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/blkdev.h> 47 #include <linux/kthread.h> 48 #include <linux/raid/pq.h> 49 #include <linux/async_tx.h> 50 #include <linux/async.h> 51 #include <linux/seq_file.h> 52 #include <linux/cpu.h> 53 #include <linux/slab.h> 54 #include "md.h" 55 #include "raid5.h" 56 #include "raid0.h" 57 #include "bitmap.h" 58 59 /* 60 * Stripe cache 61 */ 62 63 #define NR_STRIPES 256 64 #define STRIPE_SIZE PAGE_SIZE 65 #define STRIPE_SHIFT (PAGE_SHIFT - 9) 66 #define STRIPE_SECTORS (STRIPE_SIZE>>9) 67 #define IO_THRESHOLD 1 68 #define BYPASS_THRESHOLD 1 69 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head)) 70 #define HASH_MASK (NR_HASH - 1) 71 72 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK])) 73 74 /* bio's attached to a stripe+device for I/O are linked together in bi_sector 75 * order without overlap. There may be several bio's per stripe+device, and 76 * a bio could span several devices. 77 * When walking this list for a particular stripe+device, we must never proceed 78 * beyond a bio that extends past this device, as the next bio might no longer 79 * be valid. 80 * This macro is used to determine the 'next' bio in the list, given the sector 81 * of the current stripe+device 82 */ 83 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL) 84 /* 85 * The following can be used to debug the driver 86 */ 87 #define RAID5_PARANOIA 1 88 #if RAID5_PARANOIA && defined(CONFIG_SMP) 89 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock) 90 #else 91 # define CHECK_DEVLOCK() 92 #endif 93 94 #ifdef DEBUG 95 #define inline 96 #define __inline__ 97 #endif 98 99 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args))) 100 101 /* 102 * We maintain a biased count of active stripes in the bottom 16 bits of 103 * bi_phys_segments, and a count of processed stripes in the upper 16 bits 104 */ 105 static inline int raid5_bi_phys_segments(struct bio *bio) 106 { 107 return bio->bi_phys_segments & 0xffff; 108 } 109 110 static inline int raid5_bi_hw_segments(struct bio *bio) 111 { 112 return (bio->bi_phys_segments >> 16) & 0xffff; 113 } 114 115 static inline int raid5_dec_bi_phys_segments(struct bio *bio) 116 { 117 --bio->bi_phys_segments; 118 return raid5_bi_phys_segments(bio); 119 } 120 121 static inline int raid5_dec_bi_hw_segments(struct bio *bio) 122 { 123 unsigned short val = raid5_bi_hw_segments(bio); 124 125 --val; 126 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio); 127 return val; 128 } 129 130 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt) 131 { 132 bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16); 133 } 134 135 /* Find first data disk in a raid6 stripe */ 136 static inline int raid6_d0(struct stripe_head *sh) 137 { 138 if (sh->ddf_layout) 139 /* ddf always start from first device */ 140 return 0; 141 /* md starts just after Q block */ 142 if (sh->qd_idx == sh->disks - 1) 143 return 0; 144 else 145 return sh->qd_idx + 1; 146 } 147 static inline int raid6_next_disk(int disk, int raid_disks) 148 { 149 disk++; 150 return (disk < raid_disks) ? disk : 0; 151 } 152 153 /* When walking through the disks in a raid5, starting at raid6_d0, 154 * We need to map each disk to a 'slot', where the data disks are slot 155 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk 156 * is raid_disks-1. This help does that mapping. 157 */ 158 static int raid6_idx_to_slot(int idx, struct stripe_head *sh, 159 int *count, int syndrome_disks) 160 { 161 int slot = *count; 162 163 if (sh->ddf_layout) 164 (*count)++; 165 if (idx == sh->pd_idx) 166 return syndrome_disks; 167 if (idx == sh->qd_idx) 168 return syndrome_disks + 1; 169 if (!sh->ddf_layout) 170 (*count)++; 171 return slot; 172 } 173 174 static void return_io(struct bio *return_bi) 175 { 176 struct bio *bi = return_bi; 177 while (bi) { 178 179 return_bi = bi->bi_next; 180 bi->bi_next = NULL; 181 bi->bi_size = 0; 182 bio_endio(bi, 0); 183 bi = return_bi; 184 } 185 } 186 187 static void print_raid5_conf (raid5_conf_t *conf); 188 189 static int stripe_operations_active(struct stripe_head *sh) 190 { 191 return sh->check_state || sh->reconstruct_state || 192 test_bit(STRIPE_BIOFILL_RUN, &sh->state) || 193 test_bit(STRIPE_COMPUTE_RUN, &sh->state); 194 } 195 196 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh) 197 { 198 if (atomic_dec_and_test(&sh->count)) { 199 BUG_ON(!list_empty(&sh->lru)); 200 BUG_ON(atomic_read(&conf->active_stripes)==0); 201 if (test_bit(STRIPE_HANDLE, &sh->state)) { 202 if (test_bit(STRIPE_DELAYED, &sh->state)) { 203 list_add_tail(&sh->lru, &conf->delayed_list); 204 plugger_set_plug(&conf->plug); 205 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) && 206 sh->bm_seq - conf->seq_write > 0) { 207 list_add_tail(&sh->lru, &conf->bitmap_list); 208 plugger_set_plug(&conf->plug); 209 } else { 210 clear_bit(STRIPE_BIT_DELAY, &sh->state); 211 list_add_tail(&sh->lru, &conf->handle_list); 212 } 213 md_wakeup_thread(conf->mddev->thread); 214 } else { 215 BUG_ON(stripe_operations_active(sh)); 216 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 217 atomic_dec(&conf->preread_active_stripes); 218 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) 219 md_wakeup_thread(conf->mddev->thread); 220 } 221 atomic_dec(&conf->active_stripes); 222 if (!test_bit(STRIPE_EXPANDING, &sh->state)) { 223 list_add_tail(&sh->lru, &conf->inactive_list); 224 wake_up(&conf->wait_for_stripe); 225 if (conf->retry_read_aligned) 226 md_wakeup_thread(conf->mddev->thread); 227 } 228 } 229 } 230 } 231 232 static void release_stripe(struct stripe_head *sh) 233 { 234 raid5_conf_t *conf = sh->raid_conf; 235 unsigned long flags; 236 237 spin_lock_irqsave(&conf->device_lock, flags); 238 __release_stripe(conf, sh); 239 spin_unlock_irqrestore(&conf->device_lock, flags); 240 } 241 242 static inline void remove_hash(struct stripe_head *sh) 243 { 244 pr_debug("remove_hash(), stripe %llu\n", 245 (unsigned long long)sh->sector); 246 247 hlist_del_init(&sh->hash); 248 } 249 250 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh) 251 { 252 struct hlist_head *hp = stripe_hash(conf, sh->sector); 253 254 pr_debug("insert_hash(), stripe %llu\n", 255 (unsigned long long)sh->sector); 256 257 CHECK_DEVLOCK(); 258 hlist_add_head(&sh->hash, hp); 259 } 260 261 262 /* find an idle stripe, make sure it is unhashed, and return it. */ 263 static struct stripe_head *get_free_stripe(raid5_conf_t *conf) 264 { 265 struct stripe_head *sh = NULL; 266 struct list_head *first; 267 268 CHECK_DEVLOCK(); 269 if (list_empty(&conf->inactive_list)) 270 goto out; 271 first = conf->inactive_list.next; 272 sh = list_entry(first, struct stripe_head, lru); 273 list_del_init(first); 274 remove_hash(sh); 275 atomic_inc(&conf->active_stripes); 276 out: 277 return sh; 278 } 279 280 static void shrink_buffers(struct stripe_head *sh) 281 { 282 struct page *p; 283 int i; 284 int num = sh->raid_conf->pool_size; 285 286 for (i = 0; i < num ; i++) { 287 p = sh->dev[i].page; 288 if (!p) 289 continue; 290 sh->dev[i].page = NULL; 291 put_page(p); 292 } 293 } 294 295 static int grow_buffers(struct stripe_head *sh) 296 { 297 int i; 298 int num = sh->raid_conf->pool_size; 299 300 for (i = 0; i < num; i++) { 301 struct page *page; 302 303 if (!(page = alloc_page(GFP_KERNEL))) { 304 return 1; 305 } 306 sh->dev[i].page = page; 307 } 308 return 0; 309 } 310 311 static void raid5_build_block(struct stripe_head *sh, int i, int previous); 312 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous, 313 struct stripe_head *sh); 314 315 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous) 316 { 317 raid5_conf_t *conf = sh->raid_conf; 318 int i; 319 320 BUG_ON(atomic_read(&sh->count) != 0); 321 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state)); 322 BUG_ON(stripe_operations_active(sh)); 323 324 CHECK_DEVLOCK(); 325 pr_debug("init_stripe called, stripe %llu\n", 326 (unsigned long long)sh->sector); 327 328 remove_hash(sh); 329 330 sh->generation = conf->generation - previous; 331 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks; 332 sh->sector = sector; 333 stripe_set_idx(sector, conf, previous, sh); 334 sh->state = 0; 335 336 337 for (i = sh->disks; i--; ) { 338 struct r5dev *dev = &sh->dev[i]; 339 340 if (dev->toread || dev->read || dev->towrite || dev->written || 341 test_bit(R5_LOCKED, &dev->flags)) { 342 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n", 343 (unsigned long long)sh->sector, i, dev->toread, 344 dev->read, dev->towrite, dev->written, 345 test_bit(R5_LOCKED, &dev->flags)); 346 BUG(); 347 } 348 dev->flags = 0; 349 raid5_build_block(sh, i, previous); 350 } 351 insert_hash(conf, sh); 352 } 353 354 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, 355 short generation) 356 { 357 struct stripe_head *sh; 358 struct hlist_node *hn; 359 360 CHECK_DEVLOCK(); 361 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector); 362 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash) 363 if (sh->sector == sector && sh->generation == generation) 364 return sh; 365 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector); 366 return NULL; 367 } 368 369 /* 370 * Need to check if array has failed when deciding whether to: 371 * - start an array 372 * - remove non-faulty devices 373 * - add a spare 374 * - allow a reshape 375 * This determination is simple when no reshape is happening. 376 * However if there is a reshape, we need to carefully check 377 * both the before and after sections. 378 * This is because some failed devices may only affect one 379 * of the two sections, and some non-in_sync devices may 380 * be insync in the section most affected by failed devices. 381 */ 382 static int has_failed(raid5_conf_t *conf) 383 { 384 int degraded; 385 int i; 386 if (conf->mddev->reshape_position == MaxSector) 387 return conf->mddev->degraded > conf->max_degraded; 388 389 rcu_read_lock(); 390 degraded = 0; 391 for (i = 0; i < conf->previous_raid_disks; i++) { 392 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev); 393 if (!rdev || test_bit(Faulty, &rdev->flags)) 394 degraded++; 395 else if (test_bit(In_sync, &rdev->flags)) 396 ; 397 else 398 /* not in-sync or faulty. 399 * If the reshape increases the number of devices, 400 * this is being recovered by the reshape, so 401 * this 'previous' section is not in_sync. 402 * If the number of devices is being reduced however, 403 * the device can only be part of the array if 404 * we are reverting a reshape, so this section will 405 * be in-sync. 406 */ 407 if (conf->raid_disks >= conf->previous_raid_disks) 408 degraded++; 409 } 410 rcu_read_unlock(); 411 if (degraded > conf->max_degraded) 412 return 1; 413 rcu_read_lock(); 414 degraded = 0; 415 for (i = 0; i < conf->raid_disks; i++) { 416 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev); 417 if (!rdev || test_bit(Faulty, &rdev->flags)) 418 degraded++; 419 else if (test_bit(In_sync, &rdev->flags)) 420 ; 421 else 422 /* not in-sync or faulty. 423 * If reshape increases the number of devices, this 424 * section has already been recovered, else it 425 * almost certainly hasn't. 426 */ 427 if (conf->raid_disks <= conf->previous_raid_disks) 428 degraded++; 429 } 430 rcu_read_unlock(); 431 if (degraded > conf->max_degraded) 432 return 1; 433 return 0; 434 } 435 436 static void unplug_slaves(mddev_t *mddev); 437 438 static struct stripe_head * 439 get_active_stripe(raid5_conf_t *conf, sector_t sector, 440 int previous, int noblock, int noquiesce) 441 { 442 struct stripe_head *sh; 443 444 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector); 445 446 spin_lock_irq(&conf->device_lock); 447 448 do { 449 wait_event_lock_irq(conf->wait_for_stripe, 450 conf->quiesce == 0 || noquiesce, 451 conf->device_lock, /* nothing */); 452 sh = __find_stripe(conf, sector, conf->generation - previous); 453 if (!sh) { 454 if (!conf->inactive_blocked) 455 sh = get_free_stripe(conf); 456 if (noblock && sh == NULL) 457 break; 458 if (!sh) { 459 conf->inactive_blocked = 1; 460 wait_event_lock_irq(conf->wait_for_stripe, 461 !list_empty(&conf->inactive_list) && 462 (atomic_read(&conf->active_stripes) 463 < (conf->max_nr_stripes *3/4) 464 || !conf->inactive_blocked), 465 conf->device_lock, 466 md_raid5_unplug_device(conf) 467 ); 468 conf->inactive_blocked = 0; 469 } else 470 init_stripe(sh, sector, previous); 471 } else { 472 if (atomic_read(&sh->count)) { 473 BUG_ON(!list_empty(&sh->lru) 474 && !test_bit(STRIPE_EXPANDING, &sh->state)); 475 } else { 476 if (!test_bit(STRIPE_HANDLE, &sh->state)) 477 atomic_inc(&conf->active_stripes); 478 if (list_empty(&sh->lru) && 479 !test_bit(STRIPE_EXPANDING, &sh->state)) 480 BUG(); 481 list_del_init(&sh->lru); 482 } 483 } 484 } while (sh == NULL); 485 486 if (sh) 487 atomic_inc(&sh->count); 488 489 spin_unlock_irq(&conf->device_lock); 490 return sh; 491 } 492 493 static void 494 raid5_end_read_request(struct bio *bi, int error); 495 static void 496 raid5_end_write_request(struct bio *bi, int error); 497 498 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s) 499 { 500 raid5_conf_t *conf = sh->raid_conf; 501 int i, disks = sh->disks; 502 503 might_sleep(); 504 505 for (i = disks; i--; ) { 506 int rw; 507 struct bio *bi; 508 mdk_rdev_t *rdev; 509 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) 510 rw = WRITE; 511 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags)) 512 rw = READ; 513 else 514 continue; 515 516 bi = &sh->dev[i].req; 517 518 bi->bi_rw = rw; 519 if (rw == WRITE) 520 bi->bi_end_io = raid5_end_write_request; 521 else 522 bi->bi_end_io = raid5_end_read_request; 523 524 rcu_read_lock(); 525 rdev = rcu_dereference(conf->disks[i].rdev); 526 if (rdev && test_bit(Faulty, &rdev->flags)) 527 rdev = NULL; 528 if (rdev) 529 atomic_inc(&rdev->nr_pending); 530 rcu_read_unlock(); 531 532 if (rdev) { 533 if (s->syncing || s->expanding || s->expanded) 534 md_sync_acct(rdev->bdev, STRIPE_SECTORS); 535 536 set_bit(STRIPE_IO_STARTED, &sh->state); 537 538 bi->bi_bdev = rdev->bdev; 539 pr_debug("%s: for %llu schedule op %ld on disc %d\n", 540 __func__, (unsigned long long)sh->sector, 541 bi->bi_rw, i); 542 atomic_inc(&sh->count); 543 bi->bi_sector = sh->sector + rdev->data_offset; 544 bi->bi_flags = 1 << BIO_UPTODATE; 545 bi->bi_vcnt = 1; 546 bi->bi_max_vecs = 1; 547 bi->bi_idx = 0; 548 bi->bi_io_vec = &sh->dev[i].vec; 549 bi->bi_io_vec[0].bv_len = STRIPE_SIZE; 550 bi->bi_io_vec[0].bv_offset = 0; 551 bi->bi_size = STRIPE_SIZE; 552 bi->bi_next = NULL; 553 if (rw == WRITE && 554 test_bit(R5_ReWrite, &sh->dev[i].flags)) 555 atomic_add(STRIPE_SECTORS, 556 &rdev->corrected_errors); 557 generic_make_request(bi); 558 } else { 559 if (rw == WRITE) 560 set_bit(STRIPE_DEGRADED, &sh->state); 561 pr_debug("skip op %ld on disc %d for sector %llu\n", 562 bi->bi_rw, i, (unsigned long long)sh->sector); 563 clear_bit(R5_LOCKED, &sh->dev[i].flags); 564 set_bit(STRIPE_HANDLE, &sh->state); 565 } 566 } 567 } 568 569 static struct dma_async_tx_descriptor * 570 async_copy_data(int frombio, struct bio *bio, struct page *page, 571 sector_t sector, struct dma_async_tx_descriptor *tx) 572 { 573 struct bio_vec *bvl; 574 struct page *bio_page; 575 int i; 576 int page_offset; 577 struct async_submit_ctl submit; 578 enum async_tx_flags flags = 0; 579 580 if (bio->bi_sector >= sector) 581 page_offset = (signed)(bio->bi_sector - sector) * 512; 582 else 583 page_offset = (signed)(sector - bio->bi_sector) * -512; 584 585 if (frombio) 586 flags |= ASYNC_TX_FENCE; 587 init_async_submit(&submit, flags, tx, NULL, NULL, NULL); 588 589 bio_for_each_segment(bvl, bio, i) { 590 int len = bio_iovec_idx(bio, i)->bv_len; 591 int clen; 592 int b_offset = 0; 593 594 if (page_offset < 0) { 595 b_offset = -page_offset; 596 page_offset += b_offset; 597 len -= b_offset; 598 } 599 600 if (len > 0 && page_offset + len > STRIPE_SIZE) 601 clen = STRIPE_SIZE - page_offset; 602 else 603 clen = len; 604 605 if (clen > 0) { 606 b_offset += bio_iovec_idx(bio, i)->bv_offset; 607 bio_page = bio_iovec_idx(bio, i)->bv_page; 608 if (frombio) 609 tx = async_memcpy(page, bio_page, page_offset, 610 b_offset, clen, &submit); 611 else 612 tx = async_memcpy(bio_page, page, b_offset, 613 page_offset, clen, &submit); 614 } 615 /* chain the operations */ 616 submit.depend_tx = tx; 617 618 if (clen < len) /* hit end of page */ 619 break; 620 page_offset += len; 621 } 622 623 return tx; 624 } 625 626 static void ops_complete_biofill(void *stripe_head_ref) 627 { 628 struct stripe_head *sh = stripe_head_ref; 629 struct bio *return_bi = NULL; 630 raid5_conf_t *conf = sh->raid_conf; 631 int i; 632 633 pr_debug("%s: stripe %llu\n", __func__, 634 (unsigned long long)sh->sector); 635 636 /* clear completed biofills */ 637 spin_lock_irq(&conf->device_lock); 638 for (i = sh->disks; i--; ) { 639 struct r5dev *dev = &sh->dev[i]; 640 641 /* acknowledge completion of a biofill operation */ 642 /* and check if we need to reply to a read request, 643 * new R5_Wantfill requests are held off until 644 * !STRIPE_BIOFILL_RUN 645 */ 646 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) { 647 struct bio *rbi, *rbi2; 648 649 BUG_ON(!dev->read); 650 rbi = dev->read; 651 dev->read = NULL; 652 while (rbi && rbi->bi_sector < 653 dev->sector + STRIPE_SECTORS) { 654 rbi2 = r5_next_bio(rbi, dev->sector); 655 if (!raid5_dec_bi_phys_segments(rbi)) { 656 rbi->bi_next = return_bi; 657 return_bi = rbi; 658 } 659 rbi = rbi2; 660 } 661 } 662 } 663 spin_unlock_irq(&conf->device_lock); 664 clear_bit(STRIPE_BIOFILL_RUN, &sh->state); 665 666 return_io(return_bi); 667 668 set_bit(STRIPE_HANDLE, &sh->state); 669 release_stripe(sh); 670 } 671 672 static void ops_run_biofill(struct stripe_head *sh) 673 { 674 struct dma_async_tx_descriptor *tx = NULL; 675 raid5_conf_t *conf = sh->raid_conf; 676 struct async_submit_ctl submit; 677 int i; 678 679 pr_debug("%s: stripe %llu\n", __func__, 680 (unsigned long long)sh->sector); 681 682 for (i = sh->disks; i--; ) { 683 struct r5dev *dev = &sh->dev[i]; 684 if (test_bit(R5_Wantfill, &dev->flags)) { 685 struct bio *rbi; 686 spin_lock_irq(&conf->device_lock); 687 dev->read = rbi = dev->toread; 688 dev->toread = NULL; 689 spin_unlock_irq(&conf->device_lock); 690 while (rbi && rbi->bi_sector < 691 dev->sector + STRIPE_SECTORS) { 692 tx = async_copy_data(0, rbi, dev->page, 693 dev->sector, tx); 694 rbi = r5_next_bio(rbi, dev->sector); 695 } 696 } 697 } 698 699 atomic_inc(&sh->count); 700 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL); 701 async_trigger_callback(&submit); 702 } 703 704 static void mark_target_uptodate(struct stripe_head *sh, int target) 705 { 706 struct r5dev *tgt; 707 708 if (target < 0) 709 return; 710 711 tgt = &sh->dev[target]; 712 set_bit(R5_UPTODATE, &tgt->flags); 713 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags)); 714 clear_bit(R5_Wantcompute, &tgt->flags); 715 } 716 717 static void ops_complete_compute(void *stripe_head_ref) 718 { 719 struct stripe_head *sh = stripe_head_ref; 720 721 pr_debug("%s: stripe %llu\n", __func__, 722 (unsigned long long)sh->sector); 723 724 /* mark the computed target(s) as uptodate */ 725 mark_target_uptodate(sh, sh->ops.target); 726 mark_target_uptodate(sh, sh->ops.target2); 727 728 clear_bit(STRIPE_COMPUTE_RUN, &sh->state); 729 if (sh->check_state == check_state_compute_run) 730 sh->check_state = check_state_compute_result; 731 set_bit(STRIPE_HANDLE, &sh->state); 732 release_stripe(sh); 733 } 734 735 /* return a pointer to the address conversion region of the scribble buffer */ 736 static addr_conv_t *to_addr_conv(struct stripe_head *sh, 737 struct raid5_percpu *percpu) 738 { 739 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2); 740 } 741 742 static struct dma_async_tx_descriptor * 743 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu) 744 { 745 int disks = sh->disks; 746 struct page **xor_srcs = percpu->scribble; 747 int target = sh->ops.target; 748 struct r5dev *tgt = &sh->dev[target]; 749 struct page *xor_dest = tgt->page; 750 int count = 0; 751 struct dma_async_tx_descriptor *tx; 752 struct async_submit_ctl submit; 753 int i; 754 755 pr_debug("%s: stripe %llu block: %d\n", 756 __func__, (unsigned long long)sh->sector, target); 757 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags)); 758 759 for (i = disks; i--; ) 760 if (i != target) 761 xor_srcs[count++] = sh->dev[i].page; 762 763 atomic_inc(&sh->count); 764 765 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL, 766 ops_complete_compute, sh, to_addr_conv(sh, percpu)); 767 if (unlikely(count == 1)) 768 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit); 769 else 770 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit); 771 772 return tx; 773 } 774 775 /* set_syndrome_sources - populate source buffers for gen_syndrome 776 * @srcs - (struct page *) array of size sh->disks 777 * @sh - stripe_head to parse 778 * 779 * Populates srcs in proper layout order for the stripe and returns the 780 * 'count' of sources to be used in a call to async_gen_syndrome. The P 781 * destination buffer is recorded in srcs[count] and the Q destination 782 * is recorded in srcs[count+1]]. 