2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include <trace/events/block.h>
67 #define NR_STRIPES 256
68 #define STRIPE_SIZE PAGE_SIZE
69 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
70 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
71 #define IO_THRESHOLD 1
72 #define BYPASS_THRESHOLD 1
73 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
74 #define HASH_MASK (NR_HASH - 1)
76 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
78 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
79 return &conf->stripe_hashtbl[hash];
82 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
83 * order without overlap. There may be several bio's per stripe+device, and
84 * a bio could span several devices.
85 * When walking this list for a particular stripe+device, we must never proceed
86 * beyond a bio that extends past this device, as the next bio might no longer
88 * This function is used to determine the 'next' bio in the list, given the sector
89 * of the current stripe+device
91 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
93 int sectors = bio->bi_size >> 9;
94 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
101 * We maintain a biased count of active stripes in the bottom 16 bits of
102 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
104 static inline int raid5_bi_processed_stripes(struct bio *bio)
106 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
107 return (atomic_read(segments) >> 16) & 0xffff;
110 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
112 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
113 return atomic_sub_return(1, segments) & 0xffff;
116 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
118 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
119 atomic_inc(segments);
122 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
125 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
129 old = atomic_read(segments);
130 new = (old & 0xffff) | (cnt << 16);
131 } while (atomic_cmpxchg(segments, old, new) != old);
134 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
136 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
137 atomic_set(segments, cnt);
140 /* Find first data disk in a raid6 stripe */
141 static inline int raid6_d0(struct stripe_head *sh)
144 /* ddf always start from first device */
146 /* md starts just after Q block */
147 if (sh->qd_idx == sh->disks - 1)
150 return sh->qd_idx + 1;
152 static inline int raid6_next_disk(int disk, int raid_disks)
155 return (disk < raid_disks) ? disk : 0;
158 /* When walking through the disks in a raid5, starting at raid6_d0,
159 * We need to map each disk to a 'slot', where the data disks are slot
160 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
161 * is raid_disks-1. This help does that mapping.
163 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
164 int *count, int syndrome_disks)
170 if (idx == sh->pd_idx)
171 return syndrome_disks;
172 if (idx == sh->qd_idx)
173 return syndrome_disks + 1;
179 static void return_io(struct bio *return_bi)
181 struct bio *bi = return_bi;
184 return_bi = bi->bi_next;
187 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
194 static void print_raid5_conf (struct r5conf *conf);
196 static int stripe_operations_active(struct stripe_head *sh)
198 return sh->check_state || sh->reconstruct_state ||
199 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
200 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
203 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
205 BUG_ON(!list_empty(&sh->lru));
206 BUG_ON(atomic_read(&conf->active_stripes)==0);
207 if (test_bit(STRIPE_HANDLE, &sh->state)) {
208 if (test_bit(STRIPE_DELAYED, &sh->state) &&
209 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
210 list_add_tail(&sh->lru, &conf->delayed_list);
211 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
212 sh->bm_seq - conf->seq_write > 0)
213 list_add_tail(&sh->lru, &conf->bitmap_list);
215 clear_bit(STRIPE_DELAYED, &sh->state);
216 clear_bit(STRIPE_BIT_DELAY, &sh->state);
217 list_add_tail(&sh->lru, &conf->handle_list);
219 md_wakeup_thread(conf->mddev->thread);
221 BUG_ON(stripe_operations_active(sh));
222 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
223 if (atomic_dec_return(&conf->preread_active_stripes)
225 md_wakeup_thread(conf->mddev->thread);
226 atomic_dec(&conf->active_stripes);
227 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
228 list_add_tail(&sh->lru, &conf->inactive_list);
229 wake_up(&conf->wait_for_stripe);
230 if (conf->retry_read_aligned)
231 md_wakeup_thread(conf->mddev->thread);
236 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
238 if (atomic_dec_and_test(&sh->count))
239 do_release_stripe(conf, sh);
242 static void release_stripe(struct stripe_head *sh)
244 struct r5conf *conf = sh->raid_conf;
247 local_irq_save(flags);
248 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
249 do_release_stripe(conf, sh);
250 spin_unlock(&conf->device_lock);
252 local_irq_restore(flags);
255 static inline void remove_hash(struct stripe_head *sh)
257 pr_debug("remove_hash(), stripe %llu\n",
258 (unsigned long long)sh->sector);
260 hlist_del_init(&sh->hash);
263 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
265 struct hlist_head *hp = stripe_hash(conf, sh->sector);
267 pr_debug("insert_hash(), stripe %llu\n",
268 (unsigned long long)sh->sector);
270 hlist_add_head(&sh->hash, hp);
274 /* find an idle stripe, make sure it is unhashed, and return it. */
275 static struct stripe_head *get_free_stripe(struct r5conf *conf)
277 struct stripe_head *sh = NULL;
278 struct list_head *first;
280 if (list_empty(&conf->inactive_list))
282 first = conf->inactive_list.next;
283 sh = list_entry(first, struct stripe_head, lru);
284 list_del_init(first);
286 atomic_inc(&conf->active_stripes);
291 static void shrink_buffers(struct stripe_head *sh)
295 int num = sh->raid_conf->pool_size;
297 for (i = 0; i < num ; i++) {
301 sh->dev[i].page = NULL;
306 static int grow_buffers(struct stripe_head *sh)
309 int num = sh->raid_conf->pool_size;
311 for (i = 0; i < num; i++) {
314 if (!(page = alloc_page(GFP_KERNEL))) {
317 sh->dev[i].page = page;
322 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
323 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
324 struct stripe_head *sh);
326 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
328 struct r5conf *conf = sh->raid_conf;
331 BUG_ON(atomic_read(&sh->count) != 0);
332 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
333 BUG_ON(stripe_operations_active(sh));
335 pr_debug("init_stripe called, stripe %llu\n",
336 (unsigned long long)sh->sector);
340 sh->generation = conf->generation - previous;
341 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
343 stripe_set_idx(sector, conf, previous, sh);
347 for (i = sh->disks; i--; ) {
348 struct r5dev *dev = &sh->dev[i];
350 if (dev->toread || dev->read || dev->towrite || dev->written ||
351 test_bit(R5_LOCKED, &dev->flags)) {
352 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
353 (unsigned long long)sh->sector, i, dev->toread,
354 dev->read, dev->towrite, dev->written,
355 test_bit(R5_LOCKED, &dev->flags));
359 raid5_build_block(sh, i, previous);
361 insert_hash(conf, sh);
364 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
367 struct stripe_head *sh;
368 struct hlist_node *hn;
370 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
371 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
372 if (sh->sector == sector && sh->generation == generation)
374 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
379 * Need to check if array has failed when deciding whether to:
381 * - remove non-faulty devices
384 * This determination is simple when no reshape is happening.
385 * However if there is a reshape, we need to carefully check
386 * both the before and after sections.
387 * This is because some failed devices may only affect one
388 * of the two sections, and some non-in_sync devices may
389 * be insync in the section most affected by failed devices.
391 static int calc_degraded(struct r5conf *conf)
393 int degraded, degraded2;
398 for (i = 0; i < conf->previous_raid_disks; i++) {
399 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
400 if (rdev && test_bit(Faulty, &rdev->flags))
401 rdev = rcu_dereference(conf->disks[i].replacement);
402 if (!rdev || test_bit(Faulty, &rdev->flags))
404 else if (test_bit(In_sync, &rdev->flags))
407 /* not in-sync or faulty.
408 * If the reshape increases the number of devices,
409 * this is being recovered by the reshape, so
410 * this 'previous' section is not in_sync.
411 * If the number of devices is being reduced however,
412 * the device can only be part of the array if
413 * we are reverting a reshape, so this section will
416 if (conf->raid_disks >= conf->previous_raid_disks)
420 if (conf->raid_disks == conf->previous_raid_disks)
424 for (i = 0; i < conf->raid_disks; i++) {
425 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
426 if (rdev && test_bit(Faulty, &rdev->flags))
427 rdev = rcu_dereference(conf->disks[i].replacement);
428 if (!rdev || test_bit(Faulty, &rdev->flags))
430 else if (test_bit(In_sync, &rdev->flags))
433 /* not in-sync or faulty.
434 * If reshape increases the number of devices, this
435 * section has already been recovered, else it
436 * almost certainly hasn't.
438 if (conf->raid_disks <= conf->previous_raid_disks)
442 if (degraded2 > degraded)
447 static int has_failed(struct r5conf *conf)
451 if (conf->mddev->reshape_position == MaxSector)
452 return conf->mddev->degraded > conf->max_degraded;
454 degraded = calc_degraded(conf);
455 if (degraded > conf->max_degraded)
460 static struct stripe_head *
461 get_active_stripe(struct r5conf *conf, sector_t sector,
462 int previous, int noblock, int noquiesce)
464 struct stripe_head *sh;
466 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
468 spin_lock_irq(&conf->device_lock);
471 wait_event_lock_irq(conf->wait_for_stripe,
472 conf->quiesce == 0 || noquiesce,
473 conf->device_lock, /* nothing */);
474 sh = __find_stripe(conf, sector, conf->generation - previous);
476 if (!conf->inactive_blocked)
477 sh = get_free_stripe(conf);
478 if (noblock && sh == NULL)
481 conf->inactive_blocked = 1;
482 wait_event_lock_irq(conf->wait_for_stripe,
483 !list_empty(&conf->inactive_list) &&
484 (atomic_read(&conf->active_stripes)
485 < (conf->max_nr_stripes *3/4)
486 || !conf->inactive_blocked),
489 conf->inactive_blocked = 0;
491 init_stripe(sh, sector, previous);
493 if (atomic_read(&sh->count)) {
494 BUG_ON(!list_empty(&sh->lru)
495 && !test_bit(STRIPE_EXPANDING, &sh->state)
496 && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state));
498 if (!test_bit(STRIPE_HANDLE, &sh->state))
499 atomic_inc(&conf->active_stripes);
500 if (list_empty(&sh->lru) &&
501 !test_bit(STRIPE_EXPANDING, &sh->state))
503 list_del_init(&sh->lru);
506 } while (sh == NULL);
509 atomic_inc(&sh->count);
511 spin_unlock_irq(&conf->device_lock);
515 /* Determine if 'data_offset' or 'new_data_offset' should be used
516 * in this stripe_head.
518 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
520 sector_t progress = conf->reshape_progress;
521 /* Need a memory barrier to make sure we see the value
522 * of conf->generation, or ->data_offset that was set before
523 * reshape_progress was updated.
526 if (progress == MaxSector)
528 if (sh->generation == conf->generation - 1)
530 /* We are in a reshape, and this is a new-generation stripe,
531 * so use new_data_offset.
537 raid5_end_read_request(struct bio *bi, int error);
539 raid5_end_write_request(struct bio *bi, int error);
541 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
543 struct r5conf *conf = sh->raid_conf;
544 int i, disks = sh->disks;
548 for (i = disks; i--; ) {
550 int replace_only = 0;
551 struct bio *bi, *rbi;
552 struct md_rdev *rdev, *rrdev = NULL;
553 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
554 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
558 if (test_bit(R5_Discard, &sh->dev[i].flags))
560 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
562 else if (test_and_clear_bit(R5_WantReplace,
563 &sh->dev[i].flags)) {
568 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
571 bi = &sh->dev[i].req;
572 rbi = &sh->dev[i].rreq; /* For writing to replacement */
577 bi->bi_end_io = raid5_end_write_request;
578 rbi->bi_end_io = raid5_end_write_request;
580 bi->bi_end_io = raid5_end_read_request;
583 rrdev = rcu_dereference(conf->disks[i].replacement);
584 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
585 rdev = rcu_dereference(conf->disks[i].rdev);
594 /* We raced and saw duplicates */
597 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
602 if (rdev && test_bit(Faulty, &rdev->flags))
605 atomic_inc(&rdev->nr_pending);
606 if (rrdev && test_bit(Faulty, &rrdev->flags))
609 atomic_inc(&rrdev->nr_pending);
612 /* We have already checked bad blocks for reads. Now
613 * need to check for writes. We never accept write errors
614 * on the replacement, so we don't to check rrdev.
616 while ((rw & WRITE) && rdev &&
617 test_bit(WriteErrorSeen, &rdev->flags)) {
620 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
621 &first_bad, &bad_sectors);
626 set_bit(BlockedBadBlocks, &rdev->flags);
627 if (!conf->mddev->external &&
628 conf->mddev->flags) {
629 /* It is very unlikely, but we might
630 * still need to write out the
631 * bad block log - better give it
633 md_check_recovery(conf->mddev);
636 * Because md_wait_for_blocked_rdev
637 * will dec nr_pending, we must
638 * increment it first.
640 atomic_inc(&rdev->nr_pending);
641 md_wait_for_blocked_rdev(rdev, conf->mddev);
643 /* Acknowledged bad block - skip the write */
644 rdev_dec_pending(rdev, conf->mddev);
650 if (s->syncing || s->expanding || s->expanded
652 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
654 set_bit(STRIPE_IO_STARTED, &sh->state);
656 bi->bi_bdev = rdev->bdev;
657 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
658 __func__, (unsigned long long)sh->sector,
660 atomic_inc(&sh->count);
661 if (use_new_offset(conf, sh))
662 bi->bi_sector = (sh->sector
663 + rdev->new_data_offset);
665 bi->bi_sector = (sh->sector
666 + rdev->data_offset);
667 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
668 bi->bi_rw |= REQ_FLUSH;
670 bi->bi_flags = 1 << BIO_UPTODATE;
672 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
673 bi->bi_io_vec[0].bv_offset = 0;
674 bi->bi_size = STRIPE_SIZE;
677 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
678 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
679 bi, disk_devt(conf->mddev->gendisk),
681 generic_make_request(bi);
684 if (s->syncing || s->expanding || s->expanded
686 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
688 set_bit(STRIPE_IO_STARTED, &sh->state);
690 rbi->bi_bdev = rrdev->bdev;
691 pr_debug("%s: for %llu schedule op %ld on "
692 "replacement disc %d\n",
693 __func__, (unsigned long long)sh->sector,
695 atomic_inc(&sh->count);
696 if (use_new_offset(conf, sh))
697 rbi->bi_sector = (sh->sector
698 + rrdev->new_data_offset);
700 rbi->bi_sector = (sh->sector
701 + rrdev->data_offset);
702 rbi->bi_flags = 1 << BIO_UPTODATE;
704 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
705 rbi->bi_io_vec[0].bv_offset = 0;
706 rbi->bi_size = STRIPE_SIZE;
708 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
709 rbi, disk_devt(conf->mddev->gendisk),
711 generic_make_request(rbi);
713 if (!rdev && !rrdev) {
715 set_bit(STRIPE_DEGRADED, &sh->state);
716 pr_debug("skip op %ld on disc %d for sector %llu\n",
717 bi->bi_rw, i, (unsigned long long)sh->sector);
718 clear_bit(R5_LOCKED, &sh->dev[i].flags);
719 set_bit(STRIPE_HANDLE, &sh->state);
724 static struct dma_async_tx_descriptor *
725 async_copy_data(int frombio, struct bio *bio, struct page *page,
726 sector_t sector, struct dma_async_tx_descriptor *tx)
729 struct page *bio_page;
732 struct async_submit_ctl submit;
733 enum async_tx_flags flags = 0;
735 if (bio->bi_sector >= sector)
736 page_offset = (signed)(bio->bi_sector - sector) * 512;
738 page_offset = (signed)(sector - bio->bi_sector) * -512;
741 flags |= ASYNC_TX_FENCE;
742 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
744 bio_for_each_segment(bvl, bio, i) {
745 int len = bvl->bv_len;
749 if (page_offset < 0) {
750 b_offset = -page_offset;
751 page_offset += b_offset;
755 if (len > 0 && page_offset + len > STRIPE_SIZE)
756 clen = STRIPE_SIZE - page_offset;
761 b_offset += bvl->bv_offset;
762 bio_page = bvl->bv_page;
764 tx = async_memcpy(page, bio_page, page_offset,
765 b_offset, clen, &submit);
767 tx = async_memcpy(bio_page, page, b_offset,
768 page_offset, clen, &submit);
770 /* chain the operations */
771 submit.depend_tx = tx;
773 if (clen < len) /* hit end of page */
781 static void ops_complete_biofill(void *stripe_head_ref)
783 struct stripe_head *sh = stripe_head_ref;
784 struct bio *return_bi = NULL;
787 pr_debug("%s: stripe %llu\n", __func__,
788 (unsigned long long)sh->sector);
790 /* clear completed biofills */
791 for (i = sh->disks; i--; ) {
792 struct r5dev *dev = &sh->dev[i];
794 /* acknowledge completion of a biofill operation */
795 /* and check if we need to reply to a read request,
796 * new R5_Wantfill requests are held off until
797 * !STRIPE_BIOFILL_RUN
799 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
800 struct bio *rbi, *rbi2;
805 while (rbi && rbi->bi_sector <
806 dev->sector + STRIPE_SECTORS) {
807 rbi2 = r5_next_bio(rbi, dev->sector);
808 if (!raid5_dec_bi_active_stripes(rbi)) {
809 rbi->bi_next = return_bi;
816 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
818 return_io(return_bi);
820 set_bit(STRIPE_HANDLE, &sh->state);
824 static void ops_run_biofill(struct stripe_head *sh)
826 struct dma_async_tx_descriptor *tx = NULL;
827 struct async_submit_ctl submit;
830 pr_debug("%s: stripe %llu\n", __func__,
831 (unsigned long long)sh->sector);
833 for (i = sh->disks; i--; ) {
834 struct r5dev *dev = &sh->dev[i];
835 if (test_bit(R5_Wantfill, &dev->flags)) {
837 spin_lock_irq(&sh->stripe_lock);
838 dev->read = rbi = dev->toread;
840 spin_unlock_irq(&sh->stripe_lock);
841 while (rbi && rbi->bi_sector <
842 dev->sector + STRIPE_SECTORS) {
843 tx = async_copy_data(0, rbi, dev->page,
845 rbi = r5_next_bio(rbi, dev->sector);
850 atomic_inc(&sh->count);
851 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
852 async_trigger_callback(&submit);
855 static void mark_target_uptodate(struct stripe_head *sh, int target)
862 tgt = &sh->dev[target];
863 set_bit(R5_UPTODATE, &tgt->flags);
864 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
865 clear_bit(R5_Wantcompute, &tgt->flags);
868 static void ops_complete_compute(void *stripe_head_ref)
870 struct stripe_head *sh = stripe_head_ref;
872 pr_debug("%s: stripe %llu\n", __func__,
873 (unsigned long long)sh->sector);
875 /* mark the computed target(s) as uptodate */
876 mark_target_uptodate(sh, sh->ops.target);
877 mark_target_uptodate(sh, sh->ops.target2);
879 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
880 if (sh->check_state == check_state_compute_run)
881 sh->check_state = check_state_compute_result;
882 set_bit(STRIPE_HANDLE, &sh->state);
886 /* return a pointer to the address conversion region of the scribble buffer */
887 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
888 struct raid5_percpu *percpu)
890 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
893 static struct dma_async_tx_descriptor *
894 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
896 int disks = sh->disks;
897 struct page **xor_srcs = percpu->scribble;
898 int target = sh->ops.target;
899 struct r5dev *tgt = &sh->dev[target];
900 struct page *xor_dest = tgt->page;
902 struct dma_async_tx_descriptor *tx;
903 struct async_submit_ctl submit;
906 pr_debug("%s: stripe %llu block: %d\n",
907 __func__, (unsigned long long)sh->sector, target);
908 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
910 for (i = disks; i--; )
912 xor_srcs[count++] = sh->dev[i].page;
914 atomic_inc(&sh->count);
916 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
917 ops_complete_compute, sh, to_addr_conv(sh, percpu));
918 if (unlikely(count == 1))
919 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
921 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
926 /* set_syndrome_sources - populate source buffers for gen_syndrome
927 * @srcs - (struct page *) array of size sh->disks
928 * @sh - stripe_head to parse
930 * Populates srcs in proper layout order for the stripe and returns the
931 * 'count' of sources to be used in a call to async_gen_syndrome. The P
932 * destination buffer is recorded in srcs[count] and the Q destination
933 * is recorded in srcs[count+1]].
