raid5.c revision d089c6af10c2be5988f03667d6d22fe6085fbe5e
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->bm_write is the number of the last batch successfully written. 31 * conf->bm_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 bm_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/module.h> 47#include <linux/slab.h> 48#include <linux/highmem.h> 49#include <linux/bitops.h> 50#include <linux/kthread.h> 51#include <asm/atomic.h> 52#include "raid6.h" 53 54#include <linux/raid/bitmap.h> 55#include <linux/async_tx.h> 56 57/* 58 * Stripe cache 59 */ 60 61#define NR_STRIPES 256 62#define STRIPE_SIZE PAGE_SIZE 63#define STRIPE_SHIFT (PAGE_SHIFT - 9) 64#define STRIPE_SECTORS (STRIPE_SIZE>>9) 65#define IO_THRESHOLD 1 66#define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head)) 67#define HASH_MASK (NR_HASH - 1) 68 69#define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK])) 70 71/* bio's attached to a stripe+device for I/O are linked together in bi_sector 72 * order without overlap. There may be several bio's per stripe+device, and 73 * a bio could span several devices. 74 * When walking this list for a particular stripe+device, we must never proceed 75 * beyond a bio that extends past this device, as the next bio might no longer 76 * be valid. 77 * This macro is used to determine the 'next' bio in the list, given the sector 78 * of the current stripe+device 79 */ 80#define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL) 81/* 82 * The following can be used to debug the driver 83 */ 84#define RAID5_PARANOIA 1 85#if RAID5_PARANOIA && defined(CONFIG_SMP) 86# define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock) 87#else 88# define CHECK_DEVLOCK() 89#endif 90 91#ifdef DEBUG 92#define inline 93#define __inline__ 94#endif 95 96#if !RAID6_USE_EMPTY_ZERO_PAGE 97/* In .bss so it's zeroed */ 98const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256))); 99#endif 100 101static inline int raid6_next_disk(int disk, int raid_disks) 102{ 103 disk++; 104 return (disk < raid_disks) ? disk : 0; 105} 106 107static void return_io(struct bio *return_bi) 108{ 109 struct bio *bi = return_bi; 110 while (bi) { 111 112 return_bi = bi->bi_next; 113 bi->bi_next = NULL; 114 bi->bi_size = 0; 115 bi->bi_end_io(bi, 116 test_bit(BIO_UPTODATE, &bi->bi_flags) 117 ? 0 : -EIO); 118 bi = return_bi; 119 } 120} 121 122static void print_raid5_conf (raid5_conf_t *conf); 123 124static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh) 125{ 126 if (atomic_dec_and_test(&sh->count)) { 127 BUG_ON(!list_empty(&sh->lru)); 128 BUG_ON(atomic_read(&conf->active_stripes)==0); 129 if (test_bit(STRIPE_HANDLE, &sh->state)) { 130 if (test_bit(STRIPE_DELAYED, &sh->state)) { 131 list_add_tail(&sh->lru, &conf->delayed_list); 132 blk_plug_device(conf->mddev->queue); 133 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) && 134 sh->bm_seq - conf->seq_write > 0) { 135 list_add_tail(&sh->lru, &conf->bitmap_list); 136 blk_plug_device(conf->mddev->queue); 137 } else { 138 clear_bit(STRIPE_BIT_DELAY, &sh->state); 139 list_add_tail(&sh->lru, &conf->handle_list); 140 } 141 md_wakeup_thread(conf->mddev->thread); 142 } else { 143 BUG_ON(sh->ops.pending); 144 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 145 atomic_dec(&conf->preread_active_stripes); 146 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) 147 md_wakeup_thread(conf->mddev->thread); 148 } 149 atomic_dec(&conf->active_stripes); 150 if (!test_bit(STRIPE_EXPANDING, &sh->state)) { 151 list_add_tail(&sh->lru, &conf->inactive_list); 152 wake_up(&conf->wait_for_stripe); 153 if (conf->retry_read_aligned) 154 md_wakeup_thread(conf->mddev->thread); 155 } 156 } 157 } 158} 159static void release_stripe(struct stripe_head *sh) 160{ 161 raid5_conf_t *conf = sh->raid_conf; 162 unsigned long flags; 163 164 spin_lock_irqsave(&conf->device_lock, flags); 165 __release_stripe(conf, sh); 166 spin_unlock_irqrestore(&conf->device_lock, flags); 167} 168 169static inline void remove_hash(struct stripe_head *sh) 170{ 171 pr_debug("remove_hash(), stripe %llu\n", 172 (unsigned long long)sh->sector); 173 174 hlist_del_init(&sh->hash); 175} 176 177static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh) 178{ 179 struct hlist_head *hp = stripe_hash(conf, sh->sector); 180 181 pr_debug("insert_hash(), stripe %llu\n", 182 (unsigned long long)sh->sector); 183 184 CHECK_DEVLOCK(); 185 hlist_add_head(&sh->hash, hp); 186} 187 188 189/* find an idle stripe, make sure it is unhashed, and return it. */ 190static struct stripe_head *get_free_stripe(raid5_conf_t *conf) 191{ 192 struct stripe_head *sh = NULL; 193 struct list_head *first; 194 195 CHECK_DEVLOCK(); 196 if (list_empty(&conf->inactive_list)) 197 goto out; 198 first = conf->inactive_list.next; 199 sh = list_entry(first, struct stripe_head, lru); 200 list_del_init(first); 201 remove_hash(sh); 202 atomic_inc(&conf->active_stripes); 203out: 204 return sh; 205} 206 207static void shrink_buffers(struct stripe_head *sh, int num) 208{ 209 struct page *p; 210 int i; 211 212 for (i=0; i<num ; i++) { 213 p = sh->dev[i].page; 214 if (!p) 215 continue; 216 sh->dev[i].page = NULL; 217 put_page(p); 218 } 219} 220 221static int grow_buffers(struct stripe_head *sh, int num) 222{ 223 int i; 224 225 for (i=0; i<num; i++) { 226 struct page *page; 227 228 if (!(page = alloc_page(GFP_KERNEL))) { 229 return 1; 230 } 231 sh->dev[i].page = page; 232 } 233 return 0; 234} 235 236static void raid5_build_block (struct stripe_head *sh, int i); 237 238static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks) 239{ 240 raid5_conf_t *conf = sh->raid_conf; 241 int i; 242 243 BUG_ON(atomic_read(&sh->count) != 0); 244 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state)); 245 BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete); 246 247 CHECK_DEVLOCK(); 248 pr_debug("init_stripe called, stripe %llu\n", 249 (unsigned long long)sh->sector); 250 251 remove_hash(sh); 252 253 sh->sector = sector; 254 sh->pd_idx = pd_idx; 255 sh->state = 0; 256 257 sh->disks = disks; 258 259 for (i = sh->disks; i--; ) { 260 struct r5dev *dev = &sh->dev[i]; 261 262 if (dev->toread || dev->read || dev->towrite || dev->written || 263 test_bit(R5_LOCKED, &dev->flags)) { 264 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n", 265 (unsigned long long)sh->sector, i, dev->toread, 266 dev->read, dev->towrite, dev->written, 267 test_bit(R5_LOCKED, &dev->flags)); 268 BUG(); 269 } 270 dev->flags = 0; 271 raid5_build_block(sh, i); 272 } 273 insert_hash(conf, sh); 274} 275 276static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks) 277{ 278 struct stripe_head *sh; 279 struct hlist_node *hn; 280 281 CHECK_DEVLOCK(); 282 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector); 283 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash) 284 if (sh->sector == sector && sh->disks == disks) 285 return sh; 286 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector); 287 return NULL; 288} 289 290static void unplug_slaves(mddev_t *mddev); 291static void raid5_unplug_device(struct request_queue *q); 292 293static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks, 294 int pd_idx, int noblock) 295{ 296 struct stripe_head *sh; 297 298 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector); 299 300 spin_lock_irq(&conf->device_lock); 301 302 do { 303 wait_event_lock_irq(conf->wait_for_stripe, 304 conf->quiesce == 0, 305 conf->device_lock, /* nothing */); 306 sh = __find_stripe(conf, sector, disks); 307 if (!sh) { 308 if (!conf->inactive_blocked) 309 sh = get_free_stripe(conf); 310 if (noblock && sh == NULL) 311 break; 312 if (!sh) { 313 conf->inactive_blocked = 1; 314 wait_event_lock_irq(conf->wait_for_stripe, 315 !list_empty(&conf->inactive_list) && 316 (atomic_read(&conf->active_stripes) 317 < (conf->max_nr_stripes *3/4) 318 || !conf->inactive_blocked), 319 conf->device_lock, 320 raid5_unplug_device(conf->mddev->queue) 321 ); 322 conf->inactive_blocked = 0; 323 } else 324 init_stripe(sh, sector, pd_idx, disks); 325 } else { 326 if (atomic_read(&sh->count)) { 327 BUG_ON(!list_empty(&sh->lru)); 328 } else { 329 if (!test_bit(STRIPE_HANDLE, &sh->state)) 330 atomic_inc(&conf->active_stripes); 331 if (list_empty(&sh->lru) && 332 !test_bit(STRIPE_EXPANDING, &sh->state)) 333 BUG(); 334 list_del_init(&sh->lru); 335 } 336 } 337 } while (sh == NULL); 338 339 if (sh) 340 atomic_inc(&sh->count); 341 342 spin_unlock_irq(&conf->device_lock); 343 return sh; 344} 345 346/* test_and_ack_op() ensures that we only dequeue an operation once */ 347#define test_and_ack_op(op, pend) \ 348do { \ 349 if (test_bit(op, &sh->ops.pending) && \ 350 !test_bit(op, &sh->ops.complete)) { \ 351 if (test_and_set_bit(op, &sh->ops.ack)) \ 352 clear_bit(op, &pend); \ 353 else \ 354 ack++; \ 355 } else \ 356 clear_bit(op, &pend); \ 357} while (0) 358 359/* find new work to run, do not resubmit work that is already 360 * in flight 361 */ 362static unsigned long get_stripe_work(struct stripe_head *sh) 363{ 364 unsigned long pending; 365 int ack = 0; 366 367 pending = sh->ops.pending; 368 369 test_and_ack_op(STRIPE_OP_BIOFILL, pending); 370 test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending); 371 test_and_ack_op(STRIPE_OP_PREXOR, pending); 372 test_and_ack_op(STRIPE_OP_BIODRAIN, pending); 373 test_and_ack_op(STRIPE_OP_POSTXOR, pending); 374 test_and_ack_op(STRIPE_OP_CHECK, pending); 375 if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending)) 376 ack++; 377 378 sh->ops.count -= ack; 379 if (unlikely(sh->ops.count < 0)) { 380 printk(KERN_ERR "pending: %#lx ops.pending: %#lx ops.ack: %#lx " 381 "ops.complete: %#lx\n", pending, sh->ops.pending, 382 sh->ops.ack, sh->ops.complete); 383 BUG(); 384 } 385 386 return pending; 387} 388 389static void 390raid5_end_read_request(struct bio *bi, int error); 391static void 392raid5_end_write_request(struct bio *bi, int error); 393 394static void ops_run_io(struct stripe_head *sh) 395{ 396 raid5_conf_t *conf = sh->raid_conf; 397 int i, disks = sh->disks; 398 399 might_sleep(); 400 401 for (i = disks; i--; ) { 402 int rw; 403 struct bio *bi; 404 mdk_rdev_t *rdev; 405 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) 406 rw = WRITE; 407 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags)) 408 rw = READ; 409 else 410 continue; 411 412 bi = &sh->dev[i].req; 413 414 bi->bi_rw = rw; 415 if (rw == WRITE) 416 bi->bi_end_io = raid5_end_write_request; 417 else 418 bi->bi_end_io = raid5_end_read_request; 419 420 rcu_read_lock(); 421 rdev = rcu_dereference(conf->disks[i].rdev); 422 if (rdev && test_bit(Faulty, &rdev->flags)) 423 rdev = NULL; 424 if (rdev) 425 atomic_inc(&rdev->nr_pending); 426 rcu_read_unlock(); 427 428 if (rdev) { 429 if (test_bit(STRIPE_SYNCING, &sh->state) || 430 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) || 431 test_bit(STRIPE_EXPAND_READY, &sh->state)) 432 md_sync_acct(rdev->bdev, STRIPE_SECTORS); 433 434 bi->bi_bdev = rdev->bdev; 435 pr_debug("%s: for %llu schedule op %ld on disc %d\n", 436 __FUNCTION__, (unsigned long long)sh->sector, 437 bi->bi_rw, i); 438 atomic_inc(&sh->count); 439 bi->bi_sector = sh->sector + rdev->data_offset; 440 bi->bi_flags = 1 << BIO_UPTODATE; 441 bi->bi_vcnt = 1; 442 bi->bi_max_vecs = 1; 443 bi->bi_idx = 0; 444 bi->bi_io_vec = &sh->dev[i].vec; 445 bi->bi_io_vec[0].bv_len = STRIPE_SIZE; 446 bi->bi_io_vec[0].bv_offset = 0; 447 bi->bi_size = STRIPE_SIZE; 448 bi->bi_next = NULL; 449 if (rw == WRITE && 450 test_bit(R5_ReWrite, &sh->dev[i].flags)) 451 atomic_add(STRIPE_SECTORS, 452 &rdev->corrected_errors); 453 generic_make_request(bi); 454 } else { 455 if (rw == WRITE) 456 set_bit(STRIPE_DEGRADED, &sh->state); 457 pr_debug("skip op %ld on disc %d for sector %llu\n", 458 bi->bi_rw, i, (unsigned long long)sh->sector); 459 clear_bit(R5_LOCKED, &sh->dev[i].flags); 460 set_bit(STRIPE_HANDLE, &sh->state); 461 } 462 } 463} 464 465static struct dma_async_tx_descriptor * 466async_copy_data(int frombio, struct bio *bio, struct page *page, 467 sector_t sector, struct dma_async_tx_descriptor *tx) 468{ 469 struct bio_vec *bvl; 470 struct page *bio_page; 471 int i; 472 int page_offset; 473 474 if (bio->bi_sector >= sector) 475 page_offset = (signed)(bio->bi_sector - sector) * 512; 476 else 477 page_offset = (signed)(sector - bio->bi_sector) * -512; 478 bio_for_each_segment(bvl, bio, i) { 479 int len = bio_iovec_idx(bio, i)->bv_len; 480 int clen; 481 int b_offset = 0; 482 483 if (page_offset < 0) { 484 b_offset = -page_offset; 485 page_offset += b_offset; 486 len -= b_offset; 487 } 488 489 if (len > 0 && page_offset + len > STRIPE_SIZE) 490 clen = STRIPE_SIZE - page_offset; 491 else 492 clen = len; 493 494 if (clen > 0) { 495 b_offset += bio_iovec_idx(bio, i)->bv_offset; 496 bio_page = bio_iovec_idx(bio, i)->bv_page; 497 if (frombio) 498 tx = async_memcpy(page, bio_page, page_offset, 499 b_offset, clen, 500 ASYNC_TX_DEP_ACK, 501 tx, NULL, NULL); 502 else 503 tx = async_memcpy(bio_page, page, b_offset, 504 page_offset, clen, 505 ASYNC_TX_DEP_ACK, 506 tx, NULL, NULL); 507 } 508 if (clen < len) /* hit end of page */ 509 break; 510 page_offset += len; 511 } 512 513 return tx; 514} 515 516static void ops_complete_biofill(void *stripe_head_ref) 517{ 518 struct stripe_head *sh = stripe_head_ref; 519 struct bio *return_bi = NULL; 520 raid5_conf_t *conf = sh->raid_conf; 521 int i; 522 523 pr_debug("%s: stripe %llu\n", __FUNCTION__, 524 (unsigned long long)sh->sector); 525 526 /* clear completed biofills */ 527 for (i = sh->disks; i--; ) { 528 struct r5dev *dev = &sh->dev[i]; 529 530 /* acknowledge completion of a biofill operation */ 531 /* and check if we need to reply to a read request, 532 * new R5_Wantfill requests are held off until 533 * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending) 534 */ 535 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) { 536 struct bio *rbi, *rbi2; 537 538 /* The access to dev->read is outside of the 539 * spin_lock_irq(&conf->device_lock), but is protected 540 * by the STRIPE_OP_BIOFILL pending bit 541 */ 542 BUG_ON(!dev->read); 543 rbi = dev->read; 544 dev->read = NULL; 545 while (rbi && rbi->bi_sector < 546 dev->sector + STRIPE_SECTORS) { 547 rbi2 = r5_next_bio(rbi, dev->sector); 548 spin_lock_irq(&conf->device_lock); 549 if (--rbi->bi_phys_segments == 0) { 550 rbi->bi_next = return_bi; 551 return_bi = rbi; 552 } 553 spin_unlock_irq(&conf->device_lock); 554 rbi = rbi2; 555 } 556 } 557 } 558 set_bit(STRIPE_OP_BIOFILL, &sh->ops.complete); 559 560 return_io(return_bi); 561 562 set_bit(STRIPE_HANDLE, &sh->state); 563 release_stripe(sh); 564} 565 566static void ops_run_biofill(struct stripe_head *sh) 567{ 568 struct dma_async_tx_descriptor *tx = NULL; 569 raid5_conf_t *conf = sh->raid_conf; 570 int i; 571 572 pr_debug("%s: stripe %llu\n", __FUNCTION__, 573 (unsigned long long)sh->sector); 574 575 for (i = sh->disks; i--; ) { 576 struct r5dev *dev = &sh->dev[i]; 577 if (test_bit(R5_Wantfill, &dev->flags)) { 578 struct bio *rbi; 579 spin_lock_irq(&conf->device_lock); 580 dev->read = rbi = dev->toread; 581 dev->toread = NULL; 582 spin_unlock_irq(&conf->device_lock); 583 while (rbi && rbi->bi_sector < 584 dev->sector + STRIPE_SECTORS) { 585 tx = async_copy_data(0, rbi, dev->page, 586 dev->sector, tx); 587 rbi = r5_next_bio(rbi, dev->sector); 588 } 589 } 590 } 591 592 atomic_inc(&sh->count); 593 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx, 594 ops_complete_biofill, sh); 595} 596 597static void ops_complete_compute5(void *stripe_head_ref) 598{ 599 struct stripe_head *sh = stripe_head_ref; 600 int target = sh->ops.target; 601 struct r5dev *tgt = &sh->dev[target]; 602 603 pr_debug("%s: stripe %llu\n", __FUNCTION__, 604 (unsigned long long)sh->sector); 605 606 set_bit(R5_UPTODATE, &tgt->flags); 607 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags)); 608 clear_bit(R5_Wantcompute, &tgt->flags); 609 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete); 610 set_bit(STRIPE_HANDLE, &sh->state); 611 release_stripe(sh); 612} 613 614static struct dma_async_tx_descriptor * 615ops_run_compute5(struct stripe_head *sh, unsigned long pending) 616{ 617 /* kernel stack size limits the total number of disks */ 618 int disks = sh->disks; 619 struct page *xor_srcs[disks]; 620 int target = sh->ops.target; 621 struct r5dev *tgt = &sh->dev[target]; 622 struct page *xor_dest = tgt->page; 623 int count = 0; 624 struct dma_async_tx_descriptor *tx; 625 int i; 626 627 pr_debug("%s: stripe %llu block: %d\n", 628 __FUNCTION__, (unsigned long long)sh->sector, target); 629 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags)); 630 631 for (i = disks; i--; ) 632 if (i != target) 633 xor_srcs[count++] = sh->dev[i].page; 634 635 atomic_inc(&sh->count); 636 637 if (unlikely(count == 1)) 638 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, 639 0, NULL, ops_complete_compute5, sh); 640 else 641 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, 642 ASYNC_TX_XOR_ZERO_DST, NULL, 643 ops_complete_compute5, sh); 644 645 /* ack now if postxor is not set to be run */ 646 if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending)) 647 async_tx_ack(tx); 648 649 return tx; 650} 651 652static void ops_complete_prexor(void *stripe_head_ref) 653{ 654 struct stripe_head *sh = stripe_head_ref; 655 656 pr_debug("%s: stripe %llu\n", __FUNCTION__, 657 (unsigned long long)sh->sector); 658 659 set_bit(STRIPE_OP_PREXOR, &sh->ops.complete); 660} 661 662static struct dma_async_tx_descriptor * 663ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx) 664{ 665 /* kernel stack size limits the total number of disks */ 666 int disks = sh->disks; 667 struct page *xor_srcs[disks]; 668 int count = 0, pd_idx = sh->pd_idx, i; 669 670 /* existing parity data subtracted */ 671 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page; 672 673 pr_debug("%s: stripe %llu\n", __FUNCTION__, 674 (unsigned long long)sh->sector); 675 676 for (i = disks; i--; ) { 677 struct r5dev *dev = &sh->dev[i]; 678 /* Only process blocks that are known to be uptodate */ 679 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags)) 680 xor_srcs[count++] = dev->page; 681 } 682 683 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, 684 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx, 685 ops_complete_prexor, sh); 686 687 return tx; 688} 689 690static struct dma_async_tx_descriptor * 691ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx, 692 unsigned