1 /* 2 * raid1.c : Multiple Devices driver for Linux 3 * 4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat 5 * 6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman 7 * 8 * RAID-1 management functions. 9 * 10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000 11 * 12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk> 13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au> 14 * 15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support 16 * bitmapped intelligence in resync: 17 * 18 * - bitmap marked during normal i/o 19 * - bitmap used to skip nondirty blocks during sync 20 * 21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology: 22 * - persistent bitmap code 23 * 24 * This program is free software; you can redistribute it and/or modify 25 * it under the terms of the GNU General Public License as published by 26 * the Free Software Foundation; either version 2, or (at your option) 27 * any later version. 28 * 29 * You should have received a copy of the GNU General Public License 30 * (for example /usr/src/linux/COPYING); if not, write to the Free 31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. 32 */ 33 34 #include <linux/slab.h> 35 #include <linux/delay.h> 36 #include <linux/blkdev.h> 37 #include <linux/module.h> 38 #include <linux/seq_file.h> 39 #include <linux/ratelimit.h> 40 #include "md.h" 41 #include "raid1.h" 42 #include "bitmap.h" 43 44 /* 45 * Number of guaranteed r1bios in case of extreme VM load: 46 */ 47 #define NR_RAID1_BIOS 256 48 49 /* when we get a read error on a read-only array, we redirect to another 50 * device without failing the first device, or trying to over-write to 51 * correct the read error. To keep track of bad blocks on a per-bio 52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer 53 */ 54 #define IO_BLOCKED ((struct bio *)1) 55 /* When we successfully write to a known bad-block, we need to remove the 56 * bad-block marking which must be done from process context. So we record 57 * the success by setting devs[n].bio to IO_MADE_GOOD 58 */ 59 #define IO_MADE_GOOD ((struct bio *)2) 60 61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) 62 63 /* When there are this many requests queue to be written by 64 * the raid1 thread, we become 'congested' to provide back-pressure 65 * for writeback. 66 */ 67 static int max_queued_requests = 1024; 68 69 static void allow_barrier(struct r1conf *conf, sector_t start_next_window, 70 sector_t bi_sector); 71 static void lower_barrier(struct r1conf *conf); 72 73 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data) 74 { 75 struct pool_info *pi = data; 76 int size = offsetof(struct r1bio, bios[pi->raid_disks]); 77 78 /* allocate a r1bio with room for raid_disks entries in the bios array */ 79 return kzalloc(size, gfp_flags); 80 } 81 82 static void r1bio_pool_free(void *r1_bio, void *data) 83 { 84 kfree(r1_bio); 85 } 86 87 #define RESYNC_BLOCK_SIZE (64*1024) 88 #define RESYNC_DEPTH 32 89 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9) 90 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE) 91 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH) 92 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9) 93 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW) 94 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9) 95 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS) 96 97 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data) 98 { 99 struct pool_info *pi = data; 100 struct r1bio *r1_bio; 101 struct bio *bio; 102 int need_pages; 103 int i, j; 104 105 r1_bio = r1bio_pool_alloc(gfp_flags, pi); 106 if (!r1_bio) 107 return NULL; 108 109 /* 110 * Allocate bios : 1 for reading, n-1 for writing 111 */ 112 for (j = pi->raid_disks ; j-- ; ) { 113 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES); 114 if (!bio) 115 goto out_free_bio; 116 r1_bio->bios[j] = bio; 117 } 118 /* 119 * Allocate RESYNC_PAGES data pages and attach them to 120 * the first bio. 121 * If this is a user-requested check/repair, allocate 122 * RESYNC_PAGES for each bio. 123 */ 124 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) 125 need_pages = pi->raid_disks; 126 else 127 need_pages = 1; 128 for (j = 0; j < need_pages; j++) { 129 bio = r1_bio->bios[j]; 130 bio->bi_vcnt = RESYNC_PAGES; 131 132 if (bio_alloc_pages(bio, gfp_flags)) 133 goto out_free_pages; 134 } 135 /* If not user-requests, copy the page pointers to all bios */ 136 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) { 137 for (i=0; i<RESYNC_PAGES ; i++) 138 for (j=1; j<pi->raid_disks; j++) 139 r1_bio->bios[j]->bi_io_vec[i].bv_page = 140 r1_bio->bios[0]->bi_io_vec[i].bv_page; 141 } 142 143 r1_bio->master_bio = NULL; 144 145 return r1_bio; 146 147 out_free_pages: 148 while (--j >= 0) 149 bio_free_pages(r1_bio->bios[j]); 150 151 out_free_bio: 152 while (++j < pi->raid_disks) 153 bio_put(r1_bio->bios[j]); 154 r1bio_pool_free(r1_bio, data); 155 return NULL; 156 } 157 158 static void r1buf_pool_free(void *__r1_bio, void *data) 159 { 160 struct pool_info *pi = data; 161 int i,j; 162 struct r1bio *r1bio = __r1_bio; 163 164 for (i = 0; i < RESYNC_PAGES; i++) 165 for (j = pi->raid_disks; j-- ;) { 166 if (j == 0 || 167 r1bio->bios[j]->bi_io_vec[i].bv_page != 168 r1bio->bios[0]->bi_io_vec[i].bv_page) 169 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page); 170 } 171 for (i=0 ; i < pi->raid_disks; i++) 172 bio_put(r1bio->bios[i]); 173 174 r1bio_pool_free(r1bio, data); 175 } 176 177 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio) 178 { 179 int i; 180 181 for (i = 0; i < conf->raid_disks * 2; i++) { 182 struct bio **bio = r1_bio->bios + i; 183 if (!BIO_SPECIAL(*bio)) 184 bio_put(*bio); 185 *bio = NULL; 186 } 187 } 188 189 static void free_r1bio(struct r1bio *r1_bio) 190 { 191 struct r1conf *conf = r1_bio->mddev->private; 192 193 put_all_bios(conf, r1_bio); 194 mempool_free(r1_bio, conf->r1bio_pool); 195 } 196 197 static void put_buf(struct r1bio *r1_bio) 198 { 199 struct r1conf *conf = r1_bio->mddev->private; 200 int i; 201 202 for (i = 0; i < conf->raid_disks * 2; i++) { 203 struct bio *bio = r1_bio->bios[i]; 204 if (bio->bi_end_io) 205 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev); 206 } 207 208 mempool_free(r1_bio, conf->r1buf_pool); 209 210 lower_barrier(conf); 211 } 212 213 static void reschedule_retry(struct r1bio *r1_bio) 214 { 215 unsigned long flags; 216 struct mddev *mddev = r1_bio->mddev; 217 struct r1conf *conf = mddev->private; 218 219 spin_lock_irqsave(&conf->device_lock, flags); 220 list_add(&r1_bio->retry_list, &conf->retry_list); 221 conf->nr_queued ++; 222 spin_unlock_irqrestore(&conf->device_lock, flags); 223 224 wake_up(&conf->wait_barrier); 225 md_wakeup_thread(mddev->thread); 226 } 227 228 /* 229 * raid_end_bio_io() is called when we have finished servicing a mirrored 230 * operation and are ready to return a success/failure code to the buffer 231 * cache layer. 232 */ 233 static void call_bio_endio(struct r1bio *r1_bio) 234 { 235 struct bio *bio = r1_bio->master_bio; 236 int done; 237 struct r1conf *conf = r1_bio->mddev->private; 238 sector_t start_next_window = r1_bio->start_next_window; 239 sector_t bi_sector = bio->bi_iter.bi_sector; 240 241 if (bio->bi_phys_segments) { 242 unsigned long flags; 243 spin_lock_irqsave(&conf->device_lock, flags); 244 bio->bi_phys_segments--; 245 done = (bio->bi_phys_segments == 0); 246 spin_unlock_irqrestore(&conf->device_lock, flags); 247 /* 248 * make_request() might be waiting for 249 * bi_phys_segments to decrease 250 */ 251 wake_up(&conf->wait_barrier); 252 } else 253 done = 1; 254 255 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 256 bio->bi_error = -EIO; 257 258 if (done) { 259 bio_endio(bio); 260 /* 261 * Wake up any possible resync thread that waits for the device 262 * to go idle. 263 */ 264 allow_barrier(conf, start_next_window, bi_sector); 265 } 266 } 267 268 static void raid_end_bio_io(struct r1bio *r1_bio) 269 { 270 struct bio *bio = r1_bio->master_bio; 271 272 /* if nobody has done the final endio yet, do it now */ 273 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { 274 pr_debug("raid1: sync end %s on sectors %llu-%llu\n", 275 (bio_data_dir(bio) == WRITE) ? "write" : "read", 276 (unsigned long long) bio->bi_iter.bi_sector, 277 (unsigned long long) bio_end_sector(bio) - 1); 278 279 call_bio_endio(r1_bio); 280 } 281 free_r1bio(r1_bio); 282 } 283 284 /* 285 * Update disk head position estimator based on IRQ completion info. 286 */ 287 static inline void update_head_pos(int disk, struct r1bio *r1_bio) 288 { 289 struct r1conf *conf = r1_bio->mddev->private; 290 291 conf->mirrors[disk].head_position = 292 r1_bio->sector + (r1_bio->sectors); 293 } 294 295 /* 296 * Find the disk number which triggered given bio 297 */ 298 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio) 299 { 300 int mirror; 301 struct r1conf *conf = r1_bio->mddev->private; 302 int raid_disks = conf->raid_disks; 303 304 for (mirror = 0; mirror < raid_disks * 2; mirror++) 305 if (r1_bio->bios[mirror] == bio) 306 break; 307 308 BUG_ON(mirror == raid_disks * 2); 309 update_head_pos(mirror, r1_bio); 310 311 return mirror; 312 } 313 314 static void raid1_end_read_request(struct bio *bio) 315 { 316 int uptodate = !bio->bi_error; 317 struct r1bio *r1_bio = bio->bi_private; 318 struct r1conf *conf = r1_bio->mddev->private; 319 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev; 320 321 /* 322 * this branch is our 'one mirror IO has finished' event handler: 323 */ 324 update_head_pos(r1_bio->read_disk, r1_bio); 325 326 if (uptodate) 327 set_bit(R1BIO_Uptodate, &r1_bio->state); 328 else { 329 /* If all other devices have failed, we want to return 330 * the error upwards rather than fail the last device. 331 * Here we redefine "uptodate" to mean "Don't want to retry" 332 */ 333 unsigned long flags; 334 spin_lock_irqsave(&conf->device_lock, flags); 335 if (r1_bio->mddev->degraded == conf->raid_disks || 336 (r1_bio->mddev->degraded == conf->raid_disks-1 && 337 test_bit(In_sync, &rdev->flags))) 338 uptodate = 1; 339 spin_unlock_irqrestore(&conf->device_lock, flags); 340 } 341 342 if (uptodate) { 343 raid_end_bio_io(r1_bio); 344 rdev_dec_pending(rdev, conf->mddev); 345 } else { 346 /* 347 * oops, read error: 348 */ 349 char b[BDEVNAME_SIZE]; 350 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n", 351 mdname(conf->mddev), 352 bdevname(rdev->bdev, b), 353 (unsigned long long)r1_bio->sector); 354 set_bit(R1BIO_ReadError, &r1_bio->state); 355 reschedule_retry(r1_bio); 356 /* don't drop the reference on read_disk yet */ 357 } 358 } 359 360 static void close_write(struct r1bio *r1_bio) 361 { 362 /* it really is the end of this request */ 363 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 364 /* free extra copy of the data pages */ 365 int i = r1_bio->behind_page_count; 366 while (i--) 367 safe_put_page(r1_bio->behind_bvecs[i].bv_page); 368 kfree(r1_bio->behind_bvecs); 369 r1_bio->behind_bvecs = NULL; 370 } 371 /* clear the bitmap if all writes complete successfully */ 372 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector, 373 r1_bio->sectors, 374 !test_bit(R1BIO_Degraded, &r1_bio->state), 375 test_bit(R1BIO_BehindIO, &r1_bio->state)); 376 md_write_end(r1_bio->mddev); 377 } 378 379 static void r1_bio_write_done(struct r1bio *r1_bio) 380 { 381 if (!atomic_dec_and_test(&r1_bio->remaining)) 382 return; 383 384 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 385 reschedule_retry(r1_bio); 386 else { 387 close_write(r1_bio); 388 if (test_bit(R1BIO_MadeGood, &r1_bio->state)) 389 reschedule_retry(r1_bio); 390 else 391 raid_end_bio_io(r1_bio); 392 } 393 } 394 395 static void raid1_end_write_request(struct bio *bio) 396 { 397 struct r1bio *r1_bio = bio->bi_private; 398 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state); 399 struct r1conf *conf = r1_bio->mddev->private; 400 struct bio *to_put = NULL; 401 int mirror = find_bio_disk(r1_bio, bio); 402 struct md_rdev *rdev = conf->mirrors[mirror].rdev; 403 bool discard_error; 404 405 discard_error = bio->bi_error && bio_op(bio) == REQ_OP_DISCARD; 406 407 /* 408 * 'one mirror IO has finished' event handler: 409 */ 410 if (bio->bi_error && !discard_error) { 411 set_bit(WriteErrorSeen, &rdev->flags); 412 if (!test_and_set_bit(WantReplacement, &rdev->flags)) 413 set_bit(MD_RECOVERY_NEEDED, & 414 conf->mddev->recovery); 415 416 set_bit(R1BIO_WriteError, &r1_bio->state); 417 } else { 418 /* 419 * Set R1BIO_Uptodate in our master bio, so that we 420 * will return a good error code for to the higher 421 * levels even if IO on some other mirrored buffer 422 * fails. 