1/* 2 * raid10.c : Multiple Devices driver for Linux 3 * 4 * Copyright (C) 2000-2004 Neil Brown 5 * 6 * RAID-10 support for md. 7 * 8 * Base on code in raid1.c. See raid1.c for futher copyright information. 9 * 10 * 11 * This program is free software; you can redistribute it and/or modify 12 * it under the terms of the GNU General Public License as published by 13 * the Free Software Foundation; either version 2, or (at your option) 14 * any later version. 15 * 16 * You should have received a copy of the GNU General Public License 17 * (for example /usr/src/linux/COPYING); if not, write to the Free 18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. 19 */ 20 21#include <linux/slab.h> 22#include <linux/delay.h> 23#include <linux/blkdev.h> 24#include <linux/seq_file.h> 25#include "md.h" 26#include "raid10.h" 27#include "raid0.h" 28#include "bitmap.h" 29 30/* 31 * RAID10 provides a combination of RAID0 and RAID1 functionality. 32 * The layout of data is defined by 33 * chunk_size 34 * raid_disks 35 * near_copies (stored in low byte of layout) 36 * far_copies (stored in second byte of layout) 37 * far_offset (stored in bit 16 of layout ) 38 * 39 * The data to be stored is divided into chunks using chunksize. 40 * Each device is divided into far_copies sections. 41 * In each section, chunks are laid out in a style similar to raid0, but 42 * near_copies copies of each chunk is stored (each on a different drive). 43 * The starting device for each section is offset near_copies from the starting 44 * device of the previous section. 45 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different 46 * drive. 47 * near_copies and far_copies must be at least one, and their product is at most 48 * raid_disks. 49 * 50 * If far_offset is true, then the far_copies are handled a bit differently. 51 * The copies are still in different stripes, but instead of be very far apart 52 * on disk, there are adjacent stripes. 53 */ 54 55/* 56 * Number of guaranteed r10bios in case of extreme VM load: 57 */ 58#define NR_RAID10_BIOS 256 59 60static void unplug_slaves(mddev_t *mddev); 61 62static void allow_barrier(conf_t *conf); 63static void lower_barrier(conf_t *conf); 64 65static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data) 66{ 67 conf_t *conf = data; 68 r10bio_t *r10_bio; 69 int size = offsetof(struct r10bio_s, devs[conf->copies]); 70 71 /* allocate a r10bio with room for raid_disks entries in the bios array */ 72 r10_bio = kzalloc(size, gfp_flags); 73 if (!r10_bio && conf->mddev) 74 unplug_slaves(conf->mddev); 75 76 return r10_bio; 77} 78 79static void r10bio_pool_free(void *r10_bio, void *data) 80{ 81 kfree(r10_bio); 82} 83 84/* Maximum size of each resync request */ 85#define RESYNC_BLOCK_SIZE (64*1024) 86#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE) 87/* amount of memory to reserve for resync requests */ 88#define RESYNC_WINDOW (1024*1024) 89/* maximum number of concurrent requests, memory permitting */ 90#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE) 91 92/* 93 * When performing a resync, we need to read and compare, so 94 * we need as many pages are there are copies. 95 * When performing a recovery, we need 2 bios, one for read, 96 * one for write (we recover only one drive per r10buf) 97 * 98 */ 99static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data) 100{ 101 conf_t *conf = data; 102 struct page *page; 103 r10bio_t *r10_bio; 104 struct bio *bio; 105 int i, j; 106 int nalloc; 107 108 r10_bio = r10bio_pool_alloc(gfp_flags, conf); 109 if (!r10_bio) { 110 unplug_slaves(conf->mddev); 111 return NULL; 112 } 113 114 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery)) 115 nalloc = conf->copies; /* resync */ 116 else 117 nalloc = 2; /* recovery */ 118 119 /* 120 * Allocate bios. 121 */ 122 for (j = nalloc ; j-- ; ) { 123 bio = bio_alloc(gfp_flags, RESYNC_PAGES); 124 if (!bio) 125 goto out_free_bio; 126 r10_bio->devs[j].bio = bio; 127 } 128 /* 129 * Allocate RESYNC_PAGES data pages and attach them 130 * where needed. 131 */ 132 for (j = 0 ; j < nalloc; j++) { 133 bio = r10_bio->devs[j].bio; 134 for (i = 0; i < RESYNC_PAGES; i++) { 135 page = alloc_page(gfp_flags); 136 if (unlikely(!page)) 137 goto out_free_pages; 138 139 bio->bi_io_vec[i].bv_page = page; 140 } 141 } 142 143 return r10_bio; 144 145out_free_pages: 146 for ( ; i > 0 ; i--) 147 safe_put_page(bio->bi_io_vec[i-1].bv_page); 148 while (j--) 149 for (i = 0; i < RESYNC_PAGES ; i++) 150 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page); 151 j = -1; 152out_free_bio: 153 while ( ++j < nalloc ) 154 bio_put(r10_bio->devs[j].bio); 155 r10bio_pool_free(r10_bio, conf); 156 return NULL; 157} 158 159static void r10buf_pool_free(void *__r10_bio, void *data) 160{ 161 int i; 162 conf_t *conf = data; 163 r10bio_t *r10bio = __r10_bio; 164 int j; 165 166 for (j=0; j < conf->copies; j++) { 167 struct bio *bio = r10bio->devs[j].bio; 168 if (bio) { 169 for (i = 0; i < RESYNC_PAGES; i++) { 170 safe_put_page(bio->bi_io_vec[i].bv_page); 171 bio->bi_io_vec[i].bv_page = NULL; 172 } 173 bio_put(bio); 174 } 175 } 176 r10bio_pool_free(r10bio, conf); 177} 178 179static void put_all_bios(conf_t *conf, r10bio_t *r10_bio) 180{ 181 int i; 182 183 for (i = 0; i < conf->copies; i++) { 184 struct bio **bio = & r10_bio->devs[i].bio; 185 if (*bio && *bio != IO_BLOCKED) 186 bio_put(*bio); 187 *bio = NULL; 188 } 189} 190 191static void free_r10bio(r10bio_t *r10_bio) 192{ 193 conf_t *conf = r10_bio->mddev->private; 194 195 /* 196 * Wake up any possible resync thread that waits for the device 197 * to go idle. 198 */ 199 allow_barrier(conf); 200 201 put_all_bios(conf, r10_bio); 202 mempool_free(r10_bio, conf->r10bio_pool); 203} 204 205static void put_buf(r10bio_t *r10_bio) 206{ 207 conf_t *conf = r10_bio->mddev->private; 208 209 mempool_free(r10_bio, conf->r10buf_pool); 210 211 lower_barrier(conf); 212} 213 214static void reschedule_retry(r10bio_t *r10_bio) 215{ 216 unsigned long flags; 217 mddev_t *mddev = r10_bio->mddev; 218 conf_t *conf = mddev->private; 219 220 spin_lock_irqsave(&conf->device_lock, flags); 221 list_add(&r10_bio->retry_list, &conf->retry_list); 222 conf->nr_queued ++; 223 spin_unlock_irqrestore(&conf->device_lock, flags); 224 225 /* wake up frozen array... */ 226 wake_up(&conf->wait_barrier); 227 228 md_wakeup_thread(mddev->thread); 229} 230 231/* 232 * raid_end_bio_io() is called when we have finished servicing a mirrored 233 * operation and are ready to return a success/failure code to the buffer 234 * cache layer. 235 */ 236static void raid_end_bio_io(r10bio_t *r10_bio) 237{ 238 struct bio *bio = r10_bio->master_bio; 239 240 bio_endio(bio, 241 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO); 242 free_r10bio(r10_bio); 243} 244 245/* 246 * Update disk head position estimator based on IRQ completion info. 247 */ 248static inline void update_head_pos(int slot, r10bio_t *r10_bio) 249{ 250 conf_t *conf = r10_bio->mddev->private; 251 252 conf->mirrors[r10_bio->devs[slot].devnum].head_position = 253 r10_bio->devs[slot].addr + (r10_bio->sectors); 254} 255 256static void raid10_end_read_request(struct bio *bio, int error) 257{ 258 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 259 r10bio_t *r10_bio = bio->bi_private; 260 int slot, dev; 261 conf_t *conf = r10_bio->mddev->private; 262 263 264 slot = r10_bio->read_slot; 265 dev = r10_bio->devs[slot].devnum; 266 /* 267 * this branch is our 'one mirror IO has finished' event handler: 268 */ 269 update_head_pos(slot, r10_bio); 270 271 if (uptodate) { 272 /* 273 * Set R10BIO_Uptodate in our master bio, so that 274 * we will return a good error code to the higher 275 * levels even if IO on some other mirrored buffer fails. 276 * 277 * The 'master' represents the composite IO operation to 278 * user-side. So if something waits for IO, then it will 279 * wait for the 'master' bio. 280 */ 281 set_bit(R10BIO_Uptodate, &r10_bio->state); 282 raid_end_bio_io(r10_bio); 283 } else { 284 /* 285 * oops, read error: 286 */ 287 char b[BDEVNAME_SIZE]; 288 if (printk_ratelimit()) 289 printk(KERN_ERR "md/raid10:%s: %s: rescheduling sector %llu\n", 290 mdname(conf->mddev), 291 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector); 292 reschedule_retry(r10_bio); 293 } 294 295 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev); 296} 297 298static void raid10_end_write_request(struct bio *bio, int error) 299{ 300 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 301 r10bio_t *r10_bio = bio->bi_private; 302 int slot, dev; 303 conf_t *conf = r10_bio->mddev->private; 304 305 for (slot = 0; slot < conf->copies; slot++) 306 if (r10_bio->devs[slot].bio == bio) 307 break; 308 dev = r10_bio->devs[slot].devnum; 309 310 /* 311 * this branch is our 'one mirror IO has finished' event handler: 312 */ 313 if (!uptodate) { 314 md_error(r10_bio->mddev, conf->mirrors[dev].rdev); 315 /* an I/O failed, we can't clear the bitmap */ 316 set_bit(R10BIO_Degraded, &r10_bio->state); 317 } else 318 /* 319 * Set R10BIO_Uptodate in our master bio, so that 320 * we will return a good error code for to the higher 321 * levels even if IO on some other mirrored buffer fails. 