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_phys_segments(struct bio *bio)
104 return bio->bi_phys_segments & 0xffff;
107 static inline int raid5_bi_hw_segments(struct bio *bio)
109 return (bio->bi_phys_segments >> 16) & 0xffff;
112 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
114 --bio->bi_phys_segments;
115 return raid5_bi_phys_segments(bio);
118 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
120 unsigned short val = raid5_bi_hw_segments(bio);
123 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
127 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
129 bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
132 /* Find first data disk in a raid6 stripe */
133 static inline int raid6_d0(struct stripe_head *sh)
136 /* ddf always start from first device */
138 /* md starts just after Q block */
139 if (sh->qd_idx == sh->disks - 1)
142 return sh->qd_idx + 1;
144 static inline int raid6_next_disk(int disk, int raid_disks)
147 return (disk < raid_disks) ? disk : 0;
150 /* When walking through the disks in a raid5, starting at raid6_d0,
151 * We need to map each disk to a 'slot', where the data disks are slot
152 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153 * is raid_disks-1. This help does that mapping.
155 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
156 int *count, int syndrome_disks)
162 if (idx == sh->pd_idx)
163 return syndrome_disks;
164 if (idx == sh->qd_idx)
165 return syndrome_disks + 1;
171 static void return_io(struct bio *return_bi)
173 struct bio *bi = return_bi;
176 return_bi = bi->bi_next;
184 static void print_raid5_conf (struct r5conf *conf);
186 static int stripe_operations_active(struct stripe_head *sh)
188 return sh->check_state || sh->reconstruct_state ||
189 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
190 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
193 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
195 if (atomic_dec_and_test(&sh->count)) {
196 BUG_ON(!list_empty(&sh->lru));
197 BUG_ON(atomic_read(&conf->active_stripes)==0);
198 if (test_bit(STRIPE_HANDLE, &sh->state)) {
199 if (test_bit(STRIPE_DELAYED, &sh->state))
200 list_add_tail(&sh->lru, &conf->delayed_list);
201 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
202 sh->bm_seq - conf->seq_write > 0)
203 list_add_tail(&sh->lru, &conf->bitmap_list);
205 clear_bit(STRIPE_BIT_DELAY, &sh->state);
206 list_add_tail(&sh->lru, &conf->handle_list);
208 md_wakeup_thread(conf->mddev->thread);
210 BUG_ON(stripe_operations_active(sh));
211 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
212 atomic_dec(&conf->preread_active_stripes);
213 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
214 md_wakeup_thread(conf->mddev->thread);
216 atomic_dec(&conf->active_stripes);
217 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
218 list_add_tail(&sh->lru, &conf->inactive_list);
219 wake_up(&conf->wait_for_stripe);
220 if (conf->retry_read_aligned)
221 md_wakeup_thread(conf->mddev->thread);
227 static void release_stripe(struct stripe_head *sh)
229 struct r5conf *conf = sh->raid_conf;
232 spin_lock_irqsave(&conf->device_lock, flags);
233 __release_stripe(conf, sh);
234 spin_unlock_irqrestore(&conf->device_lock, flags);
237 static inline void remove_hash(struct stripe_head *sh)
239 pr_debug("remove_hash(), stripe %llu\n",
240 (unsigned long long)sh->sector);
242 hlist_del_init(&sh->hash);
245 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
247 struct hlist_head *hp = stripe_hash(conf, sh->sector);
249 pr_debug("insert_hash(), stripe %llu\n",
250 (unsigned long long)sh->sector);
252 hlist_add_head(&sh->hash, hp);
256 /* find an idle stripe, make sure it is unhashed, and return it. */
257 static struct stripe_head *get_free_stripe(struct r5conf *conf)
259 struct stripe_head *sh = NULL;
260 struct list_head *first;
262 if (list_empty(&conf->inactive_list))
264 first = conf->inactive_list.next;
265 sh = list_entry(first, struct stripe_head, lru);
266 list_del_init(first);
268 atomic_inc(&conf->active_stripes);
273 static void shrink_buffers(struct stripe_head *sh)
277 int num = sh->raid_conf->pool_size;
279 for (i = 0; i < num ; i++) {
283 sh->dev[i].page = NULL;
288 static int grow_buffers(struct stripe_head *sh)
291 int num = sh->raid_conf->pool_size;
293 for (i = 0; i < num; i++) {
296 if (!(page = alloc_page(GFP_KERNEL))) {
299 sh->dev[i].page = page;
304 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
305 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
306 struct stripe_head *sh);
308 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
310 struct r5conf *conf = sh->raid_conf;
313 BUG_ON(atomic_read(&sh->count) != 0);
314 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
315 BUG_ON(stripe_operations_active(sh));
317 pr_debug("init_stripe called, stripe %llu\n",
318 (unsigned long long)sh->sector);
322 sh->generation = conf->generation - previous;
323 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
325 stripe_set_idx(sector, conf, previous, sh);
329 for (i = sh->disks; i--; ) {
330 struct r5dev *dev = &sh->dev[i];
332 if (dev->toread || dev->read || dev->towrite || dev->written ||
333 test_bit(R5_LOCKED, &dev->flags)) {
334 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
335 (unsigned long long)sh->sector, i, dev->toread,
336 dev->read, dev->towrite, dev->written,
337 test_bit(R5_LOCKED, &dev->flags));
341 raid5_build_block(sh, i, previous);
343 insert_hash(conf, sh);
346 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
349 struct stripe_head *sh;
350 struct hlist_node *hn;
352 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
353 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
354 if (sh->sector == sector && sh->generation == generation)
356 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
361 * Need to check if array has failed when deciding whether to:
363 * - remove non-faulty devices
366 * This determination is simple when no reshape is happening.
367 * However if there is a reshape, we need to carefully check
368 * both the before and after sections.
369 * This is because some failed devices may only affect one
370 * of the two sections, and some non-in_sync devices may
371 * be insync in the section most affected by failed devices.
373 static int has_failed(struct r5conf *conf)
377 if (conf->mddev->reshape_position == MaxSector)
378 return conf->mddev->degraded > conf->max_degraded;
382 for (i = 0; i < conf->previous_raid_disks; i++) {
383 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
384 if (!rdev || test_bit(Faulty, &rdev->flags))
386 else if (test_bit(In_sync, &rdev->flags))
389 /* not in-sync or faulty.
390 * If the reshape increases the number of devices,
391 * this is being recovered by the reshape, so
392 * this 'previous' section is not in_sync.
393 * If the number of devices is being reduced however,
394 * the device can only be part of the array if
395 * we are reverting a reshape, so this section will
398 if (conf->raid_disks >= conf->previous_raid_disks)
402 if (degraded > conf->max_degraded)
406 for (i = 0; i < conf->raid_disks; i++) {
407 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
408 if (!rdev || test_bit(Faulty, &rdev->flags))
410 else if (test_bit(In_sync, &rdev->flags))
413 /* not in-sync or faulty.
414 * If reshape increases the number of devices, this
415 * section has already been recovered, else it
416 * almost certainly hasn't.
418 if (conf->raid_disks <= conf->previous_raid_disks)
422 if (degraded > conf->max_degraded)
427 static struct stripe_head *
428 get_active_stripe(struct r5conf *conf, sector_t sector,
429 int previous, int noblock, int noquiesce)
431 struct stripe_head *sh;
433 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
435 spin_lock_irq(&conf->device_lock);
438 wait_event_lock_irq(conf->wait_for_stripe,
439 conf->quiesce == 0 || noquiesce,
440 conf->device_lock, /* nothing */);
441 sh = __find_stripe(conf, sector, conf->generation - previous);
443 if (!conf->inactive_blocked)
444 sh = get_free_stripe(conf);
445 if (noblock && sh == NULL)
448 conf->inactive_blocked = 1;
449 wait_event_lock_irq(conf->wait_for_stripe,
450 !list_empty(&conf->inactive_list) &&
451 (atomic_read(&conf->active_stripes)
452 < (conf->max_nr_stripes *3/4)
453 || !conf->inactive_blocked),
456 conf->inactive_blocked = 0;
458 init_stripe(sh, sector, previous);
460 if (atomic_read(&sh->count)) {
461 BUG_ON(!list_empty(&sh->lru)
462 && !test_bit(STRIPE_EXPANDING, &sh->state));
464 if (!test_bit(STRIPE_HANDLE, &sh->state))
465 atomic_inc(&conf->active_stripes);
466 if (list_empty(&sh->lru) &&
467 !test_bit(STRIPE_EXPANDING, &sh->state))
469 list_del_init(&sh->lru);
472 } while (sh == NULL);
475 atomic_inc(&sh->count);
477 spin_unlock_irq(&conf->device_lock);
482 raid5_end_read_request(struct bio *bi, int error);
484 raid5_end_write_request(struct bio *bi, int error);
486 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
488 struct r5conf *conf = sh->raid_conf;
489 int i, disks = sh->disks;
493 for (i = disks; i--; ) {
496 struct md_rdev *rdev;
497 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
498 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
502 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
507 bi = &sh->dev[i].req;
511 bi->bi_end_io = raid5_end_write_request;
513 bi->bi_end_io = raid5_end_read_request;
516 rdev = rcu_dereference(conf->disks[i].rdev);
517 if (rdev && test_bit(Faulty, &rdev->flags))
520 atomic_inc(&rdev->nr_pending);
523 /* We have already checked bad blocks for reads. Now
524 * need to check for writes.
526 while ((rw & WRITE) && rdev &&
527 test_bit(WriteErrorSeen, &rdev->flags)) {
530 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
531 &first_bad, &bad_sectors);
536 set_bit(BlockedBadBlocks, &rdev->flags);
537 if (!conf->mddev->external &&
538 conf->mddev->flags) {
539 /* It is very unlikely, but we might
540 * still need to write out the
541 * bad block log - better give it
543 md_check_recovery(conf->mddev);
545 md_wait_for_blocked_rdev(rdev, conf->mddev);
547 /* Acknowledged bad block - skip the write */
548 rdev_dec_pending(rdev, conf->mddev);
554 if (s->syncing || s->expanding || s->expanded)
555 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
557 set_bit(STRIPE_IO_STARTED, &sh->state);
559 bi->bi_bdev = rdev->bdev;
560 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
561 __func__, (unsigned long long)sh->sector,
563 atomic_inc(&sh->count);
564 bi->bi_sector = sh->sector + rdev->data_offset;
565 bi->bi_flags = 1 << BIO_UPTODATE;
569 bi->bi_io_vec = &sh->dev[i].vec;
570 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
571 bi->bi_io_vec[0].bv_offset = 0;
572 bi->bi_size = STRIPE_SIZE;
574 generic_make_request(bi);
577 set_bit(STRIPE_DEGRADED, &sh->state);
578 pr_debug("skip op %ld on disc %d for sector %llu\n",
579 bi->bi_rw, i, (unsigned long long)sh->sector);
580 clear_bit(R5_LOCKED, &sh->dev[i].flags);
581 set_bit(STRIPE_HANDLE, &sh->state);
586 static struct dma_async_tx_descriptor *
587 async_copy_data(int frombio, struct bio *bio, struct page *page,
588 sector_t sector, struct dma_async_tx_descriptor *tx)
591 struct page *bio_page;
594 struct async_submit_ctl submit;
595 enum async_tx_flags flags = 0;
597 if (bio->bi_sector >= sector)
598 page_offset = (signed)(bio->bi_sector - sector) * 512;
600 page_offset = (signed)(sector - bio->bi_sector) * -512;
603 flags |= ASYNC_TX_FENCE;
604 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
606 bio_for_each_segment(bvl, bio, i) {
607 int len = bvl->bv_len;
611 if (page_offset < 0) {
612 b_offset = -page_offset;
613 page_offset += b_offset;
617 if (len > 0 && page_offset + len > STRIPE_SIZE)
618 clen = STRIPE_SIZE - page_offset;
623 b_offset += bvl->bv_offset;
624 bio_page = bvl->bv_page;
626 tx = async_memcpy(page, bio_page, page_offset,
627 b_offset, clen, &submit);
629 tx = async_memcpy(bio_page, page, b_offset,
630 page_offset, clen, &submit);
632 /* chain the operations */
633 submit.depend_tx = tx;
635 if (clen < len) /* hit end of page */
643 static void ops_complete_biofill(void *stripe_head_ref)
645 struct stripe_head *sh = stripe_head_ref;
646 struct bio *return_bi = NULL;
647 struct r5conf *conf = sh->raid_conf;
650 pr_debug("%s: stripe %llu\n", __func__,
651 (unsigned long long)sh->sector);
653 /* clear completed biofills */
654 spin_lock_irq(&conf->device_lock);
655 for (i = sh->disks; i--; ) {
656 struct r5dev *dev = &sh->dev[i];
658 /* acknowledge completion of a biofill operation */
659 /* and check if we need to reply to a read request,
660 * new R5_Wantfill requests are held off until
661 * !STRIPE_BIOFILL_RUN
663 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
664 struct bio *rbi, *rbi2;
669 while (rbi && rbi->bi_sector <
670 dev->sector + STRIPE_SECTORS) {
671 rbi2 = r5_next_bio(rbi, dev->sector);
672 if (!raid5_dec_bi_phys_segments(rbi)) {
673 rbi->bi_next = return_bi;
680 spin_unlock_irq(&conf->device_lock);
681 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
683 return_io(return_bi);
685 set_bit(STRIPE_HANDLE, &sh->state);
689 static void ops_run_biofill(struct stripe_head *sh)
691 struct dma_async_tx_descriptor *tx = NULL;
692 struct r5conf *conf = sh->raid_conf;
693 struct async_submit_ctl submit;
696 pr_debug("%s: stripe %llu\n", __func__,
697 (unsigned long long)sh->sector);
699 for (i = sh->disks; i--; ) {
700 struct r5dev *dev = &sh->dev[i];
701 if (test_bit(R5_Wantfill, &dev->flags)) {
703 spin_lock_irq(&conf->device_lock);
704 dev->read = rbi = dev->toread;
706 spin_unlock_irq(&conf->device_lock);
707 while (rbi && rbi->bi_sector <
708 dev->sector + STRIPE_SECTORS) {
709 tx = async_copy_data(0, rbi, dev->page,
711 rbi = r5_next_bio(rbi, dev->sector);
716 atomic_inc(&sh->count);
717 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
718 async_trigger_callback(&submit);
721 static void mark_target_uptodate(struct stripe_head *sh, int target)
728 tgt = &sh->dev[target];
729 set_bit(R5_UPTODATE, &tgt->flags);
730 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
731 clear_bit(R5_Wantcompute, &tgt->flags);
734 static void ops_complete_compute(void *stripe_head_ref)
736 struct stripe_head *sh = stripe_head_ref;
738 pr_debug("%s: stripe %llu\n", __func__,
739 (unsigned long long)sh->sector);
741 /* mark the computed target(s) as uptodate */
742 mark_target_uptodate(sh, sh->ops.target);
743 mark_target_uptodate(sh, sh->ops.target2);
745 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
746 if (sh->check_state == check_state_compute_run)
747 sh->check_state = check_state_compute_result;
748 set_bit(STRIPE_HANDLE, &sh->state);
752 /* return a pointer to the address conversion region of the scribble buffer */
753 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
754 struct raid5_percpu *percpu)
756 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
759 static struct dma_async_tx_descriptor *
760 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
762 int disks = sh->disks;
763 struct page **xor_srcs = percpu->scribble;
764 int target = sh->ops.target;
765 struct r5dev *tgt = &sh->dev[target];
766 struct page *xor_dest = tgt->page;
768 struct dma_async_tx_descriptor *tx;
769 struct async_submit_ctl submit;
772 pr_debug("%s: stripe %llu block: %d\n",
773 __func__, (unsigned long long)sh->sector, target);
774 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
776 for (i = disks; i--; )
778 xor_srcs[count++] = sh->dev[i].page;
780 atomic_inc(&sh->count);
782 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
783 ops_complete_compute, sh, to_addr_conv(sh, percpu));
784 if (unlikely(count == 1))
785 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
787 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
792 /* set_syndrome_sources - populate source buffers for gen_syndrome
793 * @srcs - (struct page *) array of size sh->disks
794 * @sh - stripe_head to parse
796 * Populates srcs in proper layout order for the stripe and returns the
797 * 'count' of sources to be used in a call to async_gen_syndrome. The P
798 * destination buffer is recorded in srcs[count] and the Q destination
799 * is recorded in srcs[count+1]].
