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 /* 499 * This routine returns the disk from which the requested read should 500 * be done. There is a per-array 'next expected sequential IO' sector 501 * number - if this matches on the next IO then we use the last disk. 502 * There is also a per-disk 'last know head position' sector that is 503 * maintained from IRQ contexts, both the normal and the resync IO 504 * completion handlers update this position correctly. If there is no 505 * perfect sequential match then we pick the disk whose head is closest. 506 * 507 * If there are 2 mirrors in the same 2 devices, performance degrades 508 * because position is mirror, not device based. 509 * 510 * The rdev for the device selected will have nr_pending incremented. 511 */ 512 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors) 513 { 514 const sector_t this_sector = r1_bio->sector; 515 int sectors; 516 int best_good_sectors; 517 int best_disk, best_dist_disk, best_pending_disk; 518 int has_nonrot_disk; 519 int disk; 520 sector_t best_dist; 521 unsigned int min_pending; 522 struct md_rdev *rdev; 523 int choose_first; 524 int choose_next_idle; 525 526 rcu_read_lock(); 527 /* 528 * Check if we can balance. We can balance on the whole 529 * device if no resync is going on, or below the resync window. 530 * We take the first readable disk when above the resync window. 531 */ 532 retry: 533 sectors = r1_bio->sectors; 534 best_disk = -1; 535 best_dist_disk = -1; 536 best_dist = MaxSector; 537 best_pending_disk = -1; 538 min_pending = UINT_MAX; 539 best_good_sectors = 0; 540 has_nonrot_disk = 0; 541 choose_next_idle = 0; 542 543 choose_first = (conf->mddev->recovery_cp < this_sector + sectors); 544 545 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) { 546 sector_t dist; 547 sector_t first_bad; 548 int bad_sectors; 549 unsigned int pending; 550 bool nonrot; 551 552 rdev = rcu_dereference(conf->mirrors[disk].rdev); 553 if (r1_bio->bios[disk] == IO_BLOCKED 554 || rdev == NULL 555 || test_bit(Unmerged, &rdev->flags) 556 || test_bit(Faulty, &rdev->flags)) 557 continue; 558 if (!test_bit(In_sync, &rdev->flags) && 559 rdev->recovery_offset < this_sector + sectors) 560 continue; 561 if (test_bit(WriteMostly, &rdev->flags)) { 562 /* Don't balance among write-mostly, just 563 * use the first as a last resort */ 564 if (best_disk < 0) { 565 if (is_badblock(rdev, this_sector, sectors, 566 &first_bad, &bad_sectors)) { 567 if (first_bad < this_sector) 568 /* Cannot use this */ 569 continue; 570 best_good_sectors = first_bad - this_sector; 571 } else 572 best_good_sectors = sectors; 573 best_disk = disk; 574 } 575 continue; 576 } 577 /* This is a reasonable device to use. It might 578 * even be best. 579 */ 580 if (is_badblock(rdev, this_sector, sectors, 581 &first_bad, &bad_sectors)) { 582 if (best_dist < MaxSector) 583 /* already have a better device */ 584 continue; 585 if (first_bad <= this_sector) { 586 /* cannot read here. If this is the 'primary' 587 * device, then we must not read beyond 588 * bad_sectors from another device.. 589 */ 590 bad_sectors -= (this_sector - first_bad); 591 if (choose_first && sectors > bad_sectors) 592 sectors = bad_sectors; 593 if (best_good_sectors > sectors) 594 best_good_sectors = sectors; 595 596 } else { 597 sector_t good_sectors = first_bad - this_sector; 598 if (good_sectors > best_good_sectors) { 599 best_good_sectors = good_sectors; 600 best_disk = disk; 601 } 602 if (choose_first) 603 break; 604 } 605 continue; 606 } else 607 best_good_sectors = sectors; 608 609 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev)); 610 has_nonrot_disk |= nonrot; 611 pending = atomic_read(&rdev->nr_pending); 612 dist = abs(this_sector - conf->mirrors[disk].head_position); 613 if (choose_first) { 614 best_disk = disk; 615 break; 616 } 617 /* Don't change to another disk for sequential reads */ 618 if (conf->mirrors[disk].next_seq_sect == this_sector 619 || dist == 0) { 620 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9; 621 struct raid1_info *mirror = &conf->mirrors[disk]; 622 623 best_disk = disk; 624 /* 625 * If buffered sequential IO size exceeds optimal 626 * iosize, check if there is idle disk. If yes, choose 627 * the idle disk. read_balance could already choose an 628 * idle disk before noticing it's a sequential IO in 629 * this disk. This doesn't matter because this disk 630 * will idle, next time it will be utilized after the 631 * first disk has IO size exceeds optimal iosize. In 632 * this way, iosize of the first disk will be optimal 633 * iosize at least. iosize of the second disk might be 634 * small, but not a big deal since when the second disk 635 * starts IO, the first disk is likely still busy. 636 */ 637 if (nonrot && opt_iosize > 0 && 638 mirror->seq_start != MaxSector && 639 mirror->next_seq_sect > opt_iosize && 640 mirror->next_seq_sect - opt_iosize >= 641 mirror->seq_start) { 642 choose_next_idle = 1; 643 continue; 644 } 645 break; 646 } 647 /* If device is idle, use it */ 648 if (pending == 0) { 649 best_disk = disk; 650 break; 651 } 652 653 if (choose_next_idle) 654 continue; 655 656 if (min_pending > pending) { 657 min_pending = pending; 658 best_pending_disk = disk; 659 } 660 661 if (dist < best_dist) { 662 best_dist = dist; 663 best_dist_disk = disk; 664 } 665 } 666 667 /* 668 * If all disks are rotational, choose the closest disk. If any disk is 669 * non-rotational, choose the disk with less pending request even the 670 * disk is rotational, which might/might not be optimal for raids with 671 * mixed ratation/non-rotational disks depending on workload. 672 */ 673 if (best_disk == -1) { 674 if (has_nonrot_disk) 675 best_disk = best_pending_disk; 676 else 677 best_disk = best_dist_disk; 678 } 679 680 if (best_disk >= 0) { 681 rdev = rcu_dereference(conf->mirrors[best_disk].rdev); 682 if (!rdev) 683 goto retry; 684 atomic_inc(&rdev->nr_pending); 685 if (test_bit(Faulty, &rdev->flags)) { 686 /* cannot risk returning a device that failed 687 * before we inc'ed nr_pending 688 */ 689 rdev_dec_pending(rdev, conf->mddev); 690 goto retry; 691 } 692 sectors = best_good_sectors; 693 694 if (conf->mirrors[best_disk].next_seq_sect != this_sector) 695 conf->mirrors[best_disk].seq_start = this_sector; 696 697 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors; 698 } 699 rcu_read_unlock(); 700 *max_sectors = sectors; 701 702 return best_disk; 703 } 704 705 static int raid1_mergeable_bvec(struct request_queue *q, 706 struct bvec_merge_data *bvm, 707 struct bio_vec *biovec) 708 { 709 struct mddev *mddev = q->queuedata; 710 struct r1conf *conf = mddev->private; 711 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev); 712 int max = biovec->bv_len; 713 714 if (mddev->merge_check_needed) { 715 int disk; 716 rcu_read_lock(); 717 for (disk = 0; disk < conf->raid_disks * 2; disk++) { 718 struct md_rdev *rdev = rcu_dereference( 719 conf->mirrors[disk].rdev); 720 if (rdev && !test_bit(Faulty, &rdev->flags)) { 721 struct request_queue *q = 722 bdev_get_queue(rdev->bdev); 723 if (q->merge_bvec_fn) { 724 bvm->bi_sector = sector + 725 rdev->data_offset; 726 bvm->bi_bdev = rdev->bdev; 727 max = min(max, q->merge_bvec_fn( 728 q, bvm, biovec)); 729 } 730 } 731 } 732 rcu_read_unlock(); 733 } 734 return max; 735 736 } 737 738 int md_raid1_congested(struct mddev *mddev, int bits) 739 { 740 struct r1conf *conf = mddev->private; 741 int i, ret = 0; 742 743 if ((bits & (1 << BDI_async_congested)) && 744 conf->pending_count >= max_queued_requests) 745 return 1; 746 747 rcu_read_lock(); 748 for (i = 0; i < conf->raid_disks * 2; i++) { 749 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 750 if (rdev && !test_bit(Faulty, &rdev->flags)) { 751 struct request_queue *q = bdev_get_queue(rdev->bdev); 752 753 BUG_ON(!q); 754 755 /* Note the '|| 1' - when read_balance prefers 756 * non-congested targets, it can be removed 757 */ 758 if ((bits & (1<<BDI_async_congested)) || 1) 759 ret |= bdi_congested(&q->backing_dev_info, bits); 760 else 761 ret &= bdi_congested(&q->backing_dev_info, bits); 762 } 763 } 764 rcu_read_unlock(); 765 return ret; 766 } 767 EXPORT_SYMBOL_GPL(md_raid1_congested); 768 769 static int raid1_congested(void *data, int bits) 770 { 771 struct mddev *mddev = data; 772 773 return mddev_congested(mddev, bits) || 774 md_raid1_congested(mddev, bits); 775 } 776 777 static void flush_pending_writes(struct r1conf *conf) 778 { 779 /* Any writes that have been queued but are awaiting 780 * bitmap updates get flushed here. 781 */ 782 spin_lock_irq(&conf->device_lock); 783 784 if (conf->pending_bio_list.head) { 785 struct bio *bio; 786 bio = bio_list_get(&conf->pending_bio_list); 787 conf->pending_count = 0; 788 spin_unlock_irq(&conf->device_lock); 789 /* flush any pending bitmap writes to 790 * disk before proceeding w/ I/O */ 791 bitmap_unplug(conf->mddev->bitmap); 792 wake_up(&conf->wait_barrier); 793 794 while (bio) { /* submit pending writes */ 795 struct bio *next = bio->bi_next; 796 bio->bi_next = NULL; 797 if (unlikely((bio->bi_rw & REQ_DISCARD) && 798 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 799 /* Just ignore it */ 800 bio_endio(bio, 0); 801 else 802 generic_make_request(bio); 803 bio = next; 804 } 805 } else 806 spin_unlock_irq(&conf->device_lock); 807 } 808 809 /* Barriers.... 