long pending) 693{ 694 int disks = sh->disks; 695 int pd_idx = sh->pd_idx, i; 696 697 /* check if prexor is active which means only process blocks 698 * that are part of a read-modify-write (Wantprexor) 699 */ 700 int prexor = test_bit(STRIPE_OP_PREXOR, &pending); 701 702 pr_debug("%s: stripe %llu\n", __FUNCTION__, 703 (unsigned long long)sh->sector); 704 705 for (i = disks; i--; ) { 706 struct r5dev *dev = &sh->dev[i]; 707 struct bio *chosen; 708 int towrite; 709 710 towrite = 0; 711 if (prexor) { /* rmw */ 712 if (dev->towrite && 713 test_bit(R5_Wantprexor, &dev->flags)) 714 towrite = 1; 715 } else { /* rcw */ 716 if (i != pd_idx && dev->towrite && 717 test_bit(R5_LOCKED, &dev->flags)) 718 towrite = 1; 719 } 720 721 if (towrite) { 722 struct bio *wbi; 723 724 spin_lock(&sh->lock); 725 chosen = dev->towrite; 726 dev->towrite = NULL; 727 BUG_ON(dev->written); 728 wbi = dev->written = chosen; 729 spin_unlock(&sh->lock); 730 731 while (wbi && wbi->bi_sector < 732 dev->sector + STRIPE_SECTORS) { 733 tx = async_copy_data(1, wbi, dev->page, 734 dev->sector, tx); 735 wbi = r5_next_bio(wbi, dev->sector); 736 } 737 } 738 } 739 740 return tx; 741} 742 743static void ops_complete_postxor(void *stripe_head_ref) 744{ 745 struct stripe_head *sh = stripe_head_ref; 746 747 pr_debug("%s: stripe %llu\n", __FUNCTION__, 748 (unsigned long long)sh->sector); 749 750 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete); 751 set_bit(STRIPE_HANDLE, &sh->state); 752 release_stripe(sh); 753} 754 755static void ops_complete_write(void *stripe_head_ref) 756{ 757 struct stripe_head *sh = stripe_head_ref; 758 int disks = sh->disks, i, pd_idx = sh->pd_idx; 759 760 pr_debug("%s: stripe %llu\n", __FUNCTION__, 761 (unsigned long long)sh->sector); 762 763 for (i = disks; i--; ) { 764 struct r5dev *dev = &sh->dev[i]; 765 if (dev->written || i == pd_idx) 766 set_bit(R5_UPTODATE, &dev->flags); 767 } 768 769 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete); 770 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete); 771 772 set_bit(STRIPE_HANDLE, &sh->state); 773 release_stripe(sh); 774} 775 776static void 777ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx, 778 unsigned long pending) 779{ 780 /* kernel stack size limits the total number of disks */ 781 int disks = sh->disks; 782 struct page *xor_srcs[disks]; 783 784 int count = 0, pd_idx = sh->pd_idx, i; 785 struct page *xor_dest; 786 int prexor = test_bit(STRIPE_OP_PREXOR, &pending); 787 unsigned long flags; 788 dma_async_tx_callback callback; 789 790 pr_debug("%s: stripe %llu\n", __FUNCTION__, 791 (unsigned long long)sh->sector); 792 793 /* check if prexor is active which means only process blocks 794 * that are part of a read-modify-write (written) 795 */ 796 if (prexor) { 797 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page; 798 for (i = disks; i--; ) { 799 struct r5dev *dev = &sh->dev[i]; 800 if (dev->written) 801 xor_srcs[count++] = dev->page; 802 } 803 } else { 804 xor_dest = sh->dev[pd_idx].page; 805 for (i = disks; i--; ) { 806 struct r5dev *dev = &sh->dev[i]; 807 if (i != pd_idx) 808 xor_srcs[count++] = dev->page; 809 } 810 } 811 812 /* check whether this postxor is part of a write */ 813 callback = test_bit(STRIPE_OP_BIODRAIN, &pending) ? 814 ops_complete_write : ops_complete_postxor; 815 816 /* 1/ if we prexor'd then the dest is reused as a source 817 * 2/ if we did not prexor then we are redoing the parity 818 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST 819 * for the synchronous xor case 820 */ 821 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK | 822 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST); 823 824 atomic_inc(&sh->count); 825 826 if (unlikely(count == 1)) { 827 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST); 828 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, 829 flags, tx, callback, sh); 830 } else 831 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, 832 flags, tx, callback, sh); 833} 834 835static void ops_complete_check(void *stripe_head_ref) 836{ 837 struct stripe_head *sh = stripe_head_ref; 838 int pd_idx = sh->pd_idx; 839 840 pr_debug("%s: stripe %llu\n", __FUNCTION__, 841 (unsigned long long)sh->sector); 842 843 if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) && 844 sh->ops.zero_sum_result == 0) 845 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags); 846 847 set_bit(STRIPE_OP_CHECK, &sh->ops.complete); 848 set_bit(STRIPE_HANDLE, &sh->state); 849 release_stripe(sh); 850} 851 852static void ops_run_check(struct stripe_head *sh) 853{ 854 /* kernel stack size limits the total number of disks */ 855 int disks = sh->disks; 856 struct page *xor_srcs[disks]; 857 struct dma_async_tx_descriptor *tx; 858 859 int count = 0, pd_idx = sh->pd_idx, i; 860 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page; 861 862 pr_debug("%s: stripe %llu\n", __FUNCTION__, 863 (unsigned long long)sh->sector); 864 865 for (i = disks; i--; ) { 866 struct r5dev *dev = &sh->dev[i]; 867 if (i != pd_idx) 868 xor_srcs[count++] = dev->page; 869 } 870 871 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, 872 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL); 873 874 if (tx) 875 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending); 876 else 877 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending); 878 879 atomic_inc(&sh->count); 880 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx, 881 ops_complete_check, sh); 882} 883 884static void raid5_run_ops(struct stripe_head *sh, unsigned long pending) 885{ 886 int overlap_clear = 0, i, disks = sh->disks; 887 struct dma_async_tx_descriptor *tx = NULL; 888 889 if (test_bit(STRIPE_OP_BIOFILL, &pending)) { 890 ops_run_biofill(sh); 891 overlap_clear++; 892 } 893 894 if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending)) 895 tx = ops_run_compute5(sh, pending); 896 897 if (test_bit(STRIPE_OP_PREXOR, &pending)) 898 tx = ops_run_prexor(sh, tx); 899 900 if (test_bit(STRIPE_OP_BIODRAIN, &pending)) { 901 tx = ops_run_biodrain(sh, tx, pending); 902 overlap_clear++; 903 } 904 905 if (test_bit(STRIPE_OP_POSTXOR, &pending)) 906 ops_run_postxor(sh, tx, pending); 907 908 if (test_bit(STRIPE_OP_CHECK, &pending)) 909 ops_run_check(sh); 910 911 if (test_bit(STRIPE_OP_IO, &pending)) 912 ops_run_io(sh); 913 914 if (overlap_clear) 915 for (i = disks; i--; ) { 916 struct r5dev *dev = &sh->dev[i]; 917 if (test_and_clear_bit(R5_Overlap, &dev->flags)) 918 wake_up(&sh->raid_conf->wait_for_overlap); 919 } 920} 921 922static int grow_one_stripe(raid5_conf_t *conf) 923{ 924 struct stripe_head *sh; 925 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL); 926 if (!sh) 927 return 0; 928 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev)); 929 sh->raid_conf = conf; 930 spin_lock_init(&sh->lock); 931 932 if (grow_buffers(sh, conf->raid_disks)) { 933 shrink_buffers(sh, conf->raid_disks); 934 kmem_cache_free(conf->slab_cache, sh); 935 return 0; 936 } 937 sh->disks = conf->raid_disks; 938 /* we just created an active stripe so... */ 939 atomic_set(&sh->count, 1); 940 atomic_inc(&conf->active_stripes); 941 INIT_LIST_HEAD(&sh->lru); 942 release_stripe(sh); 943 return 1; 944} 945 946static int grow_stripes(raid5_conf_t *conf, int num) 947{ 948 struct kmem_cache *sc; 949 int devs = conf->raid_disks; 950 951 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev)); 952 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev)); 953 conf->active_name = 0; 954 sc = kmem_cache_create(conf->cache_name[conf->active_name], 955 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev), 956 0, 0, NULL); 957 if (!sc) 958 return 1; 959 conf->slab_cache = sc; 960 conf->pool_size = devs; 961 while (num--) 962 if (!grow_one_stripe(conf)) 963 return 1; 964 return 0; 965} 966 967#ifdef CONFIG_MD_RAID5_RESHAPE 968static int resize_stripes(raid5_conf_t *conf, int newsize) 969{ 970 /* Make all the stripes able to hold 'newsize' devices. 971 * New slots in each stripe get 'page' set to a new page. 972 * 973 * This happens in stages: 974 * 1/ create a new kmem_cache and allocate the required number of 975 * stripe_heads. 976 * 2/ gather all the old stripe_heads and tranfer the pages across 977 * to the new stripe_heads. This will have the side effect of 978 * freezing the array as once all stripe_heads have been collected, 979 * no IO will be possible. Old stripe heads are freed once their 980 * pages have been transferred over, and the old kmem_cache is 981 * freed when all stripes are done. 982 * 3/ reallocate conf->disks to be suitable bigger. If this fails, 983 * we simple return a failre status - no need to clean anything up. 984 * 4/ allocate new pages for the new slots in the new stripe_heads. 985 * If this fails, we don't bother trying the shrink the 986 * stripe_heads down again, we just leave them as they are. 987 * As each stripe_head is processed the new one is released into 988 * active service. 989 * 990 * Once step2 is started, we cannot afford to wait for a write, 991 * so we use GFP_NOIO allocations. 992 */ 993 struct stripe_head *osh, *nsh; 994 LIST_HEAD(newstripes); 995 struct disk_info *ndisks; 996 int err = 0; 997 struct kmem_cache *sc; 998 int i; 999 1000 if (newsize <= conf->pool_size) 1001 return 0; /* never bother to shrink */ 1002 1003 md_allow_write(conf->mddev); 1004 1005 /* Step 1 */ 1006 sc = kmem_cache_create(conf->cache_name[1-conf->active_name], 1007 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev), 1008 0, 0, NULL); 1009 if (!sc) 1010 return -ENOMEM; 1011 1012 for (i = conf->max_nr_stripes; i; i--) { 1013 nsh = kmem_cache_alloc(sc, GFP_KERNEL); 1014 if (!nsh) 1015 break; 1016 1017 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev)); 1018 1019 nsh->raid_conf = conf; 1020 spin_lock_init(&nsh->lock); 1021 1022 list_add(&nsh->lru, &newstripes); 1023 } 1024 if (i) { 1025 /* didn't get enough, give up */ 1026 while (!list_empty(&newstripes)) { 1027 nsh = list_entry(newstripes.next, struct stripe_head, lru); 1028 list_del(&nsh->lru); 1029 kmem_cache_free(sc, nsh); 1030 } 1031 kmem_cache_destroy(sc); 1032 return -ENOMEM; 1033 } 1034 /* Step 2 - Must use GFP_NOIO now. 1035 * OK, we have enough stripes, start collecting inactive 1036 * stripes and copying them over 1037 */ 1038 list_for_each_entry(nsh, &newstripes, lru) { 1039 spin_lock_irq(&conf->device_lock); 1040 wait_event_lock_irq(conf->wait_for_stripe, 1041 !list_empty(&conf->inactive_list), 1042 conf->device_lock, 1043 unplug_slaves(conf->mddev) 1044 ); 1045 osh = get_free_stripe(conf); 1046 spin_unlock_irq(&conf->device_lock); 1047 atomic_set(&nsh->count, 1); 1048 for(i=0; i<conf->pool_size; i++) 1049 nsh->dev[i].page = osh->dev[i].page; 1050 for( ; i<newsize; i++) 1051 nsh->dev[i].page = NULL; 1052 kmem_cache_free(conf->slab_cache, osh); 1053 } 1054 kmem_cache_destroy(conf->slab_cache); 1055 1056 /* Step 3. 1057 * At this point, we are holding all the stripes so the array 1058 * is completely stalled, so now is a good time to resize 1059 * conf->disks. 1060 */ 1061 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO); 1062 if (ndisks) { 1063 for (i=0; i<conf->raid_disks; i++) 1064 ndisks[i] = conf->disks[i]; 1065 kfree(conf->disks); 1066 conf->disks = ndisks; 1067 } else 1068 err = -ENOMEM; 1069 1070 /* Step 4, return new stripes to service */ 1071 while(!list_empty(&newstripes)) { 1072 nsh = list_entry(newstripes.next, struct stripe_head, lru); 1073 list_del_init(&nsh->lru); 1074 for (i=conf->raid_disks; i < newsize; i++) 1075 if (nsh->dev[i].page == NULL) { 1076 struct page *p = alloc_page(GFP_NOIO); 1077 nsh->dev[i].page = p; 1078 if (!p) 1079 err = -ENOMEM; 1080 } 1081 release_stripe(nsh); 1082 } 1083 /* critical section pass, GFP_NOIO no longer needed */ 1084 1085 conf->slab_cache = sc; 1086 conf->active_name = 1-conf->active_name; 1087 conf->pool_size = newsize; 1088 return err; 1089} 1090#endif 1091 1092static int drop_one_stripe(raid5_conf_t *conf) 1093{ 1094 struct stripe_head *sh; 1095 1096 spin_lock_irq(&conf->device_lock); 1097 sh = get_free_stripe(conf); 1098 spin_unlock_irq(&conf->device_lock); 1099 if (!sh) 1100 return 0; 1101 BUG_ON(atomic_read(&sh->count)); 1102 shrink_buffers(sh, conf->pool_size); 1103 kmem_cache_free(conf->slab_cache, sh); 1104 atomic_dec(&conf->active_stripes); 1105 return 1; 1106} 1107 1108static void shrink_stripes(raid5_conf_t *conf) 1109{ 1110 while (drop_one_stripe(conf)) 1111 ; 1112 1113 if (conf->slab_cache) 1114 kmem_cache_destroy(conf->slab_cache); 1115 conf->slab_cache = NULL; 1116} 1117 1118static void raid5_end_read_request(struct bio * bi, int error) 1119{ 1120 struct stripe_head *sh = bi->bi_private; 1121 raid5_conf_t *conf = sh->raid_conf; 1122 int disks = sh->disks, i; 1123 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags); 1124 char b[BDEVNAME_SIZE]; 1125 mdk_rdev_t *rdev; 1126 1127 1128 for (i=0 ; i<disks; i++) 1129 if (bi == &sh->dev[i].req) 1130 break; 1131 1132 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n", 1133 (unsigned long long)sh->sector, i, atomic_read(&sh->count), 1134 uptodate); 1135 if (i == disks) { 1136 BUG(); 1137 return; 1138 } 1139 1140 if (uptodate) { 1141 set_bit(R5_UPTODATE, &sh->dev[i].flags); 1142 if (test_bit(R5_ReadError, &sh->dev[i].flags)) { 1143 rdev = conf->disks[i].rdev; 1144 printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n", 1145 mdname(conf->mddev), STRIPE_SECTORS, 1146 (unsigned long long)sh->sector + rdev->data_offset, 1147 bdevname(rdev->bdev, b)); 1148 clear_bit(R5_ReadError, &sh->dev[i].flags); 1149 clear_bit(R5_ReWrite, &sh->dev[i].flags); 1150 } 1151 if (atomic_read(&conf->disks[i].rdev->read_errors)) 1152 atomic_set(&conf->disks[i].rdev->read_errors, 0); 1153 } else { 1154 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b); 1155 int retry = 0; 1156 rdev = conf->disks[i].rdev; 1157 1158 clear_bit(R5_UPTODATE, &sh->dev[i].flags); 1159 atomic_inc(&rdev->read_errors); 1160 if (conf->mddev->degraded) 1161 printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n", 1162 mdname(conf->mddev), 1163 (unsigned long long)sh->sector + rdev->data_offset, 1164 bdn); 1165 else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) 1166 /* Oh, no!!! */ 1167 printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n", 1168 mdname(conf->mddev), 1169 (unsigned long long)sh->sector + rdev->data_offset, 1170 bdn); 1171 else if (atomic_read(&rdev->read_errors) 1172 > conf->max_nr_stripes) 1173 printk(KERN_WARNING 1174 "raid5:%s: Too many read errors, failing device %s.\n", 1175 mdname(conf->mddev), bdn); 1176 else 1177 retry = 1; 1178 if (retry) 1179 set_bit(R5_ReadError, &sh->dev[i].flags); 1180 else { 1181 clear_bit(R5_ReadError, &sh->dev[i].flags); 1182 clear_bit(R5_ReWrite, &sh->dev[i].flags); 1183 md_error(conf->mddev, rdev); 1184 } 1185 } 1186 rdev_dec_pending(conf->disks[i].rdev, conf->mddev); 1187 clear_bit(R5_LOCKED, &sh->dev[i].flags); 1188 set_bit(STRIPE_HANDLE, &sh->state); 1189 release_stripe(sh); 1190} 1191 1192static void raid5_end_write_request (struct bio *bi, int error) 1193{ 1194 struct stripe_head *sh = bi->bi_private; 1195 raid5_conf_t *conf = sh->raid_conf; 1196 int disks = sh->disks, i; 1197 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags); 1198 1199 for (i=0 ; i<disks; i++) 1200 if (bi == &sh->dev[i].req) 1201 break; 1202 1203 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n", 1204 (unsigned long long)sh->sector, i, atomic_read(&sh->count), 1205 uptodate); 1206 if (i == disks) { 1207 BUG(); 1208 return; 1209 } 1210 1211 if (!uptodate) 1212 md_error(conf->mddev, conf->disks[i].rdev); 1213 1214 rdev_dec_pending(conf->disks[i].rdev, conf->mddev); 1215 1216 clear_bit(R5_LOCKED, &sh->dev[i].flags); 1217 set_bit(STRIPE_HANDLE, &sh->state); 1218 release_stripe(sh); 1219} 1220 1221 1222static sector_t compute_blocknr(struct stripe_head *sh, int i); 1223 1224static void raid5_build_block (struct stripe_head *sh, int i) 1225{ 1226 struct r5dev *dev = &sh->dev[i]; 1227 1228 bio_init(&dev->req); 1229 dev->req.bi_io_vec = &dev->vec; 1230 dev->req.bi_vcnt++; 1231 dev->req.bi_max_vecs++; 1232 dev->vec.bv_page = dev->page; 1233 dev->vec.bv_len = STRIPE_SIZE; 1234 dev->vec.bv_offset = 0; 1235 1236 dev->req.bi_sector = sh->sector; 1237 dev->req.bi_private = sh; 1238 1239 dev->flags = 0; 1240 dev->sector = compute_blocknr(sh, i); 1241} 1242 1243static void error(mddev_t *mddev, mdk_rdev_t *rdev) 1244{ 1245 char b[BDEVNAME_SIZE]; 1246 raid5_conf_t *conf = (raid5_conf_t *) mddev->private; 1247 pr_debug("raid5: error called\n"); 1248 1249 if (!test_bit(Faulty, &rdev->flags)) { 1250 set_bit(MD_CHANGE_DEVS, &mddev->flags); 1251 if (test_and_clear_bit(In_sync, &rdev->flags)) { 1252 unsigned long flags; 1253 spin_lock_irqsave(&conf->device_lock, flags); 1254 mddev->degraded++; 1255 spin_unlock_irqrestore(&conf->device_lock, flags); 1256 /* 1257 * if recovery was running, make sure it aborts. 1258 */ 1259 set_bit(MD_RECOVERY_ERR, &mddev->recovery); 1260 } 1261 set_bit(Faulty, &rdev->flags); 1262 printk (KERN_ALERT 1263 "raid5: Disk failure on %s, disabling device." 1264 " Operation continuing on %d devices\n", 1265 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded); 1266 } 1267} 1268 1269/* 1270 * Input: a 'big' sector number, 1271 * Output: index of the data and parity disk, and the sector # in them. 1272 */ 1273static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks, 1274 unsigned int data_disks, unsigned int * dd_idx, 1275 unsigned int * pd_idx, raid5_conf_t *conf) 1276{ 1277 long stripe; 1278 unsigned long chunk_number; 1279 unsigned int chunk_offset; 1280 sector_t new_sector; 1281 int sectors_per_chunk = conf->chunk_size >> 9; 1282 1283 /* First compute the information on this sector */ 1284 1285 /* 1286 * Compute the chunk number and the sector offset inside the chunk 1287 */ 1288 chunk_offset = sector_div(r_sector, sectors_per_chunk); 1289 chunk_number = r_sector; 1290 BUG_ON(r_sector != chunk_number); 1291 1292 /* 1293 * Compute the stripe number 1294 */ 1295 stripe = chunk_number / data_disks; 1296 1297 /* 1298 * Compute the data disk and parity disk indexes inside the stripe 1299 */ 1300 *dd_idx = chunk_number % data_disks; 1301 1302 /* 1303 * Select the parity disk based on the user selected algorithm. 