801 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
803 int disks = sh->disks;
804 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
805 int d0_idx = raid6_d0(sh);
809 for (i = 0; i < disks; i++)
815 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
817 srcs[slot] = sh->dev[i].page;
818 i = raid6_next_disk(i, disks);
819 } while (i != d0_idx);
821 return syndrome_disks;
824 static struct dma_async_tx_descriptor *
825 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
827 int disks = sh->disks;
828 struct page **blocks = percpu->scribble;
830 int qd_idx = sh->qd_idx;
831 struct dma_async_tx_descriptor *tx;
832 struct async_submit_ctl submit;
838 if (sh->ops.target < 0)
839 target = sh->ops.target2;
840 else if (sh->ops.target2 < 0)
841 target = sh->ops.target;
843 /* we should only have one valid target */
846 pr_debug("%s: stripe %llu block: %d\n",
847 __func__, (unsigned long long)sh->sector, target);
849 tgt = &sh->dev[target];
850 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
853 atomic_inc(&sh->count);
855 if (target == qd_idx) {
856 count = set_syndrome_sources(blocks, sh);
857 blocks[count] = NULL; /* regenerating p is not necessary */
858 BUG_ON(blocks[count+1] != dest); /* q should already be set */
859 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
860 ops_complete_compute, sh,
861 to_addr_conv(sh, percpu));
862 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
864 /* Compute any data- or p-drive using XOR */
866 for (i = disks; i-- ; ) {
867 if (i == target || i == qd_idx)
869 blocks[count++] = sh->dev[i].page;
872 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
873 NULL, ops_complete_compute, sh,
874 to_addr_conv(sh, percpu));
875 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
881 static struct dma_async_tx_descriptor *
882 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
884 int i, count, disks = sh->disks;
885 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
886 int d0_idx = raid6_d0(sh);
887 int faila = -1, failb = -1;
888 int target = sh->ops.target;
889 int target2 = sh->ops.target2;
890 struct r5dev *tgt = &sh->dev[target];
891 struct r5dev *tgt2 = &sh->dev[target2];
892 struct dma_async_tx_descriptor *tx;
893 struct page **blocks = percpu->scribble;
894 struct async_submit_ctl submit;
896 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
897 __func__, (unsigned long long)sh->sector, target, target2);
898 BUG_ON(target < 0 || target2 < 0);
899 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
900 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
902 /* we need to open-code set_syndrome_sources to handle the
903 * slot number conversion for 'faila' and 'failb'
905 for (i = 0; i < disks ; i++)
910 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
912 blocks[slot] = sh->dev[i].page;
918 i = raid6_next_disk(i, disks);
919 } while (i != d0_idx);
921 BUG_ON(faila == failb);
924 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
925 __func__, (unsigned long long)sh->sector, faila, failb);
927 atomic_inc(&sh->count);
929 if (failb == syndrome_disks+1) {
930 /* Q disk is one of the missing disks */
931 if (faila == syndrome_disks) {
932 /* Missing P+Q, just recompute */
933 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
934 ops_complete_compute, sh,
935 to_addr_conv(sh, percpu));
936 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
937 STRIPE_SIZE, &submit);
941 int qd_idx = sh->qd_idx;
943 /* Missing D+Q: recompute D from P, then recompute Q */
944 if (target == qd_idx)
945 data_target = target2;
947 data_target = target;
950 for (i = disks; i-- ; ) {
951 if (i == data_target || i == qd_idx)
953 blocks[count++] = sh->dev[i].page;
955 dest = sh->dev[data_target].page;
956 init_async_submit(&submit,
957 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
959 to_addr_conv(sh, percpu));
960 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
963 count = set_syndrome_sources(blocks, sh);
964 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
965 ops_complete_compute, sh,
966 to_addr_conv(sh, percpu));
967 return async_gen_syndrome(blocks, 0, count+2,
968 STRIPE_SIZE, &submit);
971 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
972 ops_complete_compute, sh,
973 to_addr_conv(sh, percpu));
974 if (failb == syndrome_disks) {
975 /* We're missing D+P. */
976 return async_raid6_datap_recov(syndrome_disks+2,
980 /* We're missing D+D. */
981 return async_raid6_2data_recov(syndrome_disks+2,
982 STRIPE_SIZE, faila, failb,
989 static void ops_complete_prexor(void *stripe_head_ref)
991 struct stripe_head *sh = stripe_head_ref;
993 pr_debug("%s: stripe %llu\n", __func__,
994 (unsigned long long)sh->sector);
997 static struct dma_async_tx_descriptor *
998 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
999 struct dma_async_tx_descriptor *tx)
1001 int disks = sh->disks;
1002 struct page **xor_srcs = percpu->scribble;
1003 int count = 0, pd_idx = sh->pd_idx, i;
1004 struct async_submit_ctl submit;
1006 /* existing parity data subtracted */
1007 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1009 pr_debug("%s: stripe %llu\n", __func__,
1010 (unsigned long long)sh->sector);
1012 for (i = disks; i--; ) {
1013 struct r5dev *dev = &sh->dev[i];
1014 /* Only process blocks that are known to be uptodate */
1015 if (test_bit(R5_Wantdrain, &dev->flags))
1016 xor_srcs[count++] = dev->page;
1019 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1020 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1021 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1026 static struct dma_async_tx_descriptor *
1027 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1029 int disks = sh->disks;
1032 pr_debug("%s: stripe %llu\n", __func__,
1033 (unsigned long long)sh->sector);
1035 for (i = disks; i--; ) {
1036 struct r5dev *dev = &sh->dev[i];
1039 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1042 spin_lock_irq(&sh->raid_conf->device_lock);
1043 chosen = dev->towrite;
1044 dev->towrite = NULL;
1045 BUG_ON(dev->written);
1046 wbi = dev->written = chosen;
1047 spin_unlock_irq(&sh->raid_conf->device_lock);
1049 while (wbi && wbi->bi_sector <
1050 dev->sector + STRIPE_SECTORS) {
1051 if (wbi->bi_rw & REQ_FUA)
1052 set_bit(R5_WantFUA, &dev->flags);
1053 tx = async_copy_data(1, wbi, dev->page,
1055 wbi = r5_next_bio(wbi, dev->sector);
1063 static void ops_complete_reconstruct(void *stripe_head_ref)
1065 struct stripe_head *sh = stripe_head_ref;
1066 int disks = sh->disks;
1067 int pd_idx = sh->pd_idx;
1068 int qd_idx = sh->qd_idx;
1072 pr_debug("%s: stripe %llu\n", __func__,
1073 (unsigned long long)sh->sector);
1075 for (i = disks; i--; )
1076 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1078 for (i = disks; i--; ) {
1079 struct r5dev *dev = &sh->dev[i];
1081 if (dev->written || i == pd_idx || i == qd_idx) {
1082 set_bit(R5_UPTODATE, &dev->flags);
1084 set_bit(R5_WantFUA, &dev->flags);
1088 if (sh->reconstruct_state == reconstruct_state_drain_run)
1089 sh->reconstruct_state = reconstruct_state_drain_result;
1090 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1091 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1093 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1094 sh->reconstruct_state = reconstruct_state_result;
1097 set_bit(STRIPE_HANDLE, &sh->state);
1102 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1103 struct dma_async_tx_descriptor *tx)
1105 int disks = sh->disks;
1106 struct page **xor_srcs = percpu->scribble;
1107 struct async_submit_ctl submit;
1108 int count = 0, pd_idx = sh->pd_idx, i;
1109 struct page *xor_dest;
1111 unsigned long flags;
1113 pr_debug("%s: stripe %llu\n", __func__,
1114 (unsigned long long)sh->sector);
1116 /* check if prexor is active which means only process blocks
1117 * that are part of a read-modify-write (written)
1119 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1121 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1122 for (i = disks; i--; ) {
1123 struct r5dev *dev = &sh->dev[i];
1125 xor_srcs[count++] = dev->page;
1128 xor_dest = sh->dev[pd_idx].page;
1129 for (i = disks; i--; ) {
1130 struct r5dev *dev = &sh->dev[i];
1132 xor_srcs[count++] = dev->page;
1136 /* 1/ if we prexor'd then the dest is reused as a source
1137 * 2/ if we did not prexor then we are redoing the parity
1138 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1139 * for the synchronous xor case
1141 flags = ASYNC_TX_ACK |
1142 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1144 atomic_inc(&sh->count);
1146 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1147 to_addr_conv(sh, percpu));
1148 if (unlikely(count == 1))
1149 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1151 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1155 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1156 struct dma_async_tx_descriptor *tx)
1158 struct async_submit_ctl submit;
1159 struct page **blocks = percpu->scribble;
1162 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1164 count = set_syndrome_sources(blocks, sh);
1166 atomic_inc(&sh->count);
1168 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1169 sh, to_addr_conv(sh, percpu));
1170 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1173 static void ops_complete_check(void *stripe_head_ref)
1175 struct stripe_head *sh = stripe_head_ref;
1177 pr_debug("%s: stripe %llu\n", __func__,
1178 (unsigned long long)sh->sector);
1180 sh->check_state = check_state_check_result;
1181 set_bit(STRIPE_HANDLE, &sh->state);
1185 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1187 int disks = sh->disks;
1188 int pd_idx = sh->pd_idx;
1189 int qd_idx = sh->qd_idx;
1190 struct page *xor_dest;
1191 struct page **xor_srcs = percpu->scribble;
1192 struct dma_async_tx_descriptor *tx;
1193 struct async_submit_ctl submit;
1197 pr_debug("%s: stripe %llu\n", __func__,
1198 (unsigned long long)sh->sector);
1201 xor_dest = sh->dev[pd_idx].page;
1202 xor_srcs[count++] = xor_dest;
1203 for (i = disks; i--; ) {
1204 if (i == pd_idx || i == qd_idx)
1206 xor_srcs[count++] = sh->dev[i].page;
1209 init_async_submit(&submit, 0, NULL, NULL, NULL,
1210 to_addr_conv(sh, percpu));
1211 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1212 &sh->ops.zero_sum_result, &submit);
1214 atomic_inc(&sh->count);
1215 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1216 tx = async_trigger_callback(&submit);
1219 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1221 struct page **srcs = percpu->scribble;
1222 struct async_submit_ctl submit;
1225 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1226 (unsigned long long)sh->sector, checkp);
1228 count = set_syndrome_sources(srcs, sh);
1232 atomic_inc(&sh->count);
1233 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1234 sh, to_addr_conv(sh, percpu));
1235 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1236 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1239 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1241 int overlap_clear = 0, i, disks = sh->disks;
1242 struct dma_async_tx_descriptor *tx = NULL;
1243 struct r5conf *conf = sh->raid_conf;
1244 int level = conf->level;
1245 struct raid5_percpu *percpu;
1249 percpu = per_cpu_ptr(conf->percpu, cpu);
1250 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1251 ops_run_biofill(sh);
1255 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1257 tx = ops_run_compute5(sh, percpu);
1259 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1260 tx = ops_run_compute6_1(sh, percpu);
1262 tx = ops_run_compute6_2(sh, percpu);
1264 /* terminate the chain if reconstruct is not set to be run */
1265 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1269 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1270 tx = ops_run_prexor(sh, percpu, tx);
1272 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1273 tx = ops_run_biodrain(sh, tx);
1277 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1279 ops_run_reconstruct5(sh, percpu, tx);
1281 ops_run_reconstruct6(sh, percpu, tx);
1284 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1285 if (sh->check_state == check_state_run)
1286 ops_run_check_p(sh, percpu);
1287 else if (sh->check_state == check_state_run_q)
1288 ops_run_check_pq(sh, percpu, 0);
1289 else if (sh->check_state == check_state_run_pq)
1290 ops_run_check_pq(sh, percpu, 1);
1296 for (i = disks; i--; ) {
1297 struct r5dev *dev = &sh->dev[i];
1298 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1299 wake_up(&sh->raid_conf->wait_for_overlap);
1304 #ifdef CONFIG_MULTICORE_RAID456
1305 static void async_run_ops(void *param, async_cookie_t cookie)
1307 struct stripe_head *sh = param;
1308 unsigned long ops_request = sh->ops.request;
1310 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1311 wake_up(&sh->ops.wait_for_ops);
1313 __raid_run_ops(sh, ops_request);
1317 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1319 /* since handle_stripe can be called outside of raid5d context
1320 * we need to ensure sh->ops.request is de-staged before another
1323 wait_event(sh->ops.wait_for_ops,
1324 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1325 sh->ops.request = ops_request;
1327 atomic_inc(&sh->count);
1328 async_schedule(async_run_ops, sh);
1331 #define raid_run_ops __raid_run_ops
1334 static int grow_one_stripe(struct r5conf *conf)
1336 struct stripe_head *sh;
1337 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1341 sh->raid_conf = conf;
1342 #ifdef CONFIG_MULTICORE_RAID456
1343 init_waitqueue_head(&sh->ops.wait_for_ops);
1346 if (grow_buffers(sh)) {
1348 kmem_cache_free(conf->slab_cache, sh);
1351 /* we just created an active stripe so... */
1352 atomic_set(&sh->count, 1);
1353 atomic_inc(&conf->active_stripes);
1354 INIT_LIST_HEAD(&sh->lru);
1359 static int grow_stripes(struct r5conf *conf, int num)
1361 struct kmem_cache *sc;
1362 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1364 if (conf->mddev->gendisk)
1365 sprintf(conf->cache_name[0],
1366 "raid%d-%s", conf->level, mdname(conf->mddev));
1368 sprintf(conf->cache_name[0],
1369 "raid%d-%p", conf->level, conf->mddev);
1370 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1372 conf->active_name = 0;
1373 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1374 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1378 conf->slab_cache = sc;
1379 conf->pool_size = devs;
1381 if (!grow_one_stripe(conf))
1387 * scribble_len - return the required size of the scribble region
1388 * @num - total number of disks in the array
1390 * The size must be enough to contain:
1391 * 1/ a struct page pointer for each device in the array +2
1392 * 2/ room to convert each entry in (1) to its corresponding dma
1393 * (dma_map_page()) or page (page_address()) address.
1395 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1396 * calculate over all devices (not just the data blocks), using zeros in place
1397 * of the P and Q blocks.
1399 static size_t scribble_len(int num)
1403 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1408 static int resize_stripes(struct r5conf *conf, int newsize)
1410 /* Make all the stripes able to hold 'newsize' devices.
1411 * New slots in each stripe get 'page' set to a new page.
1413 * This happens in stages:
1414 * 1/ create a new kmem_cache and allocate the required number of
1416 * 2/ gather all the old stripe_heads and tranfer the pages across
1417 * to the new stripe_heads. This will have the side effect of
1418 * freezing the array as once all stripe_heads have been collected,
1419 * no IO will be possible. Old stripe heads are freed once their
1420 * pages have been transferred over, and the old kmem_cache is
1421 * freed when all stripes are done.
1422 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1423 * we simple return a failre status - no need to clean anything up.
1424 * 4/ allocate new pages for the new slots in the new stripe_heads.
1425 * If this fails, we don't bother trying the shrink the
1426 * stripe_heads down again, we just leave them as they are.
1427 * As each stripe_head is processed the new one is released into
1430 * Once step2 is started, we cannot afford to wait for a write,
1431 * so we use GFP_NOIO allocations.
1433 struct stripe_head *osh, *nsh;
1434 LIST_HEAD(newstripes);
1435 struct disk_info *ndisks;
1438 struct kmem_cache *sc;
1441 if (newsize <= conf->pool_size)
1442 return 0; /* never bother to shrink */
1444 err = md_allow_write(conf->mddev);
1449 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1450 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1455 for (i = conf->max_nr_stripes; i; i--) {
1456 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1460 nsh->raid_conf = conf;
1461 #ifdef CONFIG_MULTICORE_RAID456
1462 init_waitqueue_head(&nsh->ops.wait_for_ops);
1465 list_add(&nsh->lru, &newstripes);
1468 /* didn't get enough, give up */
1469 while (!list_empty(&newstripes)) {
1470 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1471 list_del(&nsh->lru);
1472 kmem_cache_free(sc, nsh);
1474 kmem_cache_destroy(sc);
1477 /* Step 2 - Must use GFP_NOIO now.
1478 * OK, we have enough stripes, start collecting inactive
1479 * stripes and copying them over
1481 list_for_each_entry(nsh, &newstripes, lru) {
1482 spin_lock_irq(&conf->device_lock);
1483 wait_event_lock_irq(conf->wait_for_stripe,
1484 !list_empty(&conf->inactive_list),
1487 osh = get_free_stripe(conf);
1488 spin_unlock_irq(&conf->device_lock);
1489 atomic_set(&nsh->count, 1);
1490 for(i=0; i<conf->pool_size; i++)
1491 nsh->dev[i].page = osh->dev[i].page;
1492 for( ; i<newsize; i++)
1493 nsh->dev[i].page = NULL;
1494 kmem_cache_free(conf->slab_cache, osh);
1496 kmem_cache_destroy(conf->slab_cache);
1499 * At this point, we are holding all the stripes so the array
1500 * is completely stalled, so now is a good time to resize
1501 * conf->disks and the scribble region
1503 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1505 for (i=0; i<conf->raid_disks; i++)
1506 ndisks[i] = conf->disks[i];
1508 conf->disks = ndisks;
1513 conf->scribble_len = scribble_len(newsize);
1514 for_each_present_cpu(cpu) {
1515 struct raid5_percpu *percpu;
1518 percpu = per_cpu_ptr(conf->percpu, cpu);
1519 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1522 kfree(percpu->scribble);
1523 percpu->scribble = scribble;
1531 /* Step 4, return new stripes to service */
1532 while(!list_empty(&newstripes)) {
1533 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1534 list_del_init(&nsh->lru);
1536 for (i=conf->raid_disks; i < newsize; i++)
1537 if (nsh->dev[i].page == NULL) {
1538 struct page *p = alloc_page(GFP_NOIO);
1539 nsh->dev[i].page = p;
1543 release_stripe(nsh);
1545 /* critical section pass, GFP_NOIO no longer needed */
1547 conf->slab_cache = sc;
1548 conf->active_name = 1-conf->active_name;
1549 conf->pool_size = newsize;
1553 static int drop_one_stripe(struct r5conf *conf)
1555 struct stripe_head *sh;
1557 spin_lock_irq(&conf->device_lock);
1558 sh = get_free_stripe(conf);
1559 spin_unlock_irq(&conf->device_lock);
1562 BUG_ON(atomic_read(&sh->count));
1564 kmem_cache_free(conf->slab_cache, sh);
1565 atomic_dec(&conf->active_stripes);
1569 static void shrink_stripes(struct r5conf *conf)
1571 while (drop_one_stripe(conf))
1574 if (conf->slab_cache)
1575 kmem_cache_destroy(conf->slab_cache);
1576 conf->slab_cache = NULL;
1579 static void raid5_end_read_request(struct bio * bi, int error)
1581 struct stripe_head *sh = bi->bi_private;
1582 struct r5conf *conf = sh->raid_conf;
1583 int disks = sh->disks, i;
1584 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1585 char b[BDEVNAME_SIZE];
1586 struct md_rdev *rdev;
1589 for (i=0 ; i<disks; i++)
1590 if (bi == &sh->dev[i].req)
1593 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1594 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1602 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1603 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1604 rdev = conf->disks[i].rdev;
1607 "md/raid:%s: read error corrected"
1608 " (%lu sectors at %llu on %s)\n",
1609 mdname(conf->mddev), STRIPE_SECTORS,
1610 (unsigned long long)(sh->sector
1611 + rdev->data_offset),
1612 bdevname(rdev->bdev, b));
1613 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1614 clear_bit(R5_ReadError, &sh->dev[i].flags);
1615 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1617 if (atomic_read(&conf->disks[i].rdev->read_errors))
1618 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1620 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1622 rdev = conf->disks[i].rdev;
1624 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1625 atomic_inc(&rdev->read_errors);
1626 if (conf->mddev->degraded >= conf->max_degraded)
1629 "md/raid:%s: read error not correctable "
1630 "(sector %llu on %s).\n",
1631 mdname(conf->mddev),
1632 (unsigned long long)(sh->sector
1633 + rdev->data_offset),
1635 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1639 "md/raid:%s: read error NOT corrected!! "
1640 "(sector %llu on %s).\n",
1641 mdname(conf->mddev),
1642 (unsigned long long)(sh->sector
1643 + rdev->data_offset),
1645 else if (atomic_read(&rdev->read_errors)
1646 > conf->max_nr_stripes)
1648 "md/raid:%s: Too many read errors, failing device %s.\n",
1649 mdname(conf->mddev), bdn);
1653 set_bit(R5_ReadError, &sh->dev[i].flags);
1655 clear_bit(R5_ReadError, &sh->dev[i].flags);
1656 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1657 md_error(conf->mddev, rdev);
1660 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1661 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1662 set_bit(STRIPE_HANDLE, &sh->state);
1666 static void raid5_end_write_request(struct bio *bi, int error)
1668 struct stripe_head *sh = bi->bi_private;
1669 struct r5conf *conf = sh->raid_conf;
1670 int disks = sh->disks, i;
1671 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1675 for (i=0 ; i<disks; i++)
1676 if (bi == &sh->dev[i].req)
1679 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1680 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1688 set_bit(WriteErrorSeen, &conf->disks[i].rdev->flags);
1689 set_bit(R5_WriteError, &sh->dev[i].flags);
1690 } else if (is_badblock(conf->disks[i].rdev, sh->sector, STRIPE_SECTORS,
1691 &first_bad, &bad_sectors))
1692 set_bit(R5_MadeGood, &sh->dev[i].flags);
1694 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1696 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1697 set_bit(STRIPE_HANDLE, &sh->state);
1702 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1704 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1706 struct r5dev *dev = &sh->dev[i];
1708 bio_init(&dev->req);
1709 dev->req.bi_io_vec = &dev->vec;
1711 dev->req.bi_max_vecs++;
1712 dev->vec.bv_page = dev->page;
1713 dev->vec.bv_len = STRIPE_SIZE;
1714 dev->vec.bv_offset = 0;
1716 dev->req.bi_sector = sh->sector;
1717 dev->req.bi_private = sh;
1720 dev->sector = compute_blocknr(sh, i, previous);
1723 static void error(struct mddev *mddev, struct md_rdev *rdev)
1725 char b[BDEVNAME_SIZE];
1726 struct r5conf *conf = mddev->private;
1727 pr_debug("raid456: error called\n");
1729 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1730 unsigned long flags;
1731 spin_lock_irqsave(&conf->device_lock, flags);
1733 spin_unlock_irqrestore(&conf->device_lock, flags);
1735 * if recovery was running, make sure it aborts.
