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
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!
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)
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.
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
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
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
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/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <trace/events/block.h>
67 #define NR_STRIPES 256
68 #define STRIPE_SIZE PAGE_SIZE
69 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
70 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
71 #define IO_THRESHOLD 1
72 #define BYPASS_THRESHOLD 1
73 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
74 #define HASH_MASK (NR_HASH - 1)
76 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
78 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
79 return &conf->stripe_hashtbl[hash];
82 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
83 * order without overlap. There may be several bio's per stripe+device, and
84 * a bio could span several devices.
85 * When walking this list for a particular stripe+device, we must never proceed
86 * beyond a bio that extends past this device, as the next bio might no longer
88 * This function is used to determine the 'next' bio in the list, given the sector
89 * of the current stripe+device
91 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
93 int sectors = bio->bi_size >> 9;
94 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
101 * We maintain a biased count of active stripes in the bottom 16 bits of
102 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
104 static inline int raid5_bi_processed_stripes(struct bio *bio)
106 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
107 return (atomic_read(segments) >> 16) & 0xffff;
110 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
112 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
113 return atomic_sub_return(1, segments) & 0xffff;
116 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
118 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
119 atomic_inc(segments);
122 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
125 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
129 old = atomic_read(segments);
130 new = (old & 0xffff) | (cnt << 16);
131 } while (atomic_cmpxchg(segments, old, new) != old);
134 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
136 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
137 atomic_set(segments, cnt);
140 /* Find first data disk in a raid6 stripe */
141 static inline int raid6_d0(struct stripe_head *sh)
144 /* ddf always start from first device */
146 /* md starts just after Q block */
147 if (sh->qd_idx == sh->disks - 1)
150 return sh->qd_idx + 1;
152 static inline int raid6_next_disk(int disk, int raid_disks)
155 return (disk < raid_disks) ? disk : 0;
158 /* When walking through the disks in a raid5, starting at raid6_d0,
159 * We need to map each disk to a 'slot', where the data disks are slot
160 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
161 * is raid_disks-1. This help does that mapping.
163 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
164 int *count, int syndrome_disks)
170 if (idx == sh->pd_idx)
171 return syndrome_disks;
172 if (idx == sh->qd_idx)
173 return syndrome_disks + 1;
179 static void return_io(struct bio *return_bi)
181 struct bio *bi = return_bi;
184 return_bi = bi->bi_next;
192 static void print_raid5_conf (struct r5conf *conf);
194 static int stripe_operations_active(struct stripe_head *sh)
196 return sh->check_state || sh->reconstruct_state ||
197 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
198 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
201 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
203 BUG_ON(!list_empty(&sh->lru));
204 BUG_ON(atomic_read(&conf->active_stripes)==0);
205 if (test_bit(STRIPE_HANDLE, &sh->state)) {
206 if (test_bit(STRIPE_DELAYED, &sh->state) &&
207 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
208 list_add_tail(&sh->lru, &conf->delayed_list);
209 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
210 sh->bm_seq - conf->seq_write > 0)
211 list_add_tail(&sh->lru, &conf->bitmap_list);
213 clear_bit(STRIPE_DELAYED, &sh->state);
214 clear_bit(STRIPE_BIT_DELAY, &sh->state);
215 list_add_tail(&sh->lru, &conf->handle_list);
217 md_wakeup_thread(conf->mddev->thread);
219 BUG_ON(stripe_operations_active(sh));
220 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
221 if (atomic_dec_return(&conf->preread_active_stripes)
223 md_wakeup_thread(conf->mddev->thread);
224 atomic_dec(&conf->active_stripes);
225 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
226 list_add_tail(&sh->lru, &conf->inactive_list);
227 wake_up(&conf->wait_for_stripe);
228 if (conf->retry_read_aligned)
229 md_wakeup_thread(conf->mddev->thread);
234 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
236 if (atomic_dec_and_test(&sh->count))
237 do_release_stripe(conf, sh);
240 static void release_stripe(struct stripe_head *sh)
242 struct r5conf *conf = sh->raid_conf;
245 local_irq_save(flags);
246 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
247 do_release_stripe(conf, sh);
248 spin_unlock(&conf->device_lock);
250 local_irq_restore(flags);
253 static inline void remove_hash(struct stripe_head *sh)
255 pr_debug("remove_hash(), stripe %llu\n",
256 (unsigned long long)sh->sector);
258 hlist_del_init(&sh->hash);
261 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
263 struct hlist_head *hp = stripe_hash(conf, sh->sector);
265 pr_debug("insert_hash(), stripe %llu\n",
266 (unsigned long long)sh->sector);
268 hlist_add_head(&sh->hash, hp);
272 /* find an idle stripe, make sure it is unhashed, and return it. */
273 static struct stripe_head *get_free_stripe(struct r5conf *conf)
275 struct stripe_head *sh = NULL;
276 struct list_head *first;
278 if (list_empty(&conf->inactive_list))
280 first = conf->inactive_list.next;
281 sh = list_entry(first, struct stripe_head, lru);
282 list_del_init(first);
284 atomic_inc(&conf->active_stripes);
289 static void shrink_buffers(struct stripe_head *sh)
293 int num = sh->raid_conf->pool_size;
295 for (i = 0; i < num ; i++) {
299 sh->dev[i].page = NULL;
304 static int grow_buffers(struct stripe_head *sh)
307 int num = sh->raid_conf->pool_size;
309 for (i = 0; i < num; i++) {
312 if (!(page = alloc_page(GFP_KERNEL))) {
315 sh->dev[i].page = page;
320 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
321 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
322 struct stripe_head *sh);
324 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
326 struct r5conf *conf = sh->raid_conf;
329 BUG_ON(atomic_read(&sh->count) != 0);
330 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
331 BUG_ON(stripe_operations_active(sh));
333 pr_debug("init_stripe called, stripe %llu\n",
334 (unsigned long long)sh->sector);
338 sh->generation = conf->generation - previous;
339 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
341 stripe_set_idx(sector, conf, previous, sh);
345 for (i = sh->disks; i--; ) {
346 struct r5dev *dev = &sh->dev[i];
348 if (dev->toread || dev->read || dev->towrite || dev->written ||
349 test_bit(R5_LOCKED, &dev->flags)) {
350 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
351 (unsigned long long)sh->sector, i, dev->toread,
352 dev->read, dev->towrite, dev->written,
353 test_bit(R5_LOCKED, &dev->flags));
357 raid5_build_block(sh, i, previous);
359 insert_hash(conf, sh);
362 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
365 struct stripe_head *sh;
367 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
368 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
369 if (sh->sector == sector && sh->generation == generation)
371 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
376 * Need to check if array has failed when deciding whether to:
378 * - remove non-faulty devices
381 * This determination is simple when no reshape is happening.
382 * However if there is a reshape, we need to carefully check
383 * both the before and after sections.
384 * This is because some failed devices may only affect one
385 * of the two sections, and some non-in_sync devices may
386 * be insync in the section most affected by failed devices.
388 static int calc_degraded(struct r5conf *conf)
390 int degraded, degraded2;
395 for (i = 0; i < conf->previous_raid_disks; i++) {
396 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
397 if (rdev && test_bit(Faulty, &rdev->flags))
398 rdev = rcu_dereference(conf->disks[i].replacement);
399 if (!rdev || test_bit(Faulty, &rdev->flags))
401 else if (test_bit(In_sync, &rdev->flags))
404 /* not in-sync or faulty.
405 * If the reshape increases the number of devices,
406 * this is being recovered by the reshape, so
407 * this 'previous' section is not in_sync.
408 * If the number of devices is being reduced however,
409 * the device can only be part of the array if
410 * we are reverting a reshape, so this section will
413 if (conf->raid_disks >= conf->previous_raid_disks)
417 if (conf->raid_disks == conf->previous_raid_disks)
421 for (i = 0; i < conf->raid_disks; i++) {
422 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
423 if (rdev && test_bit(Faulty, &rdev->flags))
424 rdev = rcu_dereference(conf->disks[i].replacement);
425 if (!rdev || test_bit(Faulty, &rdev->flags))
427 else if (test_bit(In_sync, &rdev->flags))
430 /* not in-sync or faulty.
431 * If reshape increases the number of devices, this
432 * section has already been recovered, else it
433 * almost certainly hasn't.
435 if (conf->raid_disks <= conf->previous_raid_disks)
439 if (degraded2 > degraded)
444 static int has_failed(struct r5conf *conf)
448 if (conf->mddev->reshape_position == MaxSector)
449 return conf->mddev->degraded > conf->max_degraded;
451 degraded = calc_degraded(conf);
452 if (degraded > conf->max_degraded)
457 static struct stripe_head *
458 get_active_stripe(struct r5conf *conf, sector_t sector,
459 int previous, int noblock, int noquiesce)
461 struct stripe_head *sh;
463 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
465 spin_lock_irq(&conf->device_lock);
468 wait_event_lock_irq(conf->wait_for_stripe,
469 conf->quiesce == 0 || noquiesce,
471 sh = __find_stripe(conf, sector, conf->generation - previous);
473 if (!conf->inactive_blocked)
474 sh = get_free_stripe(conf);
475 if (noblock && sh == NULL)
478 conf->inactive_blocked = 1;
479 wait_event_lock_irq(conf->wait_for_stripe,
480 !list_empty(&conf->inactive_list) &&
481 (atomic_read(&conf->active_stripes)
482 < (conf->max_nr_stripes *3/4)
483 || !conf->inactive_blocked),
485 conf->inactive_blocked = 0;
487 init_stripe(sh, sector, previous);
489 if (atomic_read(&sh->count)) {
490 BUG_ON(!list_empty(&sh->lru)
491 && !test_bit(STRIPE_EXPANDING, &sh->state)
492 && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state));
494 if (!test_bit(STRIPE_HANDLE, &sh->state))
495 atomic_inc(&conf->active_stripes);
496 if (list_empty(&sh->lru) &&
497 !test_bit(STRIPE_EXPANDING, &sh->state))
499 list_del_init(&sh->lru);
502 } while (sh == NULL);
505 atomic_inc(&sh->count);
507 spin_unlock_irq(&conf->device_lock);
511 /* Determine if 'data_offset' or 'new_data_offset' should be used
512 * in this stripe_head.
514 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
516 sector_t progress = conf->reshape_progress;
517 /* Need a memory barrier to make sure we see the value
518 * of conf->generation, or ->data_offset that was set before
519 * reshape_progress was updated.
522 if (progress == MaxSector)
524 if (sh->generation == conf->generation - 1)
526 /* We are in a reshape, and this is a new-generation stripe,
527 * so use new_data_offset.
533 raid5_end_read_request(struct bio *bi, int error);
535 raid5_end_write_request(struct bio *bi, int error);
537 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
539 struct r5conf *conf = sh->raid_conf;
540 int i, disks = sh->disks;
544 for (i = disks; i--; ) {
546 int replace_only = 0;
547 struct bio *bi, *rbi;
548 struct md_rdev *rdev, *rrdev = NULL;
549 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
550 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
554 if (test_bit(R5_Discard, &sh->dev[i].flags))
556 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
558 else if (test_and_clear_bit(R5_WantReplace,
559 &sh->dev[i].flags)) {
564 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
567 bi = &sh->dev[i].req;
568 rbi = &sh->dev[i].rreq; /* For writing to replacement */
573 bi->bi_end_io = raid5_end_write_request;
574 rbi->bi_end_io = raid5_end_write_request;
576 bi->bi_end_io = raid5_end_read_request;
579 rrdev = rcu_dereference(conf->disks[i].replacement);
580 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
581 rdev = rcu_dereference(conf->disks[i].rdev);
590 /* We raced and saw duplicates */
593 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
598 if (rdev && test_bit(Faulty, &rdev->flags))
601 atomic_inc(&rdev->nr_pending);
602 if (rrdev && test_bit(Faulty, &rrdev->flags))
605 atomic_inc(&rrdev->nr_pending);
608 /* We have already checked bad blocks for reads. Now
609 * need to check for writes. We never accept write errors
610 * on the replacement, so we don't to check rrdev.
612 while ((rw & WRITE) && rdev &&
613 test_bit(WriteErrorSeen, &rdev->flags)) {
616 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
617 &first_bad, &bad_sectors);
622 set_bit(BlockedBadBlocks, &rdev->flags);
623 if (!conf->mddev->external &&
624 conf->mddev->flags) {
625 /* It is very unlikely, but we might
626 * still need to write out the
627 * bad block log - better give it
629 md_check_recovery(conf->mddev);
632 * Because md_wait_for_blocked_rdev
633 * will dec nr_pending, we must
634 * increment it first.
636 atomic_inc(&rdev->nr_pending);
637 md_wait_for_blocked_rdev(rdev, conf->mddev);
639 /* Acknowledged bad block - skip the write */
640 rdev_dec_pending(rdev, conf->mddev);
646 if (s->syncing || s->expanding || s->expanded
648 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
650 set_bit(STRIPE_IO_STARTED, &sh->state);
652 bi->bi_bdev = rdev->bdev;
653 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
654 __func__, (unsigned long long)sh->sector,
656 atomic_inc(&sh->count);
657 if (use_new_offset(conf, sh))
658 bi->bi_sector = (sh->sector
659 + rdev->new_data_offset);
661 bi->bi_sector = (sh->sector
662 + rdev->data_offset);
663 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
664 bi->bi_rw |= REQ_FLUSH;
666 bi->bi_flags = 1 << BIO_UPTODATE;
668 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
669 bi->bi_io_vec[0].bv_offset = 0;
670 bi->bi_size = STRIPE_SIZE;
673 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
674 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
675 bi, disk_devt(conf->mddev->gendisk),
677 generic_make_request(bi);
680 if (s->syncing || s->expanding || s->expanded
682 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
684 set_bit(STRIPE_IO_STARTED, &sh->state);
686 rbi->bi_bdev = rrdev->bdev;
687 pr_debug("%s: for %llu schedule op %ld on "
688 "replacement disc %d\n",
689 __func__, (unsigned long long)sh->sector,
691 atomic_inc(&sh->count);
692 if (use_new_offset(conf, sh))
693 rbi->bi_sector = (sh->sector
694 + rrdev->new_data_offset);
696 rbi->bi_sector = (sh->sector
697 + rrdev->data_offset);
698 rbi->bi_flags = 1 << BIO_UPTODATE;
700 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
701 rbi->bi_io_vec[0].bv_offset = 0;
702 rbi->bi_size = STRIPE_SIZE;
704 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
705 rbi, disk_devt(conf->mddev->gendisk),
707 generic_make_request(rbi);
709 if (!rdev && !rrdev) {
711 set_bit(STRIPE_DEGRADED, &sh->state);
712 pr_debug("skip op %ld on disc %d for sector %llu\n",
713 bi->bi_rw, i, (unsigned long long)sh->sector);
714 clear_bit(R5_LOCKED, &sh->dev[i].flags);
715 set_bit(STRIPE_HANDLE, &sh->state);
720 static struct dma_async_tx_descriptor *
721 async_copy_data(int frombio, struct bio *bio, struct page *page,
722 sector_t sector, struct dma_async_tx_descriptor *tx)
725 struct page *bio_page;
728 struct async_submit_ctl submit;
729 enum async_tx_flags flags = 0;
731 if (bio->bi_sector >= sector)
732 page_offset = (signed)(bio->bi_sector - sector) * 512;
734 page_offset = (signed)(sector - bio->bi_sector) * -512;
737 flags |= ASYNC_TX_FENCE;
738 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
740 bio_for_each_segment(bvl, bio, i) {
741 int len = bvl->bv_len;
745 if (page_offset < 0) {
746 b_offset = -page_offset;
747 page_offset += b_offset;
751 if (len > 0 && page_offset + len > STRIPE_SIZE)
752 clen = STRIPE_SIZE - page_offset;
757 b_offset += bvl->bv_offset;
758 bio_page = bvl->bv_page;
760 tx = async_memcpy(page, bio_page, page_offset,
761 b_offset, clen, &submit);
763 tx = async_memcpy(bio_page, page, b_offset,
764 page_offset, clen, &submit);
766 /* chain the operations */
767 submit.depend_tx = tx;
769 if (clen < len) /* hit end of page */
777 static void ops_complete_biofill(void *stripe_head_ref)
779 struct stripe_head *sh = stripe_head_ref;
780 struct bio *return_bi = NULL;
783 pr_debug("%s: stripe %llu\n", __func__,
784 (unsigned long long)sh->sector);
786 /* clear completed biofills */
787 for (i = sh->disks; i--; ) {
788 struct r5dev *dev = &sh->dev[i];
790 /* acknowledge completion of a biofill operation */
791 /* and check if we need to reply to a read request,
792 * new R5_Wantfill requests are held off until
793 * !STRIPE_BIOFILL_RUN
795 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
796 struct bio *rbi, *rbi2;
801 while (rbi && rbi->bi_sector <
802 dev->sector + STRIPE_SECTORS) {
803 rbi2 = r5_next_bio(rbi, dev->sector);
804 if (!raid5_dec_bi_active_stripes(rbi)) {
805 rbi->bi_next = return_bi;
812 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
814 return_io(return_bi);
816 set_bit(STRIPE_HANDLE, &sh->state);
820 static void ops_run_biofill(struct stripe_head *sh)
822 struct dma_async_tx_descriptor *tx = NULL;
823 struct async_submit_ctl submit;
826 pr_debug("%s: stripe %llu\n", __func__,
827 (unsigned long long)sh->sector);
829 for (i = sh->disks; i--; ) {
830 struct r5dev *dev = &sh->dev[i];
831 if (test_bit(R5_Wantfill, &dev->flags)) {
833 spin_lock_irq(&sh->stripe_lock);
834 dev->read = rbi = dev->toread;
836 spin_unlock_irq(&sh->stripe_lock);
837 while (rbi && rbi->bi_sector <
838 dev->sector + STRIPE_SECTORS) {
839 tx = async_copy_data(0, rbi, dev->page,
841 rbi = r5_next_bio(rbi, dev->sector);
846 atomic_inc(&sh->count);
847 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
848 async_trigger_callback(&submit);
851 static void mark_target_uptodate(struct stripe_head *sh, int target)
858 tgt = &sh->dev[target];
859 set_bit(R5_UPTODATE, &tgt->flags);
860 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
861 clear_bit(R5_Wantcompute, &tgt->flags);
864 static void ops_complete_compute(void *stripe_head_ref)
866 struct stripe_head *sh = stripe_head_ref;
868 pr_debug("%s: stripe %llu\n", __func__,
869 (unsigned long long)sh->sector);
871 /* mark the computed target(s) as uptodate */
872 mark_target_uptodate(sh, sh->ops.target);
873 mark_target_uptodate(sh, sh->ops.target2);
875 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
876 if (sh->check_state == check_state_compute_run)
877 sh->check_state = check_state_compute_result;
878 set_bit(STRIPE_HANDLE, &sh->state);
882 /* return a pointer to the address conversion region of the scribble buffer */
883 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
884 struct raid5_percpu *percpu)
886 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
889 static struct dma_async_tx_descriptor *
890 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
892 int disks = sh->disks;
893 struct page **xor_srcs = percpu->scribble;
894 int target = sh->ops.target;
895 struct r5dev *tgt = &sh->dev[target];
896 struct page *xor_dest = tgt->page;
898 struct dma_async_tx_descriptor *tx;
899 struct async_submit_ctl submit;
902 pr_debug("%s: stripe %llu block: %d\n",
903 __func__, (unsigned long long)sh->sector, target);
904 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
906 for (i = disks; i--; )
908 xor_srcs[count++] = sh->dev[i].page;
910 atomic_inc(&sh->count);
912 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
913 ops_complete_compute, sh, to_addr_conv(sh, percpu));
914 if (unlikely(count == 1))
915 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
917 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
922 /* set_syndrome_sources - populate source buffers for gen_syndrome
923 * @srcs - (struct page *) array of size sh->disks
924 * @sh - stripe_head to parse
926 * Populates srcs in proper layout order for the stripe and returns the
927 * 'count' of sources to be used in a call to async_gen_syndrome. The P
928 * destination buffer is recorded in srcs[count] and the Q destination
929 * is recorded in srcs[count+1]].
