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 calc_degraded(struct r5conf *conf)
375 int degraded, degraded2;
380 for (i = 0; i < conf->previous_raid_disks; i++) {
381 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
382 if (!rdev || test_bit(Faulty, &rdev->flags))
384 else if (test_bit(In_sync, &rdev->flags))
387 /* not in-sync or faulty.
388 * If the reshape increases the number of devices,
389 * this is being recovered by the reshape, so
390 * this 'previous' section is not in_sync.
391 * If the number of devices is being reduced however,
392 * the device can only be part of the array if
393 * we are reverting a reshape, so this section will
396 if (conf->raid_disks >= conf->previous_raid_disks)
400 if (conf->raid_disks == conf->previous_raid_disks)
404 for (i = 0; i < conf->raid_disks; i++) {
405 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
406 if (!rdev || test_bit(Faulty, &rdev->flags))
408 else if (test_bit(In_sync, &rdev->flags))
411 /* not in-sync or faulty.
412 * If reshape increases the number of devices, this
413 * section has already been recovered, else it
414 * almost certainly hasn't.
416 if (conf->raid_disks <= conf->previous_raid_disks)
420 if (degraded2 > degraded)
425 static int has_failed(struct r5conf *conf)
429 if (conf->mddev->reshape_position == MaxSector)
430 return conf->mddev->degraded > conf->max_degraded;
432 degraded = calc_degraded(conf);
433 if (degraded > conf->max_degraded)
438 static struct stripe_head *
439 get_active_stripe(struct r5conf *conf, sector_t sector,
440 int previous, int noblock, int noquiesce)
442 struct stripe_head *sh;
444 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
446 spin_lock_irq(&conf->device_lock);
449 wait_event_lock_irq(conf->wait_for_stripe,
450 conf->quiesce == 0 || noquiesce,
451 conf->device_lock, /* nothing */);
452 sh = __find_stripe(conf, sector, conf->generation - previous);
454 if (!conf->inactive_blocked)
455 sh = get_free_stripe(conf);
456 if (noblock && sh == NULL)
459 conf->inactive_blocked = 1;
460 wait_event_lock_irq(conf->wait_for_stripe,
461 !list_empty(&conf->inactive_list) &&
462 (atomic_read(&conf->active_stripes)
463 < (conf->max_nr_stripes *3/4)
464 || !conf->inactive_blocked),
467 conf->inactive_blocked = 0;
469 init_stripe(sh, sector, previous);
471 if (atomic_read(&sh->count)) {
472 BUG_ON(!list_empty(&sh->lru)
473 && !test_bit(STRIPE_EXPANDING, &sh->state));
475 if (!test_bit(STRIPE_HANDLE, &sh->state))
476 atomic_inc(&conf->active_stripes);
477 if (list_empty(&sh->lru) &&
478 !test_bit(STRIPE_EXPANDING, &sh->state))
480 list_del_init(&sh->lru);
483 } while (sh == NULL);
486 atomic_inc(&sh->count);
488 spin_unlock_irq(&conf->device_lock);
493 raid5_end_read_request(struct bio *bi, int error);
495 raid5_end_write_request(struct bio *bi, int error);
497 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
499 struct r5conf *conf = sh->raid_conf;
500 int i, disks = sh->disks;
504 for (i = disks; i--; ) {
507 struct md_rdev *rdev;
508 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
509 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
513 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
518 bi = &sh->dev[i].req;
522 bi->bi_end_io = raid5_end_write_request;
524 bi->bi_end_io = raid5_end_read_request;
527 rdev = rcu_dereference(conf->disks[i].rdev);
528 if (rdev && test_bit(Faulty, &rdev->flags))
531 atomic_inc(&rdev->nr_pending);
534 /* We have already checked bad blocks for reads. Now
535 * need to check for writes.
537 while ((rw & WRITE) && rdev &&
538 test_bit(WriteErrorSeen, &rdev->flags)) {
541 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
542 &first_bad, &bad_sectors);
547 set_bit(BlockedBadBlocks, &rdev->flags);
548 if (!conf->mddev->external &&
549 conf->mddev->flags) {
550 /* It is very unlikely, but we might
551 * still need to write out the
552 * bad block log - better give it
554 md_check_recovery(conf->mddev);
556 md_wait_for_blocked_rdev(rdev, conf->mddev);
558 /* Acknowledged bad block - skip the write */
559 rdev_dec_pending(rdev, conf->mddev);
565 if (s->syncing || s->expanding || s->expanded)
566 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
568 set_bit(STRIPE_IO_STARTED, &sh->state);
570 bi->bi_bdev = rdev->bdev;
571 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
572 __func__, (unsigned long long)sh->sector,
574 atomic_inc(&sh->count);
575 bi->bi_sector = sh->sector + rdev->data_offset;
576 bi->bi_flags = 1 << BIO_UPTODATE;
580 bi->bi_io_vec = &sh->dev[i].vec;
581 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
582 bi->bi_io_vec[0].bv_offset = 0;
583 bi->bi_size = STRIPE_SIZE;
585 generic_make_request(bi);
588 set_bit(STRIPE_DEGRADED, &sh->state);
589 pr_debug("skip op %ld on disc %d for sector %llu\n",
590 bi->bi_rw, i, (unsigned long long)sh->sector);
591 clear_bit(R5_LOCKED, &sh->dev[i].flags);
592 set_bit(STRIPE_HANDLE, &sh->state);
597 static struct dma_async_tx_descriptor *
598 async_copy_data(int frombio, struct bio *bio, struct page *page,
599 sector_t sector, struct dma_async_tx_descriptor *tx)
602 struct page *bio_page;
605 struct async_submit_ctl submit;
606 enum async_tx_flags flags = 0;
608 if (bio->bi_sector >= sector)
609 page_offset = (signed)(bio->bi_sector - sector) * 512;
611 page_offset = (signed)(sector - bio->bi_sector) * -512;
614 flags |= ASYNC_TX_FENCE;
615 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
617 bio_for_each_segment(bvl, bio, i) {
618 int len = bvl->bv_len;
622 if (page_offset < 0) {
623 b_offset = -page_offset;
624 page_offset += b_offset;
628 if (len > 0 && page_offset + len > STRIPE_SIZE)
629 clen = STRIPE_SIZE - page_offset;
634 b_offset += bvl->bv_offset;
635 bio_page = bvl->bv_page;
637 tx = async_memcpy(page, bio_page, page_offset,
638 b_offset, clen, &submit);
640 tx = async_memcpy(bio_page, page, b_offset,
641 page_offset, clen, &submit);
643 /* chain the operations */
644 submit.depend_tx = tx;
646 if (clen < len) /* hit end of page */
654 static void ops_complete_biofill(void *stripe_head_ref)
656 struct stripe_head *sh = stripe_head_ref;
657 struct bio *return_bi = NULL;
658 struct r5conf *conf = sh->raid_conf;
661 pr_debug("%s: stripe %llu\n", __func__,
662 (unsigned long long)sh->sector);
664 /* clear completed biofills */
665 spin_lock_irq(&conf->device_lock);
666 for (i = sh->disks; i--; ) {
667 struct r5dev *dev = &sh->dev[i];
669 /* acknowledge completion of a biofill operation */
670 /* and check if we need to reply to a read request,
671 * new R5_Wantfill requests are held off until
672 * !STRIPE_BIOFILL_RUN
674 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
675 struct bio *rbi, *rbi2;
680 while (rbi && rbi->bi_sector <
681 dev->sector + STRIPE_SECTORS) {
682 rbi2 = r5_next_bio(rbi, dev->sector);
683 if (!raid5_dec_bi_phys_segments(rbi)) {
684 rbi->bi_next = return_bi;
691 spin_unlock_irq(&conf->device_lock);
692 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
694 return_io(return_bi);
696 set_bit(STRIPE_HANDLE, &sh->state);
700 static void ops_run_biofill(struct stripe_head *sh)
702 struct dma_async_tx_descriptor *tx = NULL;
703 struct r5conf *conf = sh->raid_conf;
704 struct async_submit_ctl submit;
707 pr_debug("%s: stripe %llu\n", __func__,
708 (unsigned long long)sh->sector);
710 for (i = sh->disks; i--; ) {
711 struct r5dev *dev = &sh->dev[i];
712 if (test_bit(R5_Wantfill, &dev->flags)) {
714 spin_lock_irq(&conf->device_lock);
715 dev->read = rbi = dev->toread;
717 spin_unlock_irq(&conf->device_lock);
718 while (rbi && rbi->bi_sector <
719 dev->sector + STRIPE_SECTORS) {
720 tx = async_copy_data(0, rbi, dev->page,
722 rbi = r5_next_bio(rbi, dev->sector);
727 atomic_inc(&sh->count);
728 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
729 async_trigger_callback(&submit);
732 static void mark_target_uptodate(struct stripe_head *sh, int target)
739 tgt = &sh->dev[target];
740 set_bit(R5_UPTODATE, &tgt->flags);
741 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
742 clear_bit(R5_Wantcompute, &tgt->flags);
745 static void ops_complete_compute(void *stripe_head_ref)
747 struct stripe_head *sh = stripe_head_ref;
749 pr_debug("%s: stripe %llu\n", __func__,
750 (unsigned long long)sh->sector);
752 /* mark the computed target(s) as uptodate */
753 mark_target_uptodate(sh, sh->ops.target);
754 mark_target_uptodate(sh, sh->ops.target2);
756 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
757 if (sh->check_state == check_state_compute_run)
758 sh->check_state = check_state_compute_result;
759 set_bit(STRIPE_HANDLE, &sh->state);
763 /* return a pointer to the address conversion region of the scribble buffer */
764 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
765 struct raid5_percpu *percpu)
767 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
770 static struct dma_async_tx_descriptor *
771 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
773 int disks = sh->disks;
774 struct page **xor_srcs = percpu->scribble;
775 int target = sh->ops.target;
776 struct r5dev *tgt = &sh->dev[target];
777 struct page *xor_dest = tgt->page;
779 struct dma_async_tx_descriptor *tx;
780 struct async_submit_ctl submit;
783 pr_debug("%s: stripe %llu block: %d\n",
784 __func__, (unsigned long long)sh->sector, target);
785 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
787 for (i = disks; i--; )
789 xor_srcs[count++] = sh->dev[i].page;
791 atomic_inc(&sh->count);
793 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
794 ops_complete_compute, sh, to_addr_conv(sh, percpu));
795 if (unlikely(count == 1))
796 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
798 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
803 /* set_syndrome_sources - populate source buffers for gen_syndrome
804 * @srcs - (struct page *) array of size sh->disks
805 * @sh - stripe_head to parse
807 * Populates srcs in proper layout order for the stripe and returns the
808 * 'count' of sources to be used in a call to async_gen_syndrome. The P
809 * destination buffer is recorded in srcs[count] and the Q destination
810 * is recorded in srcs[count+1]].
812 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
814 int disks = sh->disks;
815 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
816 int d0_idx = raid6_d0(sh);
820 for (i = 0; i < disks; i++)
826 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
828 srcs[slot] = sh->dev[i].page;
829 i = raid6_next_disk(i, disks);
830 } while (i != d0_idx);
832 return syndrome_disks;
835 static struct dma_async_tx_descriptor *
836 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
838 int disks = sh->disks;
839 struct page **blocks = percpu->scribble;
841 int qd_idx = sh->qd_idx;
842 struct dma_async_tx_descriptor *tx;
843 struct async_submit_ctl submit;
849 if (sh->ops.target < 0)
850 target = sh->ops.target2;
851 else if (sh->ops.target2 < 0)
852 target = sh->ops.target;
854 /* we should only have one valid target */
857 pr_debug("%s: stripe %llu block: %d\n",
858 __func__, (unsigned long long)sh->sector, target);
860 tgt = &sh->dev[target];
861 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
864 atomic_inc(&sh->count);
866 if (target == qd_idx) {
867 count = set_syndrome_sources(blocks, sh);
868 blocks[count] = NULL; /* regenerating p is not necessary */
869 BUG_ON(blocks[count+1] != dest); /* q should already be set */
870 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
871 ops_complete_compute, sh,
872 to_addr_conv(sh, percpu));
873 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
875 /* Compute any data- or p-drive using XOR */
877 for (i = disks; i-- ; ) {
878 if (i == target || i == qd_idx)
880 blocks[count++] = sh->dev[i].page;
883 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
884 NULL, ops_complete_compute, sh,
885 to_addr_conv(sh, percpu));
886 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
892 static struct dma_async_tx_descriptor *
893 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
895 int i, count, disks = sh->disks;
896 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
897 int d0_idx = raid6_d0(sh);
898 int faila = -1, failb = -1;
899 int target = sh->ops.target;
900 int target2 = sh->ops.target2;
901 struct r5dev *tgt = &sh->dev[target];
902 struct r5dev *tgt2 = &sh->dev[target2];
903 struct dma_async_tx_descriptor *tx;
904 struct page **blocks = percpu->scribble;
905 struct async_submit_ctl submit;
907 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
908 __func__, (unsigned long long)sh->sector, target, target2);
909 BUG_ON(target < 0 || target2 < 0);
910 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
911 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
913 /* we need to open-code set_syndrome_sources to handle the
914 * slot number conversion for 'faila' and 'failb'
916 for (i = 0; i < disks ; i++)
921 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
923 blocks[slot] = sh->dev[i].page;
929 i = raid6_next_disk(i, disks);
930 } while (i != d0_idx);
932 BUG_ON(faila == failb);
935 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
936 __func__, (unsigned long long)sh->sector, faila, failb);
938 atomic_inc(&sh->count);
940 if (failb == syndrome_disks+1) {
941 /* Q disk is one of the missing disks */
942 if (faila == syndrome_disks) {
943 /* Missing P+Q, just recompute */
944 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
945 ops_complete_compute, sh,
946 to_addr_conv(sh, percpu));
947 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
948 STRIPE_SIZE, &submit);
952 int qd_idx = sh->qd_idx;
954 /* Missing D+Q: recompute D from P, then recompute Q */
955 if (target == qd_idx)
956 data_target = target2;
958 data_target = target;
961 for (i = disks; i-- ; ) {
962 if (i == data_target || i == qd_idx)
964 blocks[count++] = sh->dev[i].page;
966 dest = sh->dev[data_target].page;
967 init_async_submit(&submit,
968 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
970 to_addr_conv(sh, percpu));
971 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
974 count = set_syndrome_sources(blocks, sh);
975 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
976 ops_complete_compute, sh,
977 to_addr_conv(sh, percpu));
978 return async_gen_syndrome(blocks, 0, count+2,
979 STRIPE_SIZE, &submit);
982 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
983 ops_complete_compute, sh,
984 to_addr_conv(sh, percpu));
985 if (failb == syndrome_disks) {
986 /* We're missing D+P. */
987 return async_raid6_datap_recov(syndrome_disks+2,
991 /* We're missing D+D. */
992 return async_raid6_2data_recov(syndrome_disks+2,
993 STRIPE_SIZE, faila, failb,
1000 static void ops_complete_prexor(void *stripe_head_ref)
1002 struct stripe_head *sh = stripe_head_ref;
1004 pr_debug("%s: stripe %llu\n", __func__,
1005 (unsigned long long)sh->sector);
1008 static struct dma_async_tx_descriptor *
1009 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1010 struct dma_async_tx_descriptor *tx)
1012 int disks = sh->disks;
1013 struct page **xor_srcs = percpu->scribble;
1014 int count = 0, pd_idx = sh->pd_idx, i;
1015 struct async_submit_ctl submit;
1017 /* existing parity data subtracted */
1018 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1020 pr_debug("%s: stripe %llu\n", __func__,
1021 (unsigned long long)sh->sector);
1023 for (i = disks; i--; ) {
1024 struct r5dev *dev = &sh->dev[i];
1025 /* Only process blocks that are known to be uptodate */
1026 if (test_bit(R5_Wantdrain, &dev->flags))
1027 xor_srcs[count++] = dev->page;
1030 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1031 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1032 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1037 static struct dma_async_tx_descriptor *
1038 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1040 int disks = sh->disks;
1043 pr_debug("%s: stripe %llu\n", __func__,
1044 (unsigned long long)sh->sector);
1046 for (i = disks; i--; ) {
1047 struct r5dev *dev = &sh->dev[i];
1050 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1053 spin_lock_irq(&sh->raid_conf->device_lock);
1054 chosen = dev->towrite;
1055 dev->towrite = NULL;
1056 BUG_ON(dev->written);
1057 wbi = dev->written = chosen;
1058 spin_unlock_irq(&sh->raid_conf->device_lock);
1060 while (wbi && wbi->bi_sector <
1061 dev->sector + STRIPE_SECTORS) {
1062 if (wbi->bi_rw & REQ_FUA)
1063 set_bit(R5_WantFUA, &dev->flags);
1064 tx = async_copy_data(1, wbi, dev->page,
1066 wbi = r5_next_bio(wbi, dev->sector);
1074 static void ops_complete_reconstruct(void *stripe_head_ref)
1076 struct stripe_head *sh = stripe_head_ref;
1077 int disks = sh->disks;
1078 int pd_idx = sh->pd_idx;
1079 int qd_idx = sh->qd_idx;
1083 pr_debug("%s: stripe %llu\n", __func__,
1084 (unsigned long long)sh->sector);
1086 for (i = disks; i--; )
1087 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1089 for (i = disks; i--; ) {
1090 struct r5dev *dev = &sh->dev[i];
1092 if (dev->written || i == pd_idx || i == qd_idx) {
1093 set_bit(R5_UPTODATE, &dev->flags);
1095 set_bit(R5_WantFUA, &dev->flags);
1099 if (sh->reconstruct_state == reconstruct_state_drain_run)
1100 sh->reconstruct_state = reconstruct_state_drain_result;
1101 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1102 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1104 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1105 sh->reconstruct_state = reconstruct_state_result;
