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 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
201 list_add_tail(&sh->lru, &conf->delayed_list);
202 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
203 sh->bm_seq - conf->seq_write > 0)
204 list_add_tail(&sh->lru, &conf->bitmap_list);
206 clear_bit(STRIPE_DELAYED, &sh->state);
207 clear_bit(STRIPE_BIT_DELAY, &sh->state);
208 list_add_tail(&sh->lru, &conf->handle_list);
210 md_wakeup_thread(conf->mddev->thread);
212 BUG_ON(stripe_operations_active(sh));
213 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
214 atomic_dec(&conf->preread_active_stripes);
215 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
216 md_wakeup_thread(conf->mddev->thread);
218 atomic_dec(&conf->active_stripes);
219 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
220 list_add_tail(&sh->lru, &conf->inactive_list);
221 wake_up(&conf->wait_for_stripe);
222 if (conf->retry_read_aligned)
223 md_wakeup_thread(conf->mddev->thread);
229 static void release_stripe(struct stripe_head *sh)
231 struct r5conf *conf = sh->raid_conf;
234 spin_lock_irqsave(&conf->device_lock, flags);
235 __release_stripe(conf, sh);
236 spin_unlock_irqrestore(&conf->device_lock, flags);
239 static inline void remove_hash(struct stripe_head *sh)
241 pr_debug("remove_hash(), stripe %llu\n",
242 (unsigned long long)sh->sector);
244 hlist_del_init(&sh->hash);
247 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
249 struct hlist_head *hp = stripe_hash(conf, sh->sector);
251 pr_debug("insert_hash(), stripe %llu\n",
252 (unsigned long long)sh->sector);
254 hlist_add_head(&sh->hash, hp);
258 /* find an idle stripe, make sure it is unhashed, and return it. */
259 static struct stripe_head *get_free_stripe(struct r5conf *conf)
261 struct stripe_head *sh = NULL;
262 struct list_head *first;
264 if (list_empty(&conf->inactive_list))
266 first = conf->inactive_list.next;
267 sh = list_entry(first, struct stripe_head, lru);
268 list_del_init(first);
270 atomic_inc(&conf->active_stripes);
275 static void shrink_buffers(struct stripe_head *sh)
279 int num = sh->raid_conf->pool_size;
281 for (i = 0; i < num ; i++) {
285 sh->dev[i].page = NULL;
290 static int grow_buffers(struct stripe_head *sh)
293 int num = sh->raid_conf->pool_size;
295 for (i = 0; i < num; i++) {
298 if (!(page = alloc_page(GFP_KERNEL))) {
301 sh->dev[i].page = page;
306 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
307 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
308 struct stripe_head *sh);
310 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
312 struct r5conf *conf = sh->raid_conf;
315 BUG_ON(atomic_read(&sh->count) != 0);
316 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
317 BUG_ON(stripe_operations_active(sh));
319 pr_debug("init_stripe called, stripe %llu\n",
320 (unsigned long long)sh->sector);
324 sh->generation = conf->generation - previous;
325 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
327 stripe_set_idx(sector, conf, previous, sh);
331 for (i = sh->disks; i--; ) {
332 struct r5dev *dev = &sh->dev[i];
334 if (dev->toread || dev->read || dev->towrite || dev->written ||
335 test_bit(R5_LOCKED, &dev->flags)) {
336 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
337 (unsigned long long)sh->sector, i, dev->toread,
338 dev->read, dev->towrite, dev->written,
339 test_bit(R5_LOCKED, &dev->flags));
343 raid5_build_block(sh, i, previous);
345 insert_hash(conf, sh);
348 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
351 struct stripe_head *sh;
352 struct hlist_node *hn;
354 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
355 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
356 if (sh->sector == sector && sh->generation == generation)
358 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
363 * Need to check if array has failed when deciding whether to:
365 * - remove non-faulty devices
368 * This determination is simple when no reshape is happening.
369 * However if there is a reshape, we need to carefully check
370 * both the before and after sections.
371 * This is because some failed devices may only affect one
372 * of the two sections, and some non-in_sync devices may
373 * be insync in the section most affected by failed devices.
