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 struct bio_vec *bv; 150 151 bio_for_each_segment_all(bv, r1_bio->bios[j], i) 152 __free_page(bv->bv_page); 153 } 154 155 out_free_bio: 156 while (++j < pi->raid_disks) 157 bio_put(r1_bio->bios[j]); 158 r1bio_pool_free(r1_bio, data); 159 return NULL; 160 } 161 162 static void r1buf_pool_free(void *__r1_bio, void *data) 163 { 164 struct pool_info *pi = data; 165 int i,j; 166 struct r1bio *r1bio = __r1_bio; 167 168 for (i = 0; i < RESYNC_PAGES; i++) 169 for (j = pi->raid_disks; j-- ;) { 170 if (j == 0 || 171 r1bio->bios[j]->bi_io_vec[i].bv_page != 172 r1bio->bios[0]->bi_io_vec[i].bv_page) 173 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page); 174 } 175 for (i=0 ; i < pi->raid_disks; i++) 176 bio_put(r1bio->bios[i]); 177 178 r1bio_pool_free(r1bio, data); 179 } 180 181 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio) 182 { 183 int i; 184 185 for (i = 0; i < conf->raid_disks * 2; i++) { 186 struct bio **bio = r1_bio->bios + i; 187 if (!BIO_SPECIAL(*bio)) 188 bio_put(*bio); 189 *bio = NULL; 190 } 191 } 192 193 static void free_r1bio(struct r1bio *r1_bio) 194 { 195 struct r1conf *conf = r1_bio->mddev->private; 196 197 put_all_bios(conf, r1_bio); 198 mempool_free(r1_bio, conf->r1bio_pool); 199 } 200 201 static void put_buf(struct r1bio *r1_bio) 202 { 203 struct r1conf *conf = r1_bio->mddev->private; 204 int i; 205 206 for (i = 0; i < conf->raid_disks * 2; i++) { 207 struct bio *bio = r1_bio->bios[i]; 208 if (bio->bi_end_io) 209 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev); 210 } 211 212 mempool_free(r1_bio, conf->r1buf_pool); 213 214 lower_barrier(conf); 215 } 216 217 static void reschedule_retry(struct r1bio *r1_bio) 218 { 219 unsigned long flags; 220 struct mddev *mddev = r1_bio->mddev; 221 struct r1conf *conf = mddev->private; 222 223 spin_lock_irqsave(&conf->device_lock, flags); 224 list_add(&r1_bio->retry_list, &conf->retry_list); 225 conf->nr_queued ++; 226 spin_unlock_irqrestore(&conf->device_lock, flags); 227 228 wake_up(&conf->wait_barrier); 229 md_wakeup_thread(mddev->thread); 230 } 231 232 /* 233 * raid_end_bio_io() is called when we have finished servicing a mirrored 234 * operation and are ready to return a success/failure code to the buffer 235 * cache layer. 236 */ 237 static void call_bio_endio(struct r1bio *r1_bio) 238 { 239 struct bio *bio = r1_bio->master_bio; 240 int done; 241 struct r1conf *conf = r1_bio->mddev->private; 242 sector_t start_next_window = r1_bio->start_next_window; 243 sector_t bi_sector = bio->bi_iter.bi_sector; 244 245 if (bio->bi_phys_segments) { 246 unsigned long flags; 247 spin_lock_irqsave(&conf->device_lock, flags); 248 bio->bi_phys_segments--; 249 done = (bio->bi_phys_segments == 0); 250 spin_unlock_irqrestore(&conf->device_lock, flags); 251 /* 252 * make_request() might be waiting for 253 * bi_phys_segments to decrease 254 */ 255 wake_up(&conf->wait_barrier); 256 } else 257 done = 1; 258 259 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 260 bio->bi_error = -EIO; 261 262 if (done) { 263 bio_endio(bio); 264 /* 265 * Wake up any possible resync thread that waits for the device 266 * to go idle. 267 */ 268 allow_barrier(conf, start_next_window, bi_sector); 269 } 270 } 271 272 static void raid_end_bio_io(struct r1bio *r1_bio) 273 { 274 struct bio *bio = r1_bio->master_bio; 275 276 /* if nobody has done the final endio yet, do it now */ 277 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { 278 pr_debug("raid1: sync end %s on sectors %llu-%llu\n", 279 (bio_data_dir(bio) == WRITE) ? "write" : "read", 280 (unsigned long long) bio->bi_iter.bi_sector, 281 (unsigned long long) bio_end_sector(bio) - 1); 282 283 call_bio_endio(r1_bio); 284 } 285 free_r1bio(r1_bio); 286 } 287 288 /* 289 * Update disk head position estimator based on IRQ completion info. 290 */ 291 static inline void update_head_pos(int disk, struct r1bio *r1_bio) 292 { 293 struct r1conf *conf = r1_bio->mddev->private; 294 295 conf->mirrors[disk].head_position = 296 r1_bio->sector + (r1_bio->sectors); 297 } 298 299 /* 300 * Find the disk number which triggered given bio 301 */ 302 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio) 303 { 304 int mirror; 305 struct r1conf *conf = r1_bio->mddev->private; 306 int raid_disks = conf->raid_disks; 307 308 for (mirror = 0; mirror < raid_disks * 2; mirror++) 309 if (r1_bio->bios[mirror] == bio) 310 break; 311 312 BUG_ON(mirror == raid_disks * 2); 313 update_head_pos(mirror, r1_bio); 314 315 return mirror; 316 } 317 318 static void raid1_end_read_request(struct bio *bio) 319 { 320 int uptodate = !bio->bi_error; 321 struct r1bio *r1_bio = bio->bi_private; 322 int mirror; 323 struct r1conf *conf = r1_bio->mddev->private; 324 325 mirror = r1_bio->read_disk; 326 /* 327 * this branch is our 'one mirror IO has finished' event handler: 328 */ 329 update_head_pos(mirror, r1_bio); 330 331 if (uptodate) 332 set_bit(R1BIO_Uptodate, &r1_bio->state); 333 else { 334 /* If all other devices have failed, we want to return 335 * the error upwards rather than fail the last device. 336 * Here we redefine "uptodate" to mean "Don't want to retry" 337 */ 338 unsigned long flags; 339 spin_lock_irqsave(&conf->device_lock, flags); 340 if (r1_bio->mddev->degraded == conf->raid_disks || 341 (r1_bio->mddev->degraded == conf->raid_disks-1 && 342 test_bit(In_sync, &conf->mirrors[mirror].rdev->flags))) 343 uptodate = 1; 344 spin_unlock_irqrestore(&conf->device_lock, flags); 345 } 346 347 if (uptodate) { 348 raid_end_bio_io(r1_bio); 349 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev); 350 } else { 351 /* 352 * oops, read error: 353 */ 354 char b[BDEVNAME_SIZE]; 355 printk_ratelimited( 356 KERN_ERR "md/raid1:%s: %s: " 357 "rescheduling sector %llu\n", 358 mdname(conf->mddev), 359 bdevname(conf->mirrors[mirror].rdev->bdev, 360 b), 361 (unsigned long long)r1_bio->sector); 362 set_bit(R1BIO_ReadError, &r1_bio->state); 363 reschedule_retry(r1_bio); 364 /* don't drop the reference on read_disk yet */ 365 } 366 } 367 368 static void close_write(struct r1bio *r1_bio) 369 { 370 /* it really is the end of this request */ 371 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 372 /* free extra copy of the data pages */ 373 int i = r1_bio->behind_page_count; 374 while (i--) 375 safe_put_page(r1_bio->behind_bvecs[i].bv_page); 376 kfree(r1_bio->behind_bvecs); 377 r1_bio->behind_bvecs = NULL; 378 } 379 /* clear the bitmap if all writes complete successfully */ 380 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector, 381 r1_bio->sectors, 382 !test_bit(R1BIO_Degraded, &r1_bio->state), 383 test_bit(R1BIO_BehindIO, &r1_bio->state)); 384 md_write_end(r1_bio->mddev); 385 } 386 387 static void r1_bio_write_done(struct r1bio *r1_bio) 388 { 389 if (!atomic_dec_and_test(&r1_bio->remaining)) 390 return; 391 392 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 393 reschedule_retry(r1_bio); 394 else { 395 close_write(r1_bio); 396 if (test_bit(R1BIO_MadeGood, &r1_bio->state)) 397 reschedule_retry(r1_bio); 398 else 399 raid_end_bio_io(r1_bio); 400 } 401 } 402 403 static void raid1_end_write_request(struct bio *bio) 404 { 405 struct r1bio *r1_bio = bio->bi_private; 406 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state); 407 struct r1conf *conf = r1_bio->mddev->private; 408 struct bio *to_put = NULL; 409 410 mirror = find_bio_disk(r1_bio, bio); 411 412 /* 413 * 'one mirror IO has finished' event handler: 414 */ 415 if (bio->bi_error) { 416 set_bit(WriteErrorSeen, 417 &conf->mirrors[mirror].rdev->flags); 418 if (!test_and_set_bit(WantReplacement, 419 &conf->mirrors[mirror].rdev->flags)) 420 set_bit(MD_RECOVERY_NEEDED, & 421 conf->mddev->recovery); 422 423 set_bit(R1BIO_WriteError, &r1_bio->state); 424 } else { 425 /* 426 * Set R1BIO_Uptodate in our master bio, so that we 427 * will return a good error code for to the higher 428 * levels even if IO on some other mirrored buffer 429 * fails. 430 * 431 * The 'master' represents the composite IO operation 432 * to user-side. So if something waits for IO, then it 433 * will wait for the 'master' bio. 434 */ 435 sector_t first_bad; 436 int bad_sectors; 437 438 r1_bio->bios[mirror] = NULL; 439 to_put = bio; 440 /* 441 * Do not set R1BIO_Uptodate if the current device is 442 * rebuilding or Faulty. This is because we cannot use 443 * such device for properly reading the data back (we could 444 * potentially use it, if the current write would have felt 445 * before rdev->recovery_offset, but for simplicity we don't 446 * check this here. 447 */ 448 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) && 449 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)) 450 set_bit(R1BIO_Uptodate, &r1_bio->state); 451 452 /* Maybe we can clear some bad blocks. */ 453 if (is_badblock(conf->mirrors[mirror].rdev, 454 r1_bio->sector, r1_bio->sectors, 455 &first_bad, &bad_sectors)) { 456 r1_bio->bios[mirror] = IO_MADE_GOOD; 457 set_bit(R1BIO_MadeGood, &r1_bio->state); 458 } 459 } 460 461 if (behind) { 462 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags)) 463 atomic_dec(&r1_bio->behind_remaining); 464 465 /* 466 * In behind mode, we ACK the master bio once the I/O 467 * has safely reached all non-writemostly 468 * disks. Setting the Returned bit ensures that this 469 * gets done only once -- we don't ever want to return 470 * -EIO here, instead we'll wait 471 */ 472 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) && 473 test_bit(R1BIO_Uptodate, &r1_bio->state)) { 474 /* Maybe we can return now */ 475 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { 476 struct bio *mbio = r1_bio->master_bio; 477 pr_debug("raid1: behind end write sectors" 478 " %llu-%llu\n", 479 (unsigned long long) mbio->bi_iter.bi_sector, 480 (unsigned long long) bio_end_sector(mbio) - 1); 481 call_bio_endio(r1_bio); 482 } 483 } 484 } 485 if (r1_bio->bios[mirror] == NULL) 486 rdev_dec_pending(conf->mirrors[mirror].rdev, 487 conf->mddev); 488 489 /* 490 * Let's see if all mirrored write operations have finished 491 * already. 492 */ 493 r1_bio_write_done(r1_bio); 494 495 if (to_put) 496 bio_put(to_put); 497 } 498 499 /* 500 * This routine returns the disk from which the requested read should 501 * be done. There is a per-array 'next expected sequential IO' sector 502 * number - if this matches on the next IO then we use the last disk. 503 * There is also a per-disk 'last know head position' sector that is 504 * maintained from IRQ contexts, both the normal and the resync IO 505 * completion handlers update this position correctly. If there is no 506 * perfect sequential match then we pick the disk whose head is closest. 507 * 508 * If there are 2 mirrors in the same 2 devices, performance degrades 509 * because position is mirror, not device based. 510 * 511 * The rdev for the device selected will have nr_pending incremented. 512 */ 513 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors) 514 { 515 const sector_t this_sector = r1_bio->sector; 516 int sectors; 517 int best_good_sectors; 518 int best_disk, best_dist_disk, best_pending_disk; 519 int has_nonrot_disk; 520 int disk; 521 sector_t best_dist; 522 unsigned int min_pending; 523 struct md_rdev *rdev; 524 int choose_first; 525 int choose_next_idle; 526 527 rcu_read_lock(); 528 /* 529 * Check if we can balance. We can balance on the whole 530 * device if no resync is going on, or below the resync window. 531 * We take the first readable disk when above the resync window. 