1 /* 2 * raid1.c : Multiple Devices driver for Linux 3 * 4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat 5 * 6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman 7 * 8 * RAID-1 management functions. 9 * 10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000 11 * 12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk> 13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au> 14 * 15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support 16 * bitmapped intelligence in resync: 17 * 18 * - bitmap marked during normal i/o 19 * - bitmap used to skip nondirty blocks during sync 20 * 21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology: 22 * - persistent bitmap code 23 * 24 * This program is free software; you can redistribute it and/or modify 25 * it under the terms of the GNU General Public License as published by 26 * the Free Software Foundation; either version 2, or (at your option) 27 * any later version. 28 * 29 * You should have received a copy of the GNU General Public License 30 * (for example /usr/src/linux/COPYING); if not, write to the Free 31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. 32 */ 33 34 #include <linux/slab.h> 35 #include <linux/delay.h> 36 #include <linux/blkdev.h> 37 #include <linux/module.h> 38 #include <linux/seq_file.h> 39 #include <linux/ratelimit.h> 40 #include "md.h" 41 #include "raid1.h" 42 #include "bitmap.h" 43 44 /* 45 * Number of guaranteed r1bios in case of extreme VM load: 46 */ 47 #define NR_RAID1_BIOS 256 48 49 /* when we get a read error on a read-only array, we redirect to another 50 * device without failing the first device, or trying to over-write to 51 * correct the read error. To keep track of bad blocks on a per-bio 52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer 53 */ 54 #define IO_BLOCKED ((struct bio *)1) 55 /* When we successfully write to a known bad-block, we need to remove the 56 * bad-block marking which must be done from process context. So we record 57 * the success by setting devs[n].bio to IO_MADE_GOOD 58 */ 59 #define IO_MADE_GOOD ((struct bio *)2) 60 61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) 62 63 /* When there are this many requests queue to be written by 64 * the raid1 thread, we become 'congested' to provide back-pressure 65 * for writeback. 66 */ 67 static int max_queued_requests = 1024; 68 69 static void allow_barrier(struct r1conf *conf, sector_t start_next_window, 70 sector_t bi_sector); 71 static void lower_barrier(struct r1conf *conf); 72 73 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data) 74 { 75 struct pool_info *pi = data; 76 int size = offsetof(struct r1bio, bios[pi->raid_disks]); 77 78 /* allocate a r1bio with room for raid_disks entries in the bios array */ 79 return kzalloc(size, gfp_flags); 80 } 81 82 static void r1bio_pool_free(void *r1_bio, void *data) 83 { 84 kfree(r1_bio); 85 } 86 87 #define RESYNC_BLOCK_SIZE (64*1024) 88 #define RESYNC_DEPTH 32 89 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9) 90 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE) 91 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH) 92 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9) 93 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS) 94 95 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data) 96 { 97 struct pool_info *pi = data; 98 struct r1bio *r1_bio; 99 struct bio *bio; 100 int need_pages; 101 int i, j; 102 103 r1_bio = r1bio_pool_alloc(gfp_flags, pi); 104 if (!r1_bio) 105 return NULL; 106 107 /* 108 * Allocate bios : 1 for reading, n-1 for writing 109 */ 110 for (j = pi->raid_disks ; j-- ; ) { 111 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES); 112 if (!bio) 113 goto out_free_bio; 114 r1_bio->bios[j] = bio; 115 } 116 /* 117 * Allocate RESYNC_PAGES data pages and attach them to 118 * the first bio. 119 * If this is a user-requested check/repair, allocate 120 * RESYNC_PAGES for each bio. 121 */ 122 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) 123 need_pages = pi->raid_disks; 124 else 125 need_pages = 1; 126 for (j = 0; j < need_pages; j++) { 127 bio = r1_bio->bios[j]; 128 bio->bi_vcnt = RESYNC_PAGES; 129 130 if (bio_alloc_pages(bio, gfp_flags)) 131 goto out_free_pages; 132 } 133 /* If not user-requests, copy the page pointers to all bios */ 134 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) { 135 for (i=0; i<RESYNC_PAGES ; i++) 136 for (j=1; j<pi->raid_disks; j++) 137 r1_bio->bios[j]->bi_io_vec[i].bv_page = 138 r1_bio->bios[0]->bi_io_vec[i].bv_page; 139 } 140 141 r1_bio->master_bio = NULL; 142 143 return r1_bio; 144 145 out_free_pages: 146 while (--j >= 0) { 147 struct bio_vec *bv; 148 149 bio_for_each_segment_all(bv, r1_bio->bios[j], i) 150 __free_page(bv->bv_page); 151 } 152 153 out_free_bio: 154 while (++j < pi->raid_disks) 155 bio_put(r1_bio->bios[j]); 156 r1bio_pool_free(r1_bio, data); 157 return NULL; 158 } 159 160 static void r1buf_pool_free(void *__r1_bio, void *data) 161 { 162 struct pool_info *pi = data; 163 int i,j; 164 struct r1bio *r1bio = __r1_bio; 165 166 for (i = 0; i < RESYNC_PAGES; i++) 167 for (j = pi->raid_disks; j-- ;) { 168 if (j == 0 || 169 r1bio->bios[j]->bi_io_vec[i].bv_page != 170 r1bio->bios[0]->bi_io_vec[i].bv_page) 171 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page); 172 } 173 for (i=0 ; i < pi->raid_disks; i++) 174 bio_put(r1bio->bios[i]); 175 176 r1bio_pool_free(r1bio, data); 177 } 178 179 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio) 180 { 181 int i; 182 183 for (i = 0; i < conf->raid_disks * 2; i++) { 184 struct bio **bio = r1_bio->bios + i; 185 if (!BIO_SPECIAL(*bio)) 186 bio_put(*bio); 187 *bio = NULL; 188 } 189 } 190 191 static void free_r1bio(struct r1bio *r1_bio) 192 { 193 struct r1conf *conf = r1_bio->mddev->private; 194 195 put_all_bios(conf, r1_bio); 196 mempool_free(r1_bio, conf->r1bio_pool); 197 } 198 199 static void put_buf(struct r1bio *r1_bio) 200 { 201 struct r1conf *conf = r1_bio->mddev->private; 202 int i; 203 204 for (i = 0; i < conf->raid_disks * 2; i++) { 205 struct bio *bio = r1_bio->bios[i]; 206 if (bio->bi_end_io) 207 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev); 208 } 209 210 mempool_free(r1_bio, conf->r1buf_pool); 211 212 lower_barrier(conf); 213 } 214 215 static void reschedule_retry(struct r1bio *r1_bio) 216 { 217 unsigned long flags; 218 struct mddev *mddev = r1_bio->mddev; 219 struct r1conf *conf = mddev->private; 220 221 spin_lock_irqsave(&conf->device_lock, flags); 222 list_add(&r1_bio->retry_list, &conf->retry_list); 223 conf->nr_queued ++; 224 spin_unlock_irqrestore(&conf->device_lock, flags); 225 226 wake_up(&conf->wait_barrier); 227 md_wakeup_thread(mddev->thread); 228 } 229 230 /* 231 * raid_end_bio_io() is called when we have finished servicing a mirrored 232 * operation and are ready to return a success/failure code to the buffer 233 * cache layer. 234 */ 235 static void call_bio_endio(struct r1bio *r1_bio) 236 { 237 struct bio *bio = r1_bio->master_bio; 238 int done; 239 struct r1conf *conf = r1_bio->mddev->private; 240 sector_t start_next_window = r1_bio->start_next_window; 241 sector_t bi_sector = bio->bi_iter.bi_sector; 242 243 if (bio->bi_phys_segments) { 244 unsigned long flags; 245 spin_lock_irqsave(&conf->device_lock, flags); 246 bio->bi_phys_segments--; 247 done = (bio->bi_phys_segments == 0); 248 spin_unlock_irqrestore(&conf->device_lock, flags); 249 /* 250 * make_request() might be waiting for 251 * bi_phys_segments to decrease 252 */ 253 wake_up(&conf->wait_barrier); 254 } else 255 done = 1; 256 257 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 258 clear_bit(BIO_UPTODATE, &bio->bi_flags); 259 if (done) { 260 bio_endio(bio, 0); 261 /* 262 * Wake up any possible resync thread that waits for the device 263 * to go idle. 264 */ 265 allow_barrier(conf, start_next_window, bi_sector); 266 } 267 } 268 269 static void raid_end_bio_io(struct r1bio *r1_bio) 270 { 271 struct bio *bio = r1_bio->master_bio; 272 273 /* if nobody has done the final endio yet, do it now */ 274 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { 275 pr_debug("raid1: sync end %s on sectors %llu-%llu\n", 276 (bio_data_dir(bio) == WRITE) ? "write" : "read", 277 (unsigned long long) bio->bi_iter.bi_sector, 278 (unsigned long long) bio_end_sector(bio) - 1); 279 280 call_bio_endio(r1_bio); 281 } 282 free_r1bio(r1_bio); 283 } 284 285 /* 286 * Update disk head position estimator based on IRQ completion info. 287 */ 288 static inline void update_head_pos(int disk, struct r1bio *r1_bio) 289 { 290 struct r1conf *conf = r1_bio->mddev->private; 291 292 conf->mirrors[disk].head_position = 293 r1_bio->sector + (r1_bio->sectors); 294 } 295 296 /* 297 * Find the disk number which triggered given bio 298 */ 299 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio) 300 { 301 int mirror; 302 struct r1conf *conf = r1_bio->mddev->private; 303 int raid_disks = conf->raid_disks; 304 305 for (mirror = 0; mirror < raid_disks * 2; mirror++) 306 if (r1_bio->bios[mirror] == bio) 307 break; 308 309 BUG_ON(mirror == raid_disks * 2); 310 update_head_pos(mirror, r1_bio); 311 312 return mirror; 313 } 314 315 static void raid1_end_read_request(struct bio *bio, int error) 316 { 317 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 318 struct r1bio *r1_bio = bio->bi_private; 319 int mirror; 320 struct r1conf *conf = r1_bio->mddev->private; 321 322 mirror = r1_bio->read_disk; 323 /* 324 * this branch is our 'one mirror IO has finished' event handler: 325 */ 326 update_head_pos(mirror, r1_bio); 327 328 if (uptodate) 329 set_bit(R1BIO_Uptodate, &r1_bio->state); 330 else { 331 /* If all other devices have failed, we want to return 332 * the error upwards rather than fail the last device. 333 * Here we redefine "uptodate" to mean "Don't want to retry" 334 */ 335 unsigned long flags; 336 spin_lock_irqsave(&conf->device_lock, flags); 337 if (r1_bio->mddev->degraded == conf->raid_disks || 338 (r1_bio->mddev->degraded == conf->raid_disks-1 && 339 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))) 340 uptodate = 1; 341 spin_unlock_irqrestore(&conf->device_lock, flags); 342 } 343 344 if (uptodate) { 345 raid_end_bio_io(r1_bio); 346 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev); 347 } else { 348 /* 349 * oops, read error: 350 */ 351 char b[BDEVNAME_SIZE]; 352 printk_ratelimited( 353 KERN_ERR "md/raid1:%s: %s: " 354 "rescheduling sector %llu\n", 355 mdname(conf->mddev), 356 bdevname(conf->mirrors[mirror].rdev->bdev, 357 b), 358 (unsigned long long)r1_bio->sector); 359 set_bit(R1BIO_ReadError, &r1_bio->state); 360 reschedule_retry(r1_bio); 361 /* don't drop the reference on read_disk yet */ 362 } 363 } 364 365 static void close_write(struct r1bio *r1_bio) 366 { 367 /* it really is the end of this request */ 368 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 369 /* free extra copy of the data pages */ 370 int i = r1_bio->behind_page_count; 371 while (i--) 372 safe_put_page(r1_bio->behind_bvecs[i].bv_page); 373 kfree(r1_bio->behind_bvecs); 374 r1_bio->behind_bvecs = NULL; 375 } 376 /* clear the bitmap if all writes complete successfully */ 377 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector, 378 r1_bio->sectors, 379 !test_bit(R1BIO_Degraded, &r1_bio->state), 380 test_bit(R1BIO_BehindIO, &r1_bio->state)); 381 md_write_end(r1_bio->mddev); 382 } 383 384 static void r1_bio_write_done(struct r1bio *r1_bio) 385 { 386 if (!atomic_dec_and_test(&r1_bio->remaining)) 387 return; 388 389 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 390 reschedule_retry(r1_bio); 391 else { 392 close_write(r1_bio); 393 if (test_bit(R1BIO_MadeGood, &r1_bio->state)) 394 reschedule_retry(r1_bio); 395 else 396 raid_end_bio_io(r1_bio); 397 } 398 } 399 400 static void raid1_end_write_request(struct bio *bio, int error) 401 { 402 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 403 struct r1bio *r1_bio = bio->bi_private; 404 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state); 405 struct r1conf *conf = r1_bio->mddev->private; 406 struct bio *to_put = NULL; 407 408 mirror = find_bio_disk(r1_bio, bio); 409 410 /* 411 * 'one mirror IO has finished' event handler: 412 */ 413 if (!uptodate) { 414 set_bit(WriteErrorSeen, 415 &conf->mirrors[mirror].rdev->flags); 416 if (!test_and_set_bit(WantReplacement, 417 &conf->mirrors[mirror].rdev->flags)) 418 set_bit(MD_RECOVERY_NEEDED, & 419 conf->mddev->recovery); 420 421 set_bit(R1BIO_WriteError, &r1_bio->state); 422 } else { 423 /* 424 * Set R1BIO_Uptodate in our master bio, so that we 425 * will return a good error code for to the higher 426 * levels even if IO on some other mirrored buffer 427 * fails. 