1108 set_bit(STRIPE_HANDLE, &sh->state);
1113 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1114 struct dma_async_tx_descriptor *tx)
1116 int disks = sh->disks;
1117 struct page **xor_srcs = percpu->scribble;
1118 struct async_submit_ctl submit;
1119 int count = 0, pd_idx = sh->pd_idx, i;
1120 struct page *xor_dest;
1122 unsigned long flags;
1124 pr_debug("%s: stripe %llu\n", __func__,
1125 (unsigned long long)sh->sector);
1127 /* check if prexor is active which means only process blocks
1128 * that are part of a read-modify-write (written)
1130 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1132 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1133 for (i = disks; i--; ) {
1134 struct r5dev *dev = &sh->dev[i];
1136 xor_srcs[count++] = dev->page;
1139 xor_dest = sh->dev[pd_idx].page;
1140 for (i = disks; i--; ) {
1141 struct r5dev *dev = &sh->dev[i];
1143 xor_srcs[count++] = dev->page;
1147 /* 1/ if we prexor'd then the dest is reused as a source
1148 * 2/ if we did not prexor then we are redoing the parity
1149 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1150 * for the synchronous xor case
1152 flags = ASYNC_TX_ACK |
1153 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1155 atomic_inc(&sh->count);
1157 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1158 to_addr_conv(sh, percpu));
1159 if (unlikely(count == 1))
1160 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1162 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1166 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1167 struct dma_async_tx_descriptor *tx)
1169 struct async_submit_ctl submit;
1170 struct page **blocks = percpu->scribble;
1173 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1175 count = set_syndrome_sources(blocks, sh);
1177 atomic_inc(&sh->count);
1179 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1180 sh, to_addr_conv(sh, percpu));
1181 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1184 static void ops_complete_check(void *stripe_head_ref)
1186 struct stripe_head *sh = stripe_head_ref;
1188 pr_debug("%s: stripe %llu\n", __func__,
1189 (unsigned long long)sh->sector);
1191 sh->check_state = check_state_check_result;
1192 set_bit(STRIPE_HANDLE, &sh->state);
1196 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1198 int disks = sh->disks;
1199 int pd_idx = sh->pd_idx;
1200 int qd_idx = sh->qd_idx;
1201 struct page *xor_dest;
1202 struct page **xor_srcs = percpu->scribble;
1203 struct dma_async_tx_descriptor *tx;
1204 struct async_submit_ctl submit;
1208 pr_debug("%s: stripe %llu\n", __func__,
1209 (unsigned long long)sh->sector);
1212 xor_dest = sh->dev[pd_idx].page;
1213 xor_srcs[count++] = xor_dest;
1214 for (i = disks; i--; ) {
1215 if (i == pd_idx || i == qd_idx)
1217 xor_srcs[count++] = sh->dev[i].page;
1220 init_async_submit(&submit, 0, NULL, NULL, NULL,
1221 to_addr_conv(sh, percpu));
1222 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1223 &sh->ops.zero_sum_result, &submit);
1225 atomic_inc(&sh->count);
1226 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1227 tx = async_trigger_callback(&submit);
1230 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1232 struct page **srcs = percpu->scribble;
1233 struct async_submit_ctl submit;
1236 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1237 (unsigned long long)sh->sector, checkp);
1239 count = set_syndrome_sources(srcs, sh);
1243 atomic_inc(&sh->count);
1244 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1245 sh, to_addr_conv(sh, percpu));
1246 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1247 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1250 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1252 int overlap_clear = 0, i, disks = sh->disks;
1253 struct dma_async_tx_descriptor *tx = NULL;
1254 struct r5conf *conf = sh->raid_conf;
1255 int level = conf->level;
1256 struct raid5_percpu *percpu;
1260 percpu = per_cpu_ptr(conf->percpu, cpu);
1261 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1262 ops_run_biofill(sh);
1266 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1268 tx = ops_run_compute5(sh, percpu);
1270 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1271 tx = ops_run_compute6_1(sh, percpu);
1273 tx = ops_run_compute6_2(sh, percpu);
1275 /* terminate the chain if reconstruct is not set to be run */
1276 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1280 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1281 tx = ops_run_prexor(sh, percpu, tx);
1283 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1284 tx = ops_run_biodrain(sh, tx);
1288 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1290 ops_run_reconstruct5(sh, percpu, tx);
1292 ops_run_reconstruct6(sh, percpu, tx);
1295 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1296 if (sh->check_state == check_state_run)
1297 ops_run_check_p(sh, percpu);
1298 else if (sh->check_state == check_state_run_q)
1299 ops_run_check_pq(sh, percpu, 0);
1300 else if (sh->check_state == check_state_run_pq)
1301 ops_run_check_pq(sh, percpu, 1);
1307 for (i = disks; i--; ) {
1308 struct r5dev *dev = &sh->dev[i];
1309 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1310 wake_up(&sh->raid_conf->wait_for_overlap);
1315 #ifdef CONFIG_MULTICORE_RAID456
1316 static void async_run_ops(void *param, async_cookie_t cookie)
1318 struct stripe_head *sh = param;
1319 unsigned long ops_request = sh->ops.request;
1321 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1322 wake_up(&sh->ops.wait_for_ops);
1324 __raid_run_ops(sh, ops_request);
1328 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1330 /* since handle_stripe can be called outside of raid5d context
1331 * we need to ensure sh->ops.request is de-staged before another
1334 wait_event(sh->ops.wait_for_ops,
1335 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1336 sh->ops.request = ops_request;
1338 atomic_inc(&sh->count);
1339 async_schedule(async_run_ops, sh);
1342 #define raid_run_ops __raid_run_ops
1345 static int grow_one_stripe(struct r5conf *conf)
1347 struct stripe_head *sh;
1348 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1352 sh->raid_conf = conf;
1353 #ifdef CONFIG_MULTICORE_RAID456
1354 init_waitqueue_head(&sh->ops.wait_for_ops);
1357 if (grow_buffers(sh)) {
1359 kmem_cache_free(conf->slab_cache, sh);
1362 /* we just created an active stripe so... */
1363 atomic_set(&sh->count, 1);
1364 atomic_inc(&conf->active_stripes);
1365 INIT_LIST_HEAD(&sh->lru);
1370 static int grow_stripes(struct r5conf *conf, int num)
1372 struct kmem_cache *sc;
1373 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1375 if (conf->mddev->gendisk)
1376 sprintf(conf->cache_name[0],
1377 "raid%d-%s", conf->level, mdname(conf->mddev));
1379 sprintf(conf->cache_name[0],
1380 "raid%d-%p", conf->level, conf->mddev);
1381 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1383 conf->active_name = 0;
1384 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1385 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1389 conf->slab_cache = sc;
1390 conf->pool_size = devs;
1392 if (!grow_one_stripe(conf))
1398 * scribble_len - return the required size of the scribble region
1399 * @num - total number of disks in the array
1401 * The size must be enough to contain:
1402 * 1/ a struct page pointer for each device in the array +2
1403 * 2/ room to convert each entry in (1) to its corresponding dma
1404 * (dma_map_page()) or page (page_address()) address.
1406 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1407 * calculate over all devices (not just the data blocks), using zeros in place
1408 * of the P and Q blocks.
1410 static size_t scribble_len(int num)
1414 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1419 static int resize_stripes(struct r5conf *conf, int newsize)
1421 /* Make all the stripes able to hold 'newsize' devices.
1422 * New slots in each stripe get 'page' set to a new page.
1424 * This happens in stages:
1425 * 1/ create a new kmem_cache and allocate the required number of
1427 * 2/ gather all the old stripe_heads and tranfer the pages across
1428 * to the new stripe_heads. This will have the side effect of
1429 * freezing the array as once all stripe_heads have been collected,
1430 * no IO will be possible. Old stripe heads are freed once their
1431 * pages have been transferred over, and the old kmem_cache is
1432 * freed when all stripes are done.
1433 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1434 * we simple return a failre status - no need to clean anything up.
1435 * 4/ allocate new pages for the new slots in the new stripe_heads.
1436 * If this fails, we don't bother trying the shrink the
1437 * stripe_heads down again, we just leave them as they are.
1438 * As each stripe_head is processed the new one is released into
1441 * Once step2 is started, we cannot afford to wait for a write,
1442 * so we use GFP_NOIO allocations.
1444 struct stripe_head *osh, *nsh;
1445 LIST_HEAD(newstripes);
1446 struct disk_info *ndisks;
1449 struct kmem_cache *sc;
1452 if (newsize <= conf->pool_size)
1453 return 0; /* never bother to shrink */
1455 err = md_allow_write(conf->mddev);
1460 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1461 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1466 for (i = conf->max_nr_stripes; i; i--) {
1467 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1471 nsh->raid_conf = conf;
1472 #ifdef CONFIG_MULTICORE_RAID456
1473 init_waitqueue_head(&nsh->ops.wait_for_ops);
1476 list_add(&nsh->lru, &newstripes);
1479 /* didn't get enough, give up */
1480 while (!list_empty(&newstripes)) {
1481 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1482 list_del(&nsh->lru);
1483 kmem_cache_free(sc, nsh);
1485 kmem_cache_destroy(sc);
1488 /* Step 2 - Must use GFP_NOIO now.
1489 * OK, we have enough stripes, start collecting inactive
1490 * stripes and copying them over
1492 list_for_each_entry(nsh, &newstripes, lru) {
1493 spin_lock_irq(&conf->device_lock);
1494 wait_event_lock_irq(conf->wait_for_stripe,
1495 !list_empty(&conf->inactive_list),
1498 osh = get_free_stripe(conf);
1499 spin_unlock_irq(&conf->device_lock);
1500 atomic_set(&nsh->count, 1);
1501 for(i=0; i<conf->pool_size; i++)
1502 nsh->dev[i].page = osh->dev[i].page;
1503 for( ; i<newsize; i++)
1504 nsh->dev[i].page = NULL;
1505 kmem_cache_free(conf->slab_cache, osh);
1507 kmem_cache_destroy(conf->slab_cache);
1510 * At this point, we are holding all the stripes so the array
1511 * is completely stalled, so now is a good time to resize
1512 * conf->disks and the scribble region
1514 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1516 for (i=0; i<conf->raid_disks; i++)
1517 ndisks[i] = conf->disks[i];
1519 conf->disks = ndisks;
1524 conf->scribble_len = scribble_len(newsize);
1525 for_each_present_cpu(cpu) {
1526 struct raid5_percpu *percpu;
1529 percpu = per_cpu_ptr(conf->percpu, cpu);
1530 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1533 kfree(percpu->scribble);
1534 percpu->scribble = scribble;
1542 /* Step 4, return new stripes to service */
1543 while(!list_empty(&newstripes)) {
1544 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1545 list_del_init(&nsh->lru);
1547 for (i=conf->raid_disks; i < newsize; i++)
1548 if (nsh->dev[i].page == NULL) {
1549 struct page *p = alloc_page(GFP_NOIO);
1550 nsh->dev[i].page = p;
1554 release_stripe(nsh);
1556 /* critical section pass, GFP_NOIO no longer needed */
1558 conf->slab_cache = sc;
1559 conf->active_name = 1-conf->active_name;
1560 conf->pool_size = newsize;
1564 static int drop_one_stripe(struct r5conf *conf)
1566 struct stripe_head *sh;
1568 spin_lock_irq(&conf->device_lock);
1569 sh = get_free_stripe(conf);
1570 spin_unlock_irq(&conf->device_lock);
1573 BUG_ON(atomic_read(&sh->count));
1575 kmem_cache_free(conf->slab_cache, sh);
1576 atomic_dec(&conf->active_stripes);
1580 static void shrink_stripes(struct r5conf *conf)
1582 while (drop_one_stripe(conf))
1585 if (conf->slab_cache)
1586 kmem_cache_destroy(conf->slab_cache);
1587 conf->slab_cache = NULL;
1590 static void raid5_end_read_request(struct bio * bi, int error)
1592 struct stripe_head *sh = bi->bi_private;
1593 struct r5conf *conf = sh->raid_conf;
1594 int disks = sh->disks, i;
1595 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1596 char b[BDEVNAME_SIZE];
1597 struct md_rdev *rdev;
1600 for (i=0 ; i<disks; i++)
1601 if (bi == &sh->dev[i].req)
1604 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1605 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1613 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1614 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1615 rdev = conf->disks[i].rdev;
1618 "md/raid:%s: read error corrected"
1619 " (%lu sectors at %llu on %s)\n",
1620 mdname(conf->mddev), STRIPE_SECTORS,
1621 (unsigned long long)(sh->sector
1622 + rdev->data_offset),
1623 bdevname(rdev->bdev, b));
1624 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1625 clear_bit(R5_ReadError, &sh->dev[i].flags);
1626 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1628 if (atomic_read(&conf->disks[i].rdev->read_errors))
1629 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1631 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1633 rdev = conf->disks[i].rdev;
1635 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1636 atomic_inc(&rdev->read_errors);
1637 if (conf->mddev->degraded >= conf->max_degraded)
1640 "md/raid:%s: read error not correctable "
1641 "(sector %llu on %s).\n",
1642 mdname(conf->mddev),
1643 (unsigned long long)(sh->sector
1644 + rdev->data_offset),
1646 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1650 "md/raid:%s: read error NOT corrected!! "
1651 "(sector %llu on %s).\n",
1652 mdname(conf->mddev),
1653 (unsigned long long)(sh->sector
1654 + rdev->data_offset),
1656 else if (atomic_read(&rdev->read_errors)
1657 > conf->max_nr_stripes)
1659 "md/raid:%s: Too many read errors, failing device %s.\n",
1660 mdname(conf->mddev), bdn);
1664 set_bit(R5_ReadError, &sh->dev[i].flags);
1666 clear_bit(R5_ReadError, &sh->dev[i].flags);
1667 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1668 md_error(conf->mddev, rdev);
1671 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1672 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1673 set_bit(STRIPE_HANDLE, &sh->state);
1677 static void raid5_end_write_request(struct bio *bi, int error)
1679 struct stripe_head *sh = bi->bi_private;
1680 struct r5conf *conf = sh->raid_conf;
1681 int disks = sh->disks, i;
1682 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1686 for (i=0 ; i<disks; i++)
1687 if (bi == &sh->dev[i].req)
1690 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1691 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1699 set_bit(WriteErrorSeen, &conf->disks[i].rdev->flags);
1700 set_bit(R5_WriteError, &sh->dev[i].flags);
1701 } else if (is_badblock(conf->disks[i].rdev, sh->sector, STRIPE_SECTORS,
1702 &first_bad, &bad_sectors))
1703 set_bit(R5_MadeGood, &sh->dev[i].flags);
1705 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1707 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1708 set_bit(STRIPE_HANDLE, &sh->state);
1713 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1715 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1717 struct r5dev *dev = &sh->dev[i];
1719 bio_init(&dev->req);
1720 dev->req.bi_io_vec = &dev->vec;
1722 dev->req.bi_max_vecs++;
1723 dev->vec.bv_page = dev->page;
1724 dev->vec.bv_len = STRIPE_SIZE;
1725 dev->vec.bv_offset = 0;
1727 dev->req.bi_sector = sh->sector;
1728 dev->req.bi_private = sh;
1731 dev->sector = compute_blocknr(sh, i, previous);
1734 static void error(struct mddev *mddev, struct md_rdev *rdev)
1736 char b[BDEVNAME_SIZE];
1737 struct r5conf *conf = mddev->private;
1738 unsigned long flags;
1739 pr_debug("raid456: error called\n");
1741 spin_lock_irqsave(&conf->device_lock, flags);
1742 clear_bit(In_sync, &rdev->flags);
1743 mddev->degraded = calc_degraded(conf);
1744 spin_unlock_irqrestore(&conf->device_lock, flags);
1745 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1747 set_bit(Blocked, &rdev->flags);
1748 set_bit(Faulty, &rdev->flags);
1749 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1751 "md/raid:%s: Disk failure on %s, disabling device.\n"
1752 "md/raid:%s: Operation continuing on %d devices.\n",
1754 bdevname(rdev->bdev, b),
1756 conf->raid_disks - mddev->degraded);
1760 * Input: a 'big' sector number,
1761 * Output: index of the data and parity disk, and the sector # in them.
1763 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1764 int previous, int *dd_idx,
1765 struct stripe_head *sh)
1767 sector_t stripe, stripe2;
1768 sector_t chunk_number;
1769 unsigned int chunk_offset;
1772 sector_t new_sector;
1773 int algorithm = previous ? conf->prev_algo
1775 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1776 : conf->chunk_sectors;
1777 int raid_disks = previous ? conf->previous_raid_disks
1779 int data_disks = raid_disks - conf->max_degraded;
1781 /* First compute the information on this sector */
1784 * Compute the chunk number and the sector offset inside the chunk
1786 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1787 chunk_number = r_sector;
1790 * Compute the stripe number
1792 stripe = chunk_number;
1793 *dd_idx = sector_div(stripe, data_disks);
1796 * Select the parity disk based on the user selected algorithm.
