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;
187 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
194 static void print_raid5_conf (struct r5conf *conf);
196 static int stripe_operations_active(struct stripe_head *sh)
198 return sh->check_state || sh->reconstruct_state ||
199 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
200 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
203 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
205 BUG_ON(!list_empty(&sh->lru));
206 BUG_ON(atomic_read(&conf->active_stripes)==0);
207 if (test_bit(STRIPE_HANDLE, &sh->state)) {
208 if (test_bit(STRIPE_DELAYED, &sh->state) &&
209 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
210 list_add_tail(&sh->lru, &conf->delayed_list);
211 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
212 sh->bm_seq - conf->seq_write > 0)
213 list_add_tail(&sh->lru, &conf->bitmap_list);
215 clear_bit(STRIPE_DELAYED, &sh->state);
216 clear_bit(STRIPE_BIT_DELAY, &sh->state);
217 list_add_tail(&sh->lru, &conf->handle_list);
219 md_wakeup_thread(conf->mddev->thread);
221 BUG_ON(stripe_operations_active(sh));
222 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
223 if (atomic_dec_return(&conf->preread_active_stripes)
225 md_wakeup_thread(conf->mddev->thread);
226 atomic_dec(&conf->active_stripes);
227 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
228 list_add_tail(&sh->lru, &conf->inactive_list);
229 wake_up(&conf->wait_for_stripe);
230 if (conf->retry_read_aligned)
231 md_wakeup_thread(conf->mddev->thread);
236 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
238 if (atomic_dec_and_test(&sh->count))
239 do_release_stripe(conf, sh);
242 static void release_stripe(struct stripe_head *sh)
244 struct r5conf *conf = sh->raid_conf;
247 local_irq_save(flags);
248 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
249 do_release_stripe(conf, sh);
250 spin_unlock(&conf->device_lock);
252 local_irq_restore(flags);
255 static inline void remove_hash(struct stripe_head *sh)
257 pr_debug("remove_hash(), stripe %llu\n",
258 (unsigned long long)sh->sector);
260 hlist_del_init(&sh->hash);
263 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
265 struct hlist_head *hp = stripe_hash(conf, sh->sector);
267 pr_debug("insert_hash(), stripe %llu\n",
268 (unsigned long long)sh->sector);
270 hlist_add_head(&sh->hash, hp);
274 /* find an idle stripe, make sure it is unhashed, and return it. */
275 static struct stripe_head *get_free_stripe(struct r5conf *conf)
277 struct stripe_head *sh = NULL;
278 struct list_head *first;
280 if (list_empty(&conf->inactive_list))
282 first = conf->inactive_list.next;
283 sh = list_entry(first, struct stripe_head, lru);
284 list_del_init(first);
286 atomic_inc(&conf->active_stripes);
291 static void shrink_buffers(struct stripe_head *sh)
295 int num = sh->raid_conf->pool_size;
297 for (i = 0; i < num ; i++) {
301 sh->dev[i].page = NULL;
306 static int grow_buffers(struct stripe_head *sh)
309 int num = sh->raid_conf->pool_size;
311 for (i = 0; i < num; i++) {
314 if (!(page = alloc_page(GFP_KERNEL))) {
317 sh->dev[i].page = page;
322 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
323 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
324 struct stripe_head *sh);
326 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
328 struct r5conf *conf = sh->raid_conf;
331 BUG_ON(atomic_read(&sh->count) != 0);
332 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
333 BUG_ON(stripe_operations_active(sh));
335 pr_debug("init_stripe called, stripe %llu\n",
336 (unsigned long long)sh->sector);
340 sh->generation = conf->generation - previous;
341 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
343 stripe_set_idx(sector, conf, previous, sh);
347 for (i = sh->disks; i--; ) {
348 struct r5dev *dev = &sh->dev[i];
350 if (dev->toread || dev->read || dev->towrite || dev->written ||
351 test_bit(R5_LOCKED, &dev->flags)) {
352 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
353 (unsigned long long)sh->sector, i, dev->toread,
354 dev->read, dev->towrite, dev->written,
355 test_bit(R5_LOCKED, &dev->flags));
359 raid5_build_block(sh, i, previous);
361 insert_hash(conf, sh);
364 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
367 struct stripe_head *sh;
369 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
370 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
371 if (sh->sector == sector && sh->generation == generation)
373 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
378 * Need to check if array has failed when deciding whether to:
380 * - remove non-faulty devices
383 * This determination is simple when no reshape is happening.
384 * However if there is a reshape, we need to carefully check
385 * both the before and after sections.
386 * This is because some failed devices may only affect one
387 * of the two sections, and some non-in_sync devices may
388 * be insync in the section most affected by failed devices.
390 static int calc_degraded(struct r5conf *conf)
392 int degraded, degraded2;
397 for (i = 0; i < conf->previous_raid_disks; i++) {
398 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
399 if (rdev && test_bit(Faulty, &rdev->flags))
400 rdev = rcu_dereference(conf->disks[i].replacement);
401 if (!rdev || test_bit(Faulty, &rdev->flags))
403 else if (test_bit(In_sync, &rdev->flags))
406 /* not in-sync or faulty.
407 * If the reshape increases the number of devices,
408 * this is being recovered by the reshape, so
409 * this 'previous' section is not in_sync.
410 * If the number of devices is being reduced however,
411 * the device can only be part of the array if
412 * we are reverting a reshape, so this section will
415 if (conf->raid_disks >= conf->previous_raid_disks)
419 if (conf->raid_disks == conf->previous_raid_disks)
423 for (i = 0; i < conf->raid_disks; i++) {
424 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
425 if (rdev && test_bit(Faulty, &rdev->flags))
426 rdev = rcu_dereference(conf->disks[i].replacement);
427 if (!rdev || test_bit(Faulty, &rdev->flags))
429 else if (test_bit(In_sync, &rdev->flags))
432 /* not in-sync or faulty.
433 * If reshape increases the number of devices, this
434 * section has already been recovered, else it
435 * almost certainly hasn't.
437 if (conf->raid_disks <= conf->previous_raid_disks)
441 if (degraded2 > degraded)
446 static int has_failed(struct r5conf *conf)
450 if (conf->mddev->reshape_position == MaxSector)
451 return conf->mddev->degraded > conf->max_degraded;
453 degraded = calc_degraded(conf);
454 if (degraded > conf->max_degraded)
459 static struct stripe_head *
460 get_active_stripe(struct r5conf *conf, sector_t sector,
461 int previous, int noblock, int noquiesce)
463 struct stripe_head *sh;
465 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
467 spin_lock_irq(&conf->device_lock);
470 wait_event_lock_irq(conf->wait_for_stripe,
471 conf->quiesce == 0 || noquiesce,
473 sh = __find_stripe(conf, sector, conf->generation - previous);
475 if (!conf->inactive_blocked)
476 sh = get_free_stripe(conf);
477 if (noblock && sh == NULL)
480 conf->inactive_blocked = 1;
481 wait_event_lock_irq(conf->wait_for_stripe,
482 !list_empty(&conf->inactive_list) &&
483 (atomic_read(&conf->active_stripes)
484 < (conf->max_nr_stripes *3/4)
485 || !conf->inactive_blocked),
487 conf->inactive_blocked = 0;
489 init_stripe(sh, sector, previous);
491 if (atomic_read(&sh->count)) {
492 BUG_ON(!list_empty(&sh->lru)
493 && !test_bit(STRIPE_EXPANDING, &sh->state)
494 && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state));
496 if (!test_bit(STRIPE_HANDLE, &sh->state))
497 atomic_inc(&conf->active_stripes);
498 if (list_empty(&sh->lru) &&
499 !test_bit(STRIPE_EXPANDING, &sh->state))
501 list_del_init(&sh->lru);
504 } while (sh == NULL);
507 atomic_inc(&sh->count);
509 spin_unlock_irq(&conf->device_lock);
513 /* Determine if 'data_offset' or 'new_data_offset' should be used
514 * in this stripe_head.
516 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
518 sector_t progress = conf->reshape_progress;
519 /* Need a memory barrier to make sure we see the value
520 * of conf->generation, or ->data_offset that was set before
521 * reshape_progress was updated.
524 if (progress == MaxSector)
526 if (sh->generation == conf->generation - 1)
528 /* We are in a reshape, and this is a new-generation stripe,
529 * so use new_data_offset.
535 raid5_end_read_request(struct bio *bi, int error);
537 raid5_end_write_request(struct bio *bi, int error);
539 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
541 struct r5conf *conf = sh->raid_conf;
542 int i, disks = sh->disks;
546 for (i = disks; i--; ) {
548 int replace_only = 0;
549 struct bio *bi, *rbi;
550 struct md_rdev *rdev, *rrdev = NULL;
551 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
552 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
556 if (test_bit(R5_Discard, &sh->dev[i].flags))
558 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
560 else if (test_and_clear_bit(R5_WantReplace,
561 &sh->dev[i].flags)) {
566 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
569 bi = &sh->dev[i].req;
570 rbi = &sh->dev[i].rreq; /* For writing to replacement */
575 bi->bi_end_io = raid5_end_write_request;
576 rbi->bi_end_io = raid5_end_write_request;
578 bi->bi_end_io = raid5_end_read_request;
581 rrdev = rcu_dereference(conf->disks[i].replacement);
582 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
583 rdev = rcu_dereference(conf->disks[i].rdev);
592 /* We raced and saw duplicates */
595 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
600 if (rdev && test_bit(Faulty, &rdev->flags))
603 atomic_inc(&rdev->nr_pending);
604 if (rrdev && test_bit(Faulty, &rrdev->flags))
607 atomic_inc(&rrdev->nr_pending);
610 /* We have already checked bad blocks for reads. Now
611 * need to check for writes. We never accept write errors
612 * on the replacement, so we don't to check rrdev.
614 while ((rw & WRITE) && rdev &&
615 test_bit(WriteErrorSeen, &rdev->flags)) {
618 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
619 &first_bad, &bad_sectors);
624 set_bit(BlockedBadBlocks, &rdev->flags);
625 if (!conf->mddev->external &&
626 conf->mddev->flags) {
627 /* It is very unlikely, but we might
628 * still need to write out the
629 * bad block log - better give it
631 md_check_recovery(conf->mddev);
634 * Because md_wait_for_blocked_rdev
635 * will dec nr_pending, we must
636 * increment it first.
638 atomic_inc(&rdev->nr_pending);
639 md_wait_for_blocked_rdev(rdev, conf->mddev);
641 /* Acknowledged bad block - skip the write */
642 rdev_dec_pending(rdev, conf->mddev);
648 if (s->syncing || s->expanding || s->expanded
650 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
652 set_bit(STRIPE_IO_STARTED, &sh->state);
654 bi->bi_bdev = rdev->bdev;
655 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
656 __func__, (unsigned long long)sh->sector,
658 atomic_inc(&sh->count);
659 if (use_new_offset(conf, sh))
660 bi->bi_sector = (sh->sector
661 + rdev->new_data_offset);
663 bi->bi_sector = (sh->sector
664 + rdev->data_offset);
665 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
666 bi->bi_rw |= REQ_FLUSH;
668 bi->bi_flags = 1 << BIO_UPTODATE;
670 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
671 bi->bi_io_vec[0].bv_offset = 0;
672 bi->bi_size = STRIPE_SIZE;
675 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
676 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
677 bi, disk_devt(conf->mddev->gendisk),
679 generic_make_request(bi);
682 if (s->syncing || s->expanding || s->expanded
684 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
686 set_bit(STRIPE_IO_STARTED, &sh->state);
688 rbi->bi_bdev = rrdev->bdev;
689 pr_debug("%s: for %llu schedule op %ld on "
690 "replacement disc %d\n",
691 __func__, (unsigned long long)sh->sector,
693 atomic_inc(&sh->count);
694 if (use_new_offset(conf, sh))
695 rbi->bi_sector = (sh->sector
696 + rrdev->new_data_offset);
698 rbi->bi_sector = (sh->sector
699 + rrdev->data_offset);
700 rbi->bi_flags = 1 << BIO_UPTODATE;
702 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
703 rbi->bi_io_vec[0].bv_offset = 0;
704 rbi->bi_size = STRIPE_SIZE;
706 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
707 rbi, disk_devt(conf->mddev->gendisk),
709 generic_make_request(rbi);
711 if (!rdev && !rrdev) {
713 set_bit(STRIPE_DEGRADED, &sh->state);
714 pr_debug("skip op %ld on disc %d for sector %llu\n",
715 bi->bi_rw, i, (unsigned long long)sh->sector);
716 clear_bit(R5_LOCKED, &sh->dev[i].flags);
717 set_bit(STRIPE_HANDLE, &sh->state);
722 static struct dma_async_tx_descriptor *
723 async_copy_data(int frombio, struct bio *bio, struct page *page,
724 sector_t sector, struct dma_async_tx_descriptor *tx)
727 struct page *bio_page;
730 struct async_submit_ctl submit;
731 enum async_tx_flags flags = 0;
733 if (bio->bi_sector >= sector)
734 page_offset = (signed)(bio->bi_sector - sector) * 512;
736 page_offset = (signed)(sector - bio->bi_sector) * -512;
739 flags |= ASYNC_TX_FENCE;
740 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
742 bio_for_each_segment(bvl, bio, i) {
743 int len = bvl->bv_len;
747 if (page_offset < 0) {
748 b_offset = -page_offset;
749 page_offset += b_offset;
753 if (len > 0 && page_offset + len > STRIPE_SIZE)
754 clen = STRIPE_SIZE - page_offset;
759 b_offset += bvl->bv_offset;
760 bio_page = bvl->bv_page;
762 tx = async_memcpy(page, bio_page, page_offset,
763 b_offset, clen, &submit);
765 tx = async_memcpy(bio_page, page, b_offset,
766 page_offset, clen, &submit);
768 /* chain the operations */
769 submit.depend_tx = tx;
771 if (clen < len) /* hit end of page */
779 static void ops_complete_biofill(void *stripe_head_ref)
781 struct stripe_head *sh = stripe_head_ref;
782 struct bio *return_bi = NULL;
785 pr_debug("%s: stripe %llu\n", __func__,
786 (unsigned long long)sh->sector);
788 /* clear completed biofills */
789 for (i = sh->disks; i--; ) {
790 struct r5dev *dev = &sh->dev[i];
792 /* acknowledge completion of a biofill operation */
793 /* and check if we need to reply to a read request,
794 * new R5_Wantfill requests are held off until
795 * !STRIPE_BIOFILL_RUN
797 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
798 struct bio *rbi, *rbi2;
803 while (rbi && rbi->bi_sector <
804 dev->sector + STRIPE_SECTORS) {
805 rbi2 = r5_next_bio(rbi, dev->sector);
806 if (!raid5_dec_bi_active_stripes(rbi)) {
807 rbi->bi_next = return_bi;
814 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
816 return_io(return_bi);
818 set_bit(STRIPE_HANDLE, &sh->state);
822 static void ops_run_biofill(struct stripe_head *sh)
824 struct dma_async_tx_descriptor *tx = NULL;
825 struct async_submit_ctl submit;
828 pr_debug("%s: stripe %llu\n", __func__,
829 (unsigned long long)sh->sector);
831 for (i = sh->disks; i--; ) {
832 struct r5dev *dev = &sh->dev[i];
833 if (test_bit(R5_Wantfill, &dev->flags)) {
835 spin_lock_irq(&sh->stripe_lock);
836 dev->read = rbi = dev->toread;
838 spin_unlock_irq(&sh->stripe_lock);
839 while (rbi && rbi->bi_sector <
840 dev->sector + STRIPE_SECTORS) {
841 tx = async_copy_data(0, rbi, dev->page,
843 rbi = r5_next_bio(rbi, dev->sector);
848 atomic_inc(&sh->count);
849 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
850 async_trigger_callback(&submit);
853 static void mark_target_uptodate(struct stripe_head *sh, int target)
860 tgt = &sh->dev[target];
861 set_bit(R5_UPTODATE, &tgt->flags);
862 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
863 clear_bit(R5_Wantcompute, &tgt->flags);
866 static void ops_complete_compute(void *stripe_head_ref)
868 struct stripe_head *sh = stripe_head_ref;
870 pr_debug("%s: stripe %llu\n", __func__,
871 (unsigned long long)sh->sector);
873 /* mark the computed target(s) as uptodate */
874 mark_target_uptodate(sh, sh->ops.target);
875 mark_target_uptodate(sh, sh->ops.target2);
877 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
878 if (sh->check_state == check_state_compute_run)
879 sh->check_state = check_state_compute_result;
880 set_bit(STRIPE_HANDLE, &sh->state);
884 /* return a pointer to the address conversion region of the scribble buffer */
885 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
886 struct raid5_percpu *percpu)
888 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
891 static struct dma_async_tx_descriptor *
892 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
894 int disks = sh->disks;
895 struct page **xor_srcs = percpu->scribble;
896 int target = sh->ops.target;
897 struct r5dev *tgt = &sh->dev[target];
898 struct page *xor_dest = tgt->page;
900 struct dma_async_tx_descriptor *tx;
901 struct async_submit_ctl submit;
904 pr_debug("%s: stripe %llu block: %d\n",
905 __func__, (unsigned long long)sh->sector, target);
906 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
908 for (i = disks; i--; )
910 xor_srcs[count++] = sh->dev[i].page;
912 atomic_inc(&sh->count);
914 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
915 ops_complete_compute, sh, to_addr_conv(sh, percpu));
916 if (unlikely(count == 1))
917 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
919 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
924 /* set_syndrome_sources - populate source buffers for gen_syndrome
925 * @srcs - (struct page *) array of size sh->disks
926 * @sh - stripe_head to parse
928 * Populates srcs in proper layout order for the stripe and returns the
929 * 'count' of sources to be used in a call to async_gen_syndrome. The P
930 * destination buffer is recorded in srcs[count] and the Q destination
931 * is recorded in srcs[count+1]].
