raid5.c revision c5a3100062cf277d3edd4e6f4a1f1e403524b464
1/* 2 * raid5.c : Multiple Devices driver for Linux 3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman 4 * Copyright (C) 1999, 2000 Ingo Molnar 5 * Copyright (C) 2002, 2003 H. Peter Anvin 6 * 7 * RAID-4/5/6 management functions. 8 * Thanks to Penguin Computing for making the RAID-6 development possible 9 * by donating a test server! 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/* 22 * BITMAP UNPLUGGING: 23 * 24 * The sequencing for updating the bitmap reliably is a little 25 * subtle (and I got it wrong the first time) so it deserves some 26 * explanation. 27 * 28 * We group bitmap updates into batches. Each batch has a number. 29 * We may write out several batches at once, but that isn't very important. 30 * conf->seq_write is the number of the last batch successfully written. 31 * conf->seq_flush is the number of the last batch that was closed to 32 * new additions. 33 * When we discover that we will need to write to any block in a stripe 34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq 35 * the number of the batch it will be in. This is seq_flush+1. 36 * When we are ready to do a write, if that batch hasn't been written yet, 37 * we plug the array and queue the stripe for later. 38 * When an unplug happens, we increment bm_flush, thus closing the current 39 * batch. 40 * When we notice that bm_flush > bm_write, we write out all pending updates 41 * to the bitmap, and advance bm_write to where bm_flush was. 42 * This may occasionally write a bit out twice, but is sure never to 43 * miss any bits. 44 */ 45 46#include <linux/blkdev.h> 47#include <linux/kthread.h> 48#include <linux/raid/pq.h> 49#include <linux/async_tx.h> 50#include <linux/async.h> 51#include <linux/seq_file.h> 52#include <linux/cpu.h> 53#include <linux/slab.h> 54#include "md.h" 55#include "raid5.h" 56#include "raid0.h" 57#include "bitmap.h" 58 59/* 60 * Stripe cache 61 */ 62 63#define NR_STRIPES 256 64#define STRIPE_SIZE PAGE_SIZE 65#define STRIPE_SHIFT (PAGE_SHIFT - 9) 66#define STRIPE_SECTORS (STRIPE_SIZE>>9) 67#define IO_THRESHOLD 1 68#define BYPASS_THRESHOLD 1 69#define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head)) 70#define HASH_MASK (NR_HASH - 1) 71 72#define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK])) 73 74/* bio's attached to a stripe+device for I/O are linked together in bi_sector 75 * order without overlap. There may be several bio's per stripe+device, and 76 * a bio could span several devices. 77 * When walking this list for a particular stripe+device, we must never proceed 78 * beyond a bio that extends past this device, as the next bio might no longer 79 * be valid. 80 * This macro is used to determine the 'next' bio in the list, given the sector 81 * of the current stripe+device 82 */ 83#define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL) 84/* 85 * The following can be used to debug the driver 86 */ 87#define RAID5_PARANOIA 1 88#if RAID5_PARANOIA && defined(CONFIG_SMP) 89# define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock) 90#else 91# define CHECK_DEVLOCK() 92#endif 93 94#ifdef DEBUG 95#define inline 96#define __inline__ 97#endif 98 99#define printk_rl(args...) ((void) (printk_ratelimit() && printk(args))) 100 101/* 102 * We maintain a biased count of active stripes in the bottom 16 bits of 103 * bi_phys_segments, and a count of processed stripes in the upper 16 bits 104 */ 105static inline int raid5_bi_phys_segments(struct bio *bio) 106{ 107 return bio->bi_phys_segments & 0xffff; 108} 109 110static inline int raid5_bi_hw_segments(struct bio *bio) 111{ 112 return (bio->bi_phys_segments >> 16) & 0xffff; 113} 114 115static inline int raid5_dec_bi_phys_segments(struct bio *bio) 116{ 117 --bio->bi_phys_segments; 118 return raid5_bi_phys_segments(bio); 119} 120 121static inline int raid5_dec_bi_hw_segments(struct bio *bio) 122{ 123 unsigned short val = raid5_bi_hw_segments(bio); 124 125 --val; 126 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio); 127 return val; 128} 129 130static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt) 131{ 132 bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16); 133} 134 135/* Find first data disk in a raid6 stripe */ 136static inline int raid6_d0(struct stripe_head *sh) 137{ 138 if (sh->ddf_layout) 139 /* ddf always start from first device */ 140 return 0; 141 /* md starts just after Q block */ 142 if (sh->qd_idx == sh->disks - 1) 143 return 0; 144 else 145 return sh->qd_idx + 1; 146} 147static inline int raid6_next_disk(int disk, int raid_disks) 148{ 149 disk++; 150 return (disk < raid_disks) ? disk : 0; 151} 152 153/* When walking through the disks in a raid5, starting at raid6_d0, 154 * We need to map each disk to a 'slot', where the data disks are slot 155 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk 156 * is raid_disks-1. This help does that mapping. 157 */ 158static int raid6_idx_to_slot(int idx, struct stripe_head *sh, 159 int *count, int syndrome_disks) 160{ 161 int slot = *count; 162 163 if (sh->ddf_layout) 164 (*count)++; 165 if (idx == sh->pd_idx) 166 return syndrome_disks; 167 if (idx == sh->qd_idx) 168 return syndrome_disks + 1; 169 if (!sh->ddf_layout) 170 (*count)++; 171 return slot; 172} 173 174static void return_io(struct bio *return_bi) 175{ 176 struct bio *bi = return_bi; 177 while (bi) { 178 179 return_bi = bi->bi_next; 180 bi->bi_next = NULL; 181 bi->bi_size = 0; 182 bio_endio(bi, 0); 183 bi = return_bi; 184 } 185} 186 187static void print_raid5_conf (raid5_conf_t *conf); 188 189static int stripe_operations_active(struct stripe_head *sh) 190{ 191 return sh->check_state || sh->reconstruct_state || 192 test_bit(STRIPE_BIOFILL_RUN, &sh->state) || 193 test_bit(STRIPE_COMPUTE_RUN, &sh->state); 194} 195 196static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh) 197{ 198 if (atomic_dec_and_test(&sh->count)) { 199 BUG_ON(!list_empty(&sh->lru)); 200 BUG_ON(atomic_read(&conf->active_stripes)==0); 201 if (test_bit(STRIPE_HANDLE, &sh->state)) { 202 if (test_bit(STRIPE_DELAYED, &sh->state)) 203 list_add_tail(&sh->lru, &conf->delayed_list); 204 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) && 205 sh->bm_seq - conf->seq_write > 0) 206 list_add_tail(&sh->lru, &conf->bitmap_list); 207 else { 208 clear_bit(STRIPE_BIT_DELAY, &sh->state); 209 list_add_tail(&sh->lru, &conf->handle_list); 210 } 211 md_wakeup_thread(conf->mddev->thread); 212 } else { 213 BUG_ON(stripe_operations_active(sh)); 214 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 215 atomic_dec(&conf->preread_active_stripes); 216 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) 217 md_wakeup_thread(conf->mddev->thread); 218 } 219 atomic_dec(&conf->active_stripes); 220 if (!test_bit(STRIPE_EXPANDING, &sh->state)) { 221 list_add_tail(&sh->lru, &conf->inactive_list); 222 wake_up(&conf->wait_for_stripe); 223 if (conf->retry_read_aligned) 224 md_wakeup_thread(conf->mddev->thread); 225 } 226 } 227 } 228} 229 230static void release_stripe(struct stripe_head *sh) 231{ 232 raid5_conf_t *conf = sh->raid_conf; 233 unsigned long flags; 234 235 spin_lock_irqsave(&conf->device_lock, flags); 236 __release_stripe(conf, sh); 237 spin_unlock_irqrestore(&conf->device_lock, flags); 238} 239 240static inline void remove_hash(struct stripe_head *sh) 241{ 242 pr_debug("remove_hash(), stripe %llu\n", 243 (unsigned long long)sh->sector); 244 245 hlist_del_init(&sh->hash); 246} 247 248static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh) 249{ 250 struct hlist_head *hp = stripe_hash(conf, sh->sector); 251 252 pr_debug("insert_hash(), stripe %llu\n", 253 (unsigned long long)sh->sector); 254 255 CHECK_DEVLOCK(); 256 hlist_add_head(&sh->hash, hp); 257} 258 259 260/* find an idle stripe, make sure it is unhashed, and return it. */ 261static struct stripe_head *get_free_stripe(raid5_conf_t *conf) 262{ 263 struct stripe_head *sh = NULL; 264 struct list_head *first; 265 266 CHECK_DEVLOCK(); 267 if (list_empty(&conf->inactive_list)) 268 goto out; 269 first = conf->inactive_list.next; 270 sh = list_entry(first, struct stripe_head, lru); 271 list_del_init(first); 272 remove_hash(sh); 273 atomic_inc(&conf->active_stripes); 274out: 275 return sh; 276} 277 278static void shrink_buffers(struct stripe_head *sh) 279{ 280 struct page *p; 281 int i; 282 int num = sh->raid_conf->pool_size; 283 284 for (i = 0; i < num ; i++) { 285 p = sh->dev[i].page; 286 if (!p) 287 continue; 288 sh->dev[i].page = NULL; 289 put_page(p); 290 } 291} 292 293static int grow_buffers(struct stripe_head *sh) 294{ 295 int i; 296 int num = sh->raid_conf->pool_size; 297 298 for (i = 0; i < num; i++) { 299 struct page *page; 300 301 if (!(page = alloc_page(GFP_KERNEL))) { 302 return 1; 303 } 304 sh->dev[i].page = page; 305 } 306 return 0; 307} 308 309static void raid5_build_block(struct stripe_head *sh, int i, int previous); 310static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous, 311 struct stripe_head *sh); 312 313static void init_stripe(struct stripe_head *sh, sector_t sector, int previous) 314{ 315 raid5_conf_t *conf = sh->raid_conf; 316 int i; 317 318 BUG_ON(atomic_read(&sh->count) != 0); 319 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state)); 320 BUG_ON(stripe_operations_active(sh)); 321 322 CHECK_DEVLOCK(); 323 pr_debug("init_stripe called, stripe %llu\n", 324 (unsigned long long)sh->sector); 325 326 remove_hash(sh); 327 328 sh->generation = conf->generation - previous; 329 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks; 330 sh->sector = sector; 331 stripe_set_idx(sector, conf, previous, sh); 332 sh->state = 0; 333 334 335 for (i = sh->disks; i--; ) { 336 struct r5dev *dev = &sh->dev[i]; 337 338 if (dev->toread || dev->read || dev->towrite || dev->written || 339 test_bit(R5_LOCKED, &dev->flags)) { 340 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n", 341 (unsigned long long)sh->sector, i, dev->toread, 342 dev->read, dev->towrite, dev->written, 343 test_bit(R5_LOCKED, &dev->flags)); 344 BUG(); 345 } 346 dev->flags = 0; 347 raid5_build_block(sh, i, previous); 348 } 349 insert_hash(conf, sh); 350} 351 352static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, 353 short generation) 354{ 355 struct stripe_head *sh; 356 struct hlist_node *hn; 357 358 CHECK_DEVLOCK(); 359 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector); 360 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash) 361 if (sh->sector == sector && sh->generation == generation) 362 return sh; 363 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector); 364 return NULL; 365} 366 367/* 368 * Need to check if array has failed when deciding whether to: 369 * - start an array 370 * - remove non-faulty devices 371 * - add a spare 372 * - allow a reshape 373 * This determination is simple when no reshape is happening. 374 * However if there is a reshape, we need to carefully check 375 * both the before and after sections. 376 * This is because some failed devices may only affect one 377 * of the two sections, and some non-in_sync devices may 378 * be insync in the section most affected by failed devices. 379 */ 380static int has_failed(raid5_conf_t *conf) 381{ 382 int degraded; 383 int i; 384 if (conf->mddev->reshape_position == MaxSector) 385 return conf->mddev->degraded > conf->max_degraded; 386 387 rcu_read_lock(); 388 degraded = 0; 389 for (i = 0; i < conf->previous_raid_disks; i++) { 390 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev); 391 if (!rdev || test_bit(Faulty, &rdev->flags)) 392 degraded++; 393 else if (test_bit(In_sync, &rdev->flags)) 394 ; 395 else 396 /* not in-sync or faulty. 397 * If the reshape increases the number of devices, 398 * this is being recovered by the reshape, so 399 * this 'previous' section is not in_sync. 400 * If the number of devices is being reduced however, 401 * the device can only be part of the array if 402 * we are reverting a reshape, so this section will 403 * be in-sync. 404 */ 405 if (conf->raid_disks >= conf->previous_raid_disks) 406 degraded++; 407 } 408 rcu_read_unlock(); 409 if (degraded > conf->max_degraded) 410 return 1; 411 rcu_read_lock(); 412 degraded = 0; 413 for (i = 0; i < conf->raid_disks; i++) { 414 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev); 415 if (!rdev || test_bit(Faulty, &rdev->flags)) 416 degraded++; 417 else if (test_bit(In_sync, &rdev->flags)) 418 ; 419 else 420 /* not in-sync or faulty. 421 * If reshape increases the number of devices, this 422 * section has already been recovered, else it 423 * almost certainly hasn't. 424 */ 425 if (conf->raid_disks <= conf->previous_raid_disks) 426 degraded++; 427 } 428 rcu_read_unlock(); 429 if (degraded > conf->max_degraded) 430 return 1; 431 return 0; 432} 433 434static struct stripe_head * 435get_active_stripe(raid5_conf_t *conf, sector_t sector, 436 int previous, int noblock, int noquiesce) 437{ 438 struct stripe_head *sh; 439 440 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector); 441 442 spin_lock_irq(&conf->device_lock); 443 444 do { 445 wait_event_lock_irq(conf->wait_for_stripe, 446 conf->quiesce == 0 || noquiesce, 447 conf->device_lock, /* nothing */); 448 sh = __find_stripe(conf, sector, conf->generation - previous); 449 if (!sh) { 450 if (!conf->inactive_blocked) 451 sh = get_free_stripe(conf); 452 if (noblock && sh == NULL) 453 break; 454 if (!sh) { 455 conf->inactive_blocked = 1; 456 wait_event_lock_irq(conf->wait_for_stripe, 457 !list_empty(&conf->inactive_list) && 458 (atomic_read(&conf->active_stripes) 459 < (conf->max_nr_stripes *3/4) 460 || !conf->inactive_blocked), 461 conf->device_lock, 462 ); 463 conf->inactive_blocked = 0; 464 } else 465 init_stripe(sh, sector, previous); 466 } else { 467 if (atomic_read(&sh->count)) { 468 BUG_ON(!list_empty(&sh->lru) 469 && !test_bit(STRIPE_EXPANDING, &sh->state)); 470 } else { 471 if (!test_bit(STRIPE_HANDLE, &sh->state)) 472 atomic_inc(&conf->active_stripes); 473 if (list_empty(&sh->lru) && 474 !test_bit(STRIPE_EXPANDING, &sh->state)) 475 BUG(); 476 list_del_init(&sh->lru); 477 } 478 } 479 } while (sh == NULL); 480 481 if (sh) 482 atomic_inc(&sh->count); 483 484 spin_unlock_irq(&conf->device_lock); 485 return sh; 486} 487 488static void 489raid5_end_read_request(struct bio *bi, int error); 490static void 491raid5_end_write_request(struct bio *bi, int error); 492 493static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s) 494{ 495 raid5_conf_t *conf = sh->raid_conf; 496 int i, disks = sh->disks; 497 498 might_sleep(); 499 500 for (i = disks; i--; ) { 501 int rw; 502 struct bio *bi; 503 mdk_rdev_t *rdev; 504 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) { 505 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags)) 506 rw = WRITE_FUA; 507 else 508 rw = WRITE; 509 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags)) 510 rw = READ; 511 else 512 continue; 513 514 bi = &sh->dev[i].req; 515 516 bi->bi_rw = rw; 517 if (rw & WRITE) 518 bi->bi_end_io = raid5_end_write_request; 519 else 520 bi->bi_end_io = raid5_end_read_request; 521 522 rcu_read_lock(); 523 rdev = rcu_dereference(conf->disks[i].rdev); 524 if (rdev && test_bit(Faulty, &rdev->flags)) 525 rdev = NULL; 526 if (rdev) 527 atomic_inc(&rdev->nr_pending); 528 rcu_read_unlock(); 529 530 if (rdev) { 531 if (s->syncing || s->expanding || s->expanded) 532 md_sync_acct(rdev->bdev, STRIPE_SECTORS); 533 534 set_bit(STRIPE_IO_STARTED, &sh->state); 535 536 bi->bi_bdev = rdev->bdev; 537 pr_debug("%s: for %llu schedule op %ld on disc %d\n", 538 __func__, (unsigned long long)sh->sector, 539 bi->bi_rw, i); 540 atomic_inc(&sh->count); 541 bi->bi_sector = sh->sector + rdev->data_offset; 542 bi->bi_flags = 1 << BIO_UPTODATE; 543 bi->bi_vcnt = 1; 544 bi->bi_max_vecs = 1; 545 bi->bi_idx = 0; 546 bi->bi_io_vec = &sh->dev[i].vec; 547 bi->bi_io_vec[0].bv_len = STRIPE_SIZE; 548 bi->bi_io_vec[0].bv_offset = 0; 549 bi->bi_size = STRIPE_SIZE; 550 bi->bi_next = NULL; 551 if ((rw & WRITE) && 552 test_bit(R5_ReWrite, &sh->dev[i].flags)) 553 atomic_add(STRIPE_SECTORS, 554 &rdev->corrected_errors); 555 generic_make_request(bi); 556 } else { 557 if (rw & WRITE) 558 set_bit(STRIPE_DEGRADED, &sh->state); 559 pr_debug("skip op %ld on disc %d for sector %llu\n", 560 bi->bi_rw, i, (unsigned long long)sh->sector); 561 clear_bit(R5_LOCKED, &sh->dev[i].flags); 562 set_bit(STRIPE_HANDLE, &sh->state); 563 } 564 } 565} 566 567static struct dma_async_tx_descriptor * 568async_copy_data(int frombio, struct bio *bio, struct page *page, 569 sector_t sector, struct dma_async_tx_descriptor *tx) 570{ 571 struct bio_vec *bvl; 572 struct page *bio_page; 573 int i; 574 int page_offset; 575 struct async_submit_ctl submit; 576 enum async_tx_flags flags = 0; 577 578 if (bio->bi_sector >= sector) 579 page_offset = (signed)(bio->bi_sector - sector) * 512; 580 else 581 page_offset = (signed)(sector - bio->bi_sector) * -512; 582 583 if (frombio) 584 flags |= ASYNC_TX_FENCE; 585 init_async_submit(&submit, flags, tx, NULL, NULL, NULL); 586 587 bio_for_each_segment(bvl, bio, i) { 588 int len = bvl->bv_len; 589 int clen; 590 int b_offset = 0; 591 592 if (page_offset < 0) { 593 b_offset = -page_offset; 594 page_offset += b_offset; 595 len -= b_offset; 596 } 597 598 if (len > 0 && page_offset + len > STRIPE_SIZE) 599 clen = STRIPE_SIZE - page_offset; 600 else 601 clen = len; 602 603 if (clen > 0) { 604 b_offset += bvl->bv_offset; 605 bio_page = bvl->bv_page; 606 if (frombio) 607 tx = async_memcpy(page, bio_page, page_offset, 608 b_offset, clen, &submit); 609 else 610 tx = async_memcpy(bio_page, page, b_offset, 611 page_offset, clen, &submit); 612 } 613 /* chain the operations */ 614 submit.depend_tx = tx; 615 616 if (clen < len) /* hit end of page */ 617 break; 618 page_offset += len; 619 } 620 621 return tx; 622} 623 624static void ops_complete_biofill(void *stripe_head_ref) 625{ 626 struct stripe_head *sh = stripe_head_ref; 627 struct bio *return_bi = NULL; 628 raid5_conf_t *conf = sh->raid_conf; 629 int i; 630 631 pr_debug("%s: stripe %llu\n", __func__, 632 (unsigned long long)sh->sector); 633 634 /* clear completed biofills */ 635 spin_lock_irq(&conf->device_lock); 636 for (i = sh->disks; i--; ) { 637 struct r5dev *dev = &sh->dev[i]; 638 639 /* acknowledge completion of a biofill operation */ 640 /* and check if we need to reply to a read request, 641 * new R5_Wantfill requests are held off until 642 * !STRIPE_BIOFILL_RUN 643 */ 644 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) { 645 struct bio *rbi, *rbi2; 646 647 BUG_ON(!dev->read); 648 rbi = dev->read; 649 dev->read = NULL; 650 while (rbi && rbi->bi_sector < 651 dev->sector + STRIPE_SECTORS) { 652 rbi2 = r5_next_bio(rbi, dev->sector); 653 if (!raid5_dec_bi_phys_segments(rbi)) { 654 rbi->bi_next = return_bi; 655 return_bi = rbi; 656 } 657 rbi = rbi2; 658 } 659 } 660 } 661 spin_unlock_irq(&conf->device_lock); 662 clear_bit(STRIPE_BIOFILL_RUN, &sh->state); 663 664 return_io(return_bi); 665 666 set_bit(STRIPE_HANDLE, &sh->state); 667 release_stripe(sh); 668} 669 670static void ops_run_biofill(struct stripe_head *sh) 671{ 672 struct dma_async_tx_descriptor *tx = NULL; 673 raid5_conf_t *conf = sh->raid_conf; 674 struct async_submit_ctl submit; 675 int i; 676 677 pr_debug("%s: stripe %llu\n", __func__, 678 (unsigned long long)sh->sector); 679 680 for (i = sh->disks; i--; ) { 681 struct r5dev *dev = &sh->dev[i]; 682 if (test_bit(R5_Wantfill, &dev->flags)) { 683 struct bio *rbi; 684 spin_lock_irq(&conf->device_lock); 685 dev->read = rbi = dev->toread; 686 dev->toread = NULL; 687 spin_unlock_irq(&conf->device_lock); 688 while (rbi && rbi->bi_sector < 689 dev->sector + STRIPE_SECTORS) { 690 tx = async_copy_data(0, rbi, dev->page, 691 dev->sector, tx); 692 rbi = r5_next_bio(rbi, dev->sector); 693 } 694 } 695 } 696 697 atomic_inc(&sh->count); 698 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL); 699 async_trigger_callback(&submit); 700} 701 702static void mark_target_uptodate(struct stripe_head *sh, int target) 703{ 704 struct r5dev *tgt; 705 706 if (target < 0) 707 return; 708 709 tgt = &sh->dev[target]; 710 set_bit(R5_UPTODATE, &tgt->flags); 711 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags)); 712 clear_bit(R5_Wantcompute, &tgt->flags); 713} 714 715static void ops_complete_compute(void *stripe_head_ref) 716{ 717 struct stripe_head *sh = stripe_head_ref; 718 719 pr_debug("%s: stripe %llu\n", __func__, 720 (unsigned long long)sh->sector); 721 722 /* mark the computed target(s) as uptodate */ 723 mark_target_uptodate(sh, sh->ops.target); 724 mark_target_uptodate(sh, sh->ops.target2); 725 726 clear_bit(STRIPE_COMPUTE_RUN, &sh->state); 727 if (sh->check_state == check_state_compute_run) 728 sh->check_state = check_state_compute_result; 729 set_bit(STRIPE_HANDLE, &sh->state); 730 release_stripe(sh); 731} 732 733/* return a pointer to the address conversion region of the scribble buffer */ 734static addr_conv_t *to_addr_conv(struct stripe_head *sh, 735 struct raid5_percpu *percpu) 736{ 737 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2); 738} 739 740static struct dma_async_tx_descriptor * 741ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu) 742{ 743 int disks = sh->disks; 744 struct page **xor_srcs = percpu->scribble; 745 int target = sh->ops.target; 746 struct r5dev *tgt = &sh->dev[target]; 747 struct page *xor_dest = tgt->page; 748 int count = 0; 749 struct dma_async_tx_descriptor *tx; 750 struct async_submit_ctl submit; 751 int i; 752 753 pr_debug("%s: stripe %llu block: %d\n", 754 __func__, (unsigned long long)sh->sector, target); 755 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags)); 756 757 for (i = disks; i--; ) 758 if (i != target) 759 xor_srcs[count++] = sh->dev[i].page; 760 761 atomic_inc(&sh->count); 762 763 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL, 764 ops_complete_compute, sh, to_addr_conv(sh, percpu)); 765 if (unlikely(count == 1)) 766 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit); 767 else 768 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit); 769 770 return tx; 771} 772 773/* set_syndrome_sources - populate source buffers for gen_syndrome 774 * @srcs - (struct page *) array of size sh->disks 775 * @sh - stripe_head to parse 776 * 777 * Populates srcs in proper layout order for the stripe and returns the 778 * 'count' of sources to be used in a call to async_gen_syndrome. The P 779 * destination buffer is recorded in srcs[count] and the Q destination 780 * is recorded in srcs[count+1]]. 