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) 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++; 840 841 /* For these conditions we must wait: 842 * A: while the array is in frozen state 843 * B: while barrier >= RESYNC_DEPTH, meaning resync reach 844 * the max count which allowed. 845 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning 846 * next resync will reach to the window which normal bios are 847 * handling. 848 * D: while there are any active requests in the current window. 849 */ 850 wait_event_lock_irq(conf->wait_barrier, 851 !conf->array_frozen && 852 conf->barrier < RESYNC_DEPTH && 853 conf->current_window_requests == 0 && 854 (conf->start_next_window >= 855 conf->next_resync + RESYNC_SECTORS), 856 conf->resync_lock); 857 858 spin_unlock_irq(&conf->resync_lock); 859 } 860 861 static void lower_barrier(struct r1conf *conf) 862 { 863 unsigned long flags; 864 BUG_ON(conf->barrier <= 0); 865 spin_lock_irqsave(&conf->resync_lock, flags); 866 conf->barrier--; 867 spin_unlock_irqrestore(&conf->resync_lock, flags); 868 wake_up(&conf->wait_barrier); 869 } 870 871 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio) 872 { 873 bool wait = false; 874 875 if (conf->array_frozen || !bio) 876 wait = true; 877 else if (conf->barrier && bio_data_dir(bio) == WRITE) { 878 if (conf->next_resync < RESYNC_WINDOW_SECTORS) 879 wait = true; 880 else if ((conf->next_resync - RESYNC_WINDOW_SECTORS 881 >= bio_end_sector(bio)) || 882 (conf->next_resync + NEXT_NORMALIO_DISTANCE 883 <= bio->bi_iter.bi_sector)) 884 wait = false; 885 else 886 wait = true; 887 } 888 889 return wait; 890 } 891 892 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio) 893 { 894 sector_t sector = 0; 895 896 spin_lock_irq(&conf->resync_lock); 897 if (need_to_wait_for_sync(conf, bio)) { 898 conf->nr_waiting++; 899 /* Wait for the barrier to drop. 900 * However if there are already pending 901 * requests (preventing the barrier from 902 * rising completely), and the 903 * pre-process bio queue isn't empty, 904 * then don't wait, as we need to empty 905 * that queue to get the nr_pending 906 * count down. 907 */ 908 wait_event_lock_irq(conf->wait_barrier, 909 !conf->array_frozen && 910 (!conf->barrier || 911 ((conf->start_next_window < 912 conf->next_resync + RESYNC_SECTORS) && 913 current->bio_list && 914 !bio_list_empty(current->bio_list))), 915 conf->resync_lock); 916 conf->nr_waiting--; 917 } 918 919 if (bio && bio_data_dir(bio) == WRITE) { 920 if (bio->bi_iter.bi_sector >= 921 conf->next_resync) { 922 if (conf->start_next_window == MaxSector) 923 conf->start_next_window = 924 conf->next_resync + 925 NEXT_NORMALIO_DISTANCE; 926 927 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE) 928 <= bio->bi_iter.bi_sector) 929 conf->next_window_requests++; 930 else 931 conf->current_window_requests++; 932 sector = conf->start_next_window; 933 } 934 } 935 936 conf->nr_pending++; 937 spin_unlock_irq(&conf->resync_lock); 938 return sector; 939 } 940 941 static void allow_barrier(struct r1conf *conf, sector_t start_next_window, 942 sector_t bi_sector) 943 { 944 unsigned long flags; 945 946 spin_lock_irqsave(&conf->resync_lock, flags); 947 conf->nr_pending--; 948 if (start_next_window) { 949 if (start_next_window == conf->start_next_window) { 950 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE 951 <= bi_sector) 952 conf->next_window_requests--; 953 else 954 conf->current_window_requests--; 955 } else 956 conf->current_window_requests--; 957 958 if (!conf->current_window_requests) { 959 if (conf->next_window_requests) { 960 conf->current_window_requests = 961 conf->next_window_requests; 962 conf->next_window_requests = 0; 963 conf->start_next_window += 964 NEXT_NORMALIO_DISTANCE; 965 } else 966 conf->start_next_window = MaxSector; 967 } 968 } 969 spin_unlock_irqrestore(&conf->resync_lock, flags); 970 wake_up(&conf->wait_barrier); 971 } 972 973 static void freeze_array(struct r1conf *conf, int extra) 974 { 975 /* stop syncio and normal IO and wait for everything to 976 * go quite. 977 * We wait until nr_pending match nr_queued+extra 978 * This is called in the context of one normal IO request 979 * that has failed. Thus any sync request that might be pending 980 * will be blocked by nr_pending, and we need to wait for 981 * pending IO requests to complete or be queued for re-try. 982 * Thus the number queued (nr_queued) plus this request (extra) 983 * must match the number of pending IOs (nr_pending) before 984 * we continue. 985 */ 986 spin_lock_irq(&conf->resync_lock); 987 conf->array_frozen = 1; 988 wait_event_lock_irq_cmd(conf->wait_barrier, 989 conf->nr_pending == conf->nr_queued+extra, 990 conf->resync_lock, 991 flush_pending_writes(conf)); 992 spin_unlock_irq(&conf->resync_lock); 993 } 994 static void unfreeze_array(struct r1conf *conf) 995 { 996 /* reverse the effect of the freeze */ 997 spin_lock_irq(&conf->resync_lock); 998 conf->array_frozen = 0; 999 wake_up(&conf->wait_barrier); 1000 spin_unlock_irq(&conf->resync_lock); 1001 } 1002 1003 1004 /* duplicate the data pages for behind I/O 1005 */ 1006 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio) 1007 { 1008 int i; 1009 struct bio_vec *bvec; 1010 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec), 1011 GFP_NOIO); 1012 if (unlikely(!bvecs)) 1013 return; 1014 1015 bio_for_each_segment_all(bvec, bio, i) { 1016 bvecs[i] = *bvec; 1017 bvecs[i].bv_page = alloc_page(GFP_NOIO); 1018 if (unlikely(!bvecs[i].bv_page)) 1019 goto do_sync_io; 1020 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset, 1021 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len); 1022 kunmap(bvecs[i].bv_page); 1023 kunmap(bvec->bv_page); 1024 } 1025 r1_bio->behind_bvecs = bvecs; 1026 r1_bio->behind_page_count = bio->bi_vcnt; 1027 set_bit(R1BIO_BehindIO, &r1_bio->state); 1028 return; 1029 1030 do_sync_io: 1031 for (i = 0; i < bio->bi_vcnt; i++) 1032 if (bvecs[i].bv_page) 1033 put_page(bvecs[i].bv_page); 1034 kfree(bvecs); 1035 pr_debug("%dB behind alloc failed, doing sync I/O\n", 1036 bio->bi_iter.bi_size); 1037 } 1038 1039 struct raid1_plug_cb { 1040 struct blk_plug_cb cb; 1041 struct bio_list pending; 1042 int pending_cnt; 1043 }; 1044 1045 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule) 1046 { 1047 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, 1048 cb); 1049 struct mddev *mddev = plug->cb.data; 1050 struct r1conf *conf = mddev->private; 1051 struct bio *bio; 1052 1053 if (from_schedule || current->bio_list) { 1054 spin_lock_irq(&conf->device_lock); 1055 bio_list_merge(&conf->pending_bio_list, &plug->pending); 1056 conf->pending_count += plug->pending_cnt; 1057 spin_unlock_irq(&conf->device_lock); 1058 wake_up(&conf->wait_barrier); 1059 md_wakeup_thread(mddev->thread); 1060 kfree(plug); 1061 return; 1062 } 1063 1064 /* we aren't scheduling, so we can do the write-out directly. */ 1065 bio = bio_list_get(&plug->pending); 1066 bitmap_unplug(mddev->bitmap); 1067 wake_up(&conf->wait_barrier); 1068 1069 while (bio) { /* submit pending writes */ 1070 struct bio *next = bio->bi_next; 1071 bio->bi_next = NULL; 1072 if (unlikely((bio->bi_rw & REQ_DISCARD) && 1073 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 1074 /* Just ignore it */ 1075 bio_endio(bio, 0); 1076 else 1077 generic_make_request(bio); 1078 bio = next; 1079 } 1080 kfree(plug); 1081 } 1082 1083 static void make_request(struct mddev *mddev, struct bio * bio) 1084 { 1085 struct r1conf *conf = mddev->private; 1086 struct raid1_info *mirror; 1087 struct r1bio *r1_bio; 1088 struct bio *read_bio; 1089 int i, disks; 1090 struct bitmap *bitmap; 1091 unsigned long flags; 1092 const int rw = bio_data_dir(bio); 1093 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC); 1094 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA)); 1095 const unsigned long do_discard = (bio->bi_rw 1096 & (REQ_DISCARD | REQ_SECURE)); 1097 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME); 1098 struct md_rdev *blocked_rdev; 1099 struct blk_plug_cb *cb; 1100 struct raid1_plug_cb *plug = NULL; 1101 int first_clone; 1102 int sectors_handled; 1103 int max_sectors; 1104 sector_t start_next_window; 1105 1106 /* 1107 * Register the new request and wait if the reconstruction 1108 * thread has put up a bar for new requests. 1109 * Continue immediately if no resync is active currently. 1110 */ 1111 1112 md_write_start(mddev, bio); /* wait on superblock update early */ 1113 1114 if (bio_data_dir(bio) == WRITE && 1115 bio_end_sector(bio) > mddev->suspend_lo && 1116 bio->bi_iter.bi_sector < mddev->suspend_hi) { 1117 /* As the suspend_* range is controlled by 1118 * userspace, we want an interruptible 1119 * wait. 1120 */ 1121 DEFINE_WAIT(w); 1122 for (;;) { 1123 flush_signals(current); 1124 prepare_to_wait(&conf->wait_barrier, 1125 &w, TASK_INTERRUPTIBLE); 1126 if (bio_end_sector(bio) <= mddev->suspend_lo || 1127 bio->bi_iter.bi_sector >= mddev->suspend_hi) 1128 break; 1129 schedule(); 1130 } 1131 finish_wait(&conf->wait_barrier, &w); 1132 } 1133 1134 start_next_window = wait_barrier(conf, bio); 1135 1136 bitmap = mddev->bitmap; 1137 1138 /* 1139 * make_request() can abort the operation when READA is being 1140 * used and no empty request is available. 