428 * 429 * The 'master' represents the composite IO operation 430 * to user-side. So if something waits for IO, then it 431 * will wait for the 'master' bio. 432 */ 433 sector_t first_bad; 434 int bad_sectors; 435 436 r1_bio->bios[mirror] = NULL; 437 to_put = bio; 438 /* 439 * Do not set R1BIO_Uptodate if the current device is 440 * rebuilding or Faulty. This is because we cannot use 441 * such device for properly reading the data back (we could 442 * potentially use it, if the current write would have felt 443 * before rdev->recovery_offset, but for simplicity we don't 444 * check this here. 445 */ 446 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) && 447 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)) 448 set_bit(R1BIO_Uptodate, &r1_bio->state); 449 450 /* Maybe we can clear some bad blocks. */ 451 if (is_badblock(conf->mirrors[mirror].rdev, 452 r1_bio->sector, r1_bio->sectors, 453 &first_bad, &bad_sectors)) { 454 r1_bio->bios[mirror] = IO_MADE_GOOD; 455 set_bit(R1BIO_MadeGood, &r1_bio->state); 456 } 457 } 458 459 if (behind) { 460 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags)) 461 atomic_dec(&r1_bio->behind_remaining); 462 463 /* 464 * In behind mode, we ACK the master bio once the I/O 465 * has safely reached all non-writemostly 466 * disks. Setting the Returned bit ensures that this 467 * gets done only once -- we don't ever want to return 468 * -EIO here, instead we'll wait 469 */ 470 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) && 471 test_bit(R1BIO_Uptodate, &r1_bio->state)) { 472 /* Maybe we can return now */ 473 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { 474 struct bio *mbio = r1_bio->master_bio; 475 pr_debug("raid1: behind end write sectors" 476 " %llu-%llu\n", 477 (unsigned long long) mbio->bi_iter.bi_sector, 478 (unsigned long long) bio_end_sector(mbio) - 1); 479 call_bio_endio(r1_bio); 480 } 481 } 482 } 483 if (r1_bio->bios[mirror] == NULL) 484 rdev_dec_pending(conf->mirrors[mirror].rdev, 485 conf->mddev); 486 487 /* 488 * Let's see if all mirrored write operations have finished 489 * already. 490 */ 491 r1_bio_write_done(r1_bio); 492 493 if (to_put) 494 bio_put(to_put); 495 } 496 497 /* 498 * This routine returns the disk from which the requested read should 499 * be done. There is a per-array 'next expected sequential IO' sector 500 * number - if this matches on the next IO then we use the last disk. 501 * There is also a per-disk 'last know head position' sector that is 502 * maintained from IRQ contexts, both the normal and the resync IO 503 * completion handlers update this position correctly. If there is no 504 * perfect sequential match then we pick the disk whose head is closest. 505 * 506 * If there are 2 mirrors in the same 2 devices, performance degrades 507 * because position is mirror, not device based. 508 * 509 * The rdev for the device selected will have nr_pending incremented. 510 */ 511 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors) 512 { 513 const sector_t this_sector = r1_bio->sector; 514 int sectors; 515 int best_good_sectors; 516 int best_disk, best_dist_disk, best_pending_disk; 517 int has_nonrot_disk; 518 int disk; 519 sector_t best_dist; 520 unsigned int min_pending; 521 struct md_rdev *rdev; 522 int choose_first; 523 int choose_next_idle; 524 525 rcu_read_lock(); 526 /* 527 * Check if we can balance. We can balance on the whole 528 * device if no resync is going on, or below the resync window. 529 * We take the first readable disk when above the resync window. 530 */ 531 retry: 532 sectors = r1_bio->sectors; 533 best_disk = -1; 534 best_dist_disk = -1; 535 best_dist = MaxSector; 536 best_pending_disk = -1; 537 min_pending = UINT_MAX; 538 best_good_sectors = 0; 539 has_nonrot_disk = 0; 540 choose_next_idle = 0; 541 542 choose_first = (conf->mddev->recovery_cp < this_sector + sectors); 543 544 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) { 545 sector_t dist; 546 sector_t first_bad; 547 int bad_sectors; 548 unsigned int pending; 549 bool nonrot; 550 551 rdev = rcu_dereference(conf->mirrors[disk].rdev); 552 if (r1_bio->bios[disk] == IO_BLOCKED 553 || rdev == NULL 554 || test_bit(Unmerged, &rdev->flags) 555 || test_bit(Faulty, &rdev->flags)) 556 continue; 557 if (!test_bit(In_sync, &rdev->flags) && 558 rdev->recovery_offset < this_sector + sectors) 559 continue; 560 if (test_bit(WriteMostly, &rdev->flags)) { 561 /* Don't balance among write-mostly, just 562 * use the first as a last resort */ 563 if (best_disk < 0) { 564 if (is_badblock(rdev, this_sector, sectors, 565 &first_bad, &bad_sectors)) { 566 if (first_bad < this_sector) 567 /* Cannot use this */ 568 continue; 569 best_good_sectors = first_bad - this_sector; 570 } else 571 best_good_sectors = sectors; 572 best_disk = disk; 573 } 574 continue; 575 } 576 /* This is a reasonable device to use. It might 577 * even be best. 578 */ 579 if (is_badblock(rdev, this_sector, sectors, 580 &first_bad, &bad_sectors)) { 581 if (best_dist < MaxSector) 582 /* already have a better device */ 583 continue; 584 if (first_bad <= this_sector) { 585 /* cannot read here. If this is the 'primary' 586 * device, then we must not read beyond 587 * bad_sectors from another device.. 588 */ 589 bad_sectors -= (this_sector - first_bad); 590 if (choose_first && sectors > bad_sectors) 591 sectors = bad_sectors; 592 if (best_good_sectors > sectors) 593 best_good_sectors = sectors; 594 595 } else { 596 sector_t good_sectors = first_bad - this_sector; 597 if (good_sectors > best_good_sectors) { 598 best_good_sectors = good_sectors; 599 best_disk = disk; 600 } 601 if (choose_first) 602 break; 603 } 604 continue; 605 } else 606 best_good_sectors = sectors; 607 608 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev)); 609 has_nonrot_disk |= nonrot; 610 pending = atomic_read(&rdev->nr_pending); 611 dist = abs(this_sector - conf->mirrors[disk].head_position); 612 if (choose_first) { 613 best_disk = disk; 614 break; 615 } 616 /* Don't change to another disk for sequential reads */ 617 if (conf->mirrors[disk].next_seq_sect == this_sector 618 || dist == 0) { 619 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9; 620 struct raid1_info *mirror = &conf->mirrors[disk]; 621 622 best_disk = disk; 623 /* 624 * If buffered sequential IO size exceeds optimal 625 * iosize, check if there is idle disk. If yes, choose 626 * the idle disk. read_balance could already choose an 627 * idle disk before noticing it's a sequential IO in 628 * this disk. This doesn't matter because this disk 629 * will idle, next time it will be utilized after the 630 * first disk has IO size exceeds optimal iosize. In 631 * this way, iosize of the first disk will be optimal 632 * iosize at least. iosize of the second disk might be 633 * small, but not a big deal since when the second disk 634 * starts IO, the first disk is likely still busy. 635 */ 636 if (nonrot && opt_iosize > 0 && 637 mirror->seq_start != MaxSector && 638 mirror->next_seq_sect > opt_iosize && 639 mirror->next_seq_sect - opt_iosize >= 640 mirror->seq_start) { 641 choose_next_idle = 1; 642 continue; 643 } 644 break; 645 } 646 /* If device is idle, use it */ 647 if (pending == 0) { 648 best_disk = disk; 649 break; 650 } 651 652 if (choose_next_idle) 653 continue; 654 655 if (min_pending > pending) { 656 min_pending = pending; 657 best_pending_disk = disk; 658 } 659 660 if (dist < best_dist) { 661 best_dist = dist; 662 best_dist_disk = disk; 663 } 664 } 665 666 /* 667 * If all disks are rotational, choose the closest disk. If any disk is 668 * non-rotational, choose the disk with less pending request even the 669 * disk is rotational, which might/might not be optimal for raids with 670 * mixed ratation/non-rotational disks depending on workload. 671 */ 672 if (best_disk == -1) { 673 if (has_nonrot_disk) 674 best_disk = best_pending_disk; 675 else 676 best_disk = best_dist_disk; 677 } 678 679 if (best_disk >= 0) { 680 rdev = rcu_dereference(conf->mirrors[best_disk].rdev); 681 if (!rdev) 682 goto retry; 683 atomic_inc(&rdev->nr_pending); 684 if (test_bit(Faulty, &rdev->flags)) { 685 /* cannot risk returning a device that failed 686 * before we inc'ed nr_pending 687 */ 688 rdev_dec_pending(rdev, conf->mddev); 689 goto retry; 690 } 691 sectors = best_good_sectors; 692 693 if (conf->mirrors[best_disk].next_seq_sect != this_sector) 694 conf->mirrors[best_disk].seq_start = this_sector; 695 696 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors; 697 } 698 rcu_read_unlock(); 699 *max_sectors = sectors; 700 701 return best_disk; 702 } 703 704 static int raid1_mergeable_bvec(struct request_queue *q, 705 struct bvec_merge_data *bvm, 706 struct bio_vec *biovec) 707 { 708 struct mddev *mddev = q->queuedata; 709 struct r1conf *conf = mddev->private; 710 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev); 711 int max = biovec->bv_len; 712 713 if (mddev->merge_check_needed) { 714 int disk; 715 rcu_read_lock(); 716 for (disk = 0; disk < conf->raid_disks * 2; disk++) { 717 struct md_rdev *rdev = rcu_dereference( 718 conf->mirrors[disk].rdev); 719 if (rdev && !test_bit(Faulty, &rdev->flags)) { 720 struct request_queue *q = 721 bdev_get_queue(rdev->bdev); 722 if (q->merge_bvec_fn) { 723 bvm->bi_sector = sector + 724 rdev->data_offset; 725 bvm->bi_bdev = rdev->bdev; 726 max = min(max, q->merge_bvec_fn( 727 q, bvm, biovec)); 728 } 729 } 730 } 731 rcu_read_unlock(); 732 } 733 return max; 734 735 } 736 737 static int raid1_congested(struct mddev *mddev, int bits) 738 { 739 struct r1conf *conf = mddev->private; 740 int i, ret = 0; 741 742 if ((bits & (1 << BDI_async_congested)) && 743 conf->pending_count >= max_queued_requests) 744 return 1; 745 746 rcu_read_lock(); 747 for (i = 0; i < conf->raid_disks * 2; i++) { 748 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 749 if (rdev && !test_bit(Faulty, &rdev->flags)) { 750 struct request_queue *q = bdev_get_queue(rdev->bdev); 751 752 BUG_ON(!q); 753 754 /* Note the '|| 1' - when read_balance prefers 755 * non-congested targets, it can be removed 756 */ 757 if ((bits & (1<<BDI_async_congested)) || 1) 758 ret |= bdi_congested(&q->backing_dev_info, bits); 759 else 760 ret &= bdi_congested(&q->backing_dev_info, bits); 761 } 762 } 763 rcu_read_unlock(); 764 return ret; 765 } 766 767 static void flush_pending_writes(struct r1conf *conf) 768 { 769 /* Any writes that have been queued but are awaiting 770 * bitmap updates get flushed here. 771 */ 772 spin_lock_irq(&conf->device_lock); 773 774 if (conf->pending_bio_list.head) { 775 struct bio *bio; 776 bio = bio_list_get(&conf->pending_bio_list); 777 conf->pending_count = 0; 778 spin_unlock_irq(&conf->device_lock); 779 /* flush any pending bitmap writes to 780 * disk before proceeding w/ I/O */ 781 bitmap_unplug(conf->mddev->bitmap); 782 wake_up(&conf->wait_barrier); 783 784 while (bio) { /* submit pending writes */ 785 struct bio *next = bio->bi_next; 786 bio->bi_next = NULL; 787 if (unlikely((bio->bi_rw & REQ_DISCARD) && 788 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 789 /* Just ignore it */ 790 bio_endio(bio, 0); 791 else 792 generic_make_request(bio); 793 bio = next; 794 } 795 } else 796 spin_unlock_irq(&conf->device_lock); 797 } 798 799 /* Barriers.... 800 * Sometimes we need to suspend IO while we do something else, 801 * either some resync/recovery, or reconfigure the array. 802 * To do this we raise a 'barrier'. 