783 */ 784 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh) 785 { 786 int disks = sh->disks; 787 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2); 788 int d0_idx = raid6_d0(sh); 789 int count; 790 int i; 791 792 for (i = 0; i < disks; i++) 793 srcs[i] = NULL; 794 795 count = 0; 796 i = d0_idx; 797 do { 798 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks); 799 800 srcs[slot] = sh->dev[i].page; 801 i = raid6_next_disk(i, disks); 802 } while (i != d0_idx); 803 804 return syndrome_disks; 805 } 806 807 static struct dma_async_tx_descriptor * 808 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu) 809 { 810 int disks = sh->disks; 811 struct page **blocks = percpu->scribble; 812 int target; 813 int qd_idx = sh->qd_idx; 814 struct dma_async_tx_descriptor *tx; 815 struct async_submit_ctl submit; 816 struct r5dev *tgt; 817 struct page *dest; 818 int i; 819 int count; 820 821 if (sh->ops.target < 0) 822 target = sh->ops.target2; 823 else if (sh->ops.target2 < 0) 824 target = sh->ops.target; 825 else 826 /* we should only have one valid target */ 827 BUG(); 828 BUG_ON(target < 0); 829 pr_debug("%s: stripe %llu block: %d\n", 830 __func__, (unsigned long long)sh->sector, target); 831 832 tgt = &sh->dev[target]; 833 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags)); 834 dest = tgt->page; 835 836 atomic_inc(&sh->count); 837 838 if (target == qd_idx) { 839 count = set_syndrome_sources(blocks, sh); 840 blocks[count] = NULL; /* regenerating p is not necessary */ 841 BUG_ON(blocks[count+1] != dest); /* q should already be set */ 842 init_async_submit(&submit, ASYNC_TX_FENCE, NULL, 843 ops_complete_compute, sh, 844 to_addr_conv(sh, percpu)); 845 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit); 846 } else { 847 /* Compute any data- or p-drive using XOR */ 848 count = 0; 849 for (i = disks; i-- ; ) { 850 if (i == target || i == qd_idx) 851 continue; 852 blocks[count++] = sh->dev[i].page; 853 } 854 855 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, 856 NULL, ops_complete_compute, sh, 857 to_addr_conv(sh, percpu)); 858 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit); 859 } 860 861 return tx; 862 } 863 864 static struct dma_async_tx_descriptor * 865 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu) 866 { 867 int i, count, disks = sh->disks; 868 int syndrome_disks = sh->ddf_layout ? disks : disks-2; 869 int d0_idx = raid6_d0(sh); 870 int faila = -1, failb = -1; 871 int target = sh->ops.target; 872 int target2 = sh->ops.target2; 873 struct r5dev *tgt = &sh->dev[target]; 874 struct r5dev *tgt2 = &sh->dev[target2]; 875 struct dma_async_tx_descriptor *tx; 876 struct page **blocks = percpu->scribble; 877 struct async_submit_ctl submit; 878 879 pr_debug("%s: stripe %llu block1: %d block2: %d\n", 880 __func__, (unsigned long long)sh->sector, target, target2); 881 BUG_ON(target < 0 || target2 < 0); 882 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags)); 883 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags)); 884 885 /* we need to open-code set_syndrome_sources to handle the 886 * slot number conversion for 'faila' and 'failb' 887 */ 888 for (i = 0; i < disks ; i++) 889 blocks[i] = NULL; 890 count = 0; 891 i = d0_idx; 892 do { 893 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks); 894 895 blocks[slot] = sh->dev[i].page; 896 897 if (i == target) 898 faila = slot; 899 if (i == target2) 900 failb = slot; 901 i = raid6_next_disk(i, disks); 902 } while (i != d0_idx); 903 904 BUG_ON(faila == failb); 905 if (failb < faila) 906 swap(faila, failb); 907 pr_debug("%s: stripe: %llu faila: %d failb: %d\n", 908 __func__, (unsigned long long)sh->sector, faila, failb); 909 910 atomic_inc(&sh->count); 911 912 if (failb == syndrome_disks+1) { 913 /* Q disk is one of the missing disks */ 914 if (faila == syndrome_disks) { 915 /* Missing P+Q, just recompute */ 916 init_async_submit(&submit, ASYNC_TX_FENCE, NULL, 917 ops_complete_compute, sh, 918 to_addr_conv(sh, percpu)); 919 return async_gen_syndrome(blocks, 0, syndrome_disks+2, 920 STRIPE_SIZE, &submit); 921 } else { 922 struct page *dest; 923 int data_target; 924 int qd_idx = sh->qd_idx; 925 926 /* Missing D+Q: recompute D from P, then recompute Q */ 927 if (target == qd_idx) 928 data_target = target2; 929 else 930 data_target = target; 931 932 count = 0; 933 for (i = disks; i-- ; ) { 934 if (i == data_target || i == qd_idx) 935 continue; 936 blocks[count++] = sh->dev[i].page; 937 } 938 dest = sh->dev[data_target].page; 939 init_async_submit(&submit, 940 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, 941 NULL, NULL, NULL, 942 to_addr_conv(sh, percpu)); 943 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, 944 &submit); 945 946 count = set_syndrome_sources(blocks, sh); 947 init_async_submit(&submit, ASYNC_TX_FENCE, tx, 948 ops_complete_compute, sh, 949 to_addr_conv(sh, percpu)); 950 return async_gen_syndrome(blocks, 0, count+2, 951 STRIPE_SIZE, &submit); 952 } 953 } else { 954 init_async_submit(&submit, ASYNC_TX_FENCE, NULL, 955 ops_complete_compute, sh, 956 to_addr_conv(sh, percpu)); 957 if (failb == syndrome_disks) { 958 /* We're missing D+P. */ 959 return async_raid6_datap_recov(syndrome_disks+2, 960 STRIPE_SIZE, faila, 961 blocks, &submit); 962 } else { 963 /* We're missing D+D. */ 964 return async_raid6_2data_recov(syndrome_disks+2, 965 STRIPE_SIZE, faila, failb, 966 blocks, &submit); 967 } 968 } 969 } 970 971 972 static void ops_complete_prexor(void *stripe_head_ref) 973 { 974 struct stripe_head *sh = stripe_head_ref; 975 976 pr_debug("%s: stripe %llu\n", __func__, 977 (unsigned long long)sh->sector); 978 } 979 980 static struct dma_async_tx_descriptor * 981 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu, 982 struct dma_async_tx_descriptor *tx) 983 { 984 int disks = sh->disks; 985 struct page **xor_srcs = percpu->scribble; 986 int count = 0, pd_idx = sh->pd_idx, i; 987 struct async_submit_ctl submit; 988 989 /* existing parity data subtracted */ 990 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page; 991 992 pr_debug("%s: stripe %llu\n", __func__, 993 (unsigned long long)sh->sector); 994 995 for (i = disks; i--; ) { 996 struct r5dev *dev = &sh->dev[i]; 997 /* Only process blocks that are known to be uptodate */ 998 if (test_bit(R5_Wantdrain, &dev->flags)) 999 xor_srcs[count++] = dev->page; 1000 } 1001 1002 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx, 1003 ops_complete_prexor, sh, to_addr_conv(sh, percpu)); 1004 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit); 1005 1006 return tx; 1007 } 1008 1009 static struct dma_async_tx_descriptor * 1010 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx) 1011 { 1012 int disks = sh->disks; 1013 int i; 1014 1015 pr_debug("%s: stripe %llu\n", __func__, 1016 (unsigned long long)sh->sector); 1017 1018 for (i = disks; i--; ) { 1019 struct r5dev *dev = &sh->dev[i]; 1020 struct bio *chosen; 1021 1022 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) { 1023 struct bio *wbi; 1024 1025 spin_lock(&sh->lock); 1026 chosen = dev->towrite; 1027 dev->towrite = NULL; 1028 BUG_ON(dev->written); 1029 wbi = dev->written = chosen; 1030 spin_unlock(&sh->lock); 1031 1032 while (wbi && wbi->bi_sector < 1033 dev->sector + STRIPE_SECTORS) { 1034 tx = async_copy_data(1, wbi, dev->page, 1035 dev->sector, tx); 1036 wbi = r5_next_bio(wbi, dev->sector); 1037 } 1038 } 1039 } 1040 1041 return tx; 1042 } 1043 1044 static void ops_complete_reconstruct(void *stripe_head_ref) 1045 { 1046 struct stripe_head *sh = stripe_head_ref; 1047 int disks = sh->disks; 1048 int pd_idx = sh->pd_idx; 1049 int qd_idx = sh->qd_idx; 1050 int i; 1051 1052 pr_debug("%s: stripe %llu\n", __func__, 1053 (unsigned long long)sh->sector); 1054 1055 for (i = disks; i--; ) { 1056 struct r5dev *dev = &sh->dev[i]; 1057 1058 if (dev->written || i == pd_idx || i == qd_idx) 1059 set_bit(R5_UPTODATE, &dev->flags); 1060 } 1061 1062 if (sh->reconstruct_state == reconstruct_state_drain_run) 1063 sh->reconstruct_state = reconstruct_state_drain_result; 1064 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) 1065 sh->reconstruct_state = reconstruct_state_prexor_drain_result; 1066 else { 1067 BUG_ON(sh->reconstruct_state != reconstruct_state_run); 1068 sh->reconstruct_state = reconstruct_state_result; 1069 } 1070 1071 set_bit(STRIPE_HANDLE, &sh->state); 1072 release_stripe(sh); 1073 } 1074 1075 static void 1076 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu, 1077 struct dma_async_tx_descriptor *tx) 1078 { 1079 int disks = sh->disks; 1080 struct page **xor_srcs = percpu->scribble; 1081 struct async_submit_ctl submit; 1082 int count = 0, pd_idx = sh->pd_idx, i; 1083 struct page *xor_dest; 1084 int prexor = 0; 1085 unsigned long flags; 1086 1087 pr_debug("%s: stripe %llu\n", __func__, 1088 (unsigned long long)sh->sector); 1089 1090 /* check if prexor is active which means only process blocks 1091 * that are part of a read-modify-write (written) 1092 */ 1093 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) { 1094 prexor = 1; 1095 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page; 1096 for (i = disks; i--; ) { 1097 struct r5dev *dev = &sh->dev[i]; 1098 if (dev->written) 1099 xor_srcs[count++] = dev->page; 1100 } 1101 } else { 1102 xor_dest = sh->dev[pd_idx].page; 1103 for (i = disks; i--; ) { 1104 struct r5dev *dev = &sh->dev[i]; 1105 if (i != pd_idx) 1106 xor_srcs[count++] = dev->page; 1107 } 1108 } 1109 1110 /* 1/ if we prexor'd then the dest is reused as a source 1111 * 2/ if we did not prexor then we are redoing the parity 1112 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST 1113 * for the synchronous xor case 1114 */ 1115 flags = ASYNC_TX_ACK | 1116 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST); 1117 1118 atomic_inc(&sh->count); 1119 1120 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh, 1121 to_addr_conv(sh, percpu)); 1122 if (unlikely(count == 1)) 1123 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit); 1124 else 1125 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit); 1126 } 1127 1128 static void 1129 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu, 1130 struct dma_async_tx_descriptor *tx) 1131 { 1132 struct async_submit_ctl submit; 1133 struct page **blocks = percpu->scribble; 1134 int count; 1135 1136 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector); 1137 1138 count = set_syndrome_sources(blocks, sh); 1139 1140 atomic_inc(&sh->count); 1141 1142 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct, 1143 sh, to_addr_conv(sh, percpu)); 1144 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit); 1145 } 1146 1147 static void ops_complete_check(void *stripe_head_ref) 1148 { 1149 struct stripe_head *sh = stripe_head_ref; 1150 1151 pr_debug("%s: stripe %llu\n", __func__, 1152 (unsigned long long)sh->sector); 1153 1154 sh->check_state = check_state_check_result; 1155 set_bit(STRIPE_HANDLE, &sh->state); 1156 release_stripe(sh); 1157 } 1158 1159 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu) 1160 { 1161 int disks = sh->disks; 1162 int pd_idx = sh->pd_idx; 1163 int qd_idx = sh->qd_idx; 1164 struct page *xor_dest; 1165 struct page **xor_srcs = percpu->scribble; 1166 struct dma_async_tx_descriptor *tx; 1167 struct async_submit_ctl submit; 1168 int count; 1169 int i; 1170 1171 pr_debug("%s: stripe %llu\n", __func__, 1172 (unsigned long long)sh->sector); 1173 1174 count = 0; 1175 xor_dest = sh->dev[pd_idx].page; 1176 xor_srcs[count++] = xor_dest; 1177 for (i = disks; i--; ) { 1178 if (i == pd_idx || i == qd_idx) 1179 continue; 1180 xor_srcs[count++] = sh->dev[i].page; 1181 } 1182 1183 init_async_submit(&submit, 0, NULL, NULL, NULL, 1184 to_addr_conv(sh, percpu)); 1185 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, 1186 &sh->ops.zero_sum_result, &submit); 1187 1188 atomic_inc(&sh->count); 1189 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL); 1190 tx = async_trigger_callback(&submit); 1191 } 1192 1193 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp) 1194 { 1195 struct page **srcs = percpu->scribble; 1196 struct async_submit_ctl submit; 1197 int count; 1198 1199 pr_debug("%s: stripe %llu checkp: %d\n", __func__, 1200 (unsigned long long)sh->sector, checkp); 1201 1202 count = set_syndrome_sources(srcs, sh); 1203 if (!checkp) 1204 srcs[count] = NULL; 1205 1206 atomic_inc(&sh->count); 1207 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check, 1208 sh, to_addr_conv(sh, percpu)); 1209 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE, 1210 &sh->ops.zero_sum_result, percpu->spare_page, &submit); 1211 } 1212 1213 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request) 1214 { 1215 int overlap_clear = 0, i, disks = sh->disks; 1216 struct dma_async_tx_descriptor *tx = NULL; 1217 raid5_conf_t *conf = sh->raid_conf; 1218 int level = conf->level; 1219 struct raid5_percpu *percpu; 1220 unsigned long cpu; 1221 1222 cpu = get_cpu(); 1223 percpu = per_cpu_ptr(conf->percpu, cpu); 1224 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) { 1225 ops_run_biofill(sh); 1226 overlap_clear++; 1227 } 1228 1229 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) { 1230 if (level < 6) 1231 tx = ops_run_compute5(sh, percpu); 1232 else { 1233 if (sh->ops.target2 < 0 || sh->ops.target < 0) 1234 tx = ops_run_compute6_1(sh, percpu); 1235 else 1236 tx = ops_run_compute6_2(sh, percpu); 1237 } 1238 /* terminate the chain if reconstruct is not set to be run */ 1239 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) 1240 async_tx_ack(tx); 1241 } 1242 1243 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) 1244 tx = ops_run_prexor(sh, percpu, tx); 1245 1246 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) { 1247 tx = ops_run_biodrain(sh, tx); 1248 overlap_clear++; 1249 } 1250 1251 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) { 1252 if (level < 6) 1253 ops_run_reconstruct5(sh, percpu, tx); 1254 else 1255 ops_run_reconstruct6(sh, percpu, tx); 1256 } 1257 1258 if (test_bit(STRIPE_OP_CHECK, &ops_request)) { 1259 if (sh->check_state == check_state_run) 1260 ops_run_check_p(sh, percpu); 1261 else if (sh->check_state == check_state_run_q) 1262 ops_run_check_pq(sh, percpu, 0); 1263 else if (sh->check_state == check_state_run_pq) 1264 ops_run_check_pq(sh, percpu, 1); 1265 else 1266 BUG(); 1267 } 1268 1269 if (overlap_clear) 1270 for (i = disks; i--; ) { 1271 struct r5dev *dev = &sh->dev[i]; 1272 if (test_and_clear_bit(R5_Overlap, &dev->flags)) 1273 wake_up(&sh->raid_conf->wait_for_overlap); 1274 } 1275 put_cpu(); 1276 } 1277 1278 #ifdef CONFIG_MULTICORE_RAID456 1279 static void async_run_ops(void *param, async_cookie_t cookie) 1280 { 1281 struct stripe_head *sh = param; 1282 unsigned long ops_request = sh->ops.request; 1283 1284 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state); 1285 wake_up(&sh->ops.wait_for_ops); 1286 1287 __raid_run_ops(sh, ops_request); 1288 release_stripe(sh); 1289 } 1290 1291 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request) 1292 { 1293 /* since handle_stripe can be called outside of raid5d context 1294 * we need to ensure sh->ops.request is de-staged before another 1295 * request arrives 1296 */ 1297 wait_event(sh->ops.wait_for_ops, 1298 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state)); 1299 sh->ops.request = ops_request; 1300 1301 atomic_inc(&sh->count); 1302 async_schedule(async_run_ops, sh); 1303 } 1304 #else 1305 #define raid_run_ops __raid_run_ops 1306 #endif 1307 1308 static int grow_one_stripe(raid5_conf_t *conf) 1309 { 1310 struct stripe_head *sh; 1311 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL); 1312 if (!sh) 1313 return 0; 1314 memset(sh, 0, sizeof(*sh) + (conf->pool_size-1)*sizeof(struct r5dev)); 1315 sh->raid_conf = conf; 1316 spin_lock_init(&sh->lock); 1317 #ifdef CONFIG_MULTICORE_RAID456 1318 init_waitqueue_head(&sh->ops.wait_for_ops); 1319 #endif 1320 1321 if (grow_buffers(sh)) { 1322 shrink_buffers(sh); 1323 kmem_cache_free(conf->slab_cache, sh); 1324 return 0; 1325 } 1326 /* we just created an active stripe so... */ 1327 atomic_set(&sh->count, 1); 1328 atomic_inc(&conf->active_stripes); 1329 INIT_LIST_HEAD(&sh->lru); 1330 release_stripe(sh); 1331 return 1; 1332 } 1333 1334 static int grow_stripes(raid5_conf_t *conf, int num) 1335 { 1336 struct kmem_cache *sc; 1337 int devs = max(conf->raid_disks, conf->previous_raid_disks); 1338 1339 if (conf->mddev->gendisk) 1340 sprintf(conf->cache_name[0], 1341 "raid%d-%s", conf->level, mdname(conf->mddev)); 1342 else 1343 sprintf(conf->cache_name[0], 1344 "raid%d-%p", conf->level, conf->mddev); 1345 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]); 1346 1347 conf->active_name = 0; 1348 sc = kmem_cache_create(conf->cache_name[conf->active_name], 1349 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev), 1350 0, 0, NULL); 1351 if (!sc) 1352 return 1; 1353 conf->slab_cache = sc; 1354 conf->pool_size = devs; 1355 while (num--) 1356 if (!grow_one_stripe(conf)) 1357 return 1; 1358 return 0; 1359 } 1360 1361 /** 1362 * scribble_len - return the required size of the scribble region 1363 * @num - total number of disks in the array 1364 * 1365 * The size must be enough to contain: 1366 * 1/ a struct page pointer for each device in the array +2 1367 * 2/ room to convert each entry in (1) to its corresponding dma 1368 * (dma_map_page()) or page (page_address()) address. 1369 * 1370 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we 1371 * calculate over all devices (not just the data blocks), using zeros in place 1372 * of the P and Q blocks. 1373 */ 1374 static size_t scribble_len(int num) 1375 { 1376 size_t len; 1377 1378 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2); 1379 1380 return len; 1381 } 1382 1383 static int resize_stripes(raid5_conf_t *conf, int newsize) 1384 { 1385 /* Make all the stripes able to hold 'newsize' devices. 1386 * New slots in each stripe get 'page' set to a new page. 1387 * 1388 * This happens in stages: 1389 * 1/ create a new kmem_cache and allocate the required number of 1390 * stripe_heads. 1391 * 2/ gather all the old stripe_heads and tranfer the pages across 1392 * to the new stripe_heads. This will have the side effect of 1393 * freezing the array as once all stripe_heads have been collected, 1394 * no IO will be possible. Old stripe heads are freed once their 1395 * pages have been transferred over, and the old kmem_cache is 1396 * freed when all stripes are done. 1397 * 3/ reallocate conf->disks to be suitable bigger. If this fails, 1398 * we simple return a failre status - no need to clean anything up. 1399 * 4/ allocate new pages for the new slots in the new stripe_heads. 1400 * If this fails, we don't bother trying the shrink the 1401 * stripe_heads down again, we just leave them as they are. 1402 * As each stripe_head is processed the new one is released into 1403 * active service. 1404 * 1405 * Once step2 is started, we cannot afford to wait for a write, 1406 * so we use GFP_NOIO allocations. 1407 */ 1408 struct stripe_head *osh, *nsh; 1409 LIST_HEAD(newstripes); 1410 struct disk_info *ndisks; 1411 unsigned long cpu; 1412 int err; 1413 struct kmem_cache *sc; 1414 int i; 1415 1416 if (newsize <= conf->pool_size) 1417 return 0; /* never bother to shrink */ 1418 1419 err = md_allow_write(conf->mddev); 1420 if (err) 1421 return err; 1422 1423 /* Step 1 */ 1424 sc = kmem_cache_create(conf->cache_name[1-conf->active_name], 1425 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev), 1426 0, 0, NULL); 1427 if (!sc) 1428 return -ENOMEM; 1429 1430 for (i = conf->max_nr_stripes; i; i--) { 1431 nsh = kmem_cache_alloc(sc, GFP_KERNEL); 1432 if (!nsh) 1433 break; 1434 1435 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev)); 1436 1437 nsh->raid_conf = conf; 1438 spin_lock_init(&nsh->lock); 1439 #ifdef CONFIG_MULTICORE_RAID456 1440 init_waitqueue_head(&nsh->ops.wait_for_ops); 1441 #endif 1442 1443 list_add(&nsh->lru, &newstripes); 1444 } 1445 if (i) { 1446 /* didn't get enough, give up */ 1447 while (!list_empty(&newstripes)) { 1448 nsh = list_entry(newstripes.next, struct stripe_head, lru); 1449 list_del(&nsh->lru); 1450 kmem_cache_free(sc, nsh); 1451 } 1452 kmem_cache_destroy(sc); 1453 return -ENOMEM; 1454 } 1455 /* Step 2 - Must use GFP_NOIO now. 1456 * OK, we have enough stripes, start collecting inactive 1457 * stripes and copying them over 1458 */ 1459 list_for_each_entry(nsh, &newstripes, lru) { 1460 spin_lock_irq(&conf->device_lock); 1461 wait_event_lock_irq(conf->wait_for_stripe, 1462 !list_empty(&conf->inactive_list), 1463 conf->device_lock, 1464 unplug_slaves(conf->mddev) 1465 ); 1466 osh = get_free_stripe(conf); 1467 spin_unlock_irq(&conf->device_lock); 1468 atomic_set(&nsh->count, 1); 1469 for(i=0; i<conf->pool_size; i++) 1470 nsh->dev[i].page = osh->dev[i].page; 1471 for( ; i<newsize; i++) 1472 nsh->dev[i].page = NULL; 1473 kmem_cache_free(conf->slab_cache, osh); 1474 } 1475 kmem_cache_destroy(conf->slab_cache); 1476 1477 /* Step 3. 1478 * At this point, we are holding all the stripes so the array 1479 * is completely stalled, so now is a good time to resize 1480 * conf->disks and the scribble region 1481 */ 1482 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO); 1483 if (ndisks) { 1484 for (i=0; i<conf->raid_disks; i++) 1485 ndisks[i] = conf->disks[i]; 1486 kfree(conf->disks); 1487 conf->disks = ndisks; 1488 } else 1489 err = -ENOMEM; 1490 1491 get_online_cpus(); 1492 conf->scribble_len = scribble_len(newsize); 1493 for_each_present_cpu(cpu) { 1494 struct raid5_percpu *percpu; 1495 void *scribble; 1496 1497 percpu = per_cpu_ptr(conf->percpu, cpu); 1498 scribble = kmalloc(conf->scribble_len, GFP_NOIO); 1499 1500 if (scribble) { 1501 kfree(percpu->scribble); 1502 percpu->scribble = scribble; 1503 } else { 1504 err = -ENOMEM; 1505 break; 1506 } 1507 } 1508 put_online_cpus(); 1509 1510 /* Step 4, return new stripes to service */ 1511 while(!list_empty(&newstripes)) { 1512 nsh = list_entry(newstripes.next, struct stripe_head, lru); 1513 list_del_init(&nsh->lru); 1514 1515 for (i=conf->raid_disks; i < newsize; i++) 1516 if (nsh->dev[i].page == NULL) { 1517 struct page *p = alloc_page(GFP_NOIO); 1518 nsh->dev[i].page = p; 1519 if (!p) 1520 err = -ENOMEM; 1521 } 1522 release_stripe(nsh); 1523 } 1524 /* critical section pass, GFP_NOIO no longer needed */ 1525 1526 conf->slab_cache = sc; 1527 conf->active_name = 1-conf->active_name; 1528 conf->pool_size = newsize; 1529 return err; 1530 } 1531 1532 static int drop_one_stripe(raid5_conf_t *conf) 1533 { 1534 struct stripe_head *sh; 1535 1536 spin_lock_irq(&conf->device_lock); 1537 sh = get_free_stripe(conf); 1538 spin_unlock_irq(&conf->device_lock); 1539 if (!sh) 1540 return 0; 1541 BUG_ON(atomic_read(&sh->count)); 1542 shrink_buffers(sh); 1543 kmem_cache_free(conf->slab_cache, sh); 1544 atomic_dec(&conf->active_stripes); 1545 return 1; 1546 } 1547 1548 static void shrink_stripes(raid5_conf_t *conf) 1549 { 1550 while (drop_one_stripe(conf)) 1551 ; 1552 1553 if (conf->slab_cache) 1554 kmem_cache_destroy(conf->slab_cache); 1555 conf->slab_cache = NULL; 1556 } 1557 1558 static void raid5_end_read_request(struct bio * bi, int error) 1559 { 1560 struct stripe_head *sh = bi->bi_private; 1561 raid5_conf_t *conf = sh->raid_conf; 1562 int disks = sh->disks, i; 1563 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags); 1564 char b[BDEVNAME_SIZE]; 1565 mdk_rdev_t *rdev; 1566 1567 1568 for (i=0 ; i<disks; i++) 1569 if (bi == &sh->dev[i].req) 1570 break; 1571 1572 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n", 1573 (unsigned long long)sh->sector, i, atomic_read(&sh->count), 1574 uptodate); 1575 if (i == disks) { 1576 BUG(); 1577 return; 1578 } 1579 1580 if (uptodate) { 1581 set_bit(R5_UPTODATE, &sh->dev[i].flags); 1582 if (test_bit(R5_ReadError, &sh->dev[i].flags)) { 1583 rdev = conf->disks[i].rdev; 1584 printk_rl(KERN_INFO "md/raid:%s: read error corrected" 1585 " (%lu sectors at %llu on %s)\n", 1586 mdname(conf->mddev), STRIPE_SECTORS, 1587 (unsigned long long)(sh->sector 1588 + rdev->data_offset), 1589 bdevname(rdev->bdev, b)); 1590 clear_bit(R5_ReadError, &sh->dev[i].flags); 1591 clear_bit(R5_ReWrite, &sh->dev[i].flags); 1592 } 1593 if (atomic_read(&conf->disks[i].rdev->read_errors)) 1594 atomic_set(&conf->disks[i].rdev->read_errors, 0); 1595 } else { 1596 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b); 1597 int retry = 0; 1598 rdev = conf->disks[i].rdev; 1599 1600 clear_bit(R5_UPTODATE, &sh->dev[i].flags); 1601 atomic_inc(&rdev->read_errors); 1602 if (conf->mddev->degraded >= conf->max_degraded) 1603 printk_rl(KERN_WARNING 1604 "md/raid:%s: read error not correctable " 1605 "(sector %llu on %s).\n", 1606 mdname(conf->mddev), 1607 (unsigned long long)(sh->sector 1608 + rdev->data_offset), 1609 bdn); 1610 else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) 1611 /* Oh, no!!! */ 1612 printk_rl(KERN_WARNING 1613 "md/raid:%s: read error NOT corrected!! " 1614 "(sector %llu on %s).\n", 1615 mdname(conf->mddev), 1616 (unsigned long long)(sh->sector 1617 + rdev->data_offset), 1618 bdn); 1619 else if (atomic_read(&rdev->read_errors) 1620 > conf->max_nr_stripes) 1621 printk(KERN_WARNING 1622 "md/raid:%s: Too many read errors, failing device %s.