423 * 424 * The 'master' represents the composite IO operation 425 * to user-side. So if something waits for IO, then it 426 * will wait for the 'master' bio. 427 */ 428 sector_t first_bad; 429 int bad_sectors; 430 431 r1_bio->bios[mirror] = NULL; 432 to_put = bio; 433 /* 434 * Do not set R1BIO_Uptodate if the current device is 435 * rebuilding or Faulty. This is because we cannot use 436 * such device for properly reading the data back (we could 437 * potentially use it, if the current write would have felt 438 * before rdev->recovery_offset, but for simplicity we don't 439 * check this here. 440 */ 441 if (test_bit(In_sync, &rdev->flags) && 442 !test_bit(Faulty, &rdev->flags)) 443 set_bit(R1BIO_Uptodate, &r1_bio->state); 444 445 /* Maybe we can clear some bad blocks. */ 446 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors, 447 &first_bad, &bad_sectors) && !discard_error) { 448 r1_bio->bios[mirror] = IO_MADE_GOOD; 449 set_bit(R1BIO_MadeGood, &r1_bio->state); 450 } 451 } 452 453 if (behind) { 454 if (test_bit(WriteMostly, &rdev->flags)) 455 atomic_dec(&r1_bio->behind_remaining); 456 457 /* 458 * In behind mode, we ACK the master bio once the I/O 459 * has safely reached all non-writemostly 460 * disks. Setting the Returned bit ensures that this 461 * gets done only once -- we don't ever want to return 462 * -EIO here, instead we'll wait 463 */ 464 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) && 465 test_bit(R1BIO_Uptodate, &r1_bio->state)) { 466 /* Maybe we can return now */ 467 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { 468 struct bio *mbio = r1_bio->master_bio; 469 pr_debug("raid1: behind end write sectors" 470 " %llu-%llu\n", 471 (unsigned long long) mbio->bi_iter.bi_sector, 472 (unsigned long long) bio_end_sector(mbio) - 1); 473 call_bio_endio(r1_bio); 474 } 475 } 476 } 477 if (r1_bio->bios[mirror] == NULL) 478 rdev_dec_pending(rdev, conf->mddev); 479 480 /* 481 * Let's see if all mirrored write operations have finished 482 * already. 483 */ 484 r1_bio_write_done(r1_bio); 485 486 if (to_put) 487 bio_put(to_put); 488 } 489 490 /* 491 * This routine returns the disk from which the requested read should 492 * be done. There is a per-array 'next expected sequential IO' sector 493 * number - if this matches on the next IO then we use the last disk. 494 * There is also a per-disk 'last know head position' sector that is 495 * maintained from IRQ contexts, both the normal and the resync IO 496 * completion handlers update this position correctly. If there is no 497 * perfect sequential match then we pick the disk whose head is closest. 498 * 499 * If there are 2 mirrors in the same 2 devices, performance degrades 500 * because position is mirror, not device based. 501 * 502 * The rdev for the device selected will have nr_pending incremented. 503 */ 504 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors) 505 { 506 const sector_t this_sector = r1_bio->sector; 507 int sectors; 508 int best_good_sectors; 509 int best_disk, best_dist_disk, best_pending_disk; 510 int has_nonrot_disk; 511 int disk; 512 sector_t best_dist; 513 unsigned int min_pending; 514 struct md_rdev *rdev; 515 int choose_first; 516 int choose_next_idle; 517 518 rcu_read_lock(); 519 /* 520 * Check if we can balance. We can balance on the whole 521 * device if no resync is going on, or below the resync window. 522 * We take the first readable disk when above the resync window. 523 */ 524 retry: 525 sectors = r1_bio->sectors; 526 best_disk = -1; 527 best_dist_disk = -1; 528 best_dist = MaxSector; 529 best_pending_disk = -1; 530 min_pending = UINT_MAX; 531 best_good_sectors = 0; 532 has_nonrot_disk = 0; 533 choose_next_idle = 0; 534 535 if ((conf->mddev->recovery_cp < this_sector + sectors) || 536 (mddev_is_clustered(conf->mddev) && 537 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector, 538 this_sector + sectors))) 539 choose_first = 1; 540 else 541 choose_first = 0; 542 543 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) { 544 sector_t dist; 545 sector_t first_bad; 546 int bad_sectors; 547 unsigned int pending; 548 bool nonrot; 549 550 rdev = rcu_dereference(conf->mirrors[disk].rdev); 551 if (r1_bio->bios[disk] == IO_BLOCKED 552 || rdev == NULL 553 || test_bit(Faulty, &rdev->flags)) 554 continue; 555 if (!test_bit(In_sync, &rdev->flags) && 556 rdev->recovery_offset < this_sector + sectors) 557 continue; 558 if (test_bit(WriteMostly, &rdev->flags)) { 559 /* Don't balance among write-mostly, just 560 * use the first as a last resort */ 561 if (best_dist_disk < 0) { 562 if (is_badblock(rdev, this_sector, sectors, 563 &first_bad, &bad_sectors)) { 564 if (first_bad <= this_sector) 565 /* Cannot use this */ 566 continue; 567 best_good_sectors = first_bad - this_sector; 568 } else 569 best_good_sectors = sectors; 570 best_dist_disk = disk; 571 best_pending_disk = disk; 572 } 573 continue; 574 } 575 /* This is a reasonable device to use. It might 576 * even be best. 577 */ 578 if (is_badblock(rdev, this_sector, sectors, 579 &first_bad, &bad_sectors)) { 580 if (best_dist < MaxSector) 581 /* already have a better device */ 582 continue; 583 if (first_bad <= this_sector) { 584 /* cannot read here. If this is the 'primary' 585 * device, then we must not read beyond 586 * bad_sectors from another device.. 587 */ 588 bad_sectors -= (this_sector - first_bad); 589 if (choose_first && sectors > bad_sectors) 590 sectors = bad_sectors; 591 if (best_good_sectors > sectors) 592 best_good_sectors = sectors; 593 594 } else { 595 sector_t good_sectors = first_bad - this_sector; 596 if (good_sectors > best_good_sectors) { 597 best_good_sectors = good_sectors; 598 best_disk = disk; 599 } 600 if (choose_first) 601 break; 602 } 603 continue; 604 } else 605 best_good_sectors = sectors; 606 607 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev)); 608 has_nonrot_disk |= nonrot; 609 pending = atomic_read(&rdev->nr_pending); 610 dist = abs(this_sector - conf->mirrors[disk].head_position); 611 if (choose_first) { 612 best_disk = disk; 613 break; 614 } 615 /* Don't change to another disk for sequential reads */ 616 if (conf->mirrors[disk].next_seq_sect == this_sector 617 || dist == 0) { 618 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9; 619 struct raid1_info *mirror = &conf->mirrors[disk]; 620 621 best_disk = disk; 622 /* 623 * If buffered sequential IO size exceeds optimal 624 * iosize, check if there is idle disk. If yes, choose 625 * the idle disk. read_balance could already choose an 626 * idle disk before noticing it's a sequential IO in 627 * this disk. This doesn't matter because this disk 628 * will idle, next time it will be utilized after the 629 * first disk has IO size exceeds optimal iosize. In 630 * this way, iosize of the first disk will be optimal 631 * iosize at least. iosize of the second disk might be 632 * small, but not a big deal since when the second disk 633 * starts IO, the first disk is likely still busy. 634 */ 635 if (nonrot && opt_iosize > 0 && 636 mirror->seq_start != MaxSector && 637 mirror->next_seq_sect > opt_iosize && 638 mirror->next_seq_sect - opt_iosize >= 639 mirror->seq_start) { 640 choose_next_idle = 1; 641 continue; 642 } 643 break; 644 } 645 /* If device is idle, use it */ 646 if (pending == 0) { 647 best_disk = disk; 648 break; 649 } 650 651 if (choose_next_idle) 652 continue; 653 654 if (min_pending > pending) { 655 min_pending = pending; 656 best_pending_disk = disk; 657 } 658 659 if (dist < best_dist) { 660 best_dist = dist; 661 best_dist_disk = disk; 662 } 663 } 664 665 /* 666 * If all disks are rotational, choose the closest disk. If any disk is 667 * non-rotational, choose the disk with less pending request even the 668 * disk is rotational, which might/might not be optimal for raids with 669 * mixed ratation/non-rotational disks depending on workload. 670 */ 671 if (best_disk == -1) { 672 if (has_nonrot_disk) 673 best_disk = best_pending_disk; 674 else 675 best_disk = best_dist_disk; 676 } 677 678 if (best_disk >= 0) { 679 rdev = rcu_dereference(conf->mirrors[best_disk].rdev); 680 if (!rdev) 681 goto retry; 682 atomic_inc(&rdev->nr_pending); 683 sectors = best_good_sectors; 684 685 if (conf->mirrors[best_disk].next_seq_sect != this_sector) 686 conf->mirrors[best_disk].seq_start = this_sector; 687 688 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors; 689 } 690 rcu_read_unlock(); 691 *max_sectors = sectors; 692 693 return best_disk; 694 } 695 696 static int raid1_congested(struct mddev *mddev, int bits) 697 { 698 struct r1conf *conf = mddev->private; 699 int i, ret = 0; 700 701 if ((bits & (1 << WB_async_congested)) && 702 conf->pending_count >= max_queued_requests) 703 return 1; 704 705 rcu_read_lock(); 706 for (i = 0; i < conf->raid_disks * 2; i++) { 707 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 708 if (rdev && !test_bit(Faulty, &rdev->flags)) { 709 struct request_queue *q = bdev_get_queue(rdev->bdev); 710 711 BUG_ON(!q); 712 713 /* Note the '|| 1' - when read_balance prefers 714 * non-congested targets, it can be removed 715 */ 716 if ((bits & (1 << WB_async_congested)) || 1) 717 ret |= bdi_congested(&q->backing_dev_info, bits); 718 else 719 ret &= bdi_congested(&q->backing_dev_info, bits); 720 } 721 } 722 rcu_read_unlock(); 723 return ret; 724 } 725 726 static void flush_pending_writes(struct r1conf *conf) 727 { 728 /* Any writes that have been queued but are awaiting 729 * bitmap updates get flushed here. 730 */ 731 spin_lock_irq(&conf->device_lock); 732 733 if (conf->pending_bio_list.head) { 734 struct bio *bio; 735 bio = bio_list_get(&conf->pending_bio_list); 736 conf->pending_count = 0; 737 spin_unlock_irq(&conf->device_lock); 738 /* flush any pending bitmap writes to 739 * disk before proceeding w/ I/O */ 740 bitmap_unplug(conf->mddev->bitmap); 741 wake_up(&conf->wait_barrier); 742 743 while (bio) { /* submit pending writes */ 744 struct bio *next = bio->bi_next; 745 bio->bi_next = NULL; 746 if (unlikely((bio_op(bio) == REQ_OP_DISCARD) && 747 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 748 /* Just ignore it */ 749 bio_endio(bio); 750 else 751 generic_make_request(bio); 752 bio = next; 753 } 754 } else 755 spin_unlock_irq(&conf->device_lock); 756 } 757 758 /* Barriers.... 759 * Sometimes we need to suspend IO while we do something else, 760 * either some resync/recovery, or reconfigure the array. 761 * To do this we raise a 'barrier'. 762 * The 'barrier' is a counter that can be raised multiple times 763 * to count how many activities are happening which preclude 764 * normal IO. 765 * We can only raise the barrier if there is no pending IO. 766 * i.e. if nr_pending == 0. 767 * We choose only to raise the barrier if no-one is waiting for the 768 * barrier to go down. This means that as soon as an IO request 769 * is ready, no other operations which require a barrier will start 770 * until the IO request has had a chance. 771 * 772 * So: regular IO calls 'wait_barrier'. When that returns there 773 * is no backgroup IO happening, It must arrange to call 774 * allow_barrier when it has finished its IO. 775 * backgroup IO calls must call raise_barrier. Once that returns 776 * there is no normal IO happeing. It must arrange to call 777 * lower_barrier when the particular background IO completes. 