781 */ 782static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh) 783{ 784 int disks = sh->disks; 785 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2); 786 int d0_idx = raid6_d0(sh); 787 int count; 788 int i; 789 790 for (i = 0; i < disks; i++) 791 srcs[i] = NULL; 792 793 count = 0; 794 i = d0_idx; 795 do { 796 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks); 797 798 srcs[slot] = sh->dev[i].page; 799 i = raid6_next_disk(i, disks); 800 } while (i != d0_idx); 801 802 return syndrome_disks; 803} 804 805static struct dma_async_tx_descriptor * 806ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu) 807{ 808 int disks = sh->disks; 809 struct page **blocks = percpu->scribble; 810 int target; 811 int qd_idx = sh->qd_idx; 812 struct dma_async_tx_descriptor *tx; 813 struct async_submit_ctl submit; 814 struct r5dev *tgt; 815 struct page *dest; 816 int i; 817 int count; 818 819 if (sh->ops.target < 0) 820 target = sh->ops.target2; 821 else if (sh->ops.target2 < 0) 822 target = sh->ops.target; 823 else 824 /* we should only have one valid target */ 825 BUG(); 826 BUG_ON(target < 0); 827 pr_debug("%s: stripe %llu block: %d\n", 828 __func__, (unsigned long long)sh->sector, target); 829 830 tgt = &sh->dev[target]; 831 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags)); 832 dest = tgt->page; 833 834 atomic_inc(&sh->count); 835 836 if (target == qd_idx) { 837 count = set_syndrome_sources(blocks, sh); 838 blocks[count] = NULL; /* regenerating p is not necessary */ 839 BUG_ON(blocks[count+1] != dest); /* q should already be set */ 840 init_async_submit(&submit, ASYNC_TX_FENCE, NULL, 841 ops_complete_compute, sh, 842 to_addr_conv(sh, percpu)); 843 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit); 844 } else { 845 /* Compute any data- or p-drive using XOR */ 846 count = 0; 847 for (i = disks; i-- ; ) { 848 if (i == target || i == qd_idx) 849 continue; 850 blocks[count++] = sh->dev[i].page; 851 } 852 853 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, 854 NULL, ops_complete_compute, sh, 855 to_addr_conv(sh, percpu)); 856 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit); 857 } 858 859 return tx; 860} 861 862static struct dma_async_tx_descriptor * 863ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu) 864{ 865 int i, count, disks = sh->disks; 866 int syndrome_disks = sh->ddf_layout ? disks : disks-2; 867 int d0_idx = raid6_d0(sh); 868 int faila = -1, failb = -1; 869 int target = sh->ops.target; 870 int target2 = sh->ops.target2; 871 struct r5dev *tgt = &sh->dev[target]; 872 struct r5dev *tgt2 = &sh->dev[target2]; 873 struct dma_async_tx_descriptor *tx; 874 struct page **blocks = percpu->scribble; 875 struct async_submit_ctl submit; 876 877 pr_debug("%s: stripe %llu block1: %d block2: %d\n", 878 __func__, (unsigned long long)sh->sector, target, target2); 879 BUG_ON(target < 0 || target2 < 0); 880 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags)); 881 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags)); 882 883 /* we need to open-code set_syndrome_sources to handle the 884 * slot number conversion for 'faila' and 'failb' 885 */ 886 for (i = 0; i < disks ; i++) 887 blocks[i] = NULL; 888 count = 0; 889 i = d0_idx; 890 do { 891 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks); 892 893 blocks[slot] = sh->dev[i].page; 894 895 if (i == target) 896 faila = slot; 897 if (i == target2) 898 failb = slot; 899 i = raid6_next_disk(i, disks); 900 } while (i != d0_idx); 901 902 BUG_ON(faila == failb); 903 if (failb < faila) 904 swap(faila, failb); 905 pr_debug("%s: stripe: %llu faila: %d failb: %d\n", 906 __func__, (unsigned long long)sh->sector, faila, failb); 907 908 atomic_inc(&sh->count); 909 910 if (failb == syndrome_disks+1) { 911 /* Q disk is one of the missing disks */ 912 if (faila == syndrome_disks) { 913 /* Missing P+Q, just recompute */ 914 init_async_submit(&submit, ASYNC_TX_FENCE, NULL, 915 ops_complete_compute, sh, 916 to_addr_conv(sh, percpu)); 917 return async_gen_syndrome(blocks, 0, syndrome_disks+2, 918 STRIPE_SIZE, &submit); 919 } else { 920 struct page *dest; 921 int data_target; 922 int qd_idx = sh->qd_idx; 923 924 /* Missing D+Q: recompute D from P, then recompute Q */ 925 if (target == qd_idx) 926 data_target = target2; 927 else 928 data_target = target; 929 930 count = 0; 931 for (i = disks; i-- ; ) { 932 if (i == data_target || i == qd_idx) 933 continue; 934 blocks[count++] = sh->dev[i].page; 935 } 936 dest = sh->dev[data_target].page; 937 init_async_submit(&submit, 938 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, 939 NULL, NULL, NULL, 940 to_addr_conv(sh, percpu)); 941 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, 942 &submit); 943 944 count = set_syndrome_sources(blocks, sh); 945 init_async_submit(&submit, ASYNC_TX_FENCE, tx, 946 ops_complete_compute, sh, 947 to_addr_conv(sh, percpu)); 948 return async_gen_syndrome(blocks, 0, count+2, 949 STRIPE_SIZE, &submit); 950 } 951 } else { 952 init_async_submit(&submit, ASYNC_TX_FENCE, NULL, 953 ops_complete_compute, sh, 954 to_addr_conv(sh, percpu)); 955 if (failb == syndrome_disks) { 956 /* We're missing D+P. */ 957 return async_raid6_datap_recov(syndrome_disks+2, 958 STRIPE_SIZE, faila, 959 blocks, &submit); 960 } else { 961 /* We're missing D+D. */ 962 return async_raid6_2data_recov(syndrome_disks+2, 963 STRIPE_SIZE, faila, failb, 964 blocks, &submit); 965 } 966 } 967} 968 969 970static void ops_complete_prexor(void *stripe_head_ref) 971{ 972 struct stripe_head *sh = stripe_head_ref; 973 974 pr_debug("%s: stripe %llu\n", __func__, 975 (unsigned long long)sh->sector); 976} 977 978static struct dma_async_tx_descriptor * 979ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu, 980 struct dma_async_tx_descriptor *tx) 981{ 982 int disks = sh->disks; 983 struct page **xor_srcs = percpu->scribble; 984 int count = 0, pd_idx = sh->pd_idx, i; 985 struct async_submit_ctl submit; 986 987 /* existing parity data subtracted */ 988 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page; 989 990 pr_debug("%s: stripe %llu\n", __func__, 991 (unsigned long long)sh->sector); 992 993 for (i = disks; i--; ) { 994 struct r5dev *dev = &sh->dev[i]; 995 /* Only process blocks that are known to be uptodate */ 996 if (test_bit(R5_Wantdrain, &dev->flags)) 997 xor_srcs[count++] = dev->page; 998 } 999 1000 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx, 1001 ops_complete_prexor, sh, to_addr_conv(sh, percpu)); 1002 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit); 1003 1004 return tx; 1005} 1006 1007static struct dma_async_tx_descriptor * 1008ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx) 1009{ 1010 int disks = sh->disks; 1011 int i; 1012 1013 pr_debug("%s: stripe %llu\n", __func__, 1014 (unsigned long long)sh->sector); 1015 1016 for (i = disks; i--; ) { 1017 struct r5dev *dev = &sh->dev[i]; 1018 struct bio *chosen; 1019 1020 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) { 1021 struct bio *wbi; 1022 1023 spin_lock_irq(&sh->raid_conf->device_lock); 1024 chosen = dev->towrite; 1025 dev->towrite = NULL; 1026 BUG_ON(dev->written); 1027 wbi = dev->written = chosen; 1028 spin_unlock_irq(&sh->raid_conf->device_lock); 1029 1030 while (wbi && wbi->bi_sector < 1031 dev->sector + STRIPE_SECTORS) { 1032 if (wbi->bi_rw & REQ_FUA) 1033 set_bit(R5_WantFUA, &dev->flags); 1034 tx = async_copy_data(1, wbi, dev->page, 1035 dev->sector, tx); 1036 wbi = r5_next_bio(wbi, dev->sector); 1037 } 1038 } 1039 } 1040 1041 return tx; 1042} 1043 1044static void ops_complete_reconstruct(void *stripe_head_ref) 1045{ 1046 struct stripe_head *sh = stripe_head_ref; 1047 int disks = sh->disks; 1048 int pd_idx = sh->pd_idx; 1049 int qd_idx = sh->qd_idx; 1050 int i; 1051 bool fua = false; 1052 1053 pr_debug("%s: stripe %llu\n", __func__, 1054 (unsigned long long)sh->sector); 1055 1056 for (i = disks; i--; ) 1057 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags); 1058 1059 for (i = disks; i--; ) { 1060 struct r5dev *dev = &sh->dev[i]; 1061 1062 if (dev->written || i == pd_idx || i == qd_idx) { 1063 set_bit(R5_UPTODATE, &dev->flags); 1064 if (fua) 1065 set_bit(R5_WantFUA, &dev->flags); 1066 } 1067 } 1068 1069 if (sh->reconstruct_state == reconstruct_state_drain_run) 1070 sh->reconstruct_state = reconstruct_state_drain_result; 1071 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) 1072 sh->reconstruct_state = reconstruct_state_prexor_drain_result; 1073 else { 1074 BUG_ON(sh->reconstruct_state != reconstruct_state_run); 1075 sh->reconstruct_state = reconstruct_state_result; 1076 } 1077 1078 set_bit(STRIPE_HANDLE, &sh->state); 1079 release_stripe(sh); 1080} 1081 1082static void 1083ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu, 1084 struct dma_async_tx_descriptor *tx) 1085{ 1086 int disks = sh->disks; 1087 struct page **xor_srcs = percpu->scribble; 1088 struct async_submit_ctl submit; 1089 int count = 0, pd_idx = sh->pd_idx, i; 1090 struct page *xor_dest; 1091 int prexor = 0; 1092 unsigned long flags; 1093 1094 pr_debug("%s: stripe %llu\n", __func__, 1095 (unsigned long long)sh->sector); 1096 1097 /* check if prexor is active which means only process blocks 1098 * that are part of a read-modify-write (written) 1099 */ 1100 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) { 1101 prexor = 1; 1102 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page; 1103 for (i = disks; i--; ) { 1104 struct r5dev *dev = &sh->dev[i]; 1105 if (dev->written) 1106 xor_srcs[count++] = dev->page; 1107 } 1108 } else { 1109 xor_dest = sh->dev[pd_idx].page; 1110 for (i = disks; i--; ) { 1111 struct r5dev *dev = &sh->dev[i]; 1112 if (i != pd_idx) 1113 xor_srcs[count++] = dev->page; 1114 } 1115 } 1116 1117 /* 1/ if we prexor'd then the dest is reused as a source 1118 * 2/ if we did not prexor then we are redoing the parity 1119 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST 1120 * for the synchronous xor case 1121 */ 1122 flags = ASYNC_TX_ACK | 1123 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST); 1124 1125 atomic_inc(&sh->count); 1126 1127 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh, 1128 to_addr_conv(sh, percpu)); 1129 if (unlikely(count == 1)) 1130 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit); 1131 else 1132 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit); 1133} 1134 1135static void 1136ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu, 1137 struct dma_async_tx_descriptor *tx) 1138{ 1139 struct async_submit_ctl submit; 1140 struct page **blocks = percpu->scribble; 1141 int count; 1142 1143 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector); 1144 1145 count = set_syndrome_sources(blocks, sh); 1146 1147 atomic_inc(&sh->count); 1148 1149 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct, 1150 sh, to_addr_conv(sh, percpu)); 1151 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit); 1152} 1153 1154static void ops_complete_check(void *stripe_head_ref) 1155{ 1156 struct stripe_head *sh = stripe_head_ref; 1157 1158 pr_debug("%s: stripe %llu\n", __func__, 1159 (unsigned long long)sh->sector); 1160 1161 sh->check_state = check_state_check_result; 1162 set_bit(STRIPE_HANDLE, &sh->state); 1163 release_stripe(sh); 1164} 1165 1166static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu) 1167{ 1168 int disks = sh->disks; 1169 int pd_idx = sh->pd_idx; 1170 int qd_idx = sh->qd_idx; 1171 struct page *xor_dest; 1172 struct page **xor_srcs = percpu->scribble; 1173 struct dma_async_tx_descriptor *tx; 1174 struct async_submit_ctl submit; 1175 int count; 1176 int i; 1177 1178 pr_debug("%s: stripe %llu\n", __func__, 1179 (unsigned long long)sh->sector); 1180 1181 count = 0; 1182 xor_dest = sh->dev[pd_idx].page; 1183 xor_srcs[count++] = xor_dest; 1184 for (i = disks; i--; ) { 1185 if (i == pd_idx || i == qd_idx) 1186 continue; 1187 xor_srcs[count++] = sh->dev[i].page; 1188 } 1189 1190 init_async_submit(&submit, 0, NULL, NULL, NULL, 1191 to_addr_conv(sh, percpu)); 1192 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, 1193 &sh->ops.zero_sum_result, &submit); 1194 1195 atomic_inc(&sh->count); 1196 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL); 1197 tx = async_trigger_callback(&submit); 1198} 1199 1200static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp) 1201{ 1202 struct page **srcs = percpu->scribble; 1203 struct async_submit_ctl submit; 1204 int count; 1205 1206 pr_debug("%s: stripe %llu checkp: %d\n", __func__, 1207 (unsigned long long)sh->sector, checkp); 1208 1209 count = set_syndrome_sources(srcs, sh); 1210 if (!checkp) 1211 srcs[count] = NULL; 1212 1213 atomic_inc(&sh->count); 1214 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check, 1215 sh, to_addr_conv(sh, percpu)); 1216 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE, 1217 &sh->ops.zero_sum_result, percpu->spare_page, &submit); 1218} 1219 1220static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request) 1221{ 1222 int overlap_clear = 0, i, disks = sh->disks; 1223 struct dma_async_tx_descriptor *tx = NULL; 1224 raid5_conf_t *conf = sh->raid_conf; 1225 int level = conf->level; 1226 struct raid5_percpu *percpu; 1227 unsigned long cpu; 1228 1229 cpu = get_cpu(); 1230 percpu = per_cpu_ptr(conf->percpu, cpu); 1231 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) { 1232 ops_run_biofill(sh); 1233 overlap_clear++; 1234 } 1235 1236 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) { 1237 if (level < 6) 1238 tx = ops_run_compute5(sh, percpu); 1239 else { 1240 if (sh->ops.target2 < 0 || sh->ops.target < 0) 1241 tx = ops_run_compute6_1(sh, percpu); 1242 else 1243 tx = ops_run_compute6_2(sh, percpu); 1244 } 1245 /* terminate the chain if reconstruct is not set to be run */ 1246 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) 1247 async_tx_ack(tx); 1248 } 1249 1250 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) 1251 tx = ops_run_prexor(sh, percpu, tx); 1252 1253 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) { 1254 tx = ops_run_biodrain(sh, tx); 1255 overlap_clear++; 1256 } 1257 1258 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) { 1259 if (level < 6) 1260 ops_run_reconstruct5(sh, percpu, tx); 1261 else 1262 ops_run_reconstruct6(sh, percpu, tx); 1263 } 1264 1265 if (test_bit(STRIPE_OP_CHECK, &ops_request)) { 1266 if (sh->check_state == check_state_run) 1267 ops_run_check_p(sh, percpu); 1268 else if (sh->check_state == check_state_run_q) 1269 ops_run_check_pq(sh, percpu, 0); 1270 else if (sh->check_state == check_state_run_pq) 1271 ops_run_check_pq(sh, percpu, 1); 1272 else 1273 BUG(); 1274 } 1275 1276 if (overlap_clear) 1277 for (i = disks; i--; ) { 1278 struct r5dev *dev = &sh->dev[i]; 1279 if (test_and_clear_bit(R5_Overlap, &dev->flags)) 1280 wake_up(&sh->raid_conf->wait_for_overlap); 1281 } 1282 put_cpu(); 1283} 1284 1285#ifdef CONFIG_MULTICORE_RAID456 1286static void async_run_ops(void *param, async_cookie_t cookie) 1287{ 1288 struct stripe_head *sh = param; 1289 unsigned long ops_request = sh->ops.request; 1290 1291 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state); 1292 wake_up(&sh->ops.wait_for_ops); 1293 1294 __raid_run_ops(sh, ops_request); 1295 release_stripe(sh); 1296} 1297 1298static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request) 1299{ 1300 /* since handle_stripe can be called outside of raid5d context 1301 * we need to ensure sh->ops.request is de-staged before another 1302 * request arrives 1303 */ 1304 wait_event(sh->ops.wait_for_ops, 1305 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state)); 1306 sh->ops.request = ops_request; 1307 1308 atomic_inc(&sh->count); 1309 async_schedule(async_run_ops, sh); 1310} 1311#else 1312#define raid_run_ops __raid_run_ops 1313#endif 1314 1315static int grow_one_stripe(raid5_conf_t *conf) 1316{ 1317 struct stripe_head *sh; 1318 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL); 1319 if (!sh) 1320 return 0; 1321 1322 sh->raid_conf = conf; 1323 #ifdef CONFIG_MULTICORE_RAID456 1324 init_waitqueue_head(&sh->ops.wait_for_ops); 1325 #endif 1326 1327 if (grow_buffers(sh)) { 1328 shrink_buffers(sh); 1329 kmem_cache_free(conf->slab_cache, sh); 1330 return 0; 1331 } 1332 /* we just created an active stripe so... */ 1333 atomic_set(&sh->count, 1); 1334 atomic_inc(&conf->active_stripes); 1335 INIT_LIST_HEAD(&sh->lru); 1336 release_stripe(sh); 1337 return 1; 1338} 1339 1340static int grow_stripes(raid5_conf_t *conf, int num) 1341{ 1342 struct kmem_cache *sc; 1343 int devs = max(conf->raid_disks, conf->previous_raid_disks); 1344 1345 if (conf->mddev->gendisk) 1346 sprintf(conf->cache_name[0], 1347 "raid%d-%s", conf->level, mdname(conf->mddev)); 1348 else 1349 sprintf(conf->cache_name[0], 1350 "raid%d-%p", conf->level, conf->mddev); 1351 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]); 1352 1353 conf->active_name = 0; 1354 sc = kmem_cache_create(conf->cache_name[conf->active_name], 1355 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev), 1356 0, 0, NULL); 1357 if (!sc) 1358 return 1; 1359 conf->slab_cache = sc; 1360 conf->pool_size = devs; 1361 while (num--) 1362 if (!grow_one_stripe(conf)) 1363 return 1; 1364 return 0; 1365} 1366 1367/** 1368 * scribble_len - return the required size of the scribble region 1369 * @num - total number of disks in the array 1370 * 1371 * The size must be enough to contain: 1372 * 1/ a struct page pointer for each device in the array +2 1373 * 2/ room to convert each entry in (1) to its corresponding dma 1374 * (dma_map_page()) or page (page_address()) address. 1375 * 1376 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we 1377 * calculate over all devices (not just the data blocks), using zeros in place 1378 * of the P and Q blocks. 1379 */ 1380static size_t scribble_len(int num) 1381{ 1382 size_t len; 1383 1384 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2); 1385 1386 return len; 1387} 1388 1389static int resize_stripes(raid5_conf_t *conf, int newsize) 1390{ 1391 /* Make all the stripes able to hold 'newsize' devices. 1392 * New slots in each stripe get 'page' set to a new page. 1393 * 1394 * This happens in stages: 1395 * 1/ create a new kmem_cache and allocate the required number of 1396 * stripe_heads. 1397 * 2/ gather all the old stripe_heads and tranfer the pages across 1398 * to the new stripe_heads. This will have the side effect of 1399 * freezing the array as once all stripe_heads have been collected, 1400 * no IO will be possible. Old stripe heads are freed once their 1401 * pages have been transferred over, and the old kmem_cache is 1402 * freed when all stripes are done. 1403 * 3/ reallocate conf->disks to be suitable bigger. If this fails, 1404 * we simple return a failre status - no need to clean anything up. 1405 * 4/ allocate new pages for the new slots in the new stripe_heads. 1406 * If this fails, we don't bother trying the shrink the 1407 * stripe_heads down again, we just leave them as they are. 1408 * As each stripe_head is processed the new one is released into 1409 * active service. 1410 * 1411 * Once step2 is started, we cannot afford to wait for a write, 1412 * so we use GFP_NOIO allocations. 1413 */ 1414 struct stripe_head *osh, *nsh; 1415 LIST_HEAD(newstripes); 1416 struct disk_info *ndisks; 1417 unsigned long cpu; 1418 int err; 1419 struct kmem_cache *sc; 1420 int i; 1421 1422 if (newsize <= conf->pool_size) 1423 return 0; /* never bother to shrink */ 1424 1425 err = md_allow_write(conf->mddev); 1426 if (err) 1427 return err; 1428 1429 /* Step 1 */ 1430 sc = kmem_cache_create(conf->cache_name[1-conf->active_name], 1431 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev), 1432 0, 0, NULL); 1433 if (!sc) 1434 return -ENOMEM; 1435 1436 for (i = conf->max_nr_stripes; i; i--) { 1437 nsh = kmem_cache_zalloc(sc, GFP_KERNEL); 1438 if (!nsh) 1439 break; 1440 1441 nsh->raid_conf = conf; 1442 #ifdef CONFIG_MULTICORE_RAID456 1443 init_waitqueue_head(&nsh->ops.wait_for_ops); 1444 #endif 1445 1446 list_add(&nsh->lru, &newstripes); 1447 } 1448 if (i) { 1449 /* didn't get enough, give up */ 1450 while (!list_empty(&newstripes)) { 1451 nsh = list_entry(newstripes.next, struct stripe_head, lru); 1452 list_del(&nsh->lru); 1453 kmem_cache_free(sc, nsh); 1454 } 1455 kmem_cache_destroy(sc); 1456 return -ENOMEM; 1457 } 1458 /* Step 2 - Must use GFP_NOIO now. 1459 * OK, we have enough stripes, start collecting inactive 1460 * stripes and copying them over 1461 */ 1462 list_for_each_entry(nsh, &newstripes, lru) { 1463 spin_lock_irq(&conf->device_lock); 1464 wait_event_lock_irq(conf->wait_for_stripe, 1465 !list_empty(&conf->inactive_list), 1466 conf->device_lock, 1467 ); 1468 osh = get_free_stripe(conf); 1469 spin_unlock_irq(&conf->device_lock); 1470 atomic_set(&nsh->count, 1); 1471 for(i=0; i<conf->pool_size; i++) 1472 nsh->dev[i].page = osh->dev[i].page; 1473 for( ; i<newsize; i++) 1474 nsh->dev[i].page = NULL; 1475 kmem_cache_free(conf->slab_cache, osh); 1476 } 1477 kmem_cache_destroy(conf->slab_cache); 1478 1479 /* Step 3. 1480 * At this point, we are holding all the stripes so the array 1481 * is completely stalled, so now is a good time to resize 1482 * conf->disks and the scribble region 1483 */ 1484 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO); 1485 if (ndisks) { 1486 for (i=0; i<conf->raid_disks; i++) 1487 ndisks[i] = conf->disks[i]; 1488 kfree(conf->disks); 1489 conf->disks = ndisks; 1490 } else 1491 err = -ENOMEM; 1492 1493 get_online_cpus(); 1494 conf->scribble_len = scribble_len(newsize); 1495 for_each_present_cpu(cpu) { 1496 struct raid5_percpu *percpu; 1497 void *scribble; 1498 1499 percpu = per_cpu_ptr(conf->percpu, cpu); 1500 scribble = kmalloc(conf->scribble_len, GFP_NOIO); 1501 1502 if (scribble) { 1503 kfree(percpu->scribble); 1504 percpu->scribble = scribble; 1505 } else { 1506 err = -ENOMEM; 1507 break; 1508 } 1509 } 1510 put_online_cpus(); 1511 1512 /* Step 4, return new stripes to service */ 1513 while(!list_empty(&newstripes)) { 1514 nsh = list_entry(newstripes.next, struct stripe_head, lru); 1515 list_del_init(&nsh->lru); 1516 1517 for (i=conf->raid_disks; i < newsize; i++) 1518 if (nsh->dev[i].page == NULL) { 1519 struct page *p = alloc_page(GFP_NOIO); 1520 nsh->dev[i].page = p; 1521 if (!p) 1522 err = -ENOMEM; 1523 } 1524 release_stripe(nsh); 1525 } 1526 /* critical section pass, GFP_NOIO no longer needed */ 1527 1528 conf->slab_cache = sc; 1529 conf->active_name = 1-conf->active_name; 1530 conf->pool_size = newsize; 1531 return err; 1532} 1533 1534static int drop_one_stripe(raid5_conf_t *conf) 1535{ 1536 struct stripe_head *sh; 1537 1538 spin_lock_irq(&conf->device_lock); 1539 sh = get_free_stripe(conf); 1540 spin_unlock_irq(&conf->device_lock); 1541 if (!sh) 1542 return 0; 1543 BUG_ON(atomic_read(&sh->count)); 1544 shrink_buffers(sh); 1545 kmem_cache_free(conf->slab_cache, sh); 1546 atomic_dec(&conf->active_stripes); 1547 return 1; 1548} 1549 1550static void shrink_stripes(raid5_conf_t *conf) 1551{ 1552 while (drop_one_stripe(conf)) 1553 ; 1554 1555 if (conf->slab_cache) 1556 kmem_cache_destroy(conf->slab_cache); 1557 conf->slab_cache = NULL; 1558} 1559 1560static void raid5_end_read_request(struct bio * bi, int error) 1561{ 1562 struct stripe_head *sh = bi->bi_private; 1563 raid5_conf_t *conf = sh->raid_conf; 1564 int disks = sh->disks, i; 1565 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags); 1566 char b[BDEVNAME_SIZE]; 1567 mdk_rdev_t *rdev; 1568 1569 1570 for (i=0 ; i<disks; i++) 1571 if (bi == &sh->dev[i].req) 1572 break; 1573 1574 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n", 1575 (unsigned long long)sh->sector, i, atomic_read(&sh->count), 1576 uptodate); 1577 if (i == disks) { 1578 BUG(); 1579 return; 1580 } 1581 1582 if (uptodate) { 1583 set_bit(R5_UPTODATE, &sh->dev[i].flags); 1584 if (test_bit(R5_ReadError, &sh->dev[i].flags)) { 1585 rdev = conf->disks[i].rdev; 1586 printk_rl(KERN_INFO "md/raid:%s: read error corrected" 1587 " (%lu sectors at %llu on %s)\n", 1588 mdname(conf->mddev), STRIPE_SECTORS, 1589 (unsigned long long)(sh->sector 1590 + rdev->data_offset), 1591 bdevname(rdev->bdev, b)); 1592 clear_bit(R5_ReadError, &sh->dev[i].flags); 1593 clear_bit(R5_ReWrite, &sh->dev[i].flags); 1594 } 1595 if (atomic_read(&conf->disks[i].rdev->read_errors)) 1596 atomic_set(&conf->disks[i].rdev->read_errors, 0); 1597 } else { 1598 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b); 1599 int retry = 0; 1600 rdev = conf->disks[i].rdev; 1601 1602 clear_bit(R5_UPTODATE, &sh->dev[i].flags); 1603 atomic_inc(&rdev->read_errors); 1604 if (conf->mddev->degraded >= conf->max_degraded) 1605 printk_rl(KERN_WARNING 1606 "md/raid:%s: read error not correctable " 1607 "(sector %llu on %s).