810 * Sometimes we need to suspend IO while we do something else, 811 * either some resync/recovery, or reconfigure the array. 812 * To do this we raise a 'barrier'. 813 * The 'barrier' is a counter that can be raised multiple times 814 * to count how many activities are happening which preclude 815 * normal IO. 816 * We can only raise the barrier if there is no pending IO. 817 * i.e. if nr_pending == 0. 818 * We choose only to raise the barrier if no-one is waiting for the 819 * barrier to go down. This means that as soon as an IO request 820 * is ready, no other operations which require a barrier will start 821 * until the IO request has had a chance. 822 * 823 * So: regular IO calls 'wait_barrier'. When that returns there 824 * is no backgroup IO happening, It must arrange to call 825 * allow_barrier when it has finished its IO. 826 * backgroup IO calls must call raise_barrier. Once that returns 827 * there is no normal IO happeing. It must arrange to call 828 * lower_barrier when the particular background IO completes. 829 */ 830 static void raise_barrier(struct r1conf *conf, sector_t sector_nr) 831 { 832 spin_lock_irq(&conf->resync_lock); 833 834 /* Wait until no block IO is waiting */ 835 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting, 836 conf->resync_lock); 837 838 /* block any new IO from starting */ 839 conf->barrier++; 840 conf->next_resync = sector_nr; 841 842 /* For these conditions we must wait: 843 * A: while the array is in frozen state 844 * B: while barrier >= RESYNC_DEPTH, meaning resync reach 845 * the max count which allowed. 846 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning 847 * next resync will reach to the window which normal bios are 848 * handling. 849 * D: while there are any active requests in the current window. 850 */ 851 wait_event_lock_irq(conf->wait_barrier, 852 !conf->array_frozen && 853 conf->barrier < RESYNC_DEPTH && 854 conf->current_window_requests == 0 && 855 (conf->start_next_window >= 856 conf->next_resync + RESYNC_SECTORS), 857 conf->resync_lock); 858 859 spin_unlock_irq(&conf->resync_lock); 860 } 861 862 static void lower_barrier(struct r1conf *conf) 863 { 864 unsigned long flags; 865 BUG_ON(conf->barrier <= 0); 866 spin_lock_irqsave(&conf->resync_lock, flags); 867 conf->barrier--; 868 spin_unlock_irqrestore(&conf->resync_lock, flags); 869 wake_up(&conf->wait_barrier); 870 } 871 872 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio) 873 { 874 bool wait = false; 875 876 if (conf->array_frozen || !bio) 877 wait = true; 878 else if (conf->barrier && bio_data_dir(bio) == WRITE) { 879 if ((conf->mddev->curr_resync_completed 880 >= bio_end_sector(bio)) || 881 (conf->next_resync + NEXT_NORMALIO_DISTANCE 882 <= bio->bi_iter.bi_sector)) 883 wait = false; 884 else 885 wait = true; 886 } 887 888 return wait; 889 } 890 891 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio) 892 { 893 sector_t sector = 0; 894 895 spin_lock_irq(&conf->resync_lock); 896 if (need_to_wait_for_sync(conf, bio)) { 897 conf->nr_waiting++; 898 /* Wait for the barrier to drop. 899 * However if there are already pending 900 * requests (preventing the barrier from 901 * rising completely), and the 902 * pre-process bio queue isn't empty, 903 * then don't wait, as we need to empty 904 * that queue to get the nr_pending 905 * count down. 906 */ 907 wait_event_lock_irq(conf->wait_barrier, 908 !conf->array_frozen && 909 (!conf->barrier || 910 ((conf->start_next_window < 911 conf->next_resync + RESYNC_SECTORS) && 912 current->bio_list && 913 !bio_list_empty(current->bio_list))), 914 conf->resync_lock); 915 conf->nr_waiting--; 916 } 917 918 if (bio && bio_data_dir(bio) == WRITE) { 919 if (bio->bi_iter.bi_sector >= 920 conf->mddev->curr_resync_completed) { 921 if (conf->start_next_window == MaxSector) 922 conf->start_next_window = 923 conf->next_resync + 924 NEXT_NORMALIO_DISTANCE; 925 926 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE) 927 <= bio->bi_iter.bi_sector) 928 conf->next_window_requests++; 929 else 930 conf->current_window_requests++; 931 sector = conf->start_next_window; 932 } 933 } 934 935 conf->nr_pending++; 936 spin_unlock_irq(&conf->resync_lock); 937 return sector; 938 } 939 940 static void allow_barrier(struct r1conf *conf, sector_t start_next_window, 941 sector_t bi_sector) 942 { 943 unsigned long flags; 944 945 spin_lock_irqsave(&conf->resync_lock, flags); 946 conf->nr_pending--; 947 if (start_next_window) { 948 if (start_next_window == conf->start_next_window) { 949 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE 950 <= bi_sector) 951 conf->next_window_requests--; 952 else 953 conf->current_window_requests--; 954 } else 955 conf->current_window_requests--; 956 957 if (!conf->current_window_requests) { 958 if (conf->next_window_requests) { 959 conf->current_window_requests = 960 conf->next_window_requests; 961 conf->next_window_requests = 0; 962 conf->start_next_window += 963 NEXT_NORMALIO_DISTANCE; 964 } else 965 conf->start_next_window = MaxSector; 966 } 967 } 968 spin_unlock_irqrestore(&conf->resync_lock, flags); 969 wake_up(&conf->wait_barrier); 970 } 971 972 static void freeze_array(struct r1conf *conf, int extra) 973 { 974 /* stop syncio and normal IO and wait for everything to 975 * go quite. 976 * We wait until nr_pending match nr_queued+extra 977 * This is called in the context of one normal IO request 978 * that has failed. Thus any sync request that might be pending 979 * will be blocked by nr_pending, and we need to wait for 980 * pending IO requests to complete or be queued for re-try. 981 * Thus the number queued (nr_queued) plus this request (extra) 982 * must match the number of pending IOs (nr_pending) before 983 * we continue. 984 */ 985 spin_lock_irq(&conf->resync_lock); 986 conf->array_frozen = 1; 987 wait_event_lock_irq_cmd(conf->wait_barrier, 988 conf->nr_pending == conf->nr_queued+extra, 989 conf->resync_lock, 990 flush_pending_writes(conf)); 991 spin_unlock_irq(&conf->resync_lock); 992 } 993 static void unfreeze_array(struct r1conf *conf) 994 { 995 /* reverse the effect of the freeze */ 996 spin_lock_irq(&conf->resync_lock); 997 conf->array_frozen = 0; 998 wake_up(&conf->wait_barrier); 999 spin_unlock_irq(&conf->resync_lock); 1000 } 1001 1002 1003 /* duplicate the data pages for behind I/O 1004 */ 1005 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio) 1006 { 1007 int i; 1008 struct bio_vec *bvec; 1009 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec), 1010 GFP_NOIO); 1011 if (unlikely(!bvecs)) 1012 return; 1013 1014 bio_for_each_segment_all(bvec, bio, i) { 1015 bvecs[i] = *bvec; 1016 bvecs[i].bv_page = alloc_page(GFP_NOIO); 1017 if (unlikely(!bvecs[i].bv_page)) 1018 goto do_sync_io; 1019 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset, 1020 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len); 1021 kunmap(bvecs[i].bv_page); 1022 kunmap(bvec->bv_page); 1023 } 1024 r1_bio->behind_bvecs = bvecs; 1025 r1_bio->behind_page_count = bio->bi_vcnt; 1026 set_bit(R1BIO_BehindIO, &r1_bio->state); 1027 return; 1028 1029 do_sync_io: 1030 for (i = 0; i < bio->bi_vcnt; i++) 1031 if (bvecs[i].bv_page) 1032 put_page(bvecs[i].bv_page); 1033 kfree(bvecs); 1034 pr_debug("%dB behind alloc failed, doing sync I/O\n", 1035 bio->bi_iter.bi_size); 1036 } 1037 1038 struct raid1_plug_cb { 1039 struct blk_plug_cb cb; 1040 struct bio_list pending; 1041 int pending_cnt; 1042 }; 1043 1044 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule) 1045 { 1046 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, 1047 cb); 1048 struct mddev *mddev = plug->cb.data; 1049 struct r1conf *conf = mddev->private; 1050 struct bio *bio; 1051 1052 if (from_schedule || current->bio_list) { 1053 spin_lock_irq(&conf->device_lock); 1054 bio_list_merge(&conf->pending_bio_list, &plug->pending); 1055 conf->pending_count += plug->pending_cnt; 1056 spin_unlock_irq(&conf->device_lock); 1057 wake_up(&conf->wait_barrier); 1058 md_wakeup_thread(mddev->thread); 1059 kfree(plug); 1060 return; 1061 } 1062 1063 /* we aren't scheduling, so we can do the write-out directly. */ 1064 bio = bio_list_get(&plug->pending); 1065 bitmap_unplug(mddev->bitmap); 1066 wake_up(&conf->wait_barrier); 1067 1068 while (bio) { /* submit pending writes */ 1069 struct bio *next = bio->bi_next; 1070 bio->bi_next = NULL; 1071 if (unlikely((bio->bi_rw & REQ_DISCARD) && 1072 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 1073 /* Just ignore it */ 1074 bio_endio(bio, 0); 1075 else 1076 generic_make_request(bio); 1077 bio = next; 1078 } 1079 kfree(plug); 1080 } 1081 1082 static void make_request(struct mddev *mddev, struct bio * bio) 1083 { 1084 struct r1conf *conf = mddev->private; 1085 struct raid1_info *mirror; 1086 struct r1bio *r1_bio; 1087 struct bio *read_bio; 1088 int i, disks; 1089 struct bitmap *bitmap; 1090 unsigned long flags; 1091 const int rw = bio_data_dir(bio); 1092 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC); 1093 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA)); 1094 const unsigned long do_discard = (bio->bi_rw 1095 & (REQ_DISCARD | REQ_SECURE)); 1096 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME); 1097 struct md_rdev *blocked_rdev; 1098 struct blk_plug_cb *cb; 1099 struct raid1_plug_cb *plug = NULL; 1100 int first_clone; 1101 int sectors_handled; 1102 int max_sectors; 1103 sector_t start_next_window; 1104 1105 /* 1106 * Register the new request and wait if the reconstruction 1107 * thread has put up a bar for new requests. 