\n", 1623 mdname(conf->mddev), bdn); 1624 else 1625 retry = 1; 1626 if (retry) 1627 set_bit(R5_ReadError, &sh->dev[i].flags); 1628 else { 1629 clear_bit(R5_ReadError, &sh->dev[i].flags); 1630 clear_bit(R5_ReWrite, &sh->dev[i].flags); 1631 md_error(conf->mddev, rdev); 1632 } 1633 } 1634 rdev_dec_pending(conf->disks[i].rdev, conf->mddev); 1635 clear_bit(R5_LOCKED, &sh->dev[i].flags); 1636 set_bit(STRIPE_HANDLE, &sh->state); 1637 release_stripe(sh); 1638 } 1639 1640 static void raid5_end_write_request(struct bio *bi, int error) 1641 { 1642 struct stripe_head *sh = bi->bi_private; 1643 raid5_conf_t *conf = sh->raid_conf; 1644 int disks = sh->disks, i; 1645 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags); 1646 1647 for (i=0 ; i<disks; i++) 1648 if (bi == &sh->dev[i].req) 1649 break; 1650 1651 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n", 1652 (unsigned long long)sh->sector, i, atomic_read(&sh->count), 1653 uptodate); 1654 if (i == disks) { 1655 BUG(); 1656 return; 1657 } 1658 1659 if (!uptodate) 1660 md_error(conf->mddev, conf->disks[i].rdev); 1661 1662 rdev_dec_pending(conf->disks[i].rdev, conf->mddev); 1663 1664 clear_bit(R5_LOCKED, &sh->dev[i].flags); 1665 set_bit(STRIPE_HANDLE, &sh->state); 1666 release_stripe(sh); 1667 } 1668 1669 1670 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous); 1671 1672 static void raid5_build_block(struct stripe_head *sh, int i, int previous) 1673 { 1674 struct r5dev *dev = &sh->dev[i]; 1675 1676 bio_init(&dev->req); 1677 dev->req.bi_io_vec = &dev->vec; 1678 dev->req.bi_vcnt++; 1679 dev->req.bi_max_vecs++; 1680 dev->vec.bv_page = dev->page; 1681 dev->vec.bv_len = STRIPE_SIZE; 1682 dev->vec.bv_offset = 0; 1683 1684 dev->req.bi_sector = sh->sector; 1685 dev->req.bi_private = sh; 1686 1687 dev->flags = 0; 1688 dev->sector = compute_blocknr(sh, i, previous); 1689 } 1690 1691 static void error(mddev_t *mddev, mdk_rdev_t *rdev) 1692 { 1693 char b[BDEVNAME_SIZE]; 1694 raid5_conf_t *conf = mddev->private; 1695 pr_debug("raid456: error called\n"); 1696 1697 if (!test_bit(Faulty, &rdev->flags)) { 1698 set_bit(MD_CHANGE_DEVS, &mddev->flags); 1699 if (test_and_clear_bit(In_sync, &rdev->flags)) { 1700 unsigned long flags; 1701 spin_lock_irqsave(&conf->device_lock, flags); 1702 mddev->degraded++; 1703 spin_unlock_irqrestore(&conf->device_lock, flags); 1704 /* 1705 * if recovery was running, make sure it aborts. 1706 */ 1707 set_bit(MD_RECOVERY_INTR, &mddev->recovery); 1708 } 1709 set_bit(Faulty, &rdev->flags); 1710 printk(KERN_ALERT 1711 "md/raid:%s: Disk failure on %s, disabling device.\n" 1712 KERN_ALERT 1713 "md/raid:%s: Operation continuing on %d devices.\n", 1714 mdname(mddev), 1715 bdevname(rdev->bdev, b), 1716 mdname(mddev), 1717 conf->raid_disks - mddev->degraded); 1718 } 1719 } 1720 1721 /* 1722 * Input: a 'big' sector number, 1723 * Output: index of the data and parity disk, and the sector # in them. 1724 */ 1725 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector, 1726 int previous, int *dd_idx, 1727 struct stripe_head *sh) 1728 { 1729 sector_t stripe, stripe2; 1730 sector_t chunk_number; 1731 unsigned int chunk_offset; 1732 int pd_idx, qd_idx; 1733 int ddf_layout = 0; 1734 sector_t new_sector; 1735 int algorithm = previous ? conf->prev_algo 1736 : conf->algorithm; 1737 int sectors_per_chunk = previous ? conf->prev_chunk_sectors 1738 : conf->chunk_sectors; 1739 int raid_disks = previous ? conf->previous_raid_disks 1740 : conf->raid_disks; 1741 int data_disks = raid_disks - conf->max_degraded; 1742 1743 /* First compute the information on this sector */ 1744 1745 /* 1746 * Compute the chunk number and the sector offset inside the chunk 1747 */ 1748 chunk_offset = sector_div(r_sector, sectors_per_chunk); 1749 chunk_number = r_sector; 1750 1751 /* 1752 * Compute the stripe number 1753 */ 1754 stripe = chunk_number; 1755 *dd_idx = sector_div(stripe, data_disks); 1756 stripe2 = stripe; 1757 /* 1758 * Select the parity disk based on the user selected algorithm. 1759 */ 1760 pd_idx = qd_idx = ~0; 1761 switch(conf->level) { 1762 case 4: 1763 pd_idx = data_disks; 1764 break; 1765 case 5: 1766 switch (algorithm) { 1767 case ALGORITHM_LEFT_ASYMMETRIC: 1768 pd_idx = data_disks - sector_div(stripe2, raid_disks); 1769 if (*dd_idx >= pd_idx) 1770 (*dd_idx)++; 1771 break; 1772 case ALGORITHM_RIGHT_ASYMMETRIC: 1773 pd_idx = sector_div(stripe2, raid_disks); 1774 if (*dd_idx >= pd_idx) 1775 (*dd_idx)++; 1776 break; 1777 case ALGORITHM_LEFT_SYMMETRIC: 1778 pd_idx = data_disks - sector_div(stripe2, raid_disks); 1779 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks; 1780 break; 1781 case ALGORITHM_RIGHT_SYMMETRIC: 1782 pd_idx = sector_div(stripe2, raid_disks); 1783 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks; 1784 break; 1785 case ALGORITHM_PARITY_0: 1786 pd_idx = 0; 1787 (*dd_idx)++; 1788 break; 1789 case ALGORITHM_PARITY_N: 1790 pd_idx = data_disks; 1791 break; 1792 default: 1793 BUG(); 1794 } 1795 break; 1796 case 6: 1797 1798 switch (algorithm) { 1799 case ALGORITHM_LEFT_ASYMMETRIC: 1800 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks); 1801 qd_idx = pd_idx + 1; 1802 if (pd_idx == raid_disks-1) { 1803 (*dd_idx)++; /* Q D D D P */ 1804 qd_idx = 0; 1805 } else if (*dd_idx >= pd_idx) 1806 (*dd_idx) += 2; /* D D P Q D */ 1807 break; 1808 case ALGORITHM_RIGHT_ASYMMETRIC: 1809 pd_idx = sector_div(stripe2, raid_disks); 1810 qd_idx = pd_idx + 1; 1811 if (pd_idx == raid_disks-1) { 1812 (*dd_idx)++; /* Q D D D P */ 1813 qd_idx = 0; 1814 } else if (*dd_idx >= pd_idx) 1815 (*dd_idx) += 2; /* D D P Q D */ 1816 break; 1817 case ALGORITHM_LEFT_SYMMETRIC: 1818 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks); 1819 qd_idx = (pd_idx + 1) % raid_disks; 1820 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks; 1821 break; 1822 case ALGORITHM_RIGHT_SYMMETRIC: 1823 pd_idx = sector_div(stripe2, raid_disks); 1824 qd_idx = (pd_idx + 1) % raid_disks; 1825 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks; 1826 break; 1827 1828 case ALGORITHM_PARITY_0: 1829 pd_idx = 0; 1830 qd_idx = 1; 1831 (*dd_idx) += 2; 1832 break; 1833 case ALGORITHM_PARITY_N: 1834 pd_idx = data_disks; 1835 qd_idx = data_disks + 1; 1836 break; 1837 1838 case ALGORITHM_ROTATING_ZERO_RESTART: 1839 /* Exactly the same as RIGHT_ASYMMETRIC, but or 1840 * of blocks for computing Q is different. 1841 */ 1842 pd_idx = sector_div(stripe2, raid_disks); 1843 qd_idx = pd_idx + 1; 1844 if (pd_idx == raid_disks-1) { 1845 (*dd_idx)++; /* Q D D D P */ 1846 qd_idx = 0; 1847 } else if (*dd_idx >= pd_idx) 1848 (*dd_idx) += 2; /* D D P Q D */ 1849 ddf_layout = 1; 1850 break; 1851 1852 case ALGORITHM_ROTATING_N_RESTART: 1853 /* Same a left_asymmetric, by first stripe is 1854 * D D D P Q rather than 1855 * Q D D D P 1856 */ 1857 stripe2 += 1; 1858 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks); 1859 qd_idx = pd_idx + 1; 1860 if (pd_idx == raid_disks-1) { 1861 (*dd_idx)++; /* Q D D D P */ 1862 qd_idx = 0; 1863 } else if (*dd_idx >= pd_idx) 1864 (*dd_idx) += 2; /* D D P Q D */ 1865 ddf_layout = 1; 1866 break; 1867 1868 case ALGORITHM_ROTATING_N_CONTINUE: 1869 /* Same as left_symmetric but Q is before P */ 1870 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks); 1871 qd_idx = (pd_idx + raid_disks - 1) % raid_disks; 1872 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks; 1873 ddf_layout = 1; 1874 break; 1875 1876 case ALGORITHM_LEFT_ASYMMETRIC_6: 1877 /* RAID5 left_asymmetric, with Q on last device */ 1878 pd_idx = data_disks - sector_div(stripe2, raid_disks-1); 1879 if (*dd_idx >= pd_idx) 1880 (*dd_idx)++; 1881 qd_idx = raid_disks - 1; 1882 break; 1883 1884 case ALGORITHM_RIGHT_ASYMMETRIC_6: 1885 pd_idx = sector_div(stripe2, raid_disks-1); 1886 if (*dd_idx >= pd_idx) 1887 (*dd_idx)++; 1888 qd_idx = raid_disks - 1; 1889 break; 1890 1891 case ALGORITHM_LEFT_SYMMETRIC_6: 1892 pd_idx = data_disks - sector_div(stripe2, raid_disks-1); 1893 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1); 1894 qd_idx = raid_disks - 1; 1895 break; 1896 1897 case ALGORITHM_RIGHT_SYMMETRIC_6: 1898 pd_idx = sector_div(stripe2, raid_disks-1); 1899 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1); 1900 qd_idx = raid_disks - 1; 1901 break; 1902 1903 case ALGORITHM_PARITY_0_6: 1904 pd_idx = 0; 1905 (*dd_idx)++; 1906 qd_idx = raid_disks - 1; 1907 break; 1908 1909 default: 1910 BUG(); 1911 } 1912 break; 1913 } 1914 1915 if (sh) { 1916 sh->pd_idx = pd_idx; 1917 sh->qd_idx = qd_idx; 1918 sh->ddf_layout = ddf_layout; 1919 } 1920 /* 1921 * Finally, compute the new sector number 1922 */ 1923 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset; 1924 return new_sector; 1925 } 1926 1927 1928 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous) 1929 { 1930 raid5_conf_t *conf = sh->raid_conf; 1931 int raid_disks = sh->disks; 1932 int data_disks = raid_disks - conf->max_degraded; 1933 sector_t new_sector = sh->sector, check; 1934 int sectors_per_chunk = previous ? conf->prev_chunk_sectors 1935 : conf->chunk_sectors; 1936 int algorithm = previous ? conf->prev_algo 1937 : conf->algorithm; 1938 sector_t stripe; 1939 int chunk_offset; 1940 sector_t chunk_number; 1941 int dummy1, dd_idx = i; 1942 sector_t r_sector; 1943 struct stripe_head sh2; 1944 1945 1946 chunk_offset = sector_div(new_sector, sectors_per_chunk); 1947 stripe = new_sector; 1948 1949 if (i == sh->pd_idx) 1950 return 0; 1951 switch(conf->level) { 1952 case 4: break; 1953 case 5: 1954 switch (algorithm) { 1955 case ALGORITHM_LEFT_ASYMMETRIC: 1956 case ALGORITHM_RIGHT_ASYMMETRIC: 1957 if (i > sh->pd_idx) 1958 i--; 1959 break; 1960 case ALGORITHM_LEFT_SYMMETRIC: 1961 case ALGORITHM_RIGHT_SYMMETRIC: 1962 if (i < sh->pd_idx) 1963 i += raid_disks; 1964 i -= (sh->pd_idx + 1); 1965 break; 1966 case ALGORITHM_PARITY_0: 1967 i -= 1; 1968 break; 1969 case ALGORITHM_PARITY_N: 1970 break; 1971 default: 1972 BUG(); 1973 } 1974 break; 1975 case 6: 1976 if (i == sh->qd_idx) 1977 return 0; /* It is the Q disk */ 1978 switch (algorithm) { 1979 case ALGORITHM_LEFT_ASYMMETRIC: 1980 case ALGORITHM_RIGHT_ASYMMETRIC: 1981 case ALGORITHM_ROTATING_ZERO_RESTART: 1982 case ALGORITHM_ROTATING_N_RESTART: 1983 if (sh->pd_idx == raid_disks-1) 1984 i--; /* Q D D D P */ 1985 else if (i > sh->pd_idx) 1986 i -= 2; /* D D P Q D */ 1987 break; 1988 case ALGORITHM_LEFT_SYMMETRIC: 1989 case ALGORITHM_RIGHT_SYMMETRIC: 1990 if (sh->pd_idx == raid_disks-1) 1991 i--; /* Q D D D P */ 1992 else { 1993 /* D D P Q D */ 1994 if (i < sh->pd_idx) 1995 i += raid_disks; 1996 i -= (sh->pd_idx + 2); 1997 } 1998 break; 1999 case ALGORITHM_PARITY_0: 2000 i -= 2; 2001 break; 2002 case ALGORITHM_PARITY_N: 2003 break; 2004 case ALGORITHM_ROTATING_N_CONTINUE: 2005 /* Like left_symmetric, but P is before Q */ 2006 if (sh->pd_idx == 0) 2007 i--; /* P D D D Q */ 2008 else { 2009 /* D D Q P D */ 2010 if (i < sh->pd_idx) 2011 i += raid_disks; 2012 i -= (sh->pd_idx + 1); 2013 } 2014 break; 2015 case ALGORITHM_LEFT_ASYMMETRIC_6: 2016 case ALGORITHM_RIGHT_ASYMMETRIC_6: 2017 if (i > sh->pd_idx) 2018 i--; 2019 break; 2020 case ALGORITHM_LEFT_SYMMETRIC_6: 2021 case ALGORITHM_RIGHT_SYMMETRIC_6: 2022 if (i < sh->pd_idx) 2023 i += data_disks + 1; 2024 i -= (sh->pd_idx + 1); 2025 break; 2026 case ALGORITHM_PARITY_0_6: 2027 i -= 1; 2028 break; 2029 default: 2030 BUG(); 2031 } 2032 break; 2033 } 2034 2035 chunk_number = stripe * data_disks + i; 2036 r_sector = chunk_number * sectors_per_chunk + chunk_offset; 2037 2038 check = raid5_compute_sector(conf, r_sector, 2039 previous, &dummy1, &sh2); 2040 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx 2041 || sh2.qd_idx != sh->qd_idx) { 2042 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n", 2043 mdname(conf->mddev)); 2044 return 0; 2045 } 2046 return r_sector; 2047 } 2048 2049 2050 static void 2051 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s, 2052 int rcw, int expand) 2053 { 2054 int i, pd_idx = sh->pd_idx, disks = sh->disks; 2055 raid5_conf_t *conf = sh->raid_conf; 2056 int level = conf->level; 2057 2058 if (rcw) { 2059 /* if we are not expanding this is a proper write request, and 2060 * there will be bios with new data to be drained into the 2061 * stripe cache 2062 */ 2063 if (!expand) { 2064 sh->reconstruct_state = reconstruct_state_drain_run; 2065 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request); 2066 } else 2067 sh->reconstruct_state = reconstruct_state_run; 2068 2069 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request); 2070 2071 for (i = disks; i--; ) { 2072 struct r5dev *dev = &sh->dev[i]; 2073 2074 if (dev->towrite) { 2075 set_bit(R5_LOCKED, &dev->flags); 2076 set_bit(R5_Wantdrain, &dev->flags); 2077 if (!expand) 2078 clear_bit(R5_UPTODATE, &dev->flags); 2079 s->locked++; 2080 } 2081 } 2082 if (s->locked + conf->max_degraded == disks) 2083 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state)) 2084 atomic_inc(&conf->pending_full_writes); 2085 } else { 2086 BUG_ON(level == 6); 2087 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) || 2088 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags))); 2089 2090 sh->reconstruct_state = reconstruct_state_prexor_drain_run; 2091 set_bit(STRIPE_OP_PREXOR, &s->ops_request); 2092 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request); 2093 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request); 2094 2095 for (i = disks; i--; ) { 2096 struct r5dev *dev = &sh->dev[i]; 2097 if (i == pd_idx) 2098 continue; 2099 2100 if (dev->towrite && 2101 (test_bit(R5_UPTODATE, &dev->flags) || 2102 test_bit(R5_Wantcompute, &dev->flags))) { 2103 set_bit(R5_Wantdrain, &dev->flags); 2104 set_bit(R5_LOCKED, &dev->flags); 2105 clear_bit(R5_UPTODATE, &dev->flags); 2106 s->locked++; 2107 } 2108 } 2109 } 2110 2111 /* keep the parity disk(s) locked while asynchronous operations 2112 * are in flight 2113 */ 2114 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags); 2115 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags); 2116 s->locked++; 2117 2118 if (level == 6) { 2119 int qd_idx = sh->qd_idx; 2120 struct r5dev *dev = &sh->dev[qd_idx]; 2121 2122 set_bit(R5_LOCKED, &dev->flags); 2123 clear_bit(R5_UPTODATE, &dev->flags); 2124 s->locked++; 2125 } 2126 2127 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n", 2128 __func__, (unsigned long long)sh->sector, 2129 s->locked, s->ops_request); 2130 } 2131 2132 /* 2133 * Each stripe/dev can have one or more bion attached. 2134 * toread/towrite point to the first in a chain. 2135 * The bi_next chain must be in order. 2136 */ 2137 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite) 2138 { 2139 struct bio **bip; 2140 raid5_conf_t *conf = sh->raid_conf; 2141 int firstwrite=0; 2142 2143 pr_debug("adding bh b#%llu to stripe s#%llu\n", 2144 (unsigned long long)bi->bi_sector, 2145 (unsigned long long)sh->sector); 2146 2147 2148 spin_lock(&sh->lock); 2149 spin_lock_irq(&conf->device_lock); 2150 if (forwrite) { 2151 bip = &sh->dev[dd_idx].towrite; 2152 if (*bip == NULL && sh->dev[dd_idx].written == NULL) 2153 firstwrite = 1; 2154 } else 2155 bip = &sh->dev[dd_idx].toread; 2156 while (*bip && (*bip)->bi_sector < bi->bi_sector) { 2157 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector) 2158 goto overlap; 2159 bip = & (*bip)->bi_next; 2160 } 2161 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9)) 2162 goto overlap; 2163 2164 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next); 2165 if (*bip) 2166 bi->bi_next = *bip; 2167 *bip = bi; 2168 bi->bi_phys_segments++; 2169 spin_unlock_irq(&conf->device_lock); 2170 spin_unlock(&sh->lock); 2171 2172 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n", 2173 (unsigned long long)bi->bi_sector, 2174 (unsigned long long)sh->sector, dd_idx); 2175 2176 if (conf->mddev->bitmap && firstwrite) { 2177 bitmap_startwrite(conf->mddev->bitmap, sh->sector, 2178 STRIPE_SECTORS, 0); 2179 sh->bm_seq = conf->seq_flush+1; 2180 set_bit(STRIPE_BIT_DELAY, &sh->state); 2181 } 2182 2183 if (forwrite) { 2184 /* check if page is covered */ 2185 sector_t sector = sh->dev[dd_idx].sector; 2186 for (bi=sh->dev[dd_idx].towrite; 2187 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS && 2188 bi && bi->bi_sector <= sector; 2189 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) { 2190 if (bi->bi_sector + (bi->bi_size>>9) >= sector) 2191 sector = bi->bi_sector + (bi->bi_size>>9); 2192 } 2193 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS) 2194 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags); 2195 } 2196 return 1; 2197 2198 overlap: 2199 set_bit(R5_Overlap, &sh->dev[dd_idx].flags); 2200 spin_unlock_irq(&conf->device_lock); 2201 spin_unlock(&sh->lock); 2202 return 0; 2203 } 2204 2205 static void end_reshape(raid5_conf_t *conf); 2206 2207 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous, 2208 struct stripe_head *sh) 2209 { 2210 int sectors_per_chunk = 2211 previous ? conf->prev_chunk_sectors : conf->chunk_sectors; 2212 int dd_idx; 2213 int chunk_offset = sector_div(stripe, sectors_per_chunk); 2214 int disks = previous ? conf->previous_raid_disks : conf->raid_disks; 2215 2216 raid5_compute_sector(conf, 2217 stripe * (disks - conf->max_degraded) 2218 *sectors_per_chunk + chunk_offset, 2219 previous, 2220 &dd_idx, sh); 2221 } 2222 2223 static void 2224 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh, 2225 struct stripe_head_state *s, int disks, 2226 struct bio **return_bi) 2227 { 2228 int i; 2229 for (i = disks; i--; ) { 2230 struct bio *bi; 2231 int bitmap_end = 0; 2232 2233 if (test_bit(R5_ReadError, &sh->dev[i].flags)) { 2234 mdk_rdev_t *rdev; 2235 rcu_read_lock(); 2236 rdev = rcu_dereference(conf->disks[i].rdev); 2237 if (rdev && test_bit(In_sync, &rdev->flags)) 2238 /* multiple read failures in one stripe */ 2239 md_error(conf->mddev, rdev); 2240 rcu_read_unlock(); 2241 } 2242 spin_lock_irq(&conf->device_lock); 2243 /* fail all writes first */ 2244 bi = sh->dev[i].towrite; 2245 sh->dev[i].towrite = NULL; 2246 if (bi) { 2247 s->to_write--; 2248 bitmap_end = 1; 2249 } 2250 2251 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) 2252 wake_up(&conf->wait_for_overlap); 2253 2254 while (bi && bi->bi_sector < 2255 sh->dev[i].sector + STRIPE_SECTORS) { 2256 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector); 2257 clear_bit(BIO_UPTODATE, &bi->bi_flags); 2258 if (!raid5_dec_bi_phys_segments(bi)) { 2259 md_write_end(conf->mddev); 2260 bi->bi_next = *return_bi; 2261 *return_bi = bi; 2262 } 2263 bi = nextbi; 2264 } 2265 /* and fail all 'written' */ 2266 bi = sh->dev[i].written; 2267 sh->dev[i].written = NULL; 2268 if (bi) bitmap_end = 1; 2269 while (bi && bi->bi_sector < 2270 sh->dev[i].sector + STRIPE_SECTORS) { 2271 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector); 2272 clear_bit(BIO_UPTODATE, &bi->bi_flags); 2273 if (!raid5_dec_bi_phys_segments(bi)) { 2274 md_write_end(conf->mddev); 2275 bi->bi_next = *return_bi; 2276 *return_bi = bi; 2277 } 2278 bi = bi2; 2279 } 2280 2281 /* fail any reads if this device is non-operational and 2282 * the data has not reached the cache yet. 2283 */ 2284 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) && 2285 (!test_bit(R5_Insync, &sh->dev[i].flags) || 2286 test_bit(R5_ReadError, &sh->dev[i].flags))) { 2287 bi = sh->dev[i].toread; 2288 sh->dev[i].toread = NULL; 2289 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) 2290 wake_up(&conf->wait_for_overlap); 2291 if (bi) s->to_read--; 2292 while (bi && bi->bi_sector < 2293 sh->dev[i].sector + STRIPE_SECTORS) { 2294 struct bio *nextbi = 2295 r5_next_bio(bi, sh->dev[i].sector); 2296 clear_bit(BIO_UPTODATE, &bi->bi_flags); 2297 if (!raid5_dec_bi_phys_segments(bi)) { 2298 bi->bi_next = *return_bi; 2299 *return_bi = bi; 2300 } 2301 bi = nextbi; 2302 } 2303 } 2304 spin_unlock_irq(&conf->device_lock); 2305 if (bitmap_end) 2306 bitmap_endwrite(conf->mddev->bitmap, sh->sector, 2307 STRIPE_SECTORS, 0, 0); 2308 } 2309 2310 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state)) 2311 if (atomic_dec_and_test(&conf->pending_full_writes)) 2312 md_wakeup_thread(conf->mddev->thread); 2313 } 2314 2315 /* fetch_block5 - checks the given member device to see if its data needs 2316 * to be read or computed to satisfy a request. 2317 * 2318 * Returns 1 when no more member devices need to be checked, otherwise returns 2319 * 0 to tell the loop in handle_stripe_fill5 to continue 2320 */ 2321 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s, 2322 int disk_idx, int disks) 2323 { 2324 struct r5dev *dev = &sh->dev[disk_idx]; 2325 struct r5dev *failed_dev = &sh->dev[s->failed_num]; 2326 2327 /* is the data in this block needed, and can we get it? */ 2328 if (!test_bit(R5_LOCKED, &dev->flags) && 2329 !test_bit(R5_UPTODATE, &dev->flags) && 2330 (dev->toread || 2331 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) || 2332 s->syncing || s->expanding || 2333 (s->failed && 2334 (failed_dev->toread || 2335 (failed_dev->towrite && 2336 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) { 2337 /* We would like to get this block, possibly by computing it, 2338 * otherwise read it if the backing disk is insync 2339 */ 2340 if ((s->uptodate == disks - 1) && 2341 (s->failed && disk_idx == s->failed_num)) { 2342 set_bit(STRIPE_COMPUTE_RUN, &sh->state); 2343 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request); 2344 set_bit(R5_Wantcompute, &dev->flags); 2345 sh->ops.target = disk_idx; 2346 sh->ops.target2 = -1; 2347 s->req_compute = 1; 2348 /* Careful: from this point on 'uptodate' is in the eye 2349 * of raid_run_ops which services 'compute' operations 2350 * before writes. R5_Wantcompute flags a block that will 2351 * be R5_UPTODATE by the time it is needed for a 2352 * subsequent operation. 2353 */ 2354 s->uptodate++; 2355 return 1; /* uptodate + compute == disks */ 2356 } else if (test_bit(R5_Insync, &dev->flags)) { 2357 set_bit(R5_LOCKED, &dev->flags); 2358 set_bit(R5_Wantread, &dev->flags); 2359 s->locked++; 2360 pr_debug("Reading block %d (sync=%d)\n", disk_idx, 2361 s->syncing); 2362 } 2363 } 2364 2365 return 0; 2366 } 2367 2368 /** 2369 * handle_stripe_fill5 - read or compute data to satisfy pending requests. 2370 */ 2371 static void handle_stripe_fill5(struct stripe_head *sh, 2372 struct stripe_head_state *s, int disks) 2373 { 2374 int i; 2375 2376 /* look for blocks to read/compute, skip this if a compute 2377 * is already in flight, or if the stripe contents are in the 2378 * midst of changing due to a write 2379 */ 2380 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state && 2381 !sh->reconstruct_state) 2382 for (i = disks; i--; ) 2383 if (fetch_block5(sh, s, i, disks)) 2384 break; 2385 set_bit(STRIPE_HANDLE, &sh->state); 2386 } 2387 2388 /* fetch_block6 - checks the given member device to see if its data needs 2389 * to be read or computed to satisfy a request. 