1304 */ 1305 switch(conf->level) { 1306 case 4: 1307 *pd_idx = data_disks; 1308 break; 1309 case 5: 1310 switch (conf->algorithm) { 1311 case ALGORITHM_LEFT_ASYMMETRIC: 1312 *pd_idx = data_disks - stripe % raid_disks; 1313 if (*dd_idx >= *pd_idx) 1314 (*dd_idx)++; 1315 break; 1316 case ALGORITHM_RIGHT_ASYMMETRIC: 1317 *pd_idx = stripe % raid_disks; 1318 if (*dd_idx >= *pd_idx) 1319 (*dd_idx)++; 1320 break; 1321 case ALGORITHM_LEFT_SYMMETRIC: 1322 *pd_idx = data_disks - stripe % raid_disks; 1323 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks; 1324 break; 1325 case ALGORITHM_RIGHT_SYMMETRIC: 1326 *pd_idx = stripe % raid_disks; 1327 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks; 1328 break; 1329 default: 1330 printk(KERN_ERR "raid5: unsupported algorithm %d\n", 1331 conf->algorithm); 1332 } 1333 break; 1334 case 6: 1335 1336 /**** FIX THIS ****/ 1337 switch (conf->algorithm) { 1338 case ALGORITHM_LEFT_ASYMMETRIC: 1339 *pd_idx = raid_disks - 1 - (stripe % raid_disks); 1340 if (*pd_idx == raid_disks-1) 1341 (*dd_idx)++; /* Q D D D P */ 1342 else if (*dd_idx >= *pd_idx) 1343 (*dd_idx) += 2; /* D D P Q D */ 1344 break; 1345 case ALGORITHM_RIGHT_ASYMMETRIC: 1346 *pd_idx = stripe % raid_disks; 1347 if (*pd_idx == raid_disks-1) 1348 (*dd_idx)++; /* Q D D D P */ 1349 else if (*dd_idx >= *pd_idx) 1350 (*dd_idx) += 2; /* D D P Q D */ 1351 break; 1352 case ALGORITHM_LEFT_SYMMETRIC: 1353 *pd_idx = raid_disks - 1 - (stripe % raid_disks); 1354 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks; 1355 break; 1356 case ALGORITHM_RIGHT_SYMMETRIC: 1357 *pd_idx = stripe % raid_disks; 1358 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks; 1359 break; 1360 default: 1361 printk (KERN_CRIT "raid6: unsupported algorithm %d\n", 1362 conf->algorithm); 1363 } 1364 break; 1365 } 1366 1367 /* 1368 * Finally, compute the new sector number 1369 */ 1370 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset; 1371 return new_sector; 1372} 1373 1374 1375static sector_t compute_blocknr(struct stripe_head *sh, int i) 1376{ 1377 raid5_conf_t *conf = sh->raid_conf; 1378 int raid_disks = sh->disks; 1379 int data_disks = raid_disks - conf->max_degraded; 1380 sector_t new_sector = sh->sector, check; 1381 int sectors_per_chunk = conf->chunk_size >> 9; 1382 sector_t stripe; 1383 int chunk_offset; 1384 int chunk_number, dummy1, dummy2, dd_idx = i; 1385 sector_t r_sector; 1386 1387 1388 chunk_offset = sector_div(new_sector, sectors_per_chunk); 1389 stripe = new_sector; 1390 BUG_ON(new_sector != stripe); 1391 1392 if (i == sh->pd_idx) 1393 return 0; 1394 switch(conf->level) { 1395 case 4: break; 1396 case 5: 1397 switch (conf->algorithm) { 1398 case ALGORITHM_LEFT_ASYMMETRIC: 1399 case ALGORITHM_RIGHT_ASYMMETRIC: 1400 if (i > sh->pd_idx) 1401 i--; 1402 break; 1403 case ALGORITHM_LEFT_SYMMETRIC: 1404 case ALGORITHM_RIGHT_SYMMETRIC: 1405 if (i < sh->pd_idx) 1406 i += raid_disks; 1407 i -= (sh->pd_idx + 1); 1408 break; 1409 default: 1410 printk(KERN_ERR "raid5: unsupported algorithm %d\n", 1411 conf->algorithm); 1412 } 1413 break; 1414 case 6: 1415 if (i == raid6_next_disk(sh->pd_idx, raid_disks)) 1416 return 0; /* It is the Q disk */ 1417 switch (conf->algorithm) { 1418 case ALGORITHM_LEFT_ASYMMETRIC: 1419 case ALGORITHM_RIGHT_ASYMMETRIC: 1420 if (sh->pd_idx == raid_disks-1) 1421 i--; /* Q D D D P */ 1422 else if (i > sh->pd_idx) 1423 i -= 2; /* D D P Q D */ 1424 break; 1425 case ALGORITHM_LEFT_SYMMETRIC: 1426 case ALGORITHM_RIGHT_SYMMETRIC: 1427 if (sh->pd_idx == raid_disks-1) 1428 i--; /* Q D D D P */ 1429 else { 1430 /* D D P Q D */ 1431 if (i < sh->pd_idx) 1432 i += raid_disks; 1433 i -= (sh->pd_idx + 2); 1434 } 1435 break; 1436 default: 1437 printk (KERN_CRIT "raid6: unsupported algorithm %d\n", 1438 conf->algorithm); 1439 } 1440 break; 1441 } 1442 1443 chunk_number = stripe * data_disks + i; 1444 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset; 1445 1446 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf); 1447 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) { 1448 printk(KERN_ERR "compute_blocknr: map not correct\n"); 1449 return 0; 1450 } 1451 return r_sector; 1452} 1453 1454 1455 1456/* 1457 * Copy data between a page in the stripe cache, and one or more bion 1458 * The page could align with the middle of the bio, or there could be 1459 * several bion, each with several bio_vecs, which cover part of the page 1460 * Multiple bion are linked together on bi_next. There may be extras 1461 * at the end of this list. We ignore them. 1462 */ 1463static void copy_data(int frombio, struct bio *bio, 1464 struct page *page, 1465 sector_t sector) 1466{ 1467 char *pa = page_address(page); 1468 struct bio_vec *bvl; 1469 int i; 1470 int page_offset; 1471 1472 if (bio->bi_sector >= sector) 1473 page_offset = (signed)(bio->bi_sector - sector) * 512; 1474 else 1475 page_offset = (signed)(sector - bio->bi_sector) * -512; 1476 bio_for_each_segment(bvl, bio, i) { 1477 int len = bio_iovec_idx(bio,i)->bv_len; 1478 int clen; 1479 int b_offset = 0; 1480 1481 if (page_offset < 0) { 1482 b_offset = -page_offset; 1483 page_offset += b_offset; 1484 len -= b_offset; 1485 } 1486 1487 if (len > 0 && page_offset + len > STRIPE_SIZE) 1488 clen = STRIPE_SIZE - page_offset; 1489 else clen = len; 1490 1491 if (clen > 0) { 1492 char *ba = __bio_kmap_atomic(bio, i, KM_USER0); 1493 if (frombio) 1494 memcpy(pa+page_offset, ba+b_offset, clen); 1495 else 1496 memcpy(ba+b_offset, pa+page_offset, clen); 1497 __bio_kunmap_atomic(ba, KM_USER0); 1498 } 1499 if (clen < len) /* hit end of page */ 1500 break; 1501 page_offset += len; 1502 } 1503} 1504 1505#define check_xor() do { \ 1506 if (count == MAX_XOR_BLOCKS) { \ 1507 xor_blocks(count, STRIPE_SIZE, dest, ptr);\ 1508 count = 0; \ 1509 } \ 1510 } while(0) 1511 1512static void compute_parity6(struct stripe_head *sh, int method) 1513{ 1514 raid6_conf_t *conf = sh->raid_conf; 1515 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count; 1516 struct bio *chosen; 1517 /**** FIX THIS: This could be very bad if disks is close to 256 ****/ 1518 void *ptrs[disks]; 1519 1520 qd_idx = raid6_next_disk(pd_idx, disks); 1521 d0_idx = raid6_next_disk(qd_idx, disks); 1522 1523 pr_debug("compute_parity, stripe %llu, method %d\n", 1524 (unsigned long long)sh->sector, method); 1525 1526 switch(method) { 1527 case READ_MODIFY_WRITE: 1528 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */ 1529 case RECONSTRUCT_WRITE: 1530 for (i= disks; i-- ;) 1531 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) { 1532 chosen = sh->dev[i].towrite; 1533 sh->dev[i].towrite = NULL; 1534 1535 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) 1536 wake_up(&conf->wait_for_overlap); 1537 1538 BUG_ON(sh->dev[i].written); 1539 sh->dev[i].written = chosen; 1540 } 1541 break; 1542 case CHECK_PARITY: 1543 BUG(); /* Not implemented yet */ 1544 } 1545 1546 for (i = disks; i--;) 1547 if (sh->dev[i].written) { 1548 sector_t sector = sh->dev[i].sector; 1549 struct bio *wbi = sh->dev[i].written; 1550 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) { 1551 copy_data(1, wbi, sh->dev[i].page, sector); 1552 wbi = r5_next_bio(wbi, sector); 1553 } 1554 1555 set_bit(R5_LOCKED, &sh->dev[i].flags); 1556 set_bit(R5_UPTODATE, &sh->dev[i].flags); 1557 } 1558 1559// switch(method) { 1560// case RECONSTRUCT_WRITE: 1561// case CHECK_PARITY: 1562// case UPDATE_PARITY: 1563 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */ 1564 /* FIX: Is this ordering of drives even remotely optimal? */ 1565 count = 0; 1566 i = d0_idx; 1567 do { 1568 ptrs[count++] = page_address(sh->dev[i].page); 1569 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags)) 1570 printk("block %d/%d not uptodate on parity calc\n", i,count); 1571 i = raid6_next_disk(i, disks); 1572 } while ( i != d0_idx ); 1573// break; 1574// } 1575 1576 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs); 1577 1578 switch(method) { 1579 case RECONSTRUCT_WRITE: 1580 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags); 1581 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags); 1582 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags); 1583 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags); 1584 break; 1585 case UPDATE_PARITY: 1586 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags); 1587 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags); 1588 break; 1589 } 1590} 1591 1592 1593/* Compute one missing block */ 1594static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero) 1595{ 1596 int i, count, disks = sh->disks; 1597 void *ptr[MAX_XOR_BLOCKS], *dest, *p; 1598 int pd_idx = sh->pd_idx; 1599 int qd_idx = raid6_next_disk(pd_idx, disks); 1600 1601 pr_debug("compute_block_1, stripe %llu, idx %d\n", 1602 (unsigned long long)sh->sector, dd_idx); 1603 1604 if ( dd_idx == qd_idx ) { 1605 /* We're actually computing the Q drive */ 1606 compute_parity6(sh, UPDATE_PARITY); 1607 } else { 1608 dest = page_address(sh->dev[dd_idx].page); 1609 if (!nozero) memset(dest, 0, STRIPE_SIZE); 1610 count = 0; 1611 for (i = disks ; i--; ) { 1612 if (i == dd_idx || i == qd_idx) 1613 continue; 1614 p = page_address(sh->dev[i].page); 1615 if (test_bit(R5_UPTODATE, &sh->dev[i].flags)) 1616 ptr[count++] = p; 1617 else 1618 printk("compute_block() %d, stripe %llu, %d" 1619 " not present\n", dd_idx, 1620 (unsigned long long)sh->sector, i); 1621 1622 check_xor(); 1623 } 1624 if (count) 1625 xor_blocks(count, STRIPE_SIZE, dest, ptr); 1626 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags); 1627 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags); 1628 } 1629} 1630 1631/* Compute two missing blocks */ 1632static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2) 1633{ 1634 int i, count, disks = sh->disks; 1635 int pd_idx = sh->pd_idx; 1636 int qd_idx = raid6_next_disk(pd_idx, disks); 1637 int d0_idx = raid6_next_disk(qd_idx, disks); 1638 int faila, failb; 1639 1640 /* faila and failb are disk numbers relative to d0_idx */ 1641 /* pd_idx become disks-2 and qd_idx become disks-1 */ 1642 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx; 1643 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx; 1644 1645 BUG_ON(faila == failb); 1646 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; } 1647 1648 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n", 1649 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb); 1650 1651 if ( failb == disks-1 ) { 1652 /* Q disk is one of the missing disks */ 1653 if ( faila == disks-2 ) { 1654 /* Missing P+Q, just recompute */ 1655 compute_parity6(sh, UPDATE_PARITY); 1656 return; 1657 } else { 1658 /* We're missing D+Q; recompute D from P */ 1659 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0); 1660 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */ 1661 return; 1662 } 1663 } 1664 1665 /* We're missing D+P or D+D; build pointer table */ 1666 { 1667 /**** FIX THIS: This could be very bad if disks is close to 256 ****/ 1668 void *ptrs[disks]; 1669 1670 count = 0; 1671 i = d0_idx; 1672 do { 1673 ptrs[count++] = page_address(sh->dev[i].page); 1674 i = raid6_next_disk(i, disks); 1675 if (i != dd_idx1 && i != dd_idx2 && 1676 !test_bit(R5_UPTODATE, &sh->dev[i].flags)) 1677 printk("compute_2 with missing block %d/%d\n", count, i); 1678 } while ( i != d0_idx ); 1679 1680 if ( failb == disks-2 ) { 1681 /* We're missing D+P. */ 1682 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs); 1683 } else { 1684 /* We're missing D+D. */ 1685 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs); 1686 } 1687 1688 /* Both the above update both missing blocks */ 1689 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags); 1690 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags); 1691 } 1692} 1693 1694static int 1695handle_write_operations5(struct stripe_head *sh, int rcw, int expand) 1696{ 1697 int i, pd_idx = sh->pd_idx, disks = sh->disks; 1698 int locked = 0; 1699 1700 if (rcw) { 1701 /* if we are not expanding this is a proper write request, and 1702 * there will be bios with new data to be drained into the 1703 * stripe cache 1704 */ 1705 if (!expand) { 1706 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending); 1707 sh->ops.count++; 1708 } 1709 1710 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending); 1711 sh->ops.count++; 1712 1713 for (i = disks; i--; ) { 1714 struct r5dev *dev = &sh->dev[i]; 1715 1716 if (dev->towrite) { 1717 set_bit(R5_LOCKED, &dev->flags); 1718 if (!expand) 1719 clear_bit(R5_UPTODATE, &dev->flags); 1720 locked++; 1721 } 1722 } 1723 } else { 1724 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) || 1725 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags))); 1726 1727 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending); 1728 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending); 1729 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending); 1730 1731 sh->ops.count += 3; 1732 1733 for (i = disks; i--; ) { 1734 struct r5dev *dev = &sh->dev[i]; 1735 if (i == pd_idx) 1736 continue; 1737 1738 /* For a read-modify write there may be blocks that are 1739 * locked for reading while others are ready to be 1740 * written so we distinguish these blocks by the 1741 * R5_Wantprexor bit 1742 */ 1743 if (dev->towrite && 1744 (test_bit(R5_UPTODATE, &dev->flags) || 1745 test_bit(R5_Wantcompute, &dev->flags))) { 1746 set_bit(R5_Wantprexor, &dev->flags); 1747 set_bit(R5_LOCKED, &dev->flags); 1748 clear_bit(R5_UPTODATE, &dev->flags); 1749 locked++; 1750 } 1751 } 1752 } 1753 1754 /* keep the parity disk locked while asynchronous operations 1755 * are in flight 1756 */ 1757 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags); 1758 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags); 1759 locked++; 1760 1761 pr_debug("%s: stripe %llu locked: %d pending: %lx\n", 1762 __FUNCTION__, (unsigned long long)sh->sector, 1763 locked, sh->ops.pending); 1764 1765 return locked; 1766} 1767 1768/* 1769 * Each stripe/dev can have one or more bion attached. 1770 * toread/towrite point to the first in a chain. 1771 * The bi_next chain must be in order. 1772 */ 1773static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite) 1774{ 1775 struct bio **bip; 1776 raid5_conf_t *conf = sh->raid_conf; 1777 int firstwrite=0; 1778 1779 pr_debug("adding bh b#%llu to stripe s#%llu\n", 1780 (unsigned long long)bi->bi_sector, 1781 (unsigned long long)sh->sector); 1782 1783 1784 spin_lock(&sh->lock); 1785 spin_lock_irq(&conf->device_lock); 1786 if (forwrite) { 1787 bip = &sh->dev[dd_idx].towrite; 1788 if (*bip == NULL && sh->dev[dd_idx].written == NULL) 1789 firstwrite = 1; 1790 } else 1791 bip = &sh->dev[dd_idx].toread; 1792 while (*bip && (*bip)->bi_sector < bi->bi_sector) { 1793 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector) 1794 goto overlap; 1795 bip = & (*bip)->bi_next; 1796 } 1797 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9)) 1798 goto overlap; 1799 1800 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next); 1801 if (*bip) 1802 bi->bi_next = *bip; 1803 *bip = bi; 1804 bi->bi_phys_segments ++; 1805 spin_unlock_irq(&conf->device_lock); 1806 spin_unlock(&sh->lock); 1807 1808 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n", 1809 (unsigned long long)bi->bi_sector, 1810 (unsigned long long)sh->sector, dd_idx); 1811 1812 if (conf->mddev->bitmap && firstwrite) { 1813 bitmap_startwrite(conf->mddev->bitmap, sh->sector, 1814 STRIPE_SECTORS, 0); 1815 sh->bm_seq = conf->seq_flush+1; 1816 set_bit(STRIPE_BIT_DELAY, &sh->state); 1817 } 1818 1819 if (forwrite) { 1820 /* check if page is covered */ 1821 sector_t sector = sh->dev[dd_idx].sector; 1822 for (bi=sh->dev[dd_idx].towrite; 1823 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS && 1824 bi && bi->bi_sector <= sector; 1825 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) { 1826 if (bi->bi_sector + (bi->bi_size>>9) >= sector) 1827 sector = bi->bi_sector + (bi->bi_size>>9); 1828 } 1829 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS) 1830 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags); 1831 } 1832 return 1; 1833 1834 overlap: 1835 set_bit(R5_Overlap, &sh->dev[dd_idx].flags); 1836 spin_unlock_irq(&conf->device_lock); 1837 spin_unlock(&sh->lock); 1838 return 0; 1839} 1840 1841static void end_reshape(raid5_conf_t *conf); 1842 1843static int page_is_zero(struct page *p) 1844{ 1845 char *a = page_address(p); 1846 return ((*(u32*)a) == 0 && 1847 memcmp(a, a+4, STRIPE_SIZE-4)==0); 1848} 1849 1850static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks) 1851{ 1852 int sectors_per_chunk = conf->chunk_size >> 9; 1853 int pd_idx, dd_idx; 1854 int chunk_offset = sector_div(stripe, sectors_per_chunk); 1855 1856 raid5_compute_sector(stripe * (disks - conf->max_degraded) 1857 *sectors_per_chunk + chunk_offset, 1858 disks, disks - conf->max_degraded, 1859 &dd_idx, &pd_idx, conf); 1860 return pd_idx; 1861} 1862 1863static void 1864handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh, 1865 struct stripe_head_state *s, int disks, 1866 struct bio **return_bi) 1867{ 1868 int i; 1869 for (i = disks; i--; ) { 1870 struct bio *bi; 1871 int bitmap_end = 0; 1872 1873 if (test_bit(R5_ReadError, &sh->dev[i].flags)) { 1874 mdk_rdev_t *rdev; 1875 rcu_read_lock(); 1876 rdev = rcu_dereference(conf->disks[i].rdev); 1877 if (rdev && test_bit(In_sync, &rdev->flags)) 1878 /* multiple read failures in one stripe */ 1879 md_error(conf->mddev, rdev); 1880 rcu_read_unlock(); 1881 } 1882 spin_lock_irq(&conf->device_lock); 1883 /* fail all writes first */ 1884 bi = sh->dev[i].towrite; 1885 sh->dev[i].towrite = NULL; 1886 if (bi) { 1887 s->to_write--; 1888 bitmap_end = 1; 1889 } 1890 1891 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) 1892 wake_up(&conf->wait_for_overlap); 1893 1894 while (bi && bi->bi_sector < 1895 sh->dev[i].sector + STRIPE_SECTORS) { 1896 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector); 1897 clear_bit(BIO_UPTODATE, &bi->bi_flags); 1898 if (--bi->bi_phys_segments == 0) { 1899 md_write_end(conf->mddev); 1900 bi->bi_next = *return_bi; 1901 *return_bi = bi; 1902 } 1903 bi = nextbi; 1904 } 1905 /* and fail all 'written' */ 1906 bi = sh->dev[i].written; 1907 sh->dev[i].written = NULL; 1908 if (bi) bitmap_end = 1; 1909 while (bi && bi->bi_sector < 1910 sh->dev[i].sector + STRIPE_SECTORS) { 1911 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector); 1912 clear_bit(BIO_UPTODATE, &bi->bi_flags); 1913 if (--bi->bi_phys_segments == 0) { 1914 md_write_end(conf->mddev); 1915 bi->bi_next = *return_bi; 1916 *return_bi = bi; 1917 } 1918 bi = bi2; 1919 } 1920 1921 /* fail any reads if this device is non-operational and 1922 * the data has not reached the cache yet. 1923 */ 1924 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) && 1925 (!test_bit(R5_Insync, &sh->dev[i].flags) || 1926 test_bit(R5_ReadError, &sh->dev[i].flags))) { 1927 bi = sh->dev[i].toread; 1928 sh->dev[i].toread = NULL; 1929 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) 1930 wake_up(&conf->wait_for_overlap); 1931 if (bi) s->to_read--; 1932 while (bi && bi->bi_sector < 1933 sh->dev[i].sector + STRIPE_SECTORS) { 1934 struct bio *nextbi = 1935 r5_next_bio(bi, sh->dev[i].sector); 1936 clear_bit(BIO_UPTODATE, &bi->bi_flags); 1937 if (--bi->bi_phys_segments == 0) { 1938 bi->bi_next = *return_bi; 1939 *return_bi = bi; 1940 } 1941 bi = nextbi; 1942 } 1943 } 1944 spin_unlock_irq(&conf->device_lock); 1945 if (bitmap_end) 1946 bitmap_endwrite(conf->mddev->bitmap, sh->sector, 1947 STRIPE_SECTORS, 0, 0); 1948 } 1949 1950} 1951 1952/* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks 1953 * to process 1954 */ 1955static int __handle_issuing_new_read_requests5(struct stripe_head *sh, 1956 struct stripe_head_state *s, int disk_idx, int disks) 1957{ 1958 struct r5dev *dev = &sh->dev[disk_idx]; 1959 struct r5dev *failed_dev = &sh->dev[s->failed_num]; 1960 1961 /* don't schedule compute operations or reads on the parity block while 1962 * a check is in flight 1963 */ 1964 if ((disk_idx == sh->pd_idx) && 1965 test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) 1966 return ~0; 1967 1968 /* is the data in this block needed, and can we get it? */ 1969 if (!test_bit(R5_LOCKED, &dev->flags) && 1970 !