\n", 1608 mdname(conf->mddev), 1609 (unsigned long long)(sh->sector 1610 + rdev->data_offset), 1611 bdn); 1612 else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) 1613 /* Oh, no!!! */ 1614 printk_rl(KERN_WARNING 1615 "md/raid:%s: read error NOT corrected!! " 1616 "(sector %llu on %s).\n", 1617 mdname(conf->mddev), 1618 (unsigned long long)(sh->sector 1619 + rdev->data_offset), 1620 bdn); 1621 else if (atomic_read(&rdev->read_errors) 1622 > conf->max_nr_stripes) 1623 printk(KERN_WARNING 1624 "md/raid:%s: Too many read errors, failing device %s.\n", 1625 mdname(conf->mddev), bdn); 1626 else 1627 retry = 1; 1628 if (retry) 1629 set_bit(R5_ReadError, &sh->dev[i].flags); 1630 else { 1631 clear_bit(R5_ReadError, &sh->dev[i].flags); 1632 clear_bit(R5_ReWrite, &sh->dev[i].flags); 1633 md_error(conf->mddev, rdev); 1634 } 1635 } 1636 rdev_dec_pending(conf->disks[i].rdev, conf->mddev); 1637 clear_bit(R5_LOCKED, &sh->dev[i].flags); 1638 set_bit(STRIPE_HANDLE, &sh->state); 1639 release_stripe(sh); 1640} 1641 1642static void raid5_end_write_request(struct bio *bi, int error) 1643{ 1644 struct stripe_head *sh = bi->bi_private; 1645 raid5_conf_t *conf = sh->raid_conf; 1646 int disks = sh->disks, i; 1647 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags); 1648 1649 for (i=0 ; i<disks; i++) 1650 if (bi == &sh->dev[i].req) 1651 break; 1652 1653 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n", 1654 (unsigned long long)sh->sector, i, atomic_read(&sh->count), 1655 uptodate); 1656 if (i == disks) { 1657 BUG(); 1658 return; 1659 } 1660 1661 if (!uptodate) 1662 md_error(conf->mddev, conf->disks[i].rdev); 1663 1664 rdev_dec_pending(conf->disks[i].rdev, conf->mddev); 1665 1666 clear_bit(R5_LOCKED, &sh->dev[i].flags); 1667 set_bit(STRIPE_HANDLE, &sh->state); 1668 release_stripe(sh); 1669} 1670 1671 1672static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous); 1673 1674static void raid5_build_block(struct stripe_head *sh, int i, int previous) 1675{ 1676 struct r5dev *dev = &sh->dev[i]; 1677 1678 bio_init(&dev->req); 1679 dev->req.bi_io_vec = &dev->vec; 1680 dev->req.bi_vcnt++; 1681 dev->req.bi_max_vecs++; 1682 dev->vec.bv_page = dev->page; 1683 dev->vec.bv_len = STRIPE_SIZE; 1684 dev->vec.bv_offset = 0; 1685 1686 dev->req.bi_sector = sh->sector; 1687 dev->req.bi_private = sh; 1688 1689 dev->flags = 0; 1690 dev->sector = compute_blocknr(sh, i, previous); 1691} 1692 1693static void error(mddev_t *mddev, mdk_rdev_t *rdev) 1694{ 1695 char b[BDEVNAME_SIZE]; 1696 raid5_conf_t *conf = mddev->private; 1697 pr_debug("raid456: error called\n"); 1698 1699 if (test_and_clear_bit(In_sync, &rdev->flags)) { 1700 unsigned long flags; 1701 spin_lock_irqsave(&conf->device_lock, flags); 1702 mddev->degraded++; 1703 spin_unlock_irqrestore(&conf->device_lock, flags); 1704 /* 1705 * if recovery was running, make sure it aborts. 1706 */ 1707 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1708 } 1709 set_bit(Faulty, &rdev->flags); 1710 set_bit(MD_CHANGE_DEVS, &mddev->flags); 1711 printk(KERN_ALERT 1712 "md/raid:%s: Disk failure on %s, disabling device.\n" 1713 "md/raid:%s: Operation continuing on %d devices.\n", 1714 mdname(mddev), 1715 bdevname(rdev->bdev, b), 1716 mdname(mddev), 1717 conf->raid_disks - mddev->degraded); 1718} 1719 1720/* 1721 * Input: a 'big' sector number, 1722 * Output: index of the data and parity disk, and the sector # in them. 1723 */ 1724static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector, 1725 int previous, int *dd_idx, 1726 struct stripe_head *sh) 1727{ 1728 sector_t stripe, stripe2; 1729 sector_t chunk_number; 1730 unsigned int chunk_offset; 1731 int pd_idx, qd_idx; 1732 int ddf_layout = 0; 1733 sector_t new_sector; 1734 int algorithm = previous ? conf->prev_algo 1735 : conf->algorithm; 1736 int sectors_per_chunk = previous ? conf->prev_chunk_sectors 1737 : conf->chunk_sectors; 1738 int raid_disks = previous ? conf->previous_raid_disks 1739 : conf->raid_disks; 1740 int data_disks = raid_disks - conf->max_degraded; 1741 1742 /* First compute the information on this sector */ 1743 1744 /* 1745 * Compute the chunk number and the sector offset inside the chunk 1746 */ 1747 chunk_offset = sector_div(r_sector, sectors_per_chunk); 1748 chunk_number = r_sector; 1749 1750 /* 1751 * Compute the stripe number 1752 */ 1753 stripe = chunk_number; 1754 *dd_idx = sector_div(stripe, data_disks); 1755 stripe2 = stripe; 1756 /* 1757 * Select the parity disk based on the user selected algorithm. 1758 */ 1759 pd_idx = qd_idx = ~0; 1760 switch(conf->level) { 1761 case 4: 1762 pd_idx = data_disks; 1763 break; 1764 case 5: 1765 switch (algorithm) { 1766 case ALGORITHM_LEFT_ASYMMETRIC: 1767 pd_idx = data_disks - sector_div(stripe2, raid_disks); 1768 if (*dd_idx >= pd_idx) 1769 (*dd_idx)++; 1770 break; 1771 case ALGORITHM_RIGHT_ASYMMETRIC: 1772 pd_idx = sector_div(stripe2, raid_disks); 1773 if (*dd_idx >= pd_idx) 1774 (*dd_idx)++; 1775 break; 1776 case ALGORITHM_LEFT_SYMMETRIC: 1777 pd_idx = data_disks - sector_div(stripe2, raid_disks); 1778 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks; 1779 break; 1780 case ALGORITHM_RIGHT_SYMMETRIC: 1781 pd_idx = sector_div(stripe2, raid_disks); 1782 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks; 1783 break; 1784 case ALGORITHM_PARITY_0: 1785 pd_idx = 0; 1786 (*dd_idx)++; 1787 break; 1788 case ALGORITHM_PARITY_N: 1789 pd_idx = data_disks; 1790 break; 1791 default: 1792 BUG(); 1793 } 1794 break; 1795 case 6: 1796 1797 switch (algorithm) { 1798 case ALGORITHM_LEFT_ASYMMETRIC: 1799 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks); 1800 qd_idx = pd_idx + 1; 1801 if (pd_idx == raid_disks-1) { 1802 (*dd_idx)++; /* Q D D D P */ 1803 qd_idx = 0; 1804 } else if (*dd_idx >= pd_idx) 1805 (*dd_idx) += 2; /* D D P Q D */ 1806 break; 1807 case ALGORITHM_RIGHT_ASYMMETRIC: 1808 pd_idx = sector_div(stripe2, raid_disks); 1809 qd_idx = pd_idx + 1; 1810 if (pd_idx == raid_disks-1) { 1811 (*dd_idx)++; /* Q D D D P */ 1812 qd_idx = 0; 1813 } else if (*dd_idx >= pd_idx) 1814 (*dd_idx) += 2; /* D D P Q D */ 1815 break; 1816 case ALGORITHM_LEFT_SYMMETRIC: 1817 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks); 1818 qd_idx = (pd_idx + 1) % raid_disks; 1819 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks; 1820 break; 1821 case ALGORITHM_RIGHT_SYMMETRIC: 1822 pd_idx = sector_div(stripe2, raid_disks); 1823 qd_idx = (pd_idx + 1) % raid_disks; 1824 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks; 1825 break; 1826 1827 case ALGORITHM_PARITY_0: 1828 pd_idx = 0; 1829 qd_idx = 1; 1830 (*dd_idx) += 2; 1831 break; 1832 case ALGORITHM_PARITY_N: 1833 pd_idx = data_disks; 1834 qd_idx = data_disks + 1; 1835 break; 1836 1837 case ALGORITHM_ROTATING_ZERO_RESTART: 1838 /* Exactly the same as RIGHT_ASYMMETRIC, but or 1839 * of blocks for computing Q is different. 1840 */ 1841 pd_idx = sector_div(stripe2, raid_disks); 1842 qd_idx = pd_idx + 1; 1843 if (pd_idx == raid_disks-1) { 1844 (*dd_idx)++; /* Q D D D P */ 1845 qd_idx = 0; 1846 } else if (*dd_idx >= pd_idx) 1847 (*dd_idx) += 2; /* D D P Q D */ 1848 ddf_layout = 1; 1849 break; 1850 1851 case ALGORITHM_ROTATING_N_RESTART: 1852 /* Same a left_asymmetric, by first stripe is 1853 * D D D P Q rather than 1854 * Q D D D P 1855 */ 1856 stripe2 += 1; 1857 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks); 1858 qd_idx = pd_idx + 1; 1859 if (pd_idx == raid_disks-1) { 1860 (*dd_idx)++; /* Q D D D P */ 1861 qd_idx = 0; 1862 } else if (*dd_idx >= pd_idx) 1863 (*dd_idx) += 2; /* D D P Q D */ 1864 ddf_layout = 1; 1865 break; 1866 1867 case ALGORITHM_ROTATING_N_CONTINUE: 1868 /* Same as left_symmetric but Q is before P */ 1869 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks); 1870 qd_idx = (pd_idx + raid_disks - 1) % raid_disks; 1871 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks; 1872 ddf_layout = 1; 1873 break; 1874 1875 case ALGORITHM_LEFT_ASYMMETRIC_6: 1876 /* RAID5 left_asymmetric, with Q on last device */ 1877 pd_idx = data_disks - sector_div(stripe2, raid_disks-1); 1878 if (*dd_idx >= pd_idx) 1879 (*dd_idx)++; 1880 qd_idx = raid_disks - 1; 1881 break; 1882 1883 case ALGORITHM_RIGHT_ASYMMETRIC_6: 1884 pd_idx = sector_div(stripe2, raid_disks-1); 1885 if (*dd_idx >= pd_idx) 1886 (*dd_idx)++; 1887 qd_idx = raid_disks - 1; 1888 break; 1889 1890 case ALGORITHM_LEFT_SYMMETRIC_6: 1891 pd_idx = data_disks - sector_div(stripe2, raid_disks-1); 1892 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1); 1893 qd_idx = raid_disks - 1; 1894 break; 1895 1896 case ALGORITHM_RIGHT_SYMMETRIC_6: 1897 pd_idx = sector_div(stripe2, raid_disks-1); 1898 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1); 1899 qd_idx = raid_disks - 1; 1900 break; 1901 1902 case ALGORITHM_PARITY_0_6: 1903 pd_idx = 0; 1904 (*dd_idx)++; 1905 qd_idx = raid_disks - 1; 1906 break; 1907 1908 default: 1909 BUG(); 1910 } 1911 break; 1912 } 1913 1914 if (sh) { 1915 sh->pd_idx = pd_idx; 1916 sh->qd_idx = qd_idx; 1917 sh->ddf_layout = ddf_layout; 1918 } 1919 /* 1920 * Finally, compute the new sector number 1921 */ 1922 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset; 1923 return new_sector; 1924} 1925 1926 1927static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous) 1928{ 1929 raid5_conf_t *conf = sh->raid_conf; 1930 int raid_disks = sh->disks; 1931 int data_disks = raid_disks - conf->max_degraded; 1932 sector_t new_sector = sh->sector, check; 1933 int sectors_per_chunk = previous ? conf->prev_chunk_sectors 1934 : conf->chunk_sectors; 1935 int algorithm = previous ? conf->prev_algo 1936 : conf->algorithm; 1937 sector_t stripe; 1938 int chunk_offset; 1939 sector_t chunk_number; 1940 int dummy1, dd_idx = i; 1941 sector_t r_sector; 1942 struct stripe_head sh2; 1943 1944 1945 chunk_offset = sector_div(new_sector, sectors_per_chunk); 1946 stripe = new_sector; 1947 1948 if (i == sh->pd_idx) 1949 return 0; 1950 switch(conf->level) { 1951 case 4: break; 1952 case 5: 1953 switch (algorithm) { 1954 case ALGORITHM_LEFT_ASYMMETRIC: 1955 case ALGORITHM_RIGHT_ASYMMETRIC: 1956 if (i > sh->pd_idx) 1957 i--; 1958 break; 1959 case ALGORITHM_LEFT_SYMMETRIC: 1960 case ALGORITHM_RIGHT_SYMMETRIC: 1961 if (i < sh->pd_idx) 1962 i += raid_disks; 1963 i -= (sh->pd_idx + 1); 1964 break; 1965 case ALGORITHM_PARITY_0: 1966 i -= 1; 1967 break; 1968 case ALGORITHM_PARITY_N: 1969 break; 1970 default: 1971 BUG(); 1972 } 1973 break; 1974 case 6: 1975 if (i == sh->qd_idx) 1976 return 0; /* It is the Q disk */ 1977 switch (algorithm) { 1978 case ALGORITHM_LEFT_ASYMMETRIC: 1979 case ALGORITHM_RIGHT_ASYMMETRIC: 1980 case ALGORITHM_ROTATING_ZERO_RESTART: 1981 case ALGORITHM_ROTATING_N_RESTART: 1982 if (sh->pd_idx == raid_disks-1) 1983 i--; /* Q D D D P */ 1984 else if (i > sh->pd_idx) 1985 i -= 2; /* D D P Q D */ 1986 break; 1987 case ALGORITHM_LEFT_SYMMETRIC: 1988 case ALGORITHM_RIGHT_SYMMETRIC: 1989 if (sh->pd_idx == raid_disks-1) 1990 i--; /* Q D D D P */ 1991 else { 1992 /* D D P Q D */ 1993 if (i < sh->pd_idx) 1994 i += raid_disks; 1995 i -= (sh->pd_idx + 2); 1996 } 1997 break; 1998 case ALGORITHM_PARITY_0: 1999 i -= 2; 2000 break; 2001 case ALGORITHM_PARITY_N: 2002 break; 2003 case ALGORITHM_ROTATING_N_CONTINUE: 2004 /* Like left_symmetric, but P is before Q */ 2005 if (sh->pd_idx == 0) 2006 i--; /* P D D D Q */ 2007 else { 2008 /* D D Q P D */ 2009 if (i < sh->pd_idx) 2010 i += raid_disks; 2011 i -= (sh->pd_idx + 1); 2012 } 2013 break; 2014 case ALGORITHM_LEFT_ASYMMETRIC_6: 2015 case ALGORITHM_RIGHT_ASYMMETRIC_6: 2016 if (i > sh->pd_idx) 2017 i--; 2018 break; 2019 case ALGORITHM_LEFT_SYMMETRIC_6: 2020 case ALGORITHM_RIGHT_SYMMETRIC_6: 2021 if (i < sh->pd_idx) 2022 i += data_disks + 1; 2023 i -= (sh->pd_idx + 1); 2024 break; 2025 case ALGORITHM_PARITY_0_6: 2026 i -= 1; 2027 break; 2028 default: 2029 BUG(); 2030 } 2031 break; 2032 } 2033 2034 chunk_number = stripe * data_disks + i; 2035 r_sector = chunk_number * sectors_per_chunk + chunk_offset; 2036 2037 check = raid5_compute_sector(conf, r_sector, 2038 previous, &dummy1, &sh2); 2039 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx 2040 || sh2.qd_idx != sh->qd_idx) { 2041 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n", 2042 mdname(conf->mddev)); 2043 return 0; 2044 } 2045 return r_sector; 2046} 2047 2048 2049static void 2050schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s, 2051 int rcw, int expand) 2052{ 2053 int i, pd_idx = sh->pd_idx, disks = sh->disks; 2054 raid5_conf_t *conf = sh->raid_conf; 2055 int level = conf->level; 2056 2057 if (rcw) { 2058 /* if we are not expanding this is a proper write request, and 2059 * there will be bios with new data to be drained into the 2060 * stripe cache 2061 */ 2062 if (!expand) { 2063 sh->reconstruct_state = reconstruct_state_drain_run; 2064 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request); 2065 } else 2066 sh->reconstruct_state = reconstruct_state_run; 2067 2068 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request); 2069 2070 for (i = disks; i--; ) { 2071 struct r5dev *dev = &sh->dev[i]; 2072 2073 if (dev->towrite) { 2074 set_bit(R5_LOCKED, &dev->flags); 2075 set_bit(R5_Wantdrain, &dev->flags); 2076 if (!expand) 2077 clear_bit(R5_UPTODATE, &dev->flags); 2078 s->locked++; 2079 } 2080 } 2081 if (s->locked + conf->max_degraded == disks) 2082 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state)) 2083 atomic_inc(&conf->pending_full_writes); 2084 } else { 2085 BUG_ON(level == 6); 2086 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) || 2087 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags))); 2088 2089 sh->reconstruct_state = reconstruct_state_prexor_drain_run; 2090 set_bit(STRIPE_OP_PREXOR, &s->ops_request); 2091 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request); 2092 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request); 2093 2094 for (i = disks; i--; ) { 2095 struct r5dev *dev = &sh->dev[i]; 2096 if (i == pd_idx) 2097 continue; 2098 2099 if (dev->towrite && 2100 (test_bit(R5_UPTODATE, &dev->flags) || 2101 test_bit(R5_Wantcompute, &dev->flags))) { 2102 set_bit(R5_Wantdrain, &dev->flags); 2103 set_bit(R5_LOCKED, &dev->flags); 2104 clear_bit(R5_UPTODATE, &dev->flags); 2105 s->locked++; 2106 } 2107 } 2108 } 2109 2110 /* keep the parity disk(s) locked while asynchronous operations 2111 * are in flight 2112 */ 2113 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags); 2114 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags); 2115 s->locked++; 2116 2117 if (level == 6) { 2118 int qd_idx = sh->qd_idx; 2119 struct r5dev *dev = &sh->dev[qd_idx]; 2120 2121 set_bit(R5_LOCKED, &dev->flags); 2122 clear_bit(R5_UPTODATE, &dev->flags); 2123 s->locked++; 2124 } 2125 2126 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n", 2127 __func__, (unsigned long long)sh->sector, 2128 s->locked, s->ops_request); 2129} 2130 2131/* 2132 * Each stripe/dev can have one or more bion attached. 2133 * toread/towrite point to the first in a chain. 2134 * The bi_next chain must be in order. 2135 */ 2136static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite) 2137{ 2138 struct bio **bip; 2139 raid5_conf_t *conf = sh->raid_conf; 2140 int firstwrite=0; 2141 2142 pr_debug("adding bi b#%llu to stripe s#%llu\n", 2143 (unsigned long long)bi->bi_sector, 2144 (unsigned long long)sh->sector); 2145 2146 2147 spin_lock_irq(&conf->device_lock); 2148 if (forwrite) { 2149 bip = &sh->dev[dd_idx].towrite; 2150 if (*bip == NULL && sh->dev[dd_idx].written == NULL) 2151 firstwrite = 1; 2152 } else 2153 bip = &sh->dev[dd_idx].toread; 2154 while (*bip && (*bip)->bi_sector < bi->bi_sector) { 2155 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector) 2156 goto overlap; 2157 bip = & (*bip)->bi_next; 2158 } 2159 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9)) 2160 goto overlap; 2161 2162 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next); 2163 if (*bip) 2164 bi->bi_next = *bip; 2165 *bip = bi; 2166 bi->bi_phys_segments++; 2167 2168 if (forwrite) { 2169 /* check if page is covered */ 2170 sector_t sector = sh->dev[dd_idx].sector; 2171 for (bi=sh->dev[dd_idx].towrite; 2172 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS && 2173 bi && bi->bi_sector <= sector; 2174 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) { 2175 if (bi->bi_sector + (bi->bi_size>>9) >= sector) 2176 sector = bi->bi_sector + (bi->bi_size>>9); 2177 } 2178 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS) 2179 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags); 2180 } 2181 spin_unlock_irq(&conf->device_lock); 2182 2183 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n", 2184 (unsigned long long)(*bip)->bi_sector, 2185 (unsigned long long)sh->sector, dd_idx); 2186 2187 if (conf->mddev->bitmap && firstwrite) { 2188 bitmap_startwrite(conf->mddev->bitmap, sh->sector, 2189 STRIPE_SECTORS, 0); 2190 sh->bm_seq = conf->seq_flush+1; 2191 set_bit(STRIPE_BIT_DELAY, &sh->state); 2192 } 2193 return 1; 2194 2195 overlap: 2196 set_bit(R5_Overlap, &sh->dev[dd_idx].flags); 2197 spin_unlock_irq(&conf->device_lock); 2198 return 0; 2199} 2200 2201static void end_reshape(raid5_conf_t *conf); 2202 2203static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous, 2204 struct stripe_head *sh) 2205{ 2206 int sectors_per_chunk = 2207 previous ? conf->prev_chunk_sectors : conf->chunk_sectors; 2208 int dd_idx; 2209 int chunk_offset = sector_div(stripe, sectors_per_chunk); 2210 int disks = previous ? conf->previous_raid_disks : conf->raid_disks; 2211 2212 raid5_compute_sector(conf, 2213 stripe * (disks - conf->max_degraded) 2214 *sectors_per_chunk + chunk_offset, 2215 previous, 2216 &dd_idx, sh); 2217} 2218 2219static void 2220handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh, 2221 struct stripe_head_state *s, int disks, 2222 struct bio **return_bi) 2223{ 2224 int i; 2225 for (i = disks; i--; ) { 2226 struct bio *bi; 2227 int bitmap_end = 0; 2228 2229 if (test_bit(R5_ReadError, &sh->dev[i].flags)) { 2230 mdk_rdev_t *rdev; 2231 rcu_read_lock(); 2232 rdev = rcu_dereference(conf->disks[i].rdev); 2233 if (rdev && test_bit(In_sync, &rdev->flags)) 2234 /* multiple read failures in one stripe */ 2235 md_error(conf->mddev, rdev); 2236 rcu_read_unlock(); 2237 } 2238 spin_lock_irq(&conf->device_lock); 2239 /* fail all writes first */ 2240 bi = sh->dev[i].towrite; 2241 sh->dev[i].towrite = NULL; 2242 if (bi) { 2243 s->to_write--; 2244 bitmap_end = 1; 2245 } 2246 2247 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) 2248 wake_up(&conf->wait_for_overlap); 2249 2250 while (bi && bi->bi_sector < 2251 sh->dev[i].sector + STRIPE_SECTORS) { 2252 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector); 2253 clear_bit(BIO_UPTODATE, &bi->bi_flags); 2254 if (!raid5_dec_bi_phys_segments(bi)) { 2255 md_write_end(conf->mddev); 2256 bi->bi_next = *return_bi; 2257 *return_bi = bi; 2258 } 2259 bi = nextbi; 2260 } 2261 /* and fail all 'written' */ 2262 bi = sh->dev[i].written; 2263 sh->dev[i].written = NULL; 2264 if (bi) bitmap_end = 1; 2265 while (bi && bi->bi_sector < 2266 sh->dev[i].sector + STRIPE_SECTORS) { 2267 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector); 2268 clear_bit(BIO_UPTODATE, &bi->bi_flags); 2269 if (!raid5_dec_bi_phys_segments(bi)) { 2270 md_write_end(conf->mddev); 2271 bi->bi_next = *return_bi; 2272 *return_bi = bi; 2273 } 2274 bi = bi2; 2275 } 2276 2277 /* fail any reads if this device is non-operational and 2278 * the data has not reached the cache yet. 2279 */ 2280 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) && 2281 (!test_bit(R5_Insync, &sh->dev[i].flags) || 2282 test_bit(R5_ReadError, &sh->dev[i].flags))) { 2283 bi = sh->dev[i].toread; 2284 sh->dev[i].toread = NULL; 2285 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) 2286 wake_up(&conf->wait_for_overlap); 2287 if (bi) s->to_read--; 2288 while (bi && bi->bi_sector < 2289 sh->dev[i].sector + STRIPE_SECTORS) { 2290 struct bio *nextbi = 2291 r5_next_bio(bi, sh->dev[i].sector); 2292 clear_bit(BIO_UPTODATE, &bi->bi_flags); 2293 if (!raid5_dec_bi_phys_segments(bi)) { 2294 bi->bi_next = *return_bi; 2295 *return_bi = bi; 2296 } 2297 bi = nextbi; 2298 } 2299 } 2300 spin_unlock_irq(&conf->device_lock); 2301 if (bitmap_end) 2302 bitmap_endwrite(conf->mddev->bitmap, sh->sector, 2303 STRIPE_SECTORS, 0, 0); 2304 } 2305 2306 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state)) 2307 if (atomic_dec_and_test(&conf->pending_full_writes)) 2308 md_wakeup_thread(conf->mddev->thread); 2309} 2310 2311/* fetch_block5 - checks the given member device to see if its data needs 2312 * to be read or computed to satisfy a request. 