931 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
933 int disks = sh->disks;
934 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
935 int d0_idx = raid6_d0(sh);
939 for (i = 0; i < disks; i++)
945 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
947 srcs[slot] = sh->dev[i].page;
948 i = raid6_next_disk(i, disks);
949 } while (i != d0_idx);
951 return syndrome_disks;
954 static struct dma_async_tx_descriptor *
955 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
957 int disks = sh->disks;
958 struct page **blocks = percpu->scribble;
960 int qd_idx = sh->qd_idx;
961 struct dma_async_tx_descriptor *tx;
962 struct async_submit_ctl submit;
968 if (sh->ops.target < 0)
969 target = sh->ops.target2;
970 else if (sh->ops.target2 < 0)
971 target = sh->ops.target;
973 /* we should only have one valid target */
976 pr_debug("%s: stripe %llu block: %d\n",
977 __func__, (unsigned long long)sh->sector, target);
979 tgt = &sh->dev[target];
980 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
983 atomic_inc(&sh->count);
985 if (target == qd_idx) {
986 count = set_syndrome_sources(blocks, sh);
987 blocks[count] = NULL; /* regenerating p is not necessary */
988 BUG_ON(blocks[count+1] != dest); /* q should already be set */
989 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
990 ops_complete_compute, sh,
991 to_addr_conv(sh, percpu));
992 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
994 /* Compute any data- or p-drive using XOR */
996 for (i = disks; i-- ; ) {
997 if (i == target || i == qd_idx)
999 blocks[count++] = sh->dev[i].page;
1002 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1003 NULL, ops_complete_compute, sh,
1004 to_addr_conv(sh, percpu));
1005 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1011 static struct dma_async_tx_descriptor *
1012 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1014 int i, count, disks = sh->disks;
1015 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1016 int d0_idx = raid6_d0(sh);
1017 int faila = -1, failb = -1;
1018 int target = sh->ops.target;
1019 int target2 = sh->ops.target2;
1020 struct r5dev *tgt = &sh->dev[target];
1021 struct r5dev *tgt2 = &sh->dev[target2];
1022 struct dma_async_tx_descriptor *tx;
1023 struct page **blocks = percpu->scribble;
1024 struct async_submit_ctl submit;
1026 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1027 __func__, (unsigned long long)sh->sector, target, target2);
1028 BUG_ON(target < 0 || target2 < 0);
1029 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1030 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1032 /* we need to open-code set_syndrome_sources to handle the
1033 * slot number conversion for 'faila' and 'failb'
1035 for (i = 0; i < disks ; i++)
1040 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1042 blocks[slot] = sh->dev[i].page;
1048 i = raid6_next_disk(i, disks);
1049 } while (i != d0_idx);
1051 BUG_ON(faila == failb);
1054 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1055 __func__, (unsigned long long)sh->sector, faila, failb);
1057 atomic_inc(&sh->count);
1059 if (failb == syndrome_disks+1) {
1060 /* Q disk is one of the missing disks */
1061 if (faila == syndrome_disks) {
1062 /* Missing P+Q, just recompute */
1063 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1064 ops_complete_compute, sh,
1065 to_addr_conv(sh, percpu));
1066 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1067 STRIPE_SIZE, &submit);
1071 int qd_idx = sh->qd_idx;
1073 /* Missing D+Q: recompute D from P, then recompute Q */
1074 if (target == qd_idx)
1075 data_target = target2;
1077 data_target = target;
1080 for (i = disks; i-- ; ) {
1081 if (i == data_target || i == qd_idx)
1083 blocks[count++] = sh->dev[i].page;
1085 dest = sh->dev[data_target].page;
1086 init_async_submit(&submit,
1087 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1089 to_addr_conv(sh, percpu));
1090 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1093 count = set_syndrome_sources(blocks, sh);
1094 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1095 ops_complete_compute, sh,
1096 to_addr_conv(sh, percpu));
1097 return async_gen_syndrome(blocks, 0, count+2,
1098 STRIPE_SIZE, &submit);
1101 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1102 ops_complete_compute, sh,
1103 to_addr_conv(sh, percpu));
1104 if (failb == syndrome_disks) {
1105 /* We're missing D+P. */
1106 return async_raid6_datap_recov(syndrome_disks+2,
1110 /* We're missing D+D. */
1111 return async_raid6_2data_recov(syndrome_disks+2,
1112 STRIPE_SIZE, faila, failb,
1119 static void ops_complete_prexor(void *stripe_head_ref)
1121 struct stripe_head *sh = stripe_head_ref;
1123 pr_debug("%s: stripe %llu\n", __func__,
1124 (unsigned long long)sh->sector);
1127 static struct dma_async_tx_descriptor *
1128 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1129 struct dma_async_tx_descriptor *tx)
1131 int disks = sh->disks;
1132 struct page **xor_srcs = percpu->scribble;
1133 int count = 0, pd_idx = sh->pd_idx, i;
1134 struct async_submit_ctl submit;
1136 /* existing parity data subtracted */
1137 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1139 pr_debug("%s: stripe %llu\n", __func__,
1140 (unsigned long long)sh->sector);
1142 for (i = disks; i--; ) {
1143 struct r5dev *dev = &sh->dev[i];
1144 /* Only process blocks that are known to be uptodate */
1145 if (test_bit(R5_Wantdrain, &dev->flags))
1146 xor_srcs[count++] = dev->page;
1149 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1150 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1151 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1156 static struct dma_async_tx_descriptor *
1157 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1159 int disks = sh->disks;
1162 pr_debug("%s: stripe %llu\n", __func__,
1163 (unsigned long long)sh->sector);
1165 for (i = disks; i--; ) {
1166 struct r5dev *dev = &sh->dev[i];
1169 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1172 spin_lock_irq(&sh->stripe_lock);
1173 chosen = dev->towrite;
1174 dev->towrite = NULL;
1175 BUG_ON(dev->written);
1176 wbi = dev->written = chosen;
1177 spin_unlock_irq(&sh->stripe_lock);
1179 while (wbi && wbi->bi_sector <
1180 dev->sector + STRIPE_SECTORS) {
1181 if (wbi->bi_rw & REQ_FUA)
1182 set_bit(R5_WantFUA, &dev->flags);
1183 if (wbi->bi_rw & REQ_SYNC)
1184 set_bit(R5_SyncIO, &dev->flags);
1185 if (wbi->bi_rw & REQ_DISCARD)
1186 set_bit(R5_Discard, &dev->flags);
1188 tx = async_copy_data(1, wbi, dev->page,
1190 wbi = r5_next_bio(wbi, dev->sector);
1198 static void ops_complete_reconstruct(void *stripe_head_ref)
1200 struct stripe_head *sh = stripe_head_ref;
1201 int disks = sh->disks;
1202 int pd_idx = sh->pd_idx;
1203 int qd_idx = sh->qd_idx;
1205 bool fua = false, sync = false, discard = false;
1207 pr_debug("%s: stripe %llu\n", __func__,
1208 (unsigned long long)sh->sector);
1210 for (i = disks; i--; ) {
1211 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1212 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1213 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1216 for (i = disks; i--; ) {
1217 struct r5dev *dev = &sh->dev[i];
1219 if (dev->written || i == pd_idx || i == qd_idx) {
1221 set_bit(R5_UPTODATE, &dev->flags);
1223 set_bit(R5_WantFUA, &dev->flags);
1225 set_bit(R5_SyncIO, &dev->flags);
1229 if (sh->reconstruct_state == reconstruct_state_drain_run)
1230 sh->reconstruct_state = reconstruct_state_drain_result;
1231 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1232 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1234 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1235 sh->reconstruct_state = reconstruct_state_result;
1238 set_bit(STRIPE_HANDLE, &sh->state);
1243 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1244 struct dma_async_tx_descriptor *tx)
1246 int disks = sh->disks;
1247 struct page **xor_srcs = percpu->scribble;
1248 struct async_submit_ctl submit;
1249 int count = 0, pd_idx = sh->pd_idx, i;
1250 struct page *xor_dest;
1252 unsigned long flags;
1254 pr_debug("%s: stripe %llu\n", __func__,
1255 (unsigned long long)sh->sector);
1257 for (i = 0; i < sh->disks; i++) {
1260 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1263 if (i >= sh->disks) {
1264 atomic_inc(&sh->count);
1265 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1266 ops_complete_reconstruct(sh);
1269 /* check if prexor is active which means only process blocks
1270 * that are part of a read-modify-write (written)
1272 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1274 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1275 for (i = disks; i--; ) {
1276 struct r5dev *dev = &sh->dev[i];
1278 xor_srcs[count++] = dev->page;
1281 xor_dest = sh->dev[pd_idx].page;
1282 for (i = disks; i--; ) {
1283 struct r5dev *dev = &sh->dev[i];
1285 xor_srcs[count++] = dev->page;
1289 /* 1/ if we prexor'd then the dest is reused as a source
1290 * 2/ if we did not prexor then we are redoing the parity
1291 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1292 * for the synchronous xor case
1294 flags = ASYNC_TX_ACK |
1295 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1297 atomic_inc(&sh->count);
1299 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1300 to_addr_conv(sh, percpu));
1301 if (unlikely(count == 1))
1302 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1304 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1308 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1309 struct dma_async_tx_descriptor *tx)
1311 struct async_submit_ctl submit;
1312 struct page **blocks = percpu->scribble;
1315 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1317 for (i = 0; i < sh->disks; i++) {
1318 if (sh->pd_idx == i || sh->qd_idx == i)
1320 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1323 if (i >= sh->disks) {
1324 atomic_inc(&sh->count);
1325 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1326 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1327 ops_complete_reconstruct(sh);
1331 count = set_syndrome_sources(blocks, sh);
1333 atomic_inc(&sh->count);
1335 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1336 sh, to_addr_conv(sh, percpu));
1337 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1340 static void ops_complete_check(void *stripe_head_ref)
1342 struct stripe_head *sh = stripe_head_ref;
1344 pr_debug("%s: stripe %llu\n", __func__,
1345 (unsigned long long)sh->sector);
1347 sh->check_state = check_state_check_result;
1348 set_bit(STRIPE_HANDLE, &sh->state);
1352 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1354 int disks = sh->disks;
1355 int pd_idx = sh->pd_idx;
1356 int qd_idx = sh->qd_idx;
1357 struct page *xor_dest;
1358 struct page **xor_srcs = percpu->scribble;
1359 struct dma_async_tx_descriptor *tx;
1360 struct async_submit_ctl submit;
1364 pr_debug("%s: stripe %llu\n", __func__,
1365 (unsigned long long)sh->sector);
1368 xor_dest = sh->dev[pd_idx].page;
1369 xor_srcs[count++] = xor_dest;
1370 for (i = disks; i--; ) {
1371 if (i == pd_idx || i == qd_idx)
1373 xor_srcs[count++] = sh->dev[i].page;
1376 init_async_submit(&submit, 0, NULL, NULL, NULL,
1377 to_addr_conv(sh, percpu));
1378 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1379 &sh->ops.zero_sum_result, &submit);
1381 atomic_inc(&sh->count);
1382 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1383 tx = async_trigger_callback(&submit);
1386 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1388 struct page **srcs = percpu->scribble;
1389 struct async_submit_ctl submit;
1392 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1393 (unsigned long long)sh->sector, checkp);
1395 count = set_syndrome_sources(srcs, sh);
1399 atomic_inc(&sh->count);
1400 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1401 sh, to_addr_conv(sh, percpu));
1402 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1403 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1406 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1408 int overlap_clear = 0, i, disks = sh->disks;
1409 struct dma_async_tx_descriptor *tx = NULL;
1410 struct r5conf *conf = sh->raid_conf;
1411 int level = conf->level;
1412 struct raid5_percpu *percpu;
1416 percpu = per_cpu_ptr(conf->percpu, cpu);
1417 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1418 ops_run_biofill(sh);
1422 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1424 tx = ops_run_compute5(sh, percpu);
1426 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1427 tx = ops_run_compute6_1(sh, percpu);
1429 tx = ops_run_compute6_2(sh, percpu);
1431 /* terminate the chain if reconstruct is not set to be run */
1432 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1436 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1437 tx = ops_run_prexor(sh, percpu, tx);
1439 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1440 tx = ops_run_biodrain(sh, tx);
1444 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1446 ops_run_reconstruct5(sh, percpu, tx);
1448 ops_run_reconstruct6(sh, percpu, tx);
1451 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1452 if (sh->check_state == check_state_run)
1453 ops_run_check_p(sh, percpu);
1454 else if (sh->check_state == check_state_run_q)
1455 ops_run_check_pq(sh, percpu, 0);
1456 else if (sh->check_state == check_state_run_pq)
1457 ops_run_check_pq(sh, percpu, 1);
1463 for (i = disks; i--; ) {
1464 struct r5dev *dev = &sh->dev[i];
1465 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1466 wake_up(&sh->raid_conf->wait_for_overlap);
1471 #ifdef CONFIG_MULTICORE_RAID456
1472 static void async_run_ops(void *param, async_cookie_t cookie)
1474 struct stripe_head *sh = param;
1475 unsigned long ops_request = sh->ops.request;
1477 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1478 wake_up(&sh->ops.wait_for_ops);
1480 __raid_run_ops(sh, ops_request);
1484 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1486 /* since handle_stripe can be called outside of raid5d context
1487 * we need to ensure sh->ops.request is de-staged before another
1490 wait_event(sh->ops.wait_for_ops,
1491 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1492 sh->ops.request = ops_request;
1494 atomic_inc(&sh->count);
1495 async_schedule(async_run_ops, sh);
1498 #define raid_run_ops __raid_run_ops
1501 static int grow_one_stripe(struct r5conf *conf)
1503 struct stripe_head *sh;
1504 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1508 sh->raid_conf = conf;
1509 #ifdef CONFIG_MULTICORE_RAID456
1510 init_waitqueue_head(&sh->ops.wait_for_ops);
1513 spin_lock_init(&sh->stripe_lock);
1515 if (grow_buffers(sh)) {
1517 kmem_cache_free(conf->slab_cache, sh);
1520 /* we just created an active stripe so... */
1521 atomic_set(&sh->count, 1);
1522 atomic_inc(&conf->active_stripes);
1523 INIT_LIST_HEAD(&sh->lru);
1528 static int grow_stripes(struct r5conf *conf, int num)
1530 struct kmem_cache *sc;
1531 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1533 if (conf->mddev->gendisk)
1534 sprintf(conf->cache_name[0],
1535 "raid%d-%s", conf->level, mdname(conf->mddev));
1537 sprintf(conf->cache_name[0],
1538 "raid%d-%p", conf->level, conf->mddev);
1539 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1541 conf->active_name = 0;
1542 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1543 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1547 conf->slab_cache = sc;
1548 conf->pool_size = devs;
1550 if (!grow_one_stripe(conf))
1556 * scribble_len - return the required size of the scribble region
1557 * @num - total number of disks in the array
1559 * The size must be enough to contain:
1560 * 1/ a struct page pointer for each device in the array +2
1561 * 2/ room to convert each entry in (1) to its corresponding dma
1562 * (dma_map_page()) or page (page_address()) address.
1564 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1565 * calculate over all devices (not just the data blocks), using zeros in place
1566 * of the P and Q blocks.
1568 static size_t scribble_len(int num)
1572 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1577 static int resize_stripes(struct r5conf *conf, int newsize)
1579 /* Make all the stripes able to hold 'newsize' devices.
1580 * New slots in each stripe get 'page' set to a new page.
1582 * This happens in stages:
1583 * 1/ create a new kmem_cache and allocate the required number of
1585 * 2/ gather all the old stripe_heads and transfer the pages across
1586 * to the new stripe_heads. This will have the side effect of
1587 * freezing the array as once all stripe_heads have been collected,
1588 * no IO will be possible. Old stripe heads are freed once their
1589 * pages have been transferred over, and the old kmem_cache is
1590 * freed when all stripes are done.
1591 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1592 * we simple return a failre status - no need to clean anything up.
1593 * 4/ allocate new pages for the new slots in the new stripe_heads.
1594 * If this fails, we don't bother trying the shrink the
1595 * stripe_heads down again, we just leave them as they are.
1596 * As each stripe_head is processed the new one is released into
1599 * Once step2 is started, we cannot afford to wait for a write,
1600 * so we use GFP_NOIO allocations.
1602 struct stripe_head *osh, *nsh;
1603 LIST_HEAD(newstripes);
1604 struct disk_info *ndisks;
1607 struct kmem_cache *sc;
1610 if (newsize <= conf->pool_size)
1611 return 0; /* never bother to shrink */
1613 err = md_allow_write(conf->mddev);
1618 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1619 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1624 for (i = conf->max_nr_stripes; i; i--) {
1625 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1629 nsh->raid_conf = conf;
1630 #ifdef CONFIG_MULTICORE_RAID456
1631 init_waitqueue_head(&nsh->ops.wait_for_ops);
1633 spin_lock_init(&nsh->stripe_lock);
1635 list_add(&nsh->lru, &newstripes);
1638 /* didn't get enough, give up */
1639 while (!list_empty(&newstripes)) {
1640 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1641 list_del(&nsh->lru);
1642 kmem_cache_free(sc, nsh);
1644 kmem_cache_destroy(sc);
1647 /* Step 2 - Must use GFP_NOIO now.
1648 * OK, we have enough stripes, start collecting inactive
1649 * stripes and copying them over
1651 list_for_each_entry(nsh, &newstripes, lru) {
1652 spin_lock_irq(&conf->device_lock);
1653 wait_event_lock_irq(conf->wait_for_stripe,
1654 !list_empty(&conf->inactive_list),
1656 osh = get_free_stripe(conf);
1657 spin_unlock_irq(&conf->device_lock);
1658 atomic_set(&nsh->count, 1);
1659 for(i=0; i<conf->pool_size; i++)
1660 nsh->dev[i].page = osh->dev[i].page;
1661 for( ; i<newsize; i++)
1662 nsh->dev[i].page = NULL;
1663 kmem_cache_free(conf->slab_cache, osh);
1665 kmem_cache_destroy(conf->slab_cache);
1668 * At this point, we are holding all the stripes so the array
1669 * is completely stalled, so now is a good time to resize
1670 * conf->disks and the scribble region
1672 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1674 for (i=0; i<conf->raid_disks; i++)
1675 ndisks[i] = conf->disks[i];
1677 conf->disks = ndisks;
1682 conf->scribble_len = scribble_len(newsize);
1683 for_each_present_cpu(cpu) {
1684 struct raid5_percpu *percpu;
1687 percpu = per_cpu_ptr(conf->percpu, cpu);
1688 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1691 kfree(percpu->scribble);
1692 percpu->scribble = scribble;
1700 /* Step 4, return new stripes to service */
1701 while(!list_empty(&newstripes)) {
1702 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1703 list_del_init(&nsh->lru);
1705 for (i=conf->raid_disks; i < newsize; i++)
1706 if (nsh->dev[i].page == NULL) {
1707 struct page *p = alloc_page(GFP_NOIO);
1708 nsh->dev[i].page = p;
1712 release_stripe(nsh);
1714 /* critical section pass, GFP_NOIO no longer needed */
1716 conf->slab_cache = sc;
1717 conf->active_name = 1-conf->active_name;
1718 conf->pool_size = newsize;
1722 static int drop_one_stripe(struct r5conf *conf)
1724 struct stripe_head *sh;
1726 spin_lock_irq(&conf->device_lock);
1727 sh = get_free_stripe(conf);
1728 spin_unlock_irq(&conf->device_lock);
1731 BUG_ON(atomic_read(&sh->count));
1733 kmem_cache_free(conf->slab_cache, sh);
1734 atomic_dec(&conf->active_stripes);
1738 static void shrink_stripes(struct r5conf *conf)
1740 while (drop_one_stripe(conf))
1743 if (conf->slab_cache)
1744 kmem_cache_destroy(conf->slab_cache);
1745 conf->slab_cache = NULL;
1748 static void raid5_end_read_request(struct bio * bi, int error)
1750 struct stripe_head *sh = bi->bi_private;
1751 struct r5conf *conf = sh->raid_conf;
1752 int disks = sh->disks, i;
1753 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1754 char b[BDEVNAME_SIZE];
1755 struct md_rdev *rdev = NULL;
1758 for (i=0 ; i<disks; i++)
1759 if (bi == &sh->dev[i].req)
1762 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1763 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1769 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1770 /* If replacement finished while this request was outstanding,
1771 * 'replacement' might be NULL already.
1772 * In that case it moved down to 'rdev'.
1773 * rdev is not removed until all requests are finished.