532 */ 533 retry: 534 sectors = r1_bio->sectors; 535 best_disk = -1; 536 best_dist_disk = -1; 537 best_dist = MaxSector; 538 best_pending_disk = -1; 539 min_pending = UINT_MAX; 540 best_good_sectors = 0; 541 has_nonrot_disk = 0; 542 choose_next_idle = 0; 543 544 if ((conf->mddev->recovery_cp < this_sector + sectors) || 545 (mddev_is_clustered(conf->mddev) && 546 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector, 547 this_sector + sectors))) 548 choose_first = 1; 549 else 550 choose_first = 0; 551 552 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) { 553 sector_t dist; 554 sector_t first_bad; 555 int bad_sectors; 556 unsigned int pending; 557 bool nonrot; 558 559 rdev = rcu_dereference(conf->mirrors[disk].rdev); 560 if (r1_bio->bios[disk] == IO_BLOCKED 561 || rdev == NULL 562 || test_bit(Faulty, &rdev->flags)) 563 continue; 564 if (!test_bit(In_sync, &rdev->flags) && 565 rdev->recovery_offset < this_sector + sectors) 566 continue; 567 if (test_bit(WriteMostly, &rdev->flags)) { 568 /* Don't balance among write-mostly, just 569 * use the first as a last resort */ 570 if (best_dist_disk < 0) { 571 if (is_badblock(rdev, this_sector, sectors, 572 &first_bad, &bad_sectors)) { 573 if (first_bad <= this_sector) 574 /* Cannot use this */ 575 continue; 576 best_good_sectors = first_bad - this_sector; 577 } else 578 best_good_sectors = sectors; 579 best_dist_disk = disk; 580 best_pending_disk = disk; 581 } 582 continue; 583 } 584 /* This is a reasonable device to use. It might 585 * even be best. 586 */ 587 if (is_badblock(rdev, this_sector, sectors, 588 &first_bad, &bad_sectors)) { 589 if (best_dist < MaxSector) 590 /* already have a better device */ 591 continue; 592 if (first_bad <= this_sector) { 593 /* cannot read here. If this is the 'primary' 594 * device, then we must not read beyond 595 * bad_sectors from another device.. 596 */ 597 bad_sectors -= (this_sector - first_bad); 598 if (choose_first && sectors > bad_sectors) 599 sectors = bad_sectors; 600 if (best_good_sectors > sectors) 601 best_good_sectors = sectors; 602 603 } else { 604 sector_t good_sectors = first_bad - this_sector; 605 if (good_sectors > best_good_sectors) { 606 best_good_sectors = good_sectors; 607 best_disk = disk; 608 } 609 if (choose_first) 610 break; 611 } 612 continue; 613 } else 614 best_good_sectors = sectors; 615 616 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev)); 617 has_nonrot_disk |= nonrot; 618 pending = atomic_read(&rdev->nr_pending); 619 dist = abs(this_sector - conf->mirrors[disk].head_position); 620 if (choose_first) { 621 best_disk = disk; 622 break; 623 } 624 /* Don't change to another disk for sequential reads */ 625 if (conf->mirrors[disk].next_seq_sect == this_sector 626 || dist == 0) { 627 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9; 628 struct raid1_info *mirror = &conf->mirrors[disk]; 629 630 best_disk = disk; 631 /* 632 * If buffered sequential IO size exceeds optimal 633 * iosize, check if there is idle disk. If yes, choose 634 * the idle disk. read_balance could already choose an 635 * idle disk before noticing it's a sequential IO in 636 * this disk. This doesn't matter because this disk 637 * will idle, next time it will be utilized after the 638 * first disk has IO size exceeds optimal iosize. In 639 * this way, iosize of the first disk will be optimal 640 * iosize at least. iosize of the second disk might be 641 * small, but not a big deal since when the second disk 642 * starts IO, the first disk is likely still busy. 643 */ 644 if (nonrot && opt_iosize > 0 && 645 mirror->seq_start != MaxSector && 646 mirror->next_seq_sect > opt_iosize && 647 mirror->next_seq_sect - opt_iosize >= 648 mirror->seq_start) { 649 choose_next_idle = 1; 650 continue; 651 } 652 break; 653 } 654 /* If device is idle, use it */ 655 if (pending == 0) { 656 best_disk = disk; 657 break; 658 } 659 660 if (choose_next_idle) 661 continue; 662 663 if (min_pending > pending) { 664 min_pending = pending; 665 best_pending_disk = disk; 666 } 667 668 if (dist < best_dist) { 669 best_dist = dist; 670 best_dist_disk = disk; 671 } 672 } 673 674 /* 675 * If all disks are rotational, choose the closest disk. If any disk is 676 * non-rotational, choose the disk with less pending request even the 677 * disk is rotational, which might/might not be optimal for raids with 678 * mixed ratation/non-rotational disks depending on workload. 679 */ 680 if (best_disk == -1) { 681 if (has_nonrot_disk) 682 best_disk = best_pending_disk; 683 else 684 best_disk = best_dist_disk; 685 } 686 687 if (best_disk >= 0) { 688 rdev = rcu_dereference(conf->mirrors[best_disk].rdev); 689 if (!rdev) 690 goto retry; 691 atomic_inc(&rdev->nr_pending); 692 if (test_bit(Faulty, &rdev->flags)) { 693 /* cannot risk returning a device that failed 694 * before we inc'ed nr_pending 695 */ 696 rdev_dec_pending(rdev, conf->mddev); 697 goto retry; 698 } 699 sectors = best_good_sectors; 700 701 if (conf->mirrors[best_disk].next_seq_sect != this_sector) 702 conf->mirrors[best_disk].seq_start = this_sector; 703 704 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors; 705 } 706 rcu_read_unlock(); 707 *max_sectors = sectors; 708 709 return best_disk; 710 } 711 712 static int raid1_congested(struct mddev *mddev, int bits) 713 { 714 struct r1conf *conf = mddev->private; 715 int i, ret = 0; 716 717 if ((bits & (1 << WB_async_congested)) && 718 conf->pending_count >= max_queued_requests) 719 return 1; 720 721 rcu_read_lock(); 722 for (i = 0; i < conf->raid_disks * 2; i++) { 723 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 724 if (rdev && !test_bit(Faulty, &rdev->flags)) { 725 struct request_queue *q = bdev_get_queue(rdev->bdev); 726 727 BUG_ON(!q); 728 729 /* Note the '|| 1' - when read_balance prefers 730 * non-congested targets, it can be removed 731 */ 732 if ((bits & (1 << WB_async_congested)) || 1) 733 ret |= bdi_congested(&q->backing_dev_info, bits); 734 else 735 ret &= bdi_congested(&q->backing_dev_info, bits); 736 } 737 } 738 rcu_read_unlock(); 739 return ret; 740 } 741 742 static void flush_pending_writes(struct r1conf *conf) 743 { 744 /* Any writes that have been queued but are awaiting 745 * bitmap updates get flushed here. 746 */ 747 spin_lock_irq(&conf->device_lock); 748 749 if (conf->pending_bio_list.head) { 750 struct bio *bio; 751 bio = bio_list_get(&conf->pending_bio_list); 752 conf->pending_count = 0; 753 spin_unlock_irq(&conf->device_lock); 754 /* flush any pending bitmap writes to 755 * disk before proceeding w/ I/O */ 756 bitmap_unplug(conf->mddev->bitmap); 757 wake_up(&conf->wait_barrier); 758 759 while (bio) { /* submit pending writes */ 760 struct bio *next = bio->bi_next; 761 bio->bi_next = NULL; 762 if (unlikely((bio_op(bio) == REQ_OP_DISCARD) && 763 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 764 /* Just ignore it */ 765 bio_endio(bio); 766 else 767 generic_make_request(bio); 768 bio = next; 769 } 770 } else 771 spin_unlock_irq(&conf->device_lock); 772 } 773 774 /* Barriers.... 775 * Sometimes we need to suspend IO while we do something else, 776 * either some resync/recovery, or reconfigure the array. 777 * To do this we raise a 'barrier'. 778 * The 'barrier' is a counter that can be raised multiple times 779 * to count how many activities are happening which preclude 780 * normal IO. 781 * We can only raise the barrier if there is no pending IO. 782 * i.e. if nr_pending == 0. 783 * We choose only to raise the barrier if no-one is waiting for the 784 * barrier to go down. This means that as soon as an IO request 785 * is ready, no other operations which require a barrier will start 786 * until the IO request has had a chance. 787 * 788 * So: regular IO calls 'wait_barrier'. When that returns there 789 * is no backgroup IO happening, It must arrange to call 790 * allow_barrier when it has finished its IO. 791 * backgroup IO calls must call raise_barrier. Once that returns 792 * there is no normal IO happeing. It must arrange to call 793 * lower_barrier when the particular background IO completes. 794 */ 795 static void raise_barrier(struct r1conf *conf, sector_t sector_nr) 796 { 797 spin_lock_irq(&conf->resync_lock); 798 799 /* Wait until no block IO is waiting */ 800 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting, 801 conf->resync_lock); 802 803 /* block any new IO from starting */ 804 conf->barrier++; 805 conf->next_resync = sector_nr; 806 807 /* For these conditions we must wait: 808 * A: while the array is in frozen state 809 * B: while barrier >= RESYNC_DEPTH, meaning resync reach 810 * the max count which allowed. 811 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning 812 * next resync will reach to the window which normal bios are 813 * handling. 814 * D: while there are any active requests in the current window. 815 */ 816 wait_event_lock_irq(conf->wait_barrier, 817 !conf->array_frozen && 818 conf->barrier < RESYNC_DEPTH && 819 conf->current_window_requests == 0 && 820 (conf->start_next_window >= 821 conf->next_resync + RESYNC_SECTORS), 822 conf->resync_lock); 823 824 conf->nr_pending++; 825 spin_unlock_irq(&conf->resync_lock); 826 } 827 828 static void lower_barrier(struct r1conf *conf) 829 { 830 unsigned long flags; 831 BUG_ON(conf->barrier <= 0); 832 spin_lock_irqsave(&conf->resync_lock, flags); 833 conf->barrier--; 834 conf->nr_pending--; 835 spin_unlock_irqrestore(&conf->resync_lock, flags); 836 wake_up(&conf->wait_barrier); 837 } 838 839 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio) 840 { 841 bool wait = false; 842 843 if (conf->array_frozen || !bio) 844 wait = true; 845 else if (conf->barrier && bio_data_dir(bio) == WRITE) { 846 if ((conf->mddev->curr_resync_completed 847 >= bio_end_sector(bio)) || 848 (conf->next_resync + NEXT_NORMALIO_DISTANCE 849 <= bio->bi_iter.bi_sector)) 850 wait = false; 851 else 852 wait = true; 853 } 854 855 return wait; 856 } 857 858 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio) 859 { 860 sector_t sector = 0; 861 862 spin_lock_irq(&conf->resync_lock); 863 if (need_to_wait_for_sync(conf, bio)) { 864 conf->nr_waiting++; 865 /* Wait for the barrier to drop. 866 * However if there are already pending 867 * requests (preventing the barrier from 868 * rising completely), and the 869 * per-process bio queue isn't empty, 870 * then don't wait, as we need to empty 871 * that queue to allow conf->start_next_window 872 * to increase. 873 */ 874 wait_event_lock_irq(conf->wait_barrier, 875 !conf->array_frozen && 876 (!conf->barrier || 877 ((conf->start_next_window < 878 conf->next_resync + RESYNC_SECTORS) && 879 current->bio_list && 880 !bio_list_empty(current->bio_list))), 881 conf->resync_lock); 882 conf->nr_waiting--; 883 } 884 885 if (bio && bio_data_dir(bio) == WRITE) { 886 if (bio->bi_iter.bi_sector >= conf->next_resync) { 887 if (conf->start_next_window == MaxSector) 888 conf->start_next_window = 889 conf->next_resync + 890 NEXT_NORMALIO_DISTANCE; 891 892 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE) 893 <= bio->bi_iter.bi_sector) 894 conf->next_window_requests++; 895 else 896 conf->current_window_requests++; 897 sector = conf->start_next_window; 898 } 899 } 900 901 conf->nr_pending++; 902 spin_unlock_irq(&conf->resync_lock); 903 return sector; 904 } 905 906 static void allow_barrier(struct r1conf *conf, sector_t start_next_window, 907 sector_t bi_sector) 908 { 909 unsigned long flags; 910 911 spin_lock_irqsave(&conf->resync_lock, flags); 912 conf->nr_pending--; 913 if (start_next_window) { 914 if (start_next_window == conf->start_next_window) { 915 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE 916 <= bi_sector) 917 conf->next_window_requests--; 918 else 919 conf->current_window_requests--; 920 } else 921 conf->current_window_requests--; 922 923 if (!conf->current_window_requests) { 924 if (conf->next_window_requests) { 925 conf->current_window_requests = 926 conf->next_window_requests; 927 conf->next_window_requests = 0; 928 conf->start_next_window += 929 NEXT_NORMALIO_DISTANCE; 930 } else 931 conf->start_next_window = MaxSector; 932 } 933 } 934 spin_unlock_irqrestore(&conf->resync_lock, flags); 935 wake_up(&conf->wait_barrier); 936 } 937 938 static void freeze_array(struct r1conf *conf, int extra) 939 { 940 /* stop syncio and normal IO and wait for everything to 941 * go quite. 