2313 * 2314 * Returns 1 when no more member devices need to be checked, otherwise returns 2315 * 0 to tell the loop in handle_stripe_fill5 to continue 2316 */ 2317static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s, 2318 int disk_idx, int disks) 2319{ 2320 struct r5dev *dev = &sh->dev[disk_idx]; 2321 struct r5dev *failed_dev = &sh->dev[s->failed_num[0]]; 2322 2323 /* is the data in this block needed, and can we get it? */ 2324 if (!test_bit(R5_LOCKED, &dev->flags) && 2325 !test_bit(R5_UPTODATE, &dev->flags) && 2326 (dev->toread || 2327 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) || 2328 s->syncing || s->expanding || 2329 (s->failed && 2330 (failed_dev->toread || 2331 (failed_dev->towrite && 2332 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) { 2333 /* We would like to get this block, possibly by computing it, 2334 * otherwise read it if the backing disk is insync 2335 */ 2336 if ((s->uptodate == disks - 1) && 2337 (s->failed && disk_idx == s->failed_num[0])) { 2338 set_bit(STRIPE_COMPUTE_RUN, &sh->state); 2339 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request); 2340 set_bit(R5_Wantcompute, &dev->flags); 2341 sh->ops.target = disk_idx; 2342 sh->ops.target2 = -1; 2343 s->req_compute = 1; 2344 /* Careful: from this point on 'uptodate' is in the eye 2345 * of raid_run_ops which services 'compute' operations 2346 * before writes. R5_Wantcompute flags a block that will 2347 * be R5_UPTODATE by the time it is needed for a 2348 * subsequent operation. 2349 */ 2350 s->uptodate++; 2351 return 1; /* uptodate + compute == disks */ 2352 } else if (test_bit(R5_Insync, &dev->flags)) { 2353 set_bit(R5_LOCKED, &dev->flags); 2354 set_bit(R5_Wantread, &dev->flags); 2355 s->locked++; 2356 pr_debug("Reading block %d (sync=%d)\n", disk_idx, 2357 s->syncing); 2358 } 2359 } 2360 2361 return 0; 2362} 2363 2364/** 2365 * handle_stripe_fill5 - read or compute data to satisfy pending requests. 2366 */ 2367static void handle_stripe_fill5(struct stripe_head *sh, 2368 struct stripe_head_state *s, int disks) 2369{ 2370 int i; 2371 2372 /* look for blocks to read/compute, skip this if a compute 2373 * is already in flight, or if the stripe contents are in the 2374 * midst of changing due to a write 2375 */ 2376 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state && 2377 !sh->reconstruct_state) 2378 for (i = disks; i--; ) 2379 if (fetch_block5(sh, s, i, disks)) 2380 break; 2381 set_bit(STRIPE_HANDLE, &sh->state); 2382} 2383 2384/* fetch_block6 - checks the given member device to see if its data needs 2385 * to be read or computed to satisfy a request. 2386 * 2387 * Returns 1 when no more member devices need to be checked, otherwise returns 2388 * 0 to tell the loop in handle_stripe_fill6 to continue 2389 */ 2390static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s, 2391 int disk_idx, int disks) 2392{ 2393 struct r5dev *dev = &sh->dev[disk_idx]; 2394 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]], 2395 &sh->dev[s->failed_num[1]] }; 2396 2397 if (!test_bit(R5_LOCKED, &dev->flags) && 2398 !test_bit(R5_UPTODATE, &dev->flags) && 2399 (dev->toread || 2400 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) || 2401 s->syncing || s->expanding || 2402 (s->failed >= 1 && 2403 (fdev[0]->toread || s->to_write)) || 2404 (s->failed >= 2 && 2405 (fdev[1]->toread || s->to_write)))) { 2406 /* we would like to get this block, possibly by computing it, 2407 * otherwise read it if the backing disk is insync 2408 */ 2409 BUG_ON(test_bit(R5_Wantcompute, &dev->flags)); 2410 BUG_ON(test_bit(R5_Wantread, &dev->flags)); 2411 if ((s->uptodate == disks - 1) && 2412 (s->failed && (disk_idx == s->failed_num[0] || 2413 disk_idx == s->failed_num[1]))) { 2414 /* have disk failed, and we're requested to fetch it; 2415 * do compute it 2416 */ 2417 pr_debug("Computing stripe %llu block %d\n", 2418 (unsigned long long)sh->sector, disk_idx); 2419 set_bit(STRIPE_COMPUTE_RUN, &sh->state); 2420 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request); 2421 set_bit(R5_Wantcompute, &dev->flags); 2422 sh->ops.target = disk_idx; 2423 sh->ops.target2 = -1; /* no 2nd target */ 2424 s->req_compute = 1; 2425 s->uptodate++; 2426 return 1; 2427 } else if (s->uptodate == disks-2 && s->failed >= 2) { 2428 /* Computing 2-failure is *very* expensive; only 2429 * do it if failed >= 2 2430 */ 2431 int other; 2432 for (other = disks; other--; ) { 2433 if (other == disk_idx) 2434 continue; 2435 if (!test_bit(R5_UPTODATE, 2436 &sh->dev[other].flags)) 2437 break; 2438 } 2439 BUG_ON(other < 0); 2440 pr_debug("Computing stripe %llu blocks %d,%d\n", 2441 (unsigned long long)sh->sector, 2442 disk_idx, other); 2443 set_bit(STRIPE_COMPUTE_RUN, &sh->state); 2444 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request); 2445 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags); 2446 set_bit(R5_Wantcompute, &sh->dev[other].flags); 2447 sh->ops.target = disk_idx; 2448 sh->ops.target2 = other; 2449 s->uptodate += 2; 2450 s->req_compute = 1; 2451 return 1; 2452 } else if (test_bit(R5_Insync, &dev->flags)) { 2453 set_bit(R5_LOCKED, &dev->flags); 2454 set_bit(R5_Wantread, &dev->flags); 2455 s->locked++; 2456 pr_debug("Reading block %d (sync=%d)\n", 2457 disk_idx, s->syncing); 2458 } 2459 } 2460 2461 return 0; 2462} 2463 2464/** 2465 * handle_stripe_fill6 - read or compute data to satisfy pending requests. 2466 */ 2467static void handle_stripe_fill6(struct stripe_head *sh, 2468 struct stripe_head_state *s, 2469 int disks) 2470{ 2471 int i; 2472 2473 /* look for blocks to read/compute, skip this if a compute 2474 * is already in flight, or if the stripe contents are in the 2475 * midst of changing due to a write 2476 */ 2477 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state && 2478 !sh->reconstruct_state) 2479 for (i = disks; i--; ) 2480 if (fetch_block6(sh, s, i, disks)) 2481 break; 2482 set_bit(STRIPE_HANDLE, &sh->state); 2483} 2484 2485 2486/* handle_stripe_clean_event 2487 * any written block on an uptodate or failed drive can be returned. 2488 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but 2489 * never LOCKED, so we don't need to test 'failed' directly. 2490 */ 2491static void handle_stripe_clean_event(raid5_conf_t *conf, 2492 struct stripe_head *sh, int disks, struct bio **return_bi) 2493{ 2494 int i; 2495 struct r5dev *dev; 2496 2497 for (i = disks; i--; ) 2498 if (sh->dev[i].written) { 2499 dev = &sh->dev[i]; 2500 if (!test_bit(R5_LOCKED, &dev->flags) && 2501 test_bit(R5_UPTODATE, &dev->flags)) { 2502 /* We can return any write requests */ 2503 struct bio *wbi, *wbi2; 2504 int bitmap_end = 0; 2505 pr_debug("Return write for disc %d\n", i); 2506 spin_lock_irq(&conf->device_lock); 2507 wbi = dev->written; 2508 dev->written = NULL; 2509 while (wbi && wbi->bi_sector < 2510 dev->sector + STRIPE_SECTORS) { 2511 wbi2 = r5_next_bio(wbi, dev->sector); 2512 if (!raid5_dec_bi_phys_segments(wbi)) { 2513 md_write_end(conf->mddev); 2514 wbi->bi_next = *return_bi; 2515 *return_bi = wbi; 2516 } 2517 wbi = wbi2; 2518 } 2519 if (dev->towrite == NULL) 2520 bitmap_end = 1; 2521 spin_unlock_irq(&conf->device_lock); 2522 if (bitmap_end) 2523 bitmap_endwrite(conf->mddev->bitmap, 2524 sh->sector, 2525 STRIPE_SECTORS, 2526 !test_bit(STRIPE_DEGRADED, &sh->state), 2527 0); 2528 } 2529 } 2530 2531 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state)) 2532 if (atomic_dec_and_test(&conf->pending_full_writes)) 2533 md_wakeup_thread(conf->mddev->thread); 2534} 2535 2536static void handle_stripe_dirtying5(raid5_conf_t *conf, 2537 struct stripe_head *sh, struct stripe_head_state *s, int disks) 2538{ 2539 int rmw = 0, rcw = 0, i; 2540 for (i = disks; i--; ) { 2541 /* would I have to read this buffer for read_modify_write */ 2542 struct r5dev *dev = &sh->dev[i]; 2543 if ((dev->towrite || i == sh->pd_idx) && 2544 !test_bit(R5_LOCKED, &dev->flags) && 2545 !(test_bit(R5_UPTODATE, &dev->flags) || 2546 test_bit(R5_Wantcompute, &dev->flags))) { 2547 if (test_bit(R5_Insync, &dev->flags)) 2548 rmw++; 2549 else 2550 rmw += 2*disks; /* cannot read it */ 2551 } 2552 /* Would I have to read this buffer for reconstruct_write */ 2553 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx && 2554 !test_bit(R5_LOCKED, &dev->flags) && 2555 !(test_bit(R5_UPTODATE, &dev->flags) || 2556 test_bit(R5_Wantcompute, &dev->flags))) { 2557 if (test_bit(R5_Insync, &dev->flags)) rcw++; 2558 else 2559 rcw += 2*disks; 2560 } 2561 } 2562 pr_debug("for sector %llu, rmw=%d rcw=%d\n", 2563 (unsigned long long)sh->sector, rmw, rcw); 2564 set_bit(STRIPE_HANDLE, &sh->state); 2565 if (rmw < rcw && rmw > 0) 2566 /* prefer read-modify-write, but need to get some data */ 2567 for (i = disks; i--; ) { 2568 struct r5dev *dev = &sh->dev[i]; 2569 if ((dev->towrite || i == sh->pd_idx) && 2570 !test_bit(R5_LOCKED, &dev->flags) && 2571 !(test_bit(R5_UPTODATE, &dev->flags) || 2572 test_bit(R5_Wantcompute, &dev->flags)) && 2573 test_bit(R5_Insync, &dev->flags)) { 2574 if ( 2575 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 2576 pr_debug("Read_old block " 2577 "%d for r-m-w\n", i); 2578 set_bit(R5_LOCKED, &dev->flags); 2579 set_bit(R5_Wantread, &dev->flags); 2580 s->locked++; 2581 } else { 2582 set_bit(STRIPE_DELAYED, &sh->state); 2583 set_bit(STRIPE_HANDLE, &sh->state); 2584 } 2585 } 2586 } 2587 if (rcw <= rmw && rcw > 0) 2588 /* want reconstruct write, but need to get some data */ 2589 for (i = disks; i--; ) { 2590 struct r5dev *dev = &sh->dev[i]; 2591 if (!test_bit(R5_OVERWRITE, &dev->flags) && 2592 i != sh->pd_idx && 2593 !test_bit(R5_LOCKED, &dev->flags) && 2594 !(test_bit(R5_UPTODATE, &dev->flags) || 2595 test_bit(R5_Wantcompute, &dev->flags)) && 2596 test_bit(R5_Insync, &dev->flags)) { 2597 if ( 2598 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 2599 pr_debug("Read_old block " 2600 "%d for Reconstruct\n", i); 2601 set_bit(R5_LOCKED, &dev->flags); 2602 set_bit(R5_Wantread, &dev->flags); 2603 s->locked++; 2604 } else { 2605 set_bit(STRIPE_DELAYED, &sh->state); 2606 set_bit(STRIPE_HANDLE, &sh->state); 2607 } 2608 } 2609 } 2610 /* now if nothing is locked, and if we have enough data, 2611 * we can start a write request 2612 */ 2613 /* since handle_stripe can be called at any time we need to handle the 2614 * case where a compute block operation has been submitted and then a 2615 * subsequent call wants to start a write request. raid_run_ops only 2616 * handles the case where compute block and reconstruct are requested 2617 * simultaneously. If this is not the case then new writes need to be 2618 * held off until the compute completes. 2619 */ 2620 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) && 2621 (s->locked == 0 && (rcw == 0 || rmw == 0) && 2622 !test_bit(STRIPE_BIT_DELAY, &sh->state))) 2623 schedule_reconstruction(sh, s, rcw == 0, 0); 2624} 2625 2626static void handle_stripe_dirtying6(raid5_conf_t *conf, 2627 struct stripe_head *sh, struct stripe_head_state *s, 2628 int disks) 2629{ 2630 int rcw = 0, pd_idx = sh->pd_idx, i; 2631 int qd_idx = sh->qd_idx; 2632 2633 set_bit(STRIPE_HANDLE, &sh->state); 2634 for (i = disks; i--; ) { 2635 struct r5dev *dev = &sh->dev[i]; 2636 /* check if we haven't enough data */ 2637 if (!test_bit(R5_OVERWRITE, &dev->flags) && 2638 i != pd_idx && i != qd_idx && 2639 !test_bit(R5_LOCKED, &dev->flags) && 2640 !(test_bit(R5_UPTODATE, &dev->flags) || 2641 test_bit(R5_Wantcompute, &dev->flags))) { 2642 rcw++; 2643 if (!test_bit(R5_Insync, &dev->flags)) 2644 continue; /* it's a failed drive */ 2645 2646 if ( 2647 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 2648 pr_debug("Read_old stripe %llu " 2649 "block %d for Reconstruct\n", 2650 (unsigned long long)sh->sector, i); 2651 set_bit(R5_LOCKED, &dev->flags); 2652 set_bit(R5_Wantread, &dev->flags); 2653 s->locked++; 2654 } else { 2655 pr_debug("Request delayed stripe %llu " 2656 "block %d for Reconstruct\n", 2657 (unsigned long long)sh->sector, i); 2658 set_bit(STRIPE_DELAYED, &sh->state); 2659 set_bit(STRIPE_HANDLE, &sh->state); 2660 } 2661 } 2662 } 2663 /* now if nothing is locked, and if we have enough data, we can start a 2664 * write request 2665 */ 2666 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) && 2667 s->locked == 0 && rcw == 0 && 2668 !test_bit(STRIPE_BIT_DELAY, &sh->state)) { 2669 schedule_reconstruction(sh, s, 1, 0); 2670 } 2671} 2672 2673static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh, 2674 struct stripe_head_state *s, int disks) 2675{ 2676 struct r5dev *dev = NULL; 2677 2678 set_bit(STRIPE_HANDLE, &sh->state); 2679 2680 switch (sh->check_state) { 2681 case check_state_idle: 2682 /* start a new check operation if there are no failures */ 2683 if (s->failed == 0) { 2684 BUG_ON(s->uptodate != disks); 2685 sh->check_state = check_state_run; 2686 set_bit(STRIPE_OP_CHECK, &s->ops_request); 2687 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags); 2688 s->uptodate--; 2689 break; 2690 } 2691 dev = &sh->dev[s->failed_num[0]]; 2692 /* fall through */ 2693 case check_state_compute_result: 2694 sh->check_state = check_state_idle; 2695 if (!dev) 2696 dev = &sh->dev[sh->pd_idx]; 2697 2698 /* check that a write has not made the stripe insync */ 2699 if (test_bit(STRIPE_INSYNC, &sh->state)) 2700 break; 2701 2702 /* either failed parity check, or recovery is happening */ 2703 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags)); 2704 BUG_ON(s->uptodate != disks); 2705 2706 set_bit(R5_LOCKED, &dev->flags); 2707 s->locked++; 2708 set_bit(R5_Wantwrite, &dev->flags); 2709 2710 clear_bit(STRIPE_DEGRADED, &sh->state); 2711 set_bit(STRIPE_INSYNC, &sh->state); 2712 break; 2713 case check_state_run: 2714 break; /* we will be called again upon completion */ 2715 case check_state_check_result: 2716 sh->check_state = check_state_idle; 2717 2718 /* if a failure occurred during the check operation, leave 2719 * STRIPE_INSYNC not set and let the stripe be handled again 2720 */ 2721 if (s->failed) 2722 break; 2723 2724 /* handle a successful check operation, if parity is correct 2725 * we are done. Otherwise update the mismatch count and repair 2726 * parity if !MD_RECOVERY_CHECK 2727 */ 2728 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0) 2729 /* parity is correct (on disc, 2730 * not in buffer any more) 2731 */ 2732 set_bit(STRIPE_INSYNC, &sh->state); 2733 else { 2734 conf->mddev->resync_mismatches += STRIPE_SECTORS; 2735 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) 2736 /* don't try to repair!! */ 2737 set_bit(STRIPE_INSYNC, &sh->state); 2738 else { 2739 sh->check_state = check_state_compute_run; 2740 set_bit(STRIPE_COMPUTE_RUN, &sh->state); 2741 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request); 2742 set_bit(R5_Wantcompute, 2743 &sh->dev[sh->pd_idx].flags); 2744 sh->ops.target = sh->pd_idx; 2745 sh->ops.target2 = -1; 2746 s->uptodate++; 2747 } 2748 } 2749 break; 2750 case check_state_compute_run: 2751 break; 2752 default: 2753 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n", 2754 __func__, sh->check_state, 2755 (unsigned long long) sh->sector); 2756 BUG(); 2757 } 2758} 2759 2760 2761static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh, 2762 struct stripe_head_state *s, 2763 int disks) 2764{ 2765 int pd_idx = sh->pd_idx; 2766 int qd_idx = sh->qd_idx; 2767 struct r5dev *dev; 2768 2769 set_bit(STRIPE_HANDLE, &sh->state); 2770 2771 BUG_ON(s->failed > 2); 2772 2773 /* Want to check and possibly repair P and Q. 2774 * However there could be one 'failed' device, in which 2775 * case we can only check one of them, possibly using the 2776 * other to generate missing data 2777 */ 2778 2779 switch (sh->check_state) { 2780 case check_state_idle: 2781 /* start a new check operation if there are < 2 failures */ 2782 if (s->failed == s->q_failed) { 2783 /* The only possible failed device holds Q, so it 2784 * makes sense to check P (If anything else were failed, 2785 * we would have used P to recreate it). 2786 */ 2787 sh->check_state = check_state_run; 2788 } 2789 if (!s->q_failed && s->failed < 2) { 2790 /* Q is not failed, and we didn't use it to generate 2791 * anything, so it makes sense to check it 2792 */ 2793 if (sh->check_state == check_state_run) 2794 sh->check_state = check_state_run_pq; 2795 else 2796 sh->check_state = check_state_run_q; 2797 } 2798 2799 /* discard potentially stale zero_sum_result */ 2800 sh->ops.zero_sum_result = 0; 2801 2802 if (sh->check_state == check_state_run) { 2803 /* async_xor_zero_sum destroys the contents of P */ 2804 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags); 2805 s->uptodate--; 2806 } 2807 if (sh->check_state >= check_state_run && 2808 sh->check_state <= check_state_run_pq) { 2809 /* async_syndrome_zero_sum preserves P and Q, so 2810 * no need to mark them !uptodate here 2811 */ 2812 set_bit(STRIPE_OP_CHECK, &s->ops_request); 2813 break; 2814 } 2815 2816 /* we have 2-disk failure */ 2817 BUG_ON(s->failed != 2); 2818 /* fall through */ 2819 case check_state_compute_result: 2820 sh->check_state = check_state_idle; 2821 2822 /* check that a write has not made the stripe insync */ 2823 if (test_bit(STRIPE_INSYNC, &sh->state)) 2824 break; 2825 2826 /* now write out any block on a failed drive, 2827 * or P or Q if they were recomputed 2828 */ 2829 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */ 2830 if (s->failed == 2) { 2831 dev = &sh->dev[s->failed_num[1]]; 2832 s->locked++; 2833 set_bit(R5_LOCKED, &dev->flags); 2834 set_bit(R5_Wantwrite, &dev->flags); 2835 } 2836 if (s->failed >= 1) { 2837 dev = &sh->dev[s->failed_num[0]]; 2838 s->locked++; 2839 set_bit(R5_LOCKED, &dev->flags); 2840 set_bit(R5_Wantwrite, &dev->flags); 2841 } 2842 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) { 2843 dev = &sh->dev[pd_idx]; 2844 s->locked++; 2845 set_bit(R5_LOCKED, &dev->flags); 2846 set_bit(R5_Wantwrite, &dev->flags); 2847 } 2848 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) { 2849 dev = &sh->dev[qd_idx]; 2850 s->locked++; 2851 set_bit(R5_LOCKED, &dev->flags); 2852 set_bit(R5_Wantwrite, &dev->flags); 2853 } 2854 clear_bit(STRIPE_DEGRADED, &sh->state); 2855 2856 set_bit(STRIPE_INSYNC, &sh->state); 2857 break; 2858 case check_state_run: 2859 case check_state_run_q: 2860 case check_state_run_pq: 2861 break; /* we will be called again upon completion */ 2862 case check_state_check_result: 2863 sh->check_state = check_state_idle; 2864 2865 /* handle a successful check operation, if parity is correct 2866 * we are done. Otherwise update the mismatch count and repair 2867 * parity if !MD_RECOVERY_CHECK 2868 */ 2869 if (sh->ops.zero_sum_result == 0) { 2870 /* both parities are correct */ 2871 if (!s->failed) 2872 set_bit(STRIPE_INSYNC, &sh->state); 2873 else { 2874 /* in contrast to the raid5 case we can validate 2875 * parity, but still have a failure to write 2876 * back 2877 */ 2878 sh->check_state = check_state_compute_result; 2879 /* Returning at this point means that we may go 2880 * off and bring p and/or q uptodate again so 2881 * we make sure to check zero_sum_result again 2882 * to verify if p or q need writeback 2883 */ 2884 } 2885 } else { 2886 conf->mddev->resync_mismatches += STRIPE_SECTORS; 2887 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) 2888 /* don't try to repair!! */ 2889 set_bit(STRIPE_INSYNC, &sh->state); 2890 else { 2891 int *target = &sh->ops.target; 2892 2893 sh->ops.target = -1; 2894 sh->ops.target2 = -1; 2895 sh->check_state = check_state_compute_run; 2896 set_bit(STRIPE_COMPUTE_RUN, &sh->state); 2897 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request); 2898 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) { 2899 set_bit(R5_Wantcompute, 2900 &sh->dev[pd_idx].flags); 2901 *target = pd_idx; 2902 target = &sh->ops.target2; 2903 s->uptodate++; 2904 } 2905 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) { 2906 set_bit(R5_Wantcompute, 2907 &sh->dev[qd_idx].flags); 2908 *target = qd_idx; 2909 s->uptodate++; 2910 } 2911 } 2912 } 2913 break; 2914 case check_state_compute_run: 2915 break; 2916 default: 2917 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n", 2918 __func__, sh->check_state, 2919 (unsigned long long) sh->sector); 2920 BUG(); 2921 } 2922} 2923 2924static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh) 2925{ 2926 int i; 2927 2928 /* We have read all the blocks in this stripe and now we need to 2929 * copy some of them into a target stripe for expand. 2930 */ 2931 struct dma_async_tx_descriptor *tx = NULL; 2932 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state); 2933 for (i = 0; i < sh->disks; i++) 2934 if (i != sh->pd_idx && i != sh->qd_idx) { 2935 int dd_idx, j; 2936 struct stripe_head *sh2; 2937 struct async_submit_ctl submit; 2938 2939 sector_t bn = compute_blocknr(sh, i, 1); 2940 sector_t s = raid5_compute_sector(conf, bn, 0, 2941 &dd_idx, NULL); 2942 sh2 = get_active_stripe(conf, s, 0, 1, 1); 2943 if (sh2 == NULL) 2944 /* so far only the early blocks of this stripe 2945 * have been requested. When later blocks 2946 * get requested, we will try again 2947 */ 2948 continue; 2949 if (!test_bit(STRIPE_EXPANDING, &sh2->state) || 2950 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) { 2951 /* must have already done this block */ 2952 release_stripe(sh2); 2953 continue; 2954 } 2955 2956 /* place all the copies on one channel */ 2957 init_async_submit(&submit, 0, tx, NULL, NULL, NULL); 2958 tx = async_memcpy(sh2->dev[dd_idx].page, 2959 sh->dev[i].page, 0, 0, STRIPE_SIZE, 2960 &submit); 2961 2962 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags); 2963 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags); 2964 for (j = 0; j < conf->raid_disks; j++) 2965 if (j != sh2->pd_idx && 2966 j != sh2->qd_idx && 2967 !test_bit(R5_Expanded, &sh2->dev[j].flags)) 2968 break; 2969 if (j == conf->raid_disks) { 2970 set_bit(STRIPE_EXPAND_READY, &sh2->state); 2971 set_bit(STRIPE_HANDLE, &sh2->state); 2972 } 2973 release_stripe(sh2); 2974 2975 } 2976 /* done submitting copies, wait for them to complete */ 2977 if (tx) { 2978 async_tx_ack(tx); 2979 dma_wait_for_async_tx(tx); 2980 } 2981} 2982 2983 2984/* 2985 * handle_stripe - do things to a stripe. 2986 * 2987 * We lock the stripe and then examine the state of various bits 2988 * to see what needs to be done. 2989 * Possible results: 2990 * return some read request which now have data 2991 * return some write requests which are safely on disc 2992 * schedule a read on some buffers 2993 * schedule a write of some buffers 2994 * return confirmation of parity correctness 2995 * 2996 * buffers are taken off read_list or write_list, and bh_cache buffers 2997 * get BH_Lock set before the stripe lock is released. 2998 * 2999 */ 3000 3001static int handle_stripe5(struct stripe_head *sh, struct stripe_head_state *s) 3002{ 3003 raid5_conf_t *conf = sh->raid_conf; 3004 int disks = sh->disks, i; 3005 struct r5dev *dev; 3006 int prexor; 3007 3008 /* Now to look around and see what can be done */ 3009 rcu_read_lock(); 3010 spin_lock_irq(&conf->device_lock); 3011 for (i=disks; i--; ) { 3012 mdk_rdev_t *rdev; 3013 3014 dev = &sh->dev[i]; 3015 3016 pr_debug("check %d: state 0x%lx toread %p read %p write %p " 3017 "written %p\n", i, dev->flags, dev->toread, dev->read, 3018 dev->towrite, dev->written); 3019 3020 /* maybe we can request a biofill operation 3021 * 3022 * new wantfill requests are only permitted while 3023 * ops_complete_biofill is guaranteed to be inactive 3024 */ 3025 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread && 3026 !