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