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_sectors(bio);
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 */
573 rrdev = rcu_dereference(conf->disks[i].replacement);
574 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
575 rdev = rcu_dereference(conf->disks[i].rdev);
584 /* We raced and saw duplicates */
587 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
592 if (rdev && test_bit(Faulty, &rdev->flags))
595 atomic_inc(&rdev->nr_pending);
596 if (rrdev && test_bit(Faulty, &rrdev->flags))
599 atomic_inc(&rrdev->nr_pending);
602 /* We have already checked bad blocks for reads. Now
603 * need to check for writes. We never accept write errors
604 * on the replacement, so we don't to check rrdev.
606 while ((rw & WRITE) && rdev &&
607 test_bit(WriteErrorSeen, &rdev->flags)) {
610 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
611 &first_bad, &bad_sectors);
616 set_bit(BlockedBadBlocks, &rdev->flags);
617 if (!conf->mddev->external &&
618 conf->mddev->flags) {
619 /* It is very unlikely, but we might
620 * still need to write out the
621 * bad block log - better give it
623 md_check_recovery(conf->mddev);
626 * Because md_wait_for_blocked_rdev
627 * will dec nr_pending, we must
628 * increment it first.
630 atomic_inc(&rdev->nr_pending);
631 md_wait_for_blocked_rdev(rdev, conf->mddev);
633 /* Acknowledged bad block - skip the write */
634 rdev_dec_pending(rdev, conf->mddev);
640 if (s->syncing || s->expanding || s->expanded
642 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
644 set_bit(STRIPE_IO_STARTED, &sh->state);
647 bi->bi_bdev = rdev->bdev;
649 bi->bi_end_io = (rw & WRITE)
650 ? raid5_end_write_request
651 : raid5_end_read_request;
654 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
655 __func__, (unsigned long long)sh->sector,
657 atomic_inc(&sh->count);
658 if (use_new_offset(conf, sh))
659 bi->bi_sector = (sh->sector
660 + rdev->new_data_offset);
662 bi->bi_sector = (sh->sector
663 + rdev->data_offset);
664 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
665 bi->bi_rw |= REQ_FLUSH;
668 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
669 bi->bi_io_vec[0].bv_offset = 0;
670 bi->bi_size = STRIPE_SIZE;
672 * If this is discard request, set bi_vcnt 0. We don't
673 * want to confuse SCSI because SCSI will replace payload
675 if (rw & REQ_DISCARD)
678 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
680 if (conf->mddev->gendisk)
681 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
682 bi, disk_devt(conf->mddev->gendisk),
684 generic_make_request(bi);
687 if (s->syncing || s->expanding || s->expanded
689 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
691 set_bit(STRIPE_IO_STARTED, &sh->state);
694 rbi->bi_bdev = rrdev->bdev;
696 BUG_ON(!(rw & WRITE));
697 rbi->bi_end_io = raid5_end_write_request;
698 rbi->bi_private = sh;
700 pr_debug("%s: for %llu schedule op %ld on "
701 "replacement disc %d\n",
702 __func__, (unsigned long long)sh->sector,
704 atomic_inc(&sh->count);
705 if (use_new_offset(conf, sh))
706 rbi->bi_sector = (sh->sector
707 + rrdev->new_data_offset);
709 rbi->bi_sector = (sh->sector
710 + rrdev->data_offset);
712 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
713 rbi->bi_io_vec[0].bv_offset = 0;
714 rbi->bi_size = STRIPE_SIZE;
716 * If this is discard request, set bi_vcnt 0. We don't
717 * want to confuse SCSI because SCSI will replace payload
719 if (rw & REQ_DISCARD)
721 if (conf->mddev->gendisk)
722 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
723 rbi, disk_devt(conf->mddev->gendisk),
725 generic_make_request(rbi);
727 if (!rdev && !rrdev) {
729 set_bit(STRIPE_DEGRADED, &sh->state);
730 pr_debug("skip op %ld on disc %d for sector %llu\n",
731 bi->bi_rw, i, (unsigned long long)sh->sector);
732 clear_bit(R5_LOCKED, &sh->dev[i].flags);
733 set_bit(STRIPE_HANDLE, &sh->state);
738 static struct dma_async_tx_descriptor *
739 async_copy_data(int frombio, struct bio *bio, struct page *page,
740 sector_t sector, struct dma_async_tx_descriptor *tx)
743 struct page *bio_page;
746 struct async_submit_ctl submit;
747 enum async_tx_flags flags = 0;
749 if (bio->bi_sector >= sector)
750 page_offset = (signed)(bio->bi_sector - sector) * 512;
752 page_offset = (signed)(sector - bio->bi_sector) * -512;
755 flags |= ASYNC_TX_FENCE;
756 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
758 bio_for_each_segment(bvl, bio, i) {
759 int len = bvl->bv_len;
763 if (page_offset < 0) {
764 b_offset = -page_offset;
765 page_offset += b_offset;
769 if (len > 0 && page_offset + len > STRIPE_SIZE)
770 clen = STRIPE_SIZE - page_offset;
775 b_offset += bvl->bv_offset;
776 bio_page = bvl->bv_page;
778 tx = async_memcpy(page, bio_page, page_offset,
779 b_offset, clen, &submit);
781 tx = async_memcpy(bio_page, page, b_offset,
782 page_offset, clen, &submit);
784 /* chain the operations */
785 submit.depend_tx = tx;
787 if (clen < len) /* hit end of page */
795 static void ops_complete_biofill(void *stripe_head_ref)
797 struct stripe_head *sh = stripe_head_ref;
798 struct bio *return_bi = NULL;
801 pr_debug("%s: stripe %llu\n", __func__,
802 (unsigned long long)sh->sector);
804 /* clear completed biofills */
805 for (i = sh->disks; i--; ) {
806 struct r5dev *dev = &sh->dev[i];
808 /* acknowledge completion of a biofill operation */
809 /* and check if we need to reply to a read request,
810 * new R5_Wantfill requests are held off until
811 * !STRIPE_BIOFILL_RUN
813 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
814 struct bio *rbi, *rbi2;
819 while (rbi && rbi->bi_sector <
820 dev->sector + STRIPE_SECTORS) {
821 rbi2 = r5_next_bio(rbi, dev->sector);
822 if (!raid5_dec_bi_active_stripes(rbi)) {
823 rbi->bi_next = return_bi;
830 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
832 return_io(return_bi);
834 set_bit(STRIPE_HANDLE, &sh->state);
838 static void ops_run_biofill(struct stripe_head *sh)
840 struct dma_async_tx_descriptor *tx = NULL;
841 struct async_submit_ctl submit;
844 pr_debug("%s: stripe %llu\n", __func__,
845 (unsigned long long)sh->sector);
847 for (i = sh->disks; i--; ) {
848 struct r5dev *dev = &sh->dev[i];
849 if (test_bit(R5_Wantfill, &dev->flags)) {
851 spin_lock_irq(&sh->stripe_lock);
852 dev->read = rbi = dev->toread;
854 spin_unlock_irq(&sh->stripe_lock);
855 while (rbi && rbi->bi_sector <
856 dev->sector + STRIPE_SECTORS) {
857 tx = async_copy_data(0, rbi, dev->page,
859 rbi = r5_next_bio(rbi, dev->sector);
864 atomic_inc(&sh->count);
865 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
866 async_trigger_callback(&submit);
869 static void mark_target_uptodate(struct stripe_head *sh, int target)
876 tgt = &sh->dev[target];
877 set_bit(R5_UPTODATE, &tgt->flags);
878 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
879 clear_bit(R5_Wantcompute, &tgt->flags);
882 static void ops_complete_compute(void *stripe_head_ref)
884 struct stripe_head *sh = stripe_head_ref;
886 pr_debug("%s: stripe %llu\n", __func__,
887 (unsigned long long)sh->sector);
889 /* mark the computed target(s) as uptodate */
890 mark_target_uptodate(sh, sh->ops.target);
891 mark_target_uptodate(sh, sh->ops.target2);
893 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
894 if (sh->check_state == check_state_compute_run)
895 sh->check_state = check_state_compute_result;
896 set_bit(STRIPE_HANDLE, &sh->state);
900 /* return a pointer to the address conversion region of the scribble buffer */
901 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
902 struct raid5_percpu *percpu)
904 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
907 static struct dma_async_tx_descriptor *
908 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
910 int disks = sh->disks;
911 struct page **xor_srcs = percpu->scribble;
912 int target = sh->ops.target;
913 struct r5dev *tgt = &sh->dev[target];
914 struct page *xor_dest = tgt->page;
916 struct dma_async_tx_descriptor *tx;
917 struct async_submit_ctl submit;
920 pr_debug("%s: stripe %llu block: %d\n",
921 __func__, (unsigned long long)sh->sector, target);
922 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
924 for (i = disks; i--; )
926 xor_srcs[count++] = sh->dev[i].page;
928 atomic_inc(&sh->count);
930 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
931 ops_complete_compute, sh, to_addr_conv(sh, percpu));
932 if (unlikely(count == 1))
933 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
935 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
940 /* set_syndrome_sources - populate source buffers for gen_syndrome
941 * @srcs - (struct page *) array of size sh->disks
942 * @sh - stripe_head to parse
944 * Populates srcs in proper layout order for the stripe and returns the
945 * 'count' of sources to be used in a call to async_gen_syndrome. The P
946 * destination buffer is recorded in srcs[count] and the Q destination
947 * is recorded in srcs[count+1]].
949 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
951 int disks = sh->disks;
952 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
953 int d0_idx = raid6_d0(sh);
957 for (i = 0; i < disks; i++)
963 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
965 srcs[slot] = sh->dev[i].page;
966 i = raid6_next_disk(i, disks);
967 } while (i != d0_idx);
969 return syndrome_disks;
972 static struct dma_async_tx_descriptor *
973 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
975 int disks = sh->disks;
976 struct page **blocks = percpu->scribble;
978 int qd_idx = sh->qd_idx;
979 struct dma_async_tx_descriptor *tx;
980 struct async_submit_ctl submit;
986 if (sh->ops.target < 0)
987 target = sh->ops.target2;
988 else if (sh->ops.target2 < 0)
989 target = sh->ops.target;
991 /* we should only have one valid target */
994 pr_debug("%s: stripe %llu block: %d\n",
995 __func__, (unsigned long long)sh->sector, target);
997 tgt = &sh->dev[target];
998 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1001 atomic_inc(&sh->count);
1003 if (target == qd_idx) {
1004 count = set_syndrome_sources(blocks, sh);
1005 blocks[count] = NULL; /* regenerating p is not necessary */
1006 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1007 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1008 ops_complete_compute, sh,
1009 to_addr_conv(sh, percpu));
1010 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1012 /* Compute any data- or p-drive using XOR */
1014 for (i = disks; i-- ; ) {
1015 if (i == target || i == qd_idx)
1017 blocks[count++] = sh->dev[i].page;
1020 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1021 NULL, ops_complete_compute, sh,
1022 to_addr_conv(sh, percpu));
1023 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1029 static struct dma_async_tx_descriptor *
1030 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1032 int i, count, disks = sh->disks;
1033 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1034 int d0_idx = raid6_d0(sh);
1035 int faila = -1, failb = -1;
1036 int target = sh->ops.target;
1037 int target2 = sh->ops.target2;
1038 struct r5dev *tgt = &sh->dev[target];
1039 struct r5dev *tgt2 = &sh->dev[target2];
1040 struct dma_async_tx_descriptor *tx;
1041 struct page **blocks = percpu->scribble;
1042 struct async_submit_ctl submit;
1044 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1045 __func__, (unsigned long long)sh->sector, target, target2);
1046 BUG_ON(target < 0 || target2 < 0);
1047 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1048 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1050 /* we need to open-code set_syndrome_sources to handle the
1051 * slot number conversion for 'faila' and 'failb'
1053 for (i = 0; i < disks ; i++)
1058 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1060 blocks[slot] = sh->dev[i].page;
1066 i = raid6_next_disk(i, disks);
1067 } while (i != d0_idx);
1069 BUG_ON(faila == failb);
1072 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1073 __func__, (unsigned long long)sh->sector, faila, failb);
1075 atomic_inc(&sh->count);
1077 if (failb == syndrome_disks+1) {
1078 /* Q disk is one of the missing disks */
1079 if (faila == syndrome_disks) {
1080 /* Missing P+Q, just recompute */
1081 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1082 ops_complete_compute, sh,
1083 to_addr_conv(sh, percpu));
1084 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1085 STRIPE_SIZE, &submit);
1089 int qd_idx = sh->qd_idx;
1091 /* Missing D+Q: recompute D from P, then recompute Q */
1092 if (target == qd_idx)
1093 data_target = target2;
1095 data_target = target;
1098 for (i = disks; i-- ; ) {
1099 if (i == data_target || i == qd_idx)
1101 blocks[count++] = sh->dev[i].page;
1103 dest = sh->dev[data_target].page;
1104 init_async_submit(&submit,
1105 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1107 to_addr_conv(sh, percpu));
1108 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1111 count = set_syndrome_sources(blocks, sh);
1112 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1113 ops_complete_compute, sh,
1114 to_addr_conv(sh, percpu));
1115 return async_gen_syndrome(blocks, 0, count+2,
1116 STRIPE_SIZE, &submit);
1119 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1120 ops_complete_compute, sh,
1121 to_addr_conv(sh, percpu));
1122 if (failb == syndrome_disks) {
1123 /* We're missing D+P. */
1124 return async_raid6_datap_recov(syndrome_disks+2,
1128 /* We're missing D+D. */
1129 return async_raid6_2data_recov(syndrome_disks+2,
1130 STRIPE_SIZE, faila, failb,
1137 static void ops_complete_prexor(void *stripe_head_ref)
1139 struct stripe_head *sh = stripe_head_ref;
1141 pr_debug("%s: stripe %llu\n", __func__,
1142 (unsigned long long)sh->sector);
1145 static struct dma_async_tx_descriptor *
1146 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1147 struct dma_async_tx_descriptor *tx)
1149 int disks = sh->disks;
1150 struct page **xor_srcs = percpu->scribble;
1151 int count = 0, pd_idx = sh->pd_idx, i;
1152 struct async_submit_ctl submit;
1154 /* existing parity data subtracted */
1155 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1157 pr_debug("%s: stripe %llu\n", __func__,
1158 (unsigned long long)sh->sector);
1160 for (i = disks; i--; ) {
1161 struct r5dev *dev = &sh->dev[i];
1162 /* Only process blocks that are known to be uptodate */
1163 if (test_bit(R5_Wantdrain, &dev->flags))
1164 xor_srcs[count++] = dev->page;
1167 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1168 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1169 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1174 static struct dma_async_tx_descriptor *
1175 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1177 int disks = sh->disks;
1180 pr_debug("%s: stripe %llu\n", __func__,
1181 (unsigned long long)sh->sector);
1183 for (i = disks; i--; ) {
1184 struct r5dev *dev = &sh->dev[i];
1187 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1190 spin_lock_irq(&sh->stripe_lock);
1191 chosen = dev->towrite;
1192 dev->towrite = NULL;
1193 BUG_ON(dev->written);
1194 wbi = dev->written = chosen;
1195 spin_unlock_irq(&sh->stripe_lock);
1197 while (wbi && wbi->bi_sector <
1198 dev->sector + STRIPE_SECTORS) {
1199 if (wbi->bi_rw & REQ_FUA)
1200 set_bit(R5_WantFUA, &dev->flags);
1201 if (wbi->bi_rw & REQ_SYNC)
1202 set_bit(R5_SyncIO, &dev->flags);
1203 if (wbi->bi_rw & REQ_DISCARD)
1204 set_bit(R5_Discard, &dev->flags);
1206 tx = async_copy_data(1, wbi, dev->page,
1208 wbi = r5_next_bio(wbi, dev->sector);
1216 static void ops_complete_reconstruct(void *stripe_head_ref)
1218 struct stripe_head *sh = stripe_head_ref;
1219 int disks = sh->disks;
1220 int pd_idx = sh->pd_idx;
1221 int qd_idx = sh->qd_idx;
1223 bool fua = false, sync = false, discard = false;
1225 pr_debug("%s: stripe %llu\n", __func__,
1226 (unsigned long long)sh->sector);
1228 for (i = disks; i--; ) {
1229 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1230 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1231 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1234 for (i = disks; i--; ) {
1235 struct r5dev *dev = &sh->dev[i];
1237 if (dev->written || i == pd_idx || i == qd_idx) {
1239 set_bit(R5_UPTODATE, &dev->flags);
1241 set_bit(R5_WantFUA, &dev->flags);
1243 set_bit(R5_SyncIO, &dev->flags);
1247 if (sh->reconstruct_state == reconstruct_state_drain_run)
1248 sh->reconstruct_state = reconstruct_state_drain_result;
1249 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1250 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1252 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1253 sh->reconstruct_state = reconstruct_state_result;
1256 set_bit(STRIPE_HANDLE, &sh->state);
1261 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1262 struct dma_async_tx_descriptor *tx)
1264 int disks = sh->disks;
1265 struct page **xor_srcs = percpu->scribble;
1266 struct async_submit_ctl submit;
1267 int count = 0, pd_idx = sh->pd_idx, i;
1268 struct page *xor_dest;
1270 unsigned long flags;
1272 pr_debug("%s: stripe %llu\n", __func__,
1273 (unsigned long long)sh->sector);
1275 for (i = 0; i < sh->disks; i++) {
1278 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1281 if (i >= sh->disks) {
1282 atomic_inc(&sh->count);
1283 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1284 ops_complete_reconstruct(sh);
1287 /* check if prexor is active which means only process blocks
1288 * that are part of a read-modify-write (written)
1290 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1292 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1293 for (i = disks; i--; ) {
1294 struct r5dev *dev = &sh->dev[i];
1296 xor_srcs[count++] = dev->page;
1299 xor_dest = sh->dev[pd_idx].page;
1300 for (i = disks; i--; ) {
1301 struct r5dev *dev = &sh->dev[i];
1303 xor_srcs[count++] = dev->page;
1307 /* 1/ if we prexor'd then the dest is reused as a source
1308 * 2/ if we did not prexor then we are redoing the parity
1309 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1310 * for the synchronous xor case
1312 flags = ASYNC_TX_ACK |
1313 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1315 atomic_inc(&sh->count);
1317 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1318 to_addr_conv(sh, percpu));
1319 if (unlikely(count == 1))
1320 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1322 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1326 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1327 struct dma_async_tx_descriptor *tx)
1329 struct async_submit_ctl submit;
1330 struct page **blocks = percpu->scribble;
1333 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1335 for (i = 0; i < sh->disks; i++) {
1336 if (sh->pd_idx == i || sh->qd_idx == i)
1338 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1341 if (i >= sh->disks) {
1342 atomic_inc(&sh->count);
1343 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1344 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1345 ops_complete_reconstruct(sh);
1349 count = set_syndrome_sources(blocks, sh);
1351 atomic_inc(&sh->count);
1353 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1354 sh, to_addr_conv(sh, percpu));
1355 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1358 static void ops_complete_check(void *stripe_head_ref)
1360 struct stripe_head *sh = stripe_head_ref;
1362 pr_debug("%s: stripe %llu\n", __func__,
1363 (unsigned long long)sh->sector);
1365 sh->check_state = check_state_check_result;
1366 set_bit(STRIPE_HANDLE, &sh->state);
1370 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1372 int disks = sh->disks;
1373 int pd_idx = sh->pd_idx;
1374 int qd_idx = sh->qd_idx;
1375 struct page *xor_dest;
1376 struct page **xor_srcs = percpu->scribble;
1377 struct dma_async_tx_descriptor *tx;
1378 struct async_submit_ctl submit;
1382 pr_debug("%s: stripe %llu\n", __func__,
1383 (unsigned long long)sh->sector);
1386 xor_dest = sh->dev[pd_idx].page;
1387 xor_srcs[count++] = xor_dest;
1388 for (i = disks; i--; ) {
1389 if (i == pd_idx || i == qd_idx)
1391 xor_srcs[count++] = sh->dev[i].page;
1394 init_async_submit(&submit, 0, NULL, NULL, NULL,
1395 to_addr_conv(sh, percpu));
1396 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1397 &sh->ops.zero_sum_result, &submit);
1399 atomic_inc(&sh->count);
1400 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1401 tx = async_trigger_callback(&submit);
1404 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1406 struct page **srcs = percpu->scribble;
1407 struct async_submit_ctl submit;
1410 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1411 (unsigned long long)sh->sector, checkp);
1413 count = set_syndrome_sources(srcs, sh);
1417 atomic_inc(&sh->count);
1418 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1419 sh, to_addr_conv(sh, percpu));
1420 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1421 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1424 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1426 int overlap_clear = 0, i, disks = sh->disks;
1427 struct dma_async_tx_descriptor *tx = NULL;
1428 struct r5conf *conf = sh->raid_conf;
1429 int level = conf->level;
1430 struct raid5_percpu *percpu;
1434 percpu = per_cpu_ptr(conf->percpu, cpu);
1435 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1436 ops_run_biofill(sh);
1440 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1442 tx = ops_run_compute5(sh, percpu);
1444 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1445 tx = ops_run_compute6_1(sh, percpu);
1447 tx = ops_run_compute6_2(sh, percpu);
1449 /* terminate the chain if reconstruct is not set to be run */
1450 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1454 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1455 tx = ops_run_prexor(sh, percpu, tx);
1457 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1458 tx = ops_run_biodrain(sh, tx);
1462 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1464 ops_run_reconstruct5(sh, percpu, tx);
1466 ops_run_reconstruct6(sh, percpu, tx);
1469 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1470 if (sh->check_state == check_state_run)
1471 ops_run_check_p(sh, percpu);
1472 else if (sh->check_state == check_state_run_q)
1473 ops_run_check_pq(sh, percpu, 0);
1474 else if (sh->check_state == check_state_run_pq)
1475 ops_run_check_pq(sh, percpu, 1);
1481 for (i = disks; i--; ) {
1482 struct r5dev *dev = &sh->dev[i];
1483 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1484 wake_up(&sh->raid_conf->wait_for_overlap);
1489 static int grow_one_stripe(struct r5conf *conf)
1491 struct stripe_head *sh;
1492 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1496 sh->raid_conf = conf;
1498 spin_lock_init(&sh->stripe_lock);
1500 if (grow_buffers(sh)) {
1502 kmem_cache_free(conf->slab_cache, sh);
1505 /* we just created an active stripe so... */
1506 atomic_set(&sh->count, 1);
1507 atomic_inc(&conf->active_stripes);
1508 INIT_LIST_HEAD(&sh->lru);
1513 static int grow_stripes(struct r5conf *conf, int num)
1515 struct kmem_cache *sc;
1516 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1518 if (conf->mddev->gendisk)
1519 sprintf(conf->cache_name[0],
1520 "raid%d-%s", conf->level, mdname(conf->mddev));
1522 sprintf(conf->cache_name[0],
1523 "raid%d-%p", conf->level, conf->mddev);
1524 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1526 conf->active_name = 0;
1527 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1528 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1532 conf->slab_cache = sc;
1533 conf->pool_size = devs;
1535 if (!grow_one_stripe(conf))
1541 * scribble_len - return the required size of the scribble region
1542 * @num - total number of disks in the array
1544 * The size must be enough to contain:
1545 * 1/ a struct page pointer for each device in the array +2
1546 * 2/ room to convert each entry in (1) to its corresponding dma
1547 * (dma_map_page()) or page (page_address()) address.