935 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
937 int disks = sh->disks;
938 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
939 int d0_idx = raid6_d0(sh);
943 for (i = 0; i < disks; i++)
949 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
951 srcs[slot] = sh->dev[i].page;
952 i = raid6_next_disk(i, disks);
953 } while (i != d0_idx);
955 return syndrome_disks;
958 static struct dma_async_tx_descriptor *
959 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
961 int disks = sh->disks;
962 struct page **blocks = percpu->scribble;
964 int qd_idx = sh->qd_idx;
965 struct dma_async_tx_descriptor *tx;
966 struct async_submit_ctl submit;
972 if (sh->ops.target < 0)
973 target = sh->ops.target2;
974 else if (sh->ops.target2 < 0)
975 target = sh->ops.target;
977 /* we should only have one valid target */
980 pr_debug("%s: stripe %llu block: %d\n",
981 __func__, (unsigned long long)sh->sector, target);
983 tgt = &sh->dev[target];
984 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
987 atomic_inc(&sh->count);
989 if (target == qd_idx) {
990 count = set_syndrome_sources(blocks, sh);
991 blocks[count] = NULL; /* regenerating p is not necessary */
992 BUG_ON(blocks[count+1] != dest); /* q should already be set */
993 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
994 ops_complete_compute, sh,
995 to_addr_conv(sh, percpu));
996 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
998 /* Compute any data- or p-drive using XOR */
1000 for (i = disks; i-- ; ) {
1001 if (i == target || i == qd_idx)
1003 blocks[count++] = sh->dev[i].page;
1006 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1007 NULL, ops_complete_compute, sh,
1008 to_addr_conv(sh, percpu));
1009 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1015 static struct dma_async_tx_descriptor *
1016 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1018 int i, count, disks = sh->disks;
1019 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1020 int d0_idx = raid6_d0(sh);
1021 int faila = -1, failb = -1;
1022 int target = sh->ops.target;
1023 int target2 = sh->ops.target2;
1024 struct r5dev *tgt = &sh->dev[target];
1025 struct r5dev *tgt2 = &sh->dev[target2];
1026 struct dma_async_tx_descriptor *tx;
1027 struct page **blocks = percpu->scribble;
1028 struct async_submit_ctl submit;
1030 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1031 __func__, (unsigned long long)sh->sector, target, target2);
1032 BUG_ON(target < 0 || target2 < 0);
1033 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1034 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1036 /* we need to open-code set_syndrome_sources to handle the
1037 * slot number conversion for 'faila' and 'failb'
1039 for (i = 0; i < disks ; i++)
1044 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1046 blocks[slot] = sh->dev[i].page;
1052 i = raid6_next_disk(i, disks);
1053 } while (i != d0_idx);
1055 BUG_ON(faila == failb);
1058 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1059 __func__, (unsigned long long)sh->sector, faila, failb);
1061 atomic_inc(&sh->count);
1063 if (failb == syndrome_disks+1) {
1064 /* Q disk is one of the missing disks */
1065 if (faila == syndrome_disks) {
1066 /* Missing P+Q, just recompute */
1067 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1068 ops_complete_compute, sh,
1069 to_addr_conv(sh, percpu));
1070 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1071 STRIPE_SIZE, &submit);
1075 int qd_idx = sh->qd_idx;
1077 /* Missing D+Q: recompute D from P, then recompute Q */
1078 if (target == qd_idx)
1079 data_target = target2;
1081 data_target = target;
1084 for (i = disks; i-- ; ) {
1085 if (i == data_target || i == qd_idx)
1087 blocks[count++] = sh->dev[i].page;
1089 dest = sh->dev[data_target].page;
1090 init_async_submit(&submit,
1091 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1093 to_addr_conv(sh, percpu));
1094 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1097 count = set_syndrome_sources(blocks, sh);
1098 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1099 ops_complete_compute, sh,
1100 to_addr_conv(sh, percpu));
1101 return async_gen_syndrome(blocks, 0, count+2,
1102 STRIPE_SIZE, &submit);
1105 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1106 ops_complete_compute, sh,
1107 to_addr_conv(sh, percpu));
1108 if (failb == syndrome_disks) {
1109 /* We're missing D+P. */
1110 return async_raid6_datap_recov(syndrome_disks+2,
1114 /* We're missing D+D. */
1115 return async_raid6_2data_recov(syndrome_disks+2,
1116 STRIPE_SIZE, faila, failb,
1123 static void ops_complete_prexor(void *stripe_head_ref)
1125 struct stripe_head *sh = stripe_head_ref;
1127 pr_debug("%s: stripe %llu\n", __func__,
1128 (unsigned long long)sh->sector);
1131 static struct dma_async_tx_descriptor *
1132 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1133 struct dma_async_tx_descriptor *tx)
1135 int disks = sh->disks;
1136 struct page **xor_srcs = percpu->scribble;
1137 int count = 0, pd_idx = sh->pd_idx, i;
1138 struct async_submit_ctl submit;
1140 /* existing parity data subtracted */
1141 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1143 pr_debug("%s: stripe %llu\n", __func__,
1144 (unsigned long long)sh->sector);
1146 for (i = disks; i--; ) {
1147 struct r5dev *dev = &sh->dev[i];
1148 /* Only process blocks that are known to be uptodate */
1149 if (test_bit(R5_Wantdrain, &dev->flags))
1150 xor_srcs[count++] = dev->page;
1153 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1154 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1155 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1160 static struct dma_async_tx_descriptor *
1161 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1163 int disks = sh->disks;
1166 pr_debug("%s: stripe %llu\n", __func__,
1167 (unsigned long long)sh->sector);
1169 for (i = disks; i--; ) {
1170 struct r5dev *dev = &sh->dev[i];
1173 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1176 spin_lock_irq(&sh->stripe_lock);
1177 chosen = dev->towrite;
1178 dev->towrite = NULL;
1179 BUG_ON(dev->written);
1180 wbi = dev->written = chosen;
1181 spin_unlock_irq(&sh->stripe_lock);
1183 while (wbi && wbi->bi_sector <
1184 dev->sector + STRIPE_SECTORS) {
1185 if (wbi->bi_rw & REQ_FUA)
1186 set_bit(R5_WantFUA, &dev->flags);
1187 if (wbi->bi_rw & REQ_SYNC)
1188 set_bit(R5_SyncIO, &dev->flags);
1189 if (wbi->bi_rw & REQ_DISCARD)
1190 set_bit(R5_Discard, &dev->flags);
1192 tx = async_copy_data(1, wbi, dev->page,
1194 wbi = r5_next_bio(wbi, dev->sector);
1202 static void ops_complete_reconstruct(void *stripe_head_ref)
1204 struct stripe_head *sh = stripe_head_ref;
1205 int disks = sh->disks;
1206 int pd_idx = sh->pd_idx;
1207 int qd_idx = sh->qd_idx;
1209 bool fua = false, sync = false, discard = false;
1211 pr_debug("%s: stripe %llu\n", __func__,
1212 (unsigned long long)sh->sector);
1214 for (i = disks; i--; ) {
1215 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1216 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1217 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1220 for (i = disks; i--; ) {
1221 struct r5dev *dev = &sh->dev[i];
1223 if (dev->written || i == pd_idx || i == qd_idx) {
1225 set_bit(R5_UPTODATE, &dev->flags);
1227 set_bit(R5_WantFUA, &dev->flags);
1229 set_bit(R5_SyncIO, &dev->flags);
1233 if (sh->reconstruct_state == reconstruct_state_drain_run)
1234 sh->reconstruct_state = reconstruct_state_drain_result;
1235 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1236 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1238 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1239 sh->reconstruct_state = reconstruct_state_result;
1242 set_bit(STRIPE_HANDLE, &sh->state);
1247 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1248 struct dma_async_tx_descriptor *tx)
1250 int disks = sh->disks;
1251 struct page **xor_srcs = percpu->scribble;
1252 struct async_submit_ctl submit;
1253 int count = 0, pd_idx = sh->pd_idx, i;
1254 struct page *xor_dest;
1256 unsigned long flags;
1258 pr_debug("%s: stripe %llu\n", __func__,
1259 (unsigned long long)sh->sector);
1261 for (i = 0; i < sh->disks; i++) {
1264 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1267 if (i >= sh->disks) {
1268 atomic_inc(&sh->count);
1269 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1270 ops_complete_reconstruct(sh);
1273 /* check if prexor is active which means only process blocks
1274 * that are part of a read-modify-write (written)
1276 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1278 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1279 for (i = disks; i--; ) {
1280 struct r5dev *dev = &sh->dev[i];
1282 xor_srcs[count++] = dev->page;
1285 xor_dest = sh->dev[pd_idx].page;
1286 for (i = disks; i--; ) {
1287 struct r5dev *dev = &sh->dev[i];
1289 xor_srcs[count++] = dev->page;
1293 /* 1/ if we prexor'd then the dest is reused as a source
1294 * 2/ if we did not prexor then we are redoing the parity
1295 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1296 * for the synchronous xor case
1298 flags = ASYNC_TX_ACK |
1299 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1301 atomic_inc(&sh->count);
1303 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1304 to_addr_conv(sh, percpu));
1305 if (unlikely(count == 1))
1306 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1308 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1312 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1313 struct dma_async_tx_descriptor *tx)
1315 struct async_submit_ctl submit;
1316 struct page **blocks = percpu->scribble;
1319 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1321 for (i = 0; i < sh->disks; i++) {
1322 if (sh->pd_idx == i || sh->qd_idx == i)
1324 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1327 if (i >= sh->disks) {
1328 atomic_inc(&sh->count);
1329 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1330 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1331 ops_complete_reconstruct(sh);
1335 count = set_syndrome_sources(blocks, sh);
1337 atomic_inc(&sh->count);
1339 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1340 sh, to_addr_conv(sh, percpu));
1341 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1344 static void ops_complete_check(void *stripe_head_ref)
1346 struct stripe_head *sh = stripe_head_ref;
1348 pr_debug("%s: stripe %llu\n", __func__,
1349 (unsigned long long)sh->sector);
1351 sh->check_state = check_state_check_result;
1352 set_bit(STRIPE_HANDLE, &sh->state);
1356 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1358 int disks = sh->disks;
1359 int pd_idx = sh->pd_idx;
1360 int qd_idx = sh->qd_idx;
1361 struct page *xor_dest;
1362 struct page **xor_srcs = percpu->scribble;
1363 struct dma_async_tx_descriptor *tx;
1364 struct async_submit_ctl submit;
1368 pr_debug("%s: stripe %llu\n", __func__,
1369 (unsigned long long)sh->sector);
1372 xor_dest = sh->dev[pd_idx].page;
1373 xor_srcs[count++] = xor_dest;
1374 for (i = disks; i--; ) {
1375 if (i == pd_idx || i == qd_idx)
1377 xor_srcs[count++] = sh->dev[i].page;
1380 init_async_submit(&submit, 0, NULL, NULL, NULL,
1381 to_addr_conv(sh, percpu));
1382 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1383 &sh->ops.zero_sum_result, &submit);
1385 atomic_inc(&sh->count);
1386 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1387 tx = async_trigger_callback(&submit);
1390 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1392 struct page **srcs = percpu->scribble;
1393 struct async_submit_ctl submit;
1396 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1397 (unsigned long long)sh->sector, checkp);
1399 count = set_syndrome_sources(srcs, sh);
1403 atomic_inc(&sh->count);
1404 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1405 sh, to_addr_conv(sh, percpu));
1406 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1407 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1410 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1412 int overlap_clear = 0, i, disks = sh->disks;
1413 struct dma_async_tx_descriptor *tx = NULL;
1414 struct r5conf *conf = sh->raid_conf;
1415 int level = conf->level;
1416 struct raid5_percpu *percpu;
1420 percpu = per_cpu_ptr(conf->percpu, cpu);
1421 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1422 ops_run_biofill(sh);
1426 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1428 tx = ops_run_compute5(sh, percpu);
1430 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1431 tx = ops_run_compute6_1(sh, percpu);
1433 tx = ops_run_compute6_2(sh, percpu);
1435 /* terminate the chain if reconstruct is not set to be run */
1436 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1440 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1441 tx = ops_run_prexor(sh, percpu, tx);
1443 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1444 tx = ops_run_biodrain(sh, tx);
1448 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1450 ops_run_reconstruct5(sh, percpu, tx);
1452 ops_run_reconstruct6(sh, percpu, tx);
1455 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1456 if (sh->check_state == check_state_run)
1457 ops_run_check_p(sh, percpu);
1458 else if (sh->check_state == check_state_run_q)
1459 ops_run_check_pq(sh, percpu, 0);
1460 else if (sh->check_state == check_state_run_pq)
1461 ops_run_check_pq(sh, percpu, 1);
1467 for (i = disks; i--; ) {
1468 struct r5dev *dev = &sh->dev[i];
1469 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1470 wake_up(&sh->raid_conf->wait_for_overlap);
1475 #ifdef CONFIG_MULTICORE_RAID456
1476 static void async_run_ops(void *param, async_cookie_t cookie)
1478 struct stripe_head *sh = param;
1479 unsigned long ops_request = sh->ops.request;
1481 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1482 wake_up(&sh->ops.wait_for_ops);
1484 __raid_run_ops(sh, ops_request);
1488 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1490 /* since handle_stripe can be called outside of raid5d context
1491 * we need to ensure sh->ops.request is de-staged before another
1494 wait_event(sh->ops.wait_for_ops,
1495 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1496 sh->ops.request = ops_request;
1498 atomic_inc(&sh->count);
1499 async_schedule(async_run_ops, sh);
1502 #define raid_run_ops __raid_run_ops
1505 static int grow_one_stripe(struct r5conf *conf)
1507 struct stripe_head *sh;
1508 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1512 sh->raid_conf = conf;
1513 #ifdef CONFIG_MULTICORE_RAID456
1514 init_waitqueue_head(&sh->ops.wait_for_ops);
1517 spin_lock_init(&sh->stripe_lock);
1519 if (grow_buffers(sh)) {
1521 kmem_cache_free(conf->slab_cache, sh);
1524 /* we just created an active stripe so... */
1525 atomic_set(&sh->count, 1);
1526 atomic_inc(&conf->active_stripes);
1527 INIT_LIST_HEAD(&sh->lru);
1532 static int grow_stripes(struct r5conf *conf, int num)
1534 struct kmem_cache *sc;
1535 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1537 if (conf->mddev->gendisk)
1538 sprintf(conf->cache_name[0],
1539 "raid%d-%s", conf->level, mdname(conf->mddev));
1541 sprintf(conf->cache_name[0],
1542 "raid%d-%p", conf->level, conf->mddev);
1543 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1545 conf->active_name = 0;
1546 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1547 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1551 conf->slab_cache = sc;
1552 conf->pool_size = devs;
1554 if (!grow_one_stripe(conf))
1560 * scribble_len - return the required size of the scribble region
1561 * @num - total number of disks in the array
1563 * The size must be enough to contain:
1564 * 1/ a struct page pointer for each device in the array +2
1565 * 2/ room to convert each entry in (1) to its corresponding dma
1566 * (dma_map_page()) or page (page_address()) address.
1568 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1569 * calculate over all devices (not just the data blocks), using zeros in place
1570 * of the P and Q blocks.
1572 static size_t scribble_len(int num)
1576 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1581 static int resize_stripes(struct r5conf *conf, int newsize)
1583 /* Make all the stripes able to hold 'newsize' devices.
1584 * New slots in each stripe get 'page' set to a new page.
1586 * This happens in stages:
1587 * 1/ create a new kmem_cache and allocate the required number of
1589 * 2/ gather all the old stripe_heads and tranfer the pages across
1590 * to the new stripe_heads. This will have the side effect of
1591 * freezing the array as once all stripe_heads have been collected,
1592 * no IO will be possible. Old stripe heads are freed once their
1593 * pages have been transferred over, and the old kmem_cache is
1594 * freed when all stripes are done.
1595 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1596 * we simple return a failre status - no need to clean anything up.
1597 * 4/ allocate new pages for the new slots in the new stripe_heads.
1598 * If this fails, we don't bother trying the shrink the
1599 * stripe_heads down again, we just leave them as they are.
1600 * As each stripe_head is processed the new one is released into
1603 * Once step2 is started, we cannot afford to wait for a write,
1604 * so we use GFP_NOIO allocations.
1606 struct stripe_head *osh, *nsh;
1607 LIST_HEAD(newstripes);
1608 struct disk_info *ndisks;
1611 struct kmem_cache *sc;
1614 if (newsize <= conf->pool_size)
1615 return 0; /* never bother to shrink */
1617 err = md_allow_write(conf->mddev);
1622 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1623 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1628 for (i = conf->max_nr_stripes; i; i--) {
1629 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1633 nsh->raid_conf = conf;
1634 #ifdef CONFIG_MULTICORE_RAID456
1635 init_waitqueue_head(&nsh->ops.wait_for_ops);
1637 spin_lock_init(&nsh->stripe_lock);
1639 list_add(&nsh->lru, &newstripes);
1642 /* didn't get enough, give up */
1643 while (!list_empty(&newstripes)) {
1644 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1645 list_del(&nsh->lru);
1646 kmem_cache_free(sc, nsh);
1648 kmem_cache_destroy(sc);
1651 /* Step 2 - Must use GFP_NOIO now.
1652 * OK, we have enough stripes, start collecting inactive
1653 * stripes and copying them over
1655 list_for_each_entry(nsh, &newstripes, lru) {
1656 spin_lock_irq(&conf->device_lock);
1657 wait_event_lock_irq(conf->wait_for_stripe,
1658 !list_empty(&conf->inactive_list),
1661 osh = get_free_stripe(conf);
1662 spin_unlock_irq(&conf->device_lock);
1663 atomic_set(&nsh->count, 1);
1664 for(i=0; i<conf->pool_size; i++)
1665 nsh->dev[i].page = osh->dev[i].page;
1666 for( ; i<newsize; i++)
1667 nsh->dev[i].page = NULL;
1668 kmem_cache_free(conf->slab_cache, osh);
1670 kmem_cache_destroy(conf->slab_cache);
1673 * At this point, we are holding all the stripes so the array
1674 * is completely stalled, so now is a good time to resize
1675 * conf->disks and the scribble region
1677 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1679 for (i=0; i<conf->raid_disks; i++)
1680 ndisks[i] = conf->disks[i];
1682 conf->disks = ndisks;
1687 conf->scribble_len = scribble_len(newsize);
1688 for_each_present_cpu(cpu) {
1689 struct raid5_percpu *percpu;
1692 percpu = per_cpu_ptr(conf->percpu, cpu);
1693 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1696 kfree(percpu->scribble);
1697 percpu->scribble = scribble;
1705 /* Step 4, return new stripes to service */
1706 while(!list_empty(&newstripes)) {
1707 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1708 list_del_init(&nsh->lru);
1710 for (i=conf->raid_disks; i < newsize; i++)
1711 if (nsh->dev[i].page == NULL) {
1712 struct page *p = alloc_page(GFP_NOIO);
1713 nsh->dev[i].page = p;
1717 release_stripe(nsh);
1719 /* critical section pass, GFP_NOIO no longer needed */
1721 conf->slab_cache = sc;
1722 conf->active_name = 1-conf->active_name;
1723 conf->pool_size = newsize;
1727 static int drop_one_stripe(struct r5conf *conf)
1729 struct stripe_head *sh;
1731 spin_lock_irq(&conf->device_lock);
1732 sh = get_free_stripe(conf);
1733 spin_unlock_irq(&conf->device_lock);
1736 BUG_ON(atomic_read(&sh->count));
1738 kmem_cache_free(conf->slab_cache, sh);
1739 atomic_dec(&conf->active_stripes);
1743 static void shrink_stripes(struct r5conf *conf)
1745 while (drop_one_stripe(conf))
1748 if (conf->slab_cache)
1749 kmem_cache_destroy(conf->slab_cache);
1750 conf->slab_cache = NULL;
1753 static void raid5_end_read_request(struct bio * bi, int error)
1755 struct stripe_head *sh = bi->bi_private;
1756 struct r5conf *conf = sh->raid_conf;
1757 int disks = sh->disks, i;
1758 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1759 char b[BDEVNAME_SIZE];
1760 struct md_rdev *rdev = NULL;
1763 for (i=0 ; i<disks; i++)
1764 if (bi == &sh->dev[i].req)
1767 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1768 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1774 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1775 /* If replacement finished while this request was outstanding,
1776 * 'replacement' might be NULL already.
1777 * In that case it moved down to 'rdev'.
1778 * rdev is not removed until all requests are finished.