1141 * 1142 */ 1143 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1144 1145 r1_bio->master_bio = bio; 1146 r1_bio->sectors = bio_sectors(bio); 1147 r1_bio->state = 0; 1148 r1_bio->mddev = mddev; 1149 r1_bio->sector = bio->bi_iter.bi_sector; 1150 1151 /* We might need to issue multiple reads to different 1152 * devices if there are bad blocks around, so we keep 1153 * track of the number of reads in bio->bi_phys_segments. 1154 * If this is 0, there is only one r1_bio and no locking 1155 * will be needed when requests complete. If it is 1156 * non-zero, then it is the number of not-completed requests. 1157 */ 1158 bio->bi_phys_segments = 0; 1159 clear_bit(BIO_SEG_VALID, &bio->bi_flags); 1160 1161 if (rw == READ) { 1162 /* 1163 * read balancing logic: 1164 */ 1165 int rdisk; 1166 1167 read_again: 1168 rdisk = read_balance(conf, r1_bio, &max_sectors); 1169 1170 if (rdisk < 0) { 1171 /* couldn't find anywhere to read from */ 1172 raid_end_bio_io(r1_bio); 1173 return; 1174 } 1175 mirror = conf->mirrors + rdisk; 1176 1177 if (test_bit(WriteMostly, &mirror->rdev->flags) && 1178 bitmap) { 1179 /* Reading from a write-mostly device must 1180 * take care not to over-take any writes 1181 * that are 'behind' 1182 */ 1183 wait_event(bitmap->behind_wait, 1184 atomic_read(&bitmap->behind_writes) == 0); 1185 } 1186 r1_bio->read_disk = rdisk; 1187 r1_bio->start_next_window = 0; 1188 1189 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1190 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector, 1191 max_sectors); 1192 1193 r1_bio->bios[rdisk] = read_bio; 1194 1195 read_bio->bi_iter.bi_sector = r1_bio->sector + 1196 mirror->rdev->data_offset; 1197 read_bio->bi_bdev = mirror->rdev->bdev; 1198 read_bio->bi_end_io = raid1_end_read_request; 1199 read_bio->bi_rw = READ | do_sync; 1200 read_bio->bi_private = r1_bio; 1201 1202 if (max_sectors < r1_bio->sectors) { 1203 /* could not read all from this device, so we will 1204 * need another r1_bio. 1205 */ 1206 1207 sectors_handled = (r1_bio->sector + max_sectors 1208 - bio->bi_iter.bi_sector); 1209 r1_bio->sectors = max_sectors; 1210 spin_lock_irq(&conf->device_lock); 1211 if (bio->bi_phys_segments == 0) 1212 bio->bi_phys_segments = 2; 1213 else 1214 bio->bi_phys_segments++; 1215 spin_unlock_irq(&conf->device_lock); 1216 /* Cannot call generic_make_request directly 1217 * as that will be queued in __make_request 1218 * and subsequent mempool_alloc might block waiting 1219 * for it. So hand bio over to raid1d. 1220 */ 1221 reschedule_retry(r1_bio); 1222 1223 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1224 1225 r1_bio->master_bio = bio; 1226 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1227 r1_bio->state = 0; 1228 r1_bio->mddev = mddev; 1229 r1_bio->sector = bio->bi_iter.bi_sector + 1230 sectors_handled; 1231 goto read_again; 1232 } else 1233 generic_make_request(read_bio); 1234 return; 1235 } 1236 1237 /* 1238 * WRITE: 1239 */ 1240 if (conf->pending_count >= max_queued_requests) { 1241 md_wakeup_thread(mddev->thread); 1242 wait_event(conf->wait_barrier, 1243 conf->pending_count < max_queued_requests); 1244 } 1245 /* first select target devices under rcu_lock and 1246 * inc refcount on their rdev. Record them by setting 1247 * bios[x] to bio 1248 * If there are known/acknowledged bad blocks on any device on 1249 * which we have seen a write error, we want to avoid writing those 1250 * blocks. 1251 * This potentially requires several writes to write around 1252 * the bad blocks. Each set of writes gets it's own r1bio 1253 * with a set of bios attached. 1254 */ 1255 1256 disks = conf->raid_disks * 2; 1257 retry_write: 1258 r1_bio->start_next_window = start_next_window; 1259 blocked_rdev = NULL; 1260 rcu_read_lock(); 1261 max_sectors = r1_bio->sectors; 1262 for (i = 0; i < disks; i++) { 1263 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1264 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 1265 atomic_inc(&rdev->nr_pending); 1266 blocked_rdev = rdev; 1267 break; 1268 } 1269 r1_bio->bios[i] = NULL; 1270 if (!rdev || test_bit(Faulty, &rdev->flags) 1271 || test_bit(Unmerged, &rdev->flags)) { 1272 if (i < conf->raid_disks) 1273 set_bit(R1BIO_Degraded, &r1_bio->state); 1274 continue; 1275 } 1276 1277 atomic_inc(&rdev->nr_pending); 1278 if (test_bit(WriteErrorSeen, &rdev->flags)) { 1279 sector_t first_bad; 1280 int bad_sectors; 1281 int is_bad; 1282 1283 is_bad = is_badblock(rdev, r1_bio->sector, 1284 max_sectors, 1285 &first_bad, &bad_sectors); 1286 if (is_bad < 0) { 1287 /* mustn't write here until the bad block is 1288 * acknowledged*/ 1289 set_bit(BlockedBadBlocks, &rdev->flags); 1290 blocked_rdev = rdev; 1291 break; 1292 } 1293 if (is_bad && first_bad <= r1_bio->sector) { 1294 /* Cannot write here at all */ 1295 bad_sectors -= (r1_bio->sector - first_bad); 1296 if (bad_sectors < max_sectors) 1297 /* mustn't write more than bad_sectors 1298 * to other devices yet 1299 */ 1300 max_sectors = bad_sectors; 1301 rdev_dec_pending(rdev, mddev); 1302 /* We don't set R1BIO_Degraded as that 1303 * only applies if the disk is 1304 * missing, so it might be re-added, 1305 * and we want to know to recover this 1306 * chunk. 1307 * In this case the device is here, 1308 * and the fact that this chunk is not 1309 * in-sync is recorded in the bad 1310 * block log 1311 */ 1312 continue; 1313 } 1314 if (is_bad) { 1315 int good_sectors = first_bad - r1_bio->sector; 1316 if (good_sectors < max_sectors) 1317 max_sectors = good_sectors; 1318 } 1319 } 1320 r1_bio->bios[i] = bio; 1321 } 1322 rcu_read_unlock(); 1323 1324 if (unlikely(blocked_rdev)) { 1325 /* Wait for this device to become unblocked */ 1326 int j; 1327 sector_t old = start_next_window; 1328 1329 for (j = 0; j < i; j++) 1330 if (r1_bio->bios[j]) 1331 rdev_dec_pending(conf->mirrors[j].rdev, mddev); 1332 r1_bio->state = 0; 1333 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector); 1334 md_wait_for_blocked_rdev(blocked_rdev, mddev); 1335 start_next_window = wait_barrier(conf, bio); 1336 /* 1337 * We must make sure the multi r1bios of bio have 1338 * the same value of bi_phys_segments 1339 */ 1340 if (bio->bi_phys_segments && old && 1341 old != start_next_window) 1342 /* Wait for the former r1bio(s) to complete */ 1343 wait_event(conf->wait_barrier, 1344 bio->bi_phys_segments == 1); 1345 goto retry_write; 1346 } 1347 1348 if (max_sectors < r1_bio->sectors) { 1349 /* We are splitting this write into multiple parts, so 1350 * we need to prepare for allocating another r1_bio. 1351 */ 1352 r1_bio->sectors = max_sectors; 1353 spin_lock_irq(&conf->device_lock); 1354 if (bio->bi_phys_segments == 0) 1355 bio->bi_phys_segments = 2; 1356 else 1357 bio->bi_phys_segments++; 1358 spin_unlock_irq(&conf->device_lock); 1359 } 1360 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector; 1361 1362 atomic_set(&r1_bio->remaining, 1); 1363 atomic_set(&r1_bio->behind_remaining, 0); 1364 1365 first_clone = 1; 1366 for (i = 0; i < disks; i++) { 1367 struct bio *mbio; 1368 if (!r1_bio->bios[i]) 1369 continue; 1370 1371 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1372 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors); 1373 1374 if (first_clone) { 1375 /* do behind I/O ? 1376 * Not if there are too many, or cannot 1377 * allocate memory, or a reader on WriteMostly 1378 * is waiting for behind writes to flush */ 1379 if (bitmap && 1380 (atomic_read(&bitmap->behind_writes) 1381 < mddev->bitmap_info.max_write_behind) && 1382 !waitqueue_active(&bitmap->behind_wait)) 1383 alloc_behind_pages(mbio, r1_bio); 1384 1385 bitmap_startwrite(bitmap, r1_bio->sector, 1386 r1_bio->sectors, 1387 test_bit(R1BIO_BehindIO, 1388 &r1_bio->state)); 1389 first_clone = 0; 1390 } 1391 if (r1_bio->behind_bvecs) { 1392 struct bio_vec *bvec; 1393 int j; 1394 1395 /* 1396 * We trimmed the bio, so _all is legit 1397 */ 1398 bio_for_each_segment_all(bvec, mbio, j) 1399 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page; 1400 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags)) 1401 atomic_inc(&r1_bio->behind_remaining); 1402 } 1403 1404 r1_bio->bios[i] = mbio; 1405 1406 mbio->bi_iter.bi_sector = (r1_bio->sector + 1407 conf->mirrors[i].rdev->data_offset); 1408 mbio->bi_bdev = conf->mirrors[i].rdev->bdev; 1409 mbio->bi_end_io = raid1_end_write_request; 1410 mbio->bi_rw = 1411 WRITE | do_flush_fua | do_sync | do_discard | do_same; 1412 mbio->bi_private = r1_bio; 1413 1414 atomic_inc(&r1_bio->remaining); 1415 1416 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug)); 1417 if (cb) 1418 plug = container_of(cb, struct raid1_plug_cb, cb); 1419 else 1420 plug = NULL; 1421 spin_lock_irqsave(&conf->device_lock, flags); 1422 if (plug) { 1423 bio_list_add(&plug->pending, mbio); 1424 plug->pending_cnt++; 1425 } else { 1426 bio_list_add(&conf->pending_bio_list, mbio); 1427 conf->pending_count++; 1428 } 1429 spin_unlock_irqrestore(&conf->device_lock, flags); 1430 if (!plug) 1431 md_wakeup_thread(mddev->thread); 1432 } 1433 /* Mustn't call r1_bio_write_done before this next test, 1434 * as it could result in the bio being freed. 