1549 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1550 * calculate over all devices (not just the data blocks), using zeros in place
1551 * of the P and Q blocks.
1553 static size_t scribble_len(int num)
1557 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1562 static int resize_stripes(struct r5conf *conf, int newsize)
1564 /* Make all the stripes able to hold 'newsize' devices.
1565 * New slots in each stripe get 'page' set to a new page.
1567 * This happens in stages:
1568 * 1/ create a new kmem_cache and allocate the required number of
1570 * 2/ gather all the old stripe_heads and transfer the pages across
1571 * to the new stripe_heads. This will have the side effect of
1572 * freezing the array as once all stripe_heads have been collected,
1573 * no IO will be possible. Old stripe heads are freed once their
1574 * pages have been transferred over, and the old kmem_cache is
1575 * freed when all stripes are done.
1576 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1577 * we simple return a failre status - no need to clean anything up.
1578 * 4/ allocate new pages for the new slots in the new stripe_heads.
1579 * If this fails, we don't bother trying the shrink the
1580 * stripe_heads down again, we just leave them as they are.
1581 * As each stripe_head is processed the new one is released into
1584 * Once step2 is started, we cannot afford to wait for a write,
1585 * so we use GFP_NOIO allocations.
1587 struct stripe_head *osh, *nsh;
1588 LIST_HEAD(newstripes);
1589 struct disk_info *ndisks;
1592 struct kmem_cache *sc;
1595 if (newsize <= conf->pool_size)
1596 return 0; /* never bother to shrink */
1598 err = md_allow_write(conf->mddev);
1603 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1604 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1609 for (i = conf->max_nr_stripes; i; i--) {
1610 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1614 nsh->raid_conf = conf;
1615 spin_lock_init(&nsh->stripe_lock);
1617 list_add(&nsh->lru, &newstripes);
1620 /* didn't get enough, give up */
1621 while (!list_empty(&newstripes)) {
1622 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1623 list_del(&nsh->lru);
1624 kmem_cache_free(sc, nsh);
1626 kmem_cache_destroy(sc);
1629 /* Step 2 - Must use GFP_NOIO now.
1630 * OK, we have enough stripes, start collecting inactive
1631 * stripes and copying them over
1633 list_for_each_entry(nsh, &newstripes, lru) {
1634 spin_lock_irq(&conf->device_lock);
1635 wait_event_lock_irq(conf->wait_for_stripe,
1636 !list_empty(&conf->inactive_list),
1638 osh = get_free_stripe(conf);
1639 spin_unlock_irq(&conf->device_lock);
1640 atomic_set(&nsh->count, 1);
1641 for(i=0; i<conf->pool_size; i++)
1642 nsh->dev[i].page = osh->dev[i].page;
1643 for( ; i<newsize; i++)
1644 nsh->dev[i].page = NULL;
1645 kmem_cache_free(conf->slab_cache, osh);
1647 kmem_cache_destroy(conf->slab_cache);
1650 * At this point, we are holding all the stripes so the array
1651 * is completely stalled, so now is a good time to resize
1652 * conf->disks and the scribble region
1654 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1656 for (i=0; i<conf->raid_disks; i++)
1657 ndisks[i] = conf->disks[i];
1659 conf->disks = ndisks;
1664 conf->scribble_len = scribble_len(newsize);
1665 for_each_present_cpu(cpu) {
1666 struct raid5_percpu *percpu;
1669 percpu = per_cpu_ptr(conf->percpu, cpu);
1670 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1673 kfree(percpu->scribble);
1674 percpu->scribble = scribble;
1682 /* Step 4, return new stripes to service */
1683 while(!list_empty(&newstripes)) {
1684 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1685 list_del_init(&nsh->lru);
1687 for (i=conf->raid_disks; i < newsize; i++)
1688 if (nsh->dev[i].page == NULL) {
1689 struct page *p = alloc_page(GFP_NOIO);
1690 nsh->dev[i].page = p;
1694 release_stripe(nsh);
1696 /* critical section pass, GFP_NOIO no longer needed */
1698 conf->slab_cache = sc;
1699 conf->active_name = 1-conf->active_name;
1700 conf->pool_size = newsize;
1704 static int drop_one_stripe(struct r5conf *conf)
1706 struct stripe_head *sh;
1708 spin_lock_irq(&conf->device_lock);
1709 sh = get_free_stripe(conf);
1710 spin_unlock_irq(&conf->device_lock);
1713 BUG_ON(atomic_read(&sh->count));
1715 kmem_cache_free(conf->slab_cache, sh);
1716 atomic_dec(&conf->active_stripes);
1720 static void shrink_stripes(struct r5conf *conf)
1722 while (drop_one_stripe(conf))
1725 if (conf->slab_cache)
1726 kmem_cache_destroy(conf->slab_cache);
1727 conf->slab_cache = NULL;
1730 static void raid5_end_read_request(struct bio * bi, int error)
1732 struct stripe_head *sh = bi->bi_private;
1733 struct r5conf *conf = sh->raid_conf;
1734 int disks = sh->disks, i;
1735 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1736 char b[BDEVNAME_SIZE];
1737 struct md_rdev *rdev = NULL;
1740 for (i=0 ; i<disks; i++)
1741 if (bi == &sh->dev[i].req)
1744 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1745 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1751 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1752 /* If replacement finished while this request was outstanding,
1753 * 'replacement' might be NULL already.
1754 * In that case it moved down to 'rdev'.
1755 * rdev is not removed until all requests are finished.
1757 rdev = conf->disks[i].replacement;
1759 rdev = conf->disks[i].rdev;
1761 if (use_new_offset(conf, sh))
1762 s = sh->sector + rdev->new_data_offset;
1764 s = sh->sector + rdev->data_offset;
1766 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1767 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1768 /* Note that this cannot happen on a
1769 * replacement device. We just fail those on
1774 "md/raid:%s: read error corrected"
1775 " (%lu sectors at %llu on %s)\n",
1776 mdname(conf->mddev), STRIPE_SECTORS,
1777 (unsigned long long)s,
1778 bdevname(rdev->bdev, b));
1779 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1780 clear_bit(R5_ReadError, &sh->dev[i].flags);
1781 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1782 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1783 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1785 if (atomic_read(&rdev->read_errors))
1786 atomic_set(&rdev->read_errors, 0);
1788 const char *bdn = bdevname(rdev->bdev, b);
1792 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1793 atomic_inc(&rdev->read_errors);
1794 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1797 "md/raid:%s: read error on replacement device "
1798 "(sector %llu on %s).\n",
1799 mdname(conf->mddev),
1800 (unsigned long long)s,
1802 else if (conf->mddev->degraded >= conf->max_degraded) {
1806 "md/raid:%s: read error not correctable "
1807 "(sector %llu on %s).\n",
1808 mdname(conf->mddev),
1809 (unsigned long long)s,
1811 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1816 "md/raid:%s: read error NOT corrected!! "
1817 "(sector %llu on %s).\n",
1818 mdname(conf->mddev),
1819 (unsigned long long)s,
1821 } else if (atomic_read(&rdev->read_errors)
1822 > conf->max_nr_stripes)
1824 "md/raid:%s: Too many read errors, failing device %s.\n",
1825 mdname(conf->mddev), bdn);
1829 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1830 set_bit(R5_ReadError, &sh->dev[i].flags);
1831 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1833 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1835 clear_bit(R5_ReadError, &sh->dev[i].flags);
1836 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1838 && test_bit(In_sync, &rdev->flags)
1839 && rdev_set_badblocks(
1840 rdev, sh->sector, STRIPE_SECTORS, 0)))
1841 md_error(conf->mddev, rdev);
1844 rdev_dec_pending(rdev, conf->mddev);
1845 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1846 set_bit(STRIPE_HANDLE, &sh->state);
1850 static void raid5_end_write_request(struct bio *bi, int error)
1852 struct stripe_head *sh = bi->bi_private;
1853 struct r5conf *conf = sh->raid_conf;
1854 int disks = sh->disks, i;
1855 struct md_rdev *uninitialized_var(rdev);
1856 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1859 int replacement = 0;
1861 for (i = 0 ; i < disks; i++) {
1862 if (bi == &sh->dev[i].req) {
1863 rdev = conf->disks[i].rdev;
1866 if (bi == &sh->dev[i].rreq) {
1867 rdev = conf->disks[i].replacement;
1871 /* rdev was removed and 'replacement'
1872 * replaced it. rdev is not removed
1873 * until all requests are finished.
1875 rdev = conf->disks[i].rdev;
1879 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1880 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1889 md_error(conf->mddev, rdev);
1890 else if (is_badblock(rdev, sh->sector,
1892 &first_bad, &bad_sectors))
1893 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1896 set_bit(WriteErrorSeen, &rdev->flags);
1897 set_bit(R5_WriteError, &sh->dev[i].flags);
1898 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1899 set_bit(MD_RECOVERY_NEEDED,
1900 &rdev->mddev->recovery);
1901 } else if (is_badblock(rdev, sh->sector,
1903 &first_bad, &bad_sectors)) {
1904 set_bit(R5_MadeGood, &sh->dev[i].flags);
1905 if (test_bit(R5_ReadError, &sh->dev[i].flags))
1906 /* That was a successful write so make
1907 * sure it looks like we already did
1910 set_bit(R5_ReWrite, &sh->dev[i].flags);
1913 rdev_dec_pending(rdev, conf->mddev);
1915 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1916 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1917 set_bit(STRIPE_HANDLE, &sh->state);
1921 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1923 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1925 struct r5dev *dev = &sh->dev[i];
1927 bio_init(&dev->req);
1928 dev->req.bi_io_vec = &dev->vec;
1930 dev->req.bi_max_vecs++;
1931 dev->req.bi_private = sh;
1932 dev->vec.bv_page = dev->page;
1934 bio_init(&dev->rreq);
1935 dev->rreq.bi_io_vec = &dev->rvec;
1936 dev->rreq.bi_vcnt++;
1937 dev->rreq.bi_max_vecs++;
1938 dev->rreq.bi_private = sh;
1939 dev->rvec.bv_page = dev->page;
1942 dev->sector = compute_blocknr(sh, i, previous);
1945 static void error(struct mddev *mddev, struct md_rdev *rdev)
1947 char b[BDEVNAME_SIZE];
1948 struct r5conf *conf = mddev->private;
1949 unsigned long flags;
1950 pr_debug("raid456: error called\n");
1952 spin_lock_irqsave(&conf->device_lock, flags);
1953 clear_bit(In_sync, &rdev->flags);
1954 mddev->degraded = calc_degraded(conf);
1955 spin_unlock_irqrestore(&conf->device_lock, flags);
1956 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1958 set_bit(Blocked, &rdev->flags);
1959 set_bit(Faulty, &rdev->flags);
1960 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1962 "md/raid:%s: Disk failure on %s, disabling device.\n"
1963 "md/raid:%s: Operation continuing on %d devices.\n",
1965 bdevname(rdev->bdev, b),
1967 conf->raid_disks - mddev->degraded);
1971 * Input: a 'big' sector number,
1972 * Output: index of the data and parity disk, and the sector # in them.
1974 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1975 int previous, int *dd_idx,
1976 struct stripe_head *sh)
1978 sector_t stripe, stripe2;
1979 sector_t chunk_number;
1980 unsigned int chunk_offset;
1983 sector_t new_sector;
1984 int algorithm = previous ? conf->prev_algo
1986 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1987 : conf->chunk_sectors;
1988 int raid_disks = previous ? conf->previous_raid_disks
1990 int data_disks = raid_disks - conf->max_degraded;
1992 /* First compute the information on this sector */
1995 * Compute the chunk number and the sector offset inside the chunk
1997 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1998 chunk_number = r_sector;
2001 * Compute the stripe number
2003 stripe = chunk_number;
2004 *dd_idx = sector_div(stripe, data_disks);
2007 * Select the parity disk based on the user selected algorithm.
2009 pd_idx = qd_idx = -1;
2010 switch(conf->level) {
2012 pd_idx = data_disks;
2015 switch (algorithm) {
2016 case ALGORITHM_LEFT_ASYMMETRIC:
2017 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2018 if (*dd_idx >= pd_idx)
2021 case ALGORITHM_RIGHT_ASYMMETRIC:
2022 pd_idx = sector_div(stripe2, raid_disks);
2023 if (*dd_idx >= pd_idx)
2026 case ALGORITHM_LEFT_SYMMETRIC:
2027 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2028 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2030 case ALGORITHM_RIGHT_SYMMETRIC:
2031 pd_idx = sector_div(stripe2, raid_disks);
2032 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2034 case ALGORITHM_PARITY_0:
2038 case ALGORITHM_PARITY_N:
2039 pd_idx = data_disks;
2047 switch (algorithm) {
2048 case ALGORITHM_LEFT_ASYMMETRIC:
2049 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2050 qd_idx = pd_idx + 1;
2051 if (pd_idx == raid_disks-1) {
2052 (*dd_idx)++; /* Q D D D P */
2054 } else if (*dd_idx >= pd_idx)
2055 (*dd_idx) += 2; /* D D P Q D */
2057 case ALGORITHM_RIGHT_ASYMMETRIC:
2058 pd_idx = sector_div(stripe2, raid_disks);
2059 qd_idx = pd_idx + 1;
2060 if (pd_idx == raid_disks-1) {
2061 (*dd_idx)++; /* Q D D D P */
2063 } else if (*dd_idx >= pd_idx)
2064 (*dd_idx) += 2; /* D D P Q D */
2066 case ALGORITHM_LEFT_SYMMETRIC:
2067 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2068 qd_idx = (pd_idx + 1) % raid_disks;
2069 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2071 case ALGORITHM_RIGHT_SYMMETRIC:
2072 pd_idx = sector_div(stripe2, raid_disks);
2073 qd_idx = (pd_idx + 1) % raid_disks;
2074 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2077 case ALGORITHM_PARITY_0:
2082 case ALGORITHM_PARITY_N:
2083 pd_idx = data_disks;
2084 qd_idx = data_disks + 1;
2087 case ALGORITHM_ROTATING_ZERO_RESTART:
2088 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2089 * of blocks for computing Q is different.