1798 pd_idx = qd_idx = -1;
1799 switch(conf->level) {
1801 pd_idx = data_disks;
1804 switch (algorithm) {
1805 case ALGORITHM_LEFT_ASYMMETRIC:
1806 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1807 if (*dd_idx >= pd_idx)
1810 case ALGORITHM_RIGHT_ASYMMETRIC:
1811 pd_idx = sector_div(stripe2, raid_disks);
1812 if (*dd_idx >= pd_idx)
1815 case ALGORITHM_LEFT_SYMMETRIC:
1816 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1817 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1819 case ALGORITHM_RIGHT_SYMMETRIC:
1820 pd_idx = sector_div(stripe2, raid_disks);
1821 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1823 case ALGORITHM_PARITY_0:
1827 case ALGORITHM_PARITY_N:
1828 pd_idx = data_disks;
1836 switch (algorithm) {
1837 case ALGORITHM_LEFT_ASYMMETRIC:
1838 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1839 qd_idx = pd_idx + 1;
1840 if (pd_idx == raid_disks-1) {
1841 (*dd_idx)++; /* Q D D D P */
1843 } else if (*dd_idx >= pd_idx)
1844 (*dd_idx) += 2; /* D D P Q D */
1846 case ALGORITHM_RIGHT_ASYMMETRIC:
1847 pd_idx = sector_div(stripe2, raid_disks);
1848 qd_idx = pd_idx + 1;
1849 if (pd_idx == raid_disks-1) {
1850 (*dd_idx)++; /* Q D D D P */
1852 } else if (*dd_idx >= pd_idx)
1853 (*dd_idx) += 2; /* D D P Q D */
1855 case ALGORITHM_LEFT_SYMMETRIC:
1856 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1857 qd_idx = (pd_idx + 1) % raid_disks;
1858 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1860 case ALGORITHM_RIGHT_SYMMETRIC:
1861 pd_idx = sector_div(stripe2, raid_disks);
1862 qd_idx = (pd_idx + 1) % raid_disks;
1863 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1866 case ALGORITHM_PARITY_0:
1871 case ALGORITHM_PARITY_N:
1872 pd_idx = data_disks;
1873 qd_idx = data_disks + 1;
1876 case ALGORITHM_ROTATING_ZERO_RESTART:
1877 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1878 * of blocks for computing Q is different.
1880 pd_idx = sector_div(stripe2, raid_disks);
1881 qd_idx = pd_idx + 1;
1882 if (pd_idx == raid_disks-1) {
1883 (*dd_idx)++; /* Q D D D P */
1885 } else if (*dd_idx >= pd_idx)
1886 (*dd_idx) += 2; /* D D P Q D */
1890 case ALGORITHM_ROTATING_N_RESTART:
1891 /* Same a left_asymmetric, by first stripe is
1892 * D D D P Q rather than
1896 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1897 qd_idx = pd_idx + 1;
1898 if (pd_idx == raid_disks-1) {
1899 (*dd_idx)++; /* Q D D D P */
1901 } else if (*dd_idx >= pd_idx)
1902 (*dd_idx) += 2; /* D D P Q D */
1906 case ALGORITHM_ROTATING_N_CONTINUE:
1907 /* Same as left_symmetric but Q is before P */
1908 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1909 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1910 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1914 case ALGORITHM_LEFT_ASYMMETRIC_6:
1915 /* RAID5 left_asymmetric, with Q on last device */
1916 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1917 if (*dd_idx >= pd_idx)
1919 qd_idx = raid_disks - 1;
1922 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1923 pd_idx = sector_div(stripe2, raid_disks-1);
1924 if (*dd_idx >= pd_idx)
1926 qd_idx = raid_disks - 1;
1929 case ALGORITHM_LEFT_SYMMETRIC_6:
1930 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1931 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1932 qd_idx = raid_disks - 1;
1935 case ALGORITHM_RIGHT_SYMMETRIC_6:
1936 pd_idx = sector_div(stripe2, raid_disks-1);
1937 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1938 qd_idx = raid_disks - 1;
1941 case ALGORITHM_PARITY_0_6:
1944 qd_idx = raid_disks - 1;
1954 sh->pd_idx = pd_idx;
1955 sh->qd_idx = qd_idx;
1956 sh->ddf_layout = ddf_layout;
1959 * Finally, compute the new sector number
1961 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1966 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1968 struct r5conf *conf = sh->raid_conf;
1969 int raid_disks = sh->disks;
1970 int data_disks = raid_disks - conf->max_degraded;
1971 sector_t new_sector = sh->sector, check;
1972 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1973 : conf->chunk_sectors;
1974 int algorithm = previous ? conf->prev_algo
1978 sector_t chunk_number;
1979 int dummy1, dd_idx = i;
1981 struct stripe_head sh2;
1984 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1985 stripe = new_sector;
1987 if (i == sh->pd_idx)
1989 switch(conf->level) {
1992 switch (algorithm) {
1993 case ALGORITHM_LEFT_ASYMMETRIC:
1994 case ALGORITHM_RIGHT_ASYMMETRIC:
1998 case ALGORITHM_LEFT_SYMMETRIC:
1999 case ALGORITHM_RIGHT_SYMMETRIC:
2002 i -= (sh->pd_idx + 1);
2004 case ALGORITHM_PARITY_0:
2007 case ALGORITHM_PARITY_N:
2014 if (i == sh->qd_idx)
2015 return 0; /* It is the Q disk */
2016 switch (algorithm) {
2017 case ALGORITHM_LEFT_ASYMMETRIC:
2018 case ALGORITHM_RIGHT_ASYMMETRIC:
2019 case ALGORITHM_ROTATING_ZERO_RESTART:
2020 case ALGORITHM_ROTATING_N_RESTART:
2021 if (sh->pd_idx == raid_disks-1)
2022 i--; /* Q D D D P */
2023 else if (i > sh->pd_idx)
2024 i -= 2; /* D D P Q D */
2026 case ALGORITHM_LEFT_SYMMETRIC:
2027 case ALGORITHM_RIGHT_SYMMETRIC:
2028 if (sh->pd_idx == raid_disks-1)
2029 i--; /* Q D D D P */
2034 i -= (sh->pd_idx + 2);
2037 case ALGORITHM_PARITY_0:
2040 case ALGORITHM_PARITY_N:
2042 case ALGORITHM_ROTATING_N_CONTINUE:
2043 /* Like left_symmetric, but P is before Q */
2044 if (sh->pd_idx == 0)
2045 i--; /* P D D D Q */
2050 i -= (sh->pd_idx + 1);
2053 case ALGORITHM_LEFT_ASYMMETRIC_6:
2054 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2058 case ALGORITHM_LEFT_SYMMETRIC_6:
2059 case ALGORITHM_RIGHT_SYMMETRIC_6:
2061 i += data_disks + 1;
2062 i -= (sh->pd_idx + 1);
2064 case ALGORITHM_PARITY_0_6:
2073 chunk_number = stripe * data_disks + i;
2074 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2076 check = raid5_compute_sector(conf, r_sector,
2077 previous, &dummy1, &sh2);
2078 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2079 || sh2.qd_idx != sh->qd_idx) {
2080 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2081 mdname(conf->mddev));
2089 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2090 int rcw, int expand)
2092 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2093 struct r5conf *conf = sh->raid_conf;
2094 int level = conf->level;
2097 /* if we are not expanding this is a proper write request, and
2098 * there will be bios with new data to be drained into the
2102 sh->reconstruct_state = reconstruct_state_drain_run;
2103 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2105 sh->reconstruct_state = reconstruct_state_run;
2107 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2109 for (i = disks; i--; ) {
2110 struct r5dev *dev = &sh->dev[i];
2113 set_bit(R5_LOCKED, &dev->flags);
2114 set_bit(R5_Wantdrain, &dev->flags);
2116 clear_bit(R5_UPTODATE, &dev->flags);
2120 if (s->locked + conf->max_degraded == disks)
2121 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2122 atomic_inc(&conf->pending_full_writes);
2125 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2126 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2128 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2129 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2130 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2131 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2133 for (i = disks; i--; ) {
2134 struct r5dev *dev = &sh->dev[i];
2139 (test_bit(R5_UPTODATE, &dev->flags) ||
2140 test_bit(R5_Wantcompute, &dev->flags))) {
2141 set_bit(R5_Wantdrain, &dev->flags);
2142 set_bit(R5_LOCKED, &dev->flags);
2143 clear_bit(R5_UPTODATE, &dev->flags);
2149 /* keep the parity disk(s) locked while asynchronous operations
2152 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2153 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2157 int qd_idx = sh->qd_idx;
2158 struct r5dev *dev = &sh->dev[qd_idx];
2160 set_bit(R5_LOCKED, &dev->flags);
2161 clear_bit(R5_UPTODATE, &dev->flags);
2165 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2166 __func__, (unsigned long long)sh->sector,
2167 s->locked, s->ops_request);
2171 * Each stripe/dev can have one or more bion attached.
2172 * toread/towrite point to the first in a chain.
2173 * The bi_next chain must be in order.
2175 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2178 struct r5conf *conf = sh->raid_conf;
2181 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2182 (unsigned long long)bi->bi_sector,
2183 (unsigned long long)sh->sector);
2186 spin_lock_irq(&conf->device_lock);
2188 bip = &sh->dev[dd_idx].towrite;
2189 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2192 bip = &sh->dev[dd_idx].toread;
2193 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2194 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2196 bip = & (*bip)->bi_next;
2198 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2201 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2205 bi->bi_phys_segments++;
2208 /* check if page is covered */
2209 sector_t sector = sh->dev[dd_idx].sector;
2210 for (bi=sh->dev[dd_idx].towrite;
2211 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2212 bi && bi->bi_sector <= sector;
2213 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2214 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2215 sector = bi->bi_sector + (bi->bi_size>>9);
2217 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2218 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2220 spin_unlock_irq(&conf->device_lock);
2222 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2223 (unsigned long long)(*bip)->bi_sector,
2224 (unsigned long long)sh->sector, dd_idx);
2226 if (conf->mddev->bitmap && firstwrite) {
2227 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2229 sh->bm_seq = conf->seq_flush+1;
2230 set_bit(STRIPE_BIT_DELAY, &sh->state);
2235 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2236 spin_unlock_irq(&conf->device_lock);
2240 static void end_reshape(struct r5conf *conf);
2242 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2243 struct stripe_head *sh)
2245 int sectors_per_chunk =
2246 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2248 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2249 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2251 raid5_compute_sector(conf,
2252 stripe * (disks - conf->max_degraded)
2253 *sectors_per_chunk + chunk_offset,
2259 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2260 struct stripe_head_state *s, int disks,
2261 struct bio **return_bi)
2264 for (i = disks; i--; ) {
2268 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2269 struct md_rdev *rdev;
2271 rdev = rcu_dereference(conf->disks[i].rdev);
2272 if (rdev && test_bit(In_sync, &rdev->flags))
2273 atomic_inc(&rdev->nr_pending);
2278 if (!rdev_set_badblocks(
2282 md_error(conf->mddev, rdev);
2283 rdev_dec_pending(rdev, conf->mddev);
2286 spin_lock_irq(&conf->device_lock);
2287 /* fail all writes first */
2288 bi = sh->dev[i].towrite;
2289 sh->dev[i].towrite = NULL;
2295 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2296 wake_up(&conf->wait_for_overlap);
2298 while (bi && bi->bi_sector <
2299 sh->dev[i].sector + STRIPE_SECTORS) {
2300 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2301 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2302 if (!raid5_dec_bi_phys_segments(bi)) {
2303 md_write_end(conf->mddev);
2304 bi->bi_next = *return_bi;
2309 /* and fail all 'written' */
2310 bi = sh->dev[i].written;
2311 sh->dev[i].written = NULL;
2312 if (bi) bitmap_end = 1;
2313 while (bi && bi->bi_sector <
2314 sh->dev[i].sector + STRIPE_SECTORS) {
2315 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2316 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2317 if (!raid5_dec_bi_phys_segments(bi)) {
2318 md_write_end(conf->mddev);
2319 bi->bi_next = *return_bi;
2325 /* fail any reads if this device is non-operational and
2326 * the data has not reached the cache yet.
2328 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2329 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2330 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2331 bi = sh->dev[i].toread;
2332 sh->dev[i].toread = NULL;
2333 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2334 wake_up(&conf->wait_for_overlap);
2335 if (bi) s->to_read--;
2336 while (bi && bi->bi_sector <
2337 sh->dev[i].sector + STRIPE_SECTORS) {
2338 struct bio *nextbi =
2339 r5_next_bio(bi, sh->dev[i].sector);
2340 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2341 if (!raid5_dec_bi_phys_segments(bi)) {
2342 bi->bi_next = *return_bi;
2348 spin_unlock_irq(&conf->device_lock);
2350 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2351 STRIPE_SECTORS, 0, 0);
2352 /* If we were in the middle of a write the parity block might
2353 * still be locked - so just clear all R5_LOCKED flags
2355 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2358 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2359 if (atomic_dec_and_test(&conf->pending_full_writes))
2360 md_wakeup_thread(conf->mddev->thread);
2364 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2365 struct stripe_head_state *s)
2370 md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
2371 clear_bit(STRIPE_SYNCING, &sh->state);
2373 /* There is nothing more to do for sync/check/repair.
2374 * For recover we need to record a bad block on all
2375 * non-sync devices, or abort the recovery
2377 if (!test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery))
2379 /* During recovery devices cannot be removed, so locking and
2380 * refcounting of rdevs is not needed
2382 for (i = 0; i < conf->raid_disks; i++) {
2383 struct md_rdev *rdev = conf->disks[i].rdev;
2385 || test_bit(Faulty, &rdev->flags)
2386 || test_bit(In_sync, &rdev->flags))
2388 if (!rdev_set_badblocks(rdev, sh->sector,
2393 conf->recovery_disabled = conf->mddev->recovery_disabled;
2394 set_bit(MD_RECOVERY_INTR, &conf->mddev->recovery);
2398 /* fetch_block - checks the given member device to see if its data needs
2399 * to be read or computed to satisfy a request.
2401 * Returns 1 when no more member devices need to be checked, otherwise returns
2402 * 0 to tell the loop in handle_stripe_fill to continue
2404 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2405 int disk_idx, int disks)
2407 struct r5dev *dev = &sh->dev[disk_idx];
2408 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2409 &sh->dev[s->failed_num[1]] };
2411 /* is the data in this block needed, and can we get it? */
2412 if (!test_bit(R5_LOCKED, &dev->flags) &&
2413 !test_bit(R5_UPTODATE, &dev->flags) &&
2415 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2416 s->syncing || s->expanding ||
2417 (s->failed >= 1 && fdev[0]->toread) ||
2418 (s->failed >= 2 && fdev[1]->toread) ||
2419 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2420 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2421 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2422 /* we would like to get this block, possibly by computing it,
2423 * otherwise read it if the backing disk is insync
2425 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2426 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2427 if ((s->uptodate == disks - 1) &&
2428 (s->failed && (disk_idx == s->failed_num[0] ||
2429 disk_idx == s->failed_num[1]))) {
2430 /* have disk failed, and we're requested to fetch it;
2433 pr_debug("Computing stripe %llu block %d\n",
2434 (unsigned long long)sh->sector, disk_idx);
2435 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2436 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2437 set_bit(R5_Wantcompute, &dev->flags);
2438 sh->ops.target = disk_idx;
2439 sh->ops.target2 = -1; /* no 2nd target */
2441 /* Careful: from this point on 'uptodate' is in the eye
2442 * of raid_run_ops which services 'compute' operations
2443 * before writes. R5_Wantcompute flags a block that will
2444 * be R5_UPTODATE by the time it is needed for a
2445 * subsequent operation.
2449 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2450 /* Computing 2-failure is *very* expensive; only
2451 * do it if failed >= 2
2454 for (other = disks; other--; ) {
2455 if (other == disk_idx)
2457 if (!test_bit(R5_UPTODATE,
2458 &sh->dev[other].flags))
2462 pr_debug("Computing stripe %llu blocks %d,%d\n",
2463 (unsigned long long)sh->sector,
2465 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2466 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2467 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2468 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2469 sh->ops.target = disk_idx;
2470 sh->ops.target2 = other;
2474 } else if (test_bit(R5_Insync, &dev->flags)) {
2475 set_bit(R5_LOCKED, &dev->flags);
2476 set_bit(R5_Wantread, &dev->flags);
2478 pr_debug("Reading block %d (sync=%d)\n",
2479 disk_idx, s->syncing);
2487 * handle_stripe_fill - read or compute data to satisfy pending requests.
2489 static void handle_stripe_fill(struct stripe_head *sh,
2490 struct stripe_head_state *s,
2495 /* look for blocks to read/compute, skip this if a compute
2496 * is already in flight, or if the stripe contents are in the
2497 * midst of changing due to a write
2499 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2500 !sh->reconstruct_state)
2501 for (i = disks; i--; )
2502 if (fetch_block(sh, s, i, disks))
2504 set_bit(STRIPE_HANDLE, &sh->state);
2508 /* handle_stripe_clean_event
2509 * any written block on an uptodate or failed drive can be returned.
2510 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2511 * never LOCKED, so we don't need to test 'failed' directly.