1737 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1739 set_bit(Blocked, &rdev->flags);
1740 set_bit(Faulty, &rdev->flags);
1741 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1743 "md/raid:%s: Disk failure on %s, disabling device.\n"
1744 "md/raid:%s: Operation continuing on %d devices.\n",
1746 bdevname(rdev->bdev, b),
1748 conf->raid_disks - mddev->degraded);
1752 * Input: a 'big' sector number,
1753 * Output: index of the data and parity disk, and the sector # in them.
1755 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1756 int previous, int *dd_idx,
1757 struct stripe_head *sh)
1759 sector_t stripe, stripe2;
1760 sector_t chunk_number;
1761 unsigned int chunk_offset;
1764 sector_t new_sector;
1765 int algorithm = previous ? conf->prev_algo
1767 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1768 : conf->chunk_sectors;
1769 int raid_disks = previous ? conf->previous_raid_disks
1771 int data_disks = raid_disks - conf->max_degraded;
1773 /* First compute the information on this sector */
1776 * Compute the chunk number and the sector offset inside the chunk
1778 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1779 chunk_number = r_sector;
1782 * Compute the stripe number
1784 stripe = chunk_number;
1785 *dd_idx = sector_div(stripe, data_disks);
1788 * Select the parity disk based on the user selected algorithm.
1790 pd_idx = qd_idx = -1;
1791 switch(conf->level) {
1793 pd_idx = data_disks;
1796 switch (algorithm) {
1797 case ALGORITHM_LEFT_ASYMMETRIC:
1798 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1799 if (*dd_idx >= pd_idx)
1802 case ALGORITHM_RIGHT_ASYMMETRIC:
1803 pd_idx = sector_div(stripe2, raid_disks);
1804 if (*dd_idx >= pd_idx)
1807 case ALGORITHM_LEFT_SYMMETRIC:
1808 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1809 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1811 case ALGORITHM_RIGHT_SYMMETRIC:
1812 pd_idx = sector_div(stripe2, raid_disks);
1813 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1815 case ALGORITHM_PARITY_0:
1819 case ALGORITHM_PARITY_N:
1820 pd_idx = data_disks;
1828 switch (algorithm) {
1829 case ALGORITHM_LEFT_ASYMMETRIC:
1830 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1831 qd_idx = pd_idx + 1;
1832 if (pd_idx == raid_disks-1) {
1833 (*dd_idx)++; /* Q D D D P */
1835 } else if (*dd_idx >= pd_idx)
1836 (*dd_idx) += 2; /* D D P Q D */
1838 case ALGORITHM_RIGHT_ASYMMETRIC:
1839 pd_idx = sector_div(stripe2, raid_disks);
1840 qd_idx = pd_idx + 1;
1841 if (pd_idx == raid_disks-1) {
1842 (*dd_idx)++; /* Q D D D P */
1844 } else if (*dd_idx >= pd_idx)
1845 (*dd_idx) += 2; /* D D P Q D */
1847 case ALGORITHM_LEFT_SYMMETRIC:
1848 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1849 qd_idx = (pd_idx + 1) % raid_disks;
1850 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1852 case ALGORITHM_RIGHT_SYMMETRIC:
1853 pd_idx = sector_div(stripe2, raid_disks);
1854 qd_idx = (pd_idx + 1) % raid_disks;
1855 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1858 case ALGORITHM_PARITY_0:
1863 case ALGORITHM_PARITY_N:
1864 pd_idx = data_disks;
1865 qd_idx = data_disks + 1;
1868 case ALGORITHM_ROTATING_ZERO_RESTART:
1869 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1870 * of blocks for computing Q is different.
1872 pd_idx = sector_div(stripe2, raid_disks);
1873 qd_idx = pd_idx + 1;
1874 if (pd_idx == raid_disks-1) {
1875 (*dd_idx)++; /* Q D D D P */
1877 } else if (*dd_idx >= pd_idx)
1878 (*dd_idx) += 2; /* D D P Q D */
1882 case ALGORITHM_ROTATING_N_RESTART:
1883 /* Same a left_asymmetric, by first stripe is
1884 * D D D P Q rather than
1888 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1889 qd_idx = pd_idx + 1;
1890 if (pd_idx == raid_disks-1) {
1891 (*dd_idx)++; /* Q D D D P */
1893 } else if (*dd_idx >= pd_idx)
1894 (*dd_idx) += 2; /* D D P Q D */
1898 case ALGORITHM_ROTATING_N_CONTINUE:
1899 /* Same as left_symmetric but Q is before P */
1900 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1901 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1902 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1906 case ALGORITHM_LEFT_ASYMMETRIC_6:
1907 /* RAID5 left_asymmetric, with Q on last device */
1908 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1909 if (*dd_idx >= pd_idx)
1911 qd_idx = raid_disks - 1;
1914 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1915 pd_idx = sector_div(stripe2, raid_disks-1);
1916 if (*dd_idx >= pd_idx)
1918 qd_idx = raid_disks - 1;
1921 case ALGORITHM_LEFT_SYMMETRIC_6:
1922 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1923 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1924 qd_idx = raid_disks - 1;
1927 case ALGORITHM_RIGHT_SYMMETRIC_6:
1928 pd_idx = sector_div(stripe2, raid_disks-1);
1929 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1930 qd_idx = raid_disks - 1;
1933 case ALGORITHM_PARITY_0_6:
1936 qd_idx = raid_disks - 1;
1946 sh->pd_idx = pd_idx;
1947 sh->qd_idx = qd_idx;
1948 sh->ddf_layout = ddf_layout;
1951 * Finally, compute the new sector number
1953 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1958 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1960 struct r5conf *conf = sh->raid_conf;
1961 int raid_disks = sh->disks;
1962 int data_disks = raid_disks - conf->max_degraded;
1963 sector_t new_sector = sh->sector, check;
1964 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1965 : conf->chunk_sectors;
1966 int algorithm = previous ? conf->prev_algo
1970 sector_t chunk_number;
1971 int dummy1, dd_idx = i;
1973 struct stripe_head sh2;
1976 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1977 stripe = new_sector;
1979 if (i == sh->pd_idx)
1981 switch(conf->level) {
1984 switch (algorithm) {
1985 case ALGORITHM_LEFT_ASYMMETRIC:
1986 case ALGORITHM_RIGHT_ASYMMETRIC:
1990 case ALGORITHM_LEFT_SYMMETRIC:
1991 case ALGORITHM_RIGHT_SYMMETRIC:
1994 i -= (sh->pd_idx + 1);
1996 case ALGORITHM_PARITY_0:
1999 case ALGORITHM_PARITY_N:
2006 if (i == sh->qd_idx)
2007 return 0; /* It is the Q disk */
2008 switch (algorithm) {
2009 case ALGORITHM_LEFT_ASYMMETRIC:
2010 case ALGORITHM_RIGHT_ASYMMETRIC:
2011 case ALGORITHM_ROTATING_ZERO_RESTART:
2012 case ALGORITHM_ROTATING_N_RESTART:
2013 if (sh->pd_idx == raid_disks-1)
2014 i--; /* Q D D D P */
2015 else if (i > sh->pd_idx)
2016 i -= 2; /* D D P Q D */
2018 case ALGORITHM_LEFT_SYMMETRIC:
2019 case ALGORITHM_RIGHT_SYMMETRIC:
2020 if (sh->pd_idx == raid_disks-1)
2021 i--; /* Q D D D P */
2026 i -= (sh->pd_idx + 2);
2029 case ALGORITHM_PARITY_0:
2032 case ALGORITHM_PARITY_N:
2034 case ALGORITHM_ROTATING_N_CONTINUE:
2035 /* Like left_symmetric, but P is before Q */
2036 if (sh->pd_idx == 0)
2037 i--; /* P D D D Q */
2042 i -= (sh->pd_idx + 1);
2045 case ALGORITHM_LEFT_ASYMMETRIC_6:
2046 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2050 case ALGORITHM_LEFT_SYMMETRIC_6:
2051 case ALGORITHM_RIGHT_SYMMETRIC_6:
2053 i += data_disks + 1;
2054 i -= (sh->pd_idx + 1);
2056 case ALGORITHM_PARITY_0_6:
2065 chunk_number = stripe * data_disks + i;
2066 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2068 check = raid5_compute_sector(conf, r_sector,
2069 previous, &dummy1, &sh2);
2070 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2071 || sh2.qd_idx != sh->qd_idx) {
2072 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2073 mdname(conf->mddev));
2081 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2082 int rcw, int expand)
2084 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2085 struct r5conf *conf = sh->raid_conf;
2086 int level = conf->level;
2089 /* if we are not expanding this is a proper write request, and
2090 * there will be bios with new data to be drained into the
2094 sh->reconstruct_state = reconstruct_state_drain_run;
2095 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2097 sh->reconstruct_state = reconstruct_state_run;
2099 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2101 for (i = disks; i--; ) {
2102 struct r5dev *dev = &sh->dev[i];
2105 set_bit(R5_LOCKED, &dev->flags);
2106 set_bit(R5_Wantdrain, &dev->flags);
2108 clear_bit(R5_UPTODATE, &dev->flags);
2112 if (s->locked + conf->max_degraded == disks)
2113 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2114 atomic_inc(&conf->pending_full_writes);
2117 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2118 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2120 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2121 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2122 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2123 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2125 for (i = disks; i--; ) {
2126 struct r5dev *dev = &sh->dev[i];
2131 (test_bit(R5_UPTODATE, &dev->flags) ||
2132 test_bit(R5_Wantcompute, &dev->flags))) {
2133 set_bit(R5_Wantdrain, &dev->flags);
2134 set_bit(R5_LOCKED, &dev->flags);
2135 clear_bit(R5_UPTODATE, &dev->flags);
2141 /* keep the parity disk(s) locked while asynchronous operations
2144 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2145 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2149 int qd_idx = sh->qd_idx;
2150 struct r5dev *dev = &sh->dev[qd_idx];
2152 set_bit(R5_LOCKED, &dev->flags);
2153 clear_bit(R5_UPTODATE, &dev->flags);
2157 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2158 __func__, (unsigned long long)sh->sector,
2159 s->locked, s->ops_request);
2163 * Each stripe/dev can have one or more bion attached.
2164 * toread/towrite point to the first in a chain.
2165 * The bi_next chain must be in order.
2167 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2170 struct r5conf *conf = sh->raid_conf;
2173 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2174 (unsigned long long)bi->bi_sector,
2175 (unsigned long long)sh->sector);
2178 spin_lock_irq(&conf->device_lock);
2180 bip = &sh->dev[dd_idx].towrite;
2181 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2184 bip = &sh->dev[dd_idx].toread;
2185 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2186 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2188 bip = & (*bip)->bi_next;
2190 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2193 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2197 bi->bi_phys_segments++;
2200 /* check if page is covered */
2201 sector_t sector = sh->dev[dd_idx].sector;
2202 for (bi=sh->dev[dd_idx].towrite;
2203 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2204 bi && bi->bi_sector <= sector;
2205 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2206 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2207 sector = bi->bi_sector + (bi->bi_size>>9);
2209 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2210 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2212 spin_unlock_irq(&conf->device_lock);
2214 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2215 (unsigned long long)(*bip)->bi_sector,
2216 (unsigned long long)sh->sector, dd_idx);
2218 if (conf->mddev->bitmap && firstwrite) {
2219 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2221 sh->bm_seq = conf->seq_flush+1;
2222 set_bit(STRIPE_BIT_DELAY, &sh->state);
2227 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2228 spin_unlock_irq(&conf->device_lock);
2232 static void end_reshape(struct r5conf *conf);
2234 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2235 struct stripe_head *sh)
2237 int sectors_per_chunk =
2238 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2240 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2241 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2243 raid5_compute_sector(conf,
2244 stripe * (disks - conf->max_degraded)
2245 *sectors_per_chunk + chunk_offset,
2251 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2252 struct stripe_head_state *s, int disks,
2253 struct bio **return_bi)
2256 for (i = disks; i--; ) {
2260 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2261 struct md_rdev *rdev;
2263 rdev = rcu_dereference(conf->disks[i].rdev);
2264 if (rdev && test_bit(In_sync, &rdev->flags))
2265 atomic_inc(&rdev->nr_pending);
2270 if (!rdev_set_badblocks(
2274 md_error(conf->mddev, rdev);
2275 rdev_dec_pending(rdev, conf->mddev);
2278 spin_lock_irq(&conf->device_lock);
2279 /* fail all writes first */
2280 bi = sh->dev[i].towrite;
2281 sh->dev[i].towrite = NULL;
2287 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2288 wake_up(&conf->wait_for_overlap);
2290 while (bi && bi->bi_sector <
2291 sh->dev[i].sector + STRIPE_SECTORS) {
2292 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2293 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2294 if (!raid5_dec_bi_phys_segments(bi)) {
2295 md_write_end(conf->mddev);
2296 bi->bi_next = *return_bi;
2301 /* and fail all 'written' */
2302 bi = sh->dev[i].written;
2303 sh->dev[i].written = NULL;
2304 if (bi) bitmap_end = 1;
2305 while (bi && bi->bi_sector <
2306 sh->dev[i].sector + STRIPE_SECTORS) {
2307 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2308 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2309 if (!raid5_dec_bi_phys_segments(bi)) {
2310 md_write_end(conf->mddev);
2311 bi->bi_next = *return_bi;
2317 /* fail any reads if this device is non-operational and
2318 * the data has not reached the cache yet.
2320 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2321 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2322 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2323 bi = sh->dev[i].toread;
2324 sh->dev[i].toread = NULL;
2325 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2326 wake_up(&conf->wait_for_overlap);
2327 if (bi) s->to_read--;
2328 while (bi && bi->bi_sector <
2329 sh->dev[i].sector + STRIPE_SECTORS) {
2330 struct bio *nextbi =
2331 r5_next_bio(bi, sh->dev[i].sector);
2332 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2333 if (!raid5_dec_bi_phys_segments(bi)) {
2334 bi->bi_next = *return_bi;
2340 spin_unlock_irq(&conf->device_lock);
2342 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2343 STRIPE_SECTORS, 0, 0);
2344 /* If we were in the middle of a write the parity block might
2345 * still be locked - so just clear all R5_LOCKED flags
2347 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2350 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2351 if (atomic_dec_and_test(&conf->pending_full_writes))
2352 md_wakeup_thread(conf->mddev->thread);
2356 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2357 struct stripe_head_state *s)
2362 md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
2363 clear_bit(STRIPE_SYNCING, &sh->state);
2365 /* There is nothing more to do for sync/check/repair.
2366 * For recover we need to record a bad block on all
2367 * non-sync devices, or abort the recovery
2369 if (!test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery))
2371 /* During recovery devices cannot be removed, so locking and
2372 * refcounting of rdevs is not needed
2374 for (i = 0; i < conf->raid_disks; i++) {
2375 struct md_rdev *rdev = conf->disks[i].rdev;
2377 || test_bit(Faulty, &rdev->flags)
2378 || test_bit(In_sync, &rdev->flags))
2380 if (!rdev_set_badblocks(rdev, sh->sector,
2385 conf->recovery_disabled = conf->mddev->recovery_disabled;
2386 set_bit(MD_RECOVERY_INTR, &conf->mddev->recovery);
2390 /* fetch_block - checks the given member device to see if its data needs
2391 * to be read or computed to satisfy a request.
