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 NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS) 94 95 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data) 96 { 97 struct pool_info *pi = data; 98 struct r1bio *r1_bio; 99 struct bio *bio; 100 int need_pages; 101 int i, j; 102 103 r1_bio = r1bio_pool_alloc(gfp_flags, pi); 104 if (!r1_bio) 105 return NULL; 106 107 /* 108 * Allocate bios : 1 for reading, n-1 for writing 109 */ 110 for (j = pi->raid_disks ; j-- ; ) { 111 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES); 112 if (!bio) 113 goto out_free_bio; 114 r1_bio->bios[j] = bio; 115 } 116 /* 117 * Allocate RESYNC_PAGES data pages and attach them to 118 * the first bio. 119 * If this is a user-requested check/repair, allocate 120 * RESYNC_PAGES for each bio. 121 */ 122 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) 123 need_pages = pi->raid_disks; 124 else 125 need_pages = 1; 126 for (j = 0; j < need_pages; j++) { 127 bio = r1_bio->bios[j]; 128 bio->bi_vcnt = RESYNC_PAGES; 129 130 if (bio_alloc_pages(bio, gfp_flags)) 131 goto out_free_pages; 132 } 133 /* If not user-requests, copy the page pointers to all bios */ 134 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) { 135 for (i=0; i<RESYNC_PAGES ; i++) 136 for (j=1; j<pi->raid_disks; j++) 137 r1_bio->bios[j]->bi_io_vec[i].bv_page = 138 r1_bio->bios[0]->bi_io_vec[i].bv_page; 139 } 140 141 r1_bio->master_bio = NULL; 142 143 return r1_bio; 144 145 out_free_pages: 146 while (--j >= 0) { 147 struct bio_vec *bv; 148 149 bio_for_each_segment_all(bv, r1_bio->bios[j], i) 150 __free_page(bv->bv_page); 151 } 152 153 out_free_bio: 154 while (++j < pi->raid_disks) 155 bio_put(r1_bio->bios[j]); 156 r1bio_pool_free(r1_bio, data); 157 return NULL; 158 } 159 160 static void r1buf_pool_free(void *__r1_bio, void *data) 161 { 162 struct pool_info *pi = data; 163 int i,j; 164 struct r1bio *r1bio = __r1_bio; 165 166 for (i = 0; i < RESYNC_PAGES; i++) 167 for (j = pi->raid_disks; j-- ;) { 168 if (j == 0 || 169 r1bio->bios[j]->bi_io_vec[i].bv_page != 170 r1bio->bios[0]->bi_io_vec[i].bv_page) 171 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page); 172 } 173 for (i=0 ; i < pi->raid_disks; i++) 174 bio_put(r1bio->bios[i]); 175 176 r1bio_pool_free(r1bio, data); 177 } 178 179 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio) 180 { 181 int i; 182 183 for (i = 0; i < conf->raid_disks * 2; i++) { 184 struct bio **bio = r1_bio->bios + i; 185 if (!BIO_SPECIAL(*bio)) 186 bio_put(*bio); 187 *bio = NULL; 188 } 189 } 190 191 static void free_r1bio(struct r1bio *r1_bio) 192 { 193 struct r1conf *conf = r1_bio->mddev->private; 194 195 put_all_bios(conf, r1_bio); 196 mempool_free(r1_bio, conf->r1bio_pool); 197 } 198 199 static void put_buf(struct r1bio *r1_bio) 200 { 201 struct r1conf *conf = r1_bio->mddev->private; 202 int i; 203 204 for (i = 0; i < conf->raid_disks * 2; i++) { 205 struct bio *bio = r1_bio->bios[i]; 206 if (bio->bi_end_io) 207 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev); 208 } 209 210 mempool_free(r1_bio, conf->r1buf_pool); 211 212 lower_barrier(conf); 213 } 214 215 static void reschedule_retry(struct r1bio *r1_bio) 216 { 217 unsigned long flags; 218 struct mddev *mddev = r1_bio->mddev; 219 struct r1conf *conf = mddev->private; 220 221 spin_lock_irqsave(&conf->device_lock, flags); 222 list_add(&r1_bio->retry_list, &conf->retry_list); 223 conf->nr_queued ++; 224 spin_unlock_irqrestore(&conf->device_lock, flags); 225 226 wake_up(&conf->wait_barrier); 227 md_wakeup_thread(mddev->thread); 228 } 229 230 /* 231 * raid_end_bio_io() is called when we have finished servicing a mirrored 232 * operation and are ready to return a success/failure code to the buffer 233 * cache layer. 234 */ 235 static void call_bio_endio(struct r1bio *r1_bio) 236 { 237 struct bio *bio = r1_bio->master_bio; 238 int done; 239 struct r1conf *conf = r1_bio->mddev->private; 240 sector_t start_next_window = r1_bio->start_next_window; 241 sector_t bi_sector = bio->bi_iter.bi_sector; 242 243 if (bio->bi_phys_segments) { 244 unsigned long flags; 245 spin_lock_irqsave(&conf->device_lock, flags); 246 bio->bi_phys_segments--; 247 done = (bio->bi_phys_segments == 0); 248 spin_unlock_irqrestore(&conf->device_lock, flags); 249 /* 250 * make_request() might be waiting for 251 * bi_phys_segments to decrease 252 */ 253 wake_up(&conf->wait_barrier); 254 } else 255 done = 1; 256 257 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 258 clear_bit(BIO_UPTODATE, &bio->bi_flags); 259 if (done) { 260 bio_endio(bio, 0); 261 /* 262 * Wake up any possible resync thread that waits for the device 263 * to go idle. 264 */ 265 allow_barrier(conf, start_next_window, bi_sector); 266 } 267 } 268 269 static void raid_end_bio_io(struct r1bio *r1_bio) 270 { 271 struct bio *bio = r1_bio->master_bio; 272 273 /* if nobody has done the final endio yet, do it now */ 274 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { 275 pr_debug("raid1: sync end %s on sectors %llu-%llu\n", 276 (bio_data_dir(bio) == WRITE) ? "write" : "read", 277 (unsigned long long) bio->bi_iter.bi_sector, 278 (unsigned long long) bio_end_sector(bio) - 1); 279 280 call_bio_endio(r1_bio); 281 } 282 free_r1bio(r1_bio); 283 } 284 285 /* 286 * Update disk head position estimator based on IRQ completion info. 287 */ 288 static inline void update_head_pos(int disk, struct r1bio *r1_bio) 289 { 290 struct r1conf *conf = r1_bio->mddev->private; 291 292 conf->mirrors[disk].head_position = 293 r1_bio->sector + (r1_bio->sectors); 294 } 295 296 /* 297 * Find the disk number which triggered given bio 298 */ 299 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio) 300 { 301 int mirror; 302 struct r1conf *conf = r1_bio->mddev->private; 303 int raid_disks = conf->raid_disks; 304 305 for (mirror = 0; mirror < raid_disks * 2; mirror++) 306 if (r1_bio->bios[mirror] == bio) 307 break; 308 309 BUG_ON(mirror == raid_disks * 2); 310 update_head_pos(mirror, r1_bio); 311 312 return mirror; 313 } 314 315 static void raid1_end_read_request(struct bio *bio, int error) 316 { 317 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 318 struct r1bio *r1_bio = bio->bi_private; 319 int mirror; 320 struct r1conf *conf = r1_bio->mddev->private; 321 322 mirror = r1_bio->read_disk; 323 /* 324 * this branch is our 'one mirror IO has finished' event handler: 325 */ 326 update_head_pos(mirror, r1_bio); 327 328 if (uptodate) 329 set_bit(R1BIO_Uptodate, &r1_bio->state); 330 else { 331 /* If all other devices have failed, we want to return 332 * the error upwards rather than fail the last device. 333 * Here we redefine "uptodate" to mean "Don't want to retry" 334 */ 335 unsigned long flags; 336 spin_lock_irqsave(&conf->device_lock, flags); 337 if (r1_bio->mddev->degraded == conf->raid_disks || 338 (r1_bio->mddev->degraded == conf->raid_disks-1 && 339 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))) 340 uptodate = 1; 341 spin_unlock_irqrestore(&conf->device_lock, flags); 342 } 343 344 if (uptodate) { 345 raid_end_bio_io(r1_bio); 346 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev); 347 } else { 348 /* 349 * oops, read error: 350 */ 351 char b[BDEVNAME_SIZE]; 352 printk_ratelimited( 353 KERN_ERR "md/raid1:%s: %s: " 354 "rescheduling sector %llu\n", 355 mdname(conf->mddev), 356 bdevname(conf->mirrors[mirror].rdev->bdev, 357 b), 358 (unsigned long long)r1_bio->sector); 359 set_bit(R1BIO_ReadError, &r1_bio->state); 360 reschedule_retry(r1_bio); 361 /* don't drop the reference on read_disk yet */ 362 } 363 } 364 365 static void close_write(struct r1bio *r1_bio) 366 { 367 /* it really is the end of this request */ 368 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 369 /* free extra copy of the data pages */ 370 int i = r1_bio->behind_page_count; 371 while (i--) 372 safe_put_page(r1_bio->behind_bvecs[i].bv_page); 373 kfree(r1_bio->behind_bvecs); 374 r1_bio->behind_bvecs = NULL; 375 } 376 /* clear the bitmap if all writes complete successfully */ 377 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector, 378 r1_bio->sectors, 379 !test_bit(R1BIO_Degraded, &r1_bio->state), 380 test_bit(R1BIO_BehindIO, &r1_bio->state)); 381 md_write_end(r1_bio->mddev); 382 } 383 384 static void r1_bio_write_done(struct r1bio *r1_bio) 385 { 386 if (!atomic_dec_and_test(&r1_bio->remaining)) 387 return; 388 389 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 390 reschedule_retry(r1_bio); 391 else { 392 close_write(r1_bio); 393 if (test_bit(R1BIO_MadeGood, &r1_bio->state)) 394 reschedule_retry(r1_bio); 395 else 396 raid_end_bio_io(r1_bio); 397 } 398 } 399 400 static void raid1_end_write_request(struct bio *bio, int error) 401 { 402 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 403 struct r1bio *r1_bio = bio->bi_private; 404 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state); 405 struct r1conf *conf = r1_bio->mddev->private; 406 struct bio *to_put = NULL; 407 408 mirror = find_bio_disk(r1_bio, bio); 409 410 /* 411 * 'one mirror IO has finished' event handler: 412 */ 413 if (!uptodate) { 414 set_bit(WriteErrorSeen, 415 &conf->mirrors[mirror].rdev->flags); 416 if (!test_and_set_bit(WantReplacement, 417 &conf->mirrors[mirror].rdev->flags)) 418 set_bit(MD_RECOVERY_NEEDED, & 419 conf->mddev->recovery); 420 421 set_bit(R1BIO_WriteError, &r1_bio->state); 422 } else { 423 /* 424 * Set R1BIO_Uptodate in our master bio, so that we 425 * will return a good error code for to the higher 426 * levels even if IO on some other mirrored buffer 427 * fails. 428 * 429 * The 'master' represents the composite IO operation 430 * to user-side. So if something waits for IO, then it 431 * will wait for the 'master' bio. 432 */ 433 sector_t first_bad; 434 int bad_sectors; 435 436 r1_bio->bios[mirror] = NULL; 437 to_put = bio; 438 /* 439 * Do not set R1BIO_Uptodate if the current device is 440 * rebuilding or Faulty. This is because we cannot use 441 * such device for properly reading the data back (we could 442 * potentially use it, if the current write would have felt 443 * before rdev->recovery_offset, but for simplicity we don't 444 * check this here. 445 */ 446 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) && 447 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)) 448 set_bit(R1BIO_Uptodate, &r1_bio->state); 449 450 /* Maybe we can clear some bad blocks. */ 451 if (is_badblock(conf->mirrors[mirror].rdev, 452 r1_bio->sector, r1_bio->sectors, 453 &first_bad, &bad_sectors)) { 454 r1_bio->bios[mirror] = IO_MADE_GOOD; 455 set_bit(R1BIO_MadeGood, &r1_bio->state); 456 } 457 } 458 459 if (behind) { 460 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags)) 461 atomic_dec(&r1_bio->behind_remaining); 462 463 /* 464 * In behind mode, we ACK the master bio once the I/O 465 * has safely reached all non-writemostly 466 * disks. Setting the Returned bit ensures that this 467 * gets done only once -- we don't ever want to return 468 * -EIO here, instead we'll wait 469 */ 470 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) && 471 test_bit(R1BIO_Uptodate, &r1_bio->state)) { 472 /* Maybe we can return now */ 473 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { 474 struct bio *mbio = r1_bio->master_bio; 475 pr_debug("raid1: behind end write sectors" 476 " %llu-%llu\n", 477 (unsigned long long) mbio->bi_iter.bi_sector, 478 (unsigned long long) bio_end_sector(mbio) - 1); 479 call_bio_endio(r1_bio); 480 } 481 } 482 } 483 if (r1_bio->bios[mirror] == NULL) 484 rdev_dec_pending(conf->mirrors[mirror].rdev, 485 conf->mddev); 486 487 /* 488 * Let's see if all mirrored write operations have finished 489 * already. 