933 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
935 int disks = sh->disks;
936 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
937 int d0_idx = raid6_d0(sh);
941 for (i = 0; i < disks; i++)
947 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
949 srcs[slot] = sh->dev[i].page;
950 i = raid6_next_disk(i, disks);
951 } while (i != d0_idx);
953 return syndrome_disks;
956 static struct dma_async_tx_descriptor *
957 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
959 int disks = sh->disks;
960 struct page **blocks = percpu->scribble;
962 int qd_idx = sh->qd_idx;
963 struct dma_async_tx_descriptor *tx;
964 struct async_submit_ctl submit;
970 if (sh->ops.target < 0)
971 target = sh->ops.target2;
972 else if (sh->ops.target2 < 0)
973 target = sh->ops.target;
975 /* we should only have one valid target */
978 pr_debug("%s: stripe %llu block: %d\n",
979 __func__, (unsigned long long)sh->sector, target);
981 tgt = &sh->dev[target];
982 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
985 atomic_inc(&sh->count);
987 if (target == qd_idx) {
988 count = set_syndrome_sources(blocks, sh);
989 blocks[count] = NULL; /* regenerating p is not necessary */
990 BUG_ON(blocks[count+1] != dest); /* q should already be set */
991 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
992 ops_complete_compute, sh,
993 to_addr_conv(sh, percpu));
994 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
996 /* Compute any data- or p-drive using XOR */
998 for (i = disks; i-- ; ) {
999 if (i == target || i == qd_idx)
1001 blocks[count++] = sh->dev[i].page;
1004 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1005 NULL, ops_complete_compute, sh,
1006 to_addr_conv(sh, percpu));
1007 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1013 static struct dma_async_tx_descriptor *
1014 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1016 int i, count, disks = sh->disks;
1017 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1018 int d0_idx = raid6_d0(sh);
1019 int faila = -1, failb = -1;
1020 int target = sh->ops.target;
1021 int target2 = sh->ops.target2;
1022 struct r5dev *tgt = &sh->dev[target];
1023 struct r5dev *tgt2 = &sh->dev[target2];
1024 struct dma_async_tx_descriptor *tx;
1025 struct page **blocks = percpu->scribble;
1026 struct async_submit_ctl submit;
1028 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1029 __func__, (unsigned long long)sh->sector, target, target2);
1030 BUG_ON(target < 0 || target2 < 0);
1031 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1032 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1034 /* we need to open-code set_syndrome_sources to handle the
1035 * slot number conversion for 'faila' and 'failb'
1037 for (i = 0; i < disks ; i++)
1042 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1044 blocks[slot] = sh->dev[i].page;
1050 i = raid6_next_disk(i, disks);
1051 } while (i != d0_idx);
1053 BUG_ON(faila == failb);
1056 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1057 __func__, (unsigned long long)sh->sector, faila, failb);
1059 atomic_inc(&sh->count);
1061 if (failb == syndrome_disks+1) {
1062 /* Q disk is one of the missing disks */
1063 if (faila == syndrome_disks) {
1064 /* Missing P+Q, just recompute */
1065 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1066 ops_complete_compute, sh,
1067 to_addr_conv(sh, percpu));
1068 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1069 STRIPE_SIZE, &submit);
1073 int qd_idx = sh->qd_idx;
1075 /* Missing D+Q: recompute D from P, then recompute Q */
1076 if (target == qd_idx)
1077 data_target = target2;
1079 data_target = target;
1082 for (i = disks; i-- ; ) {
1083 if (i == data_target || i == qd_idx)
1085 blocks[count++] = sh->dev[i].page;
1087 dest = sh->dev[data_target].page;
1088 init_async_submit(&submit,
1089 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1091 to_addr_conv(sh, percpu));
1092 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1095 count = set_syndrome_sources(blocks, sh);
1096 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1097 ops_complete_compute, sh,
1098 to_addr_conv(sh, percpu));
1099 return async_gen_syndrome(blocks, 0, count+2,
1100 STRIPE_SIZE, &submit);
1103 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1104 ops_complete_compute, sh,
1105 to_addr_conv(sh, percpu));
1106 if (failb == syndrome_disks) {
1107 /* We're missing D+P. */
1108 return async_raid6_datap_recov(syndrome_disks+2,
1112 /* We're missing D+D. */
1113 return async_raid6_2data_recov(syndrome_disks+2,
1114 STRIPE_SIZE, faila, failb,
1121 static void ops_complete_prexor(void *stripe_head_ref)
1123 struct stripe_head *sh = stripe_head_ref;
1125 pr_debug("%s: stripe %llu\n", __func__,
1126 (unsigned long long)sh->sector);
1129 static struct dma_async_tx_descriptor *
1130 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1131 struct dma_async_tx_descriptor *tx)
1133 int disks = sh->disks;
1134 struct page **xor_srcs = percpu->scribble;
1135 int count = 0, pd_idx = sh->pd_idx, i;
1136 struct async_submit_ctl submit;
1138 /* existing parity data subtracted */
1139 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1141 pr_debug("%s: stripe %llu\n", __func__,
1142 (unsigned long long)sh->sector);
1144 for (i = disks; i--; ) {
1145 struct r5dev *dev = &sh->dev[i];
1146 /* Only process blocks that are known to be uptodate */
1147 if (test_bit(R5_Wantdrain, &dev->flags))
1148 xor_srcs[count++] = dev->page;
1151 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1152 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1153 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1158 static struct dma_async_tx_descriptor *
1159 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1161 int disks = sh->disks;
1164 pr_debug("%s: stripe %llu\n", __func__,
1165 (unsigned long long)sh->sector);
1167 for (i = disks; i--; ) {
1168 struct r5dev *dev = &sh->dev[i];
1171 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1174 spin_lock_irq(&sh->stripe_lock);
1175 chosen = dev->towrite;
1176 dev->towrite = NULL;
1177 BUG_ON(dev->written);
1178 wbi = dev->written = chosen;
1179 spin_unlock_irq(&sh->stripe_lock);
1181 while (wbi && wbi->bi_sector <
1182 dev->sector + STRIPE_SECTORS) {
1183 if (wbi->bi_rw & REQ_FUA)
1184 set_bit(R5_WantFUA, &dev->flags);
1185 if (wbi->bi_rw & REQ_SYNC)
1186 set_bit(R5_SyncIO, &dev->flags);
1187 if (wbi->bi_rw & REQ_DISCARD)
1188 set_bit(R5_Discard, &dev->flags);
1190 tx = async_copy_data(1, wbi, dev->page,
1192 wbi = r5_next_bio(wbi, dev->sector);
1200 static void ops_complete_reconstruct(void *stripe_head_ref)
1202 struct stripe_head *sh = stripe_head_ref;
1203 int disks = sh->disks;
1204 int pd_idx = sh->pd_idx;
1205 int qd_idx = sh->qd_idx;
1207 bool fua = false, sync = false, discard = false;
1209 pr_debug("%s: stripe %llu\n", __func__,
1210 (unsigned long long)sh->sector);
1212 for (i = disks; i--; ) {
1213 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1214 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1215 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1218 for (i = disks; i--; ) {
1219 struct r5dev *dev = &sh->dev[i];
1221 if (dev->written || i == pd_idx || i == qd_idx) {
1223 set_bit(R5_UPTODATE, &dev->flags);
1225 set_bit(R5_WantFUA, &dev->flags);
1227 set_bit(R5_SyncIO, &dev->flags);
1231 if (sh->reconstruct_state == reconstruct_state_drain_run)
1232 sh->reconstruct_state = reconstruct_state_drain_result;
1233 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1234 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1236 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1237 sh->reconstruct_state = reconstruct_state_result;
1240 set_bit(STRIPE_HANDLE, &sh->state);
1245 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1246 struct dma_async_tx_descriptor *tx)
1248 int disks = sh->disks;
1249 struct page **xor_srcs = percpu->scribble;
1250 struct async_submit_ctl submit;
1251 int count = 0, pd_idx = sh->pd_idx, i;
1252 struct page *xor_dest;
1254 unsigned long flags;
1256 pr_debug("%s: stripe %llu\n", __func__,
1257 (unsigned long long)sh->sector);
1259 for (i = 0; i < sh->disks; i++) {
1262 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1265 if (i >= sh->disks) {
1266 atomic_inc(&sh->count);
1267 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1268 ops_complete_reconstruct(sh);
1271 /* check if prexor is active which means only process blocks
1272 * that are part of a read-modify-write (written)
1274 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1276 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1277 for (i = disks; i--; ) {
1278 struct r5dev *dev = &sh->dev[i];
1280 xor_srcs[count++] = dev->page;
1283 xor_dest = sh->dev[pd_idx].page;
1284 for (i = disks; i--; ) {
1285 struct r5dev *dev = &sh->dev[i];
1287 xor_srcs[count++] = dev->page;
1291 /* 1/ if we prexor'd then the dest is reused as a source
1292 * 2/ if we did not prexor then we are redoing the parity
1293 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1294 * for the synchronous xor case
1296 flags = ASYNC_TX_ACK |
1297 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1299 atomic_inc(&sh->count);
1301 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1302 to_addr_conv(sh, percpu));
1303 if (unlikely(count == 1))
1304 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1306 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1310 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1311 struct dma_async_tx_descriptor *tx)
1313 struct async_submit_ctl submit;
1314 struct page **blocks = percpu->scribble;
1317 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1319 for (i = 0; i < sh->disks; i++) {
1320 if (sh->pd_idx == i || sh->qd_idx == i)
1322 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1325 if (i >= sh->disks) {
1326 atomic_inc(&sh->count);
1327 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1328 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1329 ops_complete_reconstruct(sh);
1333 count = set_syndrome_sources(blocks, sh);
1335 atomic_inc(&sh->count);
1337 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1338 sh, to_addr_conv(sh, percpu));
1339 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1342 static void ops_complete_check(void *stripe_head_ref)
1344 struct stripe_head *sh = stripe_head_ref;
1346 pr_debug("%s: stripe %llu\n", __func__,
1347 (unsigned long long)sh->sector);
1349 sh->check_state = check_state_check_result;
1350 set_bit(STRIPE_HANDLE, &sh->state);
1354 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1356 int disks = sh->disks;
1357 int pd_idx = sh->pd_idx;
1358 int qd_idx = sh->qd_idx;
1359 struct page *xor_dest;
1360 struct page **xor_srcs = percpu->scribble;
1361 struct dma_async_tx_descriptor *tx;
1362 struct async_submit_ctl submit;
1366 pr_debug("%s: stripe %llu\n", __func__,
1367 (unsigned long long)sh->sector);
1370 xor_dest = sh->dev[pd_idx].page;
1371 xor_srcs[count++] = xor_dest;
1372 for (i = disks; i--; ) {
1373 if (i == pd_idx || i == qd_idx)
1375 xor_srcs[count++] = sh->dev[i].page;
1378 init_async_submit(&submit, 0, NULL, NULL, NULL,
1379 to_addr_conv(sh, percpu));
1380 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1381 &sh->ops.zero_sum_result, &submit);
1383 atomic_inc(&sh->count);
1384 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1385 tx = async_trigger_callback(&submit);
1388 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1390 struct page **srcs = percpu->scribble;
1391 struct async_submit_ctl submit;
1394 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1395 (unsigned long long)sh->sector, checkp);
1397 count = set_syndrome_sources(srcs, sh);
1401 atomic_inc(&sh->count);
1402 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1403 sh, to_addr_conv(sh, percpu));
1404 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1405 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1408 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1410 int overlap_clear = 0, i, disks = sh->disks;
1411 struct dma_async_tx_descriptor *tx = NULL;
1412 struct r5conf *conf = sh->raid_conf;
1413 int level = conf->level;
1414 struct raid5_percpu *percpu;
1418 percpu = per_cpu_ptr(conf->percpu, cpu);
1419 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1420 ops_run_biofill(sh);
1424 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1426 tx = ops_run_compute5(sh, percpu);
1428 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1429 tx = ops_run_compute6_1(sh, percpu);
1431 tx = ops_run_compute6_2(sh, percpu);
1433 /* terminate the chain if reconstruct is not set to be run */
1434 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1438 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1439 tx = ops_run_prexor(sh, percpu, tx);
1441 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1442 tx = ops_run_biodrain(sh, tx);
1446 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1448 ops_run_reconstruct5(sh, percpu, tx);
1450 ops_run_reconstruct6(sh, percpu, tx);
1453 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1454 if (sh->check_state == check_state_run)
1455 ops_run_check_p(sh, percpu);
1456 else if (sh->check_state == check_state_run_q)
1457 ops_run_check_pq(sh, percpu, 0);
1458 else if (sh->check_state == check_state_run_pq)
1459 ops_run_check_pq(sh, percpu, 1);
1465 for (i = disks; i--; ) {
1466 struct r5dev *dev = &sh->dev[i];
1467 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1468 wake_up(&sh->raid_conf->wait_for_overlap);
1473 #ifdef CONFIG_MULTICORE_RAID456
1474 static void async_run_ops(void *param, async_cookie_t cookie)
1476 struct stripe_head *sh = param;
1477 unsigned long ops_request = sh->ops.request;
1479 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1480 wake_up(&sh->ops.wait_for_ops);
1482 __raid_run_ops(sh, ops_request);
1486 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1488 /* since handle_stripe can be called outside of raid5d context
1489 * we need to ensure sh->ops.request is de-staged before another
1492 wait_event(sh->ops.wait_for_ops,
1493 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1494 sh->ops.request = ops_request;
1496 atomic_inc(&sh->count);
1497 async_schedule(async_run_ops, sh);
1500 #define raid_run_ops __raid_run_ops
1503 static int grow_one_stripe(struct r5conf *conf)
1505 struct stripe_head *sh;
1506 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1510 sh->raid_conf = conf;
1511 #ifdef CONFIG_MULTICORE_RAID456
1512 init_waitqueue_head(&sh->ops.wait_for_ops);
1515 spin_lock_init(&sh->stripe_lock);
1517 if (grow_buffers(sh)) {
1519 kmem_cache_free(conf->slab_cache, sh);
1522 /* we just created an active stripe so... */
1523 atomic_set(&sh->count, 1);
1524 atomic_inc(&conf->active_stripes);
1525 INIT_LIST_HEAD(&sh->lru);
1530 static int grow_stripes(struct r5conf *conf, int num)
1532 struct kmem_cache *sc;
1533 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1535 if (conf->mddev->gendisk)
1536 sprintf(conf->cache_name[0],
1537 "raid%d-%s", conf->level, mdname(conf->mddev));
1539 sprintf(conf->cache_name[0],
1540 "raid%d-%p", conf->level, conf->mddev);
1541 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1543 conf->active_name = 0;
1544 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1545 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1549 conf->slab_cache = sc;
1550 conf->pool_size = devs;
1552 if (!grow_one_stripe(conf))
1558 * scribble_len - return the required size of the scribble region
1559 * @num - total number of disks in the array
1561 * The size must be enough to contain:
1562 * 1/ a struct page pointer for each device in the array +2
1563 * 2/ room to convert each entry in (1) to its corresponding dma
1564 * (dma_map_page()) or page (page_address()) address.
1566 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1567 * calculate over all devices (not just the data blocks), using zeros in place
1568 * of the P and Q blocks.
1570 static size_t scribble_len(int num)
1574 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1579 static int resize_stripes(struct r5conf *conf, int newsize)
1581 /* Make all the stripes able to hold 'newsize' devices.
1582 * New slots in each stripe get 'page' set to a new page.
1584 * This happens in stages:
1585 * 1/ create a new kmem_cache and allocate the required number of
1587 * 2/ gather all the old stripe_heads and transfer the pages across
1588 * to the new stripe_heads. This will have the side effect of
1589 * freezing the array as once all stripe_heads have been collected,
1590 * no IO will be possible. Old stripe heads are freed once their
1591 * pages have been transferred over, and the old kmem_cache is
1592 * freed when all stripes are done.
1593 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1594 * we simple return a failre status - no need to clean anything up.
1595 * 4/ allocate new pages for the new slots in the new stripe_heads.
1596 * If this fails, we don't bother trying the shrink the
1597 * stripe_heads down again, we just leave them as they are.
1598 * As each stripe_head is processed the new one is released into
1601 * Once step2 is started, we cannot afford to wait for a write,
1602 * so we use GFP_NOIO allocations.
1604 struct stripe_head *osh, *nsh;
1605 LIST_HEAD(newstripes);
1606 struct disk_info *ndisks;
1609 struct kmem_cache *sc;
1612 if (newsize <= conf->pool_size)
1613 return 0; /* never bother to shrink */
1615 err = md_allow_write(conf->mddev);
1620 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1621 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1626 for (i = conf->max_nr_stripes; i; i--) {
1627 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1631 nsh->raid_conf = conf;
1632 #ifdef CONFIG_MULTICORE_RAID456
1633 init_waitqueue_head(&nsh->ops.wait_for_ops);
1635 spin_lock_init(&nsh->stripe_lock);
1637 list_add(&nsh->lru, &newstripes);
1640 /* didn't get enough, give up */
1641 while (!list_empty(&newstripes)) {
1642 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1643 list_del(&nsh->lru);
1644 kmem_cache_free(sc, nsh);
1646 kmem_cache_destroy(sc);
1649 /* Step 2 - Must use GFP_NOIO now.
1650 * OK, we have enough stripes, start collecting inactive
1651 * stripes and copying them over
1653 list_for_each_entry(nsh, &newstripes, lru) {
1654 spin_lock_irq(&conf->device_lock);
1655 wait_event_lock_irq(conf->wait_for_stripe,
1656 !list_empty(&conf->inactive_list),
1658 osh = get_free_stripe(conf);
1659 spin_unlock_irq(&conf->device_lock);
1660 atomic_set(&nsh->count, 1);
1661 for(i=0; i<conf->pool_size; i++)
1662 nsh->dev[i].page = osh->dev[i].page;
1663 for( ; i<newsize; i++)
1664 nsh->dev[i].page = NULL;
1665 kmem_cache_free(conf->slab_cache, osh);
1667 kmem_cache_destroy(conf->slab_cache);
1670 * At this point, we are holding all the stripes so the array
1671 * is completely stalled, so now is a good time to resize
1672 * conf->disks and the scribble region
1674 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1676 for (i=0; i<conf->raid_disks; i++)
1677 ndisks[i] = conf->disks[i];
1679 conf->disks = ndisks;
1684 conf->scribble_len = scribble_len(newsize);
1685 for_each_present_cpu(cpu) {
1686 struct raid5_percpu *percpu;
1689 percpu = per_cpu_ptr(conf->percpu, cpu);
1690 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1693 kfree(percpu->scribble);
1694 percpu->scribble = scribble;
1702 /* Step 4, return new stripes to service */
1703 while(!list_empty(&newstripes)) {
1704 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1705 list_del_init(&nsh->lru);
1707 for (i=conf->raid_disks; i < newsize; i++)
1708 if (nsh->dev[i].page == NULL) {
1709 struct page *p = alloc_page(GFP_NOIO);
1710 nsh->dev[i].page = p;
1714 release_stripe(nsh);
1716 /* critical section pass, GFP_NOIO no longer needed */
1718 conf->slab_cache = sc;
1719 conf->active_name = 1-conf->active_name;
1720 conf->pool_size = newsize;
1724 static int drop_one_stripe(struct r5conf *conf)
1726 struct stripe_head *sh;
1728 spin_lock_irq(&conf->device_lock);
1729 sh = get_free_stripe(conf);
1730 spin_unlock_irq(&conf->device_lock);
1733 BUG_ON(atomic_read(&sh->count));
1735 kmem_cache_free(conf->slab_cache, sh);
1736 atomic_dec(&conf->active_stripes);
1740 static void shrink_stripes(struct r5conf *conf)
1742 while (drop_one_stripe(conf))
1745 if (conf->slab_cache)
1746 kmem_cache_destroy(conf->slab_cache);
1747 conf->slab_cache = NULL;
1750 static void raid5_end_read_request(struct bio * bi, int error)
1752 struct stripe_head *sh = bi->bi_private;
1753 struct r5conf *conf = sh->raid_conf;
1754 int disks = sh->disks, i;
1755 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1756 char b[BDEVNAME_SIZE];
1757 struct md_rdev *rdev = NULL;
1760 for (i=0 ; i<disks; i++)
1761 if (bi == &sh->dev[i].req)
1764 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1765 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1771 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1772 /* If replacement finished while this request was outstanding,
1773 * 'replacement' might be NULL already.
1774 * In that case it moved down to 'rdev'.
1775 * rdev is not removed until all requests are finished.
1777 rdev = conf->disks[i].replacement;
1779 rdev = conf->disks[i].rdev;
1781 if (use_new_offset(conf, sh))
1782 s = sh->sector + rdev->new_data_offset;
1784 s = sh->sector + rdev->data_offset;
1786 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1787 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1788 /* Note that this cannot happen on a
1789 * replacement device. We just fail those on
1794 "md/raid:%s: read error corrected"
1795 " (%lu sectors at %llu on %s)\n",
1796 mdname(conf->mddev), STRIPE_SECTORS,
1797 (unsigned long long)s,
1798 bdevname(rdev->bdev, b));
1799 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1800 clear_bit(R5_ReadError, &sh->dev[i].flags);
1801 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1802 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1803 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1805 if (atomic_read(&rdev->read_errors))
1806 atomic_set(&rdev->read_errors, 0);
1808 const char *bdn = bdevname(rdev->bdev, b);
1812 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1813 atomic_inc(&rdev->read_errors);
1814 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1817 "md/raid:%s: read error on replacement device "
1818 "(sector %llu on %s).\n",
1819 mdname(conf->mddev),
1820 (unsigned long long)s,
1822 else if (conf->mddev->degraded >= conf->max_degraded) {
1826 "md/raid:%s: read error not correctable "
1827 "(sector %llu on %s).\n",
1828 mdname(conf->mddev),
1829 (unsigned long long)s,
1831 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1836 "md/raid:%s: read error NOT corrected!! "
1837 "(sector %llu on %s).\n",
1838 mdname(conf->mddev),
1839 (unsigned long long)s,
1841 } else if (atomic_read(&rdev->read_errors)
1842 > conf->max_nr_stripes)
1844 "md/raid:%s: Too many read errors, failing device %s.\n",
1845 mdname(conf->mddev), bdn);
1849 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1850 set_bit(R5_ReadError, &sh->dev[i].flags);
1851 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1853 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1855 clear_bit(R5_ReadError, &sh->dev[i].flags);
1856 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1858 && test_bit(In_sync, &rdev->flags)
1859 && rdev_set_badblocks(
1860 rdev, sh->sector, STRIPE_SECTORS, 0)))
1861 md_error(conf->mddev, rdev);
1864 rdev_dec_pending(rdev, conf->mddev);
1865 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1866 set_bit(STRIPE_HANDLE, &sh->state);
1870 static void raid5_end_write_request(struct bio *bi, int error)
1872 struct stripe_head *sh = bi->bi_private;
1873 struct r5conf *conf = sh->raid_conf;
1874 int disks = sh->disks, i;
1875 struct md_rdev *uninitialized_var(rdev);
1876 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1879 int replacement = 0;
1881 for (i = 0 ; i < disks; i++) {
1882 if (bi == &sh->dev[i].req) {
1883 rdev = conf->disks[i].rdev;
1886 if (bi == &sh->dev[i].rreq) {
1887 rdev = conf->disks[i].replacement;
1891 /* rdev was removed and 'replacement'
1892 * replaced it. rdev is not removed
1893 * until all requests are finished.