2393 * Returns 1 when no more member devices need to be checked, otherwise returns
2394 * 0 to tell the loop in handle_stripe_fill to continue
2396 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2397 int disk_idx, int disks)
2399 struct r5dev *dev = &sh->dev[disk_idx];
2400 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2401 &sh->dev[s->failed_num[1]] };
2403 /* is the data in this block needed, and can we get it? */
2404 if (!test_bit(R5_LOCKED, &dev->flags) &&
2405 !test_bit(R5_UPTODATE, &dev->flags) &&
2407 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2408 s->syncing || s->expanding ||
2409 (s->failed >= 1 && fdev[0]->toread) ||
2410 (s->failed >= 2 && fdev[1]->toread) ||
2411 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2412 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2413 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2414 /* we would like to get this block, possibly by computing it,
2415 * otherwise read it if the backing disk is insync
2417 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2418 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2419 if ((s->uptodate == disks - 1) &&
2420 (s->failed && (disk_idx == s->failed_num[0] ||
2421 disk_idx == s->failed_num[1]))) {
2422 /* have disk failed, and we're requested to fetch it;
2425 pr_debug("Computing stripe %llu block %d\n",
2426 (unsigned long long)sh->sector, disk_idx);
2427 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2428 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2429 set_bit(R5_Wantcompute, &dev->flags);
2430 sh->ops.target = disk_idx;
2431 sh->ops.target2 = -1; /* no 2nd target */
2433 /* Careful: from this point on 'uptodate' is in the eye
2434 * of raid_run_ops which services 'compute' operations
2435 * before writes. R5_Wantcompute flags a block that will
2436 * be R5_UPTODATE by the time it is needed for a
2437 * subsequent operation.
2441 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2442 /* Computing 2-failure is *very* expensive; only
2443 * do it if failed >= 2
2446 for (other = disks; other--; ) {
2447 if (other == disk_idx)
2449 if (!test_bit(R5_UPTODATE,
2450 &sh->dev[other].flags))
2454 pr_debug("Computing stripe %llu blocks %d,%d\n",
2455 (unsigned long long)sh->sector,
2457 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2458 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2459 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2460 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2461 sh->ops.target = disk_idx;
2462 sh->ops.target2 = other;
2466 } else if (test_bit(R5_Insync, &dev->flags)) {
2467 set_bit(R5_LOCKED, &dev->flags);
2468 set_bit(R5_Wantread, &dev->flags);
2470 pr_debug("Reading block %d (sync=%d)\n",
2471 disk_idx, s->syncing);
2479 * handle_stripe_fill - read or compute data to satisfy pending requests.
2481 static void handle_stripe_fill(struct stripe_head *sh,
2482 struct stripe_head_state *s,
2487 /* look for blocks to read/compute, skip this if a compute
2488 * is already in flight, or if the stripe contents are in the
2489 * midst of changing due to a write
2491 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2492 !sh->reconstruct_state)
2493 for (i = disks; i--; )
2494 if (fetch_block(sh, s, i, disks))
2496 set_bit(STRIPE_HANDLE, &sh->state);
2500 /* handle_stripe_clean_event
2501 * any written block on an uptodate or failed drive can be returned.
2502 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2503 * never LOCKED, so we don't need to test 'failed' directly.
2505 static void handle_stripe_clean_event(struct r5conf *conf,
2506 struct stripe_head *sh, int disks, struct bio **return_bi)
2511 for (i = disks; i--; )
2512 if (sh->dev[i].written) {
2514 if (!test_bit(R5_LOCKED, &dev->flags) &&
2515 test_bit(R5_UPTODATE, &dev->flags)) {
2516 /* We can return any write requests */
2517 struct bio *wbi, *wbi2;
2519 pr_debug("Return write for disc %d\n", i);
2520 spin_lock_irq(&conf->device_lock);
2522 dev->written = NULL;
2523 while (wbi && wbi->bi_sector <
2524 dev->sector + STRIPE_SECTORS) {
2525 wbi2 = r5_next_bio(wbi, dev->sector);
2526 if (!raid5_dec_bi_phys_segments(wbi)) {
2527 md_write_end(conf->mddev);
2528 wbi->bi_next = *return_bi;
2533 if (dev->towrite == NULL)
2535 spin_unlock_irq(&conf->device_lock);
2537 bitmap_endwrite(conf->mddev->bitmap,
2540 !test_bit(STRIPE_DEGRADED, &sh->state),
2545 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2546 if (atomic_dec_and_test(&conf->pending_full_writes))
2547 md_wakeup_thread(conf->mddev->thread);
2550 static void handle_stripe_dirtying(struct r5conf *conf,
2551 struct stripe_head *sh,
2552 struct stripe_head_state *s,
2555 int rmw = 0, rcw = 0, i;
2556 if (conf->max_degraded == 2) {
2557 /* RAID6 requires 'rcw' in current implementation
2558 * Calculate the real rcw later - for now fake it
2559 * look like rcw is cheaper
2562 } else for (i = disks; i--; ) {
2563 /* would I have to read this buffer for read_modify_write */
2564 struct r5dev *dev = &sh->dev[i];
2565 if ((dev->towrite || i == sh->pd_idx) &&
2566 !test_bit(R5_LOCKED, &dev->flags) &&
2567 !(test_bit(R5_UPTODATE, &dev->flags) ||
2568 test_bit(R5_Wantcompute, &dev->flags))) {
2569 if (test_bit(R5_Insync, &dev->flags))
2572 rmw += 2*disks; /* cannot read it */
2574 /* Would I have to read this buffer for reconstruct_write */
2575 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2576 !test_bit(R5_LOCKED, &dev->flags) &&
2577 !(test_bit(R5_UPTODATE, &dev->flags) ||
2578 test_bit(R5_Wantcompute, &dev->flags))) {
2579 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2584 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2585 (unsigned long long)sh->sector, rmw, rcw);
2586 set_bit(STRIPE_HANDLE, &sh->state);
2587 if (rmw < rcw && rmw > 0)
2588 /* prefer read-modify-write, but need to get some data */
2589 for (i = disks; i--; ) {
2590 struct r5dev *dev = &sh->dev[i];
2591 if ((dev->towrite || i == sh->pd_idx) &&
2592 !test_bit(R5_LOCKED, &dev->flags) &&
2593 !(test_bit(R5_UPTODATE, &dev->flags) ||
2594 test_bit(R5_Wantcompute, &dev->flags)) &&
2595 test_bit(R5_Insync, &dev->flags)) {
2597 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2598 pr_debug("Read_old block "
2599 "%d for r-m-w\n", i);
2600 set_bit(R5_LOCKED, &dev->flags);
2601 set_bit(R5_Wantread, &dev->flags);
2604 set_bit(STRIPE_DELAYED, &sh->state);
2605 set_bit(STRIPE_HANDLE, &sh->state);
2609 if (rcw <= rmw && rcw > 0) {
2610 /* want reconstruct write, but need to get some data */
2612 for (i = disks; i--; ) {
2613 struct r5dev *dev = &sh->dev[i];
2614 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2615 i != sh->pd_idx && i != sh->qd_idx &&
2616 !test_bit(R5_LOCKED, &dev->flags) &&
2617 !(test_bit(R5_UPTODATE, &dev->flags) ||
2618 test_bit(R5_Wantcompute, &dev->flags))) {
2620 if (!test_bit(R5_Insync, &dev->flags))
2621 continue; /* it's a failed drive */
2623 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2624 pr_debug("Read_old block "
2625 "%d for Reconstruct\n", i);
2626 set_bit(R5_LOCKED, &dev->flags);
2627 set_bit(R5_Wantread, &dev->flags);
2630 set_bit(STRIPE_DELAYED, &sh->state);
2631 set_bit(STRIPE_HANDLE, &sh->state);
2636 /* now if nothing is locked, and if we have enough data,
2637 * we can start a write request
2639 /* since handle_stripe can be called at any time we need to handle the
2640 * case where a compute block operation has been submitted and then a
2641 * subsequent call wants to start a write request. raid_run_ops only
2642 * handles the case where compute block and reconstruct are requested
2643 * simultaneously. If this is not the case then new writes need to be
2644 * held off until the compute completes.
2646 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2647 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2648 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2649 schedule_reconstruction(sh, s, rcw == 0, 0);
2652 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2653 struct stripe_head_state *s, int disks)
2655 struct r5dev *dev = NULL;
2657 set_bit(STRIPE_HANDLE, &sh->state);
2659 switch (sh->check_state) {
2660 case check_state_idle:
2661 /* start a new check operation if there are no failures */
2662 if (s->failed == 0) {
2663 BUG_ON(s->uptodate != disks);
2664 sh->check_state = check_state_run;
2665 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2666 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2670 dev = &sh->dev[s->failed_num[0]];
2672 case check_state_compute_result:
2673 sh->check_state = check_state_idle;
2675 dev = &sh->dev[sh->pd_idx];
2677 /* check that a write has not made the stripe insync */
2678 if (test_bit(STRIPE_INSYNC, &sh->state))
2681 /* either failed parity check, or recovery is happening */
2682 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2683 BUG_ON(s->uptodate != disks);
2685 set_bit(R5_LOCKED, &dev->flags);
2687 set_bit(R5_Wantwrite, &dev->flags);
2689 clear_bit(STRIPE_DEGRADED, &sh->state);
2690 set_bit(STRIPE_INSYNC, &sh->state);
2692 case check_state_run:
2693 break; /* we will be called again upon completion */
2694 case check_state_check_result:
2695 sh->check_state = check_state_idle;
2697 /* if a failure occurred during the check operation, leave
2698 * STRIPE_INSYNC not set and let the stripe be handled again
2703 /* handle a successful check operation, if parity is correct
2704 * we are done. Otherwise update the mismatch count and repair
2705 * parity if !MD_RECOVERY_CHECK
2707 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2708 /* parity is correct (on disc,
2709 * not in buffer any more)
2711 set_bit(STRIPE_INSYNC, &sh->state);
2713 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2714 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2715 /* don't try to repair!! */
2716 set_bit(STRIPE_INSYNC, &sh->state);
2718 sh->check_state = check_state_compute_run;
2719 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2720 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2721 set_bit(R5_Wantcompute,
2722 &sh->dev[sh->pd_idx].flags);
2723 sh->ops.target = sh->pd_idx;
2724 sh->ops.target2 = -1;
2729 case check_state_compute_run:
2732 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2733 __func__, sh->check_state,
2734 (unsigned long long) sh->sector);
2740 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2741 struct stripe_head_state *s,
2744 int pd_idx = sh->pd_idx;
2745 int qd_idx = sh->qd_idx;
2748 set_bit(STRIPE_HANDLE, &sh->state);
2750 BUG_ON(s->failed > 2);
2752 /* Want to check and possibly repair P and Q.
2753 * However there could be one 'failed' device, in which
2754 * case we can only check one of them, possibly using the
2755 * other to generate missing data
2758 switch (sh->check_state) {
2759 case check_state_idle:
2760 /* start a new check operation if there are < 2 failures */
2761 if (s->failed == s->q_failed) {
2762 /* The only possible failed device holds Q, so it
2763 * makes sense to check P (If anything else were failed,
2764 * we would have used P to recreate it).
2766 sh->check_state = check_state_run;
2768 if (!s->q_failed && s->failed < 2) {
2769 /* Q is not failed, and we didn't use it to generate
2770 * anything, so it makes sense to check it
2772 if (sh->check_state == check_state_run)
2773 sh->check_state = check_state_run_pq;
2775 sh->check_state = check_state_run_q;
2778 /* discard potentially stale zero_sum_result */
2779 sh->ops.zero_sum_result = 0;
2781 if (sh->check_state == check_state_run) {
2782 /* async_xor_zero_sum destroys the contents of P */
2783 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2786 if (sh->check_state >= check_state_run &&
2787 sh->check_state <= check_state_run_pq) {
2788 /* async_syndrome_zero_sum preserves P and Q, so
2789 * no need to mark them !uptodate here
2791 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2795 /* we have 2-disk failure */
2796 BUG_ON(s->failed != 2);
2798 case check_state_compute_result:
2799 sh->check_state = check_state_idle;
2801 /* check that a write has not made the stripe insync */
2802 if (test_bit(STRIPE_INSYNC, &sh->state))
2805 /* now write out any block on a failed drive,
2806 * or P or Q if they were recomputed
2808 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2809 if (s->failed == 2) {
2810 dev = &sh->dev[s->failed_num[1]];
2812 set_bit(R5_LOCKED, &dev->flags);
2813 set_bit(R5_Wantwrite, &dev->flags);
2815 if (s->failed >= 1) {
2816 dev = &sh->dev[s->failed_num[0]];
2818 set_bit(R5_LOCKED, &dev->flags);
2819 set_bit(R5_Wantwrite, &dev->flags);
2821 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2822 dev = &sh->dev[pd_idx];
2824 set_bit(R5_LOCKED, &dev->flags);
2825 set_bit(R5_Wantwrite, &dev->flags);
2827 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2828 dev = &sh->dev[qd_idx];
2830 set_bit(R5_LOCKED, &dev->flags);
2831 set_bit(R5_Wantwrite, &dev->flags);
2833 clear_bit(STRIPE_DEGRADED, &sh->state);
2835 set_bit(STRIPE_INSYNC, &sh->state);
2837 case check_state_run:
2838 case check_state_run_q:
2839 case check_state_run_pq:
2840 break; /* we will be called again upon completion */
2841 case check_state_check_result:
2842 sh->check_state = check_state_idle;
2844 /* handle a successful check operation, if parity is correct
2845 * we are done. Otherwise update the mismatch count and repair
2846 * parity if !MD_RECOVERY_CHECK
2848 if (sh->ops.zero_sum_result == 0) {
2849 /* both parities are correct */
2851 set_bit(STRIPE_INSYNC, &sh->state);
2853 /* in contrast to the raid5 case we can validate
2854 * parity, but still have a failure to write
2857 sh->check_state = check_state_compute_result;
2858 /* Returning at this point means that we may go
2859 * off and bring p and/or q uptodate again so
2860 * we make sure to check zero_sum_result again
2861 * to verify if p or q need writeback
2865 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2866 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2867 /* don't try to repair!! */
2868 set_bit(STRIPE_INSYNC, &sh->state);
2870 int *target = &sh->ops.target;
2872 sh->ops.target = -1;
2873 sh->ops.target2 = -1;
2874 sh->check_state = check_state_compute_run;
2875 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2876 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2877 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2878 set_bit(R5_Wantcompute,
2879 &sh->dev[pd_idx].flags);
2881 target = &sh->ops.target2;
2884 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2885 set_bit(R5_Wantcompute,
2886 &sh->dev[qd_idx].flags);
2893 case check_state_compute_run:
2896 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2897 __func__, sh->check_state,
2898 (unsigned long long) sh->sector);
2903 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
2907 /* We have read all the blocks in this stripe and now we need to
2908 * copy some of them into a target stripe for expand.
2910 struct dma_async_tx_descriptor *tx = NULL;
2911 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2912 for (i = 0; i < sh->disks; i++)
2913 if (i != sh->pd_idx && i != sh->qd_idx) {
2915 struct stripe_head *sh2;
2916 struct async_submit_ctl submit;
2918 sector_t bn = compute_blocknr(sh, i, 1);
2919 sector_t s = raid5_compute_sector(conf, bn, 0,
2921 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2923 /* so far only the early blocks of this stripe
2924 * have been requested. When later blocks
2925 * get requested, we will try again
2928 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2929 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2930 /* must have already done this block */
2931 release_stripe(sh2);
2935 /* place all the copies on one channel */
2936 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2937 tx = async_memcpy(sh2->dev[dd_idx].page,
2938 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2941 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2942 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2943 for (j = 0; j < conf->raid_disks; j++)
2944 if (j != sh2->pd_idx &&
2946 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2948 if (j == conf->raid_disks) {
2949 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2950 set_bit(STRIPE_HANDLE, &sh2->state);
2952 release_stripe(sh2);
2955 /* done submitting copies, wait for them to complete */
2958 dma_wait_for_async_tx(tx);
2964 * handle_stripe - do things to a stripe.
2966 * We lock the stripe and then examine the state of various bits
2967 * to see what needs to be done.
2969 * return some read request which now have data
2970 * return some write requests which are safely on disc
2971 * schedule a read on some buffers
2972 * schedule a write of some buffers
2973 * return confirmation of parity correctness
2975 * buffers are taken off read_list or write_list, and bh_cache buffers
2976 * get BH_Lock set before the stripe lock is released.