942 * We wait until nr_pending match nr_queued+extra 943 * This is called in the context of one normal IO request 944 * that has failed. Thus any sync request that might be pending 945 * will be blocked by nr_pending, and we need to wait for 946 * pending IO requests to complete or be queued for re-try. 947 * Thus the number queued (nr_queued) plus this request (extra) 948 * must match the number of pending IOs (nr_pending) before 949 * we continue. 950 */ 951 spin_lock_irq(&conf->resync_lock); 952 conf->array_frozen = 1; 953 wait_event_lock_irq_cmd(conf->wait_barrier, 954 conf->nr_pending == conf->nr_queued+extra, 955 conf->resync_lock, 956 flush_pending_writes(conf)); 957 spin_unlock_irq(&conf->resync_lock); 958 } 959 static void unfreeze_array(struct r1conf *conf) 960 { 961 /* reverse the effect of the freeze */ 962 spin_lock_irq(&conf->resync_lock); 963 conf->array_frozen = 0; 964 wake_up(&conf->wait_barrier); 965 spin_unlock_irq(&conf->resync_lock); 966 } 967 968 /* duplicate the data pages for behind I/O 969 */ 970 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio) 971 { 972 int i; 973 struct bio_vec *bvec; 974 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec), 975 GFP_NOIO); 976 if (unlikely(!bvecs)) 977 return; 978 979 bio_for_each_segment_all(bvec, bio, i) { 980 bvecs[i] = *bvec; 981 bvecs[i].bv_page = alloc_page(GFP_NOIO); 982 if (unlikely(!bvecs[i].bv_page)) 983 goto do_sync_io; 984 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset, 985 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len); 986 kunmap(bvecs[i].bv_page); 987 kunmap(bvec->bv_page); 988 } 989 r1_bio->behind_bvecs = bvecs; 990 r1_bio->behind_page_count = bio->bi_vcnt; 991 set_bit(R1BIO_BehindIO, &r1_bio->state); 992 return; 993 994 do_sync_io: 995 for (i = 0; i < bio->bi_vcnt; i++) 996 if (bvecs[i].bv_page) 997 put_page(bvecs[i].bv_page); 998 kfree(bvecs); 999 pr_debug("%dB behind alloc failed, doing sync I/O\n", 1000 bio->bi_iter.bi_size); 1001 } 1002 1003 struct raid1_plug_cb { 1004 struct blk_plug_cb cb; 1005 struct bio_list pending; 1006 int pending_cnt; 1007 }; 1008 1009 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule) 1010 { 1011 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, 1012 cb); 1013 struct mddev *mddev = plug->cb.data; 1014 struct r1conf *conf = mddev->private; 1015 struct bio *bio; 1016 1017 if (from_schedule || current->bio_list) { 1018 spin_lock_irq(&conf->device_lock); 1019 bio_list_merge(&conf->pending_bio_list, &plug->pending); 1020 conf->pending_count += plug->pending_cnt; 1021 spin_unlock_irq(&conf->device_lock); 1022 wake_up(&conf->wait_barrier); 1023 md_wakeup_thread(mddev->thread); 1024 kfree(plug); 1025 return; 1026 } 1027 1028 /* we aren't scheduling, so we can do the write-out directly. */ 1029 bio = bio_list_get(&plug->pending); 1030 bitmap_unplug(mddev->bitmap); 1031 wake_up(&conf->wait_barrier); 1032 1033 while (bio) { /* submit pending writes */ 1034 struct bio *next = bio->bi_next; 1035 bio->bi_next = NULL; 1036 if (unlikely((bio_op(bio) == REQ_OP_DISCARD) && 1037 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 1038 /* Just ignore it */ 1039 bio_endio(bio); 1040 else 1041 generic_make_request(bio); 1042 bio = next; 1043 } 1044 kfree(plug); 1045 } 1046 1047 static void raid1_make_request(struct mddev *mddev, struct bio * bio) 1048 { 1049 struct r1conf *conf = mddev->private; 1050 struct raid1_info *mirror; 1051 struct r1bio *r1_bio; 1052 struct bio *read_bio; 1053 int i, disks; 1054 struct bitmap *bitmap; 1055 unsigned long flags; 1056 const int op = bio_op(bio); 1057 const int rw = bio_data_dir(bio); 1058 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC); 1059 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA)); 1060 const unsigned long do_sec = (bio->bi_rw & REQ_SECURE); 1061 struct md_rdev *blocked_rdev; 1062 struct blk_plug_cb *cb; 1063 struct raid1_plug_cb *plug = NULL; 1064 int first_clone; 1065 int sectors_handled; 1066 int max_sectors; 1067 sector_t start_next_window; 1068 1069 /* 1070 * Register the new request and wait if the reconstruction 1071 * thread has put up a bar for new requests. 1072 * Continue immediately if no resync is active currently. 1073 */ 1074 1075 md_write_start(mddev, bio); /* wait on superblock update early */ 1076 1077 if (bio_data_dir(bio) == WRITE && 1078 ((bio_end_sector(bio) > mddev->suspend_lo && 1079 bio->bi_iter.bi_sector < mddev->suspend_hi) || 1080 (mddev_is_clustered(mddev) && 1081 md_cluster_ops->area_resyncing(mddev, WRITE, 1082 bio->bi_iter.bi_sector, bio_end_sector(bio))))) { 1083 /* As the suspend_* range is controlled by 1084 * userspace, we want an interruptible 1085 * wait. 1086 */ 1087 DEFINE_WAIT(w); 1088 for (;;) { 1089 flush_signals(current); 1090 prepare_to_wait(&conf->wait_barrier, 1091 &w, TASK_INTERRUPTIBLE); 1092 if (bio_end_sector(bio) <= mddev->suspend_lo || 1093 bio->bi_iter.bi_sector >= mddev->suspend_hi || 1094 (mddev_is_clustered(mddev) && 1095 !md_cluster_ops->area_resyncing(mddev, WRITE, 1096 bio->bi_iter.bi_sector, bio_end_sector(bio)))) 1097 break; 1098 schedule(); 1099 } 1100 finish_wait(&conf->wait_barrier, &w); 1101 } 1102 1103 start_next_window = wait_barrier(conf, bio); 1104 1105 bitmap = mddev->bitmap; 1106 1107 /* 1108 * make_request() can abort the operation when READA is being 1109 * used and no empty request is available. 1110 * 1111 */ 1112 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1113 1114 r1_bio->master_bio = bio; 1115 r1_bio->sectors = bio_sectors(bio); 1116 r1_bio->state = 0; 1117 r1_bio->mddev = mddev; 1118 r1_bio->sector = bio->bi_iter.bi_sector; 1119 1120 /* We might need to issue multiple reads to different 1121 * devices if there are bad blocks around, so we keep 1122 * track of the number of reads in bio->bi_phys_segments. 1123 * If this is 0, there is only one r1_bio and no locking 1124 * will be needed when requests complete. If it is 1125 * non-zero, then it is the number of not-completed requests. 1126 */ 1127 bio->bi_phys_segments = 0; 1128 bio_clear_flag(bio, BIO_SEG_VALID); 1129 1130 if (rw == READ) { 1131 /* 1132 * read balancing logic: 1133 */ 1134 int rdisk; 1135 1136 read_again: 1137 rdisk = read_balance(conf, r1_bio, &max_sectors); 1138 1139 if (rdisk < 0) { 1140 /* couldn't find anywhere to read from */ 1141 raid_end_bio_io(r1_bio); 1142 return; 1143 } 1144 mirror = conf->mirrors + rdisk; 1145 1146 if (test_bit(WriteMostly, &mirror->rdev->flags) && 1147 bitmap) { 1148 /* Reading from a write-mostly device must 1149 * take care not to over-take any writes 1150 * that are 'behind' 1151 */ 1152 wait_event(bitmap->behind_wait, 1153 atomic_read(&bitmap->behind_writes) == 0); 1154 } 1155 r1_bio->read_disk = rdisk; 1156 r1_bio->start_next_window = 0; 1157 1158 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1159 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector, 1160 max_sectors); 1161 1162 r1_bio->bios[rdisk] = read_bio; 1163 1164 read_bio->bi_iter.bi_sector = r1_bio->sector + 1165 mirror->rdev->data_offset; 1166 read_bio->bi_bdev = mirror->rdev->bdev; 1167 read_bio->bi_end_io = raid1_end_read_request; 1168 bio_set_op_attrs(read_bio, op, do_sync); 1169 read_bio->bi_private = r1_bio; 1170 1171 if (max_sectors < r1_bio->sectors) { 1172 /* could not read all from this device, so we will 1173 * need another r1_bio. 1174 */ 1175 1176 sectors_handled = (r1_bio->sector + max_sectors 1177 - bio->bi_iter.bi_sector); 1178 r1_bio->sectors = max_sectors; 1179 spin_lock_irq(&conf->device_lock); 1180 if (bio->bi_phys_segments == 0) 1181 bio->bi_phys_segments = 2; 1182 else 1183 bio->bi_phys_segments++; 1184 spin_unlock_irq(&conf->device_lock); 1185 /* Cannot call generic_make_request directly 1186 * as that will be queued in __make_request 1187 * and subsequent mempool_alloc might block waiting 1188 * for it. So hand bio over to raid1d. 1189 */ 1190 reschedule_retry(r1_bio); 1191 1192 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1193 1194 r1_bio->master_bio = bio; 1195 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1196 r1_bio->state = 0; 1197 r1_bio->mddev = mddev; 1198 r1_bio->sector = bio->bi_iter.bi_sector + 1199 sectors_handled; 1200 goto read_again; 1201 } else 1202 generic_make_request(read_bio); 1203 return; 1204 } 1205 1206 /* 1207 * WRITE: 1208 */ 1209 if (conf->pending_count >= max_queued_requests) { 1210 md_wakeup_thread(mddev->thread); 1211 wait_event(conf->wait_barrier, 1212 conf->pending_count < max_queued_requests); 1213 } 1214 /* first select target devices under rcu_lock and 1215 * inc refcount on their rdev. Record them by setting 1216 * bios[x] to bio 1217 * If there are known/acknowledged bad blocks on any device on 1218 * which we have seen a write error, we want to avoid writing those 1219 * blocks. 1220 * This potentially requires several writes to write around 1221 * the bad blocks. Each set of writes gets it's own r1bio 1222 * with a set of bios attached. 1223 */ 1224 1225 disks = conf->raid_disks * 2; 1226 retry_write: 1227 r1_bio->start_next_window = start_next_window; 1228 blocked_rdev = NULL; 1229 rcu_read_lock(); 1230 max_sectors = r1_bio->sectors; 1231 for (i = 0; i < disks; i++) { 1232 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1233 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 1234 atomic_inc(&rdev->nr_pending); 1235 blocked_rdev = rdev; 1236 break; 1237 } 1238 r1_bio->bios[i] = NULL; 1239 if (!rdev || test_bit(Faulty, &rdev->flags)) { 1240 if (i < conf->raid_disks) 1241 set_bit(R1BIO_Degraded, &r1_bio->state); 1242 continue; 1243 } 1244 1245 atomic_inc(&rdev->nr_pending); 1246 if (test_bit(WriteErrorSeen, &rdev->flags)) { 1247 sector_t first_bad; 1248 int bad_sectors; 1249 int is_bad; 1250 1251 is_bad = is_badblock(rdev, r1_bio->sector, 1252 max_sectors, 1253 &first_bad, &bad_sectors); 1254 if (is_bad < 0) { 1255 /* mustn't write here until the bad block is 1256 * acknowledged*/ 1257 set_bit(BlockedBadBlocks, &rdev->flags); 1258 blocked_rdev = rdev; 1259 break; 1260 } 1261 if (is_bad && first_bad <= r1_bio->sector) { 1262 /* Cannot write here at all */ 1263 bad_sectors -= (r1_bio->sector - first_bad); 1264 if (bad_sectors < max_sectors) 1265 /* mustn't write more than bad_sectors 1266 * to other devices yet 1267 */ 1268 max_sectors = bad_sectors; 1269 rdev_dec_pending(rdev, mddev); 1270 /* We don't set R1BIO_Degraded as that 1271 * only applies if the disk is 1272 * missing, so it might be re-added, 1273 * and we want to know to recover this 1274 * chunk. 