1780 rdev = conf->disks[i].replacement;
1782 rdev = conf->disks[i].rdev;
1784 if (use_new_offset(conf, sh))
1785 s = sh->sector + rdev->new_data_offset;
1787 s = sh->sector + rdev->data_offset;
1789 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1790 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1791 /* Note that this cannot happen on a
1792 * replacement device. We just fail those on
1797 "md/raid:%s: read error corrected"
1798 " (%lu sectors at %llu on %s)\n",
1799 mdname(conf->mddev), STRIPE_SECTORS,
1800 (unsigned long long)s,
1801 bdevname(rdev->bdev, b));
1802 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1803 clear_bit(R5_ReadError, &sh->dev[i].flags);
1804 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1805 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1806 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1808 if (atomic_read(&rdev->read_errors))
1809 atomic_set(&rdev->read_errors, 0);
1811 const char *bdn = bdevname(rdev->bdev, b);
1815 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1816 atomic_inc(&rdev->read_errors);
1817 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1820 "md/raid:%s: read error on replacement device "
1821 "(sector %llu on %s).\n",
1822 mdname(conf->mddev),
1823 (unsigned long long)s,
1825 else if (conf->mddev->degraded >= conf->max_degraded) {
1829 "md/raid:%s: read error not correctable "
1830 "(sector %llu on %s).\n",
1831 mdname(conf->mddev),
1832 (unsigned long long)s,
1834 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1839 "md/raid:%s: read error NOT corrected!! "
1840 "(sector %llu on %s).\n",
1841 mdname(conf->mddev),
1842 (unsigned long long)s,
1844 } else if (atomic_read(&rdev->read_errors)
1845 > conf->max_nr_stripes)
1847 "md/raid:%s: Too many read errors, failing device %s.\n",
1848 mdname(conf->mddev), bdn);
1852 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1853 set_bit(R5_ReadError, &sh->dev[i].flags);
1854 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1856 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1858 clear_bit(R5_ReadError, &sh->dev[i].flags);
1859 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1861 && test_bit(In_sync, &rdev->flags)
1862 && rdev_set_badblocks(
1863 rdev, sh->sector, STRIPE_SECTORS, 0)))
1864 md_error(conf->mddev, rdev);
1867 rdev_dec_pending(rdev, conf->mddev);
1868 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1869 set_bit(STRIPE_HANDLE, &sh->state);
1873 static void raid5_end_write_request(struct bio *bi, int error)
1875 struct stripe_head *sh = bi->bi_private;
1876 struct r5conf *conf = sh->raid_conf;
1877 int disks = sh->disks, i;
1878 struct md_rdev *uninitialized_var(rdev);
1879 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1882 int replacement = 0;
1884 for (i = 0 ; i < disks; i++) {
1885 if (bi == &sh->dev[i].req) {
1886 rdev = conf->disks[i].rdev;
1889 if (bi == &sh->dev[i].rreq) {
1890 rdev = conf->disks[i].replacement;
1894 /* rdev was removed and 'replacement'
1895 * replaced it. rdev is not removed
1896 * until all requests are finished.
1898 rdev = conf->disks[i].rdev;
1902 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1903 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1912 md_error(conf->mddev, rdev);
1913 else if (is_badblock(rdev, sh->sector,
1915 &first_bad, &bad_sectors))
1916 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1919 set_bit(WriteErrorSeen, &rdev->flags);
1920 set_bit(R5_WriteError, &sh->dev[i].flags);
1921 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1922 set_bit(MD_RECOVERY_NEEDED,
1923 &rdev->mddev->recovery);
1924 } else if (is_badblock(rdev, sh->sector,
1926 &first_bad, &bad_sectors))
1927 set_bit(R5_MadeGood, &sh->dev[i].flags);
1929 rdev_dec_pending(rdev, conf->mddev);
1931 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1932 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1933 set_bit(STRIPE_HANDLE, &sh->state);
1937 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1939 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1941 struct r5dev *dev = &sh->dev[i];
1943 bio_init(&dev->req);
1944 dev->req.bi_io_vec = &dev->vec;
1946 dev->req.bi_max_vecs++;
1947 dev->req.bi_private = sh;
1948 dev->vec.bv_page = dev->page;
1950 bio_init(&dev->rreq);
1951 dev->rreq.bi_io_vec = &dev->rvec;
1952 dev->rreq.bi_vcnt++;
1953 dev->rreq.bi_max_vecs++;
1954 dev->rreq.bi_private = sh;
1955 dev->rvec.bv_page = dev->page;
1958 dev->sector = compute_blocknr(sh, i, previous);
1961 static void error(struct mddev *mddev, struct md_rdev *rdev)
1963 char b[BDEVNAME_SIZE];
1964 struct r5conf *conf = mddev->private;
1965 unsigned long flags;
1966 pr_debug("raid456: error called\n");
1968 spin_lock_irqsave(&conf->device_lock, flags);
1969 clear_bit(In_sync, &rdev->flags);
1970 mddev->degraded = calc_degraded(conf);
1971 spin_unlock_irqrestore(&conf->device_lock, flags);
1972 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1974 set_bit(Blocked, &rdev->flags);
1975 set_bit(Faulty, &rdev->flags);
1976 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1978 "md/raid:%s: Disk failure on %s, disabling device.\n"
1979 "md/raid:%s: Operation continuing on %d devices.\n",
1981 bdevname(rdev->bdev, b),
1983 conf->raid_disks - mddev->degraded);
1987 * Input: a 'big' sector number,
1988 * Output: index of the data and parity disk, and the sector # in them.
1990 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1991 int previous, int *dd_idx,
1992 struct stripe_head *sh)
1994 sector_t stripe, stripe2;
1995 sector_t chunk_number;
1996 unsigned int chunk_offset;
1999 sector_t new_sector;
2000 int algorithm = previous ? conf->prev_algo
2002 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2003 : conf->chunk_sectors;
2004 int raid_disks = previous ? conf->previous_raid_disks
2006 int data_disks = raid_disks - conf->max_degraded;
2008 /* First compute the information on this sector */
2011 * Compute the chunk number and the sector offset inside the chunk
2013 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2014 chunk_number = r_sector;
2017 * Compute the stripe number
2019 stripe = chunk_number;
2020 *dd_idx = sector_div(stripe, data_disks);
2023 * Select the parity disk based on the user selected algorithm.
2025 pd_idx = qd_idx = -1;
2026 switch(conf->level) {
2028 pd_idx = data_disks;
2031 switch (algorithm) {
2032 case ALGORITHM_LEFT_ASYMMETRIC:
2033 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2034 if (*dd_idx >= pd_idx)
2037 case ALGORITHM_RIGHT_ASYMMETRIC:
2038 pd_idx = sector_div(stripe2, raid_disks);
2039 if (*dd_idx >= pd_idx)
2042 case ALGORITHM_LEFT_SYMMETRIC:
2043 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2044 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2046 case ALGORITHM_RIGHT_SYMMETRIC:
2047 pd_idx = sector_div(stripe2, raid_disks);
2048 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2050 case ALGORITHM_PARITY_0:
2054 case ALGORITHM_PARITY_N:
2055 pd_idx = data_disks;
2063 switch (algorithm) {
2064 case ALGORITHM_LEFT_ASYMMETRIC:
2065 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2066 qd_idx = pd_idx + 1;
2067 if (pd_idx == raid_disks-1) {
2068 (*dd_idx)++; /* Q D D D P */
2070 } else if (*dd_idx >= pd_idx)
2071 (*dd_idx) += 2; /* D D P Q D */
2073 case ALGORITHM_RIGHT_ASYMMETRIC:
2074 pd_idx = sector_div(stripe2, raid_disks);
2075 qd_idx = pd_idx + 1;
2076 if (pd_idx == raid_disks-1) {
2077 (*dd_idx)++; /* Q D D D P */
2079 } else if (*dd_idx >= pd_idx)
2080 (*dd_idx) += 2; /* D D P Q D */
2082 case ALGORITHM_LEFT_SYMMETRIC:
2083 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2084 qd_idx = (pd_idx + 1) % raid_disks;
2085 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2087 case ALGORITHM_RIGHT_SYMMETRIC:
2088 pd_idx = sector_div(stripe2, raid_disks);
2089 qd_idx = (pd_idx + 1) % raid_disks;
2090 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2093 case ALGORITHM_PARITY_0:
2098 case ALGORITHM_PARITY_N:
2099 pd_idx = data_disks;
2100 qd_idx = data_disks + 1;
2103 case ALGORITHM_ROTATING_ZERO_RESTART:
2104 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2105 * of blocks for computing Q is different.
2107 pd_idx = 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_RESTART:
2118 /* Same a left_asymmetric, by first stripe is
2119 * D D D P Q rather than
2123 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2124 qd_idx = pd_idx + 1;
2125 if (pd_idx == raid_disks-1) {
2126 (*dd_idx)++; /* Q D D D P */
2128 } else if (*dd_idx >= pd_idx)
2129 (*dd_idx) += 2; /* D D P Q D */
2133 case ALGORITHM_ROTATING_N_CONTINUE:
2134 /* Same as left_symmetric but Q is before P */
2135 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2136 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2137 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2141 case ALGORITHM_LEFT_ASYMMETRIC_6:
2142 /* RAID5 left_asymmetric, with Q on last device */
2143 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2144 if (*dd_idx >= pd_idx)
2146 qd_idx = raid_disks - 1;
2149 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2150 pd_idx = sector_div(stripe2, raid_disks-1);
2151 if (*dd_idx >= pd_idx)
2153 qd_idx = raid_disks - 1;
2156 case ALGORITHM_LEFT_SYMMETRIC_6:
2157 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2158 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2159 qd_idx = raid_disks - 1;
2162 case ALGORITHM_RIGHT_SYMMETRIC_6:
2163 pd_idx = sector_div(stripe2, raid_disks-1);
2164 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2165 qd_idx = raid_disks - 1;
2168 case ALGORITHM_PARITY_0_6:
2171 qd_idx = raid_disks - 1;
2181 sh->pd_idx = pd_idx;
2182 sh->qd_idx = qd_idx;
2183 sh->ddf_layout = ddf_layout;
2186 * Finally, compute the new sector number
2188 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2193 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2195 struct r5conf *conf = sh->raid_conf;
2196 int raid_disks = sh->disks;
2197 int data_disks = raid_disks - conf->max_degraded;
2198 sector_t new_sector = sh->sector, check;
2199 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2200 : conf->chunk_sectors;
2201 int algorithm = previous ? conf->prev_algo
2205 sector_t chunk_number;
2206 int dummy1, dd_idx = i;
2208 struct stripe_head sh2;
2211 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2212 stripe = new_sector;
2214 if (i == sh->pd_idx)
2216 switch(conf->level) {
2219 switch (algorithm) {
2220 case ALGORITHM_LEFT_ASYMMETRIC:
2221 case ALGORITHM_RIGHT_ASYMMETRIC:
2225 case ALGORITHM_LEFT_SYMMETRIC:
2226 case ALGORITHM_RIGHT_SYMMETRIC:
2229 i -= (sh->pd_idx + 1);
2231 case ALGORITHM_PARITY_0:
2234 case ALGORITHM_PARITY_N:
2241 if (i == sh->qd_idx)
2242 return 0; /* It is the Q disk */
2243 switch (algorithm) {
2244 case ALGORITHM_LEFT_ASYMMETRIC:
2245 case ALGORITHM_RIGHT_ASYMMETRIC:
2246 case ALGORITHM_ROTATING_ZERO_RESTART:
2247 case ALGORITHM_ROTATING_N_RESTART:
2248 if (sh->pd_idx == raid_disks-1)
2249 i--; /* Q D D D P */
2250 else if (i > sh->pd_idx)
2251 i -= 2; /* D D P Q D */
2253 case ALGORITHM_LEFT_SYMMETRIC:
2254 case ALGORITHM_RIGHT_SYMMETRIC:
2255 if (sh->pd_idx == raid_disks-1)
2256 i--; /* Q D D D P */
2261 i -= (sh->pd_idx + 2);
2264 case ALGORITHM_PARITY_0:
2267 case ALGORITHM_PARITY_N:
2269 case ALGORITHM_ROTATING_N_CONTINUE:
2270 /* Like left_symmetric, but P is before Q */
2271 if (sh->pd_idx == 0)
2272 i--; /* P D D D Q */
2277 i -= (sh->pd_idx + 1);
2280 case ALGORITHM_LEFT_ASYMMETRIC_6:
2281 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2285 case ALGORITHM_LEFT_SYMMETRIC_6:
2286 case ALGORITHM_RIGHT_SYMMETRIC_6:
2288 i += data_disks + 1;
2289 i -= (sh->pd_idx + 1);
2291 case ALGORITHM_PARITY_0_6:
2300 chunk_number = stripe * data_disks + i;
2301 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2303 check = raid5_compute_sector(conf, r_sector,
2304 previous, &dummy1, &sh2);
2305 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2306 || sh2.qd_idx != sh->qd_idx) {
2307 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2308 mdname(conf->mddev));
2316 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2317 int rcw, int expand)
2319 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2320 struct r5conf *conf = sh->raid_conf;
2321 int level = conf->level;
2324 /* if we are not expanding this is a proper write request, and
2325 * there will be bios with new data to be drained into the
2329 sh->reconstruct_state = reconstruct_state_drain_run;
2330 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2332 sh->reconstruct_state = reconstruct_state_run;
2334 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2336 for (i = disks; i--; ) {
2337 struct r5dev *dev = &sh->dev[i];
2340 set_bit(R5_LOCKED, &dev->flags);
2341 set_bit(R5_Wantdrain, &dev->flags);
2343 clear_bit(R5_UPTODATE, &dev->flags);
2347 if (s->locked + conf->max_degraded == disks)
2348 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2349 atomic_inc(&conf->pending_full_writes);
2352 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2353 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2355 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2356 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2357 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2358 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2360 for (i = disks; i--; ) {
2361 struct r5dev *dev = &sh->dev[i];
2366 (test_bit(R5_UPTODATE, &dev->flags) ||
2367 test_bit(R5_Wantcompute, &dev->flags))) {
2368 set_bit(R5_Wantdrain, &dev->flags);
2369 set_bit(R5_LOCKED, &dev->flags);
2370 clear_bit(R5_UPTODATE, &dev->flags);
2376 /* keep the parity disk(s) locked while asynchronous operations
2379 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2380 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2384 int qd_idx = sh->qd_idx;
2385 struct r5dev *dev = &sh->dev[qd_idx];
2387 set_bit(R5_LOCKED, &dev->flags);
2388 clear_bit(R5_UPTODATE, &dev->flags);
2392 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2393 __func__, (unsigned long long)sh->sector,
2394 s->locked, s->ops_request);
2398 * Each stripe/dev can have one or more bion attached.
2399 * toread/towrite point to the first in a chain.
2400 * The bi_next chain must be in order.
2402 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2405 struct r5conf *conf = sh->raid_conf;
2408 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2409 (unsigned long long)bi->bi_sector,
2410 (unsigned long long)sh->sector);
2413 * If several bio share a stripe. The bio bi_phys_segments acts as a
2414 * reference count to avoid race. The reference count should already be
2415 * increased before this function is called (for example, in
2416 * make_request()), so other bio sharing this stripe will not free the
2417 * stripe. If a stripe is owned by one stripe, the stripe lock will
2420 spin_lock_irq(&sh->stripe_lock);
2422 bip = &sh->dev[dd_idx].towrite;
2426 bip = &sh->dev[dd_idx].toread;
2427 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2428 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2430 bip = & (*bip)->bi_next;
2432 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2435 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2439 raid5_inc_bi_active_stripes(bi);
2442 /* check if page is covered */
2443 sector_t sector = sh->dev[dd_idx].sector;
2444 for (bi=sh->dev[dd_idx].towrite;
2445 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2446 bi && bi->bi_sector <= sector;
2447 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2448 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2449 sector = bi->bi_sector + (bi->bi_size>>9);
2451 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2452 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2455 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2456 (unsigned long long)(*bip)->bi_sector,
2457 (unsigned long long)sh->sector, dd_idx);
2458 spin_unlock_irq(&sh->stripe_lock);
2460 if (conf->mddev->bitmap && firstwrite) {
2461 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2463 sh->bm_seq = conf->seq_flush+1;
2464 set_bit(STRIPE_BIT_DELAY, &sh->state);
2469 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2470 spin_unlock_irq(&sh->stripe_lock);
2474 static void end_reshape(struct r5conf *conf);
2476 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2477 struct stripe_head *sh)
2479 int sectors_per_chunk =
2480 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2482 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2483 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2485 raid5_compute_sector(conf,
2486 stripe * (disks - conf->max_degraded)
2487 *sectors_per_chunk + chunk_offset,
2493 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2494 struct stripe_head_state *s, int disks,
2495 struct bio **return_bi)
2498 for (i = disks; i--; ) {
2502 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2503 struct md_rdev *rdev;
2505 rdev = rcu_dereference(conf->disks[i].rdev);
2506 if (rdev && test_bit(In_sync, &rdev->flags))
2507 atomic_inc(&rdev->nr_pending);
2512 if (!rdev_set_badblocks(
2516 md_error(conf->mddev, rdev);
2517 rdev_dec_pending(rdev, conf->mddev);
2520 spin_lock_irq(&sh->stripe_lock);
2521 /* fail all writes first */
2522 bi = sh->dev[i].towrite;
2523 sh->dev[i].towrite = NULL;
2524 spin_unlock_irq(&sh->stripe_lock);
2528 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2529 wake_up(&conf->wait_for_overlap);
2531 while (bi && bi->bi_sector <
2532 sh->dev[i].sector + STRIPE_SECTORS) {
2533 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2534 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2535 if (!raid5_dec_bi_active_stripes(bi)) {
2536 md_write_end(conf->mddev);
2537 bi->bi_next = *return_bi;
2543 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2544 STRIPE_SECTORS, 0, 0);
2546 /* and fail all 'written' */
2547 bi = sh->dev[i].written;
2548 sh->dev[i].written = NULL;
2549 if (bi) bitmap_end = 1;
2550 while (bi && bi->bi_sector <
2551 sh->dev[i].sector + STRIPE_SECTORS) {
2552 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2553 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2554 if (!raid5_dec_bi_active_stripes(bi)) {
2555 md_write_end(conf->mddev);
2556 bi->bi_next = *return_bi;
2562 /* fail any reads if this device is non-operational and
2563 * the data has not reached the cache yet.
2565 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2566 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2567 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2568 spin_lock_irq(&sh->stripe_lock);
2569 bi = sh->dev[i].toread;
2570 sh->dev[i].toread = NULL;
2571 spin_unlock_irq(&sh->stripe_lock);
2572 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2573 wake_up(&conf->wait_for_overlap);
2574 while (bi && bi->bi_sector <
2575 sh->dev[i].sector + STRIPE_SECTORS) {
2576 struct bio *nextbi =
2577 r5_next_bio(bi, sh->dev[i].sector);
2578 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2579 if (!raid5_dec_bi_active_stripes(bi)) {
2580 bi->bi_next = *return_bi;
2587 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2588 STRIPE_SECTORS, 0, 0);
2589 /* If we were in the middle of a write the parity block might
2590 * still be locked - so just clear all R5_LOCKED flags
2592 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2595 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2596 if (atomic_dec_and_test(&conf->pending_full_writes))
2597 md_wakeup_thread(conf->mddev->thread);
2601 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2602 struct stripe_head_state *s)
2607 clear_bit(STRIPE_SYNCING, &sh->state);
2610 /* There is nothing more to do for sync/check/repair.
2611 * Don't even need to abort as that is handled elsewhere
2612 * if needed, and not always wanted e.g. if there is a known
2614 * For recover/replace we need to record a bad block on all
2615 * non-sync devices, or abort the recovery
2617 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2618 /* During recovery devices cannot be removed, so
2619 * locking and refcounting of rdevs is not needed
2621 for (i = 0; i < conf->raid_disks; i++) {
2622 struct md_rdev *rdev = conf->disks[i].rdev;
2624 && !test_bit(Faulty, &rdev->flags)
2625 && !test_bit(In_sync, &rdev->flags)
2626 && !rdev_set_badblocks(rdev, sh->sector,
2629 rdev = conf->disks[i].replacement;
2631 && !test_bit(Faulty, &rdev->flags)
2632 && !test_bit(In_sync, &rdev->flags)
2633 && !rdev_set_badblocks(rdev, sh->sector,
2638 conf->recovery_disabled =
2639 conf->mddev->recovery_disabled;
2641 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2644 static int want_replace(struct stripe_head *sh, int disk_idx)
2646 struct md_rdev *rdev;
2648 /* Doing recovery so rcu locking not required */
2649 rdev = sh->raid_conf->disks[disk_idx].replacement;
2651 && !test_bit(Faulty, &rdev->flags)
2652 && !test_bit(In_sync, &rdev->flags)
2653 && (rdev->recovery_offset <= sh->sector
2654 || rdev->mddev->recovery_cp <= sh->sector))
2660 /* fetch_block - checks the given member device to see if its data needs
2661 * to be read or computed to satisfy a request.