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread || 1971 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) || 1972 s->syncing || s->expanding || (s->failed && 1973 (failed_dev->toread || (failed_dev->towrite && 1974 !test_bit(R5_OVERWRITE, &failed_dev->flags) 1975 ))))) { 1976 /* 1/ We would like to get this block, possibly by computing it, 1977 * but we might not be able to. 1978 * 1979 * 2/ Since parity check operations potentially make the parity 1980 * block !uptodate it will need to be refreshed before any 1981 * compute operations on data disks are scheduled. 1982 * 1983 * 3/ We hold off parity block re-reads until check operations 1984 * have quiesced. 1985 */ 1986 if ((s->uptodate == disks - 1) && 1987 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) { 1988 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending); 1989 set_bit(R5_Wantcompute, &dev->flags); 1990 sh->ops.target = disk_idx; 1991 s->req_compute = 1; 1992 sh->ops.count++; 1993 /* Careful: from this point on 'uptodate' is in the eye 1994 * of raid5_run_ops which services 'compute' operations 1995 * before writes. R5_Wantcompute flags a block that will 1996 * be R5_UPTODATE by the time it is needed for a 1997 * subsequent operation. 1998 */ 1999 s->uptodate++; 2000 return 0; /* uptodate + compute == disks */ 2001 } else if ((s->uptodate < disks - 1) && 2002 test_bit(R5_Insync, &dev->flags)) { 2003 /* Note: we hold off compute operations while checks are 2004 * in flight, but we still prefer 'compute' over 'read' 2005 * hence we only read if (uptodate < * disks-1) 2006 */ 2007 set_bit(R5_LOCKED, &dev->flags); 2008 set_bit(R5_Wantread, &dev->flags); 2009 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending)) 2010 sh->ops.count++; 2011 s->locked++; 2012 pr_debug("Reading block %d (sync=%d)\n", disk_idx, 2013 s->syncing); 2014 } 2015 } 2016 2017 return ~0; 2018} 2019 2020static void handle_issuing_new_read_requests5(struct stripe_head *sh, 2021 struct stripe_head_state *s, int disks) 2022{ 2023 int i; 2024 2025 /* Clear completed compute operations. Parity recovery 2026 * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled 2027 * later on in this routine 2028 */ 2029 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) && 2030 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) { 2031 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete); 2032 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack); 2033 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending); 2034 } 2035 2036 /* look for blocks to read/compute, skip this if a compute 2037 * is already in flight, or if the stripe contents are in the 2038 * midst of changing due to a write 2039 */ 2040 if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) && 2041 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) && 2042 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) { 2043 for (i = disks; i--; ) 2044 if (__handle_issuing_new_read_requests5( 2045 sh, s, i, disks) == 0) 2046 break; 2047 } 2048 set_bit(STRIPE_HANDLE, &sh->state); 2049} 2050 2051static void handle_issuing_new_read_requests6(struct stripe_head *sh, 2052 struct stripe_head_state *s, struct r6_state *r6s, 2053 int disks) 2054{ 2055 int i; 2056 for (i = disks; i--; ) { 2057 struct r5dev *dev = &sh->dev[i]; 2058 if (!test_bit(R5_LOCKED, &dev->flags) && 2059 !test_bit(R5_UPTODATE, &dev->flags) && 2060 (dev->toread || (dev->towrite && 2061 !test_bit(R5_OVERWRITE, &dev->flags)) || 2062 s->syncing || s->expanding || 2063 (s->failed >= 1 && 2064 (sh->dev[r6s->failed_num[0]].toread || 2065 s->to_write)) || 2066 (s->failed >= 2 && 2067 (sh->dev[r6s->failed_num[1]].toread || 2068 s->to_write)))) { 2069 /* we would like to get this block, possibly 2070 * by computing it, but we might not be able to 2071 */ 2072 if (s->uptodate == disks-1) { 2073 pr_debug("Computing stripe %llu block %d\n", 2074 (unsigned long long)sh->sector, i); 2075 compute_block_1(sh, i, 0); 2076 s->uptodate++; 2077 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) { 2078 /* Computing 2-failure is *very* expensive; only 2079 * do it if failed >= 2 2080 */ 2081 int other; 2082 for (other = disks; other--; ) { 2083 if (other == i) 2084 continue; 2085 if (!test_bit(R5_UPTODATE, 2086 &sh->dev[other].flags)) 2087 break; 2088 } 2089 BUG_ON(other < 0); 2090 pr_debug("Computing stripe %llu blocks %d,%d\n", 2091 (unsigned long long)sh->sector, 2092 i, other); 2093 compute_block_2(sh, i, other); 2094 s->uptodate += 2; 2095 } else if (test_bit(R5_Insync, &dev->flags)) { 2096 set_bit(R5_LOCKED, &dev->flags); 2097 set_bit(R5_Wantread, &dev->flags); 2098 s->locked++; 2099 pr_debug("Reading block %d (sync=%d)\n", 2100 i, s->syncing); 2101 } 2102 } 2103 } 2104 set_bit(STRIPE_HANDLE, &sh->state); 2105} 2106 2107 2108/* handle_completed_write_requests 2109 * any written block on an uptodate or failed drive can be returned. 2110 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but 2111 * never LOCKED, so we don't need to test 'failed' directly. 2112 */ 2113static void handle_completed_write_requests(raid5_conf_t *conf, 2114 struct stripe_head *sh, int disks, struct bio **return_bi) 2115{ 2116 int i; 2117 struct r5dev *dev; 2118 2119 for (i = disks; i--; ) 2120 if (sh->dev[i].written) { 2121 dev = &sh->dev[i]; 2122 if (!test_bit(R5_LOCKED, &dev->flags) && 2123 test_bit(R5_UPTODATE, &dev->flags)) { 2124 /* We can return any write requests */ 2125 struct bio *wbi, *wbi2; 2126 int bitmap_end = 0; 2127 pr_debug("Return write for disc %d\n", i); 2128 spin_lock_irq(&conf->device_lock); 2129 wbi = dev->written; 2130 dev->written = NULL; 2131 while (wbi && wbi->bi_sector < 2132 dev->sector + STRIPE_SECTORS) { 2133 wbi2 = r5_next_bio(wbi, dev->sector); 2134 if (--wbi->bi_phys_segments == 0) { 2135 md_write_end(conf->mddev); 2136 wbi->bi_next = *return_bi; 2137 *return_bi = wbi; 2138 } 2139 wbi = wbi2; 2140 } 2141 if (dev->towrite == NULL) 2142 bitmap_end = 1; 2143 spin_unlock_irq(&conf->device_lock); 2144 if (bitmap_end) 2145 bitmap_endwrite(conf->mddev->bitmap, 2146 sh->sector, 2147 STRIPE_SECTORS, 2148 !test_bit(STRIPE_DEGRADED, &sh->state), 2149 0); 2150 } 2151 } 2152} 2153 2154static void handle_issuing_new_write_requests5(raid5_conf_t *conf, 2155 struct stripe_head *sh, struct stripe_head_state *s, int disks) 2156{ 2157 int rmw = 0, rcw = 0, i; 2158 for (i = disks; i--; ) { 2159 /* would I have to read this buffer for read_modify_write */ 2160 struct r5dev *dev = &sh->dev[i]; 2161 if ((dev->towrite || i == sh->pd_idx) && 2162 !test_bit(R5_LOCKED, &dev->flags) && 2163 !(test_bit(R5_UPTODATE, &dev->flags) || 2164 test_bit(R5_Wantcompute, &dev->flags))) { 2165 if (test_bit(R5_Insync, &dev->flags)) 2166 rmw++; 2167 else 2168 rmw += 2*disks; /* cannot read it */ 2169 } 2170 /* Would I have to read this buffer for reconstruct_write */ 2171 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx && 2172 !test_bit(R5_LOCKED, &dev->flags) && 2173 !(test_bit(R5_UPTODATE, &dev->flags) || 2174 test_bit(R5_Wantcompute, &dev->flags))) { 2175 if (test_bit(R5_Insync, &dev->flags)) rcw++; 2176 else 2177 rcw += 2*disks; 2178 } 2179 } 2180 pr_debug("for sector %llu, rmw=%d rcw=%d\n", 2181 (unsigned long long)sh->sector, rmw, rcw); 2182 set_bit(STRIPE_HANDLE, &sh->state); 2183 if (rmw < rcw && rmw > 0) 2184 /* prefer read-modify-write, but need to get some data */ 2185 for (i = disks; i--; ) { 2186 struct r5dev *dev = &sh->dev[i]; 2187 if ((dev->towrite || i == sh->pd_idx) && 2188 !test_bit(R5_LOCKED, &dev->flags) && 2189 !(test_bit(R5_UPTODATE, &dev->flags) || 2190 test_bit(R5_Wantcompute, &dev->flags)) && 2191 test_bit(R5_Insync, &dev->flags)) { 2192 if ( 2193 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 2194 pr_debug("Read_old block " 2195 "%d for r-m-w\n", i); 2196 set_bit(R5_LOCKED, &dev->flags); 2197 set_bit(R5_Wantread, &dev->flags); 2198 if (!test_and_set_bit( 2199 STRIPE_OP_IO, &sh->ops.pending)) 2200 sh->ops.count++; 2201 s->locked++; 2202 } else { 2203 set_bit(STRIPE_DELAYED, &sh->state); 2204 set_bit(STRIPE_HANDLE, &sh->state); 2205 } 2206 } 2207 } 2208 if (rcw <= rmw && rcw > 0) 2209 /* want reconstruct write, but need to get some data */ 2210 for (i = disks; i--; ) { 2211 struct r5dev *dev = &sh->dev[i]; 2212 if (!test_bit(R5_OVERWRITE, &dev->flags) && 2213 i != sh->pd_idx && 2214 !test_bit(R5_LOCKED, &dev->flags) && 2215 !(test_bit(R5_UPTODATE, &dev->flags) || 2216 test_bit(R5_Wantcompute, &dev->flags)) && 2217 test_bit(R5_Insync, &dev->flags)) { 2218 if ( 2219 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 2220 pr_debug("Read_old block " 2221 "%d for Reconstruct\n", i); 2222 set_bit(R5_LOCKED, &dev->flags); 2223 set_bit(R5_Wantread, &dev->flags); 2224 if (!test_and_set_bit( 2225 STRIPE_OP_IO, &sh->ops.pending)) 2226 sh->ops.count++; 2227 s->locked++; 2228 } else { 2229 set_bit(STRIPE_DELAYED, &sh->state); 2230 set_bit(STRIPE_HANDLE, &sh->state); 2231 } 2232 } 2233 } 2234 /* now if nothing is locked, and if we have enough data, 2235 * we can start a write request 2236 */ 2237 /* since handle_stripe can be called at any time we need to handle the 2238 * case where a compute block operation has been submitted and then a 2239 * subsequent call wants to start a write request. raid5_run_ops only 2240 * handles the case where compute block and postxor are requested 2241 * simultaneously. If this is not the case then new writes need to be 2242 * held off until the compute completes. 2243 */ 2244 if ((s->req_compute || 2245 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) && 2246 (s->locked == 0 && (rcw == 0 || rmw == 0) && 2247 !test_bit(STRIPE_BIT_DELAY, &sh->state))) 2248 s->locked += handle_write_operations5(sh, rcw == 0, 0); 2249} 2250 2251static void handle_issuing_new_write_requests6(raid5_conf_t *conf, 2252 struct stripe_head *sh, struct stripe_head_state *s, 2253 struct r6_state *r6s, int disks) 2254{ 2255 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i; 2256 int qd_idx = r6s->qd_idx; 2257 for (i = disks; i--; ) { 2258 struct r5dev *dev = &sh->dev[i]; 2259 /* Would I have to read this buffer for reconstruct_write */ 2260 if (!test_bit(R5_OVERWRITE, &dev->flags) 2261 && i != pd_idx && i != qd_idx 2262 && (!test_bit(R5_LOCKED, &dev->flags) 2263 ) && 2264 !test_bit(R5_UPTODATE, &dev->flags)) { 2265 if (test_bit(R5_Insync, &dev->flags)) rcw++; 2266 else { 2267 pr_debug("raid6: must_compute: " 2268 "disk %d flags=%#lx\n", i, dev->flags); 2269 must_compute++; 2270 } 2271 } 2272 } 2273 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n", 2274 (unsigned long long)sh->sector, rcw, must_compute); 2275 set_bit(STRIPE_HANDLE, &sh->state); 2276 2277 if (rcw > 0) 2278 /* want reconstruct write, but need to get some data */ 2279 for (i = disks; i--; ) { 2280 struct r5dev *dev = &sh->dev[i]; 2281 if (!test_bit(R5_OVERWRITE, &dev->flags) 2282 && !(s->failed == 0 && (i == pd_idx || i == qd_idx)) 2283 && !test_bit(R5_LOCKED, &dev->flags) && 2284 !test_bit(R5_UPTODATE, &dev->flags) && 2285 test_bit(R5_Insync, &dev->flags)) { 2286 if ( 2287 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 2288 pr_debug("Read_old stripe %llu " 2289 "block %d for Reconstruct\n", 2290 (unsigned long long)sh->sector, i); 2291 set_bit(R5_LOCKED, &dev->flags); 2292 set_bit(R5_Wantread, &dev->flags); 2293 s->locked++; 2294 } else { 2295 pr_debug("Request delayed stripe %llu " 2296 "block %d for Reconstruct\n", 2297 (unsigned long long)sh->sector, i); 2298 set_bit(STRIPE_DELAYED, &sh->state); 2299 set_bit(STRIPE_HANDLE, &sh->state); 2300 } 2301 } 2302 } 2303 /* now if nothing is locked, and if we have enough data, we can start a 2304 * write request 2305 */ 2306 if (s->locked == 0 && rcw == 0 && 2307 !test_bit(STRIPE_BIT_DELAY, &sh->state)) { 2308 if (must_compute > 0) { 2309 /* We have failed blocks and need to compute them */ 2310 switch (s->failed) { 2311 case 0: 2312 BUG(); 2313 case 1: 2314 compute_block_1(sh, r6s->failed_num[0], 0); 2315 break; 2316 case 2: 2317 compute_block_2(sh, r6s->failed_num[0], 2318 r6s->failed_num[1]); 2319 break; 2320 default: /* This request should have been failed? */ 2321 BUG(); 2322 } 2323 } 2324 2325 pr_debug("Computing parity for stripe %llu\n", 2326 (unsigned long long)sh->sector); 2327 compute_parity6(sh, RECONSTRUCT_WRITE); 2328 /* now every locked buffer is ready to be written */ 2329 for (i = disks; i--; ) 2330 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) { 2331 pr_debug("Writing stripe %llu block %d\n", 2332 (unsigned long long)sh->sector, i); 2333 s->locked++; 2334 set_bit(R5_Wantwrite, &sh->dev[i].flags); 2335 } 2336 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */ 2337 set_bit(STRIPE_INSYNC, &sh->state); 2338 2339 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 2340 atomic_dec(&conf->preread_active_stripes); 2341 if (atomic_read(&conf->preread_active_stripes) < 2342 IO_THRESHOLD) 2343 md_wakeup_thread(conf->mddev->thread); 2344 } 2345 } 2346} 2347 2348static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh, 2349 struct stripe_head_state *s, int disks) 2350{ 2351 set_bit(STRIPE_HANDLE, &sh->state); 2352 /* Take one of the following actions: 2353 * 1/ start a check parity operation if (uptodate == disks) 2354 * 2/ finish a check parity operation and act on the result 2355 * 3/ skip to the writeback section if we previously 2356 * initiated a recovery operation 2357 */ 2358 if (s->failed == 0 && 2359 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) { 2360 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) { 2361 BUG_ON(s->uptodate != disks); 2362 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags); 2363 sh->ops.count++; 2364 s->uptodate--; 2365 } else if ( 2366 test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) { 2367 clear_bit(STRIPE_OP_CHECK, &sh->ops.ack); 2368 clear_bit(STRIPE_OP_CHECK, &sh->ops.pending); 2369 2370 if (sh->ops.zero_sum_result == 0) 2371 /* parity is correct (on disc, 2372 * not in buffer any more) 2373 */ 2374 set_bit(STRIPE_INSYNC, &sh->state); 2375 else { 2376 conf->mddev->resync_mismatches += 2377 STRIPE_SECTORS; 2378 if (test_bit( 2379 MD_RECOVERY_CHECK, &conf->mddev->recovery)) 2380 /* don't try to repair!! */ 2381 set_bit(STRIPE_INSYNC, &sh->state); 2382 else { 2383 set_bit(STRIPE_OP_COMPUTE_BLK, 2384 &sh->ops.pending); 2385 set_bit(STRIPE_OP_MOD_REPAIR_PD, 2386 &sh->ops.pending); 2387 set_bit(R5_Wantcompute, 2388 &sh->dev[sh->pd_idx].flags); 2389 sh->ops.target = sh->pd_idx; 2390 sh->ops.count++; 2391 s->uptodate++; 2392 } 2393 } 2394 } 2395 } 2396 2397 /* check if we can clear a parity disk reconstruct */ 2398 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) && 2399 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) { 2400 2401 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending); 2402 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete); 2403 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack); 2404 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending); 2405 } 2406 2407 /* Wait for check parity and compute block operations to complete 2408 * before write-back 2409 */ 2410 if (!test_bit(STRIPE_INSYNC, &sh->state) && 2411 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) && 2412 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) { 2413 struct r5dev *dev; 2414 /* either failed parity check, or recovery is happening */ 2415 if (s->failed == 0) 2416 s->failed_num = sh->pd_idx; 2417 dev = &sh->dev[s->failed_num]; 2418 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags)); 2419 BUG_ON(s->uptodate != disks); 2420 2421 set_bit(R5_LOCKED, &dev->flags); 2422 set_bit(R5_Wantwrite, &dev->flags); 2423 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending)) 2424 sh->ops.count++; 2425 2426 clear_bit(STRIPE_DEGRADED, &sh->state); 2427 s->locked++; 2428 set_bit(STRIPE_INSYNC, &sh->state); 2429 } 2430} 2431 2432 2433static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh, 2434 struct stripe_head_state *s, 2435 struct r6_state *r6s, struct page *tmp_page, 2436 int disks) 2437{ 2438 int update_p = 0, update_q = 0; 2439 struct r5dev *dev; 2440 int pd_idx = sh->pd_idx; 2441 int qd_idx = r6s->qd_idx; 2442 2443 set_bit(STRIPE_HANDLE, &sh->state); 2444 2445 BUG_ON(s->failed > 2); 2446 BUG_ON(s->uptodate < disks); 2447 /* Want to check and possibly repair P and Q. 2448 * However there could be one 'failed' device, in which 2449 * case we can only check one of them, possibly using the 2450 * other to generate missing data 2451 */ 2452 2453 /* If !tmp_page, we cannot do the calculations, 2454 * but as we have set STRIPE_HANDLE, we will soon be called 2455 * by stripe_handle with a tmp_page - just wait until then. 2456 */ 2457 if (tmp_page) { 2458 if (s->failed == r6s->q_failed) { 2459 /* The only possible failed device holds 'Q', so it 2460 * makes sense to check P (If anything else were failed, 2461 * we would have used P to recreate it). 