1275 * In this case the device is here, 1276 * and the fact that this chunk is not 1277 * in-sync is recorded in the bad 1278 * block log 1279 */ 1280 continue; 1281 } 1282 if (is_bad) { 1283 int good_sectors = first_bad - r1_bio->sector; 1284 if (good_sectors < max_sectors) 1285 max_sectors = good_sectors; 1286 } 1287 } 1288 r1_bio->bios[i] = bio; 1289 } 1290 rcu_read_unlock(); 1291 1292 if (unlikely(blocked_rdev)) { 1293 /* Wait for this device to become unblocked */ 1294 int j; 1295 sector_t old = start_next_window; 1296 1297 for (j = 0; j < i; j++) 1298 if (r1_bio->bios[j]) 1299 rdev_dec_pending(conf->mirrors[j].rdev, mddev); 1300 r1_bio->state = 0; 1301 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector); 1302 md_wait_for_blocked_rdev(blocked_rdev, mddev); 1303 start_next_window = wait_barrier(conf, bio); 1304 /* 1305 * We must make sure the multi r1bios of bio have 1306 * the same value of bi_phys_segments 1307 */ 1308 if (bio->bi_phys_segments && old && 1309 old != start_next_window) 1310 /* Wait for the former r1bio(s) to complete */ 1311 wait_event(conf->wait_barrier, 1312 bio->bi_phys_segments == 1); 1313 goto retry_write; 1314 } 1315 1316 if (max_sectors < r1_bio->sectors) { 1317 /* We are splitting this write into multiple parts, so 1318 * we need to prepare for allocating another r1_bio. 1319 */ 1320 r1_bio->sectors = max_sectors; 1321 spin_lock_irq(&conf->device_lock); 1322 if (bio->bi_phys_segments == 0) 1323 bio->bi_phys_segments = 2; 1324 else 1325 bio->bi_phys_segments++; 1326 spin_unlock_irq(&conf->device_lock); 1327 } 1328 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector; 1329 1330 atomic_set(&r1_bio->remaining, 1); 1331 atomic_set(&r1_bio->behind_remaining, 0); 1332 1333 first_clone = 1; 1334 for (i = 0; i < disks; i++) { 1335 struct bio *mbio; 1336 if (!r1_bio->bios[i]) 1337 continue; 1338 1339 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1340 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors); 1341 1342 if (first_clone) { 1343 /* do behind I/O ? 1344 * Not if there are too many, or cannot 1345 * allocate memory, or a reader on WriteMostly 1346 * is waiting for behind writes to flush */ 1347 if (bitmap && 1348 (atomic_read(&bitmap->behind_writes) 1349 < mddev->bitmap_info.max_write_behind) && 1350 !waitqueue_active(&bitmap->behind_wait)) 1351 alloc_behind_pages(mbio, r1_bio); 1352 1353 bitmap_startwrite(bitmap, r1_bio->sector, 1354 r1_bio->sectors, 1355 test_bit(R1BIO_BehindIO, 1356 &r1_bio->state)); 1357 first_clone = 0; 1358 } 1359 if (r1_bio->behind_bvecs) { 1360 struct bio_vec *bvec; 1361 int j; 1362 1363 /* 1364 * We trimmed the bio, so _all is legit 1365 */ 1366 bio_for_each_segment_all(bvec, mbio, j) 1367 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page; 1368 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags)) 1369 atomic_inc(&r1_bio->behind_remaining); 1370 } 1371 1372 r1_bio->bios[i] = mbio; 1373 1374 mbio->bi_iter.bi_sector = (r1_bio->sector + 1375 conf->mirrors[i].rdev->data_offset); 1376 mbio->bi_bdev = conf->mirrors[i].rdev->bdev; 1377 mbio->bi_end_io = raid1_end_write_request; 1378 bio_set_op_attrs(mbio, op, do_flush_fua | do_sync | do_sec); 1379 mbio->bi_private = r1_bio; 1380 1381 atomic_inc(&r1_bio->remaining); 1382 1383 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug)); 1384 if (cb) 1385 plug = container_of(cb, struct raid1_plug_cb, cb); 1386 else 1387 plug = NULL; 1388 spin_lock_irqsave(&conf->device_lock, flags); 1389 if (plug) { 1390 bio_list_add(&plug->pending, mbio); 1391 plug->pending_cnt++; 1392 } else { 1393 bio_list_add(&conf->pending_bio_list, mbio); 1394 conf->pending_count++; 1395 } 1396 spin_unlock_irqrestore(&conf->device_lock, flags); 1397 if (!plug) 1398 md_wakeup_thread(mddev->thread); 1399 } 1400 /* Mustn't call r1_bio_write_done before this next test, 1401 * as it could result in the bio being freed. 1402 */ 1403 if (sectors_handled < bio_sectors(bio)) { 1404 r1_bio_write_done(r1_bio); 1405 /* We need another r1_bio. It has already been counted 1406 * in bio->bi_phys_segments 1407 */ 1408 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1409 r1_bio->master_bio = bio; 1410 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1411 r1_bio->state = 0; 1412 r1_bio->mddev = mddev; 1413 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled; 1414 goto retry_write; 1415 } 1416 1417 r1_bio_write_done(r1_bio); 1418 1419 /* In case raid1d snuck in to freeze_array */ 1420 wake_up(&conf->wait_barrier); 1421 } 1422 1423 static void raid1_status(struct seq_file *seq, struct mddev *mddev) 1424 { 1425 struct r1conf *conf = mddev->private; 1426 int i; 1427 1428 seq_printf(seq, " [%d/%d] [", conf->raid_disks, 1429 conf->raid_disks - mddev->degraded); 1430 rcu_read_lock(); 1431 for (i = 0; i < conf->raid_disks; i++) { 1432 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1433 seq_printf(seq, "%s", 1434 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_"); 1435 } 1436 rcu_read_unlock(); 1437 seq_printf(seq, "]"); 1438 } 1439 1440 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev) 1441 { 1442 char b[BDEVNAME_SIZE]; 1443 struct r1conf *conf = mddev->private; 1444 unsigned long flags; 1445 1446 /* 1447 * If it is not operational, then we have already marked it as dead 1448 * else if it is the last working disks, ignore the error, let the 1449 * next level up know. 1450 * else mark the drive as failed 1451 */ 1452 if (test_bit(In_sync, &rdev->flags) 1453 && (conf->raid_disks - mddev->degraded) == 1) { 1454 /* 1455 * Don't fail the drive, act as though we were just a 1456 * normal single drive. 1457 * However don't try a recovery from this drive as 1458 * it is very likely to fail. 1459 */ 1460 conf->recovery_disabled = mddev->recovery_disabled; 1461 return; 1462 } 1463 set_bit(Blocked, &rdev->flags); 1464 spin_lock_irqsave(&conf->device_lock, flags); 1465 if (test_and_clear_bit(In_sync, &rdev->flags)) { 1466 mddev->degraded++; 1467 set_bit(Faulty, &rdev->flags); 1468 } else 1469 set_bit(Faulty, &rdev->flags); 1470 spin_unlock_irqrestore(&conf->device_lock, flags); 1471 /* 1472 * if recovery is running, make sure it aborts. 1473 */ 1474 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1475 set_mask_bits(&mddev->flags, 0, 1476 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING)); 1477 printk(KERN_ALERT 1478 "md/raid1:%s: Disk failure on %s, disabling device.\n" 1479 "md/raid1:%s: Operation continuing on %d devices.\n", 1480 mdname(mddev), bdevname(rdev->bdev, b), 1481 mdname(mddev), conf->raid_disks - mddev->degraded); 1482 } 1483 1484 static void print_conf(struct r1conf *conf) 1485 { 1486 int i; 1487 1488 printk(KERN_DEBUG "RAID1 conf printout:\n"); 1489 if (!conf) { 1490 printk(KERN_DEBUG "(!conf)\n"); 1491 return; 1492 } 1493 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded, 1494 conf->raid_disks); 1495 1496 rcu_read_lock(); 1497 for (i = 0; i < conf->raid_disks; i++) { 1498 char b[BDEVNAME_SIZE]; 1499 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1500 if (rdev) 1501 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n", 1502 i, !test_bit(In_sync, &rdev->flags), 1503 !test_bit(Faulty, &rdev->flags), 1504 bdevname(rdev->bdev,b)); 1505 } 1506 rcu_read_unlock(); 1507 } 1508 1509 static void close_sync(struct r1conf *conf) 1510 { 1511 wait_barrier(conf, NULL); 1512 allow_barrier(conf, 0, 0); 1513 1514 mempool_destroy(conf->r1buf_pool); 1515 conf->r1buf_pool = NULL; 1516 1517 spin_lock_irq(&conf->resync_lock); 1518 conf->next_resync = MaxSector - 2 * NEXT_NORMALIO_DISTANCE; 1519 conf->start_next_window = MaxSector; 1520 conf->current_window_requests += 1521 conf->next_window_requests; 1522 conf->next_window_requests = 0; 1523 spin_unlock_irq(&conf->resync_lock); 1524 } 1525 1526 static int raid1_spare_active(struct mddev *mddev) 1527 { 1528 int i; 1529 struct r1conf *conf = mddev->private; 1530 int count = 0; 1531 unsigned long flags; 1532 1533 /* 1534 * Find all failed disks within the RAID1 configuration 1535 * and mark them readable. 1536 * Called under mddev lock, so rcu protection not needed. 1537 * device_lock used to avoid races with raid1_end_read_request 1538 * which expects 'In_sync' flags and ->degraded to be consistent. 1539 */ 1540 spin_lock_irqsave(&conf->device_lock, flags); 1541 for (i = 0; i < conf->raid_disks; i++) { 1542 struct md_rdev *rdev = conf->mirrors[i].rdev; 1543 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev; 1544 if (repl 1545 && !test_bit(Candidate, &repl->flags) 1546 && repl->recovery_offset == MaxSector 1547 && !test_bit(Faulty, &repl->flags) 1548 && !test_and_set_bit(In_sync, &repl->flags)) { 1549 /* replacement has just become active */ 1550 if (!rdev || 1551 !test_and_clear_bit(In_sync, &rdev->flags)) 1552 count++; 1553 if (rdev) { 1554 /* Replaced device not technically 1555 * faulty, but we need to be sure 1556 * it gets removed and never re-added 1557 */ 1558 set_bit(Faulty, &rdev->flags); 1559 sysfs_notify_dirent_safe( 1560 rdev->sysfs_state); 1561 } 1562 } 1563 if (rdev 1564 && rdev->recovery_offset == MaxSector 1565 && !test_bit(Faulty, &rdev->flags) 1566 && !test_and_set_bit(In_sync, &rdev->flags)) { 1567 count++; 1568 sysfs_notify_dirent_safe(rdev->sysfs_state); 1569 } 1570 } 1571 mddev->degraded -= count; 1572 spin_unlock_irqrestore(&conf->device_lock, flags); 1573 1574 print_conf(conf); 1575 return count; 1576 } 1577 1578 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev) 1579 { 1580 struct r1conf *conf = mddev->private; 1581 int err = -EEXIST; 1582 int mirror = 0; 1583 struct raid1_info *p; 1584 int first = 0; 1585 int last = conf->raid_disks - 1; 1586 1587 if (mddev->recovery_disabled == conf->recovery_disabled) 1588 return -EBUSY; 1589 1590 if (md_integrity_add_rdev(rdev, mddev)) 1591 return -ENXIO; 1592 1593 if (rdev->raid_disk >= 0) 1594 first = last = rdev->raid_disk; 1595 1596 /* 1597 * find the disk ... but prefer rdev->saved_raid_disk 1598 * if possible. 1599 */ 1600 if (rdev->saved_raid_disk >= 0 && 1601 rdev->saved_raid_disk >= first && 1602 conf->mirrors[rdev->saved_raid_disk].rdev == NULL) 1603 first = last = rdev->saved_raid_disk; 1604 1605 for (mirror = first; mirror <= last; mirror++) { 1606 p = conf->mirrors+mirror; 1607 if (!p->rdev) { 1608 1609 if (mddev->gendisk) 1610 disk_stack_limits(mddev->gendisk, rdev->bdev, 1611 rdev->data_offset << 9); 1612 1613 p->head_position = 0; 1614 rdev->raid_disk = mirror; 1615 err = 0; 1616 /* As all devices are equivalent, we don't need a full recovery 1617 * if this was recently any drive of the array 1618 */ 1619 if (rdev->saved_raid_disk < 0) 1620 conf->fullsync = 1; 1621 rcu_assign_pointer(p->rdev, rdev); 1622 break; 1623 } 1624 if (test_bit(WantReplacement, &p->rdev->flags) && 1625 p[conf->raid_disks].rdev == NULL) { 1626 /* Add this device as a replacement */ 1627 clear_bit(In_sync, &rdev->flags); 1628 set_bit(Replacement, &rdev->flags); 1629 rdev->raid_disk = mirror; 1630 err = 0; 1631 conf->fullsync = 1; 1632 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev); 1633 break; 1634 } 1635 } 1636 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev))) 1637 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue); 1638 print_conf(conf); 1639 return err; 1640 } 1641 1642 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev) 1643 { 1644 struct r1conf *conf = mddev->private; 1645 int err = 0; 1646 int number = rdev->raid_disk; 1647 struct raid1_info *p = conf->mirrors + number; 1648 1649 if (rdev != p->rdev) 1650 p = conf->mirrors + conf->raid_disks + number; 1651 1652 print_conf(conf); 1653 if (rdev == p->rdev) { 1654 if (test_bit(In_sync, &rdev->flags) || 1655 atomic_read(&rdev->nr_pending)) { 1656 err = -EBUSY; 1657 goto abort; 1658 } 1659 /* Only remove non-faulty devices if recovery 1660 * is not possible. 