778 */ 779 static void raise_barrier(struct r1conf *conf, sector_t sector_nr) 780 { 781 spin_lock_irq(&conf->resync_lock); 782 783 /* Wait until no block IO is waiting */ 784 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting, 785 conf->resync_lock); 786 787 /* block any new IO from starting */ 788 conf->barrier++; 789 conf->next_resync = sector_nr; 790 791 /* For these conditions we must wait: 792 * A: while the array is in frozen state 793 * B: while barrier >= RESYNC_DEPTH, meaning resync reach 794 * the max count which allowed. 795 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning 796 * next resync will reach to the window which normal bios are 797 * handling. 798 * D: while there are any active requests in the current window. 799 */ 800 wait_event_lock_irq(conf->wait_barrier, 801 !conf->array_frozen && 802 conf->barrier < RESYNC_DEPTH && 803 conf->current_window_requests == 0 && 804 (conf->start_next_window >= 805 conf->next_resync + RESYNC_SECTORS), 806 conf->resync_lock); 807 808 conf->nr_pending++; 809 spin_unlock_irq(&conf->resync_lock); 810 } 811 812 static void lower_barrier(struct r1conf *conf) 813 { 814 unsigned long flags; 815 BUG_ON(conf->barrier <= 0); 816 spin_lock_irqsave(&conf->resync_lock, flags); 817 conf->barrier--; 818 conf->nr_pending--; 819 spin_unlock_irqrestore(&conf->resync_lock, flags); 820 wake_up(&conf->wait_barrier); 821 } 822 823 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio) 824 { 825 bool wait = false; 826 827 if (conf->array_frozen || !bio) 828 wait = true; 829 else if (conf->barrier && bio_data_dir(bio) == WRITE) { 830 if ((conf->mddev->curr_resync_completed 831 >= bio_end_sector(bio)) || 832 (conf->next_resync + NEXT_NORMALIO_DISTANCE 833 <= bio->bi_iter.bi_sector)) 834 wait = false; 835 else 836 wait = true; 837 } 838 839 return wait; 840 } 841 842 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio) 843 { 844 sector_t sector = 0; 845 846 spin_lock_irq(&conf->resync_lock); 847 if (need_to_wait_for_sync(conf, bio)) { 848 conf->nr_waiting++; 849 /* Wait for the barrier to drop. 850 * However if there are already pending 851 * requests (preventing the barrier from 852 * rising completely), and the 853 * per-process bio queue isn't empty, 854 * then don't wait, as we need to empty 855 * that queue to allow conf->start_next_window 856 * to increase. 857 */ 858 wait_event_lock_irq(conf->wait_barrier, 859 !conf->array_frozen && 860 (!conf->barrier || 861 ((conf->start_next_window < 862 conf->next_resync + RESYNC_SECTORS) && 863 current->bio_list && 864 !bio_list_empty(current->bio_list))), 865 conf->resync_lock); 866 conf->nr_waiting--; 867 } 868 869 if (bio && bio_data_dir(bio) == WRITE) { 870 if (bio->bi_iter.bi_sector >= conf->next_resync) { 871 if (conf->start_next_window == MaxSector) 872 conf->start_next_window = 873 conf->next_resync + 874 NEXT_NORMALIO_DISTANCE; 875 876 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE) 877 <= bio->bi_iter.bi_sector) 878 conf->next_window_requests++; 879 else 880 conf->current_window_requests++; 881 sector = conf->start_next_window; 882 } 883 } 884 885 conf->nr_pending++; 886 spin_unlock_irq(&conf->resync_lock); 887 return sector; 888 } 889 890 static void allow_barrier(struct r1conf *conf, sector_t start_next_window, 891 sector_t bi_sector) 892 { 893 unsigned long flags; 894 895 spin_lock_irqsave(&conf->resync_lock, flags); 896 conf->nr_pending--; 897 if (start_next_window) { 898 if (start_next_window == conf->start_next_window) { 899 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE 900 <= bi_sector) 901 conf->next_window_requests--; 902 else 903 conf->current_window_requests--; 904 } else 905 conf->current_window_requests--; 906 907 if (!conf->current_window_requests) { 908 if (conf->next_window_requests) { 909 conf->current_window_requests = 910 conf->next_window_requests; 911 conf->next_window_requests = 0; 912 conf->start_next_window += 913 NEXT_NORMALIO_DISTANCE; 914 } else 915 conf->start_next_window = MaxSector; 916 } 917 } 918 spin_unlock_irqrestore(&conf->resync_lock, flags); 919 wake_up(&conf->wait_barrier); 920 } 921 922 static void freeze_array(struct r1conf *conf, int extra) 923 { 924 /* stop syncio and normal IO and wait for everything to 925 * go quite. 926 * We wait until nr_pending match nr_queued+extra 927 * This is called in the context of one normal IO request 928 * that has failed. Thus any sync request that might be pending 929 * will be blocked by nr_pending, and we need to wait for 930 * pending IO requests to complete or be queued for re-try. 931 * Thus the number queued (nr_queued) plus this request (extra) 932 * must match the number of pending IOs (nr_pending) before 933 * we continue. 934 */ 935 spin_lock_irq(&conf->resync_lock); 936 conf->array_frozen = 1; 937 wait_event_lock_irq_cmd(conf->wait_barrier, 938 conf->nr_pending == conf->nr_queued+extra, 939 conf->resync_lock, 940 flush_pending_writes(conf)); 941 spin_unlock_irq(&conf->resync_lock); 942 } 943 static void unfreeze_array(struct r1conf *conf) 944 { 945 /* reverse the effect of the freeze */ 946 spin_lock_irq(&conf->resync_lock); 947 conf->array_frozen = 0; 948 wake_up(&conf->wait_barrier); 949 spin_unlock_irq(&conf->resync_lock); 950 } 951 952 /* duplicate the data pages for behind I/O 953 */ 954 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio) 955 { 956 int i; 957 struct bio_vec *bvec; 958 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec), 959 GFP_NOIO); 960 if (unlikely(!bvecs)) 961 return; 962 963 bio_for_each_segment_all(bvec, bio, i) { 964 bvecs[i] = *bvec; 965 bvecs[i].bv_page = alloc_page(GFP_NOIO); 966 if (unlikely(!bvecs[i].bv_page)) 967 goto do_sync_io; 968 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset, 969 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len); 970 kunmap(bvecs[i].bv_page); 971 kunmap(bvec->bv_page); 972 } 973 r1_bio->behind_bvecs = bvecs; 974 r1_bio->behind_page_count = bio->bi_vcnt; 975 set_bit(R1BIO_BehindIO, &r1_bio->state); 976 return; 977 978 do_sync_io: 979 for (i = 0; i < bio->bi_vcnt; i++) 980 if (bvecs[i].bv_page) 981 put_page(bvecs[i].bv_page); 982 kfree(bvecs); 983 pr_debug("%dB behind alloc failed, doing sync I/O\n", 984 bio->bi_iter.bi_size); 985 } 986 987 struct raid1_plug_cb { 988 struct blk_plug_cb cb; 989 struct bio_list pending; 990 int pending_cnt; 991 }; 992 993 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule) 994 { 995 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, 996 cb); 997 struct mddev *mddev = plug->cb.data; 998 struct r1conf *conf = mddev->private; 999 struct bio *bio; 1000 1001 if (from_schedule || current->bio_list) { 1002 spin_lock_irq(&conf->device_lock); 1003 bio_list_merge(&conf->pending_bio_list, &plug->pending); 1004 conf->pending_count += plug->pending_cnt; 1005 spin_unlock_irq(&conf->device_lock); 1006 wake_up(&conf->wait_barrier); 1007 md_wakeup_thread(mddev->thread); 1008 kfree(plug); 1009 return; 1010 } 1011 1012 /* we aren't scheduling, so we can do the write-out directly. */ 1013 bio = bio_list_get(&plug->pending); 1014 bitmap_unplug(mddev->bitmap); 1015 wake_up(&conf->wait_barrier); 1016 1017 while (bio) { /* submit pending writes */ 1018 struct bio *next = bio->bi_next; 1019 bio->bi_next = NULL; 1020 if (unlikely((bio_op(bio) == REQ_OP_DISCARD) && 1021 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 1022 /* Just ignore it */ 1023 bio_endio(bio); 1024 else 1025 generic_make_request(bio); 1026 bio = next; 1027 } 1028 kfree(plug); 1029 } 1030 1031 static void raid1_make_request(struct mddev *mddev, struct bio * bio) 1032 { 1033 struct r1conf *conf = mddev->private; 1034 struct raid1_info *mirror; 1035 struct r1bio *r1_bio; 1036 struct bio *read_bio; 1037 int i, disks; 1038 struct bitmap *bitmap; 1039 unsigned long flags; 1040 const int op = bio_op(bio); 1041 const int rw = bio_data_dir(bio); 1042 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC); 1043 const unsigned long do_flush_fua = (bio->bi_opf & 1044 (REQ_PREFLUSH | REQ_FUA)); 1045 struct md_rdev *blocked_rdev; 1046 struct blk_plug_cb *cb; 1047 struct raid1_plug_cb *plug = NULL; 1048 int first_clone; 1049 int sectors_handled; 1050 int max_sectors; 1051 sector_t start_next_window; 1052 1053 /* 1054 * Register the new request and wait if the reconstruction 1055 * thread has put up a bar for new requests. 1056 * Continue immediately if no resync is active currently. 1057 */ 1058 1059 md_write_start(mddev, bio); /* wait on superblock update early */ 1060 1061 if (bio_data_dir(bio) == WRITE && 1062 ((bio_end_sector(bio) > mddev->suspend_lo && 1063 bio->bi_iter.bi_sector < mddev->suspend_hi) || 1064 (mddev_is_clustered(mddev) && 1065 md_cluster_ops->area_resyncing(mddev, WRITE, 1066 bio->bi_iter.bi_sector, bio_end_sector(bio))))) { 1067 /* As the suspend_* range is controlled by 1068 * userspace, we want an interruptible 1069 * wait. 1070 */ 1071 DEFINE_WAIT(w); 1072 for (;;) { 1073 flush_signals(current); 1074 prepare_to_wait(&conf->wait_barrier, 1075 &w, TASK_INTERRUPTIBLE); 1076 if (bio_end_sector(bio) <= mddev->suspend_lo || 1077 bio->bi_iter.bi_sector >= mddev->suspend_hi || 1078 (mddev_is_clustered(mddev) && 1079 !md_cluster_ops->area_resyncing(mddev, WRITE, 1080 bio->bi_iter.bi_sector, bio_end_sector(bio)))) 1081 break; 1082 schedule(); 1083 } 1084 finish_wait(&conf->wait_barrier, &w); 1085 } 1086 1087 start_next_window = wait_barrier(conf, bio); 1088 1089 bitmap = mddev->bitmap; 1090 1091 /* 1092 * make_request() can abort the operation when read-ahead is being 1093 * used and no empty request is available. 1094 * 1095 */ 1096 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1097 1098 r1_bio->master_bio = bio; 1099 r1_bio->sectors = bio_sectors(bio); 1100 r1_bio->state = 0; 1101 r1_bio->mddev = mddev; 1102 r1_bio->sector = bio->bi_iter.bi_sector; 1103 1104 /* We might need to issue multiple reads to different 1105 * devices if there are bad blocks around, so we keep 1106 * track of the number of reads in bio->bi_phys_segments. 1107 * If this is 0, there is only one r1_bio and no locking 1108 * will be needed when requests complete. If it is 1109 * non-zero, then it is the number of not-completed requests. 1110 */ 1111 bio->bi_phys_segments = 0; 1112 bio_clear_flag(bio, BIO_SEG_VALID); 1113 1114 if (rw == READ) { 1115 /* 1116 * read balancing logic: 1117 */ 1118 int rdisk; 1119 1120 read_again: 1121 rdisk = read_balance(conf, r1_bio, &max_sectors); 1122 1123 if (rdisk < 0) { 1124 /* couldn't find anywhere to read from */ 1125 raid_end_bio_io(r1_bio); 1126 return; 1127 } 1128 mirror = conf->mirrors + rdisk; 1129 1130 if (test_bit(WriteMostly, &mirror->rdev->flags) && 1131 bitmap) { 1132 /* Reading from a write-mostly device must 1133 * take care not to over-take any writes 1134 * that are 'behind' 1135 */ 1136 wait_event(bitmap->behind_wait, 1137 atomic_read(&bitmap->behind_writes) == 0); 1138 } 1139 r1_bio->read_disk = rdisk; 1140 r1_bio->start_next_window = 0; 1141 1142 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1143 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector, 1144 max_sectors); 1145 1146 r1_bio->bios[rdisk] = read_bio; 1147 1148 read_bio->bi_iter.bi_sector = r1_bio->sector + 1149 mirror->rdev->data_offset; 1150 read_bio->bi_bdev = mirror->rdev->bdev; 1151 read_bio->bi_end_io = raid1_end_read_request; 1152 bio_set_op_attrs(read_bio, op, do_sync); 1153 read_bio->bi_private = r1_bio; 1154 1155 if (max_sectors < r1_bio->sectors) { 1156 /* could not read all from this device, so we will 1157 * need another r1_bio. 1158 */ 1159 1160 sectors_handled = (r1_bio->sector + max_sectors 1161 - bio->bi_iter.bi_sector); 1162 r1_bio->sectors = max_sectors; 1163 spin_lock_irq(&conf->device_lock); 1164 if (bio->bi_phys_segments == 0) 1165 bio->bi_phys_segments = 2; 1166 else 1167 bio->bi_phys_segments++; 1168 spin_unlock_irq(&conf->device_lock); 1169 /* Cannot call generic_make_request directly 1170 * as that will be queued in __make_request 1171 * and subsequent mempool_alloc might block waiting 1172 * for it. So hand bio over to raid1d. 1173 */ 1174 reschedule_retry(r1_bio); 1175 1176 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1177 1178 r1_bio->master_bio = bio; 1179 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1180 r1_bio->state = 0; 1181 r1_bio->mddev = mddev; 1182 r1_bio->sector = bio->bi_iter.bi_sector + 1183 sectors_handled; 1184 goto read_again; 1185 } else 1186 generic_make_request(read_bio); 1187 return; 1188 } 1189 1190 /* 1191 * WRITE: 1192 */ 1193 if (conf->pending_count >= max_queued_requests) { 1194 md_wakeup_thread(mddev->thread); 1195 wait_event(conf->wait_barrier, 1196 conf->pending_count < max_queued_requests); 1197 } 1198 /* first select target devices under rcu_lock and 1199 * inc refcount on their rdev. Record them by setting 1200 * bios[x] to bio 1201 * If there are known/acknowledged bad blocks on any device on 1202 * which we have seen a write error, we want to avoid writing those 1203 * blocks. 