375 static int has_failed(struct r5conf *conf)
379 if (conf->mddev->reshape_position == MaxSector)
380 return conf->mddev->degraded > conf->max_degraded;
384 for (i = 0; i < conf->previous_raid_disks; i++) {
385 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
386 if (!rdev || test_bit(Faulty, &rdev->flags))
388 else if (test_bit(In_sync, &rdev->flags))
391 /* not in-sync or faulty.
392 * If the reshape increases the number of devices,
393 * this is being recovered by the reshape, so
394 * this 'previous' section is not in_sync.
395 * If the number of devices is being reduced however,
396 * the device can only be part of the array if
397 * we are reverting a reshape, so this section will
400 if (conf->raid_disks >= conf->previous_raid_disks)
404 if (degraded > conf->max_degraded)
408 for (i = 0; i < conf->raid_disks; i++) {
409 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
410 if (!rdev || test_bit(Faulty, &rdev->flags))
412 else if (test_bit(In_sync, &rdev->flags))
415 /* not in-sync or faulty.
416 * If reshape increases the number of devices, this
417 * section has already been recovered, else it
418 * almost certainly hasn't.
420 if (conf->raid_disks <= conf->previous_raid_disks)
424 if (degraded > conf->max_degraded)
429 static struct stripe_head *
430 get_active_stripe(struct r5conf *conf, sector_t sector,
431 int previous, int noblock, int noquiesce)
433 struct stripe_head *sh;
435 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
437 spin_lock_irq(&conf->device_lock);
440 wait_event_lock_irq(conf->wait_for_stripe,
441 conf->quiesce == 0 || noquiesce,
442 conf->device_lock, /* nothing */);
443 sh = __find_stripe(conf, sector, conf->generation - previous);
445 if (!conf->inactive_blocked)
446 sh = get_free_stripe(conf);
447 if (noblock && sh == NULL)
450 conf->inactive_blocked = 1;
451 wait_event_lock_irq(conf->wait_for_stripe,
452 !list_empty(&conf->inactive_list) &&
453 (atomic_read(&conf->active_stripes)
454 < (conf->max_nr_stripes *3/4)
455 || !conf->inactive_blocked),
458 conf->inactive_blocked = 0;
460 init_stripe(sh, sector, previous);
462 if (atomic_read(&sh->count)) {
463 BUG_ON(!list_empty(&sh->lru)
464 && !test_bit(STRIPE_EXPANDING, &sh->state));
466 if (!test_bit(STRIPE_HANDLE, &sh->state))
467 atomic_inc(&conf->active_stripes);
468 if (list_empty(&sh->lru) &&
469 !test_bit(STRIPE_EXPANDING, &sh->state))
471 list_del_init(&sh->lru);
474 } while (sh == NULL);
477 atomic_inc(&sh->count);
479 spin_unlock_irq(&conf->device_lock);
484 raid5_end_read_request(struct bio *bi, int error);
486 raid5_end_write_request(struct bio *bi, int error);
488 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
490 struct r5conf *conf = sh->raid_conf;
491 int i, disks = sh->disks;
495 for (i = disks; i--; ) {
498 struct md_rdev *rdev;
499 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
500 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
504 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
509 bi = &sh->dev[i].req;
513 bi->bi_end_io = raid5_end_write_request;
515 bi->bi_end_io = raid5_end_read_request;
518 rdev = rcu_dereference(conf->disks[i].rdev);
519 if (rdev && test_bit(Faulty, &rdev->flags))
522 atomic_inc(&rdev->nr_pending);
525 /* We have already checked bad blocks for reads. Now
526 * need to check for writes.
528 while ((rw & WRITE) && rdev &&
529 test_bit(WriteErrorSeen, &rdev->flags)) {
532 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
533 &first_bad, &bad_sectors);
538 set_bit(BlockedBadBlocks, &rdev->flags);
539 if (!conf->mddev->external &&
540 conf->mddev->flags) {
541 /* It is very unlikely, but we might
542 * still need to write out the
543 * bad block log - better give it
545 md_check_recovery(conf->mddev);
548 * Because md_wait_for_blocked_rdev
549 * will dec nr_pending, we must
550 * increment it first.