2462 */ 2463 compute_block_1(sh, pd_idx, 1); 2464 if (!page_is_zero(sh->dev[pd_idx].page)) { 2465 compute_block_1(sh, pd_idx, 0); 2466 update_p = 1; 2467 } 2468 } 2469 if (!r6s->q_failed && s->failed < 2) { 2470 /* q is not failed, and we didn't use it to generate 2471 * anything, so it makes sense to check it 2472 */ 2473 memcpy(page_address(tmp_page), 2474 page_address(sh->dev[qd_idx].page), 2475 STRIPE_SIZE); 2476 compute_parity6(sh, UPDATE_PARITY); 2477 if (memcmp(page_address(tmp_page), 2478 page_address(sh->dev[qd_idx].page), 2479 STRIPE_SIZE) != 0) { 2480 clear_bit(STRIPE_INSYNC, &sh->state); 2481 update_q = 1; 2482 } 2483 } 2484 if (update_p || update_q) { 2485 conf->mddev->resync_mismatches += STRIPE_SECTORS; 2486 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) 2487 /* don't try to repair!! */ 2488 update_p = update_q = 0; 2489 } 2490 2491 /* now write out any block on a failed drive, 2492 * or P or Q if they need it 2493 */ 2494 2495 if (s->failed == 2) { 2496 dev = &sh->dev[r6s->failed_num[1]]; 2497 s->locked++; 2498 set_bit(R5_LOCKED, &dev->flags); 2499 set_bit(R5_Wantwrite, &dev->flags); 2500 } 2501 if (s->failed >= 1) { 2502 dev = &sh->dev[r6s->failed_num[0]]; 2503 s->locked++; 2504 set_bit(R5_LOCKED, &dev->flags); 2505 set_bit(R5_Wantwrite, &dev->flags); 2506 } 2507 2508 if (update_p) { 2509 dev = &sh->dev[pd_idx]; 2510 s->locked++; 2511 set_bit(R5_LOCKED, &dev->flags); 2512 set_bit(R5_Wantwrite, &dev->flags); 2513 } 2514 if (update_q) { 2515 dev = &sh->dev[qd_idx]; 2516 s->locked++; 2517 set_bit(R5_LOCKED, &dev->flags); 2518 set_bit(R5_Wantwrite, &dev->flags); 2519 } 2520 clear_bit(STRIPE_DEGRADED, &sh->state); 2521 2522 set_bit(STRIPE_INSYNC, &sh->state); 2523 } 2524} 2525 2526static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh, 2527 struct r6_state *r6s) 2528{ 2529 int i; 2530 2531 /* We have read all the blocks in this stripe and now we need to 2532 * copy some of them into a target stripe for expand. 2533 */ 2534 struct dma_async_tx_descriptor *tx = NULL; 2535 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state); 2536 for (i = 0; i < sh->disks; i++) 2537 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) { 2538 int dd_idx, pd_idx, j; 2539 struct stripe_head *sh2; 2540 2541 sector_t bn = compute_blocknr(sh, i); 2542 sector_t s = raid5_compute_sector(bn, conf->raid_disks, 2543 conf->raid_disks - 2544 conf->max_degraded, &dd_idx, 2545 &pd_idx, conf); 2546 sh2 = get_active_stripe(conf, s, conf->raid_disks, 2547 pd_idx, 1); 2548 if (sh2 == NULL) 2549 /* so far only the early blocks of this stripe 2550 * have been requested. When later blocks 2551 * get requested, we will try again 2552 */ 2553 continue; 2554 if (!test_bit(STRIPE_EXPANDING, &sh2->state) || 2555 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) { 2556 /* must have already done this block */ 2557 release_stripe(sh2); 2558 continue; 2559 } 2560 2561 /* place all the copies on one channel */ 2562 tx = async_memcpy(sh2->dev[dd_idx].page, 2563 sh->dev[i].page, 0, 0, STRIPE_SIZE, 2564 ASYNC_TX_DEP_ACK, tx, NULL, NULL); 2565 2566 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags); 2567 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags); 2568 for (j = 0; j < conf->raid_disks; j++) 2569 if (j != sh2->pd_idx && 2570 (!r6s || j != raid6_next_disk(sh2->pd_idx, 2571 sh2->disks)) && 2572 !test_bit(R5_Expanded, &sh2->dev[j].flags)) 2573 break; 2574 if (j == conf->raid_disks) { 2575 set_bit(STRIPE_EXPAND_READY, &sh2->state); 2576 set_bit(STRIPE_HANDLE, &sh2->state); 2577 } 2578 release_stripe(sh2); 2579 2580 } 2581 /* done submitting copies, wait for them to complete */ 2582 if (tx) { 2583 async_tx_ack(tx); 2584 dma_wait_for_async_tx(tx); 2585 } 2586} 2587 2588/* 2589 * handle_stripe - do things to a stripe. 2590 * 2591 * We lock the stripe and then examine the state of various bits 2592 * to see what needs to be done. 2593 * Possible results: 2594 * return some read request which now have data 2595 * return some write requests which are safely on disc 2596 * schedule a read on some buffers 2597 * schedule a write of some buffers 2598 * return confirmation of parity correctness 2599 * 2600 * buffers are taken off read_list or write_list, and bh_cache buffers 2601 * get BH_Lock set before the stripe lock is released. 2602 * 2603 */ 2604 2605static void handle_stripe5(struct stripe_head *sh) 2606{ 2607 raid5_conf_t *conf = sh->raid_conf; 2608 int disks = sh->disks, i; 2609 struct bio *return_bi = NULL; 2610 struct stripe_head_state s; 2611 struct r5dev *dev; 2612 unsigned long pending = 0; 2613 2614 memset(&s, 0, sizeof(s)); 2615 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d " 2616 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state, 2617 atomic_read(&sh->count), sh->pd_idx, 2618 sh->ops.pending, sh->ops.ack, sh->ops.complete); 2619 2620 spin_lock(&sh->lock); 2621 clear_bit(STRIPE_HANDLE, &sh->state); 2622 clear_bit(STRIPE_DELAYED, &sh->state); 2623 2624 s.syncing = test_bit(STRIPE_SYNCING, &sh->state); 2625 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state); 2626 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state); 2627 /* Now to look around and see what can be done */ 2628 2629 /* clean-up completed biofill operations */ 2630 if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) { 2631 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending); 2632 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack); 2633 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete); 2634 } 2635 2636 rcu_read_lock(); 2637 for (i=disks; i--; ) { 2638 mdk_rdev_t *rdev; 2639 struct r5dev *dev = &sh->dev[i]; 2640 clear_bit(R5_Insync, &dev->flags); 2641 2642 pr_debug("check %d: state 0x%lx toread %p read %p write %p " 2643 "written %p\n", i, dev->flags, dev->toread, dev->read, 2644 dev->towrite, dev->written); 2645 2646 /* maybe we can request a biofill operation 2647 * 2648 * new wantfill requests are only permitted while 2649 * STRIPE_OP_BIOFILL is clear 2650 */ 2651 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread && 2652 !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)) 2653 set_bit(R5_Wantfill, &dev->flags); 2654 2655 /* now count some things */ 2656 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++; 2657 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++; 2658 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++; 2659 2660 if (test_bit(R5_Wantfill, &dev->flags)) 2661 s.to_fill++; 2662 else if (dev->toread) 2663 s.to_read++; 2664 if (dev->towrite) { 2665 s.to_write++; 2666 if (!test_bit(R5_OVERWRITE, &dev->flags)) 2667 s.non_overwrite++; 2668 } 2669 if (dev->written) 2670 s.written++; 2671 rdev = rcu_dereference(conf->disks[i].rdev); 2672 if (!rdev || !test_bit(In_sync, &rdev->flags)) { 2673 /* The ReadError flag will just be confusing now */ 2674 clear_bit(R5_ReadError, &dev->flags); 2675 clear_bit(R5_ReWrite, &dev->flags); 2676 } 2677 if (!rdev || !test_bit(In_sync, &rdev->flags) 2678 || test_bit(R5_ReadError, &dev->flags)) { 2679 s.failed++; 2680 s.failed_num = i; 2681 } else 2682 set_bit(R5_Insync, &dev->flags); 2683 } 2684 rcu_read_unlock(); 2685 2686 if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)) 2687 sh->ops.count++; 2688 2689 pr_debug("locked=%d uptodate=%d to_read=%d" 2690 " to_write=%d failed=%d failed_num=%d\n", 2691 s.locked, s.uptodate, s.to_read, s.to_write, 2692 s.failed, s.failed_num); 2693 /* check if the array has lost two devices and, if so, some requests might 2694 * need to be failed 2695 */ 2696 if (s.failed > 1 && s.to_read+s.to_write+s.written) 2697 handle_requests_to_failed_array(conf, sh, &s, disks, 2698 &return_bi); 2699 if (s.failed > 1 && s.syncing) { 2700 md_done_sync(conf->mddev, STRIPE_SECTORS,0); 2701 clear_bit(STRIPE_SYNCING, &sh->state); 2702 s.syncing = 0; 2703 } 2704 2705 /* might be able to return some write requests if the parity block 2706 * is safe, or on a failed drive 2707 */ 2708 dev = &sh->dev[sh->pd_idx]; 2709 if ( s.written && 2710 ((test_bit(R5_Insync, &dev->flags) && 2711 !test_bit(R5_LOCKED, &dev->flags) && 2712 test_bit(R5_UPTODATE, &dev->flags)) || 2713 (s.failed == 1 && s.failed_num == sh->pd_idx))) 2714 handle_completed_write_requests(conf, sh, disks, &return_bi); 2715 2716 /* Now we might consider reading some blocks, either to check/generate 2717 * parity, or to satisfy requests 2718 * or to load a block that is being partially written. 2719 */ 2720 if (s.to_read || s.non_overwrite || 2721 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding || 2722 test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) 2723 handle_issuing_new_read_requests5(sh, &s, disks); 2724 2725 /* Now we check to see if any write operations have recently 2726 * completed 2727 */ 2728 2729 /* leave prexor set until postxor is done, allows us to distinguish 2730 * a rmw from a rcw during biodrain 2731 */ 2732 if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) && 2733 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) { 2734 2735 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete); 2736 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack); 2737 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending); 2738 2739 for (i = disks; i--; ) 2740 clear_bit(R5_Wantprexor, &sh->dev[i].flags); 2741 } 2742 2743 /* if only POSTXOR is set then this is an 'expand' postxor */ 2744 if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) && 2745 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) { 2746 2747 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete); 2748 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack); 2749 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending); 2750 2751 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete); 2752 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack); 2753 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending); 2754 2755 /* All the 'written' buffers and the parity block are ready to 2756 * be written back to disk 2757 */ 2758 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags)); 2759 for (i = disks; i--; ) { 2760 dev = &sh->dev[i]; 2761 if (test_bit(R5_LOCKED, &dev->flags) && 2762 (i == sh->pd_idx || dev->written)) { 2763 pr_debug("Writing block %d\n", i); 2764 set_bit(R5_Wantwrite, &dev->flags); 2765 if (!test_and_set_bit( 2766 STRIPE_OP_IO, &sh->ops.pending)) 2767 sh->ops.count++; 2768 if (!test_bit(R5_Insync, &dev->flags) || 2769 (i == sh->pd_idx && s.failed == 0)) 2770 set_bit(STRIPE_INSYNC, &sh->state); 2771 } 2772 } 2773 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 2774 atomic_dec(&conf->preread_active_stripes); 2775 if (atomic_read(&conf->preread_active_stripes) < 2776 IO_THRESHOLD) 2777 md_wakeup_thread(conf->mddev->thread); 2778 } 2779 } 2780 2781 /* Now to consider new write requests and what else, if anything 2782 * should be read. We do not handle new writes when: 2783 * 1/ A 'write' operation (copy+xor) is already in flight. 2784 * 2/ A 'check' operation is in flight, as it may clobber the parity 2785 * block. 2786 */ 2787 if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) && 2788 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) 2789 handle_issuing_new_write_requests5(conf, sh, &s, disks); 2790 2791 /* maybe we need to check and possibly fix the parity for this stripe 2792 * Any reads will already have been scheduled, so we just see if enough 2793 * data is available. The parity check is held off while parity 2794 * dependent operations are in flight. 2795 */ 2796 if ((s.syncing && s.locked == 0 && 2797 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) && 2798 !test_bit(STRIPE_INSYNC, &sh->state)) || 2799 test_bit(STRIPE_OP_CHECK, &sh->ops.pending) || 2800 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) 2801 handle_parity_checks5(conf, sh, &s, disks); 2802 2803 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) { 2804 md_done_sync(conf->mddev, STRIPE_SECTORS,1); 2805 clear_bit(STRIPE_SYNCING, &sh->state); 2806 } 2807 2808 /* If the failed drive is just a ReadError, then we might need to progress 2809 * the repair/check process 2810 */ 2811 if (s.failed == 1 && !conf->mddev->ro && 2812 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags) 2813 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags) 2814 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags) 2815 ) { 2816 dev = &sh->dev[s.failed_num]; 2817 if (!test_bit(R5_ReWrite, &dev->flags)) { 2818 set_bit(R5_Wantwrite, &dev->flags); 2819 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending)) 2820 sh->ops.count++; 2821 set_bit(R5_ReWrite, &dev->flags); 2822 set_bit(R5_LOCKED, &dev->flags); 2823 s.locked++; 2824 } else { 2825 /* let's read it back */ 2826 set_bit(R5_Wantread, &dev->flags); 2827 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending)) 2828 sh->ops.count++; 2829 set_bit(R5_LOCKED, &dev->flags); 2830 s.locked++; 2831 } 2832 } 2833 2834 /* Finish postxor operations initiated by the expansion 2835 * process 2836 */ 2837 if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) && 2838 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) { 2839 2840 clear_bit(STRIPE_EXPANDING, &sh->state); 2841 2842 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending); 2843 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack); 2844 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete); 2845 2846 for (i = conf->raid_disks; i--; ) { 2847 set_bit(R5_Wantwrite, &sh->dev[i].flags); 2848 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending)) 2849 sh->ops.count++; 2850 } 2851 } 2852 2853 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) && 2854 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) { 2855 /* Need to write out all blocks after computing parity */ 2856 sh->disks = conf->raid_disks; 2857 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, 2858 conf->raid_disks); 2859 s.locked += handle_write_operations5(sh, 1, 1); 2860 } else if (s.expanded && 2861 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) { 2862 clear_bit(STRIPE_EXPAND_READY, &sh->state); 2863 atomic_dec(&conf->reshape_stripes); 2864 wake_up(&conf->wait_for_overlap); 2865 md_done_sync(conf->mddev, STRIPE_SECTORS, 1); 2866 } 2867 2868 if (s.expanding && s.locked == 0 && 2869 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) 2870 handle_stripe_expansion(conf, sh, NULL); 2871 2872 if (sh->ops.count) 2873 pending = get_stripe_work(sh); 2874 2875 spin_unlock(&sh->lock); 2876 2877 if (pending) 2878 raid5_run_ops(sh, pending); 2879 2880 return_io(return_bi); 2881 2882} 2883 2884static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page) 2885{ 2886 raid6_conf_t *conf = sh->raid_conf; 2887 int disks = sh->disks; 2888 struct bio *return_bi = NULL; 2889 int i, pd_idx = sh->pd_idx; 2890 struct stripe_head_state s; 2891 struct r6_state r6s; 2892 struct r5dev *dev, *pdev, *qdev; 2893 2894 r6s.qd_idx = raid6_next_disk(pd_idx, disks); 2895 pr_debug("handling stripe %llu, state=%#lx cnt=%d, " 2896 "pd_idx=%d, qd_idx=%d\n", 2897 (unsigned long long)sh->sector, sh->state, 2898 atomic_read(&sh->count), pd_idx, r6s.qd_idx); 2899 memset(&s, 0, sizeof(s)); 2900 2901 spin_lock(&sh->lock); 2902 clear_bit(STRIPE_HANDLE, &sh->state); 2903 clear_bit(STRIPE_DELAYED, &sh->state); 2904 2905 s.syncing = test_bit(STRIPE_SYNCING, &sh->state); 2906 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state); 2907 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state); 2908 /* Now to look around and see what can be done */ 2909 2910 rcu_read_lock(); 2911 for (i=disks; i--; ) { 2912 mdk_rdev_t *rdev; 2913 dev = &sh->dev[i]; 2914 clear_bit(R5_Insync, &dev->flags); 2915 2916 pr_debug("check %d: state 0x%lx read %p write %p written %p\n", 2917 i, dev->flags, dev->toread, dev->towrite, dev->written); 2918 /* maybe we can reply to a read */ 2919 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) { 2920 struct bio *rbi, *rbi2; 2921 pr_debug("Return read for disc %d\n", i); 2922 spin_lock_irq(&conf->device_lock); 2923 rbi = dev->toread; 2924 dev->toread = NULL; 2925 if (test_and_clear_bit(R5_Overlap, &dev->flags)) 2926 wake_up(&conf->wait_for_overlap); 2927 spin_unlock_irq(&conf->device_lock); 2928 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) { 2929 copy_data(0, rbi, dev->page, dev->sector); 2930 rbi2 = r5_next_bio(rbi, dev->sector); 2931 spin_lock_irq(&conf->device_lock); 2932 if (--rbi->bi_phys_segments == 0) { 2933 rbi->bi_next = return_bi; 2934 return_bi = rbi; 2935 } 2936 spin_unlock_irq(&conf->device_lock); 2937 rbi = rbi2; 2938 } 2939 } 2940 2941 /* now count some things */ 2942 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++; 2943 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++; 2944 2945 2946 if (dev->toread) 2947 s.to_read++; 2948 if (dev->towrite) { 2949 s.to_write++; 2950 if (!test_bit(R5_OVERWRITE, &dev->flags)) 2951 s.non_overwrite++; 2952 } 2953 if (dev->written) 2954 s.written++; 2955 rdev = rcu_dereference(conf->disks[i].rdev); 2956 if (!rdev || !test_bit(In_sync, &rdev->flags)) { 2957 /* The ReadError flag will just be confusing now */ 2958 clear_bit(R5_ReadError, &dev->flags); 2959 clear_bit(R5_ReWrite, &dev->flags); 2960 } 2961 if (!rdev || !test_bit(In_sync, &rdev->flags) 2962 || test_bit(R5_ReadError, &dev->flags)) { 2963 if (s.failed < 2) 2964 r6s.failed_num[s.failed] = i; 2965 s.failed++; 2966 } else 2967 set_bit(R5_Insync, &dev->flags); 2968 } 2969 rcu_read_unlock(); 2970 pr_debug("locked=%d uptodate=%d to_read=%d" 2971 " to_write=%d failed=%d failed_num=%d,%d\n", 2972 s.locked, s.uptodate, s.to_read, s.to_write, s.failed, 2973 r6s.failed_num[0], r6s.failed_num[1]); 2974 /* check if the array has lost >2 devices and, if so, some requests 2975 * might need to be failed 2976 */ 2977 if (s.failed > 2 && s.to_read+s.to_write+s.written) 2978 handle_requests_to_failed_array(conf, sh, &s, disks, 2979 &return_bi); 2980 if (s.failed > 2 && s.syncing) { 2981 md_done_sync(conf->mddev, STRIPE_SECTORS,0); 2982 clear_bit(STRIPE_SYNCING, &sh->state); 2983 s.syncing = 0; 2984 } 2985 2986 /* 2987 * might be able to return some write requests if the parity blocks 2988 * are safe, or on a failed drive 2989 */ 2990 pdev = &sh->dev[pd_idx]; 2991 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx) 2992 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx); 2993 qdev = &sh->dev[r6s.qd_idx]; 2994 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx) 2995 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx); 2996 2997 if ( s.written && 2998 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags) 2999 && !test_bit(R5_LOCKED, &pdev->flags) 3000 && test_bit(R5_UPTODATE, &pdev->flags)))) && 3001 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags) 3002 && !test_bit(R5_LOCKED, &qdev->flags) 3003 && test_bit(R5_UPTODATE, &qdev->flags))))) 3004 handle_completed_write_requests(conf, sh, disks, &return_bi); 3005 3006 /* Now we might consider reading some blocks, either to check/generate 3007 * parity, or to satisfy requests 3008 * or to load a block that is being partially written. 