1204 * This potentially requires several writes to write around 1205 * the bad blocks. Each set of writes gets it's own r1bio 1206 * with a set of bios attached. 1207 */ 1208 1209 disks = conf->raid_disks * 2; 1210 retry_write: 1211 r1_bio->start_next_window = start_next_window; 1212 blocked_rdev = NULL; 1213 rcu_read_lock(); 1214 max_sectors = r1_bio->sectors; 1215 for (i = 0; i < disks; i++) { 1216 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1217 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 1218 atomic_inc(&rdev->nr_pending); 1219 blocked_rdev = rdev; 1220 break; 1221 } 1222 r1_bio->bios[i] = NULL; 1223 if (!rdev || test_bit(Faulty, &rdev->flags)) { 1224 if (i < conf->raid_disks) 1225 set_bit(R1BIO_Degraded, &r1_bio->state); 1226 continue; 1227 } 1228 1229 atomic_inc(&rdev->nr_pending); 1230 if (test_bit(WriteErrorSeen, &rdev->flags)) { 1231 sector_t first_bad; 1232 int bad_sectors; 1233 int is_bad; 1234 1235 is_bad = is_badblock(rdev, r1_bio->sector, 1236 max_sectors, 1237 &first_bad, &bad_sectors); 1238 if (is_bad < 0) { 1239 /* mustn't write here until the bad block is 1240 * acknowledged*/ 1241 set_bit(BlockedBadBlocks, &rdev->flags); 1242 blocked_rdev = rdev; 1243 break; 1244 } 1245 if (is_bad && first_bad <= r1_bio->sector) { 1246 /* Cannot write here at all */ 1247 bad_sectors -= (r1_bio->sector - first_bad); 1248 if (bad_sectors < max_sectors) 1249 /* mustn't write more than bad_sectors 1250 * to other devices yet 1251 */ 1252 max_sectors = bad_sectors; 1253 rdev_dec_pending(rdev, mddev); 1254 /* We don't set R1BIO_Degraded as that 1255 * only applies if the disk is 1256 * missing, so it might be re-added, 1257 * and we want to know to recover this 1258 * chunk. 1259 * In this case the device is here, 1260 * and the fact that this chunk is not 1261 * in-sync is recorded in the bad 1262 * block log 1263 */ 1264 continue; 1265 } 1266 if (is_bad) { 1267 int good_sectors = first_bad - r1_bio->sector; 1268 if (good_sectors < max_sectors) 1269 max_sectors = good_sectors; 1270 } 1271 } 1272 r1_bio->bios[i] = bio; 1273 } 1274 rcu_read_unlock(); 1275 1276 if (unlikely(blocked_rdev)) { 1277 /* Wait for this device to become unblocked */ 1278 int j; 1279 sector_t old = start_next_window; 1280 1281 for (j = 0; j < i; j++) 1282 if (r1_bio->bios[j]) 1283 rdev_dec_pending(conf->mirrors[j].rdev, mddev); 1284 r1_bio->state = 0; 1285 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector); 1286 md_wait_for_blocked_rdev(blocked_rdev, mddev); 1287 start_next_window = wait_barrier(conf, bio); 1288 /* 1289 * We must make sure the multi r1bios of bio have 1290 * the same value of bi_phys_segments 1291 */ 1292 if (bio->bi_phys_segments && old && 1293 old != start_next_window) 1294 /* Wait for the former r1bio(s) to complete */ 1295 wait_event(conf->wait_barrier, 1296 bio->bi_phys_segments == 1); 1297 goto retry_write; 1298 } 1299 1300 if (max_sectors < r1_bio->sectors) { 1301 /* We are splitting this write into multiple parts, so 1302 * we need to prepare for allocating another r1_bio. 1303 */ 1304 r1_bio->sectors = max_sectors; 1305 spin_lock_irq(&conf->device_lock); 1306 if (bio->bi_phys_segments == 0) 1307 bio->bi_phys_segments = 2; 1308 else 1309 bio->bi_phys_segments++; 1310 spin_unlock_irq(&conf->device_lock); 1311 } 1312 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector; 1313 1314 atomic_set(&r1_bio->remaining, 1); 1315 atomic_set(&r1_bio->behind_remaining, 0); 1316 1317 first_clone = 1; 1318 for (i = 0; i < disks; i++) { 1319 struct bio *mbio; 1320 if (!r1_bio->bios[i]) 1321 continue; 1322 1323 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1324 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors); 1325 1326 if (first_clone) { 1327 /* do behind I/O ? 1328 * Not if there are too many, or cannot 1329 * allocate memory, or a reader on WriteMostly 1330 * is waiting for behind writes to flush */ 1331 if (bitmap && 1332 (atomic_read(&bitmap->behind_writes) 1333 < mddev->bitmap_info.max_write_behind) && 1334 !waitqueue_active(&bitmap->behind_wait)) 1335 alloc_behind_pages(mbio, r1_bio); 1336 1337 bitmap_startwrite(bitmap, r1_bio->sector, 1338 r1_bio->sectors, 1339 test_bit(R1BIO_BehindIO, 1340 &r1_bio->state)); 1341 first_clone = 0; 1342 } 1343 if (r1_bio->behind_bvecs) { 1344 struct bio_vec *bvec; 1345 int j; 1346 1347 /* 1348 * We trimmed the bio, so _all is legit 1349 */ 1350 bio_for_each_segment_all(bvec, mbio, j) 1351 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page; 1352 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags)) 1353 atomic_inc(&r1_bio->behind_remaining); 1354 } 1355 1356 r1_bio->bios[i] = mbio; 1357 1358 mbio->bi_iter.bi_sector = (r1_bio->sector + 1359 conf->mirrors[i].rdev->data_offset); 1360 mbio->bi_bdev = conf->mirrors[i].rdev->bdev; 1361 mbio->bi_end_io = raid1_end_write_request; 1362 bio_set_op_attrs(mbio, op, do_flush_fua | do_sync); 1363 mbio->bi_private = r1_bio; 1364 1365 atomic_inc(&r1_bio->remaining); 1366 1367 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug)); 1368 if (cb) 1369 plug = container_of(cb, struct raid1_plug_cb, cb); 1370 else 1371 plug = NULL; 1372 spin_lock_irqsave(&conf->device_lock, flags); 1373 if (plug) { 1374 bio_list_add(&plug->pending, mbio); 1375 plug->pending_cnt++; 1376 } else { 1377 bio_list_add(&conf->pending_bio_list, mbio); 1378 conf->pending_count++; 1379 } 1380 spin_unlock_irqrestore(&conf->device_lock, flags); 1381 if (!plug) 1382 md_wakeup_thread(mddev->thread); 1383 } 1384 /* Mustn't call r1_bio_write_done before this next test, 1385 * as it could result in the bio being freed. 1386 */ 1387 if (sectors_handled < bio_sectors(bio)) { 1388 r1_bio_write_done(r1_bio); 1389 /* We need another r1_bio. It has already been counted 1390 * in bio->bi_phys_segments 1391 */ 1392 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1393 r1_bio->master_bio = bio; 1394 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1395 r1_bio->state = 0; 1396 r1_bio->mddev = mddev; 1397 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled; 1398 goto retry_write; 1399 } 1400 1401 r1_bio_write_done(r1_bio); 1402 1403 /* In case raid1d snuck in to freeze_array */ 1404 wake_up(&conf->wait_barrier); 1405 } 1406 1407 static void raid1_status(struct seq_file *seq, struct mddev *mddev) 1408 { 1409 struct r1conf *conf = mddev->private; 1410 int i; 1411 1412 seq_printf(seq, " [%d/%d] [", conf->raid_disks, 1413 conf->raid_disks - mddev->degraded); 1414 rcu_read_lock(); 1415 for (i = 0; i < conf->raid_disks; i++) { 1416 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1417 seq_printf(seq, "%s", 1418 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_"); 1419 } 1420 rcu_read_unlock(); 1421 seq_printf(seq, "]"); 1422 } 1423 1424 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev) 1425 { 1426 char b[BDEVNAME_SIZE]; 1427 struct r1conf *conf = mddev->private; 1428 unsigned long flags; 1429 1430 /* 1431 * If it is not operational, then we have already marked it as dead 1432 * else if it is the last working disks, ignore the error, let the 1433 * next level up know. 1434 * else mark the drive as failed 1435 */ 1436 if (test_bit(In_sync, &rdev->flags) 1437 && (conf->raid_disks - mddev->degraded) == 1) { 1438 /* 1439 * Don't fail the drive, act as though we were just a 1440 * normal single drive. 1441 * However don't try a recovery from this drive as 1442 * it is very likely to fail. 1443 */ 1444 conf->recovery_disabled = mddev->recovery_disabled; 1445 return; 1446 } 1447 set_bit(Blocked, &rdev->flags); 1448 spin_lock_irqsave(&conf->device_lock, flags); 1449 if (test_and_clear_bit(In_sync, &rdev->flags)) { 1450 mddev->degraded++; 1451 set_bit(Faulty, &rdev->flags); 1452 } else 1453 set_bit(Faulty, &rdev->flags); 1454 spin_unlock_irqrestore(&conf->device_lock, flags); 1455 /* 1456 * if recovery is running, make sure it aborts. 1457 */ 1458 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1459 set_mask_bits(&mddev->flags, 0, 1460 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING)); 1461 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n" 1462 "md/raid1:%s: Operation continuing on %d devices.\n", 1463 mdname(mddev), bdevname(rdev->bdev, b), 1464 mdname(mddev), conf->raid_disks - mddev->degraded); 1465 } 1466 1467 static void print_conf(struct r1conf *conf) 1468 { 1469 int i; 1470 1471 pr_debug("RAID1 conf printout:\n"); 1472 if (!conf) { 1473 pr_debug("(!conf)\n"); 1474 return; 1475 } 1476 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded, 1477 conf->raid_disks); 1478 1479 rcu_read_lock(); 1480 for (i = 0; i < conf->raid_disks; i++) { 1481 char b[BDEVNAME_SIZE]; 1482 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1483 if (rdev) 1484 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n", 1485 i, !test_bit(In_sync, &rdev->flags), 1486 !test_bit(Faulty, &rdev->flags), 1487 bdevname(rdev->bdev,b)); 1488 } 1489 rcu_read_unlock(); 1490 } 1491 1492 static void close_sync(struct r1conf *conf) 1493 { 1494 wait_barrier(conf, NULL); 1495 allow_barrier(conf, 0, 0); 1496 1497 mempool_destroy(conf->r1buf_pool); 1498 conf->r1buf_pool = NULL; 1499 1500 spin_lock_irq(&conf->resync_lock); 1501 conf->next_resync = MaxSector - 2 * NEXT_NORMALIO_DISTANCE; 1502 conf->start_next_window = MaxSector; 1503 conf->current_window_requests += 1504 conf->next_window_requests; 1505 conf->next_window_requests = 0; 1506 spin_unlock_irq(&conf->resync_lock); 1507 } 1508 1509 static int raid1_spare_active(struct mddev *mddev) 1510 { 1511 int i; 1512 struct r1conf *conf = mddev->private; 1513 int count = 0; 1514 unsigned long flags; 1515 1516 /* 1517 * Find all failed disks within the RAID1 configuration 1518 * and mark them readable. 1519 * Called under mddev lock, so rcu protection not needed. 1520 * device_lock used to avoid races with raid1_end_read_request 1521 * which expects 'In_sync' flags and ->degraded to be consistent. 1522 */ 1523 spin_lock_irqsave(&conf->device_lock, flags); 1524 for (i = 0; i < conf->raid_disks; i++) { 1525 struct md_rdev *rdev = conf->mirrors[i].rdev; 1526 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev; 1527 if (repl 1528 && !test_bit(Candidate, &repl->flags) 1529 && repl->recovery_offset == MaxSector 1530 && !test_bit(Faulty, &repl->flags) 1531 && !test_and_set_bit(In_sync, &repl->flags)) { 1532 /* replacement has just become active */ 1533 if (!rdev || 1534 !test_and_clear_bit(In_sync, &rdev->flags)) 1535 count++; 1536 if (rdev) { 1537 /* Replaced device not technically 1538 * faulty, but we need to be sure 1539 * it gets removed and never re-added 1540 */ 1541 set_bit(Faulty, &rdev->flags); 1542 sysfs_notify_dirent_safe( 1543 rdev->sysfs_state); 1544 } 1545 } 1546 if (rdev 1547 && rdev->recovery_offset == MaxSector 1548 && !test_bit(Faulty, &rdev->flags) 1549 && !test_and_set_bit(In_sync, &rdev->flags)) { 1550 count++; 1551 sysfs_notify_dirent_safe(rdev->sysfs_state); 1552 } 1553 } 1554 mddev->degraded -= count; 1555 spin_unlock_irqrestore(&conf->device_lock, flags); 1556 1557 print_conf(conf); 1558 return count; 1559 } 1560 1561 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev) 1562 { 1563 struct r1conf *conf = mddev->private; 1564 int err = -EEXIST; 1565 int mirror = 0; 1566 struct raid1_info *p; 1567 int first = 0; 1568 int last = conf->raid_disks - 1; 1569 1570 if (mddev->recovery_disabled == conf->recovery_disabled) 1571 return -EBUSY; 1572 1573 if (md_integrity_add_rdev(rdev, mddev)) 1574 return -ENXIO; 1575 1576 if (rdev->raid_disk >= 0) 1577 first = last = rdev->raid_disk; 1578 1579 /* 1580 * find the disk ... but prefer rdev->saved_raid_disk 1581 * if possible. 