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>
65 #define NR_STRIPES 256
66 #define STRIPE_SIZE PAGE_SIZE
67 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
68 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
69 #define IO_THRESHOLD 1
70 #define BYPASS_THRESHOLD 1
71 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
72 #define HASH_MASK (NR_HASH - 1)
74 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
76 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
77 return &conf->stripe_hashtbl[hash];
80 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
81 * order without overlap. There may be several bio's per stripe+device, and
82 * a bio could span several devices.
83 * When walking this list for a particular stripe+device, we must never proceed
84 * beyond a bio that extends past this device, as the next bio might no longer
86 * This function is used to determine the 'next' bio in the list, given the sector
87 * of the current stripe+device
89 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
91 int sectors = bio->bi_size >> 9;
92 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
99 * We maintain a biased count of active stripes in the bottom 16 bits of
100 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
102 static inline int raid5_bi_processed_stripes(struct bio *bio)
104 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
105 return (atomic_read(segments) >> 16) & 0xffff;
108 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
110 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
111 return atomic_sub_return(1, segments) & 0xffff;
114 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
116 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
117 atomic_inc(segments);
120 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
123 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
127 old = atomic_read(segments);
128 new = (old & 0xffff) | (cnt << 16);
129 } while (atomic_cmpxchg(segments, old, new) != old);
132 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
134 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
135 atomic_set(segments, cnt);
138 /* Find first data disk in a raid6 stripe */
139 static inline int raid6_d0(struct stripe_head *sh)
142 /* ddf always start from first device */
144 /* md starts just after Q block */
145 if (sh->qd_idx == sh->disks - 1)
148 return sh->qd_idx + 1;
150 static inline int raid6_next_disk(int disk, int raid_disks)
153 return (disk < raid_disks) ? disk : 0;
156 /* When walking through the disks in a raid5, starting at raid6_d0,
157 * We need to map each disk to a 'slot', where the data disks are slot
158 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
159 * is raid_disks-1. This help does that mapping.
161 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
162 int *count, int syndrome_disks)
168 if (idx == sh->pd_idx)
169 return syndrome_disks;
170 if (idx == sh->qd_idx)
171 return syndrome_disks + 1;
177 static void return_io(struct bio *return_bi)
179 struct bio *bi = return_bi;
182 return_bi = bi->bi_next;
190 static void print_raid5_conf (struct r5conf *conf);
192 static int stripe_operations_active(struct stripe_head *sh)
194 return sh->check_state || sh->reconstruct_state ||
195 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
196 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
199 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
201 BUG_ON(!list_empty(&sh->lru));
202 BUG_ON(atomic_read(&conf->active_stripes)==0);
203 if (test_bit(STRIPE_HANDLE, &sh->state)) {
204 if (test_bit(STRIPE_DELAYED, &sh->state) &&
205 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
206 list_add_tail(&sh->lru, &conf->delayed_list);
207 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
208 sh->bm_seq - conf->seq_write > 0)
209 list_add_tail(&sh->lru, &conf->bitmap_list);
211 clear_bit(STRIPE_DELAYED, &sh->state);
212 clear_bit(STRIPE_BIT_DELAY, &sh->state);
213 list_add_tail(&sh->lru, &conf->handle_list);
215 md_wakeup_thread(conf->mddev->thread);
217 BUG_ON(stripe_operations_active(sh));
218 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
219 if (atomic_dec_return(&conf->preread_active_stripes)
221 md_wakeup_thread(conf->mddev->thread);
222 atomic_dec(&conf->active_stripes);
223 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
224 list_add_tail(&sh->lru, &conf->inactive_list);
225 wake_up(&conf->wait_for_stripe);
226 if (conf->retry_read_aligned)
227 md_wakeup_thread(conf->mddev->thread);
232 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
234 if (atomic_dec_and_test(&sh->count))
235 do_release_stripe(conf, sh);
238 static void release_stripe(struct stripe_head *sh)
240 struct r5conf *conf = sh->raid_conf;
243 local_irq_save(flags);
244 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
245 do_release_stripe(conf, sh);
246 spin_unlock(&conf->device_lock);
248 local_irq_restore(flags);
251 static inline void remove_hash(struct stripe_head *sh)
253 pr_debug("remove_hash(), stripe %llu\n",
254 (unsigned long long)sh->sector);
256 hlist_del_init(&sh->hash);
259 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
261 struct hlist_head *hp = stripe_hash(conf, sh->sector);
263 pr_debug("insert_hash(), stripe %llu\n",
264 (unsigned long long)sh->sector);
266 hlist_add_head(&sh->hash, hp);
270 /* find an idle stripe, make sure it is unhashed, and return it. */
271 static struct stripe_head *get_free_stripe(struct r5conf *conf)
273 struct stripe_head *sh = NULL;
274 struct list_head *first;
276 if (list_empty(&conf->inactive_list))
278 first = conf->inactive_list.next;
279 sh = list_entry(first, struct stripe_head, lru);
280 list_del_init(first);
282 atomic_inc(&conf->active_stripes);
287 static void shrink_buffers(struct stripe_head *sh)
291 int num = sh->raid_conf->pool_size;
293 for (i = 0; i < num ; i++) {
297 sh->dev[i].page = NULL;
302 static int grow_buffers(struct stripe_head *sh)
305 int num = sh->raid_conf->pool_size;
307 for (i = 0; i < num; i++) {
310 if (!(page = alloc_page(GFP_KERNEL))) {
313 sh->dev[i].page = page;
318 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
319 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
320 struct stripe_head *sh);
322 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
324 struct r5conf *conf = sh->raid_conf;
327 BUG_ON(atomic_read(&sh->count) != 0);
328 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
329 BUG_ON(stripe_operations_active(sh));
331 pr_debug("init_stripe called, stripe %llu\n",
332 (unsigned long long)sh->sector);
336 sh->generation = conf->generation - previous;
337 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
339 stripe_set_idx(sector, conf, previous, sh);
343 for (i = sh->disks; i--; ) {
344 struct r5dev *dev = &sh->dev[i];
346 if (dev->toread || dev->read || dev->towrite || dev->written ||
347 test_bit(R5_LOCKED, &dev->flags)) {
348 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
349 (unsigned long long)sh->sector, i, dev->toread,
350 dev->read, dev->towrite, dev->written,
351 test_bit(R5_LOCKED, &dev->flags));
355 raid5_build_block(sh, i, previous);
357 insert_hash(conf, sh);
360 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
363 struct stripe_head *sh;
364 struct hlist_node *hn;
366 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
367 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
368 if (sh->sector == sector && sh->generation == generation)
370 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
375 * Need to check if array has failed when deciding whether to:
377 * - remove non-faulty devices
380 * This determination is simple when no reshape is happening.
381 * However if there is a reshape, we need to carefully check
382 * both the before and after sections.
383 * This is because some failed devices may only affect one
384 * of the two sections, and some non-in_sync devices may
385 * be insync in the section most affected by failed devices.
387 static int calc_degraded(struct r5conf *conf)
389 int degraded, degraded2;
394 for (i = 0; i < conf->previous_raid_disks; i++) {
395 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
396 if (rdev && test_bit(Faulty, &rdev->flags))
397 rdev = rcu_dereference(conf->disks[i].replacement);
398 if (!rdev || test_bit(Faulty, &rdev->flags))
400 else if (test_bit(In_sync, &rdev->flags))
403 /* not in-sync or faulty.
404 * If the reshape increases the number of devices,
405 * this is being recovered by the reshape, so
406 * this 'previous' section is not in_sync.
407 * If the number of devices is being reduced however,
408 * the device can only be part of the array if
409 * we are reverting a reshape, so this section will
412 if (conf->raid_disks >= conf->previous_raid_disks)
416 if (conf->raid_disks == conf->previous_raid_disks)
420 for (i = 0; i < conf->raid_disks; i++) {
421 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
422 if (rdev && test_bit(Faulty, &rdev->flags))
423 rdev = rcu_dereference(conf->disks[i].replacement);
424 if (!rdev || test_bit(Faulty, &rdev->flags))
426 else if (test_bit(In_sync, &rdev->flags))
429 /* not in-sync or faulty.
430 * If reshape increases the number of devices, this
431 * section has already been recovered, else it
432 * almost certainly hasn't.
434 if (conf->raid_disks <= conf->previous_raid_disks)
438 if (degraded2 > degraded)
443 static int has_failed(struct r5conf *conf)
447 if (conf->mddev->reshape_position == MaxSector)
448 return conf->mddev->degraded > conf->max_degraded;
450 degraded = calc_degraded(conf);
451 if (degraded > conf->max_degraded)
456 static struct stripe_head *
457 get_active_stripe(struct r5conf *conf, sector_t sector,
458 int previous, int noblock, int noquiesce)
460 struct stripe_head *sh;
462 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
464 spin_lock_irq(&conf->device_lock);
467 wait_event_lock_irq(conf->wait_for_stripe,
468 conf->quiesce == 0 || noquiesce,
470 sh = __find_stripe(conf, sector, conf->generation - previous);
472 if (!conf->inactive_blocked)
473 sh = get_free_stripe(conf);
474 if (noblock && sh == NULL)
477 conf->inactive_blocked = 1;
478 wait_event_lock_irq(conf->wait_for_stripe,
479 !list_empty(&conf->inactive_list) &&
480 (atomic_read(&conf->active_stripes)
481 < (conf->max_nr_stripes *3/4)
482 || !conf->inactive_blocked),
484 conf->inactive_blocked = 0;
486 init_stripe(sh, sector, previous);
488 if (atomic_read(&sh->count)) {
489 BUG_ON(!list_empty(&sh->lru)
490 && !test_bit(STRIPE_EXPANDING, &sh->state)
491 && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state));
493 if (!test_bit(STRIPE_HANDLE, &sh->state))
494 atomic_inc(&conf->active_stripes);
495 if (list_empty(&sh->lru) &&
496 !test_bit(STRIPE_EXPANDING, &sh->state))
498 list_del_init(&sh->lru);
501 } while (sh == NULL);
504 atomic_inc(&sh->count);
506 spin_unlock_irq(&conf->device_lock);
510 /* Determine if 'data_offset' or 'new_data_offset' should be used
511 * in this stripe_head.
513 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
515 sector_t progress = conf->reshape_progress;
516 /* Need a memory barrier to make sure we see the value
517 * of conf->generation, or ->data_offset that was set before
518 * reshape_progress was updated.
521 if (progress == MaxSector)
523 if (sh->generation == conf->generation - 1)
525 /* We are in a reshape, and this is a new-generation stripe,
526 * so use new_data_offset.
532 raid5_end_read_request(struct bio *bi, int error);
534 raid5_end_write_request(struct bio *bi, int error);
536 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
538 struct r5conf *conf = sh->raid_conf;
539 int i, disks = sh->disks;
543 for (i = disks; i--; ) {
545 int replace_only = 0;
546 struct bio *bi, *rbi;
547 struct md_rdev *rdev, *rrdev = NULL;
548 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
549 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
553 if (test_bit(R5_Discard, &sh->dev[i].flags))
555 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
557 else if (test_and_clear_bit(R5_WantReplace,
558 &sh->dev[i].flags)) {
563 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
566 bi = &sh->dev[i].req;
567 rbi = &sh->dev[i].rreq; /* For writing to replacement */
572 bi->bi_end_io = raid5_end_write_request;
573 rbi->bi_end_io = raid5_end_write_request;
575 bi->bi_end_io = raid5_end_read_request;
578 rrdev = rcu_dereference(conf->disks[i].replacement);
579 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
580 rdev = rcu_dereference(conf->disks[i].rdev);
589 /* We raced and saw duplicates */
592 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
597 if (rdev && test_bit(Faulty, &rdev->flags))
600 atomic_inc(&rdev->nr_pending);
601 if (rrdev && test_bit(Faulty, &rrdev->flags))
604 atomic_inc(&rrdev->nr_pending);
607 /* We have already checked bad blocks for reads. Now
608 * need to check for writes. We never accept write errors
609 * on the replacement, so we don't to check rrdev.
611 while ((rw & WRITE) && rdev &&
612 test_bit(WriteErrorSeen, &rdev->flags)) {
615 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
616 &first_bad, &bad_sectors);
621 set_bit(BlockedBadBlocks, &rdev->flags);
622 if (!conf->mddev->external &&
623 conf->mddev->flags) {
624 /* It is very unlikely, but we might
625 * still need to write out the
626 * bad block log - better give it
628 md_check_recovery(conf->mddev);
631 * Because md_wait_for_blocked_rdev
632 * will dec nr_pending, we must
633 * increment it first.
635 atomic_inc(&rdev->nr_pending);
636 md_wait_for_blocked_rdev(rdev, conf->mddev);
638 /* Acknowledged bad block - skip the write */
639 rdev_dec_pending(rdev, conf->mddev);
645 if (s->syncing || s->expanding || s->expanded
647 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
649 set_bit(STRIPE_IO_STARTED, &sh->state);
651 bi->bi_bdev = rdev->bdev;
652 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
653 __func__, (unsigned long long)sh->sector,
655 atomic_inc(&sh->count);
656 if (use_new_offset(conf, sh))
657 bi->bi_sector = (sh->sector
658 + rdev->new_data_offset);
660 bi->bi_sector = (sh->sector
661 + rdev->data_offset);
662 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
663 bi->bi_rw |= REQ_FLUSH;
665 bi->bi_flags = 1 << BIO_UPTODATE;
667 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
668 bi->bi_io_vec[0].bv_offset = 0;
669 bi->bi_size = STRIPE_SIZE;
672 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
673 generic_make_request(bi);
676 if (s->syncing || s->expanding || s->expanded
678 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
680 set_bit(STRIPE_IO_STARTED, &sh->state);
682 rbi->bi_bdev = rrdev->bdev;
683 pr_debug("%s: for %llu schedule op %ld on "
684 "replacement disc %d\n",
685 __func__, (unsigned long long)sh->sector,
687 atomic_inc(&sh->count);
688 if (use_new_offset(conf, sh))
689 rbi->bi_sector = (sh->sector
690 + rrdev->new_data_offset);
692 rbi->bi_sector = (sh->sector
693 + rrdev->data_offset);
694 rbi->bi_flags = 1 << BIO_UPTODATE;
696 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
697 rbi->bi_io_vec[0].bv_offset = 0;
698 rbi->bi_size = STRIPE_SIZE;
700 generic_make_request(rbi);
702 if (!rdev && !rrdev) {
704 set_bit(STRIPE_DEGRADED, &sh->state);
705 pr_debug("skip op %ld on disc %d for sector %llu\n",
706 bi->bi_rw, i, (unsigned long long)sh->sector);
707 clear_bit(R5_LOCKED, &sh->dev[i].flags);
708 set_bit(STRIPE_HANDLE, &sh->state);
713 static struct dma_async_tx_descriptor *
714 async_copy_data(int frombio, struct bio *bio, struct page *page,
715 sector_t sector, struct dma_async_tx_descriptor *tx)
718 struct page *bio_page;
721 struct async_submit_ctl submit;
722 enum async_tx_flags flags = 0;
724 if (bio->bi_sector >= sector)
725 page_offset = (signed)(bio->bi_sector - sector) * 512;
727 page_offset = (signed)(sector - bio->bi_sector) * -512;
730 flags |= ASYNC_TX_FENCE;
731 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
733 bio_for_each_segment(bvl, bio, i) {
734 int len = bvl->bv_len;
738 if (page_offset < 0) {
739 b_offset = -page_offset;
740 page_offset += b_offset;
744 if (len > 0 && page_offset + len > STRIPE_SIZE)
745 clen = STRIPE_SIZE - page_offset;
750 b_offset += bvl->bv_offset;
751 bio_page = bvl->bv_page;
753 tx = async_memcpy(page, bio_page, page_offset,
754 b_offset, clen, &submit);
756 tx = async_memcpy(bio_page, page, b_offset,
757 page_offset, clen, &submit);
759 /* chain the operations */
760 submit.depend_tx = tx;
762 if (clen < len) /* hit end of page */
770 static void ops_complete_biofill(void *stripe_head_ref)
772 struct stripe_head *sh = stripe_head_ref;
773 struct bio *return_bi = NULL;
776 pr_debug("%s: stripe %llu\n", __func__,
777 (unsigned long long)sh->sector);
779 /* clear completed biofills */
780 for (i = sh->disks; i--; ) {
781 struct r5dev *dev = &sh->dev[i];
783 /* acknowledge completion of a biofill operation */
784 /* and check if we need to reply to a read request,
785 * new R5_Wantfill requests are held off until
786 * !STRIPE_BIOFILL_RUN
788 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
789 struct bio *rbi, *rbi2;
794 while (rbi && rbi->bi_sector <
795 dev->sector + STRIPE_SECTORS) {
796 rbi2 = r5_next_bio(rbi, dev->sector);
797 if (!raid5_dec_bi_active_stripes(rbi)) {
798 rbi->bi_next = return_bi;
805 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
807 return_io(return_bi);
809 set_bit(STRIPE_HANDLE, &sh->state);
813 static void ops_run_biofill(struct stripe_head *sh)
815 struct dma_async_tx_descriptor *tx = NULL;
816 struct async_submit_ctl submit;
819 pr_debug("%s: stripe %llu\n", __func__,
820 (unsigned long long)sh->sector);
822 for (i = sh->disks; i--; ) {
823 struct r5dev *dev = &sh->dev[i];
824 if (test_bit(R5_Wantfill, &dev->flags)) {
826 spin_lock_irq(&sh->stripe_lock);
827 dev->read = rbi = dev->toread;
829 spin_unlock_irq(&sh->stripe_lock);
830 while (rbi && rbi->bi_sector <
831 dev->sector + STRIPE_SECTORS) {
832 tx = async_copy_data(0, rbi, dev->page,
834 rbi = r5_next_bio(rbi, dev->sector);
839 atomic_inc(&sh->count);
840 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
841 async_trigger_callback(&submit);
844 static void mark_target_uptodate(struct stripe_head *sh, int target)
851 tgt = &sh->dev[target];
852 set_bit(R5_UPTODATE, &tgt->flags);
853 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
854 clear_bit(R5_Wantcompute, &tgt->flags);
857 static void ops_complete_compute(void *stripe_head_ref)
859 struct stripe_head *sh = stripe_head_ref;
861 pr_debug("%s: stripe %llu\n", __func__,
862 (unsigned long long)sh->sector);
864 /* mark the computed target(s) as uptodate */
865 mark_target_uptodate(sh, sh->ops.target);
866 mark_target_uptodate(sh, sh->ops.target2);
868 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
869 if (sh->check_state == check_state_compute_run)
870 sh->check_state = check_state_compute_result;
871 set_bit(STRIPE_HANDLE, &sh->state);
875 /* return a pointer to the address conversion region of the scribble buffer */
876 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
877 struct raid5_percpu *percpu)
879 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
882 static struct dma_async_tx_descriptor *
883 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
885 int disks = sh->disks;
886 struct page **xor_srcs = percpu->scribble;
887 int target = sh->ops.target;
888 struct r5dev *tgt = &sh->dev[target];
889 struct page *xor_dest = tgt->page;
891 struct dma_async_tx_descriptor *tx;
892 struct async_submit_ctl submit;
895 pr_debug("%s: stripe %llu block: %d\n",
896 __func__, (unsigned long long)sh->sector, target);
897 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
899 for (i = disks; i--; )
901 xor_srcs[count++] = sh->dev[i].page;
903 atomic_inc(&sh->count);
905 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
906 ops_complete_compute, sh, to_addr_conv(sh, percpu));
907 if (unlikely(count == 1))
908 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
910 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
915 /* set_syndrome_sources - populate source buffers for gen_syndrome
916 * @srcs - (struct page *) array of size sh->disks
917 * @sh - stripe_head to parse
919 * Populates srcs in proper layout order for the stripe and returns the
920 * 'count' of sources to be used in a call to async_gen_syndrome. The P
921 * destination buffer is recorded in srcs[count] and the Q destination
922 * is recorded in srcs[count+1]].