3009 */ 3010 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) || 3011 (s.syncing && (s.uptodate < disks)) || s.expanding) 3012 handle_issuing_new_read_requests6(sh, &s, &r6s, disks); 3013 3014 /* now to consider writing and what else, if anything should be read */ 3015 if (s.to_write) 3016 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks); 3017 3018 /* maybe we need to check and possibly fix the parity for this stripe 3019 * Any reads will already have been scheduled, so we just see if enough 3020 * data is available 3021 */ 3022 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state)) 3023 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks); 3024 3025 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) { 3026 md_done_sync(conf->mddev, STRIPE_SECTORS,1); 3027 clear_bit(STRIPE_SYNCING, &sh->state); 3028 } 3029 3030 /* If the failed drives are just a ReadError, then we might need 3031 * to progress the repair/check process 3032 */ 3033 if (s.failed <= 2 && !conf->mddev->ro) 3034 for (i = 0; i < s.failed; i++) { 3035 dev = &sh->dev[r6s.failed_num[i]]; 3036 if (test_bit(R5_ReadError, &dev->flags) 3037 && !test_bit(R5_LOCKED, &dev->flags) 3038 && test_bit(R5_UPTODATE, &dev->flags) 3039 ) { 3040 if (!test_bit(R5_ReWrite, &dev->flags)) { 3041 set_bit(R5_Wantwrite, &dev->flags); 3042 set_bit(R5_ReWrite, &dev->flags); 3043 set_bit(R5_LOCKED, &dev->flags); 3044 } else { 3045 /* let's read it back */ 3046 set_bit(R5_Wantread, &dev->flags); 3047 set_bit(R5_LOCKED, &dev->flags); 3048 } 3049 } 3050 } 3051 3052 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) { 3053 /* Need to write out all blocks after computing P&Q */ 3054 sh->disks = conf->raid_disks; 3055 sh->pd_idx = stripe_to_pdidx(sh->sector, conf, 3056 conf->raid_disks); 3057 compute_parity6(sh, RECONSTRUCT_WRITE); 3058 for (i = conf->raid_disks ; i-- ; ) { 3059 set_bit(R5_LOCKED, &sh->dev[i].flags); 3060 s.locked++; 3061 set_bit(R5_Wantwrite, &sh->dev[i].flags); 3062 } 3063 clear_bit(STRIPE_EXPANDING, &sh->state); 3064 } else if (s.expanded) { 3065 clear_bit(STRIPE_EXPAND_READY, &sh->state); 3066 atomic_dec(&conf->reshape_stripes); 3067 wake_up(&conf->wait_for_overlap); 3068 md_done_sync(conf->mddev, STRIPE_SECTORS, 1); 3069 } 3070 3071 if (s.expanding && s.locked == 0 && 3072 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) 3073 handle_stripe_expansion(conf, sh, &r6s); 3074 3075 spin_unlock(&sh->lock); 3076 3077 return_io(return_bi); 3078 3079 for (i=disks; i-- ;) { 3080 int rw; 3081 struct bio *bi; 3082 mdk_rdev_t *rdev; 3083 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) 3084 rw = WRITE; 3085 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags)) 3086 rw = READ; 3087 else 3088 continue; 3089 3090 bi = &sh->dev[i].req; 3091 3092 bi->bi_rw = rw; 3093 if (rw == WRITE) 3094 bi->bi_end_io = raid5_end_write_request; 3095 else 3096 bi->bi_end_io = raid5_end_read_request; 3097 3098 rcu_read_lock(); 3099 rdev = rcu_dereference(conf->disks[i].rdev); 3100 if (rdev && test_bit(Faulty, &rdev->flags)) 3101 rdev = NULL; 3102 if (rdev) 3103 atomic_inc(&rdev->nr_pending); 3104 rcu_read_unlock(); 3105 3106 if (rdev) { 3107 if (s.syncing || s.expanding || s.expanded) 3108 md_sync_acct(rdev->bdev, STRIPE_SECTORS); 3109 3110 bi->bi_bdev = rdev->bdev; 3111 pr_debug("for %llu schedule op %ld on disc %d\n", 3112 (unsigned long long)sh->sector, bi->bi_rw, i); 3113 atomic_inc(&sh->count); 3114 bi->bi_sector = sh->sector + rdev->data_offset; 3115 bi->bi_flags = 1 << BIO_UPTODATE; 3116 bi->bi_vcnt = 1; 3117 bi->bi_max_vecs = 1; 3118 bi->bi_idx = 0; 3119 bi->bi_io_vec = &sh->dev[i].vec; 3120 bi->bi_io_vec[0].bv_len = STRIPE_SIZE; 3121 bi->bi_io_vec[0].bv_offset = 0; 3122 bi->bi_size = STRIPE_SIZE; 3123 bi->bi_next = NULL; 3124 if (rw == WRITE && 3125 test_bit(R5_ReWrite, &sh->dev[i].flags)) 3126 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors); 3127 generic_make_request(bi); 3128 } else { 3129 if (rw == WRITE) 3130 set_bit(STRIPE_DEGRADED, &sh->state); 3131 pr_debug("skip op %ld on disc %d for sector %llu\n", 3132 bi->bi_rw, i, (unsigned long long)sh->sector); 3133 clear_bit(R5_LOCKED, &sh->dev[i].flags); 3134 set_bit(STRIPE_HANDLE, &sh->state); 3135 } 3136 } 3137} 3138 3139static void handle_stripe(struct stripe_head *sh, struct page *tmp_page) 3140{ 3141 if (sh->raid_conf->level == 6) 3142 handle_stripe6(sh, tmp_page); 3143 else 3144 handle_stripe5(sh); 3145} 3146 3147 3148 3149static void raid5_activate_delayed(raid5_conf_t *conf) 3150{ 3151 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) { 3152 while (!list_empty(&conf->delayed_list)) { 3153 struct list_head *l = conf->delayed_list.next; 3154 struct stripe_head *sh; 3155 sh = list_entry(l, struct stripe_head, lru); 3156 list_del_init(l); 3157 clear_bit(STRIPE_DELAYED, &sh->state); 3158 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 3159 atomic_inc(&conf->preread_active_stripes); 3160 list_add_tail(&sh->lru, &conf->handle_list); 3161 } 3162 } 3163} 3164 3165static void activate_bit_delay(raid5_conf_t *conf) 3166{ 3167 /* device_lock is held */ 3168 struct list_head head; 3169 list_add(&head, &conf->bitmap_list); 3170 list_del_init(&conf->bitmap_list); 3171 while (!list_empty(&head)) { 3172 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru); 3173 list_del_init(&sh->lru); 3174 atomic_inc(&sh->count); 3175 __release_stripe(conf, sh); 3176 } 3177} 3178 3179static void unplug_slaves(mddev_t *mddev) 3180{ 3181 raid5_conf_t *conf = mddev_to_conf(mddev); 3182 int i; 3183 3184 rcu_read_lock(); 3185 for (i=0; i<mddev->raid_disks; i++) { 3186 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev); 3187 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) { 3188 struct request_queue *r_queue = bdev_get_queue(rdev->bdev); 3189 3190 atomic_inc(&rdev->nr_pending); 3191 rcu_read_unlock(); 3192 3193 blk_unplug(r_queue); 3194 3195 rdev_dec_pending(rdev, mddev); 3196 rcu_read_lock(); 3197 } 3198 } 3199 rcu_read_unlock(); 3200} 3201 3202static void raid5_unplug_device(struct request_queue *q) 3203{ 3204 mddev_t *mddev = q->queuedata; 3205 raid5_conf_t *conf = mddev_to_conf(mddev); 3206 unsigned long flags; 3207 3208 spin_lock_irqsave(&conf->device_lock, flags); 3209 3210 if (blk_remove_plug(q)) { 3211 conf->seq_flush++; 3212 raid5_activate_delayed(conf); 3213 } 3214 md_wakeup_thread(mddev->thread); 3215 3216 spin_unlock_irqrestore(&conf->device_lock, flags); 3217 3218 unplug_slaves(mddev); 3219} 3220 3221static int raid5_congested(void *data, int bits) 3222{ 3223 mddev_t *mddev = data; 3224 raid5_conf_t *conf = mddev_to_conf(mddev); 3225 3226 /* No difference between reads and writes. Just check 3227 * how busy the stripe_cache is 3228 */ 3229 if (conf->inactive_blocked) 3230 return 1; 3231 if (conf->quiesce) 3232 return 1; 3233 if (list_empty_careful(&conf->inactive_list)) 3234 return 1; 3235 3236 return 0; 3237} 3238 3239/* We want read requests to align with chunks where possible, 3240 * but write requests don't need to. 3241 */ 3242static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec) 3243{ 3244 mddev_t *mddev = q->queuedata; 3245 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev); 3246 int max; 3247 unsigned int chunk_sectors = mddev->chunk_size >> 9; 3248 unsigned int bio_sectors = bio->bi_size >> 9; 3249 3250 if (bio_data_dir(bio) == WRITE) 3251 return biovec->bv_len; /* always allow writes to be mergeable */ 3252 3253 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9; 3254 if (max < 0) max = 0; 3255 if (max <= biovec->bv_len && bio_sectors == 0) 3256 return biovec->bv_len; 3257 else 3258 return max; 3259} 3260 3261 3262static int in_chunk_boundary(mddev_t *mddev, struct bio *bio) 3263{ 3264 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev); 3265 unsigned int chunk_sectors = mddev->chunk_size >> 9; 3266 unsigned int bio_sectors = bio->bi_size >> 9; 3267 3268 return chunk_sectors >= 3269 ((sector & (chunk_sectors - 1)) + bio_sectors); 3270} 3271 3272/* 3273 * add bio to the retry LIFO ( in O(1) ... we are in interrupt ) 3274 * later sampled by raid5d. 3275 */ 3276static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf) 3277{ 3278 unsigned long flags; 3279 3280 spin_lock_irqsave(&conf->device_lock, flags); 3281 3282 bi->bi_next = conf->retry_read_aligned_list; 3283 conf->retry_read_aligned_list = bi; 3284 3285 spin_unlock_irqrestore(&conf->device_lock, flags); 3286 md_wakeup_thread(conf->mddev->thread); 3287} 3288 3289 3290static struct bio *remove_bio_from_retry(raid5_conf_t *conf) 3291{ 3292 struct bio *bi; 3293 3294 bi = conf->retry_read_aligned; 3295 if (bi) { 3296 conf->retry_read_aligned = NULL; 3297 return bi; 3298 } 3299 bi = conf->retry_read_aligned_list; 3300 if(bi) { 3301 conf->retry_read_aligned_list = bi->bi_next; 3302 bi->bi_next = NULL; 3303 bi->bi_phys_segments = 1; /* biased count of active stripes */ 3304 bi->bi_hw_segments = 0; /* count of processed stripes */ 3305 } 3306 3307 return bi; 3308} 3309 3310 3311/* 3312 * The "raid5_align_endio" should check if the read succeeded and if it 3313 * did, call bio_endio on the original bio (having bio_put the new bio 3314 * first). 3315 * If the read failed.. 3316 */ 3317static void raid5_align_endio(struct bio *bi, int error) 3318{ 3319 struct bio* raid_bi = bi->bi_private; 3320 mddev_t *mddev; 3321 raid5_conf_t *conf; 3322 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags); 3323 mdk_rdev_t *rdev; 3324 3325 bio_put(bi); 3326 3327 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata; 3328 conf = mddev_to_conf(mddev); 3329 rdev = (void*)raid_bi->bi_next; 3330 raid_bi->bi_next = NULL; 3331 3332 rdev_dec_pending(rdev, conf->mddev); 3333 3334 if (!error && uptodate) { 3335 bio_endio(raid_bi, 0); 3336 if (atomic_dec_and_test(&conf->active_aligned_reads)) 3337 wake_up(&conf->wait_for_stripe); 3338 return; 3339 } 3340 3341 3342 pr_debug("raid5_align_endio : io error...handing IO for a retry\n"); 3343 3344 add_bio_to_retry(raid_bi, conf); 3345} 3346 3347static int bio_fits_rdev(struct bio *bi) 3348{ 3349 struct request_queue *q = bdev_get_queue(bi->bi_bdev); 3350 3351 if ((bi->bi_size>>9) > q->max_sectors) 3352 return 0; 3353 blk_recount_segments(q, bi); 3354 if (bi->bi_phys_segments > q->max_phys_segments || 3355 bi->bi_hw_segments > q->max_hw_segments) 3356 return 0; 3357 3358 if (q->merge_bvec_fn) 3359 /* it's too hard to apply the merge_bvec_fn at this stage, 3360 * just just give up 3361 */ 3362 return 0; 3363 3364 return 1; 3365} 3366 3367 3368static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio) 3369{ 3370 mddev_t *mddev = q->queuedata; 3371 raid5_conf_t *conf = mddev_to_conf(mddev); 3372 const unsigned int raid_disks = conf->raid_disks; 3373 const unsigned int data_disks = raid_disks - conf->max_degraded; 3374 unsigned int dd_idx, pd_idx; 3375 struct bio* align_bi; 3376 mdk_rdev_t *rdev; 3377 3378 if (!in_chunk_boundary(mddev, raid_bio)) { 3379 pr_debug("chunk_aligned_read : non aligned\n"); 3380 return 0; 3381 } 3382 /* 3383 * use bio_clone to make a copy of the bio 3384 */ 3385 align_bi = bio_clone(raid_bio, GFP_NOIO); 3386 if (!align_bi) 3387 return 0; 3388 /* 3389 * set bi_end_io to a new function, and set bi_private to the 3390 * original bio. 3391 */ 3392 align_bi->bi_end_io = raid5_align_endio; 3393 align_bi->bi_private = raid_bio; 3394 /* 3395 * compute position 3396 */ 3397 align_bi->bi_sector = raid5_compute_sector(raid_bio->bi_sector, 3398 raid_disks, 3399 data_disks, 3400 &dd_idx, 3401 &pd_idx, 3402 conf); 3403 3404 rcu_read_lock(); 3405 rdev = rcu_dereference(conf->disks[dd_idx].rdev); 3406 if (rdev && test_bit(In_sync, &rdev->flags)) { 3407 atomic_inc(&rdev->nr_pending); 3408 rcu_read_unlock(); 3409 raid_bio->bi_next = (void*)rdev; 3410 align_bi->bi_bdev = rdev->bdev; 3411 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID); 3412 align_bi->bi_sector += rdev->data_offset; 3413 3414 if (!bio_fits_rdev(align_bi)) { 3415 /* too big in some way */ 3416 bio_put(align_bi); 3417 rdev_dec_pending(rdev, mddev); 3418 return 0; 3419 } 3420 3421 spin_lock_irq(&conf->device_lock); 3422 wait_event_lock_irq(conf->wait_for_stripe, 3423 conf->quiesce == 0, 3424 conf->device_lock, /* nothing */); 3425 atomic_inc(&conf->active_aligned_reads); 3426 spin_unlock_irq(&conf->device_lock); 3427 3428 generic_make_request(align_bi); 3429 return 1; 3430 } else { 3431 rcu_read_unlock(); 3432 bio_put(align_bi); 3433 return 0; 3434 } 3435} 3436 3437 3438static int make_request(struct request_queue *q, struct bio * bi) 3439{ 3440 mddev_t *mddev = q->queuedata; 3441 raid5_conf_t *conf = mddev_to_conf(mddev); 3442 unsigned int dd_idx, pd_idx; 3443 sector_t new_sector; 3444 sector_t logical_sector, last_sector; 3445 struct stripe_head *sh; 3446 const int rw = bio_data_dir(bi); 3447 int remaining; 3448 3449 if (unlikely(bio_barrier(bi))) { 3450 bio_endio(bi, -EOPNOTSUPP); 3451 return 0; 3452 } 3453 3454 md_write_start(mddev, bi); 3455 3456 disk_stat_inc(mddev->gendisk, ios[rw]); 3457 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi)); 3458 3459 if (rw == READ && 3460 mddev->reshape_position == MaxSector && 3461 chunk_aligned_read(q,bi)) 3462 return 0; 3463 3464 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1); 3465 last_sector = bi->bi_sector + (bi->bi_size>>9); 3466 bi->bi_next = NULL; 3467 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */ 3468 3469 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) { 3470 DEFINE_WAIT(w); 3471 int disks, data_disks; 3472 3473 retry: 3474 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE); 3475 if (likely(conf->expand_progress == MaxSector)) 3476 disks = conf->raid_disks; 3477 else { 3478 /* spinlock is needed as expand_progress may be 3479 * 64bit on a 32bit platform, and so it might be 3480 * possible to see a half-updated value 3481 * Ofcourse expand_progress could change after 3482 * the lock is dropped, so once we get a reference 3483 * to the stripe that we think it is, we will have 3484 * to check again. 3485 */ 3486 spin_lock_irq(&conf->device_lock); 3487 disks = conf->raid_disks; 3488 if (logical_sector >= conf->expand_progress) 3489 disks = conf->previous_raid_disks; 3490 else { 3491 if (logical_sector >= conf->expand_lo) { 3492 spin_unlock_irq(&conf->device_lock); 3493 schedule(); 3494 goto retry; 3495 } 3496 } 3497 spin_unlock_irq(&conf->device_lock); 3498 } 3499 data_disks = disks - conf->max_degraded; 3500 3501 new_sector = raid5_compute_sector(logical_sector, disks, data_disks, 3502 &dd_idx, &pd_idx, conf); 3503 pr_debug("raid5: make_request, sector %llu logical %llu\n", 3504 (unsigned long long)new_sector, 3505 (unsigned long long)logical_sector); 3506 3507 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK)); 3508 if (sh) { 3509 if (unlikely(conf->expand_progress != MaxSector)) { 3510 /* expansion might have moved on while waiting for a 3511 * stripe, so we must do the range check again. 3512 * Expansion could still move past after this 3513 * test, but as we are holding a reference to 3514 * 'sh', we know that if that happens, 3515 * STRIPE_EXPANDING will get set and the expansion 3516 * won't proceed until we finish with the stripe. 3517 */ 3518 int must_retry = 0; 3519 spin_lock_irq(&conf->device_lock); 3520 if (logical_sector < conf->expand_progress && 3521 disks == conf->previous_raid_disks) 3522 /* mismatch, need to try again */ 3523 must_retry = 1; 3524 spin_unlock_irq(&conf->device_lock); 3525 if (must_retry) { 3526 release_stripe(sh); 3527 goto retry; 3528 } 3529 } 3530 /* FIXME what if we get a false positive because these 3531 * are being updated. 3532 */ 3533 if (logical_sector >= mddev->suspend_lo && 3534 logical_sector < mddev->suspend_hi) { 3535 release_stripe(sh); 3536 schedule(); 3537 goto retry; 3538 } 3539 3540 if (test_bit(STRIPE_EXPANDING, &sh->state) || 3541 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) { 3542 /* Stripe is busy expanding or 3543 * add failed due to overlap. Flush everything 3544 * and wait a while 3545 */ 3546 raid5_unplug_device(mddev->queue); 3547 release_stripe(sh); 3548 schedule(); 3549 goto retry; 3550 } 3551 finish_wait(&conf->wait_for_overlap, &w); 3552 handle_stripe(sh, NULL); 3553 release_stripe(sh); 3554 } else { 3555 /* cannot get stripe for read-ahead, just give-up */ 3556 clear_bit(BIO_UPTODATE, &bi->bi_flags); 3557 finish_wait(&conf->wait_for_overlap, &w); 3558 break; 3559 } 3560 3561 } 3562 spin_lock_irq(&conf->device_lock); 3563 remaining = --bi->bi_phys_segments; 3564 spin_unlock_irq(&conf->device_lock); 3565 if (remaining == 0) { 3566 3567 if ( rw == WRITE ) 3568 md_write_end(mddev); 3569 3570 bi->bi_end_io(bi, 3571 test_bit(BIO_UPTODATE, &bi->bi_flags) 3572 ? 0 : -EIO); 3573 } 3574 return 0; 3575} 3576 3577static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped) 3578{ 3579 /* reshaping is quite different to recovery/resync so it is 3580 * handled quite separately ... here. 3581 * 3582 * On each call to sync_request, we gather one chunk worth of 3583 * destination stripes and flag them as expanding. 3584 * Then we find all the source stripes and request reads. 3585 * As the reads complete, handle_stripe will copy the data 3586 * into the destination stripe and release that stripe. 