803 * The 'barrier' is a counter that can be raised multiple times 804 * to count how many activities are happening which preclude 805 * normal IO. 806 * We can only raise the barrier if there is no pending IO. 807 * i.e. if nr_pending == 0. 808 * We choose only to raise the barrier if no-one is waiting for the 809 * barrier to go down. This means that as soon as an IO request 810 * is ready, no other operations which require a barrier will start 811 * until the IO request has had a chance. 812 * 813 * So: regular IO calls 'wait_barrier'. When that returns there 814 * is no backgroup IO happening, It must arrange to call 815 * allow_barrier when it has finished its IO. 816 * backgroup IO calls must call raise_barrier. Once that returns 817 * there is no normal IO happeing. It must arrange to call 818 * lower_barrier when the particular background IO completes. 819 */ 820 static void raise_barrier(struct r1conf *conf, sector_t sector_nr) 821 { 822 spin_lock_irq(&conf->resync_lock); 823 824 /* Wait until no block IO is waiting */ 825 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting, 826 conf->resync_lock); 827 828 /* block any new IO from starting */ 829 conf->barrier++; 830 conf->next_resync = sector_nr; 831 832 /* For these conditions we must wait: 833 * A: while the array is in frozen state 834 * B: while barrier >= RESYNC_DEPTH, meaning resync reach 835 * the max count which allowed. 836 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning 837 * next resync will reach to the window which normal bios are 838 * handling. 839 * D: while there are any active requests in the current window. 840 */ 841 wait_event_lock_irq(conf->wait_barrier, 842 !conf->array_frozen && 843 conf->barrier < RESYNC_DEPTH && 844 conf->current_window_requests == 0 && 845 (conf->start_next_window >= 846 conf->next_resync + RESYNC_SECTORS), 847 conf->resync_lock); 848 849 conf->nr_pending++; 850 spin_unlock_irq(&conf->resync_lock); 851 } 852 853 static void lower_barrier(struct r1conf *conf) 854 { 855 unsigned long flags; 856 BUG_ON(conf->barrier <= 0); 857 spin_lock_irqsave(&conf->resync_lock, flags); 858 conf->barrier--; 859 conf->nr_pending--; 860 spin_unlock_irqrestore(&conf->resync_lock, flags); 861 wake_up(&conf->wait_barrier); 862 } 863 864 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio) 865 { 866 bool wait = false; 867 868 if (conf->array_frozen || !bio) 869 wait = true; 870 else if (conf->barrier && bio_data_dir(bio) == WRITE) { 871 if ((conf->mddev->curr_resync_completed 872 >= bio_end_sector(bio)) || 873 (conf->next_resync + NEXT_NORMALIO_DISTANCE 874 <= bio->bi_iter.bi_sector)) 875 wait = false; 876 else 877 wait = true; 878 } 879 880 return wait; 881 } 882 883 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio) 884 { 885 sector_t sector = 0; 886 887 spin_lock_irq(&conf->resync_lock); 888 if (need_to_wait_for_sync(conf, bio)) { 889 conf->nr_waiting++; 890 /* Wait for the barrier to drop. 891 * However if there are already pending 892 * requests (preventing the barrier from 893 * rising completely), and the 894 * per-process bio queue isn't empty, 895 * then don't wait, as we need to empty 896 * that queue to allow conf->start_next_window 897 * to increase. 898 */ 899 wait_event_lock_irq(conf->wait_barrier, 900 !conf->array_frozen && 901 (!conf->barrier || 902 ((conf->start_next_window < 903 conf->next_resync + RESYNC_SECTORS) && 904 current->bio_list && 905 !bio_list_empty(current->bio_list))), 906 conf->resync_lock); 907 conf->nr_waiting--; 908 } 909 910 if (bio && bio_data_dir(bio) == WRITE) { 911 if (bio->bi_iter.bi_sector >= 912 conf->mddev->curr_resync_completed) { 913 if (conf->start_next_window == MaxSector) 914 conf->start_next_window = 915 conf->next_resync + 916 NEXT_NORMALIO_DISTANCE; 917 918 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE) 919 <= bio->bi_iter.bi_sector) 920 conf->next_window_requests++; 921 else 922 conf->current_window_requests++; 923 sector = conf->start_next_window; 924 } 925 } 926 927 conf->nr_pending++; 928 spin_unlock_irq(&conf->resync_lock); 929 return sector; 930 } 931 932 static void allow_barrier(struct r1conf *conf, sector_t start_next_window, 933 sector_t bi_sector) 934 { 935 unsigned long flags; 936 937 spin_lock_irqsave(&conf->resync_lock, flags); 938 conf->nr_pending--; 939 if (start_next_window) { 940 if (start_next_window == conf->start_next_window) { 941 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE 942 <= bi_sector) 943 conf->next_window_requests--; 944 else 945 conf->current_window_requests--; 946 } else 947 conf->current_window_requests--; 948 949 if (!conf->current_window_requests) { 950 if (conf->next_window_requests) { 951 conf->current_window_requests = 952 conf->next_window_requests; 953 conf->next_window_requests = 0; 954 conf->start_next_window += 955 NEXT_NORMALIO_DISTANCE; 956 } else 957 conf->start_next_window = MaxSector; 958 } 959 } 960 spin_unlock_irqrestore(&conf->resync_lock, flags); 961 wake_up(&conf->wait_barrier); 962 } 963 964 static void freeze_array(struct r1conf *conf, int extra) 965 { 966 /* stop syncio and normal IO and wait for everything to 967 * go quite. 968 * We wait until nr_pending match nr_queued+extra 969 * This is called in the context of one normal IO request 970 * that has failed. Thus any sync request that might be pending 971 * will be blocked by nr_pending, and we need to wait for 972 * pending IO requests to complete or be queued for re-try. 973 * Thus the number queued (nr_queued) plus this request (extra) 974 * must match the number of pending IOs (nr_pending) before 975 * we continue. 976 */ 977 spin_lock_irq(&conf->resync_lock); 978 conf->array_frozen = 1; 979 wait_event_lock_irq_cmd(conf->wait_barrier, 980 conf->nr_pending == conf->nr_queued+extra, 981 conf->resync_lock, 982 flush_pending_writes(conf)); 983 spin_unlock_irq(&conf->resync_lock); 984 } 985 static void unfreeze_array(struct r1conf *conf) 986 { 987 /* reverse the effect of the freeze */ 988 spin_lock_irq(&conf->resync_lock); 989 conf->array_frozen = 0; 990 wake_up(&conf->wait_barrier); 991 spin_unlock_irq(&conf->resync_lock); 992 } 993 994 /* duplicate the data pages for behind I/O 995 */ 996 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio) 997 { 998 int i; 999 struct bio_vec *bvec; 1000 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec), 1001 GFP_NOIO); 1002 if (unlikely(!bvecs)) 1003 return; 1004 1005 bio_for_each_segment_all(bvec, bio, i) { 1006 bvecs[i] = *bvec; 1007 bvecs[i].bv_page = alloc_page(GFP_NOIO); 1008 if (unlikely(!bvecs[i].bv_page)) 1009 goto do_sync_io; 1010 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset, 1011 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len); 1012 kunmap(bvecs[i].bv_page); 1013 kunmap(bvec->bv_page); 1014 } 1015 r1_bio->behind_bvecs = bvecs; 1016 r1_bio->behind_page_count = bio->bi_vcnt; 1017 set_bit(R1BIO_BehindIO, &r1_bio->state); 1018 return; 1019 1020 do_sync_io: 1021 for (i = 0; i < bio->bi_vcnt; i++) 1022 if (bvecs[i].bv_page) 1023 put_page(bvecs[i].bv_page); 1024 kfree(bvecs); 1025 pr_debug("%dB behind alloc failed, doing sync I/O\n", 1026 bio->bi_iter.bi_size); 1027 } 1028 1029 struct raid1_plug_cb { 1030 struct blk_plug_cb cb; 1031 struct bio_list pending; 1032 int pending_cnt; 1033 }; 1034 1035 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule) 1036 { 1037 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, 1038 cb); 1039 struct mddev *mddev = plug->cb.data; 1040 struct r1conf *conf = mddev->private; 1041 struct bio *bio; 1042 1043 if (from_schedule || current->bio_list) { 1044 spin_lock_irq(&conf->device_lock); 1045 bio_list_merge(&conf->pending_bio_list, &plug->pending); 1046 conf->pending_count += plug->pending_cnt; 1047 spin_unlock_irq(&conf->device_lock); 1048 wake_up(&conf->wait_barrier); 1049 md_wakeup_thread(mddev->thread); 1050 kfree(plug); 1051 return; 1052 } 1053 1054 /* we aren't scheduling, so we can do the write-out directly. */ 1055 bio = bio_list_get(&plug->pending); 1056 bitmap_unplug(mddev->bitmap); 1057 wake_up(&conf->wait_barrier); 1058 1059 while (bio) { /* submit pending writes */ 1060 struct bio *next = bio->bi_next; 1061 bio->bi_next = NULL; 1062 if (unlikely((bio->bi_rw & REQ_DISCARD) && 1063 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 1064 /* Just ignore it */ 1065 bio_endio(bio, 0); 1066 else 1067 generic_make_request(bio); 1068 bio = next; 1069 } 1070 kfree(plug); 1071 } 1072 1073 static void make_request(struct mddev *mddev, struct bio * bio) 1074 { 1075 struct r1conf *conf = mddev->private; 1076 struct raid1_info *mirror; 1077 struct r1bio *r1_bio; 1078 struct bio *read_bio; 1079 int i, disks; 1080 struct bitmap *bitmap; 1081 unsigned long flags; 1082 const int rw = bio_data_dir(bio); 1083 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC); 1084 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA)); 1085 const unsigned long do_discard = (bio->bi_rw 1086 & (REQ_DISCARD | REQ_SECURE)); 1087 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME); 1088 struct md_rdev *blocked_rdev; 1089 struct blk_plug_cb *cb; 1090 struct raid1_plug_cb *plug = NULL; 1091 int first_clone; 1092 int sectors_handled; 1093 int max_sectors; 1094 sector_t start_next_window; 1095 1096 /* 1097 * Register the new request and wait if the reconstruction 1098 * thread has put up a bar for new requests. 1099 * Continue immediately if no resync is active currently. 1100 */ 1101 1102 md_write_start(mddev, bio); /* wait on superblock update early */ 1103 1104 if (bio_data_dir(bio) == WRITE && 1105 bio_end_sector(bio) > mddev->suspend_lo && 1106 bio->bi_iter.bi_sector < mddev->suspend_hi) { 1107 /* As the suspend_* range is controlled by 1108 * userspace, we want an interruptible 1109 * wait. 1110 */ 1111 DEFINE_WAIT(w); 1112 for (;;) { 1113 flush_signals(current); 1114 prepare_to_wait(&conf->wait_barrier, 1115 &w, TASK_INTERRUPTIBLE); 1116 if (bio_end_sector(bio) <= mddev->suspend_lo || 1117 bio->bi_iter.bi_sector >= mddev->suspend_hi) 1118 break; 1119 schedule(); 1120 } 1121 finish_wait(&conf->wait_barrier, &w); 1122 } 1123 1124 start_next_window = wait_barrier(conf, bio); 1125 1126 bitmap = mddev->bitmap; 1127 1128 /* 1129 * make_request() can abort the operation when READA is being 1130 * used and no empty request is available. 1131 * 1132 */ 1133 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1134 1135 r1_bio->master_bio = bio; 1136 r1_bio->sectors = bio_sectors(bio); 1137 r1_bio->state = 0; 1138 r1_bio->mddev = mddev; 1139 r1_bio->sector = bio->bi_iter.bi_sector; 1140 1141 /* We might need to issue multiple reads to different 1142 * devices if there are bad blocks around, so we keep 1143 * track of the number of reads in bio->bi_phys_segments. 1144 * If this is 0, there is only one r1_bio and no locking 1145 * will be needed when requests complete. If it is 1146 * non-zero, then it is the number of not-completed requests. 