322 * 323 * The 'master' represents the composite IO operation to 324 * user-side. So if something waits for IO, then it will 325 * wait for the 'master' bio. 326 */ 327 set_bit(R10BIO_Uptodate, &r10_bio->state); 328 329 update_head_pos(slot, r10_bio); 330 331 /* 332 * 333 * Let's see if all mirrored write operations have finished 334 * already. 335 */ 336 if (atomic_dec_and_test(&r10_bio->remaining)) { 337 /* clear the bitmap if all writes complete successfully */ 338 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector, 339 r10_bio->sectors, 340 !test_bit(R10BIO_Degraded, &r10_bio->state), 341 0); 342 md_write_end(r10_bio->mddev); 343 raid_end_bio_io(r10_bio); 344 } 345 346 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev); 347} 348 349 350/* 351 * RAID10 layout manager 352 * Aswell as the chunksize and raid_disks count, there are two 353 * parameters: near_copies and far_copies. 354 * near_copies * far_copies must be <= raid_disks. 355 * Normally one of these will be 1. 356 * If both are 1, we get raid0. 357 * If near_copies == raid_disks, we get raid1. 358 * 359 * Chunks are layed out in raid0 style with near_copies copies of the 360 * first chunk, followed by near_copies copies of the next chunk and 361 * so on. 362 * If far_copies > 1, then after 1/far_copies of the array has been assigned 363 * as described above, we start again with a device offset of near_copies. 364 * So we effectively have another copy of the whole array further down all 365 * the drives, but with blocks on different drives. 366 * With this layout, and block is never stored twice on the one device. 367 * 368 * raid10_find_phys finds the sector offset of a given virtual sector 369 * on each device that it is on. 370 * 371 * raid10_find_virt does the reverse mapping, from a device and a 372 * sector offset to a virtual address 373 */ 374 375static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio) 376{ 377 int n,f; 378 sector_t sector; 379 sector_t chunk; 380 sector_t stripe; 381 int dev; 382 383 int slot = 0; 384 385 /* now calculate first sector/dev */ 386 chunk = r10bio->sector >> conf->chunk_shift; 387 sector = r10bio->sector & conf->chunk_mask; 388 389 chunk *= conf->near_copies; 390 stripe = chunk; 391 dev = sector_div(stripe, conf->raid_disks); 392 if (conf->far_offset) 393 stripe *= conf->far_copies; 394 395 sector += stripe << conf->chunk_shift; 396 397 /* and calculate all the others */ 398 for (n=0; n < conf->near_copies; n++) { 399 int d = dev; 400 sector_t s = sector; 401 r10bio->devs[slot].addr = sector; 402 r10bio->devs[slot].devnum = d; 403 slot++; 404 405 for (f = 1; f < conf->far_copies; f++) { 406 d += conf->near_copies; 407 if (d >= conf->raid_disks) 408 d -= conf->raid_disks; 409 s += conf->stride; 410 r10bio->devs[slot].devnum = d; 411 r10bio->devs[slot].addr = s; 412 slot++; 413 } 414 dev++; 415 if (dev >= conf->raid_disks) { 416 dev = 0; 417 sector += (conf->chunk_mask + 1); 418 } 419 } 420 BUG_ON(slot != conf->copies); 421} 422 423static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev) 424{ 425 sector_t offset, chunk, vchunk; 426 427 offset = sector & conf->chunk_mask; 428 if (conf->far_offset) { 429 int fc; 430 chunk = sector >> conf->chunk_shift; 431 fc = sector_div(chunk, conf->far_copies); 432 dev -= fc * conf->near_copies; 433 if (dev < 0) 434 dev += conf->raid_disks; 435 } else { 436 while (sector >= conf->stride) { 437 sector -= conf->stride; 438 if (dev < conf->near_copies) 439 dev += conf->raid_disks - conf->near_copies; 440 else 441 dev -= conf->near_copies; 442 } 443 chunk = sector >> conf->chunk_shift; 444 } 445 vchunk = chunk * conf->raid_disks + dev; 446 sector_div(vchunk, conf->near_copies); 447 return (vchunk << conf->chunk_shift) + offset; 448} 449 450/** 451 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged 452 * @q: request queue 453 * @bvm: properties of new bio 454 * @biovec: the request that could be merged to it. 455 * 456 * Return amount of bytes we can accept at this offset 457 * If near_copies == raid_disk, there are no striping issues, 458 * but in that case, the function isn't called at all. 459 */ 460static int raid10_mergeable_bvec(struct request_queue *q, 461 struct bvec_merge_data *bvm, 462 struct bio_vec *biovec) 463{ 464 mddev_t *mddev = q->queuedata; 465 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev); 466 int max; 467 unsigned int chunk_sectors = mddev->chunk_sectors; 468 unsigned int bio_sectors = bvm->bi_size >> 9; 469 470 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9; 471 if (max < 0) max = 0; /* bio_add cannot handle a negative return */ 472 if (max <= biovec->bv_len && bio_sectors == 0) 473 return biovec->bv_len; 474 else 475 return max; 476} 477 478/* 479 * This routine returns the disk from which the requested read should 480 * be done. There is a per-array 'next expected sequential IO' sector 481 * number - if this matches on the next IO then we use the last disk. 482 * There is also a per-disk 'last know head position' sector that is 483 * maintained from IRQ contexts, both the normal and the resync IO 484 * completion handlers update this position correctly. If there is no 485 * perfect sequential match then we pick the disk whose head is closest. 486 * 487 * If there are 2 mirrors in the same 2 devices, performance degrades 488 * because position is mirror, not device based. 489 * 490 * The rdev for the device selected will have nr_pending incremented. 491 */ 492 493static int read_balance(conf_t *conf, r10bio_t *r10_bio) 494{ 495 const sector_t this_sector = r10_bio->sector; 496 int disk, slot, nslot; 497 const int sectors = r10_bio->sectors; 498 sector_t new_distance, current_distance; 499 mdk_rdev_t *rdev; 500 501 raid10_find_phys(conf, r10_bio); 502 rcu_read_lock(); 503 /* 504 * Check if we can balance. We can balance on the whole 505 * device if no resync is going on (recovery is ok), or below 506 * the resync window. We take the first readable disk when 507 * above the resync window. 508 */ 509 if (conf->mddev->recovery_cp < MaxSector 510 && (this_sector + sectors >= conf->next_resync)) { 511 /* make sure that disk is operational */ 512 slot = 0; 513 disk = r10_bio->devs[slot].devnum; 514 515 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL || 516 r10_bio->devs[slot].bio == IO_BLOCKED || 517 !test_bit(In_sync, &rdev->flags)) { 518 slot++; 519 if (slot == conf->copies) { 520 slot = 0; 521 disk = -1; 522 break; 523 } 524 disk = r10_bio->devs[slot].devnum; 525 } 526 goto rb_out; 527 } 528 529 530 /* make sure the disk is operational */ 531 slot = 0; 532 disk = r10_bio->devs[slot].devnum; 533 while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL || 534 r10_bio->devs[slot].bio == IO_BLOCKED || 535 !test_bit(In_sync, &rdev->flags)) { 536 slot ++; 537 if (slot == conf->copies) { 538 disk = -1; 539 goto rb_out; 540 } 541 disk = r10_bio->devs[slot].devnum; 542 } 543 544 545 current_distance = abs(r10_bio->devs[slot].addr - 546 conf->mirrors[disk].head_position); 547 548 /* Find the disk whose head is closest, 549 * or - for far > 1 - find the closest to partition beginning */ 550 551 for (nslot = slot; nslot < conf->copies; nslot++) { 552 int ndisk = r10_bio->devs[nslot].devnum; 553 554 555 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL || 556 r10_bio->devs[nslot].bio == IO_BLOCKED || 557 !test_bit(In_sync, &rdev->flags)) 558 continue; 559 560 /* This optimisation is debatable, and completely destroys 561 * sequential read speed for 'far copies' arrays. So only 562 * keep it for 'near' arrays, and review those later. 563 */ 564 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) { 565 disk = ndisk; 566 slot = nslot; 567 break; 568 } 569 570 /* for far > 1 always use the lowest address */ 571 if (conf->far_copies > 1) 572 new_distance = r10_bio->devs[nslot].addr; 573 else 574 new_distance = abs(r10_bio->devs[nslot].addr - 575 conf->mirrors[ndisk].head_position); 576 if (new_distance < current_distance) { 577 current_distance = new_distance; 578 disk = ndisk; 579 slot = nslot; 580 } 581 } 582 583rb_out: 584 r10_bio->read_slot = slot; 585/* conf->next_seq_sect = this_sector + sectors;*/ 586 587 if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL) 588 atomic_inc(&conf->mirrors[disk].rdev->nr_pending); 589 else 590 disk = -1; 591 rcu_read_unlock(); 592 593 return disk; 594} 595 596static void unplug_slaves(mddev_t *mddev) 597{ 598 conf_t *conf = mddev->private; 599 int i; 600 601 rcu_read_lock(); 602 for (i=0; i < conf->raid_disks; i++) { 603 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev); 604 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) { 605 struct request_queue *r_queue = bdev_get_queue(rdev->bdev); 606 607 atomic_inc(&rdev->nr_pending); 608 rcu_read_unlock(); 609 610 blk_unplug(r_queue); 611 612 rdev_dec_pending(rdev, mddev); 613 rcu_read_lock(); 614 } 615 } 616 rcu_read_unlock(); 617} 618 619static void raid10_unplug(struct request_queue *q) 620{ 621 mddev_t *mddev = q->queuedata; 622 623 unplug_slaves(q->queuedata); 624 md_wakeup_thread(mddev->thread); 625} 626 627static int raid10_congested(void *data, int bits) 628{ 629 mddev_t *mddev = data; 630 conf_t *conf = mddev->private; 631 int i, ret = 0; 632 633 if (mddev_congested(mddev, bits)) 634 return 1; 635 rcu_read_lock(); 636 for (i = 0; i < conf->raid_disks && ret == 0; i++) { 637 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev); 638 if (rdev && !test_bit(Faulty, &rdev->flags)) { 639 struct request_queue *q = bdev_get_queue(rdev->bdev); 640 641 ret |= bdi_congested(&q->backing_dev_info, bits); 642 } 643 } 644 rcu_read_unlock(); 645 return ret; 646} 647 648static int flush_pending_writes(conf_t *conf) 649{ 650 /* Any writes that have been queued but are awaiting 651 * bitmap updates get flushed here. 652 * We return 1 if any requests were actually submitted. 