2390 * 2391 * Returns 1 when no more member devices need to be checked, otherwise returns 2392 * 0 to tell the loop in handle_stripe_fill6 to continue 2393 */ 2394 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s, 2395 struct r6_state *r6s, int disk_idx, int disks) 2396 { 2397 struct r5dev *dev = &sh->dev[disk_idx]; 2398 struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]], 2399 &sh->dev[r6s->failed_num[1]] }; 2400 2401 if (!test_bit(R5_LOCKED, &dev->flags) && 2402 !test_bit(R5_UPTODATE, &dev->flags) && 2403 (dev->toread || 2404 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) || 2405 s->syncing || s->expanding || 2406 (s->failed >= 1 && 2407 (fdev[0]->toread || s->to_write)) || 2408 (s->failed >= 2 && 2409 (fdev[1]->toread || s->to_write)))) { 2410 /* we would like to get this block, possibly by computing it, 2411 * otherwise read it if the backing disk is insync 2412 */ 2413 BUG_ON(test_bit(R5_Wantcompute, &dev->flags)); 2414 BUG_ON(test_bit(R5_Wantread, &dev->flags)); 2415 if ((s->uptodate == disks - 1) && 2416 (s->failed && (disk_idx == r6s->failed_num[0] || 2417 disk_idx == r6s->failed_num[1]))) { 2418 /* have disk failed, and we're requested to fetch it; 2419 * do compute it 2420 */ 2421 pr_debug("Computing stripe %llu block %d\n", 2422 (unsigned long long)sh->sector, disk_idx); 2423 set_bit(STRIPE_COMPUTE_RUN, &sh->state); 2424 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request); 2425 set_bit(R5_Wantcompute, &dev->flags); 2426 sh->ops.target = disk_idx; 2427 sh->ops.target2 = -1; /* no 2nd target */ 2428 s->req_compute = 1; 2429 s->uptodate++; 2430 return 1; 2431 } else if (s->uptodate == disks-2 && s->failed >= 2) { 2432 /* Computing 2-failure is *very* expensive; only 2433 * do it if failed >= 2 2434 */ 2435 int other; 2436 for (other = disks; other--; ) { 2437 if (other == disk_idx) 2438 continue; 2439 if (!test_bit(R5_UPTODATE, 2440 &sh->dev[other].flags)) 2441 break; 2442 } 2443 BUG_ON(other < 0); 2444 pr_debug("Computing stripe %llu blocks %d,%d\n", 2445 (unsigned long long)sh->sector, 2446 disk_idx, other); 2447 set_bit(STRIPE_COMPUTE_RUN, &sh->state); 2448 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request); 2449 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags); 2450 set_bit(R5_Wantcompute, &sh->dev[other].flags); 2451 sh->ops.target = disk_idx; 2452 sh->ops.target2 = other; 2453 s->uptodate += 2; 2454 s->req_compute = 1; 2455 return 1; 2456 } else if (test_bit(R5_Insync, &dev->flags)) { 2457 set_bit(R5_LOCKED, &dev->flags); 2458 set_bit(R5_Wantread, &dev->flags); 2459 s->locked++; 2460 pr_debug("Reading block %d (sync=%d)\n", 2461 disk_idx, s->syncing); 2462 } 2463 } 2464 2465 return 0; 2466 } 2467 2468 /** 2469 * handle_stripe_fill6 - read or compute data to satisfy pending requests. 2470 */ 2471 static void handle_stripe_fill6(struct stripe_head *sh, 2472 struct stripe_head_state *s, struct r6_state *r6s, 2473 int disks) 2474 { 2475 int i; 2476 2477 /* look for blocks to read/compute, skip this if a compute 2478 * is already in flight, or if the stripe contents are in the 2479 * midst of changing due to a write 2480 */ 2481 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state && 2482 !sh->reconstruct_state) 2483 for (i = disks; i--; ) 2484 if (fetch_block6(sh, s, r6s, i, disks)) 2485 break; 2486 set_bit(STRIPE_HANDLE, &sh->state); 2487 } 2488 2489 2490 /* handle_stripe_clean_event 2491 * any written block on an uptodate or failed drive can be returned. 2492 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but 2493 * never LOCKED, so we don't need to test 'failed' directly. 2494 */ 2495 static void handle_stripe_clean_event(raid5_conf_t *conf, 2496 struct stripe_head *sh, int disks, struct bio **return_bi) 2497 { 2498 int i; 2499 struct r5dev *dev; 2500 2501 for (i = disks; i--; ) 2502 if (sh->dev[i].written) { 2503 dev = &sh->dev[i]; 2504 if (!test_bit(R5_LOCKED, &dev->flags) && 2505 test_bit(R5_UPTODATE, &dev->flags)) { 2506 /* We can return any write requests */ 2507 struct bio *wbi, *wbi2; 2508 int bitmap_end = 0; 2509 pr_debug("Return write for disc %d\n", i); 2510 spin_lock_irq(&conf->device_lock); 2511 wbi = dev->written; 2512 dev->written = NULL; 2513 while (wbi && wbi->bi_sector < 2514 dev->sector + STRIPE_SECTORS) { 2515 wbi2 = r5_next_bio(wbi, dev->sector); 2516 if (!raid5_dec_bi_phys_segments(wbi)) { 2517 md_write_end(conf->mddev); 2518 wbi->bi_next = *return_bi; 2519 *return_bi = wbi; 2520 } 2521 wbi = wbi2; 2522 } 2523 if (dev->towrite == NULL) 2524 bitmap_end = 1; 2525 spin_unlock_irq(&conf->device_lock); 2526 if (bitmap_end) 2527 bitmap_endwrite(conf->mddev->bitmap, 2528 sh->sector, 2529 STRIPE_SECTORS, 2530 !test_bit(STRIPE_DEGRADED, &sh->state), 2531 0); 2532 } 2533 } 2534 2535 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state)) 2536 if (atomic_dec_and_test(&conf->pending_full_writes)) 2537 md_wakeup_thread(conf->mddev->thread); 2538 } 2539 2540 static void handle_stripe_dirtying5(raid5_conf_t *conf, 2541 struct stripe_head *sh, struct stripe_head_state *s, int disks) 2542 { 2543 int rmw = 0, rcw = 0, i; 2544 for (i = disks; i--; ) { 2545 /* would I have to read this buffer for read_modify_write */ 2546 struct r5dev *dev = &sh->dev[i]; 2547 if ((dev->towrite || i == sh->pd_idx) && 2548 !test_bit(R5_LOCKED, &dev->flags) && 2549 !(test_bit(R5_UPTODATE, &dev->flags) || 2550 test_bit(R5_Wantcompute, &dev->flags))) { 2551 if (test_bit(R5_Insync, &dev->flags)) 2552 rmw++; 2553 else 2554 rmw += 2*disks; /* cannot read it */ 2555 } 2556 /* Would I have to read this buffer for reconstruct_write */ 2557 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx && 2558 !test_bit(R5_LOCKED, &dev->flags) && 2559 !(test_bit(R5_UPTODATE, &dev->flags) || 2560 test_bit(R5_Wantcompute, &dev->flags))) { 2561 if (test_bit(R5_Insync, &dev->flags)) rcw++; 2562 else 2563 rcw += 2*disks; 2564 } 2565 } 2566 pr_debug("for sector %llu, rmw=%d rcw=%d\n", 2567 (unsigned long long)sh->sector, rmw, rcw); 2568 set_bit(STRIPE_HANDLE, &sh->state); 2569 if (rmw < rcw && rmw > 0) 2570 /* prefer read-modify-write, but need to get some data */ 2571 for (i = disks; i--; ) { 2572 struct r5dev *dev = &sh->dev[i]; 2573 if ((dev->towrite || i == sh->pd_idx) && 2574 !test_bit(R5_LOCKED, &dev->flags) && 2575 !(test_bit(R5_UPTODATE, &dev->flags) || 2576 test_bit(R5_Wantcompute, &dev->flags)) && 2577 test_bit(R5_Insync, &dev->flags)) { 2578 if ( 2579 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 2580 pr_debug("Read_old block " 2581 "%d for r-m-w\n", i); 2582 set_bit(R5_LOCKED, &dev->flags); 2583 set_bit(R5_Wantread, &dev->flags); 2584 s->locked++; 2585 } else { 2586 set_bit(STRIPE_DELAYED, &sh->state); 2587 set_bit(STRIPE_HANDLE, &sh->state); 2588 } 2589 } 2590 } 2591 if (rcw <= rmw && rcw > 0) 2592 /* want reconstruct write, but need to get some data */ 2593 for (i = disks; i--; ) { 2594 struct r5dev *dev = &sh->dev[i]; 2595 if (!test_bit(R5_OVERWRITE, &dev->flags) && 2596 i != sh->pd_idx && 2597 !test_bit(R5_LOCKED, &dev->flags) && 2598 !(test_bit(R5_UPTODATE, &dev->flags) || 2599 test_bit(R5_Wantcompute, &dev->flags)) && 2600 test_bit(R5_Insync, &dev->flags)) { 2601 if ( 2602 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 2603 pr_debug("Read_old block " 2604 "%d for Reconstruct\n", i); 2605 set_bit(R5_LOCKED, &dev->flags); 2606 set_bit(R5_Wantread, &dev->flags); 2607 s->locked++; 2608 } else { 2609 set_bit(STRIPE_DELAYED, &sh->state); 2610 set_bit(STRIPE_HANDLE, &sh->state); 2611 } 2612 } 2613 } 2614 /* now if nothing is locked, and if we have enough data, 2615 * we can start a write request 2616 */ 2617 /* since handle_stripe can be called at any time we need to handle the 2618 * case where a compute block operation has been submitted and then a 2619 * subsequent call wants to start a write request. raid_run_ops only 2620 * handles the case where compute block and reconstruct are requested 2621 * simultaneously. If this is not the case then new writes need to be 2622 * held off until the compute completes. 2623 */ 2624 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) && 2625 (s->locked == 0 && (rcw == 0 || rmw == 0) && 2626 !test_bit(STRIPE_BIT_DELAY, &sh->state))) 2627 schedule_reconstruction(sh, s, rcw == 0, 0); 2628 } 2629 2630 static void handle_stripe_dirtying6(raid5_conf_t *conf, 2631 struct stripe_head *sh, struct stripe_head_state *s, 2632 struct r6_state *r6s, int disks) 2633 { 2634 int rcw = 0, pd_idx = sh->pd_idx, i; 2635 int qd_idx = sh->qd_idx; 2636 2637 set_bit(STRIPE_HANDLE, &sh->state); 2638 for (i = disks; i--; ) { 2639 struct r5dev *dev = &sh->dev[i]; 2640 /* check if we haven't enough data */ 2641 if (!test_bit(R5_OVERWRITE, &dev->flags) && 2642 i != pd_idx && i != qd_idx && 2643 !test_bit(R5_LOCKED, &dev->flags) && 2644 !(test_bit(R5_UPTODATE, &dev->flags) || 2645 test_bit(R5_Wantcompute, &dev->flags))) { 2646 rcw++; 2647 if (!test_bit(R5_Insync, &dev->flags)) 2648 continue; /* it's a failed drive */ 2649 2650 if ( 2651 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) { 2652 pr_debug("Read_old stripe %llu " 2653 "block %d for Reconstruct\n", 2654 (unsigned long long)sh->sector, i); 2655 set_bit(R5_LOCKED, &dev->flags); 2656 set_bit(R5_Wantread, &dev->flags); 2657 s->locked++; 2658 } else { 2659 pr_debug("Request delayed stripe %llu " 2660 "block %d for Reconstruct\n", 2661 (unsigned long long)sh->sector, i); 2662 set_bit(STRIPE_DELAYED, &sh->state); 2663 set_bit(STRIPE_HANDLE, &sh->state); 2664 } 2665 } 2666 } 2667 /* now if nothing is locked, and if we have enough data, we can start a 2668 * write request 2669 */ 2670 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) && 2671 s->locked == 0 && rcw == 0 && 2672 !test_bit(STRIPE_BIT_DELAY, &sh->state)) { 2673 schedule_reconstruction(sh, s, 1, 0); 2674 } 2675 } 2676 2677 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh, 2678 struct stripe_head_state *s, int disks) 2679 { 2680 struct r5dev *dev = NULL; 2681 2682 set_bit(STRIPE_HANDLE, &sh->state); 2683 2684 switch (sh->check_state) { 2685 case check_state_idle: 2686 /* start a new check operation if there are no failures */ 2687 if (s->failed == 0) { 2688 BUG_ON(s->uptodate != disks); 2689 sh->check_state = check_state_run; 2690 set_bit(STRIPE_OP_CHECK, &s->ops_request); 2691 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags); 2692 s->uptodate--; 2693 break; 2694 } 2695 dev = &sh->dev[s->failed_num]; 2696 /* fall through */ 2697 case check_state_compute_result: 2698 sh->check_state = check_state_idle; 2699 if (!dev) 2700 dev = &sh->dev[sh->pd_idx]; 2701 2702 /* check that a write has not made the stripe insync */ 2703 if (test_bit(STRIPE_INSYNC, &sh->state)) 2704 break; 2705 2706 /* either failed parity check, or recovery is happening */ 2707 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags)); 2708 BUG_ON(s->uptodate != disks); 2709 2710 set_bit(R5_LOCKED, &dev->flags); 2711 s->locked++; 2712 set_bit(R5_Wantwrite, &dev->flags); 2713 2714 clear_bit(STRIPE_DEGRADED, &sh->state); 2715 set_bit(STRIPE_INSYNC, &sh->state); 2716 break; 2717 case check_state_run: 2718 break; /* we will be called again upon completion */ 2719 case check_state_check_result: 2720 sh->check_state = check_state_idle; 2721 2722 /* if a failure occurred during the check operation, leave 2723 * STRIPE_INSYNC not set and let the stripe be handled again 2724 */ 2725 if (s->failed) 2726 break; 2727 2728 /* handle a successful check operation, if parity is correct 2729 * we are done. Otherwise update the mismatch count and repair 2730 * parity if !MD_RECOVERY_CHECK 2731 */ 2732 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0) 2733 /* parity is correct (on disc, 2734 * not in buffer any more) 2735 */ 2736 set_bit(STRIPE_INSYNC, &sh->state); 2737 else { 2738 conf->mddev->resync_mismatches += STRIPE_SECTORS; 2739 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) 2740 /* don't try to repair!! */ 2741 set_bit(STRIPE_INSYNC, &sh->state); 2742 else { 2743 sh->check_state = check_state_compute_run; 2744 set_bit(STRIPE_COMPUTE_RUN, &sh->state); 2745 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request); 2746 set_bit(R5_Wantcompute, 2747 &sh->dev[sh->pd_idx].flags); 2748 sh->ops.target = sh->pd_idx; 2749 sh->ops.target2 = -1; 2750 s->uptodate++; 2751 } 2752 } 2753 break; 2754 case check_state_compute_run: 2755 break; 2756 default: 2757 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n", 2758 __func__, sh->check_state, 2759 (unsigned long long) sh->sector); 2760 BUG(); 2761 } 2762 } 2763 2764 2765 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh, 2766 struct stripe_head_state *s, 2767 struct r6_state *r6s, int disks) 2768 { 2769 int pd_idx = sh->pd_idx; 2770 int qd_idx = sh->qd_idx; 2771 struct r5dev *dev; 2772 2773 set_bit(STRIPE_HANDLE, &sh->state); 2774 2775 BUG_ON(s->failed > 2); 2776 2777 /* Want to check and possibly repair P and Q. 2778 * However there could be one 'failed' device, in which 2779 * case we can only check one of them, possibly using the 2780 * other to generate missing data 2781 */ 2782 2783 switch (sh->check_state) { 2784 case check_state_idle: 2785 /* start a new check operation if there are < 2 failures */ 2786 if (s->failed == r6s->q_failed) { 2787 /* The only possible failed device holds Q, so it 2788 * makes sense to check P (If anything else were failed, 2789 * we would have used P to recreate it). 2790 */ 2791 sh->check_state = check_state_run; 2792 } 2793 if (!r6s->q_failed && s->failed < 2) { 2794 /* Q is not failed, and we didn't use it to generate 2795 * anything, so it makes sense to check it 2796 */ 2797 if (sh->check_state == check_state_run) 2798 sh->check_state = check_state_run_pq; 2799 else 2800 sh->check_state = check_state_run_q; 2801 } 2802 2803 /* discard potentially stale zero_sum_result */ 2804 sh->ops.zero_sum_result = 0; 2805 2806 if (sh->check_state == check_state_run) { 2807 /* async_xor_zero_sum destroys the contents of P */ 2808 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags); 2809 s->uptodate--; 2810 } 2811 if (sh->check_state >= check_state_run && 2812 sh->check_state <= check_state_run_pq) { 2813 /* async_syndrome_zero_sum preserves P and Q, so 2814 * no need to mark them !uptodate here 2815 */ 2816 set_bit(STRIPE_OP_CHECK, &s->ops_request); 2817 break; 2818 } 2819 2820 /* we have 2-disk failure */ 2821 BUG_ON(s->failed != 2); 2822 /* fall through */ 2823 case check_state_compute_result: 2824 sh->check_state = check_state_idle; 2825 2826 /* check that a write has not made the stripe insync */ 2827 if (test_bit(STRIPE_INSYNC, &sh->state)) 2828 break; 2829 2830 /* now write out any block on a failed drive, 2831 * or P or Q if they were recomputed 2832 */ 2833 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */ 2834 if (s->failed == 2) { 2835 dev = &sh->dev[r6s->failed_num[1]]; 2836 s->locked++; 2837 set_bit(R5_LOCKED, &dev->flags); 2838 set_bit(R5_Wantwrite, &dev->flags); 2839 } 2840 if (s->failed >= 1) { 2841 dev = &sh->dev[r6s->failed_num[0]]; 2842 s->locked++; 2843 set_bit(R5_LOCKED, &dev->flags); 2844 set_bit(R5_Wantwrite, &dev->flags); 2845 } 2846 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) { 2847 dev = &sh->dev[pd_idx]; 2848 s->locked++; 2849 set_bit(R5_LOCKED, &dev->flags); 2850 set_bit(R5_Wantwrite, &dev->flags); 2851 } 2852 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) { 2853 dev = &sh->dev[qd_idx]; 2854 s->locked++; 2855 set_bit(R5_LOCKED, &dev->flags); 2856 set_bit(R5_Wantwrite, &dev->flags); 2857 } 2858 clear_bit(STRIPE_DEGRADED, &sh->state); 2859 2860 set_bit(STRIPE_INSYNC, &sh->state); 2861 break; 2862 case check_state_run: 2863 case check_state_run_q: 2864 case check_state_run_pq: 2865 break; /* we will be called again upon completion */ 2866 case check_state_check_result: 2867 sh->check_state = check_state_idle; 2868 2869 /* handle a successful check operation, if parity is correct 2870 * we are done. Otherwise update the mismatch count and repair 2871 * parity if !MD_RECOVERY_CHECK 2872 */ 2873 if (sh->ops.zero_sum_result == 0) { 2874 /* both parities are correct */ 2875 if (!s->failed) 2876 set_bit(STRIPE_INSYNC, &sh->state); 2877 else { 2878 /* in contrast to the raid5 case we can validate 2879 * parity, but still have a failure to write 2880 * back 2881 */ 2882 sh->check_state = check_state_compute_result; 2883 /* Returning at this point means that we may go 2884 * off and bring p and/or q uptodate again so 2885 * we make sure to check zero_sum_result again 2886 * to verify if p or q need writeback 2887 */ 2888 } 2889 } else { 2890 conf->mddev->resync_mismatches += STRIPE_SECTORS; 2891 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) 2892 /* don't try to repair!! */ 2893 set_bit(STRIPE_INSYNC, &sh->state); 2894 else { 2895 int *target = &sh->ops.target; 2896 2897 sh->ops.target = -1; 2898 sh->ops.target2 = -1; 2899 sh->check_state = check_state_compute_run; 2900 set_bit(STRIPE_COMPUTE_RUN, &sh->state); 2901 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request); 2902 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) { 2903 set_bit(R5_Wantcompute, 2904 &sh->dev[pd_idx].flags); 2905 *target = pd_idx; 2906 target = &sh->ops.target2; 2907 s->uptodate++; 2908 } 2909 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) { 2910 set_bit(R5_Wantcompute, 2911 &sh->dev[qd_idx].flags); 2912 *target = qd_idx; 2913 s->uptodate++; 2914 } 2915 } 2916 } 2917 break; 2918 case check_state_compute_run: 2919 break; 2920 default: 2921 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n", 2922 __func__, sh->check_state, 2923 (unsigned long long) sh->sector); 2924 BUG(); 2925 } 2926 } 2927 2928 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh, 2929 struct r6_state *r6s) 2930 { 2931 int i; 2932 2933 /* We have read all the blocks in this stripe and now we need to 2934 * copy some of them into a target stripe for expand. 2935 */ 2936 struct dma_async_tx_descriptor *tx = NULL; 2937 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state); 2938 for (i = 0; i < sh->disks; i++) 2939 if (i != sh->pd_idx && i != sh->qd_idx) { 2940 int dd_idx, j; 2941 struct stripe_head *sh2; 2942 struct async_submit_ctl submit; 2943 2944 sector_t bn = compute_blocknr(sh, i, 1); 2945 sector_t s = raid5_compute_sector(conf, bn, 0, 2946 &dd_idx, NULL); 2947 sh2 = get_active_stripe(conf, s, 0, 1, 1); 2948 if (sh2 == NULL) 2949 /* so far only the early blocks of this stripe 2950 * have been requested. When later blocks 2951 * get requested, we will try again 2952 */ 2953 continue; 2954 if (!test_bit(STRIPE_EXPANDING, &sh2->state) || 2955 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) { 2956 /* must have already done this block */ 2957 release_stripe(sh2); 2958 continue; 2959 } 2960 2961 /* place all the copies on one channel */ 2962 init_async_submit(&submit, 0, tx, NULL, NULL, NULL); 2963 tx = async_memcpy(sh2->dev[dd_idx].page, 2964 sh->dev[i].page, 0, 0, STRIPE_SIZE, 2965 &submit); 2966 2967 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags); 2968 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags); 2969 for (j = 0; j < conf->raid_disks; j++) 2970 if (j != sh2->pd_idx && 2971 (!r6s || j != sh2->qd_idx) && 2972 !test_bit(R5_Expanded, &sh2->dev[j].flags)) 2973 break; 2974 if (j == conf->raid_disks) { 2975 set_bit(STRIPE_EXPAND_READY, &sh2->state); 2976 set_bit(STRIPE_HANDLE, &sh2->state); 2977 } 2978 release_stripe(sh2); 2979 2980 } 2981 /* done submitting copies, wait for them to complete */ 2982 if (tx) { 2983 async_tx_ack(tx); 2984 dma_wait_for_async_tx(tx); 2985 } 2986 } 2987 2988 2989 /* 2990 * handle_stripe - do things to a stripe. 2991 * 2992 * We lock the stripe and then examine the state of various bits 2993 * to see what needs to be done. 2994 * Possible results: 2995 * return some read request which now have data 2996 * return some write requests which are safely on disc 2997 * schedule a read on some buffers 2998 * schedule a write of some buffers 2999 * return confirmation of parity correctness 3000 * 3001 * buffers are taken off read_list or write_list, and bh_cache buffers 3002 * get BH_Lock set before the stripe lock is released. 3003 * 3004 */ 3005 3006 static void handle_stripe5(struct stripe_head *sh) 3007 { 3008 raid5_conf_t *conf = sh->raid_conf; 3009 int disks = sh->disks, i; 3010 struct bio *return_bi = NULL; 3011 struct stripe_head_state s; 3012 struct r5dev *dev; 3013 mdk_rdev_t *blocked_rdev = NULL; 3014 int prexor; 3015 int dec_preread_active = 0; 3016 3017 memset(&s, 0, sizeof(s)); 3018 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d " 3019 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state, 3020 atomic_read(&sh->count), sh->pd_idx, sh->check_state, 3021 sh->reconstruct_state); 3022 3023 spin_lock(&sh->lock); 3024 clear_bit(STRIPE_HANDLE, &sh->state); 3025 clear_bit(STRIPE_DELAYED, &sh->state); 3026 3027 s.syncing = test_bit(STRIPE_SYNCING, &sh->state); 3028 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state); 3029 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state); 3030 3031 /* Now to look around and see what can be done */ 3032 rcu_read_lock(); 3033 for (i=disks; i--; ) { 3034 mdk_rdev_t *rdev; 3035 3036 dev = &sh->dev[i]; 3037 3038 pr_debug("check %d: state 0x%lx toread %p read %p write %p " 3039 "written %p\n", i, dev->flags, dev->toread, dev->read, 3040 dev->towrite, dev->written); 3041 3042 /* maybe we can request a biofill operation 3043 * 3044 * new wantfill requests are only permitted while 3045 * ops_complete_biofill is guaranteed to be inactive 3046 */ 3047 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread && 3048 !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) 3049 set_bit(R5_Wantfill, &dev->flags); 3050 3051 /* now count some things */ 3052 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++; 3053 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++; 3054 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++; 3055 3056 if (test_bit(R5_Wantfill, &dev->flags)) 3057 s.to_fill++; 3058 else if (dev->toread) 3059 s.to_read++; 3060 if (dev->towrite) { 3061 s.to_write++; 3062 if (!test_bit(R5_OVERWRITE, &dev->flags)) 3063 s.non_overwrite++; 3064 } 3065 if (dev->written) 3066 s.written++; 3067 rdev = rcu_dereference(conf->disks[i].rdev); 3068 if (blocked_rdev == NULL && 3069 rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 3070 blocked_rdev = rdev; 3071 atomic_inc(&rdev->nr_pending); 3072 } 3073 clear_bit(R5_Insync, &dev->flags); 3074 if (!rdev) 3075 /* Not in-sync */; 3076 else if (test_bit(In_sync, &rdev->flags)) 3077 set_bit(R5_Insync, &dev->flags); 3078 else { 3079 /* could be in-sync depending on recovery/reshape status */ 3080 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset) 3081 set_bit(R5_Insync, &dev->flags); 3082 } 3083 if (!test_bit(R5_Insync, &dev->flags)) { 3084 /* The ReadError flag will just be confusing now */ 3085 clear_bit(R5_ReadError, &dev->flags); 3086 clear_bit(R5_ReWrite, &dev->flags); 3087 } 3088 if (test_bit(R5_ReadError, &dev->flags)) 3089 clear_bit(R5_Insync, &dev->flags); 3090 if (!test_bit(R5_Insync, &dev->flags)) { 3091 s.failed++; 3092 s.failed_num = i; 3093 } 3094 } 3095 rcu_read_unlock(); 3096 3097 if (unlikely(blocked_rdev)) { 3098 if (s.syncing || s.expanding || s.expanded || 3099 s.to_write || s.written) { 3100 set_bit(STRIPE_HANDLE, &sh->state); 3101 goto unlock; 3102 } 3103 /* There is nothing for the blocked_rdev to block */ 3104 rdev_dec_pending(blocked_rdev, conf->mddev); 3105 blocked_rdev = NULL; 3106 } 3107 3108 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) { 3109 set_bit(STRIPE_OP_BIOFILL, &s.ops_request); 3110 set_bit(STRIPE_BIOFILL_RUN, &sh->state); 3111 } 3112 3113 pr_debug("locked=%d uptodate=%d to_read=%d" 3114 " to_write=%d failed=%d failed_num=%d\n", 3115 s.locked, s.uptodate, s.to_read, s.to_write, 3116 s.failed, s.failed_num); 3117 /* check if the array has lost two devices and, if so, some requests might 3118 * need to be failed 3119 */ 3120 if (s.failed > 1 && s.to_read+s.to_write+s.written) 3121 handle_failed_stripe(conf, sh, &s, disks, &return_bi); 3122 if (s.failed > 1 && s.syncing) { 3123 md_done_sync(conf->mddev, STRIPE_SECTORS,0); 3124 clear_bit(STRIPE_SYNCING, &sh->state); 3125 s.syncing = 0; 3126 } 3127 3128 /* might be able to return some write requests if the parity block 3129 * is safe, or on a failed drive 3130 */ 3131 dev = &sh->dev[sh->pd_idx]; 3132 if ( s.written && 3133 ((test_bit(R5_Insync, &dev->flags) && 3134 !test_bit(R5_LOCKED, &dev->flags) && 3135 test_bit(R5_UPTODATE, &dev->flags)) || 3136 (s.failed == 1 && s.failed_num == sh->pd_idx))) 3137 handle_stripe_clean_event(conf, sh, disks, &return_bi); 3138 3139 /* Now we might consider reading some blocks, either to check/generate 3140 * parity, or to satisfy requests 3141 * or to load a block that is being partially written. 