490 */ 491 r1_bio_write_done(r1_bio); 492 493 if (to_put) 494 bio_put(to_put); 495 } 496 497 /* 498 * This routine returns the disk from which the requested read should 499 * be done. There is a per-array 'next expected sequential IO' sector 500 * number - if this matches on the next IO then we use the last disk. 501 * There is also a per-disk 'last know head position' sector that is 502 * maintained from IRQ contexts, both the normal and the resync IO 503 * completion handlers update this position correctly. If there is no 504 * perfect sequential match then we pick the disk whose head is closest. 505 * 506 * If there are 2 mirrors in the same 2 devices, performance degrades 507 * because position is mirror, not device based. 508 * 509 * The rdev for the device selected will have nr_pending incremented. 510 */ 511 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors) 512 { 513 const sector_t this_sector = r1_bio->sector; 514 int sectors; 515 int best_good_sectors; 516 int best_disk, best_dist_disk, best_pending_disk; 517 int has_nonrot_disk; 518 int disk; 519 sector_t best_dist; 520 unsigned int min_pending; 521 struct md_rdev *rdev; 522 int choose_first; 523 int choose_next_idle; 524 525 rcu_read_lock(); 526 /* 527 * Check if we can balance. We can balance on the whole 528 * device if no resync is going on, or below the resync window. 529 * We take the first readable disk when above the resync window. 530 */ 531 retry: 532 sectors = r1_bio->sectors; 533 best_disk = -1; 534 best_dist_disk = -1; 535 best_dist = MaxSector; 536 best_pending_disk = -1; 537 min_pending = UINT_MAX; 538 best_good_sectors = 0; 539 has_nonrot_disk = 0; 540 choose_next_idle = 0; 541 542 if ((conf->mddev->recovery_cp < this_sector + sectors) || 543 (mddev_is_clustered(conf->mddev) && 544 md_cluster_ops->area_resyncing(conf->mddev, this_sector, 545 this_sector + sectors))) 546 choose_first = 1; 547 else 548 choose_first = 0; 549 550 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) { 551 sector_t dist; 552 sector_t first_bad; 553 int bad_sectors; 554 unsigned int pending; 555 bool nonrot; 556 557 rdev = rcu_dereference(conf->mirrors[disk].rdev); 558 if (r1_bio->bios[disk] == IO_BLOCKED 559 || rdev == NULL 560 || test_bit(Unmerged, &rdev->flags) 561 || test_bit(Faulty, &rdev->flags)) 562 continue; 563 if (!test_bit(In_sync, &rdev->flags) && 564 rdev->recovery_offset < this_sector + sectors) 565 continue; 566 if (test_bit(WriteMostly, &rdev->flags)) { 567 /* Don't balance among write-mostly, just 568 * use the first as a last resort */ 569 if (best_disk < 0) { 570 if (is_badblock(rdev, this_sector, sectors, 571 &first_bad, &bad_sectors)) { 572 if (first_bad < this_sector) 573 /* Cannot use this */ 574 continue; 575 best_good_sectors = first_bad - this_sector; 576 } else 577 best_good_sectors = sectors; 578 best_disk = disk; 579 } 580 continue; 581 } 582 /* This is a reasonable device to use. It might 583 * even be best. 584 */ 585 if (is_badblock(rdev, this_sector, sectors, 586 &first_bad, &bad_sectors)) { 587 if (best_dist < MaxSector) 588 /* already have a better device */ 589 continue; 590 if (first_bad <= this_sector) { 591 /* cannot read here. If this is the 'primary' 592 * device, then we must not read beyond 593 * bad_sectors from another device.. 594 */ 595 bad_sectors -= (this_sector - first_bad); 596 if (choose_first && sectors > bad_sectors) 597 sectors = bad_sectors; 598 if (best_good_sectors > sectors) 599 best_good_sectors = sectors; 600 601 } else { 602 sector_t good_sectors = first_bad - this_sector; 603 if (good_sectors > best_good_sectors) { 604 best_good_sectors = good_sectors; 605 best_disk = disk; 606 } 607 if (choose_first) 608 break; 609 } 610 continue; 611 } else 612 best_good_sectors = sectors; 613 614 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev)); 615 has_nonrot_disk |= nonrot; 616 pending = atomic_read(&rdev->nr_pending); 617 dist = abs(this_sector - conf->mirrors[disk].head_position); 618 if (choose_first) { 619 best_disk = disk; 620 break; 621 } 622 /* Don't change to another disk for sequential reads */ 623 if (conf->mirrors[disk].next_seq_sect == this_sector 624 || dist == 0) { 625 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9; 626 struct raid1_info *mirror = &conf->mirrors[disk]; 627 628 best_disk = disk; 629 /* 630 * If buffered sequential IO size exceeds optimal 631 * iosize, check if there is idle disk. If yes, choose 632 * the idle disk. read_balance could already choose an 633 * idle disk before noticing it's a sequential IO in 634 * this disk. This doesn't matter because this disk 635 * will idle, next time it will be utilized after the 636 * first disk has IO size exceeds optimal iosize. In 637 * this way, iosize of the first disk will be optimal 638 * iosize at least. iosize of the second disk might be 639 * small, but not a big deal since when the second disk 640 * starts IO, the first disk is likely still busy. 641 */ 642 if (nonrot && opt_iosize > 0 && 643 mirror->seq_start != MaxSector && 644 mirror->next_seq_sect > opt_iosize && 645 mirror->next_seq_sect - opt_iosize >= 646 mirror->seq_start) { 647 choose_next_idle = 1; 648 continue; 649 } 650 break; 651 } 652 /* If device is idle, use it */ 653 if (pending == 0) { 654 best_disk = disk; 655 break; 656 } 657 658 if (choose_next_idle) 659 continue; 660 661 if (min_pending > pending) { 662 min_pending = pending; 663 best_pending_disk = disk; 664 } 665 666 if (dist < best_dist) { 667 best_dist = dist; 668 best_dist_disk = disk; 669 } 670 } 671 672 /* 673 * If all disks are rotational, choose the closest disk. If any disk is 674 * non-rotational, choose the disk with less pending request even the 675 * disk is rotational, which might/might not be optimal for raids with 676 * mixed ratation/non-rotational disks depending on workload. 677 */ 678 if (best_disk == -1) { 679 if (has_nonrot_disk) 680 best_disk = best_pending_disk; 681 else 682 best_disk = best_dist_disk; 683 } 684 685 if (best_disk >= 0) { 686 rdev = rcu_dereference(conf->mirrors[best_disk].rdev); 687 if (!rdev) 688 goto retry; 689 atomic_inc(&rdev->nr_pending); 690 if (test_bit(Faulty, &rdev->flags)) { 691 /* cannot risk returning a device that failed 692 * before we inc'ed nr_pending 693 */ 694 rdev_dec_pending(rdev, conf->mddev); 695 goto retry; 696 } 697 sectors = best_good_sectors; 698 699 if (conf->mirrors[best_disk].next_seq_sect != this_sector) 700 conf->mirrors[best_disk].seq_start = this_sector; 701 702 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors; 703 } 704 rcu_read_unlock(); 705 *max_sectors = sectors; 706 707 return best_disk; 708 } 709 710 static int raid1_mergeable_bvec(struct mddev *mddev, 711 struct bvec_merge_data *bvm, 712 struct bio_vec *biovec) 713 { 714 struct r1conf *conf = mddev->private; 715 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev); 716 int max = biovec->bv_len; 717 718 if (mddev->merge_check_needed) { 719 int disk; 720 rcu_read_lock(); 721 for (disk = 0; disk < conf->raid_disks * 2; disk++) { 722 struct md_rdev *rdev = rcu_dereference( 723 conf->mirrors[disk].rdev); 724 if (rdev && !test_bit(Faulty, &rdev->flags)) { 725 struct request_queue *q = 726 bdev_get_queue(rdev->bdev); 727 if (q->merge_bvec_fn) { 728 bvm->bi_sector = sector + 729 rdev->data_offset; 730 bvm->bi_bdev = rdev->bdev; 731 max = min(max, q->merge_bvec_fn( 732 q, bvm, biovec)); 733 } 734 } 735 } 736 rcu_read_unlock(); 737 } 738 return max; 739 740 } 741 742 static int raid1_congested(struct mddev *mddev, int bits) 743 { 744 struct r1conf *conf = mddev->private; 745 int i, ret = 0; 746 747 if ((bits & (1 << BDI_async_congested)) && 748 conf->pending_count >= max_queued_requests) 749 return 1; 750 751 rcu_read_lock(); 752 for (i = 0; i < conf->raid_disks * 2; i++) { 753 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 754 if (rdev && !test_bit(Faulty, &rdev->flags)) { 755 struct request_queue *q = bdev_get_queue(rdev->bdev); 756 757 BUG_ON(!q); 758 759 /* Note the '|| 1' - when read_balance prefers 760 * non-congested targets, it can be removed 761 */ 762 if ((bits & (1<<BDI_async_congested)) || 1) 763 ret |= bdi_congested(&q->backing_dev_info, bits); 764 else 765 ret &= bdi_congested(&q->backing_dev_info, bits); 766 } 767 } 768 rcu_read_unlock(); 769 return ret; 770 } 771 772 static void flush_pending_writes(struct r1conf *conf) 773 { 774 /* Any writes that have been queued but are awaiting 775 * bitmap updates get flushed here. 776 */ 777 spin_lock_irq(&conf->device_lock); 778 779 if (conf->pending_bio_list.head) { 780 struct bio *bio; 781 bio = bio_list_get(&conf->pending_bio_list); 782 conf->pending_count = 0; 783 spin_unlock_irq(&conf->device_lock); 784 /* flush any pending bitmap writes to 785 * disk before proceeding w/ I/O */ 786 bitmap_unplug(conf->mddev->bitmap); 787 wake_up(&conf->wait_barrier); 788 789 while (bio) { /* submit pending writes */ 790 struct bio *next = bio->bi_next; 791 bio->bi_next = NULL; 792 if (unlikely((bio->bi_rw & REQ_DISCARD) && 793 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 794 /* Just ignore it */ 795 bio_endio(bio, 0); 796 else 797 generic_make_request(bio); 798 bio = next; 799 } 800 } else 801 spin_unlock_irq(&conf->device_lock); 802 } 803 804 /* Barriers.... 805 * Sometimes we need to suspend IO while we do something else, 806 * either some resync/recovery, or reconfigure the array. 807 * To do this we raise a 'barrier'. 808 * The 'barrier' is a counter that can be raised multiple times 809 * to count how many activities are happening which preclude 810 * normal IO. 811 * We can only raise the barrier if there is no pending IO. 812 * i.e. if nr_pending == 0. 813 * We choose only to raise the barrier if no-one is waiting for the 814 * barrier to go down. This means that as soon as an IO request 815 * is ready, no other operations which require a barrier will start 816 * until the IO request has had a chance. 817 * 818 * So: regular IO calls 'wait_barrier'. When that returns there 819 * is no backgroup IO happening, It must arrange to call 820 * allow_barrier when it has finished its IO. 821 * backgroup IO calls must call raise_barrier. Once that returns 822 * there is no normal IO happeing. It must arrange to call 823 * lower_barrier when the particular background IO completes. 824 */ 825 static void raise_barrier(struct r1conf *conf, sector_t sector_nr) 826 { 827 spin_lock_irq(&conf->resync_lock); 828 829 /* Wait until no block IO is waiting */ 830 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting, 831 conf->resync_lock); 832 833 /* block any new IO from starting */ 834 conf->barrier++; 835 conf->next_resync = sector_nr; 836 837 /* For these conditions we must wait: 838 * A: while the array is in frozen state 839 * B: while barrier >= RESYNC_DEPTH, meaning resync reach 840 * the max count which allowed. 841 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning 842 * next resync will reach to the window which normal bios are 843 * handling. 844 * D: while there are any active requests in the current window. 