1582 */ 1583 if (rdev->saved_raid_disk >= 0 && 1584 rdev->saved_raid_disk >= first && 1585 conf->mirrors[rdev->saved_raid_disk].rdev == NULL) 1586 first = last = rdev->saved_raid_disk; 1587 1588 for (mirror = first; mirror <= last; mirror++) { 1589 p = conf->mirrors+mirror; 1590 if (!p->rdev) { 1591 1592 if (mddev->gendisk) 1593 disk_stack_limits(mddev->gendisk, rdev->bdev, 1594 rdev->data_offset << 9); 1595 1596 p->head_position = 0; 1597 rdev->raid_disk = mirror; 1598 err = 0; 1599 /* As all devices are equivalent, we don't need a full recovery 1600 * if this was recently any drive of the array 1601 */ 1602 if (rdev->saved_raid_disk < 0) 1603 conf->fullsync = 1; 1604 rcu_assign_pointer(p->rdev, rdev); 1605 break; 1606 } 1607 if (test_bit(WantReplacement, &p->rdev->flags) && 1608 p[conf->raid_disks].rdev == NULL) { 1609 /* Add this device as a replacement */ 1610 clear_bit(In_sync, &rdev->flags); 1611 set_bit(Replacement, &rdev->flags); 1612 rdev->raid_disk = mirror; 1613 err = 0; 1614 conf->fullsync = 1; 1615 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev); 1616 break; 1617 } 1618 } 1619 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev))) 1620 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue); 1621 print_conf(conf); 1622 return err; 1623 } 1624 1625 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev) 1626 { 1627 struct r1conf *conf = mddev->private; 1628 int err = 0; 1629 int number = rdev->raid_disk; 1630 struct raid1_info *p = conf->mirrors + number; 1631 1632 if (rdev != p->rdev) 1633 p = conf->mirrors + conf->raid_disks + number; 1634 1635 print_conf(conf); 1636 if (rdev == p->rdev) { 1637 if (test_bit(In_sync, &rdev->flags) || 1638 atomic_read(&rdev->nr_pending)) { 1639 err = -EBUSY; 1640 goto abort; 1641 } 1642 /* Only remove non-faulty devices if recovery 1643 * is not possible. 1644 */ 1645 if (!test_bit(Faulty, &rdev->flags) && 1646 mddev->recovery_disabled != conf->recovery_disabled && 1647 mddev->degraded < conf->raid_disks) { 1648 err = -EBUSY; 1649 goto abort; 1650 } 1651 p->rdev = NULL; 1652 if (!test_bit(RemoveSynchronized, &rdev->flags)) { 1653 synchronize_rcu(); 1654 if (atomic_read(&rdev->nr_pending)) { 1655 /* lost the race, try later */ 1656 err = -EBUSY; 1657 p->rdev = rdev; 1658 goto abort; 1659 } 1660 } 1661 if (conf->mirrors[conf->raid_disks + number].rdev) { 1662 /* We just removed a device that is being replaced. 1663 * Move down the replacement. We drain all IO before 1664 * doing this to avoid confusion. 1665 */ 1666 struct md_rdev *repl = 1667 conf->mirrors[conf->raid_disks + number].rdev; 1668 freeze_array(conf, 0); 1669 clear_bit(Replacement, &repl->flags); 1670 p->rdev = repl; 1671 conf->mirrors[conf->raid_disks + number].rdev = NULL; 1672 unfreeze_array(conf); 1673 clear_bit(WantReplacement, &rdev->flags); 1674 } else 1675 clear_bit(WantReplacement, &rdev->flags); 1676 err = md_integrity_register(mddev); 1677 } 1678 abort: 1679 1680 print_conf(conf); 1681 return err; 1682 } 1683 1684 static void end_sync_read(struct bio *bio) 1685 { 1686 struct r1bio *r1_bio = bio->bi_private; 1687 1688 update_head_pos(r1_bio->read_disk, r1_bio); 1689 1690 /* 1691 * we have read a block, now it needs to be re-written, 1692 * or re-read if the read failed. 1693 * We don't do much here, just schedule handling by raid1d 1694 */ 1695 if (!bio->bi_error) 1696 set_bit(R1BIO_Uptodate, &r1_bio->state); 1697 1698 if (atomic_dec_and_test(&r1_bio->remaining)) 1699 reschedule_retry(r1_bio); 1700 } 1701 1702 static void end_sync_write(struct bio *bio) 1703 { 1704 int uptodate = !bio->bi_error; 1705 struct r1bio *r1_bio = bio->bi_private; 1706 struct mddev *mddev = r1_bio->mddev; 1707 struct r1conf *conf = mddev->private; 1708 sector_t first_bad; 1709 int bad_sectors; 1710 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev; 1711 1712 if (!uptodate) { 1713 sector_t sync_blocks = 0; 1714 sector_t s = r1_bio->sector; 1715 long sectors_to_go = r1_bio->sectors; 1716 /* make sure these bits doesn't get cleared. */ 1717 do { 1718 bitmap_end_sync(mddev->bitmap, s, 1719 &sync_blocks, 1); 1720 s += sync_blocks; 1721 sectors_to_go -= sync_blocks; 1722 } while (sectors_to_go > 0); 1723 set_bit(WriteErrorSeen, &rdev->flags); 1724 if (!test_and_set_bit(WantReplacement, &rdev->flags)) 1725 set_bit(MD_RECOVERY_NEEDED, & 1726 mddev->recovery); 1727 set_bit(R1BIO_WriteError, &r1_bio->state); 1728 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors, 1729 &first_bad, &bad_sectors) && 1730 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev, 1731 r1_bio->sector, 1732 r1_bio->sectors, 1733 &first_bad, &bad_sectors) 1734 ) 1735 set_bit(R1BIO_MadeGood, &r1_bio->state); 1736 1737 if (atomic_dec_and_test(&r1_bio->remaining)) { 1738 int s = r1_bio->sectors; 1739 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 1740 test_bit(R1BIO_WriteError, &r1_bio->state)) 1741 reschedule_retry(r1_bio); 1742 else { 1743 put_buf(r1_bio); 1744 md_done_sync(mddev, s, uptodate); 1745 } 1746 } 1747 } 1748 1749 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector, 1750 int sectors, struct page *page, int rw) 1751 { 1752 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false)) 1753 /* success */ 1754 return 1; 1755 if (rw == WRITE) { 1756 set_bit(WriteErrorSeen, &rdev->flags); 1757 if (!test_and_set_bit(WantReplacement, 1758 &rdev->flags)) 1759 set_bit(MD_RECOVERY_NEEDED, & 1760 rdev->mddev->recovery); 1761 } 1762 /* need to record an error - either for the block or the device */ 1763 if (!rdev_set_badblocks(rdev, sector, sectors, 0)) 1764 md_error(rdev->mddev, rdev); 1765 return 0; 1766 } 1767 1768 static int fix_sync_read_error(struct r1bio *r1_bio) 1769 { 1770 /* Try some synchronous reads of other devices to get 1771 * good data, much like with normal read errors. Only 1772 * read into the pages we already have so we don't 1773 * need to re-issue the read request. 1774 * We don't need to freeze the array, because being in an 1775 * active sync request, there is no normal IO, and 1776 * no overlapping syncs. 1777 * We don't need to check is_badblock() again as we 1778 * made sure that anything with a bad block in range 1779 * will have bi_end_io clear. 1780 */ 1781 struct mddev *mddev = r1_bio->mddev; 1782 struct r1conf *conf = mddev->private; 1783 struct bio *bio = r1_bio->bios[r1_bio->read_disk]; 1784 sector_t sect = r1_bio->sector; 1785 int sectors = r1_bio->sectors; 1786 int idx = 0; 1787 1788 while(sectors) { 1789 int s = sectors; 1790 int d = r1_bio->read_disk; 1791 int success = 0; 1792 struct md_rdev *rdev; 1793 int start; 1794 1795 if (s > (PAGE_SIZE>>9)) 1796 s = PAGE_SIZE >> 9; 1797 do { 1798 if (r1_bio->bios[d]->bi_end_io == end_sync_read) { 1799 /* No rcu protection needed here devices 1800 * can only be removed when no resync is 1801 * active, and resync is currently active 1802 */ 1803 rdev = conf->mirrors[d].rdev; 1804 if (sync_page_io(rdev, sect, s<<9, 1805 bio->bi_io_vec[idx].bv_page, 1806 REQ_OP_READ, 0, false)) { 1807 success = 1; 1808 break; 1809 } 1810 } 1811 d++; 1812 if (d == conf->raid_disks * 2) 1813 d = 0; 1814 } while (!success && d != r1_bio->read_disk); 1815 1816 if (!success) { 1817 char b[BDEVNAME_SIZE]; 1818 int abort = 0; 1819 /* Cannot read from anywhere, this block is lost. 1820 * Record a bad block on each device. If that doesn't 1821 * work just disable and interrupt the recovery. 1822 * Don't fail devices as that won't really help. 1823 */ 1824 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n", 1825 mdname(mddev), 1826 bdevname(bio->bi_bdev, b), 1827 (unsigned long long)r1_bio->sector); 1828 for (d = 0; d < conf->raid_disks * 2; d++) { 1829 rdev = conf->mirrors[d].rdev; 1830 if (!rdev || test_bit(Faulty, &rdev->flags)) 1831 continue; 1832 if (!rdev_set_badblocks(rdev, sect, s, 0)) 1833 abort = 1; 1834 } 1835 if (abort) { 1836 conf->recovery_disabled = 1837 mddev->recovery_disabled; 1838 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1839 md_done_sync(mddev, r1_bio->sectors, 0); 1840 put_buf(r1_bio); 1841 return 0; 1842 } 1843 /* Try next page */ 1844 sectors -= s; 1845 sect += s; 1846 idx++; 1847 continue; 1848 } 1849 1850 start = d; 1851 /* write it back and re-read */ 1852 while (d != r1_bio->read_disk) { 1853 if (d == 0) 1854 d = conf->raid_disks * 2; 1855 d--; 1856 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1857 continue; 1858 rdev = conf->mirrors[d].rdev; 1859 if (r1_sync_page_io(rdev, sect, s, 1860 bio->bi_io_vec[idx].bv_page, 1861 WRITE) == 0) { 1862 r1_bio->bios[d]->bi_end_io = NULL; 1863 rdev_dec_pending(rdev, mddev); 1864 } 1865 } 1866 d = start; 1867 while (d != r1_bio->read_disk) { 1868 if (d == 0) 1869 d = conf->raid_disks * 2; 1870 d--; 1871 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1872 continue; 1873 rdev = conf->mirrors[d].rdev; 1874 if (r1_sync_page_io(rdev, sect, s, 1875 bio->bi_io_vec[idx].bv_page, 1876 READ) != 0) 1877 atomic_add(s, &rdev->corrected_errors); 1878 } 1879 sectors -= s; 1880 sect += s; 1881 idx ++; 1882 } 1883 set_bit(R1BIO_Uptodate, &r1_bio->state); 1884 bio->bi_error = 0; 1885 return 1; 1886 } 1887 1888 static void process_checks(struct r1bio *r1_bio) 1889 { 1890 /* We have read all readable devices. If we haven't 1891 * got the block, then there is no hope left. 1892 * If we have, then we want to do a comparison 1893 * and skip the write if everything is the same. 1894 * If any blocks failed to read, then we need to 1895 * attempt an over-write 1896 */ 1897 struct mddev *mddev = r1_bio->mddev; 1898 struct r1conf *conf = mddev->private; 1899 int primary; 1900 int i; 1901 int vcnt; 1902 1903 /* Fix variable parts of all bios */ 1904 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9); 1905 for (i = 0; i < conf->raid_disks * 2; i++) { 1906 int j; 1907 int size; 1908 int error; 1909 struct bio *b = r1_bio->bios[i]; 1910 if (b->bi_end_io != end_sync_read) 1911 continue; 1912 /* fixup the bio for reuse, but preserve errno */ 1913 error = b->bi_error; 1914 bio_reset(b); 1915 b->bi_error = error; 1916 b->bi_vcnt = vcnt; 1917 b->bi_iter.bi_size = r1_bio->sectors << 9; 1918 b->bi_iter.bi_sector = r1_bio->sector + 1919 conf->mirrors[i].rdev->data_offset; 1920 b->bi_bdev = conf->mirrors[i].rdev->bdev; 1921 b->bi_end_io = end_sync_read; 1922 b->bi_private = r1_bio; 1923 1924 size = b->bi_iter.bi_size; 1925 for (j = 0; j < vcnt ; j++) { 1926 struct bio_vec *bi; 1927 bi = &b->bi_io_vec[j]; 1928 bi->bv_offset = 0; 1929 if (size > PAGE_SIZE) 1930 bi->bv_len = PAGE_SIZE; 1931 else 1932 bi->bv_len = size; 1933 size -= PAGE_SIZE; 1934 } 1935 } 1936 for (primary = 0; primary < conf->raid_disks * 2; primary++) 1937 if (r1_bio->bios[primary]->bi_end_io == end_sync_read && 1938 !r1_bio->bios[primary]->bi_error) { 1939 r1_bio->bios[primary]->bi_end_io = NULL; 1940 rdev_dec_pending(conf->mirrors[primary].rdev, mddev); 1941 break; 1942 } 1943 r1_bio->read_disk = primary; 1944 for (i = 0; i < conf->raid_disks * 2; i++) { 1945 int j; 1946 struct bio *pbio = r1_bio->bios[primary]; 1947 struct bio *sbio = r1_bio->bios[i]; 1948 int error = sbio->bi_error; 1949 1950 if (sbio->bi_end_io != end_sync_read) 1951 continue; 1952 /* Now we can 'fixup' the error value */ 1953 sbio->bi_error = 0; 1954 1955 if (!error) { 1956 for (j = vcnt; j-- ; ) { 1957 struct page *p, *s; 1958 p = pbio->bi_io_vec[j].bv_page; 1959 s = sbio->bi_io_vec[j].bv_page; 1960 if (memcmp(page_address(p), 1961 page_address(s), 1962 sbio->bi_io_vec[j].bv_len)) 1963 break; 1964 } 1965 } else 1966 j = 0; 1967 if (j >= 0) 1968 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches); 1969 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery) 1970 && !error)) { 1971 /* No need to write to this device. */ 1972 sbio->bi_end_io = NULL; 1973 rdev_dec_pending(conf->mirrors[i].rdev, mddev); 1974 continue; 1975 } 1976 1977 bio_copy_data(sbio, pbio); 1978 } 1979 } 1980 1981 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio) 1982 { 1983 struct r1conf *conf = mddev->private; 1984 int i; 1985 int disks = conf->raid_disks * 2; 1986 struct bio *bio, *wbio; 1987 1988 bio = r1_bio->bios[r1_bio->read_disk]; 1989 1990 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 1991 /* ouch - failed to read all of that. */ 1992 if (!fix_sync_read_error(r1_bio)) 1993 return; 1994 1995 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 1996 process_checks(r1_bio); 1997 1998 /* 1999 * schedule writes 2000 */ 2001 atomic_set(&r1_bio->remaining, 1); 2002 for (i = 0; i < disks ; i++) { 2003 wbio = r1_bio->bios[i]; 2004 if (wbio->bi_end_io == NULL || 2005 (wbio->bi_end_io == end_sync_read && 2006 (i == r1_bio->read_disk || 2007 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)))) 2008 continue; 2009 2010 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0); 2011 wbio->bi_end_io = end_sync_write; 2012 atomic_inc(&r1_bio->remaining); 2013 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio)); 2014 2015 generic_make_request(wbio); 2016 } 2017 2018 if (atomic_dec_and_test(&r1_bio->remaining)) { 2019 /* if we're here, all write(s) have completed, so clean up */ 2020 int s = r1_bio->sectors; 2021 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2022 test_bit(R1BIO_WriteError, &r1_bio->state)) 2023 reschedule_retry(r1_bio); 2024 else { 2025 put_buf(r1_bio); 2026 md_done_sync(mddev, s, 1); 2027 } 2028 } 2029 } 2030 2031 /* 2032 * This is a kernel thread which: 2033 * 2034 * 1. Retries failed read operations on working mirrors. 