552 atomic_inc(&rdev->nr_pending);
553 md_wait_for_blocked_rdev(rdev, conf->mddev);
555 /* Acknowledged bad block - skip the write */
556 rdev_dec_pending(rdev, conf->mddev);
562 if (s->syncing || s->expanding || s->expanded)
563 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
565 set_bit(STRIPE_IO_STARTED, &sh->state);
567 bi->bi_bdev = rdev->bdev;
568 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
569 __func__, (unsigned long long)sh->sector,
571 atomic_inc(&sh->count);
572 bi->bi_sector = sh->sector + rdev->data_offset;
573 bi->bi_flags = 1 << BIO_UPTODATE;
577 bi->bi_io_vec = &sh->dev[i].vec;
578 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
579 bi->bi_io_vec[0].bv_offset = 0;
580 bi->bi_size = STRIPE_SIZE;
582 generic_make_request(bi);
585 set_bit(STRIPE_DEGRADED, &sh->state);
586 pr_debug("skip op %ld on disc %d for sector %llu\n",
587 bi->bi_rw, i, (unsigned long long)sh->sector);
588 clear_bit(R5_LOCKED, &sh->dev[i].flags);
589 set_bit(STRIPE_HANDLE, &sh->state);
594 static struct dma_async_tx_descriptor *
595 async_copy_data(int frombio, struct bio *bio, struct page *page,
596 sector_t sector, struct dma_async_tx_descriptor *tx)
599 struct page *bio_page;
602 struct async_submit_ctl submit;
603 enum async_tx_flags flags = 0;
605 if (bio->bi_sector >= sector)
606 page_offset = (signed)(bio->bi_sector - sector) * 512;
608 page_offset = (signed)(sector - bio->bi_sector) * -512;
611 flags |= ASYNC_TX_FENCE;
612 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
614 bio_for_each_segment(bvl, bio, i) {
615 int len = bvl->bv_len;
619 if (page_offset < 0) {
620 b_offset = -page_offset;
621 page_offset += b_offset;
625 if (len > 0 && page_offset + len > STRIPE_SIZE)
626 clen = STRIPE_SIZE - page_offset;
631 b_offset += bvl->bv_offset;
632 bio_page = bvl->bv_page;
634 tx = async_memcpy(page, bio_page, page_offset,
635 b_offset, clen, &submit);
637 tx = async_memcpy(bio_page, page, b_offset,
638 page_offset, clen, &submit);
640 /* chain the operations */
641 submit.depend_tx = tx;
643 if (clen < len) /* hit end of page */
651 static void ops_complete_biofill(void *stripe_head_ref)
653 struct stripe_head *sh = stripe_head_ref;
654 struct bio *return_bi = NULL;
655 struct r5conf *conf = sh->raid_conf;
658 pr_debug("%s: stripe %llu\n", __func__,
659 (unsigned long long)sh->sector);
661 /* clear completed biofills */
662 spin_lock_irq(&conf->device_lock);
663 for (i = sh->disks; i--; ) {
664 struct r5dev *dev = &sh->dev[i];
666 /* acknowledge completion of a biofill operation */
667 /* and check if we need to reply to a read request,
668 * new R5_Wantfill requests are held off until
669 * !STRIPE_BIOFILL_RUN
671 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
672 struct bio *rbi, *rbi2;
677 while (rbi && rbi->bi_sector <
678 dev->sector + STRIPE_SECTORS) {
679 rbi2 = r5_next_bio(rbi, dev->sector);
680 if (!raid5_dec_bi_phys_segments(rbi)) {
681 rbi->bi_next = return_bi;
688 spin_unlock_irq(&conf->device_lock);
689 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
691 return_io(return_bi);
693 set_bit(STRIPE_HANDLE, &sh->state);
697 static void ops_run_biofill(struct stripe_head *sh)
699 struct dma_async_tx_descriptor *tx = NULL;
700 struct r5conf *conf = sh->raid_conf;
701 struct async_submit_ctl submit;
704 pr_debug("%s: stripe %llu\n", __func__,
705 (unsigned long long)sh->sector);
707 for (i = sh->disks; i--; ) {
708 struct r5dev *dev = &sh->dev[i];
709 if (test_bit(R5_Wantfill, &dev->flags)) {
711 spin_lock_irq(&conf->device_lock);
712 dev->read = rbi = dev->toread;
714 spin_unlock_irq(&conf->device_lock);
715 while (rbi && rbi->bi_sector <
716 dev->sector + STRIPE_SECTORS) {
717 tx = async_copy_data(0, rbi, dev->page,
719 rbi = r5_next_bio(rbi, dev->sector);
724 atomic_inc(&sh->count);
725 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
726 async_trigger_callback(&submit);
729 static void mark_target_uptodate(struct stripe_head *sh, int target)
736 tgt = &sh->dev[target];
737 set_bit(R5_UPTODATE, &tgt->flags);
738 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
739 clear_bit(R5_Wantcompute, &tgt->flags);