845 */ 846 wait_event_lock_irq(conf->wait_barrier, 847 !conf->array_frozen && 848 conf->barrier < RESYNC_DEPTH && 849 conf->current_window_requests == 0 && 850 (conf->start_next_window >= 851 conf->next_resync + RESYNC_SECTORS), 852 conf->resync_lock); 853 854 conf->nr_pending++; 855 spin_unlock_irq(&conf->resync_lock); 856 } 857 858 static void lower_barrier(struct r1conf *conf) 859 { 860 unsigned long flags; 861 BUG_ON(conf->barrier <= 0); 862 spin_lock_irqsave(&conf->resync_lock, flags); 863 conf->barrier--; 864 conf->nr_pending--; 865 spin_unlock_irqrestore(&conf->resync_lock, flags); 866 wake_up(&conf->wait_barrier); 867 } 868 869 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio) 870 { 871 bool wait = false; 872 873 if (conf->array_frozen || !bio) 874 wait = true; 875 else if (conf->barrier && bio_data_dir(bio) == WRITE) { 876 if ((conf->mddev->curr_resync_completed 877 >= bio_end_sector(bio)) || 878 (conf->next_resync + NEXT_NORMALIO_DISTANCE 879 <= bio->bi_iter.bi_sector)) 880 wait = false; 881 else 882 wait = true; 883 } 884 885 return wait; 886 } 887 888 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio) 889 { 890 sector_t sector = 0; 891 892 spin_lock_irq(&conf->resync_lock); 893 if (need_to_wait_for_sync(conf, bio)) { 894 conf->nr_waiting++; 895 /* Wait for the barrier to drop. 896 * However if there are already pending 897 * requests (preventing the barrier from 898 * rising completely), and the 899 * per-process bio queue isn't empty, 900 * then don't wait, as we need to empty 901 * that queue to allow conf->start_next_window 902 * to increase. 903 */ 904 wait_event_lock_irq(conf->wait_barrier, 905 !conf->array_frozen && 906 (!conf->barrier || 907 ((conf->start_next_window < 908 conf->next_resync + RESYNC_SECTORS) && 909 current->bio_list && 910 !bio_list_empty(current->bio_list))), 911 conf->resync_lock); 912 conf->nr_waiting--; 913 } 914 915 if (bio && bio_data_dir(bio) == WRITE) { 916 if (bio->bi_iter.bi_sector >= 917 conf->mddev->curr_resync_completed) { 918 if (conf->start_next_window == MaxSector) 919 conf->start_next_window = 920 conf->next_resync + 921 NEXT_NORMALIO_DISTANCE; 922 923 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE) 924 <= bio->bi_iter.bi_sector) 925 conf->next_window_requests++; 926 else 927 conf->current_window_requests++; 928 sector = conf->start_next_window; 929 } 930 } 931 932 conf->nr_pending++; 933 spin_unlock_irq(&conf->resync_lock); 934 return sector; 935 } 936 937 static void allow_barrier(struct r1conf *conf, sector_t start_next_window, 938 sector_t bi_sector) 939 { 940 unsigned long flags; 941 942 spin_lock_irqsave(&conf->resync_lock, flags); 943 conf->nr_pending--; 944 if (start_next_window) { 945 if (start_next_window == conf->start_next_window) { 946 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE 947 <= bi_sector) 948 conf->next_window_requests--; 949 else 950 conf->current_window_requests--; 951 } else 952 conf->current_window_requests--; 953 954 if (!conf->current_window_requests) { 955 if (conf->next_window_requests) { 956 conf->current_window_requests = 957 conf->next_window_requests; 958 conf->next_window_requests = 0; 959 conf->start_next_window += 960 NEXT_NORMALIO_DISTANCE; 961 } else 962 conf->start_next_window = MaxSector; 963 } 964 } 965 spin_unlock_irqrestore(&conf->resync_lock, flags); 966 wake_up(&conf->wait_barrier); 967 } 968 969 static void freeze_array(struct r1conf *conf, int extra) 970 { 971 /* stop syncio and normal IO and wait for everything to 972 * go quite. 973 * We wait until nr_pending match nr_queued+extra 974 * This is called in the context of one normal IO request 975 * that has failed. Thus any sync request that might be pending 976 * will be blocked by nr_pending, and we need to wait for 977 * pending IO requests to complete or be queued for re-try. 978 * Thus the number queued (nr_queued) plus this request (extra) 979 * must match the number of pending IOs (nr_pending) before 980 * we continue. 981 */ 982 spin_lock_irq(&conf->resync_lock); 983 conf->array_frozen = 1; 984 wait_event_lock_irq_cmd(conf->wait_barrier, 985 conf->nr_pending == conf->nr_queued+extra, 986 conf->resync_lock, 987 flush_pending_writes(conf)); 988 spin_unlock_irq(&conf->resync_lock); 989 } 990 static void unfreeze_array(struct r1conf *conf) 991 { 992 /* reverse the effect of the freeze */ 993 spin_lock_irq(&conf->resync_lock); 994 conf->array_frozen = 0; 995 wake_up(&conf->wait_barrier); 996 spin_unlock_irq(&conf->resync_lock); 997 } 998 999 /* duplicate the data pages for behind I/O 1000 */ 1001 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio) 1002 { 1003 int i; 1004 struct bio_vec *bvec; 1005 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec), 1006 GFP_NOIO); 1007 if (unlikely(!bvecs)) 1008 return; 1009 1010 bio_for_each_segment_all(bvec, bio, i) { 1011 bvecs[i] = *bvec; 1012 bvecs[i].bv_page = alloc_page(GFP_NOIO); 1013 if (unlikely(!bvecs[i].bv_page)) 1014 goto do_sync_io; 1015 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset, 1016 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len); 1017 kunmap(bvecs[i].bv_page); 1018 kunmap(bvec->bv_page); 1019 } 1020 r1_bio->behind_bvecs = bvecs; 1021 r1_bio->behind_page_count = bio->bi_vcnt; 1022 set_bit(R1BIO_BehindIO, &r1_bio->state); 1023 return; 1024 1025 do_sync_io: 1026 for (i = 0; i < bio->bi_vcnt; i++) 1027 if (bvecs[i].bv_page) 1028 put_page(bvecs[i].bv_page); 1029 kfree(bvecs); 1030 pr_debug("%dB behind alloc failed, doing sync I/O\n", 1031 bio->bi_iter.bi_size); 1032 } 1033 1034 struct raid1_plug_cb { 1035 struct blk_plug_cb cb; 1036 struct bio_list pending; 1037 int pending_cnt; 1038 }; 1039 1040 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule) 1041 { 1042 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, 1043 cb); 1044 struct mddev *mddev = plug->cb.data; 1045 struct r1conf *conf = mddev->private; 1046 struct bio *bio; 1047 1048 if (from_schedule || current->bio_list) { 1049 spin_lock_irq(&conf->device_lock); 1050 bio_list_merge(&conf->pending_bio_list, &plug->pending); 1051 conf->pending_count += plug->pending_cnt; 1052 spin_unlock_irq(&conf->device_lock); 1053 wake_up(&conf->wait_barrier); 1054 md_wakeup_thread(mddev->thread); 1055 kfree(plug); 1056 return; 1057 } 1058 1059 /* we aren't scheduling, so we can do the write-out directly. */ 1060 bio = bio_list_get(&plug->pending); 1061 bitmap_unplug(mddev->bitmap); 1062 wake_up(&conf->wait_barrier); 1063 1064 while (bio) { /* submit pending writes */ 1065 struct bio *next = bio->bi_next; 1066 bio->bi_next = NULL; 1067 if (unlikely((bio->bi_rw & REQ_DISCARD) && 1068 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 1069 /* Just ignore it */ 1070 bio_endio(bio, 0); 1071 else 1072 generic_make_request(bio); 1073 bio = next; 1074 } 1075 kfree(plug); 1076 } 1077 1078 static void make_request(struct mddev *mddev, struct bio * bio) 1079 { 1080 struct r1conf *conf = mddev->private; 1081 struct raid1_info *mirror; 1082 struct r1bio *r1_bio; 1083 struct bio *read_bio; 1084 int i, disks; 1085 struct bitmap *bitmap; 1086 unsigned long flags; 1087 const int rw = bio_data_dir(bio); 1088 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC); 1089 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA)); 1090 const unsigned long do_discard = (bio->bi_rw 1091 & (REQ_DISCARD | REQ_SECURE)); 1092 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME); 1093 struct md_rdev *blocked_rdev; 1094 struct blk_plug_cb *cb; 1095 struct raid1_plug_cb *plug = NULL; 1096 int first_clone; 1097 int sectors_handled; 1098 int max_sectors; 1099 sector_t start_next_window; 1100 1101 /* 1102 * Register the new request and wait if the reconstruction 1103 * thread has put up a bar for new requests. 1104 * Continue immediately if no resync is active currently. 1105 */ 1106 1107 md_write_start(mddev, bio); /* wait on superblock update early */ 1108 1109 if (bio_data_dir(bio) == WRITE && 1110 ((bio_end_sector(bio) > mddev->suspend_lo && 1111 bio->bi_iter.bi_sector < mddev->suspend_hi) || 1112 (mddev_is_clustered(mddev) && 1113 md_cluster_ops->area_resyncing(mddev, bio->bi_iter.bi_sector, bio_end_sector(bio))))) { 1114 /* As the suspend_* range is controlled by 1115 * userspace, we want an interruptible 1116 * wait. 1117 */ 1118 DEFINE_WAIT(w); 1119 for (;;) { 1120 flush_signals(current); 1121 prepare_to_wait(&conf->wait_barrier, 1122 &w, TASK_INTERRUPTIBLE); 1123 if (bio_end_sector(bio) <= mddev->suspend_lo || 1124 bio->bi_iter.bi_sector >= mddev->suspend_hi || 1125 (mddev_is_clustered(mddev) && 1126 !md_cluster_ops->area_resyncing(mddev, 1127 bio->bi_iter.bi_sector, bio_end_sector(bio)))) 1128 break; 1129 schedule(); 1130 } 1131 finish_wait(&conf->wait_barrier, &w); 1132 } 1133 1134 start_next_window = wait_barrier(conf, bio); 1135 1136 bitmap = mddev->bitmap; 1137 1138 /* 1139 * make_request() can abort the operation when READA is being 1140 * used and no empty request is available. 1141 * 1142 */ 1143 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1144 1145 r1_bio->master_bio = bio; 1146 r1_bio->sectors = bio_sectors(bio); 1147 r1_bio->state = 0; 1148 r1_bio->mddev = mddev; 1149 r1_bio->sector = bio->bi_iter.bi_sector; 1150 1151 /* We might need to issue multiple reads to different 1152 * devices if there are bad blocks around, so we keep 1153 * track of the number of reads in bio->bi_phys_segments. 1154 * If this is 0, there is only one r1_bio and no locking 1155 * will be needed when requests complete. If it is 1156 * non-zero, then it is the number of not-completed requests. 1157 */ 1158 bio->bi_phys_segments = 0; 1159 clear_bit(BIO_SEG_VALID, &bio->bi_flags); 1160 1161 if (rw == READ) { 1162 /* 1163 * read balancing logic: 1164 */ 1165 int rdisk; 1166 1167 read_again: 1168 rdisk = read_balance(conf, r1_bio, &max_sectors); 1169 1170 if (rdisk < 0) { 1171 /* couldn't find anywhere to read from */ 1172 raid_end_bio_io(r1_bio); 1173 return; 1174 } 1175 mirror = conf->mirrors + rdisk; 1176 1177 if (test_bit(WriteMostly, &mirror->rdev->flags) && 1178 bitmap) { 1179 /* Reading from a write-mostly device must 1180 * take care not to over-take any writes 1181 * that are 'behind' 1182 */ 1183 wait_event(bitmap->behind_wait, 1184 atomic_read(&bitmap->behind_writes) == 0); 1185 } 1186 r1_bio->read_disk = rdisk; 1187 r1_bio->start_next_window = 0; 1188 1189 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1190 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector, 1191 max_sectors); 1192 1193 r1_bio->bios[rdisk] = read_bio; 1194 1195 read_bio->bi_iter.bi_sector = r1_bio->sector + 1196 mirror->rdev->data_offset; 1197 read_bio->bi_bdev = mirror->rdev->bdev; 1198 read_bio->bi_end_io = raid1_end_read_request; 1199 read_bio->bi_rw = READ | do_sync; 1200 read_bio->bi_private = r1_bio; 1201 1202 if (max_sectors < r1_bio->sectors) { 1203 /* could not read all from this device, so we will 1204 * need another r1_bio. 1205 */ 1206 1207 sectors_handled = (r1_bio->sector + max_sectors 1208 - bio->bi_iter.bi_sector); 1209 r1_bio->sectors = max_sectors; 1210 spin_lock_irq(&conf->device_lock); 1211 if (bio->bi_phys_segments == 0) 1212 bio->bi_phys_segments = 2; 1213 else 1214 bio->bi_phys_segments++; 1215 spin_unlock_irq(&conf->device_lock); 1216 /* Cannot call generic_make_request directly 1217 * as that will be queued in __make_request 1218 * and subsequent mempool_alloc might block waiting 1219 * for it. So hand bio over to raid1d. 1220 */ 1221 reschedule_retry(r1_bio); 1222 1223 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1224 1225 r1_bio->master_bio = bio; 1226 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1227 r1_bio->state = 0; 1228 r1_bio->mddev = mddev; 1229 r1_bio->sector = bio->bi_iter.bi_sector + 1230 sectors_handled; 1231 goto read_again; 1232 } else 1233 generic_make_request(read_bio); 1234 return; 1235 } 1236 1237 /* 1238 * WRITE: 1239 */ 1240 if (conf->pending_count >= max_queued_requests) { 1241 md_wakeup_thread(mddev->thread); 1242 wait_event(conf->wait_barrier, 1243 conf->pending_count < max_queued_requests); 1244 } 1245 /* first select target devices under rcu_lock and 1246 * inc refcount on their rdev. Record them by setting 1247 * bios[x] to bio 1248 * If there are known/acknowledged bad blocks on any device on 1249 * which we have seen a write error, we want to avoid writing those 1250 * blocks. 