 dm_rh_recovery_end(reg, !(read_err || write_err)); 324 } 325 326 static int recover(struct mirror_set *ms, struct dm_region *reg) 327 { 328 int r; 329 unsigned i; 330 struct dm_io_region from, to[DM_KCOPYD_MAX_REGIONS], *dest; 331 struct mirror *m; 332 unsigned long flags = 0; 333 region_t key = dm_rh_get_region_key(reg); 334 sector_t region_size = dm_rh_get_region_size(ms->rh); 335 336 /* fill in the source */ 337 m = get_default_mirror(ms); 338 from.bdev = m->dev->bdev; 339 from.sector = m->offset + dm_rh_region_to_sector(ms->rh, key); 340 if (key == (ms->nr_regions - 1)) { 341 /* 342 * The final region may be smaller than 343 * region_size. 344 */ 345 from.count = ms->ti->len & (region_size - 1); 346 if (!from.count) 347 from.count = region_size; 348 } else 349 from.count = region_size; 350 351 /* fill in the destinations */ 352 for (i = 0, dest = to; i < ms->nr_mirrors; i++) { 353 if (&ms->mirror[i] == get_default_mirror(ms)) 354 continue; 355 356 m = ms->mirror + i; 357 dest->bdev = m->dev->bdev; 358 dest->sector = m->offset + dm_rh_region_to_sector(ms->rh, key); 359 dest->count = from.count; 360 dest++; 361 } 362 363 /* hand to kcopyd */ 364 if (!errors_handled(ms)) 365 set_bit(DM_KCOPYD_IGNORE_ERROR, &flags); 366 367 r = dm_kcopyd_copy(ms->kcopyd_client, &from, ms->nr_mirrors - 1, to, 368 flags, recovery_complete, reg); 369 370 return r; 371 } 372 373 static void do_recovery(struct mirror_set *ms) 374 { 375 struct dm_region *reg; 376 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh); 377 int r; 378 379 /* 380 * Start quiescing some regions. 381 */ 382 dm_rh_recovery_prepare(ms->rh); 383 384 /* 385 * Copy any already quiesced regions. 386 */ 387 while ((reg = dm_rh_recovery_start(ms->rh))) { 388 r = recover(ms, reg); 389 if (r) 390 dm_rh_recovery_end(reg, 0); 391 } 392 393 /* 394 * Update the in sync flag. 395 */ 396 if (!ms->in_sync && 397 (log->type->get_sync_count(log) == ms->nr_regions)) { 398 /* the sync is complete */ 399 dm_table_event(ms->ti->table); 400 ms->in_sync = 1; 401 } 402 } 403 404 /*----------------------------------------------------------------- 405 * Reads 406 *---------------------------------------------------------------*/ 407 static struct mirror *choose_mirror(struct mirror_set *ms, sector_t sector) 408 { 409 struct mirror *m = get_default_mirror(ms); 410 411 do { 412 if (likely(!atomic_read(&m->error_count))) 413 return m; 414 415 if (m-- == ms->mirror) 416 m += ms->nr_mirrors; 417 } while (m != get_default_mirror(ms)); 418 419 return NULL; 420 } 421 422 static int default_ok(struct mirror *m) 423 { 424 struct mirror *default_mirror = get_default_mirror(m->ms); 425 426 return !atomic_read(&default_mirror->error_count); 427 } 428 429 static int mirror_available(struct mirror_set *ms, struct bio *bio) 430 { 431 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh); 432 region_t region = dm_rh_bio_to_region(ms->rh, bio); 433 434 if (log->type->in_sync(log, region, 0)) 435 return choose_mirror(ms, bio->bi_sector) ? 1 : 0; 436 437 return 0; 438 } 439 440 /* 441 * remap a buffer to a particular mirror. 442 */ 443 static sector_t map_sector(struct mirror *m, struct bio *bio) 444 { 445 if (unlikely(!bio->bi_size)) 446 return 0; 447 return m->offset + dm_target_offset(m->ms->ti, bio->bi_sector); 448 } 449 450 static void map_bio(struct mirror *m, struct bio *bio) 451 { 452 bio->bi_bdev = m->dev->bdev; 453 bio->bi_sector = map_sector(m, bio); 454 } 455 456 static void map_region(struct dm_io_region *io, struct mirror *m, 457 struct bio *bio) 458 { 459 io->bdev = m->dev->bdev; 460 io->sector = map_sector(m, bio); 461 io->count = bio_sectors(bio); 462 } 463 464 static void hold_bio(struct mirror_set *ms, struct bio *bio) 465 { 466 /* 467 * Lock is required to avoid race condition during suspend 468 * process. 