2980 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
2982 struct r5conf *conf = sh->raid_conf;
2983 int disks = sh->disks;
2987 memset(s, 0, sizeof(*s));
2989 s->syncing = test_bit(STRIPE_SYNCING, &sh->state);
2990 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2991 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2992 s->failed_num[0] = -1;
2993 s->failed_num[1] = -1;
2995 /* Now to look around and see what can be done */
2997 spin_lock_irq(&conf->device_lock);
2998 for (i=disks; i--; ) {
2999 struct md_rdev *rdev;
3006 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3007 i, dev->flags, dev->toread, dev->towrite, dev->written);
3008 /* maybe we can reply to a read
3010 * new wantfill requests are only permitted while
3011 * ops_complete_biofill is guaranteed to be inactive
3013 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3014 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3015 set_bit(R5_Wantfill, &dev->flags);
3017 /* now count some things */
3018 if (test_bit(R5_LOCKED, &dev->flags))
3020 if (test_bit(R5_UPTODATE, &dev->flags))
3022 if (test_bit(R5_Wantcompute, &dev->flags)) {
3024 BUG_ON(s->compute > 2);
3027 if (test_bit(R5_Wantfill, &dev->flags))
3029 else if (dev->toread)
3033 if (!test_bit(R5_OVERWRITE, &dev->flags))
3038 rdev = rcu_dereference(conf->disks[i].rdev);
3040 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3041 &first_bad, &bad_sectors);
3042 if (s->blocked_rdev == NULL
3043 && (test_bit(Blocked, &rdev->flags)
3046 set_bit(BlockedBadBlocks,
3048 s->blocked_rdev = rdev;
3049 atomic_inc(&rdev->nr_pending);
3052 clear_bit(R5_Insync, &dev->flags);
3056 /* also not in-sync */
3057 if (!test_bit(WriteErrorSeen, &rdev->flags)) {
3058 /* treat as in-sync, but with a read error
3059 * which we can now try to correct
3061 set_bit(R5_Insync, &dev->flags);
3062 set_bit(R5_ReadError, &dev->flags);
3064 } else if (test_bit(In_sync, &rdev->flags))
3065 set_bit(R5_Insync, &dev->flags);
3066 else if (!test_bit(Faulty, &rdev->flags)) {
3067 /* in sync if before recovery_offset */
3068 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3069 set_bit(R5_Insync, &dev->flags);
3071 if (test_bit(R5_WriteError, &dev->flags)) {
3072 clear_bit(R5_Insync, &dev->flags);
3073 if (!test_bit(Faulty, &rdev->flags)) {
3074 s->handle_bad_blocks = 1;
3075 atomic_inc(&rdev->nr_pending);
3077 clear_bit(R5_WriteError, &dev->flags);
3079 if (test_bit(R5_MadeGood, &dev->flags)) {
3080 if (!test_bit(Faulty, &rdev->flags)) {
3081 s->handle_bad_blocks = 1;
3082 atomic_inc(&rdev->nr_pending);
3084 clear_bit(R5_MadeGood, &dev->flags);
3086 if (!test_bit(R5_Insync, &dev->flags)) {
3087 /* The ReadError flag will just be confusing now */
3088 clear_bit(R5_ReadError, &dev->flags);
3089 clear_bit(R5_ReWrite, &dev->flags);
3091 if (test_bit(R5_ReadError, &dev->flags))
3092 clear_bit(R5_Insync, &dev->flags);
3093 if (!test_bit(R5_Insync, &dev->flags)) {
3095 s->failed_num[s->failed] = i;
3099 spin_unlock_irq(&conf->device_lock);
3103 static void handle_stripe(struct stripe_head *sh)
3105 struct stripe_head_state s;
3106 struct r5conf *conf = sh->raid_conf;
3109 int disks = sh->disks;
3110 struct r5dev *pdev, *qdev;
3112 clear_bit(STRIPE_HANDLE, &sh->state);
3113 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3114 /* already being handled, ensure it gets handled
3115 * again when current action finishes */
3116 set_bit(STRIPE_HANDLE, &sh->state);
3120 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3121 set_bit(STRIPE_SYNCING, &sh->state);
3122 clear_bit(STRIPE_INSYNC, &sh->state);
3124 clear_bit(STRIPE_DELAYED, &sh->state);
3126 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3127 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3128 (unsigned long long)sh->sector, sh->state,
3129 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3130 sh->check_state, sh->reconstruct_state);
3132 analyse_stripe(sh, &s);
3134 if (s.handle_bad_blocks) {
3135 set_bit(STRIPE_HANDLE, &sh->state);
3139 if (unlikely(s.blocked_rdev)) {
3140 if (s.syncing || s.expanding || s.expanded ||
3141 s.to_write || s.written) {
3142 set_bit(STRIPE_HANDLE, &sh->state);
3145 /* There is nothing for the blocked_rdev to block */
3146 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3147 s.blocked_rdev = NULL;
3150 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3151 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3152 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3155 pr_debug("locked=%d uptodate=%d to_read=%d"
3156 " to_write=%d failed=%d failed_num=%d,%d\n",
3157 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3158 s.failed_num[0], s.failed_num[1]);
3159 /* check if the array has lost more than max_degraded devices and,
3160 * if so, some requests might need to be failed.
3162 if (s.failed > conf->max_degraded) {
3163 sh->check_state = 0;
3164 sh->reconstruct_state = 0;
3165 if (s.to_read+s.to_write+s.written)
3166 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3168 handle_failed_sync(conf, sh, &s);
3172 * might be able to return some write requests if the parity blocks
3173 * are safe, or on a failed drive
3175 pdev = &sh->dev[sh->pd_idx];
3176 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3177 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3178 qdev = &sh->dev[sh->qd_idx];
3179 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3180 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3184 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3185 && !test_bit(R5_LOCKED, &pdev->flags)
3186 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3187 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3188 && !test_bit(R5_LOCKED, &qdev->flags)
3189 && test_bit(R5_UPTODATE, &qdev->flags)))))
3190 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3192 /* Now we might consider reading some blocks, either to check/generate
3193 * parity, or to satisfy requests
3194 * or to load a block that is being partially written.
3196 if (s.to_read || s.non_overwrite
3197 || (conf->level == 6 && s.to_write && s.failed)
3198 || (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3199 handle_stripe_fill(sh, &s, disks);
3201 /* Now we check to see if any write operations have recently
3205 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3207 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3208 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3209 sh->reconstruct_state = reconstruct_state_idle;
3211 /* All the 'written' buffers and the parity block are ready to
3212 * be written back to disk
3214 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3215 BUG_ON(sh->qd_idx >= 0 &&
3216 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3217 for (i = disks; i--; ) {
3218 struct r5dev *dev = &sh->dev[i];
3219 if (test_bit(R5_LOCKED, &dev->flags) &&
3220 (i == sh->pd_idx || i == sh->qd_idx ||
3222 pr_debug("Writing block %d\n", i);
3223 set_bit(R5_Wantwrite, &dev->flags);
3226 if (!test_bit(R5_Insync, &dev->flags) ||
3227 ((i == sh->pd_idx || i == sh->qd_idx) &&
3229 set_bit(STRIPE_INSYNC, &sh->state);
3232 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3233 s.dec_preread_active = 1;
3236 /* Now to consider new write requests and what else, if anything
3237 * should be read. We do not handle new writes when:
3238 * 1/ A 'write' operation (copy+xor) is already in flight.
3239 * 2/ A 'check' operation is in flight, as it may clobber the parity
3242 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3243 handle_stripe_dirtying(conf, sh, &s, disks);
3245 /* maybe we need to check and possibly fix the parity for this stripe
3246 * Any reads will already have been scheduled, so we just see if enough
3247 * data is available. The parity check is held off while parity
3248 * dependent operations are in flight.
3250 if (sh->check_state ||
3251 (s.syncing && s.locked == 0 &&
3252 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3253 !test_bit(STRIPE_INSYNC, &sh->state))) {
3254 if (conf->level == 6)
3255 handle_parity_checks6(conf, sh, &s, disks);
3257 handle_parity_checks5(conf, sh, &s, disks);
3260 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3261 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3262 clear_bit(STRIPE_SYNCING, &sh->state);
3265 /* If the failed drives are just a ReadError, then we might need
3266 * to progress the repair/check process
3268 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3269 for (i = 0; i < s.failed; i++) {
3270 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3271 if (test_bit(R5_ReadError, &dev->flags)
3272 && !test_bit(R5_LOCKED, &dev->flags)
3273 && test_bit(R5_UPTODATE, &dev->flags)
3275 if (!test_bit(R5_ReWrite, &dev->flags)) {
3276 set_bit(R5_Wantwrite, &dev->flags);
3277 set_bit(R5_ReWrite, &dev->flags);
3278 set_bit(R5_LOCKED, &dev->flags);
3281 /* let's read it back */
3282 set_bit(R5_Wantread, &dev->flags);
3283 set_bit(R5_LOCKED, &dev->flags);
3290 /* Finish reconstruct operations initiated by the expansion process */
3291 if (sh->reconstruct_state == reconstruct_state_result) {
3292 struct stripe_head *sh_src
3293 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3294 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3295 /* sh cannot be written until sh_src has been read.
3296 * so arrange for sh to be delayed a little
3298 set_bit(STRIPE_DELAYED, &sh->state);
3299 set_bit(STRIPE_HANDLE, &sh->state);
3300 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3302 atomic_inc(&conf->preread_active_stripes);
3303 release_stripe(sh_src);
3307 release_stripe(sh_src);
3309 sh->reconstruct_state = reconstruct_state_idle;
3310 clear_bit(STRIPE_EXPANDING, &sh->state);
3311 for (i = conf->raid_disks; i--; ) {
3312 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3313 set_bit(R5_LOCKED, &sh->dev[i].flags);
3318 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3319 !sh->reconstruct_state) {
3320 /* Need to write out all blocks after computing parity */
3321 sh->disks = conf->raid_disks;
3322 stripe_set_idx(sh->sector, conf, 0, sh);
3323 schedule_reconstruction(sh, &s, 1, 1);
3324 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3325 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3326 atomic_dec(&conf->reshape_stripes);
3327 wake_up(&conf->wait_for_overlap);
3328 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3331 if (s.expanding && s.locked == 0 &&
3332 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3333 handle_stripe_expansion(conf, sh);
3336 /* wait for this device to become unblocked */
3337 if (conf->mddev->external && unlikely(s.blocked_rdev))
3338 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3340 if (s.handle_bad_blocks)
3341 for (i = disks; i--; ) {
3342 struct md_rdev *rdev;
3343 struct r5dev *dev = &sh->dev[i];
3344 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3345 /* We own a safe reference to the rdev */
3346 rdev = conf->disks[i].rdev;
3347 if (!rdev_set_badblocks(rdev, sh->sector,
3349 md_error(conf->mddev, rdev);
3350 rdev_dec_pending(rdev, conf->mddev);
3352 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3353 rdev = conf->disks[i].rdev;
3354 rdev_clear_badblocks(rdev, sh->sector,
3356 rdev_dec_pending(rdev, conf->mddev);
3361 raid_run_ops(sh, s.ops_request);
3365 if (s.dec_preread_active) {
3366 /* We delay this until after ops_run_io so that if make_request
3367 * is waiting on a flush, it won't continue until the writes
3368 * have actually been submitted.
3370 atomic_dec(&conf->preread_active_stripes);
3371 if (atomic_read(&conf->preread_active_stripes) <
3373 md_wakeup_thread(conf->mddev->thread);
3376 return_io(s.return_bi);
3378 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3381 static void raid5_activate_delayed(struct r5conf *conf)
3383 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3384 while (!list_empty(&conf->delayed_list)) {
3385 struct list_head *l = conf->delayed_list.next;
3386 struct stripe_head *sh;
3387 sh = list_entry(l, struct stripe_head, lru);
3389 clear_bit(STRIPE_DELAYED, &sh->state);
3390 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3391 atomic_inc(&conf->preread_active_stripes);
3392 list_add_tail(&sh->lru, &conf->hold_list);
3397 static void activate_bit_delay(struct r5conf *conf)
3399 /* device_lock is held */
3400 struct list_head head;
3401 list_add(&head, &conf->bitmap_list);
3402 list_del_init(&conf->bitmap_list);
3403 while (!list_empty(&head)) {
3404 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3405 list_del_init(&sh->lru);
3406 atomic_inc(&sh->count);
3407 __release_stripe(conf, sh);
3411 int md_raid5_congested(struct mddev *mddev, int bits)
3413 struct r5conf *conf = mddev->private;
3415 /* No difference between reads and writes. Just check
3416 * how busy the stripe_cache is
3419 if (conf->inactive_blocked)
3423 if (list_empty_careful(&conf->inactive_list))
3428 EXPORT_SYMBOL_GPL(md_raid5_congested);
3430 static int raid5_congested(void *data, int bits)
3432 struct mddev *mddev = data;
3434 return mddev_congested(mddev, bits) ||
3435 md_raid5_congested(mddev, bits);
3438 /* We want read requests to align with chunks where possible,
3439 * but write requests don't need to.
3441 static int raid5_mergeable_bvec(struct request_queue *q,
3442 struct bvec_merge_data *bvm,
3443 struct bio_vec *biovec)
3445 struct mddev *mddev = q->queuedata;
3446 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3448 unsigned int chunk_sectors = mddev->chunk_sectors;
3449 unsigned int bio_sectors = bvm->bi_size >> 9;
3451 if ((bvm->bi_rw & 1) == WRITE)
3452 return biovec->bv_len; /* always allow writes to be mergeable */
3454 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3455 chunk_sectors = mddev->new_chunk_sectors;
3456 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3457 if (max < 0) max = 0;
3458 if (max <= biovec->bv_len && bio_sectors == 0)
3459 return biovec->bv_len;
3465 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3467 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3468 unsigned int chunk_sectors = mddev->chunk_sectors;
3469 unsigned int bio_sectors = bio->bi_size >> 9;
3471 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3472 chunk_sectors = mddev->new_chunk_sectors;
3473 return chunk_sectors >=
3474 ((sector & (chunk_sectors - 1)) + bio_sectors);
3478 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3479 * later sampled by raid5d.
3481 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3483 unsigned long flags;
3485 spin_lock_irqsave(&conf->device_lock, flags);
3487 bi->bi_next = conf->retry_read_aligned_list;
3488 conf->retry_read_aligned_list = bi;
3490 spin_unlock_irqrestore(&conf->device_lock, flags);
3491 md_wakeup_thread(conf->mddev->thread);
3495 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3499 bi = conf->retry_read_aligned;
3501 conf->retry_read_aligned = NULL;
3504 bi = conf->retry_read_aligned_list;
3506 conf->retry_read_aligned_list = bi->bi_next;
3509 * this sets the active strip count to 1 and the processed
3510 * strip count to zero (upper 8 bits)
3512 bi->bi_phys_segments = 1; /* biased count of active stripes */
3520 * The "raid5_align_endio" should check if the read succeeded and if it
3521 * did, call bio_endio on the original bio (having bio_put the new bio
3523 * If the read failed..
3525 static void raid5_align_endio(struct bio *bi, int error)
3527 struct bio* raid_bi = bi->bi_private;
3528 struct mddev *mddev;
3529 struct r5conf *conf;
3530 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3531 struct md_rdev *rdev;
3535 rdev = (void*)raid_bi->bi_next;
3536 raid_bi->bi_next = NULL;
3537 mddev = rdev->mddev;
3538 conf = mddev->private;
3540 rdev_dec_pending(rdev, conf->mddev);
3542 if (!error && uptodate) {
3543 bio_endio(raid_bi, 0);
3544 if (atomic_dec_and_test(&conf->active_aligned_reads))
3545 wake_up(&conf->wait_for_stripe);
3550 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3552 add_bio_to_retry(raid_bi, conf);
3555 static int bio_fits_rdev(struct bio *bi)
3557 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3559 if ((bi->bi_size>>9) > queue_max_sectors(q))
3561 blk_recount_segments(q, bi);
3562 if (bi->bi_phys_segments > queue_max_segments(q))
3565 if (q->merge_bvec_fn)
3566 /* it's too hard to apply the merge_bvec_fn at this stage,
3575 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3577 struct r5conf *conf = mddev->private;
3579 struct bio* align_bi;
3580 struct md_rdev *rdev;
3582 if (!in_chunk_boundary(mddev, raid_bio)) {
3583 pr_debug("chunk_aligned_read : non aligned\n");
3587 * use bio_clone_mddev to make a copy of the bio
3589 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3593 * set bi_end_io to a new function, and set bi_private to the
3596 align_bi->bi_end_io = raid5_align_endio;
3597 align_bi->bi_private = raid_bio;
3601 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3606 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3607 if (rdev && test_bit(In_sync, &rdev->flags)) {
3611 atomic_inc(&rdev->nr_pending);
3613 raid_bio->bi_next = (void*)rdev;
3614 align_bi->bi_bdev = rdev->bdev;
3615 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3616 align_bi->bi_sector += rdev->data_offset;
3618 if (!bio_fits_rdev(align_bi) ||
3619 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3620 &first_bad, &bad_sectors)) {
3621 /* too big in some way, or has a known bad block */
3623 rdev_dec_pending(rdev, mddev);
3627 spin_lock_irq(&conf->device_lock);
3628 wait_event_lock_irq(conf->wait_for_stripe,
3630 conf->device_lock, /* nothing */);
3631 atomic_inc(&conf->active_aligned_reads);
3632 spin_unlock_irq(&conf->device_lock);
3634 generic_make_request(align_bi);
3643 /* __get_priority_stripe - get the next stripe to process
3645 * Full stripe writes are allowed to pass preread active stripes up until
3646 * the bypass_threshold is exceeded. In general the bypass_count
3647 * increments when the handle_list is handled before the hold_list; however, it
3648 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3649 * stripe with in flight i/o. The bypass_count will be reset when the
3650 * head of the hold_list has changed, i.e. the head was promoted to the
3653 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3655 struct stripe_head *sh;
3657 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3659 list_empty(&conf->handle_list) ? "empty" : "busy",
3660 list_empty(&conf->hold_list) ? "empty" : "busy",
3661 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3663 if (!list_empty(&conf->handle_list)) {
3664 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3666 if (list_empty(&conf->hold_list))
3667 conf->bypass_count = 0;
3668 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3669 if (conf->hold_list.next == conf->last_hold)
3670 conf->bypass_count++;
3672 conf->last_hold = conf->hold_list.next;
3673 conf->bypass_count -= conf->bypass_threshold;
3674 if (conf->bypass_count < 0)
3675 conf->bypass_count = 0;
3678 } else if (!list_empty(&conf->hold_list) &&
3679 ((conf->bypass_threshold &&
3680 conf->bypass_count > conf->bypass_threshold) ||
3681 atomic_read(&conf->pending_full_writes) == 0)) {
3682 sh = list_entry(conf->hold_list.next,
3684 conf->bypass_count -= conf->bypass_threshold;
3685 if (conf->bypass_count < 0)
3686 conf->bypass_count = 0;
3690 list_del_init(&sh->lru);
3691 atomic_inc(&sh->count);
3692 BUG_ON(atomic_read(&sh->count) != 1);
3696 static void make_request(struct mddev *mddev, struct bio * bi)
3698 struct r5conf *conf = mddev->private;
3700 sector_t new_sector;
3701 sector_t logical_sector, last_sector;
3702 struct stripe_head *sh;
3703 const int rw = bio_data_dir(bi);
3707 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3708 md_flush_request(mddev, bi);
3712 md_write_start(mddev, bi);
3715 mddev->reshape_position == MaxSector &&
3716 chunk_aligned_read(mddev,bi))
3719 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3720 last_sector = bi->bi_sector + (bi->bi_size>>9);
3722 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3724 plugged = mddev_check_plugged(mddev);
3725 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3727 int disks, data_disks;
3732 disks = conf->raid_disks;
3733 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3734 if (unlikely(conf->reshape_progress != MaxSector)) {
3735 /* spinlock is needed as reshape_progress may be
3736 * 64bit on a 32bit platform, and so it might be
3737 * possible to see a half-updated value
3738 * Of course reshape_progress could change after
3739 * the lock is dropped, so once we get a reference
3740 * to the stripe that we think it is, we will have
3743 spin_lock_irq(&conf->device_lock);
3744 if (mddev->delta_disks < 0
3745 ? logical_sector < conf->reshape_progress
3746 : logical_sector >= conf->reshape_progress) {
3747 disks = conf->previous_raid_disks;
3750 if (mddev->delta_disks < 0
3751 ? logical_sector < conf->reshape_safe
3752 : logical_sector >= conf->reshape_safe) {
3753 spin_unlock_irq(&conf->device_lock);
3758 spin_unlock_irq(&conf->device_lock);
3760 data_disks = disks - conf->max_degraded;
3762 new_sector = raid5_compute_sector(conf, logical_sector,
3765 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3766 (unsigned long long)new_sector,
3767 (unsigned long long)logical_sector);
3769 sh = get_active_stripe(conf, new_sector, previous,
3770 (bi->bi_rw&RWA_MASK), 0);
3772 if (unlikely(previous)) {
3773 /* expansion might have moved on while waiting for a
3774 * stripe, so we must do the range check again.