1895 rdev = conf->disks[i].rdev;
1899 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1900 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1909 md_error(conf->mddev, rdev);
1910 else if (is_badblock(rdev, sh->sector,
1912 &first_bad, &bad_sectors))
1913 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1916 set_bit(WriteErrorSeen, &rdev->flags);
1917 set_bit(R5_WriteError, &sh->dev[i].flags);
1918 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1919 set_bit(MD_RECOVERY_NEEDED,
1920 &rdev->mddev->recovery);
1921 } else if (is_badblock(rdev, sh->sector,
1923 &first_bad, &bad_sectors))
1924 set_bit(R5_MadeGood, &sh->dev[i].flags);
1926 rdev_dec_pending(rdev, conf->mddev);
1928 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1929 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1930 set_bit(STRIPE_HANDLE, &sh->state);
1934 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1936 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1938 struct r5dev *dev = &sh->dev[i];
1940 bio_init(&dev->req);
1941 dev->req.bi_io_vec = &dev->vec;
1943 dev->req.bi_max_vecs++;
1944 dev->req.bi_private = sh;
1945 dev->vec.bv_page = dev->page;
1947 bio_init(&dev->rreq);
1948 dev->rreq.bi_io_vec = &dev->rvec;
1949 dev->rreq.bi_vcnt++;
1950 dev->rreq.bi_max_vecs++;
1951 dev->rreq.bi_private = sh;
1952 dev->rvec.bv_page = dev->page;
1955 dev->sector = compute_blocknr(sh, i, previous);
1958 static void error(struct mddev *mddev, struct md_rdev *rdev)
1960 char b[BDEVNAME_SIZE];
1961 struct r5conf *conf = mddev->private;
1962 unsigned long flags;
1963 pr_debug("raid456: error called\n");
1965 spin_lock_irqsave(&conf->device_lock, flags);
1966 clear_bit(In_sync, &rdev->flags);
1967 mddev->degraded = calc_degraded(conf);
1968 spin_unlock_irqrestore(&conf->device_lock, flags);
1969 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1971 set_bit(Blocked, &rdev->flags);
1972 set_bit(Faulty, &rdev->flags);
1973 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1975 "md/raid:%s: Disk failure on %s, disabling device.\n"
1976 "md/raid:%s: Operation continuing on %d devices.\n",
1978 bdevname(rdev->bdev, b),
1980 conf->raid_disks - mddev->degraded);
1984 * Input: a 'big' sector number,
1985 * Output: index of the data and parity disk, and the sector # in them.
1987 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1988 int previous, int *dd_idx,
1989 struct stripe_head *sh)
1991 sector_t stripe, stripe2;
1992 sector_t chunk_number;
1993 unsigned int chunk_offset;
1996 sector_t new_sector;
1997 int algorithm = previous ? conf->prev_algo
1999 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2000 : conf->chunk_sectors;
2001 int raid_disks = previous ? conf->previous_raid_disks
2003 int data_disks = raid_disks - conf->max_degraded;
2005 /* First compute the information on this sector */
2008 * Compute the chunk number and the sector offset inside the chunk
2010 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2011 chunk_number = r_sector;
2014 * Compute the stripe number
2016 stripe = chunk_number;
2017 *dd_idx = sector_div(stripe, data_disks);
2020 * Select the parity disk based on the user selected algorithm.
2022 pd_idx = qd_idx = -1;
2023 switch(conf->level) {
2025 pd_idx = data_disks;
2028 switch (algorithm) {
2029 case ALGORITHM_LEFT_ASYMMETRIC:
2030 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2031 if (*dd_idx >= pd_idx)
2034 case ALGORITHM_RIGHT_ASYMMETRIC:
2035 pd_idx = sector_div(stripe2, raid_disks);
2036 if (*dd_idx >= pd_idx)
2039 case ALGORITHM_LEFT_SYMMETRIC:
2040 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2041 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2043 case ALGORITHM_RIGHT_SYMMETRIC:
2044 pd_idx = sector_div(stripe2, raid_disks);
2045 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2047 case ALGORITHM_PARITY_0:
2051 case ALGORITHM_PARITY_N:
2052 pd_idx = data_disks;
2060 switch (algorithm) {
2061 case ALGORITHM_LEFT_ASYMMETRIC:
2062 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2063 qd_idx = pd_idx + 1;
2064 if (pd_idx == raid_disks-1) {
2065 (*dd_idx)++; /* Q D D D P */
2067 } else if (*dd_idx >= pd_idx)
2068 (*dd_idx) += 2; /* D D P Q D */
2070 case ALGORITHM_RIGHT_ASYMMETRIC:
2071 pd_idx = sector_div(stripe2, raid_disks);
2072 qd_idx = pd_idx + 1;
2073 if (pd_idx == raid_disks-1) {
2074 (*dd_idx)++; /* Q D D D P */
2076 } else if (*dd_idx >= pd_idx)
2077 (*dd_idx) += 2; /* D D P Q D */
2079 case ALGORITHM_LEFT_SYMMETRIC:
2080 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2081 qd_idx = (pd_idx + 1) % raid_disks;
2082 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2084 case ALGORITHM_RIGHT_SYMMETRIC:
2085 pd_idx = sector_div(stripe2, raid_disks);
2086 qd_idx = (pd_idx + 1) % raid_disks;
2087 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2090 case ALGORITHM_PARITY_0:
2095 case ALGORITHM_PARITY_N:
2096 pd_idx = data_disks;
2097 qd_idx = data_disks + 1;
2100 case ALGORITHM_ROTATING_ZERO_RESTART:
2101 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2102 * of blocks for computing Q is different.
2104 pd_idx = sector_div(stripe2, raid_disks);
2105 qd_idx = pd_idx + 1;
2106 if (pd_idx == raid_disks-1) {
2107 (*dd_idx)++; /* Q D D D P */
2109 } else if (*dd_idx >= pd_idx)
2110 (*dd_idx) += 2; /* D D P Q D */
2114 case ALGORITHM_ROTATING_N_RESTART:
2115 /* Same a left_asymmetric, by first stripe is
2116 * D D D P Q rather than
2120 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2121 qd_idx = pd_idx + 1;
2122 if (pd_idx == raid_disks-1) {
2123 (*dd_idx)++; /* Q D D D P */
2125 } else if (*dd_idx >= pd_idx)
2126 (*dd_idx) += 2; /* D D P Q D */
2130 case ALGORITHM_ROTATING_N_CONTINUE:
2131 /* Same as left_symmetric but Q is before P */
2132 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2133 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2134 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2138 case ALGORITHM_LEFT_ASYMMETRIC_6:
2139 /* RAID5 left_asymmetric, with Q on last device */
2140 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2141 if (*dd_idx >= pd_idx)
2143 qd_idx = raid_disks - 1;
2146 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2147 pd_idx = sector_div(stripe2, raid_disks-1);
2148 if (*dd_idx >= pd_idx)
2150 qd_idx = raid_disks - 1;
2153 case ALGORITHM_LEFT_SYMMETRIC_6:
2154 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2155 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2156 qd_idx = raid_disks - 1;
2159 case ALGORITHM_RIGHT_SYMMETRIC_6:
2160 pd_idx = sector_div(stripe2, raid_disks-1);
2161 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2162 qd_idx = raid_disks - 1;
2165 case ALGORITHM_PARITY_0_6:
2168 qd_idx = raid_disks - 1;
2178 sh->pd_idx = pd_idx;
2179 sh->qd_idx = qd_idx;
2180 sh->ddf_layout = ddf_layout;
2183 * Finally, compute the new sector number
2185 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2190 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2192 struct r5conf *conf = sh->raid_conf;
2193 int raid_disks = sh->disks;
2194 int data_disks = raid_disks - conf->max_degraded;
2195 sector_t new_sector = sh->sector, check;
2196 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2197 : conf->chunk_sectors;
2198 int algorithm = previous ? conf->prev_algo
2202 sector_t chunk_number;
2203 int dummy1, dd_idx = i;
2205 struct stripe_head sh2;
2208 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2209 stripe = new_sector;
2211 if (i == sh->pd_idx)
2213 switch(conf->level) {
2216 switch (algorithm) {
2217 case ALGORITHM_LEFT_ASYMMETRIC:
2218 case ALGORITHM_RIGHT_ASYMMETRIC:
2222 case ALGORITHM_LEFT_SYMMETRIC:
2223 case ALGORITHM_RIGHT_SYMMETRIC:
2226 i -= (sh->pd_idx + 1);
2228 case ALGORITHM_PARITY_0:
2231 case ALGORITHM_PARITY_N:
2238 if (i == sh->qd_idx)
2239 return 0; /* It is the Q disk */
2240 switch (algorithm) {
2241 case ALGORITHM_LEFT_ASYMMETRIC:
2242 case ALGORITHM_RIGHT_ASYMMETRIC:
2243 case ALGORITHM_ROTATING_ZERO_RESTART:
2244 case ALGORITHM_ROTATING_N_RESTART:
2245 if (sh->pd_idx == raid_disks-1)
2246 i--; /* Q D D D P */
2247 else if (i > sh->pd_idx)
2248 i -= 2; /* D D P Q D */
2250 case ALGORITHM_LEFT_SYMMETRIC:
2251 case ALGORITHM_RIGHT_SYMMETRIC:
2252 if (sh->pd_idx == raid_disks-1)
2253 i--; /* Q D D D P */
2258 i -= (sh->pd_idx + 2);
2261 case ALGORITHM_PARITY_0:
2264 case ALGORITHM_PARITY_N:
2266 case ALGORITHM_ROTATING_N_CONTINUE:
2267 /* Like left_symmetric, but P is before Q */
2268 if (sh->pd_idx == 0)
2269 i--; /* P D D D Q */
2274 i -= (sh->pd_idx + 1);
2277 case ALGORITHM_LEFT_ASYMMETRIC_6:
2278 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2282 case ALGORITHM_LEFT_SYMMETRIC_6:
2283 case ALGORITHM_RIGHT_SYMMETRIC_6:
2285 i += data_disks + 1;
2286 i -= (sh->pd_idx + 1);
2288 case ALGORITHM_PARITY_0_6:
2297 chunk_number = stripe * data_disks + i;
2298 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2300 check = raid5_compute_sector(conf, r_sector,
2301 previous, &dummy1, &sh2);
2302 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2303 || sh2.qd_idx != sh->qd_idx) {
2304 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2305 mdname(conf->mddev));
2313 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2314 int rcw, int expand)
2316 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2317 struct r5conf *conf = sh->raid_conf;
2318 int level = conf->level;
2321 /* if we are not expanding this is a proper write request, and
2322 * there will be bios with new data to be drained into the
2326 sh->reconstruct_state = reconstruct_state_drain_run;
2327 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2329 sh->reconstruct_state = reconstruct_state_run;
2331 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2333 for (i = disks; i--; ) {
2334 struct r5dev *dev = &sh->dev[i];
2337 set_bit(R5_LOCKED, &dev->flags);
2338 set_bit(R5_Wantdrain, &dev->flags);
2340 clear_bit(R5_UPTODATE, &dev->flags);
2344 if (s->locked + conf->max_degraded == disks)
2345 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2346 atomic_inc(&conf->pending_full_writes);
2349 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2350 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2352 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2353 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2354 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2355 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2357 for (i = disks; i--; ) {
2358 struct r5dev *dev = &sh->dev[i];
2363 (test_bit(R5_UPTODATE, &dev->flags) ||
2364 test_bit(R5_Wantcompute, &dev->flags))) {
2365 set_bit(R5_Wantdrain, &dev->flags);
2366 set_bit(R5_LOCKED, &dev->flags);
2367 clear_bit(R5_UPTODATE, &dev->flags);
2373 /* keep the parity disk(s) locked while asynchronous operations
2376 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2377 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2381 int qd_idx = sh->qd_idx;
2382 struct r5dev *dev = &sh->dev[qd_idx];
2384 set_bit(R5_LOCKED, &dev->flags);
2385 clear_bit(R5_UPTODATE, &dev->flags);
2389 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2390 __func__, (unsigned long long)sh->sector,
2391 s->locked, s->ops_request);
2395 * Each stripe/dev can have one or more bion attached.
2396 * toread/towrite point to the first in a chain.
2397 * The bi_next chain must be in order.
2399 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2402 struct r5conf *conf = sh->raid_conf;
2405 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2406 (unsigned long long)bi->bi_sector,
2407 (unsigned long long)sh->sector);
2410 * If several bio share a stripe. The bio bi_phys_segments acts as a
2411 * reference count to avoid race. The reference count should already be
2412 * increased before this function is called (for example, in
2413 * make_request()), so other bio sharing this stripe will not free the
2414 * stripe. If a stripe is owned by one stripe, the stripe lock will
2417 spin_lock_irq(&sh->stripe_lock);
2419 bip = &sh->dev[dd_idx].towrite;
2423 bip = &sh->dev[dd_idx].toread;
2424 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2425 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2427 bip = & (*bip)->bi_next;
2429 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2432 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2436 raid5_inc_bi_active_stripes(bi);
2439 /* check if page is covered */
2440 sector_t sector = sh->dev[dd_idx].sector;
2441 for (bi=sh->dev[dd_idx].towrite;
2442 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2443 bi && bi->bi_sector <= sector;
2444 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2445 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2446 sector = bi->bi_sector + (bi->bi_size>>9);
2448 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2449 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2452 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2453 (unsigned long long)(*bip)->bi_sector,
2454 (unsigned long long)sh->sector, dd_idx);
2455 spin_unlock_irq(&sh->stripe_lock);
2457 if (conf->mddev->bitmap && firstwrite) {
2458 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2460 sh->bm_seq = conf->seq_flush+1;
2461 set_bit(STRIPE_BIT_DELAY, &sh->state);
2466 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2467 spin_unlock_irq(&sh->stripe_lock);
2471 static void end_reshape(struct r5conf *conf);
2473 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2474 struct stripe_head *sh)
2476 int sectors_per_chunk =
2477 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2479 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2480 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2482 raid5_compute_sector(conf,
2483 stripe * (disks - conf->max_degraded)
2484 *sectors_per_chunk + chunk_offset,
2490 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2491 struct stripe_head_state *s, int disks,
2492 struct bio **return_bi)
2495 for (i = disks; i--; ) {
2499 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2500 struct md_rdev *rdev;
2502 rdev = rcu_dereference(conf->disks[i].rdev);
2503 if (rdev && test_bit(In_sync, &rdev->flags))
2504 atomic_inc(&rdev->nr_pending);
2509 if (!rdev_set_badblocks(
2513 md_error(conf->mddev, rdev);
2514 rdev_dec_pending(rdev, conf->mddev);
2517 spin_lock_irq(&sh->stripe_lock);
2518 /* fail all writes first */
2519 bi = sh->dev[i].towrite;
2520 sh->dev[i].towrite = NULL;
2521 spin_unlock_irq(&sh->stripe_lock);
2525 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2526 wake_up(&conf->wait_for_overlap);
2528 while (bi && bi->bi_sector <
2529 sh->dev[i].sector + STRIPE_SECTORS) {
2530 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2531 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2532 if (!raid5_dec_bi_active_stripes(bi)) {
2533 md_write_end(conf->mddev);
2534 bi->bi_next = *return_bi;
2540 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2541 STRIPE_SECTORS, 0, 0);
2543 /* and fail all 'written' */
2544 bi = sh->dev[i].written;
2545 sh->dev[i].written = NULL;
2546 if (bi) bitmap_end = 1;
2547 while (bi && bi->bi_sector <
2548 sh->dev[i].sector + STRIPE_SECTORS) {
2549 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2550 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2551 if (!raid5_dec_bi_active_stripes(bi)) {
2552 md_write_end(conf->mddev);
2553 bi->bi_next = *return_bi;
2559 /* fail any reads if this device is non-operational and
2560 * the data has not reached the cache yet.
2562 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2563 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2564 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2565 spin_lock_irq(&sh->stripe_lock);
2566 bi = sh->dev[i].toread;
2567 sh->dev[i].toread = NULL;
2568 spin_unlock_irq(&sh->stripe_lock);
2569 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2570 wake_up(&conf->wait_for_overlap);
2571 while (bi && bi->bi_sector <
2572 sh->dev[i].sector + STRIPE_SECTORS) {
2573 struct bio *nextbi =
2574 r5_next_bio(bi, sh->dev[i].sector);
2575 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2576 if (!raid5_dec_bi_active_stripes(bi)) {
2577 bi->bi_next = *return_bi;
2584 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2585 STRIPE_SECTORS, 0, 0);
2586 /* If we were in the middle of a write the parity block might
2587 * still be locked - so just clear all R5_LOCKED flags
2589 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2592 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2593 if (atomic_dec_and_test(&conf->pending_full_writes))
2594 md_wakeup_thread(conf->mddev->thread);
2598 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2599 struct stripe_head_state *s)
2604 clear_bit(STRIPE_SYNCING, &sh->state);
2607 /* There is nothing more to do for sync/check/repair.
2608 * Don't even need to abort as that is handled elsewhere
2609 * if needed, and not always wanted e.g. if there is a known
2611 * For recover/replace we need to record a bad block on all
2612 * non-sync devices, or abort the recovery
2614 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2615 /* During recovery devices cannot be removed, so
2616 * locking and refcounting of rdevs is not needed
2618 for (i = 0; i < conf->raid_disks; i++) {
2619 struct md_rdev *rdev = conf->disks[i].rdev;
2621 && !test_bit(Faulty, &rdev->flags)
2622 && !test_bit(In_sync, &rdev->flags)
2623 && !rdev_set_badblocks(rdev, sh->sector,
2626 rdev = conf->disks[i].replacement;
2628 && !test_bit(Faulty, &rdev->flags)
2629 && !test_bit(In_sync, &rdev->flags)
2630 && !rdev_set_badblocks(rdev, sh->sector,
2635 conf->recovery_disabled =
2636 conf->mddev->recovery_disabled;
2638 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2641 static int want_replace(struct stripe_head *sh, int disk_idx)
2643 struct md_rdev *rdev;
2645 /* Doing recovery so rcu locking not required */
2646 rdev = sh->raid_conf->disks[disk_idx].replacement;
2648 && !test_bit(Faulty, &rdev->flags)
2649 && !test_bit(In_sync, &rdev->flags)
2650 && (rdev->recovery_offset <= sh->sector
2651 || rdev->mddev->recovery_cp <= sh->sector))
2657 /* fetch_block - checks the given member device to see if its data needs
2658 * to be read or computed to satisfy a request.
2660 * Returns 1 when no more member devices need to be checked, otherwise returns
2661 * 0 to tell the loop in handle_stripe_fill to continue
2663 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2664 int disk_idx, int disks)
2666 struct r5dev *dev = &sh->dev[disk_idx];
2667 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2668 &sh->dev[s->failed_num[1]] };
2670 /* is the data in this block needed, and can we get it? */
2671 if (!test_bit(R5_LOCKED, &dev->flags) &&
2672 !test_bit(R5_UPTODATE, &dev->flags) &&
2674 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2675 s->syncing || s->expanding ||
2676 (s->replacing && want_replace(sh, disk_idx)) ||
2677 (s->failed >= 1 && fdev[0]->toread) ||
2678 (s->failed >= 2 && fdev[1]->toread) ||
2679 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2680 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2681 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2682 /* we would like to get this block, possibly by computing it,
2683 * otherwise read it if the backing disk is insync
2685 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2686 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2687 if ((s->uptodate == disks - 1) &&
2688 (s->failed && (disk_idx == s->failed_num[0] ||
2689 disk_idx == s->failed_num[1]))) {
2690 /* have disk failed, and we're requested to fetch it;
2693 pr_debug("Computing stripe %llu block %d\n",
2694 (unsigned long long)sh->sector, disk_idx);
2695 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2696 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2697 set_bit(R5_Wantcompute, &dev->flags);
2698 sh->ops.target = disk_idx;
2699 sh->ops.target2 = -1; /* no 2nd target */
2701 /* Careful: from this point on 'uptodate' is in the eye
2702 * of raid_run_ops which services 'compute' operations
2703 * before writes. R5_Wantcompute flags a block that will
2704 * be R5_UPTODATE by the time it is needed for a
2705 * subsequent operation.