3142 */ 3143 if (s.to_read || s.non_overwrite || 3144 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding) 3145 handle_stripe_fill5(sh, &s, disks); 3146 3147 /* Now we check to see if any write operations have recently 3148 * completed 3149 */ 3150 prexor = 0; 3151 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result) 3152 prexor = 1; 3153 if (sh->reconstruct_state == reconstruct_state_drain_result || 3154 sh->reconstruct_state == reconstruct_state_prexor_drain_result) { 3155 sh->reconstruct_state = reconstruct_state_idle; 3156 3157 /* All the 'written' buffers and the parity block are ready to 3158 * be written back to disk 3159 */ 3160 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags)); 3161 for (i = disks; i--; ) { 3162 dev = &sh->dev[i]; 3163 if (test_bit(R5_LOCKED, &dev->flags) && 3164 (i == sh->pd_idx || dev->written)) { 3165 pr_debug("Writing block %d\n", i); 3166 set_bit(R5_Wantwrite, &dev->flags); 3167 if (prexor) 3168 continue; 3169 if (!test_bit(R5_Insync, &dev->flags) || 3170 (i == sh->pd_idx && s.failed == 0)) 3171 set_bit(STRIPE_INSYNC, &sh->state); 3172 } 3173 } 3174 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 3175 dec_preread_active = 1; 3176 } 3177 3178 /* Now to consider new write requests and what else, if anything 3179 * should be read. We do not handle new writes when: 3180 * 1/ A 'write' operation (copy+xor) is already in flight. 3181 * 2/ A 'check' operation is in flight, as it may clobber the parity 3182 * block. 3183 */ 3184 if (s.to_write && !sh->reconstruct_state && !sh->check_state) 3185 handle_stripe_dirtying5(conf, sh, &s, disks); 3186 3187 /* maybe we need to check and possibly fix the parity for this stripe 3188 * Any reads will already have been scheduled, so we just see if enough 3189 * data is available. The parity check is held off while parity 3190 * dependent operations are in flight. 3191 */ 3192 if (sh->check_state || 3193 (s.syncing && s.locked == 0 && 3194 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) && 3195 !test_bit(STRIPE_INSYNC, &sh->state))) 3196 handle_parity_checks5(conf, sh, &s, disks); 3197 3198 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) { 3199 md_done_sync(conf->mddev, STRIPE_SECTORS,1); 3200 clear_bit(STRIPE_SYNCING, &sh->state); 3201 } 3202 3203 /* If the failed drive is just a ReadError, then we might need to progress 3204 * the repair/check process 3205 */ 3206 if (s.failed == 1 && !conf->mddev->ro && 3207 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags) 3208 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags) 3209 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags) 3210 ) { 3211 dev = &sh->dev[s.failed_num]; 3212 if (!test_bit(R5_ReWrite, &dev->flags)) { 3213 set_bit(R5_Wantwrite, &dev->flags); 3214 set_bit(R5_ReWrite, &dev->flags); 3215 set_bit(R5_LOCKED, &dev->flags); 3216 s.locked++; 3217 } else { 3218 /* let's read it back */ 3219 set_bit(R5_Wantread, &dev->flags); 3220 set_bit(R5_LOCKED, &dev->flags); 3221 s.locked++; 3222 } 3223 } 3224 3225 /* Finish reconstruct operations initiated by the expansion process */ 3226 if (sh->reconstruct_state == reconstruct_state_result) { 3227 struct stripe_head *sh2 3228 = get_active_stripe(conf, sh->sector, 1, 1, 1); 3229 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) { 3230 /* sh cannot be written until sh2 has been read. 3231 * so arrange for sh to be delayed a little 3232 */ 3233 set_bit(STRIPE_DELAYED, &sh->state); 3234 set_bit(STRIPE_HANDLE, &sh->state); 3235 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, 3236 &sh2->state)) 3237 atomic_inc(&conf->preread_active_stripes); 3238 release_stripe(sh2); 3239 goto unlock; 3240 } 3241 if (sh2) 3242 release_stripe(sh2); 3243 3244 sh->reconstruct_state = reconstruct_state_idle; 3245 clear_bit(STRIPE_EXPANDING, &sh->state); 3246 for (i = conf->raid_disks; i--; ) { 3247 set_bit(R5_Wantwrite, &sh->dev[i].flags); 3248 set_bit(R5_LOCKED, &sh->dev[i].flags); 3249 s.locked++; 3250 } 3251 } 3252 3253 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) && 3254 !sh->reconstruct_state) { 3255 /* Need to write out all blocks after computing parity */ 3256 sh->disks = conf->raid_disks; 3257 stripe_set_idx(sh->sector, conf, 0, sh); 3258 schedule_reconstruction(sh, &s, 1, 1); 3259 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) { 3260 clear_bit(STRIPE_EXPAND_READY, &sh->state); 3261 atomic_dec(&conf->reshape_stripes); 3262 wake_up(&conf->wait_for_overlap); 3263 md_done_sync(conf->mddev, STRIPE_SECTORS, 1); 3264 } 3265 3266 if (s.expanding && s.locked == 0 && 3267 !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) 3268 handle_stripe_expansion(conf, sh, NULL); 3269 3270 unlock: 3271 spin_unlock(&sh->lock); 3272 3273 /* wait for this device to become unblocked */ 3274 if (unlikely(blocked_rdev)) 3275 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev); 3276 3277 if (s.ops_request) 3278 raid_run_ops(sh, s.ops_request); 3279 3280 ops_run_io(sh, &s); 3281 3282 if (dec_preread_active) { 3283 /* We delay this until after ops_run_io so that if make_request 3284 * is waiting on a barrier, it won't continue until the writes 3285 * have actually been submitted. 3286 */ 3287 atomic_dec(&conf->preread_active_stripes); 3288 if (atomic_read(&conf->preread_active_stripes) < 3289 IO_THRESHOLD) 3290 md_wakeup_thread(conf->mddev->thread); 3291 } 3292 return_io(return_bi); 3293 } 3294 3295 static void handle_stripe6(struct stripe_head *sh) 3296 { 3297 raid5_conf_t *conf = sh->raid_conf; 3298 int disks = sh->disks; 3299 struct bio *return_bi = NULL; 3300 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx; 3301 struct stripe_head_state s; 3302 struct r6_state r6s; 3303 struct r5dev *dev, *pdev, *qdev; 3304 mdk_rdev_t *blocked_rdev = NULL; 3305 int dec_preread_active = 0; 3306 3307 pr_debug("handling stripe %llu, state=%#lx cnt=%d, " 3308 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n", 3309 (unsigned long long)sh->sector, sh->state, 3310 atomic_read(&sh->count), pd_idx, qd_idx, 3311 sh->check_state, sh->reconstruct_state); 3312 memset(&s, 0, sizeof(s)); 3313 3314 spin_lock(&sh->lock); 3315 clear_bit(STRIPE_HANDLE, &sh->state); 3316 clear_bit(STRIPE_DELAYED, &sh->state); 3317 3318 s.syncing = test_bit(STRIPE_SYNCING, &sh->state); 3319 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state); 3320 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state); 3321 /* Now to look around and see what can be done */ 3322 3323 rcu_read_lock(); 3324 for (i=disks; i--; ) { 3325 mdk_rdev_t *rdev; 3326 dev = &sh->dev[i]; 3327 3328 pr_debug("check %d: state 0x%lx read %p write %p written %p\n", 3329 i, dev->flags, dev->toread, dev->towrite, dev->written); 3330 /* maybe we can reply to a read 3331 * 3332 * new wantfill requests are only permitted while 3333 * ops_complete_biofill is guaranteed to be inactive 3334 */ 3335 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread && 3336 !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) 3337 set_bit(R5_Wantfill, &dev->flags); 3338 3339 /* now count some things */ 3340 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++; 3341 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++; 3342 if (test_bit(R5_Wantcompute, &dev->flags)) { 3343 s.compute++; 3344 BUG_ON(s.compute > 2); 3345 } 3346 3347 if (test_bit(R5_Wantfill, &dev->flags)) { 3348 s.to_fill++; 3349 } else if (dev->toread) 3350 s.to_read++; 3351 if (dev->towrite) { 3352 s.to_write++; 3353 if (!test_bit(R5_OVERWRITE, &dev->flags)) 3354 s.non_overwrite++; 3355 } 3356 if (dev->written) 3357 s.written++; 3358 rdev = rcu_dereference(conf->disks[i].rdev); 3359 if (blocked_rdev == NULL && 3360 rdev && unlikely(test_bit(Blocked, &rdev->flags))) { 3361 blocked_rdev = rdev; 3362 atomic_inc(&rdev->nr_pending); 3363 } 3364 clear_bit(R5_Insync, &dev->flags); 3365 if (!rdev) 3366 /* Not in-sync */; 3367 else if (test_bit(In_sync, &rdev->flags)) 3368 set_bit(R5_Insync, &dev->flags); 3369 else { 3370 /* in sync if before recovery_offset */ 3371 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset) 3372 set_bit(R5_Insync, &dev->flags); 3373 } 3374 if (!test_bit(R5_Insync, &dev->flags)) { 3375 /* The ReadError flag will just be confusing now */ 3376 clear_bit(R5_ReadError, &dev->flags); 3377 clear_bit(R5_ReWrite, &dev->flags); 3378 } 3379 if (test_bit(R5_ReadError, &dev->flags)) 3380 clear_bit(R5_Insync, &dev->flags); 3381 if (!test_bit(R5_Insync, &dev->flags)) { 3382 if (s.failed < 2) 3383 r6s.failed_num[s.failed] = i; 3384 s.failed++; 3385 } 3386 } 3387 rcu_read_unlock(); 3388 3389 if (unlikely(blocked_rdev)) { 3390 if (s.syncing || s.expanding || s.expanded || 3391 s.to_write || s.written) { 3392 set_bit(STRIPE_HANDLE, &sh->state); 3393 goto unlock; 3394 } 3395 /* There is nothing for the blocked_rdev to block */ 3396 rdev_dec_pending(blocked_rdev, conf->mddev); 3397 blocked_rdev = NULL; 3398 } 3399 3400 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) { 3401 set_bit(STRIPE_OP_BIOFILL, &s.ops_request); 3402 set_bit(STRIPE_BIOFILL_RUN, &sh->state); 3403 } 3404 3405 pr_debug("locked=%d uptodate=%d to_read=%d" 3406 " to_write=%d failed=%d failed_num=%d,%d\n", 3407 s.locked, s.uptodate, s.to_read, s.to_write, s.failed, 3408 r6s.failed_num[0], r6s.failed_num[1]); 3409 /* check if the array has lost >2 devices and, if so, some requests 3410 * might need to be failed 3411 */ 3412 if (s.failed > 2 && s.to_read+s.to_write+s.written) 3413 handle_failed_stripe(conf, sh, &s, disks, &return_bi); 3414 if (s.failed > 2 && s.syncing) { 3415 md_done_sync(conf->mddev, STRIPE_SECTORS,0); 3416 clear_bit(STRIPE_SYNCING, &sh->state); 3417 s.syncing = 0; 3418 } 3419 3420 /* 3421 * might be able to return some write requests if the parity blocks 3422 * are safe, or on a failed drive 3423 */ 3424 pdev = &sh->dev[pd_idx]; 3425 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx) 3426 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx); 3427 qdev = &sh->dev[qd_idx]; 3428 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx) 3429 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx); 3430 3431 if ( s.written && 3432 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags) 3433 && !test_bit(R5_LOCKED, &pdev->flags) 3434 && test_bit(R5_UPTODATE, &pdev->flags)))) && 3435 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags) 3436 && !test_bit(R5_LOCKED, &qdev->flags) 3437 && test_bit(R5_UPTODATE, &qdev->flags))))) 3438 handle_stripe_clean_event(conf, sh, disks, &return_bi); 3439 3440 /* Now we might consider reading some blocks, either to check/generate 3441 * parity, or to satisfy requests 3442 * or to load a block that is being partially written. 3443 */ 3444 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) || 3445 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding) 3446 handle_stripe_fill6(sh, &s, &r6s, disks); 3447 3448 /* Now we check to see if any write operations have recently 3449 * completed 3450 */ 3451 if (sh->reconstruct_state == reconstruct_state_drain_result) { 3452 3453 sh->reconstruct_state = reconstruct_state_idle; 3454 /* All the 'written' buffers and the parity blocks are ready to 3455 * be written back to disk 3456 */ 3457 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags)); 3458 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags)); 3459 for (i = disks; i--; ) { 3460 dev = &sh->dev[i]; 3461 if (test_bit(R5_LOCKED, &dev->flags) && 3462 (i == sh->pd_idx || i == qd_idx || 3463 dev->written)) { 3464 pr_debug("Writing block %d\n", i); 3465 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags)); 3466 set_bit(R5_Wantwrite, &dev->flags); 3467 if (!test_bit(R5_Insync, &dev->flags) || 3468 ((i == sh->pd_idx || i == qd_idx) && 3469 s.failed == 0)) 3470 set_bit(STRIPE_INSYNC, &sh->state); 3471 } 3472 } 3473 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 3474 dec_preread_active = 1; 3475 } 3476 3477 /* Now to consider new write requests and what else, if anything 3478 * should be read. We do not handle new writes when: 3479 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight. 3480 * 2/ A 'check' operation is in flight, as it may clobber the parity 3481 * block. 3482 */ 3483 if (s.to_write && !sh->reconstruct_state && !sh->check_state) 3484 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks); 3485 3486 /* maybe we need to check and possibly fix the parity for this stripe 3487 * Any reads will already have been scheduled, so we just see if enough 3488 * data is available. The parity check is held off while parity 3489 * dependent operations are in flight. 3490 */ 3491 if (sh->check_state || 3492 (s.syncing && s.locked == 0 && 3493 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) && 3494 !test_bit(STRIPE_INSYNC, &sh->state))) 3495 handle_parity_checks6(conf, sh, &s, &r6s, disks); 3496 3497 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) { 3498 md_done_sync(conf->mddev, STRIPE_SECTORS,1); 3499 clear_bit(STRIPE_SYNCING, &sh->state); 3500 } 3501 3502 /* If the failed drives are just a ReadError, then we might need 3503 * to progress the repair/check process 3504 */ 3505 if (s.failed <= 2 && !conf->mddev->ro) 3506 for (i = 0; i < s.failed; i++) { 3507 dev = &sh->dev[r6s.failed_num[i]]; 3508 if (test_bit(R5_ReadError, &dev->flags) 3509 && !test_bit(R5_LOCKED, &dev->flags) 3510 && test_bit(R5_UPTODATE, &dev->flags) 3511 ) { 3512 if (!test_bit(R5_ReWrite, &dev->flags)) { 3513 set_bit(R5_Wantwrite, &dev->flags); 3514 set_bit(R5_ReWrite, &dev->flags); 3515 set_bit(R5_LOCKED, &dev->flags); 3516 s.locked++; 3517 } else { 3518 /* let's read it back */ 3519 set_bit(R5_Wantread, &dev->flags); 3520 set_bit(R5_LOCKED, &dev->flags); 3521 s.locked++; 3522 } 3523 } 3524 } 3525 3526 /* Finish reconstruct operations initiated by the expansion process */ 3527 if (sh->reconstruct_state == reconstruct_state_result) { 3528 sh->reconstruct_state = reconstruct_state_idle; 3529 clear_bit(STRIPE_EXPANDING, &sh->state); 3530 for (i = conf->raid_disks; i--; ) { 3531 set_bit(R5_Wantwrite, &sh->dev[i].flags); 3532 set_bit(R5_LOCKED, &sh->dev[i].flags); 3533 s.locked++; 3534 } 3535 } 3536 3537 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) && 3538 !sh->reconstruct_state) { 3539 struct stripe_head *sh2 3540 = get_active_stripe(conf, sh->sector, 1, 1, 1); 3541 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) { 3542 /* sh cannot be written until sh2 has been read. 3543 * so arrange for sh to be delayed a little 3544 */ 3545 set_bit(STRIPE_DELAYED, &sh->state); 3546 set_bit(STRIPE_HANDLE, &sh->state); 3547 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, 3548 &sh2->state)) 3549 atomic_inc(&conf->preread_active_stripes); 3550 release_stripe(sh2); 3551 goto unlock; 3552 } 3553 if (sh2) 3554 release_stripe(sh2); 3555 3556 /* Need to write out all blocks after computing P&Q */ 3557 sh->disks = conf->raid_disks; 3558 stripe_set_idx(sh->sector, conf, 0, sh); 3559 schedule_reconstruction(sh, &s, 1, 1); 3560 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) { 3561 clear_bit(STRIPE_EXPAND_READY, &sh->state); 3562 atomic_dec(&conf->reshape_stripes); 3563 wake_up(&conf->wait_for_overlap); 3564 md_done_sync(conf->mddev, STRIPE_SECTORS, 1); 3565 } 3566 3567 if (s.expanding && s.locked == 0 && 3568 !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) 3569 handle_stripe_expansion(conf, sh, &r6s); 3570 3571 unlock: 3572 spin_unlock(&sh->lock); 3573 3574 /* wait for this device to become unblocked */ 3575 if (unlikely(blocked_rdev)) 3576 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev); 3577 3578 if (s.ops_request) 3579 raid_run_ops(sh, s.ops_request); 3580 3581 ops_run_io(sh, &s); 3582 3583 3584 if (dec_preread_active) { 3585 /* We delay this until after ops_run_io so that if make_request 3586 * is waiting on a barrier, it won't continue until the writes 3587 * have actually been submitted. 3588 */ 3589 atomic_dec(&conf->preread_active_stripes); 3590 if (atomic_read(&conf->preread_active_stripes) < 3591 IO_THRESHOLD) 3592 md_wakeup_thread(conf->mddev->thread); 3593 } 3594 3595 return_io(return_bi); 3596 } 3597 3598 static void handle_stripe(struct stripe_head *sh) 3599 { 3600 if (sh->raid_conf->level == 6) 3601 handle_stripe6(sh); 3602 else 3603 handle_stripe5(sh); 3604 } 3605 3606 static void raid5_activate_delayed(raid5_conf_t *conf) 3607 { 3608 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) { 3609 while (!list_empty(&conf->delayed_list)) { 3610 struct list_head *l = conf->delayed_list.next; 3611 struct stripe_head *sh; 3612 sh = list_entry(l, struct stripe_head, lru); 3613 list_del_init(l); 3614 clear_bit(STRIPE_DELAYED, &sh->state); 3615 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 3616 atomic_inc(&conf->preread_active_stripes); 3617 list_add_tail(&sh->lru, &conf->hold_list); 3618 } 3619 } else 3620 plugger_set_plug(&conf->plug); 3621 } 3622 3623 static void activate_bit_delay(raid5_conf_t *conf) 3624 { 3625 /* device_lock is held */ 3626 struct list_head head; 3627 list_add(&head, &conf->bitmap_list); 3628 list_del_init(&conf->bitmap_list); 3629 while (!list_empty(&head)) { 3630 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru); 3631 list_del_init(&sh->lru); 3632 atomic_inc(&sh->count); 3633 __release_stripe(conf, sh); 3634 } 3635 } 3636 3637 static void unplug_slaves(mddev_t *mddev) 3638 { 3639 raid5_conf_t *conf = mddev->private; 3640 int i; 3641 int devs = max(conf->raid_disks, conf->previous_raid_disks); 3642 3643 rcu_read_lock(); 3644 for (i = 0; i < devs; i++) { 3645 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev); 3646 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) { 3647 struct request_queue *r_queue = bdev_get_queue(rdev->bdev); 3648 3649 atomic_inc(&rdev->nr_pending); 3650 rcu_read_unlock(); 3651 3652 blk_unplug(r_queue); 3653 3654 rdev_dec_pending(rdev, mddev); 3655 rcu_read_lock(); 3656 } 3657 } 3658 rcu_read_unlock(); 3659 } 3660 3661 void md_raid5_unplug_device(raid5_conf_t *conf) 3662 { 3663 unsigned long flags; 3664 3665 spin_lock_irqsave(&conf->device_lock, flags); 3666 3667 if (plugger_remove_plug(&conf->plug)) { 3668 conf->seq_flush++; 3669 raid5_activate_delayed(conf); 3670 } 3671 md_wakeup_thread(conf->mddev->thread); 3672 3673 spin_unlock_irqrestore(&conf->device_lock, flags); 3674 3675 unplug_slaves(conf->mddev); 3676 } 3677 EXPORT_SYMBOL_GPL(md_raid5_unplug_device); 3678 3679 static void raid5_unplug(struct plug_handle *plug) 3680 { 3681 raid5_conf_t *conf = container_of(plug, raid5_conf_t, plug); 3682 md_raid5_unplug_device(conf); 3683 } 3684 3685 static void raid5_unplug_queue(struct request_queue *q) 3686 { 3687 mddev_t *mddev = q->queuedata; 3688 md_raid5_unplug_device(mddev->private); 3689 } 3690 3691 int md_raid5_congested(mddev_t *mddev, int bits) 3692 { 3693 raid5_conf_t *conf = mddev->private; 3694 3695 /* No difference between reads and writes. Just check 3696 * how busy the stripe_cache is 3697 */ 3698 3699 if (conf->inactive_blocked) 3700 return 1; 3701 if (conf->quiesce) 3702 return 1; 3703 if (list_empty_careful(&conf->inactive_list)) 3704 return 1; 3705 3706 return 0; 3707 } 3708 EXPORT_SYMBOL_GPL(md_raid5_congested); 3709 3710 static int raid5_congested(void *data, int bits) 3711 { 3712 mddev_t *mddev = data; 3713 3714 return mddev_congested(mddev, bits) || 3715 md_raid5_congested(mddev, bits); 3716 } 3717 3718 /* We want read requests to align with chunks where possible, 3719 * but write requests don't need to. 3720 */ 3721 static int raid5_mergeable_bvec(struct request_queue *q, 3722 struct bvec_merge_data *bvm, 3723 struct bio_vec *biovec) 3724 { 3725 mddev_t *mddev = q->queuedata; 3726 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev); 3727 int max; 3728 unsigned int chunk_sectors = mddev->chunk_sectors; 3729 unsigned int bio_sectors = bvm->bi_size >> 9; 3730 3731 if ((bvm->bi_rw & 1) == WRITE) 3732 return biovec->bv_len; /* always allow writes to be mergeable */ 3733 3734 if (mddev->new_chunk_sectors < mddev->chunk_sectors) 3735 chunk_sectors = mddev->new_chunk_sectors; 3736 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9; 3737 if (max < 0) max = 0; 3738 if (max <= biovec->bv_len && bio_sectors == 0) 3739 return biovec->bv_len; 3740 else 3741 return max; 3742 } 3743 3744 3745 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio) 3746 { 3747 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev); 3748 unsigned int chunk_sectors = mddev->chunk_sectors; 3749 unsigned int bio_sectors = bio->bi_size >> 9; 3750 3751 if (mddev->new_chunk_sectors < mddev->chunk_sectors) 3752 chunk_sectors = mddev->new_chunk_sectors; 3753 return chunk_sectors >= 3754 ((sector & (chunk_sectors - 1)) + bio_sectors); 3755 } 3756 3757 /* 3758 * add bio to the retry LIFO ( in O(1) ... we are in interrupt ) 3759 * later sampled by raid5d. 3760 */ 3761 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf) 3762 { 3763 unsigned long flags; 3764 3765 spin_lock_irqsave(&conf->device_lock, flags); 3766 3767 bi->bi_next = conf->retry_read_aligned_list; 3768 conf->retry_read_aligned_list = bi; 3769 3770 spin_unlock_irqrestore(&conf->device_lock, flags); 3771 md_wakeup_thread(conf->mddev->thread); 3772 } 3773 3774 3775 static struct bio *remove_bio_from_retry(raid5_conf_t *conf) 3776 { 3777 struct bio *bi; 3778 3779 bi = conf->retry_read_aligned; 3780 if (bi) { 3781 conf->retry_read_aligned = NULL; 3782 return bi; 3783 } 3784 bi = conf->retry_read_aligned_list; 3785 if(bi) { 3786 conf->retry_read_aligned_list = bi->bi_next; 3787 bi->bi_next = NULL; 3788 /* 3789 * this sets the active strip count to 1 and the processed 3790 * strip count to zero (upper 8 bits) 3791 */ 3792 bi->bi_phys_segments = 1; /* biased count of active stripes */ 3793 } 3794 3795 return bi; 3796 } 3797 3798 3799 /* 3800 * The "raid5_align_endio" should check if the read succeeded and if it 3801 * did, call bio_endio on the original bio (having bio_put the new bio 3802 * first). 3803 * If the read failed.. 