1775 rdev = conf->disks[i].replacement;
1777 rdev = conf->disks[i].rdev;
1779 if (use_new_offset(conf, sh))
1780 s = sh->sector + rdev->new_data_offset;
1782 s = sh->sector + rdev->data_offset;
1784 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1785 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1786 /* Note that this cannot happen on a
1787 * replacement device. We just fail those on
1792 "md/raid:%s: read error corrected"
1793 " (%lu sectors at %llu on %s)\n",
1794 mdname(conf->mddev), STRIPE_SECTORS,
1795 (unsigned long long)s,
1796 bdevname(rdev->bdev, b));
1797 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1798 clear_bit(R5_ReadError, &sh->dev[i].flags);
1799 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1800 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1801 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1803 if (atomic_read(&rdev->read_errors))
1804 atomic_set(&rdev->read_errors, 0);
1806 const char *bdn = bdevname(rdev->bdev, b);
1810 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1811 atomic_inc(&rdev->read_errors);
1812 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1815 "md/raid:%s: read error on replacement device "
1816 "(sector %llu on %s).\n",
1817 mdname(conf->mddev),
1818 (unsigned long long)s,
1820 else if (conf->mddev->degraded >= conf->max_degraded) {
1824 "md/raid:%s: read error not correctable "
1825 "(sector %llu on %s).\n",
1826 mdname(conf->mddev),
1827 (unsigned long long)s,
1829 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1834 "md/raid:%s: read error NOT corrected!! "
1835 "(sector %llu on %s).\n",
1836 mdname(conf->mddev),
1837 (unsigned long long)s,
1839 } else if (atomic_read(&rdev->read_errors)
1840 > conf->max_nr_stripes)
1842 "md/raid:%s: Too many read errors, failing device %s.\n",
1843 mdname(conf->mddev), bdn);
1847 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1848 set_bit(R5_ReadError, &sh->dev[i].flags);
1849 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1851 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1853 clear_bit(R5_ReadError, &sh->dev[i].flags);
1854 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1856 && test_bit(In_sync, &rdev->flags)
1857 && rdev_set_badblocks(
1858 rdev, sh->sector, STRIPE_SECTORS, 0)))
1859 md_error(conf->mddev, rdev);
1862 rdev_dec_pending(rdev, conf->mddev);
1863 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1864 set_bit(STRIPE_HANDLE, &sh->state);
1868 static void raid5_end_write_request(struct bio *bi, int error)
1870 struct stripe_head *sh = bi->bi_private;
1871 struct r5conf *conf = sh->raid_conf;
1872 int disks = sh->disks, i;
1873 struct md_rdev *uninitialized_var(rdev);
1874 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1877 int replacement = 0;
1879 for (i = 0 ; i < disks; i++) {
1880 if (bi == &sh->dev[i].req) {
1881 rdev = conf->disks[i].rdev;
1884 if (bi == &sh->dev[i].rreq) {
1885 rdev = conf->disks[i].replacement;
1889 /* rdev was removed and 'replacement'
1890 * replaced it. rdev is not removed
1891 * until all requests are finished.
1893 rdev = conf->disks[i].rdev;
1897 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1898 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1907 md_error(conf->mddev, rdev);
1908 else if (is_badblock(rdev, sh->sector,
1910 &first_bad, &bad_sectors))
1911 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1914 set_bit(WriteErrorSeen, &rdev->flags);
1915 set_bit(R5_WriteError, &sh->dev[i].flags);
1916 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1917 set_bit(MD_RECOVERY_NEEDED,
1918 &rdev->mddev->recovery);
1919 } else if (is_badblock(rdev, sh->sector,
1921 &first_bad, &bad_sectors))
1922 set_bit(R5_MadeGood, &sh->dev[i].flags);
1924 rdev_dec_pending(rdev, conf->mddev);
1926 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1927 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1928 set_bit(STRIPE_HANDLE, &sh->state);
1932 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1934 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1936 struct r5dev *dev = &sh->dev[i];
1938 bio_init(&dev->req);
1939 dev->req.bi_io_vec = &dev->vec;
1941 dev->req.bi_max_vecs++;
1942 dev->req.bi_private = sh;
1943 dev->vec.bv_page = dev->page;
1945 bio_init(&dev->rreq);
1946 dev->rreq.bi_io_vec = &dev->rvec;
1947 dev->rreq.bi_vcnt++;
1948 dev->rreq.bi_max_vecs++;
1949 dev->rreq.bi_private = sh;
1950 dev->rvec.bv_page = dev->page;
1953 dev->sector = compute_blocknr(sh, i, previous);
1956 static void error(struct mddev *mddev, struct md_rdev *rdev)
1958 char b[BDEVNAME_SIZE];
1959 struct r5conf *conf = mddev->private;
1960 unsigned long flags;
1961 pr_debug("raid456: error called\n");
1963 spin_lock_irqsave(&conf->device_lock, flags);
1964 clear_bit(In_sync, &rdev->flags);
1965 mddev->degraded = calc_degraded(conf);
1966 spin_unlock_irqrestore(&conf->device_lock, flags);
1967 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1969 set_bit(Blocked, &rdev->flags);
1970 set_bit(Faulty, &rdev->flags);
1971 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1973 "md/raid:%s: Disk failure on %s, disabling device.\n"
1974 "md/raid:%s: Operation continuing on %d devices.\n",
1976 bdevname(rdev->bdev, b),
1978 conf->raid_disks - mddev->degraded);
1982 * Input: a 'big' sector number,
1983 * Output: index of the data and parity disk, and the sector # in them.
1985 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1986 int previous, int *dd_idx,
1987 struct stripe_head *sh)
1989 sector_t stripe, stripe2;
1990 sector_t chunk_number;
1991 unsigned int chunk_offset;
1994 sector_t new_sector;
1995 int algorithm = previous ? conf->prev_algo
1997 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1998 : conf->chunk_sectors;
1999 int raid_disks = previous ? conf->previous_raid_disks
2001 int data_disks = raid_disks - conf->max_degraded;
2003 /* First compute the information on this sector */
2006 * Compute the chunk number and the sector offset inside the chunk
2008 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2009 chunk_number = r_sector;
2012 * Compute the stripe number
2014 stripe = chunk_number;
2015 *dd_idx = sector_div(stripe, data_disks);
2018 * Select the parity disk based on the user selected algorithm.
2020 pd_idx = qd_idx = -1;
2021 switch(conf->level) {
2023 pd_idx = data_disks;
2026 switch (algorithm) {
2027 case ALGORITHM_LEFT_ASYMMETRIC:
2028 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2029 if (*dd_idx >= pd_idx)
2032 case ALGORITHM_RIGHT_ASYMMETRIC:
2033 pd_idx = sector_div(stripe2, raid_disks);
2034 if (*dd_idx >= pd_idx)
2037 case ALGORITHM_LEFT_SYMMETRIC:
2038 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2039 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2041 case ALGORITHM_RIGHT_SYMMETRIC:
2042 pd_idx = sector_div(stripe2, raid_disks);
2043 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2045 case ALGORITHM_PARITY_0:
2049 case ALGORITHM_PARITY_N:
2050 pd_idx = data_disks;
2058 switch (algorithm) {
2059 case ALGORITHM_LEFT_ASYMMETRIC:
2060 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2061 qd_idx = pd_idx + 1;
2062 if (pd_idx == raid_disks-1) {
2063 (*dd_idx)++; /* Q D D D P */
2065 } else if (*dd_idx >= pd_idx)
2066 (*dd_idx) += 2; /* D D P Q D */
2068 case ALGORITHM_RIGHT_ASYMMETRIC:
2069 pd_idx = sector_div(stripe2, raid_disks);
2070 qd_idx = pd_idx + 1;
2071 if (pd_idx == raid_disks-1) {
2072 (*dd_idx)++; /* Q D D D P */
2074 } else if (*dd_idx >= pd_idx)
2075 (*dd_idx) += 2; /* D D P Q D */
2077 case ALGORITHM_LEFT_SYMMETRIC:
2078 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2079 qd_idx = (pd_idx + 1) % raid_disks;
2080 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2082 case ALGORITHM_RIGHT_SYMMETRIC:
2083 pd_idx = sector_div(stripe2, raid_disks);
2084 qd_idx = (pd_idx + 1) % raid_disks;
2085 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2088 case ALGORITHM_PARITY_0:
2093 case ALGORITHM_PARITY_N:
2094 pd_idx = data_disks;
2095 qd_idx = data_disks + 1;
2098 case ALGORITHM_ROTATING_ZERO_RESTART:
2099 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2100 * of blocks for computing Q is different.
2102 pd_idx = sector_div(stripe2, raid_disks);
2103 qd_idx = pd_idx + 1;
2104 if (pd_idx == raid_disks-1) {
2105 (*dd_idx)++; /* Q D D D P */
2107 } else if (*dd_idx >= pd_idx)
2108 (*dd_idx) += 2; /* D D P Q D */
2112 case ALGORITHM_ROTATING_N_RESTART:
2113 /* Same a left_asymmetric, by first stripe is
2114 * D D D P Q rather than
2118 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2119 qd_idx = pd_idx + 1;
2120 if (pd_idx == raid_disks-1) {
2121 (*dd_idx)++; /* Q D D D P */
2123 } else if (*dd_idx >= pd_idx)
2124 (*dd_idx) += 2; /* D D P Q D */
2128 case ALGORITHM_ROTATING_N_CONTINUE:
2129 /* Same as left_symmetric but Q is before P */
2130 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2131 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2132 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2136 case ALGORITHM_LEFT_ASYMMETRIC_6:
2137 /* RAID5 left_asymmetric, with Q on last device */
2138 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2139 if (*dd_idx >= pd_idx)
2141 qd_idx = raid_disks - 1;
2144 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2145 pd_idx = sector_div(stripe2, raid_disks-1);
2146 if (*dd_idx >= pd_idx)
2148 qd_idx = raid_disks - 1;
2151 case ALGORITHM_LEFT_SYMMETRIC_6:
2152 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2153 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2154 qd_idx = raid_disks - 1;
2157 case ALGORITHM_RIGHT_SYMMETRIC_6:
2158 pd_idx = sector_div(stripe2, raid_disks-1);
2159 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2160 qd_idx = raid_disks - 1;
2163 case ALGORITHM_PARITY_0_6:
2166 qd_idx = raid_disks - 1;
2176 sh->pd_idx = pd_idx;
2177 sh->qd_idx = qd_idx;
2178 sh->ddf_layout = ddf_layout;
2181 * Finally, compute the new sector number
2183 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2188 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2190 struct r5conf *conf = sh->raid_conf;
2191 int raid_disks = sh->disks;
2192 int data_disks = raid_disks - conf->max_degraded;
2193 sector_t new_sector = sh->sector, check;
2194 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2195 : conf->chunk_sectors;
2196 int algorithm = previous ? conf->prev_algo
2200 sector_t chunk_number;
2201 int dummy1, dd_idx = i;
2203 struct stripe_head sh2;
2206 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2207 stripe = new_sector;
2209 if (i == sh->pd_idx)
2211 switch(conf->level) {
2214 switch (algorithm) {
2215 case ALGORITHM_LEFT_ASYMMETRIC:
2216 case ALGORITHM_RIGHT_ASYMMETRIC:
2220 case ALGORITHM_LEFT_SYMMETRIC:
2221 case ALGORITHM_RIGHT_SYMMETRIC:
2224 i -= (sh->pd_idx + 1);
2226 case ALGORITHM_PARITY_0:
2229 case ALGORITHM_PARITY_N:
2236 if (i == sh->qd_idx)
2237 return 0; /* It is the Q disk */
2238 switch (algorithm) {
2239 case ALGORITHM_LEFT_ASYMMETRIC:
2240 case ALGORITHM_RIGHT_ASYMMETRIC:
2241 case ALGORITHM_ROTATING_ZERO_RESTART:
2242 case ALGORITHM_ROTATING_N_RESTART:
2243 if (sh->pd_idx == raid_disks-1)
2244 i--; /* Q D D D P */
2245 else if (i > sh->pd_idx)
2246 i -= 2; /* D D P Q D */
2248 case ALGORITHM_LEFT_SYMMETRIC:
2249 case ALGORITHM_RIGHT_SYMMETRIC:
2250 if (sh->pd_idx == raid_disks-1)
2251 i--; /* Q D D D P */
2256 i -= (sh->pd_idx + 2);
2259 case ALGORITHM_PARITY_0:
2262 case ALGORITHM_PARITY_N:
2264 case ALGORITHM_ROTATING_N_CONTINUE:
2265 /* Like left_symmetric, but P is before Q */
2266 if (sh->pd_idx == 0)
2267 i--; /* P D D D Q */
2272 i -= (sh->pd_idx + 1);
2275 case ALGORITHM_LEFT_ASYMMETRIC_6:
2276 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2280 case ALGORITHM_LEFT_SYMMETRIC_6:
2281 case ALGORITHM_RIGHT_SYMMETRIC_6:
2283 i += data_disks + 1;
2284 i -= (sh->pd_idx + 1);
2286 case ALGORITHM_PARITY_0_6:
2295 chunk_number = stripe * data_disks + i;
2296 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2298 check = raid5_compute_sector(conf, r_sector,
2299 previous, &dummy1, &sh2);
2300 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2301 || sh2.qd_idx != sh->qd_idx) {
2302 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2303 mdname(conf->mddev));
2311 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2312 int rcw, int expand)
2314 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2315 struct r5conf *conf = sh->raid_conf;
2316 int level = conf->level;
2319 /* if we are not expanding this is a proper write request, and
2320 * there will be bios with new data to be drained into the
2324 sh->reconstruct_state = reconstruct_state_drain_run;
2325 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2327 sh->reconstruct_state = reconstruct_state_run;
2329 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2331 for (i = disks; i--; ) {
2332 struct r5dev *dev = &sh->dev[i];
2335 set_bit(R5_LOCKED, &dev->flags);
2336 set_bit(R5_Wantdrain, &dev->flags);
2338 clear_bit(R5_UPTODATE, &dev->flags);
2342 if (s->locked + conf->max_degraded == disks)
2343 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2344 atomic_inc(&conf->pending_full_writes);
2347 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2348 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2350 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2351 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2352 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2353 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2355 for (i = disks; i--; ) {
2356 struct r5dev *dev = &sh->dev[i];
2361 (test_bit(R5_UPTODATE, &dev->flags) ||
2362 test_bit(R5_Wantcompute, &dev->flags))) {
2363 set_bit(R5_Wantdrain, &dev->flags);
2364 set_bit(R5_LOCKED, &dev->flags);
2365 clear_bit(R5_UPTODATE, &dev->flags);
2371 /* keep the parity disk(s) locked while asynchronous operations
2374 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2375 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2379 int qd_idx = sh->qd_idx;
2380 struct r5dev *dev = &sh->dev[qd_idx];
2382 set_bit(R5_LOCKED, &dev->flags);
2383 clear_bit(R5_UPTODATE, &dev->flags);
2387 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2388 __func__, (unsigned long long)sh->sector,
2389 s->locked, s->ops_request);
2393 * Each stripe/dev can have one or more bion attached.
2394 * toread/towrite point to the first in a chain.
2395 * The bi_next chain must be in order.
2397 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2400 struct r5conf *conf = sh->raid_conf;
2403 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2404 (unsigned long long)bi->bi_sector,
2405 (unsigned long long)sh->sector);
2408 * If several bio share a stripe. The bio bi_phys_segments acts as a
2409 * reference count to avoid race. The reference count should already be
2410 * increased before this function is called (for example, in
2411 * make_request()), so other bio sharing this stripe will not free the
2412 * stripe. If a stripe is owned by one stripe, the stripe lock will
2415 spin_lock_irq(&sh->stripe_lock);
2417 bip = &sh->dev[dd_idx].towrite;
2421 bip = &sh->dev[dd_idx].toread;
2422 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2423 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2425 bip = & (*bip)->bi_next;
2427 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2430 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2434 raid5_inc_bi_active_stripes(bi);
2437 /* check if page is covered */
2438 sector_t sector = sh->dev[dd_idx].sector;
2439 for (bi=sh->dev[dd_idx].towrite;
2440 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2441 bi && bi->bi_sector <= sector;
2442 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2443 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2444 sector = bi->bi_sector + (bi->bi_size>>9);
2446 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2447 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2450 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2451 (unsigned long long)(*bip)->bi_sector,
2452 (unsigned long long)sh->sector, dd_idx);
2453 spin_unlock_irq(&sh->stripe_lock);
2455 if (conf->mddev->bitmap && firstwrite) {
2456 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2458 sh->bm_seq = conf->seq_flush+1;
2459 set_bit(STRIPE_BIT_DELAY, &sh->state);
2464 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2465 spin_unlock_irq(&sh->stripe_lock);
2469 static void end_reshape(struct r5conf *conf);
2471 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2472 struct stripe_head *sh)
2474 int sectors_per_chunk =
2475 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2477 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2478 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2480 raid5_compute_sector(conf,
2481 stripe * (disks - conf->max_degraded)
2482 *sectors_per_chunk + chunk_offset,
2488 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2489 struct stripe_head_state *s, int disks,
2490 struct bio **return_bi)
2493 for (i = disks; i--; ) {
2497 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2498 struct md_rdev *rdev;
2500 rdev = rcu_dereference(conf->disks[i].rdev);
2501 if (rdev && test_bit(In_sync, &rdev->flags))
2502 atomic_inc(&rdev->nr_pending);
2507 if (!rdev_set_badblocks(
2511 md_error(conf->mddev, rdev);
2512 rdev_dec_pending(rdev, conf->mddev);
2515 spin_lock_irq(&sh->stripe_lock);
2516 /* fail all writes first */
2517 bi = sh->dev[i].towrite;
2518 sh->dev[i].towrite = NULL;
2519 spin_unlock_irq(&sh->stripe_lock);
2523 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2524 wake_up(&conf->wait_for_overlap);
2526 while (bi && bi->bi_sector <
2527 sh->dev[i].sector + STRIPE_SECTORS) {
2528 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2529 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2530 if (!raid5_dec_bi_active_stripes(bi)) {
2531 md_write_end(conf->mddev);
2532 bi->bi_next = *return_bi;
2538 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2539 STRIPE_SECTORS, 0, 0);
2541 /* and fail all 'written' */
2542 bi = sh->dev[i].written;
2543 sh->dev[i].written = NULL;
2544 if (bi) bitmap_end = 1;
2545 while (bi && bi->bi_sector <
2546 sh->dev[i].sector + STRIPE_SECTORS) {
2547 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2548 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2549 if (!raid5_dec_bi_active_stripes(bi)) {
2550 md_write_end(conf->mddev);
2551 bi->bi_next = *return_bi;
2557 /* fail any reads if this device is non-operational and
2558 * the data has not reached the cache yet.
2560 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2561 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2562 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2563 spin_lock_irq(&sh->stripe_lock);
2564 bi = sh->dev[i].toread;
2565 sh->dev[i].toread = NULL;
2566 spin_unlock_irq(&sh->stripe_lock);
2567 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2568 wake_up(&conf->wait_for_overlap);
2569 while (bi && bi->bi_sector <
2570 sh->dev[i].sector + STRIPE_SECTORS) {
2571 struct bio *nextbi =
2572 r5_next_bio(bi, sh->dev[i].sector);
2573 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2574 if (!raid5_dec_bi_active_stripes(bi)) {
2575 bi->bi_next = *return_bi;
2582 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2583 STRIPE_SECTORS, 0, 0);
2584 /* If we were in the middle of a write the parity block might
2585 * still be locked - so just clear all R5_LOCKED flags
2587 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2590 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2591 if (atomic_dec_and_test(&conf->pending_full_writes))
2592 md_wakeup_thread(conf->mddev->thread);
2596 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2597 struct stripe_head_state *s)
2602 clear_bit(STRIPE_SYNCING, &sh->state);
2605 /* There is nothing more to do for sync/check/repair.
2606 * Don't even need to abort as that is handled elsewhere
2607 * if needed, and not always wanted e.g. if there is a known
2609 * For recover/replace we need to record a bad block on all
2610 * non-sync devices, or abort the recovery
2612 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2613 /* During recovery devices cannot be removed, so
2614 * locking and refcounting of rdevs is not needed
2616 for (i = 0; i < conf->raid_disks; i++) {
2617 struct md_rdev *rdev = conf->disks[i].rdev;
2619 && !test_bit(Faulty, &rdev->flags)
2620 && !test_bit(In_sync, &rdev->flags)
2621 && !rdev_set_badblocks(rdev, sh->sector,
2624 rdev = conf->disks[i].replacement;
2626 && !test_bit(Faulty, &rdev->flags)
2627 && !test_bit(In_sync, &rdev->flags)
2628 && !rdev_set_badblocks(rdev, sh->sector,
2633 conf->recovery_disabled =
2634 conf->mddev->recovery_disabled;
2636 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2639 static int want_replace(struct stripe_head *sh, int disk_idx)
2641 struct md_rdev *rdev;
2643 /* Doing recovery so rcu locking not required */
2644 rdev = sh->raid_conf->disks[disk_idx].replacement;
2646 && !test_bit(Faulty, &rdev->flags)
2647 && !test_bit(In_sync, &rdev->flags)
2648 && (rdev->recovery_offset <= sh->sector
2649 || rdev->mddev->recovery_cp <= sh->sector))
2655 /* fetch_block - checks the given member device to see if its data needs
2656 * to be read or computed to satisfy a request.