2513 static void handle_stripe_clean_event(struct r5conf *conf,
2514 struct stripe_head *sh, int disks, struct bio **return_bi)
2519 for (i = disks; i--; )
2520 if (sh->dev[i].written) {
2522 if (!test_bit(R5_LOCKED, &dev->flags) &&
2523 test_bit(R5_UPTODATE, &dev->flags)) {
2524 /* We can return any write requests */
2525 struct bio *wbi, *wbi2;
2527 pr_debug("Return write for disc %d\n", i);
2528 spin_lock_irq(&conf->device_lock);
2530 dev->written = NULL;
2531 while (wbi && wbi->bi_sector <
2532 dev->sector + STRIPE_SECTORS) {
2533 wbi2 = r5_next_bio(wbi, dev->sector);
2534 if (!raid5_dec_bi_phys_segments(wbi)) {
2535 md_write_end(conf->mddev);
2536 wbi->bi_next = *return_bi;
2541 if (dev->towrite == NULL)
2543 spin_unlock_irq(&conf->device_lock);
2545 bitmap_endwrite(conf->mddev->bitmap,
2548 !test_bit(STRIPE_DEGRADED, &sh->state),
2553 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2554 if (atomic_dec_and_test(&conf->pending_full_writes))
2555 md_wakeup_thread(conf->mddev->thread);
2558 static void handle_stripe_dirtying(struct r5conf *conf,
2559 struct stripe_head *sh,
2560 struct stripe_head_state *s,
2563 int rmw = 0, rcw = 0, i;
2564 if (conf->max_degraded == 2) {
2565 /* RAID6 requires 'rcw' in current implementation
2566 * Calculate the real rcw later - for now fake it
2567 * look like rcw is cheaper
2570 } else for (i = disks; i--; ) {
2571 /* would I have to read this buffer for read_modify_write */
2572 struct r5dev *dev = &sh->dev[i];
2573 if ((dev->towrite || i == sh->pd_idx) &&
2574 !test_bit(R5_LOCKED, &dev->flags) &&
2575 !(test_bit(R5_UPTODATE, &dev->flags) ||
2576 test_bit(R5_Wantcompute, &dev->flags))) {
2577 if (test_bit(R5_Insync, &dev->flags))
2580 rmw += 2*disks; /* cannot read it */
2582 /* Would I have to read this buffer for reconstruct_write */
2583 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2584 !test_bit(R5_LOCKED, &dev->flags) &&
2585 !(test_bit(R5_UPTODATE, &dev->flags) ||
2586 test_bit(R5_Wantcompute, &dev->flags))) {
2587 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2592 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2593 (unsigned long long)sh->sector, rmw, rcw);
2594 set_bit(STRIPE_HANDLE, &sh->state);
2595 if (rmw < rcw && rmw > 0)
2596 /* prefer read-modify-write, but need to get some data */
2597 for (i = disks; i--; ) {
2598 struct r5dev *dev = &sh->dev[i];
2599 if ((dev->towrite || i == sh->pd_idx) &&
2600 !test_bit(R5_LOCKED, &dev->flags) &&
2601 !(test_bit(R5_UPTODATE, &dev->flags) ||
2602 test_bit(R5_Wantcompute, &dev->flags)) &&
2603 test_bit(R5_Insync, &dev->flags)) {
2605 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2606 pr_debug("Read_old block "
2607 "%d for r-m-w\n", i);
2608 set_bit(R5_LOCKED, &dev->flags);
2609 set_bit(R5_Wantread, &dev->flags);
2612 set_bit(STRIPE_DELAYED, &sh->state);
2613 set_bit(STRIPE_HANDLE, &sh->state);
2617 if (rcw <= rmw && rcw > 0) {
2618 /* want reconstruct write, but need to get some data */
2620 for (i = disks; i--; ) {
2621 struct r5dev *dev = &sh->dev[i];
2622 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2623 i != sh->pd_idx && i != sh->qd_idx &&
2624 !test_bit(R5_LOCKED, &dev->flags) &&
2625 !(test_bit(R5_UPTODATE, &dev->flags) ||
2626 test_bit(R5_Wantcompute, &dev->flags))) {
2628 if (!test_bit(R5_Insync, &dev->flags))
2629 continue; /* it's a failed drive */
2631 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2632 pr_debug("Read_old block "
2633 "%d for Reconstruct\n", i);
2634 set_bit(R5_LOCKED, &dev->flags);
2635 set_bit(R5_Wantread, &dev->flags);
2638 set_bit(STRIPE_DELAYED, &sh->state);
2639 set_bit(STRIPE_HANDLE, &sh->state);
2644 /* now if nothing is locked, and if we have enough data,
2645 * we can start a write request
2647 /* since handle_stripe can be called at any time we need to handle the
2648 * case where a compute block operation has been submitted and then a
2649 * subsequent call wants to start a write request. raid_run_ops only
2650 * handles the case where compute block and reconstruct are requested
2651 * simultaneously. If this is not the case then new writes need to be
2652 * held off until the compute completes.
2654 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2655 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2656 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2657 schedule_reconstruction(sh, s, rcw == 0, 0);
2660 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2661 struct stripe_head_state *s, int disks)
2663 struct r5dev *dev = NULL;
2665 set_bit(STRIPE_HANDLE, &sh->state);
2667 switch (sh->check_state) {
2668 case check_state_idle:
2669 /* start a new check operation if there are no failures */
2670 if (s->failed == 0) {
2671 BUG_ON(s->uptodate != disks);
2672 sh->check_state = check_state_run;
2673 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2674 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2678 dev = &sh->dev[s->failed_num[0]];
2680 case check_state_compute_result:
2681 sh->check_state = check_state_idle;
2683 dev = &sh->dev[sh->pd_idx];
2685 /* check that a write has not made the stripe insync */
2686 if (test_bit(STRIPE_INSYNC, &sh->state))
2689 /* either failed parity check, or recovery is happening */
2690 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2691 BUG_ON(s->uptodate != disks);
2693 set_bit(R5_LOCKED, &dev->flags);
2695 set_bit(R5_Wantwrite, &dev->flags);
2697 clear_bit(STRIPE_DEGRADED, &sh->state);
2698 set_bit(STRIPE_INSYNC, &sh->state);
2700 case check_state_run:
2701 break; /* we will be called again upon completion */
2702 case check_state_check_result:
2703 sh->check_state = check_state_idle;
2705 /* if a failure occurred during the check operation, leave
2706 * STRIPE_INSYNC not set and let the stripe be handled again
2711 /* handle a successful check operation, if parity is correct
2712 * we are done. Otherwise update the mismatch count and repair
2713 * parity if !MD_RECOVERY_CHECK
2715 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2716 /* parity is correct (on disc,
2717 * not in buffer any more)
2719 set_bit(STRIPE_INSYNC, &sh->state);
2721 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2722 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2723 /* don't try to repair!! */
2724 set_bit(STRIPE_INSYNC, &sh->state);
2726 sh->check_state = check_state_compute_run;
2727 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2728 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2729 set_bit(R5_Wantcompute,
2730 &sh->dev[sh->pd_idx].flags);
2731 sh->ops.target = sh->pd_idx;
2732 sh->ops.target2 = -1;
2737 case check_state_compute_run:
2740 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2741 __func__, sh->check_state,
2742 (unsigned long long) sh->sector);
2748 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2749 struct stripe_head_state *s,
2752 int pd_idx = sh->pd_idx;
2753 int qd_idx = sh->qd_idx;
2756 set_bit(STRIPE_HANDLE, &sh->state);
2758 BUG_ON(s->failed > 2);
2760 /* Want to check and possibly repair P and Q.
2761 * However there could be one 'failed' device, in which
2762 * case we can only check one of them, possibly using the
2763 * other to generate missing data
2766 switch (sh->check_state) {
2767 case check_state_idle:
2768 /* start a new check operation if there are < 2 failures */
2769 if (s->failed == s->q_failed) {
2770 /* The only possible failed device holds Q, so it
2771 * makes sense to check P (If anything else were failed,
2772 * we would have used P to recreate it).
2774 sh->check_state = check_state_run;
2776 if (!s->q_failed && s->failed < 2) {
2777 /* Q is not failed, and we didn't use it to generate
2778 * anything, so it makes sense to check it
2780 if (sh->check_state == check_state_run)
2781 sh->check_state = check_state_run_pq;
2783 sh->check_state = check_state_run_q;
2786 /* discard potentially stale zero_sum_result */
2787 sh->ops.zero_sum_result = 0;
2789 if (sh->check_state == check_state_run) {
2790 /* async_xor_zero_sum destroys the contents of P */
2791 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2794 if (sh->check_state >= check_state_run &&
2795 sh->check_state <= check_state_run_pq) {
2796 /* async_syndrome_zero_sum preserves P and Q, so
2797 * no need to mark them !uptodate here
2799 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2803 /* we have 2-disk failure */
2804 BUG_ON(s->failed != 2);
2806 case check_state_compute_result:
2807 sh->check_state = check_state_idle;
2809 /* check that a write has not made the stripe insync */
2810 if (test_bit(STRIPE_INSYNC, &sh->state))
2813 /* now write out any block on a failed drive,
2814 * or P or Q if they were recomputed
2816 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2817 if (s->failed == 2) {
2818 dev = &sh->dev[s->failed_num[1]];
2820 set_bit(R5_LOCKED, &dev->flags);
2821 set_bit(R5_Wantwrite, &dev->flags);
2823 if (s->failed >= 1) {
2824 dev = &sh->dev[s->failed_num[0]];
2826 set_bit(R5_LOCKED, &dev->flags);
2827 set_bit(R5_Wantwrite, &dev->flags);
2829 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2830 dev = &sh->dev[pd_idx];
2832 set_bit(R5_LOCKED, &dev->flags);
2833 set_bit(R5_Wantwrite, &dev->flags);
2835 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2836 dev = &sh->dev[qd_idx];
2838 set_bit(R5_LOCKED, &dev->flags);
2839 set_bit(R5_Wantwrite, &dev->flags);
2841 clear_bit(STRIPE_DEGRADED, &sh->state);
2843 set_bit(STRIPE_INSYNC, &sh->state);
2845 case check_state_run:
2846 case check_state_run_q:
2847 case check_state_run_pq:
2848 break; /* we will be called again upon completion */
2849 case check_state_check_result:
2850 sh->check_state = check_state_idle;
2852 /* handle a successful check operation, if parity is correct
2853 * we are done. Otherwise update the mismatch count and repair
2854 * parity if !MD_RECOVERY_CHECK
2856 if (sh->ops.zero_sum_result == 0) {
2857 /* both parities are correct */
2859 set_bit(STRIPE_INSYNC, &sh->state);
2861 /* in contrast to the raid5 case we can validate
2862 * parity, but still have a failure to write
2865 sh->check_state = check_state_compute_result;
2866 /* Returning at this point means that we may go
2867 * off and bring p and/or q uptodate again so
2868 * we make sure to check zero_sum_result again
2869 * to verify if p or q need writeback
2873 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2874 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2875 /* don't try to repair!! */
2876 set_bit(STRIPE_INSYNC, &sh->state);
2878 int *target = &sh->ops.target;
2880 sh->ops.target = -1;
2881 sh->ops.target2 = -1;
2882 sh->check_state = check_state_compute_run;
2883 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2884 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2885 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2886 set_bit(R5_Wantcompute,
2887 &sh->dev[pd_idx].flags);
2889 target = &sh->ops.target2;
2892 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2893 set_bit(R5_Wantcompute,
2894 &sh->dev[qd_idx].flags);
2901 case check_state_compute_run:
2904 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2905 __func__, sh->check_state,
2906 (unsigned long long) sh->sector);
2911 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
2915 /* We have read all the blocks in this stripe and now we need to
2916 * copy some of them into a target stripe for expand.
2918 struct dma_async_tx_descriptor *tx = NULL;
2919 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2920 for (i = 0; i < sh->disks; i++)
2921 if (i != sh->pd_idx && i != sh->qd_idx) {
2923 struct stripe_head *sh2;
2924 struct async_submit_ctl submit;
2926 sector_t bn = compute_blocknr(sh, i, 1);
2927 sector_t s = raid5_compute_sector(conf, bn, 0,
2929 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2931 /* so far only the early blocks of this stripe
2932 * have been requested. When later blocks
2933 * get requested, we will try again
2936 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2937 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2938 /* must have already done this block */
2939 release_stripe(sh2);
2943 /* place all the copies on one channel */
2944 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2945 tx = async_memcpy(sh2->dev[dd_idx].page,
2946 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2949 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2950 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2951 for (j = 0; j < conf->raid_disks; j++)
2952 if (j != sh2->pd_idx &&
2954 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2956 if (j == conf->raid_disks) {
2957 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2958 set_bit(STRIPE_HANDLE, &sh2->state);
2960 release_stripe(sh2);
2963 /* done submitting copies, wait for them to complete */
2966 dma_wait_for_async_tx(tx);
2972 * handle_stripe - do things to a stripe.
2974 * We lock the stripe and then examine the state of various bits
2975 * to see what needs to be done.
2977 * return some read request which now have data
2978 * return some write requests which are safely on disc
2979 * schedule a read on some buffers
2980 * schedule a write of some buffers
2981 * return confirmation of parity correctness
2983 * buffers are taken off read_list or write_list, and bh_cache buffers
2984 * get BH_Lock set before the stripe lock is released.
2988 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
2990 struct r5conf *conf = sh->raid_conf;
2991 int disks = sh->disks;
2995 memset(s, 0, sizeof(*s));
2997 s->syncing = test_bit(STRIPE_SYNCING, &sh->state);
2998 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2999 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3000 s->failed_num[0] = -1;
3001 s->failed_num[1] = -1;
3003 /* Now to look around and see what can be done */
3005 spin_lock_irq(&conf->device_lock);
3006 for (i=disks; i--; ) {
3007 struct md_rdev *rdev;
3014 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3015 i, dev->flags, dev->toread, dev->towrite, dev->written);
3016 /* maybe we can reply to a read
3018 * new wantfill requests are only permitted while
3019 * ops_complete_biofill is guaranteed to be inactive
3021 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3022 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3023 set_bit(R5_Wantfill, &dev->flags);
3025 /* now count some things */
3026 if (test_bit(R5_LOCKED, &dev->flags))
3028 if (test_bit(R5_UPTODATE, &dev->flags))
3030 if (test_bit(R5_Wantcompute, &dev->flags)) {
3032 BUG_ON(s->compute > 2);
3035 if (test_bit(R5_Wantfill, &dev->flags))
3037 else if (dev->toread)
3041 if (!test_bit(R5_OVERWRITE, &dev->flags))
3046 rdev = rcu_dereference(conf->disks[i].rdev);
3047 if (rdev && test_bit(Faulty, &rdev->flags))
3050 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3051 &first_bad, &bad_sectors);
3052 if (s->blocked_rdev == NULL
3053 && (test_bit(Blocked, &rdev->flags)
3056 set_bit(BlockedBadBlocks,
3058 s->blocked_rdev = rdev;
3059 atomic_inc(&rdev->nr_pending);
3062 clear_bit(R5_Insync, &dev->flags);
3066 /* also not in-sync */
3067 if (!test_bit(WriteErrorSeen, &rdev->flags)) {
3068 /* treat as in-sync, but with a read error
3069 * which we can now try to correct
3071 set_bit(R5_Insync, &dev->flags);
3072 set_bit(R5_ReadError, &dev->flags);
3074 } else if (test_bit(In_sync, &rdev->flags))
3075 set_bit(R5_Insync, &dev->flags);
3076 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3077 /* in sync if before recovery_offset */
3078 set_bit(R5_Insync, &dev->flags);
3079 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3080 test_bit(R5_Expanded, &dev->flags))
3081 /* If we've reshaped into here, we assume it is Insync.
3082 * We will shortly update recovery_offset to make
3085 set_bit(R5_Insync, &dev->flags);
3087 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3088 clear_bit(R5_Insync, &dev->flags);
3089 if (!test_bit(Faulty, &rdev->flags)) {
3090 s->handle_bad_blocks = 1;
3091 atomic_inc(&rdev->nr_pending);
3093 clear_bit(R5_WriteError, &dev->flags);
3095 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3096 if (!test_bit(Faulty, &rdev->flags)) {
3097 s->handle_bad_blocks = 1;
3098 atomic_inc(&rdev->nr_pending);
3100 clear_bit(R5_MadeGood, &dev->flags);
3102 if (!test_bit(R5_Insync, &dev->flags)) {
3103 /* The ReadError flag will just be confusing now */
3104 clear_bit(R5_ReadError, &dev->flags);
3105 clear_bit(R5_ReWrite, &dev->flags);
3107 if (test_bit(R5_ReadError, &dev->flags))
3108 clear_bit(R5_Insync, &dev->flags);
3109 if (!test_bit(R5_Insync, &dev->flags)) {
3111 s->failed_num[s->failed] = i;
3115 spin_unlock_irq(&conf->device_lock);
3119 static void handle_stripe(struct stripe_head *sh)
3121 struct stripe_head_state s;
3122 struct r5conf *conf = sh->raid_conf;
3125 int disks = sh->disks;
3126 struct r5dev *pdev, *qdev;
3128 clear_bit(STRIPE_HANDLE, &sh->state);
3129 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3130 /* already being handled, ensure it gets handled
3131 * again when current action finishes */
3132 set_bit(STRIPE_HANDLE, &sh->state);
3136 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3137 set_bit(STRIPE_SYNCING, &sh->state);
3138 clear_bit(STRIPE_INSYNC, &sh->state);
3140 clear_bit(STRIPE_DELAYED, &sh->state);
3142 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3143 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3144 (unsigned long long)sh->sector, sh->state,
3145 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3146 sh->check_state, sh->reconstruct_state);
3148 analyse_stripe(sh, &s);
3150 if (s.handle_bad_blocks) {
3151 set_bit(STRIPE_HANDLE, &sh->state);
3155 if (unlikely(s.blocked_rdev)) {
3156 if (s.syncing || s.expanding || s.expanded ||
3157 s.to_write || s.written) {
3158 set_bit(STRIPE_HANDLE, &sh->state);
3161 /* There is nothing for the blocked_rdev to block */
3162 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3163 s.blocked_rdev = NULL;
3166 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3167 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3168 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3171 pr_debug("locked=%d uptodate=%d to_read=%d"
3172 " to_write=%d failed=%d failed_num=%d,%d\n",
3173 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3174 s.failed_num[0], s.failed_num[1]);
3175 /* check if the array has lost more than max_degraded devices and,
3176 * if so, some requests might need to be failed.
