1 /* 2 * raid1.c : Multiple Devices driver for Linux 3 * 4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat 5 * 6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman 7 * 8 * RAID-1 management functions. 9 * 10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000 11 * 12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk> 13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au> 14 * 15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support 16 * bitmapped intelligence in resync: 17 * 18 * - bitmap marked during normal i/o 19 * - bitmap used to skip nondirty blocks during sync 20 * 21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology: 22 * - persistent bitmap code 23 * 24 * This program is free software; you can redistribute it and/or modify 25 * it under the terms of the GNU General Public License as published by 26 * the Free Software Foundation; either version 2, or (at your option) 27 * any later version. 28 * 29 * You should have received a copy of the GNU General Public License 30 * (for example /usr/src/linux/COPYING); if not, write to the Free 31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. 32 */ 33 34 #include <linux/slab.h> 35 #include <linux/delay.h> 36 #include <linux/blkdev.h> 37 #include <linux/module.h> 38 #include <linux/seq_file.h> 39 #include <linux/ratelimit.h> 40 #include "md.h" 41 #include "raid1.h" 42 #include "bitmap.h" 43 44 /* 45 * Number of guaranteed r1bios in case of extreme VM load: 46 */ 47 #define NR_RAID1_BIOS 256 48 49 /* when we get a read error on a read-only array, we redirect to another 50 * device without failing the first device, or trying to over-write to 51 * correct the read error. To keep track of bad blocks on a per-bio 52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer 53 */ 54 #define IO_BLOCKED ((struct bio *)1) 55 /* When we successfully write to a known bad-block, we need to remove the 56 * bad-block marking which must be done from process context. So we record 57 * the success by setting devs[n].bio to IO_MADE_GOOD 58 */ 59 #define IO_MADE_GOOD ((struct bio *)2) 60 61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) 62 63 /* When there are this many requests queue to be written by 64 * the raid1 thread, we become 'congested' to provide back-pressure 65 * for writeback. 66 */ 67 static int max_queued_requests = 1024; 68 69 static void allow_barrier(struct r1conf *conf, sector_t start_next_window, 70 sector_t bi_sector); 71 static void lower_barrier(struct r1conf *conf); 72 73 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data) 74 { 75 struct pool_info *pi = data; 76 int size = offsetof(struct r1bio, bios[pi->raid_disks]); 77 78 /* allocate a r1bio with room for raid_disks entries in the bios array */ 79 return kzalloc(size, gfp_flags); 80 } 81 82 static void r1bio_pool_free(void *r1_bio, void *data) 83 { 84 kfree(r1_bio); 85 } 86 87 #define RESYNC_BLOCK_SIZE (64*1024) 88 #define RESYNC_DEPTH 32 89 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9) 90 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE) 91 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH) 92 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9) 93 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS) 94 95 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data) 96 { 97 struct pool_info *pi = data; 98 struct r1bio *r1_bio; 99 struct bio *bio; 100 int need_pages; 101 int i, j; 102 103 r1_bio = r1bio_pool_alloc(gfp_flags, pi); 104 if (!r1_bio) 105 return NULL; 106 107 /* 108 * Allocate bios : 1 for reading, n-1 for writing 109 */ 110 for (j = pi->raid_disks ; j-- ; ) { 111 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES); 112 if (!bio) 113 goto out_free_bio; 114 r1_bio->bios[j] = bio; 115 } 116 /* 117 * Allocate RESYNC_PAGES data pages and attach them to 118 * the first bio. 119 * If this is a user-requested check/repair, allocate 120 * RESYNC_PAGES for each bio. 121 */ 122 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) 123 need_pages = pi->raid_disks; 124 else 125 need_pages = 1; 126 for (j = 0; j < need_pages; j++) { 127 bio = r1_bio->bios[j]; 128 bio->bi_vcnt = RESYNC_PAGES; 129 130 if (bio_alloc_pages(bio, gfp_flags)) 131 goto out_free_pages; 132 } 133 /* If not user-requests, copy the page pointers to all bios */ 134 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) { 135 for (i=0; i<RESYNC_PAGES ; i++) 136 for (j=1; j<pi->raid_disks; j++) 137 r1_bio->bios[j]->bi_io_vec[i].bv_page = 138 r1_bio->bios[0]->bi_io_vec[i].bv_page; 139 } 140 141 r1_bio->master_bio = NULL; 142 143 return r1_bio; 144 145 out_free_pages: 146 while (--j >= 0) { 147 struct bio_vec *bv; 148 149 bio_for_each_segment_all(bv, r1_bio->bios[j], i) 150 __free_page(bv->bv_page); 151 } 152 153 out_free_bio: 154 while (++j < pi->raid_disks) 155 bio_put(r1_bio->bios[j]); 156 r1bio_pool_free(r1_bio, data); 157 return NULL; 158 } 159 160 static void r1buf_pool_free(void *__r1_bio, void *data) 161 { 162 struct pool_info *pi = data; 163 int i,j; 164 struct r1bio *r1bio = __r1_bio; 165 166 for (i = 0; i < RESYNC_PAGES; i++) 167 for (j = pi->raid_disks; j-- ;) { 168 if (j == 0 || 169 r1bio->bios[j]->bi_io_vec[i].bv_page != 170 r1bio->bios[0]->bi_io_vec[i].bv_page) 171 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page); 172 } 173 for (i=0 ; i < pi->raid_disks; i++) 174 bio_put(r1bio->bios[i]); 175 176 r1bio_pool_free(r1bio, data); 177 } 178 179 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio) 180 { 181 int i; 182 183 for (i = 0; i < conf->raid_disks * 2; i++) { 184 struct bio **bio = r1_bio->bios + i; 185 if (!BIO_SPECIAL(*bio)) 186 bio_put(*bio); 187 *bio = NULL; 188 } 189 } 190 191 static void free_r1bio(struct r1bio *r1_bio) 192 { 193 struct r1conf *conf = r1_bio->mddev->private; 194 195 put_all_bios(conf, r1_bio); 196 mempool_free(r1_bio, conf->r1bio_pool); 197 } 198 199 static void put_buf(struct r1bio *r1_bio) 200 { 201 struct r1conf *conf = r1_bio->mddev->private; 202 int i; 203 204 for (i = 0; i < conf->raid_disks * 2; i++) { 205 struct bio *bio = r1_bio->bios[i]; 206 if (bio->bi_end_io) 207 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev); 208 } 209 210 mempool_free(r1_bio, conf->r1buf_pool); 211 212 lower_barrier(conf); 213 } 214 215 static void reschedule_retry(struct r1bio *r1_bio) 216 { 217 unsigned long flags; 218 struct mddev *mddev = r1_bio->mddev; 219 struct r1conf *conf = mddev->private; 220 221 spin_lock_irqsave(&conf->device_lock, flags); 222 list_add(&r1_bio->retry_list, &conf->retry_list); 223 conf->nr_queued ++; 224 spin_unlock_irqrestore(&conf->device_lock, flags); 225 226 wake_up(&conf->wait_barrier); 227 md_wakeup_thread(mddev->thread); 228 } 229 230 /* 231 * raid_end_bio_io() is called when we have finished servicing a mirrored 232 * operation and are ready to return a success/failure code to the buffer 233 * cache layer. 234 */ 235 static void call_bio_endio(struct r1bio *r1_bio) 236 { 237 struct bio *bio = r1_bio->master_bio; 238 int done; 239 struct r1conf *conf = r1_bio->mddev->private; 240 sector_t start_next_window = r1_bio->start_next_window; 241 sector_t bi_sector = bio->bi_iter.bi_sector; 242 243 if (bio->bi_phys_segments) { 244 unsigned long flags; 245 spin_lock_irqsave(&conf->device_lock, flags); 246 bio->bi_phys_segments--; 247 done = (bio->bi_phys_segments == 0); 248 spin_unlock_irqrestore(&conf->device_lock, flags); 249 /* 250 * make_request() might be waiting for 251 * bi_phys_segments to decrease 252 */ 253 wake_up(&conf->wait_barrier); 254 } else 255 done = 1; 256 257 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 258 clear_bit(BIO_UPTODATE, &bio->bi_flags); 259 if (done) { 260 bio_endio(bio, 0); 261 /* 262 * Wake up any possible resync thread that waits for the device 263 * to go idle. 264 */ 265 allow_barrier(conf, start_next_window, bi_sector); 266 } 267 } 268 269 static void raid_end_bio_io(struct r1bio *r1_bio) 270 { 271 struct bio *bio = r1_bio->master_bio; 272 273 /* if nobody has done the final endio yet, do it now */ 274 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { 275 pr_debug("raid1: sync end %s on sectors %llu-%llu\n", 276 (bio_data_dir(bio) == WRITE) ? "write" : "read", 277 (unsigned long long) bio->bi_iter.bi_sector, 278 (unsigned long long) bio_end_sector(bio) - 1); 279 280 call_bio_endio(r1_bio); 281 } 282 free_r1bio(r1_bio); 283 } 284 285 /* 286 * Update disk head position estimator based on IRQ completion info. 287 */ 288 static inline void update_head_pos(int disk, struct r1bio *r1_bio) 289 { 290 struct r1conf *conf = r1_bio->mddev->private; 291 292 conf->mirrors[disk].head_position = 293 r1_bio->sector + (r1_bio->sectors); 294 } 295 296 /* 297 * Find the disk number which triggered given bio 298 */ 299 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio) 300 { 301 int mirror; 302 struct r1conf *conf = r1_bio->mddev->private; 303 int raid_disks = conf->raid_disks; 304 305 for (mirror = 0; mirror < raid_disks * 2; mirror++) 306 if (r1_bio->bios[mirror] == bio) 307 break; 308 309 BUG_ON(mirror == raid_disks * 2); 310 update_head_pos(mirror, r1_bio); 311 312 return mirror; 313 } 314 315 static void raid1_end_read_request(struct bio *bio, int error) 316 { 317 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 318 struct r1bio *r1_bio = bio->bi_private; 319 int mirror; 320 struct r1conf *conf = r1_bio->mddev->private; 321 322 mirror = r1_bio->read_disk; 323 /* 324 * this branch is our 'one mirror IO has finished' event handler: 325 */ 326 update_head_pos(mirror, r1_bio); 327 328 if (uptodate) 329 set_bit(R1BIO_Uptodate, &r1_bio->state); 330 else { 331 /* If all other devices have failed, we want to return 332 * the error upwards rather than fail the last device. 333 * Here we redefine "uptodate" to mean "Don't want to retry" 334 */ 335 unsigned long flags; 336 spin_lock_irqsave(&conf->device_lock, flags); 337 if (r1_bio->mddev->degraded == conf->raid_disks || 338 (r1_bio->mddev->degraded == conf->raid_disks-1 && 339 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))) 340 uptodate = 1; 341 spin_unlock_irqrestore(&conf->device_lock, flags); 342 } 343 344 if (uptodate) { 345 raid_end_bio_io(r1_bio); 346 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev); 347 } else { 348 /* 349 * oops, read error: 350 */ 351 char b[BDEVNAME_SIZE]; 352 printk_ratelimited( 353 KERN_ERR "md/raid1:%s: %s: " 354 "rescheduling sector %llu\n", 355 mdname(conf->mddev), 356 bdevname(conf->mirrors[mirror].rdev->bdev, 357 b), 358 (unsigned long long)r1_bio->sector); 359 set_bit(R1BIO_ReadError, &r1_bio->state); 360 reschedule_retry(r1_bio); 361 /* don't drop the reference on read_disk yet */ 362 } 363 } 364 365 static void close_write(struct r1bio *r1_bio) 366 { 367 /* it really is the end of this request */ 368 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 369 /* free extra copy of the data pages */ 370 int i = r1_bio->behind_page_count; 371 while (i--) 372 safe_put_page(r1_bio->behind_bvecs[i].bv_page); 373 kfree(r1_bio->behind_bvecs); 374 r1_bio->behind_bvecs = NULL; 375 } 376 /* clear the bitmap if all writes complete successfully */ 377 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector, 378 r1_bio->sectors, 379 !test_bit(R1BIO_Degraded, &r1_bio->state), 380 test_bit(R1BIO_BehindIO, &r1_bio->state)); 381 md_write_end(r1_bio->mddev); 382 } 383 384 static void r1_bio_write_done(struct r1bio *r1_bio) 385 { 386 if (!atomic_dec_and_test(&r1_bio->remaining)) 387 return; 388 389 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 390 reschedule_retry(r1_bio); 391 else { 392 close_write(r1_bio); 393 if (test_bit(R1BIO_MadeGood, &r1_bio->state)) 394 reschedule_retry(r1_bio); 395 else 396 raid_end_bio_io(r1_bio); 397 } 398 } 399 400 static void raid1_end_write_request(struct bio *bio, int error) 401 { 402 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 403 struct r1bio *r1_bio = bio->bi_private; 404 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state); 405 struct r1conf *conf = r1_bio->mddev->private; 406 struct bio *to_put = NULL; 407 408 mirror = find_bio_disk(r1_bio, bio); 409 410 /* 411 * 'one mirror IO has finished' event handler: 412 */ 413 if (!uptodate) { 414 set_bit(WriteErrorSeen, 415 &conf->mirrors[mirror].rdev->flags); 416 if (!test_and_set_bit(WantReplacement, 417 &conf->mirrors[mirror].rdev->flags)) 418 set_bit(MD_RECOVERY_NEEDED, & 419 conf->mddev->recovery); 420 421 set_bit(R1BIO_WriteError, &r1_bio->state); 422 } else { 423 /* 424 * Set R1BIO_Uptodate in our master bio, so that we 425 * will return a good error code for to the higher 426 * levels even if IO on some other mirrored buffer 427 * fails. 