2091 pd_idx = sector_div(stripe2, raid_disks);
2092 qd_idx = pd_idx + 1;
2093 if (pd_idx == raid_disks-1) {
2094 (*dd_idx)++; /* Q D D D P */
2096 } else if (*dd_idx >= pd_idx)
2097 (*dd_idx) += 2; /* D D P Q D */
2101 case ALGORITHM_ROTATING_N_RESTART:
2102 /* Same a left_asymmetric, by first stripe is
2103 * D D D P Q rather than
2107 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2108 qd_idx = pd_idx + 1;
2109 if (pd_idx == raid_disks-1) {
2110 (*dd_idx)++; /* Q D D D P */
2112 } else if (*dd_idx >= pd_idx)
2113 (*dd_idx) += 2; /* D D P Q D */
2117 case ALGORITHM_ROTATING_N_CONTINUE:
2118 /* Same as left_symmetric but Q is before P */
2119 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2120 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2121 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2125 case ALGORITHM_LEFT_ASYMMETRIC_6:
2126 /* RAID5 left_asymmetric, with Q on last device */
2127 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2128 if (*dd_idx >= pd_idx)
2130 qd_idx = raid_disks - 1;
2133 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2134 pd_idx = sector_div(stripe2, raid_disks-1);
2135 if (*dd_idx >= pd_idx)
2137 qd_idx = raid_disks - 1;
2140 case ALGORITHM_LEFT_SYMMETRIC_6:
2141 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2142 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2143 qd_idx = raid_disks - 1;
2146 case ALGORITHM_RIGHT_SYMMETRIC_6:
2147 pd_idx = sector_div(stripe2, raid_disks-1);
2148 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2149 qd_idx = raid_disks - 1;
2152 case ALGORITHM_PARITY_0_6:
2155 qd_idx = raid_disks - 1;
2165 sh->pd_idx = pd_idx;
2166 sh->qd_idx = qd_idx;
2167 sh->ddf_layout = ddf_layout;
2170 * Finally, compute the new sector number
2172 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2177 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2179 struct r5conf *conf = sh->raid_conf;
2180 int raid_disks = sh->disks;
2181 int data_disks = raid_disks - conf->max_degraded;
2182 sector_t new_sector = sh->sector, check;
2183 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2184 : conf->chunk_sectors;
2185 int algorithm = previous ? conf->prev_algo
2189 sector_t chunk_number;
2190 int dummy1, dd_idx = i;
2192 struct stripe_head sh2;
2195 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2196 stripe = new_sector;
2198 if (i == sh->pd_idx)
2200 switch(conf->level) {
2203 switch (algorithm) {
2204 case ALGORITHM_LEFT_ASYMMETRIC:
2205 case ALGORITHM_RIGHT_ASYMMETRIC:
2209 case ALGORITHM_LEFT_SYMMETRIC:
2210 case ALGORITHM_RIGHT_SYMMETRIC:
2213 i -= (sh->pd_idx + 1);
2215 case ALGORITHM_PARITY_0:
2218 case ALGORITHM_PARITY_N:
2225 if (i == sh->qd_idx)
2226 return 0; /* It is the Q disk */
2227 switch (algorithm) {
2228 case ALGORITHM_LEFT_ASYMMETRIC:
2229 case ALGORITHM_RIGHT_ASYMMETRIC:
2230 case ALGORITHM_ROTATING_ZERO_RESTART:
2231 case ALGORITHM_ROTATING_N_RESTART:
2232 if (sh->pd_idx == raid_disks-1)
2233 i--; /* Q D D D P */
2234 else if (i > sh->pd_idx)
2235 i -= 2; /* D D P Q D */
2237 case ALGORITHM_LEFT_SYMMETRIC:
2238 case ALGORITHM_RIGHT_SYMMETRIC:
2239 if (sh->pd_idx == raid_disks-1)
2240 i--; /* Q D D D P */
2245 i -= (sh->pd_idx + 2);
2248 case ALGORITHM_PARITY_0:
2251 case ALGORITHM_PARITY_N:
2253 case ALGORITHM_ROTATING_N_CONTINUE:
2254 /* Like left_symmetric, but P is before Q */
2255 if (sh->pd_idx == 0)
2256 i--; /* P D D D Q */
2261 i -= (sh->pd_idx + 1);
2264 case ALGORITHM_LEFT_ASYMMETRIC_6:
2265 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2269 case ALGORITHM_LEFT_SYMMETRIC_6:
2270 case ALGORITHM_RIGHT_SYMMETRIC_6:
2272 i += data_disks + 1;
2273 i -= (sh->pd_idx + 1);
2275 case ALGORITHM_PARITY_0_6:
2284 chunk_number = stripe * data_disks + i;
2285 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2287 check = raid5_compute_sector(conf, r_sector,
2288 previous, &dummy1, &sh2);
2289 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2290 || sh2.qd_idx != sh->qd_idx) {
2291 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2292 mdname(conf->mddev));
2300 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2301 int rcw, int expand)
2303 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2304 struct r5conf *conf = sh->raid_conf;
2305 int level = conf->level;
2309 for (i = disks; i--; ) {
2310 struct r5dev *dev = &sh->dev[i];
2313 set_bit(R5_LOCKED, &dev->flags);
2314 set_bit(R5_Wantdrain, &dev->flags);
2316 clear_bit(R5_UPTODATE, &dev->flags);
2320 /* if we are not expanding this is a proper write request, and
2321 * there will be bios with new data to be drained into the
2326 /* False alarm, nothing to do */
2328 sh->reconstruct_state = reconstruct_state_drain_run;
2329 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2331 sh->reconstruct_state = reconstruct_state_run;
2333 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2335 if (s->locked + conf->max_degraded == disks)
2336 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2337 atomic_inc(&conf->pending_full_writes);
2340 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2341 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2343 for (i = disks; i--; ) {
2344 struct r5dev *dev = &sh->dev[i];
2349 (test_bit(R5_UPTODATE, &dev->flags) ||
2350 test_bit(R5_Wantcompute, &dev->flags))) {
2351 set_bit(R5_Wantdrain, &dev->flags);
2352 set_bit(R5_LOCKED, &dev->flags);
2353 clear_bit(R5_UPTODATE, &dev->flags);
2358 /* False alarm - nothing to do */
2360 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2361 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2362 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2363 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2366 /* keep the parity disk(s) locked while asynchronous operations
2369 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2370 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2374 int qd_idx = sh->qd_idx;
2375 struct r5dev *dev = &sh->dev[qd_idx];
2377 set_bit(R5_LOCKED, &dev->flags);
2378 clear_bit(R5_UPTODATE, &dev->flags);
2382 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2383 __func__, (unsigned long long)sh->sector,
2384 s->locked, s->ops_request);
2388 * Each stripe/dev can have one or more bion attached.
2389 * toread/towrite point to the first in a chain.
2390 * The bi_next chain must be in order.
2392 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2395 struct r5conf *conf = sh->raid_conf;
2398 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2399 (unsigned long long)bi->bi_sector,
2400 (unsigned long long)sh->sector);
2403 * If several bio share a stripe. The bio bi_phys_segments acts as a
2404 * reference count to avoid race. The reference count should already be
2405 * increased before this function is called (for example, in
2406 * make_request()), so other bio sharing this stripe will not free the
2407 * stripe. If a stripe is owned by one stripe, the stripe lock will
2410 spin_lock_irq(&sh->stripe_lock);
2412 bip = &sh->dev[dd_idx].towrite;
2416 bip = &sh->dev[dd_idx].toread;
2417 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2418 if (bio_end_sector(*bip) > bi->bi_sector)
2420 bip = & (*bip)->bi_next;
2422 if (*bip && (*bip)->bi_sector < bio_end_sector(bi))
2425 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2429 raid5_inc_bi_active_stripes(bi);
2432 /* check if page is covered */
2433 sector_t sector = sh->dev[dd_idx].sector;
2434 for (bi=sh->dev[dd_idx].towrite;
2435 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2436 bi && bi->bi_sector <= sector;
2437 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2438 if (bio_end_sector(bi) >= sector)
2439 sector = bio_end_sector(bi);
2441 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2442 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2445 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2446 (unsigned long long)(*bip)->bi_sector,
2447 (unsigned long long)sh->sector, dd_idx);
2448 spin_unlock_irq(&sh->stripe_lock);
2450 if (conf->mddev->bitmap && firstwrite) {
2451 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2453 sh->bm_seq = conf->seq_flush+1;
2454 set_bit(STRIPE_BIT_DELAY, &sh->state);
2459 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2460 spin_unlock_irq(&sh->stripe_lock);
2464 static void end_reshape(struct r5conf *conf);
2466 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2467 struct stripe_head *sh)
2469 int sectors_per_chunk =
2470 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2472 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2473 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2475 raid5_compute_sector(conf,
2476 stripe * (disks - conf->max_degraded)
2477 *sectors_per_chunk + chunk_offset,
2483 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2484 struct stripe_head_state *s, int disks,
2485 struct bio **return_bi)
2488 for (i = disks; i--; ) {
2492 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2493 struct md_rdev *rdev;
2495 rdev = rcu_dereference(conf->disks[i].rdev);
2496 if (rdev && test_bit(In_sync, &rdev->flags))
2497 atomic_inc(&rdev->nr_pending);
2502 if (!rdev_set_badblocks(
2506 md_error(conf->mddev, rdev);
2507 rdev_dec_pending(rdev, conf->mddev);
2510 spin_lock_irq(&sh->stripe_lock);
2511 /* fail all writes first */
2512 bi = sh->dev[i].towrite;
2513 sh->dev[i].towrite = NULL;
2514 spin_unlock_irq(&sh->stripe_lock);
2518 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2519 wake_up(&conf->wait_for_overlap);
2521 while (bi && bi->bi_sector <
2522 sh->dev[i].sector + STRIPE_SECTORS) {
2523 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2524 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2525 if (!raid5_dec_bi_active_stripes(bi)) {
2526 md_write_end(conf->mddev);
2527 bi->bi_next = *return_bi;
2533 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2534 STRIPE_SECTORS, 0, 0);
2536 /* and fail all 'written' */
2537 bi = sh->dev[i].written;
2538 sh->dev[i].written = NULL;
2539 if (bi) bitmap_end = 1;
2540 while (bi && bi->bi_sector <
2541 sh->dev[i].sector + STRIPE_SECTORS) {
2542 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2543 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2544 if (!raid5_dec_bi_active_stripes(bi)) {
2545 md_write_end(conf->mddev);
2546 bi->bi_next = *return_bi;
2552 /* fail any reads if this device is non-operational and
2553 * the data has not reached the cache yet.
2555 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2556 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2557 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2558 spin_lock_irq(&sh->stripe_lock);
2559 bi = sh->dev[i].toread;
2560 sh->dev[i].toread = NULL;
2561 spin_unlock_irq(&sh->stripe_lock);
2562 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2563 wake_up(&conf->wait_for_overlap);
2564 while (bi && bi->bi_sector <
2565 sh->dev[i].sector + STRIPE_SECTORS) {
2566 struct bio *nextbi =
2567 r5_next_bio(bi, sh->dev[i].sector);
2568 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2569 if (!raid5_dec_bi_active_stripes(bi)) {
2570 bi->bi_next = *return_bi;
2577 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2578 STRIPE_SECTORS, 0, 0);
2579 /* If we were in the middle of a write the parity block might
2580 * still be locked - so just clear all R5_LOCKED flags
2582 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2585 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2586 if (atomic_dec_and_test(&conf->pending_full_writes))
2587 md_wakeup_thread(conf->mddev->thread);
2591 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2592 struct stripe_head_state *s)
2597 clear_bit(STRIPE_SYNCING, &sh->state);
2598 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2599 wake_up(&conf->wait_for_overlap);
2602 /* There is nothing more to do for sync/check/repair.
2603 * Don't even need to abort as that is handled elsewhere
2604 * if needed, and not always wanted e.g. if there is a known
2606 * For recover/replace we need to record a bad block on all
2607 * non-sync devices, or abort the recovery
2609 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2610 /* During recovery devices cannot be removed, so
2611 * locking and refcounting of rdevs is not needed
2613 for (i = 0; i < conf->raid_disks; i++) {
2614 struct md_rdev *rdev = conf->disks[i].rdev;
2616 && !test_bit(Faulty, &rdev->flags)
2617 && !test_bit(In_sync, &rdev->flags)
2618 && !rdev_set_badblocks(rdev, sh->sector,
2621 rdev = conf->disks[i].replacement;
2623 && !test_bit(Faulty, &rdev->flags)
2624 && !test_bit(In_sync, &rdev->flags)
2625 && !rdev_set_badblocks(rdev, sh->sector,
2630 conf->recovery_disabled =
2631 conf->mddev->recovery_disabled;
2633 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2636 static int want_replace(struct stripe_head *sh, int disk_idx)
2638 struct md_rdev *rdev;
2640 /* Doing recovery so rcu locking not required */
2641 rdev = sh->raid_conf->disks[disk_idx].replacement;
2643 && !test_bit(Faulty, &rdev->flags)
2644 && !test_bit(In_sync, &rdev->flags)
2645 && (rdev->recovery_offset <= sh->sector
2646 || rdev->mddev->recovery_cp <= sh->sector))
2652 /* fetch_block - checks the given member device to see if its data needs
2653 * to be read or computed to satisfy a request.
2655 * Returns 1 when no more member devices need to be checked, otherwise returns
2656 * 0 to tell the loop in handle_stripe_fill to continue
2658 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2659 int disk_idx, int disks)
2661 struct r5dev *dev = &sh->dev[disk_idx];
2662 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2663 &sh->dev[s->failed_num[1]] };
2665 /* is the data in this block needed, and can we get it? */
2666 if (!test_bit(R5_LOCKED, &dev->flags) &&
2667 !test_bit(R5_UPTODATE, &dev->flags) &&
2669 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2670 s->syncing || s->expanding ||
2671 (s->replacing && want_replace(sh, disk_idx)) ||
2672 (s->failed >= 1 && fdev[0]->toread) ||
2673 (s->failed >= 2 && fdev[1]->toread) ||
2674 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2675 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2676 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2677 /* we would like to get this block, possibly by computing it,
2678 * otherwise read it if the backing disk is insync
2680 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2681 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2682 if ((s->uptodate == disks - 1) &&
2683 (s->failed && (disk_idx == s->failed_num[0] ||
2684 disk_idx == s->failed_num[1]))) {
2685 /* have disk failed, and we're requested to fetch it;
2688 pr_debug("Computing stripe %llu block %d\n",
2689 (unsigned long long)sh->sector, disk_idx);
2690 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2691 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2692 set_bit(R5_Wantcompute, &dev->flags);
2693 sh->ops.target = disk_idx;
2694 sh->ops.target2 = -1; /* no 2nd target */
2696 /* Careful: from this point on 'uptodate' is in the eye
2697 * of raid_run_ops which services 'compute' operations
2698 * before writes. R5_Wantcompute flags a block that will
2699 * be R5_UPTODATE by the time it is needed for a
2700 * subsequent operation.
2704 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2705 /* Computing 2-failure is *very* expensive; only
2706 * do it if failed >= 2
2709 for (other = disks; other--; ) {
2710 if (other == disk_idx)
2712 if (!test_bit(R5_UPTODATE,
2713 &sh->dev[other].flags))
2717 pr_debug("Computing stripe %llu blocks %d,%d\n",
2718 (unsigned long long)sh->sector,
2720 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2721 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2722 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2723 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2724 sh->ops.target = disk_idx;
2725 sh->ops.target2 = other;
2729 } else if (test_bit(R5_Insync, &dev->flags)) {
2730 set_bit(R5_LOCKED, &dev->flags);
2731 set_bit(R5_Wantread, &dev->flags);
2733 pr_debug("Reading block %d (sync=%d)\n",
2734 disk_idx, s->syncing);
2742 * handle_stripe_fill - read or compute data to satisfy pending requests.
2744 static void handle_stripe_fill(struct stripe_head *sh,
2745 struct stripe_head_state *s,
2750 /* look for blocks to read/compute, skip this if a compute
2751 * is already in flight, or if the stripe contents are in the
2752 * midst of changing due to a write
2754 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2755 !sh->reconstruct_state)
2756 for (i = disks; i--; )
2757 if (fetch_block(sh, s, i, disks))
2759 set_bit(STRIPE_HANDLE, &sh->state);
2763 /* handle_stripe_clean_event
2764 * any written block on an uptodate or failed drive can be returned.
2765 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2766 * never LOCKED, so we don't need to test 'failed' directly.
2768 static void handle_stripe_clean_event(struct r5conf *conf,
2769 struct stripe_head *sh, int disks, struct bio **return_bi)
2773 int discard_pending = 0;
2775 for (i = disks; i--; )
2776 if (sh->dev[i].written) {
2778 if (!test_bit(R5_LOCKED, &dev->flags) &&
2779 (test_bit(R5_UPTODATE, &dev->flags) ||
2780 test_bit(R5_Discard, &dev->flags))) {
2781 /* We can return any write requests */
2782 struct bio *wbi, *wbi2;
2783 pr_debug("Return write for disc %d\n", i);
2784 if (test_and_clear_bit(R5_Discard, &dev->flags))
2785 clear_bit(R5_UPTODATE, &dev->flags);
2787 dev->written = NULL;
2788 while (wbi && wbi->bi_sector <
2789 dev->sector + STRIPE_SECTORS) {
2790 wbi2 = r5_next_bio(wbi, dev->sector);
2791 if (!raid5_dec_bi_active_stripes(wbi)) {
2792 md_write_end(conf->mddev);
2793 wbi->bi_next = *return_bi;
2798 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2800 !test_bit(STRIPE_DEGRADED, &sh->state),
2802 } else if (test_bit(R5_Discard, &dev->flags))
2803 discard_pending = 1;
2805 if (!discard_pending &&
2806 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
2807 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2808 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2809 if (sh->qd_idx >= 0) {
2810 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2811 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
2813 /* now that discard is done we can proceed with any sync */
2814 clear_bit(STRIPE_DISCARD, &sh->state);
2816 * SCSI discard will change some bio fields and the stripe has
2817 * no updated data, so remove it from hash list and the stripe
2818 * will be reinitialized
2820 spin_lock_irq(&conf->device_lock);
2822 spin_unlock_irq(&conf->device_lock);
2823 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2824 set_bit(STRIPE_HANDLE, &sh->state);
2828 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2829 if (atomic_dec_and_test(&conf->pending_full_writes))
2830 md_wakeup_thread(conf->mddev->thread);
2833 static void handle_stripe_dirtying(struct r5conf *conf,
2834 struct stripe_head *sh,
2835 struct stripe_head_state *s,
2838 int rmw = 0, rcw = 0, i;
2839 sector_t recovery_cp = conf->mddev->recovery_cp;
2841 /* RAID6 requires 'rcw' in current implementation.
2842 * Otherwise, check whether resync is now happening or should start.
2843 * If yes, then the array is dirty (after unclean shutdown or
2844 * initial creation), so parity in some stripes might be inconsistent.
2845 * In this case, we need to always do reconstruct-write, to ensure
2846 * that in case of drive failure or read-error correction, we
2847 * generate correct data from the parity.