2035 * 2. Updates the raid superblock when problems encounter. 2036 * 3. Performs writes following reads for array synchronising. 2037 */ 2038 2039 static void fix_read_error(struct r1conf *conf, int read_disk, 2040 sector_t sect, int sectors) 2041 { 2042 struct mddev *mddev = conf->mddev; 2043 while(sectors) { 2044 int s = sectors; 2045 int d = read_disk; 2046 int success = 0; 2047 int start; 2048 struct md_rdev *rdev; 2049 2050 if (s > (PAGE_SIZE>>9)) 2051 s = PAGE_SIZE >> 9; 2052 2053 do { 2054 sector_t first_bad; 2055 int bad_sectors; 2056 2057 rcu_read_lock(); 2058 rdev = rcu_dereference(conf->mirrors[d].rdev); 2059 if (rdev && 2060 (test_bit(In_sync, &rdev->flags) || 2061 (!test_bit(Faulty, &rdev->flags) && 2062 rdev->recovery_offset >= sect + s)) && 2063 is_badblock(rdev, sect, s, 2064 &first_bad, &bad_sectors) == 0) { 2065 atomic_inc(&rdev->nr_pending); 2066 rcu_read_unlock(); 2067 if (sync_page_io(rdev, sect, s<<9, 2068 conf->tmppage, REQ_OP_READ, 0, false)) 2069 success = 1; 2070 rdev_dec_pending(rdev, mddev); 2071 if (success) 2072 break; 2073 } else 2074 rcu_read_unlock(); 2075 d++; 2076 if (d == conf->raid_disks * 2) 2077 d = 0; 2078 } while (!success && d != read_disk); 2079 2080 if (!success) { 2081 /* Cannot read from anywhere - mark it bad */ 2082 struct md_rdev *rdev = conf->mirrors[read_disk].rdev; 2083 if (!rdev_set_badblocks(rdev, sect, s, 0)) 2084 md_error(mddev, rdev); 2085 break; 2086 } 2087 /* write it back and re-read */ 2088 start = d; 2089 while (d != read_disk) { 2090 if (d==0) 2091 d = conf->raid_disks * 2; 2092 d--; 2093 rcu_read_lock(); 2094 rdev = rcu_dereference(conf->mirrors[d].rdev); 2095 if (rdev && 2096 !test_bit(Faulty, &rdev->flags)) { 2097 atomic_inc(&rdev->nr_pending); 2098 rcu_read_unlock(); 2099 r1_sync_page_io(rdev, sect, s, 2100 conf->tmppage, WRITE); 2101 rdev_dec_pending(rdev, mddev); 2102 } else 2103 rcu_read_unlock(); 2104 } 2105 d = start; 2106 while (d != read_disk) { 2107 char b[BDEVNAME_SIZE]; 2108 if (d==0) 2109 d = conf->raid_disks * 2; 2110 d--; 2111 rcu_read_lock(); 2112 rdev = rcu_dereference(conf->mirrors[d].rdev); 2113 if (rdev && 2114 !test_bit(Faulty, &rdev->flags)) { 2115 atomic_inc(&rdev->nr_pending); 2116 rcu_read_unlock(); 2117 if (r1_sync_page_io(rdev, sect, s, 2118 conf->tmppage, READ)) { 2119 atomic_add(s, &rdev->corrected_errors); 2120 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n", 2121 mdname(mddev), s, 2122 (unsigned long long)(sect + 2123 rdev->data_offset), 2124 bdevname(rdev->bdev, b)); 2125 } 2126 rdev_dec_pending(rdev, mddev); 2127 } else 2128 rcu_read_unlock(); 2129 } 2130 sectors -= s; 2131 sect += s; 2132 } 2133 } 2134 2135 static int narrow_write_error(struct r1bio *r1_bio, int i) 2136 { 2137 struct mddev *mddev = r1_bio->mddev; 2138 struct r1conf *conf = mddev->private; 2139 struct md_rdev *rdev = conf->mirrors[i].rdev; 2140 2141 /* bio has the data to be written to device 'i' where 2142 * we just recently had a write error. 2143 * We repeatedly clone the bio and trim down to one block, 2144 * then try the write. Where the write fails we record 2145 * a bad block. 2146 * It is conceivable that the bio doesn't exactly align with 2147 * blocks. We must handle this somehow. 2148 * 2149 * We currently own a reference on the rdev. 2150 */ 2151 2152 int block_sectors; 2153 sector_t sector; 2154 int sectors; 2155 int sect_to_write = r1_bio->sectors; 2156 int ok = 1; 2157 2158 if (rdev->badblocks.shift < 0) 2159 return 0; 2160 2161 block_sectors = roundup(1 << rdev->badblocks.shift, 2162 bdev_logical_block_size(rdev->bdev) >> 9); 2163 sector = r1_bio->sector; 2164 sectors = ((sector + block_sectors) 2165 & ~(sector_t)(block_sectors - 1)) 2166 - sector; 2167 2168 while (sect_to_write) { 2169 struct bio *wbio; 2170 if (sectors > sect_to_write) 2171 sectors = sect_to_write; 2172 /* Write at 'sector' for 'sectors'*/ 2173 2174 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 2175 unsigned vcnt = r1_bio->behind_page_count; 2176 struct bio_vec *vec = r1_bio->behind_bvecs; 2177 2178 while (!vec->bv_page) { 2179 vec++; 2180 vcnt--; 2181 } 2182 2183 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev); 2184 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec)); 2185 2186 wbio->bi_vcnt = vcnt; 2187 } else { 2188 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2189 } 2190 2191 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0); 2192 wbio->bi_iter.bi_sector = r1_bio->sector; 2193 wbio->bi_iter.bi_size = r1_bio->sectors << 9; 2194 2195 bio_trim(wbio, sector - r1_bio->sector, sectors); 2196 wbio->bi_iter.bi_sector += rdev->data_offset; 2197 wbio->bi_bdev = rdev->bdev; 2198 2199 if (submit_bio_wait(wbio) < 0) 2200 /* failure! */ 2201 ok = rdev_set_badblocks(rdev, sector, 2202 sectors, 0) 2203 && ok; 2204 2205 bio_put(wbio); 2206 sect_to_write -= sectors; 2207 sector += sectors; 2208 sectors = block_sectors; 2209 } 2210 return ok; 2211 } 2212 2213 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2214 { 2215 int m; 2216 int s = r1_bio->sectors; 2217 for (m = 0; m < conf->raid_disks * 2 ; m++) { 2218 struct md_rdev *rdev = conf->mirrors[m].rdev; 2219 struct bio *bio = r1_bio->bios[m]; 2220 if (bio->bi_end_io == NULL) 2221 continue; 2222 if (!bio->bi_error && 2223 test_bit(R1BIO_MadeGood, &r1_bio->state)) { 2224 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0); 2225 } 2226 if (bio->bi_error && 2227 test_bit(R1BIO_WriteError, &r1_bio->state)) { 2228 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0)) 2229 md_error(conf->mddev, rdev); 2230 } 2231 } 2232 put_buf(r1_bio); 2233 md_done_sync(conf->mddev, s, 1); 2234 } 2235 2236 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2237 { 2238 int m; 2239 bool fail = false; 2240 for (m = 0; m < conf->raid_disks * 2 ; m++) 2241 if (r1_bio->bios[m] == IO_MADE_GOOD) { 2242 struct md_rdev *rdev = conf->mirrors[m].rdev; 2243 rdev_clear_badblocks(rdev, 2244 r1_bio->sector, 2245 r1_bio->sectors, 0); 2246 rdev_dec_pending(rdev, conf->mddev); 2247 } else if (r1_bio->bios[m] != NULL) { 2248 /* This drive got a write error. We need to 2249 * narrow down and record precise write 2250 * errors. 2251 */ 2252 fail = true; 2253 if (!narrow_write_error(r1_bio, m)) { 2254 md_error(conf->mddev, 2255 conf->mirrors[m].rdev); 2256 /* an I/O failed, we can't clear the bitmap */ 2257 set_bit(R1BIO_Degraded, &r1_bio->state); 2258 } 2259 rdev_dec_pending(conf->mirrors[m].rdev, 2260 conf->mddev); 2261 } 2262 if (fail) { 2263 spin_lock_irq(&conf->device_lock); 2264 list_add(&r1_bio->retry_list, &conf->bio_end_io_list); 2265 conf->nr_queued++; 2266 spin_unlock_irq(&conf->device_lock); 2267 md_wakeup_thread(conf->mddev->thread); 2268 } else { 2269 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 2270 close_write(r1_bio); 2271 raid_end_bio_io(r1_bio); 2272 } 2273 } 2274 2275 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio) 2276 { 2277 int disk; 2278 int max_sectors; 2279 struct mddev *mddev = conf->mddev; 2280 struct bio *bio; 2281 char b[BDEVNAME_SIZE]; 2282 struct md_rdev *rdev; 2283 2284 clear_bit(R1BIO_ReadError, &r1_bio->state); 2285 /* we got a read error. Maybe the drive is bad. Maybe just 2286 * the block and we can fix it. 2287 * We freeze all other IO, and try reading the block from 2288 * other devices. When we find one, we re-write 2289 * and check it that fixes the read error. 2290 * This is all done synchronously while the array is 2291 * frozen 2292 */ 2293 2294 bio = r1_bio->bios[r1_bio->read_disk]; 2295 bdevname(bio->bi_bdev, b); 2296 bio_put(bio); 2297 r1_bio->bios[r1_bio->read_disk] = NULL; 2298 2299 if (mddev->ro == 0) { 2300 freeze_array(conf, 1); 2301 fix_read_error(conf, r1_bio->read_disk, 2302 r1_bio->sector, r1_bio->sectors); 2303 unfreeze_array(conf); 2304 } else { 2305 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED; 2306 } 2307 2308 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev); 2309 2310 read_more: 2311 disk = read_balance(conf, r1_bio, &max_sectors); 2312 if (disk == -1) { 2313 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n", 2314 mdname(mddev), b, (unsigned long long)r1_bio->sector); 2315 raid_end_bio_io(r1_bio); 2316 } else { 2317 const unsigned long do_sync 2318 = r1_bio->master_bio->bi_opf & REQ_SYNC; 2319 r1_bio->read_disk = disk; 2320 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2321 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector, 2322 max_sectors); 2323 r1_bio->bios[r1_bio->read_disk] = bio; 2324 rdev = conf->mirrors[disk].rdev; 2325 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n", 2326 mdname(mddev), 2327 (unsigned long long)r1_bio->sector, 2328 bdevname(rdev->bdev, b)); 2329 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset; 2330 bio->bi_bdev = rdev->bdev; 2331 bio->bi_end_io = raid1_end_read_request; 2332 bio_set_op_attrs(bio, REQ_OP_READ, do_sync); 2333 bio->bi_private = r1_bio; 2334 if (max_sectors < r1_bio->sectors) { 2335 /* Drat - have to split this up more */ 2336 struct bio *mbio = r1_bio->master_bio; 2337 int sectors_handled = (r1_bio->sector + max_sectors 2338 - mbio->bi_iter.bi_sector); 2339 r1_bio->sectors = max_sectors; 2340 spin_lock_irq(&conf->device_lock); 2341 if (mbio->bi_phys_segments == 0) 2342 mbio->bi_phys_segments = 2; 2343 else 2344 mbio->bi_phys_segments++; 2345 spin_unlock_irq(&conf->device_lock); 2346 generic_make_request(bio); 2347 bio = NULL; 2348 2349 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 2350 2351 r1_bio->master_bio = mbio; 2352 r1_bio->sectors = bio_sectors(mbio) - sectors_handled; 2353 r1_bio->state = 0; 2354 set_bit(R1BIO_ReadError, &r1_bio->state); 2355 r1_bio->mddev = mddev; 2356 r1_bio->sector = mbio->bi_iter.bi_sector + 2357 sectors_handled; 2358 2359 goto read_more; 2360 } else 2361 generic_make_request(bio); 2362 } 2363 } 2364 2365 static void raid1d(struct md_thread *thread) 2366 { 2367 struct mddev *mddev = thread->mddev; 2368 struct r1bio *r1_bio; 2369 unsigned long flags; 2370 struct r1conf *conf = mddev->private; 2371 struct list_head *head = &conf->retry_list; 2372 struct blk_plug plug; 2373 2374 md_check_recovery(mddev); 2375 2376 if (!list_empty_careful(&conf->bio_end_io_list) && 2377 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) { 