653 */ 654 int rv = 0; 655 656 spin_lock_irq(&conf->device_lock); 657 658 if (conf->pending_bio_list.head) { 659 struct bio *bio; 660 bio = bio_list_get(&conf->pending_bio_list); 661 blk_remove_plug(conf->mddev->queue); 662 spin_unlock_irq(&conf->device_lock); 663 /* flush any pending bitmap writes to disk 664 * before proceeding w/ I/O */ 665 bitmap_unplug(conf->mddev->bitmap); 666 667 while (bio) { /* submit pending writes */ 668 struct bio *next = bio->bi_next; 669 bio->bi_next = NULL; 670 generic_make_request(bio); 671 bio = next; 672 } 673 rv = 1; 674 } else 675 spin_unlock_irq(&conf->device_lock); 676 return rv; 677} 678/* Barriers.... 679 * Sometimes we need to suspend IO while we do something else, 680 * either some resync/recovery, or reconfigure the array. 681 * To do this we raise a 'barrier'. 682 * The 'barrier' is a counter that can be raised multiple times 683 * to count how many activities are happening which preclude 684 * normal IO. 685 * We can only raise the barrier if there is no pending IO. 686 * i.e. if nr_pending == 0. 687 * We choose only to raise the barrier if no-one is waiting for the 688 * barrier to go down. This means that as soon as an IO request 689 * is ready, no other operations which require a barrier will start 690 * until the IO request has had a chance. 691 * 692 * So: regular IO calls 'wait_barrier'. When that returns there 693 * is no backgroup IO happening, It must arrange to call 694 * allow_barrier when it has finished its IO. 695 * backgroup IO calls must call raise_barrier. Once that returns 696 * there is no normal IO happeing. It must arrange to call 697 * lower_barrier when the particular background IO completes. 698 */ 699 700static void raise_barrier(conf_t *conf, int force) 701{ 702 BUG_ON(force && !conf->barrier); 703 spin_lock_irq(&conf->resync_lock); 704 705 /* Wait until no block IO is waiting (unless 'force') */ 706 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting, 707 conf->resync_lock, 708 raid10_unplug(conf->mddev->queue)); 709 710 /* block any new IO from starting */ 711 conf->barrier++; 712 713 /* No wait for all pending IO to complete */ 714 wait_event_lock_irq(conf->wait_barrier, 715 !conf->nr_pending && conf->barrier < RESYNC_DEPTH, 716 conf->resync_lock, 717 raid10_unplug(conf->mddev->queue)); 718 719 spin_unlock_irq(&conf->resync_lock); 720} 721 722static void lower_barrier(conf_t *conf) 723{ 724 unsigned long flags; 725 spin_lock_irqsave(&conf->resync_lock, flags); 726 conf->barrier--; 727 spin_unlock_irqrestore(&conf->resync_lock, flags); 728 wake_up(&conf->wait_barrier); 729} 730 731static void wait_barrier(conf_t *conf) 732{ 733 spin_lock_irq(&conf->resync_lock); 734 if (conf->barrier) { 735 conf->nr_waiting++; 736 wait_event_lock_irq(conf->wait_barrier, !conf->barrier, 737 conf->resync_lock, 738 raid10_unplug(conf->mddev->queue)); 739 conf->nr_waiting--; 740 } 741 conf->nr_pending++; 742 spin_unlock_irq(&conf->resync_lock); 743} 744 745static void allow_barrier(conf_t *conf) 746{ 747 unsigned long flags; 748 spin_lock_irqsave(&conf->resync_lock, flags); 749 conf->nr_pending--; 750 spin_unlock_irqrestore(&conf->resync_lock, flags); 751 wake_up(&conf->wait_barrier); 752} 753 754static void freeze_array(conf_t *conf) 755{ 756 /* stop syncio and normal IO and wait for everything to 757 * go quiet. 758 * We increment barrier and nr_waiting, and then 759 * wait until nr_pending match nr_queued+1 760 * This is called in the context of one normal IO request 761 * that has failed. Thus any sync request that might be pending 762 * will be blocked by nr_pending, and we need to wait for 763 * pending IO requests to complete or be queued for re-try. 764 * Thus the number queued (nr_queued) plus this request (1) 765 * must match the number of pending IOs (nr_pending) before 766 * we continue. 767 */ 768 spin_lock_irq(&conf->resync_lock); 769 conf->barrier++; 770 conf->nr_waiting++; 771 wait_event_lock_irq(conf->wait_barrier, 772 conf->nr_pending == conf->nr_queued+1, 773 conf->resync_lock, 774 ({ flush_pending_writes(conf); 775 raid10_unplug(conf->mddev->queue); })); 776 spin_unlock_irq(&conf->resync_lock); 777} 778 779static void unfreeze_array(conf_t *conf) 780{ 781 /* reverse the effect of the freeze */ 782 spin_lock_irq(&conf->resync_lock); 783 conf->barrier--; 784 conf->nr_waiting--; 785 wake_up(&conf->wait_barrier); 786 spin_unlock_irq(&conf->resync_lock); 787} 788 789static int make_request(mddev_t *mddev, struct bio * bio) 790{ 791 conf_t *conf = mddev->private; 792 mirror_info_t *mirror; 793 r10bio_t *r10_bio; 794 struct bio *read_bio; 795 int i; 796 int chunk_sects = conf->chunk_mask + 1; 797 const int rw = bio_data_dir(bio); 798 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC); 799 struct bio_list bl; 800 unsigned long flags; 801 mdk_rdev_t *blocked_rdev; 802 803 if (unlikely(bio->bi_rw & REQ_HARDBARRIER)) { 804 md_barrier_request(mddev, bio); 805 return 0; 806 } 807 808 /* If this request crosses a chunk boundary, we need to 809 * split it. This will only happen for 1 PAGE (or less) requests. 810 */ 811 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9) 812 > chunk_sects && 813 conf->near_copies < conf->raid_disks)) { 814 struct bio_pair *bp; 815 /* Sanity check -- queue functions should prevent this happening */ 816 if (bio->bi_vcnt != 1 || 817 bio->bi_idx != 0) 818 goto bad_map; 819 /* This is a one page bio that upper layers 820 * refuse to split for us, so we need to split it. 821 */ 822 bp = bio_split(bio, 823 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) ); 824 825 /* Each of these 'make_request' calls will call 'wait_barrier'. 826 * If the first succeeds but the second blocks due to the resync 827 * thread raising the barrier, we will deadlock because the 828 * IO to the underlying device will be queued in generic_make_request 829 * and will never complete, so will never reduce nr_pending. 830 * So increment nr_waiting here so no new raise_barriers will 831 * succeed, and so the second wait_barrier cannot block. 832 */ 833 spin_lock_irq(&conf->resync_lock); 834 conf->nr_waiting++; 835 spin_unlock_irq(&conf->resync_lock); 836 837 if (make_request(mddev, &bp->bio1)) 838 generic_make_request(&bp->bio1); 839 if (make_request(mddev, &bp->bio2)) 840 generic_make_request(&bp->bio2); 841 842 spin_lock_irq(&conf->resync_lock); 843 conf->nr_waiting--; 844 wake_up(&conf->wait_barrier); 845 spin_unlock_irq(&conf->resync_lock); 846 847 bio_pair_release(bp); 848 return 0; 849 bad_map: 850 printk("md/raid10:%s: make_request bug: can't convert block across chunks" 851 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2, 852 (unsigned long long)bio->bi_sector, bio->bi_size >> 10); 853 854 bio_io_error(bio); 855 return 0; 856 } 857 858 md_write_start(mddev, bio); 859 860 /* 861 * Register the new request and wait if the reconstruction 862 * thread has put up a bar for new requests. 863 * Continue immediately if no resync is active currently. 864 */ 865 wait_barrier(conf); 866 867 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO); 868 869 r10_bio->master_bio = bio; 870 r10_bio->sectors = bio->bi_size >> 9; 871 872 r10_bio->mddev = mddev; 873 r10_bio->sector = bio->bi_sector; 874 r10_bio->state = 0; 875 876 if (rw == READ) { 877 /* 878 * read balancing logic: 879 */ 880 int disk = read_balance(conf, r10_bio); 881 int slot = r10_bio->read_slot; 882 if (disk < 0) { 883 raid_end_bio_io(r10_bio); 884 return 0; 885 } 886 mirror = conf->mirrors + disk; 887 888 read_bio = bio_clone(bio, GFP_NOIO); 889 890 r10_bio->devs[slot].bio = read_bio; 891 892 read_bio->bi_sector = r10_bio->devs[slot].addr + 893 mirror->rdev->data_offset; 894 read_bio->bi_bdev = mirror->rdev->bdev; 895 read_bio->bi_end_io = raid10_end_read_request; 896 read_bio->bi_rw = READ | do_sync; 897 read_bio->bi_private = r10_bio; 898 899 generic_make_request(read_bio); 900 return 0; 901 } 902 903 /* 904 * WRITE: 905 */ 906 /* first select target devices under rcu_lock and 907 * inc refcount on their rdev. Record them by setting 908 * bios[x] to bio 909 */ 910 raid10_find_phys(conf, r10_bio); 911 retry_write: 912 blocked_rdev = NULL; 913 rcu_read_lock(); 914 for (i = 0; i < conf->copies; i++) { 915 int d = r10_bio->devs[i].devnum; 916 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev); 917 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 918 atomic_inc(&rdev->nr_pending); 919 blocked_rdev = rdev; 920 break; 921 } 922 if (rdev && !test_bit(Faulty, &rdev->flags)) { 923 atomic_inc(&rdev->nr_pending); 924 r10_bio->devs[i].bio = bio; 925 } else { 926 r10_bio->devs[i].bio = NULL; 927 set_bit(R10BIO_Degraded, &r10_bio->state); 928 } 929 } 930 rcu_read_unlock(); 931 932 if (unlikely(blocked_rdev)) { 933 /* Have to wait for this device to get unblocked, then retry */ 934 int j; 935 int d; 936 937 for (j = 0; j < i; j++) 938 if (r10_bio->devs[j].bio) { 