3178 if (s.failed > conf->max_degraded) {
3179 sh->check_state = 0;
3180 sh->reconstruct_state = 0;
3181 if (s.to_read+s.to_write+s.written)
3182 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3184 handle_failed_sync(conf, sh, &s);
3188 * might be able to return some write requests if the parity blocks
3189 * are safe, or on a failed drive
3191 pdev = &sh->dev[sh->pd_idx];
3192 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3193 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3194 qdev = &sh->dev[sh->qd_idx];
3195 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3196 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3200 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3201 && !test_bit(R5_LOCKED, &pdev->flags)
3202 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3203 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3204 && !test_bit(R5_LOCKED, &qdev->flags)
3205 && test_bit(R5_UPTODATE, &qdev->flags)))))
3206 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3208 /* Now we might consider reading some blocks, either to check/generate
3209 * parity, or to satisfy requests
3210 * or to load a block that is being partially written.
3212 if (s.to_read || s.non_overwrite
3213 || (conf->level == 6 && s.to_write && s.failed)
3214 || (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3215 handle_stripe_fill(sh, &s, disks);
3217 /* Now we check to see if any write operations have recently
3221 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3223 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3224 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3225 sh->reconstruct_state = reconstruct_state_idle;
3227 /* All the 'written' buffers and the parity block are ready to
3228 * be written back to disk
3230 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3231 BUG_ON(sh->qd_idx >= 0 &&
3232 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3233 for (i = disks; i--; ) {
3234 struct r5dev *dev = &sh->dev[i];
3235 if (test_bit(R5_LOCKED, &dev->flags) &&
3236 (i == sh->pd_idx || i == sh->qd_idx ||
3238 pr_debug("Writing block %d\n", i);
3239 set_bit(R5_Wantwrite, &dev->flags);
3242 if (!test_bit(R5_Insync, &dev->flags) ||
3243 ((i == sh->pd_idx || i == sh->qd_idx) &&
3245 set_bit(STRIPE_INSYNC, &sh->state);
3248 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3249 s.dec_preread_active = 1;
3252 /* Now to consider new write requests and what else, if anything
3253 * should be read. We do not handle new writes when:
3254 * 1/ A 'write' operation (copy+xor) is already in flight.
3255 * 2/ A 'check' operation is in flight, as it may clobber the parity
3258 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3259 handle_stripe_dirtying(conf, sh, &s, disks);
3261 /* maybe we need to check and possibly fix the parity for this stripe
3262 * Any reads will already have been scheduled, so we just see if enough
3263 * data is available. The parity check is held off while parity
3264 * dependent operations are in flight.
3266 if (sh->check_state ||
3267 (s.syncing && s.locked == 0 &&
3268 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3269 !test_bit(STRIPE_INSYNC, &sh->state))) {
3270 if (conf->level == 6)
3271 handle_parity_checks6(conf, sh, &s, disks);
3273 handle_parity_checks5(conf, sh, &s, disks);
3276 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3277 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3278 clear_bit(STRIPE_SYNCING, &sh->state);
3281 /* If the failed drives are just a ReadError, then we might need
3282 * to progress the repair/check process
3284 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3285 for (i = 0; i < s.failed; i++) {
3286 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3287 if (test_bit(R5_ReadError, &dev->flags)
3288 && !test_bit(R5_LOCKED, &dev->flags)
3289 && test_bit(R5_UPTODATE, &dev->flags)
3291 if (!test_bit(R5_ReWrite, &dev->flags)) {
3292 set_bit(R5_Wantwrite, &dev->flags);
3293 set_bit(R5_ReWrite, &dev->flags);
3294 set_bit(R5_LOCKED, &dev->flags);
3297 /* let's read it back */
3298 set_bit(R5_Wantread, &dev->flags);
3299 set_bit(R5_LOCKED, &dev->flags);
3306 /* Finish reconstruct operations initiated by the expansion process */
3307 if (sh->reconstruct_state == reconstruct_state_result) {
3308 struct stripe_head *sh_src
3309 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3310 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3311 /* sh cannot be written until sh_src has been read.
3312 * so arrange for sh to be delayed a little
3314 set_bit(STRIPE_DELAYED, &sh->state);
3315 set_bit(STRIPE_HANDLE, &sh->state);
3316 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3318 atomic_inc(&conf->preread_active_stripes);
3319 release_stripe(sh_src);
3323 release_stripe(sh_src);
3325 sh->reconstruct_state = reconstruct_state_idle;
3326 clear_bit(STRIPE_EXPANDING, &sh->state);
3327 for (i = conf->raid_disks; i--; ) {
3328 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3329 set_bit(R5_LOCKED, &sh->dev[i].flags);
3334 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3335 !sh->reconstruct_state) {
3336 /* Need to write out all blocks after computing parity */
3337 sh->disks = conf->raid_disks;
3338 stripe_set_idx(sh->sector, conf, 0, sh);
3339 schedule_reconstruction(sh, &s, 1, 1);
3340 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3341 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3342 atomic_dec(&conf->reshape_stripes);
3343 wake_up(&conf->wait_for_overlap);
3344 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3347 if (s.expanding && s.locked == 0 &&
3348 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3349 handle_stripe_expansion(conf, sh);
3352 /* wait for this device to become unblocked */
3353 if (conf->mddev->external && unlikely(s.blocked_rdev))
3354 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3356 if (s.handle_bad_blocks)
3357 for (i = disks; i--; ) {
3358 struct md_rdev *rdev;
3359 struct r5dev *dev = &sh->dev[i];
3360 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3361 /* We own a safe reference to the rdev */
3362 rdev = conf->disks[i].rdev;
3363 if (!rdev_set_badblocks(rdev, sh->sector,
3365 md_error(conf->mddev, rdev);
3366 rdev_dec_pending(rdev, conf->mddev);
3368 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3369 rdev = conf->disks[i].rdev;
3370 rdev_clear_badblocks(rdev, sh->sector,
3372 rdev_dec_pending(rdev, conf->mddev);
3377 raid_run_ops(sh, s.ops_request);
3381 if (s.dec_preread_active) {
3382 /* We delay this until after ops_run_io so that if make_request
3383 * is waiting on a flush, it won't continue until the writes
3384 * have actually been submitted.
3386 atomic_dec(&conf->preread_active_stripes);
3387 if (atomic_read(&conf->preread_active_stripes) <
3389 md_wakeup_thread(conf->mddev->thread);
3392 return_io(s.return_bi);
3394 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3397 static void raid5_activate_delayed(struct r5conf *conf)
3399 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3400 while (!list_empty(&conf->delayed_list)) {
3401 struct list_head *l = conf->delayed_list.next;
3402 struct stripe_head *sh;
3403 sh = list_entry(l, struct stripe_head, lru);
3405 clear_bit(STRIPE_DELAYED, &sh->state);
3406 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3407 atomic_inc(&conf->preread_active_stripes);
3408 list_add_tail(&sh->lru, &conf->hold_list);
3413 static void activate_bit_delay(struct r5conf *conf)
3415 /* device_lock is held */
3416 struct list_head head;
3417 list_add(&head, &conf->bitmap_list);
3418 list_del_init(&conf->bitmap_list);
3419 while (!list_empty(&head)) {
3420 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3421 list_del_init(&sh->lru);
3422 atomic_inc(&sh->count);
3423 __release_stripe(conf, sh);
3427 int md_raid5_congested(struct mddev *mddev, int bits)
3429 struct r5conf *conf = mddev->private;
3431 /* No difference between reads and writes. Just check
3432 * how busy the stripe_cache is
3435 if (conf->inactive_blocked)
3439 if (list_empty_careful(&conf->inactive_list))
3444 EXPORT_SYMBOL_GPL(md_raid5_congested);
3446 static int raid5_congested(void *data, int bits)
3448 struct mddev *mddev = data;
3450 return mddev_congested(mddev, bits) ||
3451 md_raid5_congested(mddev, bits);
3454 /* We want read requests to align with chunks where possible,
3455 * but write requests don't need to.
3457 static int raid5_mergeable_bvec(struct request_queue *q,
3458 struct bvec_merge_data *bvm,
3459 struct bio_vec *biovec)
3461 struct mddev *mddev = q->queuedata;
3462 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3464 unsigned int chunk_sectors = mddev->chunk_sectors;
3465 unsigned int bio_sectors = bvm->bi_size >> 9;
3467 if ((bvm->bi_rw & 1) == WRITE)
3468 return biovec->bv_len; /* always allow writes to be mergeable */
3470 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3471 chunk_sectors = mddev->new_chunk_sectors;
3472 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3473 if (max < 0) max = 0;
3474 if (max <= biovec->bv_len && bio_sectors == 0)
3475 return biovec->bv_len;
3481 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3483 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3484 unsigned int chunk_sectors = mddev->chunk_sectors;
3485 unsigned int bio_sectors = bio->bi_size >> 9;
3487 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3488 chunk_sectors = mddev->new_chunk_sectors;
3489 return chunk_sectors >=
3490 ((sector & (chunk_sectors - 1)) + bio_sectors);
3494 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3495 * later sampled by raid5d.
3497 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3499 unsigned long flags;
3501 spin_lock_irqsave(&conf->device_lock, flags);
3503 bi->bi_next = conf->retry_read_aligned_list;
3504 conf->retry_read_aligned_list = bi;
3506 spin_unlock_irqrestore(&conf->device_lock, flags);
3507 md_wakeup_thread(conf->mddev->thread);
3511 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3515 bi = conf->retry_read_aligned;
3517 conf->retry_read_aligned = NULL;
3520 bi = conf->retry_read_aligned_list;
3522 conf->retry_read_aligned_list = bi->bi_next;
3525 * this sets the active strip count to 1 and the processed
3526 * strip count to zero (upper 8 bits)
3528 bi->bi_phys_segments = 1; /* biased count of active stripes */
3536 * The "raid5_align_endio" should check if the read succeeded and if it
3537 * did, call bio_endio on the original bio (having bio_put the new bio
3539 * If the read failed..
3541 static void raid5_align_endio(struct bio *bi, int error)
3543 struct bio* raid_bi = bi->bi_private;
3544 struct mddev *mddev;
3545 struct r5conf *conf;
3546 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3547 struct md_rdev *rdev;
3551 rdev = (void*)raid_bi->bi_next;
3552 raid_bi->bi_next = NULL;
3553 mddev = rdev->mddev;
3554 conf = mddev->private;
3556 rdev_dec_pending(rdev, conf->mddev);
3558 if (!error && uptodate) {
3559 bio_endio(raid_bi, 0);
3560 if (atomic_dec_and_test(&conf->active_aligned_reads))
3561 wake_up(&conf->wait_for_stripe);
3566 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3568 add_bio_to_retry(raid_bi, conf);
3571 static int bio_fits_rdev(struct bio *bi)
3573 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3575 if ((bi->bi_size>>9) > queue_max_sectors(q))
3577 blk_recount_segments(q, bi);
3578 if (bi->bi_phys_segments > queue_max_segments(q))
3581 if (q->merge_bvec_fn)
3582 /* it's too hard to apply the merge_bvec_fn at this stage,
3591 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3593 struct r5conf *conf = mddev->private;
3595 struct bio* align_bi;
3596 struct md_rdev *rdev;
3597 sector_t end_sector;
3599 if (!in_chunk_boundary(mddev, raid_bio)) {
3600 pr_debug("chunk_aligned_read : non aligned\n");
3604 * use bio_clone_mddev to make a copy of the bio
3606 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3610 * set bi_end_io to a new function, and set bi_private to the
3613 align_bi->bi_end_io = raid5_align_endio;
3614 align_bi->bi_private = raid_bio;
3618 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3622 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3624 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3625 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3626 rdev->recovery_offset < end_sector) {
3627 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3629 (test_bit(Faulty, &rdev->flags) ||
3630 !(test_bit(In_sync, &rdev->flags) ||
3631 rdev->recovery_offset >= end_sector)))
3638 atomic_inc(&rdev->nr_pending);
3640 raid_bio->bi_next = (void*)rdev;
3641 align_bi->bi_bdev = rdev->bdev;
3642 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3643 align_bi->bi_sector += rdev->data_offset;
3645 if (!bio_fits_rdev(align_bi) ||
3646 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3647 &first_bad, &bad_sectors)) {
3648 /* too big in some way, or has a known bad block */
3650 rdev_dec_pending(rdev, mddev);
3654 spin_lock_irq(&conf->device_lock);
3655 wait_event_lock_irq(conf->wait_for_stripe,
3657 conf->device_lock, /* nothing */);
3658 atomic_inc(&conf->active_aligned_reads);
3659 spin_unlock_irq(&conf->device_lock);
3661 generic_make_request(align_bi);
3670 /* __get_priority_stripe - get the next stripe to process
3672 * Full stripe writes are allowed to pass preread active stripes up until
3673 * the bypass_threshold is exceeded. In general the bypass_count
3674 * increments when the handle_list is handled before the hold_list; however, it
3675 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3676 * stripe with in flight i/o. The bypass_count will be reset when the
3677 * head of the hold_list has changed, i.e. the head was promoted to the
3680 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3682 struct stripe_head *sh;
3684 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3686 list_empty(&conf->handle_list) ? "empty" : "busy",
3687 list_empty(&conf->hold_list) ? "empty" : "busy",
3688 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3690 if (!list_empty(&conf->handle_list)) {
3691 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3693 if (list_empty(&conf->hold_list))
3694 conf->bypass_count = 0;
3695 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3696 if (conf->hold_list.next == conf->last_hold)
3697 conf->bypass_count++;
3699 conf->last_hold = conf->hold_list.next;
3700 conf->bypass_count -= conf->bypass_threshold;
3701 if (conf->bypass_count < 0)
3702 conf->bypass_count = 0;
3705 } else if (!list_empty(&conf->hold_list) &&
3706 ((conf->bypass_threshold &&
3707 conf->bypass_count > conf->bypass_threshold) ||
3708 atomic_read(&conf->pending_full_writes) == 0)) {
3709 sh = list_entry(conf->hold_list.next,
3711 conf->bypass_count -= conf->bypass_threshold;
3712 if (conf->bypass_count < 0)
3713 conf->bypass_count = 0;
3717 list_del_init(&sh->lru);
3718 atomic_inc(&sh->count);
3719 BUG_ON(atomic_read(&sh->count) != 1);
3723 static void make_request(struct mddev *mddev, struct bio * bi)
3725 struct r5conf *conf = mddev->private;
3727 sector_t new_sector;
3728 sector_t logical_sector, last_sector;
3729 struct stripe_head *sh;
3730 const int rw = bio_data_dir(bi);
3734 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3735 md_flush_request(mddev, bi);
3739 md_write_start(mddev, bi);
3742 mddev->reshape_position == MaxSector &&
3743 chunk_aligned_read(mddev,bi))
3746 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3747 last_sector = bi->bi_sector + (bi->bi_size>>9);
3749 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3751 plugged = mddev_check_plugged(mddev);
3752 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3754 int disks, data_disks;
3759 disks = conf->raid_disks;
3760 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3761 if (unlikely(conf->reshape_progress != MaxSector)) {
3762 /* spinlock is needed as reshape_progress may be
3763 * 64bit on a 32bit platform, and so it might be
3764 * possible to see a half-updated value
3765 * Of course reshape_progress could change after
3766 * the lock is dropped, so once we get a reference
3767 * to the stripe that we think it is, we will have
3770 spin_lock_irq(&conf->device_lock);
3771 if (mddev->delta_disks < 0
3772 ? logical_sector < conf->reshape_progress
3773 : logical_sector >= conf->reshape_progress) {
3774 disks = conf->previous_raid_disks;
3777 if (mddev->delta_disks < 0
3778 ? logical_sector < conf->reshape_safe
3779 : logical_sector >= conf->reshape_safe) {
3780 spin_unlock_irq(&conf->device_lock);
3785 spin_unlock_irq(&conf->device_lock);
3787 data_disks = disks - conf->max_degraded;
3789 new_sector = raid5_compute_sector(conf, logical_sector,
3792 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3793 (unsigned long long)new_sector,
3794 (unsigned long long)logical_sector);
3796 sh = get_active_stripe(conf, new_sector, previous,
3797 (bi->bi_rw&RWA_MASK), 0);
3799 if (unlikely(previous)) {
3800 /* expansion might have moved on while waiting for a
3801 * stripe, so we must do the range check again.