1108 * Continue immediately if no resync is active currently. 1109 */ 1110 1111 md_write_start(mddev, bio); /* wait on superblock update early */ 1112 1113 if (bio_data_dir(bio) == WRITE && 1114 bio_end_sector(bio) > mddev->suspend_lo && 1115 bio->bi_iter.bi_sector < mddev->suspend_hi) { 1116 /* As the suspend_* range is controlled by 1117 * userspace, we want an interruptible 1118 * wait. 1119 */ 1120 DEFINE_WAIT(w); 1121 for (;;) { 1122 flush_signals(current); 1123 prepare_to_wait(&conf->wait_barrier, 1124 &w, TASK_INTERRUPTIBLE); 1125 if (bio_end_sector(bio) <= mddev->suspend_lo || 1126 bio->bi_iter.bi_sector >= mddev->suspend_hi) 1127 break; 1128 schedule(); 1129 } 1130 finish_wait(&conf->wait_barrier, &w); 1131 } 1132 1133 start_next_window = wait_barrier(conf, bio); 1134 1135 bitmap = mddev->bitmap; 1136 1137 /* 1138 * make_request() can abort the operation when READA is being 1139 * used and no empty request is available. 1140 * 1141 */ 1142 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1143 1144 r1_bio->master_bio = bio; 1145 r1_bio->sectors = bio_sectors(bio); 1146 r1_bio->state = 0; 1147 r1_bio->mddev = mddev; 1148 r1_bio->sector = bio->bi_iter.bi_sector; 1149 1150 /* We might need to issue multiple reads to different 1151 * devices if there are bad blocks around, so we keep 1152 * track of the number of reads in bio->bi_phys_segments. 1153 * If this is 0, there is only one r1_bio and no locking 1154 * will be needed when requests complete. If it is 1155 * non-zero, then it is the number of not-completed requests. 1156 */ 1157 bio->bi_phys_segments = 0; 1158 clear_bit(BIO_SEG_VALID, &bio->bi_flags); 1159 1160 if (rw == READ) { 1161 /* 1162 * read balancing logic: 1163 */ 1164 int rdisk; 1165 1166 read_again: 1167 rdisk = read_balance(conf, r1_bio, &max_sectors); 1168 1169 if (rdisk < 0) { 1170 /* couldn't find anywhere to read from */ 1171 raid_end_bio_io(r1_bio); 1172 return; 1173 } 1174 mirror = conf->mirrors + rdisk; 1175 1176 if (test_bit(WriteMostly, &mirror->rdev->flags) && 1177 bitmap) { 1178 /* Reading from a write-mostly device must 1179 * take care not to over-take any writes 1180 * that are 'behind' 1181 */ 1182 wait_event(bitmap->behind_wait, 1183 atomic_read(&bitmap->behind_writes) == 0); 1184 } 1185 r1_bio->read_disk = rdisk; 1186 r1_bio->start_next_window = 0; 1187 1188 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1189 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector, 1190 max_sectors); 1191 1192 r1_bio->bios[rdisk] = read_bio; 1193 1194 read_bio->bi_iter.bi_sector = r1_bio->sector + 1195 mirror->rdev->data_offset; 1196 read_bio->bi_bdev = mirror->rdev->bdev; 1197 read_bio->bi_end_io = raid1_end_read_request; 1198 read_bio->bi_rw = READ | do_sync; 1199 read_bio->bi_private = r1_bio; 1200 1201 if (max_sectors < r1_bio->sectors) { 1202 /* could not read all from this device, so we will 1203 * need another r1_bio. 1204 */ 1205 1206 sectors_handled = (r1_bio->sector + max_sectors 1207 - bio->bi_iter.bi_sector); 1208 r1_bio->sectors = max_sectors; 1209 spin_lock_irq(&conf->device_lock); 1210 if (bio->bi_phys_segments == 0) 1211 bio->bi_phys_segments = 2; 1212 else 1213 bio->bi_phys_segments++; 1214 spin_unlock_irq(&conf->device_lock); 1215 /* Cannot call generic_make_request directly 1216 * as that will be queued in __make_request 1217 * and subsequent mempool_alloc might block waiting 1218 * for it. So hand bio over to raid1d. 1219 */ 1220 reschedule_retry(r1_bio); 1221 1222 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1223 1224 r1_bio->master_bio = bio; 1225 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1226 r1_bio->state = 0; 1227 r1_bio->mddev = mddev; 1228 r1_bio->sector = bio->bi_iter.bi_sector + 1229 sectors_handled; 1230 goto read_again; 1231 } else 1232 generic_make_request(read_bio); 1233 return; 1234 } 1235 1236 /* 1237 * WRITE: 1238 */ 1239 if (conf->pending_count >= max_queued_requests) { 1240 md_wakeup_thread(mddev->thread); 1241 wait_event(conf->wait_barrier, 1242 conf->pending_count < max_queued_requests); 1243 } 1244 /* first select target devices under rcu_lock and 1245 * inc refcount on their rdev. Record them by setting 1246 * bios[x] to bio 1247 * If there are known/acknowledged bad blocks on any device on 1248 * which we have seen a write error, we want to avoid writing those 1249 * blocks. 1250 * This potentially requires several writes to write around 1251 * the bad blocks. Each set of writes gets it's own r1bio 1252 * with a set of bios attached. 1253 */ 1254 1255 disks = conf->raid_disks * 2; 1256 retry_write: 1257 r1_bio->start_next_window = start_next_window; 1258 blocked_rdev = NULL; 1259 rcu_read_lock(); 1260 max_sectors = r1_bio->sectors; 1261 for (i = 0; i < disks; i++) { 1262 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1263 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 1264 atomic_inc(&rdev->nr_pending); 1265 blocked_rdev = rdev; 1266 break; 1267 } 1268 r1_bio->bios[i] = NULL; 1269 if (!rdev || test_bit(Faulty, &rdev->flags) 1270 || test_bit(Unmerged, &rdev->flags)) { 1271 if (i < conf->raid_disks) 1272 set_bit(R1BIO_Degraded, &r1_bio->state); 1273 continue; 1274 } 1275 1276 atomic_inc(&rdev->nr_pending); 1277 if (test_bit(WriteErrorSeen, &rdev->flags)) { 1278 sector_t first_bad; 1279 int bad_sectors; 1280 int is_bad; 1281 1282 is_bad = is_badblock(rdev, r1_bio->sector, 1283 max_sectors, 1284 &first_bad, &bad_sectors); 1285 if (is_bad < 0) { 1286 /* mustn't write here until the bad block is 1287 * acknowledged*/ 1288 set_bit(BlockedBadBlocks, &rdev->flags); 1289 blocked_rdev = rdev; 1290 break; 1291 } 1292 if (is_bad && first_bad <= r1_bio->sector) { 1293 /* Cannot write here at all */ 1294 bad_sectors -= (r1_bio->sector - first_bad); 1295 if (bad_sectors < max_sectors) 1296 /* mustn't write more than bad_sectors 1297 * to other devices yet 1298 */ 1299 max_sectors = bad_sectors; 1300 rdev_dec_pending(rdev, mddev); 1301 /* We don't set R1BIO_Degraded as that 1302 * only applies if the disk is 1303 * missing, so it might be re-added, 1304 * and we want to know to recover this 1305 * chunk. 1306 * In this case the device is here, 1307 * and the fact that this chunk is not 1308 * in-sync is recorded in the bad 1309 * block log 1310 */ 1311 continue; 1312 } 1313 if (is_bad) { 1314 int good_sectors = first_bad - r1_bio->sector; 1315 if (good_sectors < max_sectors) 1316 max_sectors = good_sectors; 1317 } 1318 } 1319 r1_bio->bios[i] = bio; 1320 } 1321 rcu_read_unlock(); 1322 1323 if (unlikely(blocked_rdev)) { 1324 /* Wait for this device to become unblocked */ 1325 int j; 1326 sector_t old = start_next_window; 1327 1328 for (j = 0; j < i; j++) 1329 if (r1_bio->bios[j]) 1330 rdev_dec_pending(conf->mirrors[j].rdev, mddev); 1331 r1_bio->state = 0; 1332 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector); 1333 md_wait_for_blocked_rdev(blocked_rdev, mddev); 1334 start_next_window = wait_barrier(conf, bio); 1335 /* 1336 * We must make sure the multi r1bios of bio have 1337 * the same value of bi_phys_segments 1338 */ 1339 if (bio->bi_phys_segments && old && 1340 old != start_next_window) 1341 /* Wait for the former r1bio(s) to complete */ 1342 wait_event(conf->wait_barrier, 1343 bio->bi_phys_segments == 1); 1344 goto retry_write; 1345 } 1346 1347 if (max_sectors < r1_bio->sectors) { 1348 /* We are splitting this write into multiple parts, so 1349 * we need to prepare for allocating another r1_bio. 1350 */ 1351 r1_bio->sectors = max_sectors; 1352 spin_lock_irq(&conf->device_lock); 1353 if (bio->bi_phys_segments == 0) 1354 bio->bi_phys_segments = 2; 1355 else 1356 bio->bi_phys_segments++; 1357 spin_unlock_irq(&conf->device_lock); 1358 } 1359 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector; 1360 1361 atomic_set(&r1_bio->remaining, 1); 1362 atomic_set(&r1_bio->behind_remaining, 0); 1363 1364 first_clone = 1; 1365 for (i = 0; i < disks; i++) { 1366 struct bio *mbio; 1367 if (!r1_bio->bios[i]) 1368 continue; 1369 1370 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1371 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors); 1372 1373 if (first_clone) { 1374 /* do behind I/O ? 1375 * Not if there are too many, or cannot 1376 * allocate memory, or a reader on WriteMostly 1377 * is waiting for behind writes to flush */ 1378 if (bitmap && 1379 (atomic_read(&bitmap->behind_writes) 1380 < mddev->bitmap_info.max_write_behind) && 1381 !waitqueue_active(&bitmap->behind_wait)) 1382 alloc_behind_pages(mbio, r1_bio); 1383 1384 bitmap_startwrite(bitmap, r1_bio->sector, 1385 r1_bio->sectors, 1386 test_bit(R1BIO_BehindIO, 1387 &r1_bio->state)); 1388 first_clone = 0; 1389 } 1390 if (r1_bio->behind_bvecs) { 1391 struct bio_vec *bvec; 1392 int j; 1393 1394 /* 1395 * We trimmed the bio, so _all is legit 1396 */ 1397 bio_for_each_segment_all(bvec, mbio, j) 1398 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page; 1399 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags)) 1400 atomic_inc(&r1_bio->behind_remaining); 1401 } 1402 1403 r1_bio->bios[i] = mbio; 1404 1405 mbio->bi_iter.bi_sector = (r1_bio->sector + 1406 conf->mirrors[i].rdev->data_offset); 1407 mbio->bi_bdev = conf->mirrors[i].rdev->bdev; 1408 mbio->bi_end_io = raid1_end_write_request; 1409 mbio->bi_rw = 1410 WRITE | do_flush_fua | do_sync | do_discard | do_same; 1411 mbio->bi_private = r1_bio; 1412 1413 atomic_inc(&r1_bio->remaining); 1414 1415 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug)); 1416 if (cb) 1417 plug = container_of(cb, struct raid1_plug_cb, cb); 1418 else 1419 plug = NULL; 1420 spin_lock_irqsave(&conf->device_lock, flags); 1421 if (plug) { 1422 bio_list_add(&plug->pending, mbio); 1423 plug->pending_cnt++; 1424 } else { 1425 bio_list_add(&conf->pending_bio_list, mbio); 1426 conf->pending_count++; 1427 } 1428 spin_unlock_irqrestore(&conf->device_lock, flags); 1429 if (!plug) 1430 md_wakeup_thread(mddev->thread); 1431 } 1432 /* Mustn't call r1_bio_write_done before this next test, 1433 * as it could result in the bio being freed. 