2709 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2710 /* Computing 2-failure is *very* expensive; only
2711 * do it if failed >= 2
2714 for (other = disks; other--; ) {
2715 if (other == disk_idx)
2717 if (!test_bit(R5_UPTODATE,
2718 &sh->dev[other].flags))
2722 pr_debug("Computing stripe %llu blocks %d,%d\n",
2723 (unsigned long long)sh->sector,
2725 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2726 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2727 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2728 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2729 sh->ops.target = disk_idx;
2730 sh->ops.target2 = other;
2734 } else if (test_bit(R5_Insync, &dev->flags)) {
2735 set_bit(R5_LOCKED, &dev->flags);
2736 set_bit(R5_Wantread, &dev->flags);
2738 pr_debug("Reading block %d (sync=%d)\n",
2739 disk_idx, s->syncing);
2747 * handle_stripe_fill - read or compute data to satisfy pending requests.
2749 static void handle_stripe_fill(struct stripe_head *sh,
2750 struct stripe_head_state *s,
2755 /* look for blocks to read/compute, skip this if a compute
2756 * is already in flight, or if the stripe contents are in the
2757 * midst of changing due to a write
2759 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2760 !sh->reconstruct_state)
2761 for (i = disks; i--; )
2762 if (fetch_block(sh, s, i, disks))
2764 set_bit(STRIPE_HANDLE, &sh->state);
2768 /* handle_stripe_clean_event
2769 * any written block on an uptodate or failed drive can be returned.
2770 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2771 * never LOCKED, so we don't need to test 'failed' directly.
2773 static void handle_stripe_clean_event(struct r5conf *conf,
2774 struct stripe_head *sh, int disks, struct bio **return_bi)
2779 for (i = disks; i--; )
2780 if (sh->dev[i].written) {
2782 if (!test_bit(R5_LOCKED, &dev->flags) &&
2783 (test_bit(R5_UPTODATE, &dev->flags) ||
2784 test_bit(R5_Discard, &dev->flags))) {
2785 /* We can return any write requests */
2786 struct bio *wbi, *wbi2;
2787 pr_debug("Return write for disc %d\n", i);
2788 if (test_and_clear_bit(R5_Discard, &dev->flags))
2789 clear_bit(R5_UPTODATE, &dev->flags);
2791 dev->written = NULL;
2792 while (wbi && wbi->bi_sector <
2793 dev->sector + STRIPE_SECTORS) {
2794 wbi2 = r5_next_bio(wbi, dev->sector);
2795 if (!raid5_dec_bi_active_stripes(wbi)) {
2796 md_write_end(conf->mddev);
2797 wbi->bi_next = *return_bi;
2802 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2804 !test_bit(STRIPE_DEGRADED, &sh->state),
2807 } else if (test_bit(R5_Discard, &sh->dev[i].flags))
2808 clear_bit(R5_Discard, &sh->dev[i].flags);
2810 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2811 if (atomic_dec_and_test(&conf->pending_full_writes))
2812 md_wakeup_thread(conf->mddev->thread);
2815 static void handle_stripe_dirtying(struct r5conf *conf,
2816 struct stripe_head *sh,
2817 struct stripe_head_state *s,
2820 int rmw = 0, rcw = 0, i;
2821 sector_t recovery_cp = conf->mddev->recovery_cp;
2823 /* RAID6 requires 'rcw' in current implementation.
2824 * Otherwise, check whether resync is now happening or should start.
2825 * If yes, then the array is dirty (after unclean shutdown or
2826 * initial creation), so parity in some stripes might be inconsistent.
2827 * In this case, we need to always do reconstruct-write, to ensure
2828 * that in case of drive failure or read-error correction, we
2829 * generate correct data from the parity.
2831 if (conf->max_degraded == 2 ||
2832 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2833 /* Calculate the real rcw later - for now make it
2834 * look like rcw is cheaper
2837 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2838 conf->max_degraded, (unsigned long long)recovery_cp,
2839 (unsigned long long)sh->sector);
2840 } else for (i = disks; i--; ) {
2841 /* would I have to read this buffer for read_modify_write */
2842 struct r5dev *dev = &sh->dev[i];
2843 if ((dev->towrite || i == sh->pd_idx) &&
2844 !test_bit(R5_LOCKED, &dev->flags) &&
2845 !(test_bit(R5_UPTODATE, &dev->flags) ||
2846 test_bit(R5_Wantcompute, &dev->flags))) {
2847 if (test_bit(R5_Insync, &dev->flags))
2850 rmw += 2*disks; /* cannot read it */
2852 /* Would I have to read this buffer for reconstruct_write */
2853 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2854 !test_bit(R5_LOCKED, &dev->flags) &&
2855 !(test_bit(R5_UPTODATE, &dev->flags) ||
2856 test_bit(R5_Wantcompute, &dev->flags))) {
2857 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2862 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2863 (unsigned long long)sh->sector, rmw, rcw);
2864 set_bit(STRIPE_HANDLE, &sh->state);
2865 if (rmw < rcw && rmw > 0) {
2866 /* prefer read-modify-write, but need to get some data */
2867 blk_add_trace_msg(conf->mddev->queue, "raid5 rmw %llu %d",
2868 (unsigned long long)sh->sector, rmw);
2869 for (i = disks; i--; ) {
2870 struct r5dev *dev = &sh->dev[i];
2871 if ((dev->towrite || i == sh->pd_idx) &&
2872 !test_bit(R5_LOCKED, &dev->flags) &&
2873 !(test_bit(R5_UPTODATE, &dev->flags) ||
2874 test_bit(R5_Wantcompute, &dev->flags)) &&
2875 test_bit(R5_Insync, &dev->flags)) {
2877 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2878 pr_debug("Read_old block "
2879 "%d for r-m-w\n", i);
2880 set_bit(R5_LOCKED, &dev->flags);
2881 set_bit(R5_Wantread, &dev->flags);
2884 set_bit(STRIPE_DELAYED, &sh->state);
2885 set_bit(STRIPE_HANDLE, &sh->state);
2890 if (rcw <= rmw && rcw > 0) {
2891 /* want reconstruct write, but need to get some data */
2894 for (i = disks; i--; ) {
2895 struct r5dev *dev = &sh->dev[i];
2896 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2897 i != sh->pd_idx && i != sh->qd_idx &&
2898 !test_bit(R5_LOCKED, &dev->flags) &&
2899 !(test_bit(R5_UPTODATE, &dev->flags) ||
2900 test_bit(R5_Wantcompute, &dev->flags))) {
2902 if (!test_bit(R5_Insync, &dev->flags))
2903 continue; /* it's a failed drive */
2905 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2906 pr_debug("Read_old block "
2907 "%d for Reconstruct\n", i);
2908 set_bit(R5_LOCKED, &dev->flags);
2909 set_bit(R5_Wantread, &dev->flags);
2913 set_bit(STRIPE_DELAYED, &sh->state);
2914 set_bit(STRIPE_HANDLE, &sh->state);
2919 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
2920 (unsigned long long)sh->sector,
2921 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
2923 /* now if nothing is locked, and if we have enough data,
2924 * we can start a write request
2926 /* since handle_stripe can be called at any time we need to handle the
2927 * case where a compute block operation has been submitted and then a
2928 * subsequent call wants to start a write request. raid_run_ops only
2929 * handles the case where compute block and reconstruct are requested
2930 * simultaneously. If this is not the case then new writes need to be
2931 * held off until the compute completes.
2933 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2934 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2935 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2936 schedule_reconstruction(sh, s, rcw == 0, 0);
2939 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2940 struct stripe_head_state *s, int disks)
2942 struct r5dev *dev = NULL;
2944 set_bit(STRIPE_HANDLE, &sh->state);
2946 switch (sh->check_state) {
2947 case check_state_idle:
2948 /* start a new check operation if there are no failures */
2949 if (s->failed == 0) {
2950 BUG_ON(s->uptodate != disks);
2951 sh->check_state = check_state_run;
2952 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2953 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2957 dev = &sh->dev[s->failed_num[0]];
2959 case check_state_compute_result:
2960 sh->check_state = check_state_idle;
2962 dev = &sh->dev[sh->pd_idx];
2964 /* check that a write has not made the stripe insync */
2965 if (test_bit(STRIPE_INSYNC, &sh->state))
2968 /* either failed parity check, or recovery is happening */
2969 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2970 BUG_ON(s->uptodate != disks);
2972 set_bit(R5_LOCKED, &dev->flags);
2974 set_bit(R5_Wantwrite, &dev->flags);
2976 clear_bit(STRIPE_DEGRADED, &sh->state);
2977 set_bit(STRIPE_INSYNC, &sh->state);
2979 case check_state_run:
2980 break; /* we will be called again upon completion */
2981 case check_state_check_result:
2982 sh->check_state = check_state_idle;
2984 /* if a failure occurred during the check operation, leave
2985 * STRIPE_INSYNC not set and let the stripe be handled again
2990 /* handle a successful check operation, if parity is correct
2991 * we are done. Otherwise update the mismatch count and repair
2992 * parity if !MD_RECOVERY_CHECK
2994 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2995 /* parity is correct (on disc,
2996 * not in buffer any more)
2998 set_bit(STRIPE_INSYNC, &sh->state);
3000 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3001 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3002 /* don't try to repair!! */
3003 set_bit(STRIPE_INSYNC, &sh->state);
3005 sh->check_state = check_state_compute_run;
3006 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3007 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3008 set_bit(R5_Wantcompute,
3009 &sh->dev[sh->pd_idx].flags);
3010 sh->ops.target = sh->pd_idx;
3011 sh->ops.target2 = -1;
3016 case check_state_compute_run:
3019 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3020 __func__, sh->check_state,
3021 (unsigned long long) sh->sector);
3027 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3028 struct stripe_head_state *s,
3031 int pd_idx = sh->pd_idx;
3032 int qd_idx = sh->qd_idx;
3035 set_bit(STRIPE_HANDLE, &sh->state);
3037 BUG_ON(s->failed > 2);
3039 /* Want to check and possibly repair P and Q.
3040 * However there could be one 'failed' device, in which
3041 * case we can only check one of them, possibly using the
3042 * other to generate missing data
3045 switch (sh->check_state) {
3046 case check_state_idle:
3047 /* start a new check operation if there are < 2 failures */
3048 if (s->failed == s->q_failed) {
3049 /* The only possible failed device holds Q, so it
3050 * makes sense to check P (If anything else were failed,
3051 * we would have used P to recreate it).
3053 sh->check_state = check_state_run;
3055 if (!s->q_failed && s->failed < 2) {
3056 /* Q is not failed, and we didn't use it to generate
3057 * anything, so it makes sense to check it
3059 if (sh->check_state == check_state_run)
3060 sh->check_state = check_state_run_pq;
3062 sh->check_state = check_state_run_q;
3065 /* discard potentially stale zero_sum_result */
3066 sh->ops.zero_sum_result = 0;
3068 if (sh->check_state == check_state_run) {
3069 /* async_xor_zero_sum destroys the contents of P */
3070 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3073 if (sh->check_state >= check_state_run &&
3074 sh->check_state <= check_state_run_pq) {
3075 /* async_syndrome_zero_sum preserves P and Q, so
3076 * no need to mark them !uptodate here
3078 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3082 /* we have 2-disk failure */
3083 BUG_ON(s->failed != 2);
3085 case check_state_compute_result:
3086 sh->check_state = check_state_idle;
3088 /* check that a write has not made the stripe insync */
3089 if (test_bit(STRIPE_INSYNC, &sh->state))
3092 /* now write out any block on a failed drive,
3093 * or P or Q if they were recomputed
3095 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3096 if (s->failed == 2) {
3097 dev = &sh->dev[s->failed_num[1]];
3099 set_bit(R5_LOCKED, &dev->flags);
3100 set_bit(R5_Wantwrite, &dev->flags);
3102 if (s->failed >= 1) {
3103 dev = &sh->dev[s->failed_num[0]];
3105 set_bit(R5_LOCKED, &dev->flags);
3106 set_bit(R5_Wantwrite, &dev->flags);
3108 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3109 dev = &sh->dev[pd_idx];
3111 set_bit(R5_LOCKED, &dev->flags);
3112 set_bit(R5_Wantwrite, &dev->flags);
3114 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3115 dev = &sh->dev[qd_idx];
3117 set_bit(R5_LOCKED, &dev->flags);
3118 set_bit(R5_Wantwrite, &dev->flags);
3120 clear_bit(STRIPE_DEGRADED, &sh->state);
3122 set_bit(STRIPE_INSYNC, &sh->state);
3124 case check_state_run:
3125 case check_state_run_q:
3126 case check_state_run_pq:
3127 break; /* we will be called again upon completion */
3128 case check_state_check_result:
3129 sh->check_state = check_state_idle;
3131 /* handle a successful check operation, if parity is correct
3132 * we are done. Otherwise update the mismatch count and repair
3133 * parity if !MD_RECOVERY_CHECK
3135 if (sh->ops.zero_sum_result == 0) {
3136 /* both parities are correct */
3138 set_bit(STRIPE_INSYNC, &sh->state);
3140 /* in contrast to the raid5 case we can validate
3141 * parity, but still have a failure to write
3144 sh->check_state = check_state_compute_result;
3145 /* Returning at this point means that we may go
3146 * off and bring p and/or q uptodate again so
3147 * we make sure to check zero_sum_result again
3148 * to verify if p or q need writeback
3152 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3153 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3154 /* don't try to repair!! */
3155 set_bit(STRIPE_INSYNC, &sh->state);
3157 int *target = &sh->ops.target;
3159 sh->ops.target = -1;
3160 sh->ops.target2 = -1;
3161 sh->check_state = check_state_compute_run;
3162 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3163 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3164 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3165 set_bit(R5_Wantcompute,
3166 &sh->dev[pd_idx].flags);
3168 target = &sh->ops.target2;
3171 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3172 set_bit(R5_Wantcompute,
3173 &sh->dev[qd_idx].flags);
3180 case check_state_compute_run:
3183 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3184 __func__, sh->check_state,
3185 (unsigned long long) sh->sector);
3190 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3194 /* We have read all the blocks in this stripe and now we need to
3195 * copy some of them into a target stripe for expand.
3197 struct dma_async_tx_descriptor *tx = NULL;
3198 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3199 for (i = 0; i < sh->disks; i++)
3200 if (i != sh->pd_idx && i != sh->qd_idx) {
3202 struct stripe_head *sh2;
3203 struct async_submit_ctl submit;
3205 sector_t bn = compute_blocknr(sh, i, 1);
3206 sector_t s = raid5_compute_sector(conf, bn, 0,
3208 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3210 /* so far only the early blocks of this stripe
3211 * have been requested. When later blocks
3212 * get requested, we will try again
3215 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3216 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3217 /* must have already done this block */
3218 release_stripe(sh2);
3222 /* place all the copies on one channel */
3223 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3224 tx = async_memcpy(sh2->dev[dd_idx].page,
3225 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3228 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3229 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3230 for (j = 0; j < conf->raid_disks; j++)
3231 if (j != sh2->pd_idx &&
3233 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3235 if (j == conf->raid_disks) {
3236 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3237 set_bit(STRIPE_HANDLE, &sh2->state);
3239 release_stripe(sh2);
3242 /* done submitting copies, wait for them to complete */
3243 async_tx_quiesce(&tx);
3247 * handle_stripe - do things to a stripe.
3249 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3250 * state of various bits to see what needs to be done.
3252 * return some read requests which now have data
3253 * return some write requests which are safely on storage
3254 * schedule a read on some buffers
3255 * schedule a write of some buffers
3256 * return confirmation of parity correctness
3260 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3262 struct r5conf *conf = sh->raid_conf;
3263 int disks = sh->disks;
3266 int do_recovery = 0;
3268 memset(s, 0, sizeof(*s));
3270 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3271 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3272 s->failed_num[0] = -1;
3273 s->failed_num[1] = -1;
3275 /* Now to look around and see what can be done */
3277 for (i=disks; i--; ) {
3278 struct md_rdev *rdev;
3285 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3287 dev->toread, dev->towrite, dev->written);
3288 /* maybe we can reply to a read
3290 * new wantfill requests are only permitted while
3291 * ops_complete_biofill is guaranteed to be inactive
3293 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3294 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3295 set_bit(R5_Wantfill, &dev->flags);
3297 /* now count some things */
3298 if (test_bit(R5_LOCKED, &dev->flags))
3300 if (test_bit(R5_UPTODATE, &dev->flags))
3302 if (test_bit(R5_Wantcompute, &dev->flags)) {
3304 BUG_ON(s->compute > 2);
3307 if (test_bit(R5_Wantfill, &dev->flags))
3309 else if (dev->toread)
3313 if (!test_bit(R5_OVERWRITE, &dev->flags))
3318 /* Prefer to use the replacement for reads, but only
3319 * if it is recovered enough and has no bad blocks.
3321 rdev = rcu_dereference(conf->disks[i].replacement);
3322 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3323 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3324 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3325 &first_bad, &bad_sectors))
3326 set_bit(R5_ReadRepl, &dev->flags);
3329 set_bit(R5_NeedReplace, &dev->flags);
3330 rdev = rcu_dereference(conf->disks[i].rdev);
3331 clear_bit(R5_ReadRepl, &dev->flags);
3333 if (rdev && test_bit(Faulty, &rdev->flags))
3336 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3337 &first_bad, &bad_sectors);
3338 if (s->blocked_rdev == NULL
3339 && (test_bit(Blocked, &rdev->flags)
3342 set_bit(BlockedBadBlocks,
3344 s->blocked_rdev = rdev;
3345 atomic_inc(&rdev->nr_pending);
3348 clear_bit(R5_Insync, &dev->flags);
3352 /* also not in-sync */
3353 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3354 test_bit(R5_UPTODATE, &dev->flags)) {
3355 /* treat as in-sync, but with a read error
3356 * which we can now try to correct
3358 set_bit(R5_Insync, &dev->flags);
3359 set_bit(R5_ReadError, &dev->flags);
3361 } else if (test_bit(In_sync, &rdev->flags))
3362 set_bit(R5_Insync, &dev->flags);
3363 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3364 /* in sync if before recovery_offset */
3365 set_bit(R5_Insync, &dev->flags);
3366 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3367 test_bit(R5_Expanded, &dev->flags))
3368 /* If we've reshaped into here, we assume it is Insync.
3369 * We will shortly update recovery_offset to make
3372 set_bit(R5_Insync, &dev->flags);
3374 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3375 /* This flag does not apply to '.replacement'
3376 * only to .rdev, so make sure to check that*/
3377 struct md_rdev *rdev2 = rcu_dereference(
3378 conf->disks[i].rdev);
3380 clear_bit(R5_Insync, &dev->flags);
3381 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3382 s->handle_bad_blocks = 1;
3383 atomic_inc(&rdev2->nr_pending);
3385 clear_bit(R5_WriteError, &dev->flags);
3387 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3388 /* This flag does not apply to '.replacement'
3389 * only to .rdev, so make sure to check that*/
3390 struct md_rdev *rdev2 = rcu_dereference(
3391 conf->disks[i].rdev);
3392 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3393 s->handle_bad_blocks = 1;
3394 atomic_inc(&rdev2->nr_pending);
3396 clear_bit(R5_MadeGood, &dev->flags);
3398 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3399 struct md_rdev *rdev2 = rcu_dereference(
3400 conf->disks[i].replacement);
3401 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3402 s->handle_bad_blocks = 1;
3403 atomic_inc(&rdev2->nr_pending);
3405 clear_bit(R5_MadeGoodRepl, &dev->flags);
3407 if (!test_bit(R5_Insync, &dev->flags)) {
3408 /* The ReadError flag will just be confusing now */
3409 clear_bit(R5_ReadError, &dev->flags);
3410 clear_bit(R5_ReWrite, &dev->flags);
3412 if (test_bit(R5_ReadError, &dev->flags))
3413 clear_bit(R5_Insync, &dev->flags);
3414 if (!test_bit(R5_Insync, &dev->flags)) {
3416 s->failed_num[s->failed] = i;
3418 if (rdev && !test_bit(Faulty, &rdev->flags))
3422 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3423 /* If there is a failed device being replaced,
3424 * we must be recovering.