1661 */ 1662 if (!test_bit(Faulty, &rdev->flags) && 1663 mddev->recovery_disabled != conf->recovery_disabled && 1664 mddev->degraded < conf->raid_disks) { 1665 err = -EBUSY; 1666 goto abort; 1667 } 1668 p->rdev = NULL; 1669 synchronize_rcu(); 1670 if (atomic_read(&rdev->nr_pending)) { 1671 /* lost the race, try later */ 1672 err = -EBUSY; 1673 p->rdev = rdev; 1674 goto abort; 1675 } else if (conf->mirrors[conf->raid_disks + number].rdev) { 1676 /* We just removed a device that is being replaced. 1677 * Move down the replacement. We drain all IO before 1678 * doing this to avoid confusion. 1679 */ 1680 struct md_rdev *repl = 1681 conf->mirrors[conf->raid_disks + number].rdev; 1682 freeze_array(conf, 0); 1683 clear_bit(Replacement, &repl->flags); 1684 p->rdev = repl; 1685 conf->mirrors[conf->raid_disks + number].rdev = NULL; 1686 unfreeze_array(conf); 1687 clear_bit(WantReplacement, &rdev->flags); 1688 } else 1689 clear_bit(WantReplacement, &rdev->flags); 1690 err = md_integrity_register(mddev); 1691 } 1692 abort: 1693 1694 print_conf(conf); 1695 return err; 1696 } 1697 1698 static void end_sync_read(struct bio *bio) 1699 { 1700 struct r1bio *r1_bio = bio->bi_private; 1701 1702 update_head_pos(r1_bio->read_disk, r1_bio); 1703 1704 /* 1705 * we have read a block, now it needs to be re-written, 1706 * or re-read if the read failed. 1707 * We don't do much here, just schedule handling by raid1d 1708 */ 1709 if (!bio->bi_error) 1710 set_bit(R1BIO_Uptodate, &r1_bio->state); 1711 1712 if (atomic_dec_and_test(&r1_bio->remaining)) 1713 reschedule_retry(r1_bio); 1714 } 1715 1716 static void end_sync_write(struct bio *bio) 1717 { 1718 int uptodate = !bio->bi_error; 1719 struct r1bio *r1_bio = bio->bi_private; 1720 struct mddev *mddev = r1_bio->mddev; 1721 struct r1conf *conf = mddev->private; 1722 int mirror=0; 1723 sector_t first_bad; 1724 int bad_sectors; 1725 1726 mirror = find_bio_disk(r1_bio, bio); 1727 1728 if (!uptodate) { 1729 sector_t sync_blocks = 0; 1730 sector_t s = r1_bio->sector; 1731 long sectors_to_go = r1_bio->sectors; 1732 /* make sure these bits doesn't get cleared. */ 1733 do { 1734 bitmap_end_sync(mddev->bitmap, s, 1735 &sync_blocks, 1); 1736 s += sync_blocks; 1737 sectors_to_go -= sync_blocks; 1738 } while (sectors_to_go > 0); 1739 set_bit(WriteErrorSeen, 1740 &conf->mirrors[mirror].rdev->flags); 1741 if (!test_and_set_bit(WantReplacement, 1742 &conf->mirrors[mirror].rdev->flags)) 1743 set_bit(MD_RECOVERY_NEEDED, & 1744 mddev->recovery); 1745 set_bit(R1BIO_WriteError, &r1_bio->state); 1746 } else if (is_badblock(conf->mirrors[mirror].rdev, 1747 r1_bio->sector, 1748 r1_bio->sectors, 1749 &first_bad, &bad_sectors) && 1750 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev, 1751 r1_bio->sector, 1752 r1_bio->sectors, 1753 &first_bad, &bad_sectors) 1754 ) 1755 set_bit(R1BIO_MadeGood, &r1_bio->state); 1756 1757 if (atomic_dec_and_test(&r1_bio->remaining)) { 1758 int s = r1_bio->sectors; 1759 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 1760 test_bit(R1BIO_WriteError, &r1_bio->state)) 1761 reschedule_retry(r1_bio); 1762 else { 1763 put_buf(r1_bio); 1764 md_done_sync(mddev, s, uptodate); 1765 } 1766 } 1767 } 1768 1769 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector, 1770 int sectors, struct page *page, int rw) 1771 { 1772 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false)) 1773 /* success */ 1774 return 1; 1775 if (rw == WRITE) { 1776 set_bit(WriteErrorSeen, &rdev->flags); 1777 if (!test_and_set_bit(WantReplacement, 1778 &rdev->flags)) 1779 set_bit(MD_RECOVERY_NEEDED, & 1780 rdev->mddev->recovery); 1781 } 1782 /* need to record an error - either for the block or the device */ 1783 if (!rdev_set_badblocks(rdev, sector, sectors, 0)) 1784 md_error(rdev->mddev, rdev); 1785 return 0; 1786 } 1787 1788 static int fix_sync_read_error(struct r1bio *r1_bio) 1789 { 1790 /* Try some synchronous reads of other devices to get 1791 * good data, much like with normal read errors. Only 1792 * read into the pages we already have so we don't 1793 * need to re-issue the read request. 1794 * We don't need to freeze the array, because being in an 1795 * active sync request, there is no normal IO, and 1796 * no overlapping syncs. 1797 * We don't need to check is_badblock() again as we 1798 * made sure that anything with a bad block in range 1799 * will have bi_end_io clear. 1800 */ 1801 struct mddev *mddev = r1_bio->mddev; 1802 struct r1conf *conf = mddev->private; 1803 struct bio *bio = r1_bio->bios[r1_bio->read_disk]; 1804 sector_t sect = r1_bio->sector; 1805 int sectors = r1_bio->sectors; 1806 int idx = 0; 1807 1808 while(sectors) { 1809 int s = sectors; 1810 int d = r1_bio->read_disk; 1811 int success = 0; 1812 struct md_rdev *rdev; 1813 int start; 1814 1815 if (s > (PAGE_SIZE>>9)) 1816 s = PAGE_SIZE >> 9; 1817 do { 1818 if (r1_bio->bios[d]->bi_end_io == end_sync_read) { 1819 /* No rcu protection needed here devices 1820 * can only be removed when no resync is 1821 * active, and resync is currently active 1822 */ 1823 rdev = conf->mirrors[d].rdev; 1824 if (sync_page_io(rdev, sect, s<<9, 1825 bio->bi_io_vec[idx].bv_page, 1826 REQ_OP_READ, 0, false)) { 1827 success = 1; 1828 break; 1829 } 1830 } 1831 d++; 1832 if (d == conf->raid_disks * 2) 1833 d = 0; 1834 } while (!success && d != r1_bio->read_disk); 1835 1836 if (!success) { 1837 char b[BDEVNAME_SIZE]; 1838 int abort = 0; 1839 /* Cannot read from anywhere, this block is lost. 1840 * Record a bad block on each device. If that doesn't 1841 * work just disable and interrupt the recovery. 1842 * Don't fail devices as that won't really help. 1843 */ 1844 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error" 1845 " for block %llu\n", 1846 mdname(mddev), 1847 bdevname(bio->bi_bdev, b), 1848 (unsigned long long)r1_bio->sector); 1849 for (d = 0; d < conf->raid_disks * 2; d++) { 1850 rdev = conf->mirrors[d].rdev; 1851 if (!rdev || test_bit(Faulty, &rdev->flags)) 1852 continue; 1853 if (!rdev_set_badblocks(rdev, sect, s, 0)) 1854 abort = 1; 1855 } 1856 if (abort) { 1857 conf->recovery_disabled = 1858 mddev->recovery_disabled; 1859 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1860 md_done_sync(mddev, r1_bio->sectors, 0); 1861 put_buf(r1_bio); 1862 return 0; 1863 } 1864 /* Try next page */ 1865 sectors -= s; 1866 sect += s; 1867 idx++; 1868 continue; 1869 } 1870 1871 start = d; 1872 /* write it back and re-read */ 1873 while (d != r1_bio->read_disk) { 1874 if (d == 0) 1875 d = conf->raid_disks * 2; 1876 d--; 1877 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1878 continue; 1879 rdev = conf->mirrors[d].rdev; 1880 if (r1_sync_page_io(rdev, sect, s, 1881 bio->bi_io_vec[idx].bv_page, 1882 WRITE) == 0) { 1883 r1_bio->bios[d]->bi_end_io = NULL; 1884 rdev_dec_pending(rdev, mddev); 1885 } 1886 } 1887 d = start; 1888 while (d != r1_bio->read_disk) { 1889 if (d == 0) 1890 d = conf->raid_disks * 2; 1891 d--; 1892 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1893 continue; 1894 rdev = conf->mirrors[d].rdev; 1895 if (r1_sync_page_io(rdev, sect, s, 1896 bio->bi_io_vec[idx].bv_page, 1897 READ) != 0) 1898 atomic_add(s, &rdev->corrected_errors); 1899 } 1900 sectors -= s; 1901 sect += s; 1902 idx ++; 1903 } 1904 set_bit(R1BIO_Uptodate, &r1_bio->state); 1905 bio->bi_error = 0; 1906 return 1; 1907 } 1908 1909 static void process_checks(struct r1bio *r1_bio) 1910 { 1911 /* We have read all readable devices. If we haven't 1912 * got the block, then there is no hope left. 1913 * If we have, then we want to do a comparison 1914 * and skip the write if everything is the same. 1915 * If any blocks failed to read, then we need to 1916 * attempt an over-write 1917 */ 1918 struct mddev *mddev = r1_bio->mddev; 1919 struct r1conf *conf = mddev->private; 1920 int primary; 1921 int i; 1922 int vcnt; 1923 1924 /* Fix variable parts of all bios */ 1925 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9); 1926 for (i = 0; i < conf->raid_disks * 2; i++) { 1927 int j; 1928 int size; 1929 int error; 1930 struct bio *b = r1_bio->bios[i]; 1931 if (b->bi_end_io != end_sync_read) 1932 continue; 1933 /* fixup the bio for reuse, but preserve errno */ 1934 error = b->bi_error; 1935 bio_reset(b); 1936 b->bi_error = error; 1937 b->bi_vcnt = vcnt; 1938 b->bi_iter.bi_size = r1_bio->sectors << 9; 1939 b->bi_iter.bi_sector = r1_bio->sector + 1940 conf->mirrors[i].rdev->data_offset; 1941 b->bi_bdev = conf->mirrors[i].rdev->bdev; 1942 b->bi_end_io = end_sync_read; 1943 b->bi_private = r1_bio; 1944 1945 size = b->bi_iter.bi_size; 1946 for (j = 0; j < vcnt ; j++) { 1947 struct bio_vec *bi; 1948 bi = &b->bi_io_vec[j]; 1949 bi->bv_offset = 0; 1950 if (size > PAGE_SIZE) 1951 bi->bv_len = PAGE_SIZE; 1952 else 1953 bi->bv_len = size; 1954 size -= PAGE_SIZE; 1955 } 1956 } 1957 for (primary = 0; primary < conf->raid_disks * 2; primary++) 1958 if (r1_bio->bios[primary]->bi_end_io == end_sync_read && 1959 !r1_bio->bios[primary]->bi_error) { 1960 r1_bio->bios[primary]->bi_end_io = NULL; 1961 rdev_dec_pending(conf->mirrors[primary].rdev, mddev); 1962 break; 1963 } 1964 r1_bio->read_disk = primary; 1965 for (i = 0; i < conf->raid_disks * 2; i++) { 1966 int j; 1967 struct bio *pbio = r1_bio->bios[primary]; 1968 struct bio *sbio = r1_bio->bios[i]; 1969 int error = sbio->bi_error; 1970 1971 if (sbio->bi_end_io != end_sync_read) 1972 continue; 1973 /* Now we can 'fixup' the error value */ 1974 sbio->bi_error = 0; 1975 1976 if (!error) { 1977 for (j = vcnt; j-- ; ) { 1978 struct page *p, *s; 1979 p = pbio->bi_io_vec[j].bv_page; 1980 s = sbio->bi_io_vec[j].bv_page; 1981 if (memcmp(page_address(p), 1982 page_address(s), 1983 sbio->bi_io_vec[j].bv_len)) 1984 break; 1985 } 1986 } else 1987 j = 0; 1988 if (j >= 0) 1989 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches); 1990 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery) 1991 && !error)) { 1992 /* No need to write to this device. */ 1993 sbio->bi_end_io = NULL; 1994 rdev_dec_pending(conf->mirrors[i].rdev, mddev); 1995 continue; 1996 } 1997 1998 bio_copy_data(sbio, pbio); 1999 } 2000 } 2001 2002 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio) 2003 { 2004 struct r1conf *conf = mddev->private; 2005 int i; 2006 int disks = conf->raid_disks * 2; 2007 struct bio *bio, *wbio; 2008 2009 bio = r1_bio->bios[r1_bio->read_disk]; 2010 2011 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 2012 /* ouch - failed to read all of that. */ 2013 if (!fix_sync_read_error(r1_bio)) 2014 return; 2015 2016 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2017 process_checks(r1_bio); 2018 2019 /* 2020 * schedule writes 2021 */ 2022 atomic_set(&r1_bio->remaining, 1); 2023 for (i = 0; i < disks ; i++) { 2024 wbio = r1_bio->bios[i]; 2025 if (wbio->bi_end_io == NULL || 2026 (wbio->bi_end_io == end_sync_read && 2027 (i == r1_bio->read_disk || 2028 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)))) 2029 continue; 2030 2031 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0); 2032 wbio->bi_end_io = end_sync_write; 2033 atomic_inc(&r1_bio->remaining); 2034 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio)); 2035 2036 generic_make_request(wbio); 2037 } 2038 2039 if (atomic_dec_and_test(&r1_bio->remaining)) { 2040 /* if we're here, all write(s) have completed, so clean up */ 2041 int s = r1_bio->sectors; 2042 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2043 test_bit(R1BIO_WriteError, &r1_bio->state)) 2044 reschedule_retry(r1_bio); 2045 else { 2046 put_buf(r1_bio); 2047 md_done_sync(mddev, s, 1); 2048 } 2049 } 2050 } 2051 2052 /* 2053 * This is a kernel thread which: 2054 * 2055 * 1. Retries failed read operations on working mirrors. 