1434 */ 1435 if (sectors_handled < bio_sectors(bio)) { 1436 r1_bio_write_done(r1_bio); 1437 /* We need another r1_bio. It has already been counted 1438 * in bio->bi_phys_segments 1439 */ 1440 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1441 r1_bio->master_bio = bio; 1442 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1443 r1_bio->state = 0; 1444 r1_bio->mddev = mddev; 1445 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled; 1446 goto retry_write; 1447 } 1448 1449 r1_bio_write_done(r1_bio); 1450 1451 /* In case raid1d snuck in to freeze_array */ 1452 wake_up(&conf->wait_barrier); 1453 } 1454 1455 static void status(struct seq_file *seq, struct mddev *mddev) 1456 { 1457 struct r1conf *conf = mddev->private; 1458 int i; 1459 1460 seq_printf(seq, " [%d/%d] [", conf->raid_disks, 1461 conf->raid_disks - mddev->degraded); 1462 rcu_read_lock(); 1463 for (i = 0; i < conf->raid_disks; i++) { 1464 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1465 seq_printf(seq, "%s", 1466 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_"); 1467 } 1468 rcu_read_unlock(); 1469 seq_printf(seq, "]"); 1470 } 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 && repl->recovery_offset == MaxSector 1575 && !test_bit(Faulty, &repl->flags) 1576 && !test_and_set_bit(In_sync, &repl->flags)) { 1577 /* replacement has just become active */ 1578 if (!rdev || 1579 !test_and_clear_bit(In_sync, &rdev->flags)) 1580 count++; 1581 if (rdev) { 1582 /* Replaced device not technically 1583 * faulty, but we need to be sure 1584 * it gets removed and never re-added 1585 */ 1586 set_bit(Faulty, &rdev->flags); 1587 sysfs_notify_dirent_safe( 1588 rdev->sysfs_state); 1589 } 1590 } 1591 if (rdev 1592 && rdev->recovery_offset == MaxSector 1593 && !test_bit(Faulty, &rdev->flags) 1594 && !test_and_set_bit(In_sync, &rdev->flags)) { 1595 count++; 1596 sysfs_notify_dirent_safe(rdev->sysfs_state); 1597 } 1598 } 1599 spin_lock_irqsave(&conf->device_lock, flags); 1600 mddev->degraded -= count; 1601 spin_unlock_irqrestore(&conf->device_lock, flags); 1602 1603 print_conf(conf); 1604 return count; 1605 } 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 1737 static void end_sync_read(struct bio *bio, int error) 1738 { 1739 struct r1bio *r1_bio = bio->bi_private; 1740 1741 update_head_pos(r1_bio->read_disk, r1_bio); 1742 1743 /* 1744 * we have read a block, now it needs to be re-written, 1745 * or re-read if the read failed. 1746 * We don't do much here, just schedule handling by raid1d 1747 */ 1748 if (test_bit(BIO_UPTODATE, &bio->bi_flags)) 1749 set_bit(R1BIO_Uptodate, &r1_bio->state); 1750 1751 if (atomic_dec_and_test(&r1_bio->remaining)) 1752 reschedule_retry(r1_bio); 1753 } 1754 1755 static void end_sync_write(struct bio *bio, int error) 1756 { 1757 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 1758 struct r1bio *r1_bio = bio->bi_private; 1759 struct mddev *mddev = r1_bio->mddev; 1760 struct r1conf *conf = mddev->private; 1761 int mirror=0; 1762 sector_t first_bad; 1763 int bad_sectors; 1764 1765 mirror = find_bio_disk(r1_bio, bio); 1766 1767 if (!uptodate) { 1768 sector_t sync_blocks = 0; 1769 sector_t s = r1_bio->sector; 1770 long sectors_to_go = r1_bio->sectors; 1771 /* make sure these bits doesn't get cleared. */ 1772 do { 1773 bitmap_end_sync(mddev->bitmap, s, 1774 &sync_blocks, 1); 1775 s += sync_blocks; 1776 sectors_to_go -= sync_blocks; 1777 } while (sectors_to_go > 0); 1778 set_bit(WriteErrorSeen, 1779 &conf->mirrors[mirror].rdev->flags); 1780 if (!test_and_set_bit(WantReplacement, 1781 &conf->mirrors[mirror].rdev->flags)) 1782 set_bit(MD_RECOVERY_NEEDED, & 1783 mddev->recovery); 1784 set_bit(R1BIO_WriteError, &r1_bio->state); 1785 } else if (is_badblock(conf->mirrors[mirror].rdev, 1786 r1_bio->sector, 1787 r1_bio->sectors, 1788 &first_bad, &bad_sectors) && 1789 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev, 1790 r1_bio->sector, 1791 r1_bio->sectors, 1792 &first_bad, &bad_sectors) 1793 ) 1794 set_bit(R1BIO_MadeGood, &r1_bio->state); 1795 1796 if (atomic_dec_and_test(&r1_bio->remaining)) { 1797 int s = r1_bio->sectors; 1798 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 1799 test_bit(R1BIO_WriteError, &r1_bio->state)) 1800 reschedule_retry(r1_bio); 1801 else { 1802 put_buf(r1_bio); 1803 md_done_sync(mddev, s, uptodate); 1804 } 1805 } 1806 } 1807 1808 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector, 1809 int sectors, struct page *page, int rw) 1810 { 1811 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false)) 1812 /* success */ 1813 return 1; 1814 if (rw == WRITE) { 1815 set_bit(WriteErrorSeen, &rdev->flags); 1816 if (!test_and_set_bit(WantReplacement, 1817 &rdev->flags)) 1818 set_bit(MD_RECOVERY_NEEDED, & 1819 rdev->mddev->recovery); 1820 } 1821 /* need to record an error - either for the block or the device */ 1822 if (!rdev_set_badblocks(rdev, sector, sectors, 0)) 1823 md_error(rdev->mddev, rdev); 1824 return 0; 1825 } 1826 1827 static int fix_sync_read_error(struct r1bio *r1_bio) 1828 { 1829 /* Try some synchronous reads of other devices to get 1830 * good data, much like with normal read errors. Only 1831 * read into the pages we already have so we don't 1832 * need to re-issue the read request. 1833 * We don't need to freeze the array, because being in an 1834 * active sync request, there is no normal IO, and 1835 * no overlapping syncs. 1836 * We don't need to check is_badblock() again as we 1837 * made sure that anything with a bad block in range 1838 * will have bi_end_io clear. 1839 */ 1840 struct mddev *mddev = r1_bio->mddev; 1841 struct r1conf *conf = mddev->private; 1842 struct bio *bio = r1_bio->bios[r1_bio->read_disk]; 1843 sector_t sect = r1_bio->sector; 1844 int sectors = r1_bio->sectors; 1845 int idx = 0; 1846 1847 while(sectors) { 1848 int s = sectors; 1849 int d = r1_bio->read_disk; 1850 int success = 0; 1851 struct md_rdev *rdev; 1852 int start; 1853 1854 if (s > (PAGE_SIZE>>9)) 1855 s = PAGE_SIZE >> 9; 1856 do { 1857 if (r1_bio->bios[d]->bi_end_io == end_sync_read) { 1858 /* No rcu protection needed here devices 1859 * can only be removed when no resync is 1860 * active, and resync is currently active 1861 */ 1862 rdev = conf->mirrors[d].rdev; 1863 if (sync_page_io(rdev, sect, s<<9, 1864 bio->bi_io_vec[idx].bv_page, 1865 READ, false)) { 1866 success = 1; 1867 break; 1868 } 1869 } 1870 d++; 1871 if (d == conf->raid_disks * 2) 1872 d = 0; 1873 } while (!success && d != r1_bio->read_disk); 1874 1875 if (!success) { 1876 char b[BDEVNAME_SIZE]; 1877 int abort = 0; 1878 /* Cannot read from anywhere, this block is lost. 1879 * Record a bad block on each device. If that doesn't 1880 * work just disable and interrupt the recovery. 1881 * Don't fail devices as that won't really help. 1882 */ 1883 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error" 1884 " for block %llu\n", 1885 mdname(mddev), 1886 bdevname(bio->bi_bdev, b), 1887 (unsigned long long)r1_bio->sector); 1888 for (d = 0; d < conf->raid_disks * 2; d++) { 1889 rdev = conf->mirrors[d].rdev; 1890 if (!rdev || test_bit(Faulty, &rdev->flags)) 1891 continue; 1892 if (!rdev_set_badblocks(rdev, sect, s, 0)) 1893 abort = 1; 1894 } 1895 if (abort) { 1896 conf->recovery_disabled = 1897 mddev->recovery_disabled; 1898 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1899 md_done_sync(mddev, r1_bio->sectors, 0); 1900 put_buf(r1_bio); 1901 return 0; 1902 } 1903 /* Try next page */ 1904 sectors -= s; 1905 sect += s; 1906 idx++; 1907 continue; 1908 } 1909 1910 start = d; 1911 /* write it back and re-read */ 1912 while (d != r1_bio->read_disk) { 1913 if (d == 0) 1914 d = conf->raid_disks * 2; 1915 d--; 1916 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1917 continue; 1918 rdev = conf->mirrors[d].rdev; 1919 if (r1_sync_page_io(rdev, sect, s, 1920 bio->bi_io_vec[idx].bv_page, 1921 WRITE) == 0) { 1922 r1_bio->bios[d]->bi_end_io = NULL; 1923 rdev_dec_pending(rdev, mddev); 1924 } 1925 } 1926 d = start; 1927 while (d != r1_bio->read_disk) { 1928 if (d == 0) 1929 d = conf->raid_disks * 2; 1930 d--; 1931 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1932 continue; 1933 rdev = conf->mirrors[d].rdev; 1934 if (r1_sync_page_io(rdev, sect, s, 1935 bio->bi_io_vec[idx].bv_page, 1936 READ) != 0) 1937 atomic_add(s, &rdev->corrected_errors); 1938 } 1939 sectors -= s; 1940 sect += s; 1941 idx ++; 1942 } 1943 set_bit(R1BIO_Uptodate, &r1_bio->state); 1944 set_bit(BIO_UPTODATE, &bio->bi_flags); 1945 return 1; 1946 } 1947 1948 static int process_checks(struct r1bio *r1_bio) 1949 { 1950 /* We have read all readable devices. If