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) 3027 set_bit(R5_Wantfill, &dev->flags); 3028 3029 /* now count some things */ 3030 if (test_bit(R5_LOCKED, &dev->flags)) 3031 s->locked++; 3032 if (test_bit(R5_UPTODATE, &dev->flags)) 3033 s->uptodate++; 3034 if (test_bit(R5_Wantcompute, &dev->flags)) 3035 s->compute++; 3036 3037 if (test_bit(R5_Wantfill, &dev->flags)) 3038 s->to_fill++; 3039 else if (dev->toread) 3040 s->to_read++; 3041 if (dev->towrite) { 3042 s->to_write++; 3043 if (!test_bit(R5_OVERWRITE, &dev->flags)) 3044 s->non_overwrite++; 3045 } 3046 if (dev->written) 3047 s->written++; 3048 rdev = rcu_dereference(conf->disks[i].rdev); 3049 if (s->blocked_rdev == NULL && 3050 rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 3051 s->blocked_rdev = rdev; 3052 atomic_inc(&rdev->nr_pending); 3053 } 3054 clear_bit(R5_Insync, &dev->flags); 3055 if (!rdev) 3056 /* Not in-sync */; 3057 else if (test_bit(In_sync, &rdev->flags)) 3058 set_bit(R5_Insync, &dev->flags); 3059 else { 3060 /* could be in-sync depending on recovery/reshape status */ 3061 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset) 3062 set_bit(R5_Insync, &dev->flags); 3063 } 3064 if (!test_bit(R5_Insync, &dev->flags)) { 3065 /* The ReadError flag will just be confusing now */ 3066 clear_bit(R5_ReadError, &dev->flags); 3067 clear_bit(R5_ReWrite, &dev->flags); 3068 } 3069 if (test_bit(R5_ReadError, &dev->flags)) 3070 clear_bit(R5_Insync, &dev->flags); 3071 if (!test_bit(R5_Insync, &dev->flags)) { 3072 s->failed++; 3073 s->failed_num[0] = i; 3074 } 3075 } 3076 spin_unlock_irq(&conf->device_lock); 3077 rcu_read_unlock(); 3078 3079 if (unlikely(s->blocked_rdev)) { 3080 if (s->syncing || s->expanding || s->expanded || 3081 s->to_write || s->written) { 3082 set_bit(STRIPE_HANDLE, &sh->state); 3083 return 1; 3084 } 3085 /* There is nothing for the blocked_rdev to block */ 3086 rdev_dec_pending(s->blocked_rdev, conf->mddev); 3087 s->blocked_rdev = NULL; 3088 } 3089 3090 if (s->to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) { 3091 set_bit(STRIPE_OP_BIOFILL, &s->ops_request); 3092 set_bit(STRIPE_BIOFILL_RUN, &sh->state); 3093 } 3094 3095 pr_debug("locked=%d uptodate=%d to_read=%d" 3096 " to_write=%d failed=%d failed_num=%d\n", 3097 s->locked, s->uptodate, s->to_read, s->to_write, 3098 s->failed, s->failed_num[0]); 3099 /* check if the array has lost two devices and, if so, some requests might 3100 * need to be failed 3101 */ 3102 if (s->failed > 1 && s->to_read+s->to_write+s->written) 3103 handle_failed_stripe(conf, sh, s, disks, &s->return_bi); 3104 if (s->failed > 1 && s->syncing) { 3105 md_done_sync(conf->mddev, STRIPE_SECTORS,0); 3106 clear_bit(STRIPE_SYNCING, &sh->state); 3107 s->syncing = 0; 3108 } 3109 3110 /* might be able to return some write requests if the parity block 3111 * is safe, or on a failed drive 3112 */ 3113 dev = &sh->dev[sh->pd_idx]; 3114 if (s->written && 3115 ((test_bit(R5_Insync, &dev->flags) && 3116 !test_bit(R5_LOCKED, &dev->flags) && 3117 test_bit(R5_UPTODATE, &dev->flags)) || 3118 (s->failed == 1 && s->failed_num[0] == sh->pd_idx))) 3119 handle_stripe_clean_event(conf, sh, disks, &s->return_bi); 3120 3121 /* Now we might consider reading some blocks, either to check/generate 3122 * parity, or to satisfy requests 3123 * or to load a block that is being partially written. 3124 */ 3125 if (s->to_read || s->non_overwrite || 3126 (s->syncing && (s->uptodate + s->compute < disks)) || s->expanding) 3127 handle_stripe_fill5(sh, s, disks); 3128 3129 /* Now we check to see if any write operations have recently 3130 * completed 3131 */ 3132 prexor = 0; 3133 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result) 3134 prexor = 1; 3135 if (sh->reconstruct_state == reconstruct_state_drain_result || 3136 sh->reconstruct_state == reconstruct_state_prexor_drain_result) { 3137 sh->reconstruct_state = reconstruct_state_idle; 3138 3139 /* All the 'written' buffers and the parity block are ready to 3140 * be written back to disk 3141 */ 3142 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags)); 3143 for (i = disks; i--; ) { 3144 dev = &sh->dev[i]; 3145 if (test_bit(R5_LOCKED, &dev->flags) && 3146 (i == sh->pd_idx || dev->written)) { 3147 pr_debug("Writing block %d\n", i); 3148 set_bit(R5_Wantwrite, &dev->flags); 3149 if (prexor) 3150 continue; 3151 if (!test_bit(R5_Insync, &dev->flags) || 3152 (i == sh->pd_idx && s->failed == 0)) 3153 set_bit(STRIPE_INSYNC, &sh->state); 3154 } 3155 } 3156 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 3157 s->dec_preread_active = 1; 3158 } 3159 3160 /* Now to consider new write requests and what else, if anything 3161 * should be read. We do not handle new writes when: 3162 * 1/ A 'write' operation (copy+xor) is already in flight. 3163 * 2/ A 'check' operation is in flight, as it may clobber the parity 3164 * block. 3165 */ 3166 if (s->to_write && !sh->reconstruct_state && !sh->check_state) 3167 handle_stripe_dirtying5(conf, sh, s, disks); 3168 3169 /* maybe we need to check and possibly fix the parity for this stripe 3170 * Any reads will already have been scheduled, so we just see if enough 3171 * data is available. The parity check is held off while parity 3172 * dependent operations are in flight. 3173 */ 3174 if (sh->check_state || 3175 (s->syncing && s->locked == 0 && 3176 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) && 3177 !test_bit(STRIPE_INSYNC, &sh->state))) 3178 handle_parity_checks5(conf, sh, s, disks); 3179 return 0; 3180} 3181 3182static int handle_stripe6(struct stripe_head *sh, struct stripe_head_state *s) 3183{ 3184 raid5_conf_t *conf = sh->raid_conf; 3185 int disks = sh->disks; 3186 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx; 3187 struct r5dev *dev, *pdev, *qdev; 3188 3189 /* Now to look around and see what can be done */ 3190 3191 rcu_read_lock(); 3192 spin_lock_irq(&conf->device_lock); 3193 for (i=disks; i--; ) { 3194 mdk_rdev_t *rdev; 3195 dev = &sh->dev[i]; 3196 3197 pr_debug("check %d: state 0x%lx read %p write %p written %p\n", 3198 i, dev->flags, dev->toread, dev->towrite, dev->written); 3199 /* maybe we can reply to a read 3200 * 3201 * new wantfill requests are only permitted while 3202 * ops_complete_biofill is guaranteed to be inactive 3203 */ 3204 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread && 3205 !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) 3206 set_bit(R5_Wantfill, &dev->flags); 3207 3208 /* now count some things */ 3209 if (test_bit(R5_LOCKED, &dev->flags)) 3210 s->locked++; 3211 if (test_bit(R5_UPTODATE, &dev->flags)) 3212 s->uptodate++; 3213 if (test_bit(R5_Wantcompute, &dev->flags)) { 3214 s->compute++; 3215 BUG_ON(s->compute > 2); 3216 } 3217 3218 if (test_bit(R5_Wantfill, &dev->flags)) { 3219 s->to_fill++; 3220 } else if (dev->toread) 3221 s->to_read++; 3222 if (dev->towrite) { 3223 s->to_write++; 3224 if (!test_bit(R5_OVERWRITE, &dev->flags)) 3225 s->non_overwrite++; 3226 } 3227 if (dev->written) 3228 s->written++; 3229 rdev = rcu_dereference(conf->disks[i].rdev); 3230 if (s->blocked_rdev == NULL && 3231 rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 3232 s->blocked_rdev = rdev; 3233 atomic_inc(&rdev->nr_pending); 3234 } 3235 clear_bit(R5_Insync, &dev->flags); 3236 if (!rdev) 3237 /* Not in-sync */; 3238 else if (test_bit(In_sync, &rdev->flags)) 3239 set_bit(R5_Insync, &dev->flags); 3240 else { 3241 /* in sync if before recovery_offset */ 3242 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset) 3243 set_bit(R5_Insync, &dev->flags); 3244 } 3245 if (!test_bit(R5_Insync, &dev->flags)) { 3246 /* The ReadError flag will just be confusing now */ 3247 clear_bit(R5_ReadError, &dev->flags); 3248 clear_bit(R5_ReWrite, &dev->flags); 3249 } 3250 if (test_bit(R5_ReadError, &dev->flags)) 3251 clear_bit(R5_Insync, &dev->flags); 3252 if (!test_bit(R5_Insync, &dev->flags)) { 3253 if (s->failed < 2) 3254 s->failed_num[s->failed] = i; 3255 s->failed++; 3256 } 3257 } 3258 spin_unlock_irq(&conf->device_lock); 3259 rcu_read_unlock(); 3260 3261 if (unlikely(s->blocked_rdev)) { 3262 if (s->syncing || s->expanding || s->expanded || 3263 s->to_write || s->written) { 3264 set_bit(STRIPE_HANDLE, &sh->state); 3265 return 1; 3266 } 3267 /* There is nothing for the blocked_rdev to block */ 3268 rdev_dec_pending(s->blocked_rdev, conf->mddev); 3269 s->blocked_rdev = NULL; 3270 } 3271 3272 if (s->to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) { 3273 set_bit(STRIPE_OP_BIOFILL, &s->ops_request); 3274 set_bit(STRIPE_BIOFILL_RUN, &sh->state); 3275 } 3276 3277 pr_debug("locked=%d uptodate=%d to_read=%d" 3278 " to_write=%d failed=%d failed_num=%d,%d\n", 3279 s->locked, s->uptodate, s->to_read, s->to_write, s->failed, 3280 s->failed_num[0], s->failed_num[1]); 3281 /* check if the array has lost >2 devices and, if so, some requests 3282 * might need to be failed 3283 */ 3284 if (s->failed > 2 && s->to_read+s->to_write+s->written) 3285 handle_failed_stripe(conf, sh, s, disks, &s->return_bi); 3286 if (s->failed > 2 && s->syncing) { 3287 md_done_sync(conf->mddev, STRIPE_SECTORS,0); 3288 clear_bit(STRIPE_SYNCING, &sh->state); 3289 s->syncing = 0; 3290 } 3291 3292 /* 3293 * might be able to return some write requests if the parity blocks 3294 * are safe, or on a failed drive 3295 */ 3296 pdev = &sh->dev[pd_idx]; 3297 s->p_failed = (s->failed >= 1 && s->failed_num[0] == pd_idx) 3298 || (s->failed >= 2 && s->failed_num[1] == pd_idx); 3299 qdev = &sh->dev[qd_idx]; 3300 s->q_failed = (s->failed >= 1 && s->failed_num[0] == qd_idx) 3301 || (s->failed >= 2 && s->failed_num[1] == qd_idx); 3302 3303 if (s->written && 3304 (s->p_failed || ((test_bit(R5_Insync, &pdev->flags) 3305 && !test_bit(R5_LOCKED, &pdev->flags) 3306 && test_bit(R5_UPTODATE, &pdev->flags)))) && 3307 (s->q_failed || ((test_bit(R5_Insync, &qdev->flags) 3308 && !test_bit(R5_LOCKED, &qdev->flags) 3309 && test_bit(R5_UPTODATE, &qdev->flags))))) 3310 handle_stripe_clean_event(conf, sh, disks, &s->return_bi); 3311 3312 /* Now we might consider reading some blocks, either to check/generate 3313 * parity, or to satisfy requests 3314 * or to load a block that is being partially written. 3315 */ 3316 if (s->to_read || s->non_overwrite || (s->to_write && s->failed) || 3317 (s->syncing && (s->uptodate + s->compute < disks)) || s->expanding) 3318 handle_stripe_fill6(sh, s, disks); 3319 3320 /* Now we check to see if any write operations have recently 3321 * completed 3322 */ 3323 if (sh->reconstruct_state == reconstruct_state_drain_result) { 3324 3325 sh->reconstruct_state = reconstruct_state_idle; 3326 /* All the 'written' buffers and the parity blocks are ready to 3327 * be written back to disk 3328 */ 3329 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags)); 3330 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags)); 3331 for (i = disks; i--; ) { 3332 dev = &sh->dev[i]; 3333 if (test_bit(R5_LOCKED, &dev->flags) && 3334 (i == sh->pd_idx || i == qd_idx || 3335 dev->written)) { 3336 pr_debug("Writing block %d\n", i); 3337 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags)); 3338 set_bit(R5_Wantwrite, &dev->flags); 3339 if (!test_bit(R5_Insync, &dev->flags) || 3340 ((i == sh->pd_idx || i == qd_idx) && 3341 s->failed == 0)) 3342 set_bit(STRIPE_INSYNC, &sh->state); 3343 } 3344 } 3345 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 3346 s->dec_preread_active = 1; 3347 } 3348 3349 /* Now to consider new write requests and what else, if anything 3350 * should be read. We do not handle new writes when: 3351 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight. 3352 * 2/ A 'check' operation is in flight, as it may clobber the parity 3353 * block. 3354 */ 3355 if (s->to_write && !sh->reconstruct_state && !sh->check_state) 3356 handle_stripe_dirtying6(conf, sh, s, disks); 3357 3358 /* maybe we need to check and possibly fix the parity for this stripe 3359 * Any reads will already have been scheduled, so we just see if enough 3360 * data is available. The parity check is held off while parity 3361 * dependent operations are in flight. 3362 */ 3363 if (sh->check_state || 3364 (s->syncing && s->locked == 0 && 3365 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) && 3366 !test_bit(STRIPE_INSYNC, &sh->state))) 3367 handle_parity_checks6(conf, sh, s, disks); 3368 return 0; 3369} 3370 3371static void handle_stripe(struct stripe_head *sh) 3372{ 3373 struct stripe_head_state s; 3374 int done; 3375 int i; 3376 raid5_conf_t *conf = sh->raid_conf; 3377 3378 clear_bit(STRIPE_HANDLE, &sh->state); 3379 if (test_and_set_bit(STRIPE_ACTIVE, &sh->state)) { 3380 /* already being handled, ensure it gets handled 3381 * again when current action finishes */ 3382 set_bit(STRIPE_HANDLE, &sh->state); 3383 return; 3384 } 3385 3386 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) { 3387 set_bit(STRIPE_SYNCING, &sh->state); 3388 clear_bit(STRIPE_INSYNC, &sh->state); 3389 } 3390 clear_bit(STRIPE_DELAYED, &sh->state); 3391 3392 pr_debug("handling stripe %llu, state=%#lx cnt=%d, " 3393 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n", 3394 (unsigned long long)sh->sector, sh->state, 3395 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx, 3396 sh->check_state, sh->reconstruct_state); 3397 memset(&s, 0, sizeof(s)); 3398 3399 s.syncing = test_bit(STRIPE_SYNCING, &sh->state); 3400 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state); 3401 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state); 3402 3403 if (conf->level == 6) 3404 done = handle_stripe6(sh, &s); 3405 else 3406 done = handle_stripe5(sh, &s); 3407 3408 if (done) 3409 goto finish; 3410 3411 3412 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) { 3413 md_done_sync(conf->mddev, STRIPE_SECTORS, 1); 3414 clear_bit(STRIPE_SYNCING, &sh->state); 3415 } 3416 3417 /* If the failed drives are just a ReadError, then we might need 3418 * to progress the repair/check process 3419 */ 3420 if (s.failed <= conf->max_degraded && !conf->mddev->ro) 3421 for (i = 0; i < s.failed; i++) { 3422 struct r5dev *dev = &sh->dev[s.failed_num[i]]; 3423 if (test_bit(R5_ReadError, &dev->flags) 3424 && !test_bit(R5_LOCKED, &dev->flags) 3425 && test_bit(R5_UPTODATE, &dev->flags) 3426 ) { 3427 if (!test_bit(R5_ReWrite, &dev->flags)) { 3428 set_bit(R5_Wantwrite, &dev->flags); 3429 set_bit(R5_ReWrite, &dev->flags); 3430 set_bit(R5_LOCKED, &dev->flags); 3431 s.locked++; 3432 } else { 3433 /* let's read it back */ 3434 set_bit(R5_Wantread, &dev->flags); 3435 set_bit(R5_LOCKED, &dev->flags); 3436 s.locked++; 3437 } 3438 } 3439 } 3440 3441 3442 /* Finish reconstruct operations initiated by the expansion process */ 3443 if (sh->reconstruct_state == reconstruct_state_result) { 3444 struct stripe_head *sh_src 3445 = get_active_stripe(conf, sh->sector, 1, 1, 1); 3446 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) { 3447 /* sh cannot be written until sh_src has been read. 3448 * so arrange for sh to be delayed a little 3449 */ 3450 set_bit(STRIPE_DELAYED, &sh->state); 3451 set_bit(STRIPE_HANDLE, &sh->state); 3452 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, 3453 &sh_src->state)) 3454 atomic_inc(&conf->preread_active_stripes); 3455 release_stripe(sh_src); 3456 goto finish; 3457 } 3458 if (sh_src) 3459 release_stripe(sh_src); 3460 3461 sh->reconstruct_state = reconstruct_state_idle; 3462 clear_bit(STRIPE_EXPANDING, &sh->state); 3463 for (i = conf->raid_disks; i--; ) { 3464 set_bit(R5_Wantwrite, &sh->dev[i].flags); 3465 set_bit(R5_LOCKED, &sh->dev[i].flags); 3466 s.locked++; 3467 } 3468 } 3469 3470 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) && 3471 !sh->reconstruct_state) { 3472 /* Need to write out all blocks after computing parity */ 3473 sh->disks = conf->raid_disks; 3474 stripe_set_idx(sh->sector, conf, 0, sh); 3475 schedule_reconstruction(sh, &s, 1, 1); 3476 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) { 3477 clear_bit(STRIPE_EXPAND_READY, &sh->state); 3478 atomic_dec(&conf->reshape_stripes); 3479 wake_up(&conf->wait_for_overlap); 3480 md_done_sync(conf->mddev, STRIPE_SECTORS, 1); 3481 } 3482 3483 if (s.expanding && s.locked == 0 && 3484 !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) 3485 handle_stripe_expansion(conf, sh); 3486 3487finish: 3488 /* wait for this device to become unblocked */ 3489 if (unlikely(s.blocked_rdev)) 3490 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev); 3491 3492 if (s.ops_request) 3493 raid_run_ops(sh, s.ops_request); 3494 3495 ops_run_io(sh, &s); 3496 3497 3498 if (s.dec_preread_active) { 3499 /* We delay this until after ops_run_io so that if make_request 3500 * is waiting on a flush, it won't continue until the writes 3501 * have actually been submitted. 3502 */ 3503 atomic_dec(&conf->preread_active_stripes); 3504 if (atomic_read(&conf->preread_active_stripes) < 3505 IO_THRESHOLD) 3506 md_wakeup_thread(conf->mddev->thread); 3507 } 3508 3509 return_io(s.return_bi); 3510 3511 clear_bit(STRIPE_ACTIVE, &sh->state); 3512} 3513 3514static void raid5_activate_delayed(raid5_conf_t *conf) 3515{ 3516 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) { 3517 while (!list_empty(&conf->delayed_list)) { 3518 struct list_head *l = conf->delayed_list.next; 3519 struct stripe_head *sh; 3520 sh = list_entry(l, struct stripe_head, lru); 3521 list_del_init(l); 3522 clear_bit(STRIPE_DELAYED, &sh->state); 3523 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 3524 atomic_inc(&conf->preread_active_stripes); 3525 list_add_tail(&sh->lru, &conf->hold_list); 3526 } 3527 } 3528} 3529 3530static void activate_bit_delay(raid5_conf_t *conf) 3531{ 3532 /* device_lock is held */ 3533 struct list_head head; 3534 list_add(&head, &conf->bitmap_list); 3535 list_del_init(&conf->bitmap_list); 3536 while (!list_empty(&head)) { 3537 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru); 3538 list_del_init(&sh->lru); 3539 atomic_inc(&sh->count); 3540 __release_stripe(conf, sh); 3541 } 3542} 3543 3544int md_raid5_congested(mddev_t *mddev, int bits) 3545{ 3546 raid5_conf_t *conf = mddev->private; 3547 3548 /* No difference between reads and writes. Just check 3549 * how busy the stripe_cache is 3550 */ 3551 3552 if (conf->inactive_blocked) 3553 return 1; 3554 if (conf->quiesce) 3555 return 1; 3556 if (list_empty_careful(&conf->inactive_list)) 3557 return 1; 3558 3559 return 0; 3560} 3561EXPORT_SYMBOL_GPL(md_raid5_congested); 3562 3563static int raid5_congested(void *data, int bits) 3564{ 3565 mddev_t *mddev = data; 3566 3567 return mddev_congested(mddev, bits) || 3568 md_raid5_congested(mddev, bits); 3569} 3570 3571/* We want read requests to align with chunks where possible, 3572 * but write requests don't need to. 3573 */ 3574static int raid5_mergeable_bvec(struct request_queue *q, 3575 struct bvec_merge_data *bvm, 3576 struct bio_vec *biovec) 3577{ 3578 mddev_t *mddev = q->queuedata; 3579 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev); 3580 int max; 3581 unsigned int chunk_sectors = mddev->chunk_sectors; 3582 unsigned int bio_sectors = bvm->bi_size >> 9; 3583 3584 if ((bvm->bi_rw & 1) == WRITE) 3585 return biovec->bv_len; /* always allow writes to be mergeable */ 3586 3587 if (mddev->new_chunk_sectors < mddev->chunk_sectors) 3588 chunk_sectors = mddev->new_chunk_sectors; 3589 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9; 3590 if (max < 0) max = 0; 3591 if (max <= biovec->bv_len && bio_sectors == 0) 3592 return biovec->bv_len; 3593 else 3594 return max; 3595} 3596 3597 3598static int in_chunk_boundary(mddev_t *mddev, struct bio *bio) 3599{ 3600 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev); 3601 unsigned int chunk_sectors = mddev->chunk_sectors; 3602 unsigned int bio_sectors = bio->bi_size >> 9; 3603 3604 if (mddev->new_chunk_sectors < mddev->chunk_sectors) 3605 chunk_sectors = mddev->new_chunk_sectors; 3606 return chunk_sectors >= 3607 ((sector & (chunk_sectors - 1)) + bio_sectors); 3608} 3609 3610/* 3611 * add bio to the retry LIFO ( in O(1) ... we are in interrupt ) 3612 * later sampled by raid5d. 3613 */ 3614static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf) 3615{ 3616 unsigned long flags; 3617 3618 spin_lock_irqsave(&conf->device_lock, flags); 3619 3620 bi->bi_next = conf->retry_read_aligned_list; 3621 conf->retry_read_aligned_list = bi; 3622 3623 spin_unlock_irqrestore(&conf->device_lock, flags); 3624 md_wakeup_thread(conf->mddev->thread); 3625} 3626 3627 3628static struct bio *remove_bio_from_retry(raid5_conf_t *conf) 3629{ 3630 struct bio *bi; 3631 3632 bi = conf->retry_read_aligned; 3633 if (bi) { 3634 conf->retry_read_aligned = NULL; 3635 return bi; 3636 } 3637 bi = conf->retry_read_aligned_list; 3638 if(bi) { 3639 conf->retry_read_aligned_list = bi->bi_next; 3640 bi->bi_next = NULL; 3641 /* 3642 * this sets the active strip count to 1 and the processed 3643 * strip count to zero (upper 8 bits) 3644 */ 3645 bi->bi_phys_segments = 1; /* biased count of active stripes */ 3646 } 3647 3648 return bi; 3649} 3650 3651 3652/* 3653 * The "raid5_align_endio" should check if the read succeeded and if it 3654 * did, call bio_endio on the original bio (having bio_put the new bio 3655 * first). 3656 * If the read failed.. 