428 * 429 * The 'master' represents the composite IO operation 430 * to user-side. So if something waits for IO, then it 431 * will wait for the 'master' bio. 432 */ 433 sector_t first_bad; 434 int bad_sectors; 435 436 r1_bio->bios[mirror] = NULL; 437 to_put = bio; 438 /* 439 * Do not set R1BIO_Uptodate if the current device is 440 * rebuilding or Faulty. This is because we cannot use 441 * such device for properly reading the data back (we could 442 * potentially use it, if the current write would have felt 443 * before rdev->recovery_offset, but for simplicity we don't 444 * check this here. 445 */ 446 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) && 447 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)) 448 set_bit(R1BIO_Uptodate, &r1_bio->state); 449 450 /* Maybe we can clear some bad blocks. */ 451 if (is_badblock(conf->mirrors[mirror].rdev, 452 r1_bio->sector, r1_bio->sectors, 453 &first_bad, &bad_sectors)) { 454 r1_bio->bios[mirror] = IO_MADE_GOOD; 455 set_bit(R1BIO_MadeGood, &r1_bio->state); 456 } 457 } 458 459 if (behind) { 460 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags)) 461 atomic_dec(&r1_bio->behind_remaining); 462 463 /* 464 * In behind mode, we ACK the master bio once the I/O 465 * has safely reached all non-writemostly 466 * disks. Setting the Returned bit ensures that this 467 * gets done only once -- we don't ever want to return 468 * -EIO here, instead we'll wait 469 */ 470 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) && 471 test_bit(R1BIO_Uptodate, &r1_bio->state)) { 472 /* Maybe we can return now */ 473 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { 474 struct bio *mbio = r1_bio->master_bio; 475 pr_debug("raid1: behind end write sectors" 476 " %llu-%llu\n", 477 (unsigned long long) mbio->bi_iter.bi_sector, 478 (unsigned long long) bio_end_sector(mbio) - 1); 479 call_bio_endio(r1_bio); 480 } 481 } 482 } 483 if (r1_bio->bios[mirror] == NULL) 484 rdev_dec_pending(conf->mirrors[mirror].rdev, 485 conf->mddev); 486 487 /* 488 * Let's see if all mirrored write operations have finished 489 * already. 490 */ 491 r1_bio_write_done(r1_bio); 492 493 if (to_put) 494 bio_put(to_put); 495 } 496 497 498 /* 499 * This routine returns the disk from which the requested read should 500 * be done. There is a per-array 'next expected sequential IO' sector 501 * number - if this matches on the next IO then we use the last disk. 502 * There is also a per-disk 'last know head position' sector that is 503 * maintained from IRQ contexts, both the normal and the resync IO 504 * completion handlers update this position correctly. If there is no 505 * perfect sequential match then we pick the disk whose head is closest. 506 * 507 * If there are 2 mirrors in the same 2 devices, performance degrades 508 * because position is mirror, not device based. 509 * 510 * The rdev for the device selected will have nr_pending incremented. 511 */ 512 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors) 513 { 514 const sector_t this_sector = r1_bio->sector; 515 int sectors; 516 int best_good_sectors; 517 int best_disk, best_dist_disk, best_pending_disk; 518 int has_nonrot_disk; 519 int disk; 520 sector_t best_dist; 521 unsigned int min_pending; 522 struct md_rdev *rdev; 523 int choose_first; 524 int choose_next_idle; 525 526 rcu_read_lock(); 527 /* 528 * Check if we can balance. We can balance on the whole 529 * device if no resync is going on, or below the resync window. 530 * We take the first readable disk when above the resync window. 531 */ 532 retry: 533 sectors = r1_bio->sectors; 534 best_disk = -1; 535 best_dist_disk = -1; 536 best_dist = MaxSector; 537 best_pending_disk = -1; 538 min_pending = UINT_MAX; 539 best_good_sectors = 0; 540 has_nonrot_disk = 0; 541 choose_next_idle = 0; 542 543 if (conf->mddev->recovery_cp < MaxSector && 544 (this_sector + sectors >= conf->next_resync)) 545 choose_first = 1; 546 else 547 choose_first = 0; 548 549 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) { 550 sector_t dist; 551 sector_t first_bad; 552 int bad_sectors; 553 unsigned int pending; 554 bool nonrot; 555 556 rdev = rcu_dereference(conf->mirrors[disk].rdev); 557 if (r1_bio->bios[disk] == IO_BLOCKED 558 || rdev == NULL 559 || test_bit(Unmerged, &rdev->flags) 560 || test_bit(Faulty, &rdev->flags)) 561 continue; 562 if (!test_bit(In_sync, &rdev->flags) && 563 rdev->recovery_offset < this_sector + sectors) 564 continue; 565 if (test_bit(WriteMostly, &rdev->flags)) { 566 /* Don't balance among write-mostly, just 567 * use the first as a last resort */ 568 if (best_disk < 0) { 569 if (is_badblock(rdev, this_sector, sectors, 570 &first_bad, &bad_sectors)) { 571 if (first_bad < this_sector) 572 /* Cannot use this */ 573 continue; 574 best_good_sectors = first_bad - this_sector; 575 } else 576 best_good_sectors = sectors; 577 best_disk = disk; 578 } 579 continue; 580 } 581 /* This is a reasonable device to use. It might 582 * even be best. 583 */ 584 if (is_badblock(rdev, this_sector, sectors, 585 &first_bad, &bad_sectors)) { 586 if (best_dist < MaxSector) 587 /* already have a better device */ 588 continue; 589 if (first_bad <= this_sector) { 590 /* cannot read here. If this is the 'primary' 591 * device, then we must not read beyond 592 * bad_sectors from another device.. 593 */ 594 bad_sectors -= (this_sector - first_bad); 595 if (choose_first && sectors > bad_sectors) 596 sectors = bad_sectors; 597 if (best_good_sectors > sectors) 598 best_good_sectors = sectors; 599 600 } else { 601 sector_t good_sectors = first_bad - this_sector; 602 if (good_sectors > best_good_sectors) { 603 best_good_sectors = good_sectors; 604 best_disk = disk; 605 } 606 if (choose_first) 607 break; 608 } 609 continue; 610 } else 611 best_good_sectors = sectors; 612 613 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev)); 614 has_nonrot_disk |= nonrot; 615 pending = atomic_read(&rdev->nr_pending); 616 dist = abs(this_sector - conf->mirrors[disk].head_position); 617 if (choose_first) { 618 best_disk = disk; 619 break; 620 } 621 /* Don't change to another disk for sequential reads */ 622 if (conf->mirrors[disk].next_seq_sect == this_sector 623 || dist == 0) { 624 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9; 625 struct raid1_info *mirror = &conf->mirrors[disk]; 626 627 best_disk = disk; 628 /* 629 * If buffered sequential IO size exceeds optimal 630 * iosize, check if there is idle disk. If yes, choose 631 * the idle disk. read_balance could already choose an 632 * idle disk before noticing it's a sequential IO in 633 * this disk. This doesn't matter because this disk 634 * will idle, next time it will be utilized after the 635 * first disk has IO size exceeds optimal iosize. In 636 * this way, iosize of the first disk will be optimal 637 * iosize at least. iosize of the second disk might be 638 * small, but not a big deal since when the second disk 639 * starts IO, the first disk is likely still busy. 640 */ 641 if (nonrot && opt_iosize > 0 && 642 mirror->seq_start != MaxSector && 643 mirror->next_seq_sect > opt_iosize && 644 mirror->next_seq_sect - opt_iosize >= 645 mirror->seq_start) { 646 choose_next_idle = 1; 647 continue; 648 } 649 break; 650 } 651 /* If device is idle, use it */ 652 if (pending == 0) { 653 best_disk = disk; 654 break; 655 } 656 657 if (choose_next_idle) 658 continue; 659 660 if (min_pending > pending) { 661 min_pending = pending; 662 best_pending_disk = disk; 663 } 664 665 if (dist < best_dist) { 666 best_dist = dist; 667 best_dist_disk = disk; 668 } 669 } 670 671 /* 672 * If all disks are rotational, choose the closest disk. If any disk is 673 * non-rotational, choose the disk with less pending request even the 674 * disk is rotational, which might/might not be optimal for raids with 675 * mixed ratation/non-rotational disks depending on workload. 676 */ 677 if (best_disk == -1) { 678 if (has_nonrot_disk) 679 best_disk = best_pending_disk; 680 else 681 best_disk = best_dist_disk; 682 } 683 684 if (best_disk >= 0) { 685 rdev = rcu_dereference(conf->mirrors[best_disk].rdev); 686 if (!rdev) 687 goto retry; 688 atomic_inc(&rdev->nr_pending); 689 if (test_bit(Faulty, &rdev->flags)) { 690 /* cannot risk returning a device that failed 691 * before we inc'ed nr_pending 692 */ 693 rdev_dec_pending(rdev, conf->mddev); 694 goto retry; 695 } 696 sectors = best_good_sectors; 697 698 if (conf->mirrors[best_disk].next_seq_sect != this_sector) 699 conf->mirrors[best_disk].seq_start = this_sector; 700 701 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors; 702 } 703 rcu_read_unlock(); 704 *max_sectors = sectors; 705 706 return best_disk; 707 } 708 709 static int raid1_mergeable_bvec(struct request_queue *q, 710 struct bvec_merge_data *bvm, 711 struct bio_vec *biovec) 712 { 713 struct mddev *mddev = q->queuedata; 714 struct r1conf *conf = mddev->private; 715 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev); 716 int max = biovec->bv_len; 717 718 if (mddev->merge_check_needed) { 719 int disk; 720 rcu_read_lock(); 721 for (disk = 0; disk < conf->raid_disks * 2; disk++) { 722 struct md_rdev *rdev = rcu_dereference( 723 conf->mirrors[disk].rdev); 724 if (rdev && !test_bit(Faulty, &rdev->flags)) { 725 struct request_queue *q = 726 bdev_get_queue(rdev->bdev); 727 if (q->merge_bvec_fn) { 728 bvm->bi_sector = sector + 729 rdev->data_offset; 730 bvm->bi_bdev = rdev->bdev; 731 max = min(max, q->merge_bvec_fn( 732 q, bvm, biovec)); 733 } 734 } 735 } 736 rcu_read_unlock(); 737 } 738 return max; 739 740 } 741 742 int md_raid1_congested(struct mddev *mddev, int bits) 743 { 744 struct r1conf *conf = mddev->private; 745 int i, ret = 0; 746 747 if ((bits & (1 << BDI_async_congested)) && 748 conf->pending_count >= max_queued_requests) 749 return 1; 750 751 rcu_read_lock(); 752 for (i = 0; i < conf->raid_disks * 2; i++) { 753 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 754 if (rdev && !test_bit(Faulty, &rdev->flags)) { 755 struct request_queue *q = bdev_get_queue(rdev->bdev); 756 757 BUG_ON(!q); 758 759 /* Note the '|| 1' - when read_balance prefers 760 * non-congested targets, it can be removed 761 */ 762 if ((bits & (1<<BDI_async_congested)) || 1) 763 ret |= bdi_congested(&q->backing_dev_info, bits); 764 else 765 ret &= bdi_congested(&q->backing_dev_info, bits); 766 } 767 } 768 rcu_read_unlock(); 769 return ret; 770 } 771 EXPORT_SYMBOL_GPL(md_raid1_congested); 772 773 static int raid1_congested(void *data, int bits) 774 { 775 struct mddev *mddev = data; 776 777 return mddev_congested(mddev, bits) || 778 md_raid1_congested(mddev, bits); 779 } 780 781 static void flush_pending_writes(struct r1conf *conf) 782 { 783 /* Any writes that have been queued but are awaiting 784 * bitmap updates get flushed here. 785 */ 786 spin_lock_irq(&conf->device_lock); 787 788 if (conf->pending_bio_list.head) { 789 struct bio *bio; 790 bio = bio_list_get(&conf->pending_bio_list); 791 conf->pending_count = 0; 792 spin_unlock_irq(&conf->device_lock); 793 /* flush any pending bitmap writes to 794 * disk before proceeding w/ I/O */ 795 bitmap_unplug(conf->mddev->bitmap); 796 wake_up(&conf->wait_barrier); 797 798 while (bio) { /* submit pending writes */ 799 struct bio *next = bio->bi_next; 800 bio->bi_next = NULL; 801 if (unlikely((bio->bi_rw & REQ_DISCARD) && 802 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 803 /* Just ignore it */ 804 bio_endio(bio, 0); 805 else 806 generic_make_request(bio); 807 bio = next; 808 } 809 } else 810 spin_unlock_irq(&conf->device_lock); 811 } 812 813 /* Barriers.... 