3587 */ 3588 raid5_conf_t *conf = (raid5_conf_t *) mddev->private; 3589 struct stripe_head *sh; 3590 int pd_idx; 3591 sector_t first_sector, last_sector; 3592 int raid_disks = conf->previous_raid_disks; 3593 int data_disks = raid_disks - conf->max_degraded; 3594 int new_data_disks = conf->raid_disks - conf->max_degraded; 3595 int i; 3596 int dd_idx; 3597 sector_t writepos, safepos, gap; 3598 3599 if (sector_nr == 0 && 3600 conf->expand_progress != 0) { 3601 /* restarting in the middle, skip the initial sectors */ 3602 sector_nr = conf->expand_progress; 3603 sector_div(sector_nr, new_data_disks); 3604 *skipped = 1; 3605 return sector_nr; 3606 } 3607 3608 /* we update the metadata when there is more than 3Meg 3609 * in the block range (that is rather arbitrary, should 3610 * probably be time based) or when the data about to be 3611 * copied would over-write the source of the data at 3612 * the front of the range. 3613 * i.e. one new_stripe forward from expand_progress new_maps 3614 * to after where expand_lo old_maps to 3615 */ 3616 writepos = conf->expand_progress + 3617 conf->chunk_size/512*(new_data_disks); 3618 sector_div(writepos, new_data_disks); 3619 safepos = conf->expand_lo; 3620 sector_div(safepos, data_disks); 3621 gap = conf->expand_progress - conf->expand_lo; 3622 3623 if (writepos >= safepos || 3624 gap > (new_data_disks)*3000*2 /*3Meg*/) { 3625 /* Cannot proceed until we've updated the superblock... */ 3626 wait_event(conf->wait_for_overlap, 3627 atomic_read(&conf->reshape_stripes)==0); 3628 mddev->reshape_position = conf->expand_progress; 3629 set_bit(MD_CHANGE_DEVS, &mddev->flags); 3630 md_wakeup_thread(mddev->thread); 3631 wait_event(mddev->sb_wait, mddev->flags == 0 || 3632 kthread_should_stop()); 3633 spin_lock_irq(&conf->device_lock); 3634 conf->expand_lo = mddev->reshape_position; 3635 spin_unlock_irq(&conf->device_lock); 3636 wake_up(&conf->wait_for_overlap); 3637 } 3638 3639 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) { 3640 int j; 3641 int skipped = 0; 3642 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks); 3643 sh = get_active_stripe(conf, sector_nr+i, 3644 conf->raid_disks, pd_idx, 0); 3645 set_bit(STRIPE_EXPANDING, &sh->state); 3646 atomic_inc(&conf->reshape_stripes); 3647 /* If any of this stripe is beyond the end of the old 3648 * array, then we need to zero those blocks 3649 */ 3650 for (j=sh->disks; j--;) { 3651 sector_t s; 3652 if (j == sh->pd_idx) 3653 continue; 3654 if (conf->level == 6 && 3655 j == raid6_next_disk(sh->pd_idx, sh->disks)) 3656 continue; 3657 s = compute_blocknr(sh, j); 3658 if (s < (mddev->array_size<<1)) { 3659 skipped = 1; 3660 continue; 3661 } 3662 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE); 3663 set_bit(R5_Expanded, &sh->dev[j].flags); 3664 set_bit(R5_UPTODATE, &sh->dev[j].flags); 3665 } 3666 if (!skipped) { 3667 set_bit(STRIPE_EXPAND_READY, &sh->state); 3668 set_bit(STRIPE_HANDLE, &sh->state); 3669 } 3670 release_stripe(sh); 3671 } 3672 spin_lock_irq(&conf->device_lock); 3673 conf->expand_progress = (sector_nr + i) * new_data_disks; 3674 spin_unlock_irq(&conf->device_lock); 3675 /* Ok, those stripe are ready. We can start scheduling 3676 * reads on the source stripes. 3677 * The source stripes are determined by mapping the first and last 3678 * block on the destination stripes. 3679 */ 3680 first_sector = 3681 raid5_compute_sector(sector_nr*(new_data_disks), 3682 raid_disks, data_disks, 3683 &dd_idx, &pd_idx, conf); 3684 last_sector = 3685 raid5_compute_sector((sector_nr+conf->chunk_size/512) 3686 *(new_data_disks) -1, 3687 raid_disks, data_disks, 3688 &dd_idx, &pd_idx, conf); 3689 if (last_sector >= (mddev->size<<1)) 3690 last_sector = (mddev->size<<1)-1; 3691 while (first_sector <= last_sector) { 3692 pd_idx = stripe_to_pdidx(first_sector, conf, 3693 conf->previous_raid_disks); 3694 sh = get_active_stripe(conf, first_sector, 3695 conf->previous_raid_disks, pd_idx, 0); 3696 set_bit(STRIPE_EXPAND_SOURCE, &sh->state); 3697 set_bit(STRIPE_HANDLE, &sh->state); 3698 release_stripe(sh); 3699 first_sector += STRIPE_SECTORS; 3700 } 3701 /* If this takes us to the resync_max point where we have to pause, 3702 * then we need to write out the superblock. 3703 */ 3704 sector_nr += conf->chunk_size>>9; 3705 if (sector_nr >= mddev->resync_max) { 3706 /* Cannot proceed until we've updated the superblock... */ 3707 wait_event(conf->wait_for_overlap, 3708 atomic_read(&conf->reshape_stripes) == 0); 3709 mddev->reshape_position = conf->expand_progress; 3710 set_bit(MD_CHANGE_DEVS, &mddev->flags); 3711 md_wakeup_thread(mddev->thread); 3712 wait_event(mddev->sb_wait, 3713 !test_bit(MD_CHANGE_DEVS, &mddev->flags) 3714 || kthread_should_stop()); 3715 spin_lock_irq(&conf->device_lock); 3716 conf->expand_lo = mddev->reshape_position; 3717 spin_unlock_irq(&conf->device_lock); 3718 wake_up(&conf->wait_for_overlap); 3719 } 3720 return conf->chunk_size>>9; 3721} 3722 3723/* FIXME go_faster isn't used */ 3724static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster) 3725{ 3726 raid5_conf_t *conf = (raid5_conf_t *) mddev->private; 3727 struct stripe_head *sh; 3728 int pd_idx; 3729 int raid_disks = conf->raid_disks; 3730 sector_t max_sector = mddev->size << 1; 3731 int sync_blocks; 3732 int still_degraded = 0; 3733 int i; 3734 3735 if (sector_nr >= max_sector) { 3736 /* just being told to finish up .. nothing much to do */ 3737 unplug_slaves(mddev); 3738 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) { 3739 end_reshape(conf); 3740 return 0; 3741 } 3742 3743 if (mddev->curr_resync < max_sector) /* aborted */ 3744 bitmap_end_sync(mddev->bitmap, mddev->curr_resync, 3745 &sync_blocks, 1); 3746 else /* completed sync */ 3747 conf->fullsync = 0; 3748 bitmap_close_sync(mddev->bitmap); 3749 3750 return 0; 3751 } 3752 3753 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) 3754 return reshape_request(mddev, sector_nr, skipped); 3755 3756 /* No need to check resync_max as we never do more than one 3757 * stripe, and as resync_max will always be on a chunk boundary, 3758 * if the check in md_do_sync didn't fire, there is no chance 3759 * of overstepping resync_max here 3760 */ 3761 3762 /* if there is too many failed drives and we are trying 3763 * to resync, then assert that we are finished, because there is 3764 * nothing we can do. 3765 */ 3766 if (mddev->degraded >= conf->max_degraded && 3767 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) { 3768 sector_t rv = (mddev->size << 1) - sector_nr; 3769 *skipped = 1; 3770 return rv; 3771 } 3772 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) && 3773 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && 3774 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) { 3775 /* we can skip this block, and probably more */ 3776 sync_blocks /= STRIPE_SECTORS; 3777 *skipped = 1; 3778 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */ 3779 } 3780 3781 3782 bitmap_cond_end_sync(mddev->bitmap, sector_nr); 3783 3784 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks); 3785 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1); 3786 if (sh == NULL) { 3787 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0); 3788 /* make sure we don't swamp the stripe cache if someone else 3789 * is trying to get access 3790 */ 3791 schedule_timeout_uninterruptible(1); 3792 } 3793 /* Need to check if array will still be degraded after recovery/resync 3794 * We don't need to check the 'failed' flag as when that gets set, 3795 * recovery aborts. 3796 */ 3797 for (i=0; i<mddev->raid_disks; i++) 3798 if (conf->disks[i].rdev == NULL) 3799 still_degraded = 1; 3800 3801 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded); 3802 3803 spin_lock(&sh->lock); 3804 set_bit(STRIPE_SYNCING, &sh->state); 3805 clear_bit(STRIPE_INSYNC, &sh->state); 3806 spin_unlock(&sh->lock); 3807 3808 handle_stripe(sh, NULL); 3809 release_stripe(sh); 3810 3811 return STRIPE_SECTORS; 3812} 3813 3814static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio) 3815{ 3816 /* We may not be able to submit a whole bio at once as there 3817 * may not be enough stripe_heads available. 3818 * We cannot pre-allocate enough stripe_heads as we may need 3819 * more than exist in the cache (if we allow ever large chunks). 3820 * So we do one stripe head at a time and record in 3821 * ->bi_hw_segments how many have been done. 3822 * 3823 * We *know* that this entire raid_bio is in one chunk, so 3824 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector. 3825 */ 3826 struct stripe_head *sh; 3827 int dd_idx, pd_idx; 3828 sector_t sector, logical_sector, last_sector; 3829 int scnt = 0; 3830 int remaining; 3831 int handled = 0; 3832 3833 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1); 3834 sector = raid5_compute_sector( logical_sector, 3835 conf->raid_disks, 3836 conf->raid_disks - conf->max_degraded, 3837 &dd_idx, 3838 &pd_idx, 3839 conf); 3840 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9); 3841 3842 for (; logical_sector < last_sector; 3843 logical_sector += STRIPE_SECTORS, 3844 sector += STRIPE_SECTORS, 3845 scnt++) { 3846 3847 if (scnt < raid_bio->bi_hw_segments) 3848 /* already done this stripe */ 3849 continue; 3850 3851 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1); 3852 3853 if (!sh) { 3854 /* failed to get a stripe - must wait */ 3855 raid_bio->bi_hw_segments = scnt; 3856 conf->retry_read_aligned = raid_bio; 3857 return handled; 3858 } 3859 3860 set_bit(R5_ReadError, &sh->dev[dd_idx].flags); 3861 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) { 3862 release_stripe(sh); 3863 raid_bio->bi_hw_segments = scnt; 3864 conf->retry_read_aligned = raid_bio; 3865 return handled; 3866 } 3867 3868 handle_stripe(sh, NULL); 3869 release_stripe(sh); 3870 handled++; 3871 } 3872 spin_lock_irq(&conf->device_lock); 3873 remaining = --raid_bio->bi_phys_segments; 3874 spin_unlock_irq(&conf->device_lock); 3875 if (remaining == 0) { 3876 3877 raid_bio->bi_end_io(raid_bio, 3878 test_bit(BIO_UPTODATE, &raid_bio->bi_flags) 3879 ? 0 : -EIO); 3880 } 3881 if (atomic_dec_and_test(&conf->active_aligned_reads)) 3882 wake_up(&conf->wait_for_stripe); 3883 return handled; 3884} 3885 3886 3887 3888/* 3889 * This is our raid5 kernel thread. 3890 * 3891 * We scan the hash table for stripes which can be handled now. 3892 * During the scan, completed stripes are saved for us by the interrupt 3893 * handler, so that they will not have to wait for our next wakeup. 3894 */ 3895static void raid5d (mddev_t *mddev) 3896{ 3897 struct stripe_head *sh; 3898 raid5_conf_t *conf = mddev_to_conf(mddev); 3899 int handled; 3900 3901 pr_debug("+++ raid5d active\n"); 3902 3903 md_check_recovery(mddev); 3904 3905 handled = 0; 3906 spin_lock_irq(&conf->device_lock); 3907 while (1) { 3908 struct list_head *first; 3909 struct bio *bio; 3910 3911 if (conf->seq_flush != conf->seq_write) { 3912 int seq = conf->seq_flush; 3913 spin_unlock_irq(&conf->device_lock); 3914 bitmap_unplug(mddev->bitmap); 3915 spin_lock_irq(&conf->device_lock); 3916 conf->seq_write = seq; 3917 activate_bit_delay(conf); 3918 } 3919 3920 if (list_empty(&conf->handle_list) && 3921 atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD && 3922 !blk_queue_plugged(mddev->queue) && 3923 !list_empty(&conf->delayed_list)) 3924 raid5_activate_delayed(conf); 3925 3926 while ((bio = remove_bio_from_retry(conf))) { 3927 int ok; 3928 spin_unlock_irq(&conf->device_lock); 3929 ok = retry_aligned_read(conf, bio); 3930 spin_lock_irq(&conf->device_lock); 3931 if (!ok) 3932 break; 3933 handled++; 3934 } 3935 3936 if (list_empty(&conf->handle_list)) { 3937 async_tx_issue_pending_all(); 3938 break; 3939 } 3940 3941 first = conf->handle_list.next; 3942 sh = list_entry(first, struct stripe_head, lru); 3943 3944 list_del_init(first); 3945 atomic_inc(&sh->count); 3946 BUG_ON(atomic_read(&sh->count)!= 1); 3947 spin_unlock_irq(&conf->device_lock); 3948 3949 handled++; 3950 handle_stripe(sh, conf->spare_page); 3951 release_stripe(sh); 3952 3953 spin_lock_irq(&conf->device_lock); 3954 } 3955 pr_debug("%d stripes handled\n", handled); 3956 3957 spin_unlock_irq(&conf->device_lock); 3958 3959 unplug_slaves(mddev); 3960 3961 pr_debug("--- raid5d inactive\n"); 3962} 3963 3964static ssize_t 3965raid5_show_stripe_cache_size(mddev_t *mddev, char *page) 3966{ 3967 raid5_conf_t *conf = mddev_to_conf(mddev); 3968 if (conf) 3969 return sprintf(page, "%d\n", conf->max_nr_stripes); 3970 else 3971 return 0; 3972} 3973 3974static ssize_t 3975raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len) 3976{ 3977 raid5_conf_t *conf = mddev_to_conf(mddev); 3978 char *end; 3979 int new; 3980 if (len >= PAGE_SIZE) 3981 return -EINVAL; 3982 if (!conf) 3983 return -ENODEV; 3984 3985 new = simple_strtoul(page, &end, 10); 3986 if (!*page || (*end && *end != '\n') ) 3987 return -EINVAL; 3988 if (new <= 16 || new > 32768) 3989 return -EINVAL; 3990 while (new < conf->max_nr_stripes) { 3991 if (drop_one_stripe(conf)) 3992 conf->max_nr_stripes--; 3993 else 3994 break; 3995 } 3996 md_allow_write(mddev); 3997 while (new > conf->max_nr_stripes) { 3998 if (grow_one_stripe(conf)) 3999 conf->max_nr_stripes++; 4000 else break; 4001 } 4002 return len; 4003} 4004 4005static struct md_sysfs_entry 4006raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR, 4007 raid5_show_stripe_cache_size, 4008 raid5_store_stripe_cache_size); 4009 4010static ssize_t 4011stripe_cache_active_show(mddev_t *mddev, char *page) 4012{ 4013 raid5_conf_t *conf = mddev_to_conf(mddev); 4014 if (conf) 4015 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes)); 4016 else 4017 return 0; 4018} 4019 4020static struct md_sysfs_entry 4021raid5_stripecache_active = __ATTR_RO(stripe_cache_active); 4022 4023static struct attribute *raid5_attrs[] = { 4024 &raid5_stripecache_size.attr, 4025 &raid5_stripecache_active.attr, 4026 NULL, 4027}; 4028static struct attribute_group raid5_attrs_group = { 4029 .name = NULL, 4030 .attrs = raid5_attrs, 4031}; 4032 4033static int run(mddev_t *mddev) 4034{ 4035 raid5_conf_t *conf; 4036 int raid_disk, memory; 4037 mdk_rdev_t *rdev; 4038 struct disk_info *disk; 4039 struct list_head *tmp; 4040 int working_disks = 0; 4041 4042 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) { 4043 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n", 4044 mdname(mddev), mddev->level); 4045 return -EIO; 4046 } 4047 4048 if (mddev->reshape_position != MaxSector) { 4049 /* Check that we can continue the reshape. 4050 * Currently only disks can change, it must 4051 * increase, and we must be past the point where 4052 * a stripe over-writes itself 4053 */ 4054 sector_t here_new, here_old; 4055 int old_disks; 4056 int max_degraded = (mddev->level == 5 ? 1 : 2); 4057 4058 if (mddev->new_level != mddev->level || 4059 mddev->new_layout != mddev->layout || 4060 mddev->new_chunk != mddev->chunk_size) { 4061 printk(KERN_ERR "raid5: %s: unsupported reshape " 4062 "required - aborting.\n", 4063 mdname(mddev)); 4064 return -EINVAL; 4065 } 4066 if (mddev->delta_disks <= 0) { 4067 printk(KERN_ERR "raid5: %s: unsupported reshape " 4068 "(reduce disks) required - aborting.\n", 4069 mdname(mddev)); 4070 return -EINVAL; 4071 } 4072 old_disks = mddev->raid_disks - mddev->delta_disks; 4073 /* reshape_position must be on a new-stripe boundary, and one 4074 * further up in new geometry must map after here in old 4075 * geometry. 4076 */ 4077 here_new = mddev->reshape_position; 4078 if (sector_div(here_new, (mddev->chunk_size>>9)* 4079 (mddev->raid_disks - max_degraded))) { 4080 printk(KERN_ERR "raid5: reshape_position not " 4081 "on a stripe boundary\n"); 4082 return -EINVAL; 4083 } 4084 /* here_new is the stripe we will write to */ 4085 here_old = mddev->reshape_position; 4086 sector_div(here_old, (mddev->chunk_size>>9)* 4087 (old_disks-max_degraded)); 4088 /* here_old is the first stripe that we might need to read 4089 * from */ 4090 if (here_new >= here_old) { 4091 /* Reading from the same stripe as writing to - bad */ 4092 printk(KERN_ERR "raid5: reshape_position too early for " 4093 "auto-recovery - aborting.\n"); 4094 return -EINVAL; 4095 } 4096 printk(KERN_INFO "raid5: reshape will continue\n"); 4097 /* OK, we should be able to continue; */ 4098 } 4099 4100 4101 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL); 4102 if ((conf = mddev->private) == NULL) 4103 goto abort; 4104 if (mddev->reshape_position == MaxSector) { 4105 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks; 4106 } else { 4107 conf->raid_disks = mddev->raid_disks; 4108 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks; 4109 } 4110 4111 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info), 4112 GFP_KERNEL); 4113 if (!conf->disks) 4114 goto abort; 4115 4116 conf->mddev = mddev; 4117 4118 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL) 4119 goto abort; 4120 4121 if (mddev->level == 6) { 4122 conf->spare_page = alloc_page(GFP_KERNEL); 4123 if (!conf->spare_page) 4124 goto abort; 4125 } 4126 spin_lock_init(&conf->device_lock); 4127 init_waitqueue_head(&conf->wait_for_stripe); 4128 init_waitqueue_head(&conf->wait_for_overlap); 4129 INIT_LIST_HEAD(&conf->handle_list); 4130 INIT_LIST_HEAD(&conf->delayed_list); 4131 INIT_LIST_HEAD(&conf->bitmap_list); 4132 INIT_LIST_HEAD(&conf->inactive_list); 4133 atomic_set(&conf->active_stripes, 0); 4134 atomic_set(&conf->preread_active_stripes, 0); 4135 atomic_set(&conf->active_aligned_reads, 0); 4136 4137 pr_debug("raid5: run(%s) called.\n", mdname(mddev)); 4138 4139 rdev_for_each(rdev, tmp, mddev) { 4140 raid_disk = rdev->raid_disk; 4141 if (raid_disk >= conf->raid_disks 4142 || raid_disk < 0) 4143 continue; 4144 disk = conf->disks + raid_disk; 4145 4146 disk->rdev = rdev; 4147 4148 if (test_bit(In_sync, &rdev->flags)) { 4149 char b[BDEVNAME_SIZE]; 4150 printk(KERN_INFO "raid5: device %s operational as raid" 4151 " disk %d\n", bdevname(rdev->bdev,b), 4152 raid_disk); 4153 working_disks++; 4154 } 4155 } 4156 4157 /* 4158 * 0 for a fully functional array, 1 or 2 for a degraded array. 