469 */ 470 spin_lock_irq(&ms->lock); 471 472 if (atomic_read(&ms->suspend)) { 473 spin_unlock_irq(&ms->lock); 474 475 /* 476 * If device is suspended, complete the bio. 477 */ 478 if (dm_noflush_suspending(ms->ti)) 479 bio_endio(bio, DM_ENDIO_REQUEUE); 480 else 481 bio_endio(bio, -EIO); 482 return; 483 } 484 485 /* 486 * Hold bio until the suspend is complete. 487 */ 488 bio_list_add(&ms->holds, bio); 489 spin_unlock_irq(&ms->lock); 490 } 491 492 /*----------------------------------------------------------------- 493 * Reads 494 *---------------------------------------------------------------*/ 495 static void read_callback(unsigned long error, void *context) 496 { 497 struct bio *bio = context; 498 struct mirror *m; 499 500 m = bio_get_m(bio); 501 bio_set_m(bio, NULL); 502 503 if (likely(!error)) { 504 bio_endio(bio, 0); 505 return; 506 } 507 508 fail_mirror(m, DM_RAID1_READ_ERROR); 509 510 if (likely(default_ok(m)) || mirror_available(m->ms, bio)) { 511 DMWARN_LIMIT("Read failure on mirror device %s. " 512 "Trying alternative device.", 513 m->dev->name); 514 queue_bio(m->ms, bio, bio_rw(bio)); 515 return; 516 } 517 518 DMERR_LIMIT("Read failure on mirror device %s. Failing I/O.", 519 m->dev->name); 520 bio_endio(bio, -EIO); 521 } 522 523 /* Asynchronous read. */ 524 static void read_async_bio(struct mirror *m, struct bio *bio) 525 { 526 struct dm_io_region io; 527 struct dm_io_request io_req = { 528 .bi_rw = READ, 529 .mem.type = DM_IO_BVEC, 530 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx, 531 .notify.fn = read_callback, 532 .notify.context = bio, 533 .client = m->ms->io_client, 534 }; 535 536 map_region(&io, m, bio); 537 bio_set_m(bio, m); 538 BUG_ON(dm_io(&io_req, 1, &io, NULL)); 539 } 540 541 static inline int region_in_sync(struct mirror_set *ms, region_t region, 542 int may_block) 543 { 544 int state = dm_rh_get_state(ms->rh, region, may_block); 545 return state == DM_RH_CLEAN || state == DM_RH_DIRTY; 546 } 547 548 static void do_reads(struct mirror_set *ms, struct bio_list *reads) 549 { 550 region_t region; 551 struct bio *bio; 552 struct mirror *m; 553 554 while ((bio = bio_list_pop(reads))) { 555 region = dm_rh_bio_to_region(ms->rh, bio); 556 m = get_default_mirror(ms); 557 558 /* 559 * We can only read balance if the region is in sync. 560 */ 561 if (likely(region_in_sync(ms, region, 1))) 562 m = choose_mirror(ms, bio->bi_sector); 563 else if (m && atomic_read(&m->error_count)) 564 m = NULL; 565 566 if (likely(m)) 567 read_async_bio(m, bio); 568 else 569 bio_endio(bio, -EIO); 570 } 571 } 572 573 /*----------------------------------------------------------------- 574 * Writes. 575 * 576 * We do different things with the write io depending on the 577 * state of the region that it's in: 578 * 579 * SYNC: increment pending, use kcopyd to write to *all* mirrors 580 * RECOVERING: delay the io until recovery completes 581 * NOSYNC: increment pending, just write to the default mirror 582 *---------------------------------------------------------------*/ 583 584 585 static void write_callback(unsigned long error, void *context) 586 { 587 unsigned i, ret = 0; 588 struct bio *bio = (struct bio *) context; 589 struct mirror_set *ms; 590 int should_wake = 0; 591 unsigned long flags; 592 593 ms = bio_get_m(bio)->ms; 594 bio_set_m(bio, NULL); 595 596 /* 597 * NOTE: We don't decrement the pending count here, 598 * instead it is done by the targets endio function. 599 * This way we handle both writes to SYNC and NOSYNC 600 * regions with the same code. 601 */ 602 if (likely(!error)) { 603 bio_endio(bio, ret); 604 return; 605 } 606 607 for (i = 0; i < ms->nr_mirrors; i++) 608 if (test_bit(i, &error)) 609 fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR); 610 611 /* 612 * Need to raise event. Since raising 613 * events can block, we need to do it in 614 * the main thread. 