2658 * Returns 1 when no more member devices need to be checked, otherwise returns
2659 * 0 to tell the loop in handle_stripe_fill to continue
2661 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2662 int disk_idx, int disks)
2664 struct r5dev *dev = &sh->dev[disk_idx];
2665 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2666 &sh->dev[s->failed_num[1]] };
2668 /* is the data in this block needed, and can we get it? */
2669 if (!test_bit(R5_LOCKED, &dev->flags) &&
2670 !test_bit(R5_UPTODATE, &dev->flags) &&
2672 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2673 s->syncing || s->expanding ||
2674 (s->replacing && want_replace(sh, disk_idx)) ||
2675 (s->failed >= 1 && fdev[0]->toread) ||
2676 (s->failed >= 2 && fdev[1]->toread) ||
2677 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2678 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2679 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2680 /* we would like to get this block, possibly by computing it,
2681 * otherwise read it if the backing disk is insync
2683 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2684 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2685 if ((s->uptodate == disks - 1) &&
2686 (s->failed && (disk_idx == s->failed_num[0] ||
2687 disk_idx == s->failed_num[1]))) {
2688 /* have disk failed, and we're requested to fetch it;
2691 pr_debug("Computing stripe %llu block %d\n",
2692 (unsigned long long)sh->sector, disk_idx);
2693 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2694 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2695 set_bit(R5_Wantcompute, &dev->flags);
2696 sh->ops.target = disk_idx;
2697 sh->ops.target2 = -1; /* no 2nd target */
2699 /* Careful: from this point on 'uptodate' is in the eye
2700 * of raid_run_ops which services 'compute' operations
2701 * before writes. R5_Wantcompute flags a block that will
2702 * be R5_UPTODATE by the time it is needed for a
2703 * subsequent operation.
2707 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2708 /* Computing 2-failure is *very* expensive; only
2709 * do it if failed >= 2
2712 for (other = disks; other--; ) {
2713 if (other == disk_idx)
2715 if (!test_bit(R5_UPTODATE,
2716 &sh->dev[other].flags))
2720 pr_debug("Computing stripe %llu blocks %d,%d\n",
2721 (unsigned long long)sh->sector,
2723 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2724 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2725 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2726 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2727 sh->ops.target = disk_idx;
2728 sh->ops.target2 = other;
2732 } else if (test_bit(R5_Insync, &dev->flags)) {
2733 set_bit(R5_LOCKED, &dev->flags);
2734 set_bit(R5_Wantread, &dev->flags);
2736 pr_debug("Reading block %d (sync=%d)\n",
2737 disk_idx, s->syncing);
2745 * handle_stripe_fill - read or compute data to satisfy pending requests.
2747 static void handle_stripe_fill(struct stripe_head *sh,
2748 struct stripe_head_state *s,
2753 /* look for blocks to read/compute, skip this if a compute
2754 * is already in flight, or if the stripe contents are in the
2755 * midst of changing due to a write
2757 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2758 !sh->reconstruct_state)
2759 for (i = disks; i--; )
2760 if (fetch_block(sh, s, i, disks))
2762 set_bit(STRIPE_HANDLE, &sh->state);
2766 /* handle_stripe_clean_event
2767 * any written block on an uptodate or failed drive can be returned.
2768 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2769 * never LOCKED, so we don't need to test 'failed' directly.
2771 static void handle_stripe_clean_event(struct r5conf *conf,
2772 struct stripe_head *sh, int disks, struct bio **return_bi)
2777 for (i = disks; i--; )
2778 if (sh->dev[i].written) {
2780 if (!test_bit(R5_LOCKED, &dev->flags) &&
2781 (test_bit(R5_UPTODATE, &dev->flags) ||
2782 test_bit(R5_Discard, &dev->flags))) {
2783 /* We can return any write requests */
2784 struct bio *wbi, *wbi2;
2785 pr_debug("Return write for disc %d\n", i);
2786 if (test_and_clear_bit(R5_Discard, &dev->flags))
2787 clear_bit(R5_UPTODATE, &dev->flags);
2789 dev->written = NULL;
2790 while (wbi && wbi->bi_sector <
2791 dev->sector + STRIPE_SECTORS) {
2792 wbi2 = r5_next_bio(wbi, dev->sector);
2793 if (!raid5_dec_bi_active_stripes(wbi)) {
2794 md_write_end(conf->mddev);
2795 wbi->bi_next = *return_bi;
2800 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2802 !test_bit(STRIPE_DEGRADED, &sh->state),
2805 } else if (test_bit(R5_Discard, &sh->dev[i].flags))
2806 clear_bit(R5_Discard, &sh->dev[i].flags);
2808 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2809 if (atomic_dec_and_test(&conf->pending_full_writes))
2810 md_wakeup_thread(conf->mddev->thread);
2813 static void handle_stripe_dirtying(struct r5conf *conf,
2814 struct stripe_head *sh,
2815 struct stripe_head_state *s,
2818 int rmw = 0, rcw = 0, i;
2819 sector_t recovery_cp = conf->mddev->recovery_cp;
2821 /* RAID6 requires 'rcw' in current implementation.
2822 * Otherwise, check whether resync is now happening or should start.
2823 * If yes, then the array is dirty (after unclean shutdown or
2824 * initial creation), so parity in some stripes might be inconsistent.
2825 * In this case, we need to always do reconstruct-write, to ensure
2826 * that in case of drive failure or read-error correction, we
2827 * generate correct data from the parity.
2829 if (conf->max_degraded == 2 ||
2830 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2831 /* Calculate the real rcw later - for now make it
2832 * look like rcw is cheaper
2835 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2836 conf->max_degraded, (unsigned long long)recovery_cp,
2837 (unsigned long long)sh->sector);
2838 } else for (i = disks; i--; ) {
2839 /* would I have to read this buffer for read_modify_write */
2840 struct r5dev *dev = &sh->dev[i];
2841 if ((dev->towrite || i == sh->pd_idx) &&
2842 !test_bit(R5_LOCKED, &dev->flags) &&
2843 !(test_bit(R5_UPTODATE, &dev->flags) ||
2844 test_bit(R5_Wantcompute, &dev->flags))) {
2845 if (test_bit(R5_Insync, &dev->flags))
2848 rmw += 2*disks; /* cannot read it */
2850 /* Would I have to read this buffer for reconstruct_write */
2851 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2852 !test_bit(R5_LOCKED, &dev->flags) &&
2853 !(test_bit(R5_UPTODATE, &dev->flags) ||
2854 test_bit(R5_Wantcompute, &dev->flags))) {
2855 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2860 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2861 (unsigned long long)sh->sector, rmw, rcw);
2862 set_bit(STRIPE_HANDLE, &sh->state);
2863 if (rmw < rcw && rmw > 0) {
2864 /* prefer read-modify-write, but need to get some data */
2865 blk_add_trace_msg(conf->mddev->queue, "raid5 rmw %llu %d",
2866 (unsigned long long)sh->sector, rmw);
2867 for (i = disks; i--; ) {
2868 struct r5dev *dev = &sh->dev[i];
2869 if ((dev->towrite || i == sh->pd_idx) &&
2870 !test_bit(R5_LOCKED, &dev->flags) &&
2871 !(test_bit(R5_UPTODATE, &dev->flags) ||
2872 test_bit(R5_Wantcompute, &dev->flags)) &&
2873 test_bit(R5_Insync, &dev->flags)) {
2875 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2876 pr_debug("Read_old block "
2877 "%d for r-m-w\n", i);
2878 set_bit(R5_LOCKED, &dev->flags);
2879 set_bit(R5_Wantread, &dev->flags);
2882 set_bit(STRIPE_DELAYED, &sh->state);
2883 set_bit(STRIPE_HANDLE, &sh->state);
2888 if (rcw <= rmw && rcw > 0) {
2889 /* want reconstruct write, but need to get some data */
2892 for (i = disks; i--; ) {
2893 struct r5dev *dev = &sh->dev[i];
2894 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2895 i != sh->pd_idx && i != sh->qd_idx &&
2896 !test_bit(R5_LOCKED, &dev->flags) &&
2897 !(test_bit(R5_UPTODATE, &dev->flags) ||
2898 test_bit(R5_Wantcompute, &dev->flags))) {
2900 if (!test_bit(R5_Insync, &dev->flags))
2901 continue; /* it's a failed drive */
2903 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2904 pr_debug("Read_old block "
2905 "%d for Reconstruct\n", i);
2906 set_bit(R5_LOCKED, &dev->flags);
2907 set_bit(R5_Wantread, &dev->flags);
2911 set_bit(STRIPE_DELAYED, &sh->state);
2912 set_bit(STRIPE_HANDLE, &sh->state);
2917 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
2918 (unsigned long long)sh->sector,
2919 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
2921 /* now if nothing is locked, and if we have enough data,
2922 * we can start a write request
2924 /* since handle_stripe can be called at any time we need to handle the
2925 * case where a compute block operation has been submitted and then a
2926 * subsequent call wants to start a write request. raid_run_ops only
2927 * handles the case where compute block and reconstruct are requested
2928 * simultaneously. If this is not the case then new writes need to be
2929 * held off until the compute completes.
2931 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2932 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2933 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2934 schedule_reconstruction(sh, s, rcw == 0, 0);
2937 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2938 struct stripe_head_state *s, int disks)
2940 struct r5dev *dev = NULL;
2942 set_bit(STRIPE_HANDLE, &sh->state);
2944 switch (sh->check_state) {
2945 case check_state_idle:
2946 /* start a new check operation if there are no failures */
2947 if (s->failed == 0) {
2948 BUG_ON(s->uptodate != disks);
2949 sh->check_state = check_state_run;
2950 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2951 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2955 dev = &sh->dev[s->failed_num[0]];
2957 case check_state_compute_result:
2958 sh->check_state = check_state_idle;
2960 dev = &sh->dev[sh->pd_idx];
2962 /* check that a write has not made the stripe insync */
2963 if (test_bit(STRIPE_INSYNC, &sh->state))
2966 /* either failed parity check, or recovery is happening */
2967 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2968 BUG_ON(s->uptodate != disks);
2970 set_bit(R5_LOCKED, &dev->flags);
2972 set_bit(R5_Wantwrite, &dev->flags);
2974 clear_bit(STRIPE_DEGRADED, &sh->state);
2975 set_bit(STRIPE_INSYNC, &sh->state);
2977 case check_state_run:
2978 break; /* we will be called again upon completion */
2979 case check_state_check_result:
2980 sh->check_state = check_state_idle;
2982 /* if a failure occurred during the check operation, leave
2983 * STRIPE_INSYNC not set and let the stripe be handled again
2988 /* handle a successful check operation, if parity is correct
2989 * we are done. Otherwise update the mismatch count and repair
2990 * parity if !MD_RECOVERY_CHECK
2992 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2993 /* parity is correct (on disc,
2994 * not in buffer any more)
2996 set_bit(STRIPE_INSYNC, &sh->state);
2998 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
2999 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3000 /* don't try to repair!! */
3001 set_bit(STRIPE_INSYNC, &sh->state);
3003 sh->check_state = check_state_compute_run;
3004 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3005 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3006 set_bit(R5_Wantcompute,
3007 &sh->dev[sh->pd_idx].flags);
3008 sh->ops.target = sh->pd_idx;
3009 sh->ops.target2 = -1;
3014 case check_state_compute_run:
3017 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3018 __func__, sh->check_state,
3019 (unsigned long long) sh->sector);
3025 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3026 struct stripe_head_state *s,
3029 int pd_idx = sh->pd_idx;
3030 int qd_idx = sh->qd_idx;
3033 set_bit(STRIPE_HANDLE, &sh->state);
3035 BUG_ON(s->failed > 2);
3037 /* Want to check and possibly repair P and Q.
3038 * However there could be one 'failed' device, in which
3039 * case we can only check one of them, possibly using the
3040 * other to generate missing data
3043 switch (sh->check_state) {
3044 case check_state_idle:
3045 /* start a new check operation if there are < 2 failures */
3046 if (s->failed == s->q_failed) {
3047 /* The only possible failed device holds Q, so it
3048 * makes sense to check P (If anything else were failed,
3049 * we would have used P to recreate it).
3051 sh->check_state = check_state_run;
3053 if (!s->q_failed && s->failed < 2) {
3054 /* Q is not failed, and we didn't use it to generate
3055 * anything, so it makes sense to check it
3057 if (sh->check_state == check_state_run)
3058 sh->check_state = check_state_run_pq;
3060 sh->check_state = check_state_run_q;
3063 /* discard potentially stale zero_sum_result */
3064 sh->ops.zero_sum_result = 0;
3066 if (sh->check_state == check_state_run) {
3067 /* async_xor_zero_sum destroys the contents of P */
3068 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3071 if (sh->check_state >= check_state_run &&
3072 sh->check_state <= check_state_run_pq) {
3073 /* async_syndrome_zero_sum preserves P and Q, so
3074 * no need to mark them !uptodate here
3076 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3080 /* we have 2-disk failure */
3081 BUG_ON(s->failed != 2);
3083 case check_state_compute_result:
3084 sh->check_state = check_state_idle;
3086 /* check that a write has not made the stripe insync */
3087 if (test_bit(STRIPE_INSYNC, &sh->state))
3090 /* now write out any block on a failed drive,
3091 * or P or Q if they were recomputed
3093 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3094 if (s->failed == 2) {
3095 dev = &sh->dev[s->failed_num[1]];
3097 set_bit(R5_LOCKED, &dev->flags);
3098 set_bit(R5_Wantwrite, &dev->flags);
3100 if (s->failed >= 1) {
3101 dev = &sh->dev[s->failed_num[0]];
3103 set_bit(R5_LOCKED, &dev->flags);
3104 set_bit(R5_Wantwrite, &dev->flags);
3106 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3107 dev = &sh->dev[pd_idx];
3109 set_bit(R5_LOCKED, &dev->flags);
3110 set_bit(R5_Wantwrite, &dev->flags);
3112 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3113 dev = &sh->dev[qd_idx];
3115 set_bit(R5_LOCKED, &dev->flags);
3116 set_bit(R5_Wantwrite, &dev->flags);
3118 clear_bit(STRIPE_DEGRADED, &sh->state);
3120 set_bit(STRIPE_INSYNC, &sh->state);
3122 case check_state_run:
3123 case check_state_run_q:
3124 case check_state_run_pq:
3125 break; /* we will be called again upon completion */
3126 case check_state_check_result:
3127 sh->check_state = check_state_idle;
3129 /* handle a successful check operation, if parity is correct
3130 * we are done. Otherwise update the mismatch count and repair
3131 * parity if !MD_RECOVERY_CHECK
3133 if (sh->ops.zero_sum_result == 0) {
3134 /* both parities are correct */
3136 set_bit(STRIPE_INSYNC, &sh->state);
3138 /* in contrast to the raid5 case we can validate
3139 * parity, but still have a failure to write
3142 sh->check_state = check_state_compute_result;
3143 /* Returning at this point means that we may go
3144 * off and bring p and/or q uptodate again so
3145 * we make sure to check zero_sum_result again
3146 * to verify if p or q need writeback
3150 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3151 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3152 /* don't try to repair!! */
3153 set_bit(STRIPE_INSYNC, &sh->state);
3155 int *target = &sh->ops.target;
3157 sh->ops.target = -1;
3158 sh->ops.target2 = -1;
3159 sh->check_state = check_state_compute_run;
3160 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3161 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3162 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3163 set_bit(R5_Wantcompute,
3164 &sh->dev[pd_idx].flags);
3166 target = &sh->ops.target2;
3169 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3170 set_bit(R5_Wantcompute,
3171 &sh->dev[qd_idx].flags);
3178 case check_state_compute_run:
3181 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3182 __func__, sh->check_state,
3183 (unsigned long long) sh->sector);
3188 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3192 /* We have read all the blocks in this stripe and now we need to
3193 * copy some of them into a target stripe for expand.
3195 struct dma_async_tx_descriptor *tx = NULL;
3196 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3197 for (i = 0; i < sh->disks; i++)
3198 if (i != sh->pd_idx && i != sh->qd_idx) {
3200 struct stripe_head *sh2;
3201 struct async_submit_ctl submit;
3203 sector_t bn = compute_blocknr(sh, i, 1);
3204 sector_t s = raid5_compute_sector(conf, bn, 0,
3206 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3208 /* so far only the early blocks of this stripe
3209 * have been requested. When later blocks
3210 * get requested, we will try again
3213 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3214 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3215 /* must have already done this block */
3216 release_stripe(sh2);
3220 /* place all the copies on one channel */
3221 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3222 tx = async_memcpy(sh2->dev[dd_idx].page,
3223 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3226 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3227 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3228 for (j = 0; j < conf->raid_disks; j++)
3229 if (j != sh2->pd_idx &&
3231 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3233 if (j == conf->raid_disks) {
3234 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3235 set_bit(STRIPE_HANDLE, &sh2->state);
3237 release_stripe(sh2);
3240 /* done submitting copies, wait for them to complete */
3241 async_tx_quiesce(&tx);
3245 * handle_stripe - do things to a stripe.
3247 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3248 * state of various bits to see what needs to be done.
3250 * return some read requests which now have data
3251 * return some write requests which are safely on storage
3252 * schedule a read on some buffers
3253 * schedule a write of some buffers
3254 * return confirmation of parity correctness
3258 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3260 struct r5conf *conf = sh->raid_conf;
3261 int disks = sh->disks;
3264 int do_recovery = 0;
3266 memset(s, 0, sizeof(*s));
3268 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3269 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3270 s->failed_num[0] = -1;
3271 s->failed_num[1] = -1;
3273 /* Now to look around and see what can be done */
3275 for (i=disks; i--; ) {
3276 struct md_rdev *rdev;
3283 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3285 dev->toread, dev->towrite, dev->written);
3286 /* maybe we can reply to a read
3288 * new wantfill requests are only permitted while
3289 * ops_complete_biofill is guaranteed to be inactive
3291 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3292 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3293 set_bit(R5_Wantfill, &dev->flags);
3295 /* now count some things */
3296 if (test_bit(R5_LOCKED, &dev->flags))
3298 if (test_bit(R5_UPTODATE, &dev->flags))
3300 if (test_bit(R5_Wantcompute, &dev->flags)) {
3302 BUG_ON(s->compute > 2);
3305 if (test_bit(R5_Wantfill, &dev->flags))
3307 else if (dev->toread)
3311 if (!test_bit(R5_OVERWRITE, &dev->flags))
3316 /* Prefer to use the replacement for reads, but only
3317 * if it is recovered enough and has no bad blocks.
3319 rdev = rcu_dereference(conf->disks[i].replacement);
3320 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3321 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3322 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3323 &first_bad, &bad_sectors))
3324 set_bit(R5_ReadRepl, &dev->flags);
3327 set_bit(R5_NeedReplace, &dev->flags);
3328 rdev = rcu_dereference(conf->disks[i].rdev);
3329 clear_bit(R5_ReadRepl, &dev->flags);
3331 if (rdev && test_bit(Faulty, &rdev->flags))
3334 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3335 &first_bad, &bad_sectors);
3336 if (s->blocked_rdev == NULL
3337 && (test_bit(Blocked, &rdev->flags)
3340 set_bit(BlockedBadBlocks,
3342 s->blocked_rdev = rdev;
3343 atomic_inc(&rdev->nr_pending);
3346 clear_bit(R5_Insync, &dev->flags);
3350 /* also not in-sync */
3351 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3352 test_bit(R5_UPTODATE, &dev->flags)) {
3353 /* treat as in-sync, but with a read error
3354 * which we can now try to correct
3356 set_bit(R5_Insync, &dev->flags);
3357 set_bit(R5_ReadError, &dev->flags);
3359 } else if (test_bit(In_sync, &rdev->flags))
3360 set_bit(R5_Insync, &dev->flags);
3361 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3362 /* in sync if before recovery_offset */
3363 set_bit(R5_Insync, &dev->flags);
3364 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3365 test_bit(R5_Expanded, &dev->flags))
3366 /* If we've reshaped into here, we assume it is Insync.