2849 if (conf->max_degraded == 2 ||
2850 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2851 /* Calculate the real rcw later - for now make it
2852 * look like rcw is cheaper
2855 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2856 conf->max_degraded, (unsigned long long)recovery_cp,
2857 (unsigned long long)sh->sector);
2858 } else for (i = disks; i--; ) {
2859 /* would I have to read this buffer for read_modify_write */
2860 struct r5dev *dev = &sh->dev[i];
2861 if ((dev->towrite || i == sh->pd_idx) &&
2862 !test_bit(R5_LOCKED, &dev->flags) &&
2863 !(test_bit(R5_UPTODATE, &dev->flags) ||
2864 test_bit(R5_Wantcompute, &dev->flags))) {
2865 if (test_bit(R5_Insync, &dev->flags))
2868 rmw += 2*disks; /* cannot read it */
2870 /* Would I have to read this buffer for reconstruct_write */
2871 if (!test_bit(R5_OVERWRITE, &dev->flags) && 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 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2880 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2881 (unsigned long long)sh->sector, rmw, rcw);
2882 set_bit(STRIPE_HANDLE, &sh->state);
2883 if (rmw < rcw && rmw > 0) {
2884 /* prefer read-modify-write, but need to get some data */
2885 if (conf->mddev->queue)
2886 blk_add_trace_msg(conf->mddev->queue,
2887 "raid5 rmw %llu %d",
2888 (unsigned long long)sh->sector, rmw);
2889 for (i = disks; i--; ) {
2890 struct r5dev *dev = &sh->dev[i];
2891 if ((dev->towrite || i == sh->pd_idx) &&
2892 !test_bit(R5_LOCKED, &dev->flags) &&
2893 !(test_bit(R5_UPTODATE, &dev->flags) ||
2894 test_bit(R5_Wantcompute, &dev->flags)) &&
2895 test_bit(R5_Insync, &dev->flags)) {
2897 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2898 pr_debug("Read_old block "
2899 "%d for r-m-w\n", i);
2900 set_bit(R5_LOCKED, &dev->flags);
2901 set_bit(R5_Wantread, &dev->flags);
2904 set_bit(STRIPE_DELAYED, &sh->state);
2905 set_bit(STRIPE_HANDLE, &sh->state);
2910 if (rcw <= rmw && rcw > 0) {
2911 /* want reconstruct write, but need to get some data */
2914 for (i = disks; i--; ) {
2915 struct r5dev *dev = &sh->dev[i];
2916 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2917 i != sh->pd_idx && i != sh->qd_idx &&
2918 !test_bit(R5_LOCKED, &dev->flags) &&
2919 !(test_bit(R5_UPTODATE, &dev->flags) ||
2920 test_bit(R5_Wantcompute, &dev->flags))) {
2922 if (!test_bit(R5_Insync, &dev->flags))
2923 continue; /* it's a failed drive */
2925 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2926 pr_debug("Read_old block "
2927 "%d for Reconstruct\n", i);
2928 set_bit(R5_LOCKED, &dev->flags);
2929 set_bit(R5_Wantread, &dev->flags);
2933 set_bit(STRIPE_DELAYED, &sh->state);
2934 set_bit(STRIPE_HANDLE, &sh->state);
2938 if (rcw && conf->mddev->queue)
2939 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
2940 (unsigned long long)sh->sector,
2941 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
2943 /* now if nothing is locked, and if we have enough data,
2944 * we can start a write request
2946 /* since handle_stripe can be called at any time we need to handle the
2947 * case where a compute block operation has been submitted and then a
2948 * subsequent call wants to start a write request. raid_run_ops only
2949 * handles the case where compute block and reconstruct are requested
2950 * simultaneously. If this is not the case then new writes need to be
2951 * held off until the compute completes.
2953 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2954 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2955 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2956 schedule_reconstruction(sh, s, rcw == 0, 0);
2959 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2960 struct stripe_head_state *s, int disks)
2962 struct r5dev *dev = NULL;
2964 set_bit(STRIPE_HANDLE, &sh->state);
2966 switch (sh->check_state) {
2967 case check_state_idle:
2968 /* start a new check operation if there are no failures */
2969 if (s->failed == 0) {
2970 BUG_ON(s->uptodate != disks);
2971 sh->check_state = check_state_run;
2972 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2973 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2977 dev = &sh->dev[s->failed_num[0]];
2979 case check_state_compute_result:
2980 sh->check_state = check_state_idle;
2982 dev = &sh->dev[sh->pd_idx];
2984 /* check that a write has not made the stripe insync */
2985 if (test_bit(STRIPE_INSYNC, &sh->state))
2988 /* either failed parity check, or recovery is happening */
2989 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2990 BUG_ON(s->uptodate != disks);
2992 set_bit(R5_LOCKED, &dev->flags);
2994 set_bit(R5_Wantwrite, &dev->flags);
2996 clear_bit(STRIPE_DEGRADED, &sh->state);
2997 set_bit(STRIPE_INSYNC, &sh->state);
2999 case check_state_run:
3000 break; /* we will be called again upon completion */
3001 case check_state_check_result:
3002 sh->check_state = check_state_idle;
3004 /* if a failure occurred during the check operation, leave
3005 * STRIPE_INSYNC not set and let the stripe be handled again
3010 /* handle a successful check operation, if parity is correct
3011 * we are done. Otherwise update the mismatch count and repair
3012 * parity if !MD_RECOVERY_CHECK
3014 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3015 /* parity is correct (on disc,
3016 * not in buffer any more)
3018 set_bit(STRIPE_INSYNC, &sh->state);
3020 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3021 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3022 /* don't try to repair!! */
3023 set_bit(STRIPE_INSYNC, &sh->state);
3025 sh->check_state = check_state_compute_run;
3026 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3027 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3028 set_bit(R5_Wantcompute,
3029 &sh->dev[sh->pd_idx].flags);
3030 sh->ops.target = sh->pd_idx;
3031 sh->ops.target2 = -1;
3036 case check_state_compute_run:
3039 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3040 __func__, sh->check_state,
3041 (unsigned long long) sh->sector);
3047 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3048 struct stripe_head_state *s,
3051 int pd_idx = sh->pd_idx;
3052 int qd_idx = sh->qd_idx;
3055 set_bit(STRIPE_HANDLE, &sh->state);
3057 BUG_ON(s->failed > 2);
3059 /* Want to check and possibly repair P and Q.
3060 * However there could be one 'failed' device, in which
3061 * case we can only check one of them, possibly using the
3062 * other to generate missing data
3065 switch (sh->check_state) {
3066 case check_state_idle:
3067 /* start a new check operation if there are < 2 failures */
3068 if (s->failed == s->q_failed) {
3069 /* The only possible failed device holds Q, so it
3070 * makes sense to check P (If anything else were failed,
3071 * we would have used P to recreate it).
3073 sh->check_state = check_state_run;
3075 if (!s->q_failed && s->failed < 2) {
3076 /* Q is not failed, and we didn't use it to generate
3077 * anything, so it makes sense to check it
3079 if (sh->check_state == check_state_run)
3080 sh->check_state = check_state_run_pq;
3082 sh->check_state = check_state_run_q;
3085 /* discard potentially stale zero_sum_result */
3086 sh->ops.zero_sum_result = 0;
3088 if (sh->check_state == check_state_run) {
3089 /* async_xor_zero_sum destroys the contents of P */
3090 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3093 if (sh->check_state >= check_state_run &&
3094 sh->check_state <= check_state_run_pq) {
3095 /* async_syndrome_zero_sum preserves P and Q, so
3096 * no need to mark them !uptodate here
3098 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3102 /* we have 2-disk failure */
3103 BUG_ON(s->failed != 2);
3105 case check_state_compute_result:
3106 sh->check_state = check_state_idle;
3108 /* check that a write has not made the stripe insync */
3109 if (test_bit(STRIPE_INSYNC, &sh->state))
3112 /* now write out any block on a failed drive,
3113 * or P or Q if they were recomputed
3115 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3116 if (s->failed == 2) {
3117 dev = &sh->dev[s->failed_num[1]];
3119 set_bit(R5_LOCKED, &dev->flags);
3120 set_bit(R5_Wantwrite, &dev->flags);
3122 if (s->failed >= 1) {
3123 dev = &sh->dev[s->failed_num[0]];
3125 set_bit(R5_LOCKED, &dev->flags);
3126 set_bit(R5_Wantwrite, &dev->flags);
3128 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3129 dev = &sh->dev[pd_idx];
3131 set_bit(R5_LOCKED, &dev->flags);
3132 set_bit(R5_Wantwrite, &dev->flags);
3134 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3135 dev = &sh->dev[qd_idx];
3137 set_bit(R5_LOCKED, &dev->flags);
3138 set_bit(R5_Wantwrite, &dev->flags);
3140 clear_bit(STRIPE_DEGRADED, &sh->state);
3142 set_bit(STRIPE_INSYNC, &sh->state);
3144 case check_state_run:
3145 case check_state_run_q:
3146 case check_state_run_pq:
3147 break; /* we will be called again upon completion */
3148 case check_state_check_result:
3149 sh->check_state = check_state_idle;
3151 /* handle a successful check operation, if parity is correct
3152 * we are done. Otherwise update the mismatch count and repair
3153 * parity if !MD_RECOVERY_CHECK
3155 if (sh->ops.zero_sum_result == 0) {
3156 /* both parities are correct */
3158 set_bit(STRIPE_INSYNC, &sh->state);
3160 /* in contrast to the raid5 case we can validate
3161 * parity, but still have a failure to write
3164 sh->check_state = check_state_compute_result;
3165 /* Returning at this point means that we may go
3166 * off and bring p and/or q uptodate again so
3167 * we make sure to check zero_sum_result again
3168 * to verify if p or q need writeback
3172 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3173 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3174 /* don't try to repair!! */
3175 set_bit(STRIPE_INSYNC, &sh->state);
3177 int *target = &sh->ops.target;
3179 sh->ops.target = -1;
3180 sh->ops.target2 = -1;
3181 sh->check_state = check_state_compute_run;
3182 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3183 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3184 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3185 set_bit(R5_Wantcompute,
3186 &sh->dev[pd_idx].flags);
3188 target = &sh->ops.target2;
3191 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3192 set_bit(R5_Wantcompute,
3193 &sh->dev[qd_idx].flags);
3200 case check_state_compute_run:
3203 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3204 __func__, sh->check_state,
3205 (unsigned long long) sh->sector);
3210 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3214 /* We have read all the blocks in this stripe and now we need to
3215 * copy some of them into a target stripe for expand.
3217 struct dma_async_tx_descriptor *tx = NULL;
3218 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3219 for (i = 0; i < sh->disks; i++)
3220 if (i != sh->pd_idx && i != sh->qd_idx) {
3222 struct stripe_head *sh2;
3223 struct async_submit_ctl submit;
3225 sector_t bn = compute_blocknr(sh, i, 1);
3226 sector_t s = raid5_compute_sector(conf, bn, 0,
3228 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3230 /* so far only the early blocks of this stripe
3231 * have been requested. When later blocks
3232 * get requested, we will try again
3235 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3236 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3237 /* must have already done this block */
3238 release_stripe(sh2);
3242 /* place all the copies on one channel */
3243 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3244 tx = async_memcpy(sh2->dev[dd_idx].page,
3245 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3248 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3249 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3250 for (j = 0; j < conf->raid_disks; j++)
3251 if (j != sh2->pd_idx &&
3253 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3255 if (j == conf->raid_disks) {
3256 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3257 set_bit(STRIPE_HANDLE, &sh2->state);
3259 release_stripe(sh2);
3262 /* done submitting copies, wait for them to complete */
3263 async_tx_quiesce(&tx);
3267 * handle_stripe - do things to a stripe.
3269 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3270 * state of various bits to see what needs to be done.
3272 * return some read requests which now have data
3273 * return some write requests which are safely on storage
3274 * schedule a read on some buffers
3275 * schedule a write of some buffers
3276 * return confirmation of parity correctness
3280 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3282 struct r5conf *conf = sh->raid_conf;
3283 int disks = sh->disks;
3286 int do_recovery = 0;
3288 memset(s, 0, sizeof(*s));
3290 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3291 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3292 s->failed_num[0] = -1;
3293 s->failed_num[1] = -1;
3295 /* Now to look around and see what can be done */
3297 for (i=disks; i--; ) {
3298 struct md_rdev *rdev;
3305 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3307 dev->toread, dev->towrite, dev->written);
3308 /* maybe we can reply to a read
3310 * new wantfill requests are only permitted while
3311 * ops_complete_biofill is guaranteed to be inactive
3313 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3314 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3315 set_bit(R5_Wantfill, &dev->flags);
3317 /* now count some things */
3318 if (test_bit(R5_LOCKED, &dev->flags))
3320 if (test_bit(R5_UPTODATE, &dev->flags))
3322 if (test_bit(R5_Wantcompute, &dev->flags)) {
3324 BUG_ON(s->compute > 2);
3327 if (test_bit(R5_Wantfill, &dev->flags))
3329 else if (dev->toread)
3333 if (!test_bit(R5_OVERWRITE, &dev->flags))
3338 /* Prefer to use the replacement for reads, but only
3339 * if it is recovered enough and has no bad blocks.
3341 rdev = rcu_dereference(conf->disks[i].replacement);
3342 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3343 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3344 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3345 &first_bad, &bad_sectors))
3346 set_bit(R5_ReadRepl, &dev->flags);
3349 set_bit(R5_NeedReplace, &dev->flags);
3350 rdev = rcu_dereference(conf->disks[i].rdev);
3351 clear_bit(R5_ReadRepl, &dev->flags);
3353 if (rdev && test_bit(Faulty, &rdev->flags))
3356 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3357 &first_bad, &bad_sectors);
3358 if (s->blocked_rdev == NULL
3359 && (test_bit(Blocked, &rdev->flags)
3362 set_bit(BlockedBadBlocks,
3364 s->blocked_rdev = rdev;
3365 atomic_inc(&rdev->nr_pending);
3368 clear_bit(R5_Insync, &dev->flags);
3372 /* also not in-sync */
3373 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3374 test_bit(R5_UPTODATE, &dev->flags)) {
3375 /* treat as in-sync, but with a read error
3376 * which we can now try to correct
3378 set_bit(R5_Insync, &dev->flags);
3379 set_bit(R5_ReadError, &dev->flags);
3381 } else if (test_bit(In_sync, &rdev->flags))
3382 set_bit(R5_Insync, &dev->flags);
3383 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3384 /* in sync if before recovery_offset */
3385 set_bit(R5_Insync, &dev->flags);
3386 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3387 test_bit(R5_Expanded, &dev->flags))
3388 /* If we've reshaped into here, we assume it is Insync.
3389 * We will shortly update recovery_offset to make
3392 set_bit(R5_Insync, &dev->flags);
3394 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3395 /* This flag does not apply to '.replacement'
3396 * only to .rdev, so make sure to check that*/
3397 struct md_rdev *rdev2 = rcu_dereference(
3398 conf->disks[i].rdev);
3400 clear_bit(R5_Insync, &dev->flags);
3401 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3402 s->handle_bad_blocks = 1;
3403 atomic_inc(&rdev2->nr_pending);
3405 clear_bit(R5_WriteError, &dev->flags);
3407 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3408 /* This flag does not apply to '.replacement'
3409 * only to .rdev, so make sure to check that*/
3410 struct md_rdev *rdev2 = rcu_dereference(
3411 conf->disks[i].rdev);
3412 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3413 s->handle_bad_blocks = 1;
3414 atomic_inc(&rdev2->nr_pending);
3416 clear_bit(R5_MadeGood, &dev->flags);
3418 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3419 struct md_rdev *rdev2 = rcu_dereference(
3420 conf->disks[i].replacement);
3421 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3422 s->handle_bad_blocks = 1;
3423 atomic_inc(&rdev2->nr_pending);
3425 clear_bit(R5_MadeGoodRepl, &dev->flags);
3427 if (!test_bit(R5_Insync, &dev->flags)) {
3428 /* The ReadError flag will just be confusing now */
3429 clear_bit(R5_ReadError, &dev->flags);
3430 clear_bit(R5_ReWrite, &dev->flags);
3432 if (test_bit(R5_ReadError, &dev->flags))
3433 clear_bit(R5_Insync, &dev->flags);
3434 if (!test_bit(R5_Insync, &dev->flags)) {
3436 s->failed_num[s->failed] = i;
3438 if (rdev && !test_bit(Faulty, &rdev->flags))
3442 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3443 /* If there is a failed device being replaced,
3444 * we must be recovering.
3445 * else if we are after recovery_cp, we must be syncing
3446 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3447 * else we can only be replacing
3448 * sync and recovery both need to read all devices, and so
3449 * use the same flag.
3452 sh->sector >= conf->mddev->recovery_cp ||
3453 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3461 static void handle_stripe(struct stripe_head *sh)
3463 struct stripe_head_state s;
3464 struct r5conf *conf = sh->raid_conf;
3467 int disks = sh->disks;
3468 struct r5dev *pdev, *qdev;
3470 clear_bit(STRIPE_HANDLE, &sh->state);
3471 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3472 /* already being handled, ensure it gets handled
3473 * again when current action finishes */
3474 set_bit(STRIPE_HANDLE, &sh->state);
3478 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3479 spin_lock(&sh->stripe_lock);
3480 /* Cannot process 'sync' concurrently with 'discard' */
3481 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3482 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3483 set_bit(STRIPE_SYNCING, &sh->state);
3484 clear_bit(STRIPE_INSYNC, &sh->state);
3485 clear_bit(STRIPE_REPLACED, &sh->state);
3487 spin_unlock(&sh->stripe_lock);
3489 clear_bit(STRIPE_DELAYED, &sh->state);
3491 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3492 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3493 (unsigned long long)sh->sector, sh->state,
3494 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3495 sh->check_state, sh->reconstruct_state);
3497 analyse_stripe(sh, &s);
3499 if (s.handle_bad_blocks) {
3500 set_bit(STRIPE_HANDLE, &sh->state);
3504 if (unlikely(s.blocked_rdev)) {
3505 if (s.syncing || s.expanding || s.expanded ||
3506 s.replacing || s.to_write || s.written) {
3507 set_bit(STRIPE_HANDLE, &sh->state);
3510 /* There is nothing for the blocked_rdev to block */
3511 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3512 s.blocked_rdev = NULL;
3515 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3516 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3517 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3520 pr_debug("locked=%d uptodate=%d to_read=%d"
3521 " to_write=%d failed=%d failed_num=%d,%d\n",
3522 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3523 s.failed_num[0], s.failed_num[1]);
3524 /* check if the array has lost more than max_degraded devices and,
3525 * if so, some requests might need to be failed.
3527 if (s.failed > conf->max_degraded) {
3528 sh->check_state = 0;
3529 sh->reconstruct_state = 0;
3530 if (s.to_read+s.to_write+s.written)
3531 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3532 if (s.syncing + s.replacing)
3533 handle_failed_sync(conf, sh, &s);
3536 /* Now we check to see if any write operations have recently
3540 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3542 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3543 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3544 sh->reconstruct_state = reconstruct_state_idle;
3546 /* All the 'written' buffers and the parity block are ready to
3547 * be written back to disk
3549 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3550 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3551 BUG_ON(sh->qd_idx >= 0 &&
3552 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3553 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3554 for (i = disks; i--; ) {
3555 struct r5dev *dev = &sh->dev[i];
3556 if (test_bit(R5_LOCKED, &dev->flags) &&
3557 (i == sh->pd_idx || i == sh->qd_idx ||
3559 pr_debug("Writing block %d\n", i);
3560 set_bit(R5_Wantwrite, &dev->flags);
3563 if (!test_bit(R5_Insync, &dev->flags) ||
3564 ((i == sh->pd_idx || i == sh->qd_idx) &&
3566 set_bit(STRIPE_INSYNC, &sh->state);
3569 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3570 s.dec_preread_active = 1;
3574 * might be able to return some write requests if the parity blocks
3575 * are safe, or on a failed drive
3577 pdev = &sh->dev[sh->pd_idx];
3578 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3579 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3580 qdev = &sh->dev[sh->qd_idx];
3581 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3582 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3586 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3587 && !test_bit(R5_LOCKED, &pdev->flags)
3588 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3589 test_bit(R5_Discard, &pdev->flags))))) &&
3590 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3591 && !test_bit(R5_LOCKED, &qdev->flags)
3592 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3593 test_bit(R5_Discard, &qdev->flags))))))
3594 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3596 /* Now we might consider reading some blocks, either to check/generate
3597 * parity, or to satisfy requests
3598 * or to load a block that is being partially written.
3600 if (s.to_read || s.non_overwrite
3601 || (conf->level == 6 && s.to_write && s.failed)
3602 || (s.syncing && (s.uptodate + s.compute < disks))
3605 handle_stripe_fill(sh, &s, disks);
3607 /* Now to consider new write requests and what else, if anything
3608 * should be read. We do not handle new writes when:
3609 * 1/ A 'write' operation (copy+xor) is already in flight.