1435 */ 1436 if (sectors_handled < bio_sectors(bio)) { 1437 r1_bio_write_done(r1_bio); 1438 /* We need another r1_bio. It has already been counted 1439 * in bio->bi_phys_segments 1440 */ 1441 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1442 r1_bio->master_bio = bio; 1443 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1444 r1_bio->state = 0; 1445 r1_bio->mddev = mddev; 1446 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled; 1447 goto retry_write; 1448 } 1449 1450 r1_bio_write_done(r1_bio); 1451 1452 /* In case raid1d snuck in to freeze_array */ 1453 wake_up(&conf->wait_barrier); 1454 } 1455 1456 static void status(struct seq_file *seq, struct mddev *mddev) 1457 { 1458 struct r1conf *conf = mddev->private; 1459 int i; 1460 1461 seq_printf(seq, " [%d/%d] [", conf->raid_disks, 1462 conf->raid_disks - mddev->degraded); 1463 rcu_read_lock(); 1464 for (i = 0; i < conf->raid_disks; i++) { 1465 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1466 seq_printf(seq, "%s", 1467 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_"); 1468 } 1469 rcu_read_unlock(); 1470 seq_printf(seq, "]"); 1471 } 1472 1473 1474 static void error(struct mddev *mddev, struct md_rdev *rdev) 1475 { 1476 char b[BDEVNAME_SIZE]; 1477 struct r1conf *conf = mddev->private; 1478 1479 /* 1480 * If it is not operational, then we have already marked it as dead 1481 * else if it is the last working disks, ignore the error, let the 1482 * next level up know. 1483 * else mark the drive as failed 1484 */ 1485 if (test_bit(In_sync, &rdev->flags) 1486 && (conf->raid_disks - mddev->degraded) == 1) { 1487 /* 1488 * Don't fail the drive, act as though we were just a 1489 * normal single drive. 1490 * However don't try a recovery from this drive as 1491 * it is very likely to fail. 1492 */ 1493 conf->recovery_disabled = mddev->recovery_disabled; 1494 return; 1495 } 1496 set_bit(Blocked, &rdev->flags); 1497 if (test_and_clear_bit(In_sync, &rdev->flags)) { 1498 unsigned long flags; 1499 spin_lock_irqsave(&conf->device_lock, flags); 1500 mddev->degraded++; 1501 set_bit(Faulty, &rdev->flags); 1502 spin_unlock_irqrestore(&conf->device_lock, flags); 1503 } else 1504 set_bit(Faulty, &rdev->flags); 1505 /* 1506 * if recovery is running, make sure it aborts. 1507 */ 1508 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1509 set_bit(MD_CHANGE_DEVS, &mddev->flags); 1510 printk(KERN_ALERT 1511 "md/raid1:%s: Disk failure on %s, disabling device.\n" 1512 "md/raid1:%s: Operation continuing on %d devices.\n", 1513 mdname(mddev), bdevname(rdev->bdev, b), 1514 mdname(mddev), conf->raid_disks - mddev->degraded); 1515 } 1516 1517 static void print_conf(struct r1conf *conf) 1518 { 1519 int i; 1520 1521 printk(KERN_DEBUG "RAID1 conf printout:\n"); 1522 if (!conf) { 1523 printk(KERN_DEBUG "(!conf)\n"); 1524 return; 1525 } 1526 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded, 1527 conf->raid_disks); 1528 1529 rcu_read_lock(); 1530 for (i = 0; i < conf->raid_disks; i++) { 1531 char b[BDEVNAME_SIZE]; 1532 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1533 if (rdev) 1534 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n", 1535 i, !test_bit(In_sync, &rdev->flags), 1536 !test_bit(Faulty, &rdev->flags), 1537 bdevname(rdev->bdev,b)); 1538 } 1539 rcu_read_unlock(); 1540 } 1541 1542 static void close_sync(struct r1conf *conf) 1543 { 1544 wait_barrier(conf, NULL); 1545 allow_barrier(conf, 0, 0); 1546 1547 mempool_destroy(conf->r1buf_pool); 1548 conf->r1buf_pool = NULL; 1549 1550 spin_lock_irq(&conf->resync_lock); 1551 conf->next_resync = 0; 1552 conf->start_next_window = MaxSector; 1553 conf->current_window_requests += 1554 conf->next_window_requests; 1555 conf->next_window_requests = 0; 1556 spin_unlock_irq(&conf->resync_lock); 1557 } 1558 1559 static int raid1_spare_active(struct mddev *mddev) 1560 { 1561 int i; 1562 struct r1conf *conf = mddev->private; 1563 int count = 0; 1564 unsigned long flags; 1565 1566 /* 1567 * Find all failed disks within the RAID1 configuration 1568 * and mark them readable. 1569 * Called under mddev lock, so rcu protection not needed. 1570 */ 1571 for (i = 0; i < conf->raid_disks; i++) { 1572 struct md_rdev *rdev = conf->mirrors[i].rdev; 1573 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev; 1574 if (repl 1575 && repl->recovery_offset == MaxSector 1576 && !test_bit(Faulty, &repl->flags) 1577 && !test_and_set_bit(In_sync, &repl->flags)) { 1578 /* replacement has just become active */ 1579 if (!rdev || 1580 !test_and_clear_bit(In_sync, &rdev->flags)) 1581 count++; 1582 if (rdev) { 1583 /* Replaced device not technically 1584 * faulty, but we need to be sure 1585 * it gets removed and never re-added 1586 */ 1587 set_bit(Faulty, &rdev->flags); 1588 sysfs_notify_dirent_safe( 1589 rdev->sysfs_state); 1590 } 1591 } 1592 if (rdev 1593 && rdev->recovery_offset == MaxSector 1594 && !test_bit(Faulty, &rdev->flags) 1595 && !test_and_set_bit(In_sync, &rdev->flags)) { 1596 count++; 1597 sysfs_notify_dirent_safe(rdev->sysfs_state); 1598 } 1599 } 1600 spin_lock_irqsave(&conf->device_lock, flags); 1601 mddev->degraded -= count; 1602 spin_unlock_irqrestore(&conf->device_lock, flags); 1603 1604 print_conf(conf); 1605 return count; 1606 } 1607 1608 1609 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev) 1610 { 1611 struct r1conf *conf = mddev->private; 1612 int err = -EEXIST; 1613 int mirror = 0; 1614 struct raid1_info *p; 1615 int first = 0; 1616 int last = conf->raid_disks - 1; 1617 struct request_queue *q = bdev_get_queue(rdev->bdev); 1618 1619 if (mddev->recovery_disabled == conf->recovery_disabled) 1620 return -EBUSY; 1621 1622 if (rdev->raid_disk >= 0) 1623 first = last = rdev->raid_disk; 1624 1625 if (q->merge_bvec_fn) { 1626 set_bit(Unmerged, &rdev->flags); 1627 mddev->merge_check_needed = 1; 1628 } 1629 1630 for (mirror = first; mirror <= last; mirror++) { 1631 p = conf->mirrors+mirror; 1632 if (!p->rdev) { 1633 1634 if (mddev->gendisk) 1635 disk_stack_limits(mddev->gendisk, rdev->bdev, 1636 rdev->data_offset << 9); 1637 1638 p->head_position = 0; 1639 rdev->raid_disk = mirror; 1640 err = 0; 1641 /* As all devices are equivalent, we don't need a full recovery 1642 * if this was recently any drive of the array 1643 */ 1644 if (rdev->saved_raid_disk < 0) 1645 conf->fullsync = 1; 1646 rcu_assign_pointer(p->rdev, rdev); 1647 break; 1648 } 1649 if (test_bit(WantReplacement, &p->rdev->flags) && 1650 p[conf->raid_disks].rdev == NULL) { 1651 /* Add this device as a replacement */ 1652 clear_bit(In_sync, &rdev->flags); 1653 set_bit(Replacement, &rdev->flags); 1654 rdev->raid_disk = mirror; 1655 err = 0; 1656 conf->fullsync = 1; 1657 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev); 1658 break; 1659 } 1660 } 1661 if (err == 0 && test_bit(Unmerged, &rdev->flags)) { 1662 /* Some requests might not have seen this new 1663 * merge_bvec_fn. We must wait for them to complete 1664 * before merging the device fully. 1665 * First we make sure any code which has tested 1666 * our function has submitted the request, then 1667 * we wait for all outstanding requests to complete. 1668 */ 1669 synchronize_sched(); 1670 freeze_array(conf, 0); 1671 unfreeze_array(conf); 1672 clear_bit(Unmerged, &rdev->flags); 1673 } 1674 md_integrity_add_rdev(rdev, mddev); 1675 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev))) 1676 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue); 1677 print_conf(conf); 1678 return err; 1679 } 1680 1681 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev) 1682 { 1683 struct r1conf *conf = mddev->private; 1684 int err = 0; 1685 int number = rdev->raid_disk; 1686 struct raid1_info *p = conf->mirrors + number; 1687 1688 if (rdev != p->rdev) 1689 p = conf->mirrors + conf->raid_disks + number; 1690 1691 print_conf(conf); 1692 if (rdev == p->rdev) { 1693 if (test_bit(In_sync, &rdev->flags) || 1694 atomic_read(&rdev->nr_pending)) { 1695 err = -EBUSY; 1696 goto abort; 1697 } 1698 /* Only remove non-faulty devices if recovery 1699 * is not possible. 1700 */ 1701 if (!test_bit(Faulty, &rdev->flags) && 1702 mddev->recovery_disabled != conf->recovery_disabled && 1703 mddev->degraded < conf->raid_disks) { 1704 err = -EBUSY; 1705 goto abort; 1706 } 1707 p->rdev = NULL; 1708 synchronize_rcu(); 1709 if (atomic_read(&rdev->nr_pending)) { 1710 /* lost the race, try later */ 1711 err = -EBUSY; 1712 p->rdev = rdev; 1713 goto abort; 1714 } else if (conf->mirrors[conf->raid_disks + number].rdev) { 1715 /* We just removed a device that is being replaced. 