3775 * Expansion could still move past after this
3776 * test, but as we are holding a reference to
3777 * 'sh', we know that if that happens,
3778 * STRIPE_EXPANDING will get set and the expansion
3779 * won't proceed until we finish with the stripe.
3782 spin_lock_irq(&conf->device_lock);
3783 if (mddev->delta_disks < 0
3784 ? logical_sector >= conf->reshape_progress
3785 : logical_sector < conf->reshape_progress)
3786 /* mismatch, need to try again */
3788 spin_unlock_irq(&conf->device_lock);
3797 logical_sector >= mddev->suspend_lo &&
3798 logical_sector < mddev->suspend_hi) {
3800 /* As the suspend_* range is controlled by
3801 * userspace, we want an interruptible
3804 flush_signals(current);
3805 prepare_to_wait(&conf->wait_for_overlap,
3806 &w, TASK_INTERRUPTIBLE);
3807 if (logical_sector >= mddev->suspend_lo &&
3808 logical_sector < mddev->suspend_hi)
3813 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3814 !add_stripe_bio(sh, bi, dd_idx, rw)) {
3815 /* Stripe is busy expanding or
3816 * add failed due to overlap. Flush everything
3819 md_wakeup_thread(mddev->thread);
3824 finish_wait(&conf->wait_for_overlap, &w);
3825 set_bit(STRIPE_HANDLE, &sh->state);
3826 clear_bit(STRIPE_DELAYED, &sh->state);
3827 if ((bi->bi_rw & REQ_SYNC) &&
3828 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3829 atomic_inc(&conf->preread_active_stripes);
3832 /* cannot get stripe for read-ahead, just give-up */
3833 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3834 finish_wait(&conf->wait_for_overlap, &w);
3840 md_wakeup_thread(mddev->thread);
3842 spin_lock_irq(&conf->device_lock);
3843 remaining = raid5_dec_bi_phys_segments(bi);
3844 spin_unlock_irq(&conf->device_lock);
3845 if (remaining == 0) {
3848 md_write_end(mddev);
3854 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
3856 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
3858 /* reshaping is quite different to recovery/resync so it is
3859 * handled quite separately ... here.
3861 * On each call to sync_request, we gather one chunk worth of
3862 * destination stripes and flag them as expanding.
3863 * Then we find all the source stripes and request reads.
3864 * As the reads complete, handle_stripe will copy the data
3865 * into the destination stripe and release that stripe.
3867 struct r5conf *conf = mddev->private;
3868 struct stripe_head *sh;
3869 sector_t first_sector, last_sector;
3870 int raid_disks = conf->previous_raid_disks;
3871 int data_disks = raid_disks - conf->max_degraded;
3872 int new_data_disks = conf->raid_disks - conf->max_degraded;
3875 sector_t writepos, readpos, safepos;
3876 sector_t stripe_addr;
3877 int reshape_sectors;
3878 struct list_head stripes;
3880 if (sector_nr == 0) {
3881 /* If restarting in the middle, skip the initial sectors */
3882 if (mddev->delta_disks < 0 &&
3883 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3884 sector_nr = raid5_size(mddev, 0, 0)
3885 - conf->reshape_progress;
3886 } else if (mddev->delta_disks >= 0 &&
3887 conf->reshape_progress > 0)
3888 sector_nr = conf->reshape_progress;
3889 sector_div(sector_nr, new_data_disks);
3891 mddev->curr_resync_completed = sector_nr;
3892 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3898 /* We need to process a full chunk at a time.
3899 * If old and new chunk sizes differ, we need to process the
3902 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3903 reshape_sectors = mddev->new_chunk_sectors;
3905 reshape_sectors = mddev->chunk_sectors;
3907 /* we update the metadata when there is more than 3Meg
3908 * in the block range (that is rather arbitrary, should
3909 * probably be time based) or when the data about to be
3910 * copied would over-write the source of the data at
3911 * the front of the range.
3912 * i.e. one new_stripe along from reshape_progress new_maps
3913 * to after where reshape_safe old_maps to
3915 writepos = conf->reshape_progress;
3916 sector_div(writepos, new_data_disks);
3917 readpos = conf->reshape_progress;
3918 sector_div(readpos, data_disks);
3919 safepos = conf->reshape_safe;
3920 sector_div(safepos, data_disks);
3921 if (mddev->delta_disks < 0) {
3922 writepos -= min_t(sector_t, reshape_sectors, writepos);
3923 readpos += reshape_sectors;
3924 safepos += reshape_sectors;
3926 writepos += reshape_sectors;
3927 readpos -= min_t(sector_t, reshape_sectors, readpos);
3928 safepos -= min_t(sector_t, reshape_sectors, safepos);
3931 /* 'writepos' is the most advanced device address we might write.
3932 * 'readpos' is the least advanced device address we might read.
3933 * 'safepos' is the least address recorded in the metadata as having
3935 * If 'readpos' is behind 'writepos', then there is no way that we can
3936 * ensure safety in the face of a crash - that must be done by userspace
3937 * making a backup of the data. So in that case there is no particular
3938 * rush to update metadata.
3939 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3940 * update the metadata to advance 'safepos' to match 'readpos' so that
3941 * we can be safe in the event of a crash.
3942 * So we insist on updating metadata if safepos is behind writepos and
3943 * readpos is beyond writepos.
3944 * In any case, update the metadata every 10 seconds.
3945 * Maybe that number should be configurable, but I'm not sure it is
3946 * worth it.... maybe it could be a multiple of safemode_delay???
3948 if ((mddev->delta_disks < 0
3949 ? (safepos > writepos && readpos < writepos)
3950 : (safepos < writepos && readpos > writepos)) ||
3951 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3952 /* Cannot proceed until we've updated the superblock... */
3953 wait_event(conf->wait_for_overlap,
3954 atomic_read(&conf->reshape_stripes)==0);
3955 mddev->reshape_position = conf->reshape_progress;
3956 mddev->curr_resync_completed = sector_nr;
3957 conf->reshape_checkpoint = jiffies;
3958 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3959 md_wakeup_thread(mddev->thread);
3960 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3961 kthread_should_stop());
3962 spin_lock_irq(&conf->device_lock);
3963 conf->reshape_safe = mddev->reshape_position;
3964 spin_unlock_irq(&conf->device_lock);
3965 wake_up(&conf->wait_for_overlap);
3966 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3969 if (mddev->delta_disks < 0) {
3970 BUG_ON(conf->reshape_progress == 0);
3971 stripe_addr = writepos;
3972 BUG_ON((mddev->dev_sectors &
3973 ~((sector_t)reshape_sectors - 1))
3974 - reshape_sectors - stripe_addr
3977 BUG_ON(writepos != sector_nr + reshape_sectors);
3978 stripe_addr = sector_nr;
3980 INIT_LIST_HEAD(&stripes);
3981 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3983 int skipped_disk = 0;
3984 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
3985 set_bit(STRIPE_EXPANDING, &sh->state);
3986 atomic_inc(&conf->reshape_stripes);
3987 /* If any of this stripe is beyond the end of the old
3988 * array, then we need to zero those blocks
3990 for (j=sh->disks; j--;) {
3992 if (j == sh->pd_idx)
3994 if (conf->level == 6 &&
3997 s = compute_blocknr(sh, j, 0);
3998 if (s < raid5_size(mddev, 0, 0)) {
4002 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4003 set_bit(R5_Expanded, &sh->dev[j].flags);
4004 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4006 if (!skipped_disk) {
4007 set_bit(STRIPE_EXPAND_READY, &sh->state);
4008 set_bit(STRIPE_HANDLE, &sh->state);
4010 list_add(&sh->lru, &stripes);
4012 spin_lock_irq(&conf->device_lock);
4013 if (mddev->delta_disks < 0)
4014 conf->reshape_progress -= reshape_sectors * new_data_disks;
4016 conf->reshape_progress += reshape_sectors * new_data_disks;
4017 spin_unlock_irq(&conf->device_lock);
4018 /* Ok, those stripe are ready. We can start scheduling
4019 * reads on the source stripes.
4020 * The source stripes are determined by mapping the first and last
4021 * block on the destination stripes.
4024 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4027 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4028 * new_data_disks - 1),
4030 if (last_sector >= mddev->dev_sectors)
4031 last_sector = mddev->dev_sectors - 1;
4032 while (first_sector <= last_sector) {
4033 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4034 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4035 set_bit(STRIPE_HANDLE, &sh->state);
4037 first_sector += STRIPE_SECTORS;
4039 /* Now that the sources are clearly marked, we can release
4040 * the destination stripes
4042 while (!list_empty(&stripes)) {
4043 sh = list_entry(stripes.next, struct stripe_head, lru);
4044 list_del_init(&sh->lru);
4047 /* If this takes us to the resync_max point where we have to pause,
4048 * then we need to write out the superblock.
4050 sector_nr += reshape_sectors;
4051 if ((sector_nr - mddev->curr_resync_completed) * 2
4052 >= mddev->resync_max - mddev->curr_resync_completed) {
4053 /* Cannot proceed until we've updated the superblock... */
4054 wait_event(conf->wait_for_overlap,
4055 atomic_read(&conf->reshape_stripes) == 0);
4056 mddev->reshape_position = conf->reshape_progress;
4057 mddev->curr_resync_completed = sector_nr;
4058 conf->reshape_checkpoint = jiffies;
4059 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4060 md_wakeup_thread(mddev->thread);
4061 wait_event(mddev->sb_wait,
4062 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4063 || kthread_should_stop());
4064 spin_lock_irq(&conf->device_lock);
4065 conf->reshape_safe = mddev->reshape_position;
4066 spin_unlock_irq(&conf->device_lock);
4067 wake_up(&conf->wait_for_overlap);
4068 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4070 return reshape_sectors;
4073 /* FIXME go_faster isn't used */
4074 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4076 struct r5conf *conf = mddev->private;
4077 struct stripe_head *sh;
4078 sector_t max_sector = mddev->dev_sectors;
4079 sector_t sync_blocks;
4080 int still_degraded = 0;
4083 if (sector_nr >= max_sector) {
4084 /* just being told to finish up .. nothing much to do */
4086 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4091 if (mddev->curr_resync < max_sector) /* aborted */
4092 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4094 else /* completed sync */
4096 bitmap_close_sync(mddev->bitmap);
4101 /* Allow raid5_quiesce to complete */
4102 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4104 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4105 return reshape_request(mddev, sector_nr, skipped);
4107 /* No need to check resync_max as we never do more than one
4108 * stripe, and as resync_max will always be on a chunk boundary,
4109 * if the check in md_do_sync didn't fire, there is no chance
4110 * of overstepping resync_max here
4113 /* if there is too many failed drives and we are trying
4114 * to resync, then assert that we are finished, because there is
4115 * nothing we can do.
4117 if (mddev->degraded >= conf->max_degraded &&
4118 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4119 sector_t rv = mddev->dev_sectors - sector_nr;
4123 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4124 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4125 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4126 /* we can skip this block, and probably more */
4127 sync_blocks /= STRIPE_SECTORS;
4129 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4133 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4135 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4137 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4138 /* make sure we don't swamp the stripe cache if someone else
4139 * is trying to get access
4141 schedule_timeout_uninterruptible(1);
4143 /* Need to check if array will still be degraded after recovery/resync
4144 * We don't need to check the 'failed' flag as when that gets set,
4147 for (i = 0; i < conf->raid_disks; i++)
4148 if (conf->disks[i].rdev == NULL)
4151 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4153 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4158 return STRIPE_SECTORS;
4161 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4163 /* We may not be able to submit a whole bio at once as there
4164 * may not be enough stripe_heads available.
4165 * We cannot pre-allocate enough stripe_heads as we may need
4166 * more than exist in the cache (if we allow ever large chunks).
4167 * So we do one stripe head at a time and record in
4168 * ->bi_hw_segments how many have been done.
4170 * We *know* that this entire raid_bio is in one chunk, so
4171 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4173 struct stripe_head *sh;
4175 sector_t sector, logical_sector, last_sector;
4180 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4181 sector = raid5_compute_sector(conf, logical_sector,
4183 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4185 for (; logical_sector < last_sector;
4186 logical_sector += STRIPE_SECTORS,
4187 sector += STRIPE_SECTORS,
4190 if (scnt < raid5_bi_hw_segments(raid_bio))
4191 /* already done this stripe */
4194 sh = get_active_stripe(conf, sector, 0, 1, 0);
4197 /* failed to get a stripe - must wait */
4198 raid5_set_bi_hw_segments(raid_bio, scnt);
4199 conf->retry_read_aligned = raid_bio;
4203 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4204 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4206 raid5_set_bi_hw_segments(raid_bio, scnt);
4207 conf->retry_read_aligned = raid_bio;
4215 spin_lock_irq(&conf->device_lock);
4216 remaining = raid5_dec_bi_phys_segments(raid_bio);
4217 spin_unlock_irq(&conf->device_lock);
4219 bio_endio(raid_bio, 0);
4220 if (atomic_dec_and_test(&conf->active_aligned_reads))
4221 wake_up(&conf->wait_for_stripe);
4227 * This is our raid5 kernel thread.
4229 * We scan the hash table for stripes which can be handled now.
4230 * During the scan, completed stripes are saved for us by the interrupt
4231 * handler, so that they will not have to wait for our next wakeup.