3610 * 2/ A 'check' operation is in flight, as it may clobber the parity
3613 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3614 handle_stripe_dirtying(conf, sh, &s, disks);
3616 /* maybe we need to check and possibly fix the parity for this stripe
3617 * Any reads will already have been scheduled, so we just see if enough
3618 * data is available. The parity check is held off while parity
3619 * dependent operations are in flight.
3621 if (sh->check_state ||
3622 (s.syncing && s.locked == 0 &&
3623 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3624 !test_bit(STRIPE_INSYNC, &sh->state))) {
3625 if (conf->level == 6)
3626 handle_parity_checks6(conf, sh, &s, disks);
3628 handle_parity_checks5(conf, sh, &s, disks);
3631 if ((s.replacing || s.syncing) && s.locked == 0
3632 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3633 && !test_bit(STRIPE_REPLACED, &sh->state)) {
3634 /* Write out to replacement devices where possible */
3635 for (i = 0; i < conf->raid_disks; i++)
3636 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3637 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
3638 set_bit(R5_WantReplace, &sh->dev[i].flags);
3639 set_bit(R5_LOCKED, &sh->dev[i].flags);
3643 set_bit(STRIPE_INSYNC, &sh->state);
3644 set_bit(STRIPE_REPLACED, &sh->state);
3646 if ((s.syncing || s.replacing) && s.locked == 0 &&
3647 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3648 test_bit(STRIPE_INSYNC, &sh->state)) {
3649 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3650 clear_bit(STRIPE_SYNCING, &sh->state);
3651 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3652 wake_up(&conf->wait_for_overlap);
3655 /* If the failed drives are just a ReadError, then we might need
3656 * to progress the repair/check process
3658 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3659 for (i = 0; i < s.failed; i++) {
3660 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3661 if (test_bit(R5_ReadError, &dev->flags)
3662 && !test_bit(R5_LOCKED, &dev->flags)
3663 && test_bit(R5_UPTODATE, &dev->flags)
3665 if (!test_bit(R5_ReWrite, &dev->flags)) {
3666 set_bit(R5_Wantwrite, &dev->flags);
3667 set_bit(R5_ReWrite, &dev->flags);
3668 set_bit(R5_LOCKED, &dev->flags);
3671 /* let's read it back */
3672 set_bit(R5_Wantread, &dev->flags);
3673 set_bit(R5_LOCKED, &dev->flags);
3680 /* Finish reconstruct operations initiated by the expansion process */
3681 if (sh->reconstruct_state == reconstruct_state_result) {
3682 struct stripe_head *sh_src
3683 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3684 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3685 /* sh cannot be written until sh_src has been read.
3686 * so arrange for sh to be delayed a little
3688 set_bit(STRIPE_DELAYED, &sh->state);
3689 set_bit(STRIPE_HANDLE, &sh->state);
3690 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3692 atomic_inc(&conf->preread_active_stripes);
3693 release_stripe(sh_src);
3697 release_stripe(sh_src);
3699 sh->reconstruct_state = reconstruct_state_idle;
3700 clear_bit(STRIPE_EXPANDING, &sh->state);
3701 for (i = conf->raid_disks; i--; ) {
3702 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3703 set_bit(R5_LOCKED, &sh->dev[i].flags);
3708 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3709 !sh->reconstruct_state) {
3710 /* Need to write out all blocks after computing parity */
3711 sh->disks = conf->raid_disks;
3712 stripe_set_idx(sh->sector, conf, 0, sh);
3713 schedule_reconstruction(sh, &s, 1, 1);
3714 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3715 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3716 atomic_dec(&conf->reshape_stripes);
3717 wake_up(&conf->wait_for_overlap);
3718 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3721 if (s.expanding && s.locked == 0 &&
3722 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3723 handle_stripe_expansion(conf, sh);
3726 /* wait for this device to become unblocked */
3727 if (unlikely(s.blocked_rdev)) {
3728 if (conf->mddev->external)
3729 md_wait_for_blocked_rdev(s.blocked_rdev,
3732 /* Internal metadata will immediately
3733 * be written by raid5d, so we don't
3734 * need to wait here.
3736 rdev_dec_pending(s.blocked_rdev,
3740 if (s.handle_bad_blocks)
3741 for (i = disks; i--; ) {
3742 struct md_rdev *rdev;
3743 struct r5dev *dev = &sh->dev[i];
3744 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3745 /* We own a safe reference to the rdev */
3746 rdev = conf->disks[i].rdev;
3747 if (!rdev_set_badblocks(rdev, sh->sector,
3749 md_error(conf->mddev, rdev);
3750 rdev_dec_pending(rdev, conf->mddev);
3752 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3753 rdev = conf->disks[i].rdev;
3754 rdev_clear_badblocks(rdev, sh->sector,
3756 rdev_dec_pending(rdev, conf->mddev);
3758 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3759 rdev = conf->disks[i].replacement;
3761 /* rdev have been moved down */
3762 rdev = conf->disks[i].rdev;
3763 rdev_clear_badblocks(rdev, sh->sector,
3765 rdev_dec_pending(rdev, conf->mddev);
3770 raid_run_ops(sh, s.ops_request);
3774 if (s.dec_preread_active) {
3775 /* We delay this until after ops_run_io so that if make_request
3776 * is waiting on a flush, it won't continue until the writes
3777 * have actually been submitted.
3779 atomic_dec(&conf->preread_active_stripes);
3780 if (atomic_read(&conf->preread_active_stripes) <
3782 md_wakeup_thread(conf->mddev->thread);
3785 return_io(s.return_bi);
3787 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3790 static void raid5_activate_delayed(struct r5conf *conf)
3792 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3793 while (!list_empty(&conf->delayed_list)) {
3794 struct list_head *l = conf->delayed_list.next;
3795 struct stripe_head *sh;
3796 sh = list_entry(l, struct stripe_head, lru);
3798 clear_bit(STRIPE_DELAYED, &sh->state);
3799 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3800 atomic_inc(&conf->preread_active_stripes);
3801 list_add_tail(&sh->lru, &conf->hold_list);
3806 static void activate_bit_delay(struct r5conf *conf)
3808 /* device_lock is held */
3809 struct list_head head;
3810 list_add(&head, &conf->bitmap_list);
3811 list_del_init(&conf->bitmap_list);
3812 while (!list_empty(&head)) {
3813 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3814 list_del_init(&sh->lru);
3815 atomic_inc(&sh->count);
3816 __release_stripe(conf, sh);
3820 int md_raid5_congested(struct mddev *mddev, int bits)
3822 struct r5conf *conf = mddev->private;
3824 /* No difference between reads and writes. Just check
3825 * how busy the stripe_cache is
3828 if (conf->inactive_blocked)
3832 if (list_empty_careful(&conf->inactive_list))
3837 EXPORT_SYMBOL_GPL(md_raid5_congested);
3839 static int raid5_congested(void *data, int bits)
3841 struct mddev *mddev = data;
3843 return mddev_congested(mddev, bits) ||
3844 md_raid5_congested(mddev, bits);
3847 /* We want read requests to align with chunks where possible,
3848 * but write requests don't need to.
3850 static int raid5_mergeable_bvec(struct request_queue *q,
3851 struct bvec_merge_data *bvm,
3852 struct bio_vec *biovec)
3854 struct mddev *mddev = q->queuedata;
3855 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3857 unsigned int chunk_sectors = mddev->chunk_sectors;
3858 unsigned int bio_sectors = bvm->bi_size >> 9;
3860 if ((bvm->bi_rw & 1) == WRITE)
3861 return biovec->bv_len; /* always allow writes to be mergeable */
3863 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3864 chunk_sectors = mddev->new_chunk_sectors;
3865 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3866 if (max < 0) max = 0;
3867 if (max <= biovec->bv_len && bio_sectors == 0)
3868 return biovec->bv_len;
3874 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3876 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3877 unsigned int chunk_sectors = mddev->chunk_sectors;
3878 unsigned int bio_sectors = bio_sectors(bio);
3880 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3881 chunk_sectors = mddev->new_chunk_sectors;
3882 return chunk_sectors >=
3883 ((sector & (chunk_sectors - 1)) + bio_sectors);
3887 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3888 * later sampled by raid5d.
3890 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3892 unsigned long flags;
3894 spin_lock_irqsave(&conf->device_lock, flags);
3896 bi->bi_next = conf->retry_read_aligned_list;
3897 conf->retry_read_aligned_list = bi;
3899 spin_unlock_irqrestore(&conf->device_lock, flags);
3900 md_wakeup_thread(conf->mddev->thread);
3904 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3908 bi = conf->retry_read_aligned;
3910 conf->retry_read_aligned = NULL;
3913 bi = conf->retry_read_aligned_list;
3915 conf->retry_read_aligned_list = bi->bi_next;
3918 * this sets the active strip count to 1 and the processed
3919 * strip count to zero (upper 8 bits)
3921 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
3929 * The "raid5_align_endio" should check if the read succeeded and if it
3930 * did, call bio_endio on the original bio (having bio_put the new bio
3932 * If the read failed..
3934 static void raid5_align_endio(struct bio *bi, int error)
3936 struct bio* raid_bi = bi->bi_private;
3937 struct mddev *mddev;
3938 struct r5conf *conf;
3939 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3940 struct md_rdev *rdev;
3944 rdev = (void*)raid_bi->bi_next;
3945 raid_bi->bi_next = NULL;
3946 mddev = rdev->mddev;
3947 conf = mddev->private;
3949 rdev_dec_pending(rdev, conf->mddev);
3951 if (!error && uptodate) {
3952 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
3954 bio_endio(raid_bi, 0);
3955 if (atomic_dec_and_test(&conf->active_aligned_reads))
3956 wake_up(&conf->wait_for_stripe);
3961 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3963 add_bio_to_retry(raid_bi, conf);
3966 static int bio_fits_rdev(struct bio *bi)
3968 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3970 if (bio_sectors(bi) > queue_max_sectors(q))
3972 blk_recount_segments(q, bi);
3973 if (bi->bi_phys_segments > queue_max_segments(q))
3976 if (q->merge_bvec_fn)
3977 /* it's too hard to apply the merge_bvec_fn at this stage,
3986 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3988 struct r5conf *conf = mddev->private;
3990 struct bio* align_bi;
3991 struct md_rdev *rdev;
3992 sector_t end_sector;
3994 if (!in_chunk_boundary(mddev, raid_bio)) {
3995 pr_debug("chunk_aligned_read : non aligned\n");
3999 * use bio_clone_mddev to make a copy of the bio
4001 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4005 * set bi_end_io to a new function, and set bi_private to the
4008 align_bi->bi_end_io = raid5_align_endio;
4009 align_bi->bi_private = raid_bio;
4013 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
4017 end_sector = bio_end_sector(align_bi);
4019 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4020 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4021 rdev->recovery_offset < end_sector) {
4022 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4024 (test_bit(Faulty, &rdev->flags) ||
4025 !(test_bit(In_sync, &rdev->flags) ||
4026 rdev->recovery_offset >= end_sector)))
4033 atomic_inc(&rdev->nr_pending);
4035 raid_bio->bi_next = (void*)rdev;
4036 align_bi->bi_bdev = rdev->bdev;
4037 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4039 if (!bio_fits_rdev(align_bi) ||
4040 is_badblock(rdev, align_bi->bi_sector, bio_sectors(align_bi),
4041 &first_bad, &bad_sectors)) {
4042 /* too big in some way, or has a known bad block */
4044 rdev_dec_pending(rdev, mddev);
4048 /* No reshape active, so we can trust rdev->data_offset */
4049 align_bi->bi_sector += rdev->data_offset;
4051 spin_lock_irq(&conf->device_lock);
4052 wait_event_lock_irq(conf->wait_for_stripe,
4055 atomic_inc(&conf->active_aligned_reads);
4056 spin_unlock_irq(&conf->device_lock);
4059 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4060 align_bi, disk_devt(mddev->gendisk),
4061 raid_bio->bi_sector);
4062 generic_make_request(align_bi);
4071 /* __get_priority_stripe - get the next stripe to process
4073 * Full stripe writes are allowed to pass preread active stripes up until
4074 * the bypass_threshold is exceeded. In general the bypass_count
4075 * increments when the handle_list is handled before the hold_list; however, it
4076 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4077 * stripe with in flight i/o. The bypass_count will be reset when the
4078 * head of the hold_list has changed, i.e. the head was promoted to the
4081 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
4083 struct stripe_head *sh;
4085 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4087 list_empty(&conf->handle_list) ? "empty" : "busy",
4088 list_empty(&conf->hold_list) ? "empty" : "busy",
4089 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4091 if (!list_empty(&conf->handle_list)) {
4092 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4094 if (list_empty(&conf->hold_list))
4095 conf->bypass_count = 0;
4096 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4097 if (conf->hold_list.next == conf->last_hold)
4098 conf->bypass_count++;
4100 conf->last_hold = conf->hold_list.next;
4101 conf->bypass_count -= conf->bypass_threshold;
4102 if (conf->bypass_count < 0)
4103 conf->bypass_count = 0;
4106 } else if (!list_empty(&conf->hold_list) &&
4107 ((conf->bypass_threshold &&
4108 conf->bypass_count > conf->bypass_threshold) ||
4109 atomic_read(&conf->pending_full_writes) == 0)) {
4110 sh = list_entry(conf->hold_list.next,
4112 conf->bypass_count -= conf->bypass_threshold;
4113 if (conf->bypass_count < 0)
4114 conf->bypass_count = 0;
4118 list_del_init(&sh->lru);
4119 atomic_inc(&sh->count);
4120 BUG_ON(atomic_read(&sh->count) != 1);
4124 struct raid5_plug_cb {
4125 struct blk_plug_cb cb;
4126 struct list_head list;
4129 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4131 struct raid5_plug_cb *cb = container_of(
4132 blk_cb, struct raid5_plug_cb, cb);
4133 struct stripe_head *sh;
4134 struct mddev *mddev = cb->cb.data;
4135 struct r5conf *conf = mddev->private;
4138 if (cb->list.next && !list_empty(&cb->list)) {
4139 spin_lock_irq(&conf->device_lock);
4140 while (!list_empty(&cb->list)) {
4141 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4142 list_del_init(&sh->lru);
4144 * avoid race release_stripe_plug() sees
4145 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4146 * is still in our list
4148 smp_mb__before_clear_bit();
4149 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4150 __release_stripe(conf, sh);
4153 spin_unlock_irq(&conf->device_lock);
4156 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4160 static void release_stripe_plug(struct mddev *mddev,
4161 struct stripe_head *sh)
4163 struct blk_plug_cb *blk_cb = blk_check_plugged(
4164 raid5_unplug, mddev,
4165 sizeof(struct raid5_plug_cb));
4166 struct raid5_plug_cb *cb;
4173 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4175 if (cb->list.next == NULL)
4176 INIT_LIST_HEAD(&cb->list);
4178 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4179 list_add_tail(&sh->lru, &cb->list);
4184 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4186 struct r5conf *conf = mddev->private;
4187 sector_t logical_sector, last_sector;
4188 struct stripe_head *sh;
4192 if (mddev->reshape_position != MaxSector)
4193 /* Skip discard while reshape is happening */
4196 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4197 last_sector = bi->bi_sector + (bi->bi_size>>9);
4200 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4202 stripe_sectors = conf->chunk_sectors *
4203 (conf->raid_disks - conf->max_degraded);
4204 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4206 sector_div(last_sector, stripe_sectors);
4208 logical_sector *= conf->chunk_sectors;
4209 last_sector *= conf->chunk_sectors;
4211 for (; logical_sector < last_sector;
4212 logical_sector += STRIPE_SECTORS) {
4216 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4217 prepare_to_wait(&conf->wait_for_overlap, &w,
4218 TASK_UNINTERRUPTIBLE);
4219 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4220 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4225 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4226 spin_lock_irq(&sh->stripe_lock);
4227 for (d = 0; d < conf->raid_disks; d++) {
4228 if (d == sh->pd_idx || d == sh->qd_idx)
4230 if (sh->dev[d].towrite || sh->dev[d].toread) {
4231 set_bit(R5_Overlap, &sh->dev[d].flags);
4232 spin_unlock_irq(&sh->stripe_lock);
4238 set_bit(STRIPE_DISCARD, &sh->state);
4239 finish_wait(&conf->wait_for_overlap, &w);
4240 for (d = 0; d < conf->raid_disks; d++) {
4241 if (d == sh->pd_idx || d == sh->qd_idx)
4243 sh->dev[d].towrite = bi;
4244 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4245 raid5_inc_bi_active_stripes(bi);
4247 spin_unlock_irq(&sh->stripe_lock);
4248 if (conf->mddev->bitmap) {
4250 d < conf->raid_disks - conf->max_degraded;
4252 bitmap_startwrite(mddev->bitmap,
4256 sh->bm_seq = conf->seq_flush + 1;
4257 set_bit(STRIPE_BIT_DELAY, &sh->state);
4260 set_bit(STRIPE_HANDLE, &sh->state);
4261 clear_bit(STRIPE_DELAYED, &sh->state);
4262 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4263 atomic_inc(&conf->preread_active_stripes);
4264 release_stripe_plug(mddev, sh);
4267 remaining = raid5_dec_bi_active_stripes(bi);
4268 if (remaining == 0) {
4269 md_write_end(mddev);
4274 static void make_request(struct mddev *mddev, struct bio * bi)
4276 struct r5conf *conf = mddev->private;
4278 sector_t new_sector;
4279 sector_t logical_sector, last_sector;
4280 struct stripe_head *sh;
4281 const int rw = bio_data_dir(bi);
4284 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4285 md_flush_request(mddev, bi);
4289 md_write_start(mddev, bi);
4292 mddev->reshape_position == MaxSector &&
4293 chunk_aligned_read(mddev,bi))
4296 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4297 make_discard_request(mddev, bi);
4301 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4302 last_sector = bio_end_sector(bi);
4304 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4306 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4312 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4313 if (unlikely(conf->reshape_progress != MaxSector)) {
4314 /* spinlock is needed as reshape_progress may be
4315 * 64bit on a 32bit platform, and so it might be
4316 * possible to see a half-updated value
4317 * Of course reshape_progress could change after
4318 * the lock is dropped, so once we get a reference
4319 * to the stripe that we think it is, we will have
4322 spin_lock_irq(&conf->device_lock);
4323 if (mddev->reshape_backwards
4324 ? logical_sector < conf->reshape_progress
4325 : logical_sector >= conf->reshape_progress) {
4328 if (mddev->reshape_backwards
4329 ? logical_sector < conf->reshape_safe
4330 : logical_sector >= conf->reshape_safe) {
4331 spin_unlock_irq(&conf->device_lock);
4336 spin_unlock_irq(&conf->device_lock);
4339 new_sector = raid5_compute_sector(conf, logical_sector,
4342 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4343 (unsigned long long)new_sector,
4344 (unsigned long long)logical_sector);
4346 sh = get_active_stripe(conf, new_sector, previous,
4347 (bi->bi_rw&RWA_MASK), 0);
4349 if (unlikely(previous)) {
4350 /* expansion might have moved on while waiting for a
4351 * stripe, so we must do the range check again.
4352 * Expansion could still move past after this
4353 * test, but as we are holding a reference to
4354 * 'sh', we know that if that happens,
4355 * STRIPE_EXPANDING will get set and the expansion
4356 * won't proceed until we finish with the stripe.