1716 * Move down the replacement. We drain all IO before 1717 * doing this to avoid confusion. 1718 */ 1719 struct md_rdev *repl = 1720 conf->mirrors[conf->raid_disks + number].rdev; 1721 freeze_array(conf, 0); 1722 clear_bit(Replacement, &repl->flags); 1723 p->rdev = repl; 1724 conf->mirrors[conf->raid_disks + number].rdev = NULL; 1725 unfreeze_array(conf); 1726 clear_bit(WantReplacement, &rdev->flags); 1727 } else 1728 clear_bit(WantReplacement, &rdev->flags); 1729 err = md_integrity_register(mddev); 1730 } 1731 abort: 1732 1733 print_conf(conf); 1734 return err; 1735 } 1736 1737 1738 static void end_sync_read(struct bio *bio, int error) 1739 { 1740 struct r1bio *r1_bio = bio->bi_private; 1741 1742 update_head_pos(r1_bio->read_disk, r1_bio); 1743 1744 /* 1745 * we have read a block, now it needs to be re-written, 1746 * or re-read if the read failed. 1747 * We don't do much here, just schedule handling by raid1d 1748 */ 1749 if (test_bit(BIO_UPTODATE, &bio->bi_flags)) 1750 set_bit(R1BIO_Uptodate, &r1_bio->state); 1751 1752 if (atomic_dec_and_test(&r1_bio->remaining)) 1753 reschedule_retry(r1_bio); 1754 } 1755 1756 static void end_sync_write(struct bio *bio, int error) 1757 { 1758 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 1759 struct r1bio *r1_bio = bio->bi_private; 1760 struct mddev *mddev = r1_bio->mddev; 1761 struct r1conf *conf = mddev->private; 1762 int mirror=0; 1763 sector_t first_bad; 1764 int bad_sectors; 1765 1766 mirror = find_bio_disk(r1_bio, bio); 1767 1768 if (!uptodate) { 1769 sector_t sync_blocks = 0; 1770 sector_t s = r1_bio->sector; 1771 long sectors_to_go = r1_bio->sectors; 1772 /* make sure these bits doesn't get cleared. */ 1773 do { 1774 bitmap_end_sync(mddev->bitmap, s, 1775 &sync_blocks, 1); 1776 s += sync_blocks; 1777 sectors_to_go -= sync_blocks; 1778 } while (sectors_to_go > 0); 1779 set_bit(WriteErrorSeen, 1780 &conf->mirrors[mirror].rdev->flags); 1781 if (!test_and_set_bit(WantReplacement, 1782 &conf->mirrors[mirror].rdev->flags)) 1783 set_bit(MD_RECOVERY_NEEDED, & 1784 mddev->recovery); 1785 set_bit(R1BIO_WriteError, &r1_bio->state); 1786 } else if (is_badblock(conf->mirrors[mirror].rdev, 1787 r1_bio->sector, 1788 r1_bio->sectors, 1789 &first_bad, &bad_sectors) && 1790 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev, 1791 r1_bio->sector, 1792 r1_bio->sectors, 1793 &first_bad, &bad_sectors) 1794 ) 1795 set_bit(R1BIO_MadeGood, &r1_bio->state); 1796 1797 if (atomic_dec_and_test(&r1_bio->remaining)) { 1798 int s = r1_bio->sectors; 1799 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 1800 test_bit(R1BIO_WriteError, &r1_bio->state)) 1801 reschedule_retry(r1_bio); 1802 else { 1803 put_buf(r1_bio); 1804 md_done_sync(mddev, s, uptodate); 1805 } 1806 } 1807 } 1808 1809 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector, 1810 int sectors, struct page *page, int rw) 1811 { 1812 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false)) 1813 /* success */ 1814 return 1; 1815 if (rw == WRITE) { 1816 set_bit(WriteErrorSeen, &rdev->flags); 1817 if (!test_and_set_bit(WantReplacement, 1818 &rdev->flags)) 1819 set_bit(MD_RECOVERY_NEEDED, & 1820 rdev->mddev->recovery); 1821 } 1822 /* need to record an error - either for the block or the device */ 1823 if (!rdev_set_badblocks(rdev, sector, sectors, 0)) 1824 md_error(rdev->mddev, rdev); 1825 return 0; 1826 } 1827 1828 static int fix_sync_read_error(struct r1bio *r1_bio) 1829 { 1830 /* Try some synchronous reads of other devices to get 1831 * good data, much like with normal read errors. Only 1832 * read into the pages we already have so we don't 1833 * need to re-issue the read request. 1834 * We don't need to freeze the array, because being in an 1835 * active sync request, there is no normal IO, and 1836 * no overlapping syncs. 1837 * We don't need to check is_badblock() again as we 1838 * made sure that anything with a bad block in range 1839 * will have bi_end_io clear. 1840 */ 1841 struct mddev *mddev = r1_bio->mddev; 1842 struct r1conf *conf = mddev->private; 1843 struct bio *bio = r1_bio->bios[r1_bio->read_disk]; 1844 sector_t sect = r1_bio->sector; 1845 int sectors = r1_bio->sectors; 1846 int idx = 0; 1847 1848 while(sectors) { 1849 int s = sectors; 1850 int d = r1_bio->read_disk; 1851 int success = 0; 1852 struct md_rdev *rdev; 1853 int start; 1854 1855 if (s > (PAGE_SIZE>>9)) 1856 s = PAGE_SIZE >> 9; 1857 do { 1858 if (r1_bio->bios[d]->bi_end_io == end_sync_read) { 1859 /* No rcu protection needed here devices 1860 * can only be removed when no resync is 1861 * active, and resync is currently active 1862 */ 1863 rdev = conf->mirrors[d].rdev; 1864 if (sync_page_io(rdev, sect, s<<9, 1865 bio->bi_io_vec[idx].bv_page, 1866 READ, false)) { 1867 success = 1; 1868 break; 1869 } 1870 } 1871 d++; 1872 if (d == conf->raid_disks * 2) 1873 d = 0; 1874 } while (!success && d != r1_bio->read_disk); 1875 1876 if (!success) { 1877 char b[BDEVNAME_SIZE]; 1878 int abort = 0; 1879 /* Cannot read from anywhere, this block is lost. 1880 * Record a bad block on each device. If that doesn't 1881 * work just disable and interrupt the recovery. 1882 * Don't fail devices as that won't really help. 1883 */ 1884 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error" 1885 " for block %llu\n", 1886 mdname(mddev), 1887 bdevname(bio->bi_bdev, b), 1888 (unsigned long long)r1_bio->sector); 1889 for (d = 0; d < conf->raid_disks * 2; d++) { 1890 rdev = conf->mirrors[d].rdev; 1891 if (!rdev || test_bit(Faulty, &rdev->flags)) 1892 continue; 1893 if (!rdev_set_badblocks(rdev, sect, s, 0)) 1894 abort = 1; 1895 } 1896 if (abort) { 1897 conf->recovery_disabled = 1898 mddev->recovery_disabled; 1899 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1900 md_done_sync(mddev, r1_bio->sectors, 0); 1901 put_buf(r1_bio); 1902 return 0; 1903 } 1904 /* Try next page */ 1905 sectors -= s; 1906 sect += s; 1907 idx++; 1908 continue; 1909 } 1910 1911 start = d; 1912 /* write it back and re-read */ 1913 while (d != r1_bio->read_disk) { 1914 if (d == 0) 1915 d = conf->raid_disks * 2; 1916 d--; 1917 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1918 continue; 1919 rdev = conf->mirrors[d].rdev; 1920 if (r1_sync_page_io(rdev, sect, s, 1921 bio->bi_io_vec[idx].bv_page, 1922 WRITE) == 0) { 1923 r1_bio->bios[d]->bi_end_io = NULL; 1924 rdev_dec_pending(rdev, mddev); 1925 } 1926 } 1927 d = start; 1928 while (d != r1_bio->read_disk) { 1929 if (d == 0) 1930 d = conf->raid_disks * 2; 1931 d--; 1932 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1933 continue; 1934 rdev = conf->mirrors[d].rdev; 1935 if (r1_sync_page_io(rdev, sect, s, 1936 bio->bi_io_vec[idx].bv_page, 1937 READ) != 0) 1938 atomic_add(s, &rdev->corrected_errors); 1939 } 1940 sectors -= s; 1941 sect += s; 1942 idx ++; 1943 } 1944 set_bit(R1BIO_Uptodate, &r1_bio->state); 1945 set_bit(BIO_UPTODATE, &bio->bi_flags); 1946 return 1; 1947 } 1948 1949 static int process_checks(struct r1bio *r1_bio) 1950 { 1951 /* We have read all readable devices. If we haven't 1952 * got the block, then there is no hope left. 1953 * If we have, then we want to do a comparison 1954 * and skip the write if everything is the same. 1955 * If any blocks failed to read, then we need to 1956 * attempt an over-write 1957 */ 1958 struct mddev *mddev = r1_bio->mddev; 1959 struct r1conf *conf = mddev->private; 1960 int primary; 1961 int i; 1962 int vcnt; 1963 1964 /* Fix variable parts of all bios */ 1965 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9); 1966 for (i = 0; i < conf->raid_disks * 2; i++) { 1967 int j; 1968 int size; 1969 int uptodate; 1970 struct bio *b = r1_bio->bios[i]; 1971 if (b->bi_end_io != end_sync_read) 1972 continue; 1973 /* fixup the bio for reuse, but preserve BIO_UPTODATE */ 1974 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags); 1975 bio_reset(b); 1976 if (!uptodate) 1977 clear_bit(BIO_UPTODATE, &b->bi_flags); 1978 b->bi_vcnt = vcnt; 1979 b->bi_iter.bi_size = r1_bio->sectors << 9; 1980 b->bi_iter.bi_sector = r1_bio->sector + 1981 conf->mirrors[i].rdev->data_offset; 1982 b->bi_bdev = conf->mirrors[i].rdev->bdev; 1983 b->bi_end_io = end_sync_read; 1984 b->bi_private = r1_bio; 1985 1986 size = b->bi_iter.bi_size; 1987 for (j = 0; j < vcnt ; j++) { 1988 struct bio_vec *bi; 1989 bi = &b->bi_io_vec[j]; 1990 bi->bv_offset = 0; 1991 if (size > PAGE_SIZE) 1992 bi->bv_len = PAGE_SIZE; 1993 else 1994 bi->bv_len = size; 1995 size -= PAGE_SIZE; 1996 } 1997 } 1998 for (primary = 0; primary < conf->raid_disks * 2; primary++) 1999 if (r1_bio->bios[primary]->bi_end_io == end_sync_read && 2000 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) { 2001 r1_bio->bios[primary]->bi_end_io = NULL; 2002 rdev_dec_pending(conf->mirrors[primary].rdev, mddev); 2003 break; 2004 } 2005 r1_bio->read_disk = primary; 2006 for (i = 0; i < conf->raid_disks * 2; i++) { 2007 int j; 2008 struct bio *pbio = r1_bio->bios[primary]; 2009 struct bio *sbio = r1_bio->bios[i]; 2010 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags); 2011 2012 if (sbio->bi_end_io != end_sync_read) 2013 continue; 2014 /* Now we can 'fixup' the BIO_UPTODATE flag */ 2015 set_bit(BIO_UPTODATE, &sbio->bi_flags); 2016 2017 if (uptodate) { 2018 for (j = vcnt; j-- ; ) { 2019 struct page *p, *s; 2020 p = pbio->bi_io_vec[j].bv_page; 2021 s = sbio->bi_io_vec[j].bv_page; 2022 if (memcmp(page_address(p), 2023 page_address(s), 2024 sbio->bi_io_vec[j].bv_len)) 2025 break; 2026 } 2027 } else 2028 j = 0; 2029 if (j >= 0) 2030 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches); 2031 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery) 2032 && uptodate)) { 2033 /* No need to write to this device. */ 2034 sbio->bi_end_io = NULL; 2035 rdev_dec_pending(conf->mirrors[i].rdev, mddev); 2036 continue; 2037 } 2038 2039 bio_copy_data(sbio, pbio); 2040 } 2041 return 0; 2042 } 2043 2044 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio) 2045 { 2046 struct r1conf *conf = mddev->private; 2047 int i; 2048 int disks = conf->raid_disks * 2; 2049 struct bio *bio, *wbio; 2050 2051 bio = r1_bio->bios[r1_bio->read_disk]; 2052 2053 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 2054 /* ouch - failed to read all of that. */ 2055 if (!fix_sync_read_error(r1_bio)) 2056 return; 2057 2058 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2059 if (process_checks(r1_bio) < 0) 2060 return; 2061 /* 2062 * schedule writes 2063 */ 2064 atomic_set(&r1_bio->remaining, 1); 2065 for (i = 0; i < disks ; i++) { 2066 wbio = r1_bio->bios[i]; 2067 if (wbio->bi_end_io == NULL || 2068 (wbio->bi_end_io == end_sync_read && 2069 (i == r1_bio->read_disk || 2070 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)))) 2071 continue; 2072 2073 wbio->bi_rw = WRITE; 2074 wbio->bi_end_io = end_sync_write; 2075 atomic_inc(&r1_bio->remaining); 2076 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio)); 2077 2078 generic_make_request(wbio); 2079 } 2080 2081 if (atomic_dec_and_test(&r1_bio->remaining)) { 2082 /* if we're here, all write(s) have completed, so clean up */ 2083 int s = r1_bio->sectors; 2084 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2085 test_bit(R1BIO_WriteError, &r1_bio->state)) 2086 reschedule_retry(r1_bio); 2087 else { 2088 put_buf(r1_bio); 2089 md_done_sync(mddev, s, 1); 2090 } 2091 } 2092 } 2093 2094 /* 2095 * This is a kernel thread which: 2096 * 2097 * 1. Retries failed read operations on working mirrors. 2098 * 2. Updates the raid superblock when problems encounter. 2099 * 3. Performs writes following reads for array synchronising. 2100 */ 2101 2102 static void fix_read_error(struct r1conf *conf, int read_disk, 2103 sector_t sect, int sectors) 2104 { 2105 struct mddev *mddev = conf->mddev; 2106 while(sectors) { 2107 int s = sectors; 2108 int d = read_disk; 2109 int success = 0; 2110 int start; 2111 struct md_rdev *rdev; 2112 2113 if (s > (PAGE_SIZE>>9)) 2114 s = PAGE_SIZE >> 9; 2115 2116 do { 2117 /* Note: no rcu protection needed here 2118 * as this is synchronous in the raid1d thread 2119 * which is the thread that might remove 2120 * a device. If raid1d ever becomes multi-threaded.... 2121 */ 2122 sector_t first_bad; 2123 int bad_sectors; 2124 2125 rdev = conf->mirrors[d].rdev; 2126 if (rdev && 2127 (test_bit(In_sync, &rdev->flags) || 2128 (!test_bit(Faulty, &rdev->flags) && 2129 rdev->recovery_offset >= sect + s)) && 2130 is_badblock(rdev, sect, s, 2131 &first_bad, &bad_sectors) == 0 && 2132 sync_page_io(rdev, sect, s<<9, 2133 conf->tmppage, READ, false)) 2134 success = 1; 2135 else { 2136 d++; 2137 if (d == conf->raid_disks * 2) 2138 d = 0; 2139 } 2140 } while (!success && d != read_disk); 2141 2142 if (!success) { 2143 /* Cannot read from anywhere - mark it bad */ 2144 struct md_rdev *rdev = conf->mirrors[read_disk].rdev; 2145 if (!rdev_set_badblocks(rdev, sect, s, 0)) 2146 md_error(mddev, rdev); 2147 break; 2148 } 2149 /* write it back and re-read */ 2150 start = d; 2151 while (d != read_disk) { 2152 if (d==0) 2153 d = conf->raid_disks * 2; 2154 d--; 2155 rdev = conf->mirrors[d].rdev; 2156 if (rdev && 2157 test_bit(In_sync, &rdev->flags)) 2158 r1_sync_page_io(rdev, sect, s, 2159 conf->tmppage, WRITE); 2160 } 2161 d = start; 2162 while (d != read_disk) { 2163 char b[BDEVNAME_SIZE]; 2164 if (d==0) 2165 d = conf->raid_disks * 2; 2166 d--; 2167 rdev = conf->mirrors[d].rdev; 2168 if (rdev && 2169 test_bit(In_sync, &rdev->flags)) { 2170 if (r1_sync_page_io(rdev, sect, s, 2171 conf->tmppage, READ)) { 2172 atomic_add(s, &rdev->corrected_errors); 2173 printk(KERN_INFO 2174 "md/raid1:%s: read error corrected " 2175 "(%d sectors at %llu on %s)\n", 2176 mdname(mddev), s, 2177 (unsigned long long)(sect + 2178 rdev->data_offset), 2179 bdevname(rdev->bdev, b)); 2180 } 2181 } 2182 } 2183 sectors -= s; 2184 sect += s; 2185 } 2186 } 2187 2188 static int narrow_write_error(struct r1bio *r1_bio, int i) 2189 { 2190 struct mddev *mddev = r1_bio->mddev; 2191 struct r1conf *conf = mddev->private; 2192 struct md_rdev *rdev = conf->mirrors[i].rdev; 2193 2194 /* bio has the data to be written to device 'i' where 2195 * we just recently had a write error. 2196 * We repeatedly clone the bio and trim down to one block, 2197 * then try the write. Where the write fails we record 2198 * a bad block. 2199 * It is conceivable that the bio doesn't exactly align with 2200 * blocks. We must handle this somehow. 2201 * 2202 * We currently own a reference on the rdev. 2203 */ 2204 2205 int block_sectors; 2206 sector_t sector; 2207 int sectors; 2208 int sect_to_write = r1_bio->sectors; 2209 int ok = 1; 2210 2211 if (rdev->badblocks.shift < 0) 2212 return 0; 2213 2214 block_sectors = 1 << rdev->badblocks.shift; 2215 sector = r1_bio->sector; 2216 sectors = ((sector + block_sectors) 2217 & ~(sector_t)(block_sectors - 1)) 2218 - sector; 2219 2220 while (sect_to_write) { 2221 struct bio *wbio; 2222 if (sectors > sect_to_write) 2223 sectors = sect_to_write; 2224 /* Write at 'sector' for 'sectors'*/ 2225 2226 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 2227 unsigned vcnt = r1_bio->behind_page_count; 2228 struct bio_vec *vec = r1_bio->behind_bvecs; 2229 2230 while (!vec->bv_page) { 2231 vec++; 2232 vcnt--; 2233 } 2234 2235 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev); 2236 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec)); 2237 2238 wbio->bi_vcnt = vcnt; 2239 } else { 2240 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2241 } 2242 2243 wbio->bi_rw = WRITE; 2244 wbio->bi_iter.bi_sector = r1_bio->sector; 2245 wbio->bi_iter.bi_size = r1_bio->sectors << 9; 2246 2247 bio_trim(wbio, sector - r1_bio->sector, sectors); 2248 wbio->bi_iter.bi_sector += rdev->data_offset; 2249 wbio->bi_bdev = rdev->bdev; 2250 if (submit_bio_wait(WRITE, wbio) == 0) 2251 /* failure! */ 2252 ok = rdev_set_badblocks(rdev, sector, 2253 sectors, 0) 2254 && ok; 2255 2256 bio_put(wbio); 2257 sect_to_write -= sectors; 2258 sector += sectors; 2259 sectors = block_sectors; 2260 } 2261 return ok; 2262 } 2263 2264 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2265 { 2266 int m; 2267 int s = r1_bio->sectors; 2268 for (m = 0; m < conf->raid_disks * 2 ; m++) { 2269 struct md_rdev *rdev = conf->mirrors[m].rdev; 2270 struct bio *bio = r1_bio->bios[m]; 2271 if (bio->bi_end_io == NULL) 2272 continue; 2273 if (test_bit(BIO_UPTODATE, &bio->bi_flags) && 2274 test_bit(R1BIO_MadeGood, &r1_bio->state)) { 2275 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0); 2276 } 2277 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) && 2278 test_bit(R1BIO_WriteError, &r1_bio->state)) { 2279 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0)) 2280 md_error(conf->mddev, rdev); 2281 } 2282 } 2283 put_buf(r1_bio); 2284 md_done_sync(conf->mddev, s, 1); 2285 } 2286 2287 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2288 { 2289 int m; 2290 for (m = 0; m < conf->raid_disks * 2 ; m++) 2291 if (r1_bio->bios[m] == IO_MADE_GOOD) { 2292 struct md_rdev *rdev = conf->mirrors[m].rdev; 2293 rdev_clear_badblocks(rdev, 2294 r1_bio->sector, 2295 r1_bio->sectors, 0); 2296 rdev_dec_pending(rdev, conf->mddev); 2297 } else if (r1_bio->bios[m] != NULL) { 2298 /* This drive got a write error. We need to 2299 * narrow down and record precise write 2300 * errors. 2301 */ 2302 if (!narrow_write_error(r1_bio, m)) { 2303 md_error(conf->mddev, 2304 conf->mirrors[m].rdev); 2305 /* an I/O failed, we can't clear the bitmap */ 2306 set_bit(R1BIO_Degraded, &r1_bio->state); 2307 } 2308 rdev_dec_pending(conf->mirrors[m].rdev, 2309 conf->mddev); 2310 } 2311 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 2312 close_write(r1_bio); 2313 raid_end_bio_io(r1_bio); 2314 } 2315 2316 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio) 2317 { 2318 int disk; 2319 int max_sectors; 2320 struct mddev *mddev = conf->mddev; 2321 struct bio *bio; 2322 char b[BDEVNAME_SIZE]; 2323 struct md_rdev *rdev; 2324 2325 clear_bit(R1BIO_ReadError, &r1_bio->state); 2326 /* we got a read error. Maybe the drive is bad. Maybe just 2327 * the block and we can fix it. 2328 * We freeze all other IO, and try reading the block from 2329 * other devices. When we find one, we re-write 2330 * and check it that fixes the read error. 