4233 static void raid5d(struct mddev *mddev)
4235 struct stripe_head *sh;
4236 struct r5conf *conf = mddev->private;
4238 struct blk_plug plug;
4240 pr_debug("+++ raid5d active\n");
4242 md_check_recovery(mddev);
4244 blk_start_plug(&plug);
4246 spin_lock_irq(&conf->device_lock);
4250 if (atomic_read(&mddev->plug_cnt) == 0 &&
4251 !list_empty(&conf->bitmap_list)) {
4252 /* Now is a good time to flush some bitmap updates */
4254 spin_unlock_irq(&conf->device_lock);
4255 bitmap_unplug(mddev->bitmap);
4256 spin_lock_irq(&conf->device_lock);
4257 conf->seq_write = conf->seq_flush;
4258 activate_bit_delay(conf);
4260 if (atomic_read(&mddev->plug_cnt) == 0)
4261 raid5_activate_delayed(conf);
4263 while ((bio = remove_bio_from_retry(conf))) {
4265 spin_unlock_irq(&conf->device_lock);
4266 ok = retry_aligned_read(conf, bio);
4267 spin_lock_irq(&conf->device_lock);
4273 sh = __get_priority_stripe(conf);
4277 spin_unlock_irq(&conf->device_lock);
4284 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4285 md_check_recovery(mddev);
4287 spin_lock_irq(&conf->device_lock);
4289 pr_debug("%d stripes handled\n", handled);
4291 spin_unlock_irq(&conf->device_lock);
4293 async_tx_issue_pending_all();
4294 blk_finish_plug(&plug);
4296 pr_debug("--- raid5d inactive\n");
4300 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4302 struct r5conf *conf = mddev->private;
4304 return sprintf(page, "%d\n", conf->max_nr_stripes);
4310 raid5_set_cache_size(struct mddev *mddev, int size)
4312 struct r5conf *conf = mddev->private;
4315 if (size <= 16 || size > 32768)
4317 while (size < conf->max_nr_stripes) {
4318 if (drop_one_stripe(conf))
4319 conf->max_nr_stripes--;
4323 err = md_allow_write(mddev);
4326 while (size > conf->max_nr_stripes) {
4327 if (grow_one_stripe(conf))
4328 conf->max_nr_stripes++;
4333 EXPORT_SYMBOL(raid5_set_cache_size);
4336 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4338 struct r5conf *conf = mddev->private;
4342 if (len >= PAGE_SIZE)
4347 if (strict_strtoul(page, 10, &new))
4349 err = raid5_set_cache_size(mddev, new);
4355 static struct md_sysfs_entry
4356 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4357 raid5_show_stripe_cache_size,
4358 raid5_store_stripe_cache_size);
4361 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4363 struct r5conf *conf = mddev->private;
4365 return sprintf(page, "%d\n", conf->bypass_threshold);
4371 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4373 struct r5conf *conf = mddev->private;
4375 if (len >= PAGE_SIZE)
4380 if (strict_strtoul(page, 10, &new))
4382 if (new > conf->max_nr_stripes)
4384 conf->bypass_threshold = new;
4388 static struct md_sysfs_entry
4389 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4391 raid5_show_preread_threshold,
4392 raid5_store_preread_threshold);
4395 stripe_cache_active_show(struct mddev *mddev, char *page)
4397 struct r5conf *conf = mddev->private;
4399 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4404 static struct md_sysfs_entry
4405 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4407 static struct attribute *raid5_attrs[] = {
4408 &raid5_stripecache_size.attr,
4409 &raid5_stripecache_active.attr,
4410 &raid5_preread_bypass_threshold.attr,
4413 static struct attribute_group raid5_attrs_group = {
4415 .attrs = raid5_attrs,
4419 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4421 struct r5conf *conf = mddev->private;
4424 sectors = mddev->dev_sectors;
4426 /* size is defined by the smallest of previous and new size */
4427 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4429 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4430 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4431 return sectors * (raid_disks - conf->max_degraded);
4434 static void raid5_free_percpu(struct r5conf *conf)
4436 struct raid5_percpu *percpu;
4443 for_each_possible_cpu(cpu) {
4444 percpu = per_cpu_ptr(conf->percpu, cpu);
4445 safe_put_page(percpu->spare_page);
4446 kfree(percpu->scribble);
4448 #ifdef CONFIG_HOTPLUG_CPU
4449 unregister_cpu_notifier(&conf->cpu_notify);
4453 free_percpu(conf->percpu);
4456 static void free_conf(struct r5conf *conf)
4458 shrink_stripes(conf);
4459 raid5_free_percpu(conf);
4461 kfree(conf->stripe_hashtbl);
4465 #ifdef CONFIG_HOTPLUG_CPU
4466 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4469 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4470 long cpu = (long)hcpu;
4471 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4474 case CPU_UP_PREPARE:
4475 case CPU_UP_PREPARE_FROZEN:
4476 if (conf->level == 6 && !percpu->spare_page)
4477 percpu->spare_page = alloc_page(GFP_KERNEL);
4478 if (!percpu->scribble)
4479 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4481 if (!percpu->scribble ||
4482 (conf->level == 6 && !percpu->spare_page)) {
4483 safe_put_page(percpu->spare_page);
4484 kfree(percpu->scribble);
4485 pr_err("%s: failed memory allocation for cpu%ld\n",
4487 return notifier_from_errno(-ENOMEM);
4491 case CPU_DEAD_FROZEN:
4492 safe_put_page(percpu->spare_page);
4493 kfree(percpu->scribble);
4494 percpu->spare_page = NULL;
4495 percpu->scribble = NULL;
4504 static int raid5_alloc_percpu(struct r5conf *conf)
4507 struct page *spare_page;
4508 struct raid5_percpu __percpu *allcpus;
4512 allcpus = alloc_percpu(struct raid5_percpu);
4515 conf->percpu = allcpus;
4519 for_each_present_cpu(cpu) {
4520 if (conf->level == 6) {
4521 spare_page = alloc_page(GFP_KERNEL);
4526 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4528 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4533 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4535 #ifdef CONFIG_HOTPLUG_CPU
4536 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4537 conf->cpu_notify.priority = 0;
4539 err = register_cpu_notifier(&conf->cpu_notify);
4546 static struct r5conf *setup_conf(struct mddev *mddev)
4548 struct r5conf *conf;
4549 int raid_disk, memory, max_disks;
4550 struct md_rdev *rdev;
4551 struct disk_info *disk;
4553 if (mddev->new_level != 5
4554 && mddev->new_level != 4
4555 && mddev->new_level != 6) {
4556 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4557 mdname(mddev), mddev->new_level);
4558 return ERR_PTR(-EIO);
4560 if ((mddev->new_level == 5
4561 && !algorithm_valid_raid5(mddev->new_layout)) ||
4562 (mddev->new_level == 6
4563 && !algorithm_valid_raid6(mddev->new_layout))) {
4564 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4565 mdname(mddev), mddev->new_layout);
4566 return ERR_PTR(-EIO);
4568 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4569 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4570 mdname(mddev), mddev->raid_disks);
4571 return ERR_PTR(-EINVAL);
4574 if (!mddev->new_chunk_sectors ||
4575 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4576 !is_power_of_2(mddev->new_chunk_sectors)) {
4577 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4578 mdname(mddev), mddev->new_chunk_sectors << 9);
4579 return ERR_PTR(-EINVAL);
4582 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4585 spin_lock_init(&conf->device_lock);
4586 init_waitqueue_head(&conf->wait_for_stripe);
4587 init_waitqueue_head(&conf->wait_for_overlap);
4588 INIT_LIST_HEAD(&conf->handle_list);
4589 INIT_LIST_HEAD(&conf->hold_list);
4590 INIT_LIST_HEAD(&conf->delayed_list);
4591 INIT_LIST_HEAD(&conf->bitmap_list);
4592 INIT_LIST_HEAD(&conf->inactive_list);
4593 atomic_set(&conf->active_stripes, 0);
4594 atomic_set(&conf->preread_active_stripes, 0);
4595 atomic_set(&conf->active_aligned_reads, 0);
4596 conf->bypass_threshold = BYPASS_THRESHOLD;
4597 conf->recovery_disabled = mddev->recovery_disabled - 1;
4599 conf->raid_disks = mddev->raid_disks;
4600 if (mddev->reshape_position == MaxSector)
4601 conf->previous_raid_disks = mddev->raid_disks;
4603 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4604 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4605 conf->scribble_len = scribble_len(max_disks);
4607 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4612 conf->mddev = mddev;
4614 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4617 conf->level = mddev->new_level;
4618 if (raid5_alloc_percpu(conf) != 0)
4621 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4623 list_for_each_entry(rdev, &mddev->disks, same_set) {
4624 raid_disk = rdev->raid_disk;
4625 if (raid_disk >= max_disks
4628 disk = conf->disks + raid_disk;
4632 if (test_bit(In_sync, &rdev->flags)) {
4633 char b[BDEVNAME_SIZE];
4634 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4636 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4637 } else if (rdev->saved_raid_disk != raid_disk)
4638 /* Cannot rely on bitmap to complete recovery */
4642 conf->chunk_sectors = mddev->new_chunk_sectors;
4643 conf->level = mddev->new_level;
4644 if (conf->level == 6)
4645 conf->max_degraded = 2;
4647 conf->max_degraded = 1;
4648 conf->algorithm = mddev->new_layout;
4649 conf->max_nr_stripes = NR_STRIPES;
4650 conf->reshape_progress = mddev->reshape_position;
4651 if (conf->reshape_progress != MaxSector) {
4652 conf->prev_chunk_sectors = mddev->chunk_sectors;
4653 conf->prev_algo = mddev->layout;
4656 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4657 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4658 if (grow_stripes(conf, conf->max_nr_stripes)) {
4660 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4661 mdname(mddev), memory);
4664 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4665 mdname(mddev), memory);
4667 conf->thread = md_register_thread(raid5d, mddev, NULL);
4668 if (!conf->thread) {
4670 "md/raid:%s: couldn't allocate thread.\n",
4680 return ERR_PTR(-EIO);
4682 return ERR_PTR(-ENOMEM);
4686 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4689 case ALGORITHM_PARITY_0:
4690 if (raid_disk < max_degraded)
4693 case ALGORITHM_PARITY_N:
4694 if (raid_disk >= raid_disks - max_degraded)
4697 case ALGORITHM_PARITY_0_6:
4698 if (raid_disk == 0 ||
4699 raid_disk == raid_disks - 1)
4702 case ALGORITHM_LEFT_ASYMMETRIC_6:
4703 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4704 case ALGORITHM_LEFT_SYMMETRIC_6:
4705 case ALGORITHM_RIGHT_SYMMETRIC_6:
4706 if (raid_disk == raid_disks - 1)
4712 static int run(struct mddev *mddev)
4714 struct r5conf *conf;
4715 int working_disks = 0;
4716 int dirty_parity_disks = 0;
4717 struct md_rdev *rdev;
4718 sector_t reshape_offset = 0;
4720 if (mddev->recovery_cp != MaxSector)
4721 printk(KERN_NOTICE "md/raid:%s: not clean"
4722 " -- starting background reconstruction\n",
4724 if (mddev->reshape_position != MaxSector) {
4725 /* Check that we can continue the reshape.
4726 * Currently only disks can change, it must
4727 * increase, and we must be past the point where
4728 * a stripe over-writes itself
4730 sector_t here_new, here_old;
4732 int max_degraded = (mddev->level == 6 ? 2 : 1);
4734 if (mddev->new_level != mddev->level) {
4735 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4736 "required - aborting.\n",
4740 old_disks = mddev->raid_disks - mddev->delta_disks;
4741 /* reshape_position must be on a new-stripe boundary, and one
4742 * further up in new geometry must map after here in old
4745 here_new = mddev->reshape_position;
4746 if (sector_div(here_new, mddev->new_chunk_sectors *
4747 (mddev->raid_disks - max_degraded))) {
4748 printk(KERN_ERR "md/raid:%s: reshape_position not "
4749 "on a stripe boundary\n", mdname(mddev));
4752 reshape_offset = here_new * mddev->new_chunk_sectors;
4753 /* here_new is the stripe we will write to */
4754 here_old = mddev->reshape_position;
4755 sector_div(here_old, mddev->chunk_sectors *
4756 (old_disks-max_degraded));
4757 /* here_old is the first stripe that we might need to read
4759 if (mddev->delta_disks == 0) {
4760 /* We cannot be sure it is safe to start an in-place
4761 * reshape. It is only safe if user-space if monitoring
4762 * and taking constant backups.
4763 * mdadm always starts a situation like this in
4764 * readonly mode so it can take control before
4765 * allowing any writes. So just check for that.
4767 if ((here_new * mddev->new_chunk_sectors !=
4768 here_old * mddev->chunk_sectors) ||
4770 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4771 " in read-only mode - aborting\n",
4775 } else if (mddev->delta_disks < 0
4776 ? (here_new * mddev->new_chunk_sectors <=
4777 here_old * mddev->chunk_sectors)
4778 : (here_new * mddev->new_chunk_sectors >=
4779 here_old * mddev->chunk_sectors)) {
4780 /* Reading from the same stripe as writing to - bad */
4781 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4782 "auto-recovery - aborting.\n",
4786 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4788 /* OK, we should be able to continue; */
4790 BUG_ON(mddev->level != mddev->new_level);
4791 BUG_ON(mddev->layout != mddev->new_layout);
4792 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4793 BUG_ON(mddev->delta_disks != 0);
4796 if (mddev->private == NULL)
4797 conf = setup_conf(mddev);
4799 conf = mddev->private;
4802 return PTR_ERR(conf);
4804 mddev->thread = conf->thread;
4805 conf->thread = NULL;
4806 mddev->private = conf;
4809 * 0 for a fully functional array, 1 or 2 for a degraded array.
4811 list_for_each_entry(rdev, &mddev->disks, same_set) {
4812 if (rdev->raid_disk < 0)
4814 if (test_bit(In_sync, &rdev->flags)) {
4818 /* This disc is not fully in-sync. However if it
4819 * just stored parity (beyond the recovery_offset),
4820 * when we don't need to be concerned about the
4821 * array being dirty.
4822 * When reshape goes 'backwards', we never have
4823 * partially completed devices, so we only need
4824 * to worry about reshape going forwards.
4826 /* Hack because v0.91 doesn't store recovery_offset properly. */
4827 if (mddev->major_version == 0 &&
4828 mddev->minor_version > 90)
4829 rdev->recovery_offset = reshape_offset;
4831 if (rdev->recovery_offset < reshape_offset) {
4832 /* We need to check old and new layout */
4833 if (!only_parity(rdev->raid_disk,
4836 conf->max_degraded))
4839 if (!only_parity(rdev->raid_disk,
4841 conf->previous_raid_disks,
4842 conf->max_degraded))
4844 dirty_parity_disks++;
4847 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4850 if (has_failed(conf)) {
4851 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4852 " (%d/%d failed)\n",
4853 mdname(mddev), mddev->degraded, conf->raid_disks);
4857 /* device size must be a multiple of chunk size */
4858 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4859 mddev->resync_max_sectors = mddev->dev_sectors;
4861 if (mddev->degraded > dirty_parity_disks &&
4862 mddev->recovery_cp != MaxSector) {
4863 if (mddev->ok_start_degraded)
4865 "md/raid:%s: starting dirty degraded array"
4866 " - data corruption possible.\n",
4870 "md/raid:%s: cannot start dirty degraded array.\n",
4876 if (mddev->degraded == 0)
4877 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4878 " devices, algorithm %d\n", mdname(mddev), conf->level,
4879 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4882 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4883 " out of %d devices, algorithm %d\n",
4884 mdname(mddev), conf->level,
4885 mddev->raid_disks - mddev->degraded,
4886 mddev->raid_disks, mddev->new_layout);
4888 print_raid5_conf(conf);
4890 if (conf->reshape_progress != MaxSector) {
4891 conf->reshape_safe = conf->reshape_progress;
4892 atomic_set(&conf->reshape_stripes, 0);
4893 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4894 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4895 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4896 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4897 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4902 /* Ok, everything is just fine now */
4903 if (mddev->to_remove == &raid5_attrs_group)
4904 mddev->to_remove = NULL;
4905 else if (mddev->kobj.sd &&
4906 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4908 "raid5: failed to create sysfs attributes for %s\n",
4910 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4914 /* read-ahead size must cover two whole stripes, which
4915 * is 2 * (datadisks) * chunksize where 'n' is the
4916 * number of raid devices
4918 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4919 int stripe = data_disks *
4920 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4921 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4922 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4924 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4926 mddev->queue->backing_dev_info.congested_data = mddev;
4927 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4929 chunk_size = mddev->chunk_sectors << 9;
4930 blk_queue_io_min(mddev->queue, chunk_size);
4931 blk_queue_io_opt(mddev->queue, chunk_size *
4932 (conf->raid_disks - conf->max_degraded));
4934 list_for_each_entry(rdev, &mddev->disks, same_set)
4935 disk_stack_limits(mddev->gendisk, rdev->bdev,
4936 rdev->data_offset << 9);
4941 md_unregister_thread(&mddev->thread);
4942 print_raid5_conf(conf);
4944 mddev->private = NULL;
4945 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
4949 static int stop(struct mddev *mddev)
4951 struct r5conf *conf = mddev->private;
4953 md_unregister_thread(&mddev->thread);
4955 mddev->queue->backing_dev_info.congested_fn = NULL;
4957 mddev->private = NULL;
4958 mddev->to_remove = &raid5_attrs_group;
4962 static void status(struct seq_file *seq, struct mddev *mddev)
4964 struct r5conf *conf = mddev->private;
4967 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4968 mddev->chunk_sectors / 2, mddev->layout);
4969 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4970 for (i = 0; i < conf->raid_disks; i++)
4971 seq_printf (seq, "%s",
4972 conf->disks[i].rdev &&
4973 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4974 seq_printf (seq, "]");
4977 static void print_raid5_conf (struct r5conf *conf)
4980 struct disk_info *tmp;
4982 printk(KERN_DEBUG "RAID conf printout:\n");
4984 printk("(conf==NULL)\n");
4987 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
4989 conf->raid_disks - conf->mddev->degraded);
4991 for (i = 0; i < conf->raid_disks; i++) {
4992 char b[BDEVNAME_SIZE];
4993 tmp = conf->disks + i;
4995 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
4996 i, !test_bit(Faulty, &tmp->rdev->flags),
4997 bdevname(tmp->rdev->bdev, b));
5001 static int raid5_spare_active(struct mddev *mddev)
5004 struct r5conf *conf = mddev->private;
5005 struct disk_info *tmp;
5007 unsigned long flags;
5009 for (i = 0; i < conf->raid_disks; i++) {
5010 tmp = conf->disks + i;
5012 && tmp->rdev->recovery_offset == MaxSector
5013 && !test_bit(Faulty, &tmp->rdev->flags)
5014 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5016 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5019 spin_lock_irqsave(&conf->device_lock, flags);
5020 mddev->degraded -= count;
5021 spin_unlock_irqrestore(&conf->device_lock, flags);
5022 print_raid5_conf(conf);
5026 static int raid5_remove_disk(struct mddev *mddev, int number)
5028 struct r5conf *conf = mddev->private;
5030 struct md_rdev *rdev;
5031 struct disk_info *p = conf->disks + number;
5033 print_raid5_conf(conf);
5036 if (number >= conf->raid_disks &&
5037 conf->reshape_progress == MaxSector)
5038 clear_bit(In_sync, &rdev->flags);
5040 if (test_bit(In_sync, &rdev->flags) ||
5041 atomic_read(&rdev->nr_pending)) {
5045 /* Only remove non-faulty devices if recovery
5048 if (!test_bit(Faulty, &rdev->flags) &&
5049 mddev->recovery_disabled != conf->recovery_disabled &&
5050 !has_failed(conf) &&
5051 number < conf->raid_disks) {
5057 if (atomic_read(&rdev->nr_pending)) {
5058 /* lost the race, try later */
5065 print_raid5_conf(conf);
5069 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5071 struct r5conf *conf = mddev->private;
5074 struct disk_info *p;
5076 int last = conf->raid_disks - 1;
5078 if (mddev->recovery_disabled == conf->recovery_disabled)
5081 if (has_failed(conf))
5082 /* no point adding a device */
5085 if (rdev->raid_disk >= 0)
5086 first = last = rdev->raid_disk;
5089 * find the disk ... but prefer rdev->saved_raid_disk
5092 if (rdev->saved_raid_disk >= 0 &&
5093 rdev->saved_raid_disk >= first &&
5094 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5095 disk = rdev->saved_raid_disk;
5098 for ( ; disk <= last ; disk++)
5099 if ((p=conf->disks + disk)->rdev == NULL) {
5100 clear_bit(In_sync, &rdev->flags);
5101 rdev->raid_disk = disk;
5103 if (rdev->saved_raid_disk != disk)
5105 rcu_assign_pointer(p->rdev, rdev);
5108 print_raid5_conf(conf);
5112 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5114 /* no resync is happening, and there is enough space
5115 * on all devices, so we can resize.