742 static void ops_complete_compute(void *stripe_head_ref)
744 struct stripe_head *sh = stripe_head_ref;
746 pr_debug("%s: stripe %llu\n", __func__,
747 (unsigned long long)sh->sector);
749 /* mark the computed target(s) as uptodate */
750 mark_target_uptodate(sh, sh->ops.target);
751 mark_target_uptodate(sh, sh->ops.target2);
753 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
754 if (sh->check_state == check_state_compute_run)
755 sh->check_state = check_state_compute_result;
756 set_bit(STRIPE_HANDLE, &sh->state);
760 /* return a pointer to the address conversion region of the scribble buffer */
761 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
762 struct raid5_percpu *percpu)
764 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
767 static struct dma_async_tx_descriptor *
768 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
770 int disks = sh->disks;
771 struct page **xor_srcs = percpu->scribble;
772 int target = sh->ops.target;
773 struct r5dev *tgt = &sh->dev[target];
774 struct page *xor_dest = tgt->page;
776 struct dma_async_tx_descriptor *tx;
777 struct async_submit_ctl submit;
780 pr_debug("%s: stripe %llu block: %d\n",
781 __func__, (unsigned long long)sh->sector, target);
782 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
784 for (i = disks; i--; )
786 xor_srcs[count++] = sh->dev[i].page;
788 atomic_inc(&sh->count);
790 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
791 ops_complete_compute, sh, to_addr_conv(sh, percpu));
792 if (unlikely(count == 1))
793 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
795 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
800 /* set_syndrome_sources - populate source buffers for gen_syndrome
801 * @srcs - (struct page *) array of size sh->disks
802 * @sh - stripe_head to parse
804 * Populates srcs in proper layout order for the stripe and returns the
805 * 'count' of sources to be used in a call to async_gen_syndrome. The P
806 * destination buffer is recorded in srcs[count] and the Q destination
807 * is recorded in srcs[count+1]].
809 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
811 int disks = sh->disks;
812 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
813 int d0_idx = raid6_d0(sh);
817 for (i = 0; i < disks; i++)
823 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
825 srcs[slot] = sh->dev[i].page;
826 i = raid6_next_disk(i, disks);
827 } while (i != d0_idx);
829 return syndrome_disks;
832 static struct dma_async_tx_descriptor *
833 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
835 int disks = sh->disks;
836 struct page **blocks = percpu->scribble;
838 int qd_idx = sh->qd_idx;
839 struct dma_async_tx_descriptor *tx;
840 struct async_submit_ctl submit;
846 if (sh->ops.target < 0)
847 target = sh->ops.target2;
848 else if (sh->ops.target2 < 0)
849 target = sh->ops.target;
851 /* we should only have one valid target */
854 pr_debug("%s: stripe %llu block: %d\n",
855 __func__, (unsigned long long)sh->sector, target);
857 tgt = &sh->dev[target];
858 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
861 atomic_inc(&sh->count);
863 if (target == qd_idx) {
864 count = set_syndrome_sources(blocks, sh);
865 blocks[count] = NULL; /* regenerating p is not necessary */
866 BUG_ON(blocks[count+1] != dest); /* q should already be set */
867 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
868 ops_complete_compute, sh,
869 to_addr_conv(sh, percpu));
870 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
872 /* Compute any data- or p-drive using XOR */
874 for (i = disks; i-- ; ) {
875 if (i == target || i == qd_idx)
877 blocks[count++] = sh->dev[i].page;
880 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
881 NULL, ops_complete_compute, sh,
882 to_addr_conv(sh, percpu));
883 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
889 static struct dma_async_tx_descriptor *
890 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
892 int i, count, disks = sh->disks;
893 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
894 int d0_idx = raid6_d0(sh);
895 int faila = -1, failb = -1;
896 int target = sh->ops.target;
897 int target2 = sh->ops.target2;
