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