2663 * Returns 1 when no more member devices need to be checked, otherwise returns
2664 * 0 to tell the loop in handle_stripe_fill to continue
2666 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2667 int disk_idx, int disks)
2669 struct r5dev *dev = &sh->dev[disk_idx];
2670 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2671 &sh->dev[s->failed_num[1]] };
2673 /* is the data in this block needed, and can we get it? */
2674 if (!test_bit(R5_LOCKED, &dev->flags) &&
2675 !test_bit(R5_UPTODATE, &dev->flags) &&
2677 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2678 s->syncing || s->expanding ||
2679 (s->replacing && want_replace(sh, disk_idx)) ||
2680 (s->failed >= 1 && fdev[0]->toread) ||
2681 (s->failed >= 2 && fdev[1]->toread) ||
2682 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2683 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2684 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2685 /* we would like to get this block, possibly by computing it,
2686 * otherwise read it if the backing disk is insync
2688 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2689 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2690 if ((s->uptodate == disks - 1) &&
2691 (s->failed && (disk_idx == s->failed_num[0] ||
2692 disk_idx == s->failed_num[1]))) {
2693 /* have disk failed, and we're requested to fetch it;
2696 pr_debug("Computing stripe %llu block %d\n",
2697 (unsigned long long)sh->sector, disk_idx);
2698 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2699 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2700 set_bit(R5_Wantcompute, &dev->flags);
2701 sh->ops.target = disk_idx;
2702 sh->ops.target2 = -1; /* no 2nd target */
2704 /* Careful: from this point on 'uptodate' is in the eye
2705 * of raid_run_ops which services 'compute' operations
2706 * before writes. R5_Wantcompute flags a block that will
2707 * be R5_UPTODATE by the time it is needed for a
2708 * subsequent operation.
2712 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2713 /* Computing 2-failure is *very* expensive; only
2714 * do it if failed >= 2
2717 for (other = disks; other--; ) {
2718 if (other == disk_idx)
2720 if (!test_bit(R5_UPTODATE,
2721 &sh->dev[other].flags))
2725 pr_debug("Computing stripe %llu blocks %d,%d\n",
2726 (unsigned long long)sh->sector,
2728 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2729 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2730 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2731 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2732 sh->ops.target = disk_idx;
2733 sh->ops.target2 = other;
2737 } else if (test_bit(R5_Insync, &dev->flags)) {
2738 set_bit(R5_LOCKED, &dev->flags);
2739 set_bit(R5_Wantread, &dev->flags);
2741 pr_debug("Reading block %d (sync=%d)\n",
2742 disk_idx, s->syncing);
2750 * handle_stripe_fill - read or compute data to satisfy pending requests.
2752 static void handle_stripe_fill(struct stripe_head *sh,
2753 struct stripe_head_state *s,
2758 /* look for blocks to read/compute, skip this if a compute
2759 * is already in flight, or if the stripe contents are in the
2760 * midst of changing due to a write
2762 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2763 !sh->reconstruct_state)
2764 for (i = disks; i--; )
2765 if (fetch_block(sh, s, i, disks))
2767 set_bit(STRIPE_HANDLE, &sh->state);
2771 /* handle_stripe_clean_event
2772 * any written block on an uptodate or failed drive can be returned.
2773 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2774 * never LOCKED, so we don't need to test 'failed' directly.
2776 static void handle_stripe_clean_event(struct r5conf *conf,
2777 struct stripe_head *sh, int disks, struct bio **return_bi)
2782 for (i = disks; i--; )
2783 if (sh->dev[i].written) {
2785 if (!test_bit(R5_LOCKED, &dev->flags) &&
2786 (test_bit(R5_UPTODATE, &dev->flags) ||
2787 test_bit(R5_Discard, &dev->flags))) {
2788 /* We can return any write requests */
2789 struct bio *wbi, *wbi2;
2790 pr_debug("Return write for disc %d\n", i);
2791 if (test_and_clear_bit(R5_Discard, &dev->flags))
2792 clear_bit(R5_UPTODATE, &dev->flags);
2794 dev->written = NULL;
2795 while (wbi && wbi->bi_sector <
2796 dev->sector + STRIPE_SECTORS) {
2797 wbi2 = r5_next_bio(wbi, dev->sector);
2798 if (!raid5_dec_bi_active_stripes(wbi)) {
2799 md_write_end(conf->mddev);
2800 wbi->bi_next = *return_bi;
2805 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2807 !test_bit(STRIPE_DEGRADED, &sh->state),
2810 } else if (test_bit(R5_Discard, &sh->dev[i].flags))
2811 clear_bit(R5_Discard, &sh->dev[i].flags);
2813 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2814 if (atomic_dec_and_test(&conf->pending_full_writes))
2815 md_wakeup_thread(conf->mddev->thread);
2818 static void handle_stripe_dirtying(struct r5conf *conf,
2819 struct stripe_head *sh,
2820 struct stripe_head_state *s,
2823 int rmw = 0, rcw = 0, i;
2824 sector_t recovery_cp = conf->mddev->recovery_cp;
2826 /* RAID6 requires 'rcw' in current implementation.
2827 * Otherwise, check whether resync is now happening or should start.
2828 * If yes, then the array is dirty (after unclean shutdown or
2829 * initial creation), so parity in some stripes might be inconsistent.
2830 * In this case, we need to always do reconstruct-write, to ensure
2831 * that in case of drive failure or read-error correction, we
2832 * generate correct data from the parity.
2834 if (conf->max_degraded == 2 ||
2835 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2836 /* Calculate the real rcw later - for now make it
2837 * look like rcw is cheaper
2840 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2841 conf->max_degraded, (unsigned long long)recovery_cp,
2842 (unsigned long long)sh->sector);
2843 } else for (i = disks; i--; ) {
2844 /* would I have to read this buffer for read_modify_write */
2845 struct r5dev *dev = &sh->dev[i];
2846 if ((dev->towrite || i == sh->pd_idx) &&
2847 !test_bit(R5_LOCKED, &dev->flags) &&
2848 !(test_bit(R5_UPTODATE, &dev->flags) ||
2849 test_bit(R5_Wantcompute, &dev->flags))) {
2850 if (test_bit(R5_Insync, &dev->flags))
2853 rmw += 2*disks; /* cannot read it */
2855 /* Would I have to read this buffer for reconstruct_write */
2856 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2857 !test_bit(R5_LOCKED, &dev->flags) &&
2858 !(test_bit(R5_UPTODATE, &dev->flags) ||
2859 test_bit(R5_Wantcompute, &dev->flags))) {
2860 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2865 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2866 (unsigned long long)sh->sector, rmw, rcw);
2867 set_bit(STRIPE_HANDLE, &sh->state);
2868 if (rmw < rcw && rmw > 0) {
2869 /* prefer read-modify-write, but need to get some data */
2870 blk_add_trace_msg(conf->mddev->queue, "raid5 rmw %llu %d",
2871 (unsigned long long)sh->sector, rmw);
2872 for (i = disks; i--; ) {
2873 struct r5dev *dev = &sh->dev[i];
2874 if ((dev->towrite || i == sh->pd_idx) &&
2875 !test_bit(R5_LOCKED, &dev->flags) &&
2876 !(test_bit(R5_UPTODATE, &dev->flags) ||
2877 test_bit(R5_Wantcompute, &dev->flags)) &&
2878 test_bit(R5_Insync, &dev->flags)) {
2880 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2881 pr_debug("Read_old block "
2882 "%d for r-m-w\n", i);
2883 set_bit(R5_LOCKED, &dev->flags);
2884 set_bit(R5_Wantread, &dev->flags);
2887 set_bit(STRIPE_DELAYED, &sh->state);
2888 set_bit(STRIPE_HANDLE, &sh->state);
2893 if (rcw <= rmw && rcw > 0) {
2894 /* want reconstruct write, but need to get some data */
2897 for (i = disks; i--; ) {
2898 struct r5dev *dev = &sh->dev[i];
2899 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2900 i != sh->pd_idx && i != sh->qd_idx &&
2901 !test_bit(R5_LOCKED, &dev->flags) &&
2902 !(test_bit(R5_UPTODATE, &dev->flags) ||
2903 test_bit(R5_Wantcompute, &dev->flags))) {
2905 if (!test_bit(R5_Insync, &dev->flags))
2906 continue; /* it's a failed drive */
2908 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2909 pr_debug("Read_old block "
2910 "%d for Reconstruct\n", i);
2911 set_bit(R5_LOCKED, &dev->flags);
2912 set_bit(R5_Wantread, &dev->flags);
2916 set_bit(STRIPE_DELAYED, &sh->state);
2917 set_bit(STRIPE_HANDLE, &sh->state);
2922 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
2923 (unsigned long long)sh->sector,
2924 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
2926 /* now if nothing is locked, and if we have enough data,
2927 * we can start a write request
2929 /* since handle_stripe can be called at any time we need to handle the
2930 * case where a compute block operation has been submitted and then a
2931 * subsequent call wants to start a write request. raid_run_ops only
2932 * handles the case where compute block and reconstruct are requested
2933 * simultaneously. If this is not the case then new writes need to be
2934 * held off until the compute completes.
2936 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2937 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2938 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2939 schedule_reconstruction(sh, s, rcw == 0, 0);
2942 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2943 struct stripe_head_state *s, int disks)
2945 struct r5dev *dev = NULL;
2947 set_bit(STRIPE_HANDLE, &sh->state);
2949 switch (sh->check_state) {
2950 case check_state_idle:
2951 /* start a new check operation if there are no failures */
2952 if (s->failed == 0) {
2953 BUG_ON(s->uptodate != disks);
2954 sh->check_state = check_state_run;
2955 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2956 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2960 dev = &sh->dev[s->failed_num[0]];
2962 case check_state_compute_result:
2963 sh->check_state = check_state_idle;
2965 dev = &sh->dev[sh->pd_idx];
2967 /* check that a write has not made the stripe insync */
2968 if (test_bit(STRIPE_INSYNC, &sh->state))
2971 /* either failed parity check, or recovery is happening */
2972 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2973 BUG_ON(s->uptodate != disks);
2975 set_bit(R5_LOCKED, &dev->flags);
2977 set_bit(R5_Wantwrite, &dev->flags);
2979 clear_bit(STRIPE_DEGRADED, &sh->state);
2980 set_bit(STRIPE_INSYNC, &sh->state);
2982 case check_state_run:
2983 break; /* we will be called again upon completion */
2984 case check_state_check_result:
2985 sh->check_state = check_state_idle;
2987 /* if a failure occurred during the check operation, leave
2988 * STRIPE_INSYNC not set and let the stripe be handled again
2993 /* handle a successful check operation, if parity is correct
2994 * we are done. Otherwise update the mismatch count and repair
2995 * parity if !MD_RECOVERY_CHECK
2997 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2998 /* parity is correct (on disc,
2999 * not in buffer any more)
3001 set_bit(STRIPE_INSYNC, &sh->state);
3003 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3004 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3005 /* don't try to repair!! */
3006 set_bit(STRIPE_INSYNC, &sh->state);
3008 sh->check_state = check_state_compute_run;
3009 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3010 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3011 set_bit(R5_Wantcompute,
3012 &sh->dev[sh->pd_idx].flags);
3013 sh->ops.target = sh->pd_idx;
3014 sh->ops.target2 = -1;
3019 case check_state_compute_run:
3022 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3023 __func__, sh->check_state,
3024 (unsigned long long) sh->sector);
3030 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3031 struct stripe_head_state *s,
3034 int pd_idx = sh->pd_idx;
3035 int qd_idx = sh->qd_idx;
3038 set_bit(STRIPE_HANDLE, &sh->state);
3040 BUG_ON(s->failed > 2);
3042 /* Want to check and possibly repair P and Q.
3043 * However there could be one 'failed' device, in which
3044 * case we can only check one of them, possibly using the
3045 * other to generate missing data
3048 switch (sh->check_state) {
3049 case check_state_idle:
3050 /* start a new check operation if there are < 2 failures */
3051 if (s->failed == s->q_failed) {
3052 /* The only possible failed device holds Q, so it
3053 * makes sense to check P (If anything else were failed,
3054 * we would have used P to recreate it).
3056 sh->check_state = check_state_run;
3058 if (!s->q_failed && s->failed < 2) {
3059 /* Q is not failed, and we didn't use it to generate
3060 * anything, so it makes sense to check it
3062 if (sh->check_state == check_state_run)
3063 sh->check_state = check_state_run_pq;
3065 sh->check_state = check_state_run_q;
3068 /* discard potentially stale zero_sum_result */
3069 sh->ops.zero_sum_result = 0;
3071 if (sh->check_state == check_state_run) {
3072 /* async_xor_zero_sum destroys the contents of P */
3073 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3076 if (sh->check_state >= check_state_run &&
3077 sh->check_state <= check_state_run_pq) {
3078 /* async_syndrome_zero_sum preserves P and Q, so
3079 * no need to mark them !uptodate here
3081 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3085 /* we have 2-disk failure */
3086 BUG_ON(s->failed != 2);
3088 case check_state_compute_result:
3089 sh->check_state = check_state_idle;
3091 /* check that a write has not made the stripe insync */
3092 if (test_bit(STRIPE_INSYNC, &sh->state))
3095 /* now write out any block on a failed drive,
3096 * or P or Q if they were recomputed
3098 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3099 if (s->failed == 2) {
3100 dev = &sh->dev[s->failed_num[1]];
3102 set_bit(R5_LOCKED, &dev->flags);
3103 set_bit(R5_Wantwrite, &dev->flags);
3105 if (s->failed >= 1) {
3106 dev = &sh->dev[s->failed_num[0]];
3108 set_bit(R5_LOCKED, &dev->flags);
3109 set_bit(R5_Wantwrite, &dev->flags);
3111 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3112 dev = &sh->dev[pd_idx];
3114 set_bit(R5_LOCKED, &dev->flags);
3115 set_bit(R5_Wantwrite, &dev->flags);
3117 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3118 dev = &sh->dev[qd_idx];
3120 set_bit(R5_LOCKED, &dev->flags);
3121 set_bit(R5_Wantwrite, &dev->flags);
3123 clear_bit(STRIPE_DEGRADED, &sh->state);
3125 set_bit(STRIPE_INSYNC, &sh->state);
3127 case check_state_run:
3128 case check_state_run_q:
3129 case check_state_run_pq:
3130 break; /* we will be called again upon completion */
3131 case check_state_check_result:
3132 sh->check_state = check_state_idle;
3134 /* handle a successful check operation, if parity is correct
3135 * we are done. Otherwise update the mismatch count and repair
3136 * parity if !MD_RECOVERY_CHECK
3138 if (sh->ops.zero_sum_result == 0) {
3139 /* both parities are correct */
3141 set_bit(STRIPE_INSYNC, &sh->state);
3143 /* in contrast to the raid5 case we can validate
3144 * parity, but still have a failure to write
3147 sh->check_state = check_state_compute_result;
3148 /* Returning at this point means that we may go
3149 * off and bring p and/or q uptodate again so
3150 * we make sure to check zero_sum_result again
3151 * to verify if p or q need writeback
3155 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3156 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3157 /* don't try to repair!! */
3158 set_bit(STRIPE_INSYNC, &sh->state);
3160 int *target = &sh->ops.target;
3162 sh->ops.target = -1;
3163 sh->ops.target2 = -1;
3164 sh->check_state = check_state_compute_run;
3165 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3166 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3167 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3168 set_bit(R5_Wantcompute,
3169 &sh->dev[pd_idx].flags);
3171 target = &sh->ops.target2;
3174 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3175 set_bit(R5_Wantcompute,
3176 &sh->dev[qd_idx].flags);
3183 case check_state_compute_run:
3186 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3187 __func__, sh->check_state,
3188 (unsigned long long) sh->sector);
3193 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3197 /* We have read all the blocks in this stripe and now we need to
3198 * copy some of them into a target stripe for expand.
3200 struct dma_async_tx_descriptor *tx = NULL;
3201 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3202 for (i = 0; i < sh->disks; i++)
3203 if (i != sh->pd_idx && i != sh->qd_idx) {
3205 struct stripe_head *sh2;
3206 struct async_submit_ctl submit;
3208 sector_t bn = compute_blocknr(sh, i, 1);
3209 sector_t s = raid5_compute_sector(conf, bn, 0,
3211 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3213 /* so far only the early blocks of this stripe
3214 * have been requested. When later blocks
3215 * get requested, we will try again
3218 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3219 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3220 /* must have already done this block */
3221 release_stripe(sh2);
3225 /* place all the copies on one channel */
3226 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3227 tx = async_memcpy(sh2->dev[dd_idx].page,
3228 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3231 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3232 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3233 for (j = 0; j < conf->raid_disks; j++)
3234 if (j != sh2->pd_idx &&
3236 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3238 if (j == conf->raid_disks) {
3239 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3240 set_bit(STRIPE_HANDLE, &sh2->state);
3242 release_stripe(sh2);
3245 /* done submitting copies, wait for them to complete */
3246 async_tx_quiesce(&tx);
3250 * handle_stripe - do things to a stripe.
3252 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3253 * state of various bits to see what needs to be done.
3255 * return some read requests which now have data
3256 * return some write requests which are safely on storage
3257 * schedule a read on some buffers
3258 * schedule a write of some buffers
3259 * return confirmation of parity correctness
3263 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3265 struct r5conf *conf = sh->raid_conf;
3266 int disks = sh->disks;
3269 int do_recovery = 0;
3271 memset(s, 0, sizeof(*s));
3273 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3274 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3275 s->failed_num[0] = -1;
3276 s->failed_num[1] = -1;
3278 /* Now to look around and see what can be done */
3280 for (i=disks; i--; ) {
3281 struct md_rdev *rdev;
3288 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3290 dev->toread, dev->towrite, dev->written);
3291 /* maybe we can reply to a read
3293 * new wantfill requests are only permitted while
3294 * ops_complete_biofill is guaranteed to be inactive
3296 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3297 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3298 set_bit(R5_Wantfill, &dev->flags);
3300 /* now count some things */
3301 if (test_bit(R5_LOCKED, &dev->flags))
3303 if (test_bit(R5_UPTODATE, &dev->flags))
3305 if (test_bit(R5_Wantcompute, &dev->flags)) {
3307 BUG_ON(s->compute > 2);
3310 if (test_bit(R5_Wantfill, &dev->flags))
3312 else if (dev->toread)
3316 if (!test_bit(R5_OVERWRITE, &dev->flags))
3321 /* Prefer to use the replacement for reads, but only
3322 * if it is recovered enough and has no bad blocks.
3324 rdev = rcu_dereference(conf->disks[i].replacement);
3325 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3326 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3327 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3328 &first_bad, &bad_sectors))
3329 set_bit(R5_ReadRepl, &dev->flags);
3332 set_bit(R5_NeedReplace, &dev->flags);
3333 rdev = rcu_dereference(conf->disks[i].rdev);
3334 clear_bit(R5_ReadRepl, &dev->flags);
3336 if (rdev && test_bit(Faulty, &rdev->flags))
3339 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3340 &first_bad, &bad_sectors);
3341 if (s->blocked_rdev == NULL
3342 && (test_bit(Blocked, &rdev->flags)
3345 set_bit(BlockedBadBlocks,
3347 s->blocked_rdev = rdev;
3348 atomic_inc(&rdev->nr_pending);
3351 clear_bit(R5_Insync, &dev->flags);
3355 /* also not in-sync */
3356 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3357 test_bit(R5_UPTODATE, &dev->flags)) {
3358 /* treat as in-sync, but with a read error
3359 * which we can now try to correct
3361 set_bit(R5_Insync, &dev->flags);
3362 set_bit(R5_ReadError, &dev->flags);
3364 } else if (test_bit(In_sync, &rdev->flags))
3365 set_bit(R5_Insync, &dev->flags);
3366 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3367 /* in sync if before recovery_offset */
3368 set_bit(R5_Insync, &dev->flags);
3369 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3370 test_bit(R5_Expanded, &dev->flags))
3371 /* If we've reshaped into here, we assume it is Insync.