924 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
926 int disks = sh->disks;
927 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
928 int d0_idx = raid6_d0(sh);
932 for (i = 0; i < disks; i++)
938 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
940 srcs[slot] = sh->dev[i].page;
941 i = raid6_next_disk(i, disks);
942 } while (i != d0_idx);
944 return syndrome_disks;
947 static struct dma_async_tx_descriptor *
948 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
950 int disks = sh->disks;
951 struct page **blocks = percpu->scribble;
953 int qd_idx = sh->qd_idx;
954 struct dma_async_tx_descriptor *tx;
955 struct async_submit_ctl submit;
961 if (sh->ops.target < 0)
962 target = sh->ops.target2;
963 else if (sh->ops.target2 < 0)
964 target = sh->ops.target;
966 /* we should only have one valid target */
969 pr_debug("%s: stripe %llu block: %d\n",
970 __func__, (unsigned long long)sh->sector, target);
972 tgt = &sh->dev[target];
973 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
976 atomic_inc(&sh->count);
978 if (target == qd_idx) {
979 count = set_syndrome_sources(blocks, sh);
980 blocks[count] = NULL; /* regenerating p is not necessary */
981 BUG_ON(blocks[count+1] != dest); /* q should already be set */
982 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
983 ops_complete_compute, sh,
984 to_addr_conv(sh, percpu));
985 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
987 /* Compute any data- or p-drive using XOR */
989 for (i = disks; i-- ; ) {
990 if (i == target || i == qd_idx)
992 blocks[count++] = sh->dev[i].page;
995 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
996 NULL, ops_complete_compute, sh,
997 to_addr_conv(sh, percpu));
998 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1004 static struct dma_async_tx_descriptor *
1005 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1007 int i, count, disks = sh->disks;
1008 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1009 int d0_idx = raid6_d0(sh);
1010 int faila = -1, failb = -1;
1011 int target = sh->ops.target;
1012 int target2 = sh->ops.target2;
1013 struct r5dev *tgt = &sh->dev[target];
1014 struct r5dev *tgt2 = &sh->dev[target2];
1015 struct dma_async_tx_descriptor *tx;
1016 struct page **blocks = percpu->scribble;
1017 struct async_submit_ctl submit;
1019 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1020 __func__, (unsigned long long)sh->sector, target, target2);
1021 BUG_ON(target < 0 || target2 < 0);
1022 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1023 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1025 /* we need to open-code set_syndrome_sources to handle the
1026 * slot number conversion for 'faila' and 'failb'
1028 for (i = 0; i < disks ; i++)
1033 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1035 blocks[slot] = sh->dev[i].page;
1041 i = raid6_next_disk(i, disks);
1042 } while (i != d0_idx);
1044 BUG_ON(faila == failb);
1047 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1048 __func__, (unsigned long long)sh->sector, faila, failb);
1050 atomic_inc(&sh->count);
1052 if (failb == syndrome_disks+1) {
1053 /* Q disk is one of the missing disks */
1054 if (faila == syndrome_disks) {
1055 /* Missing P+Q, just recompute */
1056 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1057 ops_complete_compute, sh,
1058 to_addr_conv(sh, percpu));
1059 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1060 STRIPE_SIZE, &submit);
1064 int qd_idx = sh->qd_idx;
1066 /* Missing D+Q: recompute D from P, then recompute Q */
1067 if (target == qd_idx)
1068 data_target = target2;
1070 data_target = target;
1073 for (i = disks; i-- ; ) {
1074 if (i == data_target || i == qd_idx)
1076 blocks[count++] = sh->dev[i].page;
1078 dest = sh->dev[data_target].page;
1079 init_async_submit(&submit,
1080 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1082 to_addr_conv(sh, percpu));
1083 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1086 count = set_syndrome_sources(blocks, sh);
1087 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1088 ops_complete_compute, sh,
1089 to_addr_conv(sh, percpu));
1090 return async_gen_syndrome(blocks, 0, count+2,
1091 STRIPE_SIZE, &submit);
1094 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1095 ops_complete_compute, sh,
1096 to_addr_conv(sh, percpu));
1097 if (failb == syndrome_disks) {
1098 /* We're missing D+P. */
1099 return async_raid6_datap_recov(syndrome_disks+2,
1103 /* We're missing D+D. */
1104 return async_raid6_2data_recov(syndrome_disks+2,
1105 STRIPE_SIZE, faila, failb,
1112 static void ops_complete_prexor(void *stripe_head_ref)
1114 struct stripe_head *sh = stripe_head_ref;
1116 pr_debug("%s: stripe %llu\n", __func__,
1117 (unsigned long long)sh->sector);
1120 static struct dma_async_tx_descriptor *
1121 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1122 struct dma_async_tx_descriptor *tx)
1124 int disks = sh->disks;
1125 struct page **xor_srcs = percpu->scribble;
1126 int count = 0, pd_idx = sh->pd_idx, i;
1127 struct async_submit_ctl submit;
1129 /* existing parity data subtracted */
1130 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1132 pr_debug("%s: stripe %llu\n", __func__,
1133 (unsigned long long)sh->sector);
1135 for (i = disks; i--; ) {
1136 struct r5dev *dev = &sh->dev[i];
1137 /* Only process blocks that are known to be uptodate */
1138 if (test_bit(R5_Wantdrain, &dev->flags))
1139 xor_srcs[count++] = dev->page;
1142 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1143 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1144 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1149 static struct dma_async_tx_descriptor *
1150 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1152 int disks = sh->disks;
1155 pr_debug("%s: stripe %llu\n", __func__,
1156 (unsigned long long)sh->sector);
1158 for (i = disks; i--; ) {
1159 struct r5dev *dev = &sh->dev[i];
1162 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1165 spin_lock_irq(&sh->stripe_lock);
1166 chosen = dev->towrite;
1167 dev->towrite = NULL;
1168 BUG_ON(dev->written);
1169 wbi = dev->written = chosen;
1170 spin_unlock_irq(&sh->stripe_lock);
1172 while (wbi && wbi->bi_sector <
1173 dev->sector + STRIPE_SECTORS) {
1174 if (wbi->bi_rw & REQ_FUA)
1175 set_bit(R5_WantFUA, &dev->flags);
1176 if (wbi->bi_rw & REQ_SYNC)
1177 set_bit(R5_SyncIO, &dev->flags);
1178 if (wbi->bi_rw & REQ_DISCARD)
1179 set_bit(R5_Discard, &dev->flags);
1181 tx = async_copy_data(1, wbi, dev->page,
1183 wbi = r5_next_bio(wbi, dev->sector);
1191 static void ops_complete_reconstruct(void *stripe_head_ref)
1193 struct stripe_head *sh = stripe_head_ref;
1194 int disks = sh->disks;
1195 int pd_idx = sh->pd_idx;
1196 int qd_idx = sh->qd_idx;
1198 bool fua = false, sync = false, discard = false;
1200 pr_debug("%s: stripe %llu\n", __func__,
1201 (unsigned long long)sh->sector);
1203 for (i = disks; i--; ) {
1204 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1205 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1206 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1209 for (i = disks; i--; ) {
1210 struct r5dev *dev = &sh->dev[i];
1212 if (dev->written || i == pd_idx || i == qd_idx) {
1214 set_bit(R5_UPTODATE, &dev->flags);
1216 set_bit(R5_WantFUA, &dev->flags);
1218 set_bit(R5_SyncIO, &dev->flags);
1222 if (sh->reconstruct_state == reconstruct_state_drain_run)
1223 sh->reconstruct_state = reconstruct_state_drain_result;
1224 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1225 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1227 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1228 sh->reconstruct_state = reconstruct_state_result;
1231 set_bit(STRIPE_HANDLE, &sh->state);
1236 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1237 struct dma_async_tx_descriptor *tx)
1239 int disks = sh->disks;
1240 struct page **xor_srcs = percpu->scribble;
1241 struct async_submit_ctl submit;
1242 int count = 0, pd_idx = sh->pd_idx, i;
1243 struct page *xor_dest;
1245 unsigned long flags;
1247 pr_debug("%s: stripe %llu\n", __func__,
1248 (unsigned long long)sh->sector);
1250 for (i = 0; i < sh->disks; i++) {
1253 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1256 if (i >= sh->disks) {
1257 atomic_inc(&sh->count);
1258 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1259 ops_complete_reconstruct(sh);
1262 /* check if prexor is active which means only process blocks
1263 * that are part of a read-modify-write (written)
1265 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1267 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1268 for (i = disks; i--; ) {
1269 struct r5dev *dev = &sh->dev[i];
1271 xor_srcs[count++] = dev->page;
1274 xor_dest = sh->dev[pd_idx].page;
1275 for (i = disks; i--; ) {
1276 struct r5dev *dev = &sh->dev[i];
1278 xor_srcs[count++] = dev->page;
1282 /* 1/ if we prexor'd then the dest is reused as a source
1283 * 2/ if we did not prexor then we are redoing the parity
1284 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1285 * for the synchronous xor case
1287 flags = ASYNC_TX_ACK |
1288 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1290 atomic_inc(&sh->count);
1292 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1293 to_addr_conv(sh, percpu));
1294 if (unlikely(count == 1))
1295 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1297 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1301 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1302 struct dma_async_tx_descriptor *tx)
1304 struct async_submit_ctl submit;
1305 struct page **blocks = percpu->scribble;
1308 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1310 for (i = 0; i < sh->disks; i++) {
1311 if (sh->pd_idx == i || sh->qd_idx == i)
1313 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1316 if (i >= sh->disks) {
1317 atomic_inc(&sh->count);
1318 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1319 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1320 ops_complete_reconstruct(sh);
1324 count = set_syndrome_sources(blocks, sh);
1326 atomic_inc(&sh->count);
1328 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1329 sh, to_addr_conv(sh, percpu));
1330 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1333 static void ops_complete_check(void *stripe_head_ref)
1335 struct stripe_head *sh = stripe_head_ref;
1337 pr_debug("%s: stripe %llu\n", __func__,
1338 (unsigned long long)sh->sector);
1340 sh->check_state = check_state_check_result;
1341 set_bit(STRIPE_HANDLE, &sh->state);
1345 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1347 int disks = sh->disks;
1348 int pd_idx = sh->pd_idx;
1349 int qd_idx = sh->qd_idx;
1350 struct page *xor_dest;
1351 struct page **xor_srcs = percpu->scribble;
1352 struct dma_async_tx_descriptor *tx;
1353 struct async_submit_ctl submit;
1357 pr_debug("%s: stripe %llu\n", __func__,
1358 (unsigned long long)sh->sector);
1361 xor_dest = sh->dev[pd_idx].page;
1362 xor_srcs[count++] = xor_dest;
1363 for (i = disks; i--; ) {
1364 if (i == pd_idx || i == qd_idx)
1366 xor_srcs[count++] = sh->dev[i].page;
1369 init_async_submit(&submit, 0, NULL, NULL, NULL,
1370 to_addr_conv(sh, percpu));
1371 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1372 &sh->ops.zero_sum_result, &submit);
1374 atomic_inc(&sh->count);
1375 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1376 tx = async_trigger_callback(&submit);
1379 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1381 struct page **srcs = percpu->scribble;
1382 struct async_submit_ctl submit;
1385 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1386 (unsigned long long)sh->sector, checkp);
1388 count = set_syndrome_sources(srcs, sh);
1392 atomic_inc(&sh->count);
1393 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1394 sh, to_addr_conv(sh, percpu));
1395 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1396 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1399 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1401 int overlap_clear = 0, i, disks = sh->disks;
1402 struct dma_async_tx_descriptor *tx = NULL;
1403 struct r5conf *conf = sh->raid_conf;
1404 int level = conf->level;
1405 struct raid5_percpu *percpu;
1409 percpu = per_cpu_ptr(conf->percpu, cpu);
1410 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1411 ops_run_biofill(sh);
1415 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1417 tx = ops_run_compute5(sh, percpu);
1419 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1420 tx = ops_run_compute6_1(sh, percpu);
1422 tx = ops_run_compute6_2(sh, percpu);
1424 /* terminate the chain if reconstruct is not set to be run */
1425 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1429 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1430 tx = ops_run_prexor(sh, percpu, tx);
1432 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1433 tx = ops_run_biodrain(sh, tx);
1437 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1439 ops_run_reconstruct5(sh, percpu, tx);
1441 ops_run_reconstruct6(sh, percpu, tx);
1444 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1445 if (sh->check_state == check_state_run)
1446 ops_run_check_p(sh, percpu);
1447 else if (sh->check_state == check_state_run_q)
1448 ops_run_check_pq(sh, percpu, 0);
1449 else if (sh->check_state == check_state_run_pq)
1450 ops_run_check_pq(sh, percpu, 1);
1456 for (i = disks; i--; ) {
1457 struct r5dev *dev = &sh->dev[i];
1458 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1459 wake_up(&sh->raid_conf->wait_for_overlap);
1464 #ifdef CONFIG_MULTICORE_RAID456
1465 static void async_run_ops(void *param, async_cookie_t cookie)
1467 struct stripe_head *sh = param;
1468 unsigned long ops_request = sh->ops.request;
1470 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1471 wake_up(&sh->ops.wait_for_ops);
1473 __raid_run_ops(sh, ops_request);
1477 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1479 /* since handle_stripe can be called outside of raid5d context
1480 * we need to ensure sh->ops.request is de-staged before another
1483 wait_event(sh->ops.wait_for_ops,
1484 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1485 sh->ops.request = ops_request;
1487 atomic_inc(&sh->count);
1488 async_schedule(async_run_ops, sh);
1491 #define raid_run_ops __raid_run_ops
1494 static int grow_one_stripe(struct r5conf *conf)
1496 struct stripe_head *sh;
1497 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1501 sh->raid_conf = conf;
1502 #ifdef CONFIG_MULTICORE_RAID456
1503 init_waitqueue_head(&sh->ops.wait_for_ops);
1506 spin_lock_init(&sh->stripe_lock);
1508 if (grow_buffers(sh)) {
1510 kmem_cache_free(conf->slab_cache, sh);
1513 /* we just created an active stripe so... */
1514 atomic_set(&sh->count, 1);
1515 atomic_inc(&conf->active_stripes);
1516 INIT_LIST_HEAD(&sh->lru);
1521 static int grow_stripes(struct r5conf *conf, int num)
1523 struct kmem_cache *sc;
1524 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1526 if (conf->mddev->gendisk)
1527 sprintf(conf->cache_name[0],
1528 "raid%d-%s", conf->level, mdname(conf->mddev));
1530 sprintf(conf->cache_name[0],
1531 "raid%d-%p", conf->level, conf->mddev);
1532 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1534 conf->active_name = 0;
1535 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1536 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1540 conf->slab_cache = sc;
1541 conf->pool_size = devs;
1543 if (!grow_one_stripe(conf))
1549 * scribble_len - return the required size of the scribble region
1550 * @num - total number of disks in the array
1552 * The size must be enough to contain:
1553 * 1/ a struct page pointer for each device in the array +2
1554 * 2/ room to convert each entry in (1) to its corresponding dma
1555 * (dma_map_page()) or page (page_address()) address.
1557 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1558 * calculate over all devices (not just the data blocks), using zeros in place
1559 * of the P and Q blocks.
1561 static size_t scribble_len(int num)
1565 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1570 static int resize_stripes(struct r5conf *conf, int newsize)
1572 /* Make all the stripes able to hold 'newsize' devices.
1573 * New slots in each stripe get 'page' set to a new page.
1575 * This happens in stages:
1576 * 1/ create a new kmem_cache and allocate the required number of
1578 * 2/ gather all the old stripe_heads and transfer the pages across
1579 * to the new stripe_heads. This will have the side effect of
1580 * freezing the array as once all stripe_heads have been collected,
1581 * no IO will be possible. Old stripe heads are freed once their
1582 * pages have been transferred over, and the old kmem_cache is
1583 * freed when all stripes are done.
1584 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1585 * we simple return a failre status - no need to clean anything up.
1586 * 4/ allocate new pages for the new slots in the new stripe_heads.
1587 * If this fails, we don't bother trying the shrink the
1588 * stripe_heads down again, we just leave them as they are.
1589 * As each stripe_head is processed the new one is released into
1592 * Once step2 is started, we cannot afford to wait for a write,
1593 * so we use GFP_NOIO allocations.
1595 struct stripe_head *osh, *nsh;
1596 LIST_HEAD(newstripes);
1597 struct disk_info *ndisks;
1600 struct kmem_cache *sc;
1603 if (newsize <= conf->pool_size)
1604 return 0; /* never bother to shrink */
1606 err = md_allow_write(conf->mddev);
1611 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1612 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1617 for (i = conf->max_nr_stripes; i; i--) {
1618 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1622 nsh->raid_conf = conf;
1623 #ifdef CONFIG_MULTICORE_RAID456
1624 init_waitqueue_head(&nsh->ops.wait_for_ops);
1626 spin_lock_init(&nsh->stripe_lock);
1628 list_add(&nsh->lru, &newstripes);
1631 /* didn't get enough, give up */
1632 while (!list_empty(&newstripes)) {
1633 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1634 list_del(&nsh->lru);
1635 kmem_cache_free(sc, nsh);
1637 kmem_cache_destroy(sc);
1640 /* Step 2 - Must use GFP_NOIO now.
1641 * OK, we have enough stripes, start collecting inactive
1642 * stripes and copying them over
1644 list_for_each_entry(nsh, &newstripes, lru) {
1645 spin_lock_irq(&conf->device_lock);
1646 wait_event_lock_irq(conf->wait_for_stripe,
1647 !list_empty(&conf->inactive_list),
1649 osh = get_free_stripe(conf);
1650 spin_unlock_irq(&conf->device_lock);
1651 atomic_set(&nsh->count, 1);
1652 for(i=0; i<conf->pool_size; i++)
1653 nsh->dev[i].page = osh->dev[i].page;
1654 for( ; i<newsize; i++)
1655 nsh->dev[i].page = NULL;
1656 kmem_cache_free(conf->slab_cache, osh);
1658 kmem_cache_destroy(conf->slab_cache);
1661 * At this point, we are holding all the stripes so the array
1662 * is completely stalled, so now is a good time to resize
1663 * conf->disks and the scribble region
1665 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1667 for (i=0; i<conf->raid_disks; i++)
1668 ndisks[i] = conf->disks[i];
1670 conf->disks = ndisks;
1675 conf->scribble_len = scribble_len(newsize);
1676 for_each_present_cpu(cpu) {
1677 struct raid5_percpu *percpu;
1680 percpu = per_cpu_ptr(conf->percpu, cpu);
1681 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1684 kfree(percpu->scribble);
1685 percpu->scribble = scribble;
1693 /* Step 4, return new stripes to service */
1694 while(!list_empty(&newstripes)) {
1695 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1696 list_del_init(&nsh->lru);
1698 for (i=conf->raid_disks; i < newsize; i++)
1699 if (nsh->dev[i].page == NULL) {
1700 struct page *p = alloc_page(GFP_NOIO);
1701 nsh->dev[i].page = p;
1705 release_stripe(nsh);
1707 /* critical section pass, GFP_NOIO no longer needed */
1709 conf->slab_cache = sc;
1710 conf->active_name = 1-conf->active_name;
1711 conf->pool_size = newsize;
1715 static int drop_one_stripe(struct r5conf *conf)
1717 struct stripe_head *sh;
1719 spin_lock_irq(&conf->device_lock);
1720 sh = get_free_stripe(conf);
1721 spin_unlock_irq(&conf->device_lock);
1724 BUG_ON(atomic_read(&sh->count));
1726 kmem_cache_free(conf->slab_cache, sh);
1727 atomic_dec(&conf->active_stripes);
1731 static void shrink_stripes(struct r5conf *conf)
1733 while (drop_one_stripe(conf))
1736 if (conf->slab_cache)
1737 kmem_cache_destroy(conf->slab_cache);
1738 conf->slab_cache = NULL;
1741 static void raid5_end_read_request(struct bio * bi, int error)
1743 struct stripe_head *sh = bi->bi_private;
1744 struct r5conf *conf = sh->raid_conf;
1745 int disks = sh->disks, i;
1746 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1747 char b[BDEVNAME_SIZE];
1748 struct md_rdev *rdev = NULL;
1751 for (i=0 ; i<disks; i++)
1752 if (bi == &sh->dev[i].req)
1755 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1756 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1762 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1763 /* If replacement finished while this request was outstanding,
1764 * 'replacement' might be NULL already.
1765 * In that case it moved down to 'rdev'.
1766 * rdev is not removed until all requests are finished.