2331 * This is all done synchronously while the array is 2332 * frozen 2333 */ 2334 if (mddev->ro == 0) { 2335 freeze_array(conf, 1); 2336 fix_read_error(conf, r1_bio->read_disk, 2337 r1_bio->sector, r1_bio->sectors); 2338 unfreeze_array(conf); 2339 } else 2340 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev); 2341 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev); 2342 2343 bio = r1_bio->bios[r1_bio->read_disk]; 2344 bdevname(bio->bi_bdev, b); 2345 read_more: 2346 disk = read_balance(conf, r1_bio, &max_sectors); 2347 if (disk == -1) { 2348 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O" 2349 " read error for block %llu\n", 2350 mdname(mddev), b, (unsigned long long)r1_bio->sector); 2351 raid_end_bio_io(r1_bio); 2352 } else { 2353 const unsigned long do_sync 2354 = r1_bio->master_bio->bi_rw & REQ_SYNC; 2355 if (bio) { 2356 r1_bio->bios[r1_bio->read_disk] = 2357 mddev->ro ? IO_BLOCKED : NULL; 2358 bio_put(bio); 2359 } 2360 r1_bio->read_disk = disk; 2361 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2362 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector, 2363 max_sectors); 2364 r1_bio->bios[r1_bio->read_disk] = bio; 2365 rdev = conf->mirrors[disk].rdev; 2366 printk_ratelimited(KERN_ERR 2367 "md/raid1:%s: redirecting sector %llu" 2368 " to other mirror: %s\n", 2369 mdname(mddev), 2370 (unsigned long long)r1_bio->sector, 2371 bdevname(rdev->bdev, b)); 2372 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset; 2373 bio->bi_bdev = rdev->bdev; 2374 bio->bi_end_io = raid1_end_read_request; 2375 bio->bi_rw = READ | do_sync; 2376 bio->bi_private = r1_bio; 2377 if (max_sectors < r1_bio->sectors) { 2378 /* Drat - have to split this up more */ 2379 struct bio *mbio = r1_bio->master_bio; 2380 int sectors_handled = (r1_bio->sector + max_sectors 2381 - mbio->bi_iter.bi_sector); 2382 r1_bio->sectors = max_sectors; 2383 spin_lock_irq(&conf->device_lock); 2384 if (mbio->bi_phys_segments == 0) 2385 mbio->bi_phys_segments = 2; 2386 else 2387 mbio->bi_phys_segments++; 2388 spin_unlock_irq(&conf->device_lock); 2389 generic_make_request(bio); 2390 bio = NULL; 2391 2392 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 2393 2394 r1_bio->master_bio = mbio; 2395 r1_bio->sectors = bio_sectors(mbio) - sectors_handled; 2396 r1_bio->state = 0; 2397 set_bit(R1BIO_ReadError, &r1_bio->state); 2398 r1_bio->mddev = mddev; 2399 r1_bio->sector = mbio->bi_iter.bi_sector + 2400 sectors_handled; 2401 2402 goto read_more; 2403 } else 2404 generic_make_request(bio); 2405 } 2406 } 2407 2408 static void raid1d(struct md_thread *thread) 2409 { 2410 struct mddev *mddev = thread->mddev; 2411 struct r1bio *r1_bio; 2412 unsigned long flags; 2413 struct r1conf *conf = mddev->private; 2414 struct list_head *head = &conf->retry_list; 2415 struct blk_plug plug; 2416 2417 md_check_recovery(mddev); 2418 2419 blk_start_plug(&plug); 2420 for (;;) { 2421 2422 flush_pending_writes(conf); 2423 2424 spin_lock_irqsave(&conf->device_lock, flags); 2425 if (list_empty(head)) { 2426 spin_unlock_irqrestore(&conf->device_lock, flags); 2427 break; 2428 } 2429 r1_bio = list_entry(head->prev, struct r1bio, retry_list); 2430 list_del(head->prev); 2431 conf->nr_queued--; 2432 spin_unlock_irqrestore(&conf->device_lock, flags); 2433 2434 mddev = r1_bio->mddev; 2435 conf = mddev->private; 2436 if (test_bit(R1BIO_IsSync, &r1_bio->state)) { 2437 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2438 test_bit(R1BIO_WriteError, &r1_bio->state)) 2439 handle_sync_write_finished(conf, r1_bio); 2440 else 2441 sync_request_write(mddev, r1_bio); 2442 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2443 test_bit(R1BIO_WriteError, &r1_bio->state)) 2444 handle_write_finished(conf, r1_bio); 2445 else if (test_bit(R1BIO_ReadError, &r1_bio->state)) 2446 handle_read_error(conf, r1_bio); 2447 else 2448 /* just a partial read to be scheduled from separate 2449 * context 2450 */ 2451 generic_make_request(r1_bio->bios[r1_bio->read_disk]); 2452 2453 cond_resched(); 2454 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) 2455 md_check_recovery(mddev); 2456 } 2457 blk_finish_plug(&plug); 2458 } 2459 2460 2461 static int init_resync(struct r1conf *conf) 2462 { 2463 int buffs; 2464 2465 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; 2466 BUG_ON(conf->r1buf_pool); 2467 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, 2468 conf->poolinfo); 2469 if (!conf->r1buf_pool) 2470 return -ENOMEM; 2471 conf->next_resync = 0; 2472 return 0; 2473 } 2474 2475 /* 2476 * perform a "sync" on one "block" 2477 * 2478 * We need to make sure that no normal I/O request - particularly write 2479 * requests - conflict with active sync requests. 2480 * 2481 * This is achieved by tracking pending requests and a 'barrier' concept 2482 * that can be installed to exclude normal IO requests. 2483 */ 2484 2485 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster) 2486 { 2487 struct r1conf *conf = mddev->private; 2488 struct r1bio *r1_bio; 2489 struct bio *bio; 2490 sector_t max_sector, nr_sectors; 2491 int disk = -1; 2492 int i; 2493 int wonly = -1; 2494 int write_targets = 0, read_targets = 0; 2495 sector_t sync_blocks; 2496 int still_degraded = 0; 2497 int good_sectors = RESYNC_SECTORS; 2498 int min_bad = 0; /* number of sectors that are bad in all devices */ 2499 2500 if (!conf->r1buf_pool) 2501 if (init_resync(conf)) 2502 return 0; 2503 2504 max_sector = mddev->dev_sectors; 2505 if (sector_nr >= max_sector) { 2506 /* If we aborted, we need to abort the 2507 * sync on the 'current' bitmap chunk (there will 2508 * only be one in raid1 resync. 2509 * We can find the current addess in mddev->curr_resync 2510 */ 2511 if (mddev->curr_resync < max_sector) /* aborted */ 2512 bitmap_end_sync(mddev->bitmap, mddev->curr_resync, 2513 &sync_blocks, 1); 2514 else /* completed sync */ 2515 conf->fullsync = 0; 2516 2517 bitmap_close_sync(mddev->bitmap); 2518 close_sync(conf); 2519 return 0; 2520 } 2521 2522 if (mddev->bitmap == NULL && 2523 mddev->recovery_cp == MaxSector && 2524 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && 2525 conf->fullsync == 0) { 2526 *skipped = 1; 2527 return max_sector - sector_nr; 2528 } 2529 /* before building a request, check if we can skip these blocks.. 2530 * This call the bitmap_start_sync doesn't actually record anything 2531 */ 2532 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) && 2533 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2534 /* We can skip this block, and probably several more */ 2535 *skipped = 1; 2536 return sync_blocks; 2537 } 2538 /* 2539 * If there is non-resync activity waiting for a turn, 2540 * and resync is going fast enough, 2541 * then let it though before starting on this new sync request. 2542 */ 2543 if (!go_faster && conf->nr_waiting) 2544 msleep_interruptible(1000); 2545 2546 bitmap_cond_end_sync(mddev->bitmap, sector_nr); 2547 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO); 2548 raise_barrier(conf); 2549 2550 conf->next_resync = sector_nr; 2551 2552 rcu_read_lock(); 2553 /* 2554 * If we get a correctably read error during resync or recovery, 2555 * we might want to read from a different device. So we 2556 * flag all drives that could conceivably be read from for READ, 2557 * and any others (which will be non-In_sync devices) for WRITE. 2558 * If a read fails, we try reading from something else for which READ 2559 * is OK. 2560 */ 2561 2562 r1_bio->mddev = mddev; 2563 r1_bio->sector = sector_nr; 2564 r1_bio->state = 0; 2565 set_bit(R1BIO_IsSync, &r1_bio->state); 2566 2567 for (i = 0; i < conf->raid_disks * 2; i++) { 2568 struct md_rdev *rdev; 2569 bio = r1_bio->bios[i]; 2570 bio_reset(bio); 2571 2572 rdev = rcu_dereference(conf->mirrors[i].rdev); 2573 if (rdev == NULL || 2574 test_bit(Faulty, &rdev->flags)) { 2575 if (i < conf->raid_disks) 2576 still_degraded = 1; 2577 } else if (!test_bit(In_sync, &rdev->flags)) { 2578 bio->bi_rw = WRITE; 2579 bio->bi_end_io = end_sync_write; 2580 write_targets ++; 2581 } else { 2582 /* may need to read from here */ 2583 sector_t first_bad = MaxSector; 2584 int bad_sectors; 2585 2586 if (is_badblock(rdev, sector_nr, good_sectors, 2587 &first_bad, &bad_sectors)) { 2588 if (first_bad > sector_nr) 2589 good_sectors = first_bad - sector_nr; 2590 else { 2591 bad_sectors -= (sector_nr - first_bad); 2592 if (min_bad == 0 || 2593 min_bad > bad_sectors) 2594 min_bad = bad_sectors; 2595 } 2596 } 2597 if (sector_nr < first_bad) { 2598 if (test_bit(WriteMostly, &rdev->flags)) { 2599 if (wonly < 0) 2600 wonly = i; 2601 } else { 2602 if (disk < 0) 2603 disk = i; 2604 } 2605 bio->bi_rw = READ; 2606 bio->bi_end_io = end_sync_read; 2607 read_targets++; 2608 } else if (!test_bit(WriteErrorSeen, &rdev->flags) && 2609 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && 2610 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) { 2611 /* 2612 * The device is suitable for reading (InSync), 2613 * but has bad block(s) here. Let's try to correct them, 2614 * if we are doing resync or repair. Otherwise, leave 2615 * this device alone for this sync request. 