3804 */ 3805 static void raid5_align_endio(struct bio *bi, int error) 3806 { 3807 struct bio* raid_bi = bi->bi_private; 3808 mddev_t *mddev; 3809 raid5_conf_t *conf; 3810 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags); 3811 mdk_rdev_t *rdev; 3812 3813 bio_put(bi); 3814 3815 rdev = (void*)raid_bi->bi_next; 3816 raid_bi->bi_next = NULL; 3817 mddev = rdev->mddev; 3818 conf = mddev->private; 3819 3820 rdev_dec_pending(rdev, conf->mddev); 3821 3822 if (!error && uptodate) { 3823 bio_endio(raid_bi, 0); 3824 if (atomic_dec_and_test(&conf->active_aligned_reads)) 3825 wake_up(&conf->wait_for_stripe); 3826 return; 3827 } 3828 3829 3830 pr_debug("raid5_align_endio : io error...handing IO for a retry\n"); 3831 3832 add_bio_to_retry(raid_bi, conf); 3833 } 3834 3835 static int bio_fits_rdev(struct bio *bi) 3836 { 3837 struct request_queue *q = bdev_get_queue(bi->bi_bdev); 3838 3839 if ((bi->bi_size>>9) > queue_max_sectors(q)) 3840 return 0; 3841 blk_recount_segments(q, bi); 3842 if (bi->bi_phys_segments > queue_max_segments(q)) 3843 return 0; 3844 3845 if (q->merge_bvec_fn) 3846 /* it's too hard to apply the merge_bvec_fn at this stage, 3847 * just just give up 3848 */ 3849 return 0; 3850 3851 return 1; 3852 } 3853 3854 3855 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio) 3856 { 3857 raid5_conf_t *conf = mddev->private; 3858 int dd_idx; 3859 struct bio* align_bi; 3860 mdk_rdev_t *rdev; 3861 3862 if (!in_chunk_boundary(mddev, raid_bio)) { 3863 pr_debug("chunk_aligned_read : non aligned\n"); 3864 return 0; 3865 } 3866 /* 3867 * use bio_clone to make a copy of the bio 3868 */ 3869 align_bi = bio_clone(raid_bio, GFP_NOIO); 3870 if (!align_bi) 3871 return 0; 3872 /* 3873 * set bi_end_io to a new function, and set bi_private to the 3874 * original bio. 3875 */ 3876 align_bi->bi_end_io = raid5_align_endio; 3877 align_bi->bi_private = raid_bio; 3878 /* 3879 * compute position 3880 */ 3881 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector, 3882 0, 3883 &dd_idx, NULL); 3884 3885 rcu_read_lock(); 3886 rdev = rcu_dereference(conf->disks[dd_idx].rdev); 3887 if (rdev && test_bit(In_sync, &rdev->flags)) { 3888 atomic_inc(&rdev->nr_pending); 3889 rcu_read_unlock(); 3890 raid_bio->bi_next = (void*)rdev; 3891 align_bi->bi_bdev = rdev->bdev; 3892 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID); 3893 align_bi->bi_sector += rdev->data_offset; 3894 3895 if (!bio_fits_rdev(align_bi)) { 3896 /* too big in some way */ 3897 bio_put(align_bi); 3898 rdev_dec_pending(rdev, mddev); 3899 return 0; 3900 } 3901 3902 spin_lock_irq(&conf->device_lock); 3903 wait_event_lock_irq(conf->wait_for_stripe, 3904 conf->quiesce == 0, 3905 conf->device_lock, /* nothing */); 3906 atomic_inc(&conf->active_aligned_reads); 3907 spin_unlock_irq(&conf->device_lock); 3908 3909 generic_make_request(align_bi); 3910 return 1; 3911 } else { 3912 rcu_read_unlock(); 3913 bio_put(align_bi); 3914 return 0; 3915 } 3916 } 3917 3918 /* __get_priority_stripe - get the next stripe to process 3919 * 3920 * Full stripe writes are allowed to pass preread active stripes up until 3921 * the bypass_threshold is exceeded. In general the bypass_count 3922 * increments when the handle_list is handled before the hold_list; however, it 3923 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a 3924 * stripe with in flight i/o. The bypass_count will be reset when the 3925 * head of the hold_list has changed, i.e. the head was promoted to the 3926 * handle_list. 3927 */ 3928 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf) 3929 { 3930 struct stripe_head *sh; 3931 3932 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n", 3933 __func__, 3934 list_empty(&conf->handle_list) ? "empty" : "busy", 3935 list_empty(&conf->hold_list) ? "empty" : "busy", 3936 atomic_read(&conf->pending_full_writes), conf->bypass_count); 3937 3938 if (!list_empty(&conf->handle_list)) { 3939 sh = list_entry(conf->handle_list.next, typeof(*sh), lru); 3940 3941 if (list_empty(&conf->hold_list)) 3942 conf->bypass_count = 0; 3943 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) { 3944 if (conf->hold_list.next == conf->last_hold) 3945 conf->bypass_count++; 3946 else { 3947 conf->last_hold = conf->hold_list.next; 3948 conf->bypass_count -= conf->bypass_threshold; 3949 if (conf->bypass_count < 0) 3950 conf->bypass_count = 0; 3951 } 3952 } 3953 } else if (!list_empty(&conf->hold_list) && 3954 ((conf->bypass_threshold && 3955 conf->bypass_count > conf->bypass_threshold) || 3956 atomic_read(&conf->pending_full_writes) == 0)) { 3957 sh = list_entry(conf->hold_list.next, 3958 typeof(*sh), lru); 3959 conf->bypass_count -= conf->bypass_threshold; 3960 if (conf->bypass_count < 0) 3961 conf->bypass_count = 0; 3962 } else 3963 return NULL; 3964 3965 list_del_init(&sh->lru); 3966 atomic_inc(&sh->count); 3967 BUG_ON(atomic_read(&sh->count) != 1); 3968 return sh; 3969 } 3970 3971 static int make_request(mddev_t *mddev, struct bio * bi) 3972 { 3973 raid5_conf_t *conf = mddev->private; 3974 int dd_idx; 3975 sector_t new_sector; 3976 sector_t logical_sector, last_sector; 3977 struct stripe_head *sh; 3978 const int rw = bio_data_dir(bi); 3979 int remaining; 3980 3981 if (unlikely(bi->bi_rw & REQ_HARDBARRIER)) { 3982 /* Drain all pending writes. We only really need 3983 * to ensure they have been submitted, but this is 3984 * easier. 3985 */ 3986 mddev->pers->quiesce(mddev, 1); 3987 mddev->pers->quiesce(mddev, 0); 3988 md_barrier_request(mddev, bi); 3989 return 0; 3990 } 3991 3992 md_write_start(mddev, bi); 3993 3994 if (rw == READ && 3995 mddev->reshape_position == MaxSector && 3996 chunk_aligned_read(mddev,bi)) 3997 return 0; 3998 3999 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1); 4000 last_sector = bi->bi_sector + (bi->bi_size>>9); 4001 bi->bi_next = NULL; 4002 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */ 4003 4004 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) { 4005 DEFINE_WAIT(w); 4006 int disks, data_disks; 4007 int previous; 4008 4009 retry: 4010 previous = 0; 4011 disks = conf->raid_disks; 4012 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE); 4013 if (unlikely(conf->reshape_progress != MaxSector)) { 4014 /* spinlock is needed as reshape_progress may be 4015 * 64bit on a 32bit platform, and so it might be 4016 * possible to see a half-updated value 4017 * Ofcourse reshape_progress could change after 4018 * the lock is dropped, so once we get a reference 4019 * to the stripe that we think it is, we will have 4020 * to check again. 4021 */ 4022 spin_lock_irq(&conf->device_lock); 4023 if (mddev->delta_disks < 0 4024 ? logical_sector < conf->reshape_progress 4025 : logical_sector >= conf->reshape_progress) { 4026 disks = conf->previous_raid_disks; 4027 previous = 1; 4028 } else { 4029 if (mddev->delta_disks < 0 4030 ? logical_sector < conf->reshape_safe 4031 : logical_sector >= conf->reshape_safe) { 4032 spin_unlock_irq(&conf->device_lock); 4033 schedule(); 4034 goto retry; 4035 } 4036 } 4037 spin_unlock_irq(&conf->device_lock); 4038 } 4039 data_disks = disks - conf->max_degraded; 4040 4041 new_sector = raid5_compute_sector(conf, logical_sector, 4042 previous, 4043 &dd_idx, NULL); 4044 pr_debug("raid456: make_request, sector %llu logical %llu\n", 4045 (unsigned long long)new_sector, 4046 (unsigned long long)logical_sector); 4047 4048 sh = get_active_stripe(conf, new_sector, previous, 4049 (bi->bi_rw&RWA_MASK), 0); 4050 if (sh) { 4051 if (unlikely(previous)) { 4052 /* expansion might have moved on while waiting for a 4053 * stripe, so we must do the range check again. 4054 * Expansion could still move past after this 4055 * test, but as we are holding a reference to 4056 * 'sh', we know that if that happens, 4057 * STRIPE_EXPANDING will get set and the expansion 4058 * won't proceed until we finish with the stripe. 4059 */ 4060 int must_retry = 0; 4061 spin_lock_irq(&conf->device_lock); 4062 if (mddev->delta_disks < 0 4063 ? logical_sector >= conf->reshape_progress 4064 : logical_sector < conf->reshape_progress) 4065 /* mismatch, need to try again */ 4066 must_retry = 1; 4067 spin_unlock_irq(&conf->device_lock); 4068 if (must_retry) { 4069 release_stripe(sh); 4070 schedule(); 4071 goto retry; 4072 } 4073 } 4074 4075 if (bio_data_dir(bi) == WRITE && 4076 logical_sector >= mddev->suspend_lo && 4077 logical_sector < mddev->suspend_hi) { 4078 release_stripe(sh); 4079 /* As the suspend_* range is controlled by 4080 * userspace, we want an interruptible 4081 * wait. 4082 */ 4083 flush_signals(current); 4084 prepare_to_wait(&conf->wait_for_overlap, 4085 &w, TASK_INTERRUPTIBLE); 4086 if (logical_sector >= mddev->suspend_lo && 4087 logical_sector < mddev->suspend_hi) 4088 schedule(); 4089 goto retry; 4090 } 4091 4092 if (test_bit(STRIPE_EXPANDING, &sh->state) || 4093 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) { 4094 /* Stripe is busy expanding or 4095 * add failed due to overlap. Flush everything 4096 * and wait a while 4097 */ 4098 md_raid5_unplug_device(conf); 4099 release_stripe(sh); 4100 schedule(); 4101 goto retry; 4102 } 4103 finish_wait(&conf->wait_for_overlap, &w); 4104 set_bit(STRIPE_HANDLE, &sh->state); 4105 clear_bit(STRIPE_DELAYED, &sh->state); 4106 if (mddev->barrier && 4107 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 4108 atomic_inc(&conf->preread_active_stripes); 4109 release_stripe(sh); 4110 } else { 4111 /* cannot get stripe for read-ahead, just give-up */ 4112 clear_bit(BIO_UPTODATE, &bi->bi_flags); 4113 finish_wait(&conf->wait_for_overlap, &w); 4114 break; 4115 } 4116 4117 } 4118 spin_lock_irq(&conf->device_lock); 4119 remaining = raid5_dec_bi_phys_segments(bi); 4120 spin_unlock_irq(&conf->device_lock); 4121 if (remaining == 0) { 4122 4123 if ( rw == WRITE ) 4124 md_write_end(mddev); 4125 4126 bio_endio(bi, 0); 4127 } 4128 4129 if (mddev->barrier) { 4130 /* We need to wait for the stripes to all be handled. 4131 * So: wait for preread_active_stripes to drop to 0. 4132 */ 4133 wait_event(mddev->thread->wqueue, 4134 atomic_read(&conf->preread_active_stripes) == 0); 4135 } 4136 return 0; 4137 } 4138 4139 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks); 4140 4141 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped) 4142 { 4143 /* reshaping is quite different to recovery/resync so it is 4144 * handled quite separately ... here. 4145 * 4146 * On each call to sync_request, we gather one chunk worth of 4147 * destination stripes and flag them as expanding. 4148 * Then we find all the source stripes and request reads. 4149 * As the reads complete, handle_stripe will copy the data 4150 * into the destination stripe and release that stripe. 4151 */ 4152 raid5_conf_t *conf = mddev->private; 4153 struct stripe_head *sh; 4154 sector_t first_sector, last_sector; 4155 int raid_disks = conf->previous_raid_disks; 4156 int data_disks = raid_disks - conf->max_degraded; 4157 int new_data_disks = conf->raid_disks - conf->max_degraded; 4158 int i; 4159 int dd_idx; 4160 sector_t writepos, readpos, safepos; 4161 sector_t stripe_addr; 4162 int reshape_sectors; 4163 struct list_head stripes; 4164 4165 if (sector_nr == 0) { 4166 /* If restarting in the middle, skip the initial sectors */ 4167 if (mddev->delta_disks < 0 && 4168 conf->reshape_progress < raid5_size(mddev, 0, 0)) { 4169 sector_nr = raid5_size(mddev, 0, 0) 4170 - conf->reshape_progress; 4171 } else if (mddev->delta_disks >= 0 && 4172 conf->reshape_progress > 0) 4173 sector_nr = conf->reshape_progress; 4174 sector_div(sector_nr, new_data_disks); 4175 if (sector_nr) { 4176 mddev->curr_resync_completed = sector_nr; 4177 sysfs_notify(&mddev->kobj, NULL, "sync_completed"); 4178 *skipped = 1; 4179 return sector_nr; 4180 } 4181 } 4182 4183 /* We need to process a full chunk at a time. 4184 * If old and new chunk sizes differ, we need to process the 4185 * largest of these 4186 */ 4187 if (mddev->new_chunk_sectors > mddev->chunk_sectors) 4188 reshape_sectors = mddev->new_chunk_sectors; 4189 else 4190 reshape_sectors = mddev->chunk_sectors; 4191 4192 /* we update the metadata when there is more than 3Meg 4193 * in the block range (that is rather arbitrary, should 4194 * probably be time based) or when the data about to be 4195 * copied would over-write the source of the data at 4196 * the front of the range. 4197 * i.e. one new_stripe along from reshape_progress new_maps 4198 * to after where reshape_safe old_maps to 4199 */ 4200 writepos = conf->reshape_progress; 4201 sector_div(writepos, new_data_disks); 4202 readpos = conf->reshape_progress; 4203 sector_div(readpos, data_disks); 4204 safepos = conf->reshape_safe; 4205 sector_div(safepos, data_disks); 4206 if (mddev->delta_disks < 0) { 4207 writepos -= min_t(sector_t, reshape_sectors, writepos); 4208 readpos += reshape_sectors; 4209 safepos += reshape_sectors; 4210 } else { 4211 writepos += reshape_sectors; 4212 readpos -= min_t(sector_t, reshape_sectors, readpos); 4213 safepos -= min_t(sector_t, reshape_sectors, safepos); 4214 } 4215 4216 /* 'writepos' is the most advanced device address we might write. 4217 * 'readpos' is the least advanced device address we might read. 4218 * 'safepos' is the least address recorded in the metadata as having 4219 * been reshaped. 4220 * If 'readpos' is behind 'writepos', then there is no way that we can 4221 * ensure safety in the face of a crash - that must be done by userspace 4222 * making a backup of the data. So in that case there is no particular 4223 * rush to update metadata. 4224 * Otherwise if 'safepos' is behind 'writepos', then we really need to 4225 * update the metadata to advance 'safepos' to match 'readpos' so that 4226 * we can be safe in the event of a crash. 4227 * So we insist on updating metadata if safepos is behind writepos and 4228 * readpos is beyond writepos. 4229 * In any case, update the metadata every 10 seconds. 4230 * Maybe that number should be configurable, but I'm not sure it is 4231 * worth it.... maybe it could be a multiple of safemode_delay??? 4232 */ 4233 if ((mddev->delta_disks < 0 4234 ? (safepos > writepos && readpos < writepos) 4235 : (safepos < writepos && readpos > writepos)) || 4236 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) { 4237 /* Cannot proceed until we've updated the superblock... */ 4238 wait_event(conf->wait_for_overlap, 4239 atomic_read(&conf->reshape_stripes)==0); 4240 mddev->reshape_position = conf->reshape_progress; 4241 mddev->curr_resync_completed = mddev->curr_resync; 4242 conf->reshape_checkpoint = jiffies; 4243 set_bit(MD_CHANGE_DEVS, &mddev->flags); 4244 md_wakeup_thread(mddev->thread); 4245 wait_event(mddev->sb_wait, mddev->flags == 0 || 4246 kthread_should_stop()); 4247 spin_lock_irq(&conf->device_lock); 4248 conf->reshape_safe = mddev->reshape_position; 4249 spin_unlock_irq(&conf->device_lock); 4250 wake_up(&conf->wait_for_overlap); 4251 sysfs_notify(&mddev->kobj, NULL, "sync_completed"); 4252 } 4253 4254 if (mddev->delta_disks < 0) { 4255 BUG_ON(conf->reshape_progress == 0); 4256 stripe_addr = writepos; 4257 BUG_ON((mddev->dev_sectors & 4258 ~((sector_t)reshape_sectors - 1)) 4259 - reshape_sectors - stripe_addr 4260 != sector_nr); 4261 } else { 4262 BUG_ON(writepos != sector_nr + reshape_sectors); 4263 stripe_addr = sector_nr; 4264 } 4265 INIT_LIST_HEAD(&stripes); 4266 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) { 4267 int j; 4268 int skipped_disk = 0; 4269 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1); 4270 set_bit(STRIPE_EXPANDING, &sh->state); 4271 atomic_inc(&conf->reshape_stripes); 4272 /* If any of this stripe is beyond the end of the old 4273 * array, then we need to zero those blocks 4274 */ 4275 for (j=sh->disks; j--;) { 4276 sector_t s; 4277 if (j == sh->pd_idx) 4278 continue; 4279 if (conf->level == 6 && 4280 j == sh->qd_idx) 4281 continue; 4282 s = compute_blocknr(sh, j, 0); 4283 if (s < raid5_size(mddev, 0, 0)) { 4284 skipped_disk = 1; 4285 continue; 4286 } 4287 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE); 4288 set_bit(R5_Expanded, &sh->dev[j].flags); 4289 set_bit(R5_UPTODATE, &sh->dev[j].flags); 4290 } 4291 if (!skipped_disk) { 4292 set_bit(STRIPE_EXPAND_READY, &sh->state); 4293 set_bit(STRIPE_HANDLE, &sh->state); 4294 } 4295 list_add(&sh->lru, &stripes); 4296 } 4297 spin_lock_irq(&conf->device_lock); 4298 if (mddev->delta_disks < 0) 4299 conf->reshape_progress -= reshape_sectors * new_data_disks; 4300 else 4301 conf->reshape_progress += reshape_sectors * new_data_disks; 4302 spin_unlock_irq(&conf->device_lock); 4303 /* Ok, those stripe are ready. We can start scheduling 4304 * reads on the source stripes. 4305 * The source stripes are determined by mapping the first and last 4306 * block on the destination stripes. 4307 */ 4308 first_sector = 4309 raid5_compute_sector(conf, stripe_addr*(new_data_disks), 4310 1, &dd_idx, NULL); 4311 last_sector = 4312 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors) 4313 * new_data_disks - 1), 4314 1, &dd_idx, NULL); 4315 if (last_sector >= mddev->dev_sectors) 4316 last_sector = mddev->dev_sectors - 1; 4317 while (first_sector <= last_sector) { 4318 sh = get_active_stripe(conf, first_sector, 1, 0, 1); 4319 set_bit(STRIPE_EXPAND_SOURCE, &sh->state); 4320 set_bit(STRIPE_HANDLE, &sh->state); 4321 release_stripe(sh); 4322 first_sector += STRIPE_SECTORS; 4323 } 4324 /* Now that the sources are clearly marked, we can release 4325 * the destination stripes 4326 */ 4327 while (!list_empty(&stripes)) { 4328 sh = list_entry(stripes.next, struct stripe_head, lru); 4329 list_del_init(&sh->lru); 4330 release_stripe(sh); 4331 } 4332 /* If this takes us to the resync_max point where we have to pause, 4333 * then we need to write out the superblock. 4334 */ 4335 sector_nr += reshape_sectors; 4336 if ((sector_nr - mddev->curr_resync_completed) * 2 4337 >= mddev->resync_max - mddev->curr_resync_completed) { 4338 /* Cannot proceed until we've updated the superblock... */ 4339 wait_event(conf->wait_for_overlap, 4340 atomic_read(&conf->reshape_stripes) == 0); 4341 mddev->reshape_position = conf->reshape_progress; 4342 mddev->curr_resync_completed = mddev->curr_resync + reshape_sectors; 4343 conf->reshape_checkpoint = jiffies; 4344 set_bit(MD_CHANGE_DEVS, &mddev->flags); 4345 md_wakeup_thread(mddev->thread); 4346 wait_event(mddev->sb_wait, 4347 !test_bit(MD_CHANGE_DEVS, &mddev->flags) 4348 || kthread_should_stop()); 4349 spin_lock_irq(&conf->device_lock); 4350 conf->reshape_safe = mddev->reshape_position; 4351 spin_unlock_irq(&conf->device_lock); 4352 wake_up(&conf->wait_for_overlap); 4353 sysfs_notify(&mddev->kobj, NULL, "sync_completed"); 4354 } 4355 return reshape_sectors; 4356 } 4357 4358 /* FIXME go_faster isn't used */ 4359 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster) 4360 { 4361 raid5_conf_t *conf = mddev->private; 4362 struct stripe_head *sh; 4363 sector_t max_sector = mddev->dev_sectors; 4364 int sync_blocks; 4365 int still_degraded = 0; 4366 int i; 4367 4368 if (sector_nr >= max_sector) { 4369 /* just being told to finish up .. nothing much to do */ 4370 unplug_slaves(mddev); 4371 4372 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) { 4373 end_reshape(conf); 4374 return 0; 4375 } 4376 4377 if (mddev->curr_resync < max_sector) /* aborted */ 4378 bitmap_end_sync(mddev->bitmap, mddev->curr_resync, 4379 &sync_blocks, 1); 4380 else /* completed sync */ 4381 conf->fullsync = 0; 4382 bitmap_close_sync(mddev->bitmap); 4383 4384 return 0; 4385 } 4386 4387 /* Allow raid5_quiesce to complete */ 4388 wait_event(conf->wait_for_overlap, conf->quiesce != 2); 4389 4390 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) 4391 return reshape_request(mddev, sector_nr, skipped); 4392 4393 /* No need to check resync_max as we never do more than one 4394 * stripe, and as resync_max will always be on a chunk boundary, 4395 * if the check in md_do_sync didn't fire, there is no chance 4396 * of overstepping resync_max here 4397 */ 4398 4399 /* if there is too many failed drives and we are trying 4400 * to resync, then assert that we are finished, because there is 4401 * nothing we can do. 4402 */ 4403 if (mddev->degraded >= conf->max_degraded && 4404 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) { 4405 sector_t rv = mddev->dev_sectors - sector_nr; 4406 *skipped = 1; 4407 return rv; 4408 } 4409 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) && 4410 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && 4411 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) { 4412 /* we can skip this block, and probably more */ 4413 sync_blocks /= STRIPE_SECTORS; 4414 *skipped = 1; 4415 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */ 4416 } 4417 4418 4419 bitmap_cond_end_sync(mddev->bitmap, sector_nr); 4420 4421 sh = get_active_stripe(conf, sector_nr, 0, 1, 0); 4422 if (sh == NULL) { 4423 sh = get_active_stripe(conf, sector_nr, 0, 0, 0); 4424 /* make sure we don't swamp the stripe cache if someone else 4425 * is trying to get access 4426 */ 4427 schedule_timeout_uninterruptible(1); 4428 } 4429 /* Need to check if array will still be degraded after recovery/resync 4430 * We don't need to check the 'failed' flag as when that gets set, 4431 * recovery aborts. 4432 */ 4433 for (i = 0; i < conf->raid_disks; i++) 4434 if (conf->disks[i].rdev == NULL) 4435 still_degraded = 1; 4436 4437 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded); 4438 4439 spin_lock(&sh->lock); 4440 set_bit(STRIPE_SYNCING, &sh->state); 4441 clear_bit(STRIPE_INSYNC, &sh->state); 4442 spin_unlock(&sh->lock); 4443 4444 handle_stripe(sh); 4445 release_stripe(sh); 4446 4447 return STRIPE_SECTORS; 4448 } 4449 4450 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio) 4451 { 4452 /* We may not be able to submit a whole bio at once as there 4453 * may not be enough stripe_heads available. 4454 * We cannot pre-allocate enough stripe_heads as we may need 4455 * more than exist in the cache (if we allow ever large chunks). 4456 * So we do one stripe head at a time and record in 4457 * ->bi_hw_segments how many have been done. 4458 * 4459 * We *know* that this entire raid_bio is in one chunk, so 4460 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector. 4461 */ 4462 struct stripe_head *sh; 4463 int dd_idx; 4464 sector_t sector, logical_sector, last_sector; 4465 int scnt = 0; 4466 int remaining; 4467 int handled = 0; 4468 4469 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1); 4470 sector = raid5_compute_sector(conf, logical_sector, 4471 0, &dd_idx, NULL); 4472 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9); 4473 4474 for (; logical_sector < last_sector; 4475 logical_sector += STRIPE_SECTORS, 4476 sector += STRIPE_SECTORS, 4477 scnt++) { 4478 4479 if (scnt < raid5_bi_hw_segments(raid_bio)) 4480 /* already done this stripe */ 4481 continue; 4482 4483 sh = get_active_stripe(conf, sector, 0, 1, 0); 4484 4485 if (!sh) { 4486 /* failed to get a stripe - must wait */ 4487 raid5_set_bi_hw_segments(raid_bio, scnt); 4488 conf->retry_read_aligned = raid_bio; 4489 return handled; 4490 } 4491 4492 set_bit(R5_ReadError, &sh->dev[dd_idx].flags); 4493 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) { 4494 release_stripe(sh); 4495 raid5_set_bi_hw_segments(raid_bio, scnt); 4496 conf->retry_read_aligned = raid_bio; 4497 return handled; 4498 } 4499 4500 handle_stripe(sh); 4501 release_stripe(sh); 4502 handled++; 4503 } 4504 spin_lock_irq(&conf->device_lock); 4505 remaining = raid5_dec_bi_phys_segments(raid_bio); 4506 spin_unlock_irq(&conf->device_lock); 4507 if (remaining == 0) 4508 bio_endio(raid_bio, 0); 4509 if (atomic_dec_and_test(&conf->active_aligned_reads)) 4510 wake_up(&conf->wait_for_stripe); 4511 return handled; 4512 } 4513 4514 4515 /* 4516 * This is our raid5 kernel thread. 