898 struct r5dev *tgt = &sh->dev[target];
899 struct r5dev *tgt2 = &sh->dev[target2];
900 struct dma_async_tx_descriptor *tx;
901 struct page **blocks = percpu->scribble;
902 struct async_submit_ctl submit;
904 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
905 __func__, (unsigned long long)sh->sector, target, target2);
906 BUG_ON(target < 0 || target2 < 0);
907 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
908 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
910 /* we need to open-code set_syndrome_sources to handle the
911 * slot number conversion for 'faila' and 'failb'
913 for (i = 0; i < disks ; i++)
918 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
920 blocks[slot] = sh->dev[i].page;
926 i = raid6_next_disk(i, disks);
927 } while (i != d0_idx);
929 BUG_ON(faila == failb);
932 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
933 __func__, (unsigned long long)sh->sector, faila, failb);
935 atomic_inc(&sh->count);
937 if (failb == syndrome_disks+1) {
938 /* Q disk is one of the missing disks */
939 if (faila == syndrome_disks) {
940 /* Missing P+Q, just recompute */
941 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
942 ops_complete_compute, sh,
943 to_addr_conv(sh, percpu));
944 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
945 STRIPE_SIZE, &submit);
949 int qd_idx = sh->qd_idx;
951 /* Missing D+Q: recompute D from P, then recompute Q */
952 if (target == qd_idx)
953 data_target = target2;
955 data_target = target;
958 for (i = disks; i-- ; ) {
959 if (i == data_target || i == qd_idx)
961 blocks[count++] = sh->dev[i].page;
963 dest = sh->dev[data_target].page;
964 init_async_submit(&submit,
965 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
967 to_addr_conv(sh, percpu));
968 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
971 count = set_syndrome_sources(blocks, sh);
972 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
973 ops_complete_compute, sh,
974 to_addr_conv(sh, percpu));
975 return async_gen_syndrome(blocks, 0, count+2,
976 STRIPE_SIZE, &submit);
979 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
980 ops_complete_compute, sh,
981 to_addr_conv(sh, percpu));
982 if (failb == syndrome_disks) {
983 /* We're missing D+P. */
984 return async_raid6_datap_recov(syndrome_disks+2,
988 /* We're missing D+D. */
989 return async_raid6_2data_recov(syndrome_disks+2,
990 STRIPE_SIZE, faila, failb,
997 static void ops_complete_prexor(void *stripe_head_ref)
999 struct stripe_head *sh = stripe_head_ref;
1001 pr_debug("%s: stripe %llu\n", __func__,
1002 (unsigned long long)sh->sector);
1005 static struct dma_async_tx_descriptor *
1006 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1007 struct dma_async_tx_descriptor *tx)
1009 int disks = sh->disks;
1010 struct page **xor_srcs = percpu->scribble;
1011 int count = 0, pd_idx = sh->pd_idx, i;
1012 struct async_submit_ctl submit;
1014 /* existing parity data subtracted */
1015 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1017 pr_debug("%s: stripe %llu\n", __func__,
1018 (unsigned long long)sh->sector);
1020 for (i = disks; i--; ) {
1021 struct r5dev *dev = &sh->dev[i];
1022 /* Only process blocks that are known to be uptodate */
1023 if (test_bit(R5_Wantdrain, &dev->flags))
1024 xor_srcs[count++] = dev->page;
1027 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1028 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1029 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1034 static struct dma_async_tx_descriptor *
1035 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1037 int disks = sh->disks;
1040 pr_debug("%s: stripe %llu\n", __func__,
1041 (unsigned long long)sh->sector);
1043 for (i = disks; i--; ) {
1044 struct r5dev *dev = &sh->dev[i];
1047 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1050 spin_lock_irq(&sh->raid_conf->device_lock);
1051 chosen = dev->towrite;
1052 dev->towrite = NULL;
1053 BUG_ON(dev->written);
1054 wbi = dev->written = chosen;
1055 spin_unlock_irq(&sh->raid_conf->device_lock);
1057 while (wbi && wbi->bi_sector <
1058 dev->sector + STRIPE_SECTORS) {
1059 if (wbi->bi_rw & REQ_FUA)
1060 set_bit(R5_WantFUA, &dev->flags);