2616 */ 2617 bio->bi_rw = WRITE; 2618 bio->bi_end_io = end_sync_write; 2619 write_targets++; 2620 } 2621 } 2622 if (bio->bi_end_io) { 2623 atomic_inc(&rdev->nr_pending); 2624 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset; 2625 bio->bi_bdev = rdev->bdev; 2626 bio->bi_private = r1_bio; 2627 } 2628 } 2629 rcu_read_unlock(); 2630 if (disk < 0) 2631 disk = wonly; 2632 r1_bio->read_disk = disk; 2633 2634 if (read_targets == 0 && min_bad > 0) { 2635 /* These sectors are bad on all InSync devices, so we 2636 * need to mark them bad on all write targets 2637 */ 2638 int ok = 1; 2639 for (i = 0 ; i < conf->raid_disks * 2 ; i++) 2640 if (r1_bio->bios[i]->bi_end_io == end_sync_write) { 2641 struct md_rdev *rdev = conf->mirrors[i].rdev; 2642 ok = rdev_set_badblocks(rdev, sector_nr, 2643 min_bad, 0 2644 ) && ok; 2645 } 2646 set_bit(MD_CHANGE_DEVS, &mddev->flags); 2647 *skipped = 1; 2648 put_buf(r1_bio); 2649 2650 if (!ok) { 2651 /* Cannot record the badblocks, so need to 2652 * abort the resync. 2653 * If there are multiple read targets, could just 2654 * fail the really bad ones ??? 2655 */ 2656 conf->recovery_disabled = mddev->recovery_disabled; 2657 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 2658 return 0; 2659 } else 2660 return min_bad; 2661 2662 } 2663 if (min_bad > 0 && min_bad < good_sectors) { 2664 /* only resync enough to reach the next bad->good 2665 * transition */ 2666 good_sectors = min_bad; 2667 } 2668 2669 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0) 2670 /* extra read targets are also write targets */ 2671 write_targets += read_targets-1; 2672 2673 if (write_targets == 0 || read_targets == 0) { 2674 /* There is nowhere to write, so all non-sync 2675 * drives must be failed - so we are finished 2676 */ 2677 sector_t rv; 2678 if (min_bad > 0) 2679 max_sector = sector_nr + min_bad; 2680 rv = max_sector - sector_nr; 2681 *skipped = 1; 2682 put_buf(r1_bio); 2683 return rv; 2684 } 2685 2686 if (max_sector > mddev->resync_max) 2687 max_sector = mddev->resync_max; /* Don't do IO beyond here */ 2688 if (max_sector > sector_nr + good_sectors) 2689 max_sector = sector_nr + good_sectors; 2690 nr_sectors = 0; 2691 sync_blocks = 0; 2692 do { 2693 struct page *page; 2694 int len = PAGE_SIZE; 2695 if (sector_nr + (len>>9) > max_sector) 2696 len = (max_sector - sector_nr) << 9; 2697 if (len == 0) 2698 break; 2699 if (sync_blocks == 0) { 2700 if (!bitmap_start_sync(mddev->bitmap, sector_nr, 2701 &sync_blocks, still_degraded) && 2702 !conf->fullsync && 2703 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2704 break; 2705 BUG_ON(sync_blocks < (PAGE_SIZE>>9)); 2706 if ((len >> 9) > sync_blocks) 2707 len = sync_blocks<<9; 2708 } 2709 2710 for (i = 0 ; i < conf->raid_disks * 2; i++) { 2711 bio = r1_bio->bios[i]; 2712 if (bio->bi_end_io) { 2713 page = bio->bi_io_vec[bio->bi_vcnt].bv_page; 2714 if (bio_add_page(bio, page, len, 0) == 0) { 2715 /* stop here */ 2716 bio->bi_io_vec[bio->bi_vcnt].bv_page = page; 2717 while (i > 0) { 2718 i--; 2719 bio = r1_bio->bios[i]; 2720 if (bio->bi_end_io==NULL) 2721 continue; 2722 /* remove last page from this bio */ 2723 bio->bi_vcnt--; 2724 bio->bi_iter.bi_size -= len; 2725 bio->bi_flags &= ~(1<< BIO_SEG_VALID); 2726 } 2727 goto bio_full; 2728 } 2729 } 2730 } 2731 nr_sectors += len>>9; 2732 sector_nr += len>>9; 2733 sync_blocks -= (len>>9); 2734 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES); 2735 bio_full: 2736 r1_bio->sectors = nr_sectors; 2737 2738 /* For a user-requested sync, we read all readable devices and do a 2739 * compare 2740 */ 2741 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2742 atomic_set(&r1_bio->remaining, read_targets); 2743 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) { 2744 bio = r1_bio->bios[i]; 2745 if (bio->bi_end_io == end_sync_read) { 2746 read_targets--; 2747 md_sync_acct(bio->bi_bdev, nr_sectors); 2748 generic_make_request(bio); 2749 } 2750 } 2751 } else { 2752 atomic_set(&r1_bio->remaining, 1); 2753 bio = r1_bio->bios[r1_bio->read_disk]; 2754 md_sync_acct(bio->bi_bdev, nr_sectors); 2755 generic_make_request(bio); 2756 2757 } 2758 return nr_sectors; 2759 } 2760 2761 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks) 2762 { 2763 if (sectors) 2764 return sectors; 2765 2766 return mddev->dev_sectors; 2767 } 2768 2769 static struct r1conf *setup_conf(struct mddev *mddev) 2770 { 2771 struct r1conf *conf; 2772 int i; 2773 struct raid1_info *disk; 2774 struct md_rdev *rdev; 2775 int err = -ENOMEM; 2776 2777 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL); 2778 if (!conf) 2779 goto abort; 2780 2781 conf->mirrors = kzalloc(sizeof(struct raid1_info) 2782 * mddev->raid_disks * 2, 2783 GFP_KERNEL); 2784 if (!conf->mirrors) 2785 goto abort; 2786 2787 conf->tmppage = alloc_page(GFP_KERNEL); 2788 if (!conf->tmppage) 2789 goto abort; 2790 2791 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL); 2792 if (!conf->poolinfo) 2793 goto abort; 2794 conf->poolinfo->raid_disks = mddev->raid_disks * 2; 2795 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 2796 r1bio_pool_free, 2797 conf->poolinfo); 2798 if (!conf->r1bio_pool) 2799 goto abort; 2800 2801 conf->poolinfo->mddev = mddev; 2802 2803 err = -EINVAL; 2804 spin_lock_init(&conf->device_lock); 2805 rdev_for_each(rdev, mddev) { 2806 struct request_queue *q; 2807 int disk_idx = rdev->raid_disk; 2808 if (disk_idx >= mddev->raid_disks 2809 || disk_idx < 0) 2810 continue; 2811 if (test_bit(Replacement, &rdev->flags)) 2812 disk = conf->mirrors + mddev->raid_disks + disk_idx; 2813 else 2814 disk = conf->mirrors + disk_idx; 2815 2816 if (disk->rdev) 2817 goto abort; 2818 disk->rdev = rdev; 2819 q = bdev_get_queue(rdev->bdev); 2820 if (q->merge_bvec_fn) 2821 mddev->merge_check_needed = 1; 2822 2823 disk->head_position = 0; 2824 disk->seq_start = MaxSector; 2825 } 2826 conf->raid_disks = mddev->raid_disks; 2827 conf->mddev = mddev; 2828 INIT_LIST_HEAD(&conf->retry_list); 2829 2830 spin_lock_init(&conf->resync_lock); 2831 init_waitqueue_head(&conf->wait_barrier); 2832 2833 bio_list_init(&conf->pending_bio_list); 2834 conf->pending_count = 0; 2835 conf->recovery_disabled = mddev->recovery_disabled - 1; 2836 2837 conf->start_next_window = MaxSector; 2838 conf->current_window_requests = conf->next_window_requests = 0; 2839 2840 err = -EIO; 2841 for (i = 0; i < conf->raid_disks * 2; i++) { 2842 2843 disk = conf->mirrors + i; 2844 2845 if (i < conf->raid_disks && 2846 disk[conf->raid_disks].rdev) { 2847 /* This slot has a replacement. */ 2848 if (!disk->rdev) { 2849 /* No original, just make the replacement 2850 * a recovering spare 2851 */ 2852 disk->rdev = 2853 disk[conf->raid_disks].rdev; 2854 disk[conf->raid_disks].rdev = NULL; 2855 } else if (!test_bit(In_sync, &disk->rdev->flags)) 2856 /* Original is not in_sync - bad */ 2857 goto abort; 2858 } 2859 2860 if (!disk->rdev || 2861 !test_bit(In_sync, &disk->rdev->flags)) { 2862 disk->head_position = 0; 2863 if (disk->rdev && 2864 (disk->rdev->saved_raid_disk < 0)) 2865 conf->fullsync = 1; 2866 } 2867 } 2868 2869 err = -ENOMEM; 2870 conf->thread = md_register_thread(raid1d, mddev, "raid1"); 2871 if (!conf->thread) { 2872 printk(KERN_ERR 2873 "md/raid1:%s: couldn't allocate thread\n", 2874 mdname(mddev)); 2875 goto abort; 2876 } 2877 2878 return conf; 2879 2880 abort: 2881 if (conf) { 2882 if (conf->r1bio_pool) 2883 mempool_destroy(conf->r1bio_pool); 2884 kfree(conf->mirrors); 2885 safe_put_page(conf->tmppage); 2886 kfree(conf->poolinfo); 2887 kfree(conf); 2888 } 2889 return ERR_PTR(err); 2890 } 2891 2892 static int stop(struct mddev *mddev); 2893 static int run(struct mddev *mddev) 2894 { 2895 struct r1conf *conf; 2896 int i; 2897 struct md_rdev *rdev; 2898 int ret; 2899 bool discard_supported = false; 2900 2901 if (mddev->level != 1) { 2902 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n", 2903 mdname(mddev), mddev->level); 2904 return -EIO; 2905 } 2906 if (mddev->reshape_position != MaxSector) { 2907 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n", 2908 mdname(mddev)); 2909 return -EIO; 2910 } 2911 /* 2912 * copy the already verified devices into our private RAID1 2913 * bookkeeping area. [whatever we allocate in run(), 2914 * should be freed in stop()] 2915 */ 2916 if (mddev->private == NULL) 2917 conf = setup_conf(mddev); 2918 else 2919 conf = mddev->private; 2920 2921 if (IS_ERR(conf)) 2922 return PTR_ERR(conf); 2923 2924 if (mddev->queue) 2925 blk_queue_max_write_same_sectors(mddev->queue, 0); 2926 2927 rdev_for_each(rdev, mddev) { 2928 if (!mddev->gendisk) 2929 continue; 2930 disk_stack_limits(mddev->gendisk, rdev->bdev, 2931 rdev->data_offset << 9); 2932 if (blk_queue_discard(bdev_get_queue(rdev->bdev))) 2933 discard_supported = true; 2934 } 2935 2936 mddev->degraded = 0; 2937 for (i=0; i < conf->raid_disks; i++) 2938 if (conf->mirrors[i].rdev == NULL || 2939 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) || 2940 test_bit(Faulty, &conf->mirrors[i].rdev->flags)) 2941 mddev->degraded++; 2942 2943 if (conf->raid_disks - mddev->degraded == 1) 2944 mddev->recovery_cp = MaxSector; 2945 2946 if (mddev->recovery_cp != MaxSector) 2947 printk(KERN_NOTICE "md/raid1:%s: not clean" 2948 " -- starting background reconstruction\n", 2949 mdname(mddev)); 2950 printk(KERN_INFO 2951 "md/raid1:%s: active with %d out of %d mirrors\n", 2952 mdname(mddev), mddev->raid_disks - mddev->degraded, 2953 mddev->raid_disks); 2954 2955 /* 2956 * Ok, everything is just fine now 2957 */ 2958 mddev->thread = conf->thread; 2959 conf->thread = NULL; 2960 mddev->private = conf; 2961 2962 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0)); 2963 2964 if (mddev->queue) { 2965 mddev->queue->backing_dev_info.congested_fn = raid1_congested; 2966 mddev->queue->backing_dev_info.congested_data = mddev; 2967 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec); 2968 2969 if (discard_supported) 2970 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, 2971 mddev->queue); 2972 else 2973 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, 2974 mddev->queue); 2975 } 2976 2977 ret = md_integrity_register(mddev); 2978 if (ret) 2979 stop(mddev); 2980 return ret; 2981 } 2982 2983 static int stop(struct mddev *mddev) 2984 { 2985 struct r1conf *conf = mddev->private; 2986 struct bitmap *bitmap = mddev->bitmap; 2987 2988 /* wait for behind writes to complete */ 2989 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) { 2990 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n", 2991 mdname(mddev)); 2992 /* need to kick something here to make sure I/O goes? */ 2993 wait_event(bitmap->behind_wait, 2994 atomic_read(&bitmap->behind_writes) == 0); 2995 } 2996 2997 freeze_array(conf, 0); 2998 unfreeze_array(conf); 2999 3000 md_unregister_thread(&mddev->thread); 3001 if (conf->r1bio_pool) 3002 mempool_destroy(conf->r1bio_pool); 3003 kfree(conf->mirrors); 3004 safe_put_page(conf->tmppage); 3005 kfree(conf->poolinfo); 3006 kfree(conf); 3007 mddev->private = NULL; 3008 return 0; 3009 } 3010 3011 static int raid1_resize(struct mddev *mddev, sector_t sectors) 3012 { 3013 /* no resync is happening, and there is enough space 3014 * on all devices, so we can resize. 3015 * We need to make sure resync covers any new space. 3016 * If the array is shrinking we should possibly wait until 3017 * any io in the removed space completes, but it hardly seems 3018 * worth it. 3019 */ 3020 sector_t newsize = raid1_size(mddev, sectors, 0); 3021 if (mddev->external_size && 3022 mddev->array_sectors > newsize) 3023 return -EINVAL; 3024 if (mddev->bitmap) { 3025 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0); 3026 if (ret) 3027 return ret; 3028 } 3029 md_set_array_sectors(mddev, newsize); 3030 set_capacity(mddev->gendisk, mddev->array_sectors); 3031 revalidate_disk(mddev->gendisk); 3032 if (sectors > mddev->dev_sectors && 3033 mddev->recovery_cp > mddev->dev_sectors) { 3034 mddev->recovery_cp = mddev->dev_sectors; 3035 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3036 } 3037 mddev->dev_sectors = sectors; 3038 mddev->resync_max_sectors = sectors; 3039 return 0; 3040 } 3041 3042 static int raid1_reshape(struct mddev *mddev) 3043 { 3044 /* We need to: 3045 * 1/ resize the r1bio_pool 3046 * 2/ resize conf->mirrors 3047 * 3048 * We allocate a new r1bio_pool if we can. 3049 * Then raise a device barrier and wait until all IO stops. 3050 * Then resize conf->mirrors and swap in the new r1bio pool. 3051 * 3052 * At the same time, we "pack" the devices so that all the missing 3053 * devices have the higher raid_disk numbers. 3054 */ 3055 mempool_t *newpool, *oldpool; 3056 struct pool_info *newpoolinfo; 3057 struct raid1_info *newmirrors; 3058 struct r1conf *conf = mddev->private; 3059 int cnt, raid_disks; 3060 unsigned long flags; 3061 int d, d2, err; 3062 3063 /* Cannot change chunk_size, layout, or level */ 3064 if (mddev->chunk_sectors != mddev->new_chunk_sectors || 3065 mddev->layout != mddev->new_layout || 3066 mddev->level != mddev->new_level) { 3067 mddev->new_chunk_sectors = mddev->chunk_sectors; 3068 mddev->new_layout = mddev->layout; 3069 mddev->new_level = mddev->level; 3070 return -EINVAL; 3071 } 3072 3073 err = md_allow_write(mddev); 3074 if (err) 3075 return err; 3076 3077 raid_disks = mddev->raid_disks + mddev->delta_disks; 3078 3079 if (raid_disks < conf->raid_disks) { 3080 cnt=0; 3081 for (d= 0; d < conf->raid_disks; d++) 3082 if (conf->mirrors[d].rdev) 3083 cnt++; 3084 if (cnt > raid_disks) 3085 return -EBUSY; 3086 } 3087 3088 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL); 3089 if (!newpoolinfo) 3090 return -ENOMEM; 3091 newpoolinfo->mddev = mddev; 3092 newpoolinfo->raid_disks = raid_disks * 2; 3093 3094 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 3095 r1bio_pool_free, newpoolinfo); 3096 if (!newpool) { 3097 kfree(newpoolinfo); 3098 return -ENOMEM; 3099 } 3100 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2, 3101 GFP_KERNEL); 3102 if (!newmirrors) { 3103 kfree(newpoolinfo); 3104 mempool_destroy(newpool); 3105 return -ENOMEM; 3106 } 3107 3108 freeze_array(conf, 0); 3109 3110 /* ok, everything is stopped */ 3111 oldpool = conf->r1bio_pool; 3112 conf->r1bio_pool = newpool; 3113 3114 for (d = d2 = 0; d < conf->raid_disks; d++) { 3115 struct md_rdev *rdev = conf->mirrors[d].rdev; 3116 if (rdev && rdev->raid_disk != d2) { 3117 sysfs_unlink_rdev(mddev, rdev); 3118 rdev->raid_disk = d2; 3119 sysfs_unlink_rdev(mddev, rdev); 3120 if (sysfs_link_rdev(mddev, rdev)) 3121 printk(KERN_WARNING 3122 "md/raid1:%s: cannot register rd%d\n", 3123 mdname(mddev), rdev->raid_disk); 3124 } 3125 if (rdev) 3126 newmirrors[d2++].rdev = rdev; 3127 } 3128 kfree(conf->mirrors); 3129 conf->mirrors = newmirrors; 3130 kfree(conf->poolinfo); 3131 conf->poolinfo = newpoolinfo; 3132 3133 spin_lock_irqsave(&conf->device_lock, flags); 3134 mddev->degraded += (raid_disks - conf->raid_disks); 3135 spin_unlock_irqrestore(&conf->device_lock, flags); 3136 conf->raid_disks = mddev->raid_disks = raid_disks; 3137 mddev->delta_disks = 0; 3138 3139 unfreeze_array(conf); 3140 3141 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3142 md_wakeup_thread(mddev->thread); 3143 3144 mempool_destroy(oldpool); 3145 return 0; 3146 } 3147 3148 static void raid1_quiesce(struct mddev *mddev, int state) 3149 { 3150 struct r1conf *conf = mddev->private; 3151 3152 switch(state) { 3153 case 2: /* wake for suspend */ 3154 wake_up(&conf->wait_barrier); 3155 break; 3156 case 1: 3157 freeze_array(conf, 0); 3158 break; 3159 case 0: 3160 unfreeze_array(conf); 3161 break; 3162 } 3163 } 3164 3165 static void *raid1_takeover(struct mddev *mddev) 3166 { 3167 /* raid1 can take over: 3168 * raid5 with 2 devices, any layout or chunk size 3169 */ 3170 if (mddev->level == 5 && mddev->raid_disks == 2) { 3171 struct r1conf *conf; 3172 mddev->new_level = 1; 3173 mddev->new_layout = 0; 3174 mddev->new_chunk_sectors = 0; 3175 conf = setup_conf(mddev); 3176 if (!IS_ERR(conf)) 3177 /* Array must appear to be quiesced */ 3178 conf->array_frozen = 1; 3179 return conf; 3180 } 3181 return ERR_PTR(-EINVAL); 3182 } 3183 3184 static struct md_personality raid1_personality = 3185 { 3186 .name = "raid1", 3187 .level = 1, 3188 .owner = THIS_MODULE, 3189 .make_request = make_request, 3190 .run = run, 3191 .stop = stop, 3192 .status = status, 3193 .error_handler = error, 3194 .hot_add_disk = raid1_add_disk, 3195 .hot_remove_disk= raid1_remove_disk, 3196 .spare_active = raid1_spare_active, 3197 .sync_request = sync_request, 3198 .resize = raid1_resize, 3199 .size = raid1_size, 3200 .check_reshape = raid1_reshape, 3201 .quiesce = raid1_quiesce, 3202 .takeover = raid1_takeover, 3203 }; 3204 3205 static int __init raid_init(void) 3206 { 3207 return register_md_personality(&raid1_personality); 3208 } 3209 3210 static void raid_exit(void) 3211 { 3212 unregister_md_personality(&raid1_personality); 3213 } 3214 3215 module_init(raid_init); 3216 module_exit(raid_exit); 3217 MODULE_LICENSE("GPL"); 3218 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD"); 3219 MODULE_ALIAS("md-personality-3"); /* RAID1 */ 3220 MODULE_ALIAS("md-raid1"); 3221 MODULE_ALIAS("md-level-1"); 3222 3223 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR); 3224