4359 spin_lock_irq(&conf->device_lock);
4360 if (mddev->reshape_backwards
4361 ? logical_sector >= conf->reshape_progress
4362 : logical_sector < conf->reshape_progress)
4363 /* mismatch, need to try again */
4365 spin_unlock_irq(&conf->device_lock);
4374 logical_sector >= mddev->suspend_lo &&
4375 logical_sector < mddev->suspend_hi) {
4377 /* As the suspend_* range is controlled by
4378 * userspace, we want an interruptible
4381 flush_signals(current);
4382 prepare_to_wait(&conf->wait_for_overlap,
4383 &w, TASK_INTERRUPTIBLE);
4384 if (logical_sector >= mddev->suspend_lo &&
4385 logical_sector < mddev->suspend_hi)
4390 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4391 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4392 /* Stripe is busy expanding or
4393 * add failed due to overlap. Flush everything
4396 md_wakeup_thread(mddev->thread);
4401 finish_wait(&conf->wait_for_overlap, &w);
4402 set_bit(STRIPE_HANDLE, &sh->state);
4403 clear_bit(STRIPE_DELAYED, &sh->state);
4404 if ((bi->bi_rw & REQ_SYNC) &&
4405 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4406 atomic_inc(&conf->preread_active_stripes);
4407 release_stripe_plug(mddev, sh);
4409 /* cannot get stripe for read-ahead, just give-up */
4410 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4411 finish_wait(&conf->wait_for_overlap, &w);
4416 remaining = raid5_dec_bi_active_stripes(bi);
4417 if (remaining == 0) {
4420 md_write_end(mddev);
4422 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4428 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4430 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4432 /* reshaping is quite different to recovery/resync so it is
4433 * handled quite separately ... here.
4435 * On each call to sync_request, we gather one chunk worth of
4436 * destination stripes and flag them as expanding.
4437 * Then we find all the source stripes and request reads.
4438 * As the reads complete, handle_stripe will copy the data
4439 * into the destination stripe and release that stripe.
4441 struct r5conf *conf = mddev->private;
4442 struct stripe_head *sh;
4443 sector_t first_sector, last_sector;
4444 int raid_disks = conf->previous_raid_disks;
4445 int data_disks = raid_disks - conf->max_degraded;
4446 int new_data_disks = conf->raid_disks - conf->max_degraded;
4449 sector_t writepos, readpos, safepos;
4450 sector_t stripe_addr;
4451 int reshape_sectors;
4452 struct list_head stripes;
4454 if (sector_nr == 0) {
4455 /* If restarting in the middle, skip the initial sectors */
4456 if (mddev->reshape_backwards &&
4457 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4458 sector_nr = raid5_size(mddev, 0, 0)
4459 - conf->reshape_progress;
4460 } else if (!mddev->reshape_backwards &&
4461 conf->reshape_progress > 0)
4462 sector_nr = conf->reshape_progress;
4463 sector_div(sector_nr, new_data_disks);
4465 mddev->curr_resync_completed = sector_nr;
4466 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4472 /* We need to process a full chunk at a time.
4473 * If old and new chunk sizes differ, we need to process the
4476 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4477 reshape_sectors = mddev->new_chunk_sectors;
4479 reshape_sectors = mddev->chunk_sectors;
4481 /* We update the metadata at least every 10 seconds, or when
4482 * the data about to be copied would over-write the source of
4483 * the data at the front of the range. i.e. one new_stripe
4484 * along from reshape_progress new_maps to after where
4485 * reshape_safe old_maps to
4487 writepos = conf->reshape_progress;
4488 sector_div(writepos, new_data_disks);
4489 readpos = conf->reshape_progress;
4490 sector_div(readpos, data_disks);
4491 safepos = conf->reshape_safe;
4492 sector_div(safepos, data_disks);
4493 if (mddev->reshape_backwards) {
4494 writepos -= min_t(sector_t, reshape_sectors, writepos);
4495 readpos += reshape_sectors;
4496 safepos += reshape_sectors;
4498 writepos += reshape_sectors;
4499 readpos -= min_t(sector_t, reshape_sectors, readpos);
4500 safepos -= min_t(sector_t, reshape_sectors, safepos);
4503 /* Having calculated the 'writepos' possibly use it
4504 * to set 'stripe_addr' which is where we will write to.
4506 if (mddev->reshape_backwards) {
4507 BUG_ON(conf->reshape_progress == 0);
4508 stripe_addr = writepos;
4509 BUG_ON((mddev->dev_sectors &
4510 ~((sector_t)reshape_sectors - 1))
4511 - reshape_sectors - stripe_addr
4514 BUG_ON(writepos != sector_nr + reshape_sectors);
4515 stripe_addr = sector_nr;
4518 /* 'writepos' is the most advanced device address we might write.
4519 * 'readpos' is the least advanced device address we might read.
4520 * 'safepos' is the least address recorded in the metadata as having
4522 * If there is a min_offset_diff, these are adjusted either by
4523 * increasing the safepos/readpos if diff is negative, or
4524 * increasing writepos if diff is positive.
4525 * If 'readpos' is then behind 'writepos', there is no way that we can
4526 * ensure safety in the face of a crash - that must be done by userspace
4527 * making a backup of the data. So in that case there is no particular
4528 * rush to update metadata.
4529 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4530 * update the metadata to advance 'safepos' to match 'readpos' so that
4531 * we can be safe in the event of a crash.
4532 * So we insist on updating metadata if safepos is behind writepos and
4533 * readpos is beyond writepos.
4534 * In any case, update the metadata every 10 seconds.
4535 * Maybe that number should be configurable, but I'm not sure it is
4536 * worth it.... maybe it could be a multiple of safemode_delay???
4538 if (conf->min_offset_diff < 0) {
4539 safepos += -conf->min_offset_diff;
4540 readpos += -conf->min_offset_diff;
4542 writepos += conf->min_offset_diff;
4544 if ((mddev->reshape_backwards
4545 ? (safepos > writepos && readpos < writepos)
4546 : (safepos < writepos && readpos > writepos)) ||
4547 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4548 /* Cannot proceed until we've updated the superblock... */
4549 wait_event(conf->wait_for_overlap,
4550 atomic_read(&conf->reshape_stripes)==0);
4551 mddev->reshape_position = conf->reshape_progress;
4552 mddev->curr_resync_completed = sector_nr;
4553 conf->reshape_checkpoint = jiffies;
4554 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4555 md_wakeup_thread(mddev->thread);
4556 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4557 kthread_should_stop());
4558 spin_lock_irq(&conf->device_lock);
4559 conf->reshape_safe = mddev->reshape_position;
4560 spin_unlock_irq(&conf->device_lock);
4561 wake_up(&conf->wait_for_overlap);
4562 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4565 INIT_LIST_HEAD(&stripes);
4566 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4568 int skipped_disk = 0;
4569 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4570 set_bit(STRIPE_EXPANDING, &sh->state);
4571 atomic_inc(&conf->reshape_stripes);
4572 /* If any of this stripe is beyond the end of the old
4573 * array, then we need to zero those blocks
4575 for (j=sh->disks; j--;) {
4577 if (j == sh->pd_idx)
4579 if (conf->level == 6 &&
4582 s = compute_blocknr(sh, j, 0);
4583 if (s < raid5_size(mddev, 0, 0)) {
4587 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4588 set_bit(R5_Expanded, &sh->dev[j].flags);
4589 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4591 if (!skipped_disk) {
4592 set_bit(STRIPE_EXPAND_READY, &sh->state);
4593 set_bit(STRIPE_HANDLE, &sh->state);
4595 list_add(&sh->lru, &stripes);
4597 spin_lock_irq(&conf->device_lock);
4598 if (mddev->reshape_backwards)
4599 conf->reshape_progress -= reshape_sectors * new_data_disks;
4601 conf->reshape_progress += reshape_sectors * new_data_disks;
4602 spin_unlock_irq(&conf->device_lock);
4603 /* Ok, those stripe are ready. We can start scheduling
4604 * reads on the source stripes.
4605 * The source stripes are determined by mapping the first and last
4606 * block on the destination stripes.
4609 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4612 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4613 * new_data_disks - 1),
4615 if (last_sector >= mddev->dev_sectors)
4616 last_sector = mddev->dev_sectors - 1;
4617 while (first_sector <= last_sector) {
4618 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4619 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4620 set_bit(STRIPE_HANDLE, &sh->state);
4622 first_sector += STRIPE_SECTORS;
4624 /* Now that the sources are clearly marked, we can release
4625 * the destination stripes
4627 while (!list_empty(&stripes)) {
4628 sh = list_entry(stripes.next, struct stripe_head, lru);
4629 list_del_init(&sh->lru);
4632 /* If this takes us to the resync_max point where we have to pause,
4633 * then we need to write out the superblock.
4635 sector_nr += reshape_sectors;
4636 if ((sector_nr - mddev->curr_resync_completed) * 2
4637 >= mddev->resync_max - mddev->curr_resync_completed) {
4638 /* Cannot proceed until we've updated the superblock... */
4639 wait_event(conf->wait_for_overlap,
4640 atomic_read(&conf->reshape_stripes) == 0);
4641 mddev->reshape_position = conf->reshape_progress;
4642 mddev->curr_resync_completed = sector_nr;
4643 conf->reshape_checkpoint = jiffies;
4644 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4645 md_wakeup_thread(mddev->thread);
4646 wait_event(mddev->sb_wait,
4647 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4648 || kthread_should_stop());
4649 spin_lock_irq(&conf->device_lock);
4650 conf->reshape_safe = mddev->reshape_position;
4651 spin_unlock_irq(&conf->device_lock);
4652 wake_up(&conf->wait_for_overlap);
4653 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4655 return reshape_sectors;
4658 /* FIXME go_faster isn't used */
4659 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4661 struct r5conf *conf = mddev->private;
4662 struct stripe_head *sh;
4663 sector_t max_sector = mddev->dev_sectors;
4664 sector_t sync_blocks;
4665 int still_degraded = 0;
4668 if (sector_nr >= max_sector) {
4669 /* just being told to finish up .. nothing much to do */
4671 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4676 if (mddev->curr_resync < max_sector) /* aborted */
4677 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4679 else /* completed sync */
4681 bitmap_close_sync(mddev->bitmap);
4686 /* Allow raid5_quiesce to complete */
4687 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4689 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4690 return reshape_request(mddev, sector_nr, skipped);
4692 /* No need to check resync_max as we never do more than one
4693 * stripe, and as resync_max will always be on a chunk boundary,
4694 * if the check in md_do_sync didn't fire, there is no chance
4695 * of overstepping resync_max here
4698 /* if there is too many failed drives and we are trying
4699 * to resync, then assert that we are finished, because there is
4700 * nothing we can do.
4702 if (mddev->degraded >= conf->max_degraded &&
4703 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4704 sector_t rv = mddev->dev_sectors - sector_nr;
4708 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4710 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4711 sync_blocks >= STRIPE_SECTORS) {
4712 /* we can skip this block, and probably more */
4713 sync_blocks /= STRIPE_SECTORS;
4715 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4718 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4720 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4722 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4723 /* make sure we don't swamp the stripe cache if someone else
4724 * is trying to get access
4726 schedule_timeout_uninterruptible(1);
4728 /* Need to check if array will still be degraded after recovery/resync
4729 * We don't need to check the 'failed' flag as when that gets set,
4732 for (i = 0; i < conf->raid_disks; i++)
4733 if (conf->disks[i].rdev == NULL)
4736 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4738 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4743 return STRIPE_SECTORS;
4746 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4748 /* We may not be able to submit a whole bio at once as there
4749 * may not be enough stripe_heads available.
4750 * We cannot pre-allocate enough stripe_heads as we may need
4751 * more than exist in the cache (if we allow ever large chunks).
4752 * So we do one stripe head at a time and record in
4753 * ->bi_hw_segments how many have been done.
4755 * We *know* that this entire raid_bio is in one chunk, so
4756 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4758 struct stripe_head *sh;
4760 sector_t sector, logical_sector, last_sector;
4765 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4766 sector = raid5_compute_sector(conf, logical_sector,
4768 last_sector = bio_end_sector(raid_bio);
4770 for (; logical_sector < last_sector;
4771 logical_sector += STRIPE_SECTORS,
4772 sector += STRIPE_SECTORS,
4775 if (scnt < raid5_bi_processed_stripes(raid_bio))
4776 /* already done this stripe */
4779 sh = get_active_stripe(conf, sector, 0, 1, 0);
4782 /* failed to get a stripe - must wait */
4783 raid5_set_bi_processed_stripes(raid_bio, scnt);
4784 conf->retry_read_aligned = raid_bio;
4788 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4790 raid5_set_bi_processed_stripes(raid_bio, scnt);
4791 conf->retry_read_aligned = raid_bio;
4795 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4800 remaining = raid5_dec_bi_active_stripes(raid_bio);
4801 if (remaining == 0) {
4802 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4804 bio_endio(raid_bio, 0);
4806 if (atomic_dec_and_test(&conf->active_aligned_reads))
4807 wake_up(&conf->wait_for_stripe);
4811 #define MAX_STRIPE_BATCH 8
4812 static int handle_active_stripes(struct r5conf *conf)
4814 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4815 int i, batch_size = 0;
4817 while (batch_size < MAX_STRIPE_BATCH &&
4818 (sh = __get_priority_stripe(conf)) != NULL)
4819 batch[batch_size++] = sh;
4821 if (batch_size == 0)
4823 spin_unlock_irq(&conf->device_lock);
4825 for (i = 0; i < batch_size; i++)
4826 handle_stripe(batch[i]);
4830 spin_lock_irq(&conf->device_lock);
4831 for (i = 0; i < batch_size; i++)
4832 __release_stripe(conf, batch[i]);
4837 * This is our raid5 kernel thread.
4839 * We scan the hash table for stripes which can be handled now.
4840 * During the scan, completed stripes are saved for us by the interrupt
4841 * handler, so that they will not have to wait for our next wakeup.