1147 */ 1148 bio->bi_phys_segments = 0; 1149 clear_bit(BIO_SEG_VALID, &bio->bi_flags); 1150 1151 if (rw == READ) { 1152 /* 1153 * read balancing logic: 1154 */ 1155 int rdisk; 1156 1157 read_again: 1158 rdisk = read_balance(conf, r1_bio, &max_sectors); 1159 1160 if (rdisk < 0) { 1161 /* couldn't find anywhere to read from */ 1162 raid_end_bio_io(r1_bio); 1163 return; 1164 } 1165 mirror = conf->mirrors + rdisk; 1166 1167 if (test_bit(WriteMostly, &mirror->rdev->flags) && 1168 bitmap) { 1169 /* Reading from a write-mostly device must 1170 * take care not to over-take any writes 1171 * that are 'behind' 1172 */ 1173 wait_event(bitmap->behind_wait, 1174 atomic_read(&bitmap->behind_writes) == 0); 1175 } 1176 r1_bio->read_disk = rdisk; 1177 r1_bio->start_next_window = 0; 1178 1179 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1180 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector, 1181 max_sectors); 1182 1183 r1_bio->bios[rdisk] = read_bio; 1184 1185 read_bio->bi_iter.bi_sector = r1_bio->sector + 1186 mirror->rdev->data_offset; 1187 read_bio->bi_bdev = mirror->rdev->bdev; 1188 read_bio->bi_end_io = raid1_end_read_request; 1189 read_bio->bi_rw = READ | do_sync; 1190 read_bio->bi_private = r1_bio; 1191 1192 if (max_sectors < r1_bio->sectors) { 1193 /* could not read all from this device, so we will 1194 * need another r1_bio. 1195 */ 1196 1197 sectors_handled = (r1_bio->sector + max_sectors 1198 - bio->bi_iter.bi_sector); 1199 r1_bio->sectors = max_sectors; 1200 spin_lock_irq(&conf->device_lock); 1201 if (bio->bi_phys_segments == 0) 1202 bio->bi_phys_segments = 2; 1203 else 1204 bio->bi_phys_segments++; 1205 spin_unlock_irq(&conf->device_lock); 1206 /* Cannot call generic_make_request directly 1207 * as that will be queued in __make_request 1208 * and subsequent mempool_alloc might block waiting 1209 * for it. So hand bio over to raid1d. 1210 */ 1211 reschedule_retry(r1_bio); 1212 1213 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1214 1215 r1_bio->master_bio = bio; 1216 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1217 r1_bio->state = 0; 1218 r1_bio->mddev = mddev; 1219 r1_bio->sector = bio->bi_iter.bi_sector + 1220 sectors_handled; 1221 goto read_again; 1222 } else 1223 generic_make_request(read_bio); 1224 return; 1225 } 1226 1227 /* 1228 * WRITE: 1229 */ 1230 if (conf->pending_count >= max_queued_requests) { 1231 md_wakeup_thread(mddev->thread); 1232 wait_event(conf->wait_barrier, 1233 conf->pending_count < max_queued_requests); 1234 } 1235 /* first select target devices under rcu_lock and 1236 * inc refcount on their rdev. Record them by setting 1237 * bios[x] to bio 1238 * If there are known/acknowledged bad blocks on any device on 1239 * which we have seen a write error, we want to avoid writing those 1240 * blocks. 1241 * This potentially requires several writes to write around 1242 * the bad blocks. Each set of writes gets it's own r1bio 1243 * with a set of bios attached. 1244 */ 1245 1246 disks = conf->raid_disks * 2; 1247 retry_write: 1248 r1_bio->start_next_window = start_next_window; 1249 blocked_rdev = NULL; 1250 rcu_read_lock(); 1251 max_sectors = r1_bio->sectors; 1252 for (i = 0; i < disks; i++) { 1253 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1254 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 1255 atomic_inc(&rdev->nr_pending); 1256 blocked_rdev = rdev; 1257 break; 1258 } 1259 r1_bio->bios[i] = NULL; 1260 if (!rdev || test_bit(Faulty, &rdev->flags) 1261 || test_bit(Unmerged, &rdev->flags)) { 1262 if (i < conf->raid_disks) 1263 set_bit(R1BIO_Degraded, &r1_bio->state); 1264 continue; 1265 } 1266 1267 atomic_inc(&rdev->nr_pending); 1268 if (test_bit(WriteErrorSeen, &rdev->flags)) { 1269 sector_t first_bad; 1270 int bad_sectors; 1271 int is_bad; 1272 1273 is_bad = is_badblock(rdev, r1_bio->sector, 1274 max_sectors, 1275 &first_bad, &bad_sectors); 1276 if (is_bad < 0) { 1277 /* mustn't write here until the bad block is 1278 * acknowledged*/ 1279 set_bit(BlockedBadBlocks, &rdev->flags); 1280 blocked_rdev = rdev; 1281 break; 1282 } 1283 if (is_bad && first_bad <= r1_bio->sector) { 1284 /* Cannot write here at all */ 1285 bad_sectors -= (r1_bio->sector - first_bad); 1286 if (bad_sectors < max_sectors) 1287 /* mustn't write more than bad_sectors 1288 * to other devices yet 1289 */ 1290 max_sectors = bad_sectors; 1291 rdev_dec_pending(rdev, mddev); 1292 /* We don't set R1BIO_Degraded as that 1293 * only applies if the disk is 1294 * missing, so it might be re-added, 1295 * and we want to know to recover this 1296 * chunk. 1297 * In this case the device is here, 1298 * and the fact that this chunk is not 1299 * in-sync is recorded in the bad 1300 * block log 1301 */ 1302 continue; 1303 } 1304 if (is_bad) { 1305 int good_sectors = first_bad - r1_bio->sector; 1306 if (good_sectors < max_sectors) 1307 max_sectors = good_sectors; 1308 } 1309 } 1310 r1_bio->bios[i] = bio; 1311 } 1312 rcu_read_unlock(); 1313 1314 if (unlikely(blocked_rdev)) { 1315 /* Wait for this device to become unblocked */ 1316 int j; 1317 sector_t old = start_next_window; 1318 1319 for (j = 0; j < i; j++) 1320 if (r1_bio->bios[j]) 1321 rdev_dec_pending(conf->mirrors[j].rdev, mddev); 1322 r1_bio->state = 0; 1323 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector); 1324 md_wait_for_blocked_rdev(blocked_rdev, mddev); 1325 start_next_window = wait_barrier(conf, bio); 1326 /* 1327 * We must make sure the multi r1bios of bio have 1328 * the same value of bi_phys_segments 1329 */ 1330 if (bio->bi_phys_segments && old && 1331 old != start_next_window) 1332 /* Wait for the former r1bio(s) to complete */ 1333 wait_event(conf->wait_barrier, 1334 bio->bi_phys_segments == 1); 1335 goto retry_write; 1336 } 1337 1338 if (max_sectors < r1_bio->sectors) { 1339 /* We are splitting this write into multiple parts, so 1340 * we need to prepare for allocating another r1_bio. 1341 */ 1342 r1_bio->sectors = max_sectors; 1343 spin_lock_irq(&conf->device_lock); 1344 if (bio->bi_phys_segments == 0) 1345 bio->bi_phys_segments = 2; 1346 else 1347 bio->bi_phys_segments++; 1348 spin_unlock_irq(&conf->device_lock); 1349 } 1350 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector; 1351 1352 atomic_set(&r1_bio->remaining, 1); 1353 atomic_set(&r1_bio->behind_remaining, 0); 1354 1355 first_clone = 1; 1356 for (i = 0; i < disks; i++) { 1357 struct bio *mbio; 1358 if (!r1_bio->bios[i]) 1359 continue; 1360 1361 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1362 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors); 1363 1364 if (first_clone) { 1365 /* do behind I/O ? 1366 * Not if there are too many, or cannot 1367 * allocate memory, or a reader on WriteMostly 1368 * is waiting for behind writes to flush */ 1369 if (bitmap && 1370 (atomic_read(&bitmap->behind_writes) 1371 < mddev->bitmap_info.max_write_behind) && 1372 !waitqueue_active(&bitmap->behind_wait)) 1373 alloc_behind_pages(mbio, r1_bio); 1374 1375 bitmap_startwrite(bitmap, r1_bio->sector, 1376 r1_bio->sectors, 1377 test_bit(R1BIO_BehindIO, 1378 &r1_bio->state)); 1379 first_clone = 0; 1380 } 1381 if (r1_bio->behind_bvecs) { 1382 struct bio_vec *bvec; 1383 int j; 1384 1385 /* 1386 * We trimmed the bio, so _all is legit 1387 */ 1388 bio_for_each_segment_all(bvec, mbio, j) 1389 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page; 1390 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags)) 1391 atomic_inc(&r1_bio->behind_remaining); 1392 } 1393 1394 r1_bio->bios[i] = mbio; 1395 1396 mbio->bi_iter.bi_sector = (r1_bio->sector + 1397 conf->mirrors[i].rdev->data_offset); 1398 mbio->bi_bdev = conf->mirrors[i].rdev->bdev; 1399 mbio->bi_end_io = raid1_end_write_request; 1400 mbio->bi_rw = 1401 WRITE | do_flush_fua | do_sync | do_discard | do_same; 1402 mbio->bi_private = r1_bio; 1403 1404 atomic_inc(&r1_bio->remaining); 1405 1406 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug)); 1407 if (cb) 1408 plug = container_of(cb, struct raid1_plug_cb, cb); 1409 else 1410 plug = NULL; 1411 spin_lock_irqsave(&conf->device_lock, flags); 1412 if (plug) { 1413 bio_list_add(&plug->pending, mbio); 1414 plug->pending_cnt++; 1415 } else { 1416 bio_list_add(&conf->pending_bio_list, mbio); 1417 conf->pending_count++; 1418 } 1419 spin_unlock_irqrestore(&conf->device_lock, flags); 1420 if (!plug) 1421 md_wakeup_thread(mddev->thread); 1422 } 1423 /* Mustn't call r1_bio_write_done before this next test, 1424 * as it could result in the bio being freed. 1425 */ 1426 if (sectors_handled < bio_sectors(bio)) { 1427 r1_bio_write_done(r1_bio); 1428 /* We need another r1_bio. It has already been counted 1429 * in bio->bi_phys_segments 1430 */ 1431 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1432 r1_bio->master_bio = bio; 1433 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1434 r1_bio->state = 0; 1435 r1_bio->mddev = mddev; 1436 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled; 1437 goto retry_write; 1438 } 1439 1440 r1_bio_write_done(r1_bio); 1441 1442 /* In case raid1d snuck in to freeze_array */ 1443 wake_up(&conf->wait_barrier); 1444 } 1445 1446 static void status(struct seq_file *seq, struct mddev *mddev) 1447 { 1448 struct r1conf *conf = mddev->private; 1449 int i; 1450 1451 seq_printf(seq, " [%d/%d] [", conf->raid_disks, 1452 conf->raid_disks - mddev->degraded); 1453 rcu_read_lock(); 1454 for (i = 0; i < conf->raid_disks; i++) { 1455 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1456 seq_printf(seq, "%s", 1457 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_"); 1458 } 1459 rcu_read_unlock(); 1460 seq_printf(seq, "]"); 1461 } 1462 1463 static void error(struct mddev *mddev, struct md_rdev *rdev) 1464 { 1465 char b[BDEVNAME_SIZE]; 1466 struct r1conf *conf = mddev->private; 1467 1468 /* 1469 * If it is not operational, then we have already marked it as dead 1470 * else if it is the last working disks, ignore the error, let the 1471 * next level up know. 1472 * else mark the drive as failed 1473 */ 1474 if (test_bit(In_sync, &rdev->flags) 1475 && (conf->raid_disks - mddev->degraded) == 1) { 1476 /* 1477 * Don't fail the drive, act as though we were just a 1478 * normal single drive. 1479 * However don't try a recovery from this drive as 1480 * it is very likely to fail. 1481 */ 1482 conf->recovery_disabled = mddev->recovery_disabled; 1483 return; 1484 } 1485 set_bit(Blocked, &rdev->flags); 1486 if (test_and_clear_bit(In_sync, &rdev->flags)) { 1487 unsigned long flags; 1488 spin_lock_irqsave(&conf->device_lock, flags); 1489 mddev->degraded++; 1490 set_bit(Faulty, &rdev->flags); 1491 spin_unlock_irqrestore(&conf->device_lock, flags); 1492 } else 1493 set_bit(Faulty, &rdev->flags); 1494 /* 1495 * if recovery is running, make sure it aborts. 1496 */ 1497 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1498 set_bit(MD_CHANGE_DEVS, &mddev->flags); 1499 printk(KERN_ALERT 1500 "md/raid1:%s: Disk failure on %s, disabling device.\n" 1501 "md/raid1:%s: Operation continuing on %d devices.\n", 1502 mdname(mddev), bdevname(rdev->bdev, b), 1503 mdname(mddev), conf->raid_disks - mddev->degraded); 1504 } 1505 1506 static void print_conf(struct r1conf *conf) 1507 { 1508 int i; 1509 1510 printk(KERN_DEBUG "RAID1 conf printout:\n"); 1511 if (!conf) { 1512 printk(KERN_DEBUG "(!conf)\n"); 1513 return; 1514 } 1515 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded, 1516 conf->raid_disks); 1517 1518 rcu_read_lock(); 1519 for (i = 0; i < conf->raid_disks; i++) { 1520 char b[BDEVNAME_SIZE]; 1521 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1522 if (rdev) 1523 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n", 1524 i, !test_bit(In_sync, &rdev->flags), 1525 !test_bit(Faulty, &rdev->flags), 1526 bdevname(rdev->bdev,b)); 1527 } 1528 rcu_read_unlock(); 1529 } 1530 1531 static void close_sync(struct r1conf *conf) 1532 { 1533 wait_barrier(conf, NULL); 1534 allow_barrier(conf, 0, 0); 1535 1536 mempool_destroy(conf->r1buf_pool); 1537 conf->r1buf_pool = NULL; 1538 1539 spin_lock_irq(&conf->resync_lock); 1540 conf->next_resync = 0; 1541 conf->start_next_window = MaxSector; 1542 conf->current_window_requests += 1543 conf->next_window_requests; 1544 conf->next_window_requests = 0; 1545 spin_unlock_irq(&conf->resync_lock); 1546 } 1547 1548 static int raid1_spare_active(struct mddev *mddev) 1549 { 1550 int i; 1551 struct r1conf *conf = mddev->private; 1552 int count = 0; 1553 unsigned long flags; 1554 1555 /* 1556 * Find all failed disks within the RAID1 configuration 1557 * and mark them readable. 1558 * Called under mddev lock, so rcu protection not needed. 