2056 * 2. Updates the raid superblock when problems encounter. 2057 * 3. Performs writes following reads for array synchronising. 2058 */ 2059 2060 static void fix_read_error(struct r1conf *conf, int read_disk, 2061 sector_t sect, int sectors) 2062 { 2063 struct mddev *mddev = conf->mddev; 2064 while(sectors) { 2065 int s = sectors; 2066 int d = read_disk; 2067 int success = 0; 2068 int start; 2069 struct md_rdev *rdev; 2070 2071 if (s > (PAGE_SIZE>>9)) 2072 s = PAGE_SIZE >> 9; 2073 2074 do { 2075 /* Note: no rcu protection needed here 2076 * as this is synchronous in the raid1d thread 2077 * which is the thread that might remove 2078 * a device. If raid1d ever becomes multi-threaded.... 2079 */ 2080 sector_t first_bad; 2081 int bad_sectors; 2082 2083 rdev = conf->mirrors[d].rdev; 2084 if (rdev && 2085 (test_bit(In_sync, &rdev->flags) || 2086 (!test_bit(Faulty, &rdev->flags) && 2087 rdev->recovery_offset >= sect + s)) && 2088 is_badblock(rdev, sect, s, 2089 &first_bad, &bad_sectors) == 0 && 2090 sync_page_io(rdev, sect, s<<9, 2091 conf->tmppage, REQ_OP_READ, 0, false)) 2092 success = 1; 2093 else { 2094 d++; 2095 if (d == conf->raid_disks * 2) 2096 d = 0; 2097 } 2098 } while (!success && d != read_disk); 2099 2100 if (!success) { 2101 /* Cannot read from anywhere - mark it bad */ 2102 struct md_rdev *rdev = conf->mirrors[read_disk].rdev; 2103 if (!rdev_set_badblocks(rdev, sect, s, 0)) 2104 md_error(mddev, rdev); 2105 break; 2106 } 2107 /* write it back and re-read */ 2108 start = d; 2109 while (d != read_disk) { 2110 if (d==0) 2111 d = conf->raid_disks * 2; 2112 d--; 2113 rdev = conf->mirrors[d].rdev; 2114 if (rdev && 2115 !test_bit(Faulty, &rdev->flags)) 2116 r1_sync_page_io(rdev, sect, s, 2117 conf->tmppage, WRITE); 2118 } 2119 d = start; 2120 while (d != read_disk) { 2121 char b[BDEVNAME_SIZE]; 2122 if (d==0) 2123 d = conf->raid_disks * 2; 2124 d--; 2125 rdev = conf->mirrors[d].rdev; 2126 if (rdev && 2127 !test_bit(Faulty, &rdev->flags)) { 2128 if (r1_sync_page_io(rdev, sect, s, 2129 conf->tmppage, READ)) { 2130 atomic_add(s, &rdev->corrected_errors); 2131 printk(KERN_INFO 2132 "md/raid1:%s: read error corrected " 2133 "(%d sectors at %llu on %s)\n", 2134 mdname(mddev), s, 2135 (unsigned long long)(sect + 2136 rdev->data_offset), 2137 bdevname(rdev->bdev, b)); 2138 } 2139 } 2140 } 2141 sectors -= s; 2142 sect += s; 2143 } 2144 } 2145 2146 static int narrow_write_error(struct r1bio *r1_bio, int i) 2147 { 2148 struct mddev *mddev = r1_bio->mddev; 2149 struct r1conf *conf = mddev->private; 2150 struct md_rdev *rdev = conf->mirrors[i].rdev; 2151 2152 /* bio has the data to be written to device 'i' where 2153 * we just recently had a write error. 2154 * We repeatedly clone the bio and trim down to one block, 2155 * then try the write. Where the write fails we record 2156 * a bad block. 2157 * It is conceivable that the bio doesn't exactly align with 2158 * blocks. We must handle this somehow. 2159 * 2160 * We currently own a reference on the rdev. 2161 */ 2162 2163 int block_sectors; 2164 sector_t sector; 2165 int sectors; 2166 int sect_to_write = r1_bio->sectors; 2167 int ok = 1; 2168 2169 if (rdev->badblocks.shift < 0) 2170 return 0; 2171 2172 block_sectors = roundup(1 << rdev->badblocks.shift, 2173 bdev_logical_block_size(rdev->bdev) >> 9); 2174 sector = r1_bio->sector; 2175 sectors = ((sector + block_sectors) 2176 & ~(sector_t)(block_sectors - 1)) 2177 - sector; 2178 2179 while (sect_to_write) { 2180 struct bio *wbio; 2181 if (sectors > sect_to_write) 2182 sectors = sect_to_write; 2183 /* Write at 'sector' for 'sectors'*/ 2184 2185 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 2186 unsigned vcnt = r1_bio->behind_page_count; 2187 struct bio_vec *vec = r1_bio->behind_bvecs; 2188 2189 while (!vec->bv_page) { 2190 vec++; 2191 vcnt--; 2192 } 2193 2194 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev); 2195 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec)); 2196 2197 wbio->bi_vcnt = vcnt; 2198 } else { 2199 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2200 } 2201 2202 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0); 2203 wbio->bi_iter.bi_sector = r1_bio->sector; 2204 wbio->bi_iter.bi_size = r1_bio->sectors << 9; 2205 2206 bio_trim(wbio, sector - r1_bio->sector, sectors); 2207 wbio->bi_iter.bi_sector += rdev->data_offset; 2208 wbio->bi_bdev = rdev->bdev; 2209 2210 if (submit_bio_wait(wbio) < 0) 2211 /* failure! */ 2212 ok = rdev_set_badblocks(rdev, sector, 2213 sectors, 0) 2214 && ok; 2215 2216 bio_put(wbio); 2217 sect_to_write -= sectors; 2218 sector += sectors; 2219 sectors = block_sectors; 2220 } 2221 return ok; 2222 } 2223 2224 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2225 { 2226 int m; 2227 int s = r1_bio->sectors; 2228 for (m = 0; m < conf->raid_disks * 2 ; m++) { 2229 struct md_rdev *rdev = conf->mirrors[m].rdev; 2230 struct bio *bio = r1_bio->bios[m]; 2231 if (bio->bi_end_io == NULL) 2232 continue; 2233 if (!bio->bi_error && 2234 test_bit(R1BIO_MadeGood, &r1_bio->state)) { 2235 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0); 2236 } 2237 if (bio->bi_error && 2238 test_bit(R1BIO_WriteError, &r1_bio->state)) { 2239 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0)) 2240 md_error(conf->mddev, rdev); 2241 } 2242 } 2243 put_buf(r1_bio); 2244 md_done_sync(conf->mddev, s, 1); 2245 } 2246 2247 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2248 { 2249 int m; 2250 bool fail = false; 2251 for (m = 0; m < conf->raid_disks * 2 ; m++) 2252 if (r1_bio->bios[m] == IO_MADE_GOOD) { 2253 struct md_rdev *rdev = conf->mirrors[m].rdev; 2254 rdev_clear_badblocks(rdev, 2255 r1_bio->sector, 2256 r1_bio->sectors, 0); 2257 rdev_dec_pending(rdev, conf->mddev); 2258 } else if (r1_bio->bios[m] != NULL) { 2259 /* This drive got a write error. We need to 2260 * narrow down and record precise write 2261 * errors. 2262 */ 2263 fail = true; 2264 if (!narrow_write_error(r1_bio, m)) { 2265 md_error(conf->mddev, 2266 conf->mirrors[m].rdev); 2267 /* an I/O failed, we can't clear the bitmap */ 2268 set_bit(R1BIO_Degraded, &r1_bio->state); 2269 } 2270 rdev_dec_pending(conf->mirrors[m].rdev, 2271 conf->mddev); 2272 } 2273 if (fail) { 2274 spin_lock_irq(&conf->device_lock); 2275 list_add(&r1_bio->retry_list, &conf->bio_end_io_list); 2276 conf->nr_queued++; 2277 spin_unlock_irq(&conf->device_lock); 2278 md_wakeup_thread(conf->mddev->thread); 2279 } else { 2280 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 2281 close_write(r1_bio); 2282 raid_end_bio_io(r1_bio); 2283 } 2284 } 2285 2286 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio) 2287 { 2288 int disk; 2289 int max_sectors; 2290 struct mddev *mddev = conf->mddev; 2291 struct bio *bio; 2292 char b[BDEVNAME_SIZE]; 2293 struct md_rdev *rdev; 2294 2295 clear_bit(R1BIO_ReadError, &r1_bio->state); 2296 /* we got a read error. Maybe the drive is bad. Maybe just 2297 * the block and we can fix it. 2298 * We freeze all other IO, and try reading the block from 2299 * other devices. When we find one, we re-write 2300 * and check it that fixes the read error. 2301 * This is all done synchronously while the array is 2302 * frozen 2303 */ 2304 if (mddev->ro == 0) { 2305 freeze_array(conf, 1); 2306 fix_read_error(conf, r1_bio->read_disk, 2307 r1_bio->sector, r1_bio->sectors); 2308 unfreeze_array(conf); 2309 } else 2310 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev); 2311 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev); 2312 2313 bio = r1_bio->bios[r1_bio->read_disk]; 2314 bdevname(bio->bi_bdev, b); 2315 read_more: 2316 disk = read_balance(conf, r1_bio, &max_sectors); 2317 if (disk == -1) { 2318 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O" 2319 " read error for block %llu\n", 2320 mdname(mddev), b, (unsigned long long)r1_bio->sector); 2321 raid_end_bio_io(r1_bio); 2322 } else { 2323 const unsigned long do_sync 2324 = r1_bio->master_bio->bi_rw & REQ_SYNC; 2325 if (bio) { 2326 r1_bio->bios[r1_bio->read_disk] = 2327 mddev->ro ? IO_BLOCKED : NULL; 2328 bio_put(bio); 2329 } 2330 r1_bio->read_disk = disk; 2331 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2332 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector, 2333 max_sectors); 2334 r1_bio->bios[r1_bio->read_disk] = bio; 2335 rdev = conf->mirrors[disk].rdev; 2336 printk_ratelimited(KERN_ERR 2337 "md/raid1:%s: redirecting sector %llu" 2338 " to other mirror: %s\n", 2339 mdname(mddev), 2340 (unsigned long long)r1_bio->sector, 2341 bdevname(rdev->bdev, b)); 2342 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset; 2343 bio->bi_bdev = rdev->bdev; 2344 bio->bi_end_io = raid1_end_read_request; 2345 bio_set_op_attrs(bio, REQ_OP_READ, do_sync); 2346 bio->bi_private = r1_bio; 2347 if (max_sectors < r1_bio->sectors) { 2348 /* Drat - have to split this up more */ 2349 struct bio *mbio = r1_bio->master_bio; 2350 int sectors_handled = (r1_bio->sector + max_sectors 2351 - mbio->bi_iter.bi_sector); 2352 r1_bio->sectors = max_sectors; 2353 spin_lock_irq(&conf->device_lock); 2354 if (mbio->bi_phys_segments == 0) 2355 mbio->bi_phys_segments = 2; 2356 else 2357 mbio->bi_phys_segments++; 2358 spin_unlock_irq(&conf->device_lock); 2359 generic_make_request(bio); 2360 bio = NULL; 2361 2362 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 2363 2364 r1_bio->master_bio = mbio; 2365 r1_bio->sectors = bio_sectors(mbio) - sectors_handled; 2366 r1_bio->state = 0; 2367 set_bit(R1BIO_ReadError, &r1_bio->state); 2368 r1_bio->mddev = mddev; 2369 r1_bio->sector = mbio->bi_iter.bi_sector + 2370 sectors_handled; 2371 2372 goto read_more; 2373 } else 2374 generic_make_request(bio); 2375 } 2376 } 2377 2378 static void raid1d(struct md_thread *thread) 2379 { 2380 struct mddev *mddev = thread->mddev; 2381 struct r1bio *r1_bio; 2382 unsigned long flags; 2383 struct r1conf *conf = mddev->private; 2384 struct list_head *head = &conf->retry_list; 2385 struct blk_plug plug; 2386 2387 md_check_recovery(mddev); 2388 2389 if (!list_empty_careful(&conf->bio_end_io_list) && 2390 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) { 2391 LIST_HEAD(tmp); 2392 spin_lock_irqsave(&conf->device_lock, flags); 2393 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) { 2394 while (!list_empty(&conf->bio_end_io_list)) { 2395 list_move(conf->bio_end_io_list.prev, &tmp); 2396 conf->nr_queued--; 2397 } 2398 } 2399 spin_unlock_irqrestore(&conf->device_lock, flags); 2400 while (!list_empty(&tmp)) { 2401 r1_bio = list_first_entry(&tmp, struct r1bio, 2402 retry_list); 2403 list_del(&r1_bio->retry_list); 2404 if (mddev->degraded) 2405 set_bit(R1BIO_Degraded, &r1_bio->state); 2406 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 2407 close_write(r1_bio); 2408 raid_end_bio_io(r1_bio); 2409 } 2410 } 2411 2412 blk_start_plug(&plug); 2413 for (;;) { 2414 2415 flush_pending_writes(conf); 2416 2417 spin_lock_irqsave(&conf->device_lock, flags); 2418 if (list_empty(head)) { 2419 spin_unlock_irqrestore(&conf->device_lock, flags); 2420 break; 2421 } 2422 r1_bio = list_entry(head->prev, struct r1bio, retry_list); 2423 list_del(head->prev); 2424 conf->nr_queued--; 2425 spin_unlock_irqrestore(&conf->device_lock, flags); 2426 2427 mddev = r1_bio->mddev; 2428 conf = mddev->private; 2429 if (test_bit(R1BIO_IsSync, &r1_bio->state)) { 2430 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2431 test_bit(R1BIO_WriteError, &r1_bio->state)) 2432 handle_sync_write_finished(conf, r1_bio); 2433 else 2434 sync_request_write(mddev, r1_bio); 2435 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2436 test_bit(R1BIO_WriteError, &r1_bio->state)) 2437 handle_write_finished(conf, r1_bio); 2438 else if (test_bit(R1BIO_ReadError, &r1_bio->state)) 2439 handle_read_error(conf, r1_bio); 2440 else 2441 /* just a partial read to be scheduled from separate 2442 * context 2443 */ 2444 generic_make_request(r1_bio->bios[r1_bio->read_disk]); 2445 2446 cond_resched(); 2447 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) 2448 md_check_recovery(mddev); 2449 } 2450 blk_finish_plug(&plug); 2451 } 2452 2453 static int init_resync(struct r1conf *conf) 2454 { 2455 int buffs; 2456 2457 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; 2458 BUG_ON(conf->r1buf_pool); 2459 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, 2460 conf->poolinfo); 2461 if (!conf->r1buf_pool) 2462 return -ENOMEM; 2463 conf->next_resync = 0; 2464 return 0; 2465 } 2466 2467 /* 2468 * perform a "sync" on one "block" 2469 * 2470 * We need to make sure that no normal I/O request - particularly write 2471 * requests - conflict with active sync requests. 