4843 static void raid5d(struct md_thread *thread)
4845 struct mddev *mddev = thread->mddev;
4846 struct r5conf *conf = mddev->private;
4848 struct blk_plug plug;
4850 pr_debug("+++ raid5d active\n");
4852 md_check_recovery(mddev);
4854 blk_start_plug(&plug);
4856 spin_lock_irq(&conf->device_lock);
4862 !list_empty(&conf->bitmap_list)) {
4863 /* Now is a good time to flush some bitmap updates */
4865 spin_unlock_irq(&conf->device_lock);
4866 bitmap_unplug(mddev->bitmap);
4867 spin_lock_irq(&conf->device_lock);
4868 conf->seq_write = conf->seq_flush;
4869 activate_bit_delay(conf);
4871 raid5_activate_delayed(conf);
4873 while ((bio = remove_bio_from_retry(conf))) {
4875 spin_unlock_irq(&conf->device_lock);
4876 ok = retry_aligned_read(conf, bio);
4877 spin_lock_irq(&conf->device_lock);
4883 batch_size = handle_active_stripes(conf);
4886 handled += batch_size;
4888 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
4889 spin_unlock_irq(&conf->device_lock);
4890 md_check_recovery(mddev);
4891 spin_lock_irq(&conf->device_lock);
4894 pr_debug("%d stripes handled\n", handled);
4896 spin_unlock_irq(&conf->device_lock);
4898 async_tx_issue_pending_all();
4899 blk_finish_plug(&plug);
4901 pr_debug("--- raid5d inactive\n");
4905 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4907 struct r5conf *conf = mddev->private;
4909 return sprintf(page, "%d\n", conf->max_nr_stripes);
4915 raid5_set_cache_size(struct mddev *mddev, int size)
4917 struct r5conf *conf = mddev->private;
4920 if (size <= 16 || size > 32768)
4922 while (size < conf->max_nr_stripes) {
4923 if (drop_one_stripe(conf))
4924 conf->max_nr_stripes--;
4928 err = md_allow_write(mddev);
4931 while (size > conf->max_nr_stripes) {
4932 if (grow_one_stripe(conf))
4933 conf->max_nr_stripes++;
4938 EXPORT_SYMBOL(raid5_set_cache_size);
4941 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4943 struct r5conf *conf = mddev->private;
4947 if (len >= PAGE_SIZE)
4952 if (strict_strtoul(page, 10, &new))
4954 err = raid5_set_cache_size(mddev, new);
4960 static struct md_sysfs_entry
4961 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4962 raid5_show_stripe_cache_size,
4963 raid5_store_stripe_cache_size);
4966 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4968 struct r5conf *conf = mddev->private;
4970 return sprintf(page, "%d\n", conf->bypass_threshold);
4976 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4978 struct r5conf *conf = mddev->private;
4980 if (len >= PAGE_SIZE)
4985 if (strict_strtoul(page, 10, &new))
4987 if (new > conf->max_nr_stripes)
4989 conf->bypass_threshold = new;
4993 static struct md_sysfs_entry
4994 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4996 raid5_show_preread_threshold,
4997 raid5_store_preread_threshold);
5000 stripe_cache_active_show(struct mddev *mddev, char *page)
5002 struct r5conf *conf = mddev->private;
5004 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5009 static struct md_sysfs_entry
5010 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
5012 static struct attribute *raid5_attrs[] = {
5013 &raid5_stripecache_size.attr,
5014 &raid5_stripecache_active.attr,
5015 &raid5_preread_bypass_threshold.attr,
5018 static struct attribute_group raid5_attrs_group = {
5020 .attrs = raid5_attrs,
5024 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
5026 struct r5conf *conf = mddev->private;
5029 sectors = mddev->dev_sectors;
5031 /* size is defined by the smallest of previous and new size */
5032 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
5034 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5035 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
5036 return sectors * (raid_disks - conf->max_degraded);
5039 static void raid5_free_percpu(struct r5conf *conf)
5041 struct raid5_percpu *percpu;
5048 for_each_possible_cpu(cpu) {
5049 percpu = per_cpu_ptr(conf->percpu, cpu);
5050 safe_put_page(percpu->spare_page);
5051 kfree(percpu->scribble);
5053 #ifdef CONFIG_HOTPLUG_CPU
5054 unregister_cpu_notifier(&conf->cpu_notify);
5058 free_percpu(conf->percpu);
5061 static void free_conf(struct r5conf *conf)
5063 shrink_stripes(conf);
5064 raid5_free_percpu(conf);
5066 kfree(conf->stripe_hashtbl);
5070 #ifdef CONFIG_HOTPLUG_CPU
5071 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5074 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5075 long cpu = (long)hcpu;
5076 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5079 case CPU_UP_PREPARE:
5080 case CPU_UP_PREPARE_FROZEN:
5081 if (conf->level == 6 && !percpu->spare_page)
5082 percpu->spare_page = alloc_page(GFP_KERNEL);
5083 if (!percpu->scribble)
5084 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5086 if (!percpu->scribble ||
5087 (conf->level == 6 && !percpu->spare_page)) {
5088 safe_put_page(percpu->spare_page);
5089 kfree(percpu->scribble);
5090 pr_err("%s: failed memory allocation for cpu%ld\n",
5092 return notifier_from_errno(-ENOMEM);
5096 case CPU_DEAD_FROZEN:
5097 safe_put_page(percpu->spare_page);
5098 kfree(percpu->scribble);
5099 percpu->spare_page = NULL;
5100 percpu->scribble = NULL;
5109 static int raid5_alloc_percpu(struct r5conf *conf)
5112 struct page *spare_page;
5113 struct raid5_percpu __percpu *allcpus;
5117 allcpus = alloc_percpu(struct raid5_percpu);
5120 conf->percpu = allcpus;
5124 for_each_present_cpu(cpu) {
5125 if (conf->level == 6) {
5126 spare_page = alloc_page(GFP_KERNEL);
5131 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5133 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5138 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5140 #ifdef CONFIG_HOTPLUG_CPU
5141 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5142 conf->cpu_notify.priority = 0;
5144 err = register_cpu_notifier(&conf->cpu_notify);
5151 static struct r5conf *setup_conf(struct mddev *mddev)
5153 struct r5conf *conf;
5154 int raid_disk, memory, max_disks;
5155 struct md_rdev *rdev;
5156 struct disk_info *disk;
5159 if (mddev->new_level != 5
5160 && mddev->new_level != 4
5161 && mddev->new_level != 6) {
5162 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5163 mdname(mddev), mddev->new_level);
5164 return ERR_PTR(-EIO);
5166 if ((mddev->new_level == 5
5167 && !algorithm_valid_raid5(mddev->new_layout)) ||
5168 (mddev->new_level == 6
5169 && !algorithm_valid_raid6(mddev->new_layout))) {
5170 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5171 mdname(mddev), mddev->new_layout);
5172 return ERR_PTR(-EIO);
5174 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5175 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5176 mdname(mddev), mddev->raid_disks);
5177 return ERR_PTR(-EINVAL);
5180 if (!mddev->new_chunk_sectors ||
5181 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5182 !is_power_of_2(mddev->new_chunk_sectors)) {
5183 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5184 mdname(mddev), mddev->new_chunk_sectors << 9);
5185 return ERR_PTR(-EINVAL);
5188 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5191 spin_lock_init(&conf->device_lock);
5192 init_waitqueue_head(&conf->wait_for_stripe);
5193 init_waitqueue_head(&conf->wait_for_overlap);
5194 INIT_LIST_HEAD(&conf->handle_list);
5195 INIT_LIST_HEAD(&conf->hold_list);
5196 INIT_LIST_HEAD(&conf->delayed_list);
5197 INIT_LIST_HEAD(&conf->bitmap_list);
5198 INIT_LIST_HEAD(&conf->inactive_list);
5199 atomic_set(&conf->active_stripes, 0);
5200 atomic_set(&conf->preread_active_stripes, 0);
5201 atomic_set(&conf->active_aligned_reads, 0);
5202 conf->bypass_threshold = BYPASS_THRESHOLD;
5203 conf->recovery_disabled = mddev->recovery_disabled - 1;
5205 conf->raid_disks = mddev->raid_disks;
5206 if (mddev->reshape_position == MaxSector)
5207 conf->previous_raid_disks = mddev->raid_disks;
5209 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5210 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5211 conf->scribble_len = scribble_len(max_disks);
5213 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5218 conf->mddev = mddev;
5220 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5223 conf->level = mddev->new_level;
5224 if (raid5_alloc_percpu(conf) != 0)
5227 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5229 rdev_for_each(rdev, mddev) {
5230 raid_disk = rdev->raid_disk;
5231 if (raid_disk >= max_disks
5234 disk = conf->disks + raid_disk;
5236 if (test_bit(Replacement, &rdev->flags)) {
5237 if (disk->replacement)
5239 disk->replacement = rdev;
5246 if (test_bit(In_sync, &rdev->flags)) {
5247 char b[BDEVNAME_SIZE];
5248 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5250 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5251 } else if (rdev->saved_raid_disk != raid_disk)
5252 /* Cannot rely on bitmap to complete recovery */
5256 conf->chunk_sectors = mddev->new_chunk_sectors;
5257 conf->level = mddev->new_level;
5258 if (conf->level == 6)
5259 conf->max_degraded = 2;
5261 conf->max_degraded = 1;
5262 conf->algorithm = mddev->new_layout;
5263 conf->max_nr_stripes = NR_STRIPES;
5264 conf->reshape_progress = mddev->reshape_position;
5265 if (conf->reshape_progress != MaxSector) {
5266 conf->prev_chunk_sectors = mddev->chunk_sectors;
5267 conf->prev_algo = mddev->layout;
5270 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5271 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5272 if (grow_stripes(conf, conf->max_nr_stripes)) {
5274 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5275 mdname(mddev), memory);
5278 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5279 mdname(mddev), memory);
5281 sprintf(pers_name, "raid%d", mddev->new_level);
5282 conf->thread = md_register_thread(raid5d, mddev, pers_name);
5283 if (!conf->thread) {
5285 "md/raid:%s: couldn't allocate thread.\n",
5295 return ERR_PTR(-EIO);
5297 return ERR_PTR(-ENOMEM);
5301 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5304 case ALGORITHM_PARITY_0:
5305 if (raid_disk < max_degraded)
5308 case ALGORITHM_PARITY_N:
5309 if (raid_disk >= raid_disks - max_degraded)
5312 case ALGORITHM_PARITY_0_6:
5313 if (raid_disk == 0 ||
5314 raid_disk == raid_disks - 1)
5317 case ALGORITHM_LEFT_ASYMMETRIC_6:
5318 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5319 case ALGORITHM_LEFT_SYMMETRIC_6:
5320 case ALGORITHM_RIGHT_SYMMETRIC_6:
5321 if (raid_disk == raid_disks - 1)
5327 static int run(struct mddev *mddev)
5329 struct r5conf *conf;
5330 int working_disks = 0;
5331 int dirty_parity_disks = 0;
5332 struct md_rdev *rdev;
5333 sector_t reshape_offset = 0;
5335 long long min_offset_diff = 0;
5338 if (mddev->recovery_cp != MaxSector)
5339 printk(KERN_NOTICE "md/raid:%s: not clean"
5340 " -- starting background reconstruction\n",
5343 rdev_for_each(rdev, mddev) {
5345 if (rdev->raid_disk < 0)
5347 diff = (rdev->new_data_offset - rdev->data_offset);
5349 min_offset_diff = diff;
5351 } else if (mddev->reshape_backwards &&
5352 diff < min_offset_diff)
5353 min_offset_diff = diff;
5354 else if (!mddev->reshape_backwards &&
5355 diff > min_offset_diff)
5356 min_offset_diff = diff;
5359 if (mddev->reshape_position != MaxSector) {
5360 /* Check that we can continue the reshape.
5361 * Difficulties arise if the stripe we would write to
5362 * next is at or after the stripe we would read from next.
5363 * For a reshape that changes the number of devices, this
5364 * is only possible for a very short time, and mdadm makes
5365 * sure that time appears to have past before assembling
5366 * the array. So we fail if that time hasn't passed.
5367 * For a reshape that keeps the number of devices the same
5368 * mdadm must be monitoring the reshape can keeping the
5369 * critical areas read-only and backed up. It will start
5370 * the array in read-only mode, so we check for that.
5372 sector_t here_new, here_old;
5374 int max_degraded = (mddev->level == 6 ? 2 : 1);
5376 if (mddev->new_level != mddev->level) {
5377 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5378 "required - aborting.\n",
5382 old_disks = mddev->raid_disks - mddev->delta_disks;
5383 /* reshape_position must be on a new-stripe boundary, and one
5384 * further up in new geometry must map after here in old
5387 here_new = mddev->reshape_position;
5388 if (sector_div(here_new, mddev->new_chunk_sectors *
5389 (mddev->raid_disks - max_degraded))) {
5390 printk(KERN_ERR "md/raid:%s: reshape_position not "
5391 "on a stripe boundary\n", mdname(mddev));
5394 reshape_offset = here_new * mddev->new_chunk_sectors;
5395 /* here_new is the stripe we will write to */
5396 here_old = mddev->reshape_position;
5397 sector_div(here_old, mddev->chunk_sectors *
5398 (old_disks-max_degraded));
5399 /* here_old is the first stripe that we might need to read
5401 if (mddev->delta_disks == 0) {
5402 if ((here_new * mddev->new_chunk_sectors !=
5403 here_old * mddev->chunk_sectors)) {
5404 printk(KERN_ERR "md/raid:%s: reshape position is"
5405 " confused - aborting\n", mdname(mddev));
5408 /* We cannot be sure it is safe to start an in-place
5409 * reshape. It is only safe if user-space is monitoring
5410 * and taking constant backups.
5411 * mdadm always starts a situation like this in
5412 * readonly mode so it can take control before
5413 * allowing any writes. So just check for that.
5415 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5416 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5417 /* not really in-place - so OK */;
5418 else if (mddev->ro == 0) {
5419 printk(KERN_ERR "md/raid:%s: in-place reshape "
5420 "must be started in read-only mode "
5425 } else if (mddev->reshape_backwards
5426 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5427 here_old * mddev->chunk_sectors)
5428 : (here_new * mddev->new_chunk_sectors >=
5429 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5430 /* Reading from the same stripe as writing to - bad */
5431 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5432 "auto-recovery - aborting.\n",
5436 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5438 /* OK, we should be able to continue; */
5440 BUG_ON(mddev->level != mddev->new_level);
5441 BUG_ON(mddev->layout != mddev->new_layout);
5442 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5443 BUG_ON(mddev->delta_disks != 0);
5446 if (mddev->private == NULL)
5447 conf = setup_conf(mddev);
5449 conf = mddev->private;
5452 return PTR_ERR(conf);
5454 conf->min_offset_diff = min_offset_diff;
5455 mddev->thread = conf->thread;
5456 conf->thread = NULL;
5457 mddev->private = conf;
5459 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5461 rdev = conf->disks[i].rdev;
5462 if (!rdev && conf->disks[i].replacement) {
5463 /* The replacement is all we have yet */
5464 rdev = conf->disks[i].replacement;
5465 conf->disks[i].replacement = NULL;
5466 clear_bit(Replacement, &rdev->flags);
5467 conf->disks[i].rdev = rdev;
5471 if (conf->disks[i].replacement &&
5472 conf->reshape_progress != MaxSector) {
5473 /* replacements and reshape simply do not mix. */
5474 printk(KERN_ERR "md: cannot handle concurrent "
5475 "replacement and reshape.\n");
5478 if (test_bit(In_sync, &rdev->flags)) {
5482 /* This disc is not fully in-sync. However if it
5483 * just stored parity (beyond the recovery_offset),
5484 * when we don't need to be concerned about the
5485 * array being dirty.
5486 * When reshape goes 'backwards', we never have
5487 * partially completed devices, so we only need
5488 * to worry about reshape going forwards.
5490 /* Hack because v0.91 doesn't store recovery_offset properly. */
5491 if (mddev->major_version == 0 &&
5492 mddev->minor_version > 90)
5493 rdev->recovery_offset = reshape_offset;
5495 if (rdev->recovery_offset < reshape_offset) {
5496 /* We need to check old and new layout */
5497 if (!only_parity(rdev->raid_disk,
5500 conf->max_degraded))
5503 if (!only_parity(rdev->raid_disk,
5505 conf->previous_raid_disks,
5506 conf->max_degraded))
5508 dirty_parity_disks++;
5512 * 0 for a fully functional array, 1 or 2 for a degraded array.
5514 mddev->degraded = calc_degraded(conf);
5516 if (has_failed(conf)) {
5517 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5518 " (%d/%d failed)\n",
5519 mdname(mddev), mddev->degraded, conf->raid_disks);
5523 /* device size must be a multiple of chunk size */
5524 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5525 mddev->resync_max_sectors = mddev->dev_sectors;
5527 if (mddev->degraded > dirty_parity_disks &&
5528 mddev->recovery_cp != MaxSector) {
5529 if (mddev->ok_start_degraded)
5531 "md/raid:%s: starting dirty degraded array"
5532 " - data corruption possible.\n",
5536 "md/raid:%s: cannot start dirty degraded array.\n",
5542 if (mddev->degraded == 0)
5543 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5544 " devices, algorithm %d\n", mdname(mddev), conf->level,
5545 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5548 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5549 " out of %d devices, algorithm %d\n",
5550 mdname(mddev), conf->level,
5551 mddev->raid_disks - mddev->degraded,
5552 mddev->raid_disks, mddev->new_layout);
5554 print_raid5_conf(conf);
5556 if (conf->reshape_progress != MaxSector) {
5557 conf->reshape_safe = conf->reshape_progress;
5558 atomic_set(&conf->reshape_stripes, 0);
5559 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5560 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5561 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5562 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5563 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5568 /* Ok, everything is just fine now */
5569 if (mddev->to_remove == &raid5_attrs_group)
5570 mddev->to_remove = NULL;
5571 else if (mddev->kobj.sd &&
5572 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5574 "raid5: failed to create sysfs attributes for %s\n",
5576 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5580 bool discard_supported = true;
5581 /* read-ahead size must cover two whole stripes, which
5582 * is 2 * (datadisks) * chunksize where 'n' is the
5583 * number of raid devices
5585 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5586 int stripe = data_disks *
5587 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5588 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5589 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5591 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5593 mddev->queue->backing_dev_info.congested_data = mddev;
5594 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5596 chunk_size = mddev->chunk_sectors << 9;
5597 blk_queue_io_min(mddev->queue, chunk_size);
5598 blk_queue_io_opt(mddev->queue, chunk_size *
5599 (conf->raid_disks - conf->max_degraded));
5601 * We can only discard a whole stripe. It doesn't make sense to
5602 * discard data disk but write parity disk
5604 stripe = stripe * PAGE_SIZE;
5605 /* Round up to power of 2, as discard handling
5606 * currently assumes that */
5607 while ((stripe-1) & stripe)
5608 stripe = (stripe | (stripe-1)) + 1;
5609 mddev->queue->limits.discard_alignment = stripe;
5610 mddev->queue->limits.discard_granularity = stripe;
5612 * unaligned part of discard request will be ignored, so can't
5613 * guarantee discard_zerors_data
5615 mddev->queue->limits.discard_zeroes_data = 0;
5617 blk_queue_max_write_same_sectors(mddev->queue, 0);
5619 rdev_for_each(rdev, mddev) {
5620 disk_stack_limits(mddev->gendisk, rdev->bdev,
5621 rdev->data_offset << 9);
5622 disk_stack_limits(mddev->gendisk, rdev->bdev,
5623 rdev->new_data_offset << 9);
5625 * discard_zeroes_data is required, otherwise data
5626 * could be lost. Consider a scenario: discard a stripe
5627 * (the stripe could be inconsistent if
5628 * discard_zeroes_data is 0); write one disk of the
5629 * stripe (the stripe could be inconsistent again
5630 * depending on which disks are used to calculate
5631 * parity); the disk is broken; The stripe data of this
5634 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5635 !bdev_get_queue(rdev->bdev)->
5636 limits.discard_zeroes_data)
5637 discard_supported = false;
5640 if (discard_supported &&
5641 mddev->queue->limits.max_discard_sectors >= stripe &&
5642 mddev->queue->limits.discard_granularity >= stripe)
5643 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5646 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5652 md_unregister_thread(&mddev->thread);
5653 print_raid5_conf(conf);
5655 mddev->private = NULL;
5656 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5660 static int stop(struct mddev *mddev)
5662 struct r5conf *conf = mddev->private;
5664 md_unregister_thread(&mddev->thread);
5666 mddev->queue->backing_dev_info.congested_fn = NULL;
5668 mddev->private = NULL;
5669 mddev->to_remove = &raid5_attrs_group;
5673 static void status(struct seq_file *seq, struct mddev *mddev)
5675 struct r5conf *conf = mddev->private;
5678 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5679 mddev->chunk_sectors / 2, mddev->layout);
5680 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5681 for (i = 0; i < conf->raid_disks; i++)
5682 seq_printf (seq, "%s",
5683 conf->disks[i].rdev &&
5684 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5685 seq_printf (seq, "]");
5688 static void print_raid5_conf (struct r5conf *conf)
5691 struct disk_info *tmp;
5693 printk(KERN_DEBUG "RAID conf printout:\n");
5695 printk("(conf==NULL)\n");
5698 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5700 conf->raid_disks - conf->mddev->degraded);
5702 for (i = 0; i < conf->raid_disks; i++) {
5703 char b[BDEVNAME_SIZE];
5704 tmp = conf->disks + i;
5706 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5707 i, !test_bit(Faulty, &tmp->rdev->flags),
5708 bdevname(tmp->rdev->bdev, b));
5712 static int raid5_spare_active(struct mddev *mddev)
5715 struct r5conf *conf = mddev->private;
5716 struct disk_info *tmp;
5718 unsigned long flags;
5720 for (i = 0; i < conf->raid_disks; i++) {
5721 tmp = conf->disks + i;
5722 if (tmp->replacement
5723 && tmp->replacement->recovery_offset == MaxSector
5724 && !test_bit(Faulty, &tmp->replacement->flags)
5725 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5726 /* Replacement has just become active. */
5728 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5731 /* Replaced device not technically faulty,
5732 * but we need to be sure it gets removed
5733 * and never re-added.
5735 set_bit(Faulty, &tmp->rdev->flags);
5736 sysfs_notify_dirent_safe(
5737 tmp->rdev->sysfs_state);
5739 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5740 } else if (tmp->rdev
5741 && tmp->rdev->recovery_offset == MaxSector
5742 && !test_bit(Faulty, &tmp->rdev->flags)
5743 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5745 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5748 spin_lock_irqsave(&conf->device_lock, flags);
5749 mddev->degraded = calc_degraded(conf);
5750 spin_unlock_irqrestore(&conf->device_lock, flags);
5751 print_raid5_conf(conf);
5755 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5757 struct r5conf *conf = mddev->private;
5759 int number = rdev->raid_disk;
5760 struct md_rdev **rdevp;
5761 struct disk_info *p = conf->disks + number;
5763 print_raid5_conf(conf);
5764 if (rdev == p->rdev)
5766 else if (rdev == p->replacement)
5767 rdevp = &p->replacement;
5771 if (number >= conf->raid_disks &&
5772 conf->reshape_progress == MaxSector)
5773 clear_bit(In_sync, &rdev->flags);
5775 if (test_bit(In_sync, &rdev->flags) ||
5776 atomic_read(&rdev->nr_pending)) {
5780 /* Only remove non-faulty devices if recovery
5783 if (!test_bit(Faulty, &rdev->flags) &&
5784 mddev->recovery_disabled != conf->recovery_disabled &&
5785 !has_failed(conf) &&
5786 (!p->replacement || p->replacement == rdev) &&
5787 number < conf->raid_disks) {
5793 if (atomic_read(&rdev->nr_pending)) {
5794 /* lost the race, try later */
5797 } else if (p->replacement) {
5798 /* We must have just cleared 'rdev' */
5799 p->rdev = p->replacement;
5800 clear_bit(Replacement, &p->replacement->flags);
5801 smp_mb(); /* Make sure other CPUs may see both as identical
5802 * but will never see neither - if they are careful
5804 p->replacement = NULL;
5805 clear_bit(WantReplacement, &rdev->flags);
5807 /* We might have just removed the Replacement as faulty-
5808 * clear the bit just in case
5810 clear_bit(WantReplacement, &rdev->flags);
5813 print_raid5_conf(conf);
5817 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5819 struct r5conf *conf = mddev->private;
5822 struct disk_info *p;
5824 int last = conf->raid_disks - 1;
5826 if (mddev->recovery_disabled == conf->recovery_disabled)
5829 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5830 /* no point adding a device */
5833 if (rdev->raid_disk >= 0)
5834 first = last = rdev->raid_disk;
5837 * find the disk ... but prefer rdev->saved_raid_disk
5840 if (rdev->saved_raid_disk >= 0 &&
5841 rdev->saved_raid_disk >= first &&
5842 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5843 first = rdev->saved_raid_disk;
5845 for (disk = first; disk <= last; disk++) {
5846 p = conf->disks + disk;
5847 if (p->rdev == NULL) {
5848 clear_bit(In_sync, &rdev->flags);
5849 rdev->raid_disk = disk;
5851 if (rdev->saved_raid_disk != disk)
5853 rcu_assign_pointer(p->rdev, rdev);
5857 for (disk = first; disk <= last; disk++) {
5858 p = conf->disks + disk;
5859 if (test_bit(WantReplacement, &p->rdev->flags) &&
5860 p->replacement == NULL) {
5861 clear_bit(In_sync, &rdev->flags);
5862 set_bit(Replacement, &rdev->flags);
5863 rdev->raid_disk = disk;
5866 rcu_assign_pointer(p->replacement, rdev);
5871 print_raid5_conf(conf);
5875 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5877 /* no resync is happening, and there is enough space
5878 * on all devices, so we can resize.