3657 */ 3658static void raid5_align_endio(struct bio *bi, int error) 3659{ 3660 struct bio* raid_bi = bi->bi_private; 3661 mddev_t *mddev; 3662 raid5_conf_t *conf; 3663 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags); 3664 mdk_rdev_t *rdev; 3665 3666 bio_put(bi); 3667 3668 rdev = (void*)raid_bi->bi_next; 3669 raid_bi->bi_next = NULL; 3670 mddev = rdev->mddev; 3671 conf = mddev->private; 3672 3673 rdev_dec_pending(rdev, conf->mddev); 3674 3675 if (!error && uptodate) { 3676 bio_endio(raid_bi, 0); 3677 if (atomic_dec_and_test(&conf->active_aligned_reads)) 3678 wake_up(&conf->wait_for_stripe); 3679 return; 3680 } 3681 3682 3683 pr_debug("raid5_align_endio : io error...handing IO for a retry\n"); 3684 3685 add_bio_to_retry(raid_bi, conf); 3686} 3687 3688static int bio_fits_rdev(struct bio *bi) 3689{ 3690 struct request_queue *q = bdev_get_queue(bi->bi_bdev); 3691 3692 if ((bi->bi_size>>9) > queue_max_sectors(q)) 3693 return 0; 3694 blk_recount_segments(q, bi); 3695 if (bi->bi_phys_segments > queue_max_segments(q)) 3696 return 0; 3697 3698 if (q->merge_bvec_fn) 3699 /* it's too hard to apply the merge_bvec_fn at this stage, 3700 * just just give up 3701 */ 3702 return 0; 3703 3704 return 1; 3705} 3706 3707 3708static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio) 3709{ 3710 raid5_conf_t *conf = mddev->private; 3711 int dd_idx; 3712 struct bio* align_bi; 3713 mdk_rdev_t *rdev; 3714 3715 if (!in_chunk_boundary(mddev, raid_bio)) { 3716 pr_debug("chunk_aligned_read : non aligned\n"); 3717 return 0; 3718 } 3719 /* 3720 * use bio_clone_mddev to make a copy of the bio 3721 */ 3722 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev); 3723 if (!align_bi) 3724 return 0; 3725 /* 3726 * set bi_end_io to a new function, and set bi_private to the 3727 * original bio. 3728 */ 3729 align_bi->bi_end_io = raid5_align_endio; 3730 align_bi->bi_private = raid_bio; 3731 /* 3732 * compute position 3733 */ 3734 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector, 3735 0, 3736 &dd_idx, NULL); 3737 3738 rcu_read_lock(); 3739 rdev = rcu_dereference(conf->disks[dd_idx].rdev); 3740 if (rdev && test_bit(In_sync, &rdev->flags)) { 3741 atomic_inc(&rdev->nr_pending); 3742 rcu_read_unlock(); 3743 raid_bio->bi_next = (void*)rdev; 3744 align_bi->bi_bdev = rdev->bdev; 3745 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID); 3746 align_bi->bi_sector += rdev->data_offset; 3747 3748 if (!bio_fits_rdev(align_bi)) { 3749 /* too big in some way */ 3750 bio_put(align_bi); 3751 rdev_dec_pending(rdev, mddev); 3752 return 0; 3753 } 3754 3755 spin_lock_irq(&conf->device_lock); 3756 wait_event_lock_irq(conf->wait_for_stripe, 3757 conf->quiesce == 0, 3758 conf->device_lock, /* nothing */); 3759 atomic_inc(&conf->active_aligned_reads); 3760 spin_unlock_irq(&conf->device_lock); 3761 3762 generic_make_request(align_bi); 3763 return 1; 3764 } else { 3765 rcu_read_unlock(); 3766 bio_put(align_bi); 3767 return 0; 3768 } 3769} 3770 3771/* __get_priority_stripe - get the next stripe to process 3772 * 3773 * Full stripe writes are allowed to pass preread active stripes up until 3774 * the bypass_threshold is exceeded. In general the bypass_count 3775 * increments when the handle_list is handled before the hold_list; however, it 3776 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a 3777 * stripe with in flight i/o. The bypass_count will be reset when the 3778 * head of the hold_list has changed, i.e. the head was promoted to the 3779 * handle_list. 3780 */ 3781static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf) 3782{ 3783 struct stripe_head *sh; 3784 3785 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n", 3786 __func__, 3787 list_empty(&conf->handle_list) ? "empty" : "busy", 3788 list_empty(&conf->hold_list) ? "empty" : "busy", 3789 atomic_read(&conf->pending_full_writes), conf->bypass_count); 3790 3791 if (!list_empty(&conf->handle_list)) { 3792 sh = list_entry(conf->handle_list.next, typeof(*sh), lru); 3793 3794 if (list_empty(&conf->hold_list)) 3795 conf->bypass_count = 0; 3796 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) { 3797 if (conf->hold_list.next == conf->last_hold) 3798 conf->bypass_count++; 3799 else { 3800 conf->last_hold = conf->hold_list.next; 3801 conf->bypass_count -= conf->bypass_threshold; 3802 if (conf->bypass_count < 0) 3803 conf->bypass_count = 0; 3804 } 3805 } 3806 } else if (!list_empty(&conf->hold_list) && 3807 ((conf->bypass_threshold && 3808 conf->bypass_count > conf->bypass_threshold) || 3809 atomic_read(&conf->pending_full_writes) == 0)) { 3810 sh = list_entry(conf->hold_list.next, 3811 typeof(*sh), lru); 3812 conf->bypass_count -= conf->bypass_threshold; 3813 if (conf->bypass_count < 0) 3814 conf->bypass_count = 0; 3815 } else 3816 return NULL; 3817 3818 list_del_init(&sh->lru); 3819 atomic_inc(&sh->count); 3820 BUG_ON(atomic_read(&sh->count) != 1); 3821 return sh; 3822} 3823 3824static int make_request(mddev_t *mddev, struct bio * bi) 3825{ 3826 raid5_conf_t *conf = mddev->private; 3827 int dd_idx; 3828 sector_t new_sector; 3829 sector_t logical_sector, last_sector; 3830 struct stripe_head *sh; 3831 const int rw = bio_data_dir(bi); 3832 int remaining; 3833 int plugged; 3834 3835 if (unlikely(bi->bi_rw & REQ_FLUSH)) { 3836 md_flush_request(mddev, bi); 3837 return 0; 3838 } 3839 3840 md_write_start(mddev, bi); 3841 3842 if (rw == READ && 3843 mddev->reshape_position == MaxSector && 3844 chunk_aligned_read(mddev,bi)) 3845 return 0; 3846 3847 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1); 3848 last_sector = bi->bi_sector + (bi->bi_size>>9); 3849 bi->bi_next = NULL; 3850 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */ 3851 3852 plugged = mddev_check_plugged(mddev); 3853 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) { 3854 DEFINE_WAIT(w); 3855 int disks, data_disks; 3856 int previous; 3857 3858 retry: 3859 previous = 0; 3860 disks = conf->raid_disks; 3861 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE); 3862 if (unlikely(conf->reshape_progress != MaxSector)) { 3863 /* spinlock is needed as reshape_progress may be 3864 * 64bit on a 32bit platform, and so it might be 3865 * possible to see a half-updated value 3866 * Of course reshape_progress could change after 3867 * the lock is dropped, so once we get a reference 3868 * to the stripe that we think it is, we will have 3869 * to check again. 3870 */ 3871 spin_lock_irq(&conf->device_lock); 3872 if (mddev->delta_disks < 0 3873 ? logical_sector < conf->reshape_progress 3874 : logical_sector >= conf->reshape_progress) { 3875 disks = conf->previous_raid_disks; 3876 previous = 1; 3877 } else { 3878 if (mddev->delta_disks < 0 3879 ? logical_sector < conf->reshape_safe 3880 : logical_sector >= conf->reshape_safe) { 3881 spin_unlock_irq(&conf->device_lock); 3882 schedule(); 3883 goto retry; 3884 } 3885 } 3886 spin_unlock_irq(&conf->device_lock); 3887 } 3888 data_disks = disks - conf->max_degraded; 3889 3890 new_sector = raid5_compute_sector(conf, logical_sector, 3891 previous, 3892 &dd_idx, NULL); 3893 pr_debug("raid456: make_request, sector %llu logical %llu\n", 3894 (unsigned long long)new_sector, 3895 (unsigned long long)logical_sector); 3896 3897 sh = get_active_stripe(conf, new_sector, previous, 3898 (bi->bi_rw&RWA_MASK), 0); 3899 if (sh) { 3900 if (unlikely(previous)) { 3901 /* expansion might have moved on while waiting for a 3902 * stripe, so we must do the range check again. 3903 * Expansion could still move past after this 3904 * test, but as we are holding a reference to 3905 * 'sh', we know that if that happens, 3906 * STRIPE_EXPANDING will get set and the expansion 3907 * won't proceed until we finish with the stripe. 3908 */ 3909 int must_retry = 0; 3910 spin_lock_irq(&conf->device_lock); 3911 if (mddev->delta_disks < 0 3912 ? logical_sector >= conf->reshape_progress 3913 : logical_sector < conf->reshape_progress) 3914 /* mismatch, need to try again */ 3915 must_retry = 1; 3916 spin_unlock_irq(&conf->device_lock); 3917 if (must_retry) { 3918 release_stripe(sh); 3919 schedule(); 3920 goto retry; 3921 } 3922 } 3923 3924 if (rw == WRITE && 3925 logical_sector >= mddev->suspend_lo && 3926 logical_sector < mddev->suspend_hi) { 3927 release_stripe(sh); 3928 /* As the suspend_* range is controlled by 3929 * userspace, we want an interruptible 3930 * wait. 3931 */ 3932 flush_signals(current); 3933 prepare_to_wait(&conf->wait_for_overlap, 3934 &w, TASK_INTERRUPTIBLE); 3935 if (logical_sector >= mddev->suspend_lo && 3936 logical_sector < mddev->suspend_hi) 3937 schedule(); 3938 goto retry; 3939 } 3940 3941 if (test_bit(STRIPE_EXPANDING, &sh->state) || 3942 !add_stripe_bio(sh, bi, dd_idx, rw)) { 3943 /* Stripe is busy expanding or 3944 * add failed due to overlap. Flush everything 3945 * and wait a while 3946 */ 3947 md_wakeup_thread(mddev->thread); 3948 release_stripe(sh); 3949 schedule(); 3950 goto retry; 3951 } 3952 finish_wait(&conf->wait_for_overlap, &w); 3953 set_bit(STRIPE_HANDLE, &sh->state); 3954 clear_bit(STRIPE_DELAYED, &sh->state); 3955 if ((bi->bi_rw & REQ_SYNC) && 3956 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 3957 atomic_inc(&conf->preread_active_stripes); 3958 release_stripe(sh); 3959 } else { 3960 /* cannot get stripe for read-ahead, just give-up */ 3961 clear_bit(BIO_UPTODATE, &bi->bi_flags); 3962 finish_wait(&conf->wait_for_overlap, &w); 3963 break; 3964 } 3965 3966 } 3967 if (!plugged) 3968 md_wakeup_thread(mddev->thread); 3969 3970 spin_lock_irq(&conf->device_lock); 3971 remaining = raid5_dec_bi_phys_segments(bi); 3972 spin_unlock_irq(&conf->device_lock); 3973 if (remaining == 0) { 3974 3975 if ( rw == WRITE ) 3976 md_write_end(mddev); 3977 3978 bio_endio(bi, 0); 3979 } 3980 3981 return 0; 3982} 3983 3984static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks); 3985 3986static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped) 3987{ 3988 /* reshaping is quite different to recovery/resync so it is 3989 * handled quite separately ... here. 3990 * 3991 * On each call to sync_request, we gather one chunk worth of 3992 * destination stripes and flag them as expanding. 3993 * Then we find all the source stripes and request reads. 3994 * As the reads complete, handle_stripe will copy the data 3995 * into the destination stripe and release that stripe. 3996 */ 3997 raid5_conf_t *conf = mddev->private; 3998 struct stripe_head *sh; 3999 sector_t first_sector, last_sector; 4000 int raid_disks = conf->previous_raid_disks; 4001 int data_disks = raid_disks - conf->max_degraded; 4002 int new_data_disks = conf->raid_disks - conf->max_degraded; 4003 int i; 4004 int dd_idx; 4005 sector_t writepos, readpos, safepos; 4006 sector_t stripe_addr; 4007 int reshape_sectors; 4008 struct list_head stripes; 4009 4010 if (sector_nr == 0) { 4011 /* If restarting in the middle, skip the initial sectors */ 4012 if (mddev->delta_disks < 0 && 4013 conf->reshape_progress < raid5_size(mddev, 0, 0)) { 4014 sector_nr = raid5_size(mddev, 0, 0) 4015 - conf->reshape_progress; 4016 } else if (mddev->delta_disks >= 0 && 4017 conf->reshape_progress > 0) 4018 sector_nr = conf->reshape_progress; 4019 sector_div(sector_nr, new_data_disks); 4020 if (sector_nr) { 4021 mddev->curr_resync_completed = sector_nr; 4022 sysfs_notify(&mddev->kobj, NULL, "sync_completed"); 4023 *skipped = 1; 4024 return sector_nr; 4025 } 4026 } 4027 4028 /* We need to process a full chunk at a time. 4029 * If old and new chunk sizes differ, we need to process the 4030 * largest of these 4031 */ 4032 if (mddev->new_chunk_sectors > mddev->chunk_sectors) 4033 reshape_sectors = mddev->new_chunk_sectors; 4034 else 4035 reshape_sectors = mddev->chunk_sectors; 4036 4037 /* we update the metadata when there is more than 3Meg 4038 * in the block range (that is rather arbitrary, should 4039 * probably be time based) or when the data about to be 4040 * copied would over-write the source of the data at 4041 * the front of the range. 4042 * i.e. one new_stripe along from reshape_progress new_maps 4043 * to after where reshape_safe old_maps to 4044 */ 4045 writepos = conf->reshape_progress; 4046 sector_div(writepos, new_data_disks); 4047 readpos = conf->reshape_progress; 4048 sector_div(readpos, data_disks); 4049 safepos = conf->reshape_safe; 4050 sector_div(safepos, data_disks); 4051 if (mddev->delta_disks < 0) { 4052 writepos -= min_t(sector_t, reshape_sectors, writepos); 4053 readpos += reshape_sectors; 4054 safepos += reshape_sectors; 4055 } else { 4056 writepos += reshape_sectors; 4057 readpos -= min_t(sector_t, reshape_sectors, readpos); 4058 safepos -= min_t(sector_t, reshape_sectors, safepos); 4059 } 4060 4061 /* 'writepos' is the most advanced device address we might write. 4062 * 'readpos' is the least advanced device address we might read. 4063 * 'safepos' is the least address recorded in the metadata as having 4064 * been reshaped. 4065 * If 'readpos' is behind 'writepos', then there is no way that we can 4066 * ensure safety in the face of a crash - that must be done by userspace 4067 * making a backup of the data. So in that case there is no particular 4068 * rush to update metadata. 4069 * Otherwise if 'safepos' is behind 'writepos', then we really need to 4070 * update the metadata to advance 'safepos' to match 'readpos' so that 4071 * we can be safe in the event of a crash. 4072 * So we insist on updating metadata if safepos is behind writepos and 4073 * readpos is beyond writepos. 4074 * In any case, update the metadata every 10 seconds. 4075 * Maybe that number should be configurable, but I'm not sure it is 4076 * worth it.... maybe it could be a multiple of safemode_delay??? 4077 */ 4078 if ((mddev->delta_disks < 0 4079 ? (safepos > writepos && readpos < writepos) 4080 : (safepos < writepos && readpos > writepos)) || 4081 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) { 4082 /* Cannot proceed until we've updated the superblock... */ 4083 wait_event(conf->wait_for_overlap, 4084 atomic_read(&conf->reshape_stripes)==0); 4085 mddev->reshape_position = conf->reshape_progress; 4086 mddev->curr_resync_completed = sector_nr; 4087 conf->reshape_checkpoint = jiffies; 4088 set_bit(MD_CHANGE_DEVS, &mddev->flags); 4089 md_wakeup_thread(mddev->thread); 4090 wait_event(mddev->sb_wait, mddev->flags == 0 || 4091 kthread_should_stop()); 4092 spin_lock_irq(&conf->device_lock); 4093 conf->reshape_safe = mddev->reshape_position; 4094 spin_unlock_irq(&conf->device_lock); 4095 wake_up(&conf->wait_for_overlap); 4096 sysfs_notify(&mddev->kobj, NULL, "sync_completed"); 4097 } 4098 4099 if (mddev->delta_disks < 0) { 4100 BUG_ON(conf->reshape_progress == 0); 4101 stripe_addr = writepos; 4102 BUG_ON((mddev->dev_sectors & 4103 ~((sector_t)reshape_sectors - 1)) 4104 - reshape_sectors - stripe_addr 4105 != sector_nr); 4106 } else { 4107 BUG_ON(writepos != sector_nr + reshape_sectors); 4108 stripe_addr = sector_nr; 4109 } 4110 INIT_LIST_HEAD(&stripes); 4111 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) { 4112 int j; 4113 int skipped_disk = 0; 4114 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1); 4115 set_bit(STRIPE_EXPANDING, &sh->state); 4116 atomic_inc(&conf->reshape_stripes); 4117 /* If any of this stripe is beyond the end of the old 4118 * array, then we need to zero those blocks 4119 */ 4120 for (j=sh->disks; j--;) { 4121 sector_t s; 4122 if (j == sh->pd_idx) 4123 continue; 4124 if (conf->level == 6 && 4125 j == sh->qd_idx) 4126 continue; 4127 s = compute_blocknr(sh, j, 0); 4128 if (s < raid5_size(mddev, 0, 0)) { 4129 skipped_disk = 1; 4130 continue; 4131 } 4132 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE); 4133 set_bit(R5_Expanded, &sh->dev[j].flags); 4134 set_bit(R5_UPTODATE, &sh->dev[j].flags); 4135 } 4136 if (!skipped_disk) { 4137 set_bit(STRIPE_EXPAND_READY, &sh->state); 4138 set_bit(STRIPE_HANDLE, &sh->state); 4139 } 4140 list_add(&sh->lru, &stripes); 4141 } 4142 spin_lock_irq(&conf->device_lock); 4143 if (mddev->delta_disks < 0) 4144 conf->reshape_progress -= reshape_sectors * new_data_disks; 4145 else 4146 conf->reshape_progress += reshape_sectors * new_data_disks; 4147 spin_unlock_irq(&conf->device_lock); 4148 /* Ok, those stripe are ready. We can start scheduling 4149 * reads on the source stripes. 4150 * The source stripes are determined by mapping the first and last 4151 * block on the destination stripes. 4152 */ 4153 first_sector = 4154 raid5_compute_sector(conf, stripe_addr*(new_data_disks), 4155 1, &dd_idx, NULL); 4156 last_sector = 4157 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors) 4158 * new_data_disks - 1), 4159 1, &dd_idx, NULL); 4160 if (last_sector >= mddev->dev_sectors) 4161 last_sector = mddev->dev_sectors - 1; 4162 while (first_sector <= last_sector) { 4163 sh = get_active_stripe(conf, first_sector, 1, 0, 1); 4164 set_bit(STRIPE_EXPAND_SOURCE, &sh->state); 4165 set_bit(STRIPE_HANDLE, &sh->state); 4166 release_stripe(sh); 4167 first_sector += STRIPE_SECTORS; 4168 } 4169 /* Now that the sources are clearly marked, we can release 4170 * the destination stripes 4171 */ 4172 while (!list_empty(&stripes)) { 4173 sh = list_entry(stripes.next, struct stripe_head, lru); 4174 list_del_init(&sh->lru); 4175 release_stripe(sh); 4176 } 4177 /* If this takes us to the resync_max point where we have to pause, 4178 * then we need to write out the superblock. 4179 */ 4180 sector_nr += reshape_sectors; 4181 if ((sector_nr - mddev->curr_resync_completed) * 2 4182 >= mddev->resync_max - mddev->curr_resync_completed) { 4183 /* Cannot proceed until we've updated the superblock... */ 4184 wait_event(conf->wait_for_overlap, 4185 atomic_read(&conf->reshape_stripes) == 0); 4186 mddev->reshape_position = conf->reshape_progress; 4187 mddev->curr_resync_completed = sector_nr; 4188 conf->reshape_checkpoint = jiffies; 4189 set_bit(MD_CHANGE_DEVS, &mddev->flags); 4190 md_wakeup_thread(mddev->thread); 4191 wait_event(mddev->sb_wait, 4192 !test_bit(MD_CHANGE_DEVS, &mddev->flags) 4193 || kthread_should_stop()); 4194 spin_lock_irq(&conf->device_lock); 4195 conf->reshape_safe = mddev->reshape_position; 4196 spin_unlock_irq(&conf->device_lock); 4197 wake_up(&conf->wait_for_overlap); 4198 sysfs_notify(&mddev->kobj, NULL, "sync_completed"); 4199 } 4200 return reshape_sectors; 4201} 4202 4203/* FIXME go_faster isn't used */ 4204static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster) 4205{ 4206 raid5_conf_t *conf = mddev->private; 4207 struct stripe_head *sh; 4208 sector_t max_sector = mddev->dev_sectors; 4209 sector_t sync_blocks; 4210 int still_degraded = 0; 4211 int i; 4212 4213 if (sector_nr >= max_sector) { 4214 /* just being told to finish up .. nothing much to do */ 4215 4216 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) { 4217 end_reshape(conf); 4218 return 0; 4219 } 4220 4221 if (mddev->curr_resync < max_sector) /* aborted */ 4222 bitmap_end_sync(mddev->bitmap, mddev->curr_resync, 4223 &sync_blocks, 1); 4224 else /* completed sync */ 4225 conf->fullsync = 0; 4226 bitmap_close_sync(mddev->bitmap); 4227 4228 return 0; 4229 } 4230 4231 /* Allow raid5_quiesce to complete */ 4232 wait_event(conf->wait_for_overlap, conf->quiesce != 2); 4233 4234 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) 4235 return reshape_request(mddev, sector_nr, skipped); 4236 4237 /* No need to check resync_max as we never do more than one 4238 * stripe, and as resync_max will always be on a chunk boundary, 4239 * if the check in md_do_sync didn't fire, there is no chance 4240 * of overstepping resync_max here 4241 */ 4242 4243 /* if there is too many failed drives and we are trying 4244 * to resync, then assert that we are finished, because there is 4245 * nothing we can do. 4246 */ 4247 if (mddev->degraded >= conf->max_degraded && 4248 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) { 4249 sector_t rv = mddev->dev_sectors - sector_nr; 4250 *skipped = 1; 4251 return rv; 4252 } 4253 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) && 4254 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && 4255 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) { 4256 /* we can skip this block, and probably more */ 4257 sync_blocks /= STRIPE_SECTORS; 4258 *skipped = 1; 4259 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */ 4260 } 4261 4262 4263 bitmap_cond_end_sync(mddev->bitmap, sector_nr); 4264 4265 sh = get_active_stripe(conf, sector_nr, 0, 1, 0); 4266 if (sh == NULL) { 4267 sh = get_active_stripe(conf, sector_nr, 0, 0, 0); 4268 /* make sure we don't swamp the stripe cache if someone else 4269 * is trying to get access 4270 */ 4271 schedule_timeout_uninterruptible(1); 4272 } 4273 /* Need to check if array will still be degraded after recovery/resync 4274 * We don't need to check the 'failed' flag as when that gets set, 4275 * recovery aborts. 4276 */ 4277 for (i = 0; i < conf->raid_disks; i++) 4278 if (conf->disks[i].rdev == NULL) 4279 still_degraded = 1; 4280 4281 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded); 4282 4283 set_bit(STRIPE_SYNC_REQUESTED, &sh->state); 4284 4285 handle_stripe(sh); 4286 release_stripe(sh); 4287 4288 return STRIPE_SECTORS; 4289} 4290 4291static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio) 4292{ 4293 /* We may not be able to submit a whole bio at once as there 4294 * may not be enough stripe_heads available. 4295 * We cannot pre-allocate enough stripe_heads as we may need 4296 * more than exist in the cache (if we allow ever large chunks). 4297 * So we do one stripe head at a time and record in 4298 * ->bi_hw_segments how many have been done. 4299 * 4300 * We *know* that this entire raid_bio is in one chunk, so 4301 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector. 4302 */ 4303 struct stripe_head *sh; 4304 int dd_idx; 4305 sector_t sector, logical_sector, last_sector; 4306 int scnt = 0; 4307 int remaining; 4308 int handled = 0; 4309 4310 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1); 4311 sector = raid5_compute_sector(conf, logical_sector, 4312 0, &dd_idx, NULL); 4313 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9); 4314 4315 for (; logical_sector < last_sector; 4316 logical_sector += STRIPE_SECTORS, 4317 sector += STRIPE_SECTORS, 4318 scnt++) { 4319 4320 if (scnt < raid5_bi_hw_segments(raid_bio)) 4321 /* already done this stripe */ 4322 continue; 4323 4324 sh = get_active_stripe(conf, sector, 0, 1, 0); 4325 4326 if (!sh) { 4327 /* failed to get a stripe - must wait */ 4328 raid5_set_bi_hw_segments(raid_bio, scnt); 4329 conf->retry_read_aligned = raid_bio; 4330 return handled; 4331 } 4332 4333 set_bit(R5_ReadError, &sh->dev[dd_idx].flags); 4334 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) { 4335 release_stripe(sh); 4336 raid5_set_bi_hw_segments(raid_bio, scnt); 4337 conf->retry_read_aligned = raid_bio; 4338 return handled; 4339 } 4340 4341 handle_stripe(sh); 4342 release_stripe(sh); 4343 handled++; 4344 } 4345 spin_lock_irq(&conf->device_lock); 4346 remaining = raid5_dec_bi_phys_segments(raid_bio); 4347 spin_unlock_irq(&conf->device_lock); 4348 if (remaining == 0) 4349 bio_endio(raid_bio, 0); 4350 if (atomic_dec_and_test(&conf->active_aligned_reads)) 4351 wake_up(&conf->wait_for_stripe); 4352 return handled; 4353} 4354 4355 4356/* 4357 * This is our raid5 kernel thread. 4358 * 4359 * We scan the hash table for stripes which can be handled now. 4360 * During the scan, completed stripes are saved for us by the interrupt 4361 * handler, so that they will not have to wait for our next wakeup. 