1559 */ 1560 for (i = 0; i < conf->raid_disks; i++) { 1561 struct md_rdev *rdev = conf->mirrors[i].rdev; 1562 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev; 1563 if (repl 1564 && repl->recovery_offset == MaxSector 1565 && !test_bit(Faulty, &repl->flags) 1566 && !test_and_set_bit(In_sync, &repl->flags)) { 1567 /* replacement has just become active */ 1568 if (!rdev || 1569 !test_and_clear_bit(In_sync, &rdev->flags)) 1570 count++; 1571 if (rdev) { 1572 /* Replaced device not technically 1573 * faulty, but we need to be sure 1574 * it gets removed and never re-added 1575 */ 1576 set_bit(Faulty, &rdev->flags); 1577 sysfs_notify_dirent_safe( 1578 rdev->sysfs_state); 1579 } 1580 } 1581 if (rdev 1582 && rdev->recovery_offset == MaxSector 1583 && !test_bit(Faulty, &rdev->flags) 1584 && !test_and_set_bit(In_sync, &rdev->flags)) { 1585 count++; 1586 sysfs_notify_dirent_safe(rdev->sysfs_state); 1587 } 1588 } 1589 spin_lock_irqsave(&conf->device_lock, flags); 1590 mddev->degraded -= count; 1591 spin_unlock_irqrestore(&conf->device_lock, flags); 1592 1593 print_conf(conf); 1594 return count; 1595 } 1596 1597 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev) 1598 { 1599 struct r1conf *conf = mddev->private; 1600 int err = -EEXIST; 1601 int mirror = 0; 1602 struct raid1_info *p; 1603 int first = 0; 1604 int last = conf->raid_disks - 1; 1605 struct request_queue *q = bdev_get_queue(rdev->bdev); 1606 1607 if (mddev->recovery_disabled == conf->recovery_disabled) 1608 return -EBUSY; 1609 1610 if (rdev->raid_disk >= 0) 1611 first = last = rdev->raid_disk; 1612 1613 if (q->merge_bvec_fn) { 1614 set_bit(Unmerged, &rdev->flags); 1615 mddev->merge_check_needed = 1; 1616 } 1617 1618 for (mirror = first; mirror <= last; mirror++) { 1619 p = conf->mirrors+mirror; 1620 if (!p->rdev) { 1621 1622 if (mddev->gendisk) 1623 disk_stack_limits(mddev->gendisk, rdev->bdev, 1624 rdev->data_offset << 9); 1625 1626 p->head_position = 0; 1627 rdev->raid_disk = mirror; 1628 err = 0; 1629 /* As all devices are equivalent, we don't need a full recovery 1630 * if this was recently any drive of the array 1631 */ 1632 if (rdev->saved_raid_disk < 0) 1633 conf->fullsync = 1; 1634 rcu_assign_pointer(p->rdev, rdev); 1635 break; 1636 } 1637 if (test_bit(WantReplacement, &p->rdev->flags) && 1638 p[conf->raid_disks].rdev == NULL) { 1639 /* Add this device as a replacement */ 1640 clear_bit(In_sync, &rdev->flags); 1641 set_bit(Replacement, &rdev->flags); 1642 rdev->raid_disk = mirror; 1643 err = 0; 1644 conf->fullsync = 1; 1645 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev); 1646 break; 1647 } 1648 } 1649 if (err == 0 && test_bit(Unmerged, &rdev->flags)) { 1650 /* Some requests might not have seen this new 1651 * merge_bvec_fn. We must wait for them to complete 1652 * before merging the device fully. 1653 * First we make sure any code which has tested 1654 * our function has submitted the request, then 1655 * we wait for all outstanding requests to complete. 1656 */ 1657 synchronize_sched(); 1658 freeze_array(conf, 0); 1659 unfreeze_array(conf); 1660 clear_bit(Unmerged, &rdev->flags); 1661 } 1662 md_integrity_add_rdev(rdev, mddev); 1663 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev))) 1664 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue); 1665 print_conf(conf); 1666 return err; 1667 } 1668 1669 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev) 1670 { 1671 struct r1conf *conf = mddev->private; 1672 int err = 0; 1673 int number = rdev->raid_disk; 1674 struct raid1_info *p = conf->mirrors + number; 1675 1676 if (rdev != p->rdev) 1677 p = conf->mirrors + conf->raid_disks + number; 1678 1679 print_conf(conf); 1680 if (rdev == p->rdev) { 1681 if (test_bit(In_sync, &rdev->flags) || 1682 atomic_read(&rdev->nr_pending)) { 1683 err = -EBUSY; 1684 goto abort; 1685 } 1686 /* Only remove non-faulty devices if recovery 1687 * is not possible. 1688 */ 1689 if (!test_bit(Faulty, &rdev->flags) && 1690 mddev->recovery_disabled != conf->recovery_disabled && 1691 mddev->degraded < conf->raid_disks) { 1692 err = -EBUSY; 1693 goto abort; 1694 } 1695 p->rdev = NULL; 1696 synchronize_rcu(); 1697 if (atomic_read(&rdev->nr_pending)) { 1698 /* lost the race, try later */ 1699 err = -EBUSY; 1700 p->rdev = rdev; 1701 goto abort; 1702 } else if (conf->mirrors[conf->raid_disks + number].rdev) { 1703 /* We just removed a device that is being replaced. 1704 * Move down the replacement. We drain all IO before 1705 * doing this to avoid confusion. 1706 */ 1707 struct md_rdev *repl = 1708 conf->mirrors[conf->raid_disks + number].rdev; 1709 freeze_array(conf, 0); 1710 clear_bit(Replacement, &repl->flags); 1711 p->rdev = repl; 1712 conf->mirrors[conf->raid_disks + number].rdev = NULL; 1713 unfreeze_array(conf); 1714 clear_bit(WantReplacement, &rdev->flags); 1715 } else 1716 clear_bit(WantReplacement, &rdev->flags); 1717 err = md_integrity_register(mddev); 1718 } 1719 abort: 1720 1721 print_conf(conf); 1722 return err; 1723 } 1724 1725 static void end_sync_read(struct bio *bio, int error) 1726 { 1727 struct r1bio *r1_bio = bio->bi_private; 1728 1729 update_head_pos(r1_bio->read_disk, r1_bio); 1730 1731 /* 1732 * we have read a block, now it needs to be re-written, 1733 * or re-read if the read failed. 1734 * We don't do much here, just schedule handling by raid1d 1735 */ 1736 if (test_bit(BIO_UPTODATE, &bio->bi_flags)) 1737 set_bit(R1BIO_Uptodate, &r1_bio->state); 1738 1739 if (atomic_dec_and_test(&r1_bio->remaining)) 1740 reschedule_retry(r1_bio); 1741 } 1742 1743 static void end_sync_write(struct bio *bio, int error) 1744 { 1745 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 1746 struct r1bio *r1_bio = bio->bi_private; 1747 struct mddev *mddev = r1_bio->mddev; 1748 struct r1conf *conf = mddev->private; 1749 int mirror=0; 1750 sector_t first_bad; 1751 int bad_sectors; 1752 1753 mirror = find_bio_disk(r1_bio, bio); 1754 1755 if (!uptodate) { 1756 sector_t sync_blocks = 0; 1757 sector_t s = r1_bio->sector; 1758 long sectors_to_go = r1_bio->sectors; 1759 /* make sure these bits doesn't get cleared. */ 1760 do { 1761 bitmap_end_sync(mddev->bitmap, s, 1762 &sync_blocks, 1); 1763 s += sync_blocks; 1764 sectors_to_go -= sync_blocks; 1765 } while (sectors_to_go > 0); 1766 set_bit(WriteErrorSeen, 1767 &conf->mirrors[mirror].rdev->flags); 1768 if (!test_and_set_bit(WantReplacement, 1769 &conf->mirrors[mirror].rdev->flags)) 1770 set_bit(MD_RECOVERY_NEEDED, & 1771 mddev->recovery); 1772 set_bit(R1BIO_WriteError, &r1_bio->state); 1773 } else if (is_badblock(conf->mirrors[mirror].rdev, 1774 r1_bio->sector, 1775 r1_bio->sectors, 1776 &first_bad, &bad_sectors) && 1777 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev, 1778 r1_bio->sector, 1779 r1_bio->sectors, 1780 &first_bad, &bad_sectors) 1781 ) 1782 set_bit(R1BIO_MadeGood, &r1_bio->state); 1783 1784 if (atomic_dec_and_test(&r1_bio->remaining)) { 1785 int s = r1_bio->sectors; 1786 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 1787 test_bit(R1BIO_WriteError, &r1_bio->state)) 1788 reschedule_retry(r1_bio); 1789 else { 1790 put_buf(r1_bio); 1791 md_done_sync(mddev, s, uptodate); 1792 } 1793 } 1794 } 1795 1796 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector, 1797 int sectors, struct page *page, int rw) 1798 { 1799 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false)) 1800 /* success */ 1801 return 1; 1802 if (rw == WRITE) { 1803 set_bit(WriteErrorSeen, &rdev->flags); 1804 if (!test_and_set_bit(WantReplacement, 1805 &rdev->flags)) 1806 set_bit(MD_RECOVERY_NEEDED, & 1807 rdev->mddev->recovery); 1808 } 1809 /* need to record an error - either for the block or the device */ 1810 if (!rdev_set_badblocks(rdev, sector, sectors, 0)) 1811 md_error(rdev->mddev, rdev); 1812 return 0; 1813 } 1814 1815 static int fix_sync_read_error(struct r1bio *r1_bio) 1816 { 1817 /* Try some synchronous reads of other devices to get 1818 * good data, much like with normal read errors. Only 1819 * read into the pages we already have so we don't 1820 * need to re-issue the read request. 1821 * We don't need to freeze the array, because being in an 1822 * active sync request, there is no normal IO, and 1823 * no overlapping syncs. 1824 * We don't need to check is_badblock() again as we 1825 * made sure that anything with a bad block in range 1826 * will have bi_end_io clear. 1827 */ 1828 struct mddev *mddev = r1_bio->mddev; 1829 struct r1conf *conf = mddev->private; 1830 struct bio *bio = r1_bio->bios[r1_bio->read_disk]; 1831 sector_t sect = r1_bio->sector; 1832 int sectors = r1_bio->sectors; 1833 int idx = 0; 1834 1835 while(sectors) { 1836 int s = sectors; 1837 int d = r1_bio->read_disk; 1838 int success = 0; 1839 struct md_rdev *rdev; 1840 int start; 1841 1842 if (s > (PAGE_SIZE>>9)) 1843 s = PAGE_SIZE >> 9; 1844 do { 1845 if (r1_bio->bios[d]->bi_end_io == end_sync_read) { 1846 /* No rcu protection needed here devices 1847 * can only be removed when no resync is 1848 * active, and resync is currently active 1849 */ 1850 rdev = conf->mirrors[d].rdev; 1851 if (sync_page_io(rdev, sect, s<<9, 1852 bio->bi_io_vec[idx].bv_page, 1853 READ, false)) { 1854 success = 1; 1855 break; 1856 } 1857 } 1858 d++; 1859 if (d == conf->raid_disks * 2) 1860 d = 0; 1861 } while (!success && d != r1_bio->read_disk); 1862 1863 if (!success) { 1864 char b[BDEVNAME_SIZE]; 1865 int abort = 0; 1866 /* Cannot read from anywhere, this block is lost. 1867 * Record a bad block on each device. If that doesn't 1868 * work just disable and interrupt the recovery. 1869 * Don't fail devices as that won't really help. 1870 */ 1871 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error" 1872 " for block %llu\n", 1873 mdname(mddev), 1874 bdevname(bio->bi_bdev, b), 1875 (unsigned long long)r1_bio->sector); 1876 for (d = 0; d < conf->raid_disks * 2; d++) { 1877 rdev = conf->mirrors[d].rdev; 1878 if (!rdev || test_bit(Faulty, &rdev->flags)) 1879 continue; 1880 if (!rdev_set_badblocks(rdev, sect, s, 0)) 1881 abort = 1; 1882 } 1883 if (abort) { 1884 conf->recovery_disabled = 1885 mddev->recovery_disabled; 1886 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1887 md_done_sync(mddev, r1_bio->sectors, 0); 1888 put_buf(r1_bio); 1889 return 0; 1890 } 1891 /* Try next page */ 1892 sectors -= s; 1893 sect += s; 1894 idx++; 1895 continue; 1896 } 1897 1898 start = d; 1899 /* write it back and re-read */ 1900 while (d != r1_bio->read_disk) { 1901 if (d == 0) 1902 d = conf->raid_disks * 2; 1903 d--; 1904 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1905 continue; 1906 rdev = conf->mirrors[d].rdev; 1907 if (r1_sync_page_io(rdev, sect, s, 1908 bio->bi_io_vec[idx].bv_page, 1909 WRITE) == 0) { 1910 r1_bio->bios[d]->bi_end_io = NULL; 1911 rdev_dec_pending(rdev, mddev); 1912 } 1913 } 1914 d = start; 1915 while (d != r1_bio->read_disk) { 1916 if (d == 0) 1917 d = conf->raid_disks * 2; 1918 d--; 1919 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1920 continue; 1921 rdev = conf->mirrors[d].rdev; 1922 if (r1_sync_page_io(rdev, sect, s, 1923 bio->bi_io_vec[idx].bv_page, 1924 READ) != 0) 1925 atomic_add(s, &rdev->corrected_errors); 1926 } 1927 sectors -= s; 1928 sect += s; 1929 idx ++; 1930 } 1931 set_bit(R1BIO_Uptodate, &r1_bio->state); 1932 set_bit(BIO_UPTODATE, &bio->bi_flags); 1933 return 1; 1934 } 1935 1936 static void process_checks(struct r1bio *r1_bio) 1937 { 1938 /* We have read all readable devices. If we haven't 1939 * got the block, then there is no hope left. 1940 * If we have, then we want to do a comparison 1941 * and skip the write if everything is the same. 1942 * If any blocks failed to read, then we need to 1943 * attempt an over-write 1944 */ 1945 struct mddev *mddev = r1_bio->mddev; 1946 struct r1conf *conf = mddev->private; 1947 int primary; 1948 int i; 1949 int vcnt; 1950 1951 /* Fix variable parts of all bios */ 1952 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9); 1953 for (i = 0; i < conf->raid_disks * 2; i++) { 1954 int j; 1955 int size; 1956 int uptodate; 1957 struct bio *b = r1_bio->bios[i]; 1958 if (b->bi_end_io != end_sync_read) 1959 continue; 1960 /* fixup the bio for reuse, but preserve BIO_UPTODATE */ 1961 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags); 1962 bio_reset(b); 1963 if (!uptodate) 1964 clear_bit(BIO_UPTODATE, &b->bi_flags); 1965 b->bi_vcnt = vcnt; 1966 b->bi_iter.bi_size = r1_bio->sectors << 9; 1967 b->bi_iter.bi_sector = r1_bio->sector + 1968 conf->mirrors[i].rdev->data_offset; 1969 b->bi_bdev = conf->mirrors[i].rdev->bdev; 1970 b->bi_end_io = end_sync_read; 1971 b->bi_private = r1_bio; 1972 1973 size = b->bi_iter.bi_size; 