1251 * This potentially requires several writes to write around 1252 * the bad blocks. Each set of writes gets it's own r1bio 1253 * with a set of bios attached. 1254 */ 1255 1256 disks = conf->raid_disks * 2; 1257 retry_write: 1258 r1_bio->start_next_window = start_next_window; 1259 blocked_rdev = NULL; 1260 rcu_read_lock(); 1261 max_sectors = r1_bio->sectors; 1262 for (i = 0; i < disks; i++) { 1263 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1264 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 1265 atomic_inc(&rdev->nr_pending); 1266 blocked_rdev = rdev; 1267 break; 1268 } 1269 r1_bio->bios[i] = NULL; 1270 if (!rdev || test_bit(Faulty, &rdev->flags) 1271 || test_bit(Unmerged, &rdev->flags)) { 1272 if (i < conf->raid_disks) 1273 set_bit(R1BIO_Degraded, &r1_bio->state); 1274 continue; 1275 } 1276 1277 atomic_inc(&rdev->nr_pending); 1278 if (test_bit(WriteErrorSeen, &rdev->flags)) { 1279 sector_t first_bad; 1280 int bad_sectors; 1281 int is_bad; 1282 1283 is_bad = is_badblock(rdev, r1_bio->sector, 1284 max_sectors, 1285 &first_bad, &bad_sectors); 1286 if (is_bad < 0) { 1287 /* mustn't write here until the bad block is 1288 * acknowledged*/ 1289 set_bit(BlockedBadBlocks, &rdev->flags); 1290 blocked_rdev = rdev; 1291 break; 1292 } 1293 if (is_bad && first_bad <= r1_bio->sector) { 1294 /* Cannot write here at all */ 1295 bad_sectors -= (r1_bio->sector - first_bad); 1296 if (bad_sectors < max_sectors) 1297 /* mustn't write more than bad_sectors 1298 * to other devices yet 1299 */ 1300 max_sectors = bad_sectors; 1301 rdev_dec_pending(rdev, mddev); 1302 /* We don't set R1BIO_Degraded as that 1303 * only applies if the disk is 1304 * missing, so it might be re-added, 1305 * and we want to know to recover this 1306 * chunk. 1307 * In this case the device is here, 1308 * and the fact that this chunk is not 1309 * in-sync is recorded in the bad 1310 * block log 1311 */ 1312 continue; 1313 } 1314 if (is_bad) { 1315 int good_sectors = first_bad - r1_bio->sector; 1316 if (good_sectors < max_sectors) 1317 max_sectors = good_sectors; 1318 } 1319 } 1320 r1_bio->bios[i] = bio; 1321 } 1322 rcu_read_unlock(); 1323 1324 if (unlikely(blocked_rdev)) { 1325 /* Wait for this device to become unblocked */ 1326 int j; 1327 sector_t old = start_next_window; 1328 1329 for (j = 0; j < i; j++) 1330 if (r1_bio->bios[j]) 1331 rdev_dec_pending(conf->mirrors[j].rdev, mddev); 1332 r1_bio->state = 0; 1333 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector); 1334 md_wait_for_blocked_rdev(blocked_rdev, mddev); 1335 start_next_window = wait_barrier(conf, bio); 1336 /* 1337 * We must make sure the multi r1bios of bio have 1338 * the same value of bi_phys_segments 1339 */ 1340 if (bio->bi_phys_segments && old && 1341 old != start_next_window) 1342 /* Wait for the former r1bio(s) to complete */ 1343 wait_event(conf->wait_barrier, 1344 bio->bi_phys_segments == 1); 1345 goto retry_write; 1346 } 1347 1348 if (max_sectors < r1_bio->sectors) { 1349 /* We are splitting this write into multiple parts, so 1350 * we need to prepare for allocating another r1_bio. 1351 */ 1352 r1_bio->sectors = max_sectors; 1353 spin_lock_irq(&conf->device_lock); 1354 if (bio->bi_phys_segments == 0) 1355 bio->bi_phys_segments = 2; 1356 else 1357 bio->bi_phys_segments++; 1358 spin_unlock_irq(&conf->device_lock); 1359 } 1360 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector; 1361 1362 atomic_set(&r1_bio->remaining, 1); 1363 atomic_set(&r1_bio->behind_remaining, 0); 1364 1365 first_clone = 1; 1366 for (i = 0; i < disks; i++) { 1367 struct bio *mbio; 1368 if (!r1_bio->bios[i]) 1369 continue; 1370 1371 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1372 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors); 1373 1374 if (first_clone) { 1375 /* do behind I/O ? 1376 * Not if there are too many, or cannot 1377 * allocate memory, or a reader on WriteMostly 1378 * is waiting for behind writes to flush */ 1379 if (bitmap && 1380 (atomic_read(&bitmap->behind_writes) 1381 < mddev->bitmap_info.max_write_behind) && 1382 !waitqueue_active(&bitmap->behind_wait)) 1383 alloc_behind_pages(mbio, r1_bio); 1384 1385 bitmap_startwrite(bitmap, r1_bio->sector, 1386 r1_bio->sectors, 1387 test_bit(R1BIO_BehindIO, 1388 &r1_bio->state)); 1389 first_clone = 0; 1390 } 1391 if (r1_bio->behind_bvecs) { 1392 struct bio_vec *bvec; 1393 int j; 1394 1395 /* 1396 * We trimmed the bio, so _all is legit 1397 */ 1398 bio_for_each_segment_all(bvec, mbio, j) 1399 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page; 1400 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags)) 1401 atomic_inc(&r1_bio->behind_remaining); 1402 } 1403 1404 r1_bio->bios[i] = mbio; 1405 1406 mbio->bi_iter.bi_sector = (r1_bio->sector + 1407 conf->mirrors[i].rdev->data_offset); 1408 mbio->bi_bdev = conf->mirrors[i].rdev->bdev; 1409 mbio->bi_end_io = raid1_end_write_request; 1410 mbio->bi_rw = 1411 WRITE | do_flush_fua | do_sync | do_discard | do_same; 1412 mbio->bi_private = r1_bio; 1413 1414 atomic_inc(&r1_bio->remaining); 1415 1416 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug)); 1417 if (cb) 1418 plug = container_of(cb, struct raid1_plug_cb, cb); 1419 else 1420 plug = NULL; 1421 spin_lock_irqsave(&conf->device_lock, flags); 1422 if (plug) { 1423 bio_list_add(&plug->pending, mbio); 1424 plug->pending_cnt++; 1425 } else { 1426 bio_list_add(&conf->pending_bio_list, mbio); 1427 conf->pending_count++; 1428 } 1429 spin_unlock_irqrestore(&conf->device_lock, flags); 1430 if (!plug) 1431 md_wakeup_thread(mddev->thread); 1432 } 1433 /* Mustn't call r1_bio_write_done before this next test, 1434 * as it could result in the bio being freed. 1435 */ 1436 if (sectors_handled < bio_sectors(bio)) { 1437 r1_bio_write_done(r1_bio); 1438 /* We need another r1_bio. It has already been counted 1439 * in bio->bi_phys_segments 1440 */ 1441 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1442 r1_bio->master_bio = bio; 1443 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1444 r1_bio->state = 0; 1445 r1_bio->mddev = mddev; 1446 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled; 1447 goto retry_write; 1448 } 1449 1450 r1_bio_write_done(r1_bio); 1451 1452 /* In case raid1d snuck in to freeze_array */ 1453 wake_up(&conf->wait_barrier); 1454 } 1455 1456 static void status(struct seq_file *seq, struct mddev *mddev) 1457 { 1458 struct r1conf *conf = mddev->private; 1459 int i; 1460 1461 seq_printf(seq, " [%d/%d] [", conf->raid_disks, 1462 conf->raid_disks - mddev->degraded); 1463 rcu_read_lock(); 1464 for (i = 0; i < conf->raid_disks; i++) { 1465 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1466 seq_printf(seq, "%s", 1467 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_"); 1468 } 1469 rcu_read_unlock(); 1470 seq_printf(seq, "]"); 1471 } 1472 1473 static void error(struct mddev *mddev, struct md_rdev *rdev) 1474 { 1475 char b[BDEVNAME_SIZE]; 1476 struct r1conf *conf = mddev->private; 1477 1478 /* 1479 * If it is not operational, then we have already marked it as dead 1480 * else if it is the last working disks, ignore the error, let the 1481 * next level up know. 1482 * else mark the drive as failed 1483 */ 1484 if (test_bit(In_sync, &rdev->flags) 1485 && (conf->raid_disks - mddev->degraded) == 1) { 1486 /* 1487 * Don't fail the drive, act as though we were just a 1488 * normal single drive. 1489 * However don't try a recovery from this drive as 1490 * it is very likely to fail. 1491 */ 1492 conf->recovery_disabled = mddev->recovery_disabled; 1493 return; 1494 } 1495 set_bit(Blocked, &rdev->flags); 1496 if (test_and_clear_bit(In_sync, &rdev->flags)) { 1497 unsigned long flags; 1498 spin_lock_irqsave(&conf->device_lock, flags); 1499 mddev->degraded++; 1500 set_bit(Faulty, &rdev->flags); 1501 spin_unlock_irqrestore(&conf->device_lock, flags); 1502 } else 1503 set_bit(Faulty, &rdev->flags); 1504 /* 1505 * if recovery is running, make sure it aborts. 1506 */ 1507 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1508 set_bit(MD_CHANGE_DEVS, &mddev->flags); 1509 printk(KERN_ALERT 1510 "md/raid1:%s: Disk failure on %s, disabling device.\n" 1511 "md/raid1:%s: Operation continuing on %d devices.\n", 1512 mdname(mddev), bdevname(rdev->bdev, b), 1513 mdname(mddev), conf->raid_disks - mddev->degraded); 1514 } 1515 1516 static void print_conf(struct r1conf *conf) 1517 { 1518 int i; 1519 1520 printk(KERN_DEBUG "RAID1 conf printout:\n"); 1521 if (!conf) { 1522 printk(KERN_DEBUG "(!conf)\n"); 1523 return; 1524 } 1525 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded, 1526 conf->raid_disks); 1527 1528 rcu_read_lock(); 1529 for (i = 0; i < conf->raid_disks; i++) { 1530 char b[BDEVNAME_SIZE]; 1531 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1532 if (rdev) 1533 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n", 1534 i, !test_bit(In_sync, &rdev->flags), 1535 !test_bit(Faulty, &rdev->flags), 1536 bdevname(rdev->bdev,b)); 1537 } 1538 rcu_read_unlock(); 1539 } 1540 1541 static void close_sync(struct r1conf *conf) 1542 { 1543 wait_barrier(conf, NULL); 1544 allow_barrier(conf, 0, 0); 1545 1546 mempool_destroy(conf->r1buf_pool); 1547 conf->r1buf_pool = NULL; 1548 1549 spin_lock_irq(&conf->resync_lock); 1550 conf->next_resync = 0; 1551 conf->start_next_window = MaxSector; 1552 conf->current_window_requests += 1553 conf->next_window_requests; 1554 conf->next_window_requests = 0; 1555 spin_unlock_irq(&conf->resync_lock); 1556 } 1557 1558 static int raid1_spare_active(struct mddev *mddev) 1559 { 1560 int i; 1561 struct r1conf *conf = mddev->private; 1562 int count = 0; 1563 unsigned long flags; 1564 1565 /* 1566 * Find all failed disks within the RAID1 configuration 1567 * and mark them readable. 1568 * Called under mddev lock, so rcu protection not needed. 1569 */ 1570 for (i = 0; i < conf->raid_disks; i++) { 1571 struct md_rdev *rdev = conf->mirrors[i].rdev; 1572 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev; 1573 if (repl 1574 && !test_bit(Candidate, &repl->flags) 1575 && repl->recovery_offset == MaxSector 1576 && !test_bit(Faulty, &repl->flags) 1577 && !test_and_set_bit(In_sync, &repl->flags)) { 1578 /* replacement has just become active */ 1579 if (!rdev || 1580 !test_and_clear_bit(In_sync, &rdev->flags)) 1581 count++; 1582 if (rdev) { 1583 /* Replaced device not technically 1584 * faulty, but we need to be sure 1585 * it gets removed and never re-added 1586 */ 1587 set_bit(Faulty, &rdev->flags); 1588 sysfs_notify_dirent_safe( 1589 rdev->sysfs_state); 1590 } 1591 } 1592 if (rdev 1593 && rdev->recovery_offset == MaxSector 1594 && !test_bit(Faulty, &rdev->flags) 1595 && !test_and_set_bit(In_sync, &rdev->flags)) { 1596 count++; 1597 sysfs_notify_dirent_safe(rdev->sysfs_state); 1598 } 1599 } 1600 spin_lock_irqsave(&conf->device_lock, flags); 1601 mddev->degraded -= count; 1602 spin_unlock_irqrestore(&conf->device_lock, flags); 1603 1604 print_conf(conf); 1605 return count; 1606 } 1607 1608 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev) 1609 { 1610 struct r1conf *conf = mddev->private; 1611 int err = -EEXIST; 1612 int mirror = 0; 1613 struct raid1_info *p; 1614 int first = 0; 1615 int last = conf->raid_disks - 1; 1616 struct request_queue *q = bdev_get_queue(rdev->bdev); 1617 1618 if (mddev->recovery_disabled == conf->recovery_disabled) 1619 return -EBUSY; 1620 1621 if (rdev->raid_disk >= 0) 1622 first = last = rdev->raid_disk; 1623 1624 if (q->merge_bvec_fn) { 1625 set_bit(Unmerged, &rdev->flags); 1626 mddev->merge_check_needed = 1; 1627 } 1628 1629 for (mirror = first; mirror <= last; mirror++) { 1630 p = conf->mirrors+mirror; 1631 if (!p->rdev) { 1632 1633 if (mddev->gendisk) 1634 disk_stack_limits(mddev->gendisk, rdev->bdev, 1635 rdev->data_offset << 9); 1636 1637 p->head_position = 0; 1638 rdev->raid_disk = mirror; 1639 err = 0; 1640 /* As all devices are equivalent, we don't need a full recovery 1641 * if this was recently any drive of the array 1642 */ 1643 if (rdev->saved_raid_disk < 0) 1644 conf->fullsync = 1; 1645 rcu_assign_pointer(p->rdev, rdev); 1646 break; 1647 } 1648 if (test_bit(WantReplacement, &p->rdev->flags) && 1649 p[conf->raid_disks].rdev == NULL) { 1650 /* Add this device as a replacement */ 1651 clear_bit(In_sync, &rdev->flags); 1652 set_bit(Replacement, &rdev->flags); 1653 rdev->raid_disk = mirror; 1654 err = 0; 1655 conf->fullsync = 1; 1656 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev); 1657 break; 1658 } 1659 } 1660 if (err == 0 && test_bit(Unmerged, &rdev->flags)) { 1661 /* Some requests might not have seen this new 1662 * merge_bvec_fn. We must wait for them to complete 1663 * before merging the device fully. 