1768 rdev = conf->disks[i].replacement;
1770 rdev = conf->disks[i].rdev;
1772 if (use_new_offset(conf, sh))
1773 s = sh->sector + rdev->new_data_offset;
1775 s = sh->sector + rdev->data_offset;
1777 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1778 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1779 /* Note that this cannot happen on a
1780 * replacement device. We just fail those on
1785 "md/raid:%s: read error corrected"
1786 " (%lu sectors at %llu on %s)\n",
1787 mdname(conf->mddev), STRIPE_SECTORS,
1788 (unsigned long long)s,
1789 bdevname(rdev->bdev, b));
1790 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1791 clear_bit(R5_ReadError, &sh->dev[i].flags);
1792 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1793 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1794 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1796 if (atomic_read(&rdev->read_errors))
1797 atomic_set(&rdev->read_errors, 0);
1799 const char *bdn = bdevname(rdev->bdev, b);
1803 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1804 atomic_inc(&rdev->read_errors);
1805 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1808 "md/raid:%s: read error on replacement device "
1809 "(sector %llu on %s).\n",
1810 mdname(conf->mddev),
1811 (unsigned long long)s,
1813 else if (conf->mddev->degraded >= conf->max_degraded) {
1817 "md/raid:%s: read error not correctable "
1818 "(sector %llu on %s).\n",
1819 mdname(conf->mddev),
1820 (unsigned long long)s,
1822 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1827 "md/raid:%s: read error NOT corrected!! "
1828 "(sector %llu on %s).\n",
1829 mdname(conf->mddev),
1830 (unsigned long long)s,
1832 } else if (atomic_read(&rdev->read_errors)
1833 > conf->max_nr_stripes)
1835 "md/raid:%s: Too many read errors, failing device %s.\n",
1836 mdname(conf->mddev), bdn);
1840 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1841 set_bit(R5_ReadError, &sh->dev[i].flags);
1842 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1844 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1846 clear_bit(R5_ReadError, &sh->dev[i].flags);
1847 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1849 && test_bit(In_sync, &rdev->flags)
1850 && rdev_set_badblocks(
1851 rdev, sh->sector, STRIPE_SECTORS, 0)))
1852 md_error(conf->mddev, rdev);
1855 rdev_dec_pending(rdev, conf->mddev);
1856 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1857 set_bit(STRIPE_HANDLE, &sh->state);
1861 static void raid5_end_write_request(struct bio *bi, int error)
1863 struct stripe_head *sh = bi->bi_private;
1864 struct r5conf *conf = sh->raid_conf;
1865 int disks = sh->disks, i;
1866 struct md_rdev *uninitialized_var(rdev);
1867 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1870 int replacement = 0;
1872 for (i = 0 ; i < disks; i++) {
1873 if (bi == &sh->dev[i].req) {
1874 rdev = conf->disks[i].rdev;
1877 if (bi == &sh->dev[i].rreq) {
1878 rdev = conf->disks[i].replacement;
1882 /* rdev was removed and 'replacement'
1883 * replaced it. rdev is not removed
1884 * until all requests are finished.
1886 rdev = conf->disks[i].rdev;
1890 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1891 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1900 md_error(conf->mddev, rdev);
1901 else if (is_badblock(rdev, sh->sector,
1903 &first_bad, &bad_sectors))
1904 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1907 set_bit(WriteErrorSeen, &rdev->flags);
1908 set_bit(R5_WriteError, &sh->dev[i].flags);
1909 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1910 set_bit(MD_RECOVERY_NEEDED,
1911 &rdev->mddev->recovery);
1912 } else if (is_badblock(rdev, sh->sector,
1914 &first_bad, &bad_sectors))
1915 set_bit(R5_MadeGood, &sh->dev[i].flags);
1917 rdev_dec_pending(rdev, conf->mddev);
1919 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1920 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1921 set_bit(STRIPE_HANDLE, &sh->state);
1925 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1927 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1929 struct r5dev *dev = &sh->dev[i];
1931 bio_init(&dev->req);
1932 dev->req.bi_io_vec = &dev->vec;
1934 dev->req.bi_max_vecs++;
1935 dev->req.bi_private = sh;
1936 dev->vec.bv_page = dev->page;
1938 bio_init(&dev->rreq);
1939 dev->rreq.bi_io_vec = &dev->rvec;
1940 dev->rreq.bi_vcnt++;
1941 dev->rreq.bi_max_vecs++;
1942 dev->rreq.bi_private = sh;
1943 dev->rvec.bv_page = dev->page;
1946 dev->sector = compute_blocknr(sh, i, previous);
1949 static void error(struct mddev *mddev, struct md_rdev *rdev)
1951 char b[BDEVNAME_SIZE];
1952 struct r5conf *conf = mddev->private;
1953 unsigned long flags;
1954 pr_debug("raid456: error called\n");
1956 spin_lock_irqsave(&conf->device_lock, flags);
1957 clear_bit(In_sync, &rdev->flags);
1958 mddev->degraded = calc_degraded(conf);
1959 spin_unlock_irqrestore(&conf->device_lock, flags);
1960 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1962 set_bit(Blocked, &rdev->flags);
1963 set_bit(Faulty, &rdev->flags);
1964 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1966 "md/raid:%s: Disk failure on %s, disabling device.\n"
1967 "md/raid:%s: Operation continuing on %d devices.\n",
1969 bdevname(rdev->bdev, b),
1971 conf->raid_disks - mddev->degraded);
1975 * Input: a 'big' sector number,
1976 * Output: index of the data and parity disk, and the sector # in them.
1978 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1979 int previous, int *dd_idx,
1980 struct stripe_head *sh)
1982 sector_t stripe, stripe2;
1983 sector_t chunk_number;
1984 unsigned int chunk_offset;
1987 sector_t new_sector;
1988 int algorithm = previous ? conf->prev_algo
1990 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1991 : conf->chunk_sectors;
1992 int raid_disks = previous ? conf->previous_raid_disks
1994 int data_disks = raid_disks - conf->max_degraded;
1996 /* First compute the information on this sector */
1999 * Compute the chunk number and the sector offset inside the chunk
2001 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2002 chunk_number = r_sector;
2005 * Compute the stripe number
2007 stripe = chunk_number;
2008 *dd_idx = sector_div(stripe, data_disks);
2011 * Select the parity disk based on the user selected algorithm.
2013 pd_idx = qd_idx = -1;
2014 switch(conf->level) {
2016 pd_idx = data_disks;
2019 switch (algorithm) {
2020 case ALGORITHM_LEFT_ASYMMETRIC:
2021 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2022 if (*dd_idx >= pd_idx)
2025 case ALGORITHM_RIGHT_ASYMMETRIC:
2026 pd_idx = sector_div(stripe2, raid_disks);
2027 if (*dd_idx >= pd_idx)
2030 case ALGORITHM_LEFT_SYMMETRIC:
2031 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2032 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2034 case ALGORITHM_RIGHT_SYMMETRIC:
2035 pd_idx = sector_div(stripe2, raid_disks);
2036 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2038 case ALGORITHM_PARITY_0:
2042 case ALGORITHM_PARITY_N:
2043 pd_idx = data_disks;
2051 switch (algorithm) {
2052 case ALGORITHM_LEFT_ASYMMETRIC:
2053 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2054 qd_idx = pd_idx + 1;
2055 if (pd_idx == raid_disks-1) {
2056 (*dd_idx)++; /* Q D D D P */
2058 } else if (*dd_idx >= pd_idx)
2059 (*dd_idx) += 2; /* D D P Q D */
2061 case ALGORITHM_RIGHT_ASYMMETRIC:
2062 pd_idx = sector_div(stripe2, raid_disks);
2063 qd_idx = pd_idx + 1;
2064 if (pd_idx == raid_disks-1) {
2065 (*dd_idx)++; /* Q D D D P */
2067 } else if (*dd_idx >= pd_idx)
2068 (*dd_idx) += 2; /* D D P Q D */
2070 case ALGORITHM_LEFT_SYMMETRIC:
2071 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2072 qd_idx = (pd_idx + 1) % raid_disks;
2073 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2075 case ALGORITHM_RIGHT_SYMMETRIC:
2076 pd_idx = sector_div(stripe2, raid_disks);
2077 qd_idx = (pd_idx + 1) % raid_disks;
2078 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2081 case ALGORITHM_PARITY_0:
2086 case ALGORITHM_PARITY_N:
2087 pd_idx = data_disks;
2088 qd_idx = data_disks + 1;
2091 case ALGORITHM_ROTATING_ZERO_RESTART:
2092 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2093 * of blocks for computing Q is different.
2095 pd_idx = sector_div(stripe2, raid_disks);
2096 qd_idx = pd_idx + 1;
2097 if (pd_idx == raid_disks-1) {
2098 (*dd_idx)++; /* Q D D D P */
2100 } else if (*dd_idx >= pd_idx)
2101 (*dd_idx) += 2; /* D D P Q D */
2105 case ALGORITHM_ROTATING_N_RESTART:
2106 /* Same a left_asymmetric, by first stripe is
2107 * D D D P Q rather than
2111 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2112 qd_idx = pd_idx + 1;
2113 if (pd_idx == raid_disks-1) {
2114 (*dd_idx)++; /* Q D D D P */
2116 } else if (*dd_idx >= pd_idx)
2117 (*dd_idx) += 2; /* D D P Q D */
2121 case ALGORITHM_ROTATING_N_CONTINUE:
2122 /* Same as left_symmetric but Q is before P */
2123 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2124 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2125 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2129 case ALGORITHM_LEFT_ASYMMETRIC_6:
2130 /* RAID5 left_asymmetric, with Q on last device */
2131 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2132 if (*dd_idx >= pd_idx)
2134 qd_idx = raid_disks - 1;
2137 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2138 pd_idx = sector_div(stripe2, raid_disks-1);
2139 if (*dd_idx >= pd_idx)
2141 qd_idx = raid_disks - 1;
2144 case ALGORITHM_LEFT_SYMMETRIC_6:
2145 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2146 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2147 qd_idx = raid_disks - 1;
2150 case ALGORITHM_RIGHT_SYMMETRIC_6:
2151 pd_idx = sector_div(stripe2, raid_disks-1);
2152 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2153 qd_idx = raid_disks - 1;
2156 case ALGORITHM_PARITY_0_6:
2159 qd_idx = raid_disks - 1;
2169 sh->pd_idx = pd_idx;
2170 sh->qd_idx = qd_idx;
2171 sh->ddf_layout = ddf_layout;
2174 * Finally, compute the new sector number
2176 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2181 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2183 struct r5conf *conf = sh->raid_conf;
2184 int raid_disks = sh->disks;
2185 int data_disks = raid_disks - conf->max_degraded;
2186 sector_t new_sector = sh->sector, check;
2187 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2188 : conf->chunk_sectors;
2189 int algorithm = previous ? conf->prev_algo
2193 sector_t chunk_number;
2194 int dummy1, dd_idx = i;
2196 struct stripe_head sh2;
2199 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2200 stripe = new_sector;
2202 if (i == sh->pd_idx)
2204 switch(conf->level) {
2207 switch (algorithm) {
2208 case ALGORITHM_LEFT_ASYMMETRIC:
2209 case ALGORITHM_RIGHT_ASYMMETRIC:
2213 case ALGORITHM_LEFT_SYMMETRIC:
2214 case ALGORITHM_RIGHT_SYMMETRIC:
2217 i -= (sh->pd_idx + 1);
2219 case ALGORITHM_PARITY_0:
2222 case ALGORITHM_PARITY_N:
2229 if (i == sh->qd_idx)
2230 return 0; /* It is the Q disk */
2231 switch (algorithm) {
2232 case ALGORITHM_LEFT_ASYMMETRIC:
2233 case ALGORITHM_RIGHT_ASYMMETRIC:
2234 case ALGORITHM_ROTATING_ZERO_RESTART:
2235 case ALGORITHM_ROTATING_N_RESTART:
2236 if (sh->pd_idx == raid_disks-1)
2237 i--; /* Q D D D P */
2238 else if (i > sh->pd_idx)
2239 i -= 2; /* D D P Q D */
2241 case ALGORITHM_LEFT_SYMMETRIC:
2242 case ALGORITHM_RIGHT_SYMMETRIC:
2243 if (sh->pd_idx == raid_disks-1)
2244 i--; /* Q D D D P */
2249 i -= (sh->pd_idx + 2);
2252 case ALGORITHM_PARITY_0:
2255 case ALGORITHM_PARITY_N:
2257 case ALGORITHM_ROTATING_N_CONTINUE:
2258 /* Like left_symmetric, but P is before Q */
2259 if (sh->pd_idx == 0)
2260 i--; /* P D D D Q */
2265 i -= (sh->pd_idx + 1);
2268 case ALGORITHM_LEFT_ASYMMETRIC_6:
2269 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2273 case ALGORITHM_LEFT_SYMMETRIC_6:
2274 case ALGORITHM_RIGHT_SYMMETRIC_6:
2276 i += data_disks + 1;
2277 i -= (sh->pd_idx + 1);
2279 case ALGORITHM_PARITY_0_6:
2288 chunk_number = stripe * data_disks + i;
2289 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2291 check = raid5_compute_sector(conf, r_sector,
2292 previous, &dummy1, &sh2);
2293 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2294 || sh2.qd_idx != sh->qd_idx) {
2295 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2296 mdname(conf->mddev));
2304 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2305 int rcw, int expand)
2307 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2308 struct r5conf *conf = sh->raid_conf;
2309 int level = conf->level;
2312 /* if we are not expanding this is a proper write request, and
2313 * there will be bios with new data to be drained into the
2317 sh->reconstruct_state = reconstruct_state_drain_run;
2318 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2320 sh->reconstruct_state = reconstruct_state_run;
2322 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2324 for (i = disks; i--; ) {
2325 struct r5dev *dev = &sh->dev[i];
2328 set_bit(R5_LOCKED, &dev->flags);
2329 set_bit(R5_Wantdrain, &dev->flags);
2331 clear_bit(R5_UPTODATE, &dev->flags);
2335 if (s->locked + conf->max_degraded == disks)
2336 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2337 atomic_inc(&conf->pending_full_writes);
2340 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2341 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2343 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2344 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2345 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2346 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2348 for (i = disks; i--; ) {
2349 struct r5dev *dev = &sh->dev[i];
2354 (test_bit(R5_UPTODATE, &dev->flags) ||
2355 test_bit(R5_Wantcompute, &dev->flags))) {
2356 set_bit(R5_Wantdrain, &dev->flags);
2357 set_bit(R5_LOCKED, &dev->flags);
2358 clear_bit(R5_UPTODATE, &dev->flags);
2364 /* keep the parity disk(s) locked while asynchronous operations
2367 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2368 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2372 int qd_idx = sh->qd_idx;
2373 struct r5dev *dev = &sh->dev[qd_idx];
2375 set_bit(R5_LOCKED, &dev->flags);
2376 clear_bit(R5_UPTODATE, &dev->flags);
2380 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2381 __func__, (unsigned long long)sh->sector,
2382 s->locked, s->ops_request);
2386 * Each stripe/dev can have one or more bion attached.
2387 * toread/towrite point to the first in a chain.
2388 * The bi_next chain must be in order.
2390 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2393 struct r5conf *conf = sh->raid_conf;
2396 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2397 (unsigned long long)bi->bi_sector,
2398 (unsigned long long)sh->sector);
2401 * If several bio share a stripe. The bio bi_phys_segments acts as a
2402 * reference count to avoid race. The reference count should already be
2403 * increased before this function is called (for example, in
2404 * make_request()), so other bio sharing this stripe will not free the
2405 * stripe. If a stripe is owned by one stripe, the stripe lock will
2408 spin_lock_irq(&sh->stripe_lock);
2410 bip = &sh->dev[dd_idx].towrite;
2414 bip = &sh->dev[dd_idx].toread;
2415 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2416 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2418 bip = & (*bip)->bi_next;
2420 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2423 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2427 raid5_inc_bi_active_stripes(bi);
2430 /* check if page is covered */
2431 sector_t sector = sh->dev[dd_idx].sector;
2432 for (bi=sh->dev[dd_idx].towrite;
2433 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2434 bi && bi->bi_sector <= sector;
2435 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2436 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2437 sector = bi->bi_sector + (bi->bi_size>>9);
2439 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2440 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2443 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2444 (unsigned long long)(*bip)->bi_sector,
2445 (unsigned long long)sh->sector, dd_idx);
2446 spin_unlock_irq(&sh->stripe_lock);
2448 if (conf->mddev->bitmap && firstwrite) {
2449 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2451 sh->bm_seq = conf->seq_flush+1;
2452 set_bit(STRIPE_BIT_DELAY, &sh->state);
2457 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2458 spin_unlock_irq(&sh->stripe_lock);
2462 static void end_reshape(struct r5conf *conf);
2464 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2465 struct stripe_head *sh)
2467 int sectors_per_chunk =
2468 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2470 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2471 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2473 raid5_compute_sector(conf,
2474 stripe * (disks - conf->max_degraded)
2475 *sectors_per_chunk + chunk_offset,
2481 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2482 struct stripe_head_state *s, int disks,
2483 struct bio **return_bi)
2486 for (i = disks; i--; ) {
2490 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2491 struct md_rdev *rdev;
2493 rdev = rcu_dereference(conf->disks[i].rdev);
2494 if (rdev && test_bit(In_sync, &rdev->flags))
2495 atomic_inc(&rdev->nr_pending);
2500 if (!rdev_set_badblocks(
2504 md_error(conf->mddev, rdev);
2505 rdev_dec_pending(rdev, conf->mddev);
2508 spin_lock_irq(&sh->stripe_lock);
2509 /* fail all writes first */
2510 bi = sh->dev[i].towrite;
2511 sh->dev[i].towrite = NULL;
2512 spin_unlock_irq(&sh->stripe_lock);
2516 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2517 wake_up(&conf->wait_for_overlap);
2519 while (bi && bi->bi_sector <
2520 sh->dev[i].sector + STRIPE_SECTORS) {
2521 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2522 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2523 if (!raid5_dec_bi_active_stripes(bi)) {
2524 md_write_end(conf->mddev);
2525 bi->bi_next = *return_bi;
2531 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2532 STRIPE_SECTORS, 0, 0);
2534 /* and fail all 'written' */
2535 bi = sh->dev[i].written;
2536 sh->dev[i].written = NULL;
2537 if (bi) bitmap_end = 1;
2538 while (bi && bi->bi_sector <
2539 sh->dev[i].sector + STRIPE_SECTORS) {
2540 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2541 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2542 if (!raid5_dec_bi_active_stripes(bi)) {
2543 md_write_end(conf->mddev);
2544 bi->bi_next = *return_bi;
2550 /* fail any reads if this device is non-operational and
2551 * the data has not reached the cache yet.
2553 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2554 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2555 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2556 spin_lock_irq(&sh->stripe_lock);
2557 bi = sh->dev[i].toread;
2558 sh->dev[i].toread = NULL;
2559 spin_unlock_irq(&sh->stripe_lock);
2560 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2561 wake_up(&conf->wait_for_overlap);
2562 while (bi && bi->bi_sector <
2563 sh->dev[i].sector + STRIPE_SECTORS) {
2564 struct bio *nextbi =
2565 r5_next_bio(bi, sh->dev[i].sector);
2566 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2567 if (!raid5_dec_bi_active_stripes(bi)) {
2568 bi->bi_next = *return_bi;
2575 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2576 STRIPE_SECTORS, 0, 0);
2577 /* If we were in the middle of a write the parity block might
2578 * still be locked - so just clear all R5_LOCKED flags
2580 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2583 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2584 if (atomic_dec_and_test(&conf->pending_full_writes))
2585 md_wakeup_thread(conf->mddev->thread);
2589 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2590 struct stripe_head_state *s)
2595 clear_bit(STRIPE_SYNCING, &sh->state);
2598 /* There is nothing more to do for sync/check/repair.
2599 * Don't even need to abort as that is handled elsewhere
2600 * if needed, and not always wanted e.g. if there is a known
2602 * For recover/replace we need to record a bad block on all
2603 * non-sync devices, or abort the recovery
2605 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2606 /* During recovery devices cannot be removed, so
2607 * locking and refcounting of rdevs is not needed
2609 for (i = 0; i < conf->raid_disks; i++) {
2610 struct md_rdev *rdev = conf->disks[i].rdev;
2612 && !test_bit(Faulty, &rdev->flags)
2613 && !test_bit(In_sync, &rdev->flags)
2614 && !rdev_set_badblocks(rdev, sh->sector,
2617 rdev = conf->disks[i].replacement;
2619 && !test_bit(Faulty, &rdev->flags)
2620 && !test_bit(In_sync, &rdev->flags)
2621 && !rdev_set_badblocks(rdev, sh->sector,
2626 conf->recovery_disabled =
2627 conf->mddev->recovery_disabled;
2629 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2632 static int want_replace(struct stripe_head *sh, int disk_idx)
2634 struct md_rdev *rdev;
2636 /* Doing recovery so rcu locking not required */
2637 rdev = sh->raid_conf->disks[disk_idx].replacement;
2639 && !test_bit(Faulty, &rdev->flags)
2640 && !test_bit(In_sync, &rdev->flags)
2641 && (rdev->recovery_offset <= sh->sector
2642 || rdev->mddev->recovery_cp <= sh->sector))
2648 /* fetch_block - checks the given member device to see if its data needs
2649 * to be read or computed to satisfy a request.