3367 * We will shortly update recovery_offset to make
3370 set_bit(R5_Insync, &dev->flags);
3372 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3373 /* This flag does not apply to '.replacement'
3374 * only to .rdev, so make sure to check that*/
3375 struct md_rdev *rdev2 = rcu_dereference(
3376 conf->disks[i].rdev);
3378 clear_bit(R5_Insync, &dev->flags);
3379 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3380 s->handle_bad_blocks = 1;
3381 atomic_inc(&rdev2->nr_pending);
3383 clear_bit(R5_WriteError, &dev->flags);
3385 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3386 /* This flag does not apply to '.replacement'
3387 * only to .rdev, so make sure to check that*/
3388 struct md_rdev *rdev2 = rcu_dereference(
3389 conf->disks[i].rdev);
3390 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3391 s->handle_bad_blocks = 1;
3392 atomic_inc(&rdev2->nr_pending);
3394 clear_bit(R5_MadeGood, &dev->flags);
3396 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3397 struct md_rdev *rdev2 = rcu_dereference(
3398 conf->disks[i].replacement);
3399 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3400 s->handle_bad_blocks = 1;
3401 atomic_inc(&rdev2->nr_pending);
3403 clear_bit(R5_MadeGoodRepl, &dev->flags);
3405 if (!test_bit(R5_Insync, &dev->flags)) {
3406 /* The ReadError flag will just be confusing now */
3407 clear_bit(R5_ReadError, &dev->flags);
3408 clear_bit(R5_ReWrite, &dev->flags);
3410 if (test_bit(R5_ReadError, &dev->flags))
3411 clear_bit(R5_Insync, &dev->flags);
3412 if (!test_bit(R5_Insync, &dev->flags)) {
3414 s->failed_num[s->failed] = i;
3416 if (rdev && !test_bit(Faulty, &rdev->flags))
3420 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3421 /* If there is a failed device being replaced,
3422 * we must be recovering.
3423 * else if we are after recovery_cp, we must be syncing
3424 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3425 * else we can only be replacing
3426 * sync and recovery both need to read all devices, and so
3427 * use the same flag.
3430 sh->sector >= conf->mddev->recovery_cp ||
3431 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3439 static void handle_stripe(struct stripe_head *sh)
3441 struct stripe_head_state s;
3442 struct r5conf *conf = sh->raid_conf;
3445 int disks = sh->disks;
3446 struct r5dev *pdev, *qdev;
3448 clear_bit(STRIPE_HANDLE, &sh->state);
3449 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3450 /* already being handled, ensure it gets handled
3451 * again when current action finishes */
3452 set_bit(STRIPE_HANDLE, &sh->state);
3456 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3457 set_bit(STRIPE_SYNCING, &sh->state);
3458 clear_bit(STRIPE_INSYNC, &sh->state);
3460 clear_bit(STRIPE_DELAYED, &sh->state);
3462 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3463 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3464 (unsigned long long)sh->sector, sh->state,
3465 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3466 sh->check_state, sh->reconstruct_state);
3468 analyse_stripe(sh, &s);
3470 if (s.handle_bad_blocks) {
3471 set_bit(STRIPE_HANDLE, &sh->state);
3475 if (unlikely(s.blocked_rdev)) {
3476 if (s.syncing || s.expanding || s.expanded ||
3477 s.replacing || s.to_write || s.written) {
3478 set_bit(STRIPE_HANDLE, &sh->state);
3481 /* There is nothing for the blocked_rdev to block */
3482 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3483 s.blocked_rdev = NULL;
3486 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3487 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3488 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3491 pr_debug("locked=%d uptodate=%d to_read=%d"
3492 " to_write=%d failed=%d failed_num=%d,%d\n",
3493 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3494 s.failed_num[0], s.failed_num[1]);
3495 /* check if the array has lost more than max_degraded devices and,
3496 * if so, some requests might need to be failed.
3498 if (s.failed > conf->max_degraded) {
3499 sh->check_state = 0;
3500 sh->reconstruct_state = 0;
3501 if (s.to_read+s.to_write+s.written)
3502 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3503 if (s.syncing + s.replacing)
3504 handle_failed_sync(conf, sh, &s);
3507 /* Now we check to see if any write operations have recently
3511 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3513 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3514 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3515 sh->reconstruct_state = reconstruct_state_idle;
3517 /* All the 'written' buffers and the parity block are ready to
3518 * be written back to disk
3520 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3521 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3522 BUG_ON(sh->qd_idx >= 0 &&
3523 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3524 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3525 for (i = disks; i--; ) {
3526 struct r5dev *dev = &sh->dev[i];
3527 if (test_bit(R5_LOCKED, &dev->flags) &&
3528 (i == sh->pd_idx || i == sh->qd_idx ||
3530 pr_debug("Writing block %d\n", i);
3531 set_bit(R5_Wantwrite, &dev->flags);
3534 if (!test_bit(R5_Insync, &dev->flags) ||
3535 ((i == sh->pd_idx || i == sh->qd_idx) &&
3537 set_bit(STRIPE_INSYNC, &sh->state);
3540 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3541 s.dec_preread_active = 1;
3545 * might be able to return some write requests if the parity blocks
3546 * are safe, or on a failed drive
3548 pdev = &sh->dev[sh->pd_idx];
3549 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3550 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3551 qdev = &sh->dev[sh->qd_idx];
3552 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3553 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3557 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3558 && !test_bit(R5_LOCKED, &pdev->flags)
3559 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3560 test_bit(R5_Discard, &pdev->flags))))) &&
3561 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3562 && !test_bit(R5_LOCKED, &qdev->flags)
3563 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3564 test_bit(R5_Discard, &qdev->flags))))))
3565 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3567 /* Now we might consider reading some blocks, either to check/generate
3568 * parity, or to satisfy requests
3569 * or to load a block that is being partially written.
3571 if (s.to_read || s.non_overwrite
3572 || (conf->level == 6 && s.to_write && s.failed)
3573 || (s.syncing && (s.uptodate + s.compute < disks))
3576 handle_stripe_fill(sh, &s, disks);
3578 /* Now to consider new write requests and what else, if anything
3579 * should be read. We do not handle new writes when:
3580 * 1/ A 'write' operation (copy+xor) is already in flight.
3581 * 2/ A 'check' operation is in flight, as it may clobber the parity
3584 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3585 handle_stripe_dirtying(conf, sh, &s, disks);
3587 /* maybe we need to check and possibly fix the parity for this stripe
3588 * Any reads will already have been scheduled, so we just see if enough
3589 * data is available. The parity check is held off while parity
3590 * dependent operations are in flight.
3592 if (sh->check_state ||
3593 (s.syncing && s.locked == 0 &&
3594 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3595 !test_bit(STRIPE_INSYNC, &sh->state))) {
3596 if (conf->level == 6)
3597 handle_parity_checks6(conf, sh, &s, disks);
3599 handle_parity_checks5(conf, sh, &s, disks);
3602 if (s.replacing && s.locked == 0
3603 && !test_bit(STRIPE_INSYNC, &sh->state)) {
3604 /* Write out to replacement devices where possible */
3605 for (i = 0; i < conf->raid_disks; i++)
3606 if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3607 test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3608 set_bit(R5_WantReplace, &sh->dev[i].flags);
3609 set_bit(R5_LOCKED, &sh->dev[i].flags);
3612 set_bit(STRIPE_INSYNC, &sh->state);
3614 if ((s.syncing || s.replacing) && s.locked == 0 &&
3615 test_bit(STRIPE_INSYNC, &sh->state)) {
3616 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3617 clear_bit(STRIPE_SYNCING, &sh->state);
3620 /* If the failed drives are just a ReadError, then we might need
3621 * to progress the repair/check process
3623 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3624 for (i = 0; i < s.failed; i++) {
3625 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3626 if (test_bit(R5_ReadError, &dev->flags)
3627 && !test_bit(R5_LOCKED, &dev->flags)
3628 && test_bit(R5_UPTODATE, &dev->flags)
3630 if (!test_bit(R5_ReWrite, &dev->flags)) {
3631 set_bit(R5_Wantwrite, &dev->flags);
3632 set_bit(R5_ReWrite, &dev->flags);
3633 set_bit(R5_LOCKED, &dev->flags);
3636 /* let's read it back */
3637 set_bit(R5_Wantread, &dev->flags);
3638 set_bit(R5_LOCKED, &dev->flags);
3645 /* Finish reconstruct operations initiated by the expansion process */
3646 if (sh->reconstruct_state == reconstruct_state_result) {
3647 struct stripe_head *sh_src
3648 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3649 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3650 /* sh cannot be written until sh_src has been read.
3651 * so arrange for sh to be delayed a little
3653 set_bit(STRIPE_DELAYED, &sh->state);
3654 set_bit(STRIPE_HANDLE, &sh->state);
3655 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3657 atomic_inc(&conf->preread_active_stripes);
3658 release_stripe(sh_src);
3662 release_stripe(sh_src);
3664 sh->reconstruct_state = reconstruct_state_idle;
3665 clear_bit(STRIPE_EXPANDING, &sh->state);
3666 for (i = conf->raid_disks; i--; ) {
3667 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3668 set_bit(R5_LOCKED, &sh->dev[i].flags);
3673 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3674 !sh->reconstruct_state) {
3675 /* Need to write out all blocks after computing parity */
3676 sh->disks = conf->raid_disks;
3677 stripe_set_idx(sh->sector, conf, 0, sh);
3678 schedule_reconstruction(sh, &s, 1, 1);
3679 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3680 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3681 atomic_dec(&conf->reshape_stripes);
3682 wake_up(&conf->wait_for_overlap);
3683 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3686 if (s.expanding && s.locked == 0 &&
3687 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3688 handle_stripe_expansion(conf, sh);
3691 /* wait for this device to become unblocked */
3692 if (unlikely(s.blocked_rdev)) {
3693 if (conf->mddev->external)
3694 md_wait_for_blocked_rdev(s.blocked_rdev,
3697 /* Internal metadata will immediately
3698 * be written by raid5d, so we don't
3699 * need to wait here.
3701 rdev_dec_pending(s.blocked_rdev,
3705 if (s.handle_bad_blocks)
3706 for (i = disks; i--; ) {
3707 struct md_rdev *rdev;
3708 struct r5dev *dev = &sh->dev[i];
3709 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3710 /* We own a safe reference to the rdev */
3711 rdev = conf->disks[i].rdev;
3712 if (!rdev_set_badblocks(rdev, sh->sector,
3714 md_error(conf->mddev, rdev);
3715 rdev_dec_pending(rdev, conf->mddev);
3717 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3718 rdev = conf->disks[i].rdev;
3719 rdev_clear_badblocks(rdev, sh->sector,
3721 rdev_dec_pending(rdev, conf->mddev);
3723 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3724 rdev = conf->disks[i].replacement;
3726 /* rdev have been moved down */
3727 rdev = conf->disks[i].rdev;
3728 rdev_clear_badblocks(rdev, sh->sector,
3730 rdev_dec_pending(rdev, conf->mddev);
3735 raid_run_ops(sh, s.ops_request);
3739 if (s.dec_preread_active) {
3740 /* We delay this until after ops_run_io so that if make_request
3741 * is waiting on a flush, it won't continue until the writes
3742 * have actually been submitted.
3744 atomic_dec(&conf->preread_active_stripes);
3745 if (atomic_read(&conf->preread_active_stripes) <
3747 md_wakeup_thread(conf->mddev->thread);
3750 return_io(s.return_bi);
3752 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3755 static void raid5_activate_delayed(struct r5conf *conf)
3757 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3758 while (!list_empty(&conf->delayed_list)) {
3759 struct list_head *l = conf->delayed_list.next;
3760 struct stripe_head *sh;
3761 sh = list_entry(l, struct stripe_head, lru);
3763 clear_bit(STRIPE_DELAYED, &sh->state);
3764 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3765 atomic_inc(&conf->preread_active_stripes);
3766 list_add_tail(&sh->lru, &conf->hold_list);
3771 static void activate_bit_delay(struct r5conf *conf)
3773 /* device_lock is held */
3774 struct list_head head;
3775 list_add(&head, &conf->bitmap_list);
3776 list_del_init(&conf->bitmap_list);
3777 while (!list_empty(&head)) {
3778 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3779 list_del_init(&sh->lru);
3780 atomic_inc(&sh->count);
3781 __release_stripe(conf, sh);
3785 int md_raid5_congested(struct mddev *mddev, int bits)
3787 struct r5conf *conf = mddev->private;
3789 /* No difference between reads and writes. Just check
3790 * how busy the stripe_cache is
3793 if (conf->inactive_blocked)
3797 if (list_empty_careful(&conf->inactive_list))
3802 EXPORT_SYMBOL_GPL(md_raid5_congested);
3804 static int raid5_congested(void *data, int bits)
3806 struct mddev *mddev = data;
3808 return mddev_congested(mddev, bits) ||
3809 md_raid5_congested(mddev, bits);
3812 /* We want read requests to align with chunks where possible,
3813 * but write requests don't need to.
3815 static int raid5_mergeable_bvec(struct request_queue *q,
3816 struct bvec_merge_data *bvm,
3817 struct bio_vec *biovec)
3819 struct mddev *mddev = q->queuedata;
3820 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3822 unsigned int chunk_sectors = mddev->chunk_sectors;
3823 unsigned int bio_sectors = bvm->bi_size >> 9;
3825 if ((bvm->bi_rw & 1) == WRITE)
3826 return biovec->bv_len; /* always allow writes to be mergeable */
3828 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3829 chunk_sectors = mddev->new_chunk_sectors;
3830 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3831 if (max < 0) max = 0;
3832 if (max <= biovec->bv_len && bio_sectors == 0)
3833 return biovec->bv_len;
3839 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3841 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3842 unsigned int chunk_sectors = mddev->chunk_sectors;
3843 unsigned int bio_sectors = bio->bi_size >> 9;
3845 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3846 chunk_sectors = mddev->new_chunk_sectors;
3847 return chunk_sectors >=
3848 ((sector & (chunk_sectors - 1)) + bio_sectors);
3852 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3853 * later sampled by raid5d.
3855 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3857 unsigned long flags;
3859 spin_lock_irqsave(&conf->device_lock, flags);
3861 bi->bi_next = conf->retry_read_aligned_list;
3862 conf->retry_read_aligned_list = bi;
3864 spin_unlock_irqrestore(&conf->device_lock, flags);
3865 md_wakeup_thread(conf->mddev->thread);
3869 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3873 bi = conf->retry_read_aligned;
3875 conf->retry_read_aligned = NULL;
3878 bi = conf->retry_read_aligned_list;
3880 conf->retry_read_aligned_list = bi->bi_next;
3883 * this sets the active strip count to 1 and the processed
3884 * strip count to zero (upper 8 bits)
3886 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
3894 * The "raid5_align_endio" should check if the read succeeded and if it
3895 * did, call bio_endio on the original bio (having bio_put the new bio
3897 * If the read failed..
3899 static void raid5_align_endio(struct bio *bi, int error)
3901 struct bio* raid_bi = bi->bi_private;
3902 struct mddev *mddev;
3903 struct r5conf *conf;
3904 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3905 struct md_rdev *rdev;
3909 rdev = (void*)raid_bi->bi_next;
3910 raid_bi->bi_next = NULL;
3911 mddev = rdev->mddev;
3912 conf = mddev->private;
3914 rdev_dec_pending(rdev, conf->mddev);
3916 if (!error && uptodate) {
3917 bio_endio(raid_bi, 0);
3918 if (atomic_dec_and_test(&conf->active_aligned_reads))
3919 wake_up(&conf->wait_for_stripe);
3924 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3926 add_bio_to_retry(raid_bi, conf);
3929 static int bio_fits_rdev(struct bio *bi)
3931 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3933 if ((bi->bi_size>>9) > queue_max_sectors(q))
3935 blk_recount_segments(q, bi);
3936 if (bi->bi_phys_segments > queue_max_segments(q))
3939 if (q->merge_bvec_fn)
3940 /* it's too hard to apply the merge_bvec_fn at this stage,
3949 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3951 struct r5conf *conf = mddev->private;
3953 struct bio* align_bi;
3954 struct md_rdev *rdev;
3955 sector_t end_sector;
3957 if (!in_chunk_boundary(mddev, raid_bio)) {
3958 pr_debug("chunk_aligned_read : non aligned\n");
3962 * use bio_clone_mddev to make a copy of the bio
3964 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3968 * set bi_end_io to a new function, and set bi_private to the
3971 align_bi->bi_end_io = raid5_align_endio;
3972 align_bi->bi_private = raid_bio;
3976 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3980 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3982 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3983 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3984 rdev->recovery_offset < end_sector) {
3985 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3987 (test_bit(Faulty, &rdev->flags) ||
3988 !(test_bit(In_sync, &rdev->flags) ||
3989 rdev->recovery_offset >= end_sector)))
3996 atomic_inc(&rdev->nr_pending);
3998 raid_bio->bi_next = (void*)rdev;
3999 align_bi->bi_bdev = rdev->bdev;
4000 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4002 if (!bio_fits_rdev(align_bi) ||
4003 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
4004 &first_bad, &bad_sectors)) {
4005 /* too big in some way, or has a known bad block */
4007 rdev_dec_pending(rdev, mddev);
4011 /* No reshape active, so we can trust rdev->data_offset */
4012 align_bi->bi_sector += rdev->data_offset;
4014 spin_lock_irq(&conf->device_lock);
4015 wait_event_lock_irq(conf->wait_for_stripe,
4018 atomic_inc(&conf->active_aligned_reads);
4019 spin_unlock_irq(&conf->device_lock);
4021 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4022 align_bi, disk_devt(mddev->gendisk),
4023 raid_bio->bi_sector);
4024 generic_make_request(align_bi);
4033 /* __get_priority_stripe - get the next stripe to process
4035 * Full stripe writes are allowed to pass preread active stripes up until
4036 * the bypass_threshold is exceeded. In general the bypass_count
4037 * increments when the handle_list is handled before the hold_list; however, it
4038 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4039 * stripe with in flight i/o. The bypass_count will be reset when the
4040 * head of the hold_list has changed, i.e. the head was promoted to the
4043 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
4045 struct stripe_head *sh;
4047 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4049 list_empty(&conf->handle_list) ? "empty" : "busy",
4050 list_empty(&conf->hold_list) ? "empty" : "busy",
4051 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4053 if (!list_empty(&conf->handle_list)) {
4054 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4056 if (list_empty(&conf->hold_list))
4057 conf->bypass_count = 0;
4058 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4059 if (conf->hold_list.next == conf->last_hold)
4060 conf->bypass_count++;
4062 conf->last_hold = conf->hold_list.next;
4063 conf->bypass_count -= conf->bypass_threshold;
4064 if (conf->bypass_count < 0)
4065 conf->bypass_count = 0;
4068 } else if (!list_empty(&conf->hold_list) &&
4069 ((conf->bypass_threshold &&
4070 conf->bypass_count > conf->bypass_threshold) ||
4071 atomic_read(&conf->pending_full_writes) == 0)) {
4072 sh = list_entry(conf->hold_list.next,
4074 conf->bypass_count -= conf->bypass_threshold;
4075 if (conf->bypass_count < 0)
4076 conf->bypass_count = 0;
4080 list_del_init(&sh->lru);
4081 atomic_inc(&sh->count);
4082 BUG_ON(atomic_read(&sh->count) != 1);
4086 struct raid5_plug_cb {
4087 struct blk_plug_cb cb;
4088 struct list_head list;
4091 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4093 struct raid5_plug_cb *cb = container_of(
4094 blk_cb, struct raid5_plug_cb, cb);
4095 struct stripe_head *sh;
4096 struct mddev *mddev = cb->cb.data;
4097 struct r5conf *conf = mddev->private;
4100 if (cb->list.next && !list_empty(&cb->list)) {
4101 spin_lock_irq(&conf->device_lock);
4102 while (!list_empty(&cb->list)) {
4103 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4104 list_del_init(&sh->lru);
4106 * avoid race release_stripe_plug() sees
4107 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4108 * is still in our list
4110 smp_mb__before_clear_bit();
4111 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4112 __release_stripe(conf, sh);
4115 spin_unlock_irq(&conf->device_lock);
4117 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4121 static void release_stripe_plug(struct mddev *mddev,
4122 struct stripe_head *sh)
4124 struct blk_plug_cb *blk_cb = blk_check_plugged(
4125 raid5_unplug, mddev,
4126 sizeof(struct raid5_plug_cb));
4127 struct raid5_plug_cb *cb;
4134 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4136 if (cb->list.next == NULL)
4137 INIT_LIST_HEAD(&cb->list);
4139 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4140 list_add_tail(&sh->lru, &cb->list);
4145 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4147 struct r5conf *conf = mddev->private;
4148 sector_t logical_sector, last_sector;
4149 struct stripe_head *sh;
4153 if (mddev->reshape_position != MaxSector)
4154 /* Skip discard while reshape is happening */
4157 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4158 last_sector = bi->bi_sector + (bi->bi_size>>9);
4161 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4163 stripe_sectors = conf->chunk_sectors *
4164 (conf->raid_disks - conf->max_degraded);
4165 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4167 sector_div(last_sector, stripe_sectors);
4169 logical_sector *= conf->chunk_sectors;
4170 last_sector *= conf->chunk_sectors;
4172 for (; logical_sector < last_sector;
4173 logical_sector += STRIPE_SECTORS) {
4177 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4178 prepare_to_wait(&conf->wait_for_overlap, &w,
4179 TASK_UNINTERRUPTIBLE);
4180 spin_lock_irq(&sh->stripe_lock);
4181 for (d = 0; d < conf->raid_disks; d++) {
4182 if (d == sh->pd_idx || d == sh->qd_idx)
4184 if (sh->dev[d].towrite || sh->dev[d].toread) {
4185 set_bit(R5_Overlap, &sh->dev[d].flags);
4186 spin_unlock_irq(&sh->stripe_lock);
4192 finish_wait(&conf->wait_for_overlap, &w);
4193 for (d = 0; d < conf->raid_disks; d++) {
4194 if (d == sh->pd_idx || d == sh->qd_idx)
4196 sh->dev[d].towrite = bi;
4197 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4198 raid5_inc_bi_active_stripes(bi);
4200 spin_unlock_irq(&sh->stripe_lock);
4201 if (conf->mddev->bitmap) {
4203 d < conf->raid_disks - conf->max_degraded;
4205 bitmap_startwrite(mddev->bitmap,
4209 sh->bm_seq = conf->seq_flush + 1;
4210 set_bit(STRIPE_BIT_DELAY, &sh->state);
4213 set_bit(STRIPE_HANDLE, &sh->state);
4214 clear_bit(STRIPE_DELAYED, &sh->state);
4215 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4216 atomic_inc(&conf->preread_active_stripes);
4217 release_stripe_plug(mddev, sh);
4220 remaining = raid5_dec_bi_active_stripes(bi);
4221 if (remaining == 0) {
4222 md_write_end(mddev);
4227 static void make_request(struct mddev *mddev, struct bio * bi)
4229 struct r5conf *conf = mddev->private;
4231 sector_t new_sector;
4232 sector_t logical_sector, last_sector;
4233 struct stripe_head *sh;
4234 const int rw = bio_data_dir(bi);
4237 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4238 md_flush_request(mddev, bi);
4242 md_write_start(mddev, bi);
4245 mddev->reshape_position == MaxSector &&
4246 chunk_aligned_read(mddev,bi))
4249 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4250 make_discard_request(mddev, bi);
4254 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4255 last_sector = bi->bi_sector + (bi->bi_size>>9);
4257 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4259 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4265 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4266 if (unlikely(conf->reshape_progress != MaxSector)) {
4267 /* spinlock is needed as reshape_progress may be
4268 * 64bit on a 32bit platform, and so it might be
4269 * possible to see a half-updated value
4270 * Of course reshape_progress could change after
4271 * the lock is dropped, so once we get a reference
4272 * to the stripe that we think it is, we will have
4275 spin_lock_irq(&conf->device_lock);
4276 if (mddev->reshape_backwards
4277 ? logical_sector < conf->reshape_progress
4278 : logical_sector >= conf->reshape_progress) {
4281 if (mddev->reshape_backwards
4282 ? logical_sector < conf->reshape_safe
4283 : logical_sector >= conf->reshape_safe) {
4284 spin_unlock_irq(&conf->device_lock);
4289 spin_unlock_irq(&conf->device_lock);
4292 new_sector = raid5_compute_sector(conf, logical_sector,
4295 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4296 (unsigned long long)new_sector,
4297 (unsigned long long)logical_sector);
4299 sh = get_active_stripe(conf, new_sector, previous,
4300 (bi->bi_rw&RWA_MASK), 0);
4302 if (unlikely(previous)) {
4303 /* expansion might have moved on while waiting for a
4304 * stripe, so we must do the range check again.