4362 */ 4363static void raid5d(mddev_t *mddev) 4364{ 4365 struct stripe_head *sh; 4366 raid5_conf_t *conf = mddev->private; 4367 int handled; 4368 struct blk_plug plug; 4369 4370 pr_debug("+++ raid5d active\n"); 4371 4372 md_check_recovery(mddev); 4373 4374 blk_start_plug(&plug); 4375 handled = 0; 4376 spin_lock_irq(&conf->device_lock); 4377 while (1) { 4378 struct bio *bio; 4379 4380 if (atomic_read(&mddev->plug_cnt) == 0 && 4381 !list_empty(&conf->bitmap_list)) { 4382 /* Now is a good time to flush some bitmap updates */ 4383 conf->seq_flush++; 4384 spin_unlock_irq(&conf->device_lock); 4385 bitmap_unplug(mddev->bitmap); 4386 spin_lock_irq(&conf->device_lock); 4387 conf->seq_write = conf->seq_flush; 4388 activate_bit_delay(conf); 4389 } 4390 if (atomic_read(&mddev->plug_cnt) == 0) 4391 raid5_activate_delayed(conf); 4392 4393 while ((bio = remove_bio_from_retry(conf))) { 4394 int ok; 4395 spin_unlock_irq(&conf->device_lock); 4396 ok = retry_aligned_read(conf, bio); 4397 spin_lock_irq(&conf->device_lock); 4398 if (!ok) 4399 break; 4400 handled++; 4401 } 4402 4403 sh = __get_priority_stripe(conf); 4404 4405 if (!sh) 4406 break; 4407 spin_unlock_irq(&conf->device_lock); 4408 4409 handled++; 4410 handle_stripe(sh); 4411 release_stripe(sh); 4412 cond_resched(); 4413 4414 spin_lock_irq(&conf->device_lock); 4415 } 4416 pr_debug("%d stripes handled\n", handled); 4417 4418 spin_unlock_irq(&conf->device_lock); 4419 4420 async_tx_issue_pending_all(); 4421 blk_finish_plug(&plug); 4422 4423 pr_debug("--- raid5d inactive\n"); 4424} 4425 4426static ssize_t 4427raid5_show_stripe_cache_size(mddev_t *mddev, char *page) 4428{ 4429 raid5_conf_t *conf = mddev->private; 4430 if (conf) 4431 return sprintf(page, "%d\n", conf->max_nr_stripes); 4432 else 4433 return 0; 4434} 4435 4436int 4437raid5_set_cache_size(mddev_t *mddev, int size) 4438{ 4439 raid5_conf_t *conf = mddev->private; 4440 int err; 4441 4442 if (size <= 16 || size > 32768) 4443 return -EINVAL; 4444 while (size < conf->max_nr_stripes) { 4445 if (drop_one_stripe(conf)) 4446 conf->max_nr_stripes--; 4447 else 4448 break; 4449 } 4450 err = md_allow_write(mddev); 4451 if (err) 4452 return err; 4453 while (size > conf->max_nr_stripes) { 4454 if (grow_one_stripe(conf)) 4455 conf->max_nr_stripes++; 4456 else break; 4457 } 4458 return 0; 4459} 4460EXPORT_SYMBOL(raid5_set_cache_size); 4461 4462static ssize_t 4463raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len) 4464{ 4465 raid5_conf_t *conf = mddev->private; 4466 unsigned long new; 4467 int err; 4468 4469 if (len >= PAGE_SIZE) 4470 return -EINVAL; 4471 if (!conf) 4472 return -ENODEV; 4473 4474 if (strict_strtoul(page, 10, &new)) 4475 return -EINVAL; 4476 err = raid5_set_cache_size(mddev, new); 4477 if (err) 4478 return err; 4479 return len; 4480} 4481 4482static struct md_sysfs_entry 4483raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR, 4484 raid5_show_stripe_cache_size, 4485 raid5_store_stripe_cache_size); 4486 4487static ssize_t 4488raid5_show_preread_threshold(mddev_t *mddev, char *page) 4489{ 4490 raid5_conf_t *conf = mddev->private; 4491 if (conf) 4492 return sprintf(page, "%d\n", conf->bypass_threshold); 4493 else 4494 return 0; 4495} 4496 4497static ssize_t 4498raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len) 4499{ 4500 raid5_conf_t *conf = mddev->private; 4501 unsigned long new; 4502 if (len >= PAGE_SIZE) 4503 return -EINVAL; 4504 if (!conf) 4505 return -ENODEV; 4506 4507 if (strict_strtoul(page, 10, &new)) 4508 return -EINVAL; 4509 if (new > conf->max_nr_stripes) 4510 return -EINVAL; 4511 conf->bypass_threshold = new; 4512 return len; 4513} 4514 4515static struct md_sysfs_entry 4516raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold, 4517 S_IRUGO | S_IWUSR, 4518 raid5_show_preread_threshold, 4519 raid5_store_preread_threshold); 4520 4521static ssize_t 4522stripe_cache_active_show(mddev_t *mddev, char *page) 4523{ 4524 raid5_conf_t *conf = mddev->private; 4525 if (conf) 4526 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes)); 4527 else 4528 return 0; 4529} 4530 4531static struct md_sysfs_entry 4532raid5_stripecache_active = __ATTR_RO(stripe_cache_active); 4533 4534static struct attribute *raid5_attrs[] = { 4535 &raid5_stripecache_size.attr, 4536 &raid5_stripecache_active.attr, 4537 &raid5_preread_bypass_threshold.attr, 4538 NULL, 4539}; 4540static struct attribute_group raid5_attrs_group = { 4541 .name = NULL, 4542 .attrs = raid5_attrs, 4543}; 4544 4545static sector_t 4546raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks) 4547{ 4548 raid5_conf_t *conf = mddev->private; 4549 4550 if (!sectors) 4551 sectors = mddev->dev_sectors; 4552 if (!raid_disks) 4553 /* size is defined by the smallest of previous and new size */ 4554 raid_disks = min(conf->raid_disks, conf->previous_raid_disks); 4555 4556 sectors &= ~((sector_t)mddev->chunk_sectors - 1); 4557 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1); 4558 return sectors * (raid_disks - conf->max_degraded); 4559} 4560 4561static void raid5_free_percpu(raid5_conf_t *conf) 4562{ 4563 struct raid5_percpu *percpu; 4564 unsigned long cpu; 4565 4566 if (!conf->percpu) 4567 return; 4568 4569 get_online_cpus(); 4570 for_each_possible_cpu(cpu) { 4571 percpu = per_cpu_ptr(conf->percpu, cpu); 4572 safe_put_page(percpu->spare_page); 4573 kfree(percpu->scribble); 4574 } 4575#ifdef CONFIG_HOTPLUG_CPU 4576 unregister_cpu_notifier(&conf->cpu_notify); 4577#endif 4578 put_online_cpus(); 4579 4580 free_percpu(conf->percpu); 4581} 4582 4583static void free_conf(raid5_conf_t *conf) 4584{ 4585 shrink_stripes(conf); 4586 raid5_free_percpu(conf); 4587 kfree(conf->disks); 4588 kfree(conf->stripe_hashtbl); 4589 kfree(conf); 4590} 4591 4592#ifdef CONFIG_HOTPLUG_CPU 4593static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action, 4594 void *hcpu) 4595{ 4596 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify); 4597 long cpu = (long)hcpu; 4598 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu); 4599 4600 switch (action) { 4601 case CPU_UP_PREPARE: 4602 case CPU_UP_PREPARE_FROZEN: 4603 if (conf->level == 6 && !percpu->spare_page) 4604 percpu->spare_page = alloc_page(GFP_KERNEL); 4605 if (!percpu->scribble) 4606 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL); 4607 4608 if (!percpu->scribble || 4609 (conf->level == 6 && !percpu->spare_page)) { 4610 safe_put_page(percpu->spare_page); 4611 kfree(percpu->scribble); 4612 pr_err("%s: failed memory allocation for cpu%ld\n", 4613 __func__, cpu); 4614 return notifier_from_errno(-ENOMEM); 4615 } 4616 break; 4617 case CPU_DEAD: 4618 case CPU_DEAD_FROZEN: 4619 safe_put_page(percpu->spare_page); 4620 kfree(percpu->scribble); 4621 percpu->spare_page = NULL; 4622 percpu->scribble = NULL; 4623 break; 4624 default: 4625 break; 4626 } 4627 return NOTIFY_OK; 4628} 4629#endif 4630 4631static int raid5_alloc_percpu(raid5_conf_t *conf) 4632{ 4633 unsigned long cpu; 4634 struct page *spare_page; 4635 struct raid5_percpu __percpu *allcpus; 4636 void *scribble; 4637 int err; 4638 4639 allcpus = alloc_percpu(struct raid5_percpu); 4640 if (!allcpus) 4641 return -ENOMEM; 4642 conf->percpu = allcpus; 4643 4644 get_online_cpus(); 4645 err = 0; 4646 for_each_present_cpu(cpu) { 4647 if (conf->level == 6) { 4648 spare_page = alloc_page(GFP_KERNEL); 4649 if (!spare_page) { 4650 err = -ENOMEM; 4651 break; 4652 } 4653 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page; 4654 } 4655 scribble = kmalloc(conf->scribble_len, GFP_KERNEL); 4656 if (!scribble) { 4657 err = -ENOMEM; 4658 break; 4659 } 4660 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble; 4661 } 4662#ifdef CONFIG_HOTPLUG_CPU 4663 conf->cpu_notify.notifier_call = raid456_cpu_notify; 4664 conf->cpu_notify.priority = 0; 4665 if (err == 0) 4666 err = register_cpu_notifier(&conf->cpu_notify); 4667#endif 4668 put_online_cpus(); 4669 4670 return err; 4671} 4672 4673static raid5_conf_t *setup_conf(mddev_t *mddev) 4674{ 4675 raid5_conf_t *conf; 4676 int raid_disk, memory, max_disks; 4677 mdk_rdev_t *rdev; 4678 struct disk_info *disk; 4679 4680 if (mddev->new_level != 5 4681 && mddev->new_level != 4 4682 && mddev->new_level != 6) { 4683 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n", 4684 mdname(mddev), mddev->new_level); 4685 return ERR_PTR(-EIO); 4686 } 4687 if ((mddev->new_level == 5 4688 && !algorithm_valid_raid5(mddev->new_layout)) || 4689 (mddev->new_level == 6 4690 && !algorithm_valid_raid6(mddev->new_layout))) { 4691 printk(KERN_ERR "md/raid:%s: layout %d not supported\n", 4692 mdname(mddev), mddev->new_layout); 4693 return ERR_PTR(-EIO); 4694 } 4695 if (mddev->new_level == 6 && mddev->raid_disks < 4) { 4696 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n", 4697 mdname(mddev), mddev->raid_disks); 4698 return ERR_PTR(-EINVAL); 4699 } 4700 4701 if (!mddev->new_chunk_sectors || 4702 (mddev->new_chunk_sectors << 9) % PAGE_SIZE || 4703 !is_power_of_2(mddev->new_chunk_sectors)) { 4704 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n", 4705 mdname(mddev), mddev->new_chunk_sectors << 9); 4706 return ERR_PTR(-EINVAL); 4707 } 4708 4709 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL); 4710 if (conf == NULL) 4711 goto abort; 4712 spin_lock_init(&conf->device_lock); 4713 init_waitqueue_head(&conf->wait_for_stripe); 4714 init_waitqueue_head(&conf->wait_for_overlap); 4715 INIT_LIST_HEAD(&conf->handle_list); 4716 INIT_LIST_HEAD(&conf->hold_list); 4717 INIT_LIST_HEAD(&conf->delayed_list); 4718 INIT_LIST_HEAD(&conf->bitmap_list); 4719 INIT_LIST_HEAD(&conf->inactive_list); 4720 atomic_set(&conf->active_stripes, 0); 4721 atomic_set(&conf->preread_active_stripes, 0); 4722 atomic_set(&conf->active_aligned_reads, 0); 4723 conf->bypass_threshold = BYPASS_THRESHOLD; 4724 4725 conf->raid_disks = mddev->raid_disks; 4726 if (mddev->reshape_position == MaxSector) 4727 conf->previous_raid_disks = mddev->raid_disks; 4728 else 4729 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks; 4730 max_disks = max(conf->raid_disks, conf->previous_raid_disks); 4731 conf->scribble_len = scribble_len(max_disks); 4732 4733 conf->disks = kzalloc(max_disks * sizeof(struct disk_info), 4734 GFP_KERNEL); 4735 if (!conf->disks) 4736 goto abort; 4737 4738 conf->mddev = mddev; 4739 4740 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL) 4741 goto abort; 4742 4743 conf->level = mddev->new_level; 4744 if (raid5_alloc_percpu(conf) != 0) 4745 goto abort; 4746 4747 pr_debug("raid456: run(%s) called.\n", mdname(mddev)); 4748 4749 list_for_each_entry(rdev, &mddev->disks, same_set) { 4750 raid_disk = rdev->raid_disk; 4751 if (raid_disk >= max_disks 4752 || raid_disk < 0) 4753 continue; 4754 disk = conf->disks + raid_disk; 4755 4756 disk->rdev = rdev; 4757 4758 if (test_bit(In_sync, &rdev->flags)) { 4759 char b[BDEVNAME_SIZE]; 4760 printk(KERN_INFO "md/raid:%s: device %s operational as raid" 4761 " disk %d\n", 4762 mdname(mddev), bdevname(rdev->bdev, b), raid_disk); 4763 } else if (rdev->saved_raid_disk != raid_disk) 4764 /* Cannot rely on bitmap to complete recovery */ 4765 conf->fullsync = 1; 4766 } 4767 4768 conf->chunk_sectors = mddev->new_chunk_sectors; 4769 conf->level = mddev->new_level; 4770 if (conf->level == 6) 4771 conf->max_degraded = 2; 4772 else 4773 conf->max_degraded = 1; 4774 conf->algorithm = mddev->new_layout; 4775 conf->max_nr_stripes = NR_STRIPES; 4776 conf->reshape_progress = mddev->reshape_position; 4777 if (conf->reshape_progress != MaxSector) { 4778 conf->prev_chunk_sectors = mddev->chunk_sectors; 4779 conf->prev_algo = mddev->layout; 4780 } 4781 4782 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) + 4783 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024; 4784 if (grow_stripes(conf, conf->max_nr_stripes)) { 4785 printk(KERN_ERR 4786 "md/raid:%s: couldn't allocate %dkB for buffers\n", 4787 mdname(mddev), memory); 4788 goto abort; 4789 } else 4790 printk(KERN_INFO "md/raid:%s: allocated %dkB\n", 4791 mdname(mddev), memory); 4792 4793 conf->thread = md_register_thread(raid5d, mddev, NULL); 4794 if (!conf->thread) { 4795 printk(KERN_ERR 4796 "md/raid:%s: couldn't allocate thread.\n", 4797 mdname(mddev)); 4798 goto abort; 4799 } 4800 4801 return conf; 4802 4803 abort: 4804 if (conf) { 4805 free_conf(conf); 4806 return ERR_PTR(-EIO); 4807 } else 4808 return ERR_PTR(-ENOMEM); 4809} 4810 4811 4812static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded) 4813{ 4814 switch (algo) { 4815 case ALGORITHM_PARITY_0: 4816 if (raid_disk < max_degraded) 4817 return 1; 4818 break; 4819 case ALGORITHM_PARITY_N: 4820 if (raid_disk >= raid_disks - max_degraded) 4821 return 1; 4822 break; 4823 case ALGORITHM_PARITY_0_6: 4824 if (raid_disk == 0 || 4825 raid_disk == raid_disks - 1) 4826 return 1; 4827 break; 4828 case ALGORITHM_LEFT_ASYMMETRIC_6: 4829 case ALGORITHM_RIGHT_ASYMMETRIC_6: 4830 case ALGORITHM_LEFT_SYMMETRIC_6: 4831 case ALGORITHM_RIGHT_SYMMETRIC_6: 4832 if (raid_disk == raid_disks - 1) 4833 return 1; 4834 } 4835 return 0; 4836} 4837 4838static int run(mddev_t *mddev) 4839{ 4840 raid5_conf_t *conf; 4841 int working_disks = 0; 4842 int dirty_parity_disks = 0; 4843 mdk_rdev_t *rdev; 4844 sector_t reshape_offset = 0; 4845 4846 if (mddev->recovery_cp != MaxSector) 4847 printk(KERN_NOTICE "md/raid:%s: not clean" 4848 " -- starting background reconstruction\n", 4849 mdname(mddev)); 4850 if (mddev->reshape_position != MaxSector) { 4851 /* Check that we can continue the reshape. 4852 * Currently only disks can change, it must 4853 * increase, and we must be past the point where 4854 * a stripe over-writes itself 4855 */ 4856 sector_t here_new, here_old; 4857 int old_disks; 4858 int max_degraded = (mddev->level == 6 ? 2 : 1); 4859 4860 if (mddev->new_level != mddev->level) { 4861 printk(KERN_ERR "md/raid:%s: unsupported reshape " 4862 "required - aborting.\n", 4863 mdname(mddev)); 4864 return -EINVAL; 4865 } 4866 old_disks = mddev->raid_disks - mddev->delta_disks; 4867 /* reshape_position must be on a new-stripe boundary, and one 4868 * further up in new geometry must map after here in old 4869 * geometry. 4870 */ 4871 here_new = mddev->reshape_position; 4872 if (sector_div(here_new, mddev->new_chunk_sectors * 4873 (mddev->raid_disks - max_degraded))) { 4874 printk(KERN_ERR "md/raid:%s: reshape_position not " 4875 "on a stripe boundary\n", mdname(mddev)); 4876 return -EINVAL; 4877 } 4878 reshape_offset = here_new * mddev->new_chunk_sectors; 4879 /* here_new is the stripe we will write to */ 4880 here_old = mddev->reshape_position; 4881 sector_div(here_old, mddev->chunk_sectors * 4882 (old_disks-max_degraded)); 4883 /* here_old is the first stripe that we might need to read 4884 * from */ 4885 if (mddev->delta_disks == 0) { 4886 /* We cannot be sure it is safe to start an in-place 4887 * reshape. It is only safe if user-space if monitoring 4888 * and taking constant backups. 4889 * mdadm always starts a situation like this in 4890 * readonly mode so it can take control before 4891 * allowing any writes. So just check for that. 4892 */ 4893 if ((here_new * mddev->new_chunk_sectors != 4894 here_old * mddev->chunk_sectors) || 4895 mddev->ro == 0) { 4896 printk(KERN_ERR "md/raid:%s: in-place reshape must be started" 4897 " in read-only mode - aborting\n", 4898 mdname(mddev)); 4899 return -EINVAL; 4900 } 4901 } else if (mddev->delta_disks < 0 4902 ? (here_new * mddev->new_chunk_sectors <= 4903 here_old * mddev->chunk_sectors) 4904 : (here_new * mddev->new_chunk_sectors >= 4905 here_old * mddev->chunk_sectors)) { 4906 /* Reading from the same stripe as writing to - bad */ 4907 printk(KERN_ERR "md/raid:%s: reshape_position too early for " 4908 "auto-recovery - aborting.\n", 4909 mdname(mddev)); 4910 return -EINVAL; 4911 } 4912 printk(KERN_INFO "md/raid:%s: reshape will continue\n", 4913 mdname(mddev)); 4914 /* OK, we should be able to continue; */ 4915 } else { 4916 BUG_ON(mddev->level != mddev->new_level); 4917 BUG_ON(mddev->layout != mddev->new_layout); 4918 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors); 4919 BUG_ON(mddev->delta_disks != 0); 4920 } 4921 4922 if (mddev->private == NULL) 4923 conf = setup_conf(mddev); 4924 else 4925 conf = mddev->private; 4926 4927 if (IS_ERR(conf)) 4928 return PTR_ERR(conf); 4929 4930 mddev->thread = conf->thread; 4931 conf->thread = NULL; 4932 mddev->private = conf; 4933 4934 /* 4935 * 0 for a fully functional array, 1 or 2 for a degraded array. 4936 */ 4937 list_for_each_entry(rdev, &mddev->disks, same_set) { 4938 if (rdev->raid_disk < 0) 4939 continue; 4940 if (test_bit(In_sync, &rdev->flags)) { 4941 working_disks++; 4942 continue; 4943 } 4944 /* This disc is not fully in-sync. However if it 4945 * just stored parity (beyond the recovery_offset), 4946 * when we don't need to be concerned about the 4947 * array being dirty. 4948 * When reshape goes 'backwards', we never have 4949 * partially completed devices, so we only need 4950 * to worry about reshape going forwards. 4951 */ 4952 /* Hack because v0.91 doesn't store recovery_offset properly. */ 4953 if (mddev->major_version == 0 && 4954 mddev->minor_version > 90) 4955 rdev->recovery_offset = reshape_offset; 4956 4957 if (rdev->recovery_offset < reshape_offset) { 4958 /* We need to check old and new layout */ 4959 if (!only_parity(rdev->raid_disk, 4960 conf->algorithm, 4961 conf->raid_disks, 4962 conf->max_degraded)) 4963 continue; 4964 } 4965 if (!only_parity(rdev->raid_disk, 4966 conf->prev_algo, 4967 conf->previous_raid_disks, 4968 conf->max_degraded)) 4969 continue; 4970 dirty_parity_disks++; 4971 } 4972 4973 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks) 4974 - working_disks); 4975 4976 if (has_failed(conf)) { 4977 printk(KERN_ERR "md/raid:%s: not enough operational devices" 4978 " (%d/%d failed)\n", 4979 mdname(mddev), mddev->degraded, conf->raid_disks); 4980 goto abort; 4981 } 4982 4983 /* device size must be a multiple of chunk size */ 4984 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1); 4985 mddev->resync_max_sectors = mddev->dev_sectors; 4986 4987 if (mddev->degraded > dirty_parity_disks && 4988 mddev->recovery_cp != MaxSector) { 4989 if (mddev->ok_start_degraded) 4990 printk(KERN_WARNING 4991 "md/raid:%s: starting dirty degraded array" 4992 " - data corruption possible.\n", 4993 mdname(mddev)); 4994 else { 4995 printk(KERN_ERR 4996 "md/raid:%s: cannot start dirty degraded array.\n", 4997 mdname(mddev)); 4998 goto abort; 4999 } 5000 } 5001 5002 if (mddev->degraded == 0) 5003 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d" 5004 " devices, algorithm %d\n", mdname(mddev), conf->level, 5005 mddev->raid_disks-mddev->degraded, mddev->raid_disks, 5006 mddev->new_layout); 5007 else 5008 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d" 5009 " out of %d devices, algorithm %d\n", 5010 mdname(mddev), conf->level, 5011 mddev->raid_disks - mddev->degraded, 5012 mddev->raid_disks, mddev->new_layout); 5013 5014 print_raid5_conf(conf); 5015 5016 if (conf->reshape_progress != MaxSector) { 5017 conf->reshape_safe = conf->reshape_progress; 5018 atomic_set(&conf->reshape_stripes, 0); 5019 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery); 5020 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery); 5021 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery); 5022 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery); 5023 mddev->sync_thread = md_register_thread(md_do_sync, mddev, 5024 "reshape"); 5025 } 5026 5027 5028 /* Ok, everything is just fine now */ 5029 if (mddev->to_remove == &raid5_attrs_group) 5030 mddev->to_remove = NULL; 5031 else if (mddev->kobj.sd && 5032 sysfs_create_group(&mddev->kobj, &raid5_attrs_group)) 5033 printk(KERN_WARNING 5034 "raid5: failed to create sysfs attributes for %s\n", 5035 mdname(mddev)); 5036 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0)); 5037 5038 if (mddev->queue) { 5039 int chunk_size; 5040 /* read-ahead size must cover two whole stripes, which 5041 * is 2 * (datadisks) * chunksize where 'n' is the 5042 * number of raid devices 5043 */ 5044 int data_disks = conf->previous_raid_disks - conf->max_degraded; 5045 int stripe = data_disks * 5046 ((mddev->chunk_sectors << 9) / PAGE_SIZE); 5047 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe) 5048 mddev->queue->backing_dev_info.ra_pages = 2 * stripe; 5049 5050 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec); 5051 5052 mddev->queue->backing_dev_info.congested_data = mddev; 5053 mddev->queue->backing_dev_info.congested_fn = raid5_congested; 5054 5055 chunk_size = mddev->chunk_sectors << 9; 5056 blk_queue_io_min(mddev->queue, chunk_size); 5057 blk_queue_io_opt(mddev->queue, chunk_size * 5058 (conf->raid_disks - conf->max_degraded)); 5059 5060 list_for_each_entry(rdev, &mddev->disks, same_set) 5061 disk_stack_limits(mddev->gendisk, rdev->bdev, 5062 rdev->data_offset << 9); 5063 } 5064 5065 return 0; 5066abort: 5067 md_unregister_thread(mddev->thread); 5068 mddev->thread = NULL; 5069 if (conf) { 5070 print_raid5_conf(conf); 5071 free_conf(conf); 5072 } 5073 mddev->private = NULL; 5074 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev)); 5075 return -EIO; 5076} 5077 5078static int stop(mddev_t *mddev) 5079{ 5080 raid5_conf_t *conf = mddev->private; 5081 5082 md_unregister_thread(mddev->thread); 5083 mddev->thread = NULL; 5084 if (mddev->queue) 5085 mddev->queue->backing_dev_info.congested_fn = NULL; 5086 free_conf(conf); 5087 mddev->private = NULL; 5088 mddev->to_remove = &raid5_attrs_group; 5089 return 0; 5090} 5091 5092#ifdef DEBUG 5093static void print_sh(struct seq_file *seq, struct stripe_head *sh) 5094{ 5095 int i; 5096 5097 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n", 5098 (unsigned long long)sh->sector, sh->pd_idx, sh->state); 5099 seq_printf(seq, "sh %llu, count %d.\n", 5100 (unsigned long long)sh->sector, atomic_read(&sh->count)); 5101 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector); 5102 for (i = 0; i < sh->disks; i++) { 5103 seq_printf(seq, "(cache%d: %p %ld) ", 5104 i, sh->dev[i].page, sh->dev[i].flags); 5105 } 5106 seq_printf(seq, "\n"); 5107} 5108 5109static void printall(struct seq_file *seq, raid5_conf_t *conf) 5110{ 5111 struct stripe_head *sh; 5112 struct hlist_node *hn; 5113 int i; 5114 5115 spin_lock_irq(&conf->device_lock); 5116 for (i = 0; i < NR_HASH; i++) { 5117 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) { 5118 if (sh->raid_conf != conf) 5119 continue; 5120 print_sh(seq, sh); 5121 } 5122 } 5123 spin_unlock_irq(&conf->device_lock); 5124} 5125#endif 5126 5127static void status(struct seq_file *seq, mddev_t *mddev) 5128{ 5129 raid5_conf_t *conf = mddev->private; 5130 int i; 5131 5132 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level, 5133 mddev->chunk_sectors / 2, mddev->layout); 5134 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded); 5135 for (i = 0; i < conf->raid_disks; i++) 5136 seq_printf (seq, "%s", 5137 conf->disks[i].rdev && 5138 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_"); 5139 seq_printf (seq, "]"); 5140#ifdef DEBUG 5141 seq_printf (seq, "\n"); 5142 printall(seq, conf); 5143#endif 5144} 5145 5146static void print_raid5_conf (raid5_conf_t *conf) 5147{ 5148 int i; 5149 struct disk_info *tmp; 5150 5151 printk(KERN_DEBUG "RAID conf printout:\n"); 5152 if (!conf) { 5153 printk("(conf==NULL)\n"); 5154 return; 5155 } 5156 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level, 5157 conf->raid_disks, 5158 conf->raid_disks - conf->mddev->degraded); 5159 5160 for (i = 0; i < conf->raid_disks; i++) { 5161 char b[BDEVNAME_SIZE]; 5162 tmp = conf->disks + i; 5163 if (tmp->rdev) 5164 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n", 5165 i, !test_bit(Faulty, &tmp->rdev->flags), 5166 bdevname(tmp->rdev->bdev, b)); 5167 } 5168} 5169 5170static int raid5_spare_active(mddev_t *mddev) 5171{ 5172 int i; 5173 raid5_conf_t *conf = mddev->private; 5174 struct disk_info *tmp; 5175 int count = 0; 5176 unsigned long flags; 5177 5178 for (i = 0; i < conf->raid_disks; i++) { 5179 tmp = conf->disks + i; 5180 if (tmp->rdev 5181 && tmp->rdev->recovery_offset == MaxSector 5182 && !test_bit(Faulty, &tmp->rdev->flags) 5183 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) { 5184 count++; 5185 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state); 5186 } 5187 } 5188 spin_lock_irqsave(&conf->device_lock, flags); 5189 mddev->degraded -= count; 5190 spin_unlock_irqrestore(&conf->device_lock, flags); 5191 print_raid5_conf(conf); 5192 return count; 5193} 5194 5195static int raid5_remove_disk(mddev_t *mddev, int number) 5196{ 5197 raid5_conf_t *conf = mddev->private; 5198 int err = 0; 5199 mdk_rdev_t *rdev; 5200 struct disk_info *p = conf->disks + number; 5201 5202 print_raid5_conf(conf); 5203 rdev = p->rdev; 5204 if (rdev) { 5205 if (number >= conf->raid_disks && 5206 conf->reshape_progress == MaxSector) 5207 clear_bit(In_sync, &rdev->flags); 5208 5209 if (test_bit(In_sync, &rdev->flags) || 5210 atomic_read(&rdev->nr_pending)) { 5211 err = -EBUSY; 5212 goto abort; 5213 } 5214 /* Only remove non-faulty devices if recovery 5215 * isn't possible. 