814 * Sometimes we need to suspend IO while we do something else, 815 * either some resync/recovery, or reconfigure the array. 816 * To do this we raise a 'barrier'. 817 * The 'barrier' is a counter that can be raised multiple times 818 * to count how many activities are happening which preclude 819 * normal IO. 820 * We can only raise the barrier if there is no pending IO. 821 * i.e. if nr_pending == 0. 822 * We choose only to raise the barrier if no-one is waiting for the 823 * barrier to go down. This means that as soon as an IO request 824 * is ready, no other operations which require a barrier will start 825 * until the IO request has had a chance. 826 * 827 * So: regular IO calls 'wait_barrier'. When that returns there 828 * is no backgroup IO happening, It must arrange to call 829 * allow_barrier when it has finished its IO. 830 * backgroup IO calls must call raise_barrier. Once that returns 831 * there is no normal IO happeing. It must arrange to call 832 * lower_barrier when the particular background IO completes. 833 */ 834 static void raise_barrier(struct r1conf *conf) 835 { 836 spin_lock_irq(&conf->resync_lock); 837 838 /* Wait until no block IO is waiting */ 839 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting, 840 conf->resync_lock); 841 842 /* block any new IO from starting */ 843 conf->barrier++; 844 845 /* For these conditions we must wait: 846 * A: while the array is in frozen state 847 * B: while barrier >= RESYNC_DEPTH, meaning resync reach 848 * the max count which allowed. 849 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning 850 * next resync will reach to the window which normal bios are 851 * handling. 852 */ 853 wait_event_lock_irq(conf->wait_barrier, 854 !conf->array_frozen && 855 conf->barrier < RESYNC_DEPTH && 856 (conf->start_next_window >= 857 conf->next_resync + RESYNC_SECTORS), 858 conf->resync_lock); 859 860 spin_unlock_irq(&conf->resync_lock); 861 } 862 863 static void lower_barrier(struct r1conf *conf) 864 { 865 unsigned long flags; 866 BUG_ON(conf->barrier <= 0); 867 spin_lock_irqsave(&conf->resync_lock, flags); 868 conf->barrier--; 869 spin_unlock_irqrestore(&conf->resync_lock, flags); 870 wake_up(&conf->wait_barrier); 871 } 872 873 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio) 874 { 875 bool wait = false; 876 877 if (conf->array_frozen || !bio) 878 wait = true; 879 else if (conf->barrier && bio_data_dir(bio) == WRITE) { 880 if (conf->next_resync < RESYNC_WINDOW_SECTORS) 881 wait = true; 882 else if ((conf->next_resync - RESYNC_WINDOW_SECTORS 883 >= bio_end_sector(bio)) || 884 (conf->next_resync + NEXT_NORMALIO_DISTANCE 885 <= bio->bi_iter.bi_sector)) 886 wait = false; 887 else 888 wait = true; 889 } 890 891 return wait; 892 } 893 894 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio) 895 { 896 sector_t sector = 0; 897 898 spin_lock_irq(&conf->resync_lock); 899 if (need_to_wait_for_sync(conf, bio)) { 900 conf->nr_waiting++; 901 /* Wait for the barrier to drop. 902 * However if there are already pending 903 * requests (preventing the barrier from 904 * rising completely), and the 905 * pre-process bio queue isn't empty, 906 * then don't wait, as we need to empty 907 * that queue to get the nr_pending 908 * count down. 909 */ 910 wait_event_lock_irq(conf->wait_barrier, 911 !conf->array_frozen && 912 (!conf->barrier || 913 ((conf->start_next_window < 914 conf->next_resync + RESYNC_SECTORS) && 915 current->bio_list && 916 !bio_list_empty(current->bio_list))), 917 conf->resync_lock); 918 conf->nr_waiting--; 919 } 920 921 if (bio && bio_data_dir(bio) == WRITE) { 922 if (conf->next_resync + NEXT_NORMALIO_DISTANCE 923 <= bio->bi_iter.bi_sector) { 924 if (conf->start_next_window == MaxSector) 925 conf->start_next_window = 926 conf->next_resync + 927 NEXT_NORMALIO_DISTANCE; 928 929 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE) 930 <= bio->bi_iter.bi_sector) 931 conf->next_window_requests++; 932 else 933 conf->current_window_requests++; 934 sector = conf->start_next_window; 935 } 936 } 937 938 conf->nr_pending++; 939 spin_unlock_irq(&conf->resync_lock); 940 return sector; 941 } 942 943 static void allow_barrier(struct r1conf *conf, sector_t start_next_window, 944 sector_t bi_sector) 945 { 946 unsigned long flags; 947 948 spin_lock_irqsave(&conf->resync_lock, flags); 949 conf->nr_pending--; 950 if (start_next_window) { 951 if (start_next_window == conf->start_next_window) { 952 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE 953 <= bi_sector) 954 conf->next_window_requests--; 955 else 956 conf->current_window_requests--; 957 } else 958 conf->current_window_requests--; 959 960 if (!conf->current_window_requests) { 961 if (conf->next_window_requests) { 962 conf->current_window_requests = 963 conf->next_window_requests; 964 conf->next_window_requests = 0; 965 conf->start_next_window += 966 NEXT_NORMALIO_DISTANCE; 967 } else 968 conf->start_next_window = MaxSector; 969 } 970 } 971 spin_unlock_irqrestore(&conf->resync_lock, flags); 972 wake_up(&conf->wait_barrier); 973 } 974 975 static void freeze_array(struct r1conf *conf, int extra) 976 { 977 /* stop syncio and normal IO and wait for everything to 978 * go quite. 979 * We wait until nr_pending match nr_queued+extra 980 * This is called in the context of one normal IO request 981 * that has failed. Thus any sync request that might be pending 982 * will be blocked by nr_pending, and we need to wait for 983 * pending IO requests to complete or be queued for re-try. 984 * Thus the number queued (nr_queued) plus this request (extra) 985 * must match the number of pending IOs (nr_pending) before 986 * we continue. 987 */ 988 spin_lock_irq(&conf->resync_lock); 989 conf->array_frozen = 1; 990 wait_event_lock_irq_cmd(conf->wait_barrier, 991 conf->nr_pending == conf->nr_queued+extra, 992 conf->resync_lock, 993 flush_pending_writes(conf)); 994 spin_unlock_irq(&conf->resync_lock); 995 } 996 static void unfreeze_array(struct r1conf *conf) 997 { 998 /* reverse the effect of the freeze */ 999 spin_lock_irq(&conf->resync_lock); 1000 conf->array_frozen = 0; 1001 wake_up(&conf->wait_barrier); 1002 spin_unlock_irq(&conf->resync_lock); 1003 } 1004 1005 1006 /* duplicate the data pages for behind I/O 1007 */ 1008 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio) 1009 { 1010 int i; 1011 struct bio_vec *bvec; 1012 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec), 1013 GFP_NOIO); 1014 if (unlikely(!bvecs)) 1015 return; 1016 1017 bio_for_each_segment_all(bvec, bio, i) { 1018 bvecs[i] = *bvec; 1019 bvecs[i].bv_page = alloc_page(GFP_NOIO); 1020 if (unlikely(!bvecs[i].bv_page)) 1021 goto do_sync_io; 1022 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset, 1023 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len); 1024 kunmap(bvecs[i].bv_page); 1025 kunmap(bvec->bv_page); 1026 } 1027 r1_bio->behind_bvecs = bvecs; 1028 r1_bio->behind_page_count = bio->bi_vcnt; 1029 set_bit(R1BIO_BehindIO, &r1_bio->state); 1030 return; 1031 1032 do_sync_io: 1033 for (i = 0; i < bio->bi_vcnt; i++) 1034 if (bvecs[i].bv_page) 1035 put_page(bvecs[i].bv_page); 1036 kfree(bvecs); 1037 pr_debug("%dB behind alloc failed, doing sync I/O\n", 1038 bio->bi_iter.bi_size); 1039 } 1040 1041 struct raid1_plug_cb { 1042 struct blk_plug_cb cb; 1043 struct bio_list pending; 1044 int pending_cnt; 1045 }; 1046 1047 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule) 1048 { 1049 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, 1050 cb); 1051 struct mddev *mddev = plug->cb.data; 1052 struct r1conf *conf = mddev->private; 1053 struct bio *bio; 1054 1055 if (from_schedule || current->bio_list) { 1056 spin_lock_irq(&conf->device_lock); 1057 bio_list_merge(&conf->pending_bio_list, &plug->pending); 1058 conf->pending_count += plug->pending_cnt; 1059 spin_unlock_irq(&conf->device_lock); 1060 wake_up(&conf->wait_barrier); 1061 md_wakeup_thread(mddev->thread); 1062 kfree(plug); 1063 return; 1064 } 1065 1066 /* we aren't scheduling, so we can do the write-out directly. */ 1067 bio = bio_list_get(&plug->pending); 1068 bitmap_unplug(mddev->bitmap); 1069 wake_up(&conf->wait_barrier); 1070 1071 while (bio) { /* submit pending writes */ 1072 struct bio *next = bio->bi_next; 1073 bio->bi_next = NULL; 1074 if (unlikely((bio->bi_rw & REQ_DISCARD) && 1075 !blk_queue_discard(bdev_get_queue(bio->bi_bdev)))) 1076 /* Just ignore it */ 1077 bio_endio(bio, 0); 1078 else 1079 generic_make_request(bio); 1080 bio = next; 1081 } 1082 kfree(plug); 1083 } 1084 1085 static void make_request(struct mddev *mddev, struct bio * bio) 1086 { 1087 struct r1conf *conf = mddev->private; 1088 struct raid1_info *mirror; 1089 struct r1bio *r1_bio; 1090 struct bio *read_bio; 1091 int i, disks; 1092 struct bitmap *bitmap; 1093 unsigned long flags; 1094 const int rw = bio_data_dir(bio); 1095 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC); 1096 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA)); 1097 const unsigned long do_discard = (bio->bi_rw 1098 & (REQ_DISCARD | REQ_SECURE)); 1099 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME); 1100 struct md_rdev *blocked_rdev; 1101 struct blk_plug_cb *cb; 1102 struct raid1_plug_cb *plug = NULL; 1103 int first_clone; 1104 int sectors_handled; 1105 int max_sectors; 1106 sector_t start_next_window; 1107 1108 /* 1109 * Register the new request and wait if the reconstruction 1110 * thread has put up a bar for new requests. 1111 * Continue immediately if no resync is active currently. 1112 */ 1113 1114 md_write_start(mddev, bio); /* wait on superblock update early */ 1115 1116 if (bio_data_dir(bio) == WRITE && 1117 bio_end_sector(bio) > mddev->suspend_lo && 1118 bio->bi_iter.bi_sector < mddev->suspend_hi) { 1119 /* As the suspend_* range is controlled by 1120 * userspace, we want an interruptible 1121 * wait. 1122 */ 1123 DEFINE_WAIT(w); 1124 for (;;) { 1125 flush_signals(current); 1126 prepare_to_wait(&conf->wait_barrier, 1127 &w, TASK_INTERRUPTIBLE); 1128 if (bio_end_sector(bio) <= mddev->suspend_lo || 1129 bio->bi_iter.bi_sector >= mddev->suspend_hi) 1130 break; 1131 schedule(); 1132 } 1133 finish_wait(&conf->wait_barrier, &w); 1134 } 1135 1136 start_next_window = wait_barrier(conf, bio); 1137 1138 bitmap = mddev->bitmap; 1139 1140 /* 1141 * make_request() can abort the operation when READA is being 1142 * used and no empty request is available. 1143 * 1144 */ 1145 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1146 1147 r1_bio->master_bio = bio; 1148 r1_bio->sectors = bio_sectors(bio); 1149 r1_bio->state = 0; 1150 r1_bio->mddev = mddev; 1151 r1_bio->sector = bio->bi_iter.bi_sector; 1152 1153 /* We might need to issue multiple reads to different 1154 * devices if there are bad blocks around, so we keep 1155 * track of the number of reads in bio->bi_phys_segments. 1156 * If this is 0, there is only one r1_bio and no locking 1157 * will be needed when requests complete. If it is 1158 * non-zero, then it is the number of not-completed requests. 