4305 * Expansion could still move past after this
4306 * test, but as we are holding a reference to
4307 * 'sh', we know that if that happens,
4308 * STRIPE_EXPANDING will get set and the expansion
4309 * won't proceed until we finish with the stripe.
4312 spin_lock_irq(&conf->device_lock);
4313 if (mddev->reshape_backwards
4314 ? logical_sector >= conf->reshape_progress
4315 : logical_sector < conf->reshape_progress)
4316 /* mismatch, need to try again */
4318 spin_unlock_irq(&conf->device_lock);
4327 logical_sector >= mddev->suspend_lo &&
4328 logical_sector < mddev->suspend_hi) {
4330 /* As the suspend_* range is controlled by
4331 * userspace, we want an interruptible
4334 flush_signals(current);
4335 prepare_to_wait(&conf->wait_for_overlap,
4336 &w, TASK_INTERRUPTIBLE);
4337 if (logical_sector >= mddev->suspend_lo &&
4338 logical_sector < mddev->suspend_hi)
4343 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4344 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4345 /* Stripe is busy expanding or
4346 * add failed due to overlap. Flush everything
4349 md_wakeup_thread(mddev->thread);
4354 finish_wait(&conf->wait_for_overlap, &w);
4355 set_bit(STRIPE_HANDLE, &sh->state);
4356 clear_bit(STRIPE_DELAYED, &sh->state);
4357 if ((bi->bi_rw & REQ_SYNC) &&
4358 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4359 atomic_inc(&conf->preread_active_stripes);
4360 release_stripe_plug(mddev, sh);
4362 /* cannot get stripe for read-ahead, just give-up */
4363 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4364 finish_wait(&conf->wait_for_overlap, &w);
4369 remaining = raid5_dec_bi_active_stripes(bi);
4370 if (remaining == 0) {
4373 md_write_end(mddev);
4379 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4381 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4383 /* reshaping is quite different to recovery/resync so it is
4384 * handled quite separately ... here.
4386 * On each call to sync_request, we gather one chunk worth of
4387 * destination stripes and flag them as expanding.
4388 * Then we find all the source stripes and request reads.
4389 * As the reads complete, handle_stripe will copy the data
4390 * into the destination stripe and release that stripe.
4392 struct r5conf *conf = mddev->private;
4393 struct stripe_head *sh;
4394 sector_t first_sector, last_sector;
4395 int raid_disks = conf->previous_raid_disks;
4396 int data_disks = raid_disks - conf->max_degraded;
4397 int new_data_disks = conf->raid_disks - conf->max_degraded;
4400 sector_t writepos, readpos, safepos;
4401 sector_t stripe_addr;
4402 int reshape_sectors;
4403 struct list_head stripes;
4405 if (sector_nr == 0) {
4406 /* If restarting in the middle, skip the initial sectors */
4407 if (mddev->reshape_backwards &&
4408 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4409 sector_nr = raid5_size(mddev, 0, 0)
4410 - conf->reshape_progress;
4411 } else if (!mddev->reshape_backwards &&
4412 conf->reshape_progress > 0)
4413 sector_nr = conf->reshape_progress;
4414 sector_div(sector_nr, new_data_disks);
4416 mddev->curr_resync_completed = sector_nr;
4417 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4423 /* We need to process a full chunk at a time.
4424 * If old and new chunk sizes differ, we need to process the
4427 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4428 reshape_sectors = mddev->new_chunk_sectors;
4430 reshape_sectors = mddev->chunk_sectors;
4432 /* We update the metadata at least every 10 seconds, or when
4433 * the data about to be copied would over-write the source of
4434 * the data at the front of the range. i.e. one new_stripe
4435 * along from reshape_progress new_maps to after where
4436 * reshape_safe old_maps to
4438 writepos = conf->reshape_progress;
4439 sector_div(writepos, new_data_disks);
4440 readpos = conf->reshape_progress;
4441 sector_div(readpos, data_disks);
4442 safepos = conf->reshape_safe;
4443 sector_div(safepos, data_disks);
4444 if (mddev->reshape_backwards) {
4445 writepos -= min_t(sector_t, reshape_sectors, writepos);
4446 readpos += reshape_sectors;
4447 safepos += reshape_sectors;
4449 writepos += reshape_sectors;
4450 readpos -= min_t(sector_t, reshape_sectors, readpos);
4451 safepos -= min_t(sector_t, reshape_sectors, safepos);
4454 /* Having calculated the 'writepos' possibly use it
4455 * to set 'stripe_addr' which is where we will write to.
4457 if (mddev->reshape_backwards) {
4458 BUG_ON(conf->reshape_progress == 0);
4459 stripe_addr = writepos;
4460 BUG_ON((mddev->dev_sectors &
4461 ~((sector_t)reshape_sectors - 1))
4462 - reshape_sectors - stripe_addr
4465 BUG_ON(writepos != sector_nr + reshape_sectors);
4466 stripe_addr = sector_nr;
4469 /* 'writepos' is the most advanced device address we might write.
4470 * 'readpos' is the least advanced device address we might read.
4471 * 'safepos' is the least address recorded in the metadata as having
4473 * If there is a min_offset_diff, these are adjusted either by
4474 * increasing the safepos/readpos if diff is negative, or
4475 * increasing writepos if diff is positive.
4476 * If 'readpos' is then behind 'writepos', there is no way that we can
4477 * ensure safety in the face of a crash - that must be done by userspace
4478 * making a backup of the data. So in that case there is no particular
4479 * rush to update metadata.
4480 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4481 * update the metadata to advance 'safepos' to match 'readpos' so that
4482 * we can be safe in the event of a crash.
4483 * So we insist on updating metadata if safepos is behind writepos and
4484 * readpos is beyond writepos.
4485 * In any case, update the metadata every 10 seconds.
4486 * Maybe that number should be configurable, but I'm not sure it is
4487 * worth it.... maybe it could be a multiple of safemode_delay???
4489 if (conf->min_offset_diff < 0) {
4490 safepos += -conf->min_offset_diff;
4491 readpos += -conf->min_offset_diff;
4493 writepos += conf->min_offset_diff;
4495 if ((mddev->reshape_backwards
4496 ? (safepos > writepos && readpos < writepos)
4497 : (safepos < writepos && readpos > writepos)) ||
4498 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4499 /* Cannot proceed until we've updated the superblock... */
4500 wait_event(conf->wait_for_overlap,
4501 atomic_read(&conf->reshape_stripes)==0);
4502 mddev->reshape_position = conf->reshape_progress;
4503 mddev->curr_resync_completed = sector_nr;
4504 conf->reshape_checkpoint = jiffies;
4505 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4506 md_wakeup_thread(mddev->thread);
4507 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4508 kthread_should_stop());
4509 spin_lock_irq(&conf->device_lock);
4510 conf->reshape_safe = mddev->reshape_position;
4511 spin_unlock_irq(&conf->device_lock);
4512 wake_up(&conf->wait_for_overlap);
4513 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4516 INIT_LIST_HEAD(&stripes);
4517 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4519 int skipped_disk = 0;
4520 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4521 set_bit(STRIPE_EXPANDING, &sh->state);
4522 atomic_inc(&conf->reshape_stripes);
4523 /* If any of this stripe is beyond the end of the old
4524 * array, then we need to zero those blocks
4526 for (j=sh->disks; j--;) {
4528 if (j == sh->pd_idx)
4530 if (conf->level == 6 &&
4533 s = compute_blocknr(sh, j, 0);
4534 if (s < raid5_size(mddev, 0, 0)) {
4538 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4539 set_bit(R5_Expanded, &sh->dev[j].flags);
4540 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4542 if (!skipped_disk) {
4543 set_bit(STRIPE_EXPAND_READY, &sh->state);
4544 set_bit(STRIPE_HANDLE, &sh->state);
4546 list_add(&sh->lru, &stripes);
4548 spin_lock_irq(&conf->device_lock);
4549 if (mddev->reshape_backwards)
4550 conf->reshape_progress -= reshape_sectors * new_data_disks;
4552 conf->reshape_progress += reshape_sectors * new_data_disks;
4553 spin_unlock_irq(&conf->device_lock);
4554 /* Ok, those stripe are ready. We can start scheduling
4555 * reads on the source stripes.
4556 * The source stripes are determined by mapping the first and last
4557 * block on the destination stripes.
4560 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4563 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4564 * new_data_disks - 1),
4566 if (last_sector >= mddev->dev_sectors)
4567 last_sector = mddev->dev_sectors - 1;
4568 while (first_sector <= last_sector) {
4569 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4570 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4571 set_bit(STRIPE_HANDLE, &sh->state);
4573 first_sector += STRIPE_SECTORS;
4575 /* Now that the sources are clearly marked, we can release
4576 * the destination stripes
4578 while (!list_empty(&stripes)) {
4579 sh = list_entry(stripes.next, struct stripe_head, lru);
4580 list_del_init(&sh->lru);
4583 /* If this takes us to the resync_max point where we have to pause,
4584 * then we need to write out the superblock.
4586 sector_nr += reshape_sectors;
4587 if ((sector_nr - mddev->curr_resync_completed) * 2
4588 >= mddev->resync_max - mddev->curr_resync_completed) {
4589 /* Cannot proceed until we've updated the superblock... */
4590 wait_event(conf->wait_for_overlap,
4591 atomic_read(&conf->reshape_stripes) == 0);
4592 mddev->reshape_position = conf->reshape_progress;
4593 mddev->curr_resync_completed = sector_nr;
4594 conf->reshape_checkpoint = jiffies;
4595 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4596 md_wakeup_thread(mddev->thread);
4597 wait_event(mddev->sb_wait,
4598 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4599 || kthread_should_stop());
4600 spin_lock_irq(&conf->device_lock);
4601 conf->reshape_safe = mddev->reshape_position;
4602 spin_unlock_irq(&conf->device_lock);
4603 wake_up(&conf->wait_for_overlap);
4604 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4606 return reshape_sectors;
4609 /* FIXME go_faster isn't used */
4610 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4612 struct r5conf *conf = mddev->private;
4613 struct stripe_head *sh;
4614 sector_t max_sector = mddev->dev_sectors;
4615 sector_t sync_blocks;
4616 int still_degraded = 0;
4619 if (sector_nr >= max_sector) {
4620 /* just being told to finish up .. nothing much to do */
4622 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4627 if (mddev->curr_resync < max_sector) /* aborted */
4628 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4630 else /* completed sync */
4632 bitmap_close_sync(mddev->bitmap);
4637 /* Allow raid5_quiesce to complete */
4638 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4640 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4641 return reshape_request(mddev, sector_nr, skipped);
4643 /* No need to check resync_max as we never do more than one
4644 * stripe, and as resync_max will always be on a chunk boundary,
4645 * if the check in md_do_sync didn't fire, there is no chance
4646 * of overstepping resync_max here
4649 /* if there is too many failed drives and we are trying
4650 * to resync, then assert that we are finished, because there is
4651 * nothing we can do.
4653 if (mddev->degraded >= conf->max_degraded &&
4654 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4655 sector_t rv = mddev->dev_sectors - sector_nr;
4659 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4660 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4661 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4662 /* we can skip this block, and probably more */
4663 sync_blocks /= STRIPE_SECTORS;
4665 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4668 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4670 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4672 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4673 /* make sure we don't swamp the stripe cache if someone else
4674 * is trying to get access
4676 schedule_timeout_uninterruptible(1);
4678 /* Need to check if array will still be degraded after recovery/resync
4679 * We don't need to check the 'failed' flag as when that gets set,
4682 for (i = 0; i < conf->raid_disks; i++)
4683 if (conf->disks[i].rdev == NULL)
4686 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4688 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4693 return STRIPE_SECTORS;
4696 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4698 /* We may not be able to submit a whole bio at once as there
4699 * may not be enough stripe_heads available.
4700 * We cannot pre-allocate enough stripe_heads as we may need
4701 * more than exist in the cache (if we allow ever large chunks).
4702 * So we do one stripe head at a time and record in
4703 * ->bi_hw_segments how many have been done.
4705 * We *know* that this entire raid_bio is in one chunk, so
4706 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4708 struct stripe_head *sh;
4710 sector_t sector, logical_sector, last_sector;
4715 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4716 sector = raid5_compute_sector(conf, logical_sector,
4718 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4720 for (; logical_sector < last_sector;
4721 logical_sector += STRIPE_SECTORS,
4722 sector += STRIPE_SECTORS,
4725 if (scnt < raid5_bi_processed_stripes(raid_bio))
4726 /* already done this stripe */
4729 sh = get_active_stripe(conf, sector, 0, 1, 0);
4732 /* failed to get a stripe - must wait */
4733 raid5_set_bi_processed_stripes(raid_bio, scnt);
4734 conf->retry_read_aligned = raid_bio;
4738 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4740 raid5_set_bi_processed_stripes(raid_bio, scnt);
4741 conf->retry_read_aligned = raid_bio;
4745 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4750 remaining = raid5_dec_bi_active_stripes(raid_bio);
4752 bio_endio(raid_bio, 0);
4753 if (atomic_dec_and_test(&conf->active_aligned_reads))
4754 wake_up(&conf->wait_for_stripe);
4758 #define MAX_STRIPE_BATCH 8
4759 static int handle_active_stripes(struct r5conf *conf)
4761 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4762 int i, batch_size = 0;
4764 while (batch_size < MAX_STRIPE_BATCH &&
4765 (sh = __get_priority_stripe(conf)) != NULL)
4766 batch[batch_size++] = sh;
4768 if (batch_size == 0)
4770 spin_unlock_irq(&conf->device_lock);
4772 for (i = 0; i < batch_size; i++)
4773 handle_stripe(batch[i]);
4777 spin_lock_irq(&conf->device_lock);
4778 for (i = 0; i < batch_size; i++)
4779 __release_stripe(conf, batch[i]);
4784 * This is our raid5 kernel thread.
4786 * We scan the hash table for stripes which can be handled now.
4787 * During the scan, completed stripes are saved for us by the interrupt
4788 * handler, so that they will not have to wait for our next wakeup.