1061 tx = async_copy_data(1, wbi, dev->page,
1063 wbi = r5_next_bio(wbi, dev->sector);
1071 static void ops_complete_reconstruct(void *stripe_head_ref)
1073 struct stripe_head *sh = stripe_head_ref;
1074 int disks = sh->disks;
1075 int pd_idx = sh->pd_idx;
1076 int qd_idx = sh->qd_idx;
1080 pr_debug("%s: stripe %llu\n", __func__,
1081 (unsigned long long)sh->sector);
1083 for (i = disks; i--; )
1084 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1086 for (i = disks; i--; ) {
1087 struct r5dev *dev = &sh->dev[i];
1089 if (dev->written || i == pd_idx || i == qd_idx) {
1090 set_bit(R5_UPTODATE, &dev->flags);
1092 set_bit(R5_WantFUA, &dev->flags);
1096 if (sh->reconstruct_state == reconstruct_state_drain_run)
1097 sh->reconstruct_state = reconstruct_state_drain_result;
1098 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1099 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1101 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1102 sh->reconstruct_state = reconstruct_state_result;
1105 set_bit(STRIPE_HANDLE, &sh->state);
1110 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1111 struct dma_async_tx_descriptor *tx)
1113 int disks = sh->disks;
1114 struct page **xor_srcs = percpu->scribble;
1115 struct async_submit_ctl submit;
1116 int count = 0, pd_idx = sh->pd_idx, i;
1117 struct page *xor_dest;
1119 unsigned long flags;
1121 pr_debug("%s: stripe %llu\n", __func__,
1122 (unsigned long long)sh->sector);
1124 /* check if prexor is active which means only process blocks
1125 * that are part of a read-modify-write (written)
1127 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1129 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1130 for (i = disks; i--; ) {
1131 struct r5dev *dev = &sh->dev[i];
1133 xor_srcs[count++] = dev->page;
1136 xor_dest = sh->dev[pd_idx].page;
1137 for (i = disks; i--; ) {
1138 struct r5dev *dev = &sh->dev[i];
1140 xor_srcs[count++] = dev->page;
1144 /* 1/ if we prexor'd then the dest is reused as a source
1145 * 2/ if we did not prexor then we are redoing the parity
1146 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1147 * for the synchronous xor case
1149 flags = ASYNC_TX_ACK |
1150 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1152 atomic_inc(&sh->count);
1154 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1155 to_addr_conv(sh, percpu));
1156 if (unlikely(count == 1))
1157 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1159 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1163 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1164 struct dma_async_tx_descriptor *tx)
1166 struct async_submit_ctl submit;
1167 struct page **blocks = percpu->scribble;
1170 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1172 count = set_syndrome_sources(blocks, sh);
1174 atomic_inc(&sh->count);
1176 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1177 sh, to_addr_conv(sh, percpu));
1178 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1181 static void ops_complete_check(void *stripe_head_ref)
1183 struct stripe_head *sh = stripe_head_ref;
1185 pr_debug("%s: stripe %llu\n", __func__,
1186 (unsigned long long)sh->sector);
1188 sh->check_state = check_state_check_result;
1189 set_bit(STRIPE_HANDLE, &sh->state);
1193 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1195 int disks = sh->disks;
1196 int pd_idx = sh->pd_idx;
1197 int qd_idx = sh->qd_idx;
1198 struct page *xor_dest;
1199 struct page **xor_srcs = percpu->scribble;
1200 struct dma_async_tx_descriptor *tx;
1201 struct async_submit_ctl submit;
1205 pr_debug("%s: stripe %llu\n", __func__,
1206 (unsigned long long)sh->sector);
1209 xor_dest = sh->dev[pd_idx].page;
1210 xor_srcs[count++] = xor_dest;
1211 for (i = disks; i--; ) {
1212 if (i == pd_idx || i == qd_idx)
1214 xor_srcs[count++] = sh->dev[i].page;
1217 init_async_submit(&submit, 0, NULL, NULL, NULL,
1218 to_addr_conv(sh, percpu));
1219 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1220 &sh->ops.zero_sum_result, &submit);
1222 atomic_inc(&sh->count);
1223 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1224 tx = async_trigger_callback(&submit);