939 d = r10_bio->devs[j].devnum; 940 rdev_dec_pending(conf->mirrors[d].rdev, mddev); 941 } 942 allow_barrier(conf); 943 md_wait_for_blocked_rdev(blocked_rdev, mddev); 944 wait_barrier(conf); 945 goto retry_write; 946 } 947 948 atomic_set(&r10_bio->remaining, 0); 949 950 bio_list_init(&bl); 951 for (i = 0; i < conf->copies; i++) { 952 struct bio *mbio; 953 int d = r10_bio->devs[i].devnum; 954 if (!r10_bio->devs[i].bio) 955 continue; 956 957 mbio = bio_clone(bio, GFP_NOIO); 958 r10_bio->devs[i].bio = mbio; 959 960 mbio->bi_sector = r10_bio->devs[i].addr+ 961 conf->mirrors[d].rdev->data_offset; 962 mbio->bi_bdev = conf->mirrors[d].rdev->bdev; 963 mbio->bi_end_io = raid10_end_write_request; 964 mbio->bi_rw = WRITE | do_sync; 965 mbio->bi_private = r10_bio; 966 967 atomic_inc(&r10_bio->remaining); 968 bio_list_add(&bl, mbio); 969 } 970 971 if (unlikely(!atomic_read(&r10_bio->remaining))) { 972 /* the array is dead */ 973 md_write_end(mddev); 974 raid_end_bio_io(r10_bio); 975 return 0; 976 } 977 978 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0); 979 spin_lock_irqsave(&conf->device_lock, flags); 980 bio_list_merge(&conf->pending_bio_list, &bl); 981 blk_plug_device(mddev->queue); 982 spin_unlock_irqrestore(&conf->device_lock, flags); 983 984 /* In case raid10d snuck in to freeze_array */ 985 wake_up(&conf->wait_barrier); 986 987 if (do_sync) 988 md_wakeup_thread(mddev->thread); 989 990 return 0; 991} 992 993static void status(struct seq_file *seq, mddev_t *mddev) 994{ 995 conf_t *conf = mddev->private; 996 int i; 997 998 if (conf->near_copies < conf->raid_disks) 999 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2); 1000 if (conf->near_copies > 1) 1001 seq_printf(seq, " %d near-copies", conf->near_copies); 1002 if (conf->far_copies > 1) { 1003 if (conf->far_offset) 1004 seq_printf(seq, " %d offset-copies", conf->far_copies); 1005 else 1006 seq_printf(seq, " %d far-copies", conf->far_copies); 1007 } 1008 seq_printf(seq, " [%d/%d] [", conf->raid_disks, 1009 conf->raid_disks - mddev->degraded); 1010 for (i = 0; i < conf->raid_disks; i++) 1011 seq_printf(seq, "%s", 1012 conf->mirrors[i].rdev && 1013 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_"); 1014 seq_printf(seq, "]"); 1015} 1016 1017static void error(mddev_t *mddev, mdk_rdev_t *rdev) 1018{ 1019 char b[BDEVNAME_SIZE]; 1020 conf_t *conf = mddev->private; 1021 1022 /* 1023 * If it is not operational, then we have already marked it as dead 1024 * else if it is the last working disks, ignore the error, let the 1025 * next level up know. 1026 * else mark the drive as failed 1027 */ 1028 if (test_bit(In_sync, &rdev->flags) 1029 && conf->raid_disks-mddev->degraded == 1) 1030 /* 1031 * Don't fail the drive, just return an IO error. 1032 * The test should really be more sophisticated than 1033 * "working_disks == 1", but it isn't critical, and 1034 * can wait until we do more sophisticated "is the drive 1035 * really dead" tests... 1036 */ 1037 return; 1038 if (test_and_clear_bit(In_sync, &rdev->flags)) { 1039 unsigned long flags; 1040 spin_lock_irqsave(&conf->device_lock, flags); 1041 mddev->degraded++; 1042 spin_unlock_irqrestore(&conf->device_lock, flags); 1043 /* 1044 * if recovery is running, make sure it aborts. 1045 */ 1046 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1047 } 1048 set_bit(Faulty, &rdev->flags); 1049 set_bit(MD_CHANGE_DEVS, &mddev->flags); 1050 printk(KERN_ALERT "md/raid10:%s: Disk failure on %s, disabling device.\n" 1051 KERN_ALERT "md/raid10:%s: Operation continuing on %d devices.\n", 1052 mdname(mddev), bdevname(rdev->bdev, b), 1053 mdname(mddev), conf->raid_disks - mddev->degraded); 1054} 1055 1056static void print_conf(conf_t *conf) 1057{ 1058 int i; 1059 mirror_info_t *tmp; 1060 1061 printk(KERN_DEBUG "RAID10 conf printout:\n"); 1062 if (!conf) { 1063 printk(KERN_DEBUG "(!conf)\n"); 1064 return; 1065 } 1066 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded, 1067 conf->raid_disks); 1068 1069 for (i = 0; i < conf->raid_disks; i++) { 1070 char b[BDEVNAME_SIZE]; 1071 tmp = conf->mirrors + i; 1072 if (tmp->rdev) 1073 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n", 1074 i, !test_bit(In_sync, &tmp->rdev->flags), 1075 !test_bit(Faulty, &tmp->rdev->flags), 1076 bdevname(tmp->rdev->bdev,b)); 1077 } 1078} 1079 1080static void close_sync(conf_t *conf) 1081{ 1082 wait_barrier(conf); 1083 allow_barrier(conf); 1084 1085 mempool_destroy(conf->r10buf_pool); 1086 conf->r10buf_pool = NULL; 1087} 1088 1089/* check if there are enough drives for 1090 * every block to appear on atleast one 1091 */ 1092static int enough(conf_t *conf) 1093{ 1094 int first = 0; 1095 1096 do { 1097 int n = conf->copies; 1098 int cnt = 0; 1099 while (n--) { 1100 if (conf->mirrors[first].rdev) 1101 cnt++; 1102 first = (first+1) % conf->raid_disks; 1103 } 1104 if (cnt == 0) 1105 return 0; 1106 } while (first != 0); 1107 return 1; 1108} 1109 1110static int raid10_spare_active(mddev_t *mddev) 1111{ 1112 int i; 1113 conf_t *conf = mddev->private; 1114 mirror_info_t *tmp; 1115 int count = 0; 1116 unsigned long flags; 1117 1118 /* 1119 * Find all non-in_sync disks within the RAID10 configuration 1120 * and mark them in_sync 1121 */ 1122 for (i = 0; i < conf->raid_disks; i++) { 1123 tmp = conf->mirrors + i; 1124 if (tmp->rdev 1125 && !test_bit(Faulty, &tmp->rdev->flags) 1126 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) { 1127 count++; 1128 sysfs_notify_dirent(tmp->rdev->sysfs_state); 1129 } 1130 } 1131 spin_lock_irqsave(&conf->device_lock, flags); 1132 mddev->degraded -= count; 1133 spin_unlock_irqrestore(&conf->device_lock, flags); 1134 1135 print_conf(conf); 1136 return count; 1137} 1138 1139 1140static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev) 1141{ 1142 conf_t *conf = mddev->private; 1143 int err = -EEXIST; 1144 int mirror; 1145 mirror_info_t *p; 1146 int first = 0; 1147 int last = conf->raid_disks - 1; 1148 1149 if (mddev->recovery_cp < MaxSector) 1150 /* only hot-add to in-sync arrays, as recovery is 1151 * very different from resync 1152 */ 1153 return -EBUSY; 1154 if (!enough(conf)) 1155 return -EINVAL; 1156 1157 if (rdev->raid_disk >= 0) 1158 first = last = rdev->raid_disk; 1159 1160 if (rdev->saved_raid_disk >= 0 && 1161 rdev->saved_raid_disk >= first && 1162 conf->mirrors[rdev->saved_raid_disk].rdev == NULL) 1163 mirror = rdev->saved_raid_disk; 1164 else 1165 mirror = first; 1166 for ( ; mirror <= last ; mirror++) 1167 if ( !(p=conf->mirrors+mirror)->rdev) { 1168 1169 disk_stack_limits(mddev->gendisk, rdev->bdev, 1170 rdev->data_offset << 9); 1171 /* as we don't honour merge_bvec_fn, we must 1172 * never risk violating it, so limit 1173 * ->max_segments to one lying with a single 1174 * page, as a one page request is never in 1175 * violation. 1176 */ 1177 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) { 1178 blk_queue_max_segments(mddev->queue, 1); 1179 blk_queue_segment_boundary(mddev->queue, 1180 PAGE_CACHE_SIZE - 1); 1181 } 1182 1183 p->head_position = 0; 1184 rdev->raid_disk = mirror; 1185 err = 0; 1186 if (rdev->saved_raid_disk != mirror) 1187 conf->fullsync = 1; 1188 rcu_assign_pointer(p->rdev, rdev); 1189 break; 1190 } 1191 1192 md_integrity_add_rdev(rdev, mddev); 1193 print_conf(conf); 1194 return err; 1195} 1196 1197static int raid10_remove_disk(mddev_t *mddev, int number) 1198{ 1199 conf_t *conf = mddev->private; 1200 int err = 0; 1201 mdk_rdev_t *rdev; 1202 mirror_info_t *p = conf->mirrors+ number; 1203 1204 print_conf(conf); 1205 rdev = p->rdev; 1206 if (rdev) { 1207 if (test_bit(In_sync, &rdev->flags) || 1208 atomic_read(&rdev->nr_pending)) { 1209 err = -EBUSY; 1210 goto abort; 1211 } 1212 /* Only remove faulty devices in recovery 1213 * is not possible. 1214 */ 1215 if (!test_bit(Faulty, &rdev->flags) && 1216 enough(conf)) { 1217 err = -EBUSY; 1218 goto abort; 1219 } 1220 p->rdev = NULL; 1221 synchronize_rcu(); 1222 if (atomic_read(&rdev->nr_pending)) { 1223 /* lost the race, try later */ 1224 err = -EBUSY; 1225 p->rdev = rdev; 1226 goto abort; 1227 } 1228 md_integrity_register(mddev); 1229 } 1230abort: 1231 1232 print_conf(conf); 1233 return err; 1234} 1235 1236 1237static void end_sync_read(struct bio *bio, int error) 1238{ 1239 r10bio_t *r10_bio = bio->bi_private; 1240 conf_t *conf = r10_bio->mddev->private; 1241 int i,d; 1242 1243 for (i=0; i<conf->copies; i++) 1244 if (r10_bio->devs[i].bio == bio) 1245 break; 1246 BUG_ON(i == conf->copies); 1247 update_head_pos(i, r10_bio); 1248 d = r10_bio->devs[i].devnum; 1249 1250 if (test_bit(BIO_UPTODATE, &bio->bi_flags)) 1251 set_bit(R10BIO_Uptodate, &r10_bio->state); 1252 else { 1253 atomic_add(r10_bio->sectors, 1254 &conf->mirrors[d].rdev->corrected_errors); 1255 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery)) 1256 md_error(r10_bio->mddev, 1257 conf->mirrors[d].rdev); 1258 } 1259 1260 /* for reconstruct, we always reschedule after a read. 1261 * for resync, only after all reads 1262 */ 1263 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev); 1264 if (test_bit(R10BIO_IsRecover, &r10_bio->state) || 1265 atomic_dec_and_test(&r10_bio->remaining)) { 1266 /* we have read all the blocks, 1267 * do the comparison in process context in raid10d 1268 */ 1269 reschedule_retry(r10_bio); 1270 } 1271} 1272 1273static void end_sync_write(struct bio *bio, int error) 1274{ 1275 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags); 1276 r10bio_t *r10_bio = bio->bi_private; 1277 mddev_t *mddev = r10_bio->mddev; 1278 conf_t *conf = mddev->private; 1279 int i,d; 1280 1281 for (i = 0; i < conf->copies; i++) 1282 if (r10_bio->devs[i].bio == bio) 1283 break; 1284 d = r10_bio->devs[i].devnum; 1285 1286 if (!uptodate) 1287 md_error(mddev, conf->mirrors[d].rdev); 1288 1289 update_head_pos(i, r10_bio); 1290 1291 rdev_dec_pending(conf->mirrors[d].rdev, mddev); 1292 while (atomic_dec_and_test(&r10_bio->remaining)) { 1293 if (r10_bio->master_bio == NULL) { 1294 /* the primary of several recovery bios */ 1295 sector_t s = r10_bio->sectors; 1296 put_buf(r10_bio); 1297 md_done_sync(mddev, s, 1); 1298 break; 1299 } else { 1300 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio; 1301 put_buf(r10_bio); 1302 r10_bio = r10_bio2; 1303 } 1304 } 1305} 1306 1307/* 1308 * Note: sync and recover and handled very differently for raid10 1309 * This code is for resync. 