4517 * 4518 * We scan the hash table for stripes which can be handled now. 4519 * During the scan, completed stripes are saved for us by the interrupt 4520 * handler, so that they will not have to wait for our next wakeup. 4521 */ 4522 static void raid5d(mddev_t *mddev) 4523 { 4524 struct stripe_head *sh; 4525 raid5_conf_t *conf = mddev->private; 4526 int handled; 4527 4528 pr_debug("+++ raid5d active\n"); 4529 4530 md_check_recovery(mddev); 4531 4532 handled = 0; 4533 spin_lock_irq(&conf->device_lock); 4534 while (1) { 4535 struct bio *bio; 4536 4537 if (conf->seq_flush != conf->seq_write) { 4538 int seq = conf->seq_flush; 4539 spin_unlock_irq(&conf->device_lock); 4540 bitmap_unplug(mddev->bitmap); 4541 spin_lock_irq(&conf->device_lock); 4542 conf->seq_write = seq; 4543 activate_bit_delay(conf); 4544 } 4545 4546 while ((bio = remove_bio_from_retry(conf))) { 4547 int ok; 4548 spin_unlock_irq(&conf->device_lock); 4549 ok = retry_aligned_read(conf, bio); 4550 spin_lock_irq(&conf->device_lock); 4551 if (!ok) 4552 break; 4553 handled++; 4554 } 4555 4556 sh = __get_priority_stripe(conf); 4557 4558 if (!sh) 4559 break; 4560 spin_unlock_irq(&conf->device_lock); 4561 4562 handled++; 4563 handle_stripe(sh); 4564 release_stripe(sh); 4565 cond_resched(); 4566 4567 spin_lock_irq(&conf->device_lock); 4568 } 4569 pr_debug("%d stripes handled\n", handled); 4570 4571 spin_unlock_irq(&conf->device_lock); 4572 4573 async_tx_issue_pending_all(); 4574 unplug_slaves(mddev); 4575 4576 pr_debug("--- raid5d inactive\n"); 4577 } 4578 4579 static ssize_t 4580 raid5_show_stripe_cache_size(mddev_t *mddev, char *page) 4581 { 4582 raid5_conf_t *conf = mddev->private; 4583 if (conf) 4584 return sprintf(page, "%d\n", conf->max_nr_stripes); 4585 else 4586 return 0; 4587 } 4588 4589 int 4590 raid5_set_cache_size(mddev_t *mddev, int size) 4591 { 4592 raid5_conf_t *conf = mddev->private; 4593 int err; 4594 4595 if (size <= 16 || size > 32768) 4596 return -EINVAL; 4597 while (size < conf->max_nr_stripes) { 4598 if (drop_one_stripe(conf)) 4599 conf->max_nr_stripes--; 4600 else 4601 break; 4602 } 4603 err = md_allow_write(mddev); 4604 if (err) 4605 return err; 4606 while (size > conf->max_nr_stripes) { 4607 if (grow_one_stripe(conf)) 4608 conf->max_nr_stripes++; 4609 else break; 4610 } 4611 return 0; 4612 } 4613 EXPORT_SYMBOL(raid5_set_cache_size); 4614 4615 static ssize_t 4616 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len) 4617 { 4618 raid5_conf_t *conf = mddev->private; 4619 unsigned long new; 4620 int err; 4621 4622 if (len >= PAGE_SIZE) 4623 return -EINVAL; 4624 if (!conf) 4625 return -ENODEV; 4626 4627 if (strict_strtoul(page, 10, &new)) 4628 return -EINVAL; 4629 err = raid5_set_cache_size(mddev, new); 4630 if (err) 4631 return err; 4632 return len; 4633 } 4634 4635 static struct md_sysfs_entry 4636 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR, 4637 raid5_show_stripe_cache_size, 4638 raid5_store_stripe_cache_size); 4639 4640 static ssize_t 4641 raid5_show_preread_threshold(mddev_t *mddev, char *page) 4642 { 4643 raid5_conf_t *conf = mddev->private; 4644 if (conf) 4645 return sprintf(page, "%d\n", conf->bypass_threshold); 4646 else 4647 return 0; 4648 } 4649 4650 static ssize_t 4651 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len) 4652 { 4653 raid5_conf_t *conf = mddev->private; 4654 unsigned long new; 4655 if (len >= PAGE_SIZE) 4656 return -EINVAL; 4657 if (!conf) 4658 return -ENODEV; 4659 4660 if (strict_strtoul(page, 10, &new)) 4661 return -EINVAL; 4662 if (new > conf->max_nr_stripes) 4663 return -EINVAL; 4664 conf->bypass_threshold = new; 4665 return len; 4666 } 4667 4668 static struct md_sysfs_entry 4669 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold, 4670 S_IRUGO | S_IWUSR, 4671 raid5_show_preread_threshold, 4672 raid5_store_preread_threshold); 4673 4674 static ssize_t 4675 stripe_cache_active_show(mddev_t *mddev, char *page) 4676 { 4677 raid5_conf_t *conf = mddev->private; 4678 if (conf) 4679 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes)); 4680 else 4681 return 0; 4682 } 4683 4684 static struct md_sysfs_entry 4685 raid5_stripecache_active = __ATTR_RO(stripe_cache_active); 4686 4687 static struct attribute *raid5_attrs[] = { 4688 &raid5_stripecache_size.attr, 4689 &raid5_stripecache_active.attr, 4690 &raid5_preread_bypass_threshold.attr, 4691 NULL, 4692 }; 4693 static struct attribute_group raid5_attrs_group = { 4694 .name = NULL, 4695 .attrs = raid5_attrs, 4696 }; 4697 4698 static sector_t 4699 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks) 4700 { 4701 raid5_conf_t *conf = mddev->private; 4702 4703 if (!sectors) 4704 sectors = mddev->dev_sectors; 4705 if (!raid_disks) 4706 /* size is defined by the smallest of previous and new size */ 4707 raid_disks = min(conf->raid_disks, conf->previous_raid_disks); 4708 4709 sectors &= ~((sector_t)mddev->chunk_sectors - 1); 4710 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1); 4711 return sectors * (raid_disks - conf->max_degraded); 4712 } 4713 4714 static void raid5_free_percpu(raid5_conf_t *conf) 4715 { 4716 struct raid5_percpu *percpu; 4717 unsigned long cpu; 4718 4719 if (!conf->percpu) 4720 return; 4721 4722 get_online_cpus(); 4723 for_each_possible_cpu(cpu) { 4724 percpu = per_cpu_ptr(conf->percpu, cpu); 4725 safe_put_page(percpu->spare_page); 4726 kfree(percpu->scribble); 4727 } 4728 #ifdef CONFIG_HOTPLUG_CPU 4729 unregister_cpu_notifier(&conf->cpu_notify); 4730 #endif 4731 put_online_cpus(); 4732 4733 free_percpu(conf->percpu); 4734 } 4735 4736 static void free_conf(raid5_conf_t *conf) 4737 { 4738 shrink_stripes(conf); 4739 raid5_free_percpu(conf); 4740 kfree(conf->disks); 4741 kfree(conf->stripe_hashtbl); 4742 kfree(conf); 4743 } 4744 4745 #ifdef CONFIG_HOTPLUG_CPU 4746 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action, 4747 void *hcpu) 4748 { 4749 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify); 4750 long cpu = (long)hcpu; 4751 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu); 4752 4753 switch (action) { 4754 case CPU_UP_PREPARE: 4755 case CPU_UP_PREPARE_FROZEN: 4756 if (conf->level == 6 && !percpu->spare_page) 4757 percpu->spare_page = alloc_page(GFP_KERNEL); 4758 if (!percpu->scribble) 4759 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL); 4760 4761 if (!percpu->scribble || 4762 (conf->level == 6 && !percpu->spare_page)) { 4763 safe_put_page(percpu->spare_page); 4764 kfree(percpu->scribble); 4765 pr_err("%s: failed memory allocation for cpu%ld\n", 4766 __func__, cpu); 4767 return notifier_from_errno(-ENOMEM); 4768 } 4769 break; 4770 case CPU_DEAD: 4771 case CPU_DEAD_FROZEN: 4772 safe_put_page(percpu->spare_page); 4773 kfree(percpu->scribble); 4774 percpu->spare_page = NULL; 4775 percpu->scribble = NULL; 4776 break; 4777 default: 4778 break; 4779 } 4780 return NOTIFY_OK; 4781 } 4782 #endif 4783 4784 static int raid5_alloc_percpu(raid5_conf_t *conf) 4785 { 4786 unsigned long cpu; 4787 struct page *spare_page; 4788 struct raid5_percpu __percpu *allcpus; 4789 void *scribble; 4790 int err; 4791 4792 allcpus = alloc_percpu(struct raid5_percpu); 4793 if (!allcpus) 4794 return -ENOMEM; 4795 conf->percpu = allcpus; 4796 4797 get_online_cpus(); 4798 err = 0; 4799 for_each_present_cpu(cpu) { 4800 if (conf->level == 6) { 4801 spare_page = alloc_page(GFP_KERNEL); 4802 if (!spare_page) { 4803 err = -ENOMEM; 4804 break; 4805 } 4806 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page; 4807 } 4808 scribble = kmalloc(conf->scribble_len, GFP_KERNEL); 4809 if (!scribble) { 4810 err = -ENOMEM; 4811 break; 4812 } 4813 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble; 4814 } 4815 #ifdef CONFIG_HOTPLUG_CPU 4816 conf->cpu_notify.notifier_call = raid456_cpu_notify; 4817 conf->cpu_notify.priority = 0; 4818 if (err == 0) 4819 err = register_cpu_notifier(&conf->cpu_notify); 4820 #endif 4821 put_online_cpus(); 4822 4823 return err; 4824 } 4825 4826 static raid5_conf_t *setup_conf(mddev_t *mddev) 4827 { 4828 raid5_conf_t *conf; 4829 int raid_disk, memory, max_disks; 4830 mdk_rdev_t *rdev; 4831 struct disk_info *disk; 4832 4833 if (mddev->new_level != 5 4834 && mddev->new_level != 4 4835 && mddev->new_level != 6) { 4836 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n", 4837 mdname(mddev), mddev->new_level); 4838 return ERR_PTR(-EIO); 4839 } 4840 if ((mddev->new_level == 5 4841 && !algorithm_valid_raid5(mddev->new_layout)) || 4842 (mddev->new_level == 6 4843 && !algorithm_valid_raid6(mddev->new_layout))) { 4844 printk(KERN_ERR "md/raid:%s: layout %d not supported\n", 4845 mdname(mddev), mddev->new_layout); 4846 return ERR_PTR(-EIO); 4847 } 4848 if (mddev->new_level == 6 && mddev->raid_disks < 4) { 4849 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n", 4850 mdname(mddev), mddev->raid_disks); 4851 return ERR_PTR(-EINVAL); 4852 } 4853 4854 if (!mddev->new_chunk_sectors || 4855 (mddev->new_chunk_sectors << 9) % PAGE_SIZE || 4856 !is_power_of_2(mddev->new_chunk_sectors)) { 4857 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n", 4858 mdname(mddev), mddev->new_chunk_sectors << 9); 4859 return ERR_PTR(-EINVAL); 4860 } 4861 4862 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL); 4863 if (conf == NULL) 4864 goto abort; 4865 spin_lock_init(&conf->device_lock); 4866 init_waitqueue_head(&conf->wait_for_stripe); 4867 init_waitqueue_head(&conf->wait_for_overlap); 4868 INIT_LIST_HEAD(&conf->handle_list); 4869 INIT_LIST_HEAD(&conf->hold_list); 4870 INIT_LIST_HEAD(&conf->delayed_list); 4871 INIT_LIST_HEAD(&conf->bitmap_list); 4872 INIT_LIST_HEAD(&conf->inactive_list); 4873 atomic_set(&conf->active_stripes, 0); 4874 atomic_set(&conf->preread_active_stripes, 0); 4875 atomic_set(&conf->active_aligned_reads, 0); 4876 conf->bypass_threshold = BYPASS_THRESHOLD; 4877 4878 conf->raid_disks = mddev->raid_disks; 4879 if (mddev->reshape_position == MaxSector) 4880 conf->previous_raid_disks = mddev->raid_disks; 4881 else 4882 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks; 4883 max_disks = max(conf->raid_disks, conf->previous_raid_disks); 4884 conf->scribble_len = scribble_len(max_disks); 4885 4886 conf->disks = kzalloc(max_disks * sizeof(struct disk_info), 4887 GFP_KERNEL); 4888 if (!conf->disks) 4889 goto abort; 4890 4891 conf->mddev = mddev; 4892 4893 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL) 4894 goto abort; 4895 4896 conf->level = mddev->new_level; 4897 if (raid5_alloc_percpu(conf) != 0) 4898 goto abort; 4899 4900 pr_debug("raid456: run(%s) called.\n", mdname(mddev)); 4901 4902 list_for_each_entry(rdev, &mddev->disks, same_set) { 4903 raid_disk = rdev->raid_disk; 4904 if (raid_disk >= max_disks 4905 || raid_disk < 0) 4906 continue; 4907 disk = conf->disks + raid_disk; 4908 4909 disk->rdev = rdev; 4910 4911 if (test_bit(In_sync, &rdev->flags)) { 4912 char b[BDEVNAME_SIZE]; 4913 printk(KERN_INFO "md/raid:%s: device %s operational as raid" 4914 " disk %d\n", 4915 mdname(mddev), bdevname(rdev->bdev, b), raid_disk); 4916 } else 4917 /* Cannot rely on bitmap to complete recovery */ 4918 conf->fullsync = 1; 4919 } 4920 4921 conf->chunk_sectors = mddev->new_chunk_sectors; 4922 conf->level = mddev->new_level; 4923 if (conf->level == 6) 4924 conf->max_degraded = 2; 4925 else 4926 conf->max_degraded = 1; 4927 conf->algorithm = mddev->new_layout; 4928 conf->max_nr_stripes = NR_STRIPES; 4929 conf->reshape_progress = mddev->reshape_position; 4930 if (conf->reshape_progress != MaxSector) { 4931 conf->prev_chunk_sectors = mddev->chunk_sectors; 4932 conf->prev_algo = mddev->layout; 4933 } 4934 4935 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) + 4936 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024; 4937 if (grow_stripes(conf, conf->max_nr_stripes)) { 4938 printk(KERN_ERR 4939 "md/raid:%s: couldn't allocate %dkB for buffers\n", 4940 mdname(mddev), memory); 4941 goto abort; 4942 } else 4943 printk(KERN_INFO "md/raid:%s: allocated %dkB\n", 4944 mdname(mddev), memory); 4945 4946 conf->thread = md_register_thread(raid5d, mddev, NULL); 4947 if (!conf->thread) { 4948 printk(KERN_ERR 4949 "md/raid:%s: couldn't allocate thread.\n", 4950 mdname(mddev)); 4951 goto abort; 4952 } 4953 4954 return conf; 4955 4956 abort: 4957 if (conf) { 4958 free_conf(conf); 4959 return ERR_PTR(-EIO); 4960 } else 4961 return ERR_PTR(-ENOMEM); 4962 } 4963 4964 4965 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded) 4966 { 4967 switch (algo) { 4968 case ALGORITHM_PARITY_0: 4969 if (raid_disk < max_degraded) 4970 return 1; 4971 break; 4972 case ALGORITHM_PARITY_N: 4973 if (raid_disk >= raid_disks - max_degraded) 4974 return 1; 4975 break; 4976 case ALGORITHM_PARITY_0_6: 4977 if (raid_disk == 0 || 4978 raid_disk == raid_disks - 1) 4979 return 1; 4980 break; 4981 case ALGORITHM_LEFT_ASYMMETRIC_6: 4982 case ALGORITHM_RIGHT_ASYMMETRIC_6: 4983 case ALGORITHM_LEFT_SYMMETRIC_6: 4984 case ALGORITHM_RIGHT_SYMMETRIC_6: 4985 if (raid_disk == raid_disks - 1) 4986 return 1; 4987 } 4988 return 0; 4989 } 4990 4991 static int run(mddev_t *mddev) 4992 { 4993 raid5_conf_t *conf; 4994 int working_disks = 0; 4995 int dirty_parity_disks = 0; 4996 mdk_rdev_t *rdev; 4997 sector_t reshape_offset = 0; 4998 4999 if (mddev->recovery_cp != MaxSector) 5000 printk(KERN_NOTICE "md/raid:%s: not clean" 5001 " -- starting background reconstruction\n", 5002 mdname(mddev)); 5003 if (mddev->reshape_position != MaxSector) { 5004 /* Check that we can continue the reshape. 5005 * Currently only disks can change, it must 5006 * increase, and we must be past the point where 5007 * a stripe over-writes itself 5008 */ 5009 sector_t here_new, here_old; 5010 int old_disks; 5011 int max_degraded = (mddev->level == 6 ? 2 : 1); 5012 5013 if (mddev->new_level != mddev->level) { 5014 printk(KERN_ERR "md/raid:%s: unsupported reshape " 5015 "required - aborting.\n", 5016 mdname(mddev)); 5017 return -EINVAL; 5018 } 5019 old_disks = mddev->raid_disks - mddev->delta_disks; 5020 /* reshape_position must be on a new-stripe boundary, and one 5021 * further up in new geometry must map after here in old 5022 * geometry. 5023 */ 5024 here_new = mddev->reshape_position; 5025 if (sector_div(here_new, mddev->new_chunk_sectors * 5026 (mddev->raid_disks - max_degraded))) { 5027 printk(KERN_ERR "md/raid:%s: reshape_position not " 5028 "on a stripe boundary\n", mdname(mddev)); 5029 return -EINVAL; 5030 } 5031 reshape_offset = here_new * mddev->new_chunk_sectors; 5032 /* here_new is the stripe we will write to */ 5033 here_old = mddev->reshape_position; 5034 sector_div(here_old, mddev->chunk_sectors * 5035 (old_disks-max_degraded)); 5036 /* here_old is the first stripe that we might need to read 5037 * from */ 5038 if (mddev->delta_disks == 0) { 5039 /* We cannot be sure it is safe to start an in-place 5040 * reshape. It is only safe if user-space if monitoring 5041 * and taking constant backups. 5042 * mdadm always starts a situation like this in 5043 * readonly mode so it can take control before 5044 * allowing any writes. So just check for that. 5045 */ 5046 if ((here_new * mddev->new_chunk_sectors != 5047 here_old * mddev->chunk_sectors) || 5048 mddev->ro == 0) { 5049 printk(KERN_ERR "md/raid:%s: in-place reshape must be started" 5050 " in read-only mode - aborting\n", 5051 mdname(mddev)); 5052 return -EINVAL; 5053 } 5054 } else if (mddev->delta_disks < 0 5055 ? (here_new * mddev->new_chunk_sectors <= 5056 here_old * mddev->chunk_sectors) 5057 : (here_new * mddev->new_chunk_sectors >= 5058 here_old * mddev->chunk_sectors)) { 5059 /* Reading from the same stripe as writing to - bad */ 5060 printk(KERN_ERR "md/raid:%s: reshape_position too early for " 5061 "auto-recovery - aborting.\n", 5062 mdname(mddev)); 5063 return -EINVAL; 5064 } 5065 printk(KERN_INFO "md/raid:%s: reshape will continue\n", 5066 mdname(mddev)); 5067 /* OK, we should be able to continue; */ 5068 } else { 5069 BUG_ON(mddev->level != mddev->new_level); 5070 BUG_ON(mddev->layout != mddev->new_layout); 5071 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors); 5072 BUG_ON(mddev->delta_disks != 0); 5073 } 5074 5075 if (mddev->private == NULL) 5076 conf = setup_conf(mddev); 5077 else 5078 conf = mddev->private; 5079 5080 if (IS_ERR(conf)) 5081 return PTR_ERR(conf); 5082 5083 mddev->thread = conf->thread; 5084 conf->thread = NULL; 5085 mddev->private = conf; 5086 5087 /* 5088 * 0 for a fully functional array, 1 or 2 for a degraded array. 5089 */ 5090 list_for_each_entry(rdev, &mddev->disks, same_set) { 5091 if (rdev->raid_disk < 0) 5092 continue; 5093 if (test_bit(In_sync, &rdev->flags)) { 5094 working_disks++; 5095 continue; 5096 } 5097 /* This disc is not fully in-sync. However if it 5098 * just stored parity (beyond the recovery_offset), 5099 * when we don't need to be concerned about the 5100 * array being dirty. 5101 * When reshape goes 'backwards', we never have 5102 * partially completed devices, so we only need 5103 * to worry about reshape going forwards. 5104 */ 5105 /* Hack because v0.91 doesn't store recovery_offset properly. */ 5106 if (mddev->major_version == 0 && 5107 mddev->minor_version > 90) 5108 rdev->recovery_offset = reshape_offset; 5109 5110 if (rdev->recovery_offset < reshape_offset) { 5111 /* We need to check old and new layout */ 5112 if (!only_parity(rdev->raid_disk, 5113 conf->algorithm, 5114 conf->raid_disks, 5115 conf->max_degraded)) 5116 continue; 5117 } 5118 if (!only_parity(rdev->raid_disk, 5119 conf->prev_algo, 5120 conf->previous_raid_disks, 5121 conf->max_degraded)) 5122 continue; 5123 dirty_parity_disks++; 5124 } 5125 5126 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks) 5127 - working_disks); 5128 5129 if (has_failed(conf)) { 5130 printk(KERN_ERR "md/raid:%s: not enough operational devices" 5131 " (%d/%d failed)\n", 5132 mdname(mddev), mddev->degraded, conf->raid_disks); 5133 goto abort; 5134 } 5135 5136 /* device size must be a multiple of chunk size */ 5137 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1); 5138 mddev->resync_max_sectors = mddev->dev_sectors; 5139 5140 if (mddev->degraded > dirty_parity_disks && 5141 mddev->recovery_cp != MaxSector) { 5142 if (mddev->ok_start_degraded) 5143 printk(KERN_WARNING 5144 "md/raid:%s: starting dirty degraded array" 5145 " - data corruption possible.\n", 5146 mdname(mddev)); 5147 else { 5148 printk(KERN_ERR 5149 "md/raid:%s: cannot start dirty degraded array.\n", 5150 mdname(mddev)); 5151 goto abort; 5152 } 5153 } 5154 5155 if (mddev->degraded == 0) 5156 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d" 5157 " devices, algorithm %d\n", mdname(mddev), conf->level, 5158 mddev->raid_disks-mddev->degraded, mddev->raid_disks, 5159 mddev->new_layout); 5160 else 5161 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d" 5162 " out of %d devices, algorithm %d\n", 5163 mdname(mddev), conf->level, 5164 mddev->raid_disks - mddev->degraded, 5165 mddev->raid_disks, mddev->new_layout); 5166 5167 print_raid5_conf(conf); 5168 5169 if (conf->reshape_progress != MaxSector) { 5170 conf->reshape_safe = conf->reshape_progress; 5171 atomic_set(&conf->reshape_stripes, 0); 5172 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery); 5173 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery); 5174 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery); 5175 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery); 5176 mddev->sync_thread = md_register_thread(md_do_sync, mddev, 5177 "reshape"); 5178 } 5179 5180 5181 /* Ok, everything is just fine now */ 5182 if (mddev->to_remove == &raid5_attrs_group) 5183 mddev->to_remove = NULL; 5184 else if (mddev->kobj.sd && 5185 sysfs_create_group(&mddev->kobj, &raid5_attrs_group)) 5186 printk(KERN_WARNING 5187 "raid5: failed to create sysfs attributes for %s\n", 5188 mdname(mddev)); 5189 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0)); 5190 5191 plugger_init(&conf->plug, raid5_unplug); 5192 mddev->plug = &conf->plug; 5193 if (mddev->queue) { 5194 int chunk_size; 5195 /* read-ahead size must cover two whole stripes, which 5196 * is 2 * (datadisks) * chunksize where 'n' is the 5197 * number of raid devices 5198 */ 5199 int data_disks = conf->previous_raid_disks - conf->max_degraded; 5200 int stripe = data_disks * 5201 ((mddev->chunk_sectors << 9) / PAGE_SIZE); 5202 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe) 5203 mddev->queue->backing_dev_info.ra_pages = 2 * stripe; 5204 5205 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec); 5206 5207 mddev->queue->backing_dev_info.congested_data = mddev; 5208 mddev->queue->backing_dev_info.congested_fn = raid5_congested; 5209 mddev->queue->queue_lock = &conf->device_lock; 5210 mddev->queue->unplug_fn = raid5_unplug_queue; 5211 5212 chunk_size = mddev->chunk_sectors << 9; 5213 blk_queue_io_min(mddev->queue, chunk_size); 5214 blk_queue_io_opt(mddev->queue, chunk_size * 5215 (conf->raid_disks - conf->max_degraded)); 5216 5217 list_for_each_entry(rdev, &mddev->disks, same_set) 5218 disk_stack_limits(mddev->gendisk, rdev->bdev, 5219 rdev->data_offset << 9); 5220 } 5221 5222 return 0; 5223 abort: 5224 md_unregister_thread(mddev->thread); 5225 mddev->thread = NULL; 5226 if (conf) { 5227 print_raid5_conf(conf); 5228 free_conf(conf); 5229 } 5230 mddev->private = NULL; 5231 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev)); 5232 return -EIO; 5233 } 5234 5235 static int stop(mddev_t *mddev) 5236 { 5237 raid5_conf_t *conf = mddev->private; 5238 5239 md_unregister_thread(mddev->thread); 5240 mddev->thread = NULL; 5241 if (mddev->queue) 5242 mddev->queue->backing_dev_info.congested_fn = NULL; 5243 plugger_flush(&conf->plug); /* the unplug fn references 'conf'*/ 5244 free_conf(conf); 5245 mddev->private = NULL; 5246 mddev->to_remove = &raid5_attrs_group; 5247 return 0; 5248 } 5249 5250 #ifdef DEBUG 5251 static void print_sh(struct seq_file *seq, struct stripe_head *sh) 5252 { 5253 int i; 5254 5255 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n", 5256 (unsigned long long)sh->sector, sh->pd_idx, sh->state); 5257 seq_printf(seq, "sh %llu, count %d.