3372 * We will shortly update recovery_offset to make
3375 set_bit(R5_Insync, &dev->flags);
3377 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3378 /* This flag does not apply to '.replacement'
3379 * only to .rdev, so make sure to check that*/
3380 struct md_rdev *rdev2 = rcu_dereference(
3381 conf->disks[i].rdev);
3383 clear_bit(R5_Insync, &dev->flags);
3384 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3385 s->handle_bad_blocks = 1;
3386 atomic_inc(&rdev2->nr_pending);
3388 clear_bit(R5_WriteError, &dev->flags);
3390 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3391 /* This flag does not apply to '.replacement'
3392 * only to .rdev, so make sure to check that*/
3393 struct md_rdev *rdev2 = rcu_dereference(
3394 conf->disks[i].rdev);
3395 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3396 s->handle_bad_blocks = 1;
3397 atomic_inc(&rdev2->nr_pending);
3399 clear_bit(R5_MadeGood, &dev->flags);
3401 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3402 struct md_rdev *rdev2 = rcu_dereference(
3403 conf->disks[i].replacement);
3404 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3405 s->handle_bad_blocks = 1;
3406 atomic_inc(&rdev2->nr_pending);
3408 clear_bit(R5_MadeGoodRepl, &dev->flags);
3410 if (!test_bit(R5_Insync, &dev->flags)) {
3411 /* The ReadError flag will just be confusing now */
3412 clear_bit(R5_ReadError, &dev->flags);
3413 clear_bit(R5_ReWrite, &dev->flags);
3415 if (test_bit(R5_ReadError, &dev->flags))
3416 clear_bit(R5_Insync, &dev->flags);
3417 if (!test_bit(R5_Insync, &dev->flags)) {
3419 s->failed_num[s->failed] = i;
3421 if (rdev && !test_bit(Faulty, &rdev->flags))
3425 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3426 /* If there is a failed device being replaced,
3427 * we must be recovering.
3428 * else if we are after recovery_cp, we must be syncing
3429 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3430 * else we can only be replacing
3431 * sync and recovery both need to read all devices, and so
3432 * use the same flag.
3435 sh->sector >= conf->mddev->recovery_cp ||
3436 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3444 static void handle_stripe(struct stripe_head *sh)
3446 struct stripe_head_state s;
3447 struct r5conf *conf = sh->raid_conf;
3450 int disks = sh->disks;
3451 struct r5dev *pdev, *qdev;
3453 clear_bit(STRIPE_HANDLE, &sh->state);
3454 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3455 /* already being handled, ensure it gets handled
3456 * again when current action finishes */
3457 set_bit(STRIPE_HANDLE, &sh->state);
3461 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3462 set_bit(STRIPE_SYNCING, &sh->state);
3463 clear_bit(STRIPE_INSYNC, &sh->state);
3465 clear_bit(STRIPE_DELAYED, &sh->state);
3467 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3468 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3469 (unsigned long long)sh->sector, sh->state,
3470 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3471 sh->check_state, sh->reconstruct_state);
3473 analyse_stripe(sh, &s);
3475 if (s.handle_bad_blocks) {
3476 set_bit(STRIPE_HANDLE, &sh->state);
3480 if (unlikely(s.blocked_rdev)) {
3481 if (s.syncing || s.expanding || s.expanded ||
3482 s.replacing || s.to_write || s.written) {
3483 set_bit(STRIPE_HANDLE, &sh->state);
3486 /* There is nothing for the blocked_rdev to block */
3487 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3488 s.blocked_rdev = NULL;
3491 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3492 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3493 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3496 pr_debug("locked=%d uptodate=%d to_read=%d"
3497 " to_write=%d failed=%d failed_num=%d,%d\n",
3498 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3499 s.failed_num[0], s.failed_num[1]);
3500 /* check if the array has lost more than max_degraded devices and,
3501 * if so, some requests might need to be failed.
3503 if (s.failed > conf->max_degraded) {
3504 sh->check_state = 0;
3505 sh->reconstruct_state = 0;
3506 if (s.to_read+s.to_write+s.written)
3507 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3508 if (s.syncing + s.replacing)
3509 handle_failed_sync(conf, sh, &s);
3512 /* Now we check to see if any write operations have recently
3516 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3518 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3519 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3520 sh->reconstruct_state = reconstruct_state_idle;
3522 /* All the 'written' buffers and the parity block are ready to
3523 * be written back to disk
3525 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3526 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3527 BUG_ON(sh->qd_idx >= 0 &&
3528 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3529 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3530 for (i = disks; i--; ) {
3531 struct r5dev *dev = &sh->dev[i];
3532 if (test_bit(R5_LOCKED, &dev->flags) &&
3533 (i == sh->pd_idx || i == sh->qd_idx ||
3535 pr_debug("Writing block %d\n", i);
3536 set_bit(R5_Wantwrite, &dev->flags);
3539 if (!test_bit(R5_Insync, &dev->flags) ||
3540 ((i == sh->pd_idx || i == sh->qd_idx) &&
3542 set_bit(STRIPE_INSYNC, &sh->state);
3545 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3546 s.dec_preread_active = 1;
3550 * might be able to return some write requests if the parity blocks
3551 * are safe, or on a failed drive
3553 pdev = &sh->dev[sh->pd_idx];
3554 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3555 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3556 qdev = &sh->dev[sh->qd_idx];
3557 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3558 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3562 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3563 && !test_bit(R5_LOCKED, &pdev->flags)
3564 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3565 test_bit(R5_Discard, &pdev->flags))))) &&
3566 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3567 && !test_bit(R5_LOCKED, &qdev->flags)
3568 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3569 test_bit(R5_Discard, &qdev->flags))))))
3570 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3572 /* Now we might consider reading some blocks, either to check/generate
3573 * parity, or to satisfy requests
3574 * or to load a block that is being partially written.
3576 if (s.to_read || s.non_overwrite
3577 || (conf->level == 6 && s.to_write && s.failed)
3578 || (s.syncing && (s.uptodate + s.compute < disks))
3581 handle_stripe_fill(sh, &s, disks);
3583 /* Now to consider new write requests and what else, if anything
3584 * should be read. We do not handle new writes when:
3585 * 1/ A 'write' operation (copy+xor) is already in flight.
3586 * 2/ A 'check' operation is in flight, as it may clobber the parity
3589 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3590 handle_stripe_dirtying(conf, sh, &s, disks);
3592 /* maybe we need to check and possibly fix the parity for this stripe
3593 * Any reads will already have been scheduled, so we just see if enough
3594 * data is available. The parity check is held off while parity
3595 * dependent operations are in flight.
3597 if (sh->check_state ||
3598 (s.syncing && s.locked == 0 &&
3599 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3600 !test_bit(STRIPE_INSYNC, &sh->state))) {
3601 if (conf->level == 6)
3602 handle_parity_checks6(conf, sh, &s, disks);
3604 handle_parity_checks5(conf, sh, &s, disks);
3607 if (s.replacing && s.locked == 0
3608 && !test_bit(STRIPE_INSYNC, &sh->state)) {
3609 /* Write out to replacement devices where possible */
3610 for (i = 0; i < conf->raid_disks; i++)
3611 if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3612 test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3613 set_bit(R5_WantReplace, &sh->dev[i].flags);
3614 set_bit(R5_LOCKED, &sh->dev[i].flags);
3617 set_bit(STRIPE_INSYNC, &sh->state);
3619 if ((s.syncing || s.replacing) && s.locked == 0 &&
3620 test_bit(STRIPE_INSYNC, &sh->state)) {
3621 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3622 clear_bit(STRIPE_SYNCING, &sh->state);
3625 /* If the failed drives are just a ReadError, then we might need
3626 * to progress the repair/check process
3628 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3629 for (i = 0; i < s.failed; i++) {
3630 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3631 if (test_bit(R5_ReadError, &dev->flags)
3632 && !test_bit(R5_LOCKED, &dev->flags)
3633 && test_bit(R5_UPTODATE, &dev->flags)
3635 if (!test_bit(R5_ReWrite, &dev->flags)) {
3636 set_bit(R5_Wantwrite, &dev->flags);
3637 set_bit(R5_ReWrite, &dev->flags);
3638 set_bit(R5_LOCKED, &dev->flags);
3641 /* let's read it back */
3642 set_bit(R5_Wantread, &dev->flags);
3643 set_bit(R5_LOCKED, &dev->flags);
3650 /* Finish reconstruct operations initiated by the expansion process */
3651 if (sh->reconstruct_state == reconstruct_state_result) {
3652 struct stripe_head *sh_src
3653 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3654 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3655 /* sh cannot be written until sh_src has been read.
3656 * so arrange for sh to be delayed a little
3658 set_bit(STRIPE_DELAYED, &sh->state);
3659 set_bit(STRIPE_HANDLE, &sh->state);
3660 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3662 atomic_inc(&conf->preread_active_stripes);
3663 release_stripe(sh_src);
3667 release_stripe(sh_src);
3669 sh->reconstruct_state = reconstruct_state_idle;
3670 clear_bit(STRIPE_EXPANDING, &sh->state);
3671 for (i = conf->raid_disks; i--; ) {
3672 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3673 set_bit(R5_LOCKED, &sh->dev[i].flags);
3678 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3679 !sh->reconstruct_state) {
3680 /* Need to write out all blocks after computing parity */
3681 sh->disks = conf->raid_disks;
3682 stripe_set_idx(sh->sector, conf, 0, sh);
3683 schedule_reconstruction(sh, &s, 1, 1);
3684 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3685 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3686 atomic_dec(&conf->reshape_stripes);
3687 wake_up(&conf->wait_for_overlap);
3688 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3691 if (s.expanding && s.locked == 0 &&
3692 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3693 handle_stripe_expansion(conf, sh);
3696 /* wait for this device to become unblocked */
3697 if (unlikely(s.blocked_rdev)) {
3698 if (conf->mddev->external)
3699 md_wait_for_blocked_rdev(s.blocked_rdev,
3702 /* Internal metadata will immediately
3703 * be written by raid5d, so we don't
3704 * need to wait here.
3706 rdev_dec_pending(s.blocked_rdev,
3710 if (s.handle_bad_blocks)
3711 for (i = disks; i--; ) {
3712 struct md_rdev *rdev;
3713 struct r5dev *dev = &sh->dev[i];
3714 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3715 /* We own a safe reference to the rdev */
3716 rdev = conf->disks[i].rdev;
3717 if (!rdev_set_badblocks(rdev, sh->sector,
3719 md_error(conf->mddev, rdev);
3720 rdev_dec_pending(rdev, conf->mddev);
3722 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3723 rdev = conf->disks[i].rdev;
3724 rdev_clear_badblocks(rdev, sh->sector,
3726 rdev_dec_pending(rdev, conf->mddev);
3728 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3729 rdev = conf->disks[i].replacement;
3731 /* rdev have been moved down */
3732 rdev = conf->disks[i].rdev;
3733 rdev_clear_badblocks(rdev, sh->sector,
3735 rdev_dec_pending(rdev, conf->mddev);
3740 raid_run_ops(sh, s.ops_request);
3744 if (s.dec_preread_active) {
3745 /* We delay this until after ops_run_io so that if make_request
3746 * is waiting on a flush, it won't continue until the writes
3747 * have actually been submitted.
3749 atomic_dec(&conf->preread_active_stripes);
3750 if (atomic_read(&conf->preread_active_stripes) <
3752 md_wakeup_thread(conf->mddev->thread);
3755 return_io(s.return_bi);
3757 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3760 static void raid5_activate_delayed(struct r5conf *conf)
3762 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3763 while (!list_empty(&conf->delayed_list)) {
3764 struct list_head *l = conf->delayed_list.next;
3765 struct stripe_head *sh;
3766 sh = list_entry(l, struct stripe_head, lru);
3768 clear_bit(STRIPE_DELAYED, &sh->state);
3769 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3770 atomic_inc(&conf->preread_active_stripes);
3771 list_add_tail(&sh->lru, &conf->hold_list);
3776 static void activate_bit_delay(struct r5conf *conf)
3778 /* device_lock is held */
3779 struct list_head head;
3780 list_add(&head, &conf->bitmap_list);
3781 list_del_init(&conf->bitmap_list);
3782 while (!list_empty(&head)) {
3783 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3784 list_del_init(&sh->lru);
3785 atomic_inc(&sh->count);
3786 __release_stripe(conf, sh);
3790 int md_raid5_congested(struct mddev *mddev, int bits)
3792 struct r5conf *conf = mddev->private;
3794 /* No difference between reads and writes. Just check
3795 * how busy the stripe_cache is
3798 if (conf->inactive_blocked)
3802 if (list_empty_careful(&conf->inactive_list))
3807 EXPORT_SYMBOL_GPL(md_raid5_congested);
3809 static int raid5_congested(void *data, int bits)
3811 struct mddev *mddev = data;
3813 return mddev_congested(mddev, bits) ||
3814 md_raid5_congested(mddev, bits);
3817 /* We want read requests to align with chunks where possible,
3818 * but write requests don't need to.
3820 static int raid5_mergeable_bvec(struct request_queue *q,
3821 struct bvec_merge_data *bvm,
3822 struct bio_vec *biovec)
3824 struct mddev *mddev = q->queuedata;
3825 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3827 unsigned int chunk_sectors = mddev->chunk_sectors;
3828 unsigned int bio_sectors = bvm->bi_size >> 9;
3830 if ((bvm->bi_rw & 1) == WRITE)
3831 return biovec->bv_len; /* always allow writes to be mergeable */
3833 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3834 chunk_sectors = mddev->new_chunk_sectors;
3835 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3836 if (max < 0) max = 0;
3837 if (max <= biovec->bv_len && bio_sectors == 0)
3838 return biovec->bv_len;
3844 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3846 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3847 unsigned int chunk_sectors = mddev->chunk_sectors;
3848 unsigned int bio_sectors = bio->bi_size >> 9;
3850 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3851 chunk_sectors = mddev->new_chunk_sectors;
3852 return chunk_sectors >=
3853 ((sector & (chunk_sectors - 1)) + bio_sectors);
3857 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3858 * later sampled by raid5d.
3860 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3862 unsigned long flags;
3864 spin_lock_irqsave(&conf->device_lock, flags);
3866 bi->bi_next = conf->retry_read_aligned_list;
3867 conf->retry_read_aligned_list = bi;
3869 spin_unlock_irqrestore(&conf->device_lock, flags);
3870 md_wakeup_thread(conf->mddev->thread);
3874 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3878 bi = conf->retry_read_aligned;
3880 conf->retry_read_aligned = NULL;
3883 bi = conf->retry_read_aligned_list;
3885 conf->retry_read_aligned_list = bi->bi_next;
3888 * this sets the active strip count to 1 and the processed
3889 * strip count to zero (upper 8 bits)
3891 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
3899 * The "raid5_align_endio" should check if the read succeeded and if it
3900 * did, call bio_endio on the original bio (having bio_put the new bio
3902 * If the read failed..
3904 static void raid5_align_endio(struct bio *bi, int error)
3906 struct bio* raid_bi = bi->bi_private;
3907 struct mddev *mddev;
3908 struct r5conf *conf;
3909 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3910 struct md_rdev *rdev;
3914 rdev = (void*)raid_bi->bi_next;
3915 raid_bi->bi_next = NULL;
3916 mddev = rdev->mddev;
3917 conf = mddev->private;
3919 rdev_dec_pending(rdev, conf->mddev);
3921 if (!error && uptodate) {
3922 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
3924 bio_endio(raid_bi, 0);
3925 if (atomic_dec_and_test(&conf->active_aligned_reads))
3926 wake_up(&conf->wait_for_stripe);
3931 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3933 add_bio_to_retry(raid_bi, conf);
3936 static int bio_fits_rdev(struct bio *bi)
3938 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3940 if ((bi->bi_size>>9) > queue_max_sectors(q))
3942 blk_recount_segments(q, bi);
3943 if (bi->bi_phys_segments > queue_max_segments(q))
3946 if (q->merge_bvec_fn)
3947 /* it's too hard to apply the merge_bvec_fn at this stage,
3956 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3958 struct r5conf *conf = mddev->private;
3960 struct bio* align_bi;
3961 struct md_rdev *rdev;
3962 sector_t end_sector;
3964 if (!in_chunk_boundary(mddev, raid_bio)) {
3965 pr_debug("chunk_aligned_read : non aligned\n");
3969 * use bio_clone_mddev to make a copy of the bio
3971 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3975 * set bi_end_io to a new function, and set bi_private to the
3978 align_bi->bi_end_io = raid5_align_endio;
3979 align_bi->bi_private = raid_bio;
3983 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3987 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3989 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3990 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3991 rdev->recovery_offset < end_sector) {
3992 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3994 (test_bit(Faulty, &rdev->flags) ||
3995 !(test_bit(In_sync, &rdev->flags) ||
3996 rdev->recovery_offset >= end_sector)))
4003 atomic_inc(&rdev->nr_pending);
4005 raid_bio->bi_next = (void*)rdev;
4006 align_bi->bi_bdev = rdev->bdev;
4007 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4009 if (!bio_fits_rdev(align_bi) ||
4010 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
4011 &first_bad, &bad_sectors)) {
4012 /* too big in some way, or has a known bad block */
4014 rdev_dec_pending(rdev, mddev);
4018 /* No reshape active, so we can trust rdev->data_offset */
4019 align_bi->bi_sector += rdev->data_offset;
4021 spin_lock_irq(&conf->device_lock);
4022 wait_event_lock_irq(conf->wait_for_stripe,
4024 conf->device_lock, /* nothing */);
4025 atomic_inc(&conf->active_aligned_reads);
4026 spin_unlock_irq(&conf->device_lock);
4028 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4029 align_bi, disk_devt(mddev->gendisk),
4030 raid_bio->bi_sector);
4031 generic_make_request(align_bi);
4040 /* __get_priority_stripe - get the next stripe to process
4042 * Full stripe writes are allowed to pass preread active stripes up until
4043 * the bypass_threshold is exceeded. In general the bypass_count
4044 * increments when the handle_list is handled before the hold_list; however, it
4045 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4046 * stripe with in flight i/o. The bypass_count will be reset when the
4047 * head of the hold_list has changed, i.e. the head was promoted to the
4050 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
4052 struct stripe_head *sh;
4054 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4056 list_empty(&conf->handle_list) ? "empty" : "busy",
4057 list_empty(&conf->hold_list) ? "empty" : "busy",
4058 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4060 if (!list_empty(&conf->handle_list)) {
4061 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4063 if (list_empty(&conf->hold_list))
4064 conf->bypass_count = 0;
4065 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4066 if (conf->hold_list.next == conf->last_hold)
4067 conf->bypass_count++;
4069 conf->last_hold = conf->hold_list.next;
4070 conf->bypass_count -= conf->bypass_threshold;
4071 if (conf->bypass_count < 0)
4072 conf->bypass_count = 0;
4075 } else if (!list_empty(&conf->hold_list) &&
4076 ((conf->bypass_threshold &&
4077 conf->bypass_count > conf->bypass_threshold) ||
4078 atomic_read(&conf->pending_full_writes) == 0)) {
4079 sh = list_entry(conf->hold_list.next,
4081 conf->bypass_count -= conf->bypass_threshold;
4082 if (conf->bypass_count < 0)
4083 conf->bypass_count = 0;
4087 list_del_init(&sh->lru);
4088 atomic_inc(&sh->count);
4089 BUG_ON(atomic_read(&sh->count) != 1);
4093 struct raid5_plug_cb {
4094 struct blk_plug_cb cb;
4095 struct list_head list;
4098 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4100 struct raid5_plug_cb *cb = container_of(
4101 blk_cb, struct raid5_plug_cb, cb);
4102 struct stripe_head *sh;
4103 struct mddev *mddev = cb->cb.data;
4104 struct r5conf *conf = mddev->private;
4107 if (cb->list.next && !list_empty(&cb->list)) {
4108 spin_lock_irq(&conf->device_lock);
4109 while (!list_empty(&cb->list)) {
4110 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4111 list_del_init(&sh->lru);
4113 * avoid race release_stripe_plug() sees
4114 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4115 * is still in our list
4117 smp_mb__before_clear_bit();
4118 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4119 __release_stripe(conf, sh);
4122 spin_unlock_irq(&conf->device_lock);
4124 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4128 static void release_stripe_plug(struct mddev *mddev,
4129 struct stripe_head *sh)
4131 struct blk_plug_cb *blk_cb = blk_check_plugged(
4132 raid5_unplug, mddev,
4133 sizeof(struct raid5_plug_cb));
4134 struct raid5_plug_cb *cb;
4141 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4143 if (cb->list.next == NULL)
4144 INIT_LIST_HEAD(&cb->list);
4146 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4147 list_add_tail(&sh->lru, &cb->list);
4152 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4154 struct r5conf *conf = mddev->private;
4155 sector_t logical_sector, last_sector;
4156 struct stripe_head *sh;
4160 if (mddev->reshape_position != MaxSector)
4161 /* Skip discard while reshape is happening */
4164 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4165 last_sector = bi->bi_sector + (bi->bi_size>>9);
4168 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4170 stripe_sectors = conf->chunk_sectors *
4171 (conf->raid_disks - conf->max_degraded);
4172 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4174 sector_div(last_sector, stripe_sectors);
4176 logical_sector *= conf->chunk_sectors;
4177 last_sector *= conf->chunk_sectors;
4179 for (; logical_sector < last_sector;
4180 logical_sector += STRIPE_SECTORS) {
4184 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4185 prepare_to_wait(&conf->wait_for_overlap, &w,
4186 TASK_UNINTERRUPTIBLE);
4187 spin_lock_irq(&sh->stripe_lock);
4188 for (d = 0; d < conf->raid_disks; d++) {
4189 if (d == sh->pd_idx || d == sh->qd_idx)
4191 if (sh->dev[d].towrite || sh->dev[d].toread) {
4192 set_bit(R5_Overlap, &sh->dev[d].flags);
4193 spin_unlock_irq(&sh->stripe_lock);
4199 finish_wait(&conf->wait_for_overlap, &w);
4200 for (d = 0; d < conf->raid_disks; d++) {
4201 if (d == sh->pd_idx || d == sh->qd_idx)
4203 sh->dev[d].towrite = bi;
4204 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4205 raid5_inc_bi_active_stripes(bi);
4207 spin_unlock_irq(&sh->stripe_lock);
4208 if (conf->mddev->bitmap) {
4210 d < conf->raid_disks - conf->max_degraded;
4212 bitmap_startwrite(mddev->bitmap,
4216 sh->bm_seq = conf->seq_flush + 1;
4217 set_bit(STRIPE_BIT_DELAY, &sh->state);
4220 set_bit(STRIPE_HANDLE, &sh->state);
4221 clear_bit(STRIPE_DELAYED, &sh->state);
4222 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4223 atomic_inc(&conf->preread_active_stripes);
4224 release_stripe_plug(mddev, sh);
4227 remaining = raid5_dec_bi_active_stripes(bi);
4228 if (remaining == 0) {
4229 md_write_end(mddev);
4234 static void make_request(struct mddev *mddev, struct bio * bi)
4236 struct r5conf *conf = mddev->private;
4238 sector_t new_sector;
4239 sector_t logical_sector, last_sector;
4240 struct stripe_head *sh;
4241 const int rw = bio_data_dir(bi);
4244 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4245 md_flush_request(mddev, bi);
4249 md_write_start(mddev, bi);
4252 mddev->reshape_position == MaxSector &&
4253 chunk_aligned_read(mddev,bi))
4256 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4257 make_discard_request(mddev, bi);
4261 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4262 last_sector = bi->bi_sector + (bi->bi_size>>9);
4264 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4266 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4272 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4273 if (unlikely(conf->reshape_progress != MaxSector)) {
4274 /* spinlock is needed as reshape_progress may be
4275 * 64bit on a 32bit platform, and so it might be
4276 * possible to see a half-updated value
4277 * Of course reshape_progress could change after
4278 * the lock is dropped, so once we get a reference
4279 * to the stripe that we think it is, we will have
4282 spin_lock_irq(&conf->device_lock);
4283 if (mddev->reshape_backwards
4284 ? logical_sector < conf->reshape_progress
4285 : logical_sector >= conf->reshape_progress) {
4288 if (mddev->reshape_backwards
4289 ? logical_sector < conf->reshape_safe
4290 : logical_sector >= conf->reshape_safe) {
4291 spin_unlock_irq(&conf->device_lock);
4296 spin_unlock_irq(&conf->device_lock);
4299 new_sector = raid5_compute_sector(conf, logical_sector,
4302 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4303 (unsigned long long)new_sector,
4304 (unsigned long long)logical_sector);
4306 sh = get_active_stripe(conf, new_sector, previous,
4307 (bi->bi_rw&RWA_MASK), 0);
4309 if (unlikely(previous)) {
4310 /* expansion might have moved on while waiting for a
4311 * stripe, so we must do the range check again.