we haven't 1951 * got the block, then there is no hope left. 1952 * If we have, then we want to do a comparison 1953 * and skip the write if everything is the same. 1954 * If any blocks failed to read, then we need to 1955 * attempt an over-write 1956 */ 1957 struct mddev *mddev = r1_bio->mddev; 1958 struct r1conf *conf = mddev->private; 1959 int primary; 1960 int i; 1961 int vcnt; 1962 1963 /* Fix variable parts of all bios */ 1964 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9); 1965 for (i = 0; i < conf->raid_disks * 2; i++) { 1966 int j; 1967 int size; 1968 int uptodate; 1969 struct bio *b = r1_bio->bios[i]; 1970 if (b->bi_end_io != end_sync_read) 1971 continue; 1972 /* fixup the bio for reuse, but preserve BIO_UPTODATE */ 1973 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags); 1974 bio_reset(b); 1975 if (!uptodate) 1976 clear_bit(BIO_UPTODATE, &b->bi_flags); 1977 b->bi_vcnt = vcnt; 1978 b->bi_iter.bi_size = r1_bio->sectors << 9; 1979 b->bi_iter.bi_sector = r1_bio->sector + 1980 conf->mirrors[i].rdev->data_offset; 1981 b->bi_bdev = conf->mirrors[i].rdev->bdev; 1982 b->bi_end_io = end_sync_read; 1983 b->bi_private = r1_bio; 1984 1985 size = b->bi_iter.bi_size; 1986 for (j = 0; j < vcnt ; j++) { 1987 struct bio_vec *bi; 1988 bi = &b->bi_io_vec[j]; 1989 bi->bv_offset = 0; 1990 if (size > PAGE_SIZE) 1991 bi->bv_len = PAGE_SIZE; 1992 else 1993 bi->bv_len = size; 1994 size -= PAGE_SIZE; 1995 } 1996 } 1997 for (primary = 0; primary < conf->raid_disks * 2; primary++) 1998 if (r1_bio->bios[primary]->bi_end_io == end_sync_read && 1999 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) { 2000 r1_bio->bios[primary]->bi_end_io = NULL; 2001 rdev_dec_pending(conf->mirrors[primary].rdev, mddev); 2002 break; 2003 } 2004 r1_bio->read_disk = primary; 2005 for (i = 0; i < conf->raid_disks * 2; i++) { 2006 int j; 2007 struct bio *pbio = r1_bio->bios[primary]; 2008 struct bio *sbio = r1_bio->bios[i]; 2009 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags); 2010 2011 if (sbio->bi_end_io != end_sync_read) 2012 continue; 2013 /* Now we can 'fixup' the BIO_UPTODATE flag */ 2014 set_bit(BIO_UPTODATE, &sbio->bi_flags); 2015 2016 if (uptodate) { 2017 for (j = vcnt; j-- ; ) { 2018 struct page *p, *s; 2019 p = pbio->bi_io_vec[j].bv_page; 2020 s = sbio->bi_io_vec[j].bv_page; 2021 if (memcmp(page_address(p), 2022 page_address(s), 2023 sbio->bi_io_vec[j].bv_len)) 2024 break; 2025 } 2026 } else 2027 j = 0; 2028 if (j >= 0) 2029 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches); 2030 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery) 2031 && uptodate)) { 2032 /* No need to write to this device. */ 2033 sbio->bi_end_io = NULL; 2034 rdev_dec_pending(conf->mirrors[i].rdev, mddev); 2035 continue; 2036 } 2037 2038 bio_copy_data(sbio, pbio); 2039 } 2040 return 0; 2041 } 2042 2043 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio) 2044 { 2045 struct r1conf *conf = mddev->private; 2046 int i; 2047 int disks = conf->raid_disks * 2; 2048 struct bio *bio, *wbio; 2049 2050 bio = r1_bio->bios[r1_bio->read_disk]; 2051 2052 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 2053 /* ouch - failed to read all of that. */ 2054 if (!fix_sync_read_error(r1_bio)) 2055 return; 2056 2057 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2058 if (process_checks(r1_bio) < 0) 2059 return; 2060 /* 2061 * schedule writes 2062 */ 2063 atomic_set(&r1_bio->remaining, 1); 2064 for (i = 0; i < disks ; i++) { 2065 wbio = r1_bio->bios[i]; 2066 if (wbio->bi_end_io == NULL || 2067 (wbio->bi_end_io == end_sync_read && 2068 (i == r1_bio->read_disk || 2069 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)))) 2070 continue; 2071 2072 wbio->bi_rw = WRITE; 2073 wbio->bi_end_io = end_sync_write; 2074 atomic_inc(&r1_bio->remaining); 2075 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio)); 2076 2077 generic_make_request(wbio); 2078 } 2079 2080 if (atomic_dec_and_test(&r1_bio->remaining)) { 2081 /* if we're here, all write(s) have completed, so clean up */ 2082 int s = r1_bio->sectors; 2083 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2084 test_bit(R1BIO_WriteError, &r1_bio->state)) 2085 reschedule_retry(r1_bio); 2086 else { 2087 put_buf(r1_bio); 2088 md_done_sync(mddev, s, 1); 2089 } 2090 } 2091 } 2092 2093 /* 2094 * This is a kernel thread which: 2095 * 2096 * 1. Retries failed read operations on working mirrors. 2097 * 2. Updates the raid superblock when problems encounter. 2098 * 3. Performs writes following reads for array synchronising. 2099 */ 2100 2101 static void fix_read_error(struct r1conf *conf, int read_disk, 2102 sector_t sect, int sectors) 2103 { 2104 struct mddev *mddev = conf->mddev; 2105 while(sectors) { 2106 int s = sectors; 2107 int d = read_disk; 2108 int success = 0; 2109 int start; 2110 struct md_rdev *rdev; 2111 2112 if (s > (PAGE_SIZE>>9)) 2113 s = PAGE_SIZE >> 9; 2114 2115 do { 2116 /* Note: no rcu protection needed here 2117 * as this is synchronous in the raid1d thread 2118 * which is the thread that might remove 2119 * a device. If raid1d ever becomes multi-threaded.... 2120 */ 2121 sector_t first_bad; 2122 int bad_sectors; 2123 2124 rdev = conf->mirrors[d].rdev; 2125 if (rdev && 2126 (test_bit(In_sync, &rdev->flags) || 2127 (!test_bit(Faulty, &rdev->flags) && 2128 rdev->recovery_offset >= sect + s)) && 2129 is_badblock(rdev, sect, s, 2130 &first_bad, &bad_sectors) == 0 && 2131 sync_page_io(rdev, sect, s<<9, 2132 conf->tmppage, READ, false)) 2133 success = 1; 2134 else { 2135 d++; 2136 if (d == conf->raid_disks * 2) 2137 d = 0; 2138 } 2139 } while (!success && d != read_disk); 2140 2141 if (!success) { 2142 /* Cannot read from anywhere - mark it bad */ 2143 struct md_rdev *rdev = conf->mirrors[read_disk].rdev; 2144 if (!rdev_set_badblocks(rdev, sect, s, 0)) 2145 md_error(mddev, rdev); 2146 break; 2147 } 2148 /* write it back and re-read */ 2149 start = d; 2150 while (d != read_disk) { 2151 if (d==0) 2152 d = conf->raid_disks * 2; 2153 d--; 2154 rdev = conf->mirrors[d].rdev; 2155 if (rdev && 2156 test_bit(In_sync, &rdev->flags)) 2157 r1_sync_page_io(rdev, sect, s, 2158 conf->tmppage, WRITE); 2159 } 2160 d = start; 2161 while (d != read_disk) { 2162 char b[BDEVNAME_SIZE]; 2163 if (d==0) 2164 d = conf->raid_disks * 2; 2165 d--; 2166 rdev = conf->mirrors[d].rdev; 2167 if (rdev && 2168 test_bit(In_sync, &rdev->flags)) { 2169 if (r1_sync_page_io(rdev, sect, s, 2170 conf->tmppage, READ)) { 2171 atomic_add(s, &rdev->corrected_errors); 2172 printk(KERN_INFO 2173 "md/raid1:%s: read error corrected " 2174 "(%d sectors at %llu on %s)\n", 2175 mdname(mddev), s, 2176 (unsigned long long)(sect + 2177 rdev->data_offset), 2178 bdevname(rdev->bdev, b)); 2179 } 2180 } 2181 } 2182 sectors -= s; 2183 sect += s; 2184 } 2185 } 2186 2187 static int narrow_write_error(struct r1bio *r1_bio, int i) 2188 { 2189 struct mddev *mddev = r1_bio->mddev; 2190 struct r1conf *conf = mddev->private; 2191 struct md_rdev *rdev = conf->mirrors[i].rdev; 2192 2193 /* bio has the data to be written to device 'i' where 2194 * we just recently had a write error. 2195 * We repeatedly clone the bio and trim down to one block, 2196 * then try the write. Where the write fails we record 2197 * a bad block. 2198 * It is conceivable that the bio doesn't exactly align with 2199 * blocks. We must handle this somehow. 2200 * 2201 * We currently own a reference on the rdev. 2202 */ 2203 2204 int block_sectors; 2205 sector_t sector; 2206 int sectors; 2207 int sect_to_write = r1_bio->sectors; 2208 int ok = 1; 2209 2210 if (rdev->badblocks.shift < 0) 2211 return 0; 2212 2213 block_sectors = 1 << rdev->badblocks.shift; 2214 sector = r1_bio->sector; 2215 sectors = ((sector + block_sectors) 2216 & ~(sector_t)(block_sectors - 1)) 2217 - sector; 2218 2219 while (sect_to_write) { 2220 struct bio *wbio; 2221 if (sectors > sect_to_write) 2222 sectors = sect_to_write; 2223 /* Write at 'sector' for 'sectors'*/ 2224 2225 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 2226 unsigned vcnt = r1_bio->behind_page_count; 2227 struct bio_vec *vec = r1_bio->behind_bvecs; 2228 2229 while (!vec->bv_page) { 2230 vec++; 2231 vcnt--; 2232 } 2233 2234 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev); 2235 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec)); 2236 2237 wbio->bi_vcnt = vcnt; 2238 } else { 2239 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2240 } 2241 2242 wbio->bi_rw = WRITE; 2243 wbio->bi_iter.bi_sector = r1_bio->sector; 2244 wbio->bi_iter.bi_size = r1_bio->sectors << 9; 2245 2246 bio_trim(wbio, sector - r1_bio->sector, sectors); 2247 wbio->bi_iter.bi_sector += rdev->data_offset; 2248 wbio->bi_bdev = rdev->bdev; 2249 if (submit_bio_wait(WRITE, wbio) == 0) 2250 /* failure! */ 2251 ok = rdev_set_badblocks(rdev, sector, 2252 sectors, 0) 2253 && ok; 2254 2255 bio_put(wbio); 2256 sect_to_write -= sectors; 2257 sector += sectors; 2258 sectors = block_sectors; 2259 } 2260 return ok; 2261 } 2262 2263 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2264 { 2265 int m; 2266 int s = r1_bio->sectors; 2267 for (m = 0; m < conf->raid_disks * 2 ; m++) { 2268 struct md_rdev *rdev = conf->mirrors[m].rdev; 2269 struct bio *bio = r1_bio->bios[m]; 2270 if (bio->bi_end_io == NULL) 2271 continue; 2272 if (test_bit(BIO_UPTODATE, &bio->bi_flags) && 2273 test_bit(R1BIO_MadeGood, &r1_bio->state)) { 2274 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0); 2275 } 2276 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) && 2277 test_bit(R1BIO_WriteError, &r1_bio->state)) { 2278 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0)) 2279 md_error(conf->mddev, rdev); 2280 } 2281 } 2282 put_buf(r1_bio); 2283 md_done_sync(conf->mddev, s, 1); 2284 } 2285 2286 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2287 { 2288 int m; 2289 for (m = 0; m < conf->raid_disks * 2 ; m++) 2290 if (r1_bio->bios[m] == IO_MADE_GOOD) { 2291 struct md_rdev *rdev = conf->mirrors[m].rdev; 2292 rdev_clear_badblocks(rdev, 2293 r1_bio->sector, 2294 r1_bio->sectors, 0); 2295 rdev_dec_pending(rdev, conf->mddev); 2296 } else if (r1_bio->bios[m] != NULL) { 2297 /* This drive got a write error. We need to 2298 * narrow down and record precise write 2299 * errors. 