2472 * 2473 * This is achieved by tracking pending requests and a 'barrier' concept 2474 * that can be installed to exclude normal IO requests. 2475 */ 2476 2477 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr, 2478 int *skipped) 2479 { 2480 struct r1conf *conf = mddev->private; 2481 struct r1bio *r1_bio; 2482 struct bio *bio; 2483 sector_t max_sector, nr_sectors; 2484 int disk = -1; 2485 int i; 2486 int wonly = -1; 2487 int write_targets = 0, read_targets = 0; 2488 sector_t sync_blocks; 2489 int still_degraded = 0; 2490 int good_sectors = RESYNC_SECTORS; 2491 int min_bad = 0; /* number of sectors that are bad in all devices */ 2492 2493 if (!conf->r1buf_pool) 2494 if (init_resync(conf)) 2495 return 0; 2496 2497 max_sector = mddev->dev_sectors; 2498 if (sector_nr >= max_sector) { 2499 /* If we aborted, we need to abort the 2500 * sync on the 'current' bitmap chunk (there will 2501 * only be one in raid1 resync. 2502 * We can find the current addess in mddev->curr_resync 2503 */ 2504 if (mddev->curr_resync < max_sector) /* aborted */ 2505 bitmap_end_sync(mddev->bitmap, mddev->curr_resync, 2506 &sync_blocks, 1); 2507 else /* completed sync */ 2508 conf->fullsync = 0; 2509 2510 bitmap_close_sync(mddev->bitmap); 2511 close_sync(conf); 2512 2513 if (mddev_is_clustered(mddev)) { 2514 conf->cluster_sync_low = 0; 2515 conf->cluster_sync_high = 0; 2516 } 2517 return 0; 2518 } 2519 2520 if (mddev->bitmap == NULL && 2521 mddev->recovery_cp == MaxSector && 2522 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && 2523 conf->fullsync == 0) { 2524 *skipped = 1; 2525 return max_sector - sector_nr; 2526 } 2527 /* before building a request, check if we can skip these blocks.. 2528 * This call the bitmap_start_sync doesn't actually record anything 2529 */ 2530 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) && 2531 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2532 /* We can skip this block, and probably several more */ 2533 *skipped = 1; 2534 return sync_blocks; 2535 } 2536 2537 /* we are incrementing sector_nr below. To be safe, we check against 2538 * sector_nr + two times RESYNC_SECTORS 2539 */ 2540 2541 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 2542 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high)); 2543 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO); 2544 2545 raise_barrier(conf, sector_nr); 2546 2547 rcu_read_lock(); 2548 /* 2549 * If we get a correctably read error during resync or recovery, 2550 * we might want to read from a different device. So we 2551 * flag all drives that could conceivably be read from for READ, 2552 * and any others (which will be non-In_sync devices) for WRITE. 2553 * If a read fails, we try reading from something else for which READ 2554 * is OK. 2555 */ 2556 2557 r1_bio->mddev = mddev; 2558 r1_bio->sector = sector_nr; 2559 r1_bio->state = 0; 2560 set_bit(R1BIO_IsSync, &r1_bio->state); 2561 2562 for (i = 0; i < conf->raid_disks * 2; i++) { 2563 struct md_rdev *rdev; 2564 bio = r1_bio->bios[i]; 2565 bio_reset(bio); 2566 2567 rdev = rcu_dereference(conf->mirrors[i].rdev); 2568 if (rdev == NULL || 2569 test_bit(Faulty, &rdev->flags)) { 2570 if (i < conf->raid_disks) 2571 still_degraded = 1; 2572 } else if (!test_bit(In_sync, &rdev->flags)) { 2573 bio_set_op_attrs(bio, REQ_OP_WRITE, 0); 2574 bio->bi_end_io = end_sync_write; 2575 write_targets ++; 2576 } else { 2577 /* may need to read from here */ 2578 sector_t first_bad = MaxSector; 2579 int bad_sectors; 2580 2581 if (is_badblock(rdev, sector_nr, good_sectors, 2582 &first_bad, &bad_sectors)) { 2583 if (first_bad > sector_nr) 2584 good_sectors = first_bad - sector_nr; 2585 else { 2586 bad_sectors -= (sector_nr - first_bad); 2587 if (min_bad == 0 || 2588 min_bad > bad_sectors) 2589 min_bad = bad_sectors; 2590 } 2591 } 2592 if (sector_nr < first_bad) { 2593 if (test_bit(WriteMostly, &rdev->flags)) { 2594 if (wonly < 0) 2595 wonly = i; 2596 } else { 2597 if (disk < 0) 2598 disk = i; 2599 } 2600 bio_set_op_attrs(bio, REQ_OP_READ, 0); 2601 bio->bi_end_io = end_sync_read; 2602 read_targets++; 2603 } else if (!test_bit(WriteErrorSeen, &rdev->flags) && 2604 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && 2605 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) { 2606 /* 2607 * The device is suitable for reading (InSync), 2608 * but has bad block(s) here. Let's try to correct them, 2609 * if we are doing resync or repair. Otherwise, leave 2610 * this device alone for this sync request. 2611 */ 2612 bio_set_op_attrs(bio, REQ_OP_WRITE, 0); 2613 bio->bi_end_io = end_sync_write; 2614 write_targets++; 2615 } 2616 } 2617 if (bio->bi_end_io) { 2618 atomic_inc(&rdev->nr_pending); 2619 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset; 2620 bio->bi_bdev = rdev->bdev; 2621 bio->bi_private = r1_bio; 2622 } 2623 } 2624 rcu_read_unlock(); 2625 if (disk < 0) 2626 disk = wonly; 2627 r1_bio->read_disk = disk; 2628 2629 if (read_targets == 0 && min_bad > 0) { 2630 /* These sectors are bad on all InSync devices, so we 2631 * need to mark them bad on all write targets 2632 */ 2633 int ok = 1; 2634 for (i = 0 ; i < conf->raid_disks * 2 ; i++) 2635 if (r1_bio->bios[i]->bi_end_io == end_sync_write) { 2636 struct md_rdev *rdev = conf->mirrors[i].rdev; 2637 ok = rdev_set_badblocks(rdev, sector_nr, 2638 min_bad, 0 2639 ) && ok; 2640 } 2641 set_bit(MD_CHANGE_DEVS, &mddev->flags); 2642 *skipped = 1; 2643 put_buf(r1_bio); 2644 2645 if (!ok) { 2646 /* Cannot record the badblocks, so need to 2647 * abort the resync. 2648 * If there are multiple read targets, could just 2649 * fail the really bad ones ??? 2650 */ 2651 conf->recovery_disabled = mddev->recovery_disabled; 2652 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 2653 return 0; 2654 } else 2655 return min_bad; 2656 2657 } 2658 if (min_bad > 0 && min_bad < good_sectors) { 2659 /* only resync enough to reach the next bad->good 2660 * transition */ 2661 good_sectors = min_bad; 2662 } 2663 2664 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0) 2665 /* extra read targets are also write targets */ 2666 write_targets += read_targets-1; 2667 2668 if (write_targets == 0 || read_targets == 0) { 2669 /* There is nowhere to write, so all non-sync 2670 * drives must be failed - so we are finished 2671 */ 2672 sector_t rv; 2673 if (min_bad > 0) 2674 max_sector = sector_nr + min_bad; 2675 rv = max_sector - sector_nr; 2676 *skipped = 1; 2677 put_buf(r1_bio); 2678 return rv; 2679 } 2680 2681 if (max_sector > mddev->resync_max) 2682 max_sector = mddev->resync_max; /* Don't do IO beyond here */ 2683 if (max_sector > sector_nr + good_sectors) 2684 max_sector = sector_nr + good_sectors; 2685 nr_sectors = 0; 2686 sync_blocks = 0; 2687 do { 2688 struct page *page; 2689 int len = PAGE_SIZE; 2690 if (sector_nr + (len>>9) > max_sector) 2691 len = (max_sector - sector_nr) << 9; 2692 if (len == 0) 2693 break; 2694 if (sync_blocks == 0) { 2695 if (!bitmap_start_sync(mddev->bitmap, sector_nr, 2696 &sync_blocks, still_degraded) && 2697 !conf->fullsync && 2698 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2699 break; 2700 if ((len >> 9) > sync_blocks) 2701 len = sync_blocks<<9; 2702 } 2703 2704 for (i = 0 ; i < conf->raid_disks * 2; i++) { 2705 bio = r1_bio->bios[i]; 2706 if (bio->bi_end_io) { 2707 page = bio->bi_io_vec[bio->bi_vcnt].bv_page; 2708 if (bio_add_page(bio, page, len, 0) == 0) { 2709 /* stop here */ 2710 bio->bi_io_vec[bio->bi_vcnt].bv_page = page; 2711 while (i > 0) { 2712 i--; 2713 bio = r1_bio->bios[i]; 2714 if (bio->bi_end_io==NULL) 2715 continue; 2716 /* remove last page from this bio */ 2717 bio->bi_vcnt--; 2718 bio->bi_iter.bi_size -= len; 2719 bio_clear_flag(bio, BIO_SEG_VALID); 2720 } 2721 goto bio_full; 2722 } 2723 } 2724 } 2725 nr_sectors += len>>9; 2726 sector_nr += len>>9; 2727 sync_blocks -= (len>>9); 2728 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES); 2729 bio_full: 2730 r1_bio->sectors = nr_sectors; 2731 2732 if (mddev_is_clustered(mddev) && 2733 conf->cluster_sync_high < sector_nr + nr_sectors) { 2734 conf->cluster_sync_low = mddev->curr_resync_completed; 2735 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS; 2736 /* Send resync message */ 2737 md_cluster_ops->resync_info_update(mddev, 2738 conf->cluster_sync_low, 2739 conf->cluster_sync_high); 2740 } 2741 2742 /* For a user-requested sync, we read all readable devices and do a 2743 * compare 2744 */ 2745 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2746 atomic_set(&r1_bio->remaining, read_targets); 2747 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) { 2748 bio = r1_bio->bios[i]; 2749 if (bio->bi_end_io == end_sync_read) { 2750 read_targets--; 2751 md_sync_acct(bio->bi_bdev, nr_sectors); 2752 generic_make_request(bio); 2753 } 2754 } 2755 } else { 2756 atomic_set(&r1_bio->remaining, 1); 2757 bio = r1_bio->bios[r1_bio->read_disk]; 2758 md_sync_acct(bio->bi_bdev, nr_sectors); 2759 generic_make_request(bio); 2760 2761 } 2762 return nr_sectors; 2763 } 2764 2765 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks) 2766 { 2767 if (sectors) 2768 return sectors; 2769 2770 return mddev->dev_sectors; 2771 } 2772 2773 static struct r1conf *setup_conf(struct mddev *mddev) 2774 { 2775 struct r1conf *conf; 2776 int i; 2777 struct raid1_info *disk; 2778 struct md_rdev *rdev; 2779 int err = -ENOMEM; 2780 2781 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL); 2782 if (!conf) 2783 goto abort; 2784 2785 conf->mirrors = kzalloc(sizeof(struct raid1_info) 2786 * mddev->raid_disks * 2, 2787 GFP_KERNEL); 2788 if (!conf->mirrors) 2789 goto abort; 2790 2791 conf->tmppage = alloc_page(GFP_KERNEL); 2792 if (!conf->tmppage) 2793 goto abort; 2794 2795 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL); 2796 if (!conf->poolinfo) 2797 goto abort; 2798 conf->poolinfo->raid_disks = mddev->raid_disks * 2; 2799 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 2800 r1bio_pool_free, 2801 conf->poolinfo); 2802 if (!conf->r1bio_pool) 2803 goto abort; 2804 2805 conf->poolinfo->mddev = mddev; 