\n", 5258 (unsigned long long)sh->sector, atomic_read(&sh->count)); 5259 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector); 5260 for (i = 0; i < sh->disks; i++) { 5261 seq_printf(seq, "(cache%d: %p %ld) ", 5262 i, sh->dev[i].page, sh->dev[i].flags); 5263 } 5264 seq_printf(seq, "\n"); 5265 } 5266 5267 static void printall(struct seq_file *seq, raid5_conf_t *conf) 5268 { 5269 struct stripe_head *sh; 5270 struct hlist_node *hn; 5271 int i; 5272 5273 spin_lock_irq(&conf->device_lock); 5274 for (i = 0; i < NR_HASH; i++) { 5275 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) { 5276 if (sh->raid_conf != conf) 5277 continue; 5278 print_sh(seq, sh); 5279 } 5280 } 5281 spin_unlock_irq(&conf->device_lock); 5282 } 5283 #endif 5284 5285 static void status(struct seq_file *seq, mddev_t *mddev) 5286 { 5287 raid5_conf_t *conf = mddev->private; 5288 int i; 5289 5290 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level, 5291 mddev->chunk_sectors / 2, mddev->layout); 5292 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded); 5293 for (i = 0; i < conf->raid_disks; i++) 5294 seq_printf (seq, "%s", 5295 conf->disks[i].rdev && 5296 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_"); 5297 seq_printf (seq, "]"); 5298 #ifdef DEBUG 5299 seq_printf (seq, "\n"); 5300 printall(seq, conf); 5301 #endif 5302 } 5303 5304 static void print_raid5_conf (raid5_conf_t *conf) 5305 { 5306 int i; 5307 struct disk_info *tmp; 5308 5309 printk(KERN_DEBUG "RAID conf printout:\n"); 5310 if (!conf) { 5311 printk("(conf==NULL)\n"); 5312 return; 5313 } 5314 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level, 5315 conf->raid_disks, 5316 conf->raid_disks - conf->mddev->degraded); 5317 5318 for (i = 0; i < conf->raid_disks; i++) { 5319 char b[BDEVNAME_SIZE]; 5320 tmp = conf->disks + i; 5321 if (tmp->rdev) 5322 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n", 5323 i, !test_bit(Faulty, &tmp->rdev->flags), 5324 bdevname(tmp->rdev->bdev, b)); 5325 } 5326 } 5327 5328 static int raid5_spare_active(mddev_t *mddev) 5329 { 5330 int i; 5331 raid5_conf_t *conf = mddev->private; 5332 struct disk_info *tmp; 5333 int count = 0; 5334 unsigned long flags; 5335 5336 for (i = 0; i < conf->raid_disks; i++) { 5337 tmp = conf->disks + i; 5338 if (tmp->rdev 5339 && tmp->rdev->recovery_offset == MaxSector 5340 && !test_bit(Faulty, &tmp->rdev->flags) 5341 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) { 5342 count++; 5343 sysfs_notify_dirent(tmp->rdev->sysfs_state); 5344 } 5345 } 5346 spin_lock_irqsave(&conf->device_lock, flags); 5347 mddev->degraded -= count; 5348 spin_unlock_irqrestore(&conf->device_lock, flags); 5349 print_raid5_conf(conf); 5350 return count; 5351 } 5352 5353 static int raid5_remove_disk(mddev_t *mddev, int number) 5354 { 5355 raid5_conf_t *conf = mddev->private; 5356 int err = 0; 5357 mdk_rdev_t *rdev; 5358 struct disk_info *p = conf->disks + number; 5359 5360 print_raid5_conf(conf); 5361 rdev = p->rdev; 5362 if (rdev) { 5363 if (number >= conf->raid_disks && 5364 conf->reshape_progress == MaxSector) 5365 clear_bit(In_sync, &rdev->flags); 5366 5367 if (test_bit(In_sync, &rdev->flags) || 5368 atomic_read(&rdev->nr_pending)) { 5369 err = -EBUSY; 5370 goto abort; 5371 } 5372 /* Only remove non-faulty devices if recovery 5373 * isn't possible. 5374 */ 5375 if (!test_bit(Faulty, &rdev->flags) && 5376 !has_failed(conf) && 5377 number < conf->raid_disks) { 5378 err = -EBUSY; 5379 goto abort; 5380 } 5381 p->rdev = NULL; 5382 synchronize_rcu(); 5383 if (atomic_read(&rdev->nr_pending)) { 5384 /* lost the race, try later */ 5385 err = -EBUSY; 5386 p->rdev = rdev; 5387 } 5388 } 5389 abort: 5390 5391 print_raid5_conf(conf); 5392 return err; 5393 } 5394 5395 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev) 5396 { 5397 raid5_conf_t *conf = mddev->private; 5398 int err = -EEXIST; 5399 int disk; 5400 struct disk_info *p; 5401 int first = 0; 5402 int last = conf->raid_disks - 1; 5403 5404 if (has_failed(conf)) 5405 /* no point adding a device */ 5406 return -EINVAL; 5407 5408 if (rdev->raid_disk >= 0) 5409 first = last = rdev->raid_disk; 5410 5411 /* 5412 * find the disk ... but prefer rdev->saved_raid_disk 5413 * if possible. 5414 */ 5415 if (rdev->saved_raid_disk >= 0 && 5416 rdev->saved_raid_disk >= first && 5417 conf->disks[rdev->saved_raid_disk].rdev == NULL) 5418 disk = rdev->saved_raid_disk; 5419 else 5420 disk = first; 5421 for ( ; disk <= last ; disk++) 5422 if ((p=conf->disks + disk)->rdev == NULL) { 5423 clear_bit(In_sync, &rdev->flags); 5424 rdev->raid_disk = disk; 5425 err = 0; 5426 if (rdev->saved_raid_disk != disk) 5427 conf->fullsync = 1; 5428 rcu_assign_pointer(p->rdev, rdev); 5429 break; 5430 } 5431 print_raid5_conf(conf); 5432 return err; 5433 } 5434 5435 static int raid5_resize(mddev_t *mddev, sector_t sectors) 5436 { 5437 /* no resync is happening, and there is enough space 5438 * on all devices, so we can resize. 5439 * We need to make sure resync covers any new space. 5440 * If the array is shrinking we should possibly wait until 5441 * any io in the removed space completes, but it hardly seems 5442 * worth it. 5443 */ 5444 sectors &= ~((sector_t)mddev->chunk_sectors - 1); 5445 md_set_array_sectors(mddev, raid5_size(mddev, sectors, 5446 mddev->raid_disks)); 5447 if (mddev->array_sectors > 5448 raid5_size(mddev, sectors, mddev->raid_disks)) 5449 return -EINVAL; 5450 set_capacity(mddev->gendisk, mddev->array_sectors); 5451 revalidate_disk(mddev->gendisk); 5452 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) { 5453 mddev->recovery_cp = mddev->dev_sectors; 5454 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); 5455 } 5456 mddev->dev_sectors = sectors; 5457 mddev->resync_max_sectors = sectors; 5458 return 0; 5459 } 5460 5461 static int check_stripe_cache(mddev_t *mddev) 5462 { 5463 /* Can only proceed if there are plenty of stripe_heads. 5464 * We need a minimum of one full stripe,, and for sensible progress 5465 * it is best to have about 4 times that. 5466 * If we require 4 times, then the default 256 4K stripe_heads will 5467 * allow for chunk sizes up to 256K, which is probably OK. 5468 * If the chunk size is greater, user-space should request more 5469 * stripe_heads first. 5470 */ 5471 raid5_conf_t *conf = mddev->private; 5472 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4 5473 > conf->max_nr_stripes || 5474 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4 5475 > conf->max_nr_stripes) { 5476 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n", 5477 mdname(mddev), 5478 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9) 5479 / STRIPE_SIZE)*4); 5480 return 0; 5481 } 5482 return 1; 5483 } 5484 5485 static int check_reshape(mddev_t *mddev) 5486 { 5487 raid5_conf_t *conf = mddev->private; 5488 5489 if (mddev->delta_disks == 0 && 5490 mddev->new_layout == mddev->layout && 5491 mddev->new_chunk_sectors == mddev->chunk_sectors) 5492 return 0; /* nothing to do */ 5493 if (mddev->bitmap) 5494 /* Cannot grow a bitmap yet */ 5495 return -EBUSY; 5496 if (has_failed(conf)) 5497 return -EINVAL; 5498 if (mddev->delta_disks < 0) { 5499 /* We might be able to shrink, but the devices must 5500 * be made bigger first. 5501 * For raid6, 4 is the minimum size. 5502 * Otherwise 2 is the minimum 5503 */ 5504 int min = 2; 5505 if (mddev->level == 6) 5506 min = 4; 5507 if (mddev->raid_disks + mddev->delta_disks < min) 5508 return -EINVAL; 5509 } 5510 5511 if (!check_stripe_cache(mddev)) 5512 return -ENOSPC; 5513 5514 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks); 5515 } 5516 5517 static int raid5_start_reshape(mddev_t *mddev) 5518 { 5519 raid5_conf_t *conf = mddev->private; 5520 mdk_rdev_t *rdev; 5521 int spares = 0; 5522 int added_devices = 0; 5523 unsigned long flags; 5524 5525 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery)) 5526 return -EBUSY; 5527 5528 if (!check_stripe_cache(mddev)) 5529 return -ENOSPC; 5530 5531 list_for_each_entry(rdev, &mddev->disks, same_set) 5532 if (rdev->raid_disk < 0 && 5533 !test_bit(Faulty, &rdev->flags)) 5534 spares++; 5535 5536 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded) 5537 /* Not enough devices even to make a degraded array 5538 * of that size 5539 */ 5540 return -EINVAL; 5541 5542 /* Refuse to reduce size of the array. Any reductions in 5543 * array size must be through explicit setting of array_size 5544 * attribute. 5545 */ 5546 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks) 5547 < mddev->array_sectors) { 5548 printk(KERN_ERR "md/raid:%s: array size must be reduced " 5549 "before number of disks\n", mdname(mddev)); 5550 return -EINVAL; 5551 } 5552 5553 atomic_set(&conf->reshape_stripes, 0); 5554 spin_lock_irq(&conf->device_lock); 5555 conf->previous_raid_disks = conf->raid_disks; 5556 conf->raid_disks += mddev->delta_disks; 5557 conf->prev_chunk_sectors = conf->chunk_sectors; 5558 conf->chunk_sectors = mddev->new_chunk_sectors; 5559 conf->prev_algo = conf->algorithm; 5560 conf->algorithm = mddev->new_layout; 5561 if (mddev->delta_disks < 0) 5562 conf->reshape_progress = raid5_size(mddev, 0, 0); 5563 else 5564 conf->reshape_progress = 0; 5565 conf->reshape_safe = conf->reshape_progress; 5566 conf->generation++; 5567 spin_unlock_irq(&conf->device_lock); 5568 5569 /* Add some new drives, as many as will fit. 5570 * We know there are enough to make the newly sized array work. 5571 * Don't add devices if we are reducing the number of 5572 * devices in the array. This is because it is not possible 5573 * to correctly record the "partially reconstructed" state of 5574 * such devices during the reshape and confusion could result. 5575 */ 5576 if (mddev->delta_disks >= 0) 5577 list_for_each_entry(rdev, &mddev->disks, same_set) 5578 if (rdev->raid_disk < 0 && 5579 !test_bit(Faulty, &rdev->flags)) { 5580 if (raid5_add_disk(mddev, rdev) == 0) { 5581 char nm[20]; 5582 if (rdev->raid_disk >= conf->previous_raid_disks) { 5583 set_bit(In_sync, &rdev->flags); 5584 added_devices++; 5585 } else 5586 rdev->recovery_offset = 0; 5587 sprintf(nm, "rd%d", rdev->raid_disk); 5588 if (sysfs_create_link(&mddev->kobj, 5589 &rdev->kobj, nm)) 5590 /* Failure here is OK */; 5591 } else 5592 break; 5593 } 5594 5595 /* When a reshape changes the number of devices, ->degraded 5596 * is measured against the larger of the pre and post number of 5597 * devices.*/ 5598 if (mddev->delta_disks > 0) { 5599 spin_lock_irqsave(&conf->device_lock, flags); 5600 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks) 5601 - added_devices; 5602 spin_unlock_irqrestore(&conf->device_lock, flags); 5603 } 5604 mddev->raid_disks = conf->raid_disks; 5605 mddev->reshape_position = conf->reshape_progress; 5606 set_bit(MD_CHANGE_DEVS, &mddev->flags); 5607 5608 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery); 5609 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery); 5610 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery); 5611 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery); 5612 mddev->sync_thread = md_register_thread(md_do_sync, mddev, 5613 "reshape"); 5614 if (!mddev->sync_thread) { 5615 mddev->recovery = 0; 5616 spin_lock_irq(&conf->device_lock); 5617 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks; 5618 conf->reshape_progress = MaxSector; 5619 spin_unlock_irq(&conf->device_lock); 5620 return -EAGAIN; 5621 } 5622 conf->reshape_checkpoint = jiffies; 5623 md_wakeup_thread(mddev->sync_thread); 5624 md_new_event(mddev); 5625 return 0; 5626 } 5627 5628 /* This is called from the reshape thread and should make any 5629 * changes needed in 'conf' 5630 */ 5631 static void end_reshape(raid5_conf_t *conf) 5632 { 5633 5634 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) { 5635 5636 spin_lock_irq(&conf->device_lock); 5637 conf->previous_raid_disks = conf->raid_disks; 5638 conf->reshape_progress = MaxSector; 5639 spin_unlock_irq(&conf->device_lock); 5640 wake_up(&conf->wait_for_overlap); 5641 5642 /* read-ahead size must cover two whole stripes, which is 5643 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices 5644 */ 5645 if (conf->mddev->queue) { 5646 int data_disks = conf->raid_disks - conf->max_degraded; 5647 int stripe = data_disks * ((conf->chunk_sectors << 9) 5648 / PAGE_SIZE); 5649 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe) 5650 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe; 5651 } 5652 } 5653 } 5654 5655 /* This is called from the raid5d thread with mddev_lock held. 5656 * It makes config changes to the device. 5657 */ 5658 static void raid5_finish_reshape(mddev_t *mddev) 5659 { 5660 raid5_conf_t *conf = mddev->private; 5661 5662 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) { 5663 5664 if (mddev->delta_disks > 0) { 5665 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0)); 5666 set_capacity(mddev->gendisk, mddev->array_sectors); 5667 revalidate_disk(mddev->gendisk); 5668 } else { 5669 int d; 5670 mddev->degraded = conf->raid_disks; 5671 for (d = 0; d < conf->raid_disks ; d++) 5672 if (conf->disks[d].rdev && 5673 test_bit(In_sync, 5674 &conf->disks[d].rdev->flags)) 5675 mddev->degraded--; 5676 for (d = conf->raid_disks ; 5677 d < conf->raid_disks - mddev->delta_disks; 5678 d++) { 5679 mdk_rdev_t *rdev = conf->disks[d].rdev; 5680 if (rdev && raid5_remove_disk(mddev, d) == 0) { 5681 char nm[20]; 5682 sprintf(nm, "rd%d", rdev->raid_disk); 5683 sysfs_remove_link(&mddev->kobj, nm); 5684 rdev->raid_disk = -1; 5685 } 5686 } 5687 } 5688 mddev->layout = conf->algorithm; 5689 mddev->chunk_sectors = conf->chunk_sectors; 5690 mddev->reshape_position = MaxSector; 5691 mddev->delta_disks = 0; 5692 } 5693 } 5694 5695 static void raid5_quiesce(mddev_t *mddev, int state) 5696 { 5697 raid5_conf_t *conf = mddev->private; 5698 5699 switch(state) { 5700 case 2: /* resume for a suspend */ 5701 wake_up(&conf->wait_for_overlap); 5702 break; 5703 5704 case 1: /* stop all writes */ 5705 spin_lock_irq(&conf->device_lock); 5706 /* '2' tells resync/reshape to pause so that all 5707 * active stripes can drain 5708 */ 5709 conf->quiesce = 2; 5710 wait_event_lock_irq(conf->wait_for_stripe, 5711 atomic_read(&conf->active_stripes) == 0 && 5712 atomic_read(&conf->active_aligned_reads) == 0, 5713 conf->device_lock, /* nothing */); 5714 conf->quiesce = 1; 5715 spin_unlock_irq(&conf->device_lock); 5716 /* allow reshape to continue */ 5717 wake_up(&conf->wait_for_overlap); 5718 break; 5719 5720 case 0: /* re-enable writes */ 5721 spin_lock_irq(&conf->device_lock); 5722 conf->quiesce = 0; 5723 wake_up(&conf->wait_for_stripe); 5724 wake_up(&conf->wait_for_overlap); 5725 spin_unlock_irq(&conf->device_lock); 5726 break; 5727 } 5728 } 5729 5730 5731 static void *raid45_takeover_raid0(mddev_t *mddev, int level) 5732 { 5733 struct raid0_private_data *raid0_priv = mddev->private; 5734 5735 /* for raid0 takeover only one zone is supported */ 5736 if (raid0_priv->nr_strip_zones > 1) { 5737 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n", 5738 mdname(mddev)); 5739 return ERR_PTR(-EINVAL); 5740 } 5741 5742 mddev->new_level = level; 5743 mddev->new_layout = ALGORITHM_PARITY_N; 5744 mddev->new_chunk_sectors = mddev->chunk_sectors; 5745 mddev->raid_disks += 1; 5746 mddev->delta_disks = 1; 5747 /* make sure it will be not marked as dirty */ 5748 mddev->recovery_cp = MaxSector; 5749 5750 return setup_conf(mddev); 5751 } 5752 5753 5754 static void *raid5_takeover_raid1(mddev_t *mddev) 5755 { 5756 int chunksect; 5757 5758 if (mddev->raid_disks != 2 || 5759 mddev->degraded > 1) 5760 return ERR_PTR(-EINVAL); 5761 5762 /* Should check if there are write-behind devices? */ 5763 5764 chunksect = 64*2; /* 64K by default */ 5765 5766 /* The array must be an exact multiple of chunksize */ 5767 while (chunksect && (mddev->array_sectors & (chunksect-1))) 5768 chunksect >>= 1; 5769 5770 if ((chunksect<<9) < STRIPE_SIZE) 5771 /* array size does not allow a suitable chunk size */ 5772 return ERR_PTR(-EINVAL); 5773 5774 mddev->new_level = 5; 5775 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC; 5776 mddev->new_chunk_sectors = chunksect; 5777 5778 return setup_conf(mddev); 5779 } 5780 5781 static void *raid5_takeover_raid6(mddev_t *mddev) 5782 { 5783 int new_layout; 5784 5785 switch (mddev->layout) { 5786 case ALGORITHM_LEFT_ASYMMETRIC_6: 5787 new_layout = ALGORITHM_LEFT_ASYMMETRIC; 5788 break; 5789 case ALGORITHM_RIGHT_ASYMMETRIC_6: 5790 new_layout = ALGORITHM_RIGHT_ASYMMETRIC; 5791 break; 5792 case ALGORITHM_LEFT_SYMMETRIC_6: 5793 new_layout = ALGORITHM_LEFT_SYMMETRIC; 5794 break; 5795 case ALGORITHM_RIGHT_SYMMETRIC_6: 5796 new_layout = ALGORITHM_RIGHT_SYMMETRIC; 5797 break; 5798 case ALGORITHM_PARITY_0_6: 5799 new_layout = ALGORITHM_PARITY_0; 5800 break; 5801 case ALGORITHM_PARITY_N: 5802 new_layout = ALGORITHM_PARITY_N; 5803 break; 5804 default: 5805 return ERR_PTR(-EINVAL); 5806 } 5807 mddev->new_level = 5; 5808 mddev->new_layout = new_layout; 5809 mddev->delta_disks = -1; 5810 mddev->raid_disks -= 1; 5811 return setup_conf(mddev); 5812 } 5813 5814 5815 static int raid5_check_reshape(mddev_t *mddev) 5816 { 5817 /* For a 2-drive array, the layout and chunk size can be changed 5818 * immediately as not restriping is needed. 5819 * For larger arrays we record the new value - after validation 5820 * to be used by a reshape pass. 5821 */ 5822 raid5_conf_t *conf = mddev->private; 5823 int new_chunk = mddev->new_chunk_sectors; 5824 5825 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout)) 5826 return -EINVAL; 5827 if (new_chunk > 0) { 5828 if (!is_power_of_2(new_chunk)) 5829 return -EINVAL; 5830 if (new_chunk < (PAGE_SIZE>>9)) 5831 return -EINVAL; 5832 if (mddev->array_sectors & (new_chunk-1)) 5833 /* not factor of array size */ 5834 return -EINVAL; 5835 } 5836 5837 /* They look valid */ 5838 5839 if (mddev->raid_disks == 2) { 5840 /* can make the change immediately */ 5841 if (mddev->new_layout >= 0) { 5842 conf->algorithm = mddev->new_layout; 5843 mddev->layout = mddev->new_layout; 5844 } 5845 if (new_chunk > 0) { 5846 conf->chunk_sectors = new_chunk ; 5847 mddev->chunk_sectors = new_chunk; 5848 } 5849 set_bit(MD_CHANGE_DEVS, &mddev->flags); 5850 md_wakeup_thread(mddev->thread); 5851 } 5852 return check_reshape(mddev); 5853 } 5854 5855 static int raid6_check_reshape(mddev_t *mddev) 5856 { 5857 int new_chunk = mddev->new_chunk_sectors; 5858 5859 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout)) 5860 return -EINVAL; 5861 if (new_chunk > 0) { 5862 if (!is_power_of_2(new_chunk)) 5863 return -EINVAL; 5864 if (new_chunk < (PAGE_SIZE >> 9)) 5865 return -EINVAL; 5866 if (mddev->array_sectors & (new_chunk-1)) 5867 /* not factor of array size */ 5868 return -EINVAL; 5869 } 5870 5871 /* They look valid */ 5872 return check_reshape(mddev); 5873 } 5874 5875 static void *raid5_takeover(mddev_t *mddev) 5876 { 5877 /* raid5 can take over: 5878 * raid0 - if there is only one strip zone - make it a raid4 layout 5879 * raid1 - if there are two drives. We need to know the chunk size 5880 * raid4 - trivial - just use a raid4 layout. 5881 * raid6 - Providing it is a *_6 layout 5882 */ 5883 if (mddev->level == 0) 5884 return raid45_takeover_raid0(mddev, 5); 5885 if (mddev->level == 1) 5886 return raid5_takeover_raid1(mddev); 5887 if (mddev->level == 4) { 5888 mddev->new_layout = ALGORITHM_PARITY_N; 5889 mddev->new_level = 5; 5890 return setup_conf(mddev); 5891 } 5892 if (mddev->level == 6) 5893 return raid5_takeover_raid6(mddev); 5894 5895 return ERR_PTR(-EINVAL); 5896 } 5897 5898 static void *raid4_takeover(mddev_t *mddev) 5899 { 5900 /* raid4 can take over: 5901 * raid0 - if there is only one strip zone 5902 * raid5 - if layout is right 5903 */ 5904 if (mddev->level == 0) 5905 return raid45_takeover_raid0(mddev, 4); 5906 if (mddev->level == 5 && 5907 mddev->layout == ALGORITHM_PARITY_N) { 5908 mddev->new_layout = 0; 5909 mddev->new_level = 4; 5910 return setup_conf(mddev); 5911 } 5912 return ERR_PTR(-EINVAL); 5913 } 5914 5915 static struct mdk_personality raid5_personality; 5916 5917 static void *raid6_takeover(mddev_t *mddev) 5918 { 5919 /* Currently can only take over a raid5. We map the 5920 * personality to an equivalent raid6 personality 5921 * with the Q block at the end. 5922 */ 5923 int new_layout; 5924 5925 if (mddev->pers != &raid5_personality) 5926 return ERR_PTR(-EINVAL); 5927 if (mddev->degraded > 1) 5928 return ERR_PTR(-EINVAL); 5929 if (mddev->raid_disks > 253) 5930 return ERR_PTR(-EINVAL); 5931 if (mddev->raid_disks < 3) 5932 return ERR_PTR(-EINVAL); 5933 5934 switch (mddev->layout) { 5935 case ALGORITHM_LEFT_ASYMMETRIC: 5936 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6; 5937 break; 5938 case ALGORITHM_RIGHT_ASYMMETRIC: 5939 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6; 5940 break; 5941 case ALGORITHM_LEFT_SYMMETRIC: 5942 new_layout = ALGORITHM_LEFT_SYMMETRIC_6; 5943 break; 5944 case ALGORITHM_RIGHT_SYMMETRIC: 5945 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6; 5946 break; 5947 case ALGORITHM_PARITY_0: 5948 new_layout = ALGORITHM_PARITY_0_6; 5949 break; 5950 case ALGORITHM_PARITY_N: 5951 new_layout = ALGORITHM_PARITY_N; 5952 break; 5953 default: 5954 return ERR_PTR(-EINVAL); 5955 } 5956 mddev->new_level = 6; 5957 mddev->new_layout = new_layout; 5958 mddev->delta_disks = 1; 5959 mddev->raid_disks += 1; 5960 return setup_conf(mddev); 5961 } 5962 5963 5964 static struct mdk_personality raid6_personality = 5965 { 5966 .name = "raid6", 5967 .level = 6, 5968 .owner = THIS_MODULE, 5969 .make_request = make_request, 5970 .run = run, 5971 .stop = stop, 5972 .status = status, 5973 .error_handler = error, 5974 .hot_add_disk = raid5_add_disk, 5975 .hot_remove_disk= raid5_remove_disk, 5976 .spare_active = raid5_spare_active, 5977 .sync_request = sync_request, 5978 .resize = raid5_resize, 5979 .size = raid5_size, 5980 .check_reshape = raid6_check_reshape, 5981 .start_reshape = raid5_start_reshape, 5982 .finish_reshape = raid5_finish_reshape, 5983 .quiesce = raid5_quiesce, 5984 .takeover = raid6_takeover, 5985 }; 5986 static struct mdk_personality raid5_personality = 5987 { 5988 .name = "raid5", 5989 .level = 5, 5990 .owner = THIS_MODULE, 5991 .make_request = make_request, 5992 .run = run, 5993 .stop = stop, 5994 .status = status, 5995 .error_handler = error, 5996 .hot_add_disk = raid5_add_disk, 5997 .hot_remove_disk= raid5_remove_disk, 5998 .spare_active = raid5_spare_active, 5999 .sync_request = sync_request, 6000 .resize = raid5_resize, 6001 .size = raid5_size, 6002 .check_reshape = raid5_check_reshape, 6003 .start_reshape = raid5_start_reshape, 6004 .finish_reshape = raid5_finish_reshape, 6005 .quiesce = raid5_quiesce, 6006 .takeover = raid5_takeover, 6007 }; 6008 6009 static struct mdk_personality raid4_personality = 6010 { 6011 .name = "raid4", 6012 .level = 4, 6013 .owner = THIS_MODULE, 6014 .make_request = make_request, 6015 .run = run, 6016 .stop = stop, 6017 .status = status, 6018 .error_handler = error, 6019 .hot_add_disk = raid5_add_disk, 6020 .hot_remove_disk= raid5_remove_disk, 6021 .spare_active = raid5_spare_active, 6022 .sync_request = sync_request, 6023 .resize = raid5_resize, 6024 .size = raid5_size, 6025 .check_reshape = raid5_check_reshape, 6026 .start_reshape = raid5_start_reshape, 6027 .finish_reshape = raid5_finish_reshape, 6028 .quiesce = raid5_quiesce, 6029 .takeover = raid4_takeover, 6030 }; 6031 6032 static int __init raid5_init(void) 6033 { 6034 register_md_personality(&raid6_personality); 6035 register_md_personality(&raid5_personality); 6036 register_md_personality(&raid4_personality); 6037 return 0; 6038 } 6039 6040 static void raid5_exit(void) 6041 { 6042 unregister_md_personality(&raid6_personality); 6043 unregister_md_personality(&raid5_personality); 6044 unregister_md_personality(&raid4_personality); 6045 } 6046 6047 module_init(raid5_init); 6048 module_exit(raid5_exit); 6049 MODULE_LICENSE("GPL"); 6050 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD"); 6051 MODULE_ALIAS("md-personality-4"); /* RAID5 */ 6052 MODULE_ALIAS("md-raid5"); 6053 MODULE_ALIAS("md-raid4"); 6054 MODULE_ALIAS("md-level-5"); 6055 MODULE_ALIAS("md-level-4"); 6056 MODULE_ALIAS("md-personality-8"); /* RAID6 */ 6057 MODULE_ALIAS("md-raid6"); 6058 MODULE_ALIAS("md-level-6"); 6059 6060 /* This used to be two separate modules, they were: */ 6061 MODULE_ALIAS("raid5"); 6062 MODULE_ALIAS("raid6"); 6063