1310 * For resync, we read through virtual addresses and read all blocks. 1311 * If there is any error, we schedule a write. The lowest numbered 1312 * drive is authoritative. 1313 * However requests come for physical address, so we need to map. 1314 * For every physical address there are raid_disks/copies virtual addresses, 1315 * which is always are least one, but is not necessarly an integer. 1316 * This means that a physical address can span multiple chunks, so we may 1317 * have to submit multiple io requests for a single sync request. 1318 */ 1319/* 1320 * We check if all blocks are in-sync and only write to blocks that 1321 * aren't in sync 1322 */ 1323static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio) 1324{ 1325 conf_t *conf = mddev->private; 1326 int i, first; 1327 struct bio *tbio, *fbio; 1328 1329 atomic_set(&r10_bio->remaining, 1); 1330 1331 /* find the first device with a block */ 1332 for (i=0; i<conf->copies; i++) 1333 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) 1334 break; 1335 1336 if (i == conf->copies) 1337 goto done; 1338 1339 first = i; 1340 fbio = r10_bio->devs[i].bio; 1341 1342 /* now find blocks with errors */ 1343 for (i=0 ; i < conf->copies ; i++) { 1344 int j, d; 1345 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9); 1346 1347 tbio = r10_bio->devs[i].bio; 1348 1349 if (tbio->bi_end_io != end_sync_read) 1350 continue; 1351 if (i == first) 1352 continue; 1353 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) { 1354 /* We know that the bi_io_vec layout is the same for 1355 * both 'first' and 'i', so we just compare them. 1356 * All vec entries are PAGE_SIZE; 1357 */ 1358 for (j = 0; j < vcnt; j++) 1359 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page), 1360 page_address(tbio->bi_io_vec[j].bv_page), 1361 PAGE_SIZE)) 1362 break; 1363 if (j == vcnt) 1364 continue; 1365 mddev->resync_mismatches += r10_bio->sectors; 1366 } 1367 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) 1368 /* Don't fix anything. */ 1369 continue; 1370 /* Ok, we need to write this bio 1371 * First we need to fixup bv_offset, bv_len and 1372 * bi_vecs, as the read request might have corrupted these 1373 */ 1374 tbio->bi_vcnt = vcnt; 1375 tbio->bi_size = r10_bio->sectors << 9; 1376 tbio->bi_idx = 0; 1377 tbio->bi_phys_segments = 0; 1378 tbio->bi_flags &= ~(BIO_POOL_MASK - 1); 1379 tbio->bi_flags |= 1 << BIO_UPTODATE; 1380 tbio->bi_next = NULL; 1381 tbio->bi_rw = WRITE; 1382 tbio->bi_private = r10_bio; 1383 tbio->bi_sector = r10_bio->devs[i].addr; 1384 1385 for (j=0; j < vcnt ; j++) { 1386 tbio->bi_io_vec[j].bv_offset = 0; 1387 tbio->bi_io_vec[j].bv_len = PAGE_SIZE; 1388 1389 memcpy(page_address(tbio->bi_io_vec[j].bv_page), 1390 page_address(fbio->bi_io_vec[j].bv_page), 1391 PAGE_SIZE); 1392 } 1393 tbio->bi_end_io = end_sync_write; 1394 1395 d = r10_bio->devs[i].devnum; 1396 atomic_inc(&conf->mirrors[d].rdev->nr_pending); 1397 atomic_inc(&r10_bio->remaining); 1398 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9); 1399 1400 tbio->bi_sector += conf->mirrors[d].rdev->data_offset; 1401 tbio->bi_bdev = conf->mirrors[d].rdev->bdev; 1402 generic_make_request(tbio); 1403 } 1404 1405done: 1406 if (atomic_dec_and_test(&r10_bio->remaining)) { 1407 md_done_sync(mddev, r10_bio->sectors, 1); 1408 put_buf(r10_bio); 1409 } 1410} 1411 1412/* 1413 * Now for the recovery code. 1414 * Recovery happens across physical sectors. 1415 * We recover all non-is_sync drives by finding the virtual address of 1416 * each, and then choose a working drive that also has that virt address. 1417 * There is a separate r10_bio for each non-in_sync drive. 1418 * Only the first two slots are in use. The first for reading, 1419 * The second for writing. 1420 * 1421 */ 1422 1423static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio) 1424{ 1425 conf_t *conf = mddev->private; 1426 int i, d; 1427 struct bio *bio, *wbio; 1428 1429 1430 /* move the pages across to the second bio 1431 * and submit the write request 1432 */ 1433 bio = r10_bio->devs[0].bio; 1434 wbio = r10_bio->devs[1].bio; 1435 for (i=0; i < wbio->bi_vcnt; i++) { 1436 struct page *p = bio->bi_io_vec[i].bv_page; 1437 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page; 1438 wbio->bi_io_vec[i].bv_page = p; 1439 } 1440 d = r10_bio->devs[1].devnum; 1441 1442 atomic_inc(&conf->mirrors[d].rdev->nr_pending); 1443 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9); 1444 if (test_bit(R10BIO_Uptodate, &r10_bio->state)) 1445 generic_make_request(wbio); 1446 else 1447 bio_endio(wbio, -EIO); 1448} 1449 1450 1451/* 1452 * Used by fix_read_error() to decay the per rdev read_errors. 1453 * We halve the read error count for every hour that has elapsed 1454 * since the last recorded read error. 1455 * 1456 */ 1457static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev) 1458{ 1459 struct timespec cur_time_mon; 1460 unsigned long hours_since_last; 1461 unsigned int read_errors = atomic_read(&rdev->read_errors); 1462 1463 ktime_get_ts(&cur_time_mon); 1464 1465 if (rdev->last_read_error.tv_sec == 0 && 1466 rdev->last_read_error.tv_nsec == 0) { 1467 /* first time we've seen a read error */ 1468 rdev->last_read_error = cur_time_mon; 1469 return; 1470 } 1471 1472 hours_since_last = (cur_time_mon.tv_sec - 1473 rdev->last_read_error.tv_sec) / 3600; 1474 1475 rdev->last_read_error = cur_time_mon; 1476 1477 /* 1478 * if hours_since_last is > the number of bits in read_errors 1479 * just set read errors to 0. We do this to avoid 1480 * overflowing the shift of read_errors by hours_since_last. 1481 */ 1482 if (hours_since_last >= 8 * sizeof(read_errors)) 1483 atomic_set(&rdev->read_errors, 0); 1484 else 1485 atomic_set(&rdev->read_errors, read_errors >> hours_since_last); 1486} 1487 1488/* 1489 * This is a kernel thread which: 1490 * 1491 * 1. Retries failed read operations on working mirrors. 1492 * 2. Updates the raid superblock when problems encounter. 1493 * 3. Performs writes following reads for array synchronising. 1494 */ 1495 1496static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio) 1497{ 1498 int sect = 0; /* Offset from r10_bio->sector */ 1499 int sectors = r10_bio->sectors; 1500 mdk_rdev_t*rdev; 1501 int max_read_errors = atomic_read(&mddev->max_corr_read_errors); 1502 int d = r10_bio->devs[r10_bio->read_slot].devnum; 1503 1504 rcu_read_lock(); 1505 rdev = rcu_dereference(conf->mirrors[d].rdev); 1506 if (rdev) { /* If rdev is not NULL */ 1507 char b[BDEVNAME_SIZE]; 1508 int cur_read_error_count = 0; 1509 1510 bdevname(rdev->bdev, b); 1511 1512 if (test_bit(Faulty, &rdev->flags)) { 1513 rcu_read_unlock(); 1514 /* drive has already been failed, just ignore any 1515 more fix_read_error() attempts */ 1516 return; 1517 } 1518 1519 check_decay_read_errors(mddev, rdev); 1520 atomic_inc(&rdev->read_errors); 1521 cur_read_error_count = atomic_read(&rdev->read_errors); 1522 if (cur_read_error_count > max_read_errors) { 1523 rcu_read_unlock(); 1524 printk(KERN_NOTICE 1525 "md/raid10:%s: %s: Raid device exceeded " 1526 "read_error threshold " 1527 "[cur %d:max %d]\n", 1528 mdname(mddev), 1529 b, cur_read_error_count, max_read_errors); 1530 printk(KERN_NOTICE 1531 "md/raid10:%s: %s: Failing raid " 1532 "device\n", mdname(mddev), b); 1533 md_error(mddev, conf->mirrors[d].rdev); 1534 return; 1535 } 1536 } 1537 rcu_read_unlock(); 1538 1539 while(sectors) { 1540 int s = sectors; 1541 int sl = r10_bio->read_slot; 1542 int success = 0; 1543 int start; 1544 1545 if (s > (PAGE_SIZE>>9)) 1546 s = PAGE_SIZE >> 9; 1547 1548 rcu_read_lock(); 1549 do { 1550 d = r10_bio->devs[sl].devnum; 1551 rdev = rcu_dereference(conf->mirrors[d].rdev); 1552 if (rdev && 1553 test_bit(In_sync, &rdev->flags)) { 1554 atomic_inc(&rdev->nr_pending); 1555 rcu_read_unlock(); 1556 success = sync_page_io(rdev->bdev, 1557 r10_bio->devs[sl].addr + 1558 sect + rdev->data_offset, 1559 s<<9, 1560 conf->tmppage, READ); 1561 rdev_dec_pending(rdev, mddev); 1562 rcu_read_lock(); 1563 if (success) 1564 break; 1565 } 1566 sl++; 1567 if (sl == conf->copies) 1568 sl = 0; 1569 } while (!success && sl != r10_bio->read_slot); 1570 rcu_read_unlock(); 1571 1572 if (!success) { 1573 /* Cannot read from anywhere -- bye bye array */ 1574 int dn = r10_bio->devs[r10_bio->read_slot].devnum; 