4790 static void raid5d(struct md_thread *thread)
4792 struct mddev *mddev = thread->mddev;
4793 struct r5conf *conf = mddev->private;
4795 struct blk_plug plug;
4797 pr_debug("+++ raid5d active\n");
4799 md_check_recovery(mddev);
4801 blk_start_plug(&plug);
4803 spin_lock_irq(&conf->device_lock);
4809 !list_empty(&conf->bitmap_list)) {
4810 /* Now is a good time to flush some bitmap updates */
4812 spin_unlock_irq(&conf->device_lock);
4813 bitmap_unplug(mddev->bitmap);
4814 spin_lock_irq(&conf->device_lock);
4815 conf->seq_write = conf->seq_flush;
4816 activate_bit_delay(conf);
4818 raid5_activate_delayed(conf);
4820 while ((bio = remove_bio_from_retry(conf))) {
4822 spin_unlock_irq(&conf->device_lock);
4823 ok = retry_aligned_read(conf, bio);
4824 spin_lock_irq(&conf->device_lock);
4830 batch_size = handle_active_stripes(conf);
4833 handled += batch_size;
4835 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
4836 spin_unlock_irq(&conf->device_lock);
4837 md_check_recovery(mddev);
4838 spin_lock_irq(&conf->device_lock);
4841 pr_debug("%d stripes handled\n", handled);
4843 spin_unlock_irq(&conf->device_lock);
4845 async_tx_issue_pending_all();
4846 blk_finish_plug(&plug);
4848 pr_debug("--- raid5d inactive\n");
4852 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4854 struct r5conf *conf = mddev->private;
4856 return sprintf(page, "%d\n", conf->max_nr_stripes);
4862 raid5_set_cache_size(struct mddev *mddev, int size)
4864 struct r5conf *conf = mddev->private;
4867 if (size <= 16 || size > 32768)
4869 while (size < conf->max_nr_stripes) {
4870 if (drop_one_stripe(conf))
4871 conf->max_nr_stripes--;
4875 err = md_allow_write(mddev);
4878 while (size > conf->max_nr_stripes) {
4879 if (grow_one_stripe(conf))
4880 conf->max_nr_stripes++;
4885 EXPORT_SYMBOL(raid5_set_cache_size);
4888 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4890 struct r5conf *conf = mddev->private;
4894 if (len >= PAGE_SIZE)
4899 if (strict_strtoul(page, 10, &new))
4901 err = raid5_set_cache_size(mddev, new);
4907 static struct md_sysfs_entry
4908 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4909 raid5_show_stripe_cache_size,
4910 raid5_store_stripe_cache_size);
4913 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4915 struct r5conf *conf = mddev->private;
4917 return sprintf(page, "%d\n", conf->bypass_threshold);
4923 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4925 struct r5conf *conf = mddev->private;
4927 if (len >= PAGE_SIZE)
4932 if (strict_strtoul(page, 10, &new))
4934 if (new > conf->max_nr_stripes)
4936 conf->bypass_threshold = new;
4940 static struct md_sysfs_entry
4941 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4943 raid5_show_preread_threshold,
4944 raid5_store_preread_threshold);
4947 stripe_cache_active_show(struct mddev *mddev, char *page)
4949 struct r5conf *conf = mddev->private;
4951 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4956 static struct md_sysfs_entry
4957 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4959 static struct attribute *raid5_attrs[] = {
4960 &raid5_stripecache_size.attr,
4961 &raid5_stripecache_active.attr,
4962 &raid5_preread_bypass_threshold.attr,
4965 static struct attribute_group raid5_attrs_group = {
4967 .attrs = raid5_attrs,
4971 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4973 struct r5conf *conf = mddev->private;
4976 sectors = mddev->dev_sectors;
4978 /* size is defined by the smallest of previous and new size */
4979 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4981 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4982 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4983 return sectors * (raid_disks - conf->max_degraded);
4986 static void raid5_free_percpu(struct r5conf *conf)
4988 struct raid5_percpu *percpu;
4995 for_each_possible_cpu(cpu) {
4996 percpu = per_cpu_ptr(conf->percpu, cpu);
4997 safe_put_page(percpu->spare_page);
4998 kfree(percpu->scribble);
5000 #ifdef CONFIG_HOTPLUG_CPU
5001 unregister_cpu_notifier(&conf->cpu_notify);
5005 free_percpu(conf->percpu);
5008 static void free_conf(struct r5conf *conf)
5010 shrink_stripes(conf);
5011 raid5_free_percpu(conf);
5013 kfree(conf->stripe_hashtbl);
5017 #ifdef CONFIG_HOTPLUG_CPU
5018 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5021 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5022 long cpu = (long)hcpu;
5023 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5026 case CPU_UP_PREPARE:
5027 case CPU_UP_PREPARE_FROZEN:
5028 if (conf->level == 6 && !percpu->spare_page)
5029 percpu->spare_page = alloc_page(GFP_KERNEL);
5030 if (!percpu->scribble)
5031 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5033 if (!percpu->scribble ||
5034 (conf->level == 6 && !percpu->spare_page)) {
5035 safe_put_page(percpu->spare_page);
5036 kfree(percpu->scribble);
5037 pr_err("%s: failed memory allocation for cpu%ld\n",
5039 return notifier_from_errno(-ENOMEM);
5043 case CPU_DEAD_FROZEN:
5044 safe_put_page(percpu->spare_page);
5045 kfree(percpu->scribble);
5046 percpu->spare_page = NULL;
5047 percpu->scribble = NULL;
5056 static int raid5_alloc_percpu(struct r5conf *conf)
5059 struct page *spare_page;
5060 struct raid5_percpu __percpu *allcpus;
5064 allcpus = alloc_percpu(struct raid5_percpu);
5067 conf->percpu = allcpus;
5071 for_each_present_cpu(cpu) {
5072 if (conf->level == 6) {
5073 spare_page = alloc_page(GFP_KERNEL);
5078 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5080 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5085 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5087 #ifdef CONFIG_HOTPLUG_CPU
5088 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5089 conf->cpu_notify.priority = 0;
5091 err = register_cpu_notifier(&conf->cpu_notify);
5098 static struct r5conf *setup_conf(struct mddev *mddev)
5100 struct r5conf *conf;
5101 int raid_disk, memory, max_disks;
5102 struct md_rdev *rdev;
5103 struct disk_info *disk;
5106 if (mddev->new_level != 5
5107 && mddev->new_level != 4
5108 && mddev->new_level != 6) {
5109 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5110 mdname(mddev), mddev->new_level);
5111 return ERR_PTR(-EIO);
5113 if ((mddev->new_level == 5
5114 && !algorithm_valid_raid5(mddev->new_layout)) ||
5115 (mddev->new_level == 6
5116 && !algorithm_valid_raid6(mddev->new_layout))) {
5117 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5118 mdname(mddev), mddev->new_layout);
5119 return ERR_PTR(-EIO);
5121 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5122 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5123 mdname(mddev), mddev->raid_disks);
5124 return ERR_PTR(-EINVAL);
5127 if (!mddev->new_chunk_sectors ||
5128 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5129 !is_power_of_2(mddev->new_chunk_sectors)) {
5130 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5131 mdname(mddev), mddev->new_chunk_sectors << 9);
5132 return ERR_PTR(-EINVAL);
5135 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5138 spin_lock_init(&conf->device_lock);
5139 init_waitqueue_head(&conf->wait_for_stripe);
5140 init_waitqueue_head(&conf->wait_for_overlap);
5141 INIT_LIST_HEAD(&conf->handle_list);
5142 INIT_LIST_HEAD(&conf->hold_list);
5143 INIT_LIST_HEAD(&conf->delayed_list);
5144 INIT_LIST_HEAD(&conf->bitmap_list);
5145 INIT_LIST_HEAD(&conf->inactive_list);
5146 atomic_set(&conf->active_stripes, 0);
5147 atomic_set(&conf->preread_active_stripes, 0);
5148 atomic_set(&conf->active_aligned_reads, 0);
5149 conf->bypass_threshold = BYPASS_THRESHOLD;
5150 conf->recovery_disabled = mddev->recovery_disabled - 1;
5152 conf->raid_disks = mddev->raid_disks;
5153 if (mddev->reshape_position == MaxSector)
5154 conf->previous_raid_disks = mddev->raid_disks;
5156 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5157 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5158 conf->scribble_len = scribble_len(max_disks);
5160 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5165 conf->mddev = mddev;
5167 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5170 conf->level = mddev->new_level;
5171 if (raid5_alloc_percpu(conf) != 0)
5174 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5176 rdev_for_each(rdev, mddev) {
5177 raid_disk = rdev->raid_disk;
5178 if (raid_disk >= max_disks
5181 disk = conf->disks + raid_disk;
5183 if (test_bit(Replacement, &rdev->flags)) {
5184 if (disk->replacement)
5186 disk->replacement = rdev;
5193 if (test_bit(In_sync, &rdev->flags)) {
5194 char b[BDEVNAME_SIZE];
5195 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5197 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5198 } else if (rdev->saved_raid_disk != raid_disk)
5199 /* Cannot rely on bitmap to complete recovery */
5203 conf->chunk_sectors = mddev->new_chunk_sectors;
5204 conf->level = mddev->new_level;
5205 if (conf->level == 6)
5206 conf->max_degraded = 2;
5208 conf->max_degraded = 1;
5209 conf->algorithm = mddev->new_layout;
5210 conf->max_nr_stripes = NR_STRIPES;
5211 conf->reshape_progress = mddev->reshape_position;
5212 if (conf->reshape_progress != MaxSector) {
5213 conf->prev_chunk_sectors = mddev->chunk_sectors;
5214 conf->prev_algo = mddev->layout;
5217 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5218 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5219 if (grow_stripes(conf, conf->max_nr_stripes)) {
5221 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5222 mdname(mddev), memory);
5225 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5226 mdname(mddev), memory);
5228 sprintf(pers_name, "raid%d", mddev->new_level);
5229 conf->thread = md_register_thread(raid5d, mddev, pers_name);
5230 if (!conf->thread) {
5232 "md/raid:%s: couldn't allocate thread.\n",
5242 return ERR_PTR(-EIO);
5244 return ERR_PTR(-ENOMEM);
5248 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5251 case ALGORITHM_PARITY_0:
5252 if (raid_disk < max_degraded)
5255 case ALGORITHM_PARITY_N:
5256 if (raid_disk >= raid_disks - max_degraded)
5259 case ALGORITHM_PARITY_0_6:
5260 if (raid_disk == 0 ||
5261 raid_disk == raid_disks - 1)
5264 case ALGORITHM_LEFT_ASYMMETRIC_6:
5265 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5266 case ALGORITHM_LEFT_SYMMETRIC_6:
5267 case ALGORITHM_RIGHT_SYMMETRIC_6:
5268 if (raid_disk == raid_disks - 1)
5274 static int run(struct mddev *mddev)
5276 struct r5conf *conf;
5277 int working_disks = 0;
5278 int dirty_parity_disks = 0;
5279 struct md_rdev *rdev;
5280 sector_t reshape_offset = 0;
5282 long long min_offset_diff = 0;
5285 if (mddev->recovery_cp != MaxSector)
5286 printk(KERN_NOTICE "md/raid:%s: not clean"
5287 " -- starting background reconstruction\n",
5290 rdev_for_each(rdev, mddev) {
5292 if (rdev->raid_disk < 0)
5294 diff = (rdev->new_data_offset - rdev->data_offset);
5296 min_offset_diff = diff;
5298 } else if (mddev->reshape_backwards &&
5299 diff < min_offset_diff)
5300 min_offset_diff = diff;
5301 else if (!mddev->reshape_backwards &&
5302 diff > min_offset_diff)
5303 min_offset_diff = diff;
5306 if (mddev->reshape_position != MaxSector) {
5307 /* Check that we can continue the reshape.
5308 * Difficulties arise if the stripe we would write to
5309 * next is at or after the stripe we would read from next.
5310 * For a reshape that changes the number of devices, this
5311 * is only possible for a very short time, and mdadm makes
5312 * sure that time appears to have past before assembling
5313 * the array. So we fail if that time hasn't passed.
5314 * For a reshape that keeps the number of devices the same
5315 * mdadm must be monitoring the reshape can keeping the
5316 * critical areas read-only and backed up. It will start
5317 * the array in read-only mode, so we check for that.
5319 sector_t here_new, here_old;
5321 int max_degraded = (mddev->level == 6 ? 2 : 1);
5323 if (mddev->new_level != mddev->level) {
5324 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5325 "required - aborting.\n",
5329 old_disks = mddev->raid_disks - mddev->delta_disks;
5330 /* reshape_position must be on a new-stripe boundary, and one
5331 * further up in new geometry must map after here in old
5334 here_new = mddev->reshape_position;
5335 if (sector_div(here_new, mddev->new_chunk_sectors *
5336 (mddev->raid_disks - max_degraded))) {
5337 printk(KERN_ERR "md/raid:%s: reshape_position not "
5338 "on a stripe boundary\n", mdname(mddev));
5341 reshape_offset = here_new * mddev->new_chunk_sectors;
5342 /* here_new is the stripe we will write to */
5343 here_old = mddev->reshape_position;
5344 sector_div(here_old, mddev->chunk_sectors *
5345 (old_disks-max_degraded));
5346 /* here_old is the first stripe that we might need to read
5348 if (mddev->delta_disks == 0) {
5349 if ((here_new * mddev->new_chunk_sectors !=
5350 here_old * mddev->chunk_sectors)) {
5351 printk(KERN_ERR "md/raid:%s: reshape position is"
5352 " confused - aborting\n", mdname(mddev));
5355 /* We cannot be sure it is safe to start an in-place
5356 * reshape. It is only safe if user-space is monitoring
5357 * and taking constant backups.
5358 * mdadm always starts a situation like this in
5359 * readonly mode so it can take control before
5360 * allowing any writes. So just check for that.
5362 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5363 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5364 /* not really in-place - so OK */;
5365 else if (mddev->ro == 0) {
5366 printk(KERN_ERR "md/raid:%s: in-place reshape "
5367 "must be started in read-only mode "
5372 } else if (mddev->reshape_backwards
5373 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5374 here_old * mddev->chunk_sectors)
5375 : (here_new * mddev->new_chunk_sectors >=
5376 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5377 /* Reading from the same stripe as writing to - bad */
5378 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5379 "auto-recovery - aborting.\n",
5383 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5385 /* OK, we should be able to continue; */
5387 BUG_ON(mddev->level != mddev->new_level);
5388 BUG_ON(mddev->layout != mddev->new_layout);
5389 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5390 BUG_ON(mddev->delta_disks != 0);
5393 if (mddev->private == NULL)
5394 conf = setup_conf(mddev);
5396 conf = mddev->private;
5399 return PTR_ERR(conf);
5401 conf->min_offset_diff = min_offset_diff;
5402 mddev->thread = conf->thread;
5403 conf->thread = NULL;
5404 mddev->private = conf;
5406 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5408 rdev = conf->disks[i].rdev;
5409 if (!rdev && conf->disks[i].replacement) {
5410 /* The replacement is all we have yet */
5411 rdev = conf->disks[i].replacement;
5412 conf->disks[i].replacement = NULL;
5413 clear_bit(Replacement, &rdev->flags);
5414 conf->disks[i].rdev = rdev;
5418 if (conf->disks[i].replacement &&
5419 conf->reshape_progress != MaxSector) {
5420 /* replacements and reshape simply do not mix. */
5421 printk(KERN_ERR "md: cannot handle concurrent "
5422 "replacement and reshape.\n");
5425 if (test_bit(In_sync, &rdev->flags)) {
5429 /* This disc is not fully in-sync. However if it
5430 * just stored parity (beyond the recovery_offset),
5431 * when we don't need to be concerned about the
5432 * array being dirty.
5433 * When reshape goes 'backwards', we never have
5434 * partially completed devices, so we only need
5435 * to worry about reshape going forwards.
5437 /* Hack because v0.91 doesn't store recovery_offset properly. */
5438 if (mddev->major_version == 0 &&
5439 mddev->minor_version > 90)
5440 rdev->recovery_offset = reshape_offset;
5442 if (rdev->recovery_offset < reshape_offset) {
5443 /* We need to check old and new layout */
5444 if (!only_parity(rdev->raid_disk,
5447 conf->max_degraded))
5450 if (!only_parity(rdev->raid_disk,
5452 conf->previous_raid_disks,
5453 conf->max_degraded))
5455 dirty_parity_disks++;
5459 * 0 for a fully functional array, 1 or 2 for a degraded array.
5461 mddev->degraded = calc_degraded(conf);
5463 if (has_failed(conf)) {
5464 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5465 " (%d/%d failed)\n",
5466 mdname(mddev), mddev->degraded, conf->raid_disks);
5470 /* device size must be a multiple of chunk size */
5471 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5472 mddev->resync_max_sectors = mddev->dev_sectors;
5474 if (mddev->degraded > dirty_parity_disks &&
5475 mddev->recovery_cp != MaxSector) {
5476 if (mddev->ok_start_degraded)
5478 "md/raid:%s: starting dirty degraded array"
5479 " - data corruption possible.\n",
5483 "md/raid:%s: cannot start dirty degraded array.\n",
5489 if (mddev->degraded == 0)
5490 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5491 " devices, algorithm %d\n", mdname(mddev), conf->level,
5492 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5495 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5496 " out of %d devices, algorithm %d\n",
5497 mdname(mddev), conf->level,
5498 mddev->raid_disks - mddev->degraded,
5499 mddev->raid_disks, mddev->new_layout);
5501 print_raid5_conf(conf);
5503 if (conf->reshape_progress != MaxSector) {
5504 conf->reshape_safe = conf->reshape_progress;
5505 atomic_set(&conf->reshape_stripes, 0);
5506 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5507 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5508 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5509 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5510 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5515 /* Ok, everything is just fine now */
5516 if (mddev->to_remove == &raid5_attrs_group)
5517 mddev->to_remove = NULL;
5518 else if (mddev->kobj.sd &&
5519 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5521 "raid5: failed to create sysfs attributes for %s\n",
5523 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5527 bool discard_supported = true;
5528 /* read-ahead size must cover two whole stripes, which
5529 * is 2 * (datadisks) * chunksize where 'n' is the
5530 * number of raid devices
5532 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5533 int stripe = data_disks *
5534 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5535 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5536 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5538 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5540 mddev->queue->backing_dev_info.congested_data = mddev;
5541 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5543 chunk_size = mddev->chunk_sectors << 9;
5544 blk_queue_io_min(mddev->queue, chunk_size);
5545 blk_queue_io_opt(mddev->queue, chunk_size *
5546 (conf->raid_disks - conf->max_degraded));
5548 * We can only discard a whole stripe. It doesn't make sense to
5549 * discard data disk but write parity disk
5551 stripe = stripe * PAGE_SIZE;
5552 /* Round up to power of 2, as discard handling
5553 * currently assumes that */
5554 while ((stripe-1) & stripe)
5555 stripe = (stripe | (stripe-1)) + 1;
5556 mddev->queue->limits.discard_alignment = stripe;
5557 mddev->queue->limits.discard_granularity = stripe;
5559 * unaligned part of discard request will be ignored, so can't
5560 * guarantee discard_zerors_data
5562 mddev->queue->limits.discard_zeroes_data = 0;
5564 rdev_for_each(rdev, mddev) {
5565 disk_stack_limits(mddev->gendisk, rdev->bdev,
5566 rdev->data_offset << 9);
5567 disk_stack_limits(mddev->gendisk, rdev->bdev,
5568 rdev->new_data_offset << 9);
5570 * discard_zeroes_data is required, otherwise data
5571 * could be lost. Consider a scenario: discard a stripe
5572 * (the stripe could be inconsistent if
5573 * discard_zeroes_data is 0); write one disk of the
5574 * stripe (the stripe could be inconsistent again
5575 * depending on which disks are used to calculate
5576 * parity); the disk is broken; The stripe data of this
5579 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5580 !bdev_get_queue(rdev->bdev)->
5581 limits.discard_zeroes_data)
5582 discard_supported = false;
5585 if (discard_supported &&
5586 mddev->queue->limits.max_discard_sectors >= stripe &&
5587 mddev->queue->limits.discard_granularity >= stripe)
5588 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5591 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5597 md_unregister_thread(&mddev->thread);
5598 print_raid5_conf(conf);
5600 mddev->private = NULL;
5601 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5605 static int stop(struct mddev *mddev)
5607 struct r5conf *conf = mddev->private;
5609 md_unregister_thread(&mddev->thread);
5611 mddev->queue->backing_dev_info.congested_fn = NULL;
5613 mddev->private = NULL;
5614 mddev->to_remove = &raid5_attrs_group;
5618 static void status(struct seq_file *seq, struct mddev *mddev)
5620 struct r5conf *conf = mddev->private;
5623 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5624 mddev->chunk_sectors / 2, mddev->layout);
5625 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5626 for (i = 0; i < conf->raid_disks; i++)
5627 seq_printf (seq, "%s",
5628 conf->disks[i].rdev &&
5629 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5630 seq_printf (seq, "]");
5633 static void print_raid5_conf (struct r5conf *conf)
5636 struct disk_info *tmp;
5638 printk(KERN_DEBUG "RAID conf printout:\n");
5640 printk("(conf==NULL)\n");
5643 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5645 conf->raid_disks - conf->mddev->degraded);
5647 for (i = 0; i < conf->raid_disks; i++) {
5648 char b[BDEVNAME_SIZE];
5649 tmp = conf->disks + i;
5651 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5652 i, !test_bit(Faulty, &tmp->rdev->flags),
5653 bdevname(tmp->rdev->bdev, b));
5657 static int raid5_spare_active(struct mddev *mddev)
5660 struct r5conf *conf = mddev->private;
5661 struct disk_info *tmp;
5663 unsigned long flags;
5665 for (i = 0; i < conf->raid_disks; i++) {
5666 tmp = conf->disks + i;
5667 if (tmp->replacement
5668 && tmp->replacement->recovery_offset == MaxSector
5669 && !test_bit(Faulty, &tmp->replacement->flags)
5670 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5671 /* Replacement has just become active. */
5673 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5676 /* Replaced device not technically faulty,
5677 * but we need to be sure it gets removed
5678 * and never re-added.