3425 * else if we are after recovery_cp, we must be syncing
3426 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3427 * else we can only be replacing
3428 * sync and recovery both need to read all devices, and so
3429 * use the same flag.
3432 sh->sector >= conf->mddev->recovery_cp ||
3433 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3441 static void handle_stripe(struct stripe_head *sh)
3443 struct stripe_head_state s;
3444 struct r5conf *conf = sh->raid_conf;
3447 int disks = sh->disks;
3448 struct r5dev *pdev, *qdev;
3450 clear_bit(STRIPE_HANDLE, &sh->state);
3451 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3452 /* already being handled, ensure it gets handled
3453 * again when current action finishes */
3454 set_bit(STRIPE_HANDLE, &sh->state);
3458 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3459 set_bit(STRIPE_SYNCING, &sh->state);
3460 clear_bit(STRIPE_INSYNC, &sh->state);
3462 clear_bit(STRIPE_DELAYED, &sh->state);
3464 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3465 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3466 (unsigned long long)sh->sector, sh->state,
3467 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3468 sh->check_state, sh->reconstruct_state);
3470 analyse_stripe(sh, &s);
3472 if (s.handle_bad_blocks) {
3473 set_bit(STRIPE_HANDLE, &sh->state);
3477 if (unlikely(s.blocked_rdev)) {
3478 if (s.syncing || s.expanding || s.expanded ||
3479 s.replacing || s.to_write || s.written) {
3480 set_bit(STRIPE_HANDLE, &sh->state);
3483 /* There is nothing for the blocked_rdev to block */
3484 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3485 s.blocked_rdev = NULL;
3488 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3489 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3490 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3493 pr_debug("locked=%d uptodate=%d to_read=%d"
3494 " to_write=%d failed=%d failed_num=%d,%d\n",
3495 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3496 s.failed_num[0], s.failed_num[1]);
3497 /* check if the array has lost more than max_degraded devices and,
3498 * if so, some requests might need to be failed.
3500 if (s.failed > conf->max_degraded) {
3501 sh->check_state = 0;
3502 sh->reconstruct_state = 0;
3503 if (s.to_read+s.to_write+s.written)
3504 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3505 if (s.syncing + s.replacing)
3506 handle_failed_sync(conf, sh, &s);
3509 /* Now we check to see if any write operations have recently
3513 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3515 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3516 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3517 sh->reconstruct_state = reconstruct_state_idle;
3519 /* All the 'written' buffers and the parity block are ready to
3520 * be written back to disk
3522 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3523 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3524 BUG_ON(sh->qd_idx >= 0 &&
3525 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3526 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3527 for (i = disks; i--; ) {
3528 struct r5dev *dev = &sh->dev[i];
3529 if (test_bit(R5_LOCKED, &dev->flags) &&
3530 (i == sh->pd_idx || i == sh->qd_idx ||
3532 pr_debug("Writing block %d\n", i);
3533 set_bit(R5_Wantwrite, &dev->flags);
3536 if (!test_bit(R5_Insync, &dev->flags) ||
3537 ((i == sh->pd_idx || i == sh->qd_idx) &&
3539 set_bit(STRIPE_INSYNC, &sh->state);
3542 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3543 s.dec_preread_active = 1;
3547 * might be able to return some write requests if the parity blocks
3548 * are safe, or on a failed drive
3550 pdev = &sh->dev[sh->pd_idx];
3551 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3552 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3553 qdev = &sh->dev[sh->qd_idx];
3554 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3555 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3559 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3560 && !test_bit(R5_LOCKED, &pdev->flags)
3561 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3562 test_bit(R5_Discard, &pdev->flags))))) &&
3563 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3564 && !test_bit(R5_LOCKED, &qdev->flags)
3565 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3566 test_bit(R5_Discard, &qdev->flags))))))
3567 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3569 /* Now we might consider reading some blocks, either to check/generate
3570 * parity, or to satisfy requests
3571 * or to load a block that is being partially written.
3573 if (s.to_read || s.non_overwrite
3574 || (conf->level == 6 && s.to_write && s.failed)
3575 || (s.syncing && (s.uptodate + s.compute < disks))
3578 handle_stripe_fill(sh, &s, disks);
3580 /* Now to consider new write requests and what else, if anything
3581 * should be read. We do not handle new writes when:
3582 * 1/ A 'write' operation (copy+xor) is already in flight.
3583 * 2/ A 'check' operation is in flight, as it may clobber the parity
3586 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3587 handle_stripe_dirtying(conf, sh, &s, disks);
3589 /* maybe we need to check and possibly fix the parity for this stripe
3590 * Any reads will already have been scheduled, so we just see if enough
3591 * data is available. The parity check is held off while parity
3592 * dependent operations are in flight.
3594 if (sh->check_state ||
3595 (s.syncing && s.locked == 0 &&
3596 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3597 !test_bit(STRIPE_INSYNC, &sh->state))) {
3598 if (conf->level == 6)
3599 handle_parity_checks6(conf, sh, &s, disks);
3601 handle_parity_checks5(conf, sh, &s, disks);
3604 if (s.replacing && s.locked == 0
3605 && !test_bit(STRIPE_INSYNC, &sh->state)) {
3606 /* Write out to replacement devices where possible */
3607 for (i = 0; i < conf->raid_disks; i++)
3608 if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3609 test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3610 set_bit(R5_WantReplace, &sh->dev[i].flags);
3611 set_bit(R5_LOCKED, &sh->dev[i].flags);
3614 set_bit(STRIPE_INSYNC, &sh->state);
3616 if ((s.syncing || s.replacing) && s.locked == 0 &&
3617 test_bit(STRIPE_INSYNC, &sh->state)) {
3618 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3619 clear_bit(STRIPE_SYNCING, &sh->state);
3622 /* If the failed drives are just a ReadError, then we might need
3623 * to progress the repair/check process
3625 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3626 for (i = 0; i < s.failed; i++) {
3627 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3628 if (test_bit(R5_ReadError, &dev->flags)
3629 && !test_bit(R5_LOCKED, &dev->flags)
3630 && test_bit(R5_UPTODATE, &dev->flags)
3632 if (!test_bit(R5_ReWrite, &dev->flags)) {
3633 set_bit(R5_Wantwrite, &dev->flags);
3634 set_bit(R5_ReWrite, &dev->flags);
3635 set_bit(R5_LOCKED, &dev->flags);
3638 /* let's read it back */
3639 set_bit(R5_Wantread, &dev->flags);
3640 set_bit(R5_LOCKED, &dev->flags);
3647 /* Finish reconstruct operations initiated by the expansion process */
3648 if (sh->reconstruct_state == reconstruct_state_result) {
3649 struct stripe_head *sh_src
3650 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3651 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3652 /* sh cannot be written until sh_src has been read.
3653 * so arrange for sh to be delayed a little
3655 set_bit(STRIPE_DELAYED, &sh->state);
3656 set_bit(STRIPE_HANDLE, &sh->state);
3657 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3659 atomic_inc(&conf->preread_active_stripes);
3660 release_stripe(sh_src);
3664 release_stripe(sh_src);
3666 sh->reconstruct_state = reconstruct_state_idle;
3667 clear_bit(STRIPE_EXPANDING, &sh->state);
3668 for (i = conf->raid_disks; i--; ) {
3669 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3670 set_bit(R5_LOCKED, &sh->dev[i].flags);
3675 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3676 !sh->reconstruct_state) {
3677 /* Need to write out all blocks after computing parity */
3678 sh->disks = conf->raid_disks;
3679 stripe_set_idx(sh->sector, conf, 0, sh);
3680 schedule_reconstruction(sh, &s, 1, 1);
3681 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3682 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3683 atomic_dec(&conf->reshape_stripes);
3684 wake_up(&conf->wait_for_overlap);
3685 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3688 if (s.expanding && s.locked == 0 &&
3689 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3690 handle_stripe_expansion(conf, sh);
3693 /* wait for this device to become unblocked */
3694 if (unlikely(s.blocked_rdev)) {
3695 if (conf->mddev->external)
3696 md_wait_for_blocked_rdev(s.blocked_rdev,
3699 /* Internal metadata will immediately
3700 * be written by raid5d, so we don't
3701 * need to wait here.
3703 rdev_dec_pending(s.blocked_rdev,
3707 if (s.handle_bad_blocks)
3708 for (i = disks; i--; ) {
3709 struct md_rdev *rdev;
3710 struct r5dev *dev = &sh->dev[i];
3711 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3712 /* We own a safe reference to the rdev */
3713 rdev = conf->disks[i].rdev;
3714 if (!rdev_set_badblocks(rdev, sh->sector,
3716 md_error(conf->mddev, rdev);
3717 rdev_dec_pending(rdev, conf->mddev);
3719 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3720 rdev = conf->disks[i].rdev;
3721 rdev_clear_badblocks(rdev, sh->sector,
3723 rdev_dec_pending(rdev, conf->mddev);
3725 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3726 rdev = conf->disks[i].replacement;
3728 /* rdev have been moved down */
3729 rdev = conf->disks[i].rdev;
3730 rdev_clear_badblocks(rdev, sh->sector,
3732 rdev_dec_pending(rdev, conf->mddev);
3737 raid_run_ops(sh, s.ops_request);
3741 if (s.dec_preread_active) {
3742 /* We delay this until after ops_run_io so that if make_request
3743 * is waiting on a flush, it won't continue until the writes
3744 * have actually been submitted.
3746 atomic_dec(&conf->preread_active_stripes);
3747 if (atomic_read(&conf->preread_active_stripes) <
3749 md_wakeup_thread(conf->mddev->thread);
3752 return_io(s.return_bi);
3754 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3757 static void raid5_activate_delayed(struct r5conf *conf)
3759 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3760 while (!list_empty(&conf->delayed_list)) {
3761 struct list_head *l = conf->delayed_list.next;
3762 struct stripe_head *sh;
3763 sh = list_entry(l, struct stripe_head, lru);
3765 clear_bit(STRIPE_DELAYED, &sh->state);
3766 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3767 atomic_inc(&conf->preread_active_stripes);
3768 list_add_tail(&sh->lru, &conf->hold_list);
3773 static void activate_bit_delay(struct r5conf *conf)
3775 /* device_lock is held */
3776 struct list_head head;
3777 list_add(&head, &conf->bitmap_list);
3778 list_del_init(&conf->bitmap_list);
3779 while (!list_empty(&head)) {
3780 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3781 list_del_init(&sh->lru);
3782 atomic_inc(&sh->count);
3783 __release_stripe(conf, sh);
3787 int md_raid5_congested(struct mddev *mddev, int bits)
3789 struct r5conf *conf = mddev->private;
3791 /* No difference between reads and writes. Just check
3792 * how busy the stripe_cache is
3795 if (conf->inactive_blocked)
3799 if (list_empty_careful(&conf->inactive_list))
3804 EXPORT_SYMBOL_GPL(md_raid5_congested);
3806 static int raid5_congested(void *data, int bits)
3808 struct mddev *mddev = data;
3810 return mddev_congested(mddev, bits) ||
3811 md_raid5_congested(mddev, bits);
3814 /* We want read requests to align with chunks where possible,
3815 * but write requests don't need to.
3817 static int raid5_mergeable_bvec(struct request_queue *q,
3818 struct bvec_merge_data *bvm,
3819 struct bio_vec *biovec)
3821 struct mddev *mddev = q->queuedata;
3822 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3824 unsigned int chunk_sectors = mddev->chunk_sectors;
3825 unsigned int bio_sectors = bvm->bi_size >> 9;
3827 if ((bvm->bi_rw & 1) == WRITE)
3828 return biovec->bv_len; /* always allow writes to be mergeable */
3830 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3831 chunk_sectors = mddev->new_chunk_sectors;
3832 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3833 if (max < 0) max = 0;
3834 if (max <= biovec->bv_len && bio_sectors == 0)
3835 return biovec->bv_len;
3841 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3843 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3844 unsigned int chunk_sectors = mddev->chunk_sectors;
3845 unsigned int bio_sectors = bio->bi_size >> 9;
3847 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3848 chunk_sectors = mddev->new_chunk_sectors;
3849 return chunk_sectors >=
3850 ((sector & (chunk_sectors - 1)) + bio_sectors);
3854 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3855 * later sampled by raid5d.
3857 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3859 unsigned long flags;
3861 spin_lock_irqsave(&conf->device_lock, flags);
3863 bi->bi_next = conf->retry_read_aligned_list;
3864 conf->retry_read_aligned_list = bi;
3866 spin_unlock_irqrestore(&conf->device_lock, flags);
3867 md_wakeup_thread(conf->mddev->thread);
3871 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3875 bi = conf->retry_read_aligned;
3877 conf->retry_read_aligned = NULL;
3880 bi = conf->retry_read_aligned_list;
3882 conf->retry_read_aligned_list = bi->bi_next;
3885 * this sets the active strip count to 1 and the processed
3886 * strip count to zero (upper 8 bits)
3888 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
3896 * The "raid5_align_endio" should check if the read succeeded and if it
3897 * did, call bio_endio on the original bio (having bio_put the new bio
3899 * If the read failed..
3901 static void raid5_align_endio(struct bio *bi, int error)
3903 struct bio* raid_bi = bi->bi_private;
3904 struct mddev *mddev;
3905 struct r5conf *conf;
3906 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3907 struct md_rdev *rdev;
3911 rdev = (void*)raid_bi->bi_next;
3912 raid_bi->bi_next = NULL;
3913 mddev = rdev->mddev;
3914 conf = mddev->private;
3916 rdev_dec_pending(rdev, conf->mddev);
3918 if (!error && uptodate) {
3919 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
3921 bio_endio(raid_bi, 0);
3922 if (atomic_dec_and_test(&conf->active_aligned_reads))
3923 wake_up(&conf->wait_for_stripe);
3928 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3930 add_bio_to_retry(raid_bi, conf);
3933 static int bio_fits_rdev(struct bio *bi)
3935 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3937 if ((bi->bi_size>>9) > queue_max_sectors(q))
3939 blk_recount_segments(q, bi);
3940 if (bi->bi_phys_segments > queue_max_segments(q))
3943 if (q->merge_bvec_fn)
3944 /* it's too hard to apply the merge_bvec_fn at this stage,
3953 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3955 struct r5conf *conf = mddev->private;
3957 struct bio* align_bi;
3958 struct md_rdev *rdev;
3959 sector_t end_sector;
3961 if (!in_chunk_boundary(mddev, raid_bio)) {
3962 pr_debug("chunk_aligned_read : non aligned\n");
3966 * use bio_clone_mddev to make a copy of the bio
3968 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3972 * set bi_end_io to a new function, and set bi_private to the
3975 align_bi->bi_end_io = raid5_align_endio;
3976 align_bi->bi_private = raid_bio;
3980 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3984 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3986 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3987 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3988 rdev->recovery_offset < end_sector) {
3989 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3991 (test_bit(Faulty, &rdev->flags) ||
3992 !(test_bit(In_sync, &rdev->flags) ||
3993 rdev->recovery_offset >= end_sector)))
4000 atomic_inc(&rdev->nr_pending);
4002 raid_bio->bi_next = (void*)rdev;
4003 align_bi->bi_bdev = rdev->bdev;
4004 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4006 if (!bio_fits_rdev(align_bi) ||
4007 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
4008 &first_bad, &bad_sectors)) {
4009 /* too big in some way, or has a known bad block */
4011 rdev_dec_pending(rdev, mddev);
4015 /* No reshape active, so we can trust rdev->data_offset */
4016 align_bi->bi_sector += rdev->data_offset;
4018 spin_lock_irq(&conf->device_lock);
4019 wait_event_lock_irq(conf->wait_for_stripe,
4022 atomic_inc(&conf->active_aligned_reads);
4023 spin_unlock_irq(&conf->device_lock);
4025 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4026 align_bi, disk_devt(mddev->gendisk),
4027 raid_bio->bi_sector);
4028 generic_make_request(align_bi);
4037 /* __get_priority_stripe - get the next stripe to process
4039 * Full stripe writes are allowed to pass preread active stripes up until
4040 * the bypass_threshold is exceeded. In general the bypass_count
4041 * increments when the handle_list is handled before the hold_list; however, it
4042 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4043 * stripe with in flight i/o. The bypass_count will be reset when the
4044 * head of the hold_list has changed, i.e. the head was promoted to the
4047 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
4049 struct stripe_head *sh;
4051 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4053 list_empty(&conf->handle_list) ? "empty" : "busy",
4054 list_empty(&conf->hold_list) ? "empty" : "busy",
4055 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4057 if (!list_empty(&conf->handle_list)) {
4058 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4060 if (list_empty(&conf->hold_list))
4061 conf->bypass_count = 0;
4062 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4063 if (conf->hold_list.next == conf->last_hold)
4064 conf->bypass_count++;
4066 conf->last_hold = conf->hold_list.next;
4067 conf->bypass_count -= conf->bypass_threshold;
4068 if (conf->bypass_count < 0)
4069 conf->bypass_count = 0;
4072 } else if (!list_empty(&conf->hold_list) &&
4073 ((conf->bypass_threshold &&
4074 conf->bypass_count > conf->bypass_threshold) ||
4075 atomic_read(&conf->pending_full_writes) == 0)) {
4076 sh = list_entry(conf->hold_list.next,
4078 conf->bypass_count -= conf->bypass_threshold;
4079 if (conf->bypass_count < 0)
4080 conf->bypass_count = 0;
4084 list_del_init(&sh->lru);
4085 atomic_inc(&sh->count);
4086 BUG_ON(atomic_read(&sh->count) != 1);
4090 struct raid5_plug_cb {
4091 struct blk_plug_cb cb;
4092 struct list_head list;
4095 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4097 struct raid5_plug_cb *cb = container_of(
4098 blk_cb, struct raid5_plug_cb, cb);
4099 struct stripe_head *sh;
4100 struct mddev *mddev = cb->cb.data;
4101 struct r5conf *conf = mddev->private;
4104 if (cb->list.next && !list_empty(&cb->list)) {
4105 spin_lock_irq(&conf->device_lock);
4106 while (!list_empty(&cb->list)) {
4107 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4108 list_del_init(&sh->lru);
4110 * avoid race release_stripe_plug() sees
4111 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4112 * is still in our list
4114 smp_mb__before_clear_bit();
4115 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4116 __release_stripe(conf, sh);
4119 spin_unlock_irq(&conf->device_lock);
4121 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4125 static void release_stripe_plug(struct mddev *mddev,
4126 struct stripe_head *sh)
4128 struct blk_plug_cb *blk_cb = blk_check_plugged(
4129 raid5_unplug, mddev,
4130 sizeof(struct raid5_plug_cb));
4131 struct raid5_plug_cb *cb;
4138 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4140 if (cb->list.next == NULL)
4141 INIT_LIST_HEAD(&cb->list);
4143 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4144 list_add_tail(&sh->lru, &cb->list);
4149 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4151 struct r5conf *conf = mddev->private;
4152 sector_t logical_sector, last_sector;
4153 struct stripe_head *sh;
4157 if (mddev->reshape_position != MaxSector)
4158 /* Skip discard while reshape is happening */
4161 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4162 last_sector = bi->bi_sector + (bi->bi_size>>9);
4165 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4167 stripe_sectors = conf->chunk_sectors *
4168 (conf->raid_disks - conf->max_degraded);
4169 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4171 sector_div(last_sector, stripe_sectors);
4173 logical_sector *= conf->chunk_sectors;
4174 last_sector *= conf->chunk_sectors;
4176 for (; logical_sector < last_sector;
4177 logical_sector += STRIPE_SECTORS) {
4181 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4182 prepare_to_wait(&conf->wait_for_overlap, &w,
4183 TASK_UNINTERRUPTIBLE);
4184 spin_lock_irq(&sh->stripe_lock);
4185 for (d = 0; d < conf->raid_disks; d++) {
4186 if (d == sh->pd_idx || d == sh->qd_idx)
4188 if (sh->dev[d].towrite || sh->dev[d].toread) {
4189 set_bit(R5_Overlap, &sh->dev[d].flags);
4190 spin_unlock_irq(&sh->stripe_lock);
4196 finish_wait(&conf->wait_for_overlap, &w);
4197 for (d = 0; d < conf->raid_disks; d++) {
4198 if (d == sh->pd_idx || d == sh->qd_idx)
4200 sh->dev[d].towrite = bi;
4201 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4202 raid5_inc_bi_active_stripes(bi);
4204 spin_unlock_irq(&sh->stripe_lock);
4205 if (conf->mddev->bitmap) {
4207 d < conf->raid_disks - conf->max_degraded;
4209 bitmap_startwrite(mddev->bitmap,
4213 sh->bm_seq = conf->seq_flush + 1;
4214 set_bit(STRIPE_BIT_DELAY, &sh->state);
4217 set_bit(STRIPE_HANDLE, &sh->state);
4218 clear_bit(STRIPE_DELAYED, &sh->state);
4219 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4220 atomic_inc(&conf->preread_active_stripes);
4221 release_stripe_plug(mddev, sh);
4224 remaining = raid5_dec_bi_active_stripes(bi);
4225 if (remaining == 0) {
4226 md_write_end(mddev);
4231 static void make_request(struct mddev *mddev, struct bio * bi)
4233 struct r5conf *conf = mddev->private;
4235 sector_t new_sector;
4236 sector_t logical_sector, last_sector;
4237 struct stripe_head *sh;
4238 const int rw = bio_data_dir(bi);
4241 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4242 md_flush_request(mddev, bi);
4246 md_write_start(mddev, bi);
4249 mddev->reshape_position == MaxSector &&
4250 chunk_aligned_read(mddev,bi))
4253 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4254 make_discard_request(mddev, bi);
4258 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4259 last_sector = bi->bi_sector + (bi->bi_size>>9);
4261 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4263 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4269 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4270 if (unlikely(conf->reshape_progress != MaxSector)) {
4271 /* spinlock is needed as reshape_progress may be
4272 * 64bit on a 32bit platform, and so it might be
4273 * possible to see a half-updated value
4274 * Of course reshape_progress could change after
4275 * the lock is dropped, so once we get a reference
4276 * to the stripe that we think it is, we will have
4279 spin_lock_irq(&conf->device_lock);
4280 if (mddev->reshape_backwards
4281 ? logical_sector < conf->reshape_progress
4282 : logical_sector >= conf->reshape_progress) {
4285 if (mddev->reshape_backwards
4286 ? logical_sector < conf->reshape_safe
4287 : logical_sector >= conf->reshape_safe) {
4288 spin_unlock_irq(&conf->device_lock);
4293 spin_unlock_irq(&conf->device_lock);
4296 new_sector = raid5_compute_sector(conf, logical_sector,
4299 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4300 (unsigned long long)new_sector,
4301 (unsigned long long)logical_sector);
4303 sh = get_active_stripe(conf, new_sector, previous,
4304 (bi->bi_rw&RWA_MASK), 0);
4306 if (unlikely(previous)) {
4307 /* expansion might have moved on while waiting for a
4308 * stripe, so we must do the range check again.