3802 * Expansion could still move past after this
3803 * test, but as we are holding a reference to
3804 * 'sh', we know that if that happens,
3805 * STRIPE_EXPANDING will get set and the expansion
3806 * won't proceed until we finish with the stripe.
3809 spin_lock_irq(&conf->device_lock);
3810 if (mddev->delta_disks < 0
3811 ? logical_sector >= conf->reshape_progress
3812 : logical_sector < conf->reshape_progress)
3813 /* mismatch, need to try again */
3815 spin_unlock_irq(&conf->device_lock);
3824 logical_sector >= mddev->suspend_lo &&
3825 logical_sector < mddev->suspend_hi) {
3827 /* As the suspend_* range is controlled by
3828 * userspace, we want an interruptible
3831 flush_signals(current);
3832 prepare_to_wait(&conf->wait_for_overlap,
3833 &w, TASK_INTERRUPTIBLE);
3834 if (logical_sector >= mddev->suspend_lo &&
3835 logical_sector < mddev->suspend_hi)
3840 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3841 !add_stripe_bio(sh, bi, dd_idx, rw)) {
3842 /* Stripe is busy expanding or
3843 * add failed due to overlap. Flush everything
3846 md_wakeup_thread(mddev->thread);
3851 finish_wait(&conf->wait_for_overlap, &w);
3852 set_bit(STRIPE_HANDLE, &sh->state);
3853 clear_bit(STRIPE_DELAYED, &sh->state);
3854 if ((bi->bi_rw & REQ_SYNC) &&
3855 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3856 atomic_inc(&conf->preread_active_stripes);
3859 /* cannot get stripe for read-ahead, just give-up */
3860 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3861 finish_wait(&conf->wait_for_overlap, &w);
3867 md_wakeup_thread(mddev->thread);
3869 spin_lock_irq(&conf->device_lock);
3870 remaining = raid5_dec_bi_phys_segments(bi);
3871 spin_unlock_irq(&conf->device_lock);
3872 if (remaining == 0) {
3875 md_write_end(mddev);
3881 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
3883 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
3885 /* reshaping is quite different to recovery/resync so it is
3886 * handled quite separately ... here.
3888 * On each call to sync_request, we gather one chunk worth of
3889 * destination stripes and flag them as expanding.
3890 * Then we find all the source stripes and request reads.
3891 * As the reads complete, handle_stripe will copy the data
3892 * into the destination stripe and release that stripe.
3894 struct r5conf *conf = mddev->private;
3895 struct stripe_head *sh;
3896 sector_t first_sector, last_sector;
3897 int raid_disks = conf->previous_raid_disks;
3898 int data_disks = raid_disks - conf->max_degraded;
3899 int new_data_disks = conf->raid_disks - conf->max_degraded;
3902 sector_t writepos, readpos, safepos;
3903 sector_t stripe_addr;
3904 int reshape_sectors;
3905 struct list_head stripes;
3907 if (sector_nr == 0) {
3908 /* If restarting in the middle, skip the initial sectors */
3909 if (mddev->delta_disks < 0 &&
3910 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3911 sector_nr = raid5_size(mddev, 0, 0)
3912 - conf->reshape_progress;
3913 } else if (mddev->delta_disks >= 0 &&
3914 conf->reshape_progress > 0)
3915 sector_nr = conf->reshape_progress;
3916 sector_div(sector_nr, new_data_disks);
3918 mddev->curr_resync_completed = sector_nr;
3919 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3925 /* We need to process a full chunk at a time.
3926 * If old and new chunk sizes differ, we need to process the
3929 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3930 reshape_sectors = mddev->new_chunk_sectors;
3932 reshape_sectors = mddev->chunk_sectors;
3934 /* we update the metadata when there is more than 3Meg
3935 * in the block range (that is rather arbitrary, should
3936 * probably be time based) or when the data about to be
3937 * copied would over-write the source of the data at
3938 * the front of the range.
3939 * i.e. one new_stripe along from reshape_progress new_maps
3940 * to after where reshape_safe old_maps to
3942 writepos = conf->reshape_progress;
3943 sector_div(writepos, new_data_disks);
3944 readpos = conf->reshape_progress;
3945 sector_div(readpos, data_disks);
3946 safepos = conf->reshape_safe;
3947 sector_div(safepos, data_disks);
3948 if (mddev->delta_disks < 0) {
3949 writepos -= min_t(sector_t, reshape_sectors, writepos);
3950 readpos += reshape_sectors;
3951 safepos += reshape_sectors;
3953 writepos += reshape_sectors;
3954 readpos -= min_t(sector_t, reshape_sectors, readpos);
3955 safepos -= min_t(sector_t, reshape_sectors, safepos);
3958 /* 'writepos' is the most advanced device address we might write.
3959 * 'readpos' is the least advanced device address we might read.
3960 * 'safepos' is the least address recorded in the metadata as having
3962 * If 'readpos' is behind 'writepos', then there is no way that we can
3963 * ensure safety in the face of a crash - that must be done by userspace
3964 * making a backup of the data. So in that case there is no particular
3965 * rush to update metadata.
3966 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3967 * update the metadata to advance 'safepos' to match 'readpos' so that
3968 * we can be safe in the event of a crash.
3969 * So we insist on updating metadata if safepos is behind writepos and
3970 * readpos is beyond writepos.
3971 * In any case, update the metadata every 10 seconds.
3972 * Maybe that number should be configurable, but I'm not sure it is
3973 * worth it.... maybe it could be a multiple of safemode_delay???
3975 if ((mddev->delta_disks < 0
3976 ? (safepos > writepos && readpos < writepos)
3977 : (safepos < writepos && readpos > writepos)) ||
3978 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3979 /* Cannot proceed until we've updated the superblock... */
3980 wait_event(conf->wait_for_overlap,
3981 atomic_read(&conf->reshape_stripes)==0);
3982 mddev->reshape_position = conf->reshape_progress;
3983 mddev->curr_resync_completed = sector_nr;
3984 conf->reshape_checkpoint = jiffies;
3985 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3986 md_wakeup_thread(mddev->thread);
3987 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3988 kthread_should_stop());
3989 spin_lock_irq(&conf->device_lock);
3990 conf->reshape_safe = mddev->reshape_position;
3991 spin_unlock_irq(&conf->device_lock);
3992 wake_up(&conf->wait_for_overlap);
3993 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3996 if (mddev->delta_disks < 0) {
3997 BUG_ON(conf->reshape_progress == 0);
3998 stripe_addr = writepos;
3999 BUG_ON((mddev->dev_sectors &
4000 ~((sector_t)reshape_sectors - 1))
4001 - reshape_sectors - stripe_addr
4004 BUG_ON(writepos != sector_nr + reshape_sectors);
4005 stripe_addr = sector_nr;
4007 INIT_LIST_HEAD(&stripes);
4008 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4010 int skipped_disk = 0;
4011 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4012 set_bit(STRIPE_EXPANDING, &sh->state);
4013 atomic_inc(&conf->reshape_stripes);
4014 /* If any of this stripe is beyond the end of the old
4015 * array, then we need to zero those blocks
4017 for (j=sh->disks; j--;) {
4019 if (j == sh->pd_idx)
4021 if (conf->level == 6 &&
4024 s = compute_blocknr(sh, j, 0);
4025 if (s < raid5_size(mddev, 0, 0)) {
4029 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4030 set_bit(R5_Expanded, &sh->dev[j].flags);
4031 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4033 if (!skipped_disk) {
4034 set_bit(STRIPE_EXPAND_READY, &sh->state);
4035 set_bit(STRIPE_HANDLE, &sh->state);
4037 list_add(&sh->lru, &stripes);
4039 spin_lock_irq(&conf->device_lock);
4040 if (mddev->delta_disks < 0)
4041 conf->reshape_progress -= reshape_sectors * new_data_disks;
4043 conf->reshape_progress += reshape_sectors * new_data_disks;
4044 spin_unlock_irq(&conf->device_lock);
4045 /* Ok, those stripe are ready. We can start scheduling
4046 * reads on the source stripes.
4047 * The source stripes are determined by mapping the first and last
4048 * block on the destination stripes.
4051 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4054 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4055 * new_data_disks - 1),
4057 if (last_sector >= mddev->dev_sectors)
4058 last_sector = mddev->dev_sectors - 1;
4059 while (first_sector <= last_sector) {
4060 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4061 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4062 set_bit(STRIPE_HANDLE, &sh->state);
4064 first_sector += STRIPE_SECTORS;
4066 /* Now that the sources are clearly marked, we can release
4067 * the destination stripes
4069 while (!list_empty(&stripes)) {
4070 sh = list_entry(stripes.next, struct stripe_head, lru);
4071 list_del_init(&sh->lru);
4074 /* If this takes us to the resync_max point where we have to pause,
4075 * then we need to write out the superblock.
4077 sector_nr += reshape_sectors;
4078 if ((sector_nr - mddev->curr_resync_completed) * 2
4079 >= mddev->resync_max - mddev->curr_resync_completed) {
4080 /* Cannot proceed until we've updated the superblock... */
4081 wait_event(conf->wait_for_overlap,
4082 atomic_read(&conf->reshape_stripes) == 0);
4083 mddev->reshape_position = conf->reshape_progress;
4084 mddev->curr_resync_completed = sector_nr;
4085 conf->reshape_checkpoint = jiffies;
4086 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4087 md_wakeup_thread(mddev->thread);
4088 wait_event(mddev->sb_wait,
4089 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4090 || kthread_should_stop());
4091 spin_lock_irq(&conf->device_lock);
4092 conf->reshape_safe = mddev->reshape_position;
4093 spin_unlock_irq(&conf->device_lock);
4094 wake_up(&conf->wait_for_overlap);
4095 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4097 return reshape_sectors;
4100 /* FIXME go_faster isn't used */
4101 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4103 struct r5conf *conf = mddev->private;
4104 struct stripe_head *sh;
4105 sector_t max_sector = mddev->dev_sectors;
4106 sector_t sync_blocks;
4107 int still_degraded = 0;
4110 if (sector_nr >= max_sector) {
4111 /* just being told to finish up .. nothing much to do */
4113 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4118 if (mddev->curr_resync < max_sector) /* aborted */
4119 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4121 else /* completed sync */
4123 bitmap_close_sync(mddev->bitmap);
4128 /* Allow raid5_quiesce to complete */
4129 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4131 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4132 return reshape_request(mddev, sector_nr, skipped);
4134 /* No need to check resync_max as we never do more than one
4135 * stripe, and as resync_max will always be on a chunk boundary,
4136 * if the check in md_do_sync didn't fire, there is no chance
4137 * of overstepping resync_max here
4140 /* if there is too many failed drives and we are trying
4141 * to resync, then assert that we are finished, because there is
4142 * nothing we can do.
4144 if (mddev->degraded >= conf->max_degraded &&
4145 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4146 sector_t rv = mddev->dev_sectors - sector_nr;
4150 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4151 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4152 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4153 /* we can skip this block, and probably more */
4154 sync_blocks /= STRIPE_SECTORS;
4156 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4160 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4162 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4164 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4165 /* make sure we don't swamp the stripe cache if someone else
4166 * is trying to get access
4168 schedule_timeout_uninterruptible(1);
4170 /* Need to check if array will still be degraded after recovery/resync
4171 * We don't need to check the 'failed' flag as when that gets set,
4174 for (i = 0; i < conf->raid_disks; i++)
4175 if (conf->disks[i].rdev == NULL)
4178 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4180 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4185 return STRIPE_SECTORS;
4188 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4190 /* We may not be able to submit a whole bio at once as there
4191 * may not be enough stripe_heads available.
4192 * We cannot pre-allocate enough stripe_heads as we may need
4193 * more than exist in the cache (if we allow ever large chunks).
4194 * So we do one stripe head at a time and record in
4195 * ->bi_hw_segments how many have been done.
4197 * We *know* that this entire raid_bio is in one chunk, so
4198 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4200 struct stripe_head *sh;
4202 sector_t sector, logical_sector, last_sector;
4207 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4208 sector = raid5_compute_sector(conf, logical_sector,
4210 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4212 for (; logical_sector < last_sector;
4213 logical_sector += STRIPE_SECTORS,
4214 sector += STRIPE_SECTORS,
4217 if (scnt < raid5_bi_hw_segments(raid_bio))
4218 /* already done this stripe */
4221 sh = get_active_stripe(conf, sector, 0, 1, 0);
4224 /* failed to get a stripe - must wait */
4225 raid5_set_bi_hw_segments(raid_bio, scnt);
4226 conf->retry_read_aligned = raid_bio;
4230 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4231 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4233 raid5_set_bi_hw_segments(raid_bio, scnt);
4234 conf->retry_read_aligned = raid_bio;
4242 spin_lock_irq(&conf->device_lock);
4243 remaining = raid5_dec_bi_phys_segments(raid_bio);
4244 spin_unlock_irq(&conf->device_lock);
4246 bio_endio(raid_bio, 0);
4247 if (atomic_dec_and_test(&conf->active_aligned_reads))
4248 wake_up(&conf->wait_for_stripe);
4254 * This is our raid5 kernel thread.
4256 * We scan the hash table for stripes which can be handled now.
4257 * During the scan, completed stripes are saved for us by the interrupt
4258 * handler, so that they will not have to wait for our next wakeup.