2651 * Returns 1 when no more member devices need to be checked, otherwise returns
2652 * 0 to tell the loop in handle_stripe_fill to continue
2654 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2655 int disk_idx, int disks)
2657 struct r5dev *dev = &sh->dev[disk_idx];
2658 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2659 &sh->dev[s->failed_num[1]] };
2661 /* is the data in this block needed, and can we get it? */
2662 if (!test_bit(R5_LOCKED, &dev->flags) &&
2663 !test_bit(R5_UPTODATE, &dev->flags) &&
2665 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2666 s->syncing || s->expanding ||
2667 (s->replacing && want_replace(sh, disk_idx)) ||
2668 (s->failed >= 1 && fdev[0]->toread) ||
2669 (s->failed >= 2 && fdev[1]->toread) ||
2670 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2671 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2672 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2673 /* we would like to get this block, possibly by computing it,
2674 * otherwise read it if the backing disk is insync
2676 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2677 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2678 if ((s->uptodate == disks - 1) &&
2679 (s->failed && (disk_idx == s->failed_num[0] ||
2680 disk_idx == s->failed_num[1]))) {
2681 /* have disk failed, and we're requested to fetch it;
2684 pr_debug("Computing stripe %llu block %d\n",
2685 (unsigned long long)sh->sector, disk_idx);
2686 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2687 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2688 set_bit(R5_Wantcompute, &dev->flags);
2689 sh->ops.target = disk_idx;
2690 sh->ops.target2 = -1; /* no 2nd target */
2692 /* Careful: from this point on 'uptodate' is in the eye
2693 * of raid_run_ops which services 'compute' operations
2694 * before writes. R5_Wantcompute flags a block that will
2695 * be R5_UPTODATE by the time it is needed for a
2696 * subsequent operation.
2700 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2701 /* Computing 2-failure is *very* expensive; only
2702 * do it if failed >= 2
2705 for (other = disks; other--; ) {
2706 if (other == disk_idx)
2708 if (!test_bit(R5_UPTODATE,
2709 &sh->dev[other].flags))
2713 pr_debug("Computing stripe %llu blocks %d,%d\n",
2714 (unsigned long long)sh->sector,
2716 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2717 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2718 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2719 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2720 sh->ops.target = disk_idx;
2721 sh->ops.target2 = other;
2725 } else if (test_bit(R5_Insync, &dev->flags)) {
2726 set_bit(R5_LOCKED, &dev->flags);
2727 set_bit(R5_Wantread, &dev->flags);
2729 pr_debug("Reading block %d (sync=%d)\n",
2730 disk_idx, s->syncing);
2738 * handle_stripe_fill - read or compute data to satisfy pending requests.
2740 static void handle_stripe_fill(struct stripe_head *sh,
2741 struct stripe_head_state *s,
2746 /* look for blocks to read/compute, skip this if a compute
2747 * is already in flight, or if the stripe contents are in the
2748 * midst of changing due to a write
2750 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2751 !sh->reconstruct_state)
2752 for (i = disks; i--; )
2753 if (fetch_block(sh, s, i, disks))
2755 set_bit(STRIPE_HANDLE, &sh->state);
2759 /* handle_stripe_clean_event
2760 * any written block on an uptodate or failed drive can be returned.
2761 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2762 * never LOCKED, so we don't need to test 'failed' directly.
2764 static void handle_stripe_clean_event(struct r5conf *conf,
2765 struct stripe_head *sh, int disks, struct bio **return_bi)
2770 for (i = disks; i--; )
2771 if (sh->dev[i].written) {
2773 if (!test_bit(R5_LOCKED, &dev->flags) &&
2774 (test_bit(R5_UPTODATE, &dev->flags) ||
2775 test_bit(R5_Discard, &dev->flags))) {
2776 /* We can return any write requests */
2777 struct bio *wbi, *wbi2;
2778 pr_debug("Return write for disc %d\n", i);
2779 if (test_and_clear_bit(R5_Discard, &dev->flags))
2780 clear_bit(R5_UPTODATE, &dev->flags);
2782 dev->written = NULL;
2783 while (wbi && wbi->bi_sector <
2784 dev->sector + STRIPE_SECTORS) {
2785 wbi2 = r5_next_bio(wbi, dev->sector);
2786 if (!raid5_dec_bi_active_stripes(wbi)) {
2787 md_write_end(conf->mddev);
2788 wbi->bi_next = *return_bi;
2793 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2795 !test_bit(STRIPE_DEGRADED, &sh->state),
2798 } else if (test_bit(R5_Discard, &sh->dev[i].flags))
2799 clear_bit(R5_Discard, &sh->dev[i].flags);
2801 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2802 if (atomic_dec_and_test(&conf->pending_full_writes))
2803 md_wakeup_thread(conf->mddev->thread);
2806 static void handle_stripe_dirtying(struct r5conf *conf,
2807 struct stripe_head *sh,
2808 struct stripe_head_state *s,
2811 int rmw = 0, rcw = 0, i;
2812 sector_t recovery_cp = conf->mddev->recovery_cp;
2814 /* RAID6 requires 'rcw' in current implementation.
2815 * Otherwise, check whether resync is now happening or should start.
2816 * If yes, then the array is dirty (after unclean shutdown or
2817 * initial creation), so parity in some stripes might be inconsistent.
2818 * In this case, we need to always do reconstruct-write, to ensure
2819 * that in case of drive failure or read-error correction, we
2820 * generate correct data from the parity.
2822 if (conf->max_degraded == 2 ||
2823 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2824 /* Calculate the real rcw later - for now make it
2825 * look like rcw is cheaper
2828 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2829 conf->max_degraded, (unsigned long long)recovery_cp,
2830 (unsigned long long)sh->sector);
2831 } else for (i = disks; i--; ) {
2832 /* would I have to read this buffer for read_modify_write */
2833 struct r5dev *dev = &sh->dev[i];
2834 if ((dev->towrite || i == sh->pd_idx) &&
2835 !test_bit(R5_LOCKED, &dev->flags) &&
2836 !(test_bit(R5_UPTODATE, &dev->flags) ||
2837 test_bit(R5_Wantcompute, &dev->flags))) {
2838 if (test_bit(R5_Insync, &dev->flags))
2841 rmw += 2*disks; /* cannot read it */
2843 /* Would I have to read this buffer for reconstruct_write */
2844 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2845 !test_bit(R5_LOCKED, &dev->flags) &&
2846 !(test_bit(R5_UPTODATE, &dev->flags) ||
2847 test_bit(R5_Wantcompute, &dev->flags))) {
2848 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2853 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2854 (unsigned long long)sh->sector, rmw, rcw);
2855 set_bit(STRIPE_HANDLE, &sh->state);
2856 if (rmw < rcw && rmw > 0)
2857 /* prefer read-modify-write, but need to get some data */
2858 for (i = disks; i--; ) {
2859 struct r5dev *dev = &sh->dev[i];
2860 if ((dev->towrite || i == sh->pd_idx) &&
2861 !test_bit(R5_LOCKED, &dev->flags) &&
2862 !(test_bit(R5_UPTODATE, &dev->flags) ||
2863 test_bit(R5_Wantcompute, &dev->flags)) &&
2864 test_bit(R5_Insync, &dev->flags)) {
2866 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2867 pr_debug("Read_old block "
2868 "%d for r-m-w\n", i);
2869 set_bit(R5_LOCKED, &dev->flags);
2870 set_bit(R5_Wantread, &dev->flags);
2873 set_bit(STRIPE_DELAYED, &sh->state);
2874 set_bit(STRIPE_HANDLE, &sh->state);
2878 if (rcw <= rmw && rcw > 0) {
2879 /* want reconstruct write, but need to get some data */
2881 for (i = disks; i--; ) {
2882 struct r5dev *dev = &sh->dev[i];
2883 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2884 i != sh->pd_idx && i != sh->qd_idx &&
2885 !test_bit(R5_LOCKED, &dev->flags) &&
2886 !(test_bit(R5_UPTODATE, &dev->flags) ||
2887 test_bit(R5_Wantcompute, &dev->flags))) {
2889 if (!test_bit(R5_Insync, &dev->flags))
2890 continue; /* it's a failed drive */
2892 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2893 pr_debug("Read_old block "
2894 "%d for Reconstruct\n", i);
2895 set_bit(R5_LOCKED, &dev->flags);
2896 set_bit(R5_Wantread, &dev->flags);
2899 set_bit(STRIPE_DELAYED, &sh->state);
2900 set_bit(STRIPE_HANDLE, &sh->state);
2905 /* now if nothing is locked, and if we have enough data,
2906 * we can start a write request
2908 /* since handle_stripe can be called at any time we need to handle the
2909 * case where a compute block operation has been submitted and then a
2910 * subsequent call wants to start a write request. raid_run_ops only
2911 * handles the case where compute block and reconstruct are requested
2912 * simultaneously. If this is not the case then new writes need to be
2913 * held off until the compute completes.
2915 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2916 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2917 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2918 schedule_reconstruction(sh, s, rcw == 0, 0);
2921 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2922 struct stripe_head_state *s, int disks)
2924 struct r5dev *dev = NULL;
2926 set_bit(STRIPE_HANDLE, &sh->state);
2928 switch (sh->check_state) {
2929 case check_state_idle:
2930 /* start a new check operation if there are no failures */
2931 if (s->failed == 0) {
2932 BUG_ON(s->uptodate != disks);
2933 sh->check_state = check_state_run;
2934 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2935 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2939 dev = &sh->dev[s->failed_num[0]];
2941 case check_state_compute_result:
2942 sh->check_state = check_state_idle;
2944 dev = &sh->dev[sh->pd_idx];
2946 /* check that a write has not made the stripe insync */
2947 if (test_bit(STRIPE_INSYNC, &sh->state))
2950 /* either failed parity check, or recovery is happening */
2951 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2952 BUG_ON(s->uptodate != disks);
2954 set_bit(R5_LOCKED, &dev->flags);
2956 set_bit(R5_Wantwrite, &dev->flags);
2958 clear_bit(STRIPE_DEGRADED, &sh->state);
2959 set_bit(STRIPE_INSYNC, &sh->state);
2961 case check_state_run:
2962 break; /* we will be called again upon completion */
2963 case check_state_check_result:
2964 sh->check_state = check_state_idle;
2966 /* if a failure occurred during the check operation, leave
2967 * STRIPE_INSYNC not set and let the stripe be handled again
2972 /* handle a successful check operation, if parity is correct
2973 * we are done. Otherwise update the mismatch count and repair
2974 * parity if !MD_RECOVERY_CHECK
2976 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2977 /* parity is correct (on disc,
2978 * not in buffer any more)
2980 set_bit(STRIPE_INSYNC, &sh->state);
2982 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
2983 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2984 /* don't try to repair!! */
2985 set_bit(STRIPE_INSYNC, &sh->state);
2987 sh->check_state = check_state_compute_run;
2988 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2989 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2990 set_bit(R5_Wantcompute,
2991 &sh->dev[sh->pd_idx].flags);
2992 sh->ops.target = sh->pd_idx;
2993 sh->ops.target2 = -1;
2998 case check_state_compute_run:
3001 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3002 __func__, sh->check_state,
3003 (unsigned long long) sh->sector);
3009 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3010 struct stripe_head_state *s,
3013 int pd_idx = sh->pd_idx;
3014 int qd_idx = sh->qd_idx;
3017 set_bit(STRIPE_HANDLE, &sh->state);
3019 BUG_ON(s->failed > 2);
3021 /* Want to check and possibly repair P and Q.
3022 * However there could be one 'failed' device, in which
3023 * case we can only check one of them, possibly using the
3024 * other to generate missing data
3027 switch (sh->check_state) {
3028 case check_state_idle:
3029 /* start a new check operation if there are < 2 failures */
3030 if (s->failed == s->q_failed) {
3031 /* The only possible failed device holds Q, so it
3032 * makes sense to check P (If anything else were failed,
3033 * we would have used P to recreate it).
3035 sh->check_state = check_state_run;
3037 if (!s->q_failed && s->failed < 2) {
3038 /* Q is not failed, and we didn't use it to generate
3039 * anything, so it makes sense to check it
3041 if (sh->check_state == check_state_run)
3042 sh->check_state = check_state_run_pq;
3044 sh->check_state = check_state_run_q;
3047 /* discard potentially stale zero_sum_result */
3048 sh->ops.zero_sum_result = 0;
3050 if (sh->check_state == check_state_run) {
3051 /* async_xor_zero_sum destroys the contents of P */
3052 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3055 if (sh->check_state >= check_state_run &&
3056 sh->check_state <= check_state_run_pq) {
3057 /* async_syndrome_zero_sum preserves P and Q, so
3058 * no need to mark them !uptodate here
3060 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3064 /* we have 2-disk failure */
3065 BUG_ON(s->failed != 2);
3067 case check_state_compute_result:
3068 sh->check_state = check_state_idle;
3070 /* check that a write has not made the stripe insync */
3071 if (test_bit(STRIPE_INSYNC, &sh->state))
3074 /* now write out any block on a failed drive,
3075 * or P or Q if they were recomputed
3077 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3078 if (s->failed == 2) {
3079 dev = &sh->dev[s->failed_num[1]];
3081 set_bit(R5_LOCKED, &dev->flags);
3082 set_bit(R5_Wantwrite, &dev->flags);
3084 if (s->failed >= 1) {
3085 dev = &sh->dev[s->failed_num[0]];
3087 set_bit(R5_LOCKED, &dev->flags);
3088 set_bit(R5_Wantwrite, &dev->flags);
3090 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3091 dev = &sh->dev[pd_idx];
3093 set_bit(R5_LOCKED, &dev->flags);
3094 set_bit(R5_Wantwrite, &dev->flags);
3096 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3097 dev = &sh->dev[qd_idx];
3099 set_bit(R5_LOCKED, &dev->flags);
3100 set_bit(R5_Wantwrite, &dev->flags);
3102 clear_bit(STRIPE_DEGRADED, &sh->state);
3104 set_bit(STRIPE_INSYNC, &sh->state);
3106 case check_state_run:
3107 case check_state_run_q:
3108 case check_state_run_pq:
3109 break; /* we will be called again upon completion */
3110 case check_state_check_result:
3111 sh->check_state = check_state_idle;
3113 /* handle a successful check operation, if parity is correct
3114 * we are done. Otherwise update the mismatch count and repair
3115 * parity if !MD_RECOVERY_CHECK
3117 if (sh->ops.zero_sum_result == 0) {
3118 /* both parities are correct */
3120 set_bit(STRIPE_INSYNC, &sh->state);
3122 /* in contrast to the raid5 case we can validate
3123 * parity, but still have a failure to write
3126 sh->check_state = check_state_compute_result;
3127 /* Returning at this point means that we may go
3128 * off and bring p and/or q uptodate again so
3129 * we make sure to check zero_sum_result again
3130 * to verify if p or q need writeback
3134 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3135 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3136 /* don't try to repair!! */
3137 set_bit(STRIPE_INSYNC, &sh->state);
3139 int *target = &sh->ops.target;
3141 sh->ops.target = -1;
3142 sh->ops.target2 = -1;
3143 sh->check_state = check_state_compute_run;
3144 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3145 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3146 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3147 set_bit(R5_Wantcompute,
3148 &sh->dev[pd_idx].flags);
3150 target = &sh->ops.target2;
3153 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3154 set_bit(R5_Wantcompute,
3155 &sh->dev[qd_idx].flags);
3162 case check_state_compute_run:
3165 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3166 __func__, sh->check_state,
3167 (unsigned long long) sh->sector);
3172 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3176 /* We have read all the blocks in this stripe and now we need to
3177 * copy some of them into a target stripe for expand.
3179 struct dma_async_tx_descriptor *tx = NULL;
3180 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3181 for (i = 0; i < sh->disks; i++)
3182 if (i != sh->pd_idx && i != sh->qd_idx) {
3184 struct stripe_head *sh2;
3185 struct async_submit_ctl submit;
3187 sector_t bn = compute_blocknr(sh, i, 1);
3188 sector_t s = raid5_compute_sector(conf, bn, 0,
3190 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3192 /* so far only the early blocks of this stripe
3193 * have been requested. When later blocks
3194 * get requested, we will try again
3197 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3198 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3199 /* must have already done this block */
3200 release_stripe(sh2);
3204 /* place all the copies on one channel */
3205 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3206 tx = async_memcpy(sh2->dev[dd_idx].page,
3207 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3210 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3211 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3212 for (j = 0; j < conf->raid_disks; j++)
3213 if (j != sh2->pd_idx &&
3215 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3217 if (j == conf->raid_disks) {
3218 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3219 set_bit(STRIPE_HANDLE, &sh2->state);
3221 release_stripe(sh2);
3224 /* done submitting copies, wait for them to complete */
3227 dma_wait_for_async_tx(tx);
3232 * handle_stripe - do things to a stripe.
3234 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3235 * state of various bits to see what needs to be done.
3237 * return some read requests which now have data
3238 * return some write requests which are safely on storage
3239 * schedule a read on some buffers
3240 * schedule a write of some buffers
3241 * return confirmation of parity correctness
3245 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3247 struct r5conf *conf = sh->raid_conf;
3248 int disks = sh->disks;
3251 int do_recovery = 0;
3253 memset(s, 0, sizeof(*s));
3255 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3256 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3257 s->failed_num[0] = -1;
3258 s->failed_num[1] = -1;
3260 /* Now to look around and see what can be done */
3262 for (i=disks; i--; ) {
3263 struct md_rdev *rdev;
3270 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3272 dev->toread, dev->towrite, dev->written);
3273 /* maybe we can reply to a read
3275 * new wantfill requests are only permitted while
3276 * ops_complete_biofill is guaranteed to be inactive
3278 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3279 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3280 set_bit(R5_Wantfill, &dev->flags);
3282 /* now count some things */
3283 if (test_bit(R5_LOCKED, &dev->flags))
3285 if (test_bit(R5_UPTODATE, &dev->flags))
3287 if (test_bit(R5_Wantcompute, &dev->flags)) {
3289 BUG_ON(s->compute > 2);
3292 if (test_bit(R5_Wantfill, &dev->flags))
3294 else if (dev->toread)
3298 if (!test_bit(R5_OVERWRITE, &dev->flags))
3303 /* Prefer to use the replacement for reads, but only
3304 * if it is recovered enough and has no bad blocks.
3306 rdev = rcu_dereference(conf->disks[i].replacement);
3307 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3308 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3309 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3310 &first_bad, &bad_sectors))
3311 set_bit(R5_ReadRepl, &dev->flags);
3314 set_bit(R5_NeedReplace, &dev->flags);
3315 rdev = rcu_dereference(conf->disks[i].rdev);
3316 clear_bit(R5_ReadRepl, &dev->flags);
3318 if (rdev && test_bit(Faulty, &rdev->flags))
3321 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3322 &first_bad, &bad_sectors);
3323 if (s->blocked_rdev == NULL
3324 && (test_bit(Blocked, &rdev->flags)
3327 set_bit(BlockedBadBlocks,
3329 s->blocked_rdev = rdev;
3330 atomic_inc(&rdev->nr_pending);
3333 clear_bit(R5_Insync, &dev->flags);
3337 /* also not in-sync */
3338 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3339 test_bit(R5_UPTODATE, &dev->flags)) {
3340 /* treat as in-sync, but with a read error
3341 * which we can now try to correct
3343 set_bit(R5_Insync, &dev->flags);
3344 set_bit(R5_ReadError, &dev->flags);
3346 } else if (test_bit(In_sync, &rdev->flags))
3347 set_bit(R5_Insync, &dev->flags);
3348 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3349 /* in sync if before recovery_offset */
3350 set_bit(R5_Insync, &dev->flags);
3351 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3352 test_bit(R5_Expanded, &dev->flags))
3353 /* If we've reshaped into here, we assume it is Insync.