1974 for (j = 0; j < vcnt ; j++) { 1975 struct bio_vec *bi; 1976 bi = &b->bi_io_vec[j]; 1977 bi->bv_offset = 0; 1978 if (size > PAGE_SIZE) 1979 bi->bv_len = PAGE_SIZE; 1980 else 1981 bi->bv_len = size; 1982 size -= PAGE_SIZE; 1983 } 1984 } 1985 for (primary = 0; primary < conf->raid_disks * 2; primary++) 1986 if (r1_bio->bios[primary]->bi_end_io == end_sync_read && 1987 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) { 1988 r1_bio->bios[primary]->bi_end_io = NULL; 1989 rdev_dec_pending(conf->mirrors[primary].rdev, mddev); 1990 break; 1991 } 1992 r1_bio->read_disk = primary; 1993 for (i = 0; i < conf->raid_disks * 2; i++) { 1994 int j; 1995 struct bio *pbio = r1_bio->bios[primary]; 1996 struct bio *sbio = r1_bio->bios[i]; 1997 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags); 1998 1999 if (sbio->bi_end_io != end_sync_read) 2000 continue; 2001 /* Now we can 'fixup' the BIO_UPTODATE flag */ 2002 set_bit(BIO_UPTODATE, &sbio->bi_flags); 2003 2004 if (uptodate) { 2005 for (j = vcnt; j-- ; ) { 2006 struct page *p, *s; 2007 p = pbio->bi_io_vec[j].bv_page; 2008 s = sbio->bi_io_vec[j].bv_page; 2009 if (memcmp(page_address(p), 2010 page_address(s), 2011 sbio->bi_io_vec[j].bv_len)) 2012 break; 2013 } 2014 } else 2015 j = 0; 2016 if (j >= 0) 2017 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches); 2018 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery) 2019 && uptodate)) { 2020 /* No need to write to this device. */ 2021 sbio->bi_end_io = NULL; 2022 rdev_dec_pending(conf->mirrors[i].rdev, mddev); 2023 continue; 2024 } 2025 2026 bio_copy_data(sbio, pbio); 2027 } 2028 } 2029 2030 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio) 2031 { 2032 struct r1conf *conf = mddev->private; 2033 int i; 2034 int disks = conf->raid_disks * 2; 2035 struct bio *bio, *wbio; 2036 2037 bio = r1_bio->bios[r1_bio->read_disk]; 2038 2039 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 2040 /* ouch - failed to read all of that. */ 2041 if (!fix_sync_read_error(r1_bio)) 2042 return; 2043 2044 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2045 process_checks(r1_bio); 2046 2047 /* 2048 * schedule writes 2049 */ 2050 atomic_set(&r1_bio->remaining, 1); 2051 for (i = 0; i < disks ; i++) { 2052 wbio = r1_bio->bios[i]; 2053 if (wbio->bi_end_io == NULL || 2054 (wbio->bi_end_io == end_sync_read && 2055 (i == r1_bio->read_disk || 2056 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)))) 2057 continue; 2058 2059 wbio->bi_rw = WRITE; 2060 wbio->bi_end_io = end_sync_write; 2061 atomic_inc(&r1_bio->remaining); 2062 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio)); 2063 2064 generic_make_request(wbio); 2065 } 2066 2067 if (atomic_dec_and_test(&r1_bio->remaining)) { 2068 /* if we're here, all write(s) have completed, so clean up */ 2069 int s = r1_bio->sectors; 2070 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2071 test_bit(R1BIO_WriteError, &r1_bio->state)) 2072 reschedule_retry(r1_bio); 2073 else { 2074 put_buf(r1_bio); 2075 md_done_sync(mddev, s, 1); 2076 } 2077 } 2078 } 2079 2080 /* 2081 * This is a kernel thread which: 2082 * 2083 * 1. Retries failed read operations on working mirrors. 2084 * 2. Updates the raid superblock when problems encounter. 2085 * 3. Performs writes following reads for array synchronising. 2086 */ 2087 2088 static void fix_read_error(struct r1conf *conf, int read_disk, 2089 sector_t sect, int sectors) 2090 { 2091 struct mddev *mddev = conf->mddev; 2092 while(sectors) { 2093 int s = sectors; 2094 int d = read_disk; 2095 int success = 0; 2096 int start; 2097 struct md_rdev *rdev; 2098 2099 if (s > (PAGE_SIZE>>9)) 2100 s = PAGE_SIZE >> 9; 2101 2102 do { 2103 /* Note: no rcu protection needed here 2104 * as this is synchronous in the raid1d thread 2105 * which is the thread that might remove 2106 * a device. If raid1d ever becomes multi-threaded.... 2107 */ 2108 sector_t first_bad; 2109 int bad_sectors; 2110 2111 rdev = conf->mirrors[d].rdev; 2112 if (rdev && 2113 (test_bit(In_sync, &rdev->flags) || 2114 (!test_bit(Faulty, &rdev->flags) && 2115 rdev->recovery_offset >= sect + s)) && 2116 is_badblock(rdev, sect, s, 2117 &first_bad, &bad_sectors) == 0 && 2118 sync_page_io(rdev, sect, s<<9, 2119 conf->tmppage, READ, false)) 2120 success = 1; 2121 else { 2122 d++; 2123 if (d == conf->raid_disks * 2) 2124 d = 0; 2125 } 2126 } while (!success && d != read_disk); 2127 2128 if (!success) { 2129 /* Cannot read from anywhere - mark it bad */ 2130 struct md_rdev *rdev = conf->mirrors[read_disk].rdev; 2131 if (!rdev_set_badblocks(rdev, sect, s, 0)) 2132 md_error(mddev, rdev); 2133 break; 2134 } 2135 /* write it back and re-read */ 2136 start = d; 2137 while (d != read_disk) { 2138 if (d==0) 2139 d = conf->raid_disks * 2; 2140 d--; 2141 rdev = conf->mirrors[d].rdev; 2142 if (rdev && 2143 !test_bit(Faulty, &rdev->flags)) 2144 r1_sync_page_io(rdev, sect, s, 2145 conf->tmppage, WRITE); 2146 } 2147 d = start; 2148 while (d != read_disk) { 2149 char b[BDEVNAME_SIZE]; 2150 if (d==0) 2151 d = conf->raid_disks * 2; 2152 d--; 2153 rdev = conf->mirrors[d].rdev; 2154 if (rdev && 2155 !test_bit(Faulty, &rdev->flags)) { 2156 if (r1_sync_page_io(rdev, sect, s, 2157 conf->tmppage, READ)) { 2158 atomic_add(s, &rdev->corrected_errors); 2159 printk(KERN_INFO 2160 "md/raid1:%s: read error corrected " 2161 "(%d sectors at %llu on %s)\n", 2162 mdname(mddev), s, 2163 (unsigned long long)(sect + 2164 rdev->data_offset), 2165 bdevname(rdev->bdev, b)); 2166 } 2167 } 2168 } 2169 sectors -= s; 2170 sect += s; 2171 } 2172 } 2173 2174 static int narrow_write_error(struct r1bio *r1_bio, int i) 2175 { 2176 struct mddev *mddev = r1_bio->mddev; 2177 struct r1conf *conf = mddev->private; 2178 struct md_rdev *rdev = conf->mirrors[i].rdev; 2179 2180 /* bio has the data to be written to device 'i' where 2181 * we just recently had a write error. 2182 * We repeatedly clone the bio and trim down to one block, 2183 * then try the write. Where the write fails we record 2184 * a bad block. 2185 * It is conceivable that the bio doesn't exactly align with 2186 * blocks. We must handle this somehow. 2187 * 2188 * We currently own a reference on the rdev. 2189 */ 2190 2191 int block_sectors; 2192 sector_t sector; 2193 int sectors; 2194 int sect_to_write = r1_bio->sectors; 2195 int ok = 1; 2196 2197 if (rdev->badblocks.shift < 0) 2198 return 0; 2199 2200 block_sectors = 1 << rdev->badblocks.shift; 2201 sector = r1_bio->sector; 2202 sectors = ((sector + block_sectors) 2203 & ~(sector_t)(block_sectors - 1)) 2204 - sector; 2205 2206 while (sect_to_write) { 2207 struct bio *wbio; 2208 if (sectors > sect_to_write) 2209 sectors = sect_to_write; 2210 /* Write at 'sector' for 'sectors'*/ 2211 2212 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 2213 unsigned vcnt = r1_bio->behind_page_count; 2214 struct bio_vec *vec = r1_bio->behind_bvecs; 2215 2216 while (!vec->bv_page) { 2217 vec++; 2218 vcnt--; 2219 } 2220 2221 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev); 2222 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec)); 2223 2224 wbio->bi_vcnt = vcnt; 2225 } else { 2226 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2227 } 2228 2229 wbio->bi_rw = WRITE; 2230 wbio->bi_iter.bi_sector = r1_bio->sector; 2231 wbio->bi_iter.bi_size = r1_bio->sectors << 9; 2232 2233 bio_trim(wbio, sector - r1_bio->sector, sectors); 2234 wbio->bi_iter.bi_sector += rdev->data_offset; 2235 wbio->bi_bdev = rdev->bdev; 2236 if (submit_bio_wait(WRITE, wbio) == 0) 2237 /* failure! */ 2238 ok = rdev_set_badblocks(rdev, sector, 2239 sectors, 0) 2240 && ok; 2241 2242 bio_put(wbio); 2243 sect_to_write -= sectors; 2244 sector += sectors; 2245 sectors = block_sectors; 2246 } 2247 return ok; 2248 } 2249 2250 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2251 { 2252 int m; 2253 int s = r1_bio->sectors; 2254 for (m = 0; m < conf->raid_disks * 2 ; m++) { 2255 struct md_rdev *rdev = conf->mirrors[m].rdev; 2256 struct bio *bio = r1_bio->bios[m]; 2257 if (bio->bi_end_io == NULL) 2258 continue; 2259 if (test_bit(BIO_UPTODATE, &bio->bi_flags) && 2260 test_bit(R1BIO_MadeGood, &r1_bio->state)) { 2261 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0); 2262 } 2263 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) && 2264 test_bit(R1BIO_WriteError, &r1_bio->state)) { 2265 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0)) 2266 md_error(conf->mddev, rdev); 2267 } 2268 } 2269 put_buf(r1_bio); 2270 md_done_sync(conf->mddev, s, 1); 2271 } 2272 2273 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2274 { 2275 int m; 2276 for (m = 0; m < conf->raid_disks * 2 ; m++) 2277 if (r1_bio->bios[m] == IO_MADE_GOOD) { 2278 struct md_rdev *rdev = conf->mirrors[m].rdev; 2279 rdev_clear_badblocks(rdev, 2280 r1_bio->sector, 2281 r1_bio->sectors, 0); 2282 rdev_dec_pending(rdev, conf->mddev); 2283 } else if (r1_bio->bios[m] != NULL) { 2284 /* This drive got a write error. We need to 2285 * narrow down and record precise write 2286 * errors. 2287 */ 2288 if (!narrow_write_error(r1_bio, m)) { 2289 md_error(conf->mddev, 2290 conf->mirrors[m].rdev); 2291 /* an I/O failed, we can't clear the bitmap */ 2292 set_bit(R1BIO_Degraded, &r1_bio->state); 2293 } 2294 rdev_dec_pending(conf->mirrors[m].rdev, 2295 conf->mddev); 2296 } 2297 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 2298 close_write(r1_bio); 2299 raid_end_bio_io(r1_bio); 2300 } 2301 2302 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio) 2303 { 2304 int disk; 2305 int max_sectors; 2306 struct mddev *mddev = conf->mddev; 2307 struct bio *bio; 2308 char b[BDEVNAME_SIZE]; 2309 struct md_rdev *rdev; 2310 2311 clear_bit(R1BIO_ReadError, &r1_bio->state); 2312 /* we got a read error. Maybe the drive is bad. Maybe just 2313 * the block and we can fix it. 2314 * We freeze all other IO, and try reading the block from 2315 * other devices. When we find one, we re-write 2316 * and check it that fixes the read error. 2317 * This is all done synchronously while the array is 2318 * frozen 2319 */ 2320 if (mddev->ro == 0) { 2321 freeze_array(conf, 1); 2322 fix_read_error(conf, r1_bio->read_disk, 2323 r1_bio->sector, r1_bio->sectors); 2324 unfreeze_array(conf); 2325 } else 2326 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev); 2327 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev); 2328 2329 bio = r1_bio->bios[r1_bio->read_disk]; 2330 bdevname(bio->bi_bdev, b); 2331 read_more: 2332 disk = read_balance(conf, r1_bio, &max_sectors); 2333 if (disk == -1) { 2334 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O" 2335 " read error for block %llu\n", 2336 mdname(mddev), b, (unsigned long long)r1_bio->sector); 2337 raid_end_bio_io(r1_bio); 2338 } else { 2339 const unsigned long do_sync 2340 = r1_bio->master_bio->bi_rw & REQ_SYNC; 2341 if (bio) { 2342 r1_bio->bios[r1_bio->read_disk] = 2343 mddev->ro ? IO_BLOCKED : NULL; 2344 bio_put(bio); 2345 } 2346 r1_bio->read_disk = disk; 2347 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2348 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector, 2349 max_sectors); 2350 r1_bio->bios[r1_bio->read_disk] = bio; 2351 rdev = conf->mirrors[disk].rdev; 2352 printk_ratelimited(KERN_ERR 2353 "md/raid1:%s: redirecting sector %llu" 2354 " to other mirror: %s\n", 2355 mdname(mddev), 2356 (unsigned long long)r1_bio->sector, 2357 bdevname(rdev->bdev, b)); 2358 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset; 2359 bio->bi_bdev = rdev->bdev; 2360 bio->bi_end_io = raid1_end_read_request; 2361 bio->bi_rw = READ | do_sync; 2362 bio->bi_private = r1_bio; 2363 if (max_sectors < r1_bio->sectors) { 2364 /* Drat - have to split this up more */ 2365 struct bio *mbio = r1_bio->master_bio; 2366 int sectors_handled = (r1_bio->sector + max_sectors 2367 - mbio->bi_iter.bi_sector); 2368 r1_bio->sectors = max_sectors; 2369 spin_lock_irq(&conf->device_lock); 2370 if (mbio->bi_phys_segments == 0) 2371 mbio->bi_phys_segments = 2; 2372 else 2373 mbio->bi_phys_segments++; 2374 spin_unlock_irq(&conf->device_lock); 2375 generic_make_request(bio); 