4260 static void raid5d(struct mddev *mddev)
4262 struct stripe_head *sh;
4263 struct r5conf *conf = mddev->private;
4265 struct blk_plug plug;
4267 pr_debug("+++ raid5d active\n");
4269 md_check_recovery(mddev);
4271 blk_start_plug(&plug);
4273 spin_lock_irq(&conf->device_lock);
4277 if (atomic_read(&mddev->plug_cnt) == 0 &&
4278 !list_empty(&conf->bitmap_list)) {
4279 /* Now is a good time to flush some bitmap updates */
4281 spin_unlock_irq(&conf->device_lock);
4282 bitmap_unplug(mddev->bitmap);
4283 spin_lock_irq(&conf->device_lock);
4284 conf->seq_write = conf->seq_flush;
4285 activate_bit_delay(conf);
4287 if (atomic_read(&mddev->plug_cnt) == 0)
4288 raid5_activate_delayed(conf);
4290 while ((bio = remove_bio_from_retry(conf))) {
4292 spin_unlock_irq(&conf->device_lock);
4293 ok = retry_aligned_read(conf, bio);
4294 spin_lock_irq(&conf->device_lock);
4300 sh = __get_priority_stripe(conf);
4304 spin_unlock_irq(&conf->device_lock);
4311 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4312 md_check_recovery(mddev);
4314 spin_lock_irq(&conf->device_lock);
4316 pr_debug("%d stripes handled\n", handled);
4318 spin_unlock_irq(&conf->device_lock);
4320 async_tx_issue_pending_all();
4321 blk_finish_plug(&plug);
4323 pr_debug("--- raid5d inactive\n");
4327 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4329 struct r5conf *conf = mddev->private;
4331 return sprintf(page, "%d\n", conf->max_nr_stripes);
4337 raid5_set_cache_size(struct mddev *mddev, int size)
4339 struct r5conf *conf = mddev->private;
4342 if (size <= 16 || size > 32768)
4344 while (size < conf->max_nr_stripes) {
4345 if (drop_one_stripe(conf))
4346 conf->max_nr_stripes--;
4350 err = md_allow_write(mddev);
4353 while (size > conf->max_nr_stripes) {
4354 if (grow_one_stripe(conf))
4355 conf->max_nr_stripes++;
4360 EXPORT_SYMBOL(raid5_set_cache_size);
4363 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4365 struct r5conf *conf = mddev->private;
4369 if (len >= PAGE_SIZE)
4374 if (strict_strtoul(page, 10, &new))
4376 err = raid5_set_cache_size(mddev, new);
4382 static struct md_sysfs_entry
4383 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4384 raid5_show_stripe_cache_size,
4385 raid5_store_stripe_cache_size);
4388 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4390 struct r5conf *conf = mddev->private;
4392 return sprintf(page, "%d\n", conf->bypass_threshold);
4398 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4400 struct r5conf *conf = mddev->private;
4402 if (len >= PAGE_SIZE)
4407 if (strict_strtoul(page, 10, &new))
4409 if (new > conf->max_nr_stripes)
4411 conf->bypass_threshold = new;
4415 static struct md_sysfs_entry
4416 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4418 raid5_show_preread_threshold,
4419 raid5_store_preread_threshold);
4422 stripe_cache_active_show(struct mddev *mddev, char *page)
4424 struct r5conf *conf = mddev->private;
4426 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4431 static struct md_sysfs_entry
4432 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4434 static struct attribute *raid5_attrs[] = {
4435 &raid5_stripecache_size.attr,
4436 &raid5_stripecache_active.attr,
4437 &raid5_preread_bypass_threshold.attr,
4440 static struct attribute_group raid5_attrs_group = {
4442 .attrs = raid5_attrs,
4446 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4448 struct r5conf *conf = mddev->private;
4451 sectors = mddev->dev_sectors;
4453 /* size is defined by the smallest of previous and new size */
4454 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4456 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4457 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4458 return sectors * (raid_disks - conf->max_degraded);
4461 static void raid5_free_percpu(struct r5conf *conf)
4463 struct raid5_percpu *percpu;
4470 for_each_possible_cpu(cpu) {
4471 percpu = per_cpu_ptr(conf->percpu, cpu);
4472 safe_put_page(percpu->spare_page);
4473 kfree(percpu->scribble);
4475 #ifdef CONFIG_HOTPLUG_CPU
4476 unregister_cpu_notifier(&conf->cpu_notify);
4480 free_percpu(conf->percpu);
4483 static void free_conf(struct r5conf *conf)
4485 shrink_stripes(conf);
4486 raid5_free_percpu(conf);
4488 kfree(conf->stripe_hashtbl);
4492 #ifdef CONFIG_HOTPLUG_CPU
4493 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4496 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4497 long cpu = (long)hcpu;
4498 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4501 case CPU_UP_PREPARE:
4502 case CPU_UP_PREPARE_FROZEN:
4503 if (conf->level == 6 && !percpu->spare_page)
4504 percpu->spare_page = alloc_page(GFP_KERNEL);
4505 if (!percpu->scribble)
4506 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4508 if (!percpu->scribble ||
4509 (conf->level == 6 && !percpu->spare_page)) {
4510 safe_put_page(percpu->spare_page);
4511 kfree(percpu->scribble);
4512 pr_err("%s: failed memory allocation for cpu%ld\n",
4514 return notifier_from_errno(-ENOMEM);
4518 case CPU_DEAD_FROZEN:
4519 safe_put_page(percpu->spare_page);
4520 kfree(percpu->scribble);
4521 percpu->spare_page = NULL;
4522 percpu->scribble = NULL;
4531 static int raid5_alloc_percpu(struct r5conf *conf)
4534 struct page *spare_page;
4535 struct raid5_percpu __percpu *allcpus;
4539 allcpus = alloc_percpu(struct raid5_percpu);
4542 conf->percpu = allcpus;
4546 for_each_present_cpu(cpu) {
4547 if (conf->level == 6) {
4548 spare_page = alloc_page(GFP_KERNEL);
4553 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4555 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4560 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4562 #ifdef CONFIG_HOTPLUG_CPU
4563 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4564 conf->cpu_notify.priority = 0;
4566 err = register_cpu_notifier(&conf->cpu_notify);
4573 static struct r5conf *setup_conf(struct mddev *mddev)
4575 struct r5conf *conf;
4576 int raid_disk, memory, max_disks;
4577 struct md_rdev *rdev;
4578 struct disk_info *disk;
4580 if (mddev->new_level != 5
4581 && mddev->new_level != 4
4582 && mddev->new_level != 6) {
4583 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4584 mdname(mddev), mddev->new_level);
4585 return ERR_PTR(-EIO);
4587 if ((mddev->new_level == 5
4588 && !algorithm_valid_raid5(mddev->new_layout)) ||
4589 (mddev->new_level == 6
4590 && !algorithm_valid_raid6(mddev->new_layout))) {
4591 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4592 mdname(mddev), mddev->new_layout);
4593 return ERR_PTR(-EIO);
4595 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4596 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4597 mdname(mddev), mddev->raid_disks);
4598 return ERR_PTR(-EINVAL);
4601 if (!mddev->new_chunk_sectors ||
4602 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4603 !is_power_of_2(mddev->new_chunk_sectors)) {
4604 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4605 mdname(mddev), mddev->new_chunk_sectors << 9);
4606 return ERR_PTR(-EINVAL);
4609 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4612 spin_lock_init(&conf->device_lock);
4613 init_waitqueue_head(&conf->wait_for_stripe);
4614 init_waitqueue_head(&conf->wait_for_overlap);
4615 INIT_LIST_HEAD(&conf->handle_list);
4616 INIT_LIST_HEAD(&conf->hold_list);
4617 INIT_LIST_HEAD(&conf->delayed_list);
4618 INIT_LIST_HEAD(&conf->bitmap_list);
4619 INIT_LIST_HEAD(&conf->inactive_list);
4620 atomic_set(&conf->active_stripes, 0);
4621 atomic_set(&conf->preread_active_stripes, 0);
4622 atomic_set(&conf->active_aligned_reads, 0);
4623 conf->bypass_threshold = BYPASS_THRESHOLD;
4624 conf->recovery_disabled = mddev->recovery_disabled - 1;
4626 conf->raid_disks = mddev->raid_disks;
4627 if (mddev->reshape_position == MaxSector)
4628 conf->previous_raid_disks = mddev->raid_disks;
4630 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4631 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4632 conf->scribble_len = scribble_len(max_disks);
4634 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4639 conf->mddev = mddev;
4641 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4644 conf->level = mddev->new_level;
4645 if (raid5_alloc_percpu(conf) != 0)
4648 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4650 list_for_each_entry(rdev, &mddev->disks, same_set) {
4651 raid_disk = rdev->raid_disk;
4652 if (raid_disk >= max_disks
4655 disk = conf->disks + raid_disk;
4659 if (test_bit(In_sync, &rdev->flags)) {
4660 char b[BDEVNAME_SIZE];
4661 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4663 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4664 } else if (rdev->saved_raid_disk != raid_disk)
4665 /* Cannot rely on bitmap to complete recovery */
4669 conf->chunk_sectors = mddev->new_chunk_sectors;
4670 conf->level = mddev->new_level;
4671 if (conf->level == 6)
4672 conf->max_degraded = 2;
4674 conf->max_degraded = 1;
4675 conf->algorithm = mddev->new_layout;
4676 conf->max_nr_stripes = NR_STRIPES;
4677 conf->reshape_progress = mddev->reshape_position;
4678 if (conf->reshape_progress != MaxSector) {
4679 conf->prev_chunk_sectors = mddev->chunk_sectors;
4680 conf->prev_algo = mddev->layout;
4683 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4684 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4685 if (grow_stripes(conf, conf->max_nr_stripes)) {
4687 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4688 mdname(mddev), memory);
4691 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4692 mdname(mddev), memory);
4694 conf->thread = md_register_thread(raid5d, mddev, NULL);
4695 if (!conf->thread) {
4697 "md/raid:%s: couldn't allocate thread.\n",
4707 return ERR_PTR(-EIO);
4709 return ERR_PTR(-ENOMEM);
4713 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4716 case ALGORITHM_PARITY_0:
4717 if (raid_disk < max_degraded)
4720 case ALGORITHM_PARITY_N:
4721 if (raid_disk >= raid_disks - max_degraded)
4724 case ALGORITHM_PARITY_0_6:
4725 if (raid_disk == 0 ||
4726 raid_disk == raid_disks - 1)
4729 case ALGORITHM_LEFT_ASYMMETRIC_6:
4730 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4731 case ALGORITHM_LEFT_SYMMETRIC_6:
4732 case ALGORITHM_RIGHT_SYMMETRIC_6:
4733 if (raid_disk == raid_disks - 1)
4739 static int run(struct mddev *mddev)
4741 struct r5conf *conf;
4742 int working_disks = 0;
4743 int dirty_parity_disks = 0;
4744 struct md_rdev *rdev;
4745 sector_t reshape_offset = 0;
4747 if (mddev->recovery_cp != MaxSector)
4748 printk(KERN_NOTICE "md/raid:%s: not clean"
4749 " -- starting background reconstruction\n",
4751 if (mddev->reshape_position != MaxSector) {
4752 /* Check that we can continue the reshape.
4753 * Currently only disks can change, it must
4754 * increase, and we must be past the point where
4755 * a stripe over-writes itself
4757 sector_t here_new, here_old;
4759 int max_degraded = (mddev->level == 6 ? 2 : 1);
4761 if (mddev->new_level != mddev->level) {
4762 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4763 "required - aborting.\n",
4767 old_disks = mddev->raid_disks - mddev->delta_disks;
4768 /* reshape_position must be on a new-stripe boundary, and one
4769 * further up in new geometry must map after here in old
4772 here_new = mddev->reshape_position;
4773 if (sector_div(here_new, mddev->new_chunk_sectors *
4774 (mddev->raid_disks - max_degraded))) {
4775 printk(KERN_ERR "md/raid:%s: reshape_position not "
4776 "on a stripe boundary\n", mdname(mddev));
4779 reshape_offset = here_new * mddev->new_chunk_sectors;
4780 /* here_new is the stripe we will write to */
4781 here_old = mddev->reshape_position;
4782 sector_div(here_old, mddev->chunk_sectors *
4783 (old_disks-max_degraded));
4784 /* here_old is the first stripe that we might need to read
4786 if (mddev->delta_disks == 0) {
4787 /* We cannot be sure it is safe to start an in-place
4788 * reshape. It is only safe if user-space if monitoring
4789 * and taking constant backups.
4790 * mdadm always starts a situation like this in
4791 * readonly mode so it can take control before
4792 * allowing any writes. So just check for that.
4794 if ((here_new * mddev->new_chunk_sectors !=
4795 here_old * mddev->chunk_sectors) ||
4797 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4798 " in read-only mode - aborting\n",
4802 } else if (mddev->delta_disks < 0
4803 ? (here_new * mddev->new_chunk_sectors <=
4804 here_old * mddev->chunk_sectors)
4805 : (here_new * mddev->new_chunk_sectors >=
4806 here_old * mddev->chunk_sectors)) {
4807 /* Reading from the same stripe as writing to - bad */
4808 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4809 "auto-recovery - aborting.\n",
4813 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4815 /* OK, we should be able to continue; */
4817 BUG_ON(mddev->level != mddev->new_level);
4818 BUG_ON(mddev->layout != mddev->new_layout);
4819 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4820 BUG_ON(mddev->delta_disks != 0);
4823 if (mddev->private == NULL)
4824 conf = setup_conf(mddev);
4826 conf = mddev->private;
4829 return PTR_ERR(conf);
4831 mddev->thread = conf->thread;
4832 conf->thread = NULL;
4833 mddev->private = conf;
4836 * 0 for a fully functional array, 1 or 2 for a degraded array.
4838 list_for_each_entry(rdev, &mddev->disks, same_set) {
4839 if (rdev->raid_disk < 0)
4841 if (test_bit(In_sync, &rdev->flags)) {
4845 /* This disc is not fully in-sync. However if it
4846 * just stored parity (beyond the recovery_offset),
4847 * when we don't need to be concerned about the
4848 * array being dirty.
4849 * When reshape goes 'backwards', we never have
4850 * partially completed devices, so we only need
4851 * to worry about reshape going forwards.
4853 /* Hack because v0.91 doesn't store recovery_offset properly. */
4854 if (mddev->major_version == 0 &&
4855 mddev->minor_version > 90)
4856 rdev->recovery_offset = reshape_offset;
4858 if (rdev->recovery_offset < reshape_offset) {
4859 /* We need to check old and new layout */
4860 if (!only_parity(rdev->raid_disk,
4863 conf->max_degraded))
4866 if (!only_parity(rdev->raid_disk,
4868 conf->previous_raid_disks,
4869 conf->max_degraded))
4871 dirty_parity_disks++;
4874 mddev->degraded = calc_degraded(conf);
4876 if (has_failed(conf)) {
4877 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4878 " (%d/%d failed)\n",
4879 mdname(mddev), mddev->degraded, conf->raid_disks);
4883 /* device size must be a multiple of chunk size */
4884 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4885 mddev->resync_max_sectors = mddev->dev_sectors;
4887 if (mddev->degraded > dirty_parity_disks &&
4888 mddev->recovery_cp != MaxSector) {
4889 if (mddev->ok_start_degraded)
4891 "md/raid:%s: starting dirty degraded array"
4892 " - data corruption possible.\n",
4896 "md/raid:%s: cannot start dirty degraded array.\n",
4902 if (mddev->degraded == 0)
4903 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4904 " devices, algorithm %d\n", mdname(mddev), conf->level,
4905 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4908 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4909 " out of %d devices, algorithm %d\n",
4910 mdname(mddev), conf->level,
4911 mddev->raid_disks - mddev->degraded,
4912 mddev->raid_disks, mddev->new_layout);
4914 print_raid5_conf(conf);
4916 if (conf->reshape_progress != MaxSector) {
4917 conf->reshape_safe = conf->reshape_progress;
4918 atomic_set(&conf->reshape_stripes, 0);
4919 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4920 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4921 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4922 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4923 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4928 /* Ok, everything is just fine now */
4929 if (mddev->to_remove == &raid5_attrs_group)
4930 mddev->to_remove = NULL;
4931 else if (mddev->kobj.sd &&
4932 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4934 "raid5: failed to create sysfs attributes for %s\n",
4936 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4940 /* read-ahead size must cover two whole stripes, which
4941 * is 2 * (datadisks) * chunksize where 'n' is the
4942 * number of raid devices
4944 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4945 int stripe = data_disks *
4946 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4947 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4948 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4950 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4952 mddev->queue->backing_dev_info.congested_data = mddev;
4953 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4955 chunk_size = mddev->chunk_sectors << 9;
4956 blk_queue_io_min(mddev->queue, chunk_size);
4957 blk_queue_io_opt(mddev->queue, chunk_size *
4958 (conf->raid_disks - conf->max_degraded));
4960 list_for_each_entry(rdev, &mddev->disks, same_set)
4961 disk_stack_limits(mddev->gendisk, rdev->bdev,
4962 rdev->data_offset << 9);
4967 md_unregister_thread(&mddev->thread);
4968 print_raid5_conf(conf);
4970 mddev->private = NULL;
4971 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
4975 static int stop(struct mddev *mddev)
4977 struct r5conf *conf = mddev->private;
4979 md_unregister_thread(&mddev->thread);
4981 mddev->queue->backing_dev_info.congested_fn = NULL;
4983 mddev->private = NULL;
4984 mddev->to_remove = &raid5_attrs_group;
4988 static void status(struct seq_file *seq, struct mddev *mddev)
4990 struct r5conf *conf = mddev->private;
4993 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4994 mddev->chunk_sectors / 2, mddev->layout);
4995 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4996 for (i = 0; i < conf->raid_disks; i++)
4997 seq_printf (seq, "%s",
4998 conf->disks[i].rdev &&
4999 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5000 seq_printf (seq, "]");
5003 static void print_raid5_conf (struct r5conf *conf)
5006 struct disk_info *tmp;
5008 printk(KERN_DEBUG "RAID conf printout:\n");
5010 printk("(conf==NULL)\n");
5013 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5015 conf->raid_disks - conf->mddev->degraded);
5017 for (i = 0; i < conf->raid_disks; i++) {
5018 char b[BDEVNAME_SIZE];
5019 tmp = conf->disks + i;
5021 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5022 i, !test_bit(Faulty, &tmp->rdev->flags),
5023 bdevname(tmp->rdev->bdev, b));
5027 static int raid5_spare_active(struct mddev *mddev)
5030 struct r5conf *conf = mddev->private;
5031 struct disk_info *tmp;
5033 unsigned long flags;
5035 for (i = 0; i < conf->raid_disks; i++) {
5036 tmp = conf->disks + i;
5038 && tmp->rdev->recovery_offset == MaxSector
5039 && !test_bit(Faulty, &tmp->rdev->flags)
5040 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5042 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5045 spin_lock_irqsave(&conf->device_lock, flags);
5046 mddev->degraded = calc_degraded(conf);
5047 spin_unlock_irqrestore(&conf->device_lock, flags);
5048 print_raid5_conf(conf);
5052 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5054 struct r5conf *conf = mddev->private;
5056 int number = rdev->raid_disk;
5057 struct disk_info *p = conf->disks + number;
5059 print_raid5_conf(conf);
5060 if (rdev == p->rdev) {
5061 if (number >= conf->raid_disks &&
5062 conf->reshape_progress == MaxSector)
5063 clear_bit(In_sync, &rdev->flags);
5065 if (test_bit(In_sync, &rdev->flags) ||
5066 atomic_read(&rdev->nr_pending)) {
5070 /* Only remove non-faulty devices if recovery
5073 if (!test_bit(Faulty, &rdev->flags) &&
5074 mddev->recovery_disabled != conf->recovery_disabled &&
5075 !has_failed(conf) &&
5076 number < conf->raid_disks) {
5082 if (atomic_read(&rdev->nr_pending)) {
5083 /* lost the race, try later */
5090 print_raid5_conf(conf);
5094 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5096 struct r5conf *conf = mddev->private;
5099 struct disk_info *p;
5101 int last = conf->raid_disks - 1;
5103 if (mddev->recovery_disabled == conf->recovery_disabled)
5106 if (has_failed(conf))
5107 /* no point adding a device */
5110 if (rdev->raid_disk >= 0)
5111 first = last = rdev->raid_disk;
5114 * find the disk ... but prefer rdev->saved_raid_disk
5117 if (rdev->saved_raid_disk >= 0 &&
5118 rdev->saved_raid_disk >= first &&
5119 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5120 disk = rdev->saved_raid_disk;
5123 for ( ; disk <= last ; disk++)
5124 if ((p=conf->disks + disk)->rdev == NULL) {
5125 clear_bit(In_sync, &rdev->flags);
5126 rdev->raid_disk = disk;
5128 if (rdev->saved_raid_disk != disk)
5130 rcu_assign_pointer(p->rdev, rdev);
5133 print_raid5_conf(conf);
5137 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5139 /* no resync is happening, and there is enough space
5140 * on all devices, so we can resize.
5141 * We need to make sure resync covers any new space.