1575 md_error(mddev, conf->mirrors[dn].rdev); 1576 break; 1577 } 1578 1579 start = sl; 1580 /* write it back and re-read */ 1581 rcu_read_lock(); 1582 while (sl != r10_bio->read_slot) { 1583 char b[BDEVNAME_SIZE]; 1584 1585 if (sl==0) 1586 sl = conf->copies; 1587 sl--; 1588 d = r10_bio->devs[sl].devnum; 1589 rdev = rcu_dereference(conf->mirrors[d].rdev); 1590 if (rdev && 1591 test_bit(In_sync, &rdev->flags)) { 1592 atomic_inc(&rdev->nr_pending); 1593 rcu_read_unlock(); 1594 atomic_add(s, &rdev->corrected_errors); 1595 if (sync_page_io(rdev->bdev, 1596 r10_bio->devs[sl].addr + 1597 sect + rdev->data_offset, 1598 s<<9, conf->tmppage, WRITE) 1599 == 0) { 1600 /* Well, this device is dead */ 1601 printk(KERN_NOTICE 1602 "md/raid10:%s: read correction " 1603 "write failed" 1604 " (%d sectors at %llu on %s)\n", 1605 mdname(mddev), s, 1606 (unsigned long long)(sect+ 1607 rdev->data_offset), 1608 bdevname(rdev->bdev, b)); 1609 printk(KERN_NOTICE "md/raid10:%s: %s: failing " 1610 "drive\n", 1611 mdname(mddev), 1612 bdevname(rdev->bdev, b)); 1613 md_error(mddev, rdev); 1614 } 1615 rdev_dec_pending(rdev, mddev); 1616 rcu_read_lock(); 1617 } 1618 } 1619 sl = start; 1620 while (sl != r10_bio->read_slot) { 1621 1622 if (sl==0) 1623 sl = conf->copies; 1624 sl--; 1625 d = r10_bio->devs[sl].devnum; 1626 rdev = rcu_dereference(conf->mirrors[d].rdev); 1627 if (rdev && 1628 test_bit(In_sync, &rdev->flags)) { 1629 char b[BDEVNAME_SIZE]; 1630 atomic_inc(&rdev->nr_pending); 1631 rcu_read_unlock(); 1632 if (sync_page_io(rdev->bdev, 1633 r10_bio->devs[sl].addr + 1634 sect + rdev->data_offset, 1635 s<<9, conf->tmppage, 1636 READ) == 0) { 1637 /* Well, this device is dead */ 1638 printk(KERN_NOTICE 1639 "md/raid10:%s: unable to read back " 1640 "corrected sectors" 1641 " (%d sectors at %llu on %s)\n", 1642 mdname(mddev), s, 1643 (unsigned long long)(sect+ 1644 rdev->data_offset), 1645 bdevname(rdev->bdev, b)); 1646 printk(KERN_NOTICE "md/raid10:%s: %s: failing drive\n", 1647 mdname(mddev), 1648 bdevname(rdev->bdev, b)); 1649 1650 md_error(mddev, rdev); 1651 } else { 1652 printk(KERN_INFO 1653 "md/raid10:%s: read error corrected" 1654 " (%d sectors at %llu on %s)\n", 1655 mdname(mddev), s, 1656 (unsigned long long)(sect+ 1657 rdev->data_offset), 1658 bdevname(rdev->bdev, b)); 1659 } 1660 1661 rdev_dec_pending(rdev, mddev); 1662 rcu_read_lock(); 1663 } 1664 } 1665 rcu_read_unlock(); 1666 1667 sectors -= s; 1668 sect += s; 1669 } 1670} 1671 1672static void raid10d(mddev_t *mddev) 1673{ 1674 r10bio_t *r10_bio; 1675 struct bio *bio; 1676 unsigned long flags; 1677 conf_t *conf = mddev->private; 1678 struct list_head *head = &conf->retry_list; 1679 int unplug=0; 1680 mdk_rdev_t *rdev; 1681 1682 md_check_recovery(mddev); 1683 1684 for (;;) { 1685 char b[BDEVNAME_SIZE]; 1686 1687 unplug += flush_pending_writes(conf); 1688 1689 spin_lock_irqsave(&conf->device_lock, flags); 1690 if (list_empty(head)) { 1691 spin_unlock_irqrestore(&conf->device_lock, flags); 1692 break; 1693 } 1694 r10_bio = list_entry(head->prev, r10bio_t, retry_list); 1695 list_del(head->prev); 1696 conf->nr_queued--; 1697 spin_unlock_irqrestore(&conf->device_lock, flags); 1698 1699 mddev = r10_bio->mddev; 1700 conf = mddev->private; 1701 if (test_bit(R10BIO_IsSync, &r10_bio->state)) { 1702 sync_request_write(mddev, r10_bio); 1703 unplug = 1; 1704 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) { 1705 recovery_request_write(mddev, r10_bio); 1706 unplug = 1; 1707 } else { 1708 int mirror; 1709 /* we got a read error. Maybe the drive is bad. Maybe just 1710 * the block and we can fix it. 1711 * We freeze all other IO, and try reading the block from 1712 * other devices. When we find one, we re-write 1713 * and check it that fixes the read error. 1714 * This is all done synchronously while the array is 1715 * frozen. 1716 */ 1717 if (mddev->ro == 0) { 1718 freeze_array(conf); 1719 fix_read_error(conf, mddev, r10_bio); 1720 unfreeze_array(conf); 1721 } 1722 1723 bio = r10_bio->devs[r10_bio->read_slot].bio; 1724 r10_bio->devs[r10_bio->read_slot].bio = 1725 mddev->ro ? IO_BLOCKED : NULL; 1726 mirror = read_balance(conf, r10_bio); 1727 if (mirror == -1) { 1728 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O" 1729 " read error for block %llu\n", 1730 mdname(mddev), 1731 bdevname(bio->bi_bdev,b), 1732 (unsigned long long)r10_bio->sector); 1733 raid_end_bio_io(r10_bio); 1734 bio_put(bio); 1735 } else { 1736 const unsigned long do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC); 1737 bio_put(bio); 1738 rdev = conf->mirrors[mirror].rdev; 1739 if (printk_ratelimit()) 1740 printk(KERN_ERR "md/raid10:%s: %s: redirecting sector %llu to" 1741 " another mirror\n", 1742 mdname(mddev), 1743 bdevname(rdev->bdev,b), 1744 (unsigned long long)r10_bio->sector); 1745 bio = bio_clone(r10_bio->master_bio, GFP_NOIO); 1746 r10_bio->devs[r10_bio->read_slot].bio = bio; 1747 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr 1748 + rdev->data_offset; 1749 bio->bi_bdev = rdev->bdev; 1750 bio->bi_rw = READ | do_sync; 1751 bio->bi_private = r10_bio; 1752 bio->bi_end_io = raid10_end_read_request; 1753 unplug = 1; 1754 generic_make_request(bio); 1755 } 1756 } 1757 cond_resched(); 1758 } 1759 if (unplug) 1760 unplug_slaves(mddev); 1761} 1762 1763 1764static int init_resync(conf_t *conf) 1765{ 1766 int buffs; 1767 1768 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; 1769 BUG_ON(conf->r10buf_pool); 1770 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf); 1771 if (!conf->r10buf_pool) 1772 return -ENOMEM; 1773 conf->next_resync = 0; 1774 return 0; 1775} 1776 1777/* 1778 * perform a "sync" on one "block" 1779 * 1780 * We need to make sure that no normal I/O request - particularly write 1781 * requests - conflict with active sync requests. 1782 * 1783 * This is achieved by tracking pending requests and a 'barrier' concept 1784 * that can be installed to exclude normal IO requests. 1785 * 1786 * Resync and recovery are handled very differently. 1787 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery. 1788 * 1789 * For resync, we iterate over virtual addresses, read all copies, 1790 * and update if there are differences. If only one copy is live, 1791 * skip it. 1792 * For recovery, we iterate over physical addresses, read a good 1793 * value for each non-in_sync drive, and over-write. 1794 * 1795 * So, for recovery we may have several outstanding complex requests for a 1796 * given address, one for each out-of-sync device. We model this by allocating 1797 * a number of r10_bio structures, one for each out-of-sync device. 1798 * As we setup these structures, we collect all bio's together into a list 1799 * which we then process collectively to add pages, and then process again 1800 * to pass to generic_make_request. 1801 * 1802 * The r10_bio structures are linked using a borrowed master_bio pointer. 1803 * This link is counted in ->remaining. When the r10_bio that points to NULL 1804 * has its remaining count decremented to 0, the whole complex operation 1805 * is complete. 1806 * 1807 */ 1808 1809static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster) 1810{ 1811 conf_t *conf = mddev->private; 1812 r10bio_t *r10_bio; 1813 struct bio *biolist = NULL, *bio; 1814 sector_t max_sector, nr_sectors; 1815 int disk; 1816 int i; 1817 int max_sync; 1818 int sync_blocks; 1819 1820 sector_t sectors_skipped = 0; 1821 int chunks_skipped = 0; 1822 1823 if (!conf->r10buf_pool) 1824 if (init_resync(conf)) 1825 return 0; 1826 1827 skipped: 1828 max_sector = mddev->dev_sectors; 1829 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) 1830 max_sector = mddev->resync_max_sectors; 1831 if (sector_nr >= max_sector) { 1832 /* If we aborted, we need to abort the 1833 * sync on the 'current' bitmap chucks (there can 1834 * be several when recovering multiple devices). 1835 * as we may have started syncing it but not finished. 1836 * We can find the current address in 1837 * mddev->curr_resync, but for recovery, 1838 * we need to convert that to several 1839 * virtual addresses. 1840 */ 1841 if (mddev->curr_resync < max_sector) { /* aborted */ 1842 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) 1843 bitmap_end_sync(mddev->bitmap, mddev->curr_resync, 1844 &sync_blocks, 1); 1845 else for (i=0; i<conf->raid_disks; i++) { 1846 sector_t sect = 1847 raid10_find_virt(conf, mddev->curr_resync, i); 1848 bitmap_end_sync(mddev->bitmap, sect, 1849 &sync_blocks, 1); 1850 } 1851 } else /* completed sync */ 1852 conf->fullsync = 0; 1853 1854 bitmap_close_sync(mddev->bitmap); 1855 close_sync(conf); 1856 *skipped = 1; 1857 return sectors_skipped; 1858 } 1859 if (chunks_skipped >= conf->raid_disks) { 1860 /* if there has been nothing to do on any drive, 1861 * then there is nothing to do at all.. 1862 */ 1863 *skipped = 1; 1864 return (max_sector - sector_nr) + sectors_skipped; 1865 } 1866 1867 if (max_sector > mddev->resync_max) 1868 max_sector = mddev->resync_max; /* Don't do IO beyond here */ 1869 1870 /* make sure whole request will fit in a chunk - if chunks 1871 * are meaningful 1872 */ 1873 if (conf->near_copies < conf->raid_disks && 1874 max_sector > (sector_nr | conf->chunk_mask)) 1875 max_sector = (sector_nr | conf->chunk_mask) + 1; 1876 /* 1877 * If there is non-resync activity waiting for us then 1878 * put in a delay to throttle resync. 