2378 LIST_HEAD(tmp); 2379 spin_lock_irqsave(&conf->device_lock, flags); 2380 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) { 2381 while (!list_empty(&conf->bio_end_io_list)) { 2382 list_move(conf->bio_end_io_list.prev, &tmp); 2383 conf->nr_queued--; 2384 } 2385 } 2386 spin_unlock_irqrestore(&conf->device_lock, flags); 2387 while (!list_empty(&tmp)) { 2388 r1_bio = list_first_entry(&tmp, struct r1bio, 2389 retry_list); 2390 list_del(&r1_bio->retry_list); 2391 if (mddev->degraded) 2392 set_bit(R1BIO_Degraded, &r1_bio->state); 2393 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 2394 close_write(r1_bio); 2395 raid_end_bio_io(r1_bio); 2396 } 2397 } 2398 2399 blk_start_plug(&plug); 2400 for (;;) { 2401 2402 flush_pending_writes(conf); 2403 2404 spin_lock_irqsave(&conf->device_lock, flags); 2405 if (list_empty(head)) { 2406 spin_unlock_irqrestore(&conf->device_lock, flags); 2407 break; 2408 } 2409 r1_bio = list_entry(head->prev, struct r1bio, retry_list); 2410 list_del(head->prev); 2411 conf->nr_queued--; 2412 spin_unlock_irqrestore(&conf->device_lock, flags); 2413 2414 mddev = r1_bio->mddev; 2415 conf = mddev->private; 2416 if (test_bit(R1BIO_IsSync, &r1_bio->state)) { 2417 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2418 test_bit(R1BIO_WriteError, &r1_bio->state)) 2419 handle_sync_write_finished(conf, r1_bio); 2420 else 2421 sync_request_write(mddev, r1_bio); 2422 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2423 test_bit(R1BIO_WriteError, &r1_bio->state)) 2424 handle_write_finished(conf, r1_bio); 2425 else if (test_bit(R1BIO_ReadError, &r1_bio->state)) 2426 handle_read_error(conf, r1_bio); 2427 else 2428 /* just a partial read to be scheduled from separate 2429 * context 2430 */ 2431 generic_make_request(r1_bio->bios[r1_bio->read_disk]); 2432 2433 cond_resched(); 2434 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) 2435 md_check_recovery(mddev); 2436 } 2437 blk_finish_plug(&plug); 2438 } 2439 2440 static int init_resync(struct r1conf *conf) 2441 { 2442 int buffs; 2443 2444 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; 2445 BUG_ON(conf->r1buf_pool); 2446 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, 2447 conf->poolinfo); 2448 if (!conf->r1buf_pool) 2449 return -ENOMEM; 2450 conf->next_resync = 0; 2451 return 0; 2452 } 2453 2454 /* 2455 * perform a "sync" on one "block" 2456 * 2457 * We need to make sure that no normal I/O request - particularly write 2458 * requests - conflict with active sync requests. 2459 * 2460 * This is achieved by tracking pending requests and a 'barrier' concept 2461 * that can be installed to exclude normal IO requests. 2462 */ 2463 2464 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr, 2465 int *skipped) 2466 { 2467 struct r1conf *conf = mddev->private; 2468 struct r1bio *r1_bio; 2469 struct bio *bio; 2470 sector_t max_sector, nr_sectors; 2471 int disk = -1; 2472 int i; 2473 int wonly = -1; 2474 int write_targets = 0, read_targets = 0; 2475 sector_t sync_blocks; 2476 int still_degraded = 0; 2477 int good_sectors = RESYNC_SECTORS; 2478 int min_bad = 0; /* number of sectors that are bad in all devices */ 2479 2480 if (!conf->r1buf_pool) 2481 if (init_resync(conf)) 2482 return 0; 2483 2484 max_sector = mddev->dev_sectors; 2485 if (sector_nr >= max_sector) { 2486 /* If we aborted, we need to abort the 2487 * sync on the 'current' bitmap chunk (there will 2488 * only be one in raid1 resync. 2489 * We can find the current addess in mddev->curr_resync 2490 */ 2491 if (mddev->curr_resync < max_sector) /* aborted */ 2492 bitmap_end_sync(mddev->bitmap, mddev->curr_resync, 2493 &sync_blocks, 1); 2494 else /* completed sync */ 2495 conf->fullsync = 0; 2496 2497 bitmap_close_sync(mddev->bitmap); 2498 close_sync(conf); 2499 2500 if (mddev_is_clustered(mddev)) { 2501 conf->cluster_sync_low = 0; 2502 conf->cluster_sync_high = 0; 2503 } 2504 return 0; 2505 } 2506 2507 if (mddev->bitmap == NULL && 2508 mddev->recovery_cp == MaxSector && 2509 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && 2510 conf->fullsync == 0) { 2511 *skipped = 1; 2512 return max_sector - sector_nr; 2513 } 2514 /* before building a request, check if we can skip these blocks.. 2515 * This call the bitmap_start_sync doesn't actually record anything 2516 */ 2517 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) && 2518 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2519 /* We can skip this block, and probably several more */ 2520 *skipped = 1; 2521 return sync_blocks; 2522 } 2523 2524 /* 2525 * If there is non-resync activity waiting for a turn, then let it 2526 * though before starting on this new sync request. 2527 */ 2528 if (conf->nr_waiting) 2529 schedule_timeout_uninterruptible(1); 2530 2531 /* we are incrementing sector_nr below. To be safe, we check against 2532 * sector_nr + two times RESYNC_SECTORS 2533 */ 2534 2535 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 2536 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high)); 2537 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO); 2538 2539 raise_barrier(conf, sector_nr); 2540 2541 rcu_read_lock(); 2542 /* 2543 * If we get a correctably read error during resync or recovery, 2544 * we might want to read from a different device. So we 2545 * flag all drives that could conceivably be read from for READ, 2546 * and any others (which will be non-In_sync devices) for WRITE. 2547 * If a read fails, we try reading from something else for which READ 2548 * is OK. 2549 */ 2550 2551 r1_bio->mddev = mddev; 2552 r1_bio->sector = sector_nr; 2553 r1_bio->state = 0; 2554 set_bit(R1BIO_IsSync, &r1_bio->state); 2555 2556 for (i = 0; i < conf->raid_disks * 2; i++) { 2557 struct md_rdev *rdev; 2558 bio = r1_bio->bios[i]; 2559 bio_reset(bio); 2560 2561 rdev = rcu_dereference(conf->mirrors[i].rdev); 2562 if (rdev == NULL || 2563 test_bit(Faulty, &rdev->flags)) { 2564 if (i < conf->raid_disks) 2565 still_degraded = 1; 2566 } else if (!test_bit(In_sync, &rdev->flags)) { 2567 bio_set_op_attrs(bio, REQ_OP_WRITE, 0); 2568 bio->bi_end_io = end_sync_write; 2569 write_targets ++; 2570 } else { 2571 /* may need to read from here */ 2572 sector_t first_bad = MaxSector; 2573 int bad_sectors; 2574 2575 if (is_badblock(rdev, sector_nr, good_sectors, 2576 &first_bad, &bad_sectors)) { 2577 if (first_bad > sector_nr) 2578 good_sectors = first_bad - sector_nr; 2579 else { 2580 bad_sectors -= (sector_nr - first_bad); 2581 if (min_bad == 0 || 2582 min_bad > bad_sectors) 2583 min_bad = bad_sectors; 2584 } 2585 } 2586 if (sector_nr < first_bad) { 2587 if (test_bit(WriteMostly, &rdev->flags)) { 2588 if (wonly < 0) 2589 wonly = i; 2590 } else { 2591 if (disk < 0) 2592 disk = i; 2593 } 2594 bio_set_op_attrs(bio, REQ_OP_READ, 0); 2595 bio->bi_end_io = end_sync_read; 2596 read_targets++; 2597 } else if (!test_bit(WriteErrorSeen, &rdev->flags) && 2598 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && 2599 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) { 2600 /* 2601 * The device is suitable for reading (InSync), 2602 * but has bad block(s) here. Let's try to correct them, 2603 * if we are doing resync or repair. Otherwise, leave 2604 * this device alone for this sync request. 2605 */ 2606 bio_set_op_attrs(bio, REQ_OP_WRITE, 0); 2607 bio->bi_end_io = end_sync_write; 2608 write_targets++; 2609 } 2610 } 2611 if (bio->bi_end_io) { 2612 atomic_inc(&rdev->nr_pending); 2613 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset; 2614 bio->bi_bdev = rdev->bdev; 2615 bio->bi_private = r1_bio; 2616 } 2617 } 2618 rcu_read_unlock(); 2619 if (disk < 0) 2620 disk = wonly; 2621 r1_bio->read_disk = disk; 2622 2623 if (read_targets == 0 && min_bad > 0) { 2624 /* These sectors are bad on all InSync devices, so we 2625 * need to mark them bad on all write targets 2626 */ 2627 int ok = 1; 2628 for (i = 0 ; i < conf->raid_disks * 2 ; i++) 2629 if (r1_bio->bios[i]->bi_end_io == end_sync_write) { 2630 struct md_rdev *rdev = conf->mirrors[i].rdev; 2631 ok = rdev_set_badblocks(rdev, sector_nr, 2632 min_bad, 0 2633 ) && ok; 2634 } 2635 set_bit(MD_CHANGE_DEVS, &mddev->flags); 2636 *skipped = 1; 2637 put_buf(r1_bio); 2638 2639 if (!ok) { 2640 /* Cannot record the badblocks, so need to 2641 * abort the resync. 2642 * If there are multiple read targets, could just 2643 * fail the really bad ones ??? 2644 */ 2645 conf->recovery_disabled = mddev->recovery_disabled; 2646 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 2647 return 0; 2648 } else 2649 return min_bad; 2650 2651 } 2652 if (min_bad > 0 && min_bad < good_sectors) { 2653 /* only resync enough to reach the next bad->good 2654 * transition */ 2655 good_sectors = min_bad; 2656 } 2657 2658 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0) 2659 /* extra read targets are also write targets */ 2660 write_targets += read_targets-1; 2661 2662 if (write_targets == 0 || read_targets == 0) { 2663 /* There is nowhere to write, so all non-sync 2664 * drives must be failed - so we are finished 2665 */ 2666 sector_t rv; 2667 if (min_bad > 0) 2668 max_sector = sector_nr + min_bad; 2669 rv = max_sector - sector_nr; 2670 *skipped = 1; 2671 put_buf(r1_bio); 2672 return rv; 2673 } 2674 2675 if (max_sector > mddev->resync_max) 2676 max_sector = mddev->resync_max; /* Don't do IO beyond here */ 2677 if (max_sector > sector_nr + good_sectors) 2678 max_sector = sector_nr + good_sectors; 2679 nr_sectors = 0; 2680 sync_blocks = 0; 2681 do { 2682 struct page *page; 2683 int len = PAGE_SIZE; 2684 if (sector_nr + (len>>9) > max_sector) 2685 len = (max_sector - sector_nr) << 9; 2686 if (len == 0) 2687 break; 2688 if (sync_blocks == 0) { 2689 if (!bitmap_start_sync(mddev->bitmap, sector_nr, 2690 &sync_blocks, still_degraded) && 2691 !conf->fullsync && 2692 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2693 break; 2694 if ((len >> 9) > sync_blocks) 2695 len = sync_blocks<<9; 2696 } 2697 2698 for (i = 0 ; i < conf->raid_disks * 2; i++) { 2699 bio = r1_bio->bios[i]; 2700 if (bio->bi_end_io) { 2701 page = bio->bi_io_vec[bio->bi_vcnt].bv_page; 2702 if (bio_add_page(bio, page, len, 0) == 0) { 2703 /* stop here */ 2704 bio->bi_io_vec[bio->bi_vcnt].bv_page = page; 2705 while (i > 0) { 2706 i--; 2707 bio = r1_bio->bios[i]; 2708 if (bio->bi_end_io==NULL) 2709 continue; 2710 /* remove last page from this bio */ 2711 bio->bi_vcnt--; 2712 bio->bi_iter.bi_size -= len; 2713 bio_clear_flag(bio, BIO_SEG_VALID); 2714 } 2715 goto bio_full; 2716 } 2717 } 2718 } 2719 nr_sectors += len>>9; 2720 sector_nr += len>>9; 2721 sync_blocks -= (len>>9); 2722 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES); 2723 bio_full: 2724 r1_bio->sectors = nr_sectors; 2725 2726 if (mddev_is_clustered(mddev) && 2727 conf->cluster_sync_high < sector_nr + nr_sectors) { 2728 conf->cluster_sync_low = mddev->curr_resync_completed; 2729 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS; 2730 /* Send resync message */ 2731 md_cluster_ops->resync_info_update(mddev, 2732 conf->cluster_sync_low, 2733 conf->cluster_sync_high); 2734 } 2735 2736 /* For a user-requested sync, we read all readable devices and do a 2737 * compare 2738 */ 2739 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2740 atomic_set(&r1_bio->remaining, read_targets); 2741 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) { 2742 bio = r1_bio->bios[i]; 2743 if (bio->bi_end_io == end_sync_read) { 2744 read_targets--; 2745 md_sync_acct(bio->bi_bdev, nr_sectors); 2746 generic_make_request(bio); 2747 } 2748 } 2749 } else { 2750 atomic_set(&r1_bio->remaining, 1); 2751 bio = r1_bio->bios[r1_bio->read_disk]; 2752 md_sync_acct(bio->bi_bdev, nr_sectors); 2753 generic_make_request(bio); 2754 2755 } 2756 return nr_sectors; 2757 } 2758 2759 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks) 2760 { 2761 if (sectors) 2762 return sectors; 2763 2764 return mddev->dev_sectors; 2765 } 2766 2767 static struct r1conf *setup_conf(struct mddev *mddev) 2768 { 2769 struct r1conf *conf; 2770 int i; 2771 struct raid1_info *disk; 2772 struct md_rdev *rdev; 2773 int