5142 * If the array is shrinking we should possibly wait until
5143 * any io in the removed space completes, but it hardly seems
5146 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5147 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5148 mddev->raid_disks));
5149 if (mddev->array_sectors >
5150 raid5_size(mddev, sectors, mddev->raid_disks))
5152 set_capacity(mddev->gendisk, mddev->array_sectors);
5153 revalidate_disk(mddev->gendisk);
5154 if (sectors > mddev->dev_sectors &&
5155 mddev->recovery_cp > mddev->dev_sectors) {
5156 mddev->recovery_cp = mddev->dev_sectors;
5157 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5159 mddev->dev_sectors = sectors;
5160 mddev->resync_max_sectors = sectors;
5164 static int check_stripe_cache(struct mddev *mddev)
5166 /* Can only proceed if there are plenty of stripe_heads.
5167 * We need a minimum of one full stripe,, and for sensible progress
5168 * it is best to have about 4 times that.
5169 * If we require 4 times, then the default 256 4K stripe_heads will
5170 * allow for chunk sizes up to 256K, which is probably OK.
5171 * If the chunk size is greater, user-space should request more
5172 * stripe_heads first.
5174 struct r5conf *conf = mddev->private;
5175 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5176 > conf->max_nr_stripes ||
5177 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5178 > conf->max_nr_stripes) {
5179 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5181 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5188 static int check_reshape(struct mddev *mddev)
5190 struct r5conf *conf = mddev->private;
5192 if (mddev->delta_disks == 0 &&
5193 mddev->new_layout == mddev->layout &&
5194 mddev->new_chunk_sectors == mddev->chunk_sectors)
5195 return 0; /* nothing to do */
5197 /* Cannot grow a bitmap yet */
5199 if (has_failed(conf))
5201 if (mddev->delta_disks < 0) {
5202 /* We might be able to shrink, but the devices must
5203 * be made bigger first.
5204 * For raid6, 4 is the minimum size.
5205 * Otherwise 2 is the minimum
5208 if (mddev->level == 6)
5210 if (mddev->raid_disks + mddev->delta_disks < min)
5214 if (!check_stripe_cache(mddev))
5217 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5220 static int raid5_start_reshape(struct mddev *mddev)
5222 struct r5conf *conf = mddev->private;
5223 struct md_rdev *rdev;
5225 unsigned long flags;
5227 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5230 if (!check_stripe_cache(mddev))
5233 list_for_each_entry(rdev, &mddev->disks, same_set)
5234 if (!test_bit(In_sync, &rdev->flags)
5235 && !test_bit(Faulty, &rdev->flags))
5238 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5239 /* Not enough devices even to make a degraded array
5244 /* Refuse to reduce size of the array. Any reductions in
5245 * array size must be through explicit setting of array_size
5248 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5249 < mddev->array_sectors) {
5250 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5251 "before number of disks\n", mdname(mddev));
5255 atomic_set(&conf->reshape_stripes, 0);
5256 spin_lock_irq(&conf->device_lock);
5257 conf->previous_raid_disks = conf->raid_disks;
5258 conf->raid_disks += mddev->delta_disks;
5259 conf->prev_chunk_sectors = conf->chunk_sectors;
5260 conf->chunk_sectors = mddev->new_chunk_sectors;
5261 conf->prev_algo = conf->algorithm;
5262 conf->algorithm = mddev->new_layout;
5263 if (mddev->delta_disks < 0)
5264 conf->reshape_progress = raid5_size(mddev, 0, 0);
5266 conf->reshape_progress = 0;
5267 conf->reshape_safe = conf->reshape_progress;
5269 spin_unlock_irq(&conf->device_lock);
5271 /* Add some new drives, as many as will fit.
5272 * We know there are enough to make the newly sized array work.
5273 * Don't add devices if we are reducing the number of
5274 * devices in the array. This is because it is not possible
5275 * to correctly record the "partially reconstructed" state of
5276 * such devices during the reshape and confusion could result.
5278 if (mddev->delta_disks >= 0) {
5279 int added_devices = 0;
5280 list_for_each_entry(rdev, &mddev->disks, same_set)
5281 if (rdev->raid_disk < 0 &&
5282 !test_bit(Faulty, &rdev->flags)) {
5283 if (raid5_add_disk(mddev, rdev) == 0) {
5285 >= conf->previous_raid_disks) {
5286 set_bit(In_sync, &rdev->flags);
5289 rdev->recovery_offset = 0;
5291 if (sysfs_link_rdev(mddev, rdev))
5292 /* Failure here is OK */;
5294 } else if (rdev->raid_disk >= conf->previous_raid_disks
5295 && !test_bit(Faulty, &rdev->flags)) {
5296 /* This is a spare that was manually added */
5297 set_bit(In_sync, &rdev->flags);
5301 /* When a reshape changes the number of devices,
5302 * ->degraded is measured against the larger of the
5303 * pre and post number of devices.
5305 spin_lock_irqsave(&conf->device_lock, flags);
5306 mddev->degraded = calc_degraded(conf);
5307 spin_unlock_irqrestore(&conf->device_lock, flags);
5309 mddev->raid_disks = conf->raid_disks;
5310 mddev->reshape_position = conf->reshape_progress;
5311 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5313 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5314 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5315 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5316 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5317 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5319 if (!mddev->sync_thread) {
5320 mddev->recovery = 0;
5321 spin_lock_irq(&conf->device_lock);
5322 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5323 conf->reshape_progress = MaxSector;
5324 spin_unlock_irq(&conf->device_lock);
5327 conf->reshape_checkpoint = jiffies;
5328 md_wakeup_thread(mddev->sync_thread);
5329 md_new_event(mddev);
5333 /* This is called from the reshape thread and should make any
5334 * changes needed in 'conf'
5336 static void end_reshape(struct r5conf *conf)
5339 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5341 spin_lock_irq(&conf->device_lock);
5342 conf->previous_raid_disks = conf->raid_disks;
5343 conf->reshape_progress = MaxSector;
5344 spin_unlock_irq(&conf->device_lock);
5345 wake_up(&conf->wait_for_overlap);
5347 /* read-ahead size must cover two whole stripes, which is
5348 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5350 if (conf->mddev->queue) {
5351 int data_disks = conf->raid_disks - conf->max_degraded;
5352 int stripe = data_disks * ((conf->chunk_sectors << 9)
5354 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5355 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5360 /* This is called from the raid5d thread with mddev_lock held.
5361 * It makes config changes to the device.
5363 static void raid5_finish_reshape(struct mddev *mddev)
5365 struct r5conf *conf = mddev->private;
5367 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5369 if (mddev->delta_disks > 0) {
5370 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5371 set_capacity(mddev->gendisk, mddev->array_sectors);
5372 revalidate_disk(mddev->gendisk);
5375 spin_lock_irq(&conf->device_lock);
5376 mddev->degraded = calc_degraded(conf);
5377 spin_unlock_irq(&conf->device_lock);
5378 for (d = conf->raid_disks ;
5379 d < conf->raid_disks - mddev->delta_disks;
5381 struct md_rdev *rdev = conf->disks[d].rdev;
5383 raid5_remove_disk(mddev, rdev) == 0) {
5384 sysfs_unlink_rdev(mddev, rdev);
5385 rdev->raid_disk = -1;
5389 mddev->layout = conf->algorithm;
5390 mddev->chunk_sectors = conf->chunk_sectors;
5391 mddev->reshape_position = MaxSector;
5392 mddev->delta_disks = 0;
5396 static void raid5_quiesce(struct mddev *mddev, int state)
5398 struct r5conf *conf = mddev->private;
5401 case 2: /* resume for a suspend */
5402 wake_up(&conf->wait_for_overlap);
5405 case 1: /* stop all writes */
5406 spin_lock_irq(&conf->device_lock);
5407 /* '2' tells resync/reshape to pause so that all
5408 * active stripes can drain
5411 wait_event_lock_irq(conf->wait_for_stripe,
5412 atomic_read(&conf->active_stripes) == 0 &&
5413 atomic_read(&conf->active_aligned_reads) == 0,
5414 conf->device_lock, /* nothing */);
5416 spin_unlock_irq(&conf->device_lock);
5417 /* allow reshape to continue */
5418 wake_up(&conf->wait_for_overlap);
5421 case 0: /* re-enable writes */
5422 spin_lock_irq(&conf->device_lock);
5424 wake_up(&conf->wait_for_stripe);
5425 wake_up(&conf->wait_for_overlap);
5426 spin_unlock_irq(&conf->device_lock);
5432 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5434 struct r0conf *raid0_conf = mddev->private;
5437 /* for raid0 takeover only one zone is supported */
5438 if (raid0_conf->nr_strip_zones > 1) {
5439 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5441 return ERR_PTR(-EINVAL);
5444 sectors = raid0_conf->strip_zone[0].zone_end;
5445 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5446 mddev->dev_sectors = sectors;
5447 mddev->new_level = level;
5448 mddev->new_layout = ALGORITHM_PARITY_N;
5449 mddev->new_chunk_sectors = mddev->chunk_sectors;
5450 mddev->raid_disks += 1;
5451 mddev->delta_disks = 1;
5452 /* make sure it will be not marked as dirty */
5453 mddev->recovery_cp = MaxSector;
5455 return setup_conf(mddev);
5459 static void *raid5_takeover_raid1(struct mddev *mddev)
5463 if (mddev->raid_disks != 2 ||
5464 mddev->degraded > 1)
5465 return ERR_PTR(-EINVAL);
5467 /* Should check if there are write-behind devices? */
5469 chunksect = 64*2; /* 64K by default */
5471 /* The array must be an exact multiple of chunksize */
5472 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5475 if ((chunksect<<9) < STRIPE_SIZE)
5476 /* array size does not allow a suitable chunk size */
5477 return ERR_PTR(-EINVAL);
5479 mddev->new_level = 5;
5480 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5481 mddev->new_chunk_sectors = chunksect;
5483 return setup_conf(mddev);
5486 static void *raid5_takeover_raid6(struct mddev *mddev)
5490 switch (mddev->layout) {
5491 case ALGORITHM_LEFT_ASYMMETRIC_6:
5492 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5494 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5495 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5497 case ALGORITHM_LEFT_SYMMETRIC_6:
5498 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5500 case ALGORITHM_RIGHT_SYMMETRIC_6:
5501 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5503 case ALGORITHM_PARITY_0_6:
5504 new_layout = ALGORITHM_PARITY_0;
5506 case ALGORITHM_PARITY_N:
5507 new_layout = ALGORITHM_PARITY_N;
5510 return ERR_PTR(-EINVAL);
5512 mddev->new_level = 5;
5513 mddev->new_layout = new_layout;
5514 mddev->delta_disks = -1;
5515 mddev->raid_disks -= 1;
5516 return setup_conf(mddev);
5520 static int raid5_check_reshape(struct mddev *mddev)
5522 /* For a 2-drive array, the layout and chunk size can be changed
5523 * immediately as not restriping is needed.
5524 * For larger arrays we record the new value - after validation
5525 * to be used by a reshape pass.
5527 struct r5conf *conf = mddev->private;
5528 int new_chunk = mddev->new_chunk_sectors;
5530 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5532 if (new_chunk > 0) {
5533 if (!is_power_of_2(new_chunk))
5535 if (new_chunk < (PAGE_SIZE>>9))
5537 if (mddev->array_sectors & (new_chunk-1))
5538 /* not factor of array size */
5542 /* They look valid */
5544 if (mddev->raid_disks == 2) {
5545 /* can make the change immediately */
5546 if (mddev->new_layout >= 0) {
5547 conf->algorithm = mddev->new_layout;
5548 mddev->layout = mddev->new_layout;
5550 if (new_chunk > 0) {
5551 conf->chunk_sectors = new_chunk ;
5552 mddev->chunk_sectors = new_chunk;
5554 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5555 md_wakeup_thread(mddev->thread);
5557 return check_reshape(mddev);
5560 static int raid6_check_reshape(struct mddev *mddev)
5562 int new_chunk = mddev->new_chunk_sectors;
5564 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5566 if (new_chunk > 0) {
5567 if (!is_power_of_2(new_chunk))
5569 if (new_chunk < (PAGE_SIZE >> 9))
5571 if (mddev->array_sectors & (new_chunk-1))
5572 /* not factor of array size */
5576 /* They look valid */
5577 return check_reshape(mddev);
5580 static void *raid5_takeover(struct mddev *mddev)
5582 /* raid5 can take over:
5583 * raid0 - if there is only one strip zone - make it a raid4 layout
5584 * raid1 - if there are two drives. We need to know the chunk size
5585 * raid4 - trivial - just use a raid4 layout.
5586 * raid6 - Providing it is a *_6 layout
5588 if (mddev->level == 0)
5589 return raid45_takeover_raid0(mddev, 5);
5590 if (mddev->level == 1)
5591 return raid5_takeover_raid1(mddev);
5592 if (mddev->level == 4) {
5593 mddev->new_layout = ALGORITHM_PARITY_N;
5594 mddev->new_level = 5;
5595 return setup_conf(mddev);
5597 if (mddev->level == 6)
5598 return raid5_takeover_raid6(mddev);
5600 return ERR_PTR(-EINVAL);
5603 static void *raid4_takeover(struct mddev *mddev)
5605 /* raid4 can take over:
5606 * raid0 - if there is only one strip zone
5607 * raid5 - if layout is right
5609 if (mddev->level == 0)
5610 return raid45_takeover_raid0(mddev, 4);
5611 if (mddev->level == 5 &&
5612 mddev->layout == ALGORITHM_PARITY_N) {
5613 mddev->new_layout = 0;
5614 mddev->new_level = 4;
5615 return setup_conf(mddev);
5617 return ERR_PTR(-EINVAL);
5620 static struct md_personality raid5_personality;
5622 static void *raid6_takeover(struct mddev *mddev)
5624 /* Currently can only take over a raid5. We map the
5625 * personality to an equivalent raid6 personality
5626 * with the Q block at the end.
5630 if (mddev->pers != &raid5_personality)
5631 return ERR_PTR(-EINVAL);
5632 if (mddev->degraded > 1)
5633 return ERR_PTR(-EINVAL);
5634 if (mddev->raid_disks > 253)
5635 return ERR_PTR(-EINVAL);
5636 if (mddev->raid_disks < 3)
5637 return ERR_PTR(-EINVAL);
5639 switch (mddev->layout) {
5640 case ALGORITHM_LEFT_ASYMMETRIC:
5641 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5643 case ALGORITHM_RIGHT_ASYMMETRIC:
5644 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5646 case ALGORITHM_LEFT_SYMMETRIC:
5647 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5649 case ALGORITHM_RIGHT_SYMMETRIC:
5650 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5652 case ALGORITHM_PARITY_0:
5653 new_layout = ALGORITHM_PARITY_0_6;
5655 case ALGORITHM_PARITY_N:
5656 new_layout = ALGORITHM_PARITY_N;
5659 return ERR_PTR(-EINVAL);
5661 mddev->new_level = 6;
5662 mddev->new_layout = new_layout;
5663 mddev->delta_disks = 1;
5664 mddev->raid_disks += 1;
5665 return setup_conf(mddev);
5669 static struct md_personality raid6_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 = raid6_check_reshape,
5686 .start_reshape = raid5_start_reshape,
5687 .finish_reshape = raid5_finish_reshape,
5688 .quiesce = raid5_quiesce,
5689 .takeover = raid6_takeover,
5691 static struct md_personality raid5_personality =
5695 .owner = THIS_MODULE,
5696 .make_request = make_request,
5700 .error_handler = error,
5701 .hot_add_disk = raid5_add_disk,
5702 .hot_remove_disk= raid5_remove_disk,
5703 .spare_active = raid5_spare_active,
5704 .sync_request = sync_request,
5705 .resize = raid5_resize,
5707 .check_reshape = raid5_check_reshape,
5708 .start_reshape = raid5_start_reshape,
5709 .finish_reshape = raid5_finish_reshape,
5710 .quiesce = raid5_quiesce,
5711 .takeover = raid5_takeover,
5714 static struct md_personality raid4_personality =
5718 .owner = THIS_MODULE,
5719 .make_request = make_request,
5723 .error_handler = error,
5724 .hot_add_disk = raid5_add_disk,
5725 .hot_remove_disk= raid5_remove_disk,
5726 .spare_active = raid5_spare_active,
5727 .sync_request = sync_request,
5728 .resize = raid5_resize,
5730 .check_reshape = raid5_check_reshape,
5731 .start_reshape = raid5_start_reshape,
5732 .finish_reshape = raid5_finish_reshape,
5733 .quiesce = raid5_quiesce,
5734 .takeover = raid4_takeover,
5737 static int __init raid5_init(void)
5739 register_md_personality(&raid6_personality);
5740 register_md_personality(&raid5_personality);
5741 register_md_personality(&raid4_personality);
5745 static void raid5_exit(void)
5747 unregister_md_personality(&raid6_personality);
5748 unregister_md_personality(&raid5_personality);
5749 unregister_md_personality(&raid4_personality);
5752 module_init(raid5_init);
5753 module_exit(raid5_exit);
5754 MODULE_LICENSE("GPL");
5755 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5756 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5757 MODULE_ALIAS("md-raid5");
5758 MODULE_ALIAS("md-raid4");
5759 MODULE_ALIAS("md-level-5");
5760 MODULE_ALIAS("md-level-4");
5761 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5762 MODULE_ALIAS("md-raid6");
5763 MODULE_ALIAS("md-level-6");
5765 /* This used to be two separate modules, they were: */
5766 MODULE_ALIAS("raid5");
5767 MODULE_ALIAS("raid6");