1159 */ 1160 bio->bi_phys_segments = 0; 1161 clear_bit(BIO_SEG_VALID, &bio->bi_flags); 1162 1163 if (rw == READ) { 1164 /* 1165 * read balancing logic: 1166 */ 1167 int rdisk; 1168 1169 read_again: 1170 rdisk = read_balance(conf, r1_bio, &max_sectors); 1171 1172 if (rdisk < 0) { 1173 /* couldn't find anywhere to read from */ 1174 raid_end_bio_io(r1_bio); 1175 return; 1176 } 1177 mirror = conf->mirrors + rdisk; 1178 1179 if (test_bit(WriteMostly, &mirror->rdev->flags) && 1180 bitmap) { 1181 /* Reading from a write-mostly device must 1182 * take care not to over-take any writes 1183 * that are 'behind' 1184 */ 1185 wait_event(bitmap->behind_wait, 1186 atomic_read(&bitmap->behind_writes) == 0); 1187 } 1188 r1_bio->read_disk = rdisk; 1189 r1_bio->start_next_window = 0; 1190 1191 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1192 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector, 1193 max_sectors); 1194 1195 r1_bio->bios[rdisk] = read_bio; 1196 1197 read_bio->bi_iter.bi_sector = r1_bio->sector + 1198 mirror->rdev->data_offset; 1199 read_bio->bi_bdev = mirror->rdev->bdev; 1200 read_bio->bi_end_io = raid1_end_read_request; 1201 read_bio->bi_rw = READ | do_sync; 1202 read_bio->bi_private = r1_bio; 1203 1204 if (max_sectors < r1_bio->sectors) { 1205 /* could not read all from this device, so we will 1206 * need another r1_bio. 1207 */ 1208 1209 sectors_handled = (r1_bio->sector + max_sectors 1210 - bio->bi_iter.bi_sector); 1211 r1_bio->sectors = max_sectors; 1212 spin_lock_irq(&conf->device_lock); 1213 if (bio->bi_phys_segments == 0) 1214 bio->bi_phys_segments = 2; 1215 else 1216 bio->bi_phys_segments++; 1217 spin_unlock_irq(&conf->device_lock); 1218 /* Cannot call generic_make_request directly 1219 * as that will be queued in __make_request 1220 * and subsequent mempool_alloc might block waiting 1221 * for it. So hand bio over to raid1d. 1222 */ 1223 reschedule_retry(r1_bio); 1224 1225 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1226 1227 r1_bio->master_bio = bio; 1228 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1229 r1_bio->state = 0; 1230 r1_bio->mddev = mddev; 1231 r1_bio->sector = bio->bi_iter.bi_sector + 1232 sectors_handled; 1233 goto read_again; 1234 } else 1235 generic_make_request(read_bio); 1236 return; 1237 } 1238 1239 /* 1240 * WRITE: 1241 */ 1242 if (conf->pending_count >= max_queued_requests) { 1243 md_wakeup_thread(mddev->thread); 1244 wait_event(conf->wait_barrier, 1245 conf->pending_count < max_queued_requests); 1246 } 1247 /* first select target devices under rcu_lock and 1248 * inc refcount on their rdev. Record them by setting 1249 * bios[x] to bio 1250 * If there are known/acknowledged bad blocks on any device on 1251 * which we have seen a write error, we want to avoid writing those 1252 * blocks. 1253 * This potentially requires several writes to write around 1254 * the bad blocks. Each set of writes gets it's own r1bio 1255 * with a set of bios attached. 1256 */ 1257 1258 disks = conf->raid_disks * 2; 1259 retry_write: 1260 r1_bio->start_next_window = start_next_window; 1261 blocked_rdev = NULL; 1262 rcu_read_lock(); 1263 max_sectors = r1_bio->sectors; 1264 for (i = 0; i < disks; i++) { 1265 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1266 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 1267 atomic_inc(&rdev->nr_pending); 1268 blocked_rdev = rdev; 1269 break; 1270 } 1271 r1_bio->bios[i] = NULL; 1272 if (!rdev || test_bit(Faulty, &rdev->flags) 1273 || test_bit(Unmerged, &rdev->flags)) { 1274 if (i < conf->raid_disks) 1275 set_bit(R1BIO_Degraded, &r1_bio->state); 1276 continue; 1277 } 1278 1279 atomic_inc(&rdev->nr_pending); 1280 if (test_bit(WriteErrorSeen, &rdev->flags)) { 1281 sector_t first_bad; 1282 int bad_sectors; 1283 int is_bad; 1284 1285 is_bad = is_badblock(rdev, r1_bio->sector, 1286 max_sectors, 1287 &first_bad, &bad_sectors); 1288 if (is_bad < 0) { 1289 /* mustn't write here until the bad block is 1290 * acknowledged*/ 1291 set_bit(BlockedBadBlocks, &rdev->flags); 1292 blocked_rdev = rdev; 1293 break; 1294 } 1295 if (is_bad && first_bad <= r1_bio->sector) { 1296 /* Cannot write here at all */ 1297 bad_sectors -= (r1_bio->sector - first_bad); 1298 if (bad_sectors < max_sectors) 1299 /* mustn't write more than bad_sectors 1300 * to other devices yet 1301 */ 1302 max_sectors = bad_sectors; 1303 rdev_dec_pending(rdev, mddev); 1304 /* We don't set R1BIO_Degraded as that 1305 * only applies if the disk is 1306 * missing, so it might be re-added, 1307 * and we want to know to recover this 1308 * chunk. 1309 * In this case the device is here, 1310 * and the fact that this chunk is not 1311 * in-sync is recorded in the bad 1312 * block log 1313 */ 1314 continue; 1315 } 1316 if (is_bad) { 1317 int good_sectors = first_bad - r1_bio->sector; 1318 if (good_sectors < max_sectors) 1319 max_sectors = good_sectors; 1320 } 1321 } 1322 r1_bio->bios[i] = bio; 1323 } 1324 rcu_read_unlock(); 1325 1326 if (unlikely(blocked_rdev)) { 1327 /* Wait for this device to become unblocked */ 1328 int j; 1329 sector_t old = start_next_window; 1330 1331 for (j = 0; j < i; j++) 1332 if (r1_bio->bios[j]) 1333 rdev_dec_pending(conf->mirrors[j].rdev, mddev); 1334 r1_bio->state = 0; 1335 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector); 1336 md_wait_for_blocked_rdev(blocked_rdev, mddev); 1337 start_next_window = wait_barrier(conf, bio); 1338 /* 1339 * We must make sure the multi r1bios of bio have 1340 * the same value of bi_phys_segments 1341 */ 1342 if (bio->bi_phys_segments && old && 1343 old != start_next_window) 1344 /* Wait for the former r1bio(s) to complete */ 1345 wait_event(conf->wait_barrier, 1346 bio->bi_phys_segments == 1); 1347 goto retry_write; 1348 } 1349 1350 if (max_sectors < r1_bio->sectors) { 1351 /* We are splitting this write into multiple parts, so 1352 * we need to prepare for allocating another r1_bio. 1353 */ 1354 r1_bio->sectors = max_sectors; 1355 spin_lock_irq(&conf->device_lock); 1356 if (bio->bi_phys_segments == 0) 1357 bio->bi_phys_segments = 2; 1358 else 1359 bio->bi_phys_segments++; 1360 spin_unlock_irq(&conf->device_lock); 1361 } 1362 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector; 1363 1364 atomic_set(&r1_bio->remaining, 1); 1365 atomic_set(&r1_bio->behind_remaining, 0); 1366 1367 first_clone = 1; 1368 for (i = 0; i < disks; i++) { 1369 struct bio *mbio; 1370 if (!r1_bio->bios[i]) 1371 continue; 1372 1373 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev); 1374 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors); 1375 1376 if (first_clone) { 1377 /* do behind I/O ? 1378 * Not if there are too many, or cannot 1379 * allocate memory, or a reader on WriteMostly 1380 * is waiting for behind writes to flush */ 1381 if (bitmap && 1382 (atomic_read(&bitmap->behind_writes) 1383 < mddev->bitmap_info.max_write_behind) && 1384 !waitqueue_active(&bitmap->behind_wait)) 1385 alloc_behind_pages(mbio, r1_bio); 1386 1387 bitmap_startwrite(bitmap, r1_bio->sector, 1388 r1_bio->sectors, 1389 test_bit(R1BIO_BehindIO, 1390 &r1_bio->state)); 1391 first_clone = 0; 1392 } 1393 if (r1_bio->behind_bvecs) { 1394 struct bio_vec *bvec; 1395 int j; 1396 1397 /* 1398 * We trimmed the bio, so _all is legit 1399 */ 1400 bio_for_each_segment_all(bvec, mbio, j) 1401 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page; 1402 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags)) 1403 atomic_inc(&r1_bio->behind_remaining); 1404 } 1405 1406 r1_bio->bios[i] = mbio; 1407 1408 mbio->bi_iter.bi_sector = (r1_bio->sector + 1409 conf->mirrors[i].rdev->data_offset); 1410 mbio->bi_bdev = conf->mirrors[i].rdev->bdev; 1411 mbio->bi_end_io = raid1_end_write_request; 1412 mbio->bi_rw = 1413 WRITE | do_flush_fua | do_sync | do_discard | do_same; 1414 mbio->bi_private = r1_bio; 1415 1416 atomic_inc(&r1_bio->remaining); 1417 1418 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug)); 1419 if (cb) 1420 plug = container_of(cb, struct raid1_plug_cb, cb); 1421 else 1422 plug = NULL; 1423 spin_lock_irqsave(&conf->device_lock, flags); 1424 if (plug) { 1425 bio_list_add(&plug->pending, mbio); 1426 plug->pending_cnt++; 1427 } else { 1428 bio_list_add(&conf->pending_bio_list, mbio); 1429 conf->pending_count++; 1430 } 1431 spin_unlock_irqrestore(&conf->device_lock, flags); 1432 if (!plug) 1433 md_wakeup_thread(mddev->thread); 1434 } 1435 /* Mustn't call r1_bio_write_done before this next test, 1436 * as it could result in the bio being freed. 1437 */ 1438 if (sectors_handled < bio_sectors(bio)) { 1439 r1_bio_write_done(r1_bio); 1440 /* We need another r1_bio. It has already been counted 1441 * in bio->bi_phys_segments 1442 */ 1443 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 1444 r1_bio->master_bio = bio; 1445 r1_bio->sectors = bio_sectors(bio) - sectors_handled; 1446 r1_bio->state = 0; 1447 r1_bio->mddev = mddev; 1448 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled; 1449 goto retry_write; 1450 } 1451 1452 r1_bio_write_done(r1_bio); 1453 1454 /* In case raid1d snuck in to freeze_array */ 1455 wake_up(&conf->wait_barrier); 1456 } 1457 1458 static void status(struct seq_file *seq, struct mddev *mddev) 1459 { 1460 struct r1conf *conf = mddev->private; 1461 int i; 1462 1463 seq_printf(seq, " [%d/%d] [", conf->raid_disks, 1464 conf->raid_disks - mddev->degraded); 1465 rcu_read_lock(); 1466 for (i = 0; i < conf->raid_disks; i++) { 1467 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1468 seq_printf(seq, "%s", 1469 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_"); 1470 } 1471 rcu_read_unlock(); 1472 seq_printf(seq, "]"); 1473 } 1474 1475 1476 static void error(struct mddev *mddev, struct md_rdev *rdev) 1477 { 1478 char b[BDEVNAME_SIZE]; 1479 struct r1conf *conf = mddev->private; 1480 1481 /* 1482 * If it is not operational, then we have already marked it as dead 1483 * else if it is the last working disks, ignore the error, let the 1484 * next level up know. 1485 * else mark the drive as failed 1486 */ 1487 if (test_bit(In_sync, &rdev->flags) 1488 && (conf->raid_disks - mddev->degraded) == 1) { 1489 /* 1490 * Don't fail the drive, act as though we were just a 1491 * normal single drive. 1492 * However don't try a recovery from this drive as 1493 * it is very likely to fail. 1494 */ 1495 conf->recovery_disabled = mddev->recovery_disabled; 1496 return; 1497 } 1498 set_bit(Blocked, &rdev->flags); 1499 if (test_and_clear_bit(In_sync, &rdev->flags)) { 1500 unsigned long flags; 1501 spin_lock_irqsave(&conf->device_lock, flags); 1502 mddev->degraded++; 1503 set_bit(Faulty, &rdev->flags); 1504 spin_unlock_irqrestore(&conf->device_lock, flags); 1505 } else 1506 set_bit(Faulty, &rdev->flags); 1507 /* 1508 * if recovery is running, make sure it aborts. 1509 */ 1510 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1511 set_bit(MD_CHANGE_DEVS, &mddev->flags); 1512 printk(KERN_ALERT 1513 "md/raid1:%s: Disk failure on %s, disabling device.\n" 1514 "md/raid1:%s: Operation continuing on %d devices.\n", 1515 mdname(mddev), bdevname(rdev->bdev, b), 1516 mdname(mddev), conf->raid_disks - mddev->degraded); 1517 } 1518 1519 static void print_conf(struct r1conf *conf) 1520 { 1521 int i; 1522 1523 printk(KERN_DEBUG "RAID1 conf printout:\n"); 1524 if (!conf) { 1525 printk(KERN_DEBUG "(!conf)\n"); 1526 return; 1527 } 1528 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded, 1529 conf->raid_disks); 1530 1531 rcu_read_lock(); 1532 for (i = 0; i < conf->raid_disks; i++) { 1533 char b[BDEVNAME_SIZE]; 1534 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); 1535 if (rdev) 1536 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n", 1537 i, !test_bit(In_sync, &rdev->flags), 1538 !test_bit(Faulty, &rdev->flags), 1539 bdevname(rdev->bdev,b)); 1540 } 1541 rcu_read_unlock(); 1542 } 1543 1544 static void close_sync(struct r1conf *conf) 1545 { 1546 wait_barrier(conf, NULL); 1547 allow_barrier(conf, 0, 0); 1548 1549 mempool_destroy(conf->r1buf_pool); 1550 conf->r1buf_pool = NULL; 1551 1552 conf->next_resync = 0; 1553 conf->start_next_window = MaxSector; 1554 } 1555 1556 static int raid1_spare_active(struct mddev *mddev) 1557 { 1558 int i; 1559 struct r1conf *conf = mddev->private; 1560 int count = 0; 1561 unsigned long flags; 1562 1563 /* 1564 * Find all failed disks within the RAID1 configuration 1565 * and mark them readable. 1566 * Called under mddev lock, so rcu protection not needed. 