4309 * Expansion could still move past after this
4310 * test, but as we are holding a reference to
4311 * 'sh', we know that if that happens,
4312 * STRIPE_EXPANDING will get set and the expansion
4313 * won't proceed until we finish with the stripe.
4316 spin_lock_irq(&conf->device_lock);
4317 if (mddev->reshape_backwards
4318 ? logical_sector >= conf->reshape_progress
4319 : logical_sector < conf->reshape_progress)
4320 /* mismatch, need to try again */
4322 spin_unlock_irq(&conf->device_lock);
4331 logical_sector >= mddev->suspend_lo &&
4332 logical_sector < mddev->suspend_hi) {
4334 /* As the suspend_* range is controlled by
4335 * userspace, we want an interruptible
4338 flush_signals(current);
4339 prepare_to_wait(&conf->wait_for_overlap,
4340 &w, TASK_INTERRUPTIBLE);
4341 if (logical_sector >= mddev->suspend_lo &&
4342 logical_sector < mddev->suspend_hi)
4347 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4348 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4349 /* Stripe is busy expanding or
4350 * add failed due to overlap. Flush everything
4353 md_wakeup_thread(mddev->thread);
4358 finish_wait(&conf->wait_for_overlap, &w);
4359 set_bit(STRIPE_HANDLE, &sh->state);
4360 clear_bit(STRIPE_DELAYED, &sh->state);
4361 if ((bi->bi_rw & REQ_SYNC) &&
4362 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4363 atomic_inc(&conf->preread_active_stripes);
4364 release_stripe_plug(mddev, sh);
4366 /* cannot get stripe for read-ahead, just give-up */
4367 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4368 finish_wait(&conf->wait_for_overlap, &w);
4373 remaining = raid5_dec_bi_active_stripes(bi);
4374 if (remaining == 0) {
4377 md_write_end(mddev);
4379 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4385 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4387 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4389 /* reshaping is quite different to recovery/resync so it is
4390 * handled quite separately ... here.
4392 * On each call to sync_request, we gather one chunk worth of
4393 * destination stripes and flag them as expanding.
4394 * Then we find all the source stripes and request reads.
4395 * As the reads complete, handle_stripe will copy the data
4396 * into the destination stripe and release that stripe.
4398 struct r5conf *conf = mddev->private;
4399 struct stripe_head *sh;
4400 sector_t first_sector, last_sector;
4401 int raid_disks = conf->previous_raid_disks;
4402 int data_disks = raid_disks - conf->max_degraded;
4403 int new_data_disks = conf->raid_disks - conf->max_degraded;
4406 sector_t writepos, readpos, safepos;
4407 sector_t stripe_addr;
4408 int reshape_sectors;
4409 struct list_head stripes;
4411 if (sector_nr == 0) {
4412 /* If restarting in the middle, skip the initial sectors */
4413 if (mddev->reshape_backwards &&
4414 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4415 sector_nr = raid5_size(mddev, 0, 0)
4416 - conf->reshape_progress;
4417 } else if (!mddev->reshape_backwards &&
4418 conf->reshape_progress > 0)
4419 sector_nr = conf->reshape_progress;
4420 sector_div(sector_nr, new_data_disks);
4422 mddev->curr_resync_completed = sector_nr;
4423 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4429 /* We need to process a full chunk at a time.
4430 * If old and new chunk sizes differ, we need to process the
4433 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4434 reshape_sectors = mddev->new_chunk_sectors;
4436 reshape_sectors = mddev->chunk_sectors;
4438 /* We update the metadata at least every 10 seconds, or when
4439 * the data about to be copied would over-write the source of
4440 * the data at the front of the range. i.e. one new_stripe
4441 * along from reshape_progress new_maps to after where
4442 * reshape_safe old_maps to
4444 writepos = conf->reshape_progress;
4445 sector_div(writepos, new_data_disks);
4446 readpos = conf->reshape_progress;
4447 sector_div(readpos, data_disks);
4448 safepos = conf->reshape_safe;
4449 sector_div(safepos, data_disks);
4450 if (mddev->reshape_backwards) {
4451 writepos -= min_t(sector_t, reshape_sectors, writepos);
4452 readpos += reshape_sectors;
4453 safepos += reshape_sectors;
4455 writepos += reshape_sectors;
4456 readpos -= min_t(sector_t, reshape_sectors, readpos);
4457 safepos -= min_t(sector_t, reshape_sectors, safepos);
4460 /* Having calculated the 'writepos' possibly use it
4461 * to set 'stripe_addr' which is where we will write to.
4463 if (mddev->reshape_backwards) {
4464 BUG_ON(conf->reshape_progress == 0);
4465 stripe_addr = writepos;
4466 BUG_ON((mddev->dev_sectors &
4467 ~((sector_t)reshape_sectors - 1))
4468 - reshape_sectors - stripe_addr
4471 BUG_ON(writepos != sector_nr + reshape_sectors);
4472 stripe_addr = sector_nr;
4475 /* 'writepos' is the most advanced device address we might write.
4476 * 'readpos' is the least advanced device address we might read.
4477 * 'safepos' is the least address recorded in the metadata as having
4479 * If there is a min_offset_diff, these are adjusted either by
4480 * increasing the safepos/readpos if diff is negative, or
4481 * increasing writepos if diff is positive.
4482 * If 'readpos' is then behind 'writepos', there is no way that we can
4483 * ensure safety in the face of a crash - that must be done by userspace
4484 * making a backup of the data. So in that case there is no particular
4485 * rush to update metadata.
4486 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4487 * update the metadata to advance 'safepos' to match 'readpos' so that
4488 * we can be safe in the event of a crash.
4489 * So we insist on updating metadata if safepos is behind writepos and
4490 * readpos is beyond writepos.
4491 * In any case, update the metadata every 10 seconds.
4492 * Maybe that number should be configurable, but I'm not sure it is
4493 * worth it.... maybe it could be a multiple of safemode_delay???
4495 if (conf->min_offset_diff < 0) {
4496 safepos += -conf->min_offset_diff;
4497 readpos += -conf->min_offset_diff;
4499 writepos += conf->min_offset_diff;
4501 if ((mddev->reshape_backwards
4502 ? (safepos > writepos && readpos < writepos)
4503 : (safepos < writepos && readpos > writepos)) ||
4504 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4505 /* Cannot proceed until we've updated the superblock... */
4506 wait_event(conf->wait_for_overlap,
4507 atomic_read(&conf->reshape_stripes)==0);
4508 mddev->reshape_position = conf->reshape_progress;
4509 mddev->curr_resync_completed = sector_nr;
4510 conf->reshape_checkpoint = jiffies;
4511 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4512 md_wakeup_thread(mddev->thread);
4513 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4514 kthread_should_stop());
4515 spin_lock_irq(&conf->device_lock);
4516 conf->reshape_safe = mddev->reshape_position;
4517 spin_unlock_irq(&conf->device_lock);
4518 wake_up(&conf->wait_for_overlap);
4519 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4522 INIT_LIST_HEAD(&stripes);
4523 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4525 int skipped_disk = 0;
4526 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4527 set_bit(STRIPE_EXPANDING, &sh->state);
4528 atomic_inc(&conf->reshape_stripes);
4529 /* If any of this stripe is beyond the end of the old
4530 * array, then we need to zero those blocks
4532 for (j=sh->disks; j--;) {
4534 if (j == sh->pd_idx)
4536 if (conf->level == 6 &&
4539 s = compute_blocknr(sh, j, 0);
4540 if (s < raid5_size(mddev, 0, 0)) {
4544 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4545 set_bit(R5_Expanded, &sh->dev[j].flags);
4546 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4548 if (!skipped_disk) {
4549 set_bit(STRIPE_EXPAND_READY, &sh->state);
4550 set_bit(STRIPE_HANDLE, &sh->state);
4552 list_add(&sh->lru, &stripes);
4554 spin_lock_irq(&conf->device_lock);
4555 if (mddev->reshape_backwards)
4556 conf->reshape_progress -= reshape_sectors * new_data_disks;
4558 conf->reshape_progress += reshape_sectors * new_data_disks;
4559 spin_unlock_irq(&conf->device_lock);
4560 /* Ok, those stripe are ready. We can start scheduling
4561 * reads on the source stripes.
4562 * The source stripes are determined by mapping the first and last
4563 * block on the destination stripes.
4566 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4569 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4570 * new_data_disks - 1),
4572 if (last_sector >= mddev->dev_sectors)
4573 last_sector = mddev->dev_sectors - 1;
4574 while (first_sector <= last_sector) {
4575 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4576 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4577 set_bit(STRIPE_HANDLE, &sh->state);
4579 first_sector += STRIPE_SECTORS;
4581 /* Now that the sources are clearly marked, we can release
4582 * the destination stripes
4584 while (!list_empty(&stripes)) {
4585 sh = list_entry(stripes.next, struct stripe_head, lru);
4586 list_del_init(&sh->lru);
4589 /* If this takes us to the resync_max point where we have to pause,
4590 * then we need to write out the superblock.
4592 sector_nr += reshape_sectors;
4593 if ((sector_nr - mddev->curr_resync_completed) * 2
4594 >= mddev->resync_max - mddev->curr_resync_completed) {
4595 /* Cannot proceed until we've updated the superblock... */
4596 wait_event(conf->wait_for_overlap,
4597 atomic_read(&conf->reshape_stripes) == 0);
4598 mddev->reshape_position = conf->reshape_progress;
4599 mddev->curr_resync_completed = sector_nr;
4600 conf->reshape_checkpoint = jiffies;
4601 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4602 md_wakeup_thread(mddev->thread);
4603 wait_event(mddev->sb_wait,
4604 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4605 || kthread_should_stop());
4606 spin_lock_irq(&conf->device_lock);
4607 conf->reshape_safe = mddev->reshape_position;
4608 spin_unlock_irq(&conf->device_lock);
4609 wake_up(&conf->wait_for_overlap);
4610 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4612 return reshape_sectors;
4615 /* FIXME go_faster isn't used */
4616 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4618 struct r5conf *conf = mddev->private;
4619 struct stripe_head *sh;
4620 sector_t max_sector = mddev->dev_sectors;
4621 sector_t sync_blocks;
4622 int still_degraded = 0;
4625 if (sector_nr >= max_sector) {
4626 /* just being told to finish up .. nothing much to do */
4628 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4633 if (mddev->curr_resync < max_sector) /* aborted */
4634 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4636 else /* completed sync */
4638 bitmap_close_sync(mddev->bitmap);
4643 /* Allow raid5_quiesce to complete */
4644 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4646 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4647 return reshape_request(mddev, sector_nr, skipped);
4649 /* No need to check resync_max as we never do more than one
4650 * stripe, and as resync_max will always be on a chunk boundary,
4651 * if the check in md_do_sync didn't fire, there is no chance
4652 * of overstepping resync_max here
4655 /* if there is too many failed drives and we are trying
4656 * to resync, then assert that we are finished, because there is
4657 * nothing we can do.
4659 if (mddev->degraded >= conf->max_degraded &&
4660 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4661 sector_t rv = mddev->dev_sectors - sector_nr;
4665 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4666 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4667 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4668 /* we can skip this block, and probably more */
4669 sync_blocks /= STRIPE_SECTORS;
4671 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4674 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4676 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4678 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4679 /* make sure we don't swamp the stripe cache if someone else
4680 * is trying to get access
4682 schedule_timeout_uninterruptible(1);
4684 /* Need to check if array will still be degraded after recovery/resync
4685 * We don't need to check the 'failed' flag as when that gets set,
4688 for (i = 0; i < conf->raid_disks; i++)
4689 if (conf->disks[i].rdev == NULL)
4692 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4694 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4699 return STRIPE_SECTORS;
4702 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4704 /* We may not be able to submit a whole bio at once as there
4705 * may not be enough stripe_heads available.
4706 * We cannot pre-allocate enough stripe_heads as we may need
4707 * more than exist in the cache (if we allow ever large chunks).
4708 * So we do one stripe head at a time and record in
4709 * ->bi_hw_segments how many have been done.
4711 * We *know* that this entire raid_bio is in one chunk, so
4712 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4714 struct stripe_head *sh;
4716 sector_t sector, logical_sector, last_sector;
4721 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4722 sector = raid5_compute_sector(conf, logical_sector,
4724 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4726 for (; logical_sector < last_sector;
4727 logical_sector += STRIPE_SECTORS,
4728 sector += STRIPE_SECTORS,
4731 if (scnt < raid5_bi_processed_stripes(raid_bio))
4732 /* already done this stripe */
4735 sh = get_active_stripe(conf, sector, 0, 1, 0);
4738 /* failed to get a stripe - must wait */
4739 raid5_set_bi_processed_stripes(raid_bio, scnt);
4740 conf->retry_read_aligned = raid_bio;
4744 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4746 raid5_set_bi_processed_stripes(raid_bio, scnt);
4747 conf->retry_read_aligned = raid_bio;
4751 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4756 remaining = raid5_dec_bi_active_stripes(raid_bio);
4757 if (remaining == 0) {
4758 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4760 bio_endio(raid_bio, 0);
4762 if (atomic_dec_and_test(&conf->active_aligned_reads))
4763 wake_up(&conf->wait_for_stripe);
4767 #define MAX_STRIPE_BATCH 8
4768 static int handle_active_stripes(struct r5conf *conf)
4770 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4771 int i, batch_size = 0;
4773 while (batch_size < MAX_STRIPE_BATCH &&
4774 (sh = __get_priority_stripe(conf)) != NULL)
4775 batch[batch_size++] = sh;
4777 if (batch_size == 0)
4779 spin_unlock_irq(&conf->device_lock);
4781 for (i = 0; i < batch_size; i++)
4782 handle_stripe(batch[i]);
4786 spin_lock_irq(&conf->device_lock);
4787 for (i = 0; i < batch_size; i++)
4788 __release_stripe(conf, batch[i]);
4793 * This is our raid5 kernel thread.
4795 * We scan the hash table for stripes which can be handled now.
4796 * During the scan, completed stripes are saved for us by the interrupt
4797 * handler, so that they will not have to wait for our next wakeup.