615 */ 616 spin_lock_irqsave(&ms->lock, flags); 617 if (!ms->failures.head) 618 should_wake = 1; 619 bio_list_add(&ms->failures, bio); 620 spin_unlock_irqrestore(&ms->lock, flags); 621 if (should_wake) 622 wakeup_mirrord(ms); 623 } 624 625 static void do_write(struct mirror_set *ms, struct bio *bio) 626 { 627 unsigned int i; 628 struct dm_io_region io[ms->nr_mirrors], *dest = io; 629 struct mirror *m; 630 struct dm_io_request io_req = { 631 .bi_rw = WRITE | (bio->bi_rw & WRITE_FLUSH_FUA), 632 .mem.type = DM_IO_BVEC, 633 .mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx, 634 .notify.fn = write_callback, 635 .notify.context = bio, 636 .client = ms->io_client, 637 }; 638 639 if (bio->bi_rw & REQ_DISCARD) { 640 io_req.bi_rw |= REQ_DISCARD; 641 io_req.mem.type = DM_IO_KMEM; 642 io_req.mem.ptr.addr = NULL; 643 } 644 645 for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++) 646 map_region(dest++, m, bio); 647 648 /* 649 * Use default mirror because we only need it to retrieve the reference 650 * to the mirror set in write_callback(). 651 */ 652 bio_set_m(bio, get_default_mirror(ms)); 653 654 BUG_ON(dm_io(&io_req, ms->nr_mirrors, io, NULL)); 655 } 656 657 static void do_writes(struct mirror_set *ms, struct bio_list *writes) 658 { 659 int state; 660 struct bio *bio; 661 struct bio_list sync, nosync, recover, *this_list = NULL; 662 struct bio_list requeue; 663 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh); 664 region_t region; 665 666 if (!writes->head) 667 return; 668 669 /* 670 * Classify each write. 671 */ 672 bio_list_init(&sync); 673 bio_list_init(&nosync); 674 bio_list_init(&recover); 675 bio_list_init(&requeue); 676 677 while ((bio = bio_list_pop(writes))) { 678 if ((bio->bi_rw & REQ_FLUSH) || 679 (bio->bi_rw & REQ_DISCARD)) { 680 bio_list_add(&sync, bio); 681 continue; 682 } 683 684 region = dm_rh_bio_to_region(ms->rh, bio); 685 686 if (log->type->is_remote_recovering && 687 log->type->is_remote_recovering(log, region)) { 688 bio_list_add(&requeue, bio); 689 continue; 690 } 691 692 state = dm_rh_get_state(ms->rh, region, 1); 693 switch (state) { 694 case DM_RH_CLEAN: 695 case DM_RH_DIRTY: 696 this_list = &sync; 697 break; 698 699 case DM_RH_NOSYNC: 700 this_list = &nosync; 701 break; 702 703 case DM_RH_RECOVERING: 704 this_list = &recover; 705 break; 706 } 707 708 bio_list_add(this_list, bio); 709 } 710 711 /* 712 * Add bios that are delayed due to remote recovery 713 * back on to the write queue 714 */ 715 if (unlikely(requeue.head)) { 716 spin_lock_irq(&ms->lock); 717 bio_list_merge(&ms->writes, &requeue); 718 spin_unlock_irq(&ms->lock); 719 delayed_wake(ms); 720 } 721 722 /* 723 * Increment the pending counts for any regions that will 724 * be written to (writes to recover regions are going to 725 * be delayed). 726 */ 727 dm_rh_inc_pending(ms->rh, &sync); 728 dm_rh_inc_pending(ms->rh, &nosync); 729 730 /* 731 * If the flush fails on a previous call and succeeds here, 732 * we must not reset the log_failure variable. We need 733 * userspace interaction to do that. 734 */ 735 ms->log_failure = dm_rh_flush(ms->rh) ? 1 : ms->log_failure; 736 737 /* 738 * Dispatch io. 739 */ 740 if (unlikely(ms->log_failure) && errors_handled(ms)) { 741 spin_lock_irq(&ms->lock); 742 bio_list_merge(&ms->failures, &sync); 743 spin_unlock_irq(&ms->lock); 744 wakeup_mirrord(ms); 745 } else 746 while ((bio = bio_list_pop(&sync))) 747 do_write(ms, bio); 748 749 while ((bio = bio_list_pop(&recover))) 750 dm_rh_delay(ms->rh, bio); 751 752 while ((bio = bio_list_pop(&nosync))) { 753 if (unlikely(ms->leg_failure) && errors_handled(ms)) { 754 spin_lock_irq(&ms->lock); 755 bio_list_add(&ms->failures, bio); 756 