1567 */ 1568 for (i = 0; i < conf->raid_disks; i++) { 1569 struct md_rdev *rdev = conf->mirrors[i].rdev; 1570 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev; 1571 if (repl 1572 && repl->recovery_offset == MaxSector 1573 && !test_bit(Faulty, &repl->flags) 1574 && !test_and_set_bit(In_sync, &repl->flags)) { 1575 /* replacement has just become active */ 1576 if (!rdev || 1577 !test_and_clear_bit(In_sync, &rdev->flags)) 1578 count++; 1579 if (rdev) { 1580 /* Replaced device not technically 1581 * faulty, but we need to be sure 1582 * it gets removed and never re-added 1583 */ 1584 set_bit(Faulty, &rdev->flags); 1585 sysfs_notify_dirent_safe( 1586 rdev->sysfs_state); 1587 } 1588 } 1589 if (rdev 1590 && rdev->recovery_offset == MaxSector 1591 && !test_bit(Faulty, &rdev->flags) 1592 && !test_and_set_bit(In_sync, &rdev->flags)) { 1593 count++; 1594 sysfs_notify_dirent_safe(rdev->sysfs_state); 1595 } 1596 } 1597 spin_lock_irqsave(&conf->device_lock, flags); 1598 mddev->degraded -= count; 1599 spin_unlock_irqrestore(&conf->device_lock, flags); 1600 1601 print_conf(conf); 1602 return count; 1603 } 1604 1605 1606 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev) 1607 { 1608 struct r1conf *conf = mddev->private; 1609 int err = -EEXIST; 1610 int mirror = 0; 1611 struct raid1_info *p; 1612 int first = 0; 1613 int last = conf->raid_disks - 1; 1614 struct request_queue *q = bdev_get_queue(rdev->bdev); 1615 1616 if (mddev->recovery_disabled == conf->recovery_disabled) 1617 return -EBUSY; 1618 1619 if (rdev->raid_disk >= 0) 1620 first = last = rdev->raid_disk; 1621 1622 if (q->merge_bvec_fn) { 1623 set_bit(Unmerged, &rdev->flags); 1624 mddev->merge_check_needed = 1; 1625 } 1626 1627 for (mirror = first; mirror <= last; mirror++) { 1628 p = conf->mirrors+mirror; 1629 if (!p->rdev) { 1630 1631 if (mddev->gendisk) 1632 disk_stack_limits(mddev->gendisk, rdev->bdev, 1633 rdev->data_offset << 9); 1634 1635 p->head_position = 0; 1636 rdev->raid_disk = mirror; 1637 err = 0; 1638 /* As all devices are equivalent, we don't need a full recovery 1639 * if this was recently any drive of the array 1640 */ 1641 if (rdev->saved_raid_disk < 0) 1642 conf->fullsync = 1; 1643 rcu_assign_pointer(p->rdev, rdev); 1644 break; 1645 } 1646 if (test_bit(WantReplacement, &p->rdev->flags) && 1647 p[conf->raid_disks].rdev == NULL) { 1648 /* Add this device as a replacement */ 1649 clear_bit(In_sync, &rdev->flags); 1650 set_bit(Replacement, &rdev->flags); 1651 rdev->raid_disk = mirror; 1652 err = 0; 1653 conf->fullsync = 1; 1654 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev); 1655 break; 1656 } 1657 } 1658 if (err == 0 && test_bit(Unmerged, &rdev->flags)) { 1659 /* Some requests might not have seen this new 1660 * merge_bvec_fn. We must wait for them to complete 1661 * before merging the device fully. 1662 * First we make sure any code which has tested 1663 * our function has submitted the request, then 1664 * we wait for all outstanding requests to complete. 1665 */ 1666 synchronize_sched(); 1667 freeze_array(conf, 0); 1668 unfreeze_array(conf); 1669 clear_bit(Unmerged, &rdev->flags); 1670 } 1671 md_integrity_add_rdev(rdev, mddev); 1672 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev))) 1673 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue); 1674 print_conf(conf); 1675 return err; 1676 } 1677 1678 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev) 1679 { 1680 struct r1conf *conf = mddev->private; 1681 int err = 0; 1682 int number = rdev->raid_disk; 1683 struct raid1_info *p = conf->mirrors + number; 1684 1685 if (rdev != p->rdev) 1686 p = conf->mirrors + conf->raid_disks + number; 1687 1688 print_conf(conf); 1689 if (rdev == p->rdev) { 1690 if (test_bit(In_sync, &rdev->flags) || 1691 atomic_read(&rdev->nr_pending)) { 1692 err = -EBUSY; 1693 goto abort; 1694 } 1695 /* Only remove non-faulty devices if recovery 1696 * is not possible. 1697 */ 1698 if (!test_bit(Faulty, &rdev->flags) && 1699 mddev->recovery_disabled != conf->recovery_disabled && 1700 mddev->degraded < conf->raid_disks) { 1701 err = -EBUSY; 1702 goto abort; 1703 } 1704 p->rdev = NULL; 1705 synchronize_rcu(); 1706 if (atomic_read(&rdev->nr_pending)) { 1707 /* lost the race, try later */ 1708 err = -EBUSY; 1709 p->rdev = rdev; 1710 goto abort; 1711 } else if (conf->mirrors[conf->raid_disks + number].rdev) { 1712 /* We just removed a device that is being replaced. 1713 * Move down the replacement. We drain all IO before 1714 * doing this to avoid confusion. 1715 */ 1716 struct md_rdev *repl = 1717 conf->mirrors[conf->raid_disks + number].rdev; 1718 freeze_array(conf, 0); 1719 clear_bit(Replacement, &repl->flags); 1720 p->rdev = repl; 1721 conf->mirrors[conf->raid_disks + number].rdev = NULL; 1722 unfreeze_array(conf); 1723 clear_bit(WantReplacement, &rdev->flags); 1724 } else 1725 clear_bit(WantReplacement, &rdev->flags); 1726 err = md_integrity_register(mddev); 1727 } 1728 abort: 1729 1730 print_conf(conf); 1731 return err; 1732 } 1733 1734 1735 static void end_sync_read(struct bio *bio, int error) 1736 { 1737 struct r1bio *r1_bio = bio->bi_private; 1738 1739 update_head_pos(r1_bio->read_disk, r1_bio); 1740 1741 /* 1742 * we have read a block, now it needs to be re-written, 1743 * or re-read if the read failed. 1744 * We don't do much here, just schedule handling by raid1d 1745 */ 1746 if (test_bit(BIO_UPTODATE, &bio->bi_flags)) 1747 set_bit(R1BIO_Uptodate, &r1_bio->state); 1748 1749 if (atomic_dec_and_test(&r1_bio->remaining)) 1750 reschedule_retry(r1_bio); 1751 } 1752 1753 static void end_sync_write(struct bio *bio, int error) 1754 { 1755 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 1756 struct r1bio *r1_bio = bio->bi_private; 1757 struct mddev *mddev = r1_bio->mddev; 1758 struct r1conf *conf = mddev->private; 1759 int mirror=0; 1760 sector_t first_bad; 1761 int bad_sectors; 1762 1763 mirror = find_bio_disk(r1_bio, bio); 1764 1765 if (!uptodate) { 1766 sector_t sync_blocks = 0; 1767 sector_t s = r1_bio->sector; 1768 long sectors_to_go = r1_bio->sectors; 1769 /* make sure these bits doesn't get cleared. */ 1770 do { 1771 bitmap_end_sync(mddev->bitmap, s, 1772 &sync_blocks, 1); 1773 s += sync_blocks; 1774 sectors_to_go -= sync_blocks; 1775 } while (sectors_to_go > 0); 1776 set_bit(WriteErrorSeen, 1777 &conf->mirrors[mirror].rdev->flags); 1778 if (!test_and_set_bit(WantReplacement, 1779 &conf->mirrors[mirror].rdev->flags)) 1780 set_bit(MD_RECOVERY_NEEDED, & 1781 mddev->recovery); 1782 set_bit(R1BIO_WriteError, &r1_bio->state); 1783 } else if (is_badblock(conf->mirrors[mirror].rdev, 1784 r1_bio->sector, 1785 r1_bio->sectors, 1786 &first_bad, &bad_sectors) && 1787 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev, 1788 r1_bio->sector, 1789 r1_bio->sectors, 1790 &first_bad, &bad_sectors) 1791 ) 1792 set_bit(R1BIO_MadeGood, &r1_bio->state); 1793 1794 if (atomic_dec_and_test(&r1_bio->remaining)) { 1795 int s = r1_bio->sectors; 1796 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 1797 test_bit(R1BIO_WriteError, &r1_bio->state)) 1798 reschedule_retry(r1_bio); 1799 else { 1800 put_buf(r1_bio); 1801 md_done_sync(mddev, s, uptodate); 1802 } 1803 } 1804 } 1805 1806 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector, 1807 int sectors, struct page *page, int rw) 1808 { 1809 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false)) 1810 /* success */ 1811 return 1; 1812 if (rw == WRITE) { 1813 set_bit(WriteErrorSeen, &rdev->flags); 1814 if (!test_and_set_bit(WantReplacement, 1815 &rdev->flags)) 1816 set_bit(MD_RECOVERY_NEEDED, & 1817 rdev->mddev->recovery); 1818 } 1819 /* need to record an error - either for the block or the device */ 1820 if (!rdev_set_badblocks(rdev, sector, sectors, 0)) 1821 md_error(rdev->mddev, rdev); 1822 return 0; 1823 } 1824 1825 static int fix_sync_read_error(struct r1bio *r1_bio) 1826 { 1827 /* Try some synchronous reads of other devices to get 1828 * good data, much like with normal read errors. Only 1829 * read into the pages we already have so we don't 1830 * need to re-issue the read request. 1831 * We don't need to freeze the array, because being in an 1832 * active sync request, there is no normal IO, and 1833 * no overlapping syncs. 1834 * We don't need to check is_badblock() again as we 1835 * made sure that anything with a bad block in range 1836 * will have bi_end_io clear. 1837 */ 1838 struct mddev *mddev = r1_bio->mddev; 1839 struct r1conf *conf = mddev->private; 1840 struct bio *bio = r1_bio->bios[r1_bio->read_disk]; 1841 sector_t sect = r1_bio->sector; 1842 int sectors = r1_bio->sectors; 1843 int idx = 0; 1844 1845 while(sectors) { 1846 int s = sectors; 1847 int d = r1_bio->read_disk; 1848 int success = 0; 1849 struct md_rdev *rdev; 1850 int start; 1851 1852 if (s > (PAGE_SIZE>>9)) 1853 s = PAGE_SIZE >> 9; 1854 do { 1855 if (r1_bio->bios[d]->bi_end_io == end_sync_read) { 1856 /* No rcu protection needed here devices 1857 * can only be removed when no resync is 1858 * active, and resync is currently active 1859 */ 1860 rdev = conf->mirrors[d].rdev; 1861 if (sync_page_io(rdev, sect, s<<9, 1862 bio->bi_io_vec[idx].bv_page, 1863 READ, false)) { 1864 success = 1; 1865 break; 1866 } 1867 } 1868 d++; 1869 if (d == conf->raid_disks * 2) 1870 d = 0; 1871 } while (!success && d != r1_bio->read_disk); 1872 1873 if (!success) { 1874 char b[BDEVNAME_SIZE]; 1875 int abort = 0; 1876 /* Cannot read from anywhere, this block is lost. 1877 * Record a bad block on each device. If that doesn't 1878 * work just disable and interrupt the recovery. 1879 * Don't fail devices as that won't really help. 1880 */ 1881 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error" 1882 " for block %llu\n", 1883 mdname(mddev), 1884 bdevname(bio->bi_bdev, b), 1885 (unsigned long long)r1_bio->sector); 1886 for (d = 0; d < conf->raid_disks * 2; d++) { 1887 rdev = conf->mirrors[d].rdev; 1888 if (!rdev || test_bit(Faulty, &rdev->flags)) 1889 continue; 1890 if (!rdev_set_badblocks(rdev, sect, s, 0)) 1891 abort = 1; 1892 } 1893 if (abort) { 1894 conf->recovery_disabled = 1895 mddev->recovery_disabled; 1896 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1897 md_done_sync(mddev, r1_bio->sectors, 0); 1898 put_buf(r1_bio); 1899 return 0; 1900 } 1901 /* Try next page */ 1902 sectors -= s; 1903 sect += s; 1904 idx++; 1905 continue; 1906 } 1907 1908 start = d; 1909 /* write it back and re-read */ 1910 while (d != r1_bio->read_disk) { 1911 if (d == 0) 1912 d = conf->raid_disks * 2; 1913 d--; 1914 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1915 continue; 1916 rdev = conf->mirrors[d].rdev; 1917 if (r1_sync_page_io(rdev, sect, s, 1918 bio->bi_io_vec[idx].bv_page, 1919 WRITE) == 0) { 1920 r1_bio->bios[d]->bi_end_io = NULL; 1921 rdev_dec_pending(rdev, mddev); 1922 } 1923 } 1924 d = start; 1925 while (d != r1_bio->read_disk) { 1926 if (d == 0) 1927 d = conf->raid_disks * 2; 1928 d--; 1929 if (r1_bio->bios[d]->bi_end_io != end_sync_read) 1930 continue; 1931 rdev = conf->mirrors[d].rdev; 1932 if (r1_sync_page_io(rdev, sect, s, 1933 bio->bi_io_vec[idx].bv_page, 1934 READ) != 0) 1935 atomic_add(s, &rdev->corrected_errors); 1936 } 1937 sectors -= s; 1938 sect += s; 1939 idx ++; 1940 } 1941 set_bit(R1BIO_Uptodate, &r1_bio->state); 1942 set_bit(BIO_UPTODATE, &bio->bi_flags); 1943 return 1; 1944 } 1945 1946 static int process_checks(struct r1bio *r1_bio) 1947 { 1948 /* We have read all readable devices. If we haven't 1949 * got the block, then there is no hope left. 1950 * If we have, then we want to do a comparison 1951 * and skip the write if everything is the same. 1952 * If any blocks failed to read, then we need to 1953 * attempt an over-write 1954 */ 1955 struct mddev *mddev = r1_bio->mddev; 1956 struct r1conf *conf = mddev->private; 1957 int primary; 1958 int i; 1959 int vcnt; 1960 1961 /* Fix variable parts of all bios */ 1962 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9); 1963 for (i = 0; i < conf->raid_disks * 2; i++) { 1964 int j; 1965 int size; 1966 int uptodate; 1967 struct bio *b = r1_bio->bios[i]; 1968 if (b->bi_end_io != end_sync_read) 1969 continue; 1970 /* fixup the bio for reuse, but preserve BIO_UPTODATE */ 1971 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags); 1972 bio_reset(b); 1973 if (!uptodate) 1974 clear_bit(BIO_UPTODATE, &b->bi_flags); 1975 b->bi_vcnt = vcnt; 1976 b->bi_iter.bi_size = r1_bio->sectors << 9; 1977 b->bi_iter.bi_sector = r1_bio->sector + 1978 conf->mirrors[i].rdev->data_offset; 1979 b->bi_bdev = conf->mirrors[i].rdev->bdev; 1980 b->bi_end_io = end_sync_read; 1981 b->bi_private = r1_bio; 1982 1983 size = b->bi_iter.bi_size; 1984 for (j = 0; j < vcnt ; j++) { 1985 struct bio_vec *bi; 1986 bi = &b->bi_io_vec[j]; 1987 