4312 * Expansion could still move past after this
4313 * test, but as we are holding a reference to
4314 * 'sh', we know that if that happens,
4315 * STRIPE_EXPANDING will get set and the expansion
4316 * won't proceed until we finish with the stripe.
4319 spin_lock_irq(&conf->device_lock);
4320 if (mddev->reshape_backwards
4321 ? logical_sector >= conf->reshape_progress
4322 : logical_sector < conf->reshape_progress)
4323 /* mismatch, need to try again */
4325 spin_unlock_irq(&conf->device_lock);
4334 logical_sector >= mddev->suspend_lo &&
4335 logical_sector < mddev->suspend_hi) {
4337 /* As the suspend_* range is controlled by
4338 * userspace, we want an interruptible
4341 flush_signals(current);
4342 prepare_to_wait(&conf->wait_for_overlap,
4343 &w, TASK_INTERRUPTIBLE);
4344 if (logical_sector >= mddev->suspend_lo &&
4345 logical_sector < mddev->suspend_hi)
4350 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4351 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4352 /* Stripe is busy expanding or
4353 * add failed due to overlap. Flush everything
4356 md_wakeup_thread(mddev->thread);
4361 finish_wait(&conf->wait_for_overlap, &w);
4362 set_bit(STRIPE_HANDLE, &sh->state);
4363 clear_bit(STRIPE_DELAYED, &sh->state);
4364 if ((bi->bi_rw & REQ_SYNC) &&
4365 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4366 atomic_inc(&conf->preread_active_stripes);
4367 release_stripe_plug(mddev, sh);
4369 /* cannot get stripe for read-ahead, just give-up */
4370 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4371 finish_wait(&conf->wait_for_overlap, &w);
4376 remaining = raid5_dec_bi_active_stripes(bi);
4377 if (remaining == 0) {
4380 md_write_end(mddev);
4382 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4388 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4390 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4392 /* reshaping is quite different to recovery/resync so it is
4393 * handled quite separately ... here.
4395 * On each call to sync_request, we gather one chunk worth of
4396 * destination stripes and flag them as expanding.
4397 * Then we find all the source stripes and request reads.
4398 * As the reads complete, handle_stripe will copy the data
4399 * into the destination stripe and release that stripe.
4401 struct r5conf *conf = mddev->private;
4402 struct stripe_head *sh;
4403 sector_t first_sector, last_sector;
4404 int raid_disks = conf->previous_raid_disks;
4405 int data_disks = raid_disks - conf->max_degraded;
4406 int new_data_disks = conf->raid_disks - conf->max_degraded;
4409 sector_t writepos, readpos, safepos;
4410 sector_t stripe_addr;
4411 int reshape_sectors;
4412 struct list_head stripes;
4414 if (sector_nr == 0) {
4415 /* If restarting in the middle, skip the initial sectors */
4416 if (mddev->reshape_backwards &&
4417 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4418 sector_nr = raid5_size(mddev, 0, 0)
4419 - conf->reshape_progress;
4420 } else if (!mddev->reshape_backwards &&
4421 conf->reshape_progress > 0)
4422 sector_nr = conf->reshape_progress;
4423 sector_div(sector_nr, new_data_disks);
4425 mddev->curr_resync_completed = sector_nr;
4426 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4432 /* We need to process a full chunk at a time.
4433 * If old and new chunk sizes differ, we need to process the
4436 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4437 reshape_sectors = mddev->new_chunk_sectors;
4439 reshape_sectors = mddev->chunk_sectors;
4441 /* We update the metadata at least every 10 seconds, or when
4442 * the data about to be copied would over-write the source of
4443 * the data at the front of the range. i.e. one new_stripe
4444 * along from reshape_progress new_maps to after where
4445 * reshape_safe old_maps to
4447 writepos = conf->reshape_progress;
4448 sector_div(writepos, new_data_disks);
4449 readpos = conf->reshape_progress;
4450 sector_div(readpos, data_disks);
4451 safepos = conf->reshape_safe;
4452 sector_div(safepos, data_disks);
4453 if (mddev->reshape_backwards) {
4454 writepos -= min_t(sector_t, reshape_sectors, writepos);
4455 readpos += reshape_sectors;
4456 safepos += reshape_sectors;
4458 writepos += reshape_sectors;
4459 readpos -= min_t(sector_t, reshape_sectors, readpos);
4460 safepos -= min_t(sector_t, reshape_sectors, safepos);
4463 /* Having calculated the 'writepos' possibly use it
4464 * to set 'stripe_addr' which is where we will write to.
4466 if (mddev->reshape_backwards) {
4467 BUG_ON(conf->reshape_progress == 0);
4468 stripe_addr = writepos;
4469 BUG_ON((mddev->dev_sectors &
4470 ~((sector_t)reshape_sectors - 1))
4471 - reshape_sectors - stripe_addr
4474 BUG_ON(writepos != sector_nr + reshape_sectors);
4475 stripe_addr = sector_nr;
4478 /* 'writepos' is the most advanced device address we might write.
4479 * 'readpos' is the least advanced device address we might read.
4480 * 'safepos' is the least address recorded in the metadata as having
4482 * If there is a min_offset_diff, these are adjusted either by
4483 * increasing the safepos/readpos if diff is negative, or
4484 * increasing writepos if diff is positive.
4485 * If 'readpos' is then behind 'writepos', there is no way that we can
4486 * ensure safety in the face of a crash - that must be done by userspace
4487 * making a backup of the data. So in that case there is no particular
4488 * rush to update metadata.
4489 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4490 * update the metadata to advance 'safepos' to match 'readpos' so that
4491 * we can be safe in the event of a crash.
4492 * So we insist on updating metadata if safepos is behind writepos and
4493 * readpos is beyond writepos.
4494 * In any case, update the metadata every 10 seconds.
4495 * Maybe that number should be configurable, but I'm not sure it is
4496 * worth it.... maybe it could be a multiple of safemode_delay???
4498 if (conf->min_offset_diff < 0) {
4499 safepos += -conf->min_offset_diff;
4500 readpos += -conf->min_offset_diff;
4502 writepos += conf->min_offset_diff;
4504 if ((mddev->reshape_backwards
4505 ? (safepos > writepos && readpos < writepos)
4506 : (safepos < writepos && readpos > writepos)) ||
4507 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4508 /* Cannot proceed until we've updated the superblock... */
4509 wait_event(conf->wait_for_overlap,
4510 atomic_read(&conf->reshape_stripes)==0);
4511 mddev->reshape_position = conf->reshape_progress;
4512 mddev->curr_resync_completed = sector_nr;
4513 conf->reshape_checkpoint = jiffies;
4514 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4515 md_wakeup_thread(mddev->thread);
4516 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4517 kthread_should_stop());
4518 spin_lock_irq(&conf->device_lock);
4519 conf->reshape_safe = mddev->reshape_position;
4520 spin_unlock_irq(&conf->device_lock);
4521 wake_up(&conf->wait_for_overlap);
4522 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4525 INIT_LIST_HEAD(&stripes);
4526 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4528 int skipped_disk = 0;
4529 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4530 set_bit(STRIPE_EXPANDING, &sh->state);
4531 atomic_inc(&conf->reshape_stripes);
4532 /* If any of this stripe is beyond the end of the old
4533 * array, then we need to zero those blocks
4535 for (j=sh->disks; j--;) {
4537 if (j == sh->pd_idx)
4539 if (conf->level == 6 &&
4542 s = compute_blocknr(sh, j, 0);
4543 if (s < raid5_size(mddev, 0, 0)) {
4547 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4548 set_bit(R5_Expanded, &sh->dev[j].flags);
4549 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4551 if (!skipped_disk) {
4552 set_bit(STRIPE_EXPAND_READY, &sh->state);
4553 set_bit(STRIPE_HANDLE, &sh->state);
4555 list_add(&sh->lru, &stripes);
4557 spin_lock_irq(&conf->device_lock);
4558 if (mddev->reshape_backwards)
4559 conf->reshape_progress -= reshape_sectors * new_data_disks;
4561 conf->reshape_progress += reshape_sectors * new_data_disks;
4562 spin_unlock_irq(&conf->device_lock);
4563 /* Ok, those stripe are ready. We can start scheduling
4564 * reads on the source stripes.
4565 * The source stripes are determined by mapping the first and last
4566 * block on the destination stripes.
4569 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4572 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4573 * new_data_disks - 1),
4575 if (last_sector >= mddev->dev_sectors)
4576 last_sector = mddev->dev_sectors - 1;
4577 while (first_sector <= last_sector) {
4578 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4579 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4580 set_bit(STRIPE_HANDLE, &sh->state);
4582 first_sector += STRIPE_SECTORS;
4584 /* Now that the sources are clearly marked, we can release
4585 * the destination stripes
4587 while (!list_empty(&stripes)) {
4588 sh = list_entry(stripes.next, struct stripe_head, lru);
4589 list_del_init(&sh->lru);
4592 /* If this takes us to the resync_max point where we have to pause,
4593 * then we need to write out the superblock.
4595 sector_nr += reshape_sectors;
4596 if ((sector_nr - mddev->curr_resync_completed) * 2
4597 >= mddev->resync_max - mddev->curr_resync_completed) {
4598 /* Cannot proceed until we've updated the superblock... */
4599 wait_event(conf->wait_for_overlap,
4600 atomic_read(&conf->reshape_stripes) == 0);
4601 mddev->reshape_position = conf->reshape_progress;
4602 mddev->curr_resync_completed = sector_nr;
4603 conf->reshape_checkpoint = jiffies;
4604 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4605 md_wakeup_thread(mddev->thread);
4606 wait_event(mddev->sb_wait,
4607 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4608 || kthread_should_stop());
4609 spin_lock_irq(&conf->device_lock);
4610 conf->reshape_safe = mddev->reshape_position;
4611 spin_unlock_irq(&conf->device_lock);
4612 wake_up(&conf->wait_for_overlap);
4613 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4615 return reshape_sectors;
4618 /* FIXME go_faster isn't used */
4619 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4621 struct r5conf *conf = mddev->private;
4622 struct stripe_head *sh;
4623 sector_t max_sector = mddev->dev_sectors;
4624 sector_t sync_blocks;
4625 int still_degraded = 0;
4628 if (sector_nr >= max_sector) {
4629 /* just being told to finish up .. nothing much to do */
4631 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4636 if (mddev->curr_resync < max_sector) /* aborted */
4637 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4639 else /* completed sync */
4641 bitmap_close_sync(mddev->bitmap);
4646 /* Allow raid5_quiesce to complete */
4647 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4649 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4650 return reshape_request(mddev, sector_nr, skipped);
4652 /* No need to check resync_max as we never do more than one
4653 * stripe, and as resync_max will always be on a chunk boundary,
4654 * if the check in md_do_sync didn't fire, there is no chance
4655 * of overstepping resync_max here
4658 /* if there is too many failed drives and we are trying
4659 * to resync, then assert that we are finished, because there is
4660 * nothing we can do.
4662 if (mddev->degraded >= conf->max_degraded &&
4663 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4664 sector_t rv = mddev->dev_sectors - sector_nr;
4668 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4669 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4670 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4671 /* we can skip this block, and probably more */
4672 sync_blocks /= STRIPE_SECTORS;
4674 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4677 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4679 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4681 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4682 /* make sure we don't swamp the stripe cache if someone else
4683 * is trying to get access
4685 schedule_timeout_uninterruptible(1);
4687 /* Need to check if array will still be degraded after recovery/resync
4688 * We don't need to check the 'failed' flag as when that gets set,
4691 for (i = 0; i < conf->raid_disks; i++)
4692 if (conf->disks[i].rdev == NULL)
4695 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4697 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4702 return STRIPE_SECTORS;
4705 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4707 /* We may not be able to submit a whole bio at once as there
4708 * may not be enough stripe_heads available.
4709 * We cannot pre-allocate enough stripe_heads as we may need
4710 * more than exist in the cache (if we allow ever large chunks).
4711 * So we do one stripe head at a time and record in
4712 * ->bi_hw_segments how many have been done.
4714 * We *know* that this entire raid_bio is in one chunk, so
4715 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4717 struct stripe_head *sh;
4719 sector_t sector, logical_sector, last_sector;
4724 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4725 sector = raid5_compute_sector(conf, logical_sector,
4727 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4729 for (; logical_sector < last_sector;
4730 logical_sector += STRIPE_SECTORS,
4731 sector += STRIPE_SECTORS,
4734 if (scnt < raid5_bi_processed_stripes(raid_bio))
4735 /* already done this stripe */
4738 sh = get_active_stripe(conf, sector, 0, 1, 0);
4741 /* failed to get a stripe - must wait */
4742 raid5_set_bi_processed_stripes(raid_bio, scnt);
4743 conf->retry_read_aligned = raid_bio;
4747 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4749 raid5_set_bi_processed_stripes(raid_bio, scnt);
4750 conf->retry_read_aligned = raid_bio;
4754 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4759 remaining = raid5_dec_bi_active_stripes(raid_bio);
4760 if (remaining == 0) {
4761 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4763 bio_endio(raid_bio, 0);
4765 if (atomic_dec_and_test(&conf->active_aligned_reads))
4766 wake_up(&conf->wait_for_stripe);
4770 #define MAX_STRIPE_BATCH 8
4771 static int handle_active_stripes(struct r5conf *conf)
4773 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4774 int i, batch_size = 0;
4776 while (batch_size < MAX_STRIPE_BATCH &&
4777 (sh = __get_priority_stripe(conf)) != NULL)
4778 batch[batch_size++] = sh;
4780 if (batch_size == 0)
4782 spin_unlock_irq(&conf->device_lock);
4784 for (i = 0; i < batch_size; i++)
4785 handle_stripe(batch[i]);
4789 spin_lock_irq(&conf->device_lock);
4790 for (i = 0; i < batch_size; i++)
4791 __release_stripe(conf, batch[i]);
4796 * This is our raid5 kernel thread.
4798 * We scan the hash table for stripes which can be handled now.
4799 * During the scan, completed stripes are saved for us by the interrupt
4800 * handler, so that they will not have to wait for our next wakeup.