3354 * We will shortly update recovery_offset to make
3357 set_bit(R5_Insync, &dev->flags);
3359 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3360 /* This flag does not apply to '.replacement'
3361 * only to .rdev, so make sure to check that*/
3362 struct md_rdev *rdev2 = rcu_dereference(
3363 conf->disks[i].rdev);
3365 clear_bit(R5_Insync, &dev->flags);
3366 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3367 s->handle_bad_blocks = 1;
3368 atomic_inc(&rdev2->nr_pending);
3370 clear_bit(R5_WriteError, &dev->flags);
3372 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3373 /* This flag does not apply to '.replacement'
3374 * only to .rdev, so make sure to check that*/
3375 struct md_rdev *rdev2 = rcu_dereference(
3376 conf->disks[i].rdev);
3377 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3378 s->handle_bad_blocks = 1;
3379 atomic_inc(&rdev2->nr_pending);
3381 clear_bit(R5_MadeGood, &dev->flags);
3383 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3384 struct md_rdev *rdev2 = rcu_dereference(
3385 conf->disks[i].replacement);
3386 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3387 s->handle_bad_blocks = 1;
3388 atomic_inc(&rdev2->nr_pending);
3390 clear_bit(R5_MadeGoodRepl, &dev->flags);
3392 if (!test_bit(R5_Insync, &dev->flags)) {
3393 /* The ReadError flag will just be confusing now */
3394 clear_bit(R5_ReadError, &dev->flags);
3395 clear_bit(R5_ReWrite, &dev->flags);
3397 if (test_bit(R5_ReadError, &dev->flags))
3398 clear_bit(R5_Insync, &dev->flags);
3399 if (!test_bit(R5_Insync, &dev->flags)) {
3401 s->failed_num[s->failed] = i;
3403 if (rdev && !test_bit(Faulty, &rdev->flags))
3407 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3408 /* If there is a failed device being replaced,
3409 * we must be recovering.
3410 * else if we are after recovery_cp, we must be syncing
3411 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3412 * else we can only be replacing
3413 * sync and recovery both need to read all devices, and so
3414 * use the same flag.
3417 sh->sector >= conf->mddev->recovery_cp ||
3418 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3426 static void handle_stripe(struct stripe_head *sh)
3428 struct stripe_head_state s;
3429 struct r5conf *conf = sh->raid_conf;
3432 int disks = sh->disks;
3433 struct r5dev *pdev, *qdev;
3435 clear_bit(STRIPE_HANDLE, &sh->state);
3436 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3437 /* already being handled, ensure it gets handled
3438 * again when current action finishes */
3439 set_bit(STRIPE_HANDLE, &sh->state);
3443 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3444 set_bit(STRIPE_SYNCING, &sh->state);
3445 clear_bit(STRIPE_INSYNC, &sh->state);
3447 clear_bit(STRIPE_DELAYED, &sh->state);
3449 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3450 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3451 (unsigned long long)sh->sector, sh->state,
3452 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3453 sh->check_state, sh->reconstruct_state);
3455 analyse_stripe(sh, &s);
3457 if (s.handle_bad_blocks) {
3458 set_bit(STRIPE_HANDLE, &sh->state);
3462 if (unlikely(s.blocked_rdev)) {
3463 if (s.syncing || s.expanding || s.expanded ||
3464 s.replacing || s.to_write || s.written) {
3465 set_bit(STRIPE_HANDLE, &sh->state);
3468 /* There is nothing for the blocked_rdev to block */
3469 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3470 s.blocked_rdev = NULL;
3473 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3474 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3475 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3478 pr_debug("locked=%d uptodate=%d to_read=%d"
3479 " to_write=%d failed=%d failed_num=%d,%d\n",
3480 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3481 s.failed_num[0], s.failed_num[1]);
3482 /* check if the array has lost more than max_degraded devices and,
3483 * if so, some requests might need to be failed.
3485 if (s.failed > conf->max_degraded) {
3486 sh->check_state = 0;
3487 sh->reconstruct_state = 0;
3488 if (s.to_read+s.to_write+s.written)
3489 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3490 if (s.syncing + s.replacing)
3491 handle_failed_sync(conf, sh, &s);
3494 /* Now we check to see if any write operations have recently
3498 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3500 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3501 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3502 sh->reconstruct_state = reconstruct_state_idle;
3504 /* All the 'written' buffers and the parity block are ready to
3505 * be written back to disk
3507 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3508 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3509 BUG_ON(sh->qd_idx >= 0 &&
3510 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3511 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3512 for (i = disks; i--; ) {
3513 struct r5dev *dev = &sh->dev[i];
3514 if (test_bit(R5_LOCKED, &dev->flags) &&
3515 (i == sh->pd_idx || i == sh->qd_idx ||
3517 pr_debug("Writing block %d\n", i);
3518 set_bit(R5_Wantwrite, &dev->flags);
3521 if (!test_bit(R5_Insync, &dev->flags) ||
3522 ((i == sh->pd_idx || i == sh->qd_idx) &&
3524 set_bit(STRIPE_INSYNC, &sh->state);
3527 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3528 s.dec_preread_active = 1;
3532 * might be able to return some write requests if the parity blocks
3533 * are safe, or on a failed drive
3535 pdev = &sh->dev[sh->pd_idx];
3536 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3537 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3538 qdev = &sh->dev[sh->qd_idx];
3539 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3540 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3544 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3545 && !test_bit(R5_LOCKED, &pdev->flags)
3546 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3547 test_bit(R5_Discard, &pdev->flags))))) &&
3548 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3549 && !test_bit(R5_LOCKED, &qdev->flags)
3550 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3551 test_bit(R5_Discard, &qdev->flags))))))
3552 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3554 /* Now we might consider reading some blocks, either to check/generate
3555 * parity, or to satisfy requests
3556 * or to load a block that is being partially written.
3558 if (s.to_read || s.non_overwrite
3559 || (conf->level == 6 && s.to_write && s.failed)
3560 || (s.syncing && (s.uptodate + s.compute < disks))
3563 handle_stripe_fill(sh, &s, disks);
3565 /* Now to consider new write requests and what else, if anything
3566 * should be read. We do not handle new writes when:
3567 * 1/ A 'write' operation (copy+xor) is already in flight.
3568 * 2/ A 'check' operation is in flight, as it may clobber the parity
3571 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3572 handle_stripe_dirtying(conf, sh, &s, disks);
3574 /* maybe we need to check and possibly fix the parity for this stripe
3575 * Any reads will already have been scheduled, so we just see if enough
3576 * data is available. The parity check is held off while parity
3577 * dependent operations are in flight.
3579 if (sh->check_state ||
3580 (s.syncing && s.locked == 0 &&
3581 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3582 !test_bit(STRIPE_INSYNC, &sh->state))) {
3583 if (conf->level == 6)
3584 handle_parity_checks6(conf, sh, &s, disks);
3586 handle_parity_checks5(conf, sh, &s, disks);
3589 if (s.replacing && s.locked == 0
3590 && !test_bit(STRIPE_INSYNC, &sh->state)) {
3591 /* Write out to replacement devices where possible */
3592 for (i = 0; i < conf->raid_disks; i++)
3593 if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3594 test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3595 set_bit(R5_WantReplace, &sh->dev[i].flags);
3596 set_bit(R5_LOCKED, &sh->dev[i].flags);
3599 set_bit(STRIPE_INSYNC, &sh->state);
3601 if ((s.syncing || s.replacing) && s.locked == 0 &&
3602 test_bit(STRIPE_INSYNC, &sh->state)) {
3603 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3604 clear_bit(STRIPE_SYNCING, &sh->state);
3607 /* If the failed drives are just a ReadError, then we might need
3608 * to progress the repair/check process
3610 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3611 for (i = 0; i < s.failed; i++) {
3612 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3613 if (test_bit(R5_ReadError, &dev->flags)
3614 && !test_bit(R5_LOCKED, &dev->flags)
3615 && test_bit(R5_UPTODATE, &dev->flags)
3617 if (!test_bit(R5_ReWrite, &dev->flags)) {
3618 set_bit(R5_Wantwrite, &dev->flags);
3619 set_bit(R5_ReWrite, &dev->flags);
3620 set_bit(R5_LOCKED, &dev->flags);
3623 /* let's read it back */
3624 set_bit(R5_Wantread, &dev->flags);
3625 set_bit(R5_LOCKED, &dev->flags);
3632 /* Finish reconstruct operations initiated by the expansion process */
3633 if (sh->reconstruct_state == reconstruct_state_result) {
3634 struct stripe_head *sh_src
3635 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3636 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3637 /* sh cannot be written until sh_src has been read.
3638 * so arrange for sh to be delayed a little
3640 set_bit(STRIPE_DELAYED, &sh->state);
3641 set_bit(STRIPE_HANDLE, &sh->state);
3642 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3644 atomic_inc(&conf->preread_active_stripes);
3645 release_stripe(sh_src);
3649 release_stripe(sh_src);
3651 sh->reconstruct_state = reconstruct_state_idle;
3652 clear_bit(STRIPE_EXPANDING, &sh->state);
3653 for (i = conf->raid_disks; i--; ) {
3654 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3655 set_bit(R5_LOCKED, &sh->dev[i].flags);
3660 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3661 !sh->reconstruct_state) {
3662 /* Need to write out all blocks after computing parity */
3663 sh->disks = conf->raid_disks;
3664 stripe_set_idx(sh->sector, conf, 0, sh);
3665 schedule_reconstruction(sh, &s, 1, 1);
3666 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3667 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3668 atomic_dec(&conf->reshape_stripes);
3669 wake_up(&conf->wait_for_overlap);
3670 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3673 if (s.expanding && s.locked == 0 &&
3674 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3675 handle_stripe_expansion(conf, sh);
3678 /* wait for this device to become unblocked */
3679 if (unlikely(s.blocked_rdev)) {
3680 if (conf->mddev->external)
3681 md_wait_for_blocked_rdev(s.blocked_rdev,
3684 /* Internal metadata will immediately
3685 * be written by raid5d, so we don't
3686 * need to wait here.
3688 rdev_dec_pending(s.blocked_rdev,
3692 if (s.handle_bad_blocks)
3693 for (i = disks; i--; ) {
3694 struct md_rdev *rdev;
3695 struct r5dev *dev = &sh->dev[i];
3696 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3697 /* We own a safe reference to the rdev */
3698 rdev = conf->disks[i].rdev;
3699 if (!rdev_set_badblocks(rdev, sh->sector,
3701 md_error(conf->mddev, rdev);
3702 rdev_dec_pending(rdev, conf->mddev);
3704 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3705 rdev = conf->disks[i].rdev;
3706 rdev_clear_badblocks(rdev, sh->sector,
3708 rdev_dec_pending(rdev, conf->mddev);
3710 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3711 rdev = conf->disks[i].replacement;
3713 /* rdev have been moved down */
3714 rdev = conf->disks[i].rdev;
3715 rdev_clear_badblocks(rdev, sh->sector,
3717 rdev_dec_pending(rdev, conf->mddev);
3722 raid_run_ops(sh, s.ops_request);
3726 if (s.dec_preread_active) {
3727 /* We delay this until after ops_run_io so that if make_request
3728 * is waiting on a flush, it won't continue until the writes
3729 * have actually been submitted.
3731 atomic_dec(&conf->preread_active_stripes);
3732 if (atomic_read(&conf->preread_active_stripes) <
3734 md_wakeup_thread(conf->mddev->thread);
3737 return_io(s.return_bi);
3739 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3742 static void raid5_activate_delayed(struct r5conf *conf)
3744 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3745 while (!list_empty(&conf->delayed_list)) {
3746 struct list_head *l = conf->delayed_list.next;
3747 struct stripe_head *sh;
3748 sh = list_entry(l, struct stripe_head, lru);
3750 clear_bit(STRIPE_DELAYED, &sh->state);
3751 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3752 atomic_inc(&conf->preread_active_stripes);
3753 list_add_tail(&sh->lru, &conf->hold_list);
3758 static void activate_bit_delay(struct r5conf *conf)
3760 /* device_lock is held */
3761 struct list_head head;
3762 list_add(&head, &conf->bitmap_list);
3763 list_del_init(&conf->bitmap_list);
3764 while (!list_empty(&head)) {
3765 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3766 list_del_init(&sh->lru);
3767 atomic_inc(&sh->count);
3768 __release_stripe(conf, sh);
3772 int md_raid5_congested(struct mddev *mddev, int bits)
3774 struct r5conf *conf = mddev->private;
3776 /* No difference between reads and writes. Just check
3777 * how busy the stripe_cache is
3780 if (conf->inactive_blocked)
3784 if (list_empty_careful(&conf->inactive_list))
3789 EXPORT_SYMBOL_GPL(md_raid5_congested);
3791 static int raid5_congested(void *data, int bits)
3793 struct mddev *mddev = data;
3795 return mddev_congested(mddev, bits) ||
3796 md_raid5_congested(mddev, bits);
3799 /* We want read requests to align with chunks where possible,
3800 * but write requests don't need to.
3802 static int raid5_mergeable_bvec(struct request_queue *q,
3803 struct bvec_merge_data *bvm,
3804 struct bio_vec *biovec)
3806 struct mddev *mddev = q->queuedata;
3807 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3809 unsigned int chunk_sectors = mddev->chunk_sectors;
3810 unsigned int bio_sectors = bvm->bi_size >> 9;
3812 if ((bvm->bi_rw & 1) == WRITE)
3813 return biovec->bv_len; /* always allow writes to be mergeable */
3815 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3816 chunk_sectors = mddev->new_chunk_sectors;
3817 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3818 if (max < 0) max = 0;
3819 if (max <= biovec->bv_len && bio_sectors == 0)
3820 return biovec->bv_len;
3826 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3828 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3829 unsigned int chunk_sectors = mddev->chunk_sectors;
3830 unsigned int bio_sectors = bio->bi_size >> 9;
3832 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3833 chunk_sectors = mddev->new_chunk_sectors;
3834 return chunk_sectors >=
3835 ((sector & (chunk_sectors - 1)) + bio_sectors);
3839 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3840 * later sampled by raid5d.
3842 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3844 unsigned long flags;
3846 spin_lock_irqsave(&conf->device_lock, flags);
3848 bi->bi_next = conf->retry_read_aligned_list;
3849 conf->retry_read_aligned_list = bi;
3851 spin_unlock_irqrestore(&conf->device_lock, flags);
3852 md_wakeup_thread(conf->mddev->thread);
3856 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3860 bi = conf->retry_read_aligned;
3862 conf->retry_read_aligned = NULL;
3865 bi = conf->retry_read_aligned_list;
3867 conf->retry_read_aligned_list = bi->bi_next;
3870 * this sets the active strip count to 1 and the processed
3871 * strip count to zero (upper 8 bits)
3873 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
3881 * The "raid5_align_endio" should check if the read succeeded and if it
3882 * did, call bio_endio on the original bio (having bio_put the new bio
3884 * If the read failed..
3886 static void raid5_align_endio(struct bio *bi, int error)
3888 struct bio* raid_bi = bi->bi_private;
3889 struct mddev *mddev;
3890 struct r5conf *conf;
3891 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3892 struct md_rdev *rdev;
3896 rdev = (void*)raid_bi->bi_next;
3897 raid_bi->bi_next = NULL;
3898 mddev = rdev->mddev;
3899 conf = mddev->private;
3901 rdev_dec_pending(rdev, conf->mddev);
3903 if (!error && uptodate) {
3904 bio_endio(raid_bi, 0);
3905 if (atomic_dec_and_test(&conf->active_aligned_reads))
3906 wake_up(&conf->wait_for_stripe);
3911 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3913 add_bio_to_retry(raid_bi, conf);
3916 static int bio_fits_rdev(struct bio *bi)
3918 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3920 if ((bi->bi_size>>9) > queue_max_sectors(q))
3922 blk_recount_segments(q, bi);
3923 if (bi->bi_phys_segments > queue_max_segments(q))
3926 if (q->merge_bvec_fn)
3927 /* it's too hard to apply the merge_bvec_fn at this stage,
3936 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3938 struct r5conf *conf = mddev->private;
3940 struct bio* align_bi;
3941 struct md_rdev *rdev;
3942 sector_t end_sector;
3944 if (!in_chunk_boundary(mddev, raid_bio)) {
3945 pr_debug("chunk_aligned_read : non aligned\n");
3949 * use bio_clone_mddev to make a copy of the bio
3951 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3955 * set bi_end_io to a new function, and set bi_private to the
3958 align_bi->bi_end_io = raid5_align_endio;
3959 align_bi->bi_private = raid_bio;
3963 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3967 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3969 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3970 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3971 rdev->recovery_offset < end_sector) {
3972 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3974 (test_bit(Faulty, &rdev->flags) ||
3975 !(test_bit(In_sync, &rdev->flags) ||
3976 rdev->recovery_offset >= end_sector)))
3983 atomic_inc(&rdev->nr_pending);
3985 raid_bio->bi_next = (void*)rdev;
3986 align_bi->bi_bdev = rdev->bdev;
3987 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3989 if (!bio_fits_rdev(align_bi) ||
3990 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3991 &first_bad, &bad_sectors)) {
3992 /* too big in some way, or has a known bad block */
3994 rdev_dec_pending(rdev, mddev);
3998 /* No reshape active, so we can trust rdev->data_offset */
3999 align_bi->bi_sector += rdev->data_offset;
4001 spin_lock_irq(&conf->device_lock);
4002 wait_event_lock_irq(conf->wait_for_stripe,
4005 atomic_inc(&conf->active_aligned_reads);
4006 spin_unlock_irq(&conf->device_lock);
4008 generic_make_request(align_bi);
4017 /* __get_priority_stripe - get the next stripe to process
4019 * Full stripe writes are allowed to pass preread active stripes up until
4020 * the bypass_threshold is exceeded. In general the bypass_count
4021 * increments when the handle_list is handled before the hold_list; however, it
4022 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4023 * stripe with in flight i/o. The bypass_count will be reset when the
4024 * head of the hold_list has changed, i.e. the head was promoted to the
4027 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
4029 struct stripe_head *sh;
4031 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4033 list_empty(&conf->handle_list) ? "empty" : "busy",
4034 list_empty(&conf->hold_list) ? "empty" : "busy",
4035 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4037 if (!list_empty(&conf->handle_list)) {
4038 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4040 if (list_empty(&conf->hold_list))
4041 conf->bypass_count = 0;
4042 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4043 if (conf->hold_list.next == conf->last_hold)
4044 conf->bypass_count++;
4046 conf->last_hold = conf->hold_list.next;
4047 conf->bypass_count -= conf->bypass_threshold;
4048 if (conf->bypass_count < 0)
4049 conf->bypass_count = 0;
4052 } else if (!list_empty(&conf->hold_list) &&
4053 ((conf->bypass_threshold &&
4054 conf->bypass_count > conf->bypass_threshold) ||
4055 atomic_read(&conf->pending_full_writes) == 0)) {
4056 sh = list_entry(conf->hold_list.next,
4058 conf->bypass_count -= conf->bypass_threshold;
4059 if (conf->bypass_count < 0)
4060 conf->bypass_count = 0;
4064 list_del_init(&sh->lru);
4065 atomic_inc(&sh->count);
4066 BUG_ON(atomic_read(&sh->count) != 1);
4070 struct raid5_plug_cb {
4071 struct blk_plug_cb cb;
4072 struct list_head list;
4075 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4077 struct raid5_plug_cb *cb = container_of(
4078 blk_cb, struct raid5_plug_cb, cb);
4079 struct stripe_head *sh;
4080 struct mddev *mddev = cb->cb.data;
4081 struct r5conf *conf = mddev->private;
4083 if (cb->list.next && !list_empty(&cb->list)) {
4084 spin_lock_irq(&conf->device_lock);
4085 while (!list_empty(&cb->list)) {
4086 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4087 list_del_init(&sh->lru);
4089 * avoid race release_stripe_plug() sees
4090 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4091 * is still in our list
4093 smp_mb__before_clear_bit();
4094 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4095 __release_stripe(conf, sh);
4097 spin_unlock_irq(&conf->device_lock);
4102 static void release_stripe_plug(struct mddev *mddev,
4103 struct stripe_head *sh)
4105 struct blk_plug_cb *blk_cb = blk_check_plugged(
4106 raid5_unplug, mddev,
4107 sizeof(struct raid5_plug_cb));
4108 struct raid5_plug_cb *cb;
4115 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4117 if (cb->list.next == NULL)
4118 INIT_LIST_HEAD(&cb->list);
4120 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4121 list_add_tail(&sh->lru, &cb->list);
4126 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4128 struct r5conf *conf = mddev->private;
4129 sector_t logical_sector, last_sector;
4130 struct stripe_head *sh;
4134 if (mddev->reshape_position != MaxSector)
4135 /* Skip discard while reshape is happening */
4138 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4139 last_sector = bi->bi_sector + (bi->bi_size>>9);
4142 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4144 stripe_sectors = conf->chunk_sectors *
4145 (conf->raid_disks - conf->max_degraded);
4146 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4148 sector_div(last_sector, stripe_sectors);
4150 logical_sector *= conf->chunk_sectors;
4151 last_sector *= conf->chunk_sectors;
4153 for (; logical_sector < last_sector;
4154 logical_sector += STRIPE_SECTORS) {
4158 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4159 prepare_to_wait(&conf->wait_for_overlap, &w,
4160 TASK_UNINTERRUPTIBLE);
4161 spin_lock_irq(&sh->stripe_lock);
4162 for (d = 0; d < conf->raid_disks; d++) {
4163 if (d == sh->pd_idx || d == sh->qd_idx)
4165 if (sh->dev[d].towrite || sh->dev[d].toread) {
4166 set_bit(R5_Overlap, &sh->dev[d].flags);
4167 spin_unlock_irq(&sh->stripe_lock);
4173 finish_wait(&conf->wait_for_overlap, &w);
4174 for (d = 0; d < conf->raid_disks; d++) {
4175 if (d == sh->pd_idx || d == sh->qd_idx)
4177 sh->dev[d].towrite = bi;
4178 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4179 raid5_inc_bi_active_stripes(bi);
4181 spin_unlock_irq(&sh->stripe_lock);
4182 if (conf->mddev->bitmap) {
4184 d < conf->raid_disks - conf->max_degraded;
4186 bitmap_startwrite(mddev->bitmap,
4190 sh->bm_seq = conf->seq_flush + 1;
4191 set_bit(STRIPE_BIT_DELAY, &sh->state);
4194 set_bit(STRIPE_HANDLE, &sh->state);
4195 clear_bit(STRIPE_DELAYED, &sh->state);
4196 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4197 atomic_inc(&conf->preread_active_stripes);
4198 release_stripe_plug(mddev, sh);
4201 remaining = raid5_dec_bi_active_stripes(bi);
4202 if (remaining == 0) {
4203 md_write_end(mddev);
4208 static void make_request(struct mddev *mddev, struct bio * bi)
4210 struct r5conf *conf = mddev->private;
4212 sector_t new_sector;
4213 sector_t logical_sector, last_sector;
4214 struct stripe_head *sh;
4215 const int rw = bio_data_dir(bi);
4218 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4219 md_flush_request(mddev, bi);
4223 md_write_start(mddev, bi);
4226 mddev->reshape_position == MaxSector &&
4227 chunk_aligned_read(mddev,bi))
4230 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4231 make_discard_request(mddev, bi);
4235 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4236 last_sector = bi->bi_sector + (bi->bi_size>>9);
4238 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4240 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4246 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4247 if (unlikely(conf->reshape_progress != MaxSector)) {
4248 /* spinlock is needed as reshape_progress may be
4249 * 64bit on a 32bit platform, and so it might be
4250 * possible to see a half-updated value
4251 * Of course reshape_progress could change after
4252 * the lock is dropped, so once we get a reference
4253 * to the stripe that we think it is, we will have
4256 spin_lock_irq(&conf->device_lock);
4257 if (mddev->reshape_backwards
4258 ? logical_sector < conf->reshape_progress
4259 : logical_sector >= conf->reshape_progress) {
4262 if (mddev->reshape_backwards
4263 ? logical_sector < conf->reshape_safe
4264 : logical_sector >= conf->reshape_safe) {
4265 spin_unlock_irq(&conf->device_lock);
4270 spin_unlock_irq(&conf->device_lock);
4273 new_sector = raid5_compute_sector(conf, logical_sector,
4276 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4277 (unsigned long long)new_sector,
4278 (unsigned long long)logical_sector);
4280 sh = get_active_stripe(conf, new_sector, previous,
4281 (bi->bi_rw&RWA_MASK), 0);
4283 if (unlikely(previous)) {
4284 /* expansion might have moved on while waiting for a
4285 * stripe, so we must do the range check again.