spin_unlock_irq(&ms->lock); 757 wakeup_mirrord(ms); 758 } else { 759 map_bio(get_default_mirror(ms), bio); 760 generic_make_request(bio); 761 } 762 } 763 } 764 765 static void do_failures(struct mirror_set *ms, struct bio_list *failures) 766 { 767 struct bio *bio; 768 769 if (likely(!failures->head)) 770 return; 771 772 /* 773 * If the log has failed, unattempted writes are being 774 * put on the holds list. We can't issue those writes 775 * until a log has been marked, so we must store them. 776 * 777 * If a 'noflush' suspend is in progress, we can requeue 778 * the I/O's to the core. This give userspace a chance 779 * to reconfigure the mirror, at which point the core 780 * will reissue the writes. If the 'noflush' flag is 781 * not set, we have no choice but to return errors. 782 * 783 * Some writes on the failures list may have been 784 * submitted before the log failure and represent a 785 * failure to write to one of the devices. It is ok 786 * for us to treat them the same and requeue them 787 * as well. 788 */ 789 while ((bio = bio_list_pop(failures))) { 790 if (!ms->log_failure) { 791 ms->in_sync = 0; 792 dm_rh_mark_nosync(ms->rh, bio); 793 } 794 795 /* 796 * If all the legs are dead, fail the I/O. 797 * If we have been told to handle errors, hold the bio 798 * and wait for userspace to deal with the problem. 799 * Otherwise pretend that the I/O succeeded. (This would 800 * be wrong if the failed leg returned after reboot and 801 * got replicated back to the good legs.) 802 */ 803 if (!get_valid_mirror(ms)) 804 bio_endio(bio, -EIO); 805 else if (errors_handled(ms)) 806 hold_bio(ms, bio); 807 else 808 bio_endio(bio, 0); 809 } 810 } 811 812 static void trigger_event(struct work_struct *work) 813 { 814 struct mirror_set *ms = 815 container_of(work, struct mirror_set, trigger_event); 816 817 dm_table_event(ms->ti->table); 818 } 819 820 /*----------------------------------------------------------------- 821 * kmirrord 822 *---------------------------------------------------------------*/ 823 static void do_mirror(struct work_struct *work) 824 { 825 struct mirror_set *ms = container_of(work, struct mirror_set, 826 kmirrord_work); 827 struct bio_list reads, writes, failures; 828 unsigned long flags; 829 830 spin_lock_irqsave(&ms->lock, flags); 831 reads = ms->reads; 832 writes = ms->writes; 833 failures = ms->failures; 834 bio_list_init(&ms->reads); 835 bio_list_init(&ms->writes); 836 bio_list_init(&ms->failures); 837 spin_unlock_irqrestore(&ms->lock, flags); 838 839 dm_rh_update_states(ms->rh, errors_handled(ms)); 840 do_recovery(ms); 841 do_reads(ms, &reads); 842 do_writes(ms, &writes); 843 do_failures(ms, &failures); 844 } 845 846 /*----------------------------------------------------------------- 847 * Target functions 848 *---------------------------------------------------------------*/ 849 static struct mirror_set *alloc_context(unsigned int nr_mirrors, 850 uint32_t region_size, 851 struct dm_target *ti, 852 struct dm_dirty_log *dl) 853 { 854 size_t len; 855 struct mirror_set *ms = NULL; 856 857 len = sizeof(*ms) + (sizeof(ms->mirror[0]) * nr_mirrors); 858 859 ms = kzalloc(len, GFP_KERNEL); 860 if (!ms) { 861 ti->error = "Cannot allocate mirror context"; 862 return NULL; 863 } 864 865 spin_lock_init(&ms->lock); 866 bio_list_init(&ms->reads); 867 bio_list_init(&ms->writes); 868 bio_list_init(&ms->failures); 869 bio_list_init(&ms->holds); 870 871 ms->ti = ti; 872 ms->nr_mirrors = nr_mirrors; 873 ms->nr_regions = dm_sector_div_up(ti->len, region_size); 874 ms->in_sync = 0; 875 ms->log_failure = 0; 876 ms->leg_failure = 0; 877 atomic_set(&ms->suspend, 0); 878 atomic_set(&ms->default_mirror, DEFAULT_MIRROR); 879 880 ms->io_client = dm_io_client_create(); 881 if (IS_ERR(ms->io_client)) { 882 ti->error = "Error creating dm_io client"; 883 kfree(ms); 884 return NULL; 885 } 886 887 ms->rh = dm_region_hash_create(ms, dispatch_bios, wakeup_mirrord, 888 wakeup_all_recovery_waiters, 889 ms->ti->begin, MAX_RECOVERY, 890 dl, region_size, ms->nr_regions); 891 if (IS_ERR(ms->rh)) { 892 ti->error = "Error creating dirty region hash"; 893 dm_io_client_destroy(ms->io_client); 894 kfree(ms); 895 return NULL; 896 } 897 898 return ms; 899 } 900 901 static void free_context(struct mirror_set *ms, struct dm_target *ti, 902 unsigned int m) 903 { 904 while (m--) 905 dm_put_device(ti, ms->mirror[m].dev); 906 907 dm_io_client_destroy(ms->io_client); 908 dm_region_hash_destroy(ms->rh); 909 kfree(ms); 910 } 911 912 static int get_mirror(struct mirror_set *ms, struct dm_target *ti, 913 unsigned int mirror, char **argv) 914 { 915 unsigned long long offset; 916 char dummy; 917 918 if (sscanf(argv[1], "%llu%c", &offset, &dummy) != 1) { 919 ti->error = "Invalid offset"; 920 return -EINVAL; 921 } 922 923 if (dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), 924 &ms->mirror[mirror].dev)) { 925 ti->error = "Device lookup failure"; 926 return -ENXIO; 927 } 928 929 ms->mirror[mirror].ms = ms; 930 atomic_set(&(ms->mirror[mirror].error_count), 0); 931 ms->mirror[mirror].error_type = 0; 932 ms->mirror[mirror].offset = offset; 933 934 return 0; 935 } 936 937 /* 938 * Create dirty log: log_type #log_params <log_params> 939 */ 940 static struct dm_dirty_log *create_dirty_log(struct dm_target *ti, 941 unsigned argc, char **argv, 942 unsigned *args_used) 943 { 944 unsigned param_count; 945 struct dm_dirty_log *dl; 946 char dummy; 947 948 if (argc < 2) { 949 ti->error = "Insufficient mirror log arguments"; 950 return NULL; 951 } 952 953 if (sscanf(argv[1], "%u%c", ¶m_count, &dummy) != 1) { 954 ti->error = "Invalid mirror log argument count"; 955 return NULL; 956 } 957 958 *args_used = 2 + param_count; 959 960 if (argc < *args_used) { 961 ti->error = "Insufficient mirror log arguments"; 962 return NULL; 963 } 964 965 dl = dm_dirty_log_create(argv[0], ti, mirror_flush, param_count, 966 argv + 2); 967 if (!dl) { 968 ti->error = "Error creating mirror dirty log"; 969 return NULL; 970 } 971 972 return dl; 973 } 974 975 static int parse_features(struct mirror_set *ms, unsigned argc, char **argv, 976 unsigned *args_used) 977 { 978 unsigned num_features; 979 struct dm_target *ti = ms->ti; 980 char dummy; 981 982 *args_used = 0; 983 984 if (!argc) 985 return 0; 986 987 if (sscanf(argv[0], "%u%c", &num_features, &dummy) != 1) { 988 ti->error = "Invalid number of features"; 989 return -EINVAL; 990 } 991 992 argc--; 993 argv++; 994 (*args_used)++; 995 996 if (num_features > argc) { 997 ti->error = "Not enough arguments to support feature count"; 998 return -EINVAL; 999 } 1000 1001 if (!strcmp("handle_errors", argv[0])) 1002 ms->features |= DM_RAID1_HANDLE_ERRORS; 1003 else { 1004 ti->error = "Unrecognised feature requested"; 1005 return -EINVAL; 1006 } 1007 1008 (*args_used)++; 1009 1010 return 0; 1011 } 1012 1013 /* 1014 * Construct a mirror mapping: 1015 * 1016 * log_type #log_params <log_params> 1017 * #mirrors [mirror_path offset]{2,} 1018 * [#features <features>] 1019 * 1020 * log_type is "core" or "disk" 1021 * #log_params is between 1 and 3 1022 * 1023 * If present, features must be "handle_errors". 