4802 static void raid5d(struct md_thread *thread)
4804 struct mddev *mddev = thread->mddev;
4805 struct r5conf *conf = mddev->private;
4807 struct blk_plug plug;
4809 pr_debug("+++ raid5d active\n");
4811 md_check_recovery(mddev);
4813 blk_start_plug(&plug);
4815 spin_lock_irq(&conf->device_lock);
4821 !list_empty(&conf->bitmap_list)) {
4822 /* Now is a good time to flush some bitmap updates */
4824 spin_unlock_irq(&conf->device_lock);
4825 bitmap_unplug(mddev->bitmap);
4826 spin_lock_irq(&conf->device_lock);
4827 conf->seq_write = conf->seq_flush;
4828 activate_bit_delay(conf);
4830 raid5_activate_delayed(conf);
4832 while ((bio = remove_bio_from_retry(conf))) {
4834 spin_unlock_irq(&conf->device_lock);
4835 ok = retry_aligned_read(conf, bio);
4836 spin_lock_irq(&conf->device_lock);
4842 batch_size = handle_active_stripes(conf);
4845 handled += batch_size;
4847 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
4848 spin_unlock_irq(&conf->device_lock);
4849 md_check_recovery(mddev);
4850 spin_lock_irq(&conf->device_lock);
4853 pr_debug("%d stripes handled\n", handled);
4855 spin_unlock_irq(&conf->device_lock);
4857 async_tx_issue_pending_all();
4858 blk_finish_plug(&plug);
4860 pr_debug("--- raid5d inactive\n");
4864 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4866 struct r5conf *conf = mddev->private;
4868 return sprintf(page, "%d\n", conf->max_nr_stripes);
4874 raid5_set_cache_size(struct mddev *mddev, int size)
4876 struct r5conf *conf = mddev->private;
4879 if (size <= 16 || size > 32768)
4881 while (size < conf->max_nr_stripes) {
4882 if (drop_one_stripe(conf))
4883 conf->max_nr_stripes--;
4887 err = md_allow_write(mddev);
4890 while (size > conf->max_nr_stripes) {
4891 if (grow_one_stripe(conf))
4892 conf->max_nr_stripes++;
4897 EXPORT_SYMBOL(raid5_set_cache_size);
4900 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4902 struct r5conf *conf = mddev->private;
4906 if (len >= PAGE_SIZE)
4911 if (strict_strtoul(page, 10, &new))
4913 err = raid5_set_cache_size(mddev, new);
4919 static struct md_sysfs_entry
4920 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4921 raid5_show_stripe_cache_size,
4922 raid5_store_stripe_cache_size);
4925 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4927 struct r5conf *conf = mddev->private;
4929 return sprintf(page, "%d\n", conf->bypass_threshold);
4935 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4937 struct r5conf *conf = mddev->private;
4939 if (len >= PAGE_SIZE)
4944 if (strict_strtoul(page, 10, &new))
4946 if (new > conf->max_nr_stripes)
4948 conf->bypass_threshold = new;
4952 static struct md_sysfs_entry
4953 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4955 raid5_show_preread_threshold,
4956 raid5_store_preread_threshold);
4959 stripe_cache_active_show(struct mddev *mddev, char *page)
4961 struct r5conf *conf = mddev->private;
4963 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4968 static struct md_sysfs_entry
4969 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4971 static struct attribute *raid5_attrs[] = {
4972 &raid5_stripecache_size.attr,
4973 &raid5_stripecache_active.attr,
4974 &raid5_preread_bypass_threshold.attr,
4977 static struct attribute_group raid5_attrs_group = {
4979 .attrs = raid5_attrs,
4983 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4985 struct r5conf *conf = mddev->private;
4988 sectors = mddev->dev_sectors;
4990 /* size is defined by the smallest of previous and new size */
4991 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4993 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4994 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4995 return sectors * (raid_disks - conf->max_degraded);
4998 static void raid5_free_percpu(struct r5conf *conf)
5000 struct raid5_percpu *percpu;
5007 for_each_possible_cpu(cpu) {
5008 percpu = per_cpu_ptr(conf->percpu, cpu);
5009 safe_put_page(percpu->spare_page);
5010 kfree(percpu->scribble);
5012 #ifdef CONFIG_HOTPLUG_CPU
5013 unregister_cpu_notifier(&conf->cpu_notify);
5017 free_percpu(conf->percpu);
5020 static void free_conf(struct r5conf *conf)
5022 shrink_stripes(conf);
5023 raid5_free_percpu(conf);
5025 kfree(conf->stripe_hashtbl);
5029 #ifdef CONFIG_HOTPLUG_CPU
5030 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5033 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5034 long cpu = (long)hcpu;
5035 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5038 case CPU_UP_PREPARE:
5039 case CPU_UP_PREPARE_FROZEN:
5040 if (conf->level == 6 && !percpu->spare_page)
5041 percpu->spare_page = alloc_page(GFP_KERNEL);
5042 if (!percpu->scribble)
5043 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5045 if (!percpu->scribble ||
5046 (conf->level == 6 && !percpu->spare_page)) {
5047 safe_put_page(percpu->spare_page);
5048 kfree(percpu->scribble);
5049 pr_err("%s: failed memory allocation for cpu%ld\n",
5051 return notifier_from_errno(-ENOMEM);
5055 case CPU_DEAD_FROZEN:
5056 safe_put_page(percpu->spare_page);
5057 kfree(percpu->scribble);
5058 percpu->spare_page = NULL;
5059 percpu->scribble = NULL;
5068 static int raid5_alloc_percpu(struct r5conf *conf)
5071 struct page *spare_page;
5072 struct raid5_percpu __percpu *allcpus;
5076 allcpus = alloc_percpu(struct raid5_percpu);
5079 conf->percpu = allcpus;
5083 for_each_present_cpu(cpu) {
5084 if (conf->level == 6) {
5085 spare_page = alloc_page(GFP_KERNEL);
5090 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5092 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5097 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5099 #ifdef CONFIG_HOTPLUG_CPU
5100 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5101 conf->cpu_notify.priority = 0;
5103 err = register_cpu_notifier(&conf->cpu_notify);
5110 static struct r5conf *setup_conf(struct mddev *mddev)
5112 struct r5conf *conf;
5113 int raid_disk, memory, max_disks;
5114 struct md_rdev *rdev;
5115 struct disk_info *disk;
5118 if (mddev->new_level != 5
5119 && mddev->new_level != 4
5120 && mddev->new_level != 6) {
5121 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5122 mdname(mddev), mddev->new_level);
5123 return ERR_PTR(-EIO);
5125 if ((mddev->new_level == 5
5126 && !algorithm_valid_raid5(mddev->new_layout)) ||
5127 (mddev->new_level == 6
5128 && !algorithm_valid_raid6(mddev->new_layout))) {
5129 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5130 mdname(mddev), mddev->new_layout);
5131 return ERR_PTR(-EIO);
5133 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5134 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5135 mdname(mddev), mddev->raid_disks);
5136 return ERR_PTR(-EINVAL);
5139 if (!mddev->new_chunk_sectors ||
5140 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5141 !is_power_of_2(mddev->new_chunk_sectors)) {
5142 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5143 mdname(mddev), mddev->new_chunk_sectors << 9);
5144 return ERR_PTR(-EINVAL);
5147 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5150 spin_lock_init(&conf->device_lock);
5151 init_waitqueue_head(&conf->wait_for_stripe);
5152 init_waitqueue_head(&conf->wait_for_overlap);
5153 INIT_LIST_HEAD(&conf->handle_list);
5154 INIT_LIST_HEAD(&conf->hold_list);
5155 INIT_LIST_HEAD(&conf->delayed_list);
5156 INIT_LIST_HEAD(&conf->bitmap_list);
5157 INIT_LIST_HEAD(&conf->inactive_list);
5158 atomic_set(&conf->active_stripes, 0);
5159 atomic_set(&conf->preread_active_stripes, 0);
5160 atomic_set(&conf->active_aligned_reads, 0);
5161 conf->bypass_threshold = BYPASS_THRESHOLD;
5162 conf->recovery_disabled = mddev->recovery_disabled - 1;
5164 conf->raid_disks = mddev->raid_disks;
5165 if (mddev->reshape_position == MaxSector)
5166 conf->previous_raid_disks = mddev->raid_disks;
5168 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5169 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5170 conf->scribble_len = scribble_len(max_disks);
5172 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5177 conf->mddev = mddev;
5179 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5182 conf->level = mddev->new_level;
5183 if (raid5_alloc_percpu(conf) != 0)
5186 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5188 rdev_for_each(rdev, mddev) {
5189 raid_disk = rdev->raid_disk;
5190 if (raid_disk >= max_disks
5193 disk = conf->disks + raid_disk;
5195 if (test_bit(Replacement, &rdev->flags)) {
5196 if (disk->replacement)
5198 disk->replacement = rdev;
5205 if (test_bit(In_sync, &rdev->flags)) {
5206 char b[BDEVNAME_SIZE];
5207 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5209 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5210 } else if (rdev->saved_raid_disk != raid_disk)
5211 /* Cannot rely on bitmap to complete recovery */
5215 conf->chunk_sectors = mddev->new_chunk_sectors;
5216 conf->level = mddev->new_level;
5217 if (conf->level == 6)
5218 conf->max_degraded = 2;
5220 conf->max_degraded = 1;
5221 conf->algorithm = mddev->new_layout;
5222 conf->max_nr_stripes = NR_STRIPES;
5223 conf->reshape_progress = mddev->reshape_position;
5224 if (conf->reshape_progress != MaxSector) {
5225 conf->prev_chunk_sectors = mddev->chunk_sectors;
5226 conf->prev_algo = mddev->layout;
5229 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5230 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5231 if (grow_stripes(conf, conf->max_nr_stripes)) {
5233 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5234 mdname(mddev), memory);
5237 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5238 mdname(mddev), memory);
5240 sprintf(pers_name, "raid%d", mddev->new_level);
5241 conf->thread = md_register_thread(raid5d, mddev, pers_name);
5242 if (!conf->thread) {
5244 "md/raid:%s: couldn't allocate thread.\n",
5254 return ERR_PTR(-EIO);
5256 return ERR_PTR(-ENOMEM);
5260 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5263 case ALGORITHM_PARITY_0:
5264 if (raid_disk < max_degraded)
5267 case ALGORITHM_PARITY_N:
5268 if (raid_disk >= raid_disks - max_degraded)
5271 case ALGORITHM_PARITY_0_6:
5272 if (raid_disk == 0 ||
5273 raid_disk == raid_disks - 1)
5276 case ALGORITHM_LEFT_ASYMMETRIC_6:
5277 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5278 case ALGORITHM_LEFT_SYMMETRIC_6:
5279 case ALGORITHM_RIGHT_SYMMETRIC_6:
5280 if (raid_disk == raid_disks - 1)
5286 static int run(struct mddev *mddev)
5288 struct r5conf *conf;
5289 int working_disks = 0;
5290 int dirty_parity_disks = 0;
5291 struct md_rdev *rdev;
5292 sector_t reshape_offset = 0;
5294 long long min_offset_diff = 0;
5297 if (mddev->recovery_cp != MaxSector)
5298 printk(KERN_NOTICE "md/raid:%s: not clean"
5299 " -- starting background reconstruction\n",
5302 rdev_for_each(rdev, mddev) {
5304 if (rdev->raid_disk < 0)
5306 diff = (rdev->new_data_offset - rdev->data_offset);
5308 min_offset_diff = diff;
5310 } else if (mddev->reshape_backwards &&
5311 diff < min_offset_diff)
5312 min_offset_diff = diff;
5313 else if (!mddev->reshape_backwards &&
5314 diff > min_offset_diff)
5315 min_offset_diff = diff;
5318 if (mddev->reshape_position != MaxSector) {
5319 /* Check that we can continue the reshape.
5320 * Difficulties arise if the stripe we would write to
5321 * next is at or after the stripe we would read from next.
5322 * For a reshape that changes the number of devices, this
5323 * is only possible for a very short time, and mdadm makes
5324 * sure that time appears to have past before assembling
5325 * the array. So we fail if that time hasn't passed.
5326 * For a reshape that keeps the number of devices the same
5327 * mdadm must be monitoring the reshape can keeping the
5328 * critical areas read-only and backed up. It will start
5329 * the array in read-only mode, so we check for that.
5331 sector_t here_new, here_old;
5333 int max_degraded = (mddev->level == 6 ? 2 : 1);
5335 if (mddev->new_level != mddev->level) {
5336 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5337 "required - aborting.\n",
5341 old_disks = mddev->raid_disks - mddev->delta_disks;
5342 /* reshape_position must be on a new-stripe boundary, and one
5343 * further up in new geometry must map after here in old
5346 here_new = mddev->reshape_position;
5347 if (sector_div(here_new, mddev->new_chunk_sectors *
5348 (mddev->raid_disks - max_degraded))) {
5349 printk(KERN_ERR "md/raid:%s: reshape_position not "
5350 "on a stripe boundary\n", mdname(mddev));
5353 reshape_offset = here_new * mddev->new_chunk_sectors;
5354 /* here_new is the stripe we will write to */
5355 here_old = mddev->reshape_position;
5356 sector_div(here_old, mddev->chunk_sectors *
5357 (old_disks-max_degraded));
5358 /* here_old is the first stripe that we might need to read
5360 if (mddev->delta_disks == 0) {
5361 if ((here_new * mddev->new_chunk_sectors !=
5362 here_old * mddev->chunk_sectors)) {
5363 printk(KERN_ERR "md/raid:%s: reshape position is"
5364 " confused - aborting\n", mdname(mddev));
5367 /* We cannot be sure it is safe to start an in-place
5368 * reshape. It is only safe if user-space is monitoring
5369 * and taking constant backups.
5370 * mdadm always starts a situation like this in
5371 * readonly mode so it can take control before
5372 * allowing any writes. So just check for that.
5374 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5375 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5376 /* not really in-place - so OK */;
5377 else if (mddev->ro == 0) {
5378 printk(KERN_ERR "md/raid:%s: in-place reshape "
5379 "must be started in read-only mode "
5384 } else if (mddev->reshape_backwards
5385 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5386 here_old * mddev->chunk_sectors)
5387 : (here_new * mddev->new_chunk_sectors >=
5388 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5389 /* Reading from the same stripe as writing to - bad */
5390 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5391 "auto-recovery - aborting.\n",
5395 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5397 /* OK, we should be able to continue; */
5399 BUG_ON(mddev->level != mddev->new_level);
5400 BUG_ON(mddev->layout != mddev->new_layout);
5401 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5402 BUG_ON(mddev->delta_disks != 0);
5405 if (mddev->private == NULL)
5406 conf = setup_conf(mddev);
5408 conf = mddev->private;
5411 return PTR_ERR(conf);
5413 conf->min_offset_diff = min_offset_diff;
5414 mddev->thread = conf->thread;
5415 conf->thread = NULL;
5416 mddev->private = conf;
5418 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5420 rdev = conf->disks[i].rdev;
5421 if (!rdev && conf->disks[i].replacement) {
5422 /* The replacement is all we have yet */
5423 rdev = conf->disks[i].replacement;
5424 conf->disks[i].replacement = NULL;
5425 clear_bit(Replacement, &rdev->flags);
5426 conf->disks[i].rdev = rdev;
5430 if (conf->disks[i].replacement &&
5431 conf->reshape_progress != MaxSector) {
5432 /* replacements and reshape simply do not mix. */
5433 printk(KERN_ERR "md: cannot handle concurrent "
5434 "replacement and reshape.\n");
5437 if (test_bit(In_sync, &rdev->flags)) {
5441 /* This disc is not fully in-sync. However if it
5442 * just stored parity (beyond the recovery_offset),
5443 * when we don't need to be concerned about the
5444 * array being dirty.
5445 * When reshape goes 'backwards', we never have
5446 * partially completed devices, so we only need
5447 * to worry about reshape going forwards.
5449 /* Hack because v0.91 doesn't store recovery_offset properly. */
5450 if (mddev->major_version == 0 &&
5451 mddev->minor_version > 90)
5452 rdev->recovery_offset = reshape_offset;
5454 if (rdev->recovery_offset < reshape_offset) {
5455 /* We need to check old and new layout */
5456 if (!only_parity(rdev->raid_disk,
5459 conf->max_degraded))
5462 if (!only_parity(rdev->raid_disk,
5464 conf->previous_raid_disks,
5465 conf->max_degraded))
5467 dirty_parity_disks++;
5471 * 0 for a fully functional array, 1 or 2 for a degraded array.
5473 mddev->degraded = calc_degraded(conf);
5475 if (has_failed(conf)) {
5476 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5477 " (%d/%d failed)\n",
5478 mdname(mddev), mddev->degraded, conf->raid_disks);
5482 /* device size must be a multiple of chunk size */
5483 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5484 mddev->resync_max_sectors = mddev->dev_sectors;
5486 if (mddev->degraded > dirty_parity_disks &&
5487 mddev->recovery_cp != MaxSector) {
5488 if (mddev->ok_start_degraded)
5490 "md/raid:%s: starting dirty degraded array"
5491 " - data corruption possible.\n",
5495 "md/raid:%s: cannot start dirty degraded array.\n",
5501 if (mddev->degraded == 0)
5502 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5503 " devices, algorithm %d\n", mdname(mddev), conf->level,
5504 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5507 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5508 " out of %d devices, algorithm %d\n",
5509 mdname(mddev), conf->level,
5510 mddev->raid_disks - mddev->degraded,
5511 mddev->raid_disks, mddev->new_layout);
5513 print_raid5_conf(conf);
5515 if (conf->reshape_progress != MaxSector) {
5516 conf->reshape_safe = conf->reshape_progress;
5517 atomic_set(&conf->reshape_stripes, 0);
5518 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5519 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5520 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5521 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5522 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5527 /* Ok, everything is just fine now */
5528 if (mddev->to_remove == &raid5_attrs_group)
5529 mddev->to_remove = NULL;
5530 else if (mddev->kobj.sd &&
5531 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5533 "raid5: failed to create sysfs attributes for %s\n",
5535 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5539 bool discard_supported = true;
5540 /* read-ahead size must cover two whole stripes, which
5541 * is 2 * (datadisks) * chunksize where 'n' is the
5542 * number of raid devices
5544 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5545 int stripe = data_disks *
5546 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5547 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5548 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5550 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5552 mddev->queue->backing_dev_info.congested_data = mddev;
5553 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5555 chunk_size = mddev->chunk_sectors << 9;
5556 blk_queue_io_min(mddev->queue, chunk_size);
5557 blk_queue_io_opt(mddev->queue, chunk_size *
5558 (conf->raid_disks - conf->max_degraded));
5560 * We can only discard a whole stripe. It doesn't make sense to
5561 * discard data disk but write parity disk
5563 stripe = stripe * PAGE_SIZE;
5564 /* Round up to power of 2, as discard handling
5565 * currently assumes that */
5566 while ((stripe-1) & stripe)
5567 stripe = (stripe | (stripe-1)) + 1;
5568 mddev->queue->limits.discard_alignment = stripe;
5569 mddev->queue->limits.discard_granularity = stripe;
5571 * unaligned part of discard request will be ignored, so can't
5572 * guarantee discard_zerors_data
5574 mddev->queue->limits.discard_zeroes_data = 0;
5576 rdev_for_each(rdev, mddev) {
5577 disk_stack_limits(mddev->gendisk, rdev->bdev,
5578 rdev->data_offset << 9);
5579 disk_stack_limits(mddev->gendisk, rdev->bdev,
5580 rdev->new_data_offset << 9);
5582 * discard_zeroes_data is required, otherwise data
5583 * could be lost. Consider a scenario: discard a stripe
5584 * (the stripe could be inconsistent if
5585 * discard_zeroes_data is 0); write one disk of the
5586 * stripe (the stripe could be inconsistent again
5587 * depending on which disks are used to calculate
5588 * parity); the disk is broken; The stripe data of this
5591 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5592 !bdev_get_queue(rdev->bdev)->
5593 limits.discard_zeroes_data)
5594 discard_supported = false;
5597 if (discard_supported &&
5598 mddev->queue->limits.max_discard_sectors >= stripe &&
5599 mddev->queue->limits.discard_granularity >= stripe)
5600 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5603 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5609 md_unregister_thread(&mddev->thread);
5610 print_raid5_conf(conf);
5612 mddev->private = NULL;
5613 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5617 static int stop(struct mddev *mddev)
5619 struct r5conf *conf = mddev->private;
5621 md_unregister_thread(&mddev->thread);
5623 mddev->queue->backing_dev_info.congested_fn = NULL;
5625 mddev->private = NULL;
5626 mddev->to_remove = &raid5_attrs_group;
5630 static void status(struct seq_file *seq, struct mddev *mddev)
5632 struct r5conf *conf = mddev->private;
5635 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5636 mddev->chunk_sectors / 2, mddev->layout);
5637 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5638 for (i = 0; i < conf->raid_disks; i++)
5639 seq_printf (seq, "%s",
5640 conf->disks[i].rdev &&
5641 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5642 seq_printf (seq, "]");
5645 static void print_raid5_conf (struct r5conf *conf)
5648 struct disk_info *tmp;
5650 printk(KERN_DEBUG "RAID conf printout:\n");
5652 printk("(conf==NULL)\n");
5655 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5657 conf->raid_disks - conf->mddev->degraded);
5659 for (i = 0; i < conf->raid_disks; i++) {
5660 char b[BDEVNAME_SIZE];
5661 tmp = conf->disks + i;
5663 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5664 i, !test_bit(Faulty, &tmp->rdev->flags),
5665 bdevname(tmp->rdev->bdev, b));
5669 static int raid5_spare_active(struct mddev *mddev)
5672 struct r5conf *conf = mddev->private;
5673 struct disk_info *tmp;
5675 unsigned long flags;
5677 for (i = 0; i < conf->raid_disks; i++) {
5678 tmp = conf->disks + i;
5679 if (tmp->replacement
5680 && tmp->replacement->recovery_offset == MaxSector
5681 && !test_bit(Faulty, &tmp->replacement->flags)
5682 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5683 /* Replacement has just become active. */
5685 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5688 /* Replaced device not technically faulty,
5689 * but we need to be sure it gets removed
5690 * and never re-added.