4799 static void raid5d(struct md_thread *thread)
4801 struct mddev *mddev = thread->mddev;
4802 struct r5conf *conf = mddev->private;
4804 struct blk_plug plug;
4806 pr_debug("+++ raid5d active\n");
4808 md_check_recovery(mddev);
4810 blk_start_plug(&plug);
4812 spin_lock_irq(&conf->device_lock);
4818 !list_empty(&conf->bitmap_list)) {
4819 /* Now is a good time to flush some bitmap updates */
4821 spin_unlock_irq(&conf->device_lock);
4822 bitmap_unplug(mddev->bitmap);
4823 spin_lock_irq(&conf->device_lock);
4824 conf->seq_write = conf->seq_flush;
4825 activate_bit_delay(conf);
4827 raid5_activate_delayed(conf);
4829 while ((bio = remove_bio_from_retry(conf))) {
4831 spin_unlock_irq(&conf->device_lock);
4832 ok = retry_aligned_read(conf, bio);
4833 spin_lock_irq(&conf->device_lock);
4839 batch_size = handle_active_stripes(conf);
4842 handled += batch_size;
4844 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
4845 spin_unlock_irq(&conf->device_lock);
4846 md_check_recovery(mddev);
4847 spin_lock_irq(&conf->device_lock);
4850 pr_debug("%d stripes handled\n", handled);
4852 spin_unlock_irq(&conf->device_lock);
4854 async_tx_issue_pending_all();
4855 blk_finish_plug(&plug);
4857 pr_debug("--- raid5d inactive\n");
4861 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4863 struct r5conf *conf = mddev->private;
4865 return sprintf(page, "%d\n", conf->max_nr_stripes);
4871 raid5_set_cache_size(struct mddev *mddev, int size)
4873 struct r5conf *conf = mddev->private;
4876 if (size <= 16 || size > 32768)
4878 while (size < conf->max_nr_stripes) {
4879 if (drop_one_stripe(conf))
4880 conf->max_nr_stripes--;
4884 err = md_allow_write(mddev);
4887 while (size > conf->max_nr_stripes) {
4888 if (grow_one_stripe(conf))
4889 conf->max_nr_stripes++;
4894 EXPORT_SYMBOL(raid5_set_cache_size);
4897 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4899 struct r5conf *conf = mddev->private;
4903 if (len >= PAGE_SIZE)
4908 if (strict_strtoul(page, 10, &new))
4910 err = raid5_set_cache_size(mddev, new);
4916 static struct md_sysfs_entry
4917 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4918 raid5_show_stripe_cache_size,
4919 raid5_store_stripe_cache_size);
4922 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4924 struct r5conf *conf = mddev->private;
4926 return sprintf(page, "%d\n", conf->bypass_threshold);
4932 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4934 struct r5conf *conf = mddev->private;
4936 if (len >= PAGE_SIZE)
4941 if (strict_strtoul(page, 10, &new))
4943 if (new > conf->max_nr_stripes)
4945 conf->bypass_threshold = new;
4949 static struct md_sysfs_entry
4950 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4952 raid5_show_preread_threshold,
4953 raid5_store_preread_threshold);
4956 stripe_cache_active_show(struct mddev *mddev, char *page)
4958 struct r5conf *conf = mddev->private;
4960 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4965 static struct md_sysfs_entry
4966 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4968 static struct attribute *raid5_attrs[] = {
4969 &raid5_stripecache_size.attr,
4970 &raid5_stripecache_active.attr,
4971 &raid5_preread_bypass_threshold.attr,
4974 static struct attribute_group raid5_attrs_group = {
4976 .attrs = raid5_attrs,
4980 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4982 struct r5conf *conf = mddev->private;
4985 sectors = mddev->dev_sectors;
4987 /* size is defined by the smallest of previous and new size */
4988 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4990 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4991 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4992 return sectors * (raid_disks - conf->max_degraded);
4995 static void raid5_free_percpu(struct r5conf *conf)
4997 struct raid5_percpu *percpu;
5004 for_each_possible_cpu(cpu) {
5005 percpu = per_cpu_ptr(conf->percpu, cpu);
5006 safe_put_page(percpu->spare_page);
5007 kfree(percpu->scribble);
5009 #ifdef CONFIG_HOTPLUG_CPU
5010 unregister_cpu_notifier(&conf->cpu_notify);
5014 free_percpu(conf->percpu);
5017 static void free_conf(struct r5conf *conf)
5019 shrink_stripes(conf);
5020 raid5_free_percpu(conf);
5022 kfree(conf->stripe_hashtbl);
5026 #ifdef CONFIG_HOTPLUG_CPU
5027 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5030 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5031 long cpu = (long)hcpu;
5032 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5035 case CPU_UP_PREPARE:
5036 case CPU_UP_PREPARE_FROZEN:
5037 if (conf->level == 6 && !percpu->spare_page)
5038 percpu->spare_page = alloc_page(GFP_KERNEL);
5039 if (!percpu->scribble)
5040 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5042 if (!percpu->scribble ||
5043 (conf->level == 6 && !percpu->spare_page)) {
5044 safe_put_page(percpu->spare_page);
5045 kfree(percpu->scribble);
5046 pr_err("%s: failed memory allocation for cpu%ld\n",
5048 return notifier_from_errno(-ENOMEM);
5052 case CPU_DEAD_FROZEN:
5053 safe_put_page(percpu->spare_page);
5054 kfree(percpu->scribble);
5055 percpu->spare_page = NULL;
5056 percpu->scribble = NULL;
5065 static int raid5_alloc_percpu(struct r5conf *conf)
5068 struct page *spare_page;
5069 struct raid5_percpu __percpu *allcpus;
5073 allcpus = alloc_percpu(struct raid5_percpu);
5076 conf->percpu = allcpus;
5080 for_each_present_cpu(cpu) {
5081 if (conf->level == 6) {
5082 spare_page = alloc_page(GFP_KERNEL);
5087 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5089 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5094 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5096 #ifdef CONFIG_HOTPLUG_CPU
5097 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5098 conf->cpu_notify.priority = 0;
5100 err = register_cpu_notifier(&conf->cpu_notify);
5107 static struct r5conf *setup_conf(struct mddev *mddev)
5109 struct r5conf *conf;
5110 int raid_disk, memory, max_disks;
5111 struct md_rdev *rdev;
5112 struct disk_info *disk;
5115 if (mddev->new_level != 5
5116 && mddev->new_level != 4
5117 && mddev->new_level != 6) {
5118 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5119 mdname(mddev), mddev->new_level);
5120 return ERR_PTR(-EIO);
5122 if ((mddev->new_level == 5
5123 && !algorithm_valid_raid5(mddev->new_layout)) ||
5124 (mddev->new_level == 6
5125 && !algorithm_valid_raid6(mddev->new_layout))) {
5126 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5127 mdname(mddev), mddev->new_layout);
5128 return ERR_PTR(-EIO);
5130 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5131 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5132 mdname(mddev), mddev->raid_disks);
5133 return ERR_PTR(-EINVAL);
5136 if (!mddev->new_chunk_sectors ||
5137 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5138 !is_power_of_2(mddev->new_chunk_sectors)) {
5139 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5140 mdname(mddev), mddev->new_chunk_sectors << 9);
5141 return ERR_PTR(-EINVAL);
5144 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5147 spin_lock_init(&conf->device_lock);
5148 init_waitqueue_head(&conf->wait_for_stripe);
5149 init_waitqueue_head(&conf->wait_for_overlap);
5150 INIT_LIST_HEAD(&conf->handle_list);
5151 INIT_LIST_HEAD(&conf->hold_list);
5152 INIT_LIST_HEAD(&conf->delayed_list);
5153 INIT_LIST_HEAD(&conf->bitmap_list);
5154 INIT_LIST_HEAD(&conf->inactive_list);
5155 atomic_set(&conf->active_stripes, 0);
5156 atomic_set(&conf->preread_active_stripes, 0);
5157 atomic_set(&conf->active_aligned_reads, 0);
5158 conf->bypass_threshold = BYPASS_THRESHOLD;
5159 conf->recovery_disabled = mddev->recovery_disabled - 1;
5161 conf->raid_disks = mddev->raid_disks;
5162 if (mddev->reshape_position == MaxSector)
5163 conf->previous_raid_disks = mddev->raid_disks;
5165 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5166 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5167 conf->scribble_len = scribble_len(max_disks);
5169 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5174 conf->mddev = mddev;
5176 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5179 conf->level = mddev->new_level;
5180 if (raid5_alloc_percpu(conf) != 0)
5183 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5185 rdev_for_each(rdev, mddev) {
5186 raid_disk = rdev->raid_disk;
5187 if (raid_disk >= max_disks
5190 disk = conf->disks + raid_disk;
5192 if (test_bit(Replacement, &rdev->flags)) {
5193 if (disk->replacement)
5195 disk->replacement = rdev;
5202 if (test_bit(In_sync, &rdev->flags)) {
5203 char b[BDEVNAME_SIZE];
5204 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5206 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5207 } else if (rdev->saved_raid_disk != raid_disk)
5208 /* Cannot rely on bitmap to complete recovery */
5212 conf->chunk_sectors = mddev->new_chunk_sectors;
5213 conf->level = mddev->new_level;
5214 if (conf->level == 6)
5215 conf->max_degraded = 2;
5217 conf->max_degraded = 1;
5218 conf->algorithm = mddev->new_layout;
5219 conf->max_nr_stripes = NR_STRIPES;
5220 conf->reshape_progress = mddev->reshape_position;
5221 if (conf->reshape_progress != MaxSector) {
5222 conf->prev_chunk_sectors = mddev->chunk_sectors;
5223 conf->prev_algo = mddev->layout;
5226 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5227 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5228 if (grow_stripes(conf, conf->max_nr_stripes)) {
5230 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5231 mdname(mddev), memory);
5234 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5235 mdname(mddev), memory);
5237 sprintf(pers_name, "raid%d", mddev->new_level);
5238 conf->thread = md_register_thread(raid5d, mddev, pers_name);
5239 if (!conf->thread) {
5241 "md/raid:%s: couldn't allocate thread.\n",
5251 return ERR_PTR(-EIO);
5253 return ERR_PTR(-ENOMEM);
5257 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5260 case ALGORITHM_PARITY_0:
5261 if (raid_disk < max_degraded)
5264 case ALGORITHM_PARITY_N:
5265 if (raid_disk >= raid_disks - max_degraded)
5268 case ALGORITHM_PARITY_0_6:
5269 if (raid_disk == 0 ||
5270 raid_disk == raid_disks - 1)
5273 case ALGORITHM_LEFT_ASYMMETRIC_6:
5274 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5275 case ALGORITHM_LEFT_SYMMETRIC_6:
5276 case ALGORITHM_RIGHT_SYMMETRIC_6:
5277 if (raid_disk == raid_disks - 1)
5283 static int run(struct mddev *mddev)
5285 struct r5conf *conf;
5286 int working_disks = 0;
5287 int dirty_parity_disks = 0;
5288 struct md_rdev *rdev;
5289 sector_t reshape_offset = 0;
5291 long long min_offset_diff = 0;
5294 if (mddev->recovery_cp != MaxSector)
5295 printk(KERN_NOTICE "md/raid:%s: not clean"
5296 " -- starting background reconstruction\n",
5299 rdev_for_each(rdev, mddev) {
5301 if (rdev->raid_disk < 0)
5303 diff = (rdev->new_data_offset - rdev->data_offset);
5305 min_offset_diff = diff;
5307 } else if (mddev->reshape_backwards &&
5308 diff < min_offset_diff)
5309 min_offset_diff = diff;
5310 else if (!mddev->reshape_backwards &&
5311 diff > min_offset_diff)
5312 min_offset_diff = diff;
5315 if (mddev->reshape_position != MaxSector) {
5316 /* Check that we can continue the reshape.
5317 * Difficulties arise if the stripe we would write to
5318 * next is at or after the stripe we would read from next.
5319 * For a reshape that changes the number of devices, this
5320 * is only possible for a very short time, and mdadm makes
5321 * sure that time appears to have past before assembling
5322 * the array. So we fail if that time hasn't passed.
5323 * For a reshape that keeps the number of devices the same
5324 * mdadm must be monitoring the reshape can keeping the
5325 * critical areas read-only and backed up. It will start
5326 * the array in read-only mode, so we check for that.
5328 sector_t here_new, here_old;
5330 int max_degraded = (mddev->level == 6 ? 2 : 1);
5332 if (mddev->new_level != mddev->level) {
5333 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5334 "required - aborting.\n",
5338 old_disks = mddev->raid_disks - mddev->delta_disks;
5339 /* reshape_position must be on a new-stripe boundary, and one
5340 * further up in new geometry must map after here in old
5343 here_new = mddev->reshape_position;
5344 if (sector_div(here_new, mddev->new_chunk_sectors *
5345 (mddev->raid_disks - max_degraded))) {
5346 printk(KERN_ERR "md/raid:%s: reshape_position not "
5347 "on a stripe boundary\n", mdname(mddev));
5350 reshape_offset = here_new * mddev->new_chunk_sectors;
5351 /* here_new is the stripe we will write to */
5352 here_old = mddev->reshape_position;
5353 sector_div(here_old, mddev->chunk_sectors *
5354 (old_disks-max_degraded));
5355 /* here_old is the first stripe that we might need to read
5357 if (mddev->delta_disks == 0) {
5358 if ((here_new * mddev->new_chunk_sectors !=
5359 here_old * mddev->chunk_sectors)) {
5360 printk(KERN_ERR "md/raid:%s: reshape position is"
5361 " confused - aborting\n", mdname(mddev));
5364 /* We cannot be sure it is safe to start an in-place
5365 * reshape. It is only safe if user-space is monitoring
5366 * and taking constant backups.
5367 * mdadm always starts a situation like this in
5368 * readonly mode so it can take control before
5369 * allowing any writes. So just check for that.
5371 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5372 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5373 /* not really in-place - so OK */;
5374 else if (mddev->ro == 0) {
5375 printk(KERN_ERR "md/raid:%s: in-place reshape "
5376 "must be started in read-only mode "
5381 } else if (mddev->reshape_backwards
5382 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5383 here_old * mddev->chunk_sectors)
5384 : (here_new * mddev->new_chunk_sectors >=
5385 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5386 /* Reading from the same stripe as writing to - bad */
5387 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5388 "auto-recovery - aborting.\n",
5392 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5394 /* OK, we should be able to continue; */
5396 BUG_ON(mddev->level != mddev->new_level);
5397 BUG_ON(mddev->layout != mddev->new_layout);
5398 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5399 BUG_ON(mddev->delta_disks != 0);
5402 if (mddev->private == NULL)
5403 conf = setup_conf(mddev);
5405 conf = mddev->private;
5408 return PTR_ERR(conf);
5410 conf->min_offset_diff = min_offset_diff;
5411 mddev->thread = conf->thread;
5412 conf->thread = NULL;
5413 mddev->private = conf;
5415 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5417 rdev = conf->disks[i].rdev;
5418 if (!rdev && conf->disks[i].replacement) {
5419 /* The replacement is all we have yet */
5420 rdev = conf->disks[i].replacement;
5421 conf->disks[i].replacement = NULL;
5422 clear_bit(Replacement, &rdev->flags);
5423 conf->disks[i].rdev = rdev;
5427 if (conf->disks[i].replacement &&
5428 conf->reshape_progress != MaxSector) {
5429 /* replacements and reshape simply do not mix. */
5430 printk(KERN_ERR "md: cannot handle concurrent "
5431 "replacement and reshape.\n");
5434 if (test_bit(In_sync, &rdev->flags)) {
5438 /* This disc is not fully in-sync. However if it
5439 * just stored parity (beyond the recovery_offset),
5440 * when we don't need to be concerned about the
5441 * array being dirty.
5442 * When reshape goes 'backwards', we never have
5443 * partially completed devices, so we only need
5444 * to worry about reshape going forwards.
5446 /* Hack because v0.91 doesn't store recovery_offset properly. */
5447 if (mddev->major_version == 0 &&
5448 mddev->minor_version > 90)
5449 rdev->recovery_offset = reshape_offset;
5451 if (rdev->recovery_offset < reshape_offset) {
5452 /* We need to check old and new layout */
5453 if (!only_parity(rdev->raid_disk,
5456 conf->max_degraded))
5459 if (!only_parity(rdev->raid_disk,
5461 conf->previous_raid_disks,
5462 conf->max_degraded))
5464 dirty_parity_disks++;
5468 * 0 for a fully functional array, 1 or 2 for a degraded array.
5470 mddev->degraded = calc_degraded(conf);
5472 if (has_failed(conf)) {
5473 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5474 " (%d/%d failed)\n",
5475 mdname(mddev), mddev->degraded, conf->raid_disks);
5479 /* device size must be a multiple of chunk size */
5480 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5481 mddev->resync_max_sectors = mddev->dev_sectors;
5483 if (mddev->degraded > dirty_parity_disks &&
5484 mddev->recovery_cp != MaxSector) {
5485 if (mddev->ok_start_degraded)
5487 "md/raid:%s: starting dirty degraded array"
5488 " - data corruption possible.\n",
5492 "md/raid:%s: cannot start dirty degraded array.\n",
5498 if (mddev->degraded == 0)
5499 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5500 " devices, algorithm %d\n", mdname(mddev), conf->level,
5501 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5504 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5505 " out of %d devices, algorithm %d\n",
5506 mdname(mddev), conf->level,
5507 mddev->raid_disks - mddev->degraded,
5508 mddev->raid_disks, mddev->new_layout);
5510 print_raid5_conf(conf);
5512 if (conf->reshape_progress != MaxSector) {
5513 conf->reshape_safe = conf->reshape_progress;
5514 atomic_set(&conf->reshape_stripes, 0);
5515 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5516 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5517 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5518 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5519 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5524 /* Ok, everything is just fine now */
5525 if (mddev->to_remove == &raid5_attrs_group)
5526 mddev->to_remove = NULL;
5527 else if (mddev->kobj.sd &&
5528 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5530 "raid5: failed to create sysfs attributes for %s\n",
5532 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5536 bool discard_supported = true;
5537 /* read-ahead size must cover two whole stripes, which
5538 * is 2 * (datadisks) * chunksize where 'n' is the
5539 * number of raid devices
5541 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5542 int stripe = data_disks *
5543 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5544 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5545 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5547 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5549 mddev->queue->backing_dev_info.congested_data = mddev;
5550 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5552 chunk_size = mddev->chunk_sectors << 9;
5553 blk_queue_io_min(mddev->queue, chunk_size);
5554 blk_queue_io_opt(mddev->queue, chunk_size *
5555 (conf->raid_disks - conf->max_degraded));
5557 * We can only discard a whole stripe. It doesn't make sense to
5558 * discard data disk but write parity disk
5560 stripe = stripe * PAGE_SIZE;
5561 /* Round up to power of 2, as discard handling
5562 * currently assumes that */
5563 while ((stripe-1) & stripe)
5564 stripe = (stripe | (stripe-1)) + 1;
5565 mddev->queue->limits.discard_alignment = stripe;
5566 mddev->queue->limits.discard_granularity = stripe;
5568 * unaligned part of discard request will be ignored, so can't
5569 * guarantee discard_zerors_data
5571 mddev->queue->limits.discard_zeroes_data = 0;
5573 rdev_for_each(rdev, mddev) {
5574 disk_stack_limits(mddev->gendisk, rdev->bdev,
5575 rdev->data_offset << 9);
5576 disk_stack_limits(mddev->gendisk, rdev->bdev,
5577 rdev->new_data_offset << 9);
5579 * discard_zeroes_data is required, otherwise data
5580 * could be lost. Consider a scenario: discard a stripe
5581 * (the stripe could be inconsistent if
5582 * discard_zeroes_data is 0); write one disk of the
5583 * stripe (the stripe could be inconsistent again
5584 * depending on which disks are used to calculate
5585 * parity); the disk is broken; The stripe data of this
5588 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5589 !bdev_get_queue(rdev->bdev)->
5590 limits.discard_zeroes_data)
5591 discard_supported = false;
5594 if (discard_supported &&
5595 mddev->queue->limits.max_discard_sectors >= stripe &&
5596 mddev->queue->limits.discard_granularity >= stripe)
5597 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5600 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5606 md_unregister_thread(&mddev->thread);
5607 print_raid5_conf(conf);
5609 mddev->private = NULL;
5610 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5614 static int stop(struct mddev *mddev)
5616 struct r5conf *conf = mddev->private;
5618 md_unregister_thread(&mddev->thread);
5620 mddev->queue->backing_dev_info.congested_fn = NULL;
5622 mddev->private = NULL;
5623 mddev->to_remove = &raid5_attrs_group;
5627 static void status(struct seq_file *seq, struct mddev *mddev)
5629 struct r5conf *conf = mddev->private;
5632 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5633 mddev->chunk_sectors / 2, mddev->layout);
5634 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5635 for (i = 0; i < conf->raid_disks; i++)
5636 seq_printf (seq, "%s",
5637 conf->disks[i].rdev &&
5638 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5639 seq_printf (seq, "]");
5642 static void print_raid5_conf (struct r5conf *conf)
5645 struct disk_info *tmp;
5647 printk(KERN_DEBUG "RAID conf printout:\n");
5649 printk("(conf==NULL)\n");
5652 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5654 conf->raid_disks - conf->mddev->degraded);
5656 for (i = 0; i < conf->raid_disks; i++) {
5657 char b[BDEVNAME_SIZE];
5658 tmp = conf->disks + i;
5660 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5661 i, !test_bit(Faulty, &tmp->rdev->flags),
5662 bdevname(tmp->rdev->bdev, b));
5666 static int raid5_spare_active(struct mddev *mddev)
5669 struct r5conf *conf = mddev->private;
5670 struct disk_info *tmp;
5672 unsigned long flags;
5674 for (i = 0; i < conf->raid_disks; i++) {
5675 tmp = conf->disks + i;
5676 if (tmp->replacement
5677 && tmp->replacement->recovery_offset == MaxSector
5678 && !test_bit(Faulty, &tmp->replacement->flags)
5679 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5680 /* Replacement has just become active. */
5682 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5685 /* Replaced device not technically faulty,
5686 * but we need to be sure it gets removed
5687 * and never re-added.