1664 * First we make sure any code which has tested 1665 * our function has submitted the request, then 1666 * we wait for all outstanding requests to complete. 1667 */ 1668 synchronize_sched(); 1669 freeze_array(conf, 0); 1670 unfreeze_array(conf); 1671 clear_bit(Unmerged, &rdev->flags); 1672 } 1673 md_integrity_add_rdev(rdev, mddev); 1674 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev))) 1675 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue); 1676 print_conf(conf); 1677 return err; 1678 } 1679 1680 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev) 1681 { 1682 struct r1conf *conf = mddev->private; 1683 int err = 0; 1684 int number = rdev->raid_disk; 1685 struct raid1_info *p = conf->mirrors + number; 1686 1687 if (rdev != p->rdev) 1688 p = conf->mirrors + conf->raid_disks + number; 1689 1690 print_conf(conf); 1691 if (rdev == p->rdev) { 1692 if (test_bit(In_sync, &rdev->flags) || 1693 atomic_read(&rdev->nr_pending)) { 1694 err = -EBUSY; 1695 goto abort; 1696 } 1697 /* Only remove non-faulty devices if recovery 1698 * is not possible. 1699 */ 1700 if (!test_bit(Faulty, &rdev->flags) && 1701 mddev->recovery_disabled != conf->recovery_disabled && 1702 mddev->degraded < conf->raid_disks) { 1703 err = -EBUSY; 1704 goto abort; 1705 } 1706 p->rdev = NULL; 1707 synchronize_rcu(); 1708 if (atomic_read(&rdev->nr_pending)) { 1709 /* lost the race, try later */ 1710 err = -EBUSY; 1711 p->rdev = rdev; 1712 goto abort; 1713 } else if (conf->mirrors[conf->raid_disks + number].rdev) { 1714 /* We just removed a device that is being replaced. 1715 * Move down the replacement. We drain all IO before 1716 * doing this to avoid confusion. 1717 */ 1718 struct md_rdev *repl = 1719 conf->mirrors[conf->raid_disks + number].rdev; 1720 freeze_array(conf, 0); 1721 clear_bit(Replacement, &repl->flags); 1722 p->rdev = repl; 1723 conf->mirrors[conf->raid_disks + number].rdev = NULL; 1724 unfreeze_array(conf); 1725 clear_bit(WantReplacement, &rdev->flags); 1726 } else 1727 clear_bit(WantReplacement, &rdev->flags); 1728 err = md_integrity_register(mddev); 1729 } 1730 abort: 1731 1732 print_conf(conf); 1733 return err; 1734 } 1735 1736 static void end_sync_read(struct bio *bio, int error) 1737 { 1738 struct r1bio *r1_bio = bio->bi_private; 1739 1740 update_head_pos(r1_bio->read_disk, r1_bio); 1741 1742 /* 1743 * we have read a block, now it needs to be re-written, 1744 * or re-read if the read failed. 1745 * We don't do much here, just schedule handling by raid1d 1746 */ 1747 if (test_bit(BIO_UPTODATE, &bio->bi_flags)) 1748 set_bit(R1BIO_Uptodate, &r1_bio->state); 1749 1750 if (atomic_dec_and_test(&r1_bio->remaining)) 1751 reschedule_retry(r1_bio); 1752 } 1753 1754 static void end_sync_write(struct bio *bio, int error) 1755 { 1756 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 1757 struct r1bio *r1_bio = bio->bi_private; 1758 struct mddev *mddev = r1_bio->mddev; 1759 struct r1conf *conf = mddev->private; 1760 int mirror=0; 1761 sector_t first_bad; 1762 int bad_sectors; 1763 1764 mirror = find_bio_disk(r1_bio, bio); 1765 1766 if (!uptodate) { 1767 sector_t sync_blocks = 0; 1768 sector_t s = r1_bio->sector; 1769 long sectors_to_go = r1_bio->sectors; 1770 /* make sure these bits doesn't get cleared. */ 1771 do { 1772 bitmap_end_sync(mddev->bitmap, s, 1773 &sync_blocks, 1); 1774 s += sync_blocks; 1775 sectors_to_go -= sync_blocks; 1776 } while (sectors_to_go > 0); 1777 set_bit(WriteErrorSeen, 1778 &conf->mirrors[mirror].rdev->flags); 1779 if (!test_and_set_bit(WantReplacement, 1780 &conf->mirrors[mirror].rdev->flags)) 1781 set_bit(MD_RECOVERY_NEEDED, & 1782 mddev->recovery); 1783 set_bit(R1BIO_WriteError, &r1_bio->state); 1784 } else if (is_badblock(conf->mirrors[mirror].rdev, 1785 r1_bio->sector, 1786 r1_bio->sectors, 1787 &first_bad, &bad_sectors) && 1788 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev, 1789 r1_bio->sector, 1790 r1_bio->sectors, 1791 &first_bad, &bad_sectors) 1792 ) 1793 set_bit(R1BIO_MadeGood, &r1_bio->state); 1794 1795 if (atomic_dec_and_test(&r1_bio->remaining)) { 1796 int s = r1_bio->sectors; 1797 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 1798 test_bit(R1BIO_WriteError, &r1_bio->state)) 1799 reschedule_retry(r1_bio); 1800 else { 1801 put_buf(r1_bio); 1802 md_done_sync(mddev, s, uptodate); 1803 } 1804 } 1805 } 1806 1807 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector, 1808 int sectors, struct page *page, int rw) 1809 { 1810 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false)) 1811 /* success */ 1812 return 1; 1813 if (rw == WRITE) { 1814 set_bit(WriteErrorSeen, &rdev->flags); 1815 if (!test_and_set_bit(WantReplacement, 1816 &rdev->flags)) 1817 set_bit(MD_RECOVERY_NEEDED, & 1818 rdev->mddev->recovery); 1819 } 1820 /* need to record an error - either for the block or the device */ 1821 if (!rdev_set_badblocks(rdev, sector, sectors, 0)) 1822 md_error(rdev->mddev, rdev); 1823 return 0; 1824 } 1825 1826 static int fix_sync_read_error(struct r1bio *r1_bio) 1827 { 1828 /* Try some synchronous reads of other devices to get 1829 * good data, much like with normal read errors. Only 1830 * read into the pages we already have so we don't 1831 * need to re-issue the read request. 1832 * We don't need to freeze the array, because being in an 1833 * active sync request, there is no normal IO, and 1834 * no overlapping syncs. 1835 * We don't need to check is_badblock() again as we 1836 * made sure that anything with a bad block in range 1837 * will have bi_end_io clear. 1838 */ 1839 struct mddev *mddev = r1_bio->mddev; 1840 struct r1conf *conf = mddev->private; 1841 struct bio *bio = r1_bio->bios[r1_bio->read_disk]; 1842 sector_t sect = r1_bio->sector; 1843 int sectors = r1_bio->sectors; 1844 int idx = 0; 1845 1846 while(sectors) { 1847 int s = sectors; 1848 int d = r1_bio->read_disk; 1849 int success = 0; 1850 struct md_rdev *rdev; 1851 int start; 1852 1853 if (s > (PAGE_SIZE>>9)) 1854 s = PAGE_SIZE >> 9; 1855 do { 1856 if (r1_bio->bios[d]->bi_end_io == end_sync_read) { 1857 /* No rcu protection needed here devices 1858 * can only be removed when no resync is 1859 * active, and resync is currently active 1860 */ 1861 rdev = conf->mirrors[d].rdev; 1862 if (sync_page_io(rdev, sect, s<<9, 1863 bio->bi_io_vec[idx].bv_page, 1864 READ, false)) { 1865 success = 1; 1866 break; 1867 } 1868 } 1869 d++; 1870 if (d == conf->raid_disks * 2) 1871 d = 0; 1872 } while (!success && d != r1_bio->read_disk); 1873 1874 if (!success) { 1875 char b[BDEVNAME_SIZE]; 1876 int abort = 0; 1877 /* Cannot read from anywhere, this block is lost. 1878 * Record a bad block on each device. If that doesn't 1879 * work just disable and interrupt the recovery. 1880 * Don't fail devices as that won't really help. 1881 */ 1882 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error" 1883 " for block %llu\n", 1884 mdname(mddev), 1885 bdevname(bio->bi_bdev, b), 1886 (unsigned long long)r1_bio->sector); 1887 for (d = 0; d < conf->raid_disks * 2; d++) { 1888 rdev = conf->mirrors[d].rdev; 1889 if (!rdev || test_bit(Faulty, &rdev->flags)) 1890 continue; 1891 if (!rdev_set_badblocks(rdev, sect, s, 0)) 1892 abort = 1; 1893 } 1894 if (abort) { 1895 conf->recovery_disabled = 1896 mddev->recovery_disabled; 1897 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1898 md_done_sync(mddev, r1_bio->sectors, 0); 1899 put_buf(r1_bio); 1900 return 0; 1901 } 1902 /* Try next page */ 1903 sectors -= s; 1904 sect += s; 1905 idx++; 1906 continue; 1907 } 1908 1909 start = d; 1910 /* write it back and re-read */ 1911 while (d != r1_bio->read_disk) { 1912 if (d == 0) 1913 d = conf->raid_disks * 2; 1914 d--; 1915 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1916 continue; 1917 rdev = conf->mirrors[d].rdev; 1918 if (r1_sync_page_io(rdev, sect, s, 1919 bio->bi_io_vec[idx].bv_page, 1920 WRITE) == 0) { 1921 r1_bio->bios[d]->bi_end_io = NULL; 1922 rdev_dec_pending(rdev, mddev); 1923 } 1924 } 1925 d = start; 1926 while (d != r1_bio->read_disk) { 1927 if (d == 0) 1928 d = conf->raid_disks * 2; 1929 d--; 1930 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1931 continue; 1932 rdev = conf->mirrors[d].rdev; 1933 if (r1_sync_page_io(rdev, sect, s, 1934 bio->bi_io_vec[idx].bv_page, 1935 READ) != 0) 1936 atomic_add(s, &rdev->corrected_errors); 1937 } 1938 sectors -= s; 1939 sect += s; 1940 idx ++; 1941 } 1942 set_bit(R1BIO_Uptodate, &r1_bio->state); 1943 set_bit(BIO_UPTODATE, &bio->bi_flags); 1944 return 1; 1945 } 1946 1947 static void process_checks(struct r1bio *r1_bio) 1948 { 1949 /* We have read all readable devices. If we haven't 1950 * got the block, then there is no hope left. 1951 * If we have, then we want to do a comparison 1952 * and skip the write if everything is the same. 1953 * If any blocks failed to read, then we need to 1954 * attempt an over-write 1955 */ 1956 struct mddev *mddev = r1_bio->mddev; 1957 struct r1conf *conf = mddev->private; 1958 int primary; 1959 int i; 1960 int vcnt; 1961 1962 /* Fix variable parts of all bios */ 1963 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9); 1964 for (i = 0; i < conf->raid_disks * 2; i++) { 1965 int j; 1966 int size; 1967 int uptodate; 1968 struct bio *b = r1_bio->bios[i]; 1969 if (b->bi_end_io != end_sync_read) 1970 continue; 1971 /* fixup the bio for reuse, but preserve BIO_UPTODATE */ 1972 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags); 1973 bio_reset(b); 1974 if (!uptodate) 1975 clear_bit(BIO_UPTODATE, &b->bi_flags); 1976 b->bi_vcnt = vcnt; 1977 b->bi_iter.bi_size = r1_bio->sectors << 9; 1978 b->bi_iter.bi_sector = r1_bio->sector + 1979 conf->mirrors[i].rdev->data_offset; 1980 b->bi_bdev = conf->mirrors[i].rdev->bdev; 1981 b->bi_end_io = end_sync_read; 1982 b->bi_private = r1_bio; 1983 1984 size = b->bi_iter.bi_size; 1985 for (j = 0; j < vcnt ; j++) { 1986 struct bio_vec *bi; 1987 bi = &b->bi_io_vec[j]; 1988 bi->bv_offset = 0; 1989 if (size > PAGE_SIZE) 1990 bi->bv_len = PAGE_SIZE; 1991 else 1992 bi->bv_len = size; 1993 size -= PAGE_SIZE; 1994 } 1995 } 1996 for (primary = 0; primary < conf->raid_disks * 2; primary++) 1997 if (r1_bio->bios[primary]->bi_end_io == end_sync_read && 1998 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) { 1999 r1_bio->bios[primary]->bi_end_io = NULL; 2000 rdev_dec_pending(conf->mirrors[primary].rdev, mddev); 2001 break; 2002 } 2003 r1_bio->read_disk = primary; 2004 for (i = 0; i < conf->raid_disks * 2; i++) { 2005 int j; 2006 struct bio *pbio = r1_bio->bios[primary]; 2007 struct bio *sbio = r1_bio->bios[i]; 2008 