5692 set_bit(Faulty, &tmp->rdev->flags);
5693 sysfs_notify_dirent_safe(
5694 tmp->rdev->sysfs_state);
5696 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5697 } else if (tmp->rdev
5698 && tmp->rdev->recovery_offset == MaxSector
5699 && !test_bit(Faulty, &tmp->rdev->flags)
5700 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5702 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5705 spin_lock_irqsave(&conf->device_lock, flags);
5706 mddev->degraded = calc_degraded(conf);
5707 spin_unlock_irqrestore(&conf->device_lock, flags);
5708 print_raid5_conf(conf);
5712 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5714 struct r5conf *conf = mddev->private;
5716 int number = rdev->raid_disk;
5717 struct md_rdev **rdevp;
5718 struct disk_info *p = conf->disks + number;
5720 print_raid5_conf(conf);
5721 if (rdev == p->rdev)
5723 else if (rdev == p->replacement)
5724 rdevp = &p->replacement;
5728 if (number >= conf->raid_disks &&
5729 conf->reshape_progress == MaxSector)
5730 clear_bit(In_sync, &rdev->flags);
5732 if (test_bit(In_sync, &rdev->flags) ||
5733 atomic_read(&rdev->nr_pending)) {
5737 /* Only remove non-faulty devices if recovery
5740 if (!test_bit(Faulty, &rdev->flags) &&
5741 mddev->recovery_disabled != conf->recovery_disabled &&
5742 !has_failed(conf) &&
5743 (!p->replacement || p->replacement == rdev) &&
5744 number < conf->raid_disks) {
5750 if (atomic_read(&rdev->nr_pending)) {
5751 /* lost the race, try later */
5754 } else if (p->replacement) {
5755 /* We must have just cleared 'rdev' */
5756 p->rdev = p->replacement;
5757 clear_bit(Replacement, &p->replacement->flags);
5758 smp_mb(); /* Make sure other CPUs may see both as identical
5759 * but will never see neither - if they are careful
5761 p->replacement = NULL;
5762 clear_bit(WantReplacement, &rdev->flags);
5764 /* We might have just removed the Replacement as faulty-
5765 * clear the bit just in case
5767 clear_bit(WantReplacement, &rdev->flags);
5770 print_raid5_conf(conf);
5774 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5776 struct r5conf *conf = mddev->private;
5779 struct disk_info *p;
5781 int last = conf->raid_disks - 1;
5783 if (mddev->recovery_disabled == conf->recovery_disabled)
5786 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5787 /* no point adding a device */
5790 if (rdev->raid_disk >= 0)
5791 first = last = rdev->raid_disk;
5794 * find the disk ... but prefer rdev->saved_raid_disk
5797 if (rdev->saved_raid_disk >= 0 &&
5798 rdev->saved_raid_disk >= first &&
5799 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5800 first = rdev->saved_raid_disk;
5802 for (disk = first; disk <= last; disk++) {
5803 p = conf->disks + disk;
5804 if (p->rdev == NULL) {
5805 clear_bit(In_sync, &rdev->flags);
5806 rdev->raid_disk = disk;
5808 if (rdev->saved_raid_disk != disk)
5810 rcu_assign_pointer(p->rdev, rdev);
5814 for (disk = first; disk <= last; disk++) {
5815 p = conf->disks + disk;
5816 if (test_bit(WantReplacement, &p->rdev->flags) &&
5817 p->replacement == NULL) {
5818 clear_bit(In_sync, &rdev->flags);
5819 set_bit(Replacement, &rdev->flags);
5820 rdev->raid_disk = disk;
5823 rcu_assign_pointer(p->replacement, rdev);
5828 print_raid5_conf(conf);
5832 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5834 /* no resync is happening, and there is enough space
5835 * on all devices, so we can resize.
5836 * We need to make sure resync covers any new space.
5837 * If the array is shrinking we should possibly wait until
5838 * any io in the removed space completes, but it hardly seems
5842 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5843 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5844 if (mddev->external_size &&
5845 mddev->array_sectors > newsize)
5847 if (mddev->bitmap) {
5848 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5852 md_set_array_sectors(mddev, newsize);
5853 set_capacity(mddev->gendisk, mddev->array_sectors);
5854 revalidate_disk(mddev->gendisk);
5855 if (sectors > mddev->dev_sectors &&
5856 mddev->recovery_cp > mddev->dev_sectors) {
5857 mddev->recovery_cp = mddev->dev_sectors;
5858 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5860 mddev->dev_sectors = sectors;
5861 mddev->resync_max_sectors = sectors;
5865 static int check_stripe_cache(struct mddev *mddev)
5867 /* Can only proceed if there are plenty of stripe_heads.
5868 * We need a minimum of one full stripe,, and for sensible progress
5869 * it is best to have about 4 times that.
5870 * If we require 4 times, then the default 256 4K stripe_heads will
5871 * allow for chunk sizes up to 256K, which is probably OK.
5872 * If the chunk size is greater, user-space should request more
5873 * stripe_heads first.
5875 struct r5conf *conf = mddev->private;
5876 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5877 > conf->max_nr_stripes ||
5878 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5879 > conf->max_nr_stripes) {
5880 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5882 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5889 static int check_reshape(struct mddev *mddev)
5891 struct r5conf *conf = mddev->private;
5893 if (mddev->delta_disks == 0 &&
5894 mddev->new_layout == mddev->layout &&
5895 mddev->new_chunk_sectors == mddev->chunk_sectors)
5896 return 0; /* nothing to do */
5897 if (has_failed(conf))
5899 if (mddev->delta_disks < 0) {
5900 /* We might be able to shrink, but the devices must
5901 * be made bigger first.
5902 * For raid6, 4 is the minimum size.
5903 * Otherwise 2 is the minimum
5906 if (mddev->level == 6)
5908 if (mddev->raid_disks + mddev->delta_disks < min)
5912 if (!check_stripe_cache(mddev))
5915 return resize_stripes(conf, (conf->previous_raid_disks
5916 + mddev->delta_disks));
5919 static int raid5_start_reshape(struct mddev *mddev)
5921 struct r5conf *conf = mddev->private;
5922 struct md_rdev *rdev;
5924 unsigned long flags;
5926 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5929 if (!check_stripe_cache(mddev))
5932 if (has_failed(conf))
5935 rdev_for_each(rdev, mddev) {
5936 if (!test_bit(In_sync, &rdev->flags)
5937 && !test_bit(Faulty, &rdev->flags))
5941 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5942 /* Not enough devices even to make a degraded array
5947 /* Refuse to reduce size of the array. Any reductions in
5948 * array size must be through explicit setting of array_size
5951 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5952 < mddev->array_sectors) {
5953 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5954 "before number of disks\n", mdname(mddev));
5958 atomic_set(&conf->reshape_stripes, 0);
5959 spin_lock_irq(&conf->device_lock);
5960 conf->previous_raid_disks = conf->raid_disks;
5961 conf->raid_disks += mddev->delta_disks;
5962 conf->prev_chunk_sectors = conf->chunk_sectors;
5963 conf->chunk_sectors = mddev->new_chunk_sectors;
5964 conf->prev_algo = conf->algorithm;
5965 conf->algorithm = mddev->new_layout;
5967 /* Code that selects data_offset needs to see the generation update
5968 * if reshape_progress has been set - so a memory barrier needed.
5971 if (mddev->reshape_backwards)
5972 conf->reshape_progress = raid5_size(mddev, 0, 0);
5974 conf->reshape_progress = 0;
5975 conf->reshape_safe = conf->reshape_progress;
5976 spin_unlock_irq(&conf->device_lock);
5978 /* Add some new drives, as many as will fit.
5979 * We know there are enough to make the newly sized array work.
5980 * Don't add devices if we are reducing the number of
5981 * devices in the array. This is because it is not possible
5982 * to correctly record the "partially reconstructed" state of
5983 * such devices during the reshape and confusion could result.
5985 if (mddev->delta_disks >= 0) {
5986 rdev_for_each(rdev, mddev)
5987 if (rdev->raid_disk < 0 &&
5988 !test_bit(Faulty, &rdev->flags)) {
5989 if (raid5_add_disk(mddev, rdev) == 0) {
5991 >= conf->previous_raid_disks)
5992 set_bit(In_sync, &rdev->flags);
5994 rdev->recovery_offset = 0;
5996 if (sysfs_link_rdev(mddev, rdev))
5997 /* Failure here is OK */;
5999 } else if (rdev->raid_disk >= conf->previous_raid_disks
6000 && !test_bit(Faulty, &rdev->flags)) {
6001 /* This is a spare that was manually added */
6002 set_bit(In_sync, &rdev->flags);
6005 /* When a reshape changes the number of devices,
6006 * ->degraded is measured against the larger of the
6007 * pre and post number of devices.
6009 spin_lock_irqsave(&conf->device_lock, flags);
6010 mddev->degraded = calc_degraded(conf);
6011 spin_unlock_irqrestore(&conf->device_lock, flags);
6013 mddev->raid_disks = conf->raid_disks;
6014 mddev->reshape_position = conf->reshape_progress;
6015 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6017 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6018 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6019 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6020 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6021 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6023 if (!mddev->sync_thread) {
6024 mddev->recovery = 0;
6025 spin_lock_irq(&conf->device_lock);
6026 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6027 rdev_for_each(rdev, mddev)
6028 rdev->new_data_offset = rdev->data_offset;
6030 conf->reshape_progress = MaxSector;
6031 mddev->reshape_position = MaxSector;
6032 spin_unlock_irq(&conf->device_lock);
6035 conf->reshape_checkpoint = jiffies;
6036 md_wakeup_thread(mddev->sync_thread);
6037 md_new_event(mddev);
6041 /* This is called from the reshape thread and should make any
6042 * changes needed in 'conf'
6044 static void end_reshape(struct r5conf *conf)
6047 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6048 struct md_rdev *rdev;
6050 spin_lock_irq(&conf->device_lock);
6051 conf->previous_raid_disks = conf->raid_disks;
6052 rdev_for_each(rdev, conf->mddev)
6053 rdev->data_offset = rdev->new_data_offset;
6055 conf->reshape_progress = MaxSector;
6056 spin_unlock_irq(&conf->device_lock);
6057 wake_up(&conf->wait_for_overlap);
6059 /* read-ahead size must cover two whole stripes, which is
6060 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6062 if (conf->mddev->queue) {
6063 int data_disks = conf->raid_disks - conf->max_degraded;
6064 int stripe = data_disks * ((conf->chunk_sectors << 9)
6066 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6067 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6072 /* This is called from the raid5d thread with mddev_lock held.
6073 * It makes config changes to the device.
6075 static void raid5_finish_reshape(struct mddev *mddev)
6077 struct r5conf *conf = mddev->private;
6079 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6081 if (mddev->delta_disks > 0) {
6082 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6083 set_capacity(mddev->gendisk, mddev->array_sectors);
6084 revalidate_disk(mddev->gendisk);
6087 spin_lock_irq(&conf->device_lock);
6088 mddev->degraded = calc_degraded(conf);
6089 spin_unlock_irq(&conf->device_lock);
6090 for (d = conf->raid_disks ;
6091 d < conf->raid_disks - mddev->delta_disks;
6093 struct md_rdev *rdev = conf->disks[d].rdev;
6095 clear_bit(In_sync, &rdev->flags);
6096 rdev = conf->disks[d].replacement;
6098 clear_bit(In_sync, &rdev->flags);
6101 mddev->layout = conf->algorithm;
6102 mddev->chunk_sectors = conf->chunk_sectors;
6103 mddev->reshape_position = MaxSector;
6104 mddev->delta_disks = 0;
6105 mddev->reshape_backwards = 0;
6109 static void raid5_quiesce(struct mddev *mddev, int state)
6111 struct r5conf *conf = mddev->private;
6114 case 2: /* resume for a suspend */
6115 wake_up(&conf->wait_for_overlap);
6118 case 1: /* stop all writes */
6119 spin_lock_irq(&conf->device_lock);
6120 /* '2' tells resync/reshape to pause so that all
6121 * active stripes can drain
6124 wait_event_lock_irq(conf->wait_for_stripe,
6125 atomic_read(&conf->active_stripes) == 0 &&
6126 atomic_read(&conf->active_aligned_reads) == 0,
6127 conf->device_lock, /* nothing */);
6129 spin_unlock_irq(&conf->device_lock);
6130 /* allow reshape to continue */
6131 wake_up(&conf->wait_for_overlap);
6134 case 0: /* re-enable writes */
6135 spin_lock_irq(&conf->device_lock);
6137 wake_up(&conf->wait_for_stripe);
6138 wake_up(&conf->wait_for_overlap);
6139 spin_unlock_irq(&conf->device_lock);
6145 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6147 struct r0conf *raid0_conf = mddev->private;
6150 /* for raid0 takeover only one zone is supported */
6151 if (raid0_conf->nr_strip_zones > 1) {
6152 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6154 return ERR_PTR(-EINVAL);
6157 sectors = raid0_conf->strip_zone[0].zone_end;
6158 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6159 mddev->dev_sectors = sectors;
6160 mddev->new_level = level;
6161 mddev->new_layout = ALGORITHM_PARITY_N;
6162 mddev->new_chunk_sectors = mddev->chunk_sectors;
6163 mddev->raid_disks += 1;
6164 mddev->delta_disks = 1;
6165 /* make sure it will be not marked as dirty */
6166 mddev->recovery_cp = MaxSector;
6168 return setup_conf(mddev);
6172 static void *raid5_takeover_raid1(struct mddev *mddev)
6176 if (mddev->raid_disks != 2 ||
6177 mddev->degraded > 1)
6178 return ERR_PTR(-EINVAL);
6180 /* Should check if there are write-behind devices? */
6182 chunksect = 64*2; /* 64K by default */
6184 /* The array must be an exact multiple of chunksize */
6185 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6188 if ((chunksect<<9) < STRIPE_SIZE)
6189 /* array size does not allow a suitable chunk size */
6190 return ERR_PTR(-EINVAL);
6192 mddev->new_level = 5;
6193 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6194 mddev->new_chunk_sectors = chunksect;
6196 return setup_conf(mddev);
6199 static void *raid5_takeover_raid6(struct mddev *mddev)
6203 switch (mddev->layout) {
6204 case ALGORITHM_LEFT_ASYMMETRIC_6:
6205 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6207 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6208 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6210 case ALGORITHM_LEFT_SYMMETRIC_6:
6211 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6213 case ALGORITHM_RIGHT_SYMMETRIC_6:
6214 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6216 case ALGORITHM_PARITY_0_6:
6217 new_layout = ALGORITHM_PARITY_0;
6219 case ALGORITHM_PARITY_N:
6220 new_layout = ALGORITHM_PARITY_N;
6223 return ERR_PTR(-EINVAL);
6225 mddev->new_level = 5;
6226 mddev->new_layout = new_layout;
6227 mddev->delta_disks = -1;
6228 mddev->raid_disks -= 1;
6229 return setup_conf(mddev);
6233 static int raid5_check_reshape(struct mddev *mddev)
6235 /* For a 2-drive array, the layout and chunk size can be changed
6236 * immediately as not restriping is needed.
6237 * For larger arrays we record the new value - after validation
6238 * to be used by a reshape pass.
6240 struct r5conf *conf = mddev->private;
6241 int new_chunk = mddev->new_chunk_sectors;
6243 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6245 if (new_chunk > 0) {
6246 if (!is_power_of_2(new_chunk))
6248 if (new_chunk < (PAGE_SIZE>>9))
6250 if (mddev->array_sectors & (new_chunk-1))
6251 /* not factor of array size */
6255 /* They look valid */
6257 if (mddev->raid_disks == 2) {
6258 /* can make the change immediately */
6259 if (mddev->new_layout >= 0) {
6260 conf->algorithm = mddev->new_layout;
6261 mddev->layout = mddev->new_layout;
6263 if (new_chunk > 0) {
6264 conf->chunk_sectors = new_chunk ;
6265 mddev->chunk_sectors = new_chunk;
6267 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6268 md_wakeup_thread(mddev->thread);
6270 return check_reshape(mddev);
6273 static int raid6_check_reshape(struct mddev *mddev)
6275 int new_chunk = mddev->new_chunk_sectors;
6277 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6279 if (new_chunk > 0) {
6280 if (!is_power_of_2(new_chunk))
6282 if (new_chunk < (PAGE_SIZE >> 9))
6284 if (mddev->array_sectors & (new_chunk-1))
6285 /* not factor of array size */
6289 /* They look valid */
6290 return check_reshape(mddev);
6293 static void *raid5_takeover(struct mddev *mddev)
6295 /* raid5 can take over:
6296 * raid0 - if there is only one strip zone - make it a raid4 layout
6297 * raid1 - if there are two drives. We need to know the chunk size
6298 * raid4 - trivial - just use a raid4 layout.
6299 * raid6 - Providing it is a *_6 layout
6301 if (mddev->level == 0)
6302 return raid45_takeover_raid0(mddev, 5);
6303 if (mddev->level == 1)
6304 return raid5_takeover_raid1(mddev);
6305 if (mddev->level == 4) {
6306 mddev->new_layout = ALGORITHM_PARITY_N;
6307 mddev->new_level = 5;
6308 return setup_conf(mddev);
6310 if (mddev->level == 6)
6311 return raid5_takeover_raid6(mddev);
6313 return ERR_PTR(-EINVAL);
6316 static void *raid4_takeover(struct mddev *mddev)
6318 /* raid4 can take over:
6319 * raid0 - if there is only one strip zone
6320 * raid5 - if layout is right
6322 if (mddev->level == 0)
6323 return raid45_takeover_raid0(mddev, 4);
6324 if (mddev->level == 5 &&
6325 mddev->layout == ALGORITHM_PARITY_N) {
6326 mddev->new_layout = 0;
6327 mddev->new_level = 4;
6328 return setup_conf(mddev);
6330 return ERR_PTR(-EINVAL);
6333 static struct md_personality raid5_personality;
6335 static void *raid6_takeover(struct mddev *mddev)
6337 /* Currently can only take over a raid5. We map the
6338 * personality to an equivalent raid6 personality
6339 * with the Q block at the end.
6343 if (mddev->pers != &raid5_personality)
6344 return ERR_PTR(-EINVAL);
6345 if (mddev->degraded > 1)
6346 return ERR_PTR(-EINVAL);
6347 if (mddev->raid_disks > 253)
6348 return ERR_PTR(-EINVAL);
6349 if (mddev->raid_disks < 3)
6350 return ERR_PTR(-EINVAL);
6352 switch (mddev->layout) {
6353 case ALGORITHM_LEFT_ASYMMETRIC:
6354 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6356 case ALGORITHM_RIGHT_ASYMMETRIC:
6357 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6359 case ALGORITHM_LEFT_SYMMETRIC:
6360 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6362 case ALGORITHM_RIGHT_SYMMETRIC:
6363 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6365 case ALGORITHM_PARITY_0:
6366 new_layout = ALGORITHM_PARITY_0_6;
6368 case ALGORITHM_PARITY_N:
6369 new_layout = ALGORITHM_PARITY_N;
6372 return ERR_PTR(-EINVAL);
6374 mddev->new_level = 6;
6375 mddev->new_layout = new_layout;
6376 mddev->delta_disks = 1;
6377 mddev->raid_disks += 1;
6378 return setup_conf(mddev);
6382 static struct md_personality raid6_personality =
6386 .owner = THIS_MODULE,
6387 .make_request = make_request,
6391 .error_handler = error,
6392 .hot_add_disk = raid5_add_disk,
6393 .hot_remove_disk= raid5_remove_disk,
6394 .spare_active = raid5_spare_active,
6395 .sync_request = sync_request,
6396 .resize = raid5_resize,
6398 .check_reshape = raid6_check_reshape,
6399 .start_reshape = raid5_start_reshape,
6400 .finish_reshape = raid5_finish_reshape,
6401 .quiesce = raid5_quiesce,
6402 .takeover = raid6_takeover,
6404 static struct md_personality raid5_personality =
6408 .owner = THIS_MODULE,
6409 .make_request = make_request,
6413 .error_handler = error,
6414 .hot_add_disk = raid5_add_disk,
6415 .hot_remove_disk= raid5_remove_disk,
6416 .spare_active = raid5_spare_active,
6417 .sync_request = sync_request,
6418 .resize = raid5_resize,
6420 .check_reshape = raid5_check_reshape,
6421 .start_reshape = raid5_start_reshape,
6422 .finish_reshape = raid5_finish_reshape,
6423 .quiesce = raid5_quiesce,
6424 .takeover = raid5_takeover,
6427 static struct md_personality raid4_personality =
6431 .owner = THIS_MODULE,
6432 .make_request = make_request,
6436 .error_handler = error,
6437 .hot_add_disk = raid5_add_disk,
6438 .hot_remove_disk= raid5_remove_disk,
6439 .spare_active = raid5_spare_active,
6440 .sync_request = sync_request,
6441 .resize = raid5_resize,
6443 .check_reshape = raid5_check_reshape,
6444 .start_reshape = raid5_start_reshape,
6445 .finish_reshape = raid5_finish_reshape,
6446 .quiesce = raid5_quiesce,
6447 .takeover = raid4_takeover,
6450 static int __init raid5_init(void)
6452 register_md_personality(&raid6_personality);
6453 register_md_personality(&raid5_personality);
6454 register_md_personality(&raid4_personality);
6458 static void raid5_exit(void)
6460 unregister_md_personality(&raid6_personality);
6461 unregister_md_personality(&raid5_personality);
6462 unregister_md_personality(&raid4_personality);
6465 module_init(raid5_init);
6466 module_exit(raid5_exit);
6467 MODULE_LICENSE("GPL");
6468 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6469 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6470 MODULE_ALIAS("md-raid5");
6471 MODULE_ALIAS("md-raid4");
6472 MODULE_ALIAS("md-level-5");
6473 MODULE_ALIAS("md-level-4");
6474 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6475 MODULE_ALIAS("md-raid6");
6476 MODULE_ALIAS("md-level-6");
6478 /* This used to be two separate modules, they were: */
6479 MODULE_ALIAS("raid5");
6480 MODULE_ALIAS("raid6");