4286 * Expansion could still move past after this
4287 * test, but as we are holding a reference to
4288 * 'sh', we know that if that happens,
4289 * STRIPE_EXPANDING will get set and the expansion
4290 * won't proceed until we finish with the stripe.
4293 spin_lock_irq(&conf->device_lock);
4294 if (mddev->reshape_backwards
4295 ? logical_sector >= conf->reshape_progress
4296 : logical_sector < conf->reshape_progress)
4297 /* mismatch, need to try again */
4299 spin_unlock_irq(&conf->device_lock);
4308 logical_sector >= mddev->suspend_lo &&
4309 logical_sector < mddev->suspend_hi) {
4311 /* As the suspend_* range is controlled by
4312 * userspace, we want an interruptible
4315 flush_signals(current);
4316 prepare_to_wait(&conf->wait_for_overlap,
4317 &w, TASK_INTERRUPTIBLE);
4318 if (logical_sector >= mddev->suspend_lo &&
4319 logical_sector < mddev->suspend_hi)
4324 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4325 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4326 /* Stripe is busy expanding or
4327 * add failed due to overlap. Flush everything
4330 md_wakeup_thread(mddev->thread);
4335 finish_wait(&conf->wait_for_overlap, &w);
4336 set_bit(STRIPE_HANDLE, &sh->state);
4337 clear_bit(STRIPE_DELAYED, &sh->state);
4338 if ((bi->bi_rw & REQ_SYNC) &&
4339 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4340 atomic_inc(&conf->preread_active_stripes);
4341 release_stripe_plug(mddev, sh);
4343 /* cannot get stripe for read-ahead, just give-up */
4344 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4345 finish_wait(&conf->wait_for_overlap, &w);
4350 remaining = raid5_dec_bi_active_stripes(bi);
4351 if (remaining == 0) {
4354 md_write_end(mddev);
4360 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4362 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4364 /* reshaping is quite different to recovery/resync so it is
4365 * handled quite separately ... here.
4367 * On each call to sync_request, we gather one chunk worth of
4368 * destination stripes and flag them as expanding.
4369 * Then we find all the source stripes and request reads.
4370 * As the reads complete, handle_stripe will copy the data
4371 * into the destination stripe and release that stripe.
4373 struct r5conf *conf = mddev->private;
4374 struct stripe_head *sh;
4375 sector_t first_sector, last_sector;
4376 int raid_disks = conf->previous_raid_disks;
4377 int data_disks = raid_disks - conf->max_degraded;
4378 int new_data_disks = conf->raid_disks - conf->max_degraded;
4381 sector_t writepos, readpos, safepos;
4382 sector_t stripe_addr;
4383 int reshape_sectors;
4384 struct list_head stripes;
4386 if (sector_nr == 0) {
4387 /* If restarting in the middle, skip the initial sectors */
4388 if (mddev->reshape_backwards &&
4389 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4390 sector_nr = raid5_size(mddev, 0, 0)
4391 - conf->reshape_progress;
4392 } else if (!mddev->reshape_backwards &&
4393 conf->reshape_progress > 0)
4394 sector_nr = conf->reshape_progress;
4395 sector_div(sector_nr, new_data_disks);
4397 mddev->curr_resync_completed = sector_nr;
4398 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4404 /* We need to process a full chunk at a time.
4405 * If old and new chunk sizes differ, we need to process the
4408 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4409 reshape_sectors = mddev->new_chunk_sectors;
4411 reshape_sectors = mddev->chunk_sectors;
4413 /* We update the metadata at least every 10 seconds, or when
4414 * the data about to be copied would over-write the source of
4415 * the data at the front of the range. i.e. one new_stripe
4416 * along from reshape_progress new_maps to after where
4417 * reshape_safe old_maps to
4419 writepos = conf->reshape_progress;
4420 sector_div(writepos, new_data_disks);
4421 readpos = conf->reshape_progress;
4422 sector_div(readpos, data_disks);
4423 safepos = conf->reshape_safe;
4424 sector_div(safepos, data_disks);
4425 if (mddev->reshape_backwards) {
4426 writepos -= min_t(sector_t, reshape_sectors, writepos);
4427 readpos += reshape_sectors;
4428 safepos += reshape_sectors;
4430 writepos += reshape_sectors;
4431 readpos -= min_t(sector_t, reshape_sectors, readpos);
4432 safepos -= min_t(sector_t, reshape_sectors, safepos);
4435 /* Having calculated the 'writepos' possibly use it
4436 * to set 'stripe_addr' which is where we will write to.
4438 if (mddev->reshape_backwards) {
4439 BUG_ON(conf->reshape_progress == 0);
4440 stripe_addr = writepos;
4441 BUG_ON((mddev->dev_sectors &
4442 ~((sector_t)reshape_sectors - 1))
4443 - reshape_sectors - stripe_addr
4446 BUG_ON(writepos != sector_nr + reshape_sectors);
4447 stripe_addr = sector_nr;
4450 /* 'writepos' is the most advanced device address we might write.
4451 * 'readpos' is the least advanced device address we might read.
4452 * 'safepos' is the least address recorded in the metadata as having
4454 * If there is a min_offset_diff, these are adjusted either by
4455 * increasing the safepos/readpos if diff is negative, or
4456 * increasing writepos if diff is positive.
4457 * If 'readpos' is then behind 'writepos', there is no way that we can
4458 * ensure safety in the face of a crash - that must be done by userspace
4459 * making a backup of the data. So in that case there is no particular
4460 * rush to update metadata.
4461 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4462 * update the metadata to advance 'safepos' to match 'readpos' so that
4463 * we can be safe in the event of a crash.
4464 * So we insist on updating metadata if safepos is behind writepos and
4465 * readpos is beyond writepos.
4466 * In any case, update the metadata every 10 seconds.
4467 * Maybe that number should be configurable, but I'm not sure it is
4468 * worth it.... maybe it could be a multiple of safemode_delay???
4470 if (conf->min_offset_diff < 0) {
4471 safepos += -conf->min_offset_diff;
4472 readpos += -conf->min_offset_diff;
4474 writepos += conf->min_offset_diff;
4476 if ((mddev->reshape_backwards
4477 ? (safepos > writepos && readpos < writepos)
4478 : (safepos < writepos && readpos > writepos)) ||
4479 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4480 /* Cannot proceed until we've updated the superblock... */
4481 wait_event(conf->wait_for_overlap,
4482 atomic_read(&conf->reshape_stripes)==0);
4483 mddev->reshape_position = conf->reshape_progress;
4484 mddev->curr_resync_completed = sector_nr;
4485 conf->reshape_checkpoint = jiffies;
4486 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4487 md_wakeup_thread(mddev->thread);
4488 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4489 kthread_should_stop());
4490 spin_lock_irq(&conf->device_lock);
4491 conf->reshape_safe = mddev->reshape_position;
4492 spin_unlock_irq(&conf->device_lock);
4493 wake_up(&conf->wait_for_overlap);
4494 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4497 INIT_LIST_HEAD(&stripes);
4498 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4500 int skipped_disk = 0;
4501 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4502 set_bit(STRIPE_EXPANDING, &sh->state);
4503 atomic_inc(&conf->reshape_stripes);
4504 /* If any of this stripe is beyond the end of the old
4505 * array, then we need to zero those blocks
4507 for (j=sh->disks; j--;) {
4509 if (j == sh->pd_idx)
4511 if (conf->level == 6 &&
4514 s = compute_blocknr(sh, j, 0);
4515 if (s < raid5_size(mddev, 0, 0)) {
4519 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4520 set_bit(R5_Expanded, &sh->dev[j].flags);
4521 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4523 if (!skipped_disk) {
4524 set_bit(STRIPE_EXPAND_READY, &sh->state);
4525 set_bit(STRIPE_HANDLE, &sh->state);
4527 list_add(&sh->lru, &stripes);
4529 spin_lock_irq(&conf->device_lock);
4530 if (mddev->reshape_backwards)
4531 conf->reshape_progress -= reshape_sectors * new_data_disks;
4533 conf->reshape_progress += reshape_sectors * new_data_disks;
4534 spin_unlock_irq(&conf->device_lock);
4535 /* Ok, those stripe are ready. We can start scheduling
4536 * reads on the source stripes.
4537 * The source stripes are determined by mapping the first and last
4538 * block on the destination stripes.
4541 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4544 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4545 * new_data_disks - 1),
4547 if (last_sector >= mddev->dev_sectors)
4548 last_sector = mddev->dev_sectors - 1;
4549 while (first_sector <= last_sector) {
4550 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4551 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4552 set_bit(STRIPE_HANDLE, &sh->state);
4554 first_sector += STRIPE_SECTORS;
4556 /* Now that the sources are clearly marked, we can release
4557 * the destination stripes
4559 while (!list_empty(&stripes)) {
4560 sh = list_entry(stripes.next, struct stripe_head, lru);
4561 list_del_init(&sh->lru);
4564 /* If this takes us to the resync_max point where we have to pause,
4565 * then we need to write out the superblock.
4567 sector_nr += reshape_sectors;
4568 if ((sector_nr - mddev->curr_resync_completed) * 2
4569 >= mddev->resync_max - mddev->curr_resync_completed) {
4570 /* Cannot proceed until we've updated the superblock... */
4571 wait_event(conf->wait_for_overlap,
4572 atomic_read(&conf->reshape_stripes) == 0);
4573 mddev->reshape_position = conf->reshape_progress;
4574 mddev->curr_resync_completed = sector_nr;
4575 conf->reshape_checkpoint = jiffies;
4576 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4577 md_wakeup_thread(mddev->thread);
4578 wait_event(mddev->sb_wait,
4579 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4580 || kthread_should_stop());
4581 spin_lock_irq(&conf->device_lock);
4582 conf->reshape_safe = mddev->reshape_position;
4583 spin_unlock_irq(&conf->device_lock);
4584 wake_up(&conf->wait_for_overlap);
4585 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4587 return reshape_sectors;
4590 /* FIXME go_faster isn't used */
4591 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4593 struct r5conf *conf = mddev->private;
4594 struct stripe_head *sh;
4595 sector_t max_sector = mddev->dev_sectors;
4596 sector_t sync_blocks;
4597 int still_degraded = 0;
4600 if (sector_nr >= max_sector) {
4601 /* just being told to finish up .. nothing much to do */
4603 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4608 if (mddev->curr_resync < max_sector) /* aborted */
4609 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4611 else /* completed sync */
4613 bitmap_close_sync(mddev->bitmap);
4618 /* Allow raid5_quiesce to complete */
4619 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4621 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4622 return reshape_request(mddev, sector_nr, skipped);
4624 /* No need to check resync_max as we never do more than one
4625 * stripe, and as resync_max will always be on a chunk boundary,
4626 * if the check in md_do_sync didn't fire, there is no chance
4627 * of overstepping resync_max here
4630 /* if there is too many failed drives and we are trying
4631 * to resync, then assert that we are finished, because there is
4632 * nothing we can do.
4634 if (mddev->degraded >= conf->max_degraded &&
4635 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4636 sector_t rv = mddev->dev_sectors - sector_nr;
4640 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4641 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4642 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4643 /* we can skip this block, and probably more */
4644 sync_blocks /= STRIPE_SECTORS;
4646 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4649 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4651 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4653 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4654 /* make sure we don't swamp the stripe cache if someone else
4655 * is trying to get access
4657 schedule_timeout_uninterruptible(1);
4659 /* Need to check if array will still be degraded after recovery/resync
4660 * We don't need to check the 'failed' flag as when that gets set,
4663 for (i = 0; i < conf->raid_disks; i++)
4664 if (conf->disks[i].rdev == NULL)
4667 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4669 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4674 return STRIPE_SECTORS;
4677 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4679 /* We may not be able to submit a whole bio at once as there
4680 * may not be enough stripe_heads available.
4681 * We cannot pre-allocate enough stripe_heads as we may need
4682 * more than exist in the cache (if we allow ever large chunks).
4683 * So we do one stripe head at a time and record in
4684 * ->bi_hw_segments how many have been done.
4686 * We *know* that this entire raid_bio is in one chunk, so
4687 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4689 struct stripe_head *sh;
4691 sector_t sector, logical_sector, last_sector;
4696 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4697 sector = raid5_compute_sector(conf, logical_sector,
4699 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4701 for (; logical_sector < last_sector;
4702 logical_sector += STRIPE_SECTORS,
4703 sector += STRIPE_SECTORS,
4706 if (scnt < raid5_bi_processed_stripes(raid_bio))
4707 /* already done this stripe */
4710 sh = get_active_stripe(conf, sector, 0, 1, 0);
4713 /* failed to get a stripe - must wait */
4714 raid5_set_bi_processed_stripes(raid_bio, scnt);
4715 conf->retry_read_aligned = raid_bio;
4719 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4721 raid5_set_bi_processed_stripes(raid_bio, scnt);
4722 conf->retry_read_aligned = raid_bio;
4726 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4731 remaining = raid5_dec_bi_active_stripes(raid_bio);
4733 bio_endio(raid_bio, 0);
4734 if (atomic_dec_and_test(&conf->active_aligned_reads))
4735 wake_up(&conf->wait_for_stripe);
4739 #define MAX_STRIPE_BATCH 8
4740 static int handle_active_stripes(struct r5conf *conf)
4742 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4743 int i, batch_size = 0;
4745 while (batch_size < MAX_STRIPE_BATCH &&
4746 (sh = __get_priority_stripe(conf)) != NULL)
4747 batch[batch_size++] = sh;
4749 if (batch_size == 0)
4751 spin_unlock_irq(&conf->device_lock);
4753 for (i = 0; i < batch_size; i++)
4754 handle_stripe(batch[i]);
4758 spin_lock_irq(&conf->device_lock);
4759 for (i = 0; i < batch_size; i++)
4760 __release_stripe(conf, batch[i]);
4765 * This is our raid5 kernel thread.
4767 * We scan the hash table for stripes which can be handled now.
4768 * During the scan, completed stripes are saved for us by the interrupt
4769 * handler, so that they will not have to wait for our next wakeup.