5216 */ 5217 if (!test_bit(Faulty, &rdev->flags) && 5218 !has_failed(conf) && 5219 number < conf->raid_disks) { 5220 err = -EBUSY; 5221 goto abort; 5222 } 5223 p->rdev = NULL; 5224 synchronize_rcu(); 5225 if (atomic_read(&rdev->nr_pending)) { 5226 /* lost the race, try later */ 5227 err = -EBUSY; 5228 p->rdev = rdev; 5229 } 5230 } 5231abort: 5232 5233 print_raid5_conf(conf); 5234 return err; 5235} 5236 5237static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev) 5238{ 5239 raid5_conf_t *conf = mddev->private; 5240 int err = -EEXIST; 5241 int disk; 5242 struct disk_info *p; 5243 int first = 0; 5244 int last = conf->raid_disks - 1; 5245 5246 if (has_failed(conf)) 5247 /* no point adding a device */ 5248 return -EINVAL; 5249 5250 if (rdev->raid_disk >= 0) 5251 first = last = rdev->raid_disk; 5252 5253 /* 5254 * find the disk ... but prefer rdev->saved_raid_disk 5255 * if possible. 5256 */ 5257 if (rdev->saved_raid_disk >= 0 && 5258 rdev->saved_raid_disk >= first && 5259 conf->disks[rdev->saved_raid_disk].rdev == NULL) 5260 disk = rdev->saved_raid_disk; 5261 else 5262 disk = first; 5263 for ( ; disk <= last ; disk++) 5264 if ((p=conf->disks + disk)->rdev == NULL) { 5265 clear_bit(In_sync, &rdev->flags); 5266 rdev->raid_disk = disk; 5267 err = 0; 5268 if (rdev->saved_raid_disk != disk) 5269 conf->fullsync = 1; 5270 rcu_assign_pointer(p->rdev, rdev); 5271 break; 5272 } 5273 print_raid5_conf(conf); 5274 return err; 5275} 5276 5277static int raid5_resize(mddev_t *mddev, sector_t sectors) 5278{ 5279 /* no resync is happening, and there is enough space 5280 * on all devices, so we can resize. 5281 * We need to make sure resync covers any new space. 5282 * If the array is shrinking we should possibly wait until 5283 * any io in the removed space completes, but it hardly seems 5284 * worth it. 5285 */ 5286 sectors &= ~((sector_t)mddev->chunk_sectors - 1); 5287 md_set_array_sectors(mddev, raid5_size(mddev, sectors, 5288 mddev->raid_disks)); 5289 if (mddev->array_sectors > 5290 raid5_size(mddev, sectors, mddev->raid_disks)) 5291 return -EINVAL; 5292 set_capacity(mddev->gendisk, mddev->array_sectors); 5293 revalidate_disk(mddev->gendisk); 5294 if (sectors > mddev->dev_sectors && 5295 mddev->recovery_cp > mddev->dev_sectors) { 5296 mddev->recovery_cp = mddev->dev_sectors; 5297 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 5298 } 5299 mddev->dev_sectors = sectors; 5300 mddev->resync_max_sectors = sectors; 5301 return 0; 5302} 5303 5304static int check_stripe_cache(mddev_t *mddev) 5305{ 5306 /* Can only proceed if there are plenty of stripe_heads. 5307 * We need a minimum of one full stripe,, and for sensible progress 5308 * it is best to have about 4 times that. 5309 * If we require 4 times, then the default 256 4K stripe_heads will 5310 * allow for chunk sizes up to 256K, which is probably OK. 5311 * If the chunk size is greater, user-space should request more 5312 * stripe_heads first. 5313 */ 5314 raid5_conf_t *conf = mddev->private; 5315 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4 5316 > conf->max_nr_stripes || 5317 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4 5318 > conf->max_nr_stripes) { 5319 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n", 5320 mdname(mddev), 5321 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9) 5322 / STRIPE_SIZE)*4); 5323 return 0; 5324 } 5325 return 1; 5326} 5327 5328static int check_reshape(mddev_t *mddev) 5329{ 5330 raid5_conf_t *conf = mddev->private; 5331 5332 if (mddev->delta_disks == 0 && 5333 mddev->new_layout == mddev->layout && 5334 mddev->new_chunk_sectors == mddev->chunk_sectors) 5335 return 0; /* nothing to do */ 5336 if (mddev->bitmap) 5337 /* Cannot grow a bitmap yet */ 5338 return -EBUSY; 5339 if (has_failed(conf)) 5340 return -EINVAL; 5341 if (mddev->delta_disks < 0) { 5342 /* We might be able to shrink, but the devices must 5343 * be made bigger first. 5344 * For raid6, 4 is the minimum size. 5345 * Otherwise 2 is the minimum 5346 */ 5347 int min = 2; 5348 if (mddev->level == 6) 5349 min = 4; 5350 if (mddev->raid_disks + mddev->delta_disks < min) 5351 return -EINVAL; 5352 } 5353 5354 if (!check_stripe_cache(mddev)) 5355 return -ENOSPC; 5356 5357 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks); 5358} 5359 5360static int raid5_start_reshape(mddev_t *mddev) 5361{ 5362 raid5_conf_t *conf = mddev->private; 5363 mdk_rdev_t *rdev; 5364 int spares = 0; 5365 unsigned long flags; 5366 5367 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery)) 5368 return -EBUSY; 5369 5370 if (!check_stripe_cache(mddev)) 5371 return -ENOSPC; 5372 5373 list_for_each_entry(rdev, &mddev->disks, same_set) 5374 if (!test_bit(In_sync, &rdev->flags) 5375 && !test_bit(Faulty, &rdev->flags)) 5376 spares++; 5377 5378 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded) 5379 /* Not enough devices even to make a degraded array 5380 * of that size 5381 */ 5382 return -EINVAL; 5383 5384 /* Refuse to reduce size of the array. Any reductions in 5385 * array size must be through explicit setting of array_size 5386 * attribute. 5387 */ 5388 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks) 5389 < mddev->array_sectors) { 5390 printk(KERN_ERR "md/raid:%s: array size must be reduced " 5391 "before number of disks\n", mdname(mddev)); 5392 return -EINVAL; 5393 } 5394 5395 atomic_set(&conf->reshape_stripes, 0); 5396 spin_lock_irq(&conf->device_lock); 5397 conf->previous_raid_disks = conf->raid_disks; 5398 conf->raid_disks += mddev->delta_disks; 5399 conf->prev_chunk_sectors = conf->chunk_sectors; 5400 conf->chunk_sectors = mddev->new_chunk_sectors; 5401 conf->prev_algo = conf->algorithm; 5402 conf->algorithm = mddev->new_layout; 5403 if (mddev->delta_disks < 0) 5404 conf->reshape_progress = raid5_size(mddev, 0, 0); 5405 else 5406 conf->reshape_progress = 0; 5407 conf->reshape_safe = conf->reshape_progress; 5408 conf->generation++; 5409 spin_unlock_irq(&conf->device_lock); 5410 5411 /* Add some new drives, as many as will fit. 5412 * We know there are enough to make the newly sized array work. 5413 * Don't add devices if we are reducing the number of 5414 * devices in the array. This is because it is not possible 5415 * to correctly record the "partially reconstructed" state of 5416 * such devices during the reshape and confusion could result. 5417 */ 5418 if (mddev->delta_disks >= 0) { 5419 int added_devices = 0; 5420 list_for_each_entry(rdev, &mddev->disks, same_set) 5421 if (rdev->raid_disk < 0 && 5422 !test_bit(Faulty, &rdev->flags)) { 5423 if (raid5_add_disk(mddev, rdev) == 0) { 5424 char nm[20]; 5425 if (rdev->raid_disk 5426 >= conf->previous_raid_disks) { 5427 set_bit(In_sync, &rdev->flags); 5428 added_devices++; 5429 } else 5430 rdev->recovery_offset = 0; 5431 sprintf(nm, "rd%d", rdev->raid_disk); 5432 if (sysfs_create_link(&mddev->kobj, 5433 &rdev->kobj, nm)) 5434 /* Failure here is OK */; 5435 } 5436 } else if (rdev->raid_disk >= conf->previous_raid_disks 5437 && !test_bit(Faulty, &rdev->flags)) { 5438 /* This is a spare that was manually added */ 5439 set_bit(In_sync, &rdev->flags); 5440 added_devices++; 5441 } 5442 5443 /* When a reshape changes the number of devices, 5444 * ->degraded is measured against the larger of the 5445 * pre and post number of devices. 5446 */ 5447 spin_lock_irqsave(&conf->device_lock, flags); 5448 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks) 5449 - added_devices; 5450 spin_unlock_irqrestore(&conf->device_lock, flags); 5451 } 5452 mddev->raid_disks = conf->raid_disks; 5453 mddev->reshape_position = conf->reshape_progress; 5454 set_bit(MD_CHANGE_DEVS, &mddev->flags); 5455 5456 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery); 5457 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery); 5458 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery); 5459 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery); 5460 mddev->sync_thread = md_register_thread(md_do_sync, mddev, 5461 "reshape"); 5462 if (!mddev->sync_thread) { 5463 mddev->recovery = 0; 5464 spin_lock_irq(&conf->device_lock); 5465 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks; 5466 conf->reshape_progress = MaxSector; 5467 spin_unlock_irq(&conf->device_lock); 5468 return -EAGAIN; 5469 } 5470 conf->reshape_checkpoint = jiffies; 5471 md_wakeup_thread(mddev->sync_thread); 5472 md_new_event(mddev); 5473 return 0; 5474} 5475 5476/* This is called from the reshape thread and should make any 5477 * changes needed in 'conf' 5478 */ 5479static void end_reshape(raid5_conf_t *conf) 5480{ 5481 5482 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) { 5483 5484 spin_lock_irq(&conf->device_lock); 5485 conf->previous_raid_disks = conf->raid_disks; 5486 conf->reshape_progress = MaxSector; 5487 spin_unlock_irq(&conf->device_lock); 5488 wake_up(&conf->wait_for_overlap); 5489 5490 /* read-ahead size must cover two whole stripes, which is 5491 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices 5492 */ 5493 if (conf->mddev->queue) { 5494 int data_disks = conf->raid_disks - conf->max_degraded; 5495 int stripe = data_disks * ((conf->chunk_sectors << 9) 5496 / PAGE_SIZE); 5497 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe) 5498 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe; 5499 } 5500 } 5501} 5502 5503/* This is called from the raid5d thread with mddev_lock held. 5504 * It makes config changes to the device. 5505 */ 5506static void raid5_finish_reshape(mddev_t *mddev) 5507{ 5508 raid5_conf_t *conf = mddev->private; 5509 5510 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) { 5511 5512 if (mddev->delta_disks > 0) { 5513 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0)); 5514 set_capacity(mddev->gendisk, mddev->array_sectors); 5515 revalidate_disk(mddev->gendisk); 5516 } else { 5517 int d; 5518 mddev->degraded = conf->raid_disks; 5519 for (d = 0; d < conf->raid_disks ; d++) 5520 if (conf->disks[d].rdev && 5521 test_bit(In_sync, 5522 &conf->disks[d].rdev->flags)) 5523 mddev->degraded--; 5524 for (d = conf->raid_disks ; 5525 d < conf->raid_disks - mddev->delta_disks; 5526 d++) { 5527 mdk_rdev_t *rdev = conf->disks[d].rdev; 5528 if (rdev && raid5_remove_disk(mddev, d) == 0) { 5529 char nm[20]; 5530 sprintf(nm, "rd%d", rdev->raid_disk); 5531 sysfs_remove_link(&mddev->kobj, nm); 5532 rdev->raid_disk = -1; 5533 } 5534 } 5535 } 5536 mddev->layout = conf->algorithm; 5537 mddev->chunk_sectors = conf->chunk_sectors; 5538 mddev->reshape_position = MaxSector; 5539 mddev->delta_disks = 0; 5540 } 5541} 5542 5543static void raid5_quiesce(mddev_t *mddev, int state) 5544{ 5545 raid5_conf_t *conf = mddev->private; 5546 5547 switch(state) { 5548 case 2: /* resume for a suspend */ 5549 wake_up(&conf->wait_for_overlap); 5550 break; 5551 5552 case 1: /* stop all writes */ 5553 spin_lock_irq(&conf->device_lock); 5554 /* '2' tells resync/reshape to pause so that all 5555 * active stripes can drain 5556 */ 5557 conf->quiesce = 2; 5558 wait_event_lock_irq(conf->wait_for_stripe, 5559 atomic_read(&conf->active_stripes) == 0 && 5560 atomic_read(&conf->active_aligned_reads) == 0, 5561 conf->device_lock, /* nothing */); 5562 conf->quiesce = 1; 5563 spin_unlock_irq(&conf->device_lock); 5564 /* allow reshape to continue */ 5565 wake_up(&conf->wait_for_overlap); 5566 break; 5567 5568 case 0: /* re-enable writes */ 5569 spin_lock_irq(&conf->device_lock); 5570 conf->quiesce = 0; 5571 wake_up(&conf->wait_for_stripe); 5572 wake_up(&conf->wait_for_overlap); 5573 spin_unlock_irq(&conf->device_lock); 5574 break; 5575 } 5576} 5577 5578 5579static void *raid45_takeover_raid0(mddev_t *mddev, int level) 5580{ 5581 struct raid0_private_data *raid0_priv = mddev->private; 5582 sector_t sectors; 5583 5584 /* for raid0 takeover only one zone is supported */ 5585 if (raid0_priv->nr_strip_zones > 1) { 5586 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n", 5587 mdname(mddev)); 5588 return ERR_PTR(-EINVAL); 5589 } 5590 5591 sectors = raid0_priv->strip_zone[0].zone_end; 5592 sector_div(sectors, raid0_priv->strip_zone[0].nb_dev); 5593 mddev->dev_sectors = sectors; 5594 mddev->new_level = level; 5595 mddev->new_layout = ALGORITHM_PARITY_N; 5596 mddev->new_chunk_sectors = mddev->chunk_sectors; 5597 mddev->raid_disks += 1; 5598 mddev->delta_disks = 1; 5599 /* make sure it will be not marked as dirty */ 5600 mddev->recovery_cp = MaxSector; 5601 5602 return setup_conf(mddev); 5603} 5604 5605 5606static void *raid5_takeover_raid1(mddev_t *mddev) 5607{ 5608 int chunksect; 5609 5610 if (mddev->raid_disks != 2 || 5611 mddev->degraded > 1) 5612 return ERR_PTR(-EINVAL); 5613 5614 /* Should check if there are write-behind devices? */ 5615 5616 chunksect = 64*2; /* 64K by default */ 5617 5618 /* The array must be an exact multiple of chunksize */ 5619 while (chunksect && (mddev->array_sectors & (chunksect-1))) 5620 chunksect >>= 1; 5621 5622 if ((chunksect<<9) < STRIPE_SIZE) 5623 /* array size does not allow a suitable chunk size */ 5624 return ERR_PTR(-EINVAL); 5625 5626 mddev->new_level = 5; 5627 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC; 5628 mddev->new_chunk_sectors = chunksect; 5629 5630 return setup_conf(mddev); 5631} 5632 5633static void *raid5_takeover_raid6(mddev_t *mddev) 5634{ 5635 int new_layout; 5636 5637 switch (mddev->layout) { 5638 case ALGORITHM_LEFT_ASYMMETRIC_6: 5639 new_layout = ALGORITHM_LEFT_ASYMMETRIC; 5640 break; 5641 case ALGORITHM_RIGHT_ASYMMETRIC_6: 5642 new_layout = ALGORITHM_RIGHT_ASYMMETRIC; 5643 break; 5644 case ALGORITHM_LEFT_SYMMETRIC_6: 5645 new_layout = ALGORITHM_LEFT_SYMMETRIC; 5646 break; 5647 case ALGORITHM_RIGHT_SYMMETRIC_6: 5648 new_layout = ALGORITHM_RIGHT_SYMMETRIC; 5649 break; 5650 case ALGORITHM_PARITY_0_6: 5651 new_layout = ALGORITHM_PARITY_0; 5652 break; 5653 case ALGORITHM_PARITY_N: 5654 new_layout = ALGORITHM_PARITY_N; 5655 break; 5656 default: 5657 return ERR_PTR(-EINVAL); 5658 } 5659 mddev->new_level = 5; 5660 mddev->new_layout = new_layout; 5661 mddev->delta_disks = -1; 5662 mddev->raid_disks -= 1; 5663 return setup_conf(mddev); 5664} 5665 5666 5667static int raid5_check_reshape(mddev_t *mddev) 5668{ 5669 /* For a 2-drive array, the layout and chunk size can be changed 5670 * immediately as not restriping is needed. 5671 * For larger arrays we record the new value - after validation 5672 * to be used by a reshape pass. 5673 */ 5674 raid5_conf_t *conf = mddev->private; 5675 int new_chunk = mddev->new_chunk_sectors; 5676 5677 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout)) 5678 return -EINVAL; 5679 if (new_chunk > 0) { 5680 if (!is_power_of_2(new_chunk)) 5681 return -EINVAL; 5682 if (new_chunk < (PAGE_SIZE>>9)) 5683 return -EINVAL; 5684 if (mddev->array_sectors & (new_chunk-1)) 5685 /* not factor of array size */ 5686 return -EINVAL; 5687 } 5688 5689 /* They look valid */ 5690 5691 if (mddev->raid_disks == 2) { 5692 /* can make the change immediately */ 5693 if (mddev->new_layout >= 0) { 5694 conf->algorithm = mddev->new_layout; 5695 mddev->layout = mddev->new_layout; 5696 } 5697 if (new_chunk > 0) { 5698 conf->chunk_sectors = new_chunk ; 5699 mddev->chunk_sectors = new_chunk; 5700 } 5701 set_bit(MD_CHANGE_DEVS, &mddev->flags); 5702 md_wakeup_thread(mddev->thread); 5703 } 5704 return check_reshape(mddev); 5705} 5706 5707static int raid6_check_reshape(mddev_t *mddev) 5708{ 5709 int new_chunk = mddev->new_chunk_sectors; 5710 5711 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout)) 5712 return -EINVAL; 5713 if (new_chunk > 0) { 5714 if (!is_power_of_2(new_chunk)) 5715 return -EINVAL; 5716 if (new_chunk < (PAGE_SIZE >> 9)) 5717 return -EINVAL; 5718 if (mddev->array_sectors & (new_chunk-1)) 5719 /* not factor of array size */ 5720 return -EINVAL; 5721 } 5722 5723 /* They look valid */ 5724 return check_reshape(mddev); 5725} 5726 5727static void *raid5_takeover(mddev_t *mddev) 5728{ 5729 /* raid5 can take over: 5730 * raid0 - if there is only one strip zone - make it a raid4 layout 5731 * raid1 - if there are two drives. We need to know the chunk size 5732 * raid4 - trivial - just use a raid4 layout. 5733 * raid6 - Providing it is a *_6 layout 5734 */ 5735 if (mddev->level == 0) 5736 return raid45_takeover_raid0(mddev, 5); 5737 if (mddev->level == 1) 5738 return raid5_takeover_raid1(mddev); 5739 if (mddev->level == 4) { 5740 mddev->new_layout = ALGORITHM_PARITY_N; 5741 mddev->new_level = 5; 5742 return setup_conf(mddev); 5743 } 5744 if (mddev->level == 6) 5745 return raid5_takeover_raid6(mddev); 5746 5747 return ERR_PTR(-EINVAL); 5748} 5749 5750static void *raid4_takeover(mddev_t *mddev) 5751{ 5752 /* raid4 can take over: 5753 * raid0 - if there is only one strip zone 5754 * raid5 - if layout is right 5755 */ 5756 if (mddev->level == 0) 5757 return raid45_takeover_raid0(mddev, 4); 5758 if (mddev->level == 5 && 5759 mddev->layout == ALGORITHM_PARITY_N) { 5760 mddev->new_layout = 0; 5761 mddev->new_level = 4; 5762 return setup_conf(mddev); 5763 } 5764 return ERR_PTR(-EINVAL); 5765} 5766 5767static struct mdk_personality raid5_personality; 5768 5769static void *raid6_takeover(mddev_t *mddev) 5770{ 5771 /* Currently can only take over a raid5. We map the 5772 * personality to an equivalent raid6 personality 5773 * with the Q block at the end. 5774 */ 5775 int new_layout; 5776 5777 if (mddev->pers != &raid5_personality) 5778 return ERR_PTR(-EINVAL); 5779 if (mddev->degraded > 1) 5780 return ERR_PTR(-EINVAL); 5781 if (mddev->raid_disks > 253) 5782 return ERR_PTR(-EINVAL); 5783 if (mddev->raid_disks < 3) 5784 return ERR_PTR(-EINVAL); 5785 5786 switch (mddev->layout) { 5787 case ALGORITHM_LEFT_ASYMMETRIC: 5788 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6; 5789 break; 5790 case ALGORITHM_RIGHT_ASYMMETRIC: 5791 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6; 5792 break; 5793 case ALGORITHM_LEFT_SYMMETRIC: 5794 new_layout = ALGORITHM_LEFT_SYMMETRIC_6; 5795 break; 5796 case ALGORITHM_RIGHT_SYMMETRIC: 5797 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6; 5798 break; 5799 case ALGORITHM_PARITY_0: 5800 new_layout = ALGORITHM_PARITY_0_6; 5801 break; 5802 case ALGORITHM_PARITY_N: 5803 new_layout = ALGORITHM_PARITY_N; 5804 break; 5805 default: 5806 return ERR_PTR(-EINVAL); 5807 } 5808 mddev->new_level = 6; 5809 mddev->new_layout = new_layout; 5810 mddev->delta_disks = 1; 5811 mddev->raid_disks += 1; 5812 return setup_conf(mddev); 5813} 5814 5815 5816static struct mdk_personality raid6_personality = 5817{ 5818 .name = "raid6", 5819 .level = 6, 5820 .owner = THIS_MODULE, 5821 .make_request = make_request, 5822 .run = run, 5823 .stop = stop, 5824 .status = status, 5825 .error_handler = error, 5826 .hot_add_disk = raid5_add_disk, 5827 .hot_remove_disk= raid5_remove_disk, 5828 .spare_active = raid5_spare_active, 5829 .sync_request = sync_request, 5830 .resize = raid5_resize, 5831 .size = raid5_size, 5832 .check_reshape = raid6_check_reshape, 5833 .start_reshape = raid5_start_reshape, 5834 .finish_reshape = raid5_finish_reshape, 5835 .quiesce = raid5_quiesce, 5836 .takeover = raid6_takeover, 5837}; 5838static struct mdk_personality raid5_personality = 5839{ 5840 .name = "raid5", 5841 .level = 5, 5842 .owner = THIS_MODULE, 5843 .make_request = make_request, 5844 .run = run, 5845 .stop = stop, 5846 .status = status, 5847 .error_handler = error, 5848 .hot_add_disk = raid5_add_disk, 5849 .hot_remove_disk= raid5_remove_disk, 5850 .spare_active = raid5_spare_active, 5851 .sync_request = sync_request, 5852 .resize = raid5_resize, 5853 .size = raid5_size, 5854 .check_reshape = raid5_check_reshape, 5855 .start_reshape = raid5_start_reshape, 5856 .finish_reshape = raid5_finish_reshape, 5857 .quiesce = raid5_quiesce, 5858 .takeover = raid5_takeover, 5859}; 5860 5861static struct mdk_personality raid4_personality = 5862{ 5863 .name = "raid4", 5864 .level = 4, 5865 .owner = THIS_MODULE, 5866 .make_request = make_request, 5867 .run = run, 5868 .stop = stop, 5869 .status = status, 5870 .error_handler = error, 5871 .hot_add_disk = raid5_add_disk, 5872 .hot_remove_disk= raid5_remove_disk, 5873 .spare_active = raid5_spare_active, 5874 .sync_request = sync_request, 5875 .resize = raid5_resize, 5876 .size = raid5_size, 5877 .check_reshape = raid5_check_reshape, 5878 .start_reshape = raid5_start_reshape, 5879 .finish_reshape = raid5_finish_reshape, 5880 .quiesce = raid5_quiesce, 5881 .takeover = raid4_takeover, 5882}; 5883 5884static int __init raid5_init(void) 5885{ 5886 register_md_personality(&raid6_personality); 5887 register_md_personality(&raid5_personality); 5888 register_md_personality(&raid4_personality); 5889 return 0; 5890} 5891 5892static void raid5_exit(void) 5893{ 5894 unregister_md_personality(&raid6_personality); 5895 unregister_md_personality(&raid5_personality); 5896 unregister_md_personality(&raid4_personality); 5897} 5898 5899module_init(raid5_init); 5900module_exit(raid5_exit); 5901MODULE_LICENSE("GPL"); 5902MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD"); 5903MODULE_ALIAS("md-personality-4"); /* RAID5 */ 5904MODULE_ALIAS("md-raid5"); 5905MODULE_ALIAS("md-raid4"); 5906MODULE_ALIAS("md-level-5"); 5907MODULE_ALIAS("md-level-4"); 5908MODULE_ALIAS("md-personality-8"); /* RAID6 */ 5909MODULE_ALIAS("md-raid6"); 5910MODULE_ALIAS("md-level-6"); 5911 5912/* This used to be two separate modules, they were: */ 5913MODULE_ALIAS("raid5"); 5914MODULE_ALIAS("raid6"); 5915