1879 */ 1880 if (!go_faster && conf->nr_waiting) 1881 msleep_interruptible(1000); 1882 1883 /* Again, very different code for resync and recovery. 1884 * Both must result in an r10bio with a list of bios that 1885 * have bi_end_io, bi_sector, bi_bdev set, 1886 * and bi_private set to the r10bio. 1887 * For recovery, we may actually create several r10bios 1888 * with 2 bios in each, that correspond to the bios in the main one. 1889 * In this case, the subordinate r10bios link back through a 1890 * borrowed master_bio pointer, and the counter in the master 1891 * includes a ref from each subordinate. 1892 */ 1893 /* First, we decide what to do and set ->bi_end_io 1894 * To end_sync_read if we want to read, and 1895 * end_sync_write if we will want to write. 1896 */ 1897 1898 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9); 1899 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) { 1900 /* recovery... the complicated one */ 1901 int j, k; 1902 r10_bio = NULL; 1903 1904 for (i=0 ; i<conf->raid_disks; i++) 1905 if (conf->mirrors[i].rdev && 1906 !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) { 1907 int still_degraded = 0; 1908 /* want to reconstruct this device */ 1909 r10bio_t *rb2 = r10_bio; 1910 sector_t sect = raid10_find_virt(conf, sector_nr, i); 1911 int must_sync; 1912 /* Unless we are doing a full sync, we only need 1913 * to recover the block if it is set in the bitmap 1914 */ 1915 must_sync = bitmap_start_sync(mddev->bitmap, sect, 1916 &sync_blocks, 1); 1917 if (sync_blocks < max_sync) 1918 max_sync = sync_blocks; 1919 if (!must_sync && 1920 !conf->fullsync) { 1921 /* yep, skip the sync_blocks here, but don't assume 1922 * that there will never be anything to do here 1923 */ 1924 chunks_skipped = -1; 1925 continue; 1926 } 1927 1928 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO); 1929 raise_barrier(conf, rb2 != NULL); 1930 atomic_set(&r10_bio->remaining, 0); 1931 1932 r10_bio->master_bio = (struct bio*)rb2; 1933 if (rb2) 1934 atomic_inc(&rb2->remaining); 1935 r10_bio->mddev = mddev; 1936 set_bit(R10BIO_IsRecover, &r10_bio->state); 1937 r10_bio->sector = sect; 1938 1939 raid10_find_phys(conf, r10_bio); 1940 1941 /* Need to check if the array will still be 1942 * degraded 1943 */ 1944 for (j=0; j<conf->raid_disks; j++) 1945 if (conf->mirrors[j].rdev == NULL || 1946 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) { 1947 still_degraded = 1; 1948 break; 1949 } 1950 1951 must_sync = bitmap_start_sync(mddev->bitmap, sect, 1952 &sync_blocks, still_degraded); 1953 1954 for (j=0; j<conf->copies;j++) { 1955 int d = r10_bio->devs[j].devnum; 1956 if (conf->mirrors[d].rdev && 1957 test_bit(In_sync, &conf->mirrors[d].rdev->flags)) { 1958 /* This is where we read from */ 1959 bio = r10_bio->devs[0].bio; 1960 bio->bi_next = biolist; 1961 biolist = bio; 1962 bio->bi_private = r10_bio; 1963 bio->bi_end_io = end_sync_read; 1964 bio->bi_rw = READ; 1965 bio->bi_sector = r10_bio->devs[j].addr + 1966 conf->mirrors[d].rdev->data_offset; 1967 bio->bi_bdev = conf->mirrors[d].rdev->bdev; 1968 atomic_inc(&conf->mirrors[d].rdev->nr_pending); 1969 atomic_inc(&r10_bio->remaining); 1970 /* and we write to 'i' */ 1971 1972 for (k=0; k<conf->copies; k++) 1973 if (r10_bio->devs[k].devnum == i) 1974 break; 1975 BUG_ON(k == conf->copies); 1976 bio = r10_bio->devs[1].bio; 1977 bio->bi_next = biolist; 1978 biolist = bio; 1979 bio->bi_private = r10_bio; 1980 bio->bi_end_io = end_sync_write; 1981 bio->bi_rw = WRITE; 1982 bio->bi_sector = r10_bio->devs[k].addr + 1983 conf->mirrors[i].rdev->data_offset; 1984 bio->bi_bdev = conf->mirrors[i].rdev->bdev; 1985 1986 r10_bio->devs[0].devnum = d; 1987 r10_bio->devs[1].devnum = i; 1988 1989 break; 1990 } 1991 } 1992 if (j == conf->copies) { 1993 /* Cannot recover, so abort the recovery */ 1994 put_buf(r10_bio); 1995 if (rb2) 1996 atomic_dec(&rb2->remaining); 1997 r10_bio = rb2; 1998 if (!test_and_set_bit(MD_RECOVERY_INTR, 1999 &mddev->recovery)) 2000 printk(KERN_INFO "md/raid10:%s: insufficient " 2001 "working devices for recovery.\n", 2002 mdname(mddev)); 2003 break; 2004 } 2005 } 2006 if (biolist == NULL) { 2007 while (r10_bio) { 2008 r10bio_t *rb2 = r10_bio; 2009 r10_bio = (r10bio_t*) rb2->master_bio; 2010 rb2->master_bio = NULL; 2011 put_buf(rb2); 2012 } 2013 goto giveup; 2014 } 2015 } else { 2016 /* resync. Schedule a read for every block at this virt offset */ 2017 int count = 0; 2018 2019 bitmap_cond_end_sync(mddev->bitmap, sector_nr); 2020 2021 if (!bitmap_start_sync(mddev->bitmap, sector_nr, 2022 &sync_blocks, mddev->degraded) && 2023 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { 2024 /* We can skip this block */ 2025 *skipped = 1; 2026 return sync_blocks + sectors_skipped; 2027 } 2028 if (sync_blocks < max_sync) 2029 max_sync = sync_blocks; 2030 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO); 2031 2032 r10_bio->mddev = mddev; 2033 atomic_set(&r10_bio->remaining, 0); 2034 raise_barrier(conf, 0); 2035 conf->next_resync = sector_nr; 2036 2037 r10_bio->master_bio = NULL; 2038 r10_bio->sector = sector_nr; 2039 set_bit(R10BIO_IsSync, &r10_bio->state); 2040 raid10_find_phys(conf, r10_bio); 2041 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1; 2042 2043 for (i=0; i<conf->copies; i++) { 2044 int d = r10_bio->devs[i].devnum; 2045 bio = r10_bio->devs[i].bio; 2046 bio->bi_end_io = NULL; 2047 clear_bit(BIO_UPTODATE, &bio->bi_flags); 2048 if (conf->mirrors[d].rdev == NULL || 2049 test_bit(Faulty, &conf->mirrors[d].rdev->flags)) 2050 continue; 2051 atomic_inc(&conf->mirrors[d].rdev->nr_pending); 2052 atomic_inc(&r10_bio->remaining); 2053 bio->bi_next = biolist; 2054 biolist = bio; 2055 bio->bi_private = r10_bio; 2056 bio->bi_end_io = end_sync_read; 2057 bio->bi_rw = READ; 2058 bio->bi_sector = r10_bio->devs[i].addr + 2059 conf->mirrors[d].rdev->data_offset; 2060 bio->bi_bdev = conf->mirrors[d].rdev->bdev; 2061 count++; 2062 } 2063 2064 if (count < 2) { 2065 for (i=0; i<conf->copies; i++) { 2066 int d = r10_bio->devs[i].devnum; 2067 if (r10_bio->devs[i].bio->bi_end_io) 2068 rdev_dec_pending(conf->mirrors[d].rdev, mddev); 2069 } 2070 put_buf(r10_bio); 2071 biolist = NULL; 2072 goto giveup; 2073 } 2074 } 2075 2076 for (bio = biolist; bio ; bio=bio->bi_next) { 2077 2078 bio->bi_flags &= ~(BIO_POOL_MASK - 1); 2079 if (bio->bi_end_io) 2080 bio->bi_flags |= 1 << BIO_UPTODATE; 2081 bio->bi_vcnt = 0; 2082 bio->bi_idx = 0; 2083 bio->bi_phys_segments = 0; 2084 bio->bi_size = 0; 2085 } 2086 2087 nr_sectors = 0; 2088 if (sector_nr + max_sync < max_sector) 2089 max_sector = sector_nr + max_sync; 2090 do { 2091 struct page *page; 2092 int len = PAGE_SIZE; 2093 disk = 0; 2094 if (sector_nr + (len>>9) > max_sector) 2095 len = (max_sector - sector_nr) << 9; 2096 if (len == 0) 2097 break; 2098 for (bio= biolist ; bio ; bio=bio->bi_next) { 2099 page = bio->bi_io_vec[bio->bi_vcnt].bv_page; 2100 if (bio_add_page(bio, page, len, 0) == 0) { 2101 /* stop here */ 2102 struct bio *bio2; 2103 bio->bi_io_vec[bio->bi_vcnt].bv_page = page; 2104 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) { 2105 /* remove last page from this bio */ 2106 bio2->bi_vcnt--; 2107 bio2->bi_size -= len; 2108 bio2->bi_flags &= ~(1<< BIO_SEG_VALID); 2109 } 2110 goto bio_full; 2111 } 2112 disk = i; 2113 } 2114 nr_sectors += len>>9; 2115 sector_nr += len>>9; 2116 } while (biolist->bi_vcnt < RESYNC_PAGES); 2117 bio_full: 2118 r10_bio->sectors = nr_sectors; 2119 2120 while (biolist) { 2121 bio = biolist; 2122 biolist = biolist->bi_next; 2123 2124 bio->bi_next = NULL; 2125 r10_bio = bio->bi_private; 2126 r10_bio->sectors = nr_sectors; 2127 2128 if (bio->bi_end_io == end_sync_read) { 2129 md_sync_acct(bio->bi_bdev, nr_sectors); 2130 generic_make_request(bio); 2131 } 2132 } 2133 2134 if (sectors_skipped) 2135 /* pretend they weren't skipped, it makes 2136 * no important difference in this case 2137 */ 2138 md_done_sync(mddev, sectors_skipped, 1); 2139 2140 return sectors_skipped + nr_sectors; 2141 giveup: 2142 /* There is nowhere to write, so all non-sync 2143 * drives must be failed, so try the next chunk... 2144 */ 2145 if (sector_nr + max_sync < max_sector) 2146 max_sector = sector_nr + max_sync; 2147 2148 sectors_skipped += (max_sector - sector_nr); 2149 chunks_skipped ++; 2150 sector_nr = max_sector; 2151 goto skipped; 2152} 2153 2154static sector_t 2155raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks) 2156{ 2157 sector_t size; 2158 conf_t *conf = mddev->private; 2159 2160 if (!raid_disks) 2161 raid_disks = conf->raid_disks; 2162 if (!sectors) 2163 sectors = conf->dev_sectors; 2164 2165 size = sectors >> conf->chunk_shift; 2166 sector_div(size, conf->far_copies); 2167 size = size * raid_disks; 2168 sector_div(size, conf->near_copies); 2169 2170 return size << conf->chunk_shift; 2171} 2172 2173 2174static conf_t *setup_conf(mddev_t *mddev) 2175{ 2176 conf_t *conf = NULL; 2177 int nc, fc, fo; 2178 sector_t stride, size; 2179 int err = -EINVAL; 2180 2181 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) || 2182 !is_power_of_2(mddev->new_chunk_sectors)) { 2183 printk(KERN_ERR "md/raid10:%s: chunk size must be " 2184 "at least PAGE_SIZE(%ld) and be a power of 2.