4771 static void raid5d(struct md_thread *thread)
4773 struct mddev *mddev = thread->mddev;
4774 struct r5conf *conf = mddev->private;
4776 struct blk_plug plug;
4778 pr_debug("+++ raid5d active\n");
4780 md_check_recovery(mddev);
4782 blk_start_plug(&plug);
4784 spin_lock_irq(&conf->device_lock);
4790 !list_empty(&conf->bitmap_list)) {
4791 /* Now is a good time to flush some bitmap updates */
4793 spin_unlock_irq(&conf->device_lock);
4794 bitmap_unplug(mddev->bitmap);
4795 spin_lock_irq(&conf->device_lock);
4796 conf->seq_write = conf->seq_flush;
4797 activate_bit_delay(conf);
4799 raid5_activate_delayed(conf);
4801 while ((bio = remove_bio_from_retry(conf))) {
4803 spin_unlock_irq(&conf->device_lock);
4804 ok = retry_aligned_read(conf, bio);
4805 spin_lock_irq(&conf->device_lock);
4811 batch_size = handle_active_stripes(conf);
4814 handled += batch_size;
4816 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
4817 spin_unlock_irq(&conf->device_lock);
4818 md_check_recovery(mddev);
4819 spin_lock_irq(&conf->device_lock);
4822 pr_debug("%d stripes handled\n", handled);
4824 spin_unlock_irq(&conf->device_lock);
4826 async_tx_issue_pending_all();
4827 blk_finish_plug(&plug);
4829 pr_debug("--- raid5d inactive\n");
4833 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4835 struct r5conf *conf = mddev->private;
4837 return sprintf(page, "%d\n", conf->max_nr_stripes);
4843 raid5_set_cache_size(struct mddev *mddev, int size)
4845 struct r5conf *conf = mddev->private;
4848 if (size <= 16 || size > 32768)
4850 while (size < conf->max_nr_stripes) {
4851 if (drop_one_stripe(conf))
4852 conf->max_nr_stripes--;
4856 err = md_allow_write(mddev);
4859 while (size > conf->max_nr_stripes) {
4860 if (grow_one_stripe(conf))
4861 conf->max_nr_stripes++;
4866 EXPORT_SYMBOL(raid5_set_cache_size);
4869 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4871 struct r5conf *conf = mddev->private;
4875 if (len >= PAGE_SIZE)
4880 if (strict_strtoul(page, 10, &new))
4882 err = raid5_set_cache_size(mddev, new);
4888 static struct md_sysfs_entry
4889 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4890 raid5_show_stripe_cache_size,
4891 raid5_store_stripe_cache_size);
4894 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4896 struct r5conf *conf = mddev->private;
4898 return sprintf(page, "%d\n", conf->bypass_threshold);
4904 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4906 struct r5conf *conf = mddev->private;
4908 if (len >= PAGE_SIZE)
4913 if (strict_strtoul(page, 10, &new))
4915 if (new > conf->max_nr_stripes)
4917 conf->bypass_threshold = new;
4921 static struct md_sysfs_entry
4922 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4924 raid5_show_preread_threshold,
4925 raid5_store_preread_threshold);
4928 stripe_cache_active_show(struct mddev *mddev, char *page)
4930 struct r5conf *conf = mddev->private;
4932 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4937 static struct md_sysfs_entry
4938 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4940 static struct attribute *raid5_attrs[] = {
4941 &raid5_stripecache_size.attr,
4942 &raid5_stripecache_active.attr,
4943 &raid5_preread_bypass_threshold.attr,
4946 static struct attribute_group raid5_attrs_group = {
4948 .attrs = raid5_attrs,
4952 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4954 struct r5conf *conf = mddev->private;
4957 sectors = mddev->dev_sectors;
4959 /* size is defined by the smallest of previous and new size */
4960 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4962 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4963 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4964 return sectors * (raid_disks - conf->max_degraded);
4967 static void raid5_free_percpu(struct r5conf *conf)
4969 struct raid5_percpu *percpu;
4976 for_each_possible_cpu(cpu) {
4977 percpu = per_cpu_ptr(conf->percpu, cpu);
4978 safe_put_page(percpu->spare_page);
4979 kfree(percpu->scribble);
4981 #ifdef CONFIG_HOTPLUG_CPU
4982 unregister_cpu_notifier(&conf->cpu_notify);
4986 free_percpu(conf->percpu);
4989 static void free_conf(struct r5conf *conf)
4991 shrink_stripes(conf);
4992 raid5_free_percpu(conf);
4994 kfree(conf->stripe_hashtbl);
4998 #ifdef CONFIG_HOTPLUG_CPU
4999 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5002 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5003 long cpu = (long)hcpu;
5004 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5007 case CPU_UP_PREPARE:
5008 case CPU_UP_PREPARE_FROZEN:
5009 if (conf->level == 6 && !percpu->spare_page)
5010 percpu->spare_page = alloc_page(GFP_KERNEL);
5011 if (!percpu->scribble)
5012 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5014 if (!percpu->scribble ||
5015 (conf->level == 6 && !percpu->spare_page)) {
5016 safe_put_page(percpu->spare_page);
5017 kfree(percpu->scribble);
5018 pr_err("%s: failed memory allocation for cpu%ld\n",
5020 return notifier_from_errno(-ENOMEM);
5024 case CPU_DEAD_FROZEN:
5025 safe_put_page(percpu->spare_page);
5026 kfree(percpu->scribble);
5027 percpu->spare_page = NULL;
5028 percpu->scribble = NULL;
5037 static int raid5_alloc_percpu(struct r5conf *conf)
5040 struct page *spare_page;
5041 struct raid5_percpu __percpu *allcpus;
5045 allcpus = alloc_percpu(struct raid5_percpu);
5048 conf->percpu = allcpus;
5052 for_each_present_cpu(cpu) {
5053 if (conf->level == 6) {
5054 spare_page = alloc_page(GFP_KERNEL);
5059 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5061 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5066 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5068 #ifdef CONFIG_HOTPLUG_CPU
5069 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5070 conf->cpu_notify.priority = 0;
5072 err = register_cpu_notifier(&conf->cpu_notify);
5079 static struct r5conf *setup_conf(struct mddev *mddev)
5081 struct r5conf *conf;
5082 int raid_disk, memory, max_disks;
5083 struct md_rdev *rdev;
5084 struct disk_info *disk;
5087 if (mddev->new_level != 5
5088 && mddev->new_level != 4
5089 && mddev->new_level != 6) {
5090 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5091 mdname(mddev), mddev->new_level);
5092 return ERR_PTR(-EIO);
5094 if ((mddev->new_level == 5
5095 && !algorithm_valid_raid5(mddev->new_layout)) ||
5096 (mddev->new_level == 6
5097 && !algorithm_valid_raid6(mddev->new_layout))) {
5098 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5099 mdname(mddev), mddev->new_layout);
5100 return ERR_PTR(-EIO);
5102 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5103 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5104 mdname(mddev), mddev->raid_disks);
5105 return ERR_PTR(-EINVAL);
5108 if (!mddev->new_chunk_sectors ||
5109 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5110 !is_power_of_2(mddev->new_chunk_sectors)) {
5111 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5112 mdname(mddev), mddev->new_chunk_sectors << 9);
5113 return ERR_PTR(-EINVAL);
5116 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5119 spin_lock_init(&conf->device_lock);
5120 init_waitqueue_head(&conf->wait_for_stripe);
5121 init_waitqueue_head(&conf->wait_for_overlap);
5122 INIT_LIST_HEAD(&conf->handle_list);
5123 INIT_LIST_HEAD(&conf->hold_list);
5124 INIT_LIST_HEAD(&conf->delayed_list);
5125 INIT_LIST_HEAD(&conf->bitmap_list);
5126 INIT_LIST_HEAD(&conf->inactive_list);
5127 atomic_set(&conf->active_stripes, 0);
5128 atomic_set(&conf->preread_active_stripes, 0);
5129 atomic_set(&conf->active_aligned_reads, 0);
5130 conf->bypass_threshold = BYPASS_THRESHOLD;
5131 conf->recovery_disabled = mddev->recovery_disabled - 1;
5133 conf->raid_disks = mddev->raid_disks;
5134 if (mddev->reshape_position == MaxSector)
5135 conf->previous_raid_disks = mddev->raid_disks;
5137 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5138 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5139 conf->scribble_len = scribble_len(max_disks);
5141 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5146 conf->mddev = mddev;
5148 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5151 conf->level = mddev->new_level;
5152 if (raid5_alloc_percpu(conf) != 0)
5155 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5157 rdev_for_each(rdev, mddev) {
5158 raid_disk = rdev->raid_disk;
5159 if (raid_disk >= max_disks
5162 disk = conf->disks + raid_disk;
5164 if (test_bit(Replacement, &rdev->flags)) {
5165 if (disk->replacement)
5167 disk->replacement = rdev;
5174 if (test_bit(In_sync, &rdev->flags)) {
5175 char b[BDEVNAME_SIZE];
5176 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5178 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5179 } else if (rdev->saved_raid_disk != raid_disk)
5180 /* Cannot rely on bitmap to complete recovery */
5184 conf->chunk_sectors = mddev->new_chunk_sectors;
5185 conf->level = mddev->new_level;
5186 if (conf->level == 6)
5187 conf->max_degraded = 2;
5189 conf->max_degraded = 1;
5190 conf->algorithm = mddev->new_layout;
5191 conf->max_nr_stripes = NR_STRIPES;
5192 conf->reshape_progress = mddev->reshape_position;
5193 if (conf->reshape_progress != MaxSector) {
5194 conf->prev_chunk_sectors = mddev->chunk_sectors;
5195 conf->prev_algo = mddev->layout;
5198 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5199 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5200 if (grow_stripes(conf, conf->max_nr_stripes)) {
5202 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5203 mdname(mddev), memory);
5206 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5207 mdname(mddev), memory);
5209 sprintf(pers_name, "raid%d", mddev->new_level);
5210 conf->thread = md_register_thread(raid5d, mddev, pers_name);
5211 if (!conf->thread) {
5213 "md/raid:%s: couldn't allocate thread.\n",
5223 return ERR_PTR(-EIO);
5225 return ERR_PTR(-ENOMEM);
5229 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5232 case ALGORITHM_PARITY_0:
5233 if (raid_disk < max_degraded)
5236 case ALGORITHM_PARITY_N:
5237 if (raid_disk >= raid_disks - max_degraded)
5240 case ALGORITHM_PARITY_0_6:
5241 if (raid_disk == 0 ||
5242 raid_disk == raid_disks - 1)
5245 case ALGORITHM_LEFT_ASYMMETRIC_6:
5246 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5247 case ALGORITHM_LEFT_SYMMETRIC_6:
5248 case ALGORITHM_RIGHT_SYMMETRIC_6:
5249 if (raid_disk == raid_disks - 1)
5255 static int run(struct mddev *mddev)
5257 struct r5conf *conf;
5258 int working_disks = 0;
5259 int dirty_parity_disks = 0;
5260 struct md_rdev *rdev;
5261 sector_t reshape_offset = 0;
5263 long long min_offset_diff = 0;
5266 if (mddev->recovery_cp != MaxSector)
5267 printk(KERN_NOTICE "md/raid:%s: not clean"
5268 " -- starting background reconstruction\n",
5271 rdev_for_each(rdev, mddev) {
5273 if (rdev->raid_disk < 0)
5275 diff = (rdev->new_data_offset - rdev->data_offset);
5277 min_offset_diff = diff;
5279 } else if (mddev->reshape_backwards &&
5280 diff < min_offset_diff)
5281 min_offset_diff = diff;
5282 else if (!mddev->reshape_backwards &&
5283 diff > min_offset_diff)
5284 min_offset_diff = diff;
5287 if (mddev->reshape_position != MaxSector) {
5288 /* Check that we can continue the reshape.
5289 * Difficulties arise if the stripe we would write to
5290 * next is at or after the stripe we would read from next.
5291 * For a reshape that changes the number of devices, this
5292 * is only possible for a very short time, and mdadm makes
5293 * sure that time appears to have past before assembling
5294 * the array. So we fail if that time hasn't passed.
5295 * For a reshape that keeps the number of devices the same
5296 * mdadm must be monitoring the reshape can keeping the
5297 * critical areas read-only and backed up. It will start
5298 * the array in read-only mode, so we check for that.
5300 sector_t here_new, here_old;
5302 int max_degraded = (mddev->level == 6 ? 2 : 1);
5304 if (mddev->new_level != mddev->level) {
5305 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5306 "required - aborting.\n",
5310 old_disks = mddev->raid_disks - mddev->delta_disks;
5311 /* reshape_position must be on a new-stripe boundary, and one
5312 * further up in new geometry must map after here in old
5315 here_new = mddev->reshape_position;
5316 if (sector_div(here_new, mddev->new_chunk_sectors *
5317 (mddev->raid_disks - max_degraded))) {
5318 printk(KERN_ERR "md/raid:%s: reshape_position not "
5319 "on a stripe boundary\n", mdname(mddev));
5322 reshape_offset = here_new * mddev->new_chunk_sectors;
5323 /* here_new is the stripe we will write to */
5324 here_old = mddev->reshape_position;
5325 sector_div(here_old, mddev->chunk_sectors *
5326 (old_disks-max_degraded));
5327 /* here_old is the first stripe that we might need to read
5329 if (mddev->delta_disks == 0) {
5330 if ((here_new * mddev->new_chunk_sectors !=
5331 here_old * mddev->chunk_sectors)) {
5332 printk(KERN_ERR "md/raid:%s: reshape position is"
5333 " confused - aborting\n", mdname(mddev));
5336 /* We cannot be sure it is safe to start an in-place
5337 * reshape. It is only safe if user-space is monitoring
5338 * and taking constant backups.
5339 * mdadm always starts a situation like this in
5340 * readonly mode so it can take control before
5341 * allowing any writes. So just check for that.
5343 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5344 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5345 /* not really in-place - so OK */;
5346 else if (mddev->ro == 0) {
5347 printk(KERN_ERR "md/raid:%s: in-place reshape "
5348 "must be started in read-only mode "
5353 } else if (mddev->reshape_backwards
5354 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5355 here_old * mddev->chunk_sectors)
5356 : (here_new * mddev->new_chunk_sectors >=
5357 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5358 /* Reading from the same stripe as writing to - bad */
5359 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5360 "auto-recovery - aborting.\n",
5364 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5366 /* OK, we should be able to continue; */
5368 BUG_ON(mddev->level != mddev->new_level);
5369 BUG_ON(mddev->layout != mddev->new_layout);
5370 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5371 BUG_ON(mddev->delta_disks != 0);
5374 if (mddev->private == NULL)
5375 conf = setup_conf(mddev);
5377 conf = mddev->private;
5380 return PTR_ERR(conf);
5382 conf->min_offset_diff = min_offset_diff;
5383 mddev->thread = conf->thread;
5384 conf->thread = NULL;
5385 mddev->private = conf;
5387 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5389 rdev = conf->disks[i].rdev;
5390 if (!rdev && conf->disks[i].replacement) {
5391 /* The replacement is all we have yet */
5392 rdev = conf->disks[i].replacement;
5393 conf->disks[i].replacement = NULL;
5394 clear_bit(Replacement, &rdev->flags);
5395 conf->disks[i].rdev = rdev;
5399 if (conf->disks[i].replacement &&
5400 conf->reshape_progress != MaxSector) {
5401 /* replacements and reshape simply do not mix. */
5402 printk(KERN_ERR "md: cannot handle concurrent "
5403 "replacement and reshape.\n");
5406 if (test_bit(In_sync, &rdev->flags)) {
5410 /* This disc is not fully in-sync. However if it
5411 * just stored parity (beyond the recovery_offset),
5412 * when we don't need to be concerned about the
5413 * array being dirty.
5414 * When reshape goes 'backwards', we never have
5415 * partially completed devices, so we only need
5416 * to worry about reshape going forwards.
5418 /* Hack because v0.91 doesn't store recovery_offset properly. */
5419 if (mddev->major_version == 0 &&
5420 mddev->minor_version > 90)
5421 rdev->recovery_offset = reshape_offset;
5423 if (rdev->recovery_offset < reshape_offset) {
5424 /* We need to check old and new layout */
5425 if (!only_parity(rdev->raid_disk,
5428 conf->max_degraded))
5431 if (!only_parity(rdev->raid_disk,
5433 conf->previous_raid_disks,
5434 conf->max_degraded))
5436 dirty_parity_disks++;
5440 * 0 for a fully functional array, 1 or 2 for a degraded array.
5442 mddev->degraded = calc_degraded(conf);
5444 if (has_failed(conf)) {
5445 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5446 " (%d/%d failed)\n",
5447 mdname(mddev), mddev->degraded, conf->raid_disks);
5451 /* device size must be a multiple of chunk size */
5452 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5453 mddev->resync_max_sectors = mddev->dev_sectors;
5455 if (mddev->degraded > dirty_parity_disks &&
5456 mddev->recovery_cp != MaxSector) {
5457 if (mddev->ok_start_degraded)
5459 "md/raid:%s: starting dirty degraded array"
5460 " - data corruption possible.\n",
5464 "md/raid:%s: cannot start dirty degraded array.\n",
5470 if (mddev->degraded == 0)
5471 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5472 " devices, algorithm %d\n", mdname(mddev), conf->level,
5473 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5476 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5477 " out of %d devices, algorithm %d\n",
5478 mdname(mddev), conf->level,
5479 mddev->raid_disks - mddev->degraded,
5480 mddev->raid_disks, mddev->new_layout);
5482 print_raid5_conf(conf);
5484 if (conf->reshape_progress != MaxSector) {
5485 conf->reshape_safe = conf->reshape_progress;
5486 atomic_set(&conf->reshape_stripes, 0);
5487 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5488 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5489 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5490 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5491 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5496 /* Ok, everything is just fine now */
5497 if (mddev->to_remove == &raid5_attrs_group)
5498 mddev->to_remove = NULL;
5499 else if (mddev->kobj.sd &&
5500 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5502 "raid5: failed to create sysfs attributes for %s\n",
5504 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5508 bool discard_supported = true;
5509 /* read-ahead size must cover two whole stripes, which
5510 * is 2 * (datadisks) * chunksize where 'n' is the
5511 * number of raid devices
5513 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5514 int stripe = data_disks *
5515 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5516 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5517 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5519 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5521 mddev->queue->backing_dev_info.congested_data = mddev;
5522 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5524 chunk_size = mddev->chunk_sectors << 9;
5525 blk_queue_io_min(mddev->queue, chunk_size);
5526 blk_queue_io_opt(mddev->queue, chunk_size *
5527 (conf->raid_disks - conf->max_degraded));
5529 * We can only discard a whole stripe. It doesn't make sense to
5530 * discard data disk but write parity disk
5532 stripe = stripe * PAGE_SIZE;
5533 /* Round up to power of 2, as discard handling
5534 * currently assumes that */
5535 while ((stripe-1) & stripe)
5536 stripe = (stripe | (stripe-1)) + 1;
5537 mddev->queue->limits.discard_alignment = stripe;
5538 mddev->queue->limits.discard_granularity = stripe;
5540 * unaligned part of discard request will be ignored, so can't
5541 * guarantee discard_zerors_data
5543 mddev->queue->limits.discard_zeroes_data = 0;
5545 rdev_for_each(rdev, mddev) {
5546 disk_stack_limits(mddev->gendisk, rdev->bdev,
5547 rdev->data_offset << 9);
5548 disk_stack_limits(mddev->gendisk, rdev->bdev,
5549 rdev->new_data_offset << 9);
5551 * discard_zeroes_data is required, otherwise data
5552 * could be lost. Consider a scenario: discard a stripe
5553 * (the stripe could be inconsistent if
5554 * discard_zeroes_data is 0); write one disk of the
5555 * stripe (the stripe could be inconsistent again
5556 * depending on which disks are used to calculate
5557 * parity); the disk is broken; The stripe data of this
5560 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5561 !bdev_get_queue(rdev->bdev)->
5562 limits.discard_zeroes_data)
5563 discard_supported = false;
5566 if (discard_supported &&
5567 mddev->queue->limits.max_discard_sectors >= stripe &&
5568 mddev->queue->limits.discard_granularity >= stripe)
5569 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5572 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5578 md_unregister_thread(&mddev->thread);
5579 print_raid5_conf(conf);
5581 mddev->private = NULL;
5582 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5586 static int stop(struct mddev *mddev)
5588 struct r5conf *conf = mddev->private;
5590 md_unregister_thread(&mddev->thread);
5592 mddev->queue->backing_dev_info.congested_fn = NULL;
5594 mddev->private = NULL;
5595 mddev->to_remove = &raid5_attrs_group;
5599 static void status(struct seq_file *seq, struct mddev *mddev)
5601 struct r5conf *conf = mddev->private;
5604 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5605 mddev->chunk_sectors / 2, mddev->layout);
5606 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5607 for (i = 0; i < conf->raid_disks; i++)
5608 seq_printf (seq, "%s",
5609 conf->disks[i].rdev &&
5610 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5611 seq_printf (seq, "]");
5614 static void print_raid5_conf (struct r5conf *conf)
5617 struct disk_info *tmp;
5619 printk(KERN_DEBUG "RAID conf printout:\n");
5621 printk("(conf==NULL)\n");
5624 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5626 conf->raid_disks - conf->mddev->degraded);
5628 for (i = 0; i < conf->raid_disks; i++) {
5629 char b[BDEVNAME_SIZE];
5630 tmp = conf->disks + i;
5632 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5633 i, !test_bit(Faulty, &tmp->rdev->flags),
5634 bdevname(tmp->rdev->bdev, b));
5638 static int raid5_spare_active(struct mddev *mddev)
5641 struct r5conf *conf = mddev->private;
5642 struct disk_info *tmp;
5644 unsigned long flags;
5646 for (i = 0; i < conf->raid_disks; i++) {
5647 tmp = conf->disks + i;
5648 if (tmp->replacement
5649 && tmp->replacement->recovery_offset == MaxSector
5650 && !test_bit(Faulty, &tmp->replacement->flags)
5651 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5652 /* Replacement has just become active. */
5654 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5657 /* Replaced device not technically faulty,
5658 * but we need to be sure it gets removed
5659 * and never re-added.