5116 * We need to make sure resync covers any new space.
5117 * If the array is shrinking we should possibly wait until
5118 * any io in the removed space completes, but it hardly seems
5121 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5122 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5123 mddev->raid_disks));
5124 if (mddev->array_sectors >
5125 raid5_size(mddev, sectors, mddev->raid_disks))
5127 set_capacity(mddev->gendisk, mddev->array_sectors);
5128 revalidate_disk(mddev->gendisk);
5129 if (sectors > mddev->dev_sectors &&
5130 mddev->recovery_cp > mddev->dev_sectors) {
5131 mddev->recovery_cp = mddev->dev_sectors;
5132 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5134 mddev->dev_sectors = sectors;
5135 mddev->resync_max_sectors = sectors;
5139 static int check_stripe_cache(struct mddev *mddev)
5141 /* Can only proceed if there are plenty of stripe_heads.
5142 * We need a minimum of one full stripe,, and for sensible progress
5143 * it is best to have about 4 times that.
5144 * If we require 4 times, then the default 256 4K stripe_heads will
5145 * allow for chunk sizes up to 256K, which is probably OK.
5146 * If the chunk size is greater, user-space should request more
5147 * stripe_heads first.
5149 struct r5conf *conf = mddev->private;
5150 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5151 > conf->max_nr_stripes ||
5152 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5153 > conf->max_nr_stripes) {
5154 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5156 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5163 static int check_reshape(struct mddev *mddev)
5165 struct r5conf *conf = mddev->private;
5167 if (mddev->delta_disks == 0 &&
5168 mddev->new_layout == mddev->layout &&
5169 mddev->new_chunk_sectors == mddev->chunk_sectors)
5170 return 0; /* nothing to do */
5172 /* Cannot grow a bitmap yet */
5174 if (has_failed(conf))
5176 if (mddev->delta_disks < 0) {
5177 /* We might be able to shrink, but the devices must
5178 * be made bigger first.
5179 * For raid6, 4 is the minimum size.
5180 * Otherwise 2 is the minimum
5183 if (mddev->level == 6)
5185 if (mddev->raid_disks + mddev->delta_disks < min)
5189 if (!check_stripe_cache(mddev))
5192 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5195 static int raid5_start_reshape(struct mddev *mddev)
5197 struct r5conf *conf = mddev->private;
5198 struct md_rdev *rdev;
5200 unsigned long flags;
5202 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5205 if (!check_stripe_cache(mddev))
5208 list_for_each_entry(rdev, &mddev->disks, same_set)
5209 if (!test_bit(In_sync, &rdev->flags)
5210 && !test_bit(Faulty, &rdev->flags))
5213 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5214 /* Not enough devices even to make a degraded array
5219 /* Refuse to reduce size of the array. Any reductions in
5220 * array size must be through explicit setting of array_size
5223 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5224 < mddev->array_sectors) {
5225 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5226 "before number of disks\n", mdname(mddev));
5230 atomic_set(&conf->reshape_stripes, 0);
5231 spin_lock_irq(&conf->device_lock);
5232 conf->previous_raid_disks = conf->raid_disks;
5233 conf->raid_disks += mddev->delta_disks;
5234 conf->prev_chunk_sectors = conf->chunk_sectors;
5235 conf->chunk_sectors = mddev->new_chunk_sectors;
5236 conf->prev_algo = conf->algorithm;
5237 conf->algorithm = mddev->new_layout;
5238 if (mddev->delta_disks < 0)
5239 conf->reshape_progress = raid5_size(mddev, 0, 0);
5241 conf->reshape_progress = 0;
5242 conf->reshape_safe = conf->reshape_progress;
5244 spin_unlock_irq(&conf->device_lock);
5246 /* Add some new drives, as many as will fit.
5247 * We know there are enough to make the newly sized array work.
5248 * Don't add devices if we are reducing the number of
5249 * devices in the array. This is because it is not possible
5250 * to correctly record the "partially reconstructed" state of
5251 * such devices during the reshape and confusion could result.
5253 if (mddev->delta_disks >= 0) {
5254 int added_devices = 0;
5255 list_for_each_entry(rdev, &mddev->disks, same_set)
5256 if (rdev->raid_disk < 0 &&
5257 !test_bit(Faulty, &rdev->flags)) {
5258 if (raid5_add_disk(mddev, rdev) == 0) {
5260 >= conf->previous_raid_disks) {
5261 set_bit(In_sync, &rdev->flags);
5264 rdev->recovery_offset = 0;
5266 if (sysfs_link_rdev(mddev, rdev))
5267 /* Failure here is OK */;
5269 } else if (rdev->raid_disk >= conf->previous_raid_disks
5270 && !test_bit(Faulty, &rdev->flags)) {
5271 /* This is a spare that was manually added */
5272 set_bit(In_sync, &rdev->flags);
5276 /* When a reshape changes the number of devices,
5277 * ->degraded is measured against the larger of the
5278 * pre and post number of devices.
5280 spin_lock_irqsave(&conf->device_lock, flags);
5281 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5283 spin_unlock_irqrestore(&conf->device_lock, flags);
5285 mddev->raid_disks = conf->raid_disks;
5286 mddev->reshape_position = conf->reshape_progress;
5287 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5289 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5290 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5291 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5292 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5293 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5295 if (!mddev->sync_thread) {
5296 mddev->recovery = 0;
5297 spin_lock_irq(&conf->device_lock);
5298 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5299 conf->reshape_progress = MaxSector;
5300 spin_unlock_irq(&conf->device_lock);
5303 conf->reshape_checkpoint = jiffies;
5304 md_wakeup_thread(mddev->sync_thread);
5305 md_new_event(mddev);
5309 /* This is called from the reshape thread and should make any
5310 * changes needed in 'conf'
5312 static void end_reshape(struct r5conf *conf)
5315 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5317 spin_lock_irq(&conf->device_lock);
5318 conf->previous_raid_disks = conf->raid_disks;
5319 conf->reshape_progress = MaxSector;
5320 spin_unlock_irq(&conf->device_lock);
5321 wake_up(&conf->wait_for_overlap);
5323 /* read-ahead size must cover two whole stripes, which is
5324 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5326 if (conf->mddev->queue) {
5327 int data_disks = conf->raid_disks - conf->max_degraded;
5328 int stripe = data_disks * ((conf->chunk_sectors << 9)
5330 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5331 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5336 /* This is called from the raid5d thread with mddev_lock held.
5337 * It makes config changes to the device.
5339 static void raid5_finish_reshape(struct mddev *mddev)
5341 struct r5conf *conf = mddev->private;
5343 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5345 if (mddev->delta_disks > 0) {
5346 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5347 set_capacity(mddev->gendisk, mddev->array_sectors);
5348 revalidate_disk(mddev->gendisk);
5351 mddev->degraded = conf->raid_disks;
5352 for (d = 0; d < conf->raid_disks ; d++)
5353 if (conf->disks[d].rdev &&
5355 &conf->disks[d].rdev->flags))
5357 for (d = conf->raid_disks ;
5358 d < conf->raid_disks - mddev->delta_disks;
5360 struct md_rdev *rdev = conf->disks[d].rdev;
5361 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5362 sysfs_unlink_rdev(mddev, rdev);
5363 rdev->raid_disk = -1;
5367 mddev->layout = conf->algorithm;
5368 mddev->chunk_sectors = conf->chunk_sectors;
5369 mddev->reshape_position = MaxSector;
5370 mddev->delta_disks = 0;
5374 static void raid5_quiesce(struct mddev *mddev, int state)
5376 struct r5conf *conf = mddev->private;
5379 case 2: /* resume for a suspend */
5380 wake_up(&conf->wait_for_overlap);
5383 case 1: /* stop all writes */
5384 spin_lock_irq(&conf->device_lock);
5385 /* '2' tells resync/reshape to pause so that all
5386 * active stripes can drain
5389 wait_event_lock_irq(conf->wait_for_stripe,
5390 atomic_read(&conf->active_stripes) == 0 &&
5391 atomic_read(&conf->active_aligned_reads) == 0,
5392 conf->device_lock, /* nothing */);
5394 spin_unlock_irq(&conf->device_lock);
5395 /* allow reshape to continue */
5396 wake_up(&conf->wait_for_overlap);
5399 case 0: /* re-enable writes */
5400 spin_lock_irq(&conf->device_lock);
5402 wake_up(&conf->wait_for_stripe);
5403 wake_up(&conf->wait_for_overlap);
5404 spin_unlock_irq(&conf->device_lock);
5410 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5412 struct r0conf *raid0_conf = mddev->private;
5415 /* for raid0 takeover only one zone is supported */
5416 if (raid0_conf->nr_strip_zones > 1) {
5417 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5419 return ERR_PTR(-EINVAL);
5422 sectors = raid0_conf->strip_zone[0].zone_end;
5423 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5424 mddev->dev_sectors = sectors;
5425 mddev->new_level = level;
5426 mddev->new_layout = ALGORITHM_PARITY_N;
5427 mddev->new_chunk_sectors = mddev->chunk_sectors;
5428 mddev->raid_disks += 1;
5429 mddev->delta_disks = 1;
5430 /* make sure it will be not marked as dirty */
5431 mddev->recovery_cp = MaxSector;
5433 return setup_conf(mddev);
5437 static void *raid5_takeover_raid1(struct mddev *mddev)
5441 if (mddev->raid_disks != 2 ||
5442 mddev->degraded > 1)
5443 return ERR_PTR(-EINVAL);
5445 /* Should check if there are write-behind devices? */
5447 chunksect = 64*2; /* 64K by default */
5449 /* The array must be an exact multiple of chunksize */
5450 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5453 if ((chunksect<<9) < STRIPE_SIZE)
5454 /* array size does not allow a suitable chunk size */
5455 return ERR_PTR(-EINVAL);
5457 mddev->new_level = 5;
5458 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5459 mddev->new_chunk_sectors = chunksect;
5461 return setup_conf(mddev);
5464 static void *raid5_takeover_raid6(struct mddev *mddev)
5468 switch (mddev->layout) {
5469 case ALGORITHM_LEFT_ASYMMETRIC_6:
5470 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5472 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5473 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5475 case ALGORITHM_LEFT_SYMMETRIC_6:
5476 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5478 case ALGORITHM_RIGHT_SYMMETRIC_6:
5479 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5481 case ALGORITHM_PARITY_0_6:
5482 new_layout = ALGORITHM_PARITY_0;
5484 case ALGORITHM_PARITY_N:
5485 new_layout = ALGORITHM_PARITY_N;
5488 return ERR_PTR(-EINVAL);
5490 mddev->new_level = 5;
5491 mddev->new_layout = new_layout;
5492 mddev->delta_disks = -1;
5493 mddev->raid_disks -= 1;
5494 return setup_conf(mddev);
5498 static int raid5_check_reshape(struct mddev *mddev)
5500 /* For a 2-drive array, the layout and chunk size can be changed
5501 * immediately as not restriping is needed.
5502 * For larger arrays we record the new value - after validation
5503 * to be used by a reshape pass.
5505 struct r5conf *conf = mddev->private;
5506 int new_chunk = mddev->new_chunk_sectors;
5508 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5510 if (new_chunk > 0) {
5511 if (!is_power_of_2(new_chunk))
5513 if (new_chunk < (PAGE_SIZE>>9))
5515 if (mddev->array_sectors & (new_chunk-1))
5516 /* not factor of array size */
5520 /* They look valid */
5522 if (mddev->raid_disks == 2) {
5523 /* can make the change immediately */
5524 if (mddev->new_layout >= 0) {
5525 conf->algorithm = mddev->new_layout;
5526 mddev->layout = mddev->new_layout;
5528 if (new_chunk > 0) {
5529 conf->chunk_sectors = new_chunk ;
5530 mddev->chunk_sectors = new_chunk;
5532 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5533 md_wakeup_thread(mddev->thread);
5535 return check_reshape(mddev);
5538 static int raid6_check_reshape(struct mddev *mddev)
5540 int new_chunk = mddev->new_chunk_sectors;
5542 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5544 if (new_chunk > 0) {
5545 if (!is_power_of_2(new_chunk))
5547 if (new_chunk < (PAGE_SIZE >> 9))
5549 if (mddev->array_sectors & (new_chunk-1))
5550 /* not factor of array size */
5554 /* They look valid */
5555 return check_reshape(mddev);
5558 static void *raid5_takeover(struct mddev *mddev)
5560 /* raid5 can take over:
5561 * raid0 - if there is only one strip zone - make it a raid4 layout
5562 * raid1 - if there are two drives. We need to know the chunk size
5563 * raid4 - trivial - just use a raid4 layout.
5564 * raid6 - Providing it is a *_6 layout
5566 if (mddev->level == 0)
5567 return raid45_takeover_raid0(mddev, 5);
5568 if (mddev->level == 1)
5569 return raid5_takeover_raid1(mddev);
5570 if (mddev->level == 4) {
5571 mddev->new_layout = ALGORITHM_PARITY_N;
5572 mddev->new_level = 5;
5573 return setup_conf(mddev);
5575 if (mddev->level == 6)
5576 return raid5_takeover_raid6(mddev);
5578 return ERR_PTR(-EINVAL);
5581 static void *raid4_takeover(struct mddev *mddev)
5583 /* raid4 can take over:
5584 * raid0 - if there is only one strip zone
5585 * raid5 - if layout is right
5587 if (mddev->level == 0)
5588 return raid45_takeover_raid0(mddev, 4);
5589 if (mddev->level == 5 &&
5590 mddev->layout == ALGORITHM_PARITY_N) {
5591 mddev->new_layout = 0;
5592 mddev->new_level = 4;
5593 return setup_conf(mddev);
5595 return ERR_PTR(-EINVAL);
5598 static struct md_personality raid5_personality;
5600 static void *raid6_takeover(struct mddev *mddev)
5602 /* Currently can only take over a raid5. We map the
5603 * personality to an equivalent raid6 personality
5604 * with the Q block at the end.
5608 if (mddev->pers != &raid5_personality)
5609 return ERR_PTR(-EINVAL);
5610 if (mddev->degraded > 1)
5611 return ERR_PTR(-EINVAL);
5612 if (mddev->raid_disks > 253)
5613 return ERR_PTR(-EINVAL);
5614 if (mddev->raid_disks < 3)
5615 return ERR_PTR(-EINVAL);
5617 switch (mddev->layout) {
5618 case ALGORITHM_LEFT_ASYMMETRIC:
5619 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5621 case ALGORITHM_RIGHT_ASYMMETRIC:
5622 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5624 case ALGORITHM_LEFT_SYMMETRIC:
5625 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5627 case ALGORITHM_RIGHT_SYMMETRIC:
5628 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5630 case ALGORITHM_PARITY_0:
5631 new_layout = ALGORITHM_PARITY_0_6;
5633 case ALGORITHM_PARITY_N:
5634 new_layout = ALGORITHM_PARITY_N;
5637 return ERR_PTR(-EINVAL);
5639 mddev->new_level = 6;
5640 mddev->new_layout = new_layout;
5641 mddev->delta_disks = 1;
5642 mddev->raid_disks += 1;
5643 return setup_conf(mddev);
5647 static struct md_personality raid6_personality =
5651 .owner = THIS_MODULE,
5652 .make_request = make_request,
5656 .error_handler = error,
5657 .hot_add_disk = raid5_add_disk,
5658 .hot_remove_disk= raid5_remove_disk,
5659 .spare_active = raid5_spare_active,
5660 .sync_request = sync_request,
5661 .resize = raid5_resize,
5663 .check_reshape = raid6_check_reshape,
5664 .start_reshape = raid5_start_reshape,
5665 .finish_reshape = raid5_finish_reshape,
5666 .quiesce = raid5_quiesce,
5667 .takeover = raid6_takeover,
5669 static struct md_personality raid5_personality =
5673 .owner = THIS_MODULE,
5674 .make_request = make_request,
5678 .error_handler = error,
5679 .hot_add_disk = raid5_add_disk,
5680 .hot_remove_disk= raid5_remove_disk,
5681 .spare_active = raid5_spare_active,
5682 .sync_request = sync_request,
5683 .resize = raid5_resize,
5685 .check_reshape = raid5_check_reshape,
5686 .start_reshape = raid5_start_reshape,
5687 .finish_reshape = raid5_finish_reshape,
5688 .quiesce = raid5_quiesce,
5689 .takeover = raid5_takeover,
5692 static struct md_personality raid4_personality =
5696 .owner = THIS_MODULE,
5697 .make_request = make_request,
5701 .error_handler = error,
5702 .hot_add_disk = raid5_add_disk,
5703 .hot_remove_disk= raid5_remove_disk,
5704 .spare_active = raid5_spare_active,
5705 .sync_request = sync_request,
5706 .resize = raid5_resize,
5708 .check_reshape = raid5_check_reshape,
5709 .start_reshape = raid5_start_reshape,
5710 .finish_reshape = raid5_finish_reshape,
5711 .quiesce = raid5_quiesce,
5712 .takeover = raid4_takeover,
5715 static int __init raid5_init(void)
5717 register_md_personality(&raid6_personality);
5718 register_md_personality(&raid5_personality);
5719 register_md_personality(&raid4_personality);
5723 static void raid5_exit(void)
5725 unregister_md_personality(&raid6_personality);
5726 unregister_md_personality(&raid5_personality);
5727 unregister_md_personality(&raid4_personality);
5730 module_init(raid5_init);
5731 module_exit(raid5_exit);
5732 MODULE_LICENSE("GPL");
5733 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5734 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5735 MODULE_ALIAS("md-raid5");
5736 MODULE_ALIAS("md-raid4");
5737 MODULE_ALIAS("md-level-5");
5738 MODULE_ALIAS("md-level-4");
5739 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5740 MODULE_ALIAS("md-raid6");
5741 MODULE_ALIAS("md-level-6");
5743 /* This used to be two separate modules, they were: */
5744 MODULE_ALIAS("raid5");
5745 MODULE_ALIAS("raid6");