2300 */ 2301 if (!narrow_write_error(r1_bio, m)) { 2302 md_error(conf->mddev, 2303 conf->mirrors[m].rdev); 2304 /* an I/O failed, we can't clear the bitmap */ 2305 set_bit(R1BIO_Degraded, &r1_bio->state); 2306 } 2307 rdev_dec_pending(conf->mirrors[m].rdev, 2308 conf->mddev); 2309 } 2310 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 2311 close_write(r1_bio); 2312 raid_end_bio_io(r1_bio); 2313 } 2314 2315 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio) 2316 { 2317 int disk; 2318 int max_sectors; 2319 struct mddev *mddev = conf->mddev; 2320 struct bio *bio; 2321 char b[BDEVNAME_SIZE]; 2322 struct md_rdev *rdev; 2323 2324 clear_bit(R1BIO_ReadError, &r1_bio->state); 2325 /* we got a read error. Maybe the drive is bad. Maybe just 2326 * the block and we can fix it. 2327 * We freeze all other IO, and try reading the block from 2328 * other devices. When we find one, we re-write 2329 * and check it that fixes the read error. 2330 * This is all done synchronously while the array is 2331 * frozen 2332 */ 2333 if (mddev->ro == 0) { 2334 freeze_array(conf, 1); 2335 fix_read_error(conf, r1_bio->read_disk, 2336 r1_bio->sector, r1_bio->sectors); 2337 unfreeze_array(conf); 2338 } else 2339 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev); 2340 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev); 2341 2342 bio = r1_bio->bios[r1_bio->read_disk]; 2343 bdevname(bio->bi_bdev, b); 2344 read_more: 2345 disk = read_balance(conf, r1_bio, &max_sectors); 2346 if (disk == -1) { 2347 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O" 2348 " read error for block %llu\n", 2349 mdname(mddev), b, (unsigned long long)r1_bio->sector); 2350 raid_end_bio_io(r1_bio); 2351 } else { 2352 const unsigned long do_sync 2353 = r1_bio->master_bio->bi_rw & REQ_SYNC; 2354 if (bio) { 2355 r1_bio->bios[r1_bio->read_disk] = 2356 mddev->ro ? IO_BLOCKED : NULL; 2357 bio_put(bio); 2358 } 2359 r1_bio->read_disk = disk; 2360 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2361 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector, 2362 max_sectors); 2363 r1_bio->bios[r1_bio->read_disk] = bio; 2364 rdev = conf->mirrors[disk].rdev; 2365 printk_ratelimited(KERN_ERR 2366 "md/raid1:%s: redirecting sector %llu" 2367 " to other mirror: %s\n", 2368 mdname(mddev), 2369 (unsigned long long)r1_bio->sector, 2370 bdevname(rdev->bdev, b)); 2371 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset; 2372 bio->bi_bdev = rdev->bdev; 2373 bio->bi_end_io = raid1_end_read_request; 2374 bio->bi_rw = READ | do_sync; 2375 bio->bi_private = r1_bio; 2376 if (max_sectors < r1_bio->sectors) { 2377 /* Drat - have to split this up more */ 2378 struct bio *mbio = r1_bio->master_bio; 2379 int sectors_handled = (r1_bio->sector + max_sectors 2380 - mbio->bi_iter.bi_sector); 2381 r1_bio->sectors = max_sectors; 2382 spin_lock_irq(&conf->device_lock); 2383 if (mbio->bi_phys_segments == 0) 2384 mbio->bi_phys_segments = 2; 2385 else 2386 mbio->bi_phys_segments++; 2387 spin_unlock_irq(&conf->device_lock); 2388 generic_make_request(bio); 2389 bio = NULL; 2390 2391 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 2392 2393 r1_bio->master_bio = mbio; 2394 r1_bio->sectors = bio_sectors(mbio) - sectors_handled; 2395 r1_bio->state = 0; 2396 set_bit(R1BIO_ReadError, &r1_bio->state); 2397 r1_bio->mddev = mddev; 2398 r1_bio->sector = mbio->bi_iter.bi_sector + 2399 sectors_handled; 2400 2401 goto read_more; 2402 } else 2403 generic_make_request(bio); 2404 } 2405 } 2406 2407 static void raid1d(struct md_thread *thread) 2408 { 2409 struct mddev *mddev = thread->mddev; 2410 struct r1bio *r1_bio; 2411 unsigned long flags; 2412 struct r1conf *conf = mddev->private; 2413 struct list_head *head = &conf->retry_list; 2414 struct blk_plug plug; 2415 2416 md_check_recovery(mddev); 2417 2418 blk_start_plug(&plug); 2419 for (;;) { 2420 2421 flush_pending_writes(conf); 2422 2423 spin_lock_irqsave(&conf->device_lock, flags); 2424 if (list_empty(head)) { 2425 spin_unlock_irqrestore(&conf->device_lock, flags); 2426 break; 2427 } 2428 r1_bio = list_entry(head->prev, struct r1bio, retry_list); 2429 list_del(head->prev); 2430 conf->nr_queued--; 2431 spin_unlock_irqrestore(&conf->device_lock, flags); 2432 2433 mddev = r1_bio->mddev; 2434 conf = mddev->private; 2435 if (test_bit(R1BIO_IsSync, &r1_bio->state)) { 2436 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2437 test_bit(R1BIO_WriteError, &r1_bio->state)) 2438 handle_sync_write_finished(conf, r1_bio); 2439 else 2440 sync_request_write(mddev, r1_bio); 2441 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2442 test_bit(R1BIO_WriteError, &r1_bio->state)) 2443 handle_write_finished(conf, r1_bio); 2444 else if (test_bit(R1BIO_ReadError, &r1_bio->state)) 2445 handle_read_error(conf, r1_bio); 2446 else 2447 /* just a partial read to be scheduled from separate 2448 * context 2449 */ 2450 generic_make_request(r1_bio->bios[r1_bio->read_disk]); 2451 2452 cond_resched(); 2453 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) 2454 md_check_recovery(mddev); 2455 } 2456 blk_finish_plug(&plug); 2457 } 2458 2459 2460 static int init_resync(struct r1conf *conf) 2461 { 2462 int buffs; 2463 2464 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; 2465 BUG_ON(conf->r1buf_pool); 2466 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, 2467 conf->poolinfo); 2468 if (!conf->r1buf_pool) 2469 return -ENOMEM; 2470 conf->next_resync = 0; 2471 return 0; 2472 } 2473 2474 /* 2475 * perform a "sync" on one "block" 2476 * 2477 * We need to make sure that no normal I/O request - particularly write 2478 * requests - conflict with active sync requests. 2479 * 2480 * This is achieved by tracking pending requests and a 'barrier' concept 2481 * that can be installed to exclude normal IO requests. 2482 */ 2483 2484 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster) 2485 { 2486 struct r1conf *conf = mddev->private; 2487 struct r1bio *r1_bio; 2488 struct bio *bio; 2489 sector_t max_sector, nr_sectors; 2490 int disk = -1; 2491 int i; 2492 int wonly = -1; 2493 int write_targets = 0, read_targets = 0; 2494 sector_t sync_blocks; 2495 int still_degraded = 0; 2496 int good_sectors = RESYNC_SECTORS; 2497 int min_bad = 0; /* number of sectors that are bad in all devices */ 2498 2499 if (!conf->r1buf_pool) 2500 if (init_resync(conf)) 2501 return 0; 2502 2503 max_sector = mddev->dev_sectors; 2504 if (sector_nr >= max_sector) { 2505 /* If we aborted, we need to abort the 2506 * sync on the 'current' bitmap chunk (there will 2507 * only be one in raid1 resync. 2508 * We can find the current addess in mddev->curr_resync 2509 */ 2510 if (mddev->curr_resync < max_sector) /* aborted */ 2511 bitmap_end_sync(mddev->bitmap, mddev->curr_resync, 2512 &sync_blocks, 1); 2513 else /* completed sync */ 2514 conf->fullsync = 0; 2515 2516 bitmap_close_sync(mddev->bitmap); 2517 close_sync(conf); 2518 return 0; 2519 } 2520 2521 if (mddev->bitmap == NULL && 2522 mddev->recovery_cp == MaxSector && 2523 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && 2524 conf->fullsync == 0) { 2525 *skipped = 1; 2526 return max_sector - sector_nr; 2527 } 2528 /* before building a request, check if we can skip these blocks.. 2529 * This call the bitmap_start_sync doesn't actually record anything 2530 */ 2531 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) && 2532 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2533 /* We can skip this block, and probably several more */ 2534 *skipped = 1; 2535 return sync_blocks; 2536 } 2537 /* 2538 * If there is non-resync activity waiting for a turn, 2539 * and resync is going fast enough, 2540 * then let it though before starting on this new sync request. 2541 */ 2542 if (!go_faster && conf->nr_waiting) 2543 msleep_interruptible(1000); 2544 2545 bitmap_cond_end_sync(mddev->bitmap, sector_nr); 2546 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO); 2547 2548 raise_barrier(conf, sector_nr); 2549 2550 rcu_read_lock(); 2551 /* 2552 * If we get a correctably read error during resync or recovery, 2553 * we might want to read from a different device. So we 2554 * flag all drives that could conceivably be read from for READ, 2555 * and any others (which will be non-In_sync devices) for WRITE. 