5879 * We need to make sure resync covers any new space.
5880 * If the array is shrinking we should possibly wait until
5881 * any io in the removed space completes, but it hardly seems
5885 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5886 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5887 if (mddev->external_size &&
5888 mddev->array_sectors > newsize)
5890 if (mddev->bitmap) {
5891 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5895 md_set_array_sectors(mddev, newsize);
5896 set_capacity(mddev->gendisk, mddev->array_sectors);
5897 revalidate_disk(mddev->gendisk);
5898 if (sectors > mddev->dev_sectors &&
5899 mddev->recovery_cp > mddev->dev_sectors) {
5900 mddev->recovery_cp = mddev->dev_sectors;
5901 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5903 mddev->dev_sectors = sectors;
5904 mddev->resync_max_sectors = sectors;
5908 static int check_stripe_cache(struct mddev *mddev)
5910 /* Can only proceed if there are plenty of stripe_heads.
5911 * We need a minimum of one full stripe,, and for sensible progress
5912 * it is best to have about 4 times that.
5913 * If we require 4 times, then the default 256 4K stripe_heads will
5914 * allow for chunk sizes up to 256K, which is probably OK.
5915 * If the chunk size is greater, user-space should request more
5916 * stripe_heads first.
5918 struct r5conf *conf = mddev->private;
5919 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5920 > conf->max_nr_stripes ||
5921 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5922 > conf->max_nr_stripes) {
5923 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5925 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5932 static int check_reshape(struct mddev *mddev)
5934 struct r5conf *conf = mddev->private;
5936 if (mddev->delta_disks == 0 &&
5937 mddev->new_layout == mddev->layout &&
5938 mddev->new_chunk_sectors == mddev->chunk_sectors)
5939 return 0; /* nothing to do */
5940 if (has_failed(conf))
5942 if (mddev->delta_disks < 0) {
5943 /* We might be able to shrink, but the devices must
5944 * be made bigger first.
5945 * For raid6, 4 is the minimum size.
5946 * Otherwise 2 is the minimum
5949 if (mddev->level == 6)
5951 if (mddev->raid_disks + mddev->delta_disks < min)
5955 if (!check_stripe_cache(mddev))
5958 return resize_stripes(conf, (conf->previous_raid_disks
5959 + mddev->delta_disks));
5962 static int raid5_start_reshape(struct mddev *mddev)
5964 struct r5conf *conf = mddev->private;
5965 struct md_rdev *rdev;
5967 unsigned long flags;
5969 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5972 if (!check_stripe_cache(mddev))
5975 if (has_failed(conf))
5978 rdev_for_each(rdev, mddev) {
5979 if (!test_bit(In_sync, &rdev->flags)
5980 && !test_bit(Faulty, &rdev->flags))
5984 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5985 /* Not enough devices even to make a degraded array
5990 /* Refuse to reduce size of the array. Any reductions in
5991 * array size must be through explicit setting of array_size
5994 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5995 < mddev->array_sectors) {
5996 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5997 "before number of disks\n", mdname(mddev));
6001 atomic_set(&conf->reshape_stripes, 0);
6002 spin_lock_irq(&conf->device_lock);
6003 conf->previous_raid_disks = conf->raid_disks;
6004 conf->raid_disks += mddev->delta_disks;
6005 conf->prev_chunk_sectors = conf->chunk_sectors;
6006 conf->chunk_sectors = mddev->new_chunk_sectors;
6007 conf->prev_algo = conf->algorithm;
6008 conf->algorithm = mddev->new_layout;
6010 /* Code that selects data_offset needs to see the generation update
6011 * if reshape_progress has been set - so a memory barrier needed.
6014 if (mddev->reshape_backwards)
6015 conf->reshape_progress = raid5_size(mddev, 0, 0);
6017 conf->reshape_progress = 0;
6018 conf->reshape_safe = conf->reshape_progress;
6019 spin_unlock_irq(&conf->device_lock);
6021 /* Add some new drives, as many as will fit.
6022 * We know there are enough to make the newly sized array work.
6023 * Don't add devices if we are reducing the number of
6024 * devices in the array. This is because it is not possible
6025 * to correctly record the "partially reconstructed" state of
6026 * such devices during the reshape and confusion could result.
6028 if (mddev->delta_disks >= 0) {
6029 rdev_for_each(rdev, mddev)
6030 if (rdev->raid_disk < 0 &&
6031 !test_bit(Faulty, &rdev->flags)) {
6032 if (raid5_add_disk(mddev, rdev) == 0) {
6034 >= conf->previous_raid_disks)
6035 set_bit(In_sync, &rdev->flags);
6037 rdev->recovery_offset = 0;
6039 if (sysfs_link_rdev(mddev, rdev))
6040 /* Failure here is OK */;
6042 } else if (rdev->raid_disk >= conf->previous_raid_disks
6043 && !test_bit(Faulty, &rdev->flags)) {
6044 /* This is a spare that was manually added */
6045 set_bit(In_sync, &rdev->flags);
6048 /* When a reshape changes the number of devices,
6049 * ->degraded is measured against the larger of the
6050 * pre and post number of devices.
6052 spin_lock_irqsave(&conf->device_lock, flags);
6053 mddev->degraded = calc_degraded(conf);
6054 spin_unlock_irqrestore(&conf->device_lock, flags);
6056 mddev->raid_disks = conf->raid_disks;
6057 mddev->reshape_position = conf->reshape_progress;
6058 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6060 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6061 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6062 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6063 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6064 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6066 if (!mddev->sync_thread) {
6067 mddev->recovery = 0;
6068 spin_lock_irq(&conf->device_lock);
6069 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6070 rdev_for_each(rdev, mddev)
6071 rdev->new_data_offset = rdev->data_offset;
6073 conf->reshape_progress = MaxSector;
6074 mddev->reshape_position = MaxSector;
6075 spin_unlock_irq(&conf->device_lock);
6078 conf->reshape_checkpoint = jiffies;
6079 md_wakeup_thread(mddev->sync_thread);
6080 md_new_event(mddev);
6084 /* This is called from the reshape thread and should make any
6085 * changes needed in 'conf'
6087 static void end_reshape(struct r5conf *conf)
6090 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6091 struct md_rdev *rdev;
6093 spin_lock_irq(&conf->device_lock);
6094 conf->previous_raid_disks = conf->raid_disks;
6095 rdev_for_each(rdev, conf->mddev)
6096 rdev->data_offset = rdev->new_data_offset;
6098 conf->reshape_progress = MaxSector;
6099 spin_unlock_irq(&conf->device_lock);
6100 wake_up(&conf->wait_for_overlap);
6102 /* read-ahead size must cover two whole stripes, which is
6103 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6105 if (conf->mddev->queue) {
6106 int data_disks = conf->raid_disks - conf->max_degraded;
6107 int stripe = data_disks * ((conf->chunk_sectors << 9)
6109 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6110 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6115 /* This is called from the raid5d thread with mddev_lock held.
6116 * It makes config changes to the device.
6118 static void raid5_finish_reshape(struct mddev *mddev)
6120 struct r5conf *conf = mddev->private;
6122 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6124 if (mddev->delta_disks > 0) {
6125 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6126 set_capacity(mddev->gendisk, mddev->array_sectors);
6127 revalidate_disk(mddev->gendisk);
6130 spin_lock_irq(&conf->device_lock);
6131 mddev->degraded = calc_degraded(conf);
6132 spin_unlock_irq(&conf->device_lock);
6133 for (d = conf->raid_disks ;
6134 d < conf->raid_disks - mddev->delta_disks;
6136 struct md_rdev *rdev = conf->disks[d].rdev;
6138 clear_bit(In_sync, &rdev->flags);
6139 rdev = conf->disks[d].replacement;
6141 clear_bit(In_sync, &rdev->flags);
6144 mddev->layout = conf->algorithm;
6145 mddev->chunk_sectors = conf->chunk_sectors;
6146 mddev->reshape_position = MaxSector;
6147 mddev->delta_disks = 0;
6148 mddev->reshape_backwards = 0;
6152 static void raid5_quiesce(struct mddev *mddev, int state)
6154 struct r5conf *conf = mddev->private;
6157 case 2: /* resume for a suspend */
6158 wake_up(&conf->wait_for_overlap);
6161 case 1: /* stop all writes */
6162 spin_lock_irq(&conf->device_lock);
6163 /* '2' tells resync/reshape to pause so that all
6164 * active stripes can drain
6167 wait_event_lock_irq(conf->wait_for_stripe,
6168 atomic_read(&conf->active_stripes) == 0 &&
6169 atomic_read(&conf->active_aligned_reads) == 0,
6172 spin_unlock_irq(&conf->device_lock);
6173 /* allow reshape to continue */
6174 wake_up(&conf->wait_for_overlap);
6177 case 0: /* re-enable writes */
6178 spin_lock_irq(&conf->device_lock);
6180 wake_up(&conf->wait_for_stripe);
6181 wake_up(&conf->wait_for_overlap);
6182 spin_unlock_irq(&conf->device_lock);
6188 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6190 struct r0conf *raid0_conf = mddev->private;
6193 /* for raid0 takeover only one zone is supported */
6194 if (raid0_conf->nr_strip_zones > 1) {
6195 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6197 return ERR_PTR(-EINVAL);
6200 sectors = raid0_conf->strip_zone[0].zone_end;
6201 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6202 mddev->dev_sectors = sectors;
6203 mddev->new_level = level;
6204 mddev->new_layout = ALGORITHM_PARITY_N;
6205 mddev->new_chunk_sectors = mddev->chunk_sectors;
6206 mddev->raid_disks += 1;
6207 mddev->delta_disks = 1;
6208 /* make sure it will be not marked as dirty */
6209 mddev->recovery_cp = MaxSector;
6211 return setup_conf(mddev);
6215 static void *raid5_takeover_raid1(struct mddev *mddev)
6219 if (mddev->raid_disks != 2 ||
6220 mddev->degraded > 1)
6221 return ERR_PTR(-EINVAL);
6223 /* Should check if there are write-behind devices? */
6225 chunksect = 64*2; /* 64K by default */
6227 /* The array must be an exact multiple of chunksize */
6228 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6231 if ((chunksect<<9) < STRIPE_SIZE)
6232 /* array size does not allow a suitable chunk size */
6233 return ERR_PTR(-EINVAL);
6235 mddev->new_level = 5;
6236 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6237 mddev->new_chunk_sectors = chunksect;
6239 return setup_conf(mddev);
6242 static void *raid5_takeover_raid6(struct mddev *mddev)
6246 switch (mddev->layout) {
6247 case ALGORITHM_LEFT_ASYMMETRIC_6:
6248 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6250 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6251 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6253 case ALGORITHM_LEFT_SYMMETRIC_6:
6254 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6256 case ALGORITHM_RIGHT_SYMMETRIC_6:
6257 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6259 case ALGORITHM_PARITY_0_6:
6260 new_layout = ALGORITHM_PARITY_0;
6262 case ALGORITHM_PARITY_N:
6263 new_layout = ALGORITHM_PARITY_N;
6266 return ERR_PTR(-EINVAL);
6268 mddev->new_level = 5;
6269 mddev->new_layout = new_layout;
6270 mddev->delta_disks = -1;
6271 mddev->raid_disks -= 1;
6272 return setup_conf(mddev);
6276 static int raid5_check_reshape(struct mddev *mddev)
6278 /* For a 2-drive array, the layout and chunk size can be changed
6279 * immediately as not restriping is needed.
6280 * For larger arrays we record the new value - after validation
6281 * to be used by a reshape pass.
6283 struct r5conf *conf = mddev->private;
6284 int new_chunk = mddev->new_chunk_sectors;
6286 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6288 if (new_chunk > 0) {
6289 if (!is_power_of_2(new_chunk))
6291 if (new_chunk < (PAGE_SIZE>>9))
6293 if (mddev->array_sectors & (new_chunk-1))
6294 /* not factor of array size */
6298 /* They look valid */
6300 if (mddev->raid_disks == 2) {
6301 /* can make the change immediately */
6302 if (mddev->new_layout >= 0) {
6303 conf->algorithm = mddev->new_layout;
6304 mddev->layout = mddev->new_layout;
6306 if (new_chunk > 0) {
6307 conf->chunk_sectors = new_chunk ;
6308 mddev->chunk_sectors = new_chunk;
6310 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6311 md_wakeup_thread(mddev->thread);
6313 return check_reshape(mddev);
6316 static int raid6_check_reshape(struct mddev *mddev)
6318 int new_chunk = mddev->new_chunk_sectors;
6320 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6322 if (new_chunk > 0) {
6323 if (!is_power_of_2(new_chunk))
6325 if (new_chunk < (PAGE_SIZE >> 9))
6327 if (mddev->array_sectors & (new_chunk-1))
6328 /* not factor of array size */
6332 /* They look valid */
6333 return check_reshape(mddev);
6336 static void *raid5_takeover(struct mddev *mddev)
6338 /* raid5 can take over:
6339 * raid0 - if there is only one strip zone - make it a raid4 layout
6340 * raid1 - if there are two drives. We need to know the chunk size
6341 * raid4 - trivial - just use a raid4 layout.
6342 * raid6 - Providing it is a *_6 layout
6344 if (mddev->level == 0)
6345 return raid45_takeover_raid0(mddev, 5);
6346 if (mddev->level == 1)
6347 return raid5_takeover_raid1(mddev);
6348 if (mddev->level == 4) {
6349 mddev->new_layout = ALGORITHM_PARITY_N;
6350 mddev->new_level = 5;
6351 return setup_conf(mddev);
6353 if (mddev->level == 6)
6354 return raid5_takeover_raid6(mddev);
6356 return ERR_PTR(-EINVAL);
6359 static void *raid4_takeover(struct mddev *mddev)
6361 /* raid4 can take over:
6362 * raid0 - if there is only one strip zone
6363 * raid5 - if layout is right
6365 if (mddev->level == 0)
6366 return raid45_takeover_raid0(mddev, 4);
6367 if (mddev->level == 5 &&
6368 mddev->layout == ALGORITHM_PARITY_N) {
6369 mddev->new_layout = 0;
6370 mddev->new_level = 4;
6371 return setup_conf(mddev);
6373 return ERR_PTR(-EINVAL);
6376 static struct md_personality raid5_personality;
6378 static void *raid6_takeover(struct mddev *mddev)
6380 /* Currently can only take over a raid5. We map the
6381 * personality to an equivalent raid6 personality
6382 * with the Q block at the end.
6386 if (mddev->pers != &raid5_personality)
6387 return ERR_PTR(-EINVAL);
6388 if (mddev->degraded > 1)
6389 return ERR_PTR(-EINVAL);
6390 if (mddev->raid_disks > 253)
6391 return ERR_PTR(-EINVAL);
6392 if (mddev->raid_disks < 3)
6393 return ERR_PTR(-EINVAL);
6395 switch (mddev->layout) {
6396 case ALGORITHM_LEFT_ASYMMETRIC:
6397 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6399 case ALGORITHM_RIGHT_ASYMMETRIC:
6400 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6402 case ALGORITHM_LEFT_SYMMETRIC:
6403 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6405 case ALGORITHM_RIGHT_SYMMETRIC:
6406 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6408 case ALGORITHM_PARITY_0:
6409 new_layout = ALGORITHM_PARITY_0_6;
6411 case ALGORITHM_PARITY_N:
6412 new_layout = ALGORITHM_PARITY_N;
6415 return ERR_PTR(-EINVAL);
6417 mddev->new_level = 6;
6418 mddev->new_layout = new_layout;
6419 mddev->delta_disks = 1;
6420 mddev->raid_disks += 1;
6421 return setup_conf(mddev);
6425 static struct md_personality raid6_personality =
6429 .owner = THIS_MODULE,
6430 .make_request = make_request,
6434 .error_handler = error,
6435 .hot_add_disk = raid5_add_disk,
6436 .hot_remove_disk= raid5_remove_disk,
6437 .spare_active = raid5_spare_active,
6438 .sync_request = sync_request,
6439 .resize = raid5_resize,
6441 .check_reshape = raid6_check_reshape,
6442 .start_reshape = raid5_start_reshape,
6443 .finish_reshape = raid5_finish_reshape,
6444 .quiesce = raid5_quiesce,
6445 .takeover = raid6_takeover,
6447 static struct md_personality raid5_personality =
6451 .owner = THIS_MODULE,
6452 .make_request = make_request,
6456 .error_handler = error,
6457 .hot_add_disk = raid5_add_disk,
6458 .hot_remove_disk= raid5_remove_disk,
6459 .spare_active = raid5_spare_active,
6460 .sync_request = sync_request,
6461 .resize = raid5_resize,
6463 .check_reshape = raid5_check_reshape,
6464 .start_reshape = raid5_start_reshape,
6465 .finish_reshape = raid5_finish_reshape,
6466 .quiesce = raid5_quiesce,
6467 .takeover = raid5_takeover,
6470 static struct md_personality raid4_personality =
6474 .owner = THIS_MODULE,
6475 .make_request = make_request,
6479 .error_handler = error,
6480 .hot_add_disk = raid5_add_disk,
6481 .hot_remove_disk= raid5_remove_disk,
6482 .spare_active = raid5_spare_active,
6483 .sync_request = sync_request,
6484 .resize = raid5_resize,
6486 .check_reshape = raid5_check_reshape,
6487 .start_reshape = raid5_start_reshape,
6488 .finish_reshape = raid5_finish_reshape,
6489 .quiesce = raid5_quiesce,
6490 .takeover = raid4_takeover,
6493 static int __init raid5_init(void)
6495 register_md_personality(&raid6_personality);
6496 register_md_personality(&raid5_personality);
6497 register_md_personality(&raid4_personality);
6501 static void raid5_exit(void)
6503 unregister_md_personality(&raid6_personality);
6504 unregister_md_personality(&raid5_personality);
6505 unregister_md_personality(&raid4_personality);
6508 module_init(raid5_init);
6509 module_exit(raid5_exit);
6510 MODULE_LICENSE("GPL");
6511 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6512 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6513 MODULE_ALIAS("md-raid5");
6514 MODULE_ALIAS("md-raid4");
6515 MODULE_ALIAS("md-level-5");
6516 MODULE_ALIAS("md-level-4");
6517 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6518 MODULE_ALIAS("md-raid6");
6519 MODULE_ALIAS("md-level-6");
6521 /* This used to be two separate modules, they were: */
6522 MODULE_ALIAS("raid5");
6523 MODULE_ALIAS("raid6");