4159 */ 4160 mddev->degraded = conf->raid_disks - working_disks; 4161 conf->mddev = mddev; 4162 conf->chunk_size = mddev->chunk_size; 4163 conf->level = mddev->level; 4164 if (conf->level == 6) 4165 conf->max_degraded = 2; 4166 else 4167 conf->max_degraded = 1; 4168 conf->algorithm = mddev->layout; 4169 conf->max_nr_stripes = NR_STRIPES; 4170 conf->expand_progress = mddev->reshape_position; 4171 4172 /* device size must be a multiple of chunk size */ 4173 mddev->size &= ~(mddev->chunk_size/1024 -1); 4174 mddev->resync_max_sectors = mddev->size << 1; 4175 4176 if (conf->level == 6 && conf->raid_disks < 4) { 4177 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n", 4178 mdname(mddev), conf->raid_disks); 4179 goto abort; 4180 } 4181 if (!conf->chunk_size || conf->chunk_size % 4) { 4182 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n", 4183 conf->chunk_size, mdname(mddev)); 4184 goto abort; 4185 } 4186 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) { 4187 printk(KERN_ERR 4188 "raid5: unsupported parity algorithm %d for %s\n", 4189 conf->algorithm, mdname(mddev)); 4190 goto abort; 4191 } 4192 if (mddev->degraded > conf->max_degraded) { 4193 printk(KERN_ERR "raid5: not enough operational devices for %s" 4194 " (%d/%d failed)\n", 4195 mdname(mddev), mddev->degraded, conf->raid_disks); 4196 goto abort; 4197 } 4198 4199 if (mddev->degraded > 0 && 4200 mddev->recovery_cp != MaxSector) { 4201 if (mddev->ok_start_degraded) 4202 printk(KERN_WARNING 4203 "raid5: starting dirty degraded array: %s" 4204 "- data corruption possible.\n", 4205 mdname(mddev)); 4206 else { 4207 printk(KERN_ERR 4208 "raid5: cannot start dirty degraded array for %s\n", 4209 mdname(mddev)); 4210 goto abort; 4211 } 4212 } 4213 4214 { 4215 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5"); 4216 if (!mddev->thread) { 4217 printk(KERN_ERR 4218 "raid5: couldn't allocate thread for %s\n", 4219 mdname(mddev)); 4220 goto abort; 4221 } 4222 } 4223 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) + 4224 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024; 4225 if (grow_stripes(conf, conf->max_nr_stripes)) { 4226 printk(KERN_ERR 4227 "raid5: couldn't allocate %dkB for buffers\n", memory); 4228 shrink_stripes(conf); 4229 md_unregister_thread(mddev->thread); 4230 goto abort; 4231 } else 4232 printk(KERN_INFO "raid5: allocated %dkB for %s\n", 4233 memory, mdname(mddev)); 4234 4235 if (mddev->degraded == 0) 4236 printk("raid5: raid level %d set %s active with %d out of %d" 4237 " devices, algorithm %d\n", conf->level, mdname(mddev), 4238 mddev->raid_disks-mddev->degraded, mddev->raid_disks, 4239 conf->algorithm); 4240 else 4241 printk(KERN_ALERT "raid5: raid level %d set %s active with %d" 4242 " out of %d devices, algorithm %d\n", conf->level, 4243 mdname(mddev), mddev->raid_disks - mddev->degraded, 4244 mddev->raid_disks, conf->algorithm); 4245 4246 print_raid5_conf(conf); 4247 4248 if (conf->expand_progress != MaxSector) { 4249 printk("...ok start reshape thread\n"); 4250 conf->expand_lo = conf->expand_progress; 4251 atomic_set(&conf->reshape_stripes, 0); 4252 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery); 4253 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery); 4254 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery); 4255 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery); 4256 mddev->sync_thread = md_register_thread(md_do_sync, mddev, 4257 "%s_reshape"); 4258 } 4259 4260 /* read-ahead size must cover two whole stripes, which is 4261 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices 4262 */ 4263 { 4264 int data_disks = conf->previous_raid_disks - conf->max_degraded; 4265 int stripe = data_disks * 4266 (mddev->chunk_size / PAGE_SIZE); 4267 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe) 4268 mddev->queue->backing_dev_info.ra_pages = 2 * stripe; 4269 } 4270 4271 /* Ok, everything is just fine now */ 4272 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group)) 4273 printk(KERN_WARNING 4274 "raid5: failed to create sysfs attributes for %s\n", 4275 mdname(mddev)); 4276 4277 mddev->queue->unplug_fn = raid5_unplug_device; 4278 mddev->queue->backing_dev_info.congested_data = mddev; 4279 mddev->queue->backing_dev_info.congested_fn = raid5_congested; 4280 4281 mddev->array_size = mddev->size * (conf->previous_raid_disks - 4282 conf->max_degraded); 4283 4284 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec); 4285 4286 return 0; 4287abort: 4288 if (conf) { 4289 print_raid5_conf(conf); 4290 safe_put_page(conf->spare_page); 4291 kfree(conf->disks); 4292 kfree(conf->stripe_hashtbl); 4293 kfree(conf); 4294 } 4295 mddev->private = NULL; 4296 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev)); 4297 return -EIO; 4298} 4299 4300 4301 4302static int stop(mddev_t *mddev) 4303{ 4304 raid5_conf_t *conf = (raid5_conf_t *) mddev->private; 4305 4306 md_unregister_thread(mddev->thread); 4307 mddev->thread = NULL; 4308 shrink_stripes(conf); 4309 kfree(conf->stripe_hashtbl); 4310 mddev->queue->backing_dev_info.congested_fn = NULL; 4311 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/ 4312 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group); 4313 kfree(conf->disks); 4314 kfree(conf); 4315 mddev->private = NULL; 4316 return 0; 4317} 4318 4319#ifdef DEBUG 4320static void print_sh (struct seq_file *seq, struct stripe_head *sh) 4321{ 4322 int i; 4323 4324 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n", 4325 (unsigned long long)sh->sector, sh->pd_idx, sh->state); 4326 seq_printf(seq, "sh %llu, count %d.\n", 4327 (unsigned long long)sh->sector, atomic_read(&sh->count)); 4328 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector); 4329 for (i = 0; i < sh->disks; i++) { 4330 seq_printf(seq, "(cache%d: %p %ld) ", 4331 i, sh->dev[i].page, sh->dev[i].flags); 4332 } 4333 seq_printf(seq, "\n"); 4334} 4335 4336static void printall (struct seq_file *seq, raid5_conf_t *conf) 4337{ 4338 struct stripe_head *sh; 4339 struct hlist_node *hn; 4340 int i; 4341 4342 spin_lock_irq(&conf->device_lock); 4343 for (i = 0; i < NR_HASH; i++) { 4344 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) { 4345 if (sh->raid_conf != conf) 4346 continue; 4347 print_sh(seq, sh); 4348 } 4349 } 4350 spin_unlock_irq(&conf->device_lock); 4351} 4352#endif 4353 4354static void status (struct seq_file *seq, mddev_t *mddev) 4355{ 4356 raid5_conf_t *conf = (raid5_conf_t *) mddev->private; 4357 int i; 4358 4359 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout); 4360 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded); 4361 for (i = 0; i < conf->raid_disks; i++) 4362 seq_printf (seq, "%s", 4363 conf->disks[i].rdev && 4364 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_"); 4365 seq_printf (seq, "]"); 4366#ifdef DEBUG 4367 seq_printf (seq, "\n"); 4368 printall(seq, conf); 4369#endif 4370} 4371 4372static void print_raid5_conf (raid5_conf_t *conf) 4373{ 4374 int i; 4375 struct disk_info *tmp; 4376 4377 printk("RAID5 conf printout:\n"); 4378 if (!conf) { 4379 printk("(conf==NULL)\n"); 4380 return; 4381 } 4382 printk(" --- rd:%d wd:%d\n", conf->raid_disks, 4383 conf->raid_disks - conf->mddev->degraded); 4384 4385 for (i = 0; i < conf->raid_disks; i++) { 4386 char b[BDEVNAME_SIZE]; 4387 tmp = conf->disks + i; 4388 if (tmp->rdev) 4389 printk(" disk %d, o:%d, dev:%s\n", 4390 i, !test_bit(Faulty, &tmp->rdev->flags), 4391 bdevname(tmp->rdev->bdev,b)); 4392 } 4393} 4394 4395static int raid5_spare_active(mddev_t *mddev) 4396{ 4397 int i; 4398 raid5_conf_t *conf = mddev->private; 4399 struct disk_info *tmp; 4400 4401 for (i = 0; i < conf->raid_disks; i++) { 4402 tmp = conf->disks + i; 4403 if (tmp->rdev 4404 && !test_bit(Faulty, &tmp->rdev->flags) 4405 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) { 4406 unsigned long flags; 4407 spin_lock_irqsave(&conf->device_lock, flags); 4408 mddev->degraded--; 4409 spin_unlock_irqrestore(&conf->device_lock, flags); 4410 } 4411 } 4412 print_raid5_conf(conf); 4413 return 0; 4414} 4415 4416static int raid5_remove_disk(mddev_t *mddev, int number) 4417{ 4418 raid5_conf_t *conf = mddev->private; 4419 int err = 0; 4420 mdk_rdev_t *rdev; 4421 struct disk_info *p = conf->disks + number; 4422 4423 print_raid5_conf(conf); 4424 rdev = p->rdev; 4425 if (rdev) { 4426 if (test_bit(In_sync, &rdev->flags) || 4427 atomic_read(&rdev->nr_pending)) { 4428 err = -EBUSY; 4429 goto abort; 4430 } 4431 p->rdev = NULL; 4432 synchronize_rcu(); 4433 if (atomic_read(&rdev->nr_pending)) { 4434 /* lost the race, try later */ 4435 err = -EBUSY; 4436 p->rdev = rdev; 4437 } 4438 } 4439abort: 4440 4441 print_raid5_conf(conf); 4442 return err; 4443} 4444 4445static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev) 4446{ 4447 raid5_conf_t *conf = mddev->private; 4448 int found = 0; 4449 int disk; 4450 struct disk_info *p; 4451 4452 if (mddev->degraded > conf->max_degraded) 4453 /* no point adding a device */ 4454 return 0; 4455 4456 /* 4457 * find the disk ... but prefer rdev->saved_raid_disk 4458 * if possible. 4459 */ 4460 if (rdev->saved_raid_disk >= 0 && 4461 conf->disks[rdev->saved_raid_disk].rdev == NULL) 4462 disk = rdev->saved_raid_disk; 4463 else 4464 disk = 0; 4465 for ( ; disk < conf->raid_disks; disk++) 4466 if ((p=conf->disks + disk)->rdev == NULL) { 4467 clear_bit(In_sync, &rdev->flags); 4468 rdev->raid_disk = disk; 4469 found = 1; 4470 if (rdev->saved_raid_disk != disk) 4471 conf->fullsync = 1; 4472 rcu_assign_pointer(p->rdev, rdev); 4473 break; 4474 } 4475 print_raid5_conf(conf); 4476 return found; 4477} 4478 4479static int raid5_resize(mddev_t *mddev, sector_t sectors) 4480{ 4481 /* no resync is happening, and there is enough space 4482 * on all devices, so we can resize. 4483 * We need to make sure resync covers any new space. 4484 * If the array is shrinking we should possibly wait until 4485 * any io in the removed space completes, but it hardly seems 4486 * worth it. 4487 */ 4488 raid5_conf_t *conf = mddev_to_conf(mddev); 4489 4490 sectors &= ~((sector_t)mddev->chunk_size/512 - 1); 4491 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1; 4492 set_capacity(mddev->gendisk, mddev->array_size << 1); 4493 mddev->changed = 1; 4494 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) { 4495 mddev->recovery_cp = mddev->size << 1; 4496 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 4497 } 4498 mddev->size = sectors /2; 4499 mddev->resync_max_sectors = sectors; 4500 return 0; 4501} 4502 4503#ifdef CONFIG_MD_RAID5_RESHAPE 4504static int raid5_check_reshape(mddev_t *mddev) 4505{ 4506 raid5_conf_t *conf = mddev_to_conf(mddev); 4507 int err; 4508 4509 if (mddev->delta_disks < 0 || 4510 mddev->new_level != mddev->level) 4511 return -EINVAL; /* Cannot shrink array or change level yet */ 4512 if (mddev->delta_disks == 0) 4513 return 0; /* nothing to do */ 4514 4515 /* Can only proceed if there are plenty of stripe_heads. 4516 * We need a minimum of one full stripe,, and for sensible progress 4517 * it is best to have about 4 times that. 4518 * If we require 4 times, then the default 256 4K stripe_heads will 4519 * allow for chunk sizes up to 256K, which is probably OK. 4520 * If the chunk size is greater, user-space should request more 4521 * stripe_heads first. 4522 */ 4523 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes || 4524 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) { 4525 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n", 4526 (mddev->chunk_size / STRIPE_SIZE)*4); 4527 return -ENOSPC; 4528 } 4529 4530 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks); 4531 if (err) 4532 return err; 4533 4534 if (mddev->degraded > conf->max_degraded) 4535 return -EINVAL; 4536 /* looks like we might be able to manage this */ 4537 return 0; 4538} 4539 4540static int raid5_start_reshape(mddev_t *mddev) 4541{ 4542 raid5_conf_t *conf = mddev_to_conf(mddev); 4543 mdk_rdev_t *rdev; 4544 struct list_head *rtmp; 4545 int spares = 0; 4546 int added_devices = 0; 4547 unsigned long flags; 4548 4549 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery)) 4550 return -EBUSY; 4551 4552 rdev_for_each(rdev, rtmp, mddev) 4553 if (rdev->raid_disk < 0 && 4554 !test_bit(Faulty, &rdev->flags)) 4555 spares++; 4556 4557 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded) 4558 /* Not enough devices even to make a degraded array 4559 * of that size 4560 */ 4561 return -EINVAL; 4562 4563 atomic_set(&conf->reshape_stripes, 0); 4564 spin_lock_irq(&conf->device_lock); 4565 conf->previous_raid_disks = conf->raid_disks; 4566 conf->raid_disks += mddev->delta_disks; 4567 conf->expand_progress = 0; 4568 conf->expand_lo = 0; 4569 spin_unlock_irq(&conf->device_lock); 4570 4571 /* Add some new drives, as many as will fit. 4572 * We know there are enough to make the newly sized array work. 4573 */ 4574 rdev_for_each(rdev, rtmp, mddev) 4575 if (rdev->raid_disk < 0 && 4576 !test_bit(Faulty, &rdev->flags)) { 4577 if (raid5_add_disk(mddev, rdev)) { 4578 char nm[20]; 4579 set_bit(In_sync, &rdev->flags); 4580 added_devices++; 4581 rdev->recovery_offset = 0; 4582 sprintf(nm, "rd%d", rdev->raid_disk); 4583 if (sysfs_create_link(&mddev->kobj, 4584 &rdev->kobj, nm)) 4585 printk(KERN_WARNING 4586 "raid5: failed to create " 4587 " link %s for %s\n", 4588 nm, mdname(mddev)); 4589 } else 4590 break; 4591 } 4592 4593 spin_lock_irqsave(&conf->device_lock, flags); 4594 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices; 4595 spin_unlock_irqrestore(&conf->device_lock, flags); 4596 mddev->raid_disks = conf->raid_disks; 4597 mddev->reshape_position = 0; 4598 set_bit(MD_CHANGE_DEVS, &mddev->flags); 4599 4600 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery); 4601 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery); 4602 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery); 4603 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery); 4604 mddev->sync_thread = md_register_thread(md_do_sync, mddev, 4605 "%s_reshape"); 4606 if (!mddev->sync_thread) { 4607 mddev->recovery = 0; 4608 spin_lock_irq(&conf->device_lock); 4609 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks; 4610 conf->expand_progress = MaxSector; 4611 spin_unlock_irq(&conf->device_lock); 4612 return -EAGAIN; 4613 } 4614 md_wakeup_thread(mddev->sync_thread); 4615 md_new_event(mddev); 4616 return 0; 4617} 4618#endif 4619 4620static void end_reshape(raid5_conf_t *conf) 4621{ 4622 struct block_device *bdev; 4623 4624 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) { 4625 conf->mddev->array_size = conf->mddev->size * 4626 (conf->raid_disks - conf->max_degraded); 4627 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1); 4628 conf->mddev->changed = 1; 4629 4630 bdev = bdget_disk(conf->mddev->gendisk, 0); 4631 if (bdev) { 4632 mutex_lock(&bdev->bd_inode->i_mutex); 4633 i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10); 4634 mutex_unlock(&bdev->bd_inode->i_mutex); 4635 bdput(bdev); 4636 } 4637 spin_lock_irq(&conf->device_lock); 4638 conf->expand_progress = MaxSector; 4639 spin_unlock_irq(&conf->device_lock); 4640 conf->mddev->reshape_position = MaxSector; 4641 4642 /* read-ahead size must cover two whole stripes, which is 4643 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices 4644 */ 4645 { 4646 int data_disks = conf->previous_raid_disks - conf->max_degraded; 4647 int stripe = data_disks * 4648 (conf->mddev->chunk_size / PAGE_SIZE); 4649 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe) 4650 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe; 4651 } 4652 } 4653} 4654 4655static void raid5_quiesce(mddev_t *mddev, int state) 4656{ 4657 raid5_conf_t *conf = mddev_to_conf(mddev); 4658 4659 switch(state) { 4660 case 2: /* resume for a suspend */ 4661 wake_up(&conf->wait_for_overlap); 4662 break; 4663 4664 case 1: /* stop all writes */ 4665 spin_lock_irq(&conf->device_lock); 4666 conf->quiesce = 1; 4667 wait_event_lock_irq(conf->wait_for_stripe, 4668 atomic_read(&conf->active_stripes) == 0 && 4669 atomic_read(&conf->active_aligned_reads) == 0, 4670 conf->device_lock, /* nothing */); 4671 spin_unlock_irq(&conf->device_lock); 4672 break; 4673 4674 case 0: /* re-enable writes */ 4675 spin_lock_irq(&conf->device_lock); 4676 conf->quiesce = 0; 4677 wake_up(&conf->wait_for_stripe); 4678 wake_up(&conf->wait_for_overlap); 4679 spin_unlock_irq(&conf->device_lock); 4680 break; 4681 } 4682} 4683 4684static struct mdk_personality raid6_personality = 4685{ 4686 .name = "raid6", 4687 .level = 6, 4688 .owner = THIS_MODULE, 4689 .make_request = make_request, 4690 .run = run, 4691 .stop = stop, 4692 .status = status, 4693 .error_handler = error, 4694 .hot_add_disk = raid5_add_disk, 4695 .hot_remove_disk= raid5_remove_disk, 4696 .spare_active = raid5_spare_active, 4697 .sync_request = sync_request, 4698 .resize = raid5_resize, 4699#ifdef CONFIG_MD_RAID5_RESHAPE 4700 .check_reshape = raid5_check_reshape, 4701 .start_reshape = raid5_start_reshape, 4702#endif 4703 .quiesce = raid5_quiesce, 4704}; 4705static struct mdk_personality raid5_personality = 4706{ 4707 .name = "raid5", 4708 .level = 5, 4709 .owner = THIS_MODULE, 4710 .make_request = make_request, 4711 .run = run, 4712 .stop = stop, 4713 .status = status, 4714 .error_handler = error, 4715 .hot_add_disk = raid5_add_disk, 4716 .hot_remove_disk= raid5_remove_disk, 4717 .spare_active = raid5_spare_active, 4718 .sync_request = sync_request, 4719 .resize = raid5_resize, 4720#ifdef CONFIG_MD_RAID5_RESHAPE 4721 .check_reshape = raid5_check_reshape, 4722 .start_reshape = raid5_start_reshape, 4723#endif 4724 .quiesce = raid5_quiesce, 4725}; 4726 4727static struct mdk_personality raid4_personality = 4728{ 4729 .name = "raid4", 4730 .level = 4, 4731 .owner = THIS_MODULE, 4732 .make_request = make_request, 4733 .run = run, 4734 .stop = stop, 4735 .status = status, 4736 .error_handler = error, 4737 .hot_add_disk = raid5_add_disk, 4738 .hot_remove_disk= raid5_remove_disk, 4739 .spare_active = raid5_spare_active, 4740 .sync_request = sync_request, 4741 .resize = raid5_resize, 4742#ifdef CONFIG_MD_RAID5_RESHAPE 4743 .check_reshape = raid5_check_reshape, 4744 .start_reshape = raid5_start_reshape, 4745#endif 4746 .quiesce = raid5_quiesce, 4747}; 4748 4749static int __init raid5_init(void) 4750{ 4751 int e; 4752 4753 e = raid6_select_algo(); 4754 if ( e ) 4755 return e; 4756 register_md_personality(&raid6_personality); 4757 register_md_personality(&raid5_personality); 4758 register_md_personality(&raid4_personality); 4759 return 0; 4760} 4761 4762static void raid5_exit(void) 4763{ 4764 unregister_md_personality(&raid6_personality); 4765 unregister_md_personality(&raid5_personality); 4766 unregister_md_personality(&raid4_personality); 4767} 4768 4769module_init(raid5_init); 4770module_exit(raid5_exit); 4771MODULE_LICENSE("GPL"); 4772MODULE_ALIAS("md-personality-4"); /* RAID5 */ 4773MODULE_ALIAS("md-raid5"); 4774MODULE_ALIAS("md-raid4"); 4775MODULE_ALIAS("md-level-5"); 4776MODULE_ALIAS("md-level-4"); 4777MODULE_ALIAS("md-personality-8"); /* RAID6 */ 4778MODULE_ALIAS("md-raid6"); 4779MODULE_ALIAS("md-level-6"); 4780 4781/* This used to be two separate modules, they were: */ 4782MODULE_ALIAS("raid5"); 4783MODULE_ALIAS("raid6"); 4784