1024 */ 1025 static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv) 1026 { 1027 int r; 1028 unsigned int nr_mirrors, m, args_used; 1029 struct mirror_set *ms; 1030 struct dm_dirty_log *dl; 1031 char dummy; 1032 1033 dl = create_dirty_log(ti, argc, argv, &args_used); 1034 if (!dl) 1035 return -EINVAL; 1036 1037 argv += args_used; 1038 argc -= args_used; 1039 1040 if (!argc || sscanf(argv[0], "%u%c", &nr_mirrors, &dummy) != 1 || 1041 nr_mirrors < 2 || nr_mirrors > DM_KCOPYD_MAX_REGIONS + 1) { 1042 ti->error = "Invalid number of mirrors"; 1043 dm_dirty_log_destroy(dl); 1044 return -EINVAL; 1045 } 1046 1047 argv++, argc--; 1048 1049 if (argc < nr_mirrors * 2) { 1050 ti->error = "Too few mirror arguments"; 1051 dm_dirty_log_destroy(dl); 1052 return -EINVAL; 1053 } 1054 1055 ms = alloc_context(nr_mirrors, dl->type->get_region_size(dl), ti, dl); 1056 if (!ms) { 1057 dm_dirty_log_destroy(dl); 1058 return -ENOMEM; 1059 } 1060 1061 /* Get the mirror parameter sets */ 1062 for (m = 0; m < nr_mirrors; m++) { 1063 r = get_mirror(ms, ti, m, argv); 1064 if (r) { 1065 free_context(ms, ti, m); 1066 return r; 1067 } 1068 argv += 2; 1069 argc -= 2; 1070 } 1071 1072 ti->private = ms; 1073 1074 r = dm_set_target_max_io_len(ti, dm_rh_get_region_size(ms->rh)); 1075 if (r) 1076 goto err_free_context; 1077 1078 ti->num_flush_bios = 1; 1079 ti->num_discard_bios = 1; 1080 ti->per_bio_data_size = sizeof(struct dm_raid1_bio_record); 1081 ti->discard_zeroes_data_unsupported = true; 1082 1083 ms->kmirrord_wq = alloc_workqueue("kmirrord", WQ_MEM_RECLAIM, 0); 1084 if (!ms->kmirrord_wq) { 1085 DMERR("couldn't start kmirrord"); 1086 r = -ENOMEM; 1087 goto err_free_context; 1088 } 1089 INIT_WORK(&ms->kmirrord_work, do_mirror); 1090 init_timer(&ms->timer); 1091 ms->timer_pending = 0; 1092 INIT_WORK(&ms->trigger_event, trigger_event); 1093 1094 r = parse_features(ms, argc, argv, &args_used); 1095 if (r) 1096 goto err_destroy_wq; 1097 1098 argv += args_used; 1099 argc -= args_used; 1100 1101 /* 1102 * Any read-balancing addition depends on the 1103 * DM_RAID1_HANDLE_ERRORS flag being present. 1104 * This is because the decision to balance depends 1105 * on the sync state of a region. If the above 1106 * flag is not present, we ignore errors; and 1107 * the sync state may be inaccurate. 1108 */ 1109 1110 if (argc) { 1111 ti->error = "Too many mirror arguments"; 1112 r = -EINVAL; 1113 goto err_destroy_wq; 1114 } 1115 1116 ms->kcopyd_client = dm_kcopyd_client_create(&dm_kcopyd_throttle); 1117 if (IS_ERR(ms->kcopyd_client)) { 1118 r = PTR_ERR(ms->kcopyd_client); 1119 goto err_destroy_wq; 1120 } 1121 1122 wakeup_mirrord(ms); 1123 return 0; 1124 1125 err_destroy_wq: 1126 destroy_workqueue(ms->kmirrord_wq); 1127 err_free_context: 1128 free_context(ms, ti, ms->nr_mirrors); 1129 return r; 1130 } 1131 1132 static void mirror_dtr(struct dm_target *ti) 1133 { 1134 struct mirror_set *ms = (struct mirror_set *) ti->private; 1135 1136 del_timer_sync(&ms->timer); 1137 flush_workqueue(ms->kmirrord_wq); 1138 flush_work(&ms->trigger_event); 1139 dm_kcopyd_client_destroy(ms->kcopyd_client); 1140 destroy_workqueue(ms->kmirrord_wq); 1141 free_context(ms, ti, ms->nr_mirrors); 1142 } 1143 1144 /* 1145 * Mirror mapping function 1146 */ 1147 static int mirror_map(struct dm_target *ti, struct bio *bio) 1148 { 1149 int r, rw = bio_rw(bio); 1150 struct mirror *m; 1151 struct mirror_set *ms = ti->private; 1152 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh); 1153 struct dm_raid1_bio_record *bio_record = 1154 dm_per_bio_data(bio, sizeof(struct dm_raid1_bio_record)); 1155 1156 bio_record->details.bi_bdev = NULL; 1157 1158 if (rw == WRITE) { 1159 /* Save region for mirror_end_io() handler */ 1160 bio_record->write_region = dm_rh_bio_to_region(ms->rh, bio); 1161 queue_bio(ms, bio, rw); 1162 return DM_MAPIO_SUBMITTED; 1163 } 1164 1165 r = log->type->in_sync(log, dm_rh_bio_to_region(ms->rh, bio), 0); 1166 if (r < 0 && r != -EWOULDBLOCK) 1167 return r; 1168 1169 /* 1170 * If region is not in-sync queue the bio. 1171 */ 1172 if (!r || (r == -EWOULDBLOCK)) { 1173 if (rw == READA) 1174 return -EWOULDBLOCK; 1175 1176 queue_bio(ms, bio, rw); 1177 return DM_MAPIO_SUBMITTED; 1178 } 1179 1180 /* 1181 * The region is in-sync and we can perform reads directly. 