2556 * If a read fails, we try reading from something else for which READ 2557 * is OK. 2558 */ 2559 2560 r1_bio->mddev = mddev; 2561 r1_bio->sector = sector_nr; 2562 r1_bio->state = 0; 2563 set_bit(R1BIO_IsSync, &r1_bio->state); 2564 2565 for (i = 0; i < conf->raid_disks * 2; i++) { 2566 struct md_rdev *rdev; 2567 bio = r1_bio->bios[i]; 2568 bio_reset(bio); 2569 2570 rdev = rcu_dereference(conf->mirrors[i].rdev); 2571 if (rdev == NULL || 2572 test_bit(Faulty, &rdev->flags)) { 2573 if (i < conf->raid_disks) 2574 still_degraded = 1; 2575 } else if (!test_bit(In_sync, &rdev->flags)) { 2576 bio->bi_rw = WRITE; 2577 bio->bi_end_io = end_sync_write; 2578 write_targets ++; 2579 } else { 2580 /* may need to read from here */ 2581 sector_t first_bad = MaxSector; 2582 int bad_sectors; 2583 2584 if (is_badblock(rdev, sector_nr, good_sectors, 2585 &first_bad, &bad_sectors)) { 2586 if (first_bad > sector_nr) 2587 good_sectors = first_bad - sector_nr; 2588 else { 2589 bad_sectors -= (sector_nr - first_bad); 2590 if (min_bad == 0 || 2591 min_bad > bad_sectors) 2592 min_bad = bad_sectors; 2593 } 2594 } 2595 if (sector_nr < first_bad) { 2596 if (test_bit(WriteMostly, &rdev->flags)) { 2597 if (wonly < 0) 2598 wonly = i; 2599 } else { 2600 if (disk < 0) 2601 disk = i; 2602 } 2603 bio->bi_rw = READ; 2604 bio->bi_end_io = end_sync_read; 2605 read_targets++; 2606 } else if (!test_bit(WriteErrorSeen, &rdev->flags) && 2607 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && 2608 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) { 2609 /* 2610 * The device is suitable for reading (InSync), 2611 * but has bad block(s) here. Let's try to correct them, 2612 * if we are doing resync or repair. Otherwise, leave 2613 * this device alone for this sync request. 2614 */ 2615 bio->bi_rw = WRITE; 2616 bio->bi_end_io = end_sync_write; 2617 write_targets++; 2618 } 2619 } 2620 if (bio->bi_end_io) { 2621 atomic_inc(&rdev->nr_pending); 2622 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset; 2623 bio->bi_bdev = rdev->bdev; 2624 bio->bi_private = r1_bio; 2625 } 2626 } 2627 rcu_read_unlock(); 2628 if (disk < 0) 2629 disk = wonly; 2630 r1_bio->read_disk = disk; 2631 2632 if (read_targets == 0 && min_bad > 0) { 2633 /* These sectors are bad on all InSync devices, so we 2634 * need to mark them bad on all write targets 2635 */ 2636 int ok = 1; 2637 for (i = 0 ; i < conf->raid_disks * 2 ; i++) 2638 if (r1_bio->bios[i]->bi_end_io == end_sync_write) { 2639 struct md_rdev *rdev = conf->mirrors[i].rdev; 2640 ok = rdev_set_badblocks(rdev, sector_nr, 2641 min_bad, 0 2642 ) && ok; 2643 } 2644 set_bit(MD_CHANGE_DEVS, &mddev->flags); 2645 *skipped = 1; 2646 put_buf(r1_bio); 2647 2648 if (!ok) { 2649 /* Cannot record the badblocks, so need to 2650 * abort the resync. 2651 * If there are multiple read targets, could just 2652 * fail the really bad ones ??? 2653 */ 2654 conf->recovery_disabled = mddev->recovery_disabled; 2655 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 2656 return 0; 2657 } else 2658 return min_bad; 2659 2660 } 2661 if (min_bad > 0 && min_bad < good_sectors) { 2662 /* only resync enough to reach the next bad->good 2663 * transition */ 2664 good_sectors = min_bad; 2665 } 2666 2667 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0) 2668 /* extra read targets are also write targets */ 2669 write_targets += read_targets-1; 2670 2671 if (write_targets == 0 || read_targets == 0) { 2672 /* There is nowhere to write, so all non-sync 2673 * drives must be failed - so we are finished 2674 */ 2675 sector_t rv; 2676 if (min_bad > 0) 2677 max_sector = sector_nr + min_bad; 2678 rv = max_sector - sector_nr; 2679 *skipped = 1; 2680 put_buf(r1_bio); 2681 return rv; 2682 } 2683 2684 if (max_sector > mddev->resync_max) 2685 max_sector = mddev->resync_max; /* Don't do IO beyond here */ 2686 if (max_sector > sector_nr + good_sectors) 2687 max_sector = sector_nr + good_sectors; 2688 nr_sectors = 0; 2689 sync_blocks = 0; 2690 do { 2691 struct page *page; 2692 int len = PAGE_SIZE; 2693 if (sector_nr + (len>>9) > max_sector) 2694 len = (max_sector - sector_nr) << 9; 2695 if (len == 0) 2696 break; 2697 if (sync_blocks == 0) { 2698 if (!bitmap_start_sync(mddev->bitmap, sector_nr, 2699 &sync_blocks, still_degraded) && 2700 !conf->fullsync && 2701 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2702 break; 2703 BUG_ON(sync_blocks < (PAGE_SIZE>>9)); 2704 if ((len >> 9) > sync_blocks) 2705 len = sync_blocks<<9; 2706 } 2707 2708 for (i = 0 ; i < conf->raid_disks * 2; i++) { 2709 bio = r1_bio->bios[i]; 2710 if (bio->bi_end_io) { 2711 page = bio->bi_io_vec[bio->bi_vcnt].bv_page; 2712 if (bio_add_page(bio, page, len, 0) == 0) { 2713 /* stop here */ 2714 bio->bi_io_vec[bio->bi_vcnt].bv_page = page; 2715 while (i > 0) { 2716 i--; 2717 bio = r1_bio->bios[i]; 2718 if (bio->bi_end_io==NULL) 2719 continue; 2720 /* remove last page from this bio */ 2721 bio->bi_vcnt--; 2722 bio->bi_iter.bi_size -= len; 2723 bio->bi_flags &= ~(1<< BIO_SEG_VALID); 2724 } 2725 goto bio_full; 2726 } 2727 } 2728 } 2729 nr_sectors += len>>9; 2730 sector_nr += len>>9; 2731 sync_blocks -= (len>>9); 2732 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES); 2733 bio_full: 2734 r1_bio->sectors = nr_sectors; 2735 2736 /* For a user-requested sync, we read all readable devices and do a 2737 * compare 2738 */ 2739 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2740 atomic_set(&r1_bio->remaining, read_targets); 2741 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) { 2742 bio = r1_bio->bios[i]; 2743 if (bio->bi_end_io == end_sync_read) { 2744 read_targets--; 2745 md_sync_acct(bio->bi_bdev, nr_sectors); 2746 generic_make_request(bio); 2747 } 2748 } 2749 } else { 2750 atomic_set(&r1_bio->remaining, 1); 2751 bio = r1_bio->bios[r1_bio->read_disk]; 2752 md_sync_acct(bio->bi_bdev, nr_sectors); 2753 generic_make_request(bio); 2754 2755 } 2756 return nr_sectors; 2757 } 2758 2759 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks) 2760 { 2761 if (sectors) 2762 return sectors; 2763 2764 return mddev->dev_sectors; 2765 } 2766 2767 static struct r1conf *setup_conf(struct mddev *mddev) 2768 { 2769 struct r1conf *conf; 2770 int i; 2771 struct raid1_info *disk; 2772 struct md_rdev *rdev; 2773 int err = -ENOMEM; 2774 2775 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL); 2776 if (!conf) 2777 goto abort; 2778 2779 conf->mirrors = kzalloc(sizeof(struct raid1_info) 2780 * mddev->raid_disks * 2, 2781 GFP_KERNEL); 2782 if (!conf->mirrors) 2783 goto abort; 2784 2785 conf->tmppage = alloc_page(GFP_KERNEL); 2786 if (!conf->tmppage) 2787 goto abort; 2788 2789 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL); 2790 if (!conf->poolinfo) 2791 goto abort; 2792 conf->poolinfo->raid_disks = mddev->raid_disks * 2; 2793 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 2794 r1bio_pool_free, 2795 conf->poolinfo); 2796 if (!conf->r1bio_pool) 2797 goto abort; 2798 2799 conf->poolinfo->mddev = mddev; 2800 2801 err = -EINVAL; 2802 spin_lock_init(&conf->device_lock); 2803 rdev_for_each(rdev, mddev) { 2804 struct request_queue *q; 2805 int disk_idx = rdev->raid_disk; 2806 if (disk_idx >= mddev->raid_disks 2807 || disk_idx < 0) 2808 continue; 2809 if (test_bit(Replacement, &rdev->flags)) 2810 disk = conf->mirrors + mddev->raid_disks + disk_idx; 2811 else 2812 disk = conf->mirrors + disk_idx; 2813 2814 if (disk->rdev) 2815 goto abort; 2816 disk->rdev = rdev; 2817 q = bdev_get_queue(rdev->bdev); 2818 if (q->merge_bvec_fn) 2819 mddev->merge_check_needed = 1; 2820 2821 disk->head_position = 0; 2822 disk->seq_start = MaxSector; 2823 } 2824 conf->raid_disks = mddev->raid_disks; 2825 conf->mddev = mddev; 2826 INIT_LIST_HEAD(&conf->retry_list); 2827 2828 spin_lock_init(&conf->resync_lock); 2829 init_waitqueue_head(&conf->wait_barrier); 2830 2831 bio_list_init(&conf->pending_bio_list); 2832 conf->pending_count = 0; 2833 conf->recovery_disabled = mddev->recovery_disabled - 1; 2834 2835 conf->start_next_window = MaxSector; 2836 conf->current_window_requests = conf->next_window_requests = 0; 2837 2838 err = -EIO; 2839 for (i = 0; i < conf->raid_disks * 2; i++) { 2840 2841 disk = conf->mirrors + i; 2842 2843 if (i < conf->raid_disks && 2844 disk[conf->raid_disks].rdev) { 2845 /* This slot has a replacement. */ 2846 if (!disk->rdev) { 2847 /* No original, just make the replacement 2848 * a recovering spare 2849 */ 2850 disk->rdev = 2851 disk[conf->raid_disks].rdev; 2852 disk[conf->raid_disks].rdev = NULL; 2853 } else if (!test_bit(In_sync, &disk->rdev->flags)) 2854 /* Original is not in_sync - bad */ 2855 goto abort; 2856 } 2857 2858 if (!disk->rdev || 2859 !test_bit(In_sync, &disk->rdev->flags)) { 2860 disk->head_position = 0; 2861 if (disk->rdev && 2862 (disk->rdev->saved_raid_disk < 0)) 2863 conf->fullsync = 1; 2864 } 2865 } 2866 2867 err = -ENOMEM; 2868 conf->thread = md_register_thread(raid1d, mddev, "raid1"); 2869 if (!conf->thread) { 2870 printk(KERN_ERR 2871 "md/raid1:%s: couldn't allocate thread\n", 2872 mdname(mddev)); 2873 goto abort; 2874 } 2875 2876 return conf; 2877 2878 abort: 2879 if (conf) { 2880 if (conf->r1bio_pool) 2881 mempool_destroy(conf->r1bio_pool); 2882 kfree(conf->mirrors); 2883 safe_put_page(conf->tmppage); 2884 kfree(conf->poolinfo); 2885 kfree(conf); 2886 } 2887 return ERR_PTR(err); 2888 } 2889 2890 static int stop(struct mddev *mddev); 2891 static int run(struct mddev *mddev) 2892 { 2893 struct r1conf *conf; 2894 int i; 2895 struct md_rdev *rdev; 2896 int ret; 2897 bool discard_supported = false; 2898 2899 if (mddev->level != 1) { 2900 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n", 