5661 set_bit(Faulty, &tmp->rdev->flags);
5662 sysfs_notify_dirent_safe(
5663 tmp->rdev->sysfs_state);
5665 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5666 } else if (tmp->rdev
5667 && tmp->rdev->recovery_offset == MaxSector
5668 && !test_bit(Faulty, &tmp->rdev->flags)
5669 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5671 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5674 spin_lock_irqsave(&conf->device_lock, flags);
5675 mddev->degraded = calc_degraded(conf);
5676 spin_unlock_irqrestore(&conf->device_lock, flags);
5677 print_raid5_conf(conf);
5681 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5683 struct r5conf *conf = mddev->private;
5685 int number = rdev->raid_disk;
5686 struct md_rdev **rdevp;
5687 struct disk_info *p = conf->disks + number;
5689 print_raid5_conf(conf);
5690 if (rdev == p->rdev)
5692 else if (rdev == p->replacement)
5693 rdevp = &p->replacement;
5697 if (number >= conf->raid_disks &&
5698 conf->reshape_progress == MaxSector)
5699 clear_bit(In_sync, &rdev->flags);
5701 if (test_bit(In_sync, &rdev->flags) ||
5702 atomic_read(&rdev->nr_pending)) {
5706 /* Only remove non-faulty devices if recovery
5709 if (!test_bit(Faulty, &rdev->flags) &&
5710 mddev->recovery_disabled != conf->recovery_disabled &&
5711 !has_failed(conf) &&
5712 (!p->replacement || p->replacement == rdev) &&
5713 number < conf->raid_disks) {
5719 if (atomic_read(&rdev->nr_pending)) {
5720 /* lost the race, try later */
5723 } else if (p->replacement) {
5724 /* We must have just cleared 'rdev' */
5725 p->rdev = p->replacement;
5726 clear_bit(Replacement, &p->replacement->flags);
5727 smp_mb(); /* Make sure other CPUs may see both as identical
5728 * but will never see neither - if they are careful
5730 p->replacement = NULL;
5731 clear_bit(WantReplacement, &rdev->flags);
5733 /* We might have just removed the Replacement as faulty-
5734 * clear the bit just in case
5736 clear_bit(WantReplacement, &rdev->flags);
5739 print_raid5_conf(conf);
5743 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5745 struct r5conf *conf = mddev->private;
5748 struct disk_info *p;
5750 int last = conf->raid_disks - 1;
5752 if (mddev->recovery_disabled == conf->recovery_disabled)
5755 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5756 /* no point adding a device */
5759 if (rdev->raid_disk >= 0)
5760 first = last = rdev->raid_disk;
5763 * find the disk ... but prefer rdev->saved_raid_disk
5766 if (rdev->saved_raid_disk >= 0 &&
5767 rdev->saved_raid_disk >= first &&
5768 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5769 first = rdev->saved_raid_disk;
5771 for (disk = first; disk <= last; disk++) {
5772 p = conf->disks + disk;
5773 if (p->rdev == NULL) {
5774 clear_bit(In_sync, &rdev->flags);
5775 rdev->raid_disk = disk;
5777 if (rdev->saved_raid_disk != disk)
5779 rcu_assign_pointer(p->rdev, rdev);
5783 for (disk = first; disk <= last; disk++) {
5784 p = conf->disks + disk;
5785 if (test_bit(WantReplacement, &p->rdev->flags) &&
5786 p->replacement == NULL) {
5787 clear_bit(In_sync, &rdev->flags);
5788 set_bit(Replacement, &rdev->flags);
5789 rdev->raid_disk = disk;
5792 rcu_assign_pointer(p->replacement, rdev);
5797 print_raid5_conf(conf);
5801 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5803 /* no resync is happening, and there is enough space
5804 * on all devices, so we can resize.
5805 * We need to make sure resync covers any new space.
5806 * If the array is shrinking we should possibly wait until
5807 * any io in the removed space completes, but it hardly seems
5811 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5812 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5813 if (mddev->external_size &&
5814 mddev->array_sectors > newsize)
5816 if (mddev->bitmap) {
5817 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5821 md_set_array_sectors(mddev, newsize);
5822 set_capacity(mddev->gendisk, mddev->array_sectors);
5823 revalidate_disk(mddev->gendisk);
5824 if (sectors > mddev->dev_sectors &&
5825 mddev->recovery_cp > mddev->dev_sectors) {
5826 mddev->recovery_cp = mddev->dev_sectors;
5827 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5829 mddev->dev_sectors = sectors;
5830 mddev->resync_max_sectors = sectors;
5834 static int check_stripe_cache(struct mddev *mddev)
5836 /* Can only proceed if there are plenty of stripe_heads.
5837 * We need a minimum of one full stripe,, and for sensible progress
5838 * it is best to have about 4 times that.
5839 * If we require 4 times, then the default 256 4K stripe_heads will
5840 * allow for chunk sizes up to 256K, which is probably OK.
5841 * If the chunk size is greater, user-space should request more
5842 * stripe_heads first.
5844 struct r5conf *conf = mddev->private;
5845 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5846 > conf->max_nr_stripes ||
5847 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5848 > conf->max_nr_stripes) {
5849 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5851 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5858 static int check_reshape(struct mddev *mddev)
5860 struct r5conf *conf = mddev->private;
5862 if (mddev->delta_disks == 0 &&
5863 mddev->new_layout == mddev->layout &&
5864 mddev->new_chunk_sectors == mddev->chunk_sectors)
5865 return 0; /* nothing to do */
5866 if (has_failed(conf))
5868 if (mddev->delta_disks < 0) {
5869 /* We might be able to shrink, but the devices must
5870 * be made bigger first.
5871 * For raid6, 4 is the minimum size.
5872 * Otherwise 2 is the minimum
5875 if (mddev->level == 6)
5877 if (mddev->raid_disks + mddev->delta_disks < min)
5881 if (!check_stripe_cache(mddev))
5884 return resize_stripes(conf, (conf->previous_raid_disks
5885 + mddev->delta_disks));
5888 static int raid5_start_reshape(struct mddev *mddev)
5890 struct r5conf *conf = mddev->private;
5891 struct md_rdev *rdev;
5893 unsigned long flags;
5895 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5898 if (!check_stripe_cache(mddev))
5901 if (has_failed(conf))
5904 rdev_for_each(rdev, mddev) {
5905 if (!test_bit(In_sync, &rdev->flags)
5906 && !test_bit(Faulty, &rdev->flags))
5910 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5911 /* Not enough devices even to make a degraded array
5916 /* Refuse to reduce size of the array. Any reductions in
5917 * array size must be through explicit setting of array_size
5920 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5921 < mddev->array_sectors) {
5922 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5923 "before number of disks\n", mdname(mddev));
5927 atomic_set(&conf->reshape_stripes, 0);
5928 spin_lock_irq(&conf->device_lock);
5929 conf->previous_raid_disks = conf->raid_disks;
5930 conf->raid_disks += mddev->delta_disks;
5931 conf->prev_chunk_sectors = conf->chunk_sectors;
5932 conf->chunk_sectors = mddev->new_chunk_sectors;
5933 conf->prev_algo = conf->algorithm;
5934 conf->algorithm = mddev->new_layout;
5936 /* Code that selects data_offset needs to see the generation update
5937 * if reshape_progress has been set - so a memory barrier needed.
5940 if (mddev->reshape_backwards)
5941 conf->reshape_progress = raid5_size(mddev, 0, 0);
5943 conf->reshape_progress = 0;
5944 conf->reshape_safe = conf->reshape_progress;
5945 spin_unlock_irq(&conf->device_lock);
5947 /* Add some new drives, as many as will fit.
5948 * We know there are enough to make the newly sized array work.
5949 * Don't add devices if we are reducing the number of
5950 * devices in the array. This is because it is not possible
5951 * to correctly record the "partially reconstructed" state of
5952 * such devices during the reshape and confusion could result.
5954 if (mddev->delta_disks >= 0) {
5955 rdev_for_each(rdev, mddev)
5956 if (rdev->raid_disk < 0 &&
5957 !test_bit(Faulty, &rdev->flags)) {
5958 if (raid5_add_disk(mddev, rdev) == 0) {
5960 >= conf->previous_raid_disks)
5961 set_bit(In_sync, &rdev->flags);
5963 rdev->recovery_offset = 0;
5965 if (sysfs_link_rdev(mddev, rdev))
5966 /* Failure here is OK */;
5968 } else if (rdev->raid_disk >= conf->previous_raid_disks
5969 && !test_bit(Faulty, &rdev->flags)) {
5970 /* This is a spare that was manually added */
5971 set_bit(In_sync, &rdev->flags);
5974 /* When a reshape changes the number of devices,
5975 * ->degraded is measured against the larger of the
5976 * pre and post number of devices.
5978 spin_lock_irqsave(&conf->device_lock, flags);
5979 mddev->degraded = calc_degraded(conf);
5980 spin_unlock_irqrestore(&conf->device_lock, flags);
5982 mddev->raid_disks = conf->raid_disks;
5983 mddev->reshape_position = conf->reshape_progress;
5984 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5986 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5987 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5988 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5989 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5990 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5992 if (!mddev->sync_thread) {
5993 mddev->recovery = 0;
5994 spin_lock_irq(&conf->device_lock);
5995 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5996 rdev_for_each(rdev, mddev)
5997 rdev->new_data_offset = rdev->data_offset;
5999 conf->reshape_progress = MaxSector;
6000 mddev->reshape_position = MaxSector;
6001 spin_unlock_irq(&conf->device_lock);
6004 conf->reshape_checkpoint = jiffies;
6005 md_wakeup_thread(mddev->sync_thread);
6006 md_new_event(mddev);
6010 /* This is called from the reshape thread and should make any
6011 * changes needed in 'conf'
6013 static void end_reshape(struct r5conf *conf)
6016 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6017 struct md_rdev *rdev;
6019 spin_lock_irq(&conf->device_lock);
6020 conf->previous_raid_disks = conf->raid_disks;
6021 rdev_for_each(rdev, conf->mddev)
6022 rdev->data_offset = rdev->new_data_offset;
6024 conf->reshape_progress = MaxSector;
6025 spin_unlock_irq(&conf->device_lock);
6026 wake_up(&conf->wait_for_overlap);
6028 /* read-ahead size must cover two whole stripes, which is
6029 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6031 if (conf->mddev->queue) {
6032 int data_disks = conf->raid_disks - conf->max_degraded;
6033 int stripe = data_disks * ((conf->chunk_sectors << 9)
6035 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6036 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6041 /* This is called from the raid5d thread with mddev_lock held.
6042 * It makes config changes to the device.
6044 static void raid5_finish_reshape(struct mddev *mddev)
6046 struct r5conf *conf = mddev->private;
6048 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6050 if (mddev->delta_disks > 0) {
6051 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6052 set_capacity(mddev->gendisk, mddev->array_sectors);
6053 revalidate_disk(mddev->gendisk);
6056 spin_lock_irq(&conf->device_lock);
6057 mddev->degraded = calc_degraded(conf);
6058 spin_unlock_irq(&conf->device_lock);
6059 for (d = conf->raid_disks ;
6060 d < conf->raid_disks - mddev->delta_disks;
6062 struct md_rdev *rdev = conf->disks[d].rdev;
6064 clear_bit(In_sync, &rdev->flags);
6065 rdev = conf->disks[d].replacement;
6067 clear_bit(In_sync, &rdev->flags);
6070 mddev->layout = conf->algorithm;
6071 mddev->chunk_sectors = conf->chunk_sectors;
6072 mddev->reshape_position = MaxSector;
6073 mddev->delta_disks = 0;
6074 mddev->reshape_backwards = 0;
6078 static void raid5_quiesce(struct mddev *mddev, int state)
6080 struct r5conf *conf = mddev->private;
6083 case 2: /* resume for a suspend */
6084 wake_up(&conf->wait_for_overlap);
6087 case 1: /* stop all writes */
6088 spin_lock_irq(&conf->device_lock);
6089 /* '2' tells resync/reshape to pause so that all
6090 * active stripes can drain
6093 wait_event_lock_irq(conf->wait_for_stripe,
6094 atomic_read(&conf->active_stripes) == 0 &&
6095 atomic_read(&conf->active_aligned_reads) == 0,
6098 spin_unlock_irq(&conf->device_lock);
6099 /* allow reshape to continue */
6100 wake_up(&conf->wait_for_overlap);
6103 case 0: /* re-enable writes */
6104 spin_lock_irq(&conf->device_lock);
6106 wake_up(&conf->wait_for_stripe);
6107 wake_up(&conf->wait_for_overlap);
6108 spin_unlock_irq(&conf->device_lock);
6114 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6116 struct r0conf *raid0_conf = mddev->private;
6119 /* for raid0 takeover only one zone is supported */
6120 if (raid0_conf->nr_strip_zones > 1) {
6121 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6123 return ERR_PTR(-EINVAL);
6126 sectors = raid0_conf->strip_zone[0].zone_end;
6127 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6128 mddev->dev_sectors = sectors;
6129 mddev->new_level = level;
6130 mddev->new_layout = ALGORITHM_PARITY_N;
6131 mddev->new_chunk_sectors = mddev->chunk_sectors;
6132 mddev->raid_disks += 1;
6133 mddev->delta_disks = 1;
6134 /* make sure it will be not marked as dirty */
6135 mddev->recovery_cp = MaxSector;
6137 return setup_conf(mddev);
6141 static void *raid5_takeover_raid1(struct mddev *mddev)
6145 if (mddev->raid_disks != 2 ||
6146 mddev->degraded > 1)
6147 return ERR_PTR(-EINVAL);
6149 /* Should check if there are write-behind devices? */
6151 chunksect = 64*2; /* 64K by default */
6153 /* The array must be an exact multiple of chunksize */
6154 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6157 if ((chunksect<<9) < STRIPE_SIZE)
6158 /* array size does not allow a suitable chunk size */
6159 return ERR_PTR(-EINVAL);
6161 mddev->new_level = 5;
6162 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6163 mddev->new_chunk_sectors = chunksect;
6165 return setup_conf(mddev);
6168 static void *raid5_takeover_raid6(struct mddev *mddev)
6172 switch (mddev->layout) {
6173 case ALGORITHM_LEFT_ASYMMETRIC_6:
6174 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6176 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6177 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6179 case ALGORITHM_LEFT_SYMMETRIC_6:
6180 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6182 case ALGORITHM_RIGHT_SYMMETRIC_6:
6183 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6185 case ALGORITHM_PARITY_0_6:
6186 new_layout = ALGORITHM_PARITY_0;
6188 case ALGORITHM_PARITY_N:
6189 new_layout = ALGORITHM_PARITY_N;
6192 return ERR_PTR(-EINVAL);
6194 mddev->new_level = 5;
6195 mddev->new_layout = new_layout;
6196 mddev->delta_disks = -1;
6197 mddev->raid_disks -= 1;
6198 return setup_conf(mddev);
6202 static int raid5_check_reshape(struct mddev *mddev)
6204 /* For a 2-drive array, the layout and chunk size can be changed
6205 * immediately as not restriping is needed.
6206 * For larger arrays we record the new value - after validation
6207 * to be used by a reshape pass.
6209 struct r5conf *conf = mddev->private;
6210 int new_chunk = mddev->new_chunk_sectors;
6212 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6214 if (new_chunk > 0) {
6215 if (!is_power_of_2(new_chunk))
6217 if (new_chunk < (PAGE_SIZE>>9))
6219 if (mddev->array_sectors & (new_chunk-1))
6220 /* not factor of array size */
6224 /* They look valid */
6226 if (mddev->raid_disks == 2) {
6227 /* can make the change immediately */
6228 if (mddev->new_layout >= 0) {
6229 conf->algorithm = mddev->new_layout;
6230 mddev->layout = mddev->new_layout;
6232 if (new_chunk > 0) {
6233 conf->chunk_sectors = new_chunk ;
6234 mddev->chunk_sectors = new_chunk;
6236 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6237 md_wakeup_thread(mddev->thread);
6239 return check_reshape(mddev);
6242 static int raid6_check_reshape(struct mddev *mddev)
6244 int new_chunk = mddev->new_chunk_sectors;
6246 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6248 if (new_chunk > 0) {
6249 if (!is_power_of_2(new_chunk))
6251 if (new_chunk < (PAGE_SIZE >> 9))
6253 if (mddev->array_sectors & (new_chunk-1))
6254 /* not factor of array size */
6258 /* They look valid */
6259 return check_reshape(mddev);
6262 static void *raid5_takeover(struct mddev *mddev)
6264 /* raid5 can take over:
6265 * raid0 - if there is only one strip zone - make it a raid4 layout
6266 * raid1 - if there are two drives. We need to know the chunk size
6267 * raid4 - trivial - just use a raid4 layout.
6268 * raid6 - Providing it is a *_6 layout
6270 if (mddev->level == 0)
6271 return raid45_takeover_raid0(mddev, 5);
6272 if (mddev->level == 1)
6273 return raid5_takeover_raid1(mddev);
6274 if (mddev->level == 4) {
6275 mddev->new_layout = ALGORITHM_PARITY_N;
6276 mddev->new_level = 5;
6277 return setup_conf(mddev);
6279 if (mddev->level == 6)
6280 return raid5_takeover_raid6(mddev);
6282 return ERR_PTR(-EINVAL);
6285 static void *raid4_takeover(struct mddev *mddev)
6287 /* raid4 can take over:
6288 * raid0 - if there is only one strip zone
6289 * raid5 - if layout is right
6291 if (mddev->level == 0)
6292 return raid45_takeover_raid0(mddev, 4);
6293 if (mddev->level == 5 &&
6294 mddev->layout == ALGORITHM_PARITY_N) {
6295 mddev->new_layout = 0;
6296 mddev->new_level = 4;
6297 return setup_conf(mddev);
6299 return ERR_PTR(-EINVAL);
6302 static struct md_personality raid5_personality;
6304 static void *raid6_takeover(struct mddev *mddev)
6306 /* Currently can only take over a raid5. We map the
6307 * personality to an equivalent raid6 personality
6308 * with the Q block at the end.
6312 if (mddev->pers != &raid5_personality)
6313 return ERR_PTR(-EINVAL);
6314 if (mddev->degraded > 1)
6315 return ERR_PTR(-EINVAL);
6316 if (mddev->raid_disks > 253)
6317 return ERR_PTR(-EINVAL);
6318 if (mddev->raid_disks < 3)
6319 return ERR_PTR(-EINVAL);
6321 switch (mddev->layout) {
6322 case ALGORITHM_LEFT_ASYMMETRIC:
6323 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6325 case ALGORITHM_RIGHT_ASYMMETRIC:
6326 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6328 case ALGORITHM_LEFT_SYMMETRIC:
6329 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6331 case ALGORITHM_RIGHT_SYMMETRIC:
6332 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6334 case ALGORITHM_PARITY_0:
6335 new_layout = ALGORITHM_PARITY_0_6;
6337 case ALGORITHM_PARITY_N:
6338 new_layout = ALGORITHM_PARITY_N;
6341 return ERR_PTR(-EINVAL);
6343 mddev->new_level = 6;
6344 mddev->new_layout = new_layout;
6345 mddev->delta_disks = 1;
6346 mddev->raid_disks += 1;
6347 return setup_conf(mddev);
6351 static struct md_personality raid6_personality =
6355 .owner = THIS_MODULE,
6356 .make_request = make_request,
6360 .error_handler = error,
6361 .hot_add_disk = raid5_add_disk,
6362 .hot_remove_disk= raid5_remove_disk,
6363 .spare_active = raid5_spare_active,
6364 .sync_request = sync_request,
6365 .resize = raid5_resize,
6367 .check_reshape = raid6_check_reshape,
6368 .start_reshape = raid5_start_reshape,
6369 .finish_reshape = raid5_finish_reshape,
6370 .quiesce = raid5_quiesce,
6371 .takeover = raid6_takeover,
6373 static struct md_personality raid5_personality =
6377 .owner = THIS_MODULE,
6378 .make_request = make_request,
6382 .error_handler = error,
6383 .hot_add_disk = raid5_add_disk,
6384 .hot_remove_disk= raid5_remove_disk,
6385 .spare_active = raid5_spare_active,
6386 .sync_request = sync_request,
6387 .resize = raid5_resize,
6389 .check_reshape = raid5_check_reshape,
6390 .start_reshape = raid5_start_reshape,
6391 .finish_reshape = raid5_finish_reshape,
6392 .quiesce = raid5_quiesce,
6393 .takeover = raid5_takeover,
6396 static struct md_personality raid4_personality =
6400 .owner = THIS_MODULE,
6401 .make_request = make_request,
6405 .error_handler = error,
6406 .hot_add_disk = raid5_add_disk,
6407 .hot_remove_disk= raid5_remove_disk,
6408 .spare_active = raid5_spare_active,
6409 .sync_request = sync_request,
6410 .resize = raid5_resize,
6412 .check_reshape = raid5_check_reshape,
6413 .start_reshape = raid5_start_reshape,
6414 .finish_reshape = raid5_finish_reshape,
6415 .quiesce = raid5_quiesce,
6416 .takeover = raid4_takeover,
6419 static int __init raid5_init(void)
6421 register_md_personality(&raid6_personality);
6422 register_md_personality(&raid5_personality);
6423 register_md_personality(&raid4_personality);
6427 static void raid5_exit(void)
6429 unregister_md_personality(&raid6_personality);
6430 unregister_md_personality(&raid5_personality);
6431 unregister_md_personality(&raid4_personality);
6434 module_init(raid5_init);
6435 module_exit(raid5_exit);
6436 MODULE_LICENSE("GPL");
6437 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6438 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6439 MODULE_ALIAS("md-raid5");
6440 MODULE_ALIAS("md-raid4");
6441 MODULE_ALIAS("md-level-5");
6442 MODULE_ALIAS("md-level-4");
6443 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6444 MODULE_ALIAS("md-raid6");
6445 MODULE_ALIAS("md-level-6");
6447 /* This used to be two separate modules, they were: */
6448 MODULE_ALIAS("raid5");
6449 MODULE_ALIAS("raid6");