1182 * Store enough information so we can retry if it fails. 1183 */ 1184 m = choose_mirror(ms, bio->bi_sector); 1185 if (unlikely(!m)) 1186 return -EIO; 1187 1188 dm_bio_record(&bio_record->details, bio); 1189 bio_record->m = m; 1190 1191 map_bio(m, bio); 1192 1193 return DM_MAPIO_REMAPPED; 1194 } 1195 1196 static int mirror_end_io(struct dm_target *ti, struct bio *bio, int error) 1197 { 1198 int rw = bio_rw(bio); 1199 struct mirror_set *ms = (struct mirror_set *) ti->private; 1200 struct mirror *m = NULL; 1201 struct dm_bio_details *bd = NULL; 1202 struct dm_raid1_bio_record *bio_record = 1203 dm_per_bio_data(bio, sizeof(struct dm_raid1_bio_record)); 1204 1205 /* 1206 * We need to dec pending if this was a write. 1207 */ 1208 if (rw == WRITE) { 1209 if (!(bio->bi_rw & (REQ_FLUSH | REQ_DISCARD))) 1210 dm_rh_dec(ms->rh, bio_record->write_region); 1211 return error; 1212 } 1213 1214 if (error == -EOPNOTSUPP) 1215 goto out; 1216 1217 if ((error == -EWOULDBLOCK) && (bio->bi_rw & REQ_RAHEAD)) 1218 goto out; 1219 1220 if (unlikely(error)) { 1221 if (!bio_record->details.bi_bdev) { 1222 /* 1223 * There wasn't enough memory to record necessary 1224 * information for a retry or there was no other 1225 * mirror in-sync. 1226 */ 1227 DMERR_LIMIT("Mirror read failed."); 1228 return -EIO; 1229 } 1230 1231 m = bio_record->m; 1232 1233 DMERR("Mirror read failed from %s. Trying alternative device.", 1234 m->dev->name); 1235 1236 fail_mirror(m, DM_RAID1_READ_ERROR); 1237 1238 /* 1239 * A failed read is requeued for another attempt using an intact 1240 * mirror. 1241 */ 1242 if (default_ok(m) || mirror_available(ms, bio)) { 1243 bd = &bio_record->details; 1244 1245 dm_bio_restore(bd, bio); 1246 bio_record->details.bi_bdev = NULL; 1247 queue_bio(ms, bio, rw); 1248 return DM_ENDIO_INCOMPLETE; 1249 } 1250 DMERR("All replicated volumes dead, failing I/O"); 1251 } 1252 1253 out: 1254 bio_record->details.bi_bdev = NULL; 1255 1256 return error; 1257 } 1258 1259 static void mirror_presuspend(struct dm_target *ti) 1260 { 1261 struct mirror_set *ms = (struct mirror_set *) ti->private; 1262 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh); 1263 1264 struct bio_list holds; 1265 struct bio *bio; 1266 1267 atomic_set(&ms->suspend, 1); 1268 1269 /* 1270 * Process bios in the hold list to start recovery waiting 1271 * for bios in the hold list. After the process, no bio has 1272 * a chance to be added in the hold list because ms->suspend 1273 * is set. 1274 */ 1275 spin_lock_irq(&ms->lock); 1276 holds = ms->holds; 1277 bio_list_init(&ms->holds); 1278 spin_unlock_irq(&ms->lock); 1279 1280 while ((bio = bio_list_pop(&holds))) 1281 hold_bio(ms, bio); 1282 1283 /* 1284 * We must finish up all the work that we've 1285 * generated (i.e. recovery work). 1286 */ 1287 dm_rh_stop_recovery(ms->rh); 1288 1289 wait_event(_kmirrord_recovery_stopped, 1290 !dm_rh_recovery_in_flight(ms->rh)); 1291 1292 if (log->type->presuspend && log->type->presuspend(log)) 1293 /* FIXME: need better error handling */ 1294 DMWARN("log presuspend failed"); 1295 1296 /* 1297 * Now that recovery is complete/stopped and the 1298 * delayed bios are queued, we need to wait for 1299 * the worker thread to complete. This way, 1300 * we know that all of our I/O has been pushed. 