2901 mdname(mddev), mddev->level); 2902 return -EIO; 2903 } 2904 if (mddev->reshape_position != MaxSector) { 2905 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n", 2906 mdname(mddev)); 2907 return -EIO; 2908 } 2909 /* 2910 * copy the already verified devices into our private RAID1 2911 * bookkeeping area. [whatever we allocate in run(), 2912 * should be freed in stop()] 2913 */ 2914 if (mddev->private == NULL) 2915 conf = setup_conf(mddev); 2916 else 2917 conf = mddev->private; 2918 2919 if (IS_ERR(conf)) 2920 return PTR_ERR(conf); 2921 2922 if (mddev->queue) 2923 blk_queue_max_write_same_sectors(mddev->queue, 0); 2924 2925 rdev_for_each(rdev, mddev) { 2926 if (!mddev->gendisk) 2927 continue; 2928 disk_stack_limits(mddev->gendisk, rdev->bdev, 2929 rdev->data_offset << 9); 2930 if (blk_queue_discard(bdev_get_queue(rdev->bdev))) 2931 discard_supported = true; 2932 } 2933 2934 mddev->degraded = 0; 2935 for (i=0; i < conf->raid_disks; i++) 2936 if (conf->mirrors[i].rdev == NULL || 2937 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) || 2938 test_bit(Faulty, &conf->mirrors[i].rdev->flags)) 2939 mddev->degraded++; 2940 2941 if (conf->raid_disks - mddev->degraded == 1) 2942 mddev->recovery_cp = MaxSector; 2943 2944 if (mddev->recovery_cp != MaxSector) 2945 printk(KERN_NOTICE "md/raid1:%s: not clean" 2946 " -- starting background reconstruction\n", 2947 mdname(mddev)); 2948 printk(KERN_INFO 2949 "md/raid1:%s: active with %d out of %d mirrors\n", 2950 mdname(mddev), mddev->raid_disks - mddev->degraded, 2951 mddev->raid_disks); 2952 2953 /* 2954 * Ok, everything is just fine now 2955 */ 2956 mddev->thread = conf->thread; 2957 conf->thread = NULL; 2958 mddev->private = conf; 2959 2960 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0)); 2961 2962 if (mddev->queue) { 2963 mddev->queue->backing_dev_info.congested_fn = raid1_congested; 2964 mddev->queue->backing_dev_info.congested_data = mddev; 2965 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec); 2966 2967 if (discard_supported) 2968 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, 2969 mddev->queue); 2970 else 2971 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, 2972 mddev->queue); 2973 } 2974 2975 ret = md_integrity_register(mddev); 2976 if (ret) 2977 stop(mddev); 2978 return ret; 2979 } 2980 2981 static int stop(struct mddev *mddev) 2982 { 2983 struct r1conf *conf = mddev->private; 2984 struct bitmap *bitmap = mddev->bitmap; 2985 2986 /* wait for behind writes to complete */ 2987 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) { 2988 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n", 2989 mdname(mddev)); 2990 /* need to kick something here to make sure I/O goes? */ 2991 wait_event(bitmap->behind_wait, 2992 atomic_read(&bitmap->behind_writes) == 0); 2993 } 2994 2995 freeze_array(conf, 0); 2996 unfreeze_array(conf); 2997 2998 md_unregister_thread(&mddev->thread); 2999 if (conf->r1bio_pool) 3000 mempool_destroy(conf->r1bio_pool); 3001 kfree(conf->mirrors); 3002 safe_put_page(conf->tmppage); 3003 kfree(conf->poolinfo); 3004 kfree(conf); 3005 mddev->private = NULL; 3006 return 0; 3007 } 3008 3009 static int raid1_resize(struct mddev *mddev, sector_t sectors) 3010 { 3011 /* no resync is happening, and there is enough space 3012 * on all devices, so we can resize. 3013 * We need to make sure resync covers any new space. 3014 * If the array is shrinking we should possibly wait until 3015 * any io in the removed space completes, but it hardly seems 3016 * worth it. 3017 */ 3018 sector_t newsize = raid1_size(mddev, sectors, 0); 3019 if (mddev->external_size && 3020 mddev->array_sectors > newsize) 3021 return -EINVAL; 3022 if (mddev->bitmap) { 3023 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0); 3024 if (ret) 3025 return ret; 3026 } 3027 md_set_array_sectors(mddev, newsize); 3028 set_capacity(mddev->gendisk, mddev->array_sectors); 3029 revalidate_disk(mddev->gendisk); 3030 if (sectors > mddev->dev_sectors && 3031 mddev->recovery_cp > mddev->dev_sectors) { 3032 mddev->recovery_cp = mddev->dev_sectors; 3033 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3034 } 3035 mddev->dev_sectors = sectors; 3036 mddev->resync_max_sectors = sectors; 3037 return 0; 3038 } 3039 3040 static int raid1_reshape(struct mddev *mddev) 3041 { 3042 /* We need to: 3043 * 1/ resize the r1bio_pool 3044 * 2/ resize conf->mirrors 3045 * 3046 * We allocate a new r1bio_pool if we can. 3047 * Then raise a device barrier and wait until all IO stops. 3048 * Then resize conf->mirrors and swap in the new r1bio pool. 3049 * 3050 * At the same time, we "pack" the devices so that all the missing 3051 * devices have the higher raid_disk numbers. 3052 */ 3053 mempool_t *newpool, *oldpool; 3054 struct pool_info *newpoolinfo; 3055 struct raid1_info *newmirrors; 3056 struct r1conf *conf = mddev->private; 3057 int cnt, raid_disks; 3058 unsigned long flags; 3059 int d, d2, err; 3060 3061 /* Cannot change chunk_size, layout, or level */ 3062 if (mddev->chunk_sectors != mddev->new_chunk_sectors || 3063 mddev->layout != mddev->new_layout || 3064 mddev->level != mddev->new_level) { 3065 mddev->new_chunk_sectors = mddev->chunk_sectors; 3066 mddev->new_layout = mddev->layout; 3067 mddev->new_level = mddev->level; 3068 return -EINVAL; 3069 } 3070 3071 err = md_allow_write(mddev); 3072 if (err) 3073 return err; 3074 3075 raid_disks = mddev->raid_disks + mddev->delta_disks; 3076 3077 if (raid_disks < conf->raid_disks) { 3078 cnt=0; 3079 for (d= 0; d < conf->raid_disks; d++) 3080 if (conf->mirrors[d].rdev) 3081 cnt++; 3082 if (cnt > raid_disks) 3083 return -EBUSY; 3084 } 3085 3086 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL); 3087 if (!newpoolinfo) 3088 return -ENOMEM; 3089 newpoolinfo->mddev = mddev; 3090 newpoolinfo->raid_disks = raid_disks * 2; 3091 3092 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 3093 r1bio_pool_free, newpoolinfo); 3094 if (!newpool) { 3095 kfree(newpoolinfo); 3096 return -ENOMEM; 3097 } 3098 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2, 3099 GFP_KERNEL); 3100 if (!newmirrors) { 3101 kfree(newpoolinfo); 3102 mempool_destroy(newpool); 3103 return -ENOMEM; 3104 } 3105 3106 freeze_array(conf, 0); 3107 3108 /* ok, everything is stopped */ 3109 oldpool = conf->r1bio_pool; 3110 conf->r1bio_pool = newpool; 3111 3112 for (d = d2 = 0; d < conf->raid_disks; d++) { 3113 struct md_rdev *rdev = conf->mirrors[d].rdev; 3114 if (rdev && rdev->raid_disk != d2) { 3115 sysfs_unlink_rdev(mddev, rdev); 3116 rdev->raid_disk = d2; 3117 sysfs_unlink_rdev(mddev, rdev); 3118 if (sysfs_link_rdev(mddev, rdev)) 3119 printk(KERN_WARNING 3120 "md/raid1:%s: cannot register rd%d\n", 3121 mdname(mddev), rdev->raid_disk); 3122 } 3123 if (rdev) 3124 newmirrors[d2++].rdev = rdev; 3125 } 3126 kfree(conf->mirrors); 3127 conf->mirrors = newmirrors; 3128 kfree(conf->poolinfo); 3129 conf->poolinfo = newpoolinfo; 3130 3131 spin_lock_irqsave(&conf->device_lock, flags); 3132 mddev->degraded += (raid_disks - conf->raid_disks); 3133 spin_unlock_irqrestore(&conf->device_lock, flags); 3134 conf->raid_disks = mddev->raid_disks = raid_disks; 3135 mddev->delta_disks = 0; 3136 3137 unfreeze_array(conf); 3138 3139 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3140 md_wakeup_thread(mddev->thread); 3141 3142 mempool_destroy(oldpool); 3143 return 0; 3144 } 3145 3146 static void raid1_quiesce(struct mddev *mddev, int state) 3147 { 3148 struct r1conf *conf = mddev->private; 3149 3150 switch(state) { 3151 case 2: /* wake for suspend */ 3152 wake_up(&conf->wait_barrier); 3153 break; 3154 case 1: 3155 freeze_array(conf, 0); 3156 break; 3157 case 0: 3158 unfreeze_array(conf); 3159 break; 3160 } 3161 } 3162 3163 static void *raid1_takeover(struct mddev *mddev) 3164 { 3165 /* raid1 can take over: 3166 * raid5 with 2 devices, any layout or chunk size 3167 */ 3168 if (mddev->level == 5 && mddev->raid_disks == 2) { 3169 struct r1conf *conf; 3170 mddev->new_level = 1; 3171 mddev->new_layout = 0; 3172 mddev->new_chunk_sectors = 0; 3173 conf = setup_conf(mddev); 3174 if (!IS_ERR(conf)) 3175 /* Array must appear to be quiesced */ 3176 conf->array_frozen = 1; 3177 return conf; 3178 } 3179 return ERR_PTR(-EINVAL); 3180 } 3181 3182 static struct md_personality raid1_personality = 3183 { 3184 .name = "raid1", 3185 .level = 1, 3186 .owner = THIS_MODULE, 3187 .make_request = make_request, 3188 .run = run, 3189 .stop = stop, 3190 .status = status, 3191 .error_handler = error, 3192 .hot_add_disk = raid1_add_disk, 3193 .hot_remove_disk= raid1_remove_disk, 3194 .spare_active = raid1_spare_active, 3195 .sync_request = sync_request, 3196 .resize = raid1_resize, 3197 .size = raid1_size, 3198 .check_reshape = raid1_reshape, 3199 .quiesce = raid1_quiesce, 3200 .takeover = raid1_takeover, 3201 }; 3202 3203 static int __init raid_init(void) 3204 { 3205 return register_md_personality(&raid1_personality); 3206 } 3207 3208 static void raid_exit(void) 3209 { 3210 unregister_md_personality(&raid1_personality); 3211 } 3212 3213 module_init(raid_init); 3214 module_exit(raid_exit); 3215 MODULE_LICENSE("GPL"); 3216 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD"); 3217 MODULE_ALIAS("md-personality-3"); /* RAID1 */ 3218 MODULE_ALIAS("md-raid1"); 3219 MODULE_ALIAS("md-level-1"); 3220 3221 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR); 3222