2376 bio = NULL; 2377 2378 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 2379 2380 r1_bio->master_bio = mbio; 2381 r1_bio->sectors = bio_sectors(mbio) - sectors_handled; 2382 r1_bio->state = 0; 2383 set_bit(R1BIO_ReadError, &r1_bio->state); 2384 r1_bio->mddev = mddev; 2385 r1_bio->sector = mbio->bi_iter.bi_sector + 2386 sectors_handled; 2387 2388 goto read_more; 2389 } else 2390 generic_make_request(bio); 2391 } 2392 } 2393 2394 static void raid1d(struct md_thread *thread) 2395 { 2396 struct mddev *mddev = thread->mddev; 2397 struct r1bio *r1_bio; 2398 unsigned long flags; 2399 struct r1conf *conf = mddev->private; 2400 struct list_head *head = &conf->retry_list; 2401 struct blk_plug plug; 2402 2403 md_check_recovery(mddev); 2404 2405 blk_start_plug(&plug); 2406 for (;;) { 2407 2408 flush_pending_writes(conf); 2409 2410 spin_lock_irqsave(&conf->device_lock, flags); 2411 if (list_empty(head)) { 2412 spin_unlock_irqrestore(&conf->device_lock, flags); 2413 break; 2414 } 2415 r1_bio = list_entry(head->prev, struct r1bio, retry_list); 2416 list_del(head->prev); 2417 conf->nr_queued--; 2418 spin_unlock_irqrestore(&conf->device_lock, flags); 2419 2420 mddev = r1_bio->mddev; 2421 conf = mddev->private; 2422 if (test_bit(R1BIO_IsSync, &r1_bio->state)) { 2423 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2424 test_bit(R1BIO_WriteError, &r1_bio->state)) 2425 handle_sync_write_finished(conf, r1_bio); 2426 else 2427 sync_request_write(mddev, r1_bio); 2428 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2429 test_bit(R1BIO_WriteError, &r1_bio->state)) 2430 handle_write_finished(conf, r1_bio); 2431 else if (test_bit(R1BIO_ReadError, &r1_bio->state)) 2432 handle_read_error(conf, r1_bio); 2433 else 2434 /* just a partial read to be scheduled from separate 2435 * context 2436 */ 2437 generic_make_request(r1_bio->bios[r1_bio->read_disk]); 2438 2439 cond_resched(); 2440 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) 2441 md_check_recovery(mddev); 2442 } 2443 blk_finish_plug(&plug); 2444 } 2445 2446 static int init_resync(struct r1conf *conf) 2447 { 2448 int buffs; 2449 2450 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; 2451 BUG_ON(conf->r1buf_pool); 2452 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, 2453 conf->poolinfo); 2454 if (!conf->r1buf_pool) 2455 return -ENOMEM; 2456 conf->next_resync = 0; 2457 return 0; 2458 } 2459 2460 /* 2461 * perform a "sync" on one "block" 2462 * 2463 * We need to make sure that no normal I/O request - particularly write 2464 * requests - conflict with active sync requests. 2465 * 2466 * This is achieved by tracking pending requests and a 'barrier' concept 2467 * that can be installed to exclude normal IO requests. 2468 */ 2469 2470 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster) 2471 { 2472 struct r1conf *conf = mddev->private; 2473 struct r1bio *r1_bio; 2474 struct bio *bio; 2475 sector_t max_sector, nr_sectors; 2476 int disk = -1; 2477 int i; 2478 int wonly = -1; 2479 int write_targets = 0, read_targets = 0; 2480 sector_t sync_blocks; 2481 int still_degraded = 0; 2482 int good_sectors = RESYNC_SECTORS; 2483 int min_bad = 0; /* number of sectors that are bad in all devices */ 2484 2485 if (!conf->r1buf_pool) 2486 if (init_resync(conf)) 2487 return 0; 2488 2489 max_sector = mddev->dev_sectors; 2490 if (sector_nr >= max_sector) { 2491 /* If we aborted, we need to abort the 2492 * sync on the 'current' bitmap chunk (there will 2493 * only be one in raid1 resync. 2494 * We can find the current addess in mddev->curr_resync 2495 */ 2496 if (mddev->curr_resync < max_sector) /* aborted */ 2497 bitmap_end_sync(mddev->bitmap, mddev->curr_resync, 2498 &sync_blocks, 1); 2499 else /* completed sync */ 2500 conf->fullsync = 0; 2501 2502 bitmap_close_sync(mddev->bitmap); 2503 close_sync(conf); 2504 return 0; 2505 } 2506 2507 if (mddev->bitmap == NULL && 2508 mddev->recovery_cp == MaxSector && 2509 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && 2510 conf->fullsync == 0) { 2511 *skipped = 1; 2512 return max_sector - sector_nr; 2513 } 2514 /* before building a request, check if we can skip these blocks.. 2515 * This call the bitmap_start_sync doesn't actually record anything 2516 */ 2517 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) && 2518 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2519 /* We can skip this block, and probably several more */ 2520 *skipped = 1; 2521 return sync_blocks; 2522 } 2523 /* 2524 * If there is non-resync activity waiting for a turn, 2525 * and resync is going fast enough, 2526 * then let it though before starting on this new sync request. 2527 */ 2528 if (!go_faster && conf->nr_waiting) 2529 msleep_interruptible(1000); 2530 2531 bitmap_cond_end_sync(mddev->bitmap, sector_nr); 2532 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO); 2533 2534 raise_barrier(conf, sector_nr); 2535 2536 rcu_read_lock(); 2537 /* 2538 * If we get a correctably read error during resync or recovery, 2539 * we might want to read from a different device. So we 2540 * flag all drives that could conceivably be read from for READ, 2541 * and any others (which will be non-In_sync devices) for WRITE. 2542 * If a read fails, we try reading from something else for which READ 2543 * is OK. 2544 */ 2545 2546 r1_bio->mddev = mddev; 2547 r1_bio->sector = sector_nr; 2548 r1_bio->state = 0; 2549 set_bit(R1BIO_IsSync, &r1_bio->state); 2550 2551 for (i = 0; i < conf->raid_disks * 2; i++) { 2552 struct md_rdev *rdev; 2553 bio = r1_bio->bios[i]; 2554 bio_reset(bio); 2555 2556 rdev = rcu_dereference(conf->mirrors[i].rdev); 2557 if (rdev == NULL || 2558 test_bit(Faulty, &rdev->flags)) { 2559 if (i < conf->raid_disks) 2560 still_degraded = 1; 2561 } else if (!test_bit(In_sync, &rdev->flags)) { 2562 bio->bi_rw = WRITE; 2563 bio->bi_end_io = end_sync_write; 2564 write_targets ++; 2565 } else { 2566 /* may need to read from here */ 2567 sector_t first_bad = MaxSector; 2568 int bad_sectors; 2569 2570 if (is_badblock(rdev, sector_nr, good_sectors, 2571 &first_bad, &bad_sectors)) { 2572 if (first_bad > sector_nr) 2573 good_sectors = first_bad - sector_nr; 2574 else { 2575 bad_sectors -= (sector_nr - first_bad); 2576 if (min_bad == 0 || 2577 min_bad > bad_sectors) 2578 min_bad = bad_sectors; 2579 } 2580 } 2581 if (sector_nr < first_bad) { 2582 if (test_bit(WriteMostly, &rdev->flags)) { 2583 if (wonly < 0) 2584 wonly = i; 2585 } else { 2586 if (disk < 0) 2587 disk = i; 2588 } 2589 bio->bi_rw = READ; 2590 bio->bi_end_io = end_sync_read; 2591 read_targets++; 2592 } else if (!test_bit(WriteErrorSeen, &rdev->flags) && 2593 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && 2594 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) { 2595 /* 2596 * The device is suitable for reading (InSync), 2597 * but has bad block(s) here. Let's try to correct them, 2598 * if we are doing resync or repair. Otherwise, leave 2599 * this device alone for this sync request. 2600 */ 2601 bio->bi_rw = WRITE; 2602 bio->bi_end_io = end_sync_write; 2603 write_targets++; 2604 } 2605 } 2606 if (bio->bi_end_io) { 2607 atomic_inc(&rdev->nr_pending); 2608 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset; 2609 bio->bi_bdev = rdev->bdev; 2610 bio->bi_private = r1_bio; 2611 } 2612 } 2613 rcu_read_unlock(); 2614 if (disk < 0) 2615 disk = wonly; 2616 r1_bio->read_disk = disk; 2617 2618 if (read_targets == 0 && min_bad > 0) { 2619 /* These sectors are bad on all InSync devices, so we 2620 * need to mark them bad on all write targets 2621 */ 2622 int ok = 1; 2623 for (i = 0 ; i < conf->raid_disks * 2 ; i++) 2624 if (r1_bio->bios[i]->bi_end_io == end_sync_write) { 2625 struct md_rdev *rdev = conf->mirrors[i].rdev; 2626 ok = rdev_set_badblocks(rdev, sector_nr, 2627 min_bad, 0 2628 ) && ok; 2629 } 2630 set_bit(MD_CHANGE_DEVS, &mddev->flags); 2631 *skipped = 1; 2632 put_buf(r1_bio); 2633 2634 if (!ok) { 2635 /* Cannot record the badblocks, so need to 2636 * abort the resync. 2637 * If there are multiple read targets, could just 2638 * fail the really bad ones ??? 2639 */ 2640 conf->recovery_disabled = mddev->recovery_disabled; 2641 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 2642 return 0; 2643 } else 2644 return min_bad; 2645 2646 } 2647 if (min_bad > 0 && min_bad < good_sectors) { 2648 /* only resync enough to reach the next bad->good 2649 * transition */ 2650 good_sectors = min_bad; 2651 } 2652 2653 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0) 2654 /* extra read targets are also write targets */ 2655 write_targets += read_targets-1; 2656 2657 if (write_targets == 0 || read_targets == 0) { 2658 /* There is nowhere to write, so all non-sync 2659 * drives must be failed - so we are finished 2660 */ 2661 sector_t rv; 2662 if (min_bad > 0) 2663 max_sector = sector_nr + min_bad; 2664 rv = max_sector - sector_nr; 2665 *skipped = 1; 2666 put_buf(r1_bio); 2667 return rv; 2668 } 2669 2670 if (max_sector > mddev->resync_max) 2671 max_sector = mddev->resync_max; /* Don't do IO beyond here */ 2672 if (max_sector > sector_nr + good_sectors) 2673 max_sector = sector_nr + good_sectors; 2674 nr_sectors = 0; 2675 sync_blocks = 0; 2676 do { 2677 struct page *page; 2678 int len = PAGE_SIZE; 2679 if (sector_nr + (len>>9) > max_sector) 2680 len = (max_sector - sector_nr) << 9; 2681 if (len == 0) 2682 break; 2683 if (sync_blocks == 0) { 2684 if (!bitmap_start_sync(mddev->bitmap, sector_nr, 2685 &sync_blocks, still_degraded) && 2686 !conf->fullsync && 2687 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2688 break; 2689 BUG_ON(sync_blocks < (PAGE_SIZE>>9)); 2690 if ((len >> 9) > sync_blocks) 2691 len = sync_blocks<<9; 2692 } 2693 2694 for (i = 0 ; i < conf->raid_disks * 2; i++) { 2695 bio = r1_bio->bios[i]; 2696 if (bio->bi_end_io) { 2697 page = bio->bi_io_vec[bio->bi_vcnt].bv_page; 2698 if (bio_add_page(bio, page, len, 0) == 0) { 2699 /* stop here */ 2700 bio->bi_io_vec[bio->bi_vcnt].bv_page = page; 2701 while (i > 0) { 2702 i--; 2703 bio = r1_bio->bios[i]; 2704 if (bio->bi_end_io==NULL) 2705 continue; 2706 /* remove last page from this bio */ 2707 bio->bi_vcnt--; 2708 bio->bi_iter.bi_size -= len; 2709 __clear_bit(BIO_SEG_VALID, &bio->bi_flags); 2710 } 2711 goto bio_full; 2712 } 2713 } 2714 } 2715 nr_sectors += len>>9; 2716 sector_nr += len>>9; 2717 sync_blocks -= (len>>9); 2718 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES); 2719 bio_full: 2720 r1_bio->sectors = nr_sectors; 2721 2722 /* For a user-requested sync, we read all readable devices and do a 2723 * compare 2724 */ 2725 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2726 atomic_set(&r1_bio->remaining, read_targets); 2727 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) { 2728 bio = r1_bio->bios[i]; 2729 if (bio->bi_end_io == end_sync_read) { 2730 read_targets--; 2731 md_sync_acct(bio->bi_bdev, nr_sectors); 2732 generic_make_request(bio); 2733 } 2734 } 2735 } else { 2736 atomic_set(&r1_bio->remaining, 1); 2737 bio = r1_bio->bios[r1_bio->read_disk]; 2738 md_sync_acct(bio->bi_bdev, nr_sectors); 2739 generic_make_request(bio); 2740 2741 } 2742 return nr_sectors; 2743 } 2744 2745 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks) 2746 { 2747 if (sectors) 2748 return sectors; 2749 2750 return mddev->dev_sectors; 2751 } 2752 2753 static struct r1conf *setup_conf(struct mddev *mddev) 2754 { 2755 struct r1conf *conf; 2756 int i; 2757 struct raid1_info *disk; 2758 struct md_rdev *rdev; 2759 int err = -ENOMEM; 2760 2761 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL); 2762 if (!conf) 2763 goto abort; 2764 2765 conf->mirrors = kzalloc(sizeof(struct raid1_info) 2766 * mddev->raid_disks * 2, 2767 GFP_KERNEL); 2768 if (!conf->mirrors) 2769 goto abort; 2770 2771 conf->tmppage = alloc_page(GFP_KERNEL); 2772 if (!conf->tmppage) 2773 goto abort; 2774 2775 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL); 2776 if (!conf->poolinfo) 2777 goto abort; 2778 conf->poolinfo->raid_disks = mddev->raid_disks * 2; 2779 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 2780 