5680 set_bit(Faulty, &tmp->rdev->flags);
5681 sysfs_notify_dirent_safe(
5682 tmp->rdev->sysfs_state);
5684 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5685 } else if (tmp->rdev
5686 && tmp->rdev->recovery_offset == MaxSector
5687 && !test_bit(Faulty, &tmp->rdev->flags)
5688 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5690 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5693 spin_lock_irqsave(&conf->device_lock, flags);
5694 mddev->degraded = calc_degraded(conf);
5695 spin_unlock_irqrestore(&conf->device_lock, flags);
5696 print_raid5_conf(conf);
5700 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5702 struct r5conf *conf = mddev->private;
5704 int number = rdev->raid_disk;
5705 struct md_rdev **rdevp;
5706 struct disk_info *p = conf->disks + number;
5708 print_raid5_conf(conf);
5709 if (rdev == p->rdev)
5711 else if (rdev == p->replacement)
5712 rdevp = &p->replacement;
5716 if (number >= conf->raid_disks &&
5717 conf->reshape_progress == MaxSector)
5718 clear_bit(In_sync, &rdev->flags);
5720 if (test_bit(In_sync, &rdev->flags) ||
5721 atomic_read(&rdev->nr_pending)) {
5725 /* Only remove non-faulty devices if recovery
5728 if (!test_bit(Faulty, &rdev->flags) &&
5729 mddev->recovery_disabled != conf->recovery_disabled &&
5730 !has_failed(conf) &&
5731 (!p->replacement || p->replacement == rdev) &&
5732 number < conf->raid_disks) {
5738 if (atomic_read(&rdev->nr_pending)) {
5739 /* lost the race, try later */
5742 } else if (p->replacement) {
5743 /* We must have just cleared 'rdev' */
5744 p->rdev = p->replacement;
5745 clear_bit(Replacement, &p->replacement->flags);
5746 smp_mb(); /* Make sure other CPUs may see both as identical
5747 * but will never see neither - if they are careful
5749 p->replacement = NULL;
5750 clear_bit(WantReplacement, &rdev->flags);
5752 /* We might have just removed the Replacement as faulty-
5753 * clear the bit just in case
5755 clear_bit(WantReplacement, &rdev->flags);
5758 print_raid5_conf(conf);
5762 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5764 struct r5conf *conf = mddev->private;
5767 struct disk_info *p;
5769 int last = conf->raid_disks - 1;
5771 if (mddev->recovery_disabled == conf->recovery_disabled)
5774 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5775 /* no point adding a device */
5778 if (rdev->raid_disk >= 0)
5779 first = last = rdev->raid_disk;
5782 * find the disk ... but prefer rdev->saved_raid_disk
5785 if (rdev->saved_raid_disk >= 0 &&
5786 rdev->saved_raid_disk >= first &&
5787 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5788 first = rdev->saved_raid_disk;
5790 for (disk = first; disk <= last; disk++) {
5791 p = conf->disks + disk;
5792 if (p->rdev == NULL) {
5793 clear_bit(In_sync, &rdev->flags);
5794 rdev->raid_disk = disk;
5796 if (rdev->saved_raid_disk != disk)
5798 rcu_assign_pointer(p->rdev, rdev);
5802 for (disk = first; disk <= last; disk++) {
5803 p = conf->disks + disk;
5804 if (test_bit(WantReplacement, &p->rdev->flags) &&
5805 p->replacement == NULL) {
5806 clear_bit(In_sync, &rdev->flags);
5807 set_bit(Replacement, &rdev->flags);
5808 rdev->raid_disk = disk;
5811 rcu_assign_pointer(p->replacement, rdev);
5816 print_raid5_conf(conf);
5820 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5822 /* no resync is happening, and there is enough space
5823 * on all devices, so we can resize.
5824 * We need to make sure resync covers any new space.
5825 * If the array is shrinking we should possibly wait until
5826 * any io in the removed space completes, but it hardly seems
5830 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5831 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5832 if (mddev->external_size &&
5833 mddev->array_sectors > newsize)
5835 if (mddev->bitmap) {
5836 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5840 md_set_array_sectors(mddev, newsize);
5841 set_capacity(mddev->gendisk, mddev->array_sectors);
5842 revalidate_disk(mddev->gendisk);
5843 if (sectors > mddev->dev_sectors &&
5844 mddev->recovery_cp > mddev->dev_sectors) {
5845 mddev->recovery_cp = mddev->dev_sectors;
5846 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5848 mddev->dev_sectors = sectors;
5849 mddev->resync_max_sectors = sectors;
5853 static int check_stripe_cache(struct mddev *mddev)
5855 /* Can only proceed if there are plenty of stripe_heads.
5856 * We need a minimum of one full stripe,, and for sensible progress
5857 * it is best to have about 4 times that.
5858 * If we require 4 times, then the default 256 4K stripe_heads will
5859 * allow for chunk sizes up to 256K, which is probably OK.
5860 * If the chunk size is greater, user-space should request more
5861 * stripe_heads first.
5863 struct r5conf *conf = mddev->private;
5864 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5865 > conf->max_nr_stripes ||
5866 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5867 > conf->max_nr_stripes) {
5868 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5870 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5877 static int check_reshape(struct mddev *mddev)
5879 struct r5conf *conf = mddev->private;
5881 if (mddev->delta_disks == 0 &&
5882 mddev->new_layout == mddev->layout &&
5883 mddev->new_chunk_sectors == mddev->chunk_sectors)
5884 return 0; /* nothing to do */
5885 if (has_failed(conf))
5887 if (mddev->delta_disks < 0) {
5888 /* We might be able to shrink, but the devices must
5889 * be made bigger first.
5890 * For raid6, 4 is the minimum size.
5891 * Otherwise 2 is the minimum
5894 if (mddev->level == 6)
5896 if (mddev->raid_disks + mddev->delta_disks < min)
5900 if (!check_stripe_cache(mddev))
5903 return resize_stripes(conf, (conf->previous_raid_disks
5904 + mddev->delta_disks));
5907 static int raid5_start_reshape(struct mddev *mddev)
5909 struct r5conf *conf = mddev->private;
5910 struct md_rdev *rdev;
5912 unsigned long flags;
5914 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5917 if (!check_stripe_cache(mddev))
5920 if (has_failed(conf))
5923 rdev_for_each(rdev, mddev) {
5924 if (!test_bit(In_sync, &rdev->flags)
5925 && !test_bit(Faulty, &rdev->flags))
5929 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5930 /* Not enough devices even to make a degraded array
5935 /* Refuse to reduce size of the array. Any reductions in
5936 * array size must be through explicit setting of array_size
5939 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5940 < mddev->array_sectors) {
5941 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5942 "before number of disks\n", mdname(mddev));
5946 atomic_set(&conf->reshape_stripes, 0);
5947 spin_lock_irq(&conf->device_lock);
5948 conf->previous_raid_disks = conf->raid_disks;
5949 conf->raid_disks += mddev->delta_disks;
5950 conf->prev_chunk_sectors = conf->chunk_sectors;
5951 conf->chunk_sectors = mddev->new_chunk_sectors;
5952 conf->prev_algo = conf->algorithm;
5953 conf->algorithm = mddev->new_layout;
5955 /* Code that selects data_offset needs to see the generation update
5956 * if reshape_progress has been set - so a memory barrier needed.
5959 if (mddev->reshape_backwards)
5960 conf->reshape_progress = raid5_size(mddev, 0, 0);
5962 conf->reshape_progress = 0;
5963 conf->reshape_safe = conf->reshape_progress;
5964 spin_unlock_irq(&conf->device_lock);
5966 /* Add some new drives, as many as will fit.
5967 * We know there are enough to make the newly sized array work.
5968 * Don't add devices if we are reducing the number of
5969 * devices in the array. This is because it is not possible
5970 * to correctly record the "partially reconstructed" state of
5971 * such devices during the reshape and confusion could result.
5973 if (mddev->delta_disks >= 0) {
5974 rdev_for_each(rdev, mddev)
5975 if (rdev->raid_disk < 0 &&
5976 !test_bit(Faulty, &rdev->flags)) {
5977 if (raid5_add_disk(mddev, rdev) == 0) {
5979 >= conf->previous_raid_disks)
5980 set_bit(In_sync, &rdev->flags);
5982 rdev->recovery_offset = 0;
5984 if (sysfs_link_rdev(mddev, rdev))
5985 /* Failure here is OK */;
5987 } else if (rdev->raid_disk >= conf->previous_raid_disks
5988 && !test_bit(Faulty, &rdev->flags)) {
5989 /* This is a spare that was manually added */
5990 set_bit(In_sync, &rdev->flags);
5993 /* When a reshape changes the number of devices,
5994 * ->degraded is measured against the larger of the
5995 * pre and post number of devices.
5997 spin_lock_irqsave(&conf->device_lock, flags);
5998 mddev->degraded = calc_degraded(conf);
5999 spin_unlock_irqrestore(&conf->device_lock, flags);
6001 mddev->raid_disks = conf->raid_disks;
6002 mddev->reshape_position = conf->reshape_progress;
6003 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6005 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6006 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6007 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6008 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6009 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6011 if (!mddev->sync_thread) {
6012 mddev->recovery = 0;
6013 spin_lock_irq(&conf->device_lock);
6014 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6015 rdev_for_each(rdev, mddev)
6016 rdev->new_data_offset = rdev->data_offset;
6018 conf->reshape_progress = MaxSector;
6019 mddev->reshape_position = MaxSector;
6020 spin_unlock_irq(&conf->device_lock);
6023 conf->reshape_checkpoint = jiffies;
6024 md_wakeup_thread(mddev->sync_thread);
6025 md_new_event(mddev);
6029 /* This is called from the reshape thread and should make any
6030 * changes needed in 'conf'
6032 static void end_reshape(struct r5conf *conf)
6035 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6036 struct md_rdev *rdev;
6038 spin_lock_irq(&conf->device_lock);
6039 conf->previous_raid_disks = conf->raid_disks;
6040 rdev_for_each(rdev, conf->mddev)
6041 rdev->data_offset = rdev->new_data_offset;
6043 conf->reshape_progress = MaxSector;
6044 spin_unlock_irq(&conf->device_lock);
6045 wake_up(&conf->wait_for_overlap);
6047 /* read-ahead size must cover two whole stripes, which is
6048 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6050 if (conf->mddev->queue) {
6051 int data_disks = conf->raid_disks - conf->max_degraded;
6052 int stripe = data_disks * ((conf->chunk_sectors << 9)
6054 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6055 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6060 /* This is called from the raid5d thread with mddev_lock held.
6061 * It makes config changes to the device.
6063 static void raid5_finish_reshape(struct mddev *mddev)
6065 struct r5conf *conf = mddev->private;
6067 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6069 if (mddev->delta_disks > 0) {
6070 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6071 set_capacity(mddev->gendisk, mddev->array_sectors);
6072 revalidate_disk(mddev->gendisk);
6075 spin_lock_irq(&conf->device_lock);
6076 mddev->degraded = calc_degraded(conf);
6077 spin_unlock_irq(&conf->device_lock);
6078 for (d = conf->raid_disks ;
6079 d < conf->raid_disks - mddev->delta_disks;
6081 struct md_rdev *rdev = conf->disks[d].rdev;
6083 clear_bit(In_sync, &rdev->flags);
6084 rdev = conf->disks[d].replacement;
6086 clear_bit(In_sync, &rdev->flags);
6089 mddev->layout = conf->algorithm;
6090 mddev->chunk_sectors = conf->chunk_sectors;
6091 mddev->reshape_position = MaxSector;
6092 mddev->delta_disks = 0;
6093 mddev->reshape_backwards = 0;
6097 static void raid5_quiesce(struct mddev *mddev, int state)
6099 struct r5conf *conf = mddev->private;
6102 case 2: /* resume for a suspend */
6103 wake_up(&conf->wait_for_overlap);
6106 case 1: /* stop all writes */
6107 spin_lock_irq(&conf->device_lock);
6108 /* '2' tells resync/reshape to pause so that all
6109 * active stripes can drain
6112 wait_event_lock_irq(conf->wait_for_stripe,
6113 atomic_read(&conf->active_stripes) == 0 &&
6114 atomic_read(&conf->active_aligned_reads) == 0,
6117 spin_unlock_irq(&conf->device_lock);
6118 /* allow reshape to continue */
6119 wake_up(&conf->wait_for_overlap);
6122 case 0: /* re-enable writes */
6123 spin_lock_irq(&conf->device_lock);
6125 wake_up(&conf->wait_for_stripe);
6126 wake_up(&conf->wait_for_overlap);
6127 spin_unlock_irq(&conf->device_lock);
6133 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6135 struct r0conf *raid0_conf = mddev->private;
6138 /* for raid0 takeover only one zone is supported */
6139 if (raid0_conf->nr_strip_zones > 1) {
6140 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6142 return ERR_PTR(-EINVAL);
6145 sectors = raid0_conf->strip_zone[0].zone_end;
6146 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6147 mddev->dev_sectors = sectors;
6148 mddev->new_level = level;
6149 mddev->new_layout = ALGORITHM_PARITY_N;
6150 mddev->new_chunk_sectors = mddev->chunk_sectors;
6151 mddev->raid_disks += 1;
6152 mddev->delta_disks = 1;
6153 /* make sure it will be not marked as dirty */
6154 mddev->recovery_cp = MaxSector;
6156 return setup_conf(mddev);
6160 static void *raid5_takeover_raid1(struct mddev *mddev)
6164 if (mddev->raid_disks != 2 ||
6165 mddev->degraded > 1)
6166 return ERR_PTR(-EINVAL);
6168 /* Should check if there are write-behind devices? */
6170 chunksect = 64*2; /* 64K by default */
6172 /* The array must be an exact multiple of chunksize */
6173 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6176 if ((chunksect<<9) < STRIPE_SIZE)
6177 /* array size does not allow a suitable chunk size */
6178 return ERR_PTR(-EINVAL);
6180 mddev->new_level = 5;
6181 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6182 mddev->new_chunk_sectors = chunksect;
6184 return setup_conf(mddev);
6187 static void *raid5_takeover_raid6(struct mddev *mddev)
6191 switch (mddev->layout) {
6192 case ALGORITHM_LEFT_ASYMMETRIC_6:
6193 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6195 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6196 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6198 case ALGORITHM_LEFT_SYMMETRIC_6:
6199 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6201 case ALGORITHM_RIGHT_SYMMETRIC_6:
6202 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6204 case ALGORITHM_PARITY_0_6:
6205 new_layout = ALGORITHM_PARITY_0;
6207 case ALGORITHM_PARITY_N:
6208 new_layout = ALGORITHM_PARITY_N;
6211 return ERR_PTR(-EINVAL);
6213 mddev->new_level = 5;
6214 mddev->new_layout = new_layout;
6215 mddev->delta_disks = -1;
6216 mddev->raid_disks -= 1;
6217 return setup_conf(mddev);
6221 static int raid5_check_reshape(struct mddev *mddev)
6223 /* For a 2-drive array, the layout and chunk size can be changed
6224 * immediately as not restriping is needed.
6225 * For larger arrays we record the new value - after validation
6226 * to be used by a reshape pass.
6228 struct r5conf *conf = mddev->private;
6229 int new_chunk = mddev->new_chunk_sectors;
6231 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6233 if (new_chunk > 0) {
6234 if (!is_power_of_2(new_chunk))
6236 if (new_chunk < (PAGE_SIZE>>9))
6238 if (mddev->array_sectors & (new_chunk-1))
6239 /* not factor of array size */
6243 /* They look valid */
6245 if (mddev->raid_disks == 2) {
6246 /* can make the change immediately */
6247 if (mddev->new_layout >= 0) {
6248 conf->algorithm = mddev->new_layout;
6249 mddev->layout = mddev->new_layout;
6251 if (new_chunk > 0) {
6252 conf->chunk_sectors = new_chunk ;
6253 mddev->chunk_sectors = new_chunk;
6255 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6256 md_wakeup_thread(mddev->thread);
6258 return check_reshape(mddev);
6261 static int raid6_check_reshape(struct mddev *mddev)
6263 int new_chunk = mddev->new_chunk_sectors;
6265 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6267 if (new_chunk > 0) {
6268 if (!is_power_of_2(new_chunk))
6270 if (new_chunk < (PAGE_SIZE >> 9))
6272 if (mddev->array_sectors & (new_chunk-1))
6273 /* not factor of array size */
6277 /* They look valid */
6278 return check_reshape(mddev);
6281 static void *raid5_takeover(struct mddev *mddev)
6283 /* raid5 can take over:
6284 * raid0 - if there is only one strip zone - make it a raid4 layout
6285 * raid1 - if there are two drives. We need to know the chunk size
6286 * raid4 - trivial - just use a raid4 layout.
6287 * raid6 - Providing it is a *_6 layout
6289 if (mddev->level == 0)
6290 return raid45_takeover_raid0(mddev, 5);
6291 if (mddev->level == 1)
6292 return raid5_takeover_raid1(mddev);
6293 if (mddev->level == 4) {
6294 mddev->new_layout = ALGORITHM_PARITY_N;
6295 mddev->new_level = 5;
6296 return setup_conf(mddev);
6298 if (mddev->level == 6)
6299 return raid5_takeover_raid6(mddev);
6301 return ERR_PTR(-EINVAL);
6304 static void *raid4_takeover(struct mddev *mddev)
6306 /* raid4 can take over:
6307 * raid0 - if there is only one strip zone
6308 * raid5 - if layout is right
6310 if (mddev->level == 0)
6311 return raid45_takeover_raid0(mddev, 4);
6312 if (mddev->level == 5 &&
6313 mddev->layout == ALGORITHM_PARITY_N) {
6314 mddev->new_layout = 0;
6315 mddev->new_level = 4;
6316 return setup_conf(mddev);
6318 return ERR_PTR(-EINVAL);
6321 static struct md_personality raid5_personality;
6323 static void *raid6_takeover(struct mddev *mddev)
6325 /* Currently can only take over a raid5. We map the
6326 * personality to an equivalent raid6 personality
6327 * with the Q block at the end.
6331 if (mddev->pers != &raid5_personality)
6332 return ERR_PTR(-EINVAL);
6333 if (mddev->degraded > 1)
6334 return ERR_PTR(-EINVAL);
6335 if (mddev->raid_disks > 253)
6336 return ERR_PTR(-EINVAL);
6337 if (mddev->raid_disks < 3)
6338 return ERR_PTR(-EINVAL);
6340 switch (mddev->layout) {
6341 case ALGORITHM_LEFT_ASYMMETRIC:
6342 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6344 case ALGORITHM_RIGHT_ASYMMETRIC:
6345 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6347 case ALGORITHM_LEFT_SYMMETRIC:
6348 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6350 case ALGORITHM_RIGHT_SYMMETRIC:
6351 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6353 case ALGORITHM_PARITY_0:
6354 new_layout = ALGORITHM_PARITY_0_6;
6356 case ALGORITHM_PARITY_N:
6357 new_layout = ALGORITHM_PARITY_N;
6360 return ERR_PTR(-EINVAL);
6362 mddev->new_level = 6;
6363 mddev->new_layout = new_layout;
6364 mddev->delta_disks = 1;
6365 mddev->raid_disks += 1;
6366 return setup_conf(mddev);
6370 static struct md_personality raid6_personality =
6374 .owner = THIS_MODULE,
6375 .make_request = make_request,
6379 .error_handler = error,
6380 .hot_add_disk = raid5_add_disk,
6381 .hot_remove_disk= raid5_remove_disk,
6382 .spare_active = raid5_spare_active,
6383 .sync_request = sync_request,
6384 .resize = raid5_resize,
6386 .check_reshape = raid6_check_reshape,
6387 .start_reshape = raid5_start_reshape,
6388 .finish_reshape = raid5_finish_reshape,
6389 .quiesce = raid5_quiesce,
6390 .takeover = raid6_takeover,
6392 static struct md_personality raid5_personality =
6396 .owner = THIS_MODULE,
6397 .make_request = make_request,
6401 .error_handler = error,
6402 .hot_add_disk = raid5_add_disk,
6403 .hot_remove_disk= raid5_remove_disk,
6404 .spare_active = raid5_spare_active,
6405 .sync_request = sync_request,
6406 .resize = raid5_resize,
6408 .check_reshape = raid5_check_reshape,
6409 .start_reshape = raid5_start_reshape,
6410 .finish_reshape = raid5_finish_reshape,
6411 .quiesce = raid5_quiesce,
6412 .takeover = raid5_takeover,
6415 static struct md_personality raid4_personality =
6419 .owner = THIS_MODULE,
6420 .make_request = make_request,
6424 .error_handler = error,
6425 .hot_add_disk = raid5_add_disk,
6426 .hot_remove_disk= raid5_remove_disk,
6427 .spare_active = raid5_spare_active,
6428 .sync_request = sync_request,
6429 .resize = raid5_resize,
6431 .check_reshape = raid5_check_reshape,
6432 .start_reshape = raid5_start_reshape,
6433 .finish_reshape = raid5_finish_reshape,
6434 .quiesce = raid5_quiesce,
6435 .takeover = raid4_takeover,
6438 static int __init raid5_init(void)
6440 register_md_personality(&raid6_personality);
6441 register_md_personality(&raid5_personality);
6442 register_md_personality(&raid4_personality);
6446 static void raid5_exit(void)
6448 unregister_md_personality(&raid6_personality);
6449 unregister_md_personality(&raid5_personality);
6450 unregister_md_personality(&raid4_personality);
6453 module_init(raid5_init);
6454 module_exit(raid5_exit);
6455 MODULE_LICENSE("GPL");
6456 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6457 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6458 MODULE_ALIAS("md-raid5");
6459 MODULE_ALIAS("md-raid4");
6460 MODULE_ALIAS("md-level-5");
6461 MODULE_ALIAS("md-level-4");
6462 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6463 MODULE_ALIAS("md-raid6");
6464 MODULE_ALIAS("md-level-6");
6466 /* This used to be two separate modules, they were: */
6467 MODULE_ALIAS("raid5");
6468 MODULE_ALIAS("raid6");