bi->bv_offset = 0; 1988 if (size > PAGE_SIZE) 1989 bi->bv_len = PAGE_SIZE; 1990 else 1991 bi->bv_len = size; 1992 size -= PAGE_SIZE; 1993 } 1994 } 1995 for (primary = 0; primary < conf->raid_disks * 2; primary++) 1996 if (r1_bio->bios[primary]->bi_end_io == end_sync_read && 1997 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) { 1998 r1_bio->bios[primary]->bi_end_io = NULL; 1999 rdev_dec_pending(conf->mirrors[primary].rdev, mddev); 2000 break; 2001 } 2002 r1_bio->read_disk = primary; 2003 for (i = 0; i < conf->raid_disks * 2; i++) { 2004 int j; 2005 struct bio *pbio = r1_bio->bios[primary]; 2006 struct bio *sbio = r1_bio->bios[i]; 2007 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags); 2008 2009 if (sbio->bi_end_io != end_sync_read) 2010 continue; 2011 /* Now we can 'fixup' the BIO_UPTODATE flag */ 2012 set_bit(BIO_UPTODATE, &sbio->bi_flags); 2013 2014 if (uptodate) { 2015 for (j = vcnt; j-- ; ) { 2016 struct page *p, *s; 2017 p = pbio->bi_io_vec[j].bv_page; 2018 s = sbio->bi_io_vec[j].bv_page; 2019 if (memcmp(page_address(p), 2020 page_address(s), 2021 sbio->bi_io_vec[j].bv_len)) 2022 break; 2023 } 2024 } else 2025 j = 0; 2026 if (j >= 0) 2027 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches); 2028 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery) 2029 && uptodate)) { 2030 /* No need to write to this device. */ 2031 sbio->bi_end_io = NULL; 2032 rdev_dec_pending(conf->mirrors[i].rdev, mddev); 2033 continue; 2034 } 2035 2036 bio_copy_data(sbio, pbio); 2037 } 2038 return 0; 2039 } 2040 2041 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio) 2042 { 2043 struct r1conf *conf = mddev->private; 2044 int i; 2045 int disks = conf->raid_disks * 2; 2046 struct bio *bio, *wbio; 2047 2048 bio = r1_bio->bios[r1_bio->read_disk]; 2049 2050 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) 2051 /* ouch - failed to read all of that. */ 2052 if (!fix_sync_read_error(r1_bio)) 2053 return; 2054 2055 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2056 if (process_checks(r1_bio) < 0) 2057 return; 2058 /* 2059 * schedule writes 2060 */ 2061 atomic_set(&r1_bio->remaining, 1); 2062 for (i = 0; i < disks ; i++) { 2063 wbio = r1_bio->bios[i]; 2064 if (wbio->bi_end_io == NULL || 2065 (wbio->bi_end_io == end_sync_read && 2066 (i == r1_bio->read_disk || 2067 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)))) 2068 continue; 2069 2070 wbio->bi_rw = WRITE; 2071 wbio->bi_end_io = end_sync_write; 2072 atomic_inc(&r1_bio->remaining); 2073 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio)); 2074 2075 generic_make_request(wbio); 2076 } 2077 2078 if (atomic_dec_and_test(&r1_bio->remaining)) { 2079 /* if we're here, all write(s) have completed, so clean up */ 2080 int s = r1_bio->sectors; 2081 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2082 test_bit(R1BIO_WriteError, &r1_bio->state)) 2083 reschedule_retry(r1_bio); 2084 else { 2085 put_buf(r1_bio); 2086 md_done_sync(mddev, s, 1); 2087 } 2088 } 2089 } 2090 2091 /* 2092 * This is a kernel thread which: 2093 * 2094 * 1. Retries failed read operations on working mirrors. 2095 * 2. Updates the raid superblock when problems encounter. 2096 * 3. Performs writes following reads for array synchronising. 2097 */ 2098 2099 static void fix_read_error(struct r1conf *conf, int read_disk, 2100 sector_t sect, int sectors) 2101 { 2102 struct mddev *mddev = conf->mddev; 2103 while(sectors) { 2104 int s = sectors; 2105 int d = read_disk; 2106 int success = 0; 2107 int start; 2108 struct md_rdev *rdev; 2109 2110 if (s > (PAGE_SIZE>>9)) 2111 s = PAGE_SIZE >> 9; 2112 2113 do { 2114 /* Note: no rcu protection needed here 2115 * as this is synchronous in the raid1d thread 2116 * which is the thread that might remove 2117 * a device. If raid1d ever becomes multi-threaded.... 2118 */ 2119 sector_t first_bad; 2120 int bad_sectors; 2121 2122 rdev = conf->mirrors[d].rdev; 2123 if (rdev && 2124 (test_bit(In_sync, &rdev->flags) || 2125 (!test_bit(Faulty, &rdev->flags) && 2126 rdev->recovery_offset >= sect + s)) && 2127 is_badblock(rdev, sect, s, 2128 &first_bad, &bad_sectors) == 0 && 2129 sync_page_io(rdev, sect, s<<9, 2130 conf->tmppage, READ, false)) 2131 success = 1; 2132 else { 2133 d++; 2134 if (d == conf->raid_disks * 2) 2135 d = 0; 2136 } 2137 } while (!success && d != read_disk); 2138 2139 if (!success) { 2140 /* Cannot read from anywhere - mark it bad */ 2141 struct md_rdev *rdev = conf->mirrors[read_disk].rdev; 2142 if (!rdev_set_badblocks(rdev, sect, s, 0)) 2143 md_error(mddev, rdev); 2144 break; 2145 } 2146 /* write it back and re-read */ 2147 start = d; 2148 while (d != read_disk) { 2149 if (d==0) 2150 d = conf->raid_disks * 2; 2151 d--; 2152 rdev = conf->mirrors[d].rdev; 2153 if (rdev && 2154 test_bit(In_sync, &rdev->flags)) 2155 r1_sync_page_io(rdev, sect, s, 2156 conf->tmppage, WRITE); 2157 } 2158 d = start; 2159 while (d != read_disk) { 2160 char b[BDEVNAME_SIZE]; 2161 if (d==0) 2162 d = conf->raid_disks * 2; 2163 d--; 2164 rdev = conf->mirrors[d].rdev; 2165 if (rdev && 2166 test_bit(In_sync, &rdev->flags)) { 2167 if (r1_sync_page_io(rdev, sect, s, 2168 conf->tmppage, READ)) { 2169 atomic_add(s, &rdev->corrected_errors); 2170 printk(KERN_INFO 2171 "md/raid1:%s: read error corrected " 2172 "(%d sectors at %llu on %s)\n", 2173 mdname(mddev), s, 2174 (unsigned long long)(sect + 2175 rdev->data_offset), 2176 bdevname(rdev->bdev, b)); 2177 } 2178 } 2179 } 2180 sectors -= s; 2181 sect += s; 2182 } 2183 } 2184 2185 static int narrow_write_error(struct r1bio *r1_bio, int i) 2186 { 2187 struct mddev *mddev = r1_bio->mddev; 2188 struct r1conf *conf = mddev->private; 2189 struct md_rdev *rdev = conf->mirrors[i].rdev; 2190 2191 /* bio has the data to be written to device 'i' where 2192 * we just recently had a write error. 2193 * We repeatedly clone the bio and trim down to one block, 2194 * then try the write. Where the write fails we record 2195 * a bad block. 2196 * It is conceivable that the bio doesn't exactly align with 2197 * blocks. We must handle this somehow. 2198 * 2199 * We currently own a reference on the rdev. 2200 */ 2201 2202 int block_sectors; 2203 sector_t sector; 2204 int sectors; 2205 int sect_to_write = r1_bio->sectors; 2206 int ok = 1; 2207 2208 if (rdev->badblocks.shift < 0) 2209 return 0; 2210 2211 block_sectors = 1 << rdev->badblocks.shift; 2212 sector = r1_bio->sector; 2213 sectors = ((sector + block_sectors) 2214 & ~(sector_t)(block_sectors - 1)) 2215 - sector; 2216 2217 while (sect_to_write) { 2218 struct bio *wbio; 2219 if (sectors > sect_to_write) 2220 sectors = sect_to_write; 2221 /* Write at 'sector' for 'sectors'*/ 2222 2223 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { 2224 unsigned vcnt = r1_bio->behind_page_count; 2225 struct bio_vec *vec = r1_bio->behind_bvecs; 2226 2227 while (!vec->bv_page) { 2228 vec++; 2229 vcnt--; 2230 } 2231 2232 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev); 2233 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec)); 2234 2235 wbio->bi_vcnt = vcnt; 2236 } else { 2237 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2238 } 2239 2240 wbio->bi_rw = WRITE; 2241 wbio->bi_iter.bi_sector = r1_bio->sector; 2242 wbio->bi_iter.bi_size = r1_bio->sectors << 9; 2243 2244 bio_trim(wbio, sector - r1_bio->sector, sectors); 2245 wbio->bi_iter.bi_sector += rdev->data_offset; 2246 wbio->bi_bdev = rdev->bdev; 2247 if (submit_bio_wait(WRITE, wbio) == 0) 2248 /* failure! */ 2249 ok = rdev_set_badblocks(rdev, sector, 2250 sectors, 0) 2251 && ok; 2252 2253 bio_put(wbio); 2254 sect_to_write -= sectors; 2255 sector += sectors; 2256 sectors = block_sectors; 2257 } 2258 return ok; 2259 } 2260 2261 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2262 { 2263 int m; 2264 int s = r1_bio->sectors; 2265 for (m = 0; m < conf->raid_disks * 2 ; m++) { 2266 struct md_rdev *rdev = conf->mirrors[m].rdev; 2267 struct bio *bio = r1_bio->bios[m]; 2268 if (bio->bi_end_io == NULL) 2269 continue; 2270 if (test_bit(BIO_UPTODATE, &bio->bi_flags) && 2271 test_bit(R1BIO_MadeGood, &r1_bio->state)) { 2272 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0); 2273 } 2274 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) && 2275 test_bit(R1BIO_WriteError, &r1_bio->state)) { 2276 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0)) 2277 md_error(conf->mddev, rdev); 2278 } 2279 } 2280 put_buf(r1_bio); 2281 md_done_sync(conf->mddev, s, 1); 2282 } 2283 2284 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio) 2285 { 2286 int m; 2287 for (m = 0; m < conf->raid_disks * 2 ; m++) 2288 if (r1_bio->bios[m] == IO_MADE_GOOD) { 2289 struct md_rdev *rdev = conf->mirrors[m].rdev; 2290 rdev_clear_badblocks(rdev, 2291 r1_bio->sector, 2292 r1_bio->sectors, 0); 2293 rdev_dec_pending(rdev, conf->mddev); 2294 } else if (r1_bio->bios[m] != NULL) { 2295 /* This drive got a write error. We need to 2296 * narrow down and record precise write 2297 * errors. 2298 */ 2299 if (!narrow_write_error(r1_bio, m)) { 2300 md_error(conf->mddev, 2301 conf->mirrors[m].rdev); 2302 /* an I/O failed, we can't clear the bitmap */ 2303 set_bit(R1BIO_Degraded, &r1_bio->state); 2304 } 2305 rdev_dec_pending(conf->mirrors[m].rdev, 2306 conf->mddev); 2307 } 2308 if (test_bit(R1BIO_WriteError, &r1_bio->state)) 2309 close_write(r1_bio); 2310 raid_end_bio_io(r1_bio); 2311 } 2312 2313 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio) 2314 { 2315 int disk; 2316 int max_sectors; 2317 struct mddev *mddev = conf->mddev; 2318 struct bio *bio; 2319 char b[BDEVNAME_SIZE]; 2320 struct md_rdev *rdev; 2321 2322 clear_bit(R1BIO_ReadError, &r1_bio->state); 2323 /* we got a read error. Maybe the drive is bad. Maybe just 2324 * the block and we can fix it. 2325 * We freeze all other IO, and try reading the block from 2326 * other devices. When we find one, we re-write 2327 * and check it that fixes the read error. 2328 * This is all done synchronously while the array is 2329 * frozen 2330 */ 2331 if (mddev->ro == 0) { 2332 freeze_array(conf, 1); 2333 fix_read_error(conf, r1_bio->read_disk, 2334 r1_bio->sector, r1_bio->sectors); 2335 unfreeze_array(conf); 2336 } else 2337 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev); 2338 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev); 2339 2340 bio = r1_bio->bios[r1_bio->read_disk]; 2341 bdevname(bio->bi_bdev, b); 2342 read_more: 2343 disk = read_balance(conf, r1_bio, &max_sectors); 2344 if (disk == -1) { 2345 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O" 2346 " read error for block %llu\n", 2347 mdname(mddev), b, (unsigned long long)r1_bio->sector); 2348 raid_end_bio_io(r1_bio); 2349 } else { 2350 const unsigned long do_sync 2351 = r1_bio->master_bio->bi_rw & REQ_SYNC; 2352 if (bio) { 2353 r1_bio->bios[r1_bio->read_disk] = 2354 mddev->ro ? IO_BLOCKED : NULL; 2355 bio_put(bio); 2356 } 2357 r1_bio->read_disk = disk; 2358 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev); 2359 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector, 2360 max_sectors); 2361 r1_bio->bios[r1_bio->read_disk] = bio; 2362 rdev = conf->mirrors[disk].rdev; 2363 printk_ratelimited(KERN_ERR 2364 "md/raid1:%s: redirecting sector %llu" 2365 " to other mirror: %s\n", 2366 mdname(mddev), 2367 (unsigned long long)r1_bio->sector, 2368 bdevname(rdev->bdev, b)); 2369 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset; 2370 bio->bi_bdev = rdev->bdev; 2371 bio->bi_end_io = raid1_end_read_request; 2372 bio->bi_rw = READ | do_sync; 2373 bio->bi_private = r1_bio; 2374 if (max_sectors < r1_bio->sectors) { 2375 /* Drat - have to split this up more */ 2376 struct bio *mbio = r1_bio->master_bio; 2377 int sectors_handled = (r1_bio->sector + max_sectors 2378 - mbio->bi_iter.bi_sector); 2379 r1_bio->sectors = max_sectors; 2380 spin_lock_irq(&conf->device_lock); 2381 if (mbio->bi_phys_segments == 0) 2382 mbio->bi_phys_segments = 2; 2383 else 2384 mbio->bi_phys_segments++; 2385 spin_unlock_irq(&conf->device_lock); 2386 generic_make_request(bio); 2387 bio = NULL; 2388 2389 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); 