r1bio_pool_free, 2781 conf->poolinfo); 2782 if (!conf->r1bio_pool) 2783 goto abort; 2784 2785 conf->poolinfo->mddev = mddev; 2786 2787 err = -EINVAL; 2788 spin_lock_init(&conf->device_lock); 2789 rdev_for_each(rdev, mddev) { 2790 struct request_queue *q; 2791 int disk_idx = rdev->raid_disk; 2792 if (disk_idx >= mddev->raid_disks 2793 || disk_idx < 0) 2794 continue; 2795 if (test_bit(Replacement, &rdev->flags)) 2796 disk = conf->mirrors + mddev->raid_disks + disk_idx; 2797 else 2798 disk = conf->mirrors + disk_idx; 2799 2800 if (disk->rdev) 2801 goto abort; 2802 disk->rdev = rdev; 2803 q = bdev_get_queue(rdev->bdev); 2804 if (q->merge_bvec_fn) 2805 mddev->merge_check_needed = 1; 2806 2807 disk->head_position = 0; 2808 disk->seq_start = MaxSector; 2809 } 2810 conf->raid_disks = mddev->raid_disks; 2811 conf->mddev = mddev; 2812 INIT_LIST_HEAD(&conf->retry_list); 2813 2814 spin_lock_init(&conf->resync_lock); 2815 init_waitqueue_head(&conf->wait_barrier); 2816 2817 bio_list_init(&conf->pending_bio_list); 2818 conf->pending_count = 0; 2819 conf->recovery_disabled = mddev->recovery_disabled - 1; 2820 2821 conf->start_next_window = MaxSector; 2822 conf->current_window_requests = conf->next_window_requests = 0; 2823 2824 err = -EIO; 2825 for (i = 0; i < conf->raid_disks * 2; i++) { 2826 2827 disk = conf->mirrors + i; 2828 2829 if (i < conf->raid_disks && 2830 disk[conf->raid_disks].rdev) { 2831 /* This slot has a replacement. */ 2832 if (!disk->rdev) { 2833 /* No original, just make the replacement 2834 * a recovering spare 2835 */ 2836 disk->rdev = 2837 disk[conf->raid_disks].rdev; 2838 disk[conf->raid_disks].rdev = NULL; 2839 } else if (!test_bit(In_sync, &disk->rdev->flags)) 2840 /* Original is not in_sync - bad */ 2841 goto abort; 2842 } 2843 2844 if (!disk->rdev || 2845 !test_bit(In_sync, &disk->rdev->flags)) { 2846 disk->head_position = 0; 2847 if (disk->rdev && 2848 (disk->rdev->saved_raid_disk < 0)) 2849 conf->fullsync = 1; 2850 } 2851 } 2852 2853 err = -ENOMEM; 2854 conf->thread = md_register_thread(raid1d, mddev, "raid1"); 2855 if (!conf->thread) { 2856 printk(KERN_ERR 2857 "md/raid1:%s: couldn't allocate thread\n", 2858 mdname(mddev)); 2859 goto abort; 2860 } 2861 2862 return conf; 2863 2864 abort: 2865 if (conf) { 2866 if (conf->r1bio_pool) 2867 mempool_destroy(conf->r1bio_pool); 2868 kfree(conf->mirrors); 2869 safe_put_page(conf->tmppage); 2870 kfree(conf->poolinfo); 2871 kfree(conf); 2872 } 2873 return ERR_PTR(err); 2874 } 2875 2876 static int stop(struct mddev *mddev); 2877 static int run(struct mddev *mddev) 2878 { 2879 struct r1conf *conf; 2880 int i; 2881 struct md_rdev *rdev; 2882 int ret; 2883 bool discard_supported = false; 2884 2885 if (mddev->level != 1) { 2886 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n", 2887 mdname(mddev), mddev->level); 2888 return -EIO; 2889 } 2890 if (mddev->reshape_position != MaxSector) { 2891 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n", 2892 mdname(mddev)); 2893 return -EIO; 2894 } 2895 /* 2896 * copy the already verified devices into our private RAID1 2897 * bookkeeping area. [whatever we allocate in run(), 2898 * should be freed in stop()] 2899 */ 2900 if (mddev->private == NULL) 2901 conf = setup_conf(mddev); 2902 else 2903 conf = mddev->private; 2904 2905 if (IS_ERR(conf)) 2906 return PTR_ERR(conf); 2907 2908 if (mddev->queue) 2909 blk_queue_max_write_same_sectors(mddev->queue, 0); 2910 2911 rdev_for_each(rdev, mddev) { 2912 if (!mddev->gendisk) 2913 continue; 2914 disk_stack_limits(mddev->gendisk, rdev->bdev, 2915 rdev->data_offset << 9); 2916 if (blk_queue_discard(bdev_get_queue(rdev->bdev))) 2917 discard_supported = true; 2918 } 2919 2920 mddev->degraded = 0; 2921 for (i=0; i < conf->raid_disks; i++) 2922 if (conf->mirrors[i].rdev == NULL || 2923 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) || 2924 test_bit(Faulty, &conf->mirrors[i].rdev->flags)) 2925 mddev->degraded++; 2926 2927 if (conf->raid_disks - mddev->degraded == 1) 2928 mddev->recovery_cp = MaxSector; 2929 2930 if (mddev->recovery_cp != MaxSector) 2931 printk(KERN_NOTICE "md/raid1:%s: not clean" 2932 " -- starting background reconstruction\n", 2933 mdname(mddev)); 2934 printk(KERN_INFO 2935 "md/raid1:%s: active with %d out of %d mirrors\n", 2936 mdname(mddev), mddev->raid_disks - mddev->degraded, 2937 mddev->raid_disks); 2938 2939 /* 2940 * Ok, everything is just fine now 2941 */ 2942 mddev->thread = conf->thread; 2943 conf->thread = NULL; 2944 mddev->private = conf; 2945 2946 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0)); 2947 2948 if (mddev->queue) { 2949 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec); 2950 2951 if (discard_supported) 2952 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, 2953 mddev->queue); 2954 else 2955 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, 2956 mddev->queue); 2957 } 2958 2959 ret = md_integrity_register(mddev); 2960 if (ret) 2961 stop(mddev); 2962 return ret; 2963 } 2964 2965 static int stop(struct mddev *mddev) 2966 { 2967 struct r1conf *conf = mddev->private; 2968 struct bitmap *bitmap = mddev->bitmap; 2969 2970 /* wait for behind writes to complete */ 2971 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) { 2972 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n", 2973 mdname(mddev)); 2974 /* need to kick something here to make sure I/O goes? */ 2975 wait_event(bitmap->behind_wait, 2976 atomic_read(&bitmap->behind_writes) == 0); 2977 } 2978 2979 freeze_array(conf, 0); 2980 unfreeze_array(conf); 2981 2982 md_unregister_thread(&mddev->thread); 2983 if (conf->r1bio_pool) 2984 mempool_destroy(conf->r1bio_pool); 2985 kfree(conf->mirrors); 2986 safe_put_page(conf->tmppage); 2987 kfree(conf->poolinfo); 2988 kfree(conf); 2989 mddev->private = NULL; 2990 return 0; 2991 } 2992 2993 static int raid1_resize(struct mddev *mddev, sector_t sectors) 2994 { 2995 /* no resync is happening, and there is enough space 2996 * on all devices, so we can resize. 2997 * We need to make sure resync covers any new space. 2998 * If the array is shrinking we should possibly wait until 2999 * any io in the removed space completes, but it hardly seems 3000 * worth it. 3001 */ 3002 sector_t newsize = raid1_size(mddev, sectors, 0); 3003 if (mddev->external_size && 3004 mddev->array_sectors > newsize) 3005 return -EINVAL; 3006 if (mddev->bitmap) { 3007 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0); 3008 if (ret) 3009 return ret; 3010 } 3011 md_set_array_sectors(mddev, newsize); 3012 set_capacity(mddev->gendisk, mddev->array_sectors); 3013 revalidate_disk(mddev->gendisk); 3014 if (sectors > mddev->dev_sectors && 3015 mddev->recovery_cp > mddev->dev_sectors) { 3016 mddev->recovery_cp = mddev->dev_sectors; 3017 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3018 } 3019 mddev->dev_sectors = sectors; 3020 mddev->resync_max_sectors = sectors; 3021 return 0; 3022 } 3023 3024 static int raid1_reshape(struct mddev *mddev) 3025 { 3026 /* We need to: 3027 * 1/ resize the r1bio_pool 3028 * 2/ resize conf->mirrors 3029 * 3030 * We allocate a new r1bio_pool if we can. 3031 * Then raise a device barrier and wait until all IO stops. 3032 * Then resize conf->mirrors and swap in the new r1bio pool. 3033 * 3034 * At the same time, we "pack" the devices so that all the missing 3035 * devices have the higher raid_disk numbers. 3036 */ 3037 mempool_t *newpool, *oldpool; 3038 struct pool_info *newpoolinfo; 3039 struct raid1_info *newmirrors; 3040 struct r1conf *conf = mddev->private; 3041 int cnt, raid_disks; 3042 unsigned long flags; 3043 int d, d2, err; 3044 3045 /* Cannot change chunk_size, layout, or level */ 3046 if (mddev->chunk_sectors != mddev->new_chunk_sectors || 3047 mddev->layout != mddev->new_layout || 3048 mddev->level != mddev->new_level) { 3049 mddev->new_chunk_sectors = mddev->chunk_sectors; 3050 mddev->new_layout = mddev->layout; 3051 mddev->new_level = mddev->level; 3052 return -EINVAL; 3053 } 3054 3055 err = md_allow_write(mddev); 3056 if (err) 3057 return err; 3058 3059 raid_disks = mddev->raid_disks + mddev->delta_disks; 3060 3061 if (raid_disks < conf->raid_disks) { 3062 cnt=0; 3063 for (d= 0; d < conf->raid_disks; d++) 3064 if (conf->mirrors[d].rdev) 3065 cnt++; 3066 if (cnt > raid_disks) 3067 return -EBUSY; 3068 } 3069 3070 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL); 3071 if (!newpoolinfo) 3072 return -ENOMEM; 3073 newpoolinfo->mddev = mddev; 3074 newpoolinfo->raid_disks = raid_disks * 2; 3075 3076 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 3077 r1bio_pool_free, newpoolinfo); 3078 if (!newpool) { 3079 kfree(newpoolinfo); 3080 return -ENOMEM; 3081 } 3082 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2, 3083 GFP_KERNEL); 3084 if (!newmirrors) { 3085 kfree(newpoolinfo); 3086 mempool_destroy(newpool); 3087 return -ENOMEM; 3088 } 3089 3090 freeze_array(conf, 0); 3091 3092 /* ok, everything is stopped */ 3093 oldpool = conf->r1bio_pool; 3094 conf->r1bio_pool = newpool; 3095 3096 for (d = d2 = 0; d < conf->raid_disks; d++) { 3097 struct md_rdev *rdev = conf->mirrors[d].rdev; 3098 if (rdev && rdev->raid_disk != d2) { 3099 sysfs_unlink_rdev(mddev, rdev); 3100 rdev->raid_disk = d2; 3101 sysfs_unlink_rdev(mddev, rdev); 3102 if (sysfs_link_rdev(mddev, rdev)) 3103 printk(KERN_WARNING 3104 "md/raid1:%s: cannot register rd%d\n", 3105 mdname(mddev), rdev->raid_disk); 3106 } 3107 if (rdev) 3108 newmirrors[d2++].rdev = rdev; 3109 } 3110 kfree(conf->mirrors); 3111 conf->mirrors = newmirrors; 3112 kfree(conf->poolinfo); 3113 conf->poolinfo = newpoolinfo; 3114 3115 spin_lock_irqsave(&conf->device_lock, flags); 3116 mddev->degraded += (raid_disks - conf->raid_disks); 3117 spin_unlock_irqrestore(&conf->device_lock, flags); 3118 conf->raid_disks = mddev->raid_disks = raid_disks; 3119 mddev->delta_disks = 0; 3120 3121 unfreeze_array(conf); 3122 3123 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3124 md_wakeup_thread(mddev->thread); 3125 3126 mempool_destroy(oldpool); 3127 return 0; 3128 } 3129 3130 static void raid1_quiesce(struct mddev *mddev, int state) 3131 { 3132 struct r1conf *conf = mddev->private; 3133 3134 switch(state) { 3135 case 2: /* wake for suspend */ 3136 wake_up(&conf->wait_barrier); 3137 break; 3138 case 1: 3139 freeze_array(conf, 0); 3140 break; 3141 case 0: 3142 unfreeze_array(conf); 3143 break; 3144 } 3145 } 3146 3147 static void *raid1_takeover(struct mddev *mddev) 3148 { 3149 /* raid1 can take over: 3150 * raid5 with 2 devices, any layout or chunk size 3151 */ 3152 if (mddev->level == 5 && mddev->raid_disks == 2) { 3153 struct r1conf *conf; 3154 mddev->new_level = 1; 3155 mddev->new_layout = 0; 3156 mddev->new_chunk_sectors = 0; 3157 conf = setup_conf(mddev); 3158 if (!IS_ERR(conf)) 3159 /* Array must appear to be quiesced */ 3160 conf->array_frozen = 1; 3161 return conf; 3162 } 3163 return ERR_PTR(-EINVAL); 3164 } 3165 3166 static struct md_personality raid1_personality = 3167 { 3168 .name = "raid1", 3169 .level = 1, 3170 .owner = THIS_MODULE, 3171 .make_request = make_request, 3172 .run = run, 3173 .stop = stop, 3174 .status = status, 3175 .error_handler = error, 3176 .hot_add_disk = raid1_add_disk, 3177 .hot_remove_disk= raid1_remove_disk, 3178 .spare_active = raid1_spare_active, 3179 .sync_request = sync_request, 3180 .resize = raid1_resize, 3181 .size = raid1_size, 3182 .check_reshape = raid1_reshape, 3183 .quiesce = raid1_quiesce, 3184 .takeover = raid1_takeover, 3185 .congested = raid1_congested, 3186 }; 3187 3188 static int __init raid_init(void) 3189 { 3190 return register_md_personality(&raid1_personality); 3191 } 3192 3193 static void raid_exit(void) 3194 { 3195 unregister_md_personality(&raid1_personality); 3196 } 3197 3198 module_init(raid_init); 3199 module_exit(raid_exit); 3200 MODULE_LICENSE("GPL"); 3201 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD"); 3202 MODULE_ALIAS("md-personality-3"); /* RAID1 */ 3203 MODULE_ALIAS("md-raid1"); 3204 MODULE_ALIAS("md-level-1"); 3205 3206 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR); 3207