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags); 2009 2010 if (sbio->bi_end_io != end_sync_read) 2011 continue; 2012 /* Now we can 'fixup' the BIO_UPTODATE flag */ 2013 set_bit(BIO_UPTODATE, &sbio->bi_flags); 2014 2015 if (uptodate) { 2016 for (j = vcnt; j-- ; ) { 2017 struct page *p, *s; 2018 p = pbio->bi_io_vec[j].bv_page; 2019 s = sbio->bi_io_vec[j].bv_page; 2020 if (memcmp(page_address(p), 2021 page_address(s), 2022 sbio->bi_io_vec[j].bv_len)) 2023 break; 2024 } 2025 } else 2026 j = 0; 2027 if (j >= 0) 2028 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches); 2029 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery) 2030 && uptodate)) { 2031 /* No need to write to this device. */ 2032 sbio->bi_end_io = NULL; 2033 rdev_dec_pending(conf->mirrors[i].rdev, mddev); 2034 continue; 2035 } 2036 2037 bio_copy_data(sbio, pbio); 2038 } 2039 } 2040 2041 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio) 2042 { 2043 struct r1conf *conf = mddev->private; 2044 int i; 2045 int disks = conf->raid_disks * 2; 2046 struct bio *bio, *wbio; 2047 2048 bio = r1_bio->bios[r1_bio->read_disk]; 2049 2050 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 2051 /* ouch - failed to read all of that. */ 2052 if (!fix_sync_read_error(r1_bio)) 2053 return; 2054 2055 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2056 process_checks(r1_bio); 2057 2058 /* 2059 * schedule writes 2060 */ 2061 atomic_set(&r1_bio->remaining, 1); 2062 for (i = 0; i < disks ; i++) { 2063 wbio = r1_bio->bios[i]; 2064 if (wbio->bi_end_io == NULL || 2065 (wbio->bi_end_io == end_sync_read && 2066 (i == r1_bio->read_disk || 2067 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)))) 2068 continue; 2069 2070 wbio->bi_rw = WRITE; 2071 wbio->bi_end_io = end_sync_write; 2072 atomic_inc(&r1_bio->remaining); 2073 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio)); 2074 2075 generic_make_request(wbio); 2076 } 2077 2078 if (atomic_dec_and_test(&r1_bio->remaining)) { 2079 /* if we're here, all write(s) have completed, so clean up */ 2080 int s = r1_bio->sectors; 2081 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2082 test_bit(R1BIO_WriteError, &r1_bio->state)) 2083 reschedule_retry(r1_bio); 2084 else { 2085 put_buf(r1_bio); 2086 md_done_sync(mddev, s, 1); 2087 } 2088 } 2089 } 2090 2091 /* 2092 * This is a kernel thread which: 2093 * 2094 * 1. Retries failed read operations on working mirrors. 2095 * 2. Updates the raid superblock when problems encounter. 2096 * 3. Performs writes following reads for array synchronising. 2097 */ 2098 2099 static void fix_read_error(struct r1conf *conf, int read_disk, 2100 sector_t sect, int sectors) 2101 { 2102 struct mddev *mddev = conf->mddev; 2103 while(sectors) { 2104 int s = sectors; 2105 int d = read_disk; 2106 int success = 0; 2107 int start; 2108 struct md_rdev *rdev; 2109 2110 if (s > (PAGE_SIZE>>9)) 2111 s = PAGE_SIZE >> 9; 2112 2113 do { 2114 /* Note: no rcu protection needed here 2115 * as this is synchronous in the raid1d thread 2116 * which is the thread that might remove 2117 * a device. If raid1d ever becomes multi-threaded.... 2118 */ 2119 sector_t first_bad; 2120 int bad_sectors; 2121 2122 rdev = conf->mirrors[d].rdev; 2123 if (rdev && 2124 (test_bit(In_sync, &rdev->flags) || 2125 (!test_bit(Faulty, &rdev->flags) && 2126 rdev->recovery_offset >= sect + s)) && 2127 is_badblock(rdev, sect, s, 2128 &first_bad, &bad_sectors) == 0 && 2129 sync_page_io(rdev, sect, s<<9, 2130 conf->tmppage, READ, false)) 2131 success = 1; 2132 else { 2133 d++; 2134 if (d == conf->raid_disks * 2) 2135 d = 0; 2136 } 2137 } while (!success && d != read_disk); 2138 2139 if (!success) { 2140 /* Cannot read from anywhere - mark it bad */ 2141 struct md_rdev *rdev = conf->mirrors[read_disk].rdev; 2142 if (!rdev_set_badblocks(rdev, sect, s, 0)) 2143 md_error(mddev, rdev); 2144 break; 2145 } 2146 /* write it back and re-read */ 2147 start = d; 2148 while (d != read_disk) { 2149 if (d==0) 2150 d = conf->raid_disks * 2; 2151 d--; 2152 rdev = conf->mirrors[d].rdev; 2153 if (rdev && 2154 !test_bit(Faulty, &rdev->flags)) 2155 r1_sync_page_io(rdev, sect, s, 2156 conf->tmppage, WRITE); 2157 } 2158 d = start; 2159 while (d != read_disk) { 2160 char b[BDEVNAME_SIZE]; 2161 if (d==0) 2162 d = conf->raid_disks * 2; 2163 d--; 2164 rdev = conf->mirrors[d].rdev; 2165 if (rdev && 2166 !test_bit(Faulty, &rdev->flags)) { 2167 if (r1_sync_page_io(rdev, sect, s, 2168 conf->tmppage, READ)) { 2169 atomic_add(s, &rdev->corrected_errors); 2170 printk(KERN_INFO 2171 "md/raid1:%s: read error corrected " 2172 "(%d sectors at %llu on %s)\n", 2173 mdname(mddev), s, 2174 (unsigned long long)(sect + 2175 rdev->data_offset), 2176 bdevname(rdev->bdev, b)); 2177 } 2178 } 2179 } 2180 sectors -= s; 2181 sect += s; 2182 } 2183 } 2184 2185 static int narrow_write_error(struct r1bio *r1_bio, int i) 2186 { 2187 struct mddev *mddev = r1_bio->mddev; 2188 struct r1conf *conf = mddev->private; 2189 struct md_rdev *rdev = conf->mirrors[i].rdev; 2190 2191 /* bio has the data to be written to device 'i' where 2192 * we just recently had a write error. 2193 * We repeatedly clone the bio and trim down to one block, 2194 * then try the write. Where the write fails we record 2195 * a bad block. 2196 * It is conceivable that the bio doesn't exactly align with 2197 * blocks. We must handle this somehow. 2198 * 2199 * We currently own a reference on the rdev. 2200 */ 2201 2202 int block_sectors; 2203 sector_t sector; 2204 int sectors; 2205 int sect_to_write = r1_bio->sectors; 2206 int ok = 1; 2207 2208 if (rdev->badblocks.shift < 0) 2209 return 0; 2210 2211 block_sectors = roundup(1 << rdev->badblocks.shift, 2212 bdev_logical_block_size(rdev->bdev) >> 9); 2213 sector = r1_bio->sector; 2214 sectors = ((sector + block_sectors) 2215 & ~(sector_t)(block_sectors - 1)) 2216 - sector; 2217 2218 while (sect_to_write) { 2219 struct bio *wbio; 2220 if (sectors > sect_to_write) 2221 sectors = sect_to_write; 2222 /* Write at 'sector' for 'sectors'*/ 2223 2224 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 2225 unsigned vcnt = r1_bio->behind_page_count; 2226 struct bio_vec *vec = r1_bio->behind_bvecs; 2227 2228 while (!vec->bv_page) { 2229 vec++; 2230 vcnt--; 2231 } 2232 2233 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev); 2234 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec)); 2235 2236 wbio->bi_vcnt = vcnt; 2237 } else { 2238 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2239 } 2240 2241 wbio->bi_rw = WRITE; 2242 wbio->bi_iter.bi_sector = r1_bio->sector; 2243 wbio->bi_iter.bi_size = r1_bio->sectors << 9; 2244 2245 bio_trim(wbio, sector - r1_bio->sector, sectors); 2246 wbio->bi_iter.bi_sector += rdev->data_offset; 2247 wbio->bi_bdev = rdev->bdev; 2248 if (submit_bio_wait(WRITE, wbio) == 0) 2249 /* failure! */ 2250 ok = rdev_set_badblocks(rdev, sector, 2251 sectors, 0) 2252 && ok; 2253 2254 bio_put(wbio); 2255 sect_to_write -= sectors; 2256 sector += sectors; 2257 sectors = block_sectors; 2258 } 2259 return ok; 2260 } 2261 2262 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2263 { 2264 int m; 2265 int s = r1_bio->sectors; 2266 for (m = 0; m < conf->raid_disks * 2 ; m++) { 2267 struct md_rdev *rdev = conf->mirrors[m].rdev; 2268 struct bio *bio = r1_bio->bios[m]; 2269 if (bio->bi_end_io == NULL) 2270 continue; 2271 if (test_bit(BIO_UPTODATE, &bio->bi_flags) && 2272 test_bit(R1BIO_MadeGood, &r1_bio->state)) { 2273 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0); 2274 } 2275 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) && 2276 test_bit(R1BIO_WriteError, &r1_bio->state)) { 2277 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0)) 2278 md_error(conf->mddev, rdev); 2279 } 2280 } 2281 put_buf(r1_bio); 2282 md_done_sync(conf->mddev, s, 1); 2283 } 2284 2285 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2286 { 2287 int m; 2288 for (m = 0; m < conf->raid_disks * 2 ; m++) 2289 if (r1_bio->bios[m] == IO_MADE_GOOD) { 2290 struct md_rdev *rdev = conf->mirrors[m].rdev; 2291 rdev_clear_badblocks(rdev, 2292 r1_bio->sector, 2293 r1_bio->sectors, 0); 2294 rdev_dec_pending(rdev, conf->mddev); 2295 } else if (r1_bio->bios[m] != NULL) { 2296 /* This drive got a write error. We need to 2297 * narrow down and record precise write 2298 * errors. 2299 */ 2300 if (!narrow_write_error(r1_bio, m)) { 2301 md_error(conf->mddev, 2302 conf->mirrors[m].rdev); 2303 /* an I/O failed, we can't clear the bitmap */ 2304 set_bit(R1BIO_Degraded, &r1_bio->state); 2305 } 2306 rdev_dec_pending(conf->mirrors[m].rdev, 2307 conf->mddev); 2308 } 2309 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 2310 close_write(r1_bio); 2311 raid_end_bio_io(r1_bio); 2312 } 2313 2314 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio) 2315 { 2316 int disk; 2317 int max_sectors; 2318 struct mddev *mddev = conf->mddev; 2319 struct bio *bio; 2320 char b[BDEVNAME_SIZE]; 2321 struct md_rdev *rdev; 2322 2323 clear_bit(R1BIO_ReadError, &r1_bio->state); 2324 /* we got a read error. Maybe the drive is bad. Maybe just 2325 * the block and we can fix it. 2326 * We freeze all other IO, and try reading the block from 2327 * other devices. When we find one, we re-write 2328 * and check it that fixes the read error. 2329 * This is all done synchronously while the array is 2330 * frozen 2331 */ 2332 if (mddev->ro == 0) { 2333 freeze_array(conf, 1); 2334 fix_read_error(conf, r1_bio->read_disk, 2335 r1_bio->sector, r1_bio->sectors); 2336 unfreeze_array(conf); 2337 } else 2338 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev); 2339 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev); 2340 2341 bio = r1_bio->bios[r1_bio->read_disk]; 2342 bdevname(bio->bi_bdev, b); 2343 read_more: 2344 disk = read_balance(conf, r1_bio, &max_sectors); 2345 if (disk == -1) { 2346 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O" 2347 " read error for block %llu\n", 2348 mdname(mddev), b, (unsigned long long)r1_bio->sector); 2349 raid_end_bio_io(r1_bio); 2350 } else { 2351 const unsigned long do_sync 2352 = r1_bio->master_bio->bi_rw & REQ_SYNC; 2353 if (bio) { 2354 r1_bio->bios[r1_bio->read_disk] = 2355 mddev->ro ? IO_BLOCKED : NULL; 2356 bio_put(bio); 2357 } 2358 r1_bio->read_disk = disk; 2359 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2360 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector, 2361 max_sectors); 2362 r1_bio->bios[r1_bio->read_disk] = bio; 2363 rdev = conf->mirrors[disk].rdev; 2364 printk_ratelimited(KERN_ERR 2365 "md/raid1:%s: redirecting sector %llu" 2366 " to other mirror: %s\n", 2367 mdname(mddev), 2368 (unsigned long long)r1_bio->sector, 2369 bdevname(rdev->bdev, b)); 2370 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset; 2371 bio->bi_bdev = rdev->bdev; 2372 bio->bi_end_io = raid1_end_read_request; 2373 bio->bi_rw = READ | do_sync; 2374 bio->bi_private = r1_bio; 2375 if (max_sectors < r1_bio->sectors) { 2376 /* Drat - have to split this up more */ 2377 struct bio *mbio = r1_bio->master_bio; 2378 int sectors_handled = (r1_bio->sector + max_sectors 2379 - mbio->bi_iter.bi_sector); 2380 r1_bio->sectors = max_sectors; 2381 spin_lock_irq(&conf->device_lock); 2382 if (mbio->bi_phys_segments == 0) 2383 mbio->bi_phys_segments = 2; 2384 else 2385 mbio->bi_phys_segments++; 2386 spin_unlock_irq(&conf->device_lock); 2387 generic_make_request(bio); 2388 bio = NULL; 2389 2390 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 2391 2392 r1_bio->master_bio = mbio; 2393 r1_bio->sectors = bio_sectors(mbio) - sectors_handled; 2394 r1_bio->state = 0; 2395 set_bit(R1BIO_ReadError, &r1_bio->state); 2396 r1_bio->mddev = mddev; 2397 r1_bio->sector = mbio->bi_iter.bi_sector + 2398 sectors_handled; 2399 2400 goto read_more; 2401 } else 2402 generic_make_request(bio); 2403 } 2404 } 2405 2406 static void raid1d(struct md_thread *thread) 2407 { 2408 struct mddev *mddev = thread->mddev; 2409 struct r1bio *r1_bio; 2410 unsigned long flags; 2411 struct r1conf *conf = mddev->private; 2412 struct list_head *head = &conf->retry_list; 2413 struct blk_plug plug; 2414 2415 md_check_recovery(mddev); 2416 2417 blk_start_plug(&plug); 2418 for (;;) { 2419 2420 flush_pending_writes(conf); 2421 2422 spin_lock_irqsave(&conf->device_lock, flags); 2423 if (list_empty(head)) { 2424 spin_unlock_irqrestore(&conf->device_lock, flags); 2425 break; 2426 } 2427 r1_bio = list_entry(head->prev, struct r1bio, retry_list); 2428 list_del(head->prev); 2429 conf->nr_queued--; 2430 spin_unlock_irqrestore(&conf->device_lock, flags); 2431 2432 mddev = r1_bio->mddev; 2433 conf = mddev->private; 2434 if (test_bit(R1BIO_IsSync, &r1_bio->state)) { 2435 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2436 test_bit(R1BIO_WriteError, &r1_bio->state)) 2437 handle_sync_write_finished(conf, r1_bio); 2438 else 2439 sync_request_write(mddev, r1_bio); 2440 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2441 test_bit(R1BIO_WriteError, &r1_bio->state)) 2442 handle_write_finished(conf, r1_bio); 2443 else if (test_bit(R1BIO_ReadError, &r1_bio->state)) 2444 handle_read_error(conf, r1_bio); 2445 else 2446 /* just a partial read to be scheduled from separate 2447 * context 2448 */ 2449 generic_make_request(r1_bio->bios[r1_bio->read_disk]); 2450 2451 cond_resched(); 2452 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) 2453 md_check_recovery(mddev); 2454 } 2455 blk_finish_plug(&plug); 2456 } 2457 2458 static int init_resync(struct r1conf *conf) 2459 { 2460 int buffs; 2461 2462 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; 2463 BUG_ON(conf->r1buf_pool); 2464 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, 2465 conf->poolinfo); 2466 if (!conf->r1buf_pool) 2467 return -ENOMEM; 2468 conf->next_resync = 0; 2469 return 0; 2470 } 2471 2472 /* 2473 * perform a "sync" on one "block" 2474 * 2475 * We need to make sure that no normal I/O request - particularly write 2476 * requests - conflict with active sync requests. 2477 * 2478 * This is achieved by tracking pending requests and a 'barrier' concept 2479 * that can be installed to exclude normal IO requests. 2480 */ 2481 2482 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster) 2483 { 2484 struct r1conf *conf = mddev->private; 2485 struct r1bio *r1_bio; 2486 struct bio *bio; 2487 sector_t max_sector, nr_sectors; 2488 int disk = -1; 2489 int i; 2490 int wonly = -1; 2491 int write_targets = 0, read_targets = 0; 2492 sector_t sync_blocks; 2493 int still_degraded = 0; 2494 int good_sectors = RESYNC_SECTORS; 2495 int min_bad = 0; /* number of sectors that are bad in all devices */ 2496 2497 if (!conf->r1buf_pool) 2498 if (init_resync(conf)) 2499 return 0; 2500 2501 max_sector = mddev->dev_sectors; 2502 if (sector_nr >= max_sector) { 2503 /* If we aborted, we need to abort the 2504 * sync on the 'current' bitmap chunk (there will 2505 * only be one in raid1 resync. 2506 * We can find the current addess in mddev->curr_resync 2507 */ 2508 if (mddev->curr_resync < max_sector) /* aborted */ 2509 bitmap_end_sync(mddev->bitmap, mddev->curr_resync, 2510 &sync_blocks, 1); 2511 else /* completed sync */ 2512 conf->fullsync = 0; 2513 2514 bitmap_close_sync(mddev->bitmap); 2515 close_sync(conf); 2516 return 0; 2517 } 2518 2519 if (mddev->bitmap == NULL && 2520 mddev->recovery_cp == MaxSector && 2521 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && 2522 conf->fullsync == 0) { 2523 *skipped = 1; 2524 return max_sector - sector_nr; 2525 } 2526 /* before building a request, check if we can skip these blocks.. 2527 * This call the bitmap_start_sync doesn't actually record anything 2528 */ 2529 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) && 2530 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2531 /* We can skip this block, and probably several more */ 2532 *skipped = 1; 2533 return sync_blocks; 2534 } 2535 /* 2536 * If there is non-resync activity waiting for a turn, 2537 * and resync is going fast enough, 2538 * then let it though before starting on this new sync request. 2539 */ 2540 if (!go_faster && conf->nr_waiting) 2541 msleep_interruptible(1000); 2542 2543 bitmap_cond_end_sync(mddev->bitmap, sector_nr); 2544 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO); 2545 2546 raise_barrier(conf, sector_nr); 2547 2548 rcu_read_lock(); 2549 /* 2550 * If we get a correctably read error during resync or recovery, 2551 * we might want to read from a different device. So we 2552 * flag all drives that could conceivably be read from for READ, 2553 * and any others (which will be non-In_sync devices) for WRITE. 2554 * If a read fails, we try reading from something else for which READ 2555 * is OK. 2556 */ 2557 2558 r1_bio->mddev = mddev; 2559 r1_bio->sector = sector_nr; 2560 r1_bio->state = 0; 2561 set_bit(R1BIO_IsSync, &r1_bio->state); 2562 2563 for (i = 0; i < conf->raid_disks * 2; i++) { 2564 struct md_rdev *rdev; 2565 bio = r1_bio->bios[i]; 2566 bio_reset(bio); 2567 2568 rdev = rcu_dereference(conf->mirrors[i].rdev); 2569 if (rdev == NULL || 2570 test_bit(Faulty, &rdev->flags)) { 2571 if (i < conf->raid_disks) 2572 still_degraded = 1; 2573 } else if (!test_bit(In_sync, &rdev->flags)) { 2574 bio->bi_rw = WRITE; 2575 bio->bi_end_io = end_sync_write; 2576 write_targets ++; 2577 } else { 2578 /* may need to read from here */ 2579 sector_t first_bad = MaxSector; 2580 int bad_sectors; 2581 2582 if (is_badblock(rdev, sector_nr, good_sectors, 2583 &first_bad, &bad_sectors)) { 2584 if (first_bad > sector_nr) 2585 good_sectors = first_bad - sector_nr; 2586 else { 2587 bad_sectors -= (sector_nr - first_bad); 2588 if (min_bad == 0 || 2589 min_bad > bad_sectors) 2590 min_bad = bad_sectors; 2591 } 2592 } 2593 if (sector_nr < first_bad) { 2594 if (test_bit(WriteMostly, &rdev->flags)) { 2595 if (wonly < 0) 2596 wonly = i; 2597 } else { 2598 if (disk < 0) 2599 disk = i; 2600 } 2601 bio->bi_rw = READ; 2602 bio->bi_end_io = end_sync_read; 2603 read_targets++; 2604 } else if (!test_bit(WriteErrorSeen, &rdev->flags) && 2605 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && 2606 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) { 2607 /* 2608 * The device is suitable for reading (InSync), 2609 * but has bad block(s) here. Let's try to correct them, 2610 * if we are doing resync or repair. Otherwise, leave 2611 * this device alone for this sync request. 2612 */ 2613 bio->bi_rw = WRITE; 2614 bio->bi_end_io = end_sync_write; 2615 write_targets++; 2616 } 2617 } 2618 if (bio->bi_end_io) { 2619 atomic_inc(&rdev->nr_pending); 2620 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset; 2621 bio->bi_bdev = rdev->bdev; 2622 bio->bi_private = r1_bio; 2623 } 2624 } 2625 rcu_read_unlock(); 2626 if (disk < 0) 2627 disk = wonly; 2628 r1_bio->read_disk = disk; 2629 2630 if (read_targets == 0 && min_bad > 0) { 2631 /* These sectors are bad on all InSync devices, so we 2632 * need to mark them bad on all write targets 2633 */ 2634 int ok = 1; 2635 for (i = 0 ; i < conf->raid_disks * 2 ; i++) 2636 if (r1_bio->bios[i]->bi_end_io == end_sync_write) { 2637 struct md_rdev *rdev = conf->mirrors[i].rdev; 2638 ok = rdev_set_badblocks(rdev, sector_nr, 2639 min_bad, 0 2640 ) && ok; 2641 } 2642 set_bit(MD_CHANGE_DEVS, &mddev->flags); 2643 *skipped = 1; 2644 put_buf(r1_bio); 2645 2646 if (!ok) { 2647 /* Cannot record the badblocks, so need to 2648 * abort the resync. 2649 * If there are multiple read targets, could just 2650 * fail the really bad ones ??? 2651 */ 2652 conf->recovery_disabled = mddev->recovery_disabled; 2653 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 2654 return 0; 2655 } else 2656 return min_bad; 2657 2658 } 2659 if (min_bad > 0 && min_bad < good_sectors) { 2660 /* only resync enough to reach the next bad->good 2661 * transition */ 2662 good_sectors = min_bad; 2663 } 2664 2665 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0) 2666 /* extra read targets are also write targets */ 2667 write_targets += read_targets-1; 2668 2669 if (write_targets == 0 || read_targets == 0) { 2670 /* There is nowhere to write, so all non-sync 2671 * drives must be failed - so we are finished 2672 */ 2673 sector_t rv; 2674 if (min_bad > 0) 2675 max_sector = sector_nr + min_bad; 2676 rv = max_sector - sector_nr; 2677 *skipped = 1; 2678 put_buf(r1_bio); 2679 return rv; 2680 } 2681 2682 if (max_sector > mddev->resync_max) 2683 max_sector = mddev->resync_max; /* Don't do IO beyond here */ 2684 if (max_sector > sector_nr + good_sectors) 2685 max_sector = sector_nr + good_sectors; 2686 nr_sectors = 0; 2687 sync_blocks = 0; 2688 do { 2689 struct page *page; 2690 int len = PAGE_SIZE; 2691 if (sector_nr + (len>>9) > max_sector) 2692 len = (max_sector - sector_nr) << 9; 2693 if (len == 0) 2694 break; 2695 if (sync_blocks == 0) { 2696 if (!bitmap_start_sync(mddev->bitmap, sector_nr, 2697 &sync_blocks, still_degraded) && 2698 !conf->fullsync && 2699 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2700 break; 2701 BUG_ON(sync_blocks < (PAGE_SIZE>>9)); 2702 if ((len >> 9) > sync_blocks) 2703 len = sync_blocks<<9; 2704 } 2705 2706 for (i = 0 ; i < conf->raid_disks * 2; i++) { 2707 bio = r1_bio->bios[i]; 2708 if (bio->bi_end_io) { 2709 page = bio->bi_io_vec[bio->bi_vcnt].bv_page; 2710 if (bio_add_page(bio, page, len, 0) == 0) { 2711 /* stop here */ 2712 bio->bi_io_vec[bio->bi_vcnt].bv_page = page; 2713 while (i > 0) { 2714 i--; 2715 bio = r1_bio->bios[i]; 2716 if (bio->bi_end_io==NULL) 2717 continue; 2718 /* remove last page from this bio */ 2719 bio->bi_vcnt--; 2720 bio->bi_iter.bi_size -= len; 2721 __clear_bit(BIO_SEG_VALID, &bio->bi_flags); 2722 } 2723 goto bio_full; 2724 } 2725 } 2726 } 2727 nr_sectors += len>>9; 2728 sector_nr += len>>9; 2729 sync_blocks -= (len>>9); 2730 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES); 2731 bio_full: 2732 r1_bio->sectors = nr_sectors; 2733 2734 /* For a user-requested sync, we read all readable devices and do a 2735 * compare 2736 */ 2737 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2738 atomic_set(&r1_bio->remaining, read_targets); 2739 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) { 2740 bio = r1_bio->bios[i]; 2741 if (bio->bi_end_io == end_sync_read) { 2742 read_targets--; 2743 md_sync_acct(bio->bi_bdev, nr_sectors); 2744 generic_make_request(bio); 2745 } 2746 } 2747 } else { 2748 atomic_set(&r1_bio->remaining, 1); 2749 bio = r1_bio->bios[r1_bio->read_disk]; 2750 md_sync_acct(bio->bi_bdev, nr_sectors); 2751 generic_make_request(bio); 2752 2753 } 2754 return nr_sectors; 2755 } 2756 2757 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks) 2758 { 2759 if (sectors) 2760 return sectors; 2761 2762 return mddev->dev_sectors; 2763 } 2764 2765 static struct r1conf *setup_conf(struct mddev *mddev) 2766 { 2767 struct r1conf *conf; 2768 int i; 2769 struct raid1_info *disk; 2770 struct md_rdev *rdev; 2771 int err = -ENOMEM; 2772 2773 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL); 2774 if (!conf) 2775 goto abort; 2776 2777 conf->mirrors = kzalloc(sizeof(struct raid1_info) 2778 * mddev->raid_disks * 2, 2779 GFP_KERNEL); 2780 if (!conf->mirrors) 2781 goto abort; 2782 2783 conf->tmppage = alloc_page(GFP_KERNEL); 2784 if (!conf->tmppage) 2785 goto abort; 2786 2787 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL); 2788 if (!conf->poolinfo) 2789 goto abort; 2790 conf->poolinfo->raid_disks = mddev->raid_disks * 