2390 2391 r1_bio->master_bio = mbio; 2392 r1_bio->sectors = bio_sectors(mbio) - sectors_handled; 2393 r1_bio->state = 0; 2394 set_bit(R1BIO_ReadError, &r1_bio->state); 2395 r1_bio->mddev = mddev; 2396 r1_bio->sector = mbio->bi_iter.bi_sector + 2397 sectors_handled; 2398 2399 goto read_more; 2400 } else 2401 generic_make_request(bio); 2402 } 2403 } 2404 2405 static void raid1d(struct md_thread *thread) 2406 { 2407 struct mddev *mddev = thread->mddev; 2408 struct r1bio *r1_bio; 2409 unsigned long flags; 2410 struct r1conf *conf = mddev->private; 2411 struct list_head *head = &conf->retry_list; 2412 struct blk_plug plug; 2413 2414 md_check_recovery(mddev); 2415 2416 blk_start_plug(&plug); 2417 for (;;) { 2418 2419 flush_pending_writes(conf); 2420 2421 spin_lock_irqsave(&conf->device_lock, flags); 2422 if (list_empty(head)) { 2423 spin_unlock_irqrestore(&conf->device_lock, flags); 2424 break; 2425 } 2426 r1_bio = list_entry(head->prev, struct r1bio, retry_list); 2427 list_del(head->prev); 2428 conf->nr_queued--; 2429 spin_unlock_irqrestore(&conf->device_lock, flags); 2430 2431 mddev = r1_bio->mddev; 2432 conf = mddev->private; 2433 if (test_bit(R1BIO_IsSync, &r1_bio->state)) { 2434 if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2435 test_bit(R1BIO_WriteError, &r1_bio->state)) 2436 handle_sync_write_finished(conf, r1_bio); 2437 else 2438 sync_request_write(mddev, r1_bio); 2439 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) || 2440 test_bit(R1BIO_WriteError, &r1_bio->state)) 2441 handle_write_finished(conf, r1_bio); 2442 else if (test_bit(R1BIO_ReadError, &r1_bio->state)) 2443 handle_read_error(conf, r1_bio); 2444 else 2445 /* just a partial read to be scheduled from separate 2446 * context 2447 */ 2448 generic_make_request(r1_bio->bios[r1_bio->read_disk]); 2449 2450 cond_resched(); 2451 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) 2452 md_check_recovery(mddev); 2453 } 2454 blk_finish_plug(&plug); 2455 } 2456 2457 2458 static int init_resync(struct r1conf *conf) 2459 { 2460 int buffs; 2461 2462 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; 2463 BUG_ON(conf->r1buf_pool); 2464 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, 2465 conf->poolinfo); 2466 if (!conf->r1buf_pool) 2467 return -ENOMEM; 2468 conf->next_resync = 0; 2469 return 0; 2470 } 2471 2472 /* 2473 * perform a "sync" on one "block" 2474 * 2475 * We need to make sure that no normal I/O request - particularly write 2476 * requests - conflict with active sync requests. 2477 * 2478 * This is achieved by tracking pending requests and a 'barrier' concept 2479 * that can be installed to exclude normal IO requests. 2480 */ 2481 2482 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster) 2483 { 2484 struct r1conf *conf = mddev->private; 2485 struct r1bio *r1_bio; 2486 struct bio *bio; 2487 sector_t max_sector, nr_sectors; 2488 int disk = -1; 2489 int i; 2490 int wonly = -1; 2491 int write_targets = 0, read_targets = 0; 2492 sector_t sync_blocks; 2493 int still_degraded = 0; 2494 int good_sectors = RESYNC_SECTORS; 2495 int min_bad = 0; /* number of sectors that are bad in all devices */ 2496 2497 if (!conf->r1buf_pool) 2498 if (init_resync(conf)) 2499 return 0; 2500 2501 max_sector = mddev->dev_sectors; 2502 if (sector_nr >= max_sector) { 2503 /* If we aborted, we need to abort the 2504 * sync on the 'current' bitmap chunk (there will 2505 * only be one in raid1 resync. 2506 * We can find the current addess in mddev->curr_resync 2507 */ 2508 if (mddev->curr_resync < max_sector) /* aborted */ 2509 bitmap_end_sync(mddev->bitmap, mddev->curr_resync, 2510 &sync_blocks, 1); 2511 else /* completed sync */ 2512 conf->fullsync = 0; 2513 2514 bitmap_close_sync(mddev->bitmap); 2515 close_sync(conf); 2516 return 0; 2517 } 2518 2519 if (mddev->bitmap == NULL && 2520 mddev->recovery_cp == MaxSector && 2521 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && 2522 conf->fullsync == 0) { 2523 *skipped = 1; 2524 return max_sector - sector_nr; 2525 } 2526 /* before building a request, check if we can skip these blocks.. 2527 * This call the bitmap_start_sync doesn't actually record anything 2528 */ 2529 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) && 2530 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2531 /* We can skip this block, and probably several more */ 2532 *skipped = 1; 2533 return sync_blocks; 2534 } 2535 /* 2536 * If there is non-resync activity waiting for a turn, 2537 * and resync is going fast enough, 2538 * then let it though before starting on this new sync request. 2539 */ 2540 if (!go_faster && conf->nr_waiting) 2541 msleep_interruptible(1000); 2542 2543 bitmap_cond_end_sync(mddev->bitmap, sector_nr); 2544 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO); 2545 raise_barrier(conf); 2546 2547 conf->next_resync = sector_nr; 2548 2549 rcu_read_lock(); 2550 /* 2551 * If we get a correctably read error during resync or recovery, 2552 * we might want to read from a different device. So we 2553 * flag all drives that could conceivably be read from for READ, 2554 * and any others (which will be non-In_sync devices) for WRITE. 2555 * If a read fails, we try reading from something else for which READ 2556 * is OK. 2557 */ 2558 2559 r1_bio->mddev = mddev; 2560 r1_bio->sector = sector_nr; 2561 r1_bio->state = 0; 2562 set_bit(R1BIO_IsSync, &r1_bio->state); 2563 2564 for (i = 0; i < conf->raid_disks * 2; i++) { 2565 struct md_rdev *rdev; 2566 bio = r1_bio->bios[i]; 2567 bio_reset(bio); 2568 2569 rdev = rcu_dereference(conf->mirrors[i].rdev); 2570 if (rdev == NULL || 2571 test_bit(Faulty, &rdev->flags)) { 2572 if (i < conf->raid_disks) 2573 still_degraded = 1; 2574 } else if (!test_bit(In_sync, &rdev->flags)) { 2575 bio->bi_rw = WRITE; 2576 bio->bi_end_io = end_sync_write; 2577 write_targets ++; 2578 } else { 2579 /* may need to read from here */ 2580 sector_t first_bad = MaxSector; 2581 int bad_sectors; 2582 2583 if (is_badblock(rdev, sector_nr, good_sectors, 2584 &first_bad, &bad_sectors)) { 2585 if (first_bad > sector_nr) 2586 good_sectors = first_bad - sector_nr; 2587 else { 2588 bad_sectors -= (sector_nr - first_bad); 2589 if (min_bad == 0 || 2590 min_bad > bad_sectors) 2591 min_bad = bad_sectors; 2592 } 2593 } 2594 if (sector_nr < first_bad) { 2595 if (test_bit(WriteMostly, &rdev->flags)) { 2596 if (wonly < 0) 2597 wonly = i; 2598 } else { 2599 if (disk < 0) 2600 disk = i; 2601 } 2602 bio->bi_rw = READ; 2603 bio->bi_end_io = end_sync_read; 2604 read_targets++; 2605 } else if (!test_bit(WriteErrorSeen, &rdev->flags) && 2606 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && 2607 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) { 2608 /* 2609 * The device is suitable for reading (InSync), 2610 * but has bad block(s) here. Let's try to correct them, 2611 * if we are doing resync or repair. Otherwise, leave 2612 * this device alone for this sync request. 2613 */ 2614 bio->bi_rw = WRITE; 2615 bio->bi_end_io = end_sync_write; 2616 write_targets++; 2617 } 2618 } 2619 if (bio->bi_end_io) { 2620 atomic_inc(&rdev->nr_pending); 2621 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset; 2622 bio->bi_bdev = rdev->bdev; 2623 bio->bi_private = r1_bio; 2624 } 2625 } 2626 rcu_read_unlock(); 2627 if (disk < 0) 2628 disk = wonly; 2629 r1_bio->read_disk = disk; 2630 2631 if (read_targets == 0 && min_bad > 0) { 2632 /* These sectors are bad on all InSync devices, so we 2633 * need to mark them bad on all write targets 2634 */ 2635 int ok = 1; 2636 for (i = 0 ; i < conf->raid_disks * 2 ; i++) 2637 if (r1_bio->bios[i]->bi_end_io == end_sync_write) { 2638 struct md_rdev *rdev = conf->mirrors[i].rdev; 2639 ok = rdev_set_badblocks(rdev, sector_nr, 2640 min_bad, 0 2641 ) && ok; 2642 } 2643 set_bit(MD_CHANGE_DEVS, &mddev->flags); 2644 *skipped = 1; 2645 put_buf(r1_bio); 2646 2647 if (!ok) { 2648 /* Cannot record the badblocks, so need to 2649 * abort the resync. 2650 * If there are multiple read targets, could just 2651 * fail the really bad ones ??? 2652 */ 2653 conf->recovery_disabled = mddev->recovery_disabled; 2654 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 2655 return 0; 2656 } else 2657 return min_bad; 2658 2659 } 2660 if (min_bad > 0 && min_bad < good_sectors) { 2661 /* only resync enough to reach the next bad->good 2662 * transition */ 2663 good_sectors = min_bad; 2664 } 2665 2666 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0) 2667 /* extra read targets are also write targets */ 2668 write_targets += read_targets-1; 2669 2670 if (write_targets == 0 || read_targets == 0) { 2671 /* There is nowhere to write, so all non-sync 2672 * drives must be failed - so we are finished 2673 */ 2674 sector_t rv; 2675 if (min_bad > 0) 2676 max_sector = sector_nr + min_bad; 2677 rv = max_sector - sector_nr; 2678 *skipped = 1; 2679 put_buf(r1_bio); 2680 return rv; 2681 } 2682 2683 if (max_sector > mddev->resync_max) 2684 max_sector = mddev->resync_max; /* Don't do IO beyond here */ 2685 if (max_sector > sector_nr + good_sectors) 2686 max_sector = sector_nr + good_sectors; 2687 nr_sectors = 0; 2688 sync_blocks = 0; 2689 do { 2690 struct page *page; 2691 int len = PAGE_SIZE; 2692 if (sector_nr + (len>>9) > max_sector) 2693 len = (max_sector - sector_nr) << 9; 2694 if (len == 0) 2695 break; 2696 if (sync_blocks == 0) { 2697 if (!bitmap_start_sync(mddev->bitmap, sector_nr, 2698 &sync_blocks, still_degraded) && 2699 !conf->fullsync && 2700 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) 2701 break; 2702 BUG_ON(sync_blocks < (PAGE_SIZE>>9)); 2703 if ((len >> 9) > sync_blocks) 2704 len = sync_blocks<<9; 2705 } 2706 2707 for (i = 0 ; i < conf->raid_disks * 2; i++) { 2708 bio = r1_bio->bios[i]; 2709 if (bio->bi_end_io) { 2710 page = bio->bi_io_vec[bio->bi_vcnt].bv_page; 2711 if (bio_add_page(bio, page, len, 0) == 0) { 2712 /* stop here */ 2713 bio->bi_io_vec[bio->bi_vcnt].bv_page = page; 2714 while (i > 0) { 2715 i--; 2716 bio = r1_bio->bios[i]; 2717 if (bio->bi_end_io==NULL) 2718 continue; 2719 /* remove last page from this bio */ 2720 bio->bi_vcnt--; 2721 bio->bi_iter.bi_size -= len; 2722 bio->bi_flags &= ~(1<< BIO_SEG_VALID); 2723 } 2724 goto bio_full; 2725 } 2726 } 2727 } 2728 nr_sectors += len>>9; 2729 sector_nr += len>>9; 2730 sync_blocks -= (len>>9); 2731 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES); 2732 bio_full: 2733 r1_bio->sectors = nr_sectors; 2734 2735 /* For a user-requested sync, we read all readable devices and do a 2736 * compare 2737 */ 2738 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2739 atomic_set(&r1_bio->remaining, read_targets); 2740 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) { 2741 bio = r1_bio->bios[i]; 2742 if (bio->bi_end_io == end_sync_read) { 2743 read_targets--; 2744 md_sync_acct(bio->bi_bdev, nr_sectors); 2745 generic_make_request(bio); 2746 } 2747 } 2748 } else { 2749 atomic_set(&r1_bio->remaining, 1); 2750 bio = r1_bio->bios[r1_bio->read_disk]; 2751 md_sync_acct(bio->bi_bdev, nr_sectors); 2752 generic_make_request(bio); 2753 2754 } 2755 return nr_sectors; 2756 } 2757 2758 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks) 2759 { 2760 if (sectors) 2761 return sectors; 2762 2763 return mddev->dev_sectors; 2764 } 2765 2766 static struct r1conf *setup_conf(struct mddev *mddev) 2767 { 2768 struct r1conf *conf; 2769 int i; 2770 struct raid1_info *disk; 2771 struct md_rdev *rdev; 2772 int err = -ENOMEM; 2773 2774 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL); 2775 if (!conf) 2776 goto abort; 2777 2778 conf->mirrors = kzalloc(sizeof(struct raid1_info) 2779 * mddev->raid_disks * 2, 2780 GFP_KERNEL); 2781 if (!conf->mirrors) 2782 goto abort; 2783 2784 conf->tmppage = alloc_page(GFP_KERNEL); 2785 if (!conf->tmppage) 2786 goto abort; 2787 2788 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL); 2789 if (!conf->poolinfo) 2790 goto abort; 2791 conf->poolinfo->raid_disks = mddev->raid_disks * 2; 2792 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 2793 r1bio_pool_free, 2794 conf->poolinfo); 2795 if (!conf->r1bio_pool) 2796 goto