5689 set_bit(Faulty, &tmp->rdev->flags);
5690 sysfs_notify_dirent_safe(
5691 tmp->rdev->sysfs_state);
5693 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5694 } else if (tmp->rdev
5695 && tmp->rdev->recovery_offset == MaxSector
5696 && !test_bit(Faulty, &tmp->rdev->flags)
5697 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5699 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5702 spin_lock_irqsave(&conf->device_lock, flags);
5703 mddev->degraded = calc_degraded(conf);
5704 spin_unlock_irqrestore(&conf->device_lock, flags);
5705 print_raid5_conf(conf);
5709 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5711 struct r5conf *conf = mddev->private;
5713 int number = rdev->raid_disk;
5714 struct md_rdev **rdevp;
5715 struct disk_info *p = conf->disks + number;
5717 print_raid5_conf(conf);
5718 if (rdev == p->rdev)
5720 else if (rdev == p->replacement)
5721 rdevp = &p->replacement;
5725 if (number >= conf->raid_disks &&
5726 conf->reshape_progress == MaxSector)
5727 clear_bit(In_sync, &rdev->flags);
5729 if (test_bit(In_sync, &rdev->flags) ||
5730 atomic_read(&rdev->nr_pending)) {
5734 /* Only remove non-faulty devices if recovery
5737 if (!test_bit(Faulty, &rdev->flags) &&
5738 mddev->recovery_disabled != conf->recovery_disabled &&
5739 !has_failed(conf) &&
5740 (!p->replacement || p->replacement == rdev) &&
5741 number < conf->raid_disks) {
5747 if (atomic_read(&rdev->nr_pending)) {
5748 /* lost the race, try later */
5751 } else if (p->replacement) {
5752 /* We must have just cleared 'rdev' */
5753 p->rdev = p->replacement;
5754 clear_bit(Replacement, &p->replacement->flags);
5755 smp_mb(); /* Make sure other CPUs may see both as identical
5756 * but will never see neither - if they are careful
5758 p->replacement = NULL;
5759 clear_bit(WantReplacement, &rdev->flags);
5761 /* We might have just removed the Replacement as faulty-
5762 * clear the bit just in case
5764 clear_bit(WantReplacement, &rdev->flags);
5767 print_raid5_conf(conf);
5771 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5773 struct r5conf *conf = mddev->private;
5776 struct disk_info *p;
5778 int last = conf->raid_disks - 1;
5780 if (mddev->recovery_disabled == conf->recovery_disabled)
5783 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5784 /* no point adding a device */
5787 if (rdev->raid_disk >= 0)
5788 first = last = rdev->raid_disk;
5791 * find the disk ... but prefer rdev->saved_raid_disk
5794 if (rdev->saved_raid_disk >= 0 &&
5795 rdev->saved_raid_disk >= first &&
5796 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5797 first = rdev->saved_raid_disk;
5799 for (disk = first; disk <= last; disk++) {
5800 p = conf->disks + disk;
5801 if (p->rdev == NULL) {
5802 clear_bit(In_sync, &rdev->flags);
5803 rdev->raid_disk = disk;
5805 if (rdev->saved_raid_disk != disk)
5807 rcu_assign_pointer(p->rdev, rdev);
5811 for (disk = first; disk <= last; disk++) {
5812 p = conf->disks + disk;
5813 if (test_bit(WantReplacement, &p->rdev->flags) &&
5814 p->replacement == NULL) {
5815 clear_bit(In_sync, &rdev->flags);
5816 set_bit(Replacement, &rdev->flags);
5817 rdev->raid_disk = disk;
5820 rcu_assign_pointer(p->replacement, rdev);
5825 print_raid5_conf(conf);
5829 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5831 /* no resync is happening, and there is enough space
5832 * on all devices, so we can resize.
5833 * We need to make sure resync covers any new space.
5834 * If the array is shrinking we should possibly wait until
5835 * any io in the removed space completes, but it hardly seems
5839 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5840 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5841 if (mddev->external_size &&
5842 mddev->array_sectors > newsize)
5844 if (mddev->bitmap) {
5845 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5849 md_set_array_sectors(mddev, newsize);
5850 set_capacity(mddev->gendisk, mddev->array_sectors);
5851 revalidate_disk(mddev->gendisk);
5852 if (sectors > mddev->dev_sectors &&
5853 mddev->recovery_cp > mddev->dev_sectors) {
5854 mddev->recovery_cp = mddev->dev_sectors;
5855 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5857 mddev->dev_sectors = sectors;
5858 mddev->resync_max_sectors = sectors;
5862 static int check_stripe_cache(struct mddev *mddev)
5864 /* Can only proceed if there are plenty of stripe_heads.
5865 * We need a minimum of one full stripe,, and for sensible progress
5866 * it is best to have about 4 times that.
5867 * If we require 4 times, then the default 256 4K stripe_heads will
5868 * allow for chunk sizes up to 256K, which is probably OK.
5869 * If the chunk size is greater, user-space should request more
5870 * stripe_heads first.
5872 struct r5conf *conf = mddev->private;
5873 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5874 > conf->max_nr_stripes ||
5875 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5876 > conf->max_nr_stripes) {
5877 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5879 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5886 static int check_reshape(struct mddev *mddev)
5888 struct r5conf *conf = mddev->private;
5890 if (mddev->delta_disks == 0 &&
5891 mddev->new_layout == mddev->layout &&
5892 mddev->new_chunk_sectors == mddev->chunk_sectors)
5893 return 0; /* nothing to do */
5894 if (has_failed(conf))
5896 if (mddev->delta_disks < 0) {
5897 /* We might be able to shrink, but the devices must
5898 * be made bigger first.
5899 * For raid6, 4 is the minimum size.
5900 * Otherwise 2 is the minimum
5903 if (mddev->level == 6)
5905 if (mddev->raid_disks + mddev->delta_disks < min)
5909 if (!check_stripe_cache(mddev))
5912 return resize_stripes(conf, (conf->previous_raid_disks
5913 + mddev->delta_disks));
5916 static int raid5_start_reshape(struct mddev *mddev)
5918 struct r5conf *conf = mddev->private;
5919 struct md_rdev *rdev;
5921 unsigned long flags;
5923 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5926 if (!check_stripe_cache(mddev))
5929 if (has_failed(conf))
5932 rdev_for_each(rdev, mddev) {
5933 if (!test_bit(In_sync, &rdev->flags)
5934 && !test_bit(Faulty, &rdev->flags))
5938 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5939 /* Not enough devices even to make a degraded array
5944 /* Refuse to reduce size of the array. Any reductions in
5945 * array size must be through explicit setting of array_size
5948 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5949 < mddev->array_sectors) {
5950 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5951 "before number of disks\n", mdname(mddev));
5955 atomic_set(&conf->reshape_stripes, 0);
5956 spin_lock_irq(&conf->device_lock);
5957 conf->previous_raid_disks = conf->raid_disks;
5958 conf->raid_disks += mddev->delta_disks;
5959 conf->prev_chunk_sectors = conf->chunk_sectors;
5960 conf->chunk_sectors = mddev->new_chunk_sectors;
5961 conf->prev_algo = conf->algorithm;
5962 conf->algorithm = mddev->new_layout;
5964 /* Code that selects data_offset needs to see the generation update
5965 * if reshape_progress has been set - so a memory barrier needed.
5968 if (mddev->reshape_backwards)
5969 conf->reshape_progress = raid5_size(mddev, 0, 0);
5971 conf->reshape_progress = 0;
5972 conf->reshape_safe = conf->reshape_progress;
5973 spin_unlock_irq(&conf->device_lock);
5975 /* Add some new drives, as many as will fit.
5976 * We know there are enough to make the newly sized array work.
5977 * Don't add devices if we are reducing the number of
5978 * devices in the array. This is because it is not possible
5979 * to correctly record the "partially reconstructed" state of
5980 * such devices during the reshape and confusion could result.
5982 if (mddev->delta_disks >= 0) {
5983 rdev_for_each(rdev, mddev)
5984 if (rdev->raid_disk < 0 &&
5985 !test_bit(Faulty, &rdev->flags)) {
5986 if (raid5_add_disk(mddev, rdev) == 0) {
5988 >= conf->previous_raid_disks)
5989 set_bit(In_sync, &rdev->flags);
5991 rdev->recovery_offset = 0;
5993 if (sysfs_link_rdev(mddev, rdev))
5994 /* Failure here is OK */;
5996 } else if (rdev->raid_disk >= conf->previous_raid_disks
5997 && !test_bit(Faulty, &rdev->flags)) {
5998 /* This is a spare that was manually added */
5999 set_bit(In_sync, &rdev->flags);
6002 /* When a reshape changes the number of devices,
6003 * ->degraded is measured against the larger of the
6004 * pre and post number of devices.
6006 spin_lock_irqsave(&conf->device_lock, flags);
6007 mddev->degraded = calc_degraded(conf);
6008 spin_unlock_irqrestore(&conf->device_lock, flags);
6010 mddev->raid_disks = conf->raid_disks;
6011 mddev->reshape_position = conf->reshape_progress;
6012 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6014 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6015 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6016 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6017 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6018 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6020 if (!mddev->sync_thread) {
6021 mddev->recovery = 0;
6022 spin_lock_irq(&conf->device_lock);
6023 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6024 rdev_for_each(rdev, mddev)
6025 rdev->new_data_offset = rdev->data_offset;
6027 conf->reshape_progress = MaxSector;
6028 mddev->reshape_position = MaxSector;
6029 spin_unlock_irq(&conf->device_lock);
6032 conf->reshape_checkpoint = jiffies;
6033 md_wakeup_thread(mddev->sync_thread);
6034 md_new_event(mddev);
6038 /* This is called from the reshape thread and should make any
6039 * changes needed in 'conf'
6041 static void end_reshape(struct r5conf *conf)
6044 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6045 struct md_rdev *rdev;
6047 spin_lock_irq(&conf->device_lock);
6048 conf->previous_raid_disks = conf->raid_disks;
6049 rdev_for_each(rdev, conf->mddev)
6050 rdev->data_offset = rdev->new_data_offset;
6052 conf->reshape_progress = MaxSector;
6053 spin_unlock_irq(&conf->device_lock);
6054 wake_up(&conf->wait_for_overlap);
6056 /* read-ahead size must cover two whole stripes, which is
6057 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6059 if (conf->mddev->queue) {
6060 int data_disks = conf->raid_disks - conf->max_degraded;
6061 int stripe = data_disks * ((conf->chunk_sectors << 9)
6063 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6064 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6069 /* This is called from the raid5d thread with mddev_lock held.
6070 * It makes config changes to the device.
6072 static void raid5_finish_reshape(struct mddev *mddev)
6074 struct r5conf *conf = mddev->private;
6076 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6078 if (mddev->delta_disks > 0) {
6079 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6080 set_capacity(mddev->gendisk, mddev->array_sectors);
6081 revalidate_disk(mddev->gendisk);
6084 spin_lock_irq(&conf->device_lock);
6085 mddev->degraded = calc_degraded(conf);
6086 spin_unlock_irq(&conf->device_lock);
6087 for (d = conf->raid_disks ;
6088 d < conf->raid_disks - mddev->delta_disks;
6090 struct md_rdev *rdev = conf->disks[d].rdev;
6092 clear_bit(In_sync, &rdev->flags);
6093 rdev = conf->disks[d].replacement;
6095 clear_bit(In_sync, &rdev->flags);
6098 mddev->layout = conf->algorithm;
6099 mddev->chunk_sectors = conf->chunk_sectors;
6100 mddev->reshape_position = MaxSector;
6101 mddev->delta_disks = 0;
6102 mddev->reshape_backwards = 0;
6106 static void raid5_quiesce(struct mddev *mddev, int state)
6108 struct r5conf *conf = mddev->private;
6111 case 2: /* resume for a suspend */
6112 wake_up(&conf->wait_for_overlap);
6115 case 1: /* stop all writes */
6116 spin_lock_irq(&conf->device_lock);
6117 /* '2' tells resync/reshape to pause so that all
6118 * active stripes can drain
6121 wait_event_lock_irq(conf->wait_for_stripe,
6122 atomic_read(&conf->active_stripes) == 0 &&
6123 atomic_read(&conf->active_aligned_reads) == 0,
6126 spin_unlock_irq(&conf->device_lock);
6127 /* allow reshape to continue */
6128 wake_up(&conf->wait_for_overlap);
6131 case 0: /* re-enable writes */
6132 spin_lock_irq(&conf->device_lock);
6134 wake_up(&conf->wait_for_stripe);
6135 wake_up(&conf->wait_for_overlap);
6136 spin_unlock_irq(&conf->device_lock);
6142 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6144 struct r0conf *raid0_conf = mddev->private;
6147 /* for raid0 takeover only one zone is supported */
6148 if (raid0_conf->nr_strip_zones > 1) {
6149 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6151 return ERR_PTR(-EINVAL);
6154 sectors = raid0_conf->strip_zone[0].zone_end;
6155 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6156 mddev->dev_sectors = sectors;
6157 mddev->new_level = level;
6158 mddev->new_layout = ALGORITHM_PARITY_N;
6159 mddev->new_chunk_sectors = mddev->chunk_sectors;
6160 mddev->raid_disks += 1;
6161 mddev->delta_disks = 1;
6162 /* make sure it will be not marked as dirty */
6163 mddev->recovery_cp = MaxSector;
6165 return setup_conf(mddev);
6169 static void *raid5_takeover_raid1(struct mddev *mddev)
6173 if (mddev->raid_disks != 2 ||
6174 mddev->degraded > 1)
6175 return ERR_PTR(-EINVAL);
6177 /* Should check if there are write-behind devices? */
6179 chunksect = 64*2; /* 64K by default */
6181 /* The array must be an exact multiple of chunksize */
6182 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6185 if ((chunksect<<9) < STRIPE_SIZE)
6186 /* array size does not allow a suitable chunk size */
6187 return ERR_PTR(-EINVAL);
6189 mddev->new_level = 5;
6190 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6191 mddev->new_chunk_sectors = chunksect;
6193 return setup_conf(mddev);
6196 static void *raid5_takeover_raid6(struct mddev *mddev)
6200 switch (mddev->layout) {
6201 case ALGORITHM_LEFT_ASYMMETRIC_6:
6202 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6204 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6205 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6207 case ALGORITHM_LEFT_SYMMETRIC_6:
6208 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6210 case ALGORITHM_RIGHT_SYMMETRIC_6:
6211 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6213 case ALGORITHM_PARITY_0_6:
6214 new_layout = ALGORITHM_PARITY_0;
6216 case ALGORITHM_PARITY_N:
6217 new_layout = ALGORITHM_PARITY_N;
6220 return ERR_PTR(-EINVAL);
6222 mddev->new_level = 5;
6223 mddev->new_layout = new_layout;
6224 mddev->delta_disks = -1;
6225 mddev->raid_disks -= 1;
6226 return setup_conf(mddev);
6230 static int raid5_check_reshape(struct mddev *mddev)
6232 /* For a 2-drive array, the layout and chunk size can be changed
6233 * immediately as not restriping is needed.
6234 * For larger arrays we record the new value - after validation
6235 * to be used by a reshape pass.
6237 struct r5conf *conf = mddev->private;
6238 int new_chunk = mddev->new_chunk_sectors;
6240 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6242 if (new_chunk > 0) {
6243 if (!is_power_of_2(new_chunk))
6245 if (new_chunk < (PAGE_SIZE>>9))
6247 if (mddev->array_sectors & (new_chunk-1))
6248 /* not factor of array size */
6252 /* They look valid */
6254 if (mddev->raid_disks == 2) {
6255 /* can make the change immediately */
6256 if (mddev->new_layout >= 0) {
6257 conf->algorithm = mddev->new_layout;
6258 mddev->layout = mddev->new_layout;
6260 if (new_chunk > 0) {
6261 conf->chunk_sectors = new_chunk ;
6262 mddev->chunk_sectors = new_chunk;
6264 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6265 md_wakeup_thread(mddev->thread);
6267 return check_reshape(mddev);
6270 static int raid6_check_reshape(struct mddev *mddev)
6272 int new_chunk = mddev->new_chunk_sectors;
6274 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6276 if (new_chunk > 0) {
6277 if (!is_power_of_2(new_chunk))
6279 if (new_chunk < (PAGE_SIZE >> 9))
6281 if (mddev->array_sectors & (new_chunk-1))
6282 /* not factor of array size */
6286 /* They look valid */
6287 return check_reshape(mddev);
6290 static void *raid5_takeover(struct mddev *mddev)
6292 /* raid5 can take over:
6293 * raid0 - if there is only one strip zone - make it a raid4 layout
6294 * raid1 - if there are two drives. We need to know the chunk size
6295 * raid4 - trivial - just use a raid4 layout.
6296 * raid6 - Providing it is a *_6 layout
6298 if (mddev->level == 0)
6299 return raid45_takeover_raid0(mddev, 5);
6300 if (mddev->level == 1)
6301 return raid5_takeover_raid1(mddev);
6302 if (mddev->level == 4) {
6303 mddev->new_layout = ALGORITHM_PARITY_N;
6304 mddev->new_level = 5;
6305 return setup_conf(mddev);
6307 if (mddev->level == 6)
6308 return raid5_takeover_raid6(mddev);
6310 return ERR_PTR(-EINVAL);
6313 static void *raid4_takeover(struct mddev *mddev)
6315 /* raid4 can take over:
6316 * raid0 - if there is only one strip zone
6317 * raid5 - if layout is right
6319 if (mddev->level == 0)
6320 return raid45_takeover_raid0(mddev, 4);
6321 if (mddev->level == 5 &&
6322 mddev->layout == ALGORITHM_PARITY_N) {
6323 mddev->new_layout = 0;
6324 mddev->new_level = 4;
6325 return setup_conf(mddev);
6327 return ERR_PTR(-EINVAL);
6330 static struct md_personality raid5_personality;
6332 static void *raid6_takeover(struct mddev *mddev)
6334 /* Currently can only take over a raid5. We map the
6335 * personality to an equivalent raid6 personality
6336 * with the Q block at the end.
6340 if (mddev->pers != &raid5_personality)
6341 return ERR_PTR(-EINVAL);
6342 if (mddev->degraded > 1)
6343 return ERR_PTR(-EINVAL);
6344 if (mddev->raid_disks > 253)
6345 return ERR_PTR(-EINVAL);
6346 if (mddev->raid_disks < 3)
6347 return ERR_PTR(-EINVAL);
6349 switch (mddev->layout) {
6350 case ALGORITHM_LEFT_ASYMMETRIC:
6351 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6353 case ALGORITHM_RIGHT_ASYMMETRIC:
6354 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6356 case ALGORITHM_LEFT_SYMMETRIC:
6357 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6359 case ALGORITHM_RIGHT_SYMMETRIC:
6360 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6362 case ALGORITHM_PARITY_0:
6363 new_layout = ALGORITHM_PARITY_0_6;
6365 case ALGORITHM_PARITY_N:
6366 new_layout = ALGORITHM_PARITY_N;
6369 return ERR_PTR(-EINVAL);
6371 mddev->new_level = 6;
6372 mddev->new_layout = new_layout;
6373 mddev->delta_disks = 1;
6374 mddev->raid_disks += 1;
6375 return setup_conf(mddev);
6379 static struct md_personality raid6_personality =
6383 .owner = THIS_MODULE,
6384 .make_request = make_request,
6388 .error_handler = error,
6389 .hot_add_disk = raid5_add_disk,
6390 .hot_remove_disk= raid5_remove_disk,
6391 .spare_active = raid5_spare_active,
6392 .sync_request = sync_request,
6393 .resize = raid5_resize,
6395 .check_reshape = raid6_check_reshape,
6396 .start_reshape = raid5_start_reshape,
6397 .finish_reshape = raid5_finish_reshape,
6398 .quiesce = raid5_quiesce,
6399 .takeover = raid6_takeover,
6401 static struct md_personality raid5_personality =
6405 .owner = THIS_MODULE,
6406 .make_request = make_request,
6410 .error_handler = error,
6411 .hot_add_disk = raid5_add_disk,
6412 .hot_remove_disk= raid5_remove_disk,
6413 .spare_active = raid5_spare_active,
6414 .sync_request = sync_request,
6415 .resize = raid5_resize,
6417 .check_reshape = raid5_check_reshape,
6418 .start_reshape = raid5_start_reshape,
6419 .finish_reshape = raid5_finish_reshape,
6420 .quiesce = raid5_quiesce,
6421 .takeover = raid5_takeover,
6424 static struct md_personality raid4_personality =
6428 .owner = THIS_MODULE,
6429 .make_request = make_request,
6433 .error_handler = error,
6434 .hot_add_disk = raid5_add_disk,
6435 .hot_remove_disk= raid5_remove_disk,
6436 .spare_active = raid5_spare_active,
6437 .sync_request = sync_request,
6438 .resize = raid5_resize,
6440 .check_reshape = raid5_check_reshape,
6441 .start_reshape = raid5_start_reshape,
6442 .finish_reshape = raid5_finish_reshape,
6443 .quiesce = raid5_quiesce,
6444 .takeover = raid4_takeover,
6447 static int __init raid5_init(void)
6449 register_md_personality(&raid6_personality);
6450 register_md_personality(&raid5_personality);
6451 register_md_personality(&raid4_personality);
6455 static void raid5_exit(void)
6457 unregister_md_personality(&raid6_personality);
6458 unregister_md_personality(&raid5_personality);
6459 unregister_md_personality(&raid4_personality);
6462 module_init(raid5_init);
6463 module_exit(raid5_exit);
6464 MODULE_LICENSE("GPL");
6465 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6466 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6467 MODULE_ALIAS("md-raid5");
6468 MODULE_ALIAS("md-raid4");
6469 MODULE_ALIAS("md-level-5");
6470 MODULE_ALIAS("md-level-4");
6471 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6472 MODULE_ALIAS("md-raid6");
6473 MODULE_ALIAS("md-level-6");
6475 /* This used to be two separate modules, they were: */
6476 MODULE_ALIAS("raid5");
6477 MODULE_ALIAS("raid6");