1301 */ 1302 flush_workqueue(ms->kmirrord_wq); 1303 } 1304 1305 static void mirror_postsuspend(struct dm_target *ti) 1306 { 1307 struct mirror_set *ms = ti->private; 1308 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh); 1309 1310 if (log->type->postsuspend && log->type->postsuspend(log)) 1311 /* FIXME: need better error handling */ 1312 DMWARN("log postsuspend failed"); 1313 } 1314 1315 static void mirror_resume(struct dm_target *ti) 1316 { 1317 struct mirror_set *ms = ti->private; 1318 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh); 1319 1320 atomic_set(&ms->suspend, 0); 1321 if (log->type->resume && log->type->resume(log)) 1322 /* FIXME: need better error handling */ 1323 DMWARN("log resume failed"); 1324 dm_rh_start_recovery(ms->rh); 1325 } 1326 1327 /* 1328 * device_status_char 1329 * @m: mirror device/leg we want the status of 1330 * 1331 * We return one character representing the most severe error 1332 * we have encountered. 1333 * A => Alive - No failures 1334 * D => Dead - A write failure occurred leaving mirror out-of-sync 1335 * S => Sync - A sychronization failure occurred, mirror out-of-sync 1336 * R => Read - A read failure occurred, mirror data unaffected 1337 * 1338 * Returns: <char> 1339 */ 1340 static char device_status_char(struct mirror *m) 1341 { 1342 if (!atomic_read(&(m->error_count))) 1343 return 'A'; 1344 1345 return (test_bit(DM_RAID1_FLUSH_ERROR, &(m->error_type))) ? 'F' : 1346 (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' : 1347 (test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' : 1348 (test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U'; 1349 } 1350 1351 1352 static void mirror_status(struct dm_target *ti, status_type_t type, 1353 unsigned status_flags, char *result, unsigned maxlen) 1354 { 1355 unsigned int m, sz = 0; 1356 struct mirror_set *ms = (struct mirror_set *) ti->private; 1357 struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh); 1358 char buffer[ms->nr_mirrors + 1]; 1359 1360 switch (type) { 1361 case STATUSTYPE_INFO: 1362 DMEMIT("%d ", ms->nr_mirrors); 1363 for (m = 0; m < ms->nr_mirrors; m++) { 1364 DMEMIT("%s ", ms->mirror[m].dev->name); 1365 buffer[m] = device_status_char(&(ms->mirror[m])); 1366 } 1367 buffer[m] = '\0'; 1368 1369 DMEMIT("%llu/%llu 1 %s ", 1370 (unsigned long long)log->type->get_sync_count(log), 1371 (unsigned long long)ms->nr_regions, buffer); 1372 1373 sz += log->type->status(log, type, result+sz, maxlen-sz); 1374 1375 break; 1376 1377 case STATUSTYPE_TABLE: 1378 sz = log->type->status(log, type, result, maxlen); 1379 1380 DMEMIT("%d", ms->nr_mirrors); 1381 for (m = 0; m < ms->nr_mirrors; m++) 1382 DMEMIT(" %s %llu", ms->mirror[m].dev->name, 1383 (unsigned long long)ms->mirror[m].offset); 1384 1385 if (ms->features & DM_RAID1_HANDLE_ERRORS) 1386 DMEMIT(" 1 handle_errors"); 1387 } 1388 } 1389 1390 static int mirror_iterate_devices(struct dm_target *ti, 1391 iterate_devices_callout_fn fn, void *data) 1392 { 1393 struct mirror_set *ms = ti->private; 1394 int ret = 0; 1395 unsigned i; 1396 1397 for (i = 0; !ret && i < ms->nr_mirrors; i++) 1398 ret = fn(ti, ms->mirror[i].dev, 1399 ms->mirror[i].offset, ti->len, data); 1400 1401 return ret; 1402 } 1403 1404 static struct target_type mirror_target = { 1405 .name = "mirror", 1406 .version = {1, 13, 2}, 1407 .module = THIS_MODULE, 1408 .ctr = mirror_ctr, 1409 .dtr = mirror_dtr, 1410 .map = mirror_map, 1411 .end_io = mirror_end_io, 1412 .presuspend = mirror_presuspend, 1413 .postsuspend = mirror_postsuspend, 1414 .resume = mirror_resume, 1415 .status = mirror_status, 1416 .iterate_devices = mirror_iterate_devices, 1417 }; 1418 1419 static int __init dm_mirror_init(void) 1420 { 1421 int r; 1422 1423 r = dm_register_target(&mirror_target); 1424 if (r < 0) { 1425 DMERR("Failed to register mirror target"); 1426 goto bad_target; 1427 } 1428 1429 return 0; 1430 1431 bad_target: 1432 return r; 1433 } 1434 1435 static void __exit dm_mirror_exit(void) 1436 { 1437 dm_unregister_target(&mirror_target); 1438 } 1439 1440 /* Module hooks */ 1441 module_init(dm_mirror_init); 1442 module_exit(dm_mirror_exit); 1443 1444 MODULE_DESCRIPTION(DM_NAME " mirror target"); 1445 MODULE_AUTHOR("Joe Thornber"); 1446 MODULE_LICENSE("GPL"); 1447