2; 2791 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 2792 r1bio_pool_free, 2793 conf->poolinfo); 2794 if (!conf->r1bio_pool) 2795 goto abort; 2796 2797 conf->poolinfo->mddev = mddev; 2798 2799 err = -EINVAL; 2800 spin_lock_init(&conf->device_lock); 2801 rdev_for_each(rdev, mddev) { 2802 struct request_queue *q; 2803 int disk_idx = rdev->raid_disk; 2804 if (disk_idx >= mddev->raid_disks 2805 || disk_idx < 0) 2806 continue; 2807 if (test_bit(Replacement, &rdev->flags)) 2808 disk = conf->mirrors + mddev->raid_disks + disk_idx; 2809 else 2810 disk = conf->mirrors + disk_idx; 2811 2812 if (disk->rdev) 2813 goto abort; 2814 disk->rdev = rdev; 2815 q = bdev_get_queue(rdev->bdev); 2816 if (q->merge_bvec_fn) 2817 mddev->merge_check_needed = 1; 2818 2819 disk->head_position = 0; 2820 disk->seq_start = MaxSector; 2821 } 2822 conf->raid_disks = mddev->raid_disks; 2823 conf->mddev = mddev; 2824 INIT_LIST_HEAD(&conf->retry_list); 2825 2826 spin_lock_init(&conf->resync_lock); 2827 init_waitqueue_head(&conf->wait_barrier); 2828 2829 bio_list_init(&conf->pending_bio_list); 2830 conf->pending_count = 0; 2831 conf->recovery_disabled = mddev->recovery_disabled - 1; 2832 2833 conf->start_next_window = MaxSector; 2834 conf->current_window_requests = conf->next_window_requests = 0; 2835 2836 err = -EIO; 2837 for (i = 0; i < conf->raid_disks * 2; i++) { 2838 2839 disk = conf->mirrors + i; 2840 2841 if (i < conf->raid_disks && 2842 disk[conf->raid_disks].rdev) { 2843 /* This slot has a replacement. */ 2844 if (!disk->rdev) { 2845 /* No original, just make the replacement 2846 * a recovering spare 2847 */ 2848 disk->rdev = 2849 disk[conf->raid_disks].rdev; 2850 disk[conf->raid_disks].rdev = NULL; 2851 } else if (!test_bit(In_sync, &disk->rdev->flags)) 2852 /* Original is not in_sync - bad */ 2853 goto abort; 2854 } 2855 2856 if (!disk->rdev || 2857 !test_bit(In_sync, &disk->rdev->flags)) { 2858 disk->head_position = 0; 2859 if (disk->rdev && 2860 (disk->rdev->saved_raid_disk < 0)) 2861 conf->fullsync = 1; 2862 } 2863 } 2864 2865 err = -ENOMEM; 2866 conf->thread = md_register_thread(raid1d, mddev, "raid1"); 2867 if (!conf->thread) { 2868 printk(KERN_ERR 2869 "md/raid1:%s: couldn't allocate thread\n", 2870 mdname(mddev)); 2871 goto abort; 2872 } 2873 2874 return conf; 2875 2876 abort: 2877 if (conf) { 2878 if (conf->r1bio_pool) 2879 mempool_destroy(conf->r1bio_pool); 2880 kfree(conf->mirrors); 2881 safe_put_page(conf->tmppage); 2882 kfree(conf->poolinfo); 2883 kfree(conf); 2884 } 2885 return ERR_PTR(err); 2886 } 2887 2888 static void raid1_free(struct mddev *mddev, void *priv); 2889 static int run(struct mddev *mddev) 2890 { 2891 struct r1conf *conf; 2892 int i; 2893 struct md_rdev *rdev; 2894 int ret; 2895 bool discard_supported = false; 2896 2897 if (mddev->level != 1) { 2898 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n", 2899 mdname(mddev), mddev->level); 2900 return -EIO; 2901 } 2902 if (mddev->reshape_position != MaxSector) { 2903 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n", 2904 mdname(mddev)); 2905 return -EIO; 2906 } 2907 /* 2908 * copy the already verified devices into our private RAID1 2909 * bookkeeping area. [whatever we allocate in run(), 2910 * should be freed in raid1_free()] 2911 */ 2912 if (mddev->private == NULL) 2913 conf = setup_conf(mddev); 2914 else 2915 conf = mddev->private; 2916 2917 if (IS_ERR(conf)) 2918 return PTR_ERR(conf); 2919 2920 if (mddev->queue) 2921 blk_queue_max_write_same_sectors(mddev->queue, 0); 2922 2923 rdev_for_each(rdev, mddev) { 2924 if (!mddev->gendisk) 2925 continue; 2926 disk_stack_limits(mddev->gendisk, rdev->bdev, 2927 rdev->data_offset << 9); 2928 if (blk_queue_discard(bdev_get_queue(rdev->bdev))) 2929 discard_supported = true; 2930 } 2931 2932 mddev->degraded = 0; 2933 for (i=0; i < conf->raid_disks; i++) 2934 if (conf->mirrors[i].rdev == NULL || 2935 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) || 2936 test_bit(Faulty, &conf->mirrors[i].rdev->flags)) 2937 mddev->degraded++; 2938 2939 if (conf->raid_disks - mddev->degraded == 1) 2940 mddev->recovery_cp = MaxSector; 2941 2942 if (mddev->recovery_cp != MaxSector) 2943 printk(KERN_NOTICE "md/raid1:%s: not clean" 2944 " -- starting background reconstruction\n", 2945 mdname(mddev)); 2946 printk(KERN_INFO 2947 "md/raid1:%s: active with %d out of %d mirrors\n", 2948 mdname(mddev), mddev->raid_disks - mddev->degraded, 2949 mddev->raid_disks); 2950 2951 /* 2952 * Ok, everything is just fine now 2953 */ 2954 mddev->thread = conf->thread; 2955 conf->thread = NULL; 2956 mddev->private = conf; 2957 2958 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0)); 2959 2960 if (mddev->queue) { 2961 if (discard_supported) 2962 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, 2963 mddev->queue); 2964 else 2965 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, 2966 mddev->queue); 2967 } 2968 2969 ret = md_integrity_register(mddev); 2970 if (ret) { 2971 md_unregister_thread(&mddev->thread); 2972 raid1_free(mddev, conf); 2973 } 2974 return ret; 2975 } 2976 2977 static void raid1_free(struct mddev *mddev, void *priv) 2978 { 2979 struct r1conf *conf = priv; 2980 2981 if (conf->r1bio_pool) 2982 mempool_destroy(conf->r1bio_pool); 2983 kfree(conf->mirrors); 2984 safe_put_page(conf->tmppage); 2985 kfree(conf->poolinfo); 2986 kfree(conf); 2987 } 2988 2989 static int raid1_resize(struct mddev *mddev, sector_t sectors) 2990 { 2991 /* no resync is happening, and there is enough space 2992 * on all devices, so we can resize. 2993 * We need to make sure resync covers any new space. 2994 * If the array is shrinking we should possibly wait until 2995 * any io in the removed space completes, but it hardly seems 2996 * worth it. 2997 */ 2998 sector_t newsize = raid1_size(mddev, sectors, 0); 2999 if (mddev->external_size && 3000 mddev->array_sectors > newsize) 3001 return -EINVAL; 3002 if (mddev->bitmap) { 3003 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0); 3004 if (ret) 3005 return ret; 3006 } 3007 md_set_array_sectors(mddev, newsize); 3008 set_capacity(mddev->gendisk, mddev->array_sectors); 3009 revalidate_disk(mddev->gendisk); 3010 if (sectors > mddev->dev_sectors && 3011 mddev->recovery_cp > mddev->dev_sectors) { 3012 mddev->recovery_cp = mddev->dev_sectors; 3013 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3014 } 3015 mddev->dev_sectors = sectors; 3016 mddev->resync_max_sectors = sectors; 3017 return 0; 3018 } 3019 3020 static int raid1_reshape(struct mddev *mddev) 3021 { 3022 /* We need to: 3023 * 1/ resize the r1bio_pool 3024 * 2/ resize conf->mirrors 3025 * 3026 * We allocate a new r1bio_pool if we can. 3027 * Then raise a device barrier and wait until all IO stops. 3028 * Then resize conf->mirrors and swap in the new r1bio pool. 3029 * 3030 * At the same time, we "pack" the devices so that all the missing 3031 * devices have the higher raid_disk numbers. 3032 */ 3033 mempool_t *newpool, *oldpool; 3034 struct pool_info *newpoolinfo; 3035 struct raid1_info *newmirrors; 3036 struct r1conf *conf = mddev->private; 3037 int cnt, raid_disks; 3038 unsigned long flags; 3039 int d, d2, err; 3040 3041 /* Cannot change chunk_size, layout, or level */ 3042 if (mddev->chunk_sectors != mddev->new_chunk_sectors || 3043 mddev->layout != mddev->new_layout || 3044 mddev->level != mddev->new_level) { 3045 mddev->new_chunk_sectors = mddev->chunk_sectors; 3046 mddev->new_layout = mddev->layout; 3047 mddev->new_level = mddev->level; 3048 return -EINVAL; 3049 } 3050 3051 err = md_allow_write(mddev); 3052 if (err) 3053 return err; 3054 3055 raid_disks = mddev->raid_disks + mddev->delta_disks; 3056 3057 if (raid_disks < conf->raid_disks) { 3058 cnt=0; 3059 for (d= 0; d < conf->raid_disks; d++) 3060 if (conf->mirrors[d].rdev) 3061 cnt++; 3062 if (cnt > raid_disks) 3063 return -EBUSY; 3064 } 3065 3066 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL); 3067 if (!newpoolinfo) 3068 return -ENOMEM; 3069 newpoolinfo->mddev = mddev; 3070 newpoolinfo->raid_disks = raid_disks * 2; 3071 3072 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 3073 r1bio_pool_free, newpoolinfo); 3074 if (!newpool) { 3075 kfree(newpoolinfo); 3076 return -ENOMEM; 3077 } 3078 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2, 3079 GFP_KERNEL); 3080 if (!newmirrors) { 3081 kfree(newpoolinfo); 3082 mempool_destroy(newpool); 3083 return -ENOMEM; 3084 } 3085 3086 freeze_array(conf, 0); 3087 3088 /* ok, everything is stopped */ 3089 oldpool = conf->r1bio_pool; 3090 conf->r1bio_pool = newpool; 3091 3092 for (d = d2 = 0; d < conf->raid_disks; d++) { 3093 struct md_rdev *rdev = conf->mirrors[d].rdev; 3094 if (rdev && rdev->raid_disk != d2) { 3095 sysfs_unlink_rdev(mddev, rdev); 3096 rdev->raid_disk = d2; 3097 sysfs_unlink_rdev(mddev, rdev); 3098 if (sysfs_link_rdev(mddev, rdev)) 3099 printk(KERN_WARNING 3100 "md/raid1:%s: cannot register rd%d\n", 3101 mdname(mddev), rdev->raid_disk); 3102 } 3103 if (rdev) 3104 newmirrors[d2++].rdev = rdev; 3105 } 3106 kfree(conf->mirrors); 3107 conf->mirrors = newmirrors; 3108 kfree(conf->poolinfo); 3109 conf->poolinfo = newpoolinfo; 3110 3111 spin_lock_irqsave(&conf->device_lock, flags); 3112 mddev->degraded += (raid_disks - conf->raid_disks); 3113 spin_unlock_irqrestore(&conf->device_lock, flags); 3114 conf->raid_disks = mddev->raid_disks = raid_disks; 3115 mddev->delta_disks = 0; 3116 3117 unfreeze_array(conf); 3118 3119 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3120 md_wakeup_thread(mddev->thread); 3121 3122 mempool_destroy(oldpool); 3123 return 0; 3124 } 3125 3126 static void raid1_quiesce(struct mddev *mddev, int state) 3127 { 3128 struct r1conf *conf = mddev->private; 3129 3130 switch(state) { 3131 case 2: /* wake for suspend */ 3132 wake_up(&conf->wait_barrier); 3133 break; 3134 case 1: 3135 freeze_array(conf, 0); 3136 break; 3137 case 0: 3138 unfreeze_array(conf); 3139 break; 3140 } 3141 } 3142 3143 static void *raid1_takeover(struct mddev *mddev) 3144 { 3145 /* raid1 can take over: 3146 * raid5 with 2 devices, any layout or chunk size 3147 */ 3148 if (mddev->level == 5 && mddev->raid_disks == 2) { 3149 struct r1conf *conf; 3150 mddev->new_level = 1; 3151 mddev->new_layout = 0; 3152 mddev->new_chunk_sectors = 0; 3153 conf = setup_conf(mddev); 3154 if (!IS_ERR(conf)) 3155 /* Array must appear to be quiesced */ 3156 conf->array_frozen = 1; 3157 return conf; 3158 } 3159 return ERR_PTR(-EINVAL); 3160 } 3161 3162 static struct md_personality raid1_personality = 3163 { 3164 .name = "raid1", 3165 .level = 1, 3166 .owner = THIS_MODULE, 3167 .make_request = make_request, 3168 .run = run, 3169 .free = raid1_free, 3170 .status = status, 3171 .error_handler = error, 3172 .hot_add_disk = raid1_add_disk, 3173 .hot_remove_disk= raid1_remove_disk, 3174 .spare_active = raid1_spare_active, 3175 .sync_request = sync_request, 3176 .resize = raid1_resize, 3177 .size = raid1_size, 3178 .check_reshape = raid1_reshape, 3179 .quiesce = raid1_quiesce, 3180 .takeover = raid1_takeover, 3181 .congested = raid1_congested, 3182 .mergeable_bvec = raid1_mergeable_bvec, 3183 }; 3184 3185 static int __init raid_init(void) 3186 { 3187 return register_md_personality(&raid1_personality); 3188 } 3189 3190 static void raid_exit(void) 3191 { 3192 unregister_md_personality(&raid1_personality); 3193 } 3194 3195 module_init(raid_init); 3196 module_exit(raid_exit); 3197 MODULE_LICENSE("GPL"); 3198 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD"); 3199 MODULE_ALIAS("md-personality-3"); /* RAID1 */ 3200 MODULE_ALIAS("md-raid1"); 3201 MODULE_ALIAS("md-level-1"); 3202 3203 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR); 3204