\n", 2185 mdname(mddev), PAGE_SIZE); 2186 goto out; 2187 } 2188 2189 nc = mddev->new_layout & 255; 2190 fc = (mddev->new_layout >> 8) & 255; 2191 fo = mddev->new_layout & (1<<16); 2192 2193 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks || 2194 (mddev->new_layout >> 17)) { 2195 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n", 2196 mdname(mddev), mddev->new_layout); 2197 goto out; 2198 } 2199 2200 err = -ENOMEM; 2201 conf = kzalloc(sizeof(conf_t), GFP_KERNEL); 2202 if (!conf) 2203 goto out; 2204 2205 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks, 2206 GFP_KERNEL); 2207 if (!conf->mirrors) 2208 goto out; 2209 2210 conf->tmppage = alloc_page(GFP_KERNEL); 2211 if (!conf->tmppage) 2212 goto out; 2213 2214 2215 conf->raid_disks = mddev->raid_disks; 2216 conf->near_copies = nc; 2217 conf->far_copies = fc; 2218 conf->copies = nc*fc; 2219 conf->far_offset = fo; 2220 conf->chunk_mask = mddev->new_chunk_sectors - 1; 2221 conf->chunk_shift = ffz(~mddev->new_chunk_sectors); 2222 2223 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc, 2224 r10bio_pool_free, conf); 2225 if (!conf->r10bio_pool) 2226 goto out; 2227 2228 size = mddev->dev_sectors >> conf->chunk_shift; 2229 sector_div(size, fc); 2230 size = size * conf->raid_disks; 2231 sector_div(size, nc); 2232 /* 'size' is now the number of chunks in the array */ 2233 /* calculate "used chunks per device" in 'stride' */ 2234 stride = size * conf->copies; 2235 2236 /* We need to round up when dividing by raid_disks to 2237 * get the stride size. 2238 */ 2239 stride += conf->raid_disks - 1; 2240 sector_div(stride, conf->raid_disks); 2241 2242 conf->dev_sectors = stride << conf->chunk_shift; 2243 2244 if (fo) 2245 stride = 1; 2246 else 2247 sector_div(stride, fc); 2248 conf->stride = stride << conf->chunk_shift; 2249 2250 2251 spin_lock_init(&conf->device_lock); 2252 INIT_LIST_HEAD(&conf->retry_list); 2253 2254 spin_lock_init(&conf->resync_lock); 2255 init_waitqueue_head(&conf->wait_barrier); 2256 2257 conf->thread = md_register_thread(raid10d, mddev, NULL); 2258 if (!conf->thread) 2259 goto out; 2260 2261 conf->mddev = mddev; 2262 return conf; 2263 2264 out: 2265 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n", 2266 mdname(mddev)); 2267 if (conf) { 2268 if (conf->r10bio_pool) 2269 mempool_destroy(conf->r10bio_pool); 2270 kfree(conf->mirrors); 2271 safe_put_page(conf->tmppage); 2272 kfree(conf); 2273 } 2274 return ERR_PTR(err); 2275} 2276 2277static int run(mddev_t *mddev) 2278{ 2279 conf_t *conf; 2280 int i, disk_idx, chunk_size; 2281 mirror_info_t *disk; 2282 mdk_rdev_t *rdev; 2283 sector_t size; 2284 2285 /* 2286 * copy the already verified devices into our private RAID10 2287 * bookkeeping area. [whatever we allocate in run(), 2288 * should be freed in stop()] 2289 */ 2290 2291 if (mddev->private == NULL) { 2292 conf = setup_conf(mddev); 2293 if (IS_ERR(conf)) 2294 return PTR_ERR(conf); 2295 mddev->private = conf; 2296 } 2297 conf = mddev->private; 2298 if (!conf) 2299 goto out; 2300 2301 mddev->queue->queue_lock = &conf->device_lock; 2302 2303 mddev->thread = conf->thread; 2304 conf->thread = NULL; 2305 2306 chunk_size = mddev->chunk_sectors << 9; 2307 blk_queue_io_min(mddev->queue, chunk_size); 2308 if (conf->raid_disks % conf->near_copies) 2309 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks); 2310 else 2311 blk_queue_io_opt(mddev->queue, chunk_size * 2312 (conf->raid_disks / conf->near_copies)); 2313 2314 list_for_each_entry(rdev, &mddev->disks, same_set) { 2315 disk_idx = rdev->raid_disk; 2316 if (disk_idx >= conf->raid_disks 2317 || disk_idx < 0) 2318 continue; 2319 disk = conf->mirrors + disk_idx; 2320 2321 disk->rdev = rdev; 2322 disk_stack_limits(mddev->gendisk, rdev->bdev, 2323 rdev->data_offset << 9); 2324 /* as we don't honour merge_bvec_fn, we must never risk 2325 * violating it, so limit max_segments to 1 lying 2326 * within a single page. 2327 */ 2328 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) { 2329 blk_queue_max_segments(mddev->queue, 1); 2330 blk_queue_segment_boundary(mddev->queue, 2331 PAGE_CACHE_SIZE - 1); 2332 } 2333 2334 disk->head_position = 0; 2335 } 2336 /* need to check that every block has at least one working mirror */ 2337 if (!enough(conf)) { 2338 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n", 2339 mdname(mddev)); 2340 goto out_free_conf; 2341 } 2342 2343 mddev->degraded = 0; 2344 for (i = 0; i < conf->raid_disks; i++) { 2345 2346 disk = conf->mirrors + i; 2347 2348 if (!disk->rdev || 2349 !test_bit(In_sync, &disk->rdev->flags)) { 2350 disk->head_position = 0; 2351 mddev->degraded++; 2352 if (disk->rdev) 2353 conf->fullsync = 1; 2354 } 2355 } 2356 2357 if (mddev->recovery_cp != MaxSector) 2358 printk(KERN_NOTICE "md/raid10:%s: not clean" 2359 " -- starting background reconstruction\n", 2360 mdname(mddev)); 2361 printk(KERN_INFO 2362 "md/raid10:%s: active with %d out of %d devices\n", 2363 mdname(mddev), conf->raid_disks - mddev->degraded, 2364 conf->raid_disks); 2365 /* 2366 * Ok, everything is just fine now 2367 */ 2368 mddev->dev_sectors = conf->dev_sectors; 2369 size = raid10_size(mddev, 0, 0); 2370 md_set_array_sectors(mddev, size); 2371 mddev->resync_max_sectors = size; 2372 2373 mddev->queue->unplug_fn = raid10_unplug; 2374 mddev->queue->backing_dev_info.congested_fn = raid10_congested; 2375 mddev->queue->backing_dev_info.congested_data = mddev; 2376 2377 /* Calculate max read-ahead size. 2378 * We need to readahead at least twice a whole stripe.... 2379 * maybe... 2380 */ 2381 { 2382 int stripe = conf->raid_disks * 2383 ((mddev->chunk_sectors << 9) / PAGE_SIZE); 2384 stripe /= conf->near_copies; 2385 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe) 2386 mddev->queue->backing_dev_info.ra_pages = 2* stripe; 2387 } 2388 2389 if (conf->near_copies < conf->raid_disks) 2390 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec); 2391 md_integrity_register(mddev); 2392 return 0; 2393 2394out_free_conf: 2395 md_unregister_thread(mddev->thread); 2396 if (conf->r10bio_pool) 2397 mempool_destroy(conf->r10bio_pool); 2398 safe_put_page(conf->tmppage); 2399 kfree(conf->mirrors); 2400 kfree(conf); 2401 mddev->private = NULL; 2402out: 2403 return -EIO; 2404} 2405 2406static int stop(mddev_t *mddev) 2407{ 2408 conf_t *conf = mddev->private; 2409 2410 raise_barrier(conf, 0); 2411 lower_barrier(conf); 2412 2413 md_unregister_thread(mddev->thread); 2414 mddev->thread = NULL; 2415 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/ 2416 if (conf->r10bio_pool) 2417 mempool_destroy(conf->r10bio_pool); 2418 kfree(conf->mirrors); 2419 kfree(conf); 2420 mddev->private = NULL; 2421 return 0; 2422} 2423 2424static void raid10_quiesce(mddev_t *mddev, int state) 2425{ 2426 conf_t *conf = mddev->private; 2427 2428 switch(state) { 2429 case 1: 2430 raise_barrier(conf, 0); 2431 break; 2432 case 0: 2433 lower_barrier(conf); 2434 break; 2435 } 2436} 2437 2438static void *raid10_takeover_raid0(mddev_t *mddev) 2439{ 2440 mdk_rdev_t *rdev; 2441 conf_t *conf; 2442 2443 if (mddev->degraded > 0) { 2444 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n", 2445 mdname(mddev)); 2446 return ERR_PTR(-EINVAL); 2447 } 2448 2449 /* Set new parameters */ 2450 mddev->new_level = 10; 2451 /* new layout: far_copies = 1, near_copies = 2 */ 2452 mddev->new_layout = (1<<8) + 2; 2453 mddev->new_chunk_sectors = mddev->chunk_sectors; 2454 mddev->delta_disks = mddev->raid_disks; 2455 mddev->raid_disks *= 2; 2456 /* make sure it will be not marked as dirty */ 2457 mddev->recovery_cp = MaxSector; 2458 2459 conf = setup_conf(mddev); 2460 if (!IS_ERR(conf)) 2461 list_for_each_entry(rdev, &mddev->disks, same_set) 2462 if (rdev->raid_disk >= 0) 2463 rdev->new_raid_disk = rdev->raid_disk * 2; 2464 2465 return conf; 2466} 2467 2468static void *raid10_takeover(mddev_t *mddev) 2469{ 2470 struct raid0_private_data *raid0_priv; 2471 2472 /* raid10 can take over: 2473 * raid0 - providing it has only two drives 2474 */ 2475 if (mddev->level == 0) { 2476 /* for raid0 takeover only one zone is supported */ 2477 raid0_priv = mddev->private; 2478 if (raid0_priv->nr_strip_zones > 1) { 2479 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0" 2480 " with more than one zone.\n", 2481 mdname(mddev)); 2482 return ERR_PTR(-EINVAL); 2483 } 2484 return raid10_takeover_raid0(mddev); 2485 } 2486 return ERR_PTR(-EINVAL); 2487} 2488 2489static struct mdk_personality raid10_personality = 2490{ 2491 .name = "raid10", 2492 .level = 10, 2493 .owner = THIS_MODULE, 2494 .make_request = make_request, 2495 .run = run, 2496 .stop = stop, 2497 .status = status, 2498 .error_handler = error, 2499 .hot_add_disk = raid10_add_disk, 2500 .hot_remove_disk= raid10_remove_disk, 2501 .spare_active = raid10_spare_active, 2502 .sync_request = sync_request, 2503 .quiesce = raid10_quiesce, 2504 .size = raid10_size, 2505 .takeover = raid10_takeover, 2506}; 2507 2508static int __init raid_init(void) 2509{ 2510 return register_md_personality(&raid10_personality); 2511} 2512 2513static void raid_exit(void) 2514{ 2515 unregister_md_personality(&raid10_personality); 2516} 2517 2518module_init(raid_init); 2519module_exit(raid_exit); 2520MODULE_LICENSE("GPL"); 2521MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD"); 2522MODULE_ALIAS("md-personality-9"); /* RAID10 */ 2523MODULE_ALIAS("md-raid10"); 2524MODULE_ALIAS("md-level-10"); 2525