abort; 2797 2798 conf->poolinfo->mddev = mddev; 2799 2800 err = -EINVAL; 2801 spin_lock_init(&conf->device_lock); 2802 rdev_for_each(rdev, mddev) { 2803 struct request_queue *q; 2804 int disk_idx = rdev->raid_disk; 2805 if (disk_idx >= mddev->raid_disks 2806 || disk_idx < 0) 2807 continue; 2808 if (test_bit(Replacement, &rdev->flags)) 2809 disk = conf->mirrors + mddev->raid_disks + disk_idx; 2810 else 2811 disk = conf->mirrors + disk_idx; 2812 2813 if (disk->rdev) 2814 goto abort; 2815 disk->rdev = rdev; 2816 q = bdev_get_queue(rdev->bdev); 2817 if (q->merge_bvec_fn) 2818 mddev->merge_check_needed = 1; 2819 2820 disk->head_position = 0; 2821 disk->seq_start = MaxSector; 2822 } 2823 conf->raid_disks = mddev->raid_disks; 2824 conf->mddev = mddev; 2825 INIT_LIST_HEAD(&conf->retry_list); 2826 2827 spin_lock_init(&conf->resync_lock); 2828 init_waitqueue_head(&conf->wait_barrier); 2829 2830 bio_list_init(&conf->pending_bio_list); 2831 conf->pending_count = 0; 2832 conf->recovery_disabled = mddev->recovery_disabled - 1; 2833 2834 conf->start_next_window = MaxSector; 2835 conf->current_window_requests = conf->next_window_requests = 0; 2836 2837 err = -EIO; 2838 for (i = 0; i < conf->raid_disks * 2; i++) { 2839 2840 disk = conf->mirrors + i; 2841 2842 if (i < conf->raid_disks && 2843 disk[conf->raid_disks].rdev) { 2844 /* This slot has a replacement. */ 2845 if (!disk->rdev) { 2846 /* No original, just make the replacement 2847 * a recovering spare 2848 */ 2849 disk->rdev = 2850 disk[conf->raid_disks].rdev; 2851 disk[conf->raid_disks].rdev = NULL; 2852 } else if (!test_bit(In_sync, &disk->rdev->flags)) 2853 /* Original is not in_sync - bad */ 2854 goto abort; 2855 } 2856 2857 if (!disk->rdev || 2858 !test_bit(In_sync, &disk->rdev->flags)) { 2859 disk->head_position = 0; 2860 if (disk->rdev && 2861 (disk->rdev->saved_raid_disk < 0)) 2862 conf->fullsync = 1; 2863 } 2864 } 2865 2866 err = -ENOMEM; 2867 conf->thread = md_register_thread(raid1d, mddev, "raid1"); 2868 if (!conf->thread) { 2869 printk(KERN_ERR 2870 "md/raid1:%s: couldn't allocate thread\n", 2871 mdname(mddev)); 2872 goto abort; 2873 } 2874 2875 return conf; 2876 2877 abort: 2878 if (conf) { 2879 if (conf->r1bio_pool) 2880 mempool_destroy(conf->r1bio_pool); 2881 kfree(conf->mirrors); 2882 safe_put_page(conf->tmppage); 2883 kfree(conf->poolinfo); 2884 kfree(conf); 2885 } 2886 return ERR_PTR(err); 2887 } 2888 2889 static int stop(struct mddev *mddev); 2890 static int run(struct mddev *mddev) 2891 { 2892 struct r1conf *conf; 2893 int i; 2894 struct md_rdev *rdev; 2895 int ret; 2896 bool discard_supported = false; 2897 2898 if (mddev->level != 1) { 2899 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n", 2900 mdname(mddev), mddev->level); 2901 return -EIO; 2902 } 2903 if (mddev->reshape_position != MaxSector) { 2904 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n", 2905 mdname(mddev)); 2906 return -EIO; 2907 } 2908 /* 2909 * copy the already verified devices into our private RAID1 2910 * bookkeeping area. [whatever we allocate in run(), 2911 * should be freed in stop()] 2912 */ 2913 if (mddev->private == NULL) 2914 conf = setup_conf(mddev); 2915 else 2916 conf = mddev->private; 2917 2918 if (IS_ERR(conf)) 2919 return PTR_ERR(conf); 2920 2921 if (mddev->queue) 2922 blk_queue_max_write_same_sectors(mddev->queue, 0); 2923 2924 rdev_for_each(rdev, mddev) { 2925 if (!mddev->gendisk) 2926 continue; 2927 disk_stack_limits(mddev->gendisk, rdev->bdev, 2928 rdev->data_offset << 9); 2929 if (blk_queue_discard(bdev_get_queue(rdev->bdev))) 2930 discard_supported = true; 2931 } 2932 2933 mddev->degraded = 0; 2934 for (i=0; i < conf->raid_disks; i++) 2935 if (conf->mirrors[i].rdev == NULL || 2936 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) || 2937 test_bit(Faulty, &conf->mirrors[i].rdev->flags)) 2938 mddev->degraded++; 2939 2940 if (conf->raid_disks - mddev->degraded == 1) 2941 mddev->recovery_cp = MaxSector; 2942 2943 if (mddev->recovery_cp != MaxSector) 2944 printk(KERN_NOTICE "md/raid1:%s: not clean" 2945 " -- starting background reconstruction\n", 2946 mdname(mddev)); 2947 printk(KERN_INFO 2948 "md/raid1:%s: active with %d out of %d mirrors\n", 2949 mdname(mddev), mddev->raid_disks - mddev->degraded, 2950 mddev->raid_disks); 2951 2952 /* 2953 * Ok, everything is just fine now 2954 */ 2955 mddev->thread = conf->thread; 2956 conf->thread = NULL; 2957 mddev->private = conf; 2958 2959 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0)); 2960 2961 if (mddev->queue) { 2962 mddev->queue->backing_dev_info.congested_fn = raid1_congested; 2963 mddev->queue->backing_dev_info.congested_data = mddev; 2964 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec); 2965 2966 if (discard_supported) 2967 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, 2968 mddev->queue); 2969 else 2970 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, 2971 mddev->queue); 2972 } 2973 2974 ret = md_integrity_register(mddev); 2975 if (ret) 2976 stop(mddev); 2977 return ret; 2978 } 2979 2980 static int stop(struct mddev *mddev) 2981 { 2982 struct r1conf *conf = mddev->private; 2983 struct bitmap *bitmap = mddev->bitmap; 2984 2985 /* wait for behind writes to complete */ 2986 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) { 2987 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n", 2988 mdname(mddev)); 2989 /* need to kick something here to make sure I/O goes? */ 2990 wait_event(bitmap->behind_wait, 2991 atomic_read(&bitmap->behind_writes) == 0); 2992 } 2993 2994 freeze_array(conf, 0); 2995 unfreeze_array(conf); 2996 2997 md_unregister_thread(&mddev->thread); 2998 if (conf->r1bio_pool) 2999 mempool_destroy(conf->r1bio_pool); 3000 kfree(conf->mirrors); 3001 safe_put_page(conf->tmppage); 3002 kfree(conf->poolinfo); 3003 kfree(conf); 3004 mddev->private = NULL; 3005 return 0; 3006 } 3007 3008 static int raid1_resize(struct mddev *mddev, sector_t sectors) 3009 { 3010 /* no resync is happening, and there is enough space 3011 * on all devices, so we can resize. 3012 * We need to make sure resync covers any new space. 3013 * If the array is shrinking we should possibly wait until 3014 * any io in the removed space completes, but it hardly seems 3015 * worth it. 3016 */ 3017 sector_t newsize = raid1_size(mddev, sectors, 0); 3018 if (mddev->external_size && 3019 mddev->array_sectors > newsize) 3020 return -EINVAL; 3021 if (mddev->bitmap) { 3022 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0); 3023 if (ret) 3024 return ret; 3025 } 3026 md_set_array_sectors(mddev, newsize); 3027 set_capacity(mddev->gendisk, mddev->array_sectors); 3028 revalidate_disk(mddev->gendisk); 3029 if (sectors > mddev->dev_sectors && 3030 mddev->recovery_cp > mddev->dev_sectors) { 3031 mddev->recovery_cp = mddev->dev_sectors; 3032 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3033 } 3034 mddev->dev_sectors = sectors; 3035 mddev->resync_max_sectors = sectors; 3036 return 0; 3037 } 3038 3039 static int raid1_reshape(struct mddev *mddev) 3040 { 3041 /* We need to: 3042 * 1/ resize the r1bio_pool 3043 * 2/ resize conf->mirrors 3044 * 3045 * We allocate a new r1bio_pool if we can. 3046 * Then raise a device barrier and wait until all IO stops. 3047 * Then resize conf->mirrors and swap in the new r1bio pool. 3048 * 3049 * At the same time, we "pack" the devices so that all the missing 3050 * devices have the higher raid_disk numbers. 3051 */ 3052 mempool_t *newpool, *oldpool; 3053 struct pool_info *newpoolinfo; 3054 struct raid1_info *newmirrors; 3055 struct r1conf *conf = mddev->private; 3056 int cnt, raid_disks; 3057 unsigned long flags; 3058 int d, d2, err; 3059 3060 /* Cannot change chunk_size, layout, or level */ 3061 if (mddev->chunk_sectors != mddev->new_chunk_sectors || 3062 mddev->layout != mddev->new_layout || 3063 mddev->level != mddev->new_level) { 3064 mddev->new_chunk_sectors = mddev->chunk_sectors; 3065 mddev->new_layout = mddev->layout; 3066 mddev->new_level = mddev->level; 3067 return -EINVAL; 3068 } 3069 3070 err = md_allow_write(mddev); 3071 if (err) 3072 return err; 3073 3074 raid_disks = mddev->raid_disks + mddev->delta_disks; 3075 3076 if (raid_disks < conf->raid_disks) { 3077 cnt=0; 3078 for (d= 0; d < conf->raid_disks; d++) 3079 if (conf->mirrors[d].rdev) 3080 cnt++; 3081 if (cnt > raid_disks) 3082 return -EBUSY; 3083 } 3084 3085 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL); 3086 if (!newpoolinfo) 3087 return -ENOMEM; 3088 newpoolinfo->mddev = mddev; 3089 newpoolinfo->raid_disks = raid_disks * 2; 3090 3091 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, 3092 r1bio_pool_free, newpoolinfo); 3093 if (!newpool) { 3094 kfree(newpoolinfo); 3095 return -ENOMEM; 3096 } 3097 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2, 3098 GFP_KERNEL); 3099 if (!newmirrors) { 3100 kfree(newpoolinfo); 3101 mempool_destroy(newpool); 3102 return -ENOMEM; 3103 } 3104 3105 freeze_array(conf, 0); 3106 3107 /* ok, everything is stopped */ 3108 oldpool = conf->r1bio_pool; 3109 conf->r1bio_pool = newpool; 3110 3111 for (d = d2 = 0; d < conf->raid_disks; d++) { 3112 struct md_rdev *rdev = conf->mirrors[d].rdev; 3113 if (rdev && rdev->raid_disk != d2) { 3114 sysfs_unlink_rdev(mddev, rdev); 3115 rdev->raid_disk = d2; 3116 sysfs_unlink_rdev(mddev, rdev); 3117 if (sysfs_link_rdev(mddev, rdev)) 3118 printk(KERN_WARNING 3119 "md/raid1:%s: cannot register rd%d\n", 3120 mdname(mddev), rdev->raid_disk); 3121 } 3122 if (rdev) 3123 newmirrors[d2++].rdev = rdev; 3124 } 3125 kfree(conf->mirrors); 3126 conf->mirrors = newmirrors; 3127 kfree(conf->poolinfo); 3128 conf->poolinfo = newpoolinfo; 3129 3130 spin_lock_irqsave(&conf->device_lock, flags); 3131 mddev->degraded += (raid_disks - conf->raid_disks); 3132 spin_unlock_irqrestore(&conf->device_lock, flags); 3133 conf->raid_disks = mddev->raid_disks = raid_disks; 3134 mddev->delta_disks = 0; 3135 3136 unfreeze_array(conf); 3137 3138 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 3139 md_wakeup_thread(mddev->thread); 3140 3141 mempool_destroy(oldpool); 3142 return 0; 3143 } 3144 3145 static void raid1_quiesce(struct mddev *mddev, int state) 3146 { 3147 struct r1conf *conf = mddev->private; 3148 3149 switch(state) { 3150 case 2: /* wake for suspend */ 3151 wake_up(&conf->wait_barrier); 3152 break; 3153 case 1: 3154 freeze_array(conf, 0); 3155 break; 3156 case 0: 3157 unfreeze_array(conf); 3158 break; 3159 } 3160 } 3161 3162 static void *raid1_takeover(struct mddev *mddev) 3163 { 3164 /* raid1 can take over: 3165 * raid5 with 2 devices, any layout or chunk size 3166 */ 3167 if (mddev->level == 5 && mddev->raid_disks == 2) { 3168 struct r1conf *conf; 3169 mddev->new_level = 1; 3170 mddev->new_layout = 0; 3171 mddev->new_chunk_sectors = 0; 3172 conf = setup_conf(mddev); 3173 if (!IS_ERR(conf)) 3174 /* Array must appear to be quiesced */ 3175 conf->array_frozen = 1; 3176 return conf; 3177 } 3178 return ERR_PTR(-EINVAL); 3179 } 3180 3181 static struct md_personality raid1_personality = 3182 { 3183 .name = "raid1", 3184 .level = 1, 3185 .owner = THIS_MODULE, 3186 .make_request = make_request, 3187 .run = run, 3188 .stop = stop, 3189 .status = status, 3190 .error_handler = error, 3191 .hot_add_disk = raid1_add_disk, 3192 .hot_remove_disk= raid1_remove_disk, 3193 .spare_active = raid1_spare_active, 3194 .sync_request = sync_request, 3195 .resize = raid1_resize, 3196 .size = raid1_size, 3197 .check_reshape = raid1_reshape, 3198 .quiesce = raid1_quiesce, 3199 .takeover = raid1_takeover, 3200 }; 3201 3202 static int __init raid_init(void) 3203 { 3204 return register_md_personality(&raid1_personality); 3205 } 3206 3207 static void raid_exit(void) 3208 { 3209 unregister_md_personality(&raid1_personality); 3210 } 3211 3212 module_init(raid_init); 3213 module_exit(raid_exit); 3214 MODULE_LICENSE("GPL"); 3215 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD"); 3216 MODULE_ALIAS("md-personality-3"); /* RAID1 */ 3217 MODULE_ALIAS("md-raid1"); 3218 MODULE_ALIAS("md-level-1"); 3219 3220 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR); 3221