2806 2807 err = -EINVAL; 2808 spin_lock_init(&conf->device_lock); 2809 rdev_for_each(rdev, mddev) { 2810 struct request_queue *q; 2811 int disk_idx = rdev->raid_disk; 2812 if (disk_idx >= mddev->raid_disks 2813 || disk_idx < 0) 2814 continue; 2815 if (test_bit(Replacement, &rdev->flags)) 2816 disk = conf->mirrors + mddev->raid_disks + disk_idx; 2817 else 2818 disk = conf->mirrors + disk_idx; 2819 2820 if (disk->rdev) 2821 goto abort; 2822 disk->rdev = rdev; 2823 q = bdev_get_queue(rdev->bdev); 2824 2825 disk->head_position = 0; 2826 disk->seq_start = MaxSector; 2827 } 2828 conf->raid_disks = mddev->raid_disks; 2829 conf->mddev = mddev; 2830 INIT_LIST_HEAD(&conf->retry_list); 2831 INIT_LIST_HEAD(&conf->bio_end_io_list); 2832 2833 spin_lock_init(&conf->resync_lock); 2834 init_waitqueue_head(&conf->wait_barrier); 2835 2836 bio_list_init(&conf->pending_bio_list); 2837 conf->pending_count = 0; 2838 conf->recovery_disabled = mddev->recovery_disabled - 1; 2839 2840 conf->start_next_window = MaxSector; 2841 conf->current_window_requests = conf->next_window_requests = 0; 2842 2843 err = -EIO; 2844 for (i = 0; i < conf->raid_disks * 2; i++) { 2845 2846 disk = conf->mirrors + i; 2847 2848 if (i < conf->raid_disks && 2849 disk[conf->raid_disks].rdev) { 2850 /* This slot has a replacement. */ 2851 if (!disk->rdev) { 2852 /* No original, just make the replacement 2853 * a recovering spare 2854 */ 2855 disk->rdev = 2856 disk[conf->raid_disks].rdev; 2857 disk[conf->raid_disks].rdev = NULL; 2858 } else if (!test_bit(In_sync, &disk->rdev->flags)) 2859 /* Original is not in_sync - bad */ 2860 goto abort; 2861 } 2862 2863 if (!disk->rdev || 2864 !test_bit(In_sync, &disk->rdev->flags)) { 2865 disk->head_position = 0; 2866 if (disk->rdev && 2867 (disk->rdev->saved_raid_disk < 0)) 2868 conf->fullsync = 1; 2869 } 2870 } 2871 2872 err = -ENOMEM; 2873 conf->thread = md_register_thread(raid1d, mddev, "raid1"); 2874 if (!conf->thread) { 2875 printk(KERN_ERR 2876 "md/raid1:%s: couldn't allocate thread\n", 2877 mdname(mddev)); 2878 goto abort; 2879 } 2880 2881 return conf; 2882 2883 abort: 2884 if (conf) { 2885 mempool_destroy(conf->r1bio_pool); 2886 kfree(conf->mirrors); 2887 safe_put_page(conf->tmppage); 2888 kfree(conf->poolinfo); 2889 kfree(conf); 2890 } 2891 return ERR_PTR(err); 2892 } 2893 2894 static void raid1_free(struct mddev *mddev, void *priv); 2895 static int raid1_run(struct mddev *mddev) 2896 { 2897 struct r1conf *conf; 2898 int i; 2899 struct md_rdev *rdev; 2900 int ret; 2901 bool discard_supported = false; 2902 2903 if (mddev->level != 1) { 2904 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n", 2905 mdname(mddev), mddev->level); 2906 return -EIO; 2907 } 2908 if (mddev->reshape_position != MaxSector) { 2909 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n", 2910 mdname(mddev)); 2911 return -EIO; 2912 } 2913 /* 2914 * copy the already verified devices into our private RAID1 2915 * bookkeeping area. [whatever we allocate in run(), 2916 * should be freed in raid1_free()] 2917 */ 2918 if (mddev->private == NULL) 2919 conf = setup_conf(mddev); 2920 else 2921 conf = mddev->private; 2922 2923 if (IS_ERR(conf)) 2924 return PTR_ERR(conf); 2925 2926 if (mddev->queue) 2927 blk_queue_max_write_same_sectors(mddev->queue, 0); 2928 2929 rdev_for_each(rdev, mddev) { 2930 if (!mddev->gendisk) 2931 continue; 2932 disk_stack_limits(mddev->gendisk, rdev->bdev, 2933 rdev->data_offset << 9); 2934 if (blk_queue_discard(bdev_get_queue(rdev->bdev))) 2935 discard_supported = true; 2936 } 2937 2938 mddev->degraded = 0; 2939 for (i=0; i < conf->raid_disks; i++) 2940 if (conf->mirrors[i].rdev == NULL || 2941 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) || 2942 test_bit(Faulty, &conf->mirrors[i].rdev->flags)) 2943 mddev->degraded++; 2944 2945 if (conf->raid_disks - mddev->degraded == 1) 2946 mddev->recovery_cp = MaxSector; 2947 2948 if (mddev->recovery_cp != MaxSector) 2949 printk(KERN_NOTICE "md/raid1:%s: not clean" 2950 " -- starting background reconstruction\n", 2951 mdname(mddev)); 2952 printk(KERN_INFO 2953 "md/raid1:%s: active with %d out of %d mirrors\n", 2954 mdname(mddev), mddev->raid_disks - mddev->degraded, 2955 mddev->raid_disks); 2956 2957 /* 2958 * Ok, everything is just fine now 2959 */ 2960 mddev->thread = conf->thread; 2961 conf->thread = NULL; 2962 mddev->private = conf; 2963 2964 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0)); 2965 2966 if (mddev->queue) { 2967 if (discard_supported) 2968 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, 2969 mddev->queue); 2970 else 2971 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, 2972 mddev->queue); 2973 } 2974 2975 ret = md_integrity_register(mddev); 2976 if (ret) { 2977 md_unregister_thread(&mddev->thread); 2978 raid1_free(mddev, conf); 2979 } 2980 return ret; 2981 } 2982 2983 static void raid1_free(struct mddev *mddev, void *priv) 2984 { 2985 struct r1conf *conf = priv; 2986 2987 mempool_destroy(conf->r1bio_pool); 2988 kfree(conf->mirrors); 2989 safe_put_page(conf->tmppage); 2990 kfree(conf->poolinfo); 2991 kfree(conf); 2992 } 2993 2994 static int raid1_resize(struct mddev *mddev, sector_t sectors) 2995 { 2996 /* no resync is happening, and there is enough space 2997 * on all devices, so we can resize. 2998 * We need to make sure resync covers any new space. 2999 * If the array is shrinking we should possibly wait until 3000 * any io in the removed space completes, but it hardly seems 3001 * worth it. 3002 */ 3003 sector_t newsize = raid1_size(mddev, sectors, 0); 3004 if (mddev->external_size && 3005 mddev->array_sectors > newsize) 3006 return -EINVAL; 3007 if (mddev->bitmap) { 3008 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0); 3009 if (ret) 3010 return ret; 3011 } 3012 md_set_array_sectors(mddev, newsize); 3013 set_capacity(mddev->gendisk, mddev->array_sectors); 3014 revalidate_disk(mddev->gendisk); 3015 if (sectors > mddev->dev_sectors && 3016 mddev->recovery_cp > mddev->dev_sectors) { 3017 mddev->recovery_cp = mddev->dev_sectors; 3018 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3019 } 3020 mddev->dev_sectors = sectors; 3021 mddev->resync_max_sectors = sectors; 3022 return 0; 3023 } 3024 3025 static int raid1_reshape(struct mddev *mddev) 3026 { 3027 /* We need to: 3028 * 1/ resize the r1bio_pool 3029 * 2/ resize conf->mirrors 3030 * 3031 * We allocate a new r1bio_pool if we can. 3032 * Then raise a device barrier and wait until all IO stops. 3033 * Then resize conf->mirrors and swap in the new r1bio pool. 3034 * 3035 * At the same time, we "pack" the devices so that all the missing 3036 * devices have the higher raid_disk numbers. 3037 */ 3038 mempool_t *newpool, *oldpool; 3039 struct pool_info *newpoolinfo; 3040 struct raid1_info *newmirrors; 3041 struct r1conf *conf = mddev->private; 3042 int cnt, raid_disks; 3043 unsigned long flags; 3044 int d, d2, err; 3045 3046 /* Cannot change chunk_size, layout, or level */ 3047 if (mddev->chunk_sectors != mddev->new_chunk_sectors || 3048 mddev->layout != mddev->new_layout || 3049 mddev->level != mddev->new_level) { 3050 mddev->new_chunk_sectors = mddev->chunk_sectors; 3051 mddev->new_layout = mddev->layout; 3052 mddev->new_level = mddev->level; 3053 return -EINVAL; 3054 } 3055 3056 if (!mddev_is_clustered(mddev)) { 3057 err = md_allow_write(mddev); 3058 if (err) 3059 return err; 3060 } 3061 3062 raid_disks = mddev->raid_disks + mddev->delta_disks; 3063 3064 if (raid_disks < conf->raid_disks) { 3065 cnt=0; 3066 for (d= 0; d < conf->raid_disks; d++) 3067 if (conf->mirrors[d].rdev) 3068 cnt++; 3069 if (cnt > raid_disks) 3070 return -EBUSY; 3071 } 3072 3073 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL); 3074 if (!newpoolinfo) 3075 return -ENOMEM; 3076 newpoolinfo->mddev = mddev; 3077 newpoolinfo->raid_disks = raid_disks * 2; 3078 3079 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 3080 r1bio_pool_free, newpoolinfo); 3081 if (!newpool) { 3082 kfree(newpoolinfo); 3083 return -ENOMEM; 3084 } 3085 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2, 3086 GFP_KERNEL); 3087 if (!newmirrors) { 3088 kfree(newpoolinfo); 3089 mempool_destroy(newpool); 3090 return -ENOMEM; 3091 } 3092 3093 freeze_array(conf, 0); 3094 3095 /* ok, everything is stopped */ 3096 oldpool = conf->r1bio_pool; 3097 conf->r1bio_pool = newpool; 3098 3099 for (d = d2 = 0; d < conf->raid_disks; d++) { 3100 struct md_rdev *rdev = conf->mirrors[d].rdev; 3101 if (rdev && rdev->raid_disk != d2) { 3102 sysfs_unlink_rdev(mddev, rdev); 3103 rdev->raid_disk = d2; 3104 sysfs_unlink_rdev(mddev, rdev); 3105 if (sysfs_link_rdev(mddev, rdev)) 3106 printk(KERN_WARNING 3107 "md/raid1:%s: cannot register rd%d\n", 3108 mdname(mddev), rdev->raid_disk); 3109 } 3110 if (rdev) 3111 newmirrors[d2++].rdev = rdev; 3112 } 3113 kfree(conf->mirrors); 3114 conf->mirrors = newmirrors; 3115 kfree(conf->poolinfo); 3116 conf->poolinfo = newpoolinfo; 3117 3118 spin_lock_irqsave(&conf->device_lock, flags); 3119 mddev->degraded += (raid_disks - conf->raid_disks); 3120 spin_unlock_irqrestore(&conf->device_lock, flags); 3121 conf->raid_disks = mddev->raid_disks = raid_disks; 3122 mddev->delta_disks = 0; 3123 3124 unfreeze_array(conf); 3125 3126 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery); 3127 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3128 md_wakeup_thread(mddev->thread); 3129 3130 mempool_destroy(oldpool); 3131 return 0; 3132 } 3133 3134 static void raid1_quiesce(struct mddev *mddev, int state) 3135 { 3136 struct r1conf *conf = mddev->private; 3137 3138 switch(state) { 3139 case 2: /* wake for suspend */ 3140 wake_up(&conf->wait_barrier); 3141 break; 3142 case 1: 3143 freeze_array(conf, 0); 3144 break; 3145 case 0: 3146 unfreeze_array(conf); 3147 break; 3148 } 3149 } 3150 3151 static void *raid1_takeover(struct mddev *mddev) 3152 { 3153 /* raid1 can take over: 3154 * raid5 with 2 devices, any layout or chunk size 3155 */ 3156 if (mddev->level == 5 && mddev->raid_disks == 2) { 3157 struct r1conf *conf; 3158 mddev->new_level = 1; 3159 mddev->new_layout = 0; 3160 mddev->new_chunk_sectors = 0; 3161 conf = setup_conf(mddev); 3162 if (!IS_ERR(conf)) 3163 /* Array must appear to be quiesced */ 3164 conf->array_frozen = 1; 3165 return conf; 3166 } 3167 return ERR_PTR(-EINVAL); 3168 } 3169 3170 static struct md_personality raid1_personality = 3171 { 3172 .name = "raid1", 3173 .level = 1, 3174 .owner = THIS_MODULE, 3175 .make_request = raid1_make_request, 3176 .run = raid1_run, 3177 .free = raid1_free, 3178 .status = raid1_status, 3179 .error_handler = raid1_error, 3180 .hot_add_disk = raid1_add_disk, 3181 .hot_remove_disk= raid1_remove_disk, 3182 .spare_active = raid1_spare_active, 3183 .sync_request = raid1_sync_request, 3184 .resize = raid1_resize, 3185 .size = raid1_size, 3186 .check_reshape = raid1_reshape, 3187 .quiesce = raid1_quiesce, 3188 .takeover = raid1_takeover, 3189 .congested = raid1_congested, 3190 }; 3191 3192 static int __init raid_init(void) 3193 { 3194 return register_md_personality(&raid1_personality); 3195 } 3196 3197 static void raid_exit(void) 3198 { 3199 unregister_md_personality(&raid1_personality); 3200 } 3201 3202 module_init(raid_init); 3203 module_exit(raid_exit); 3204 MODULE_LICENSE("GPL"); 3205 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD"); 3206 MODULE_ALIAS("md-personality-3"); /* RAID1 */ 3207 MODULE_ALIAS("md-raid1"); 3208 MODULE_ALIAS("md-level-1"); 3209 3210 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR); 3211