err = -ENOMEM; 2774 2775 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL); 2776 if (!conf) 2777 goto abort; 2778 2779 conf->mirrors = kzalloc(sizeof(struct raid1_info) 2780 * mddev->raid_disks * 2, 2781 GFP_KERNEL); 2782 if (!conf->mirrors) 2783 goto abort; 2784 2785 conf->tmppage = alloc_page(GFP_KERNEL); 2786 if (!conf->tmppage) 2787 goto abort; 2788 2789 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL); 2790 if (!conf->poolinfo) 2791 goto abort; 2792 conf->poolinfo->raid_disks = mddev->raid_disks * 2; 2793 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 2794 r1bio_pool_free, 2795 conf->poolinfo); 2796 if (!conf->r1bio_pool) 2797 goto abort; 2798 2799 conf->poolinfo->mddev = mddev; 2800 2801 err = -EINVAL; 2802 spin_lock_init(&conf->device_lock); 2803 rdev_for_each(rdev, mddev) { 2804 struct request_queue *q; 2805 int disk_idx = rdev->raid_disk; 2806 if (disk_idx >= mddev->raid_disks 2807 || disk_idx < 0) 2808 continue; 2809 if (test_bit(Replacement, &rdev->flags)) 2810 disk = conf->mirrors + mddev->raid_disks + disk_idx; 2811 else 2812 disk = conf->mirrors + disk_idx; 2813 2814 if (disk->rdev) 2815 goto abort; 2816 disk->rdev = rdev; 2817 q = bdev_get_queue(rdev->bdev); 2818 2819 disk->head_position = 0; 2820 disk->seq_start = MaxSector; 2821 } 2822 conf->raid_disks = mddev->raid_disks; 2823 conf->mddev = mddev; 2824 INIT_LIST_HEAD(&conf->retry_list); 2825 INIT_LIST_HEAD(&conf->bio_end_io_list); 2826 2827 spin_lock_init(&conf->resync_lock); 2828 init_waitqueue_head(&conf->wait_barrier); 2829 2830 bio_list_init(&conf->pending_bio_list); 2831 conf->pending_count = 0; 2832 conf->recovery_disabled = mddev->recovery_disabled - 1; 2833 2834 conf->start_next_window = MaxSector; 2835 conf->current_window_requests = conf->next_window_requests = 0; 2836 2837 err = -EIO; 2838 for (i = 0; i < conf->raid_disks * 2; i++) { 2839 2840 disk = conf->mirrors + i; 2841 2842 if (i < conf->raid_disks && 2843 disk[conf->raid_disks].rdev) { 2844 /* This slot has a replacement. */ 2845 if (!disk->rdev) { 2846 /* No original, just make the replacement 2847 * a recovering spare 2848 */ 2849 disk->rdev = 2850 disk[conf->raid_disks].rdev; 2851 disk[conf->raid_disks].rdev = NULL; 2852 } else if (!test_bit(In_sync, &disk->rdev->flags)) 2853 /* Original is not in_sync - bad */ 2854 goto abort; 2855 } 2856 2857 if (!disk->rdev || 2858 !test_bit(In_sync, &disk->rdev->flags)) { 2859 disk->head_position = 0; 2860 if (disk->rdev && 2861 (disk->rdev->saved_raid_disk < 0)) 2862 conf->fullsync = 1; 2863 } 2864 } 2865 2866 err = -ENOMEM; 2867 conf->thread = md_register_thread(raid1d, mddev, "raid1"); 2868 if (!conf->thread) 2869 goto abort; 2870 2871 return conf; 2872 2873 abort: 2874 if (conf) { 2875 mempool_destroy(conf->r1bio_pool); 2876 kfree(conf->mirrors); 2877 safe_put_page(conf->tmppage); 2878 kfree(conf->poolinfo); 2879 kfree(conf); 2880 } 2881 return ERR_PTR(err); 2882 } 2883 2884 static void raid1_free(struct mddev *mddev, void *priv); 2885 static int raid1_run(struct mddev *mddev) 2886 { 2887 struct r1conf *conf; 2888 int i; 2889 struct md_rdev *rdev; 2890 int ret; 2891 bool discard_supported = false; 2892 2893 if (mddev->level != 1) { 2894 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n", 2895 mdname(mddev), mddev->level); 2896 return -EIO; 2897 } 2898 if (mddev->reshape_position != MaxSector) { 2899 pr_warn("md/raid1:%s: reshape_position set but not supported\n", 2900 mdname(mddev)); 2901 return -EIO; 2902 } 2903 /* 2904 * copy the already verified devices into our private RAID1 2905 * bookkeeping area. [whatever we allocate in run(), 2906 * should be freed in raid1_free()] 2907 */ 2908 if (mddev->private == NULL) 2909 conf = setup_conf(mddev); 2910 else 2911 conf = mddev->private; 2912 2913 if (IS_ERR(conf)) 2914 return PTR_ERR(conf); 2915 2916 if (mddev->queue) 2917 blk_queue_max_write_same_sectors(mddev->queue, 0); 2918 2919 rdev_for_each(rdev, mddev) { 2920 if (!mddev->gendisk) 2921 continue; 2922 disk_stack_limits(mddev->gendisk, rdev->bdev, 2923 rdev->data_offset << 9); 2924 if (blk_queue_discard(bdev_get_queue(rdev->bdev))) 2925 discard_supported = true; 2926 } 2927 2928 mddev->degraded = 0; 2929 for (i=0; i < conf->raid_disks; i++) 2930 if (conf->mirrors[i].rdev == NULL || 2931 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) || 2932 test_bit(Faulty, &conf->mirrors[i].rdev->flags)) 2933 mddev->degraded++; 2934 2935 if (conf->raid_disks - mddev->degraded == 1) 2936 mddev->recovery_cp = MaxSector; 2937 2938 if (mddev->recovery_cp != MaxSector) 2939 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n", 2940 mdname(mddev)); 2941 pr_info("md/raid1:%s: active with %d out of %d mirrors\n", 2942 mdname(mddev), mddev->raid_disks - mddev->degraded, 2943 mddev->raid_disks); 2944 2945 /* 2946 * Ok, everything is just fine now 2947 */ 2948 mddev->thread = conf->thread; 2949 conf->thread = NULL; 2950 mddev->private = conf; 2951 2952 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0)); 2953 2954 if (mddev->queue) { 2955 if (discard_supported) 2956 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, 2957 mddev->queue); 2958 else 2959 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, 2960 mddev->queue); 2961 } 2962 2963 ret = md_integrity_register(mddev); 2964 if (ret) { 2965 md_unregister_thread(&mddev->thread); 2966 raid1_free(mddev, conf); 2967 } 2968 return ret; 2969 } 2970 2971 static void raid1_free(struct mddev *mddev, void *priv) 2972 { 2973 struct r1conf *conf = priv; 2974 2975 mempool_destroy(conf->r1bio_pool); 2976 kfree(conf->mirrors); 2977 safe_put_page(conf->tmppage); 2978 kfree(conf->poolinfo); 2979 kfree(conf); 2980 } 2981 2982 static int raid1_resize(struct mddev *mddev, sector_t sectors) 2983 { 2984 /* no resync is happening, and there is enough space 2985 * on all devices, so we can resize. 2986 * We need to make sure resync covers any new space. 2987 * If the array is shrinking we should possibly wait until 2988 * any io in the removed space completes, but it hardly seems 2989 * worth it. 2990 */ 2991 sector_t newsize = raid1_size(mddev, sectors, 0); 2992 if (mddev->external_size && 2993 mddev->array_sectors > newsize) 2994 return -EINVAL; 2995 if (mddev->bitmap) { 2996 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0); 2997 if (ret) 2998 return ret; 2999 } 3000 md_set_array_sectors(mddev, newsize); 3001 set_capacity(mddev->gendisk, mddev->array_sectors); 3002 revalidate_disk(mddev->gendisk); 3003 if (sectors > mddev->dev_sectors && 3004 mddev->recovery_cp > mddev->dev_sectors) { 3005 mddev->recovery_cp = mddev->dev_sectors; 3006 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3007 } 3008 mddev->dev_sectors = sectors; 3009 mddev->resync_max_sectors = sectors; 3010 return 0; 3011 } 3012 3013 static int raid1_reshape(struct mddev *mddev) 3014 { 3015 /* We need to: 3016 * 1/ resize the r1bio_pool 3017 * 2/ resize conf->mirrors 3018 * 3019 * We allocate a new r1bio_pool if we can. 3020 * Then raise a device barrier and wait until all IO stops. 3021 * Then resize conf->mirrors and swap in the new r1bio pool. 3022 * 3023 * At the same time, we "pack" the devices so that all the missing 3024 * devices have the higher raid_disk numbers. 3025 */ 3026 mempool_t *newpool, *oldpool; 3027 struct pool_info *newpoolinfo; 3028 struct raid1_info *newmirrors; 3029 struct r1conf *conf = mddev->private; 3030 int cnt, raid_disks; 3031 unsigned long flags; 3032 int d, d2, err; 3033 3034 /* Cannot change chunk_size, layout, or level */ 3035 if (mddev->chunk_sectors != mddev->new_chunk_sectors || 3036 mddev->layout != mddev->new_layout || 3037 mddev->level != mddev->new_level) { 3038 mddev->new_chunk_sectors = mddev->chunk_sectors; 3039 mddev->new_layout = mddev->layout; 3040 mddev->new_level = mddev->level; 3041 return -EINVAL; 3042 } 3043 3044 if (!mddev_is_clustered(mddev)) { 3045 err = md_allow_write(mddev); 3046 if (err) 3047 return err; 3048 } 3049 3050 raid_disks = mddev->raid_disks + mddev->delta_disks; 3051 3052 if (raid_disks < conf->raid_disks) { 3053 cnt=0; 3054 for (d= 0; d < conf->raid_disks; d++) 3055 if (conf->mirrors[d].rdev) 3056 cnt++; 3057 if (cnt > raid_disks) 3058 return -EBUSY; 3059 } 3060 3061 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL); 3062 if (!newpoolinfo) 3063 return -ENOMEM; 3064 newpoolinfo->mddev = mddev; 3065 newpoolinfo->raid_disks = raid_disks * 2; 3066 3067 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 3068 r1bio_pool_free, newpoolinfo); 3069 if (!newpool) { 3070 kfree(newpoolinfo); 3071 return -ENOMEM; 3072 } 3073 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2, 3074 GFP_KERNEL); 3075 if (!newmirrors) { 3076 kfree(newpoolinfo); 3077 mempool_destroy(newpool); 3078 return -ENOMEM; 3079 } 3080 3081 freeze_array(conf, 0); 3082 3083 /* ok, everything is stopped */ 3084 oldpool = conf->r1bio_pool; 3085 conf->r1bio_pool = newpool; 3086 3087 for (d = d2 = 0; d < conf->raid_disks; d++) { 3088 struct md_rdev *rdev = conf->mirrors[d].rdev; 3089 if (rdev && rdev->raid_disk != d2) { 3090 sysfs_unlink_rdev(mddev, rdev); 3091 rdev->raid_disk = d2; 3092 sysfs_unlink_rdev(mddev, rdev); 3093 if (sysfs_link_rdev(mddev, rdev)) 3094 pr_warn("md/raid1:%s: cannot register rd%d\n", 3095 mdname(mddev), rdev->raid_disk); 3096 } 3097 if (rdev) 3098 newmirrors[d2++].rdev = rdev; 3099 } 3100 kfree(conf->mirrors); 3101 conf->mirrors = newmirrors; 3102 kfree(conf->poolinfo); 3103 conf->poolinfo = newpoolinfo; 3104 3105 spin_lock_irqsave(&conf->device_lock, flags); 3106 mddev->degraded += (raid_disks - conf->raid_disks); 3107 spin_unlock_irqrestore(&conf->device_lock, flags); 3108 conf->raid_disks = mddev->raid_disks = raid_disks; 3109 mddev->delta_disks = 0; 3110 3111 unfreeze_array(conf); 3112 3113 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery); 3114 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3115 md_wakeup_thread(mddev->thread); 3116 3117 mempool_destroy(oldpool); 3118 return 0; 3119 } 3120 3121 static void raid1_quiesce(struct mddev *mddev, int state) 3122 { 3123 struct r1conf *conf = mddev->private; 3124 3125 switch(state) { 3126 case 2: /* wake for suspend */ 3127 wake_up(&conf->wait_barrier); 3128 break; 3129 case 1: 3130 freeze_array(conf, 0); 3131 break; 3132 case 0: 3133 unfreeze_array(conf); 3134 break; 3135 } 3136 } 3137 3138 static void *raid1_takeover(struct mddev *mddev) 3139 { 3140 /* raid1 can take over: 3141 * raid5 with 2 devices, any layout or chunk size 3142 */ 3143 if (mddev->level == 5 && mddev->raid_disks == 2) { 3144 struct r1conf *conf; 3145 mddev->new_level = 1; 3146 mddev->new_layout = 0; 3147 mddev->new_chunk_sectors = 0; 3148 conf = setup_conf(mddev); 3149 if (!IS_ERR(conf)) 3150 /* Array must appear to be quiesced */ 3151 conf->array_frozen = 1; 3152 return conf; 3153 } 3154 return ERR_PTR(-EINVAL); 3155 } 3156 3157 static struct md_personality raid1_personality = 3158 { 3159 .name = "raid1", 3160 .level = 1, 3161 .owner = THIS_MODULE, 3162 .make_request = raid1_make_request, 3163 .run = raid1_run, 3164 .free = raid1_free, 3165 .status = raid1_status, 3166 .error_handler = raid1_error, 3167 .hot_add_disk = raid1_add_disk, 3168 .hot_remove_disk= raid1_remove_disk, 3169 .spare_active = raid1_spare_active, 3170 .sync_request = raid1_sync_request, 3171 .resize = raid1_resize, 3172 .size = raid1_size, 3173 .check_reshape = raid1_reshape, 3174 .quiesce = raid1_quiesce, 3175 .takeover = raid1_takeover, 3176 .congested = raid1_congested, 3177 }; 3178 3179 static int __init raid_init(void) 3180 { 3181 return register_md_personality(&raid1_personality); 3182 } 3183 3184 static void raid_exit(void) 3185 { 3186 unregister_md_personality(&raid1_personality); 3187 } 3188 3189 module_init(raid_init); 3190 module_exit(raid_exit); 3191 MODULE_LICENSE("GPL"); 3192 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD"); 3193 MODULE_ALIAS("md-personality-3"); /* RAID1 */ 3194 MODULE_ALIAS("md-raid1"); 3195 MODULE_ALIAS("md-level-1"); 3196 3197 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR); 3198