1227 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1229 struct page **srcs = percpu->scribble;
1230 struct async_submit_ctl submit;
1233 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1234 (unsigned long long)sh->sector, checkp);
1236 count = set_syndrome_sources(srcs, sh);
1240 atomic_inc(&sh->count);
1241 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1242 sh, to_addr_conv(sh, percpu));
1243 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1244 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1247 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1249 int overlap_clear = 0, i, disks = sh->disks;
1250 struct dma_async_tx_descriptor *tx = NULL;
1251 struct r5conf *conf = sh->raid_conf;
1252 int level = conf->level;
1253 struct raid5_percpu *percpu;
1257 percpu = per_cpu_ptr(conf->percpu, cpu);
1258 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1259 ops_run_biofill(sh);
1263 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1265 tx = ops_run_compute5(sh, percpu);
1267 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1268 tx = ops_run_compute6_1(sh, percpu);
1270 tx = ops_run_compute6_2(sh, percpu);
1272 /* terminate the chain if reconstruct is not set to be run */
1273 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1277 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1278 tx = ops_run_prexor(sh, percpu, tx);
1280 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1281 tx = ops_run_biodrain(sh, tx);
1285 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1287 ops_run_reconstruct5(sh, percpu, tx);
1289 ops_run_reconstruct6(sh, percpu, tx);
1292 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1293 if (sh->check_state == check_state_run)
1294 ops_run_check_p(sh, percpu);
1295 else if (sh->check_state == check_state_run_q)
1296 ops_run_check_pq(sh, percpu, 0);
1297 else if (sh->check_state == check_state_run_pq)
1298 ops_run_check_pq(sh, percpu, 1);
1304 for (i = disks; i--; ) {
1305 struct r5dev *dev = &sh->dev[i];
1306 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1307 wake_up(&sh->raid_conf->wait_for_overlap);
1312 #ifdef CONFIG_MULTICORE_RAID456
1313 static void async_run_ops(void *param, async_cookie_t cookie)
1315 struct stripe_head *sh = param;
1316 unsigned long ops_request = sh->ops.request;
1318 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1319 wake_up(&sh->ops.wait_for_ops);
1321 __raid_run_ops(sh, ops_request);
1325 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1327 /* since handle_stripe can be called outside of raid5d context
1328 * we need to ensure sh->ops.request is de-staged before another
1331 wait_event(sh->ops.wait_for_ops,
1332 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1333 sh->ops.request = ops_request;
1335 atomic_inc(&sh->count);
1336 async_schedule(async_run_ops, sh);
1339 #define raid_run_ops __raid_run_ops
1342 static int grow_one_stripe(struct r5conf *conf)
1344 struct stripe_head *sh;
1345 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1349 sh->raid_conf = conf;
1350 #ifdef CONFIG_MULTICORE_RAID456
1351 init_waitqueue_head(&sh->ops.wait_for_ops);
1354 if (grow_buffers(sh)) {
1356 kmem_cache_free(conf->slab_cache, sh);
1359 /* we just created an active stripe so... */
1360 atomic_set(&sh->count, 1);
1361 atomic_inc(&conf->active_stripes);
1362 INIT_LIST_HEAD(&sh->lru);
1367 static int grow_stripes(struct r5conf *conf, int num)
1369 struct kmem_cache *sc;
1370 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1372 if (conf->mddev->gendisk)
1373 sprintf(conf->cache_name[0],
1374 "raid%d-%s", conf->level, mdname(conf->mddev));
1376 sprintf(conf->cache_name[0],
1377 "raid%d-%p", conf->level, conf->mddev);
1378 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1380 conf->active_name = 0;
1381 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1382 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1386 conf->slab_cache = sc;
1387 conf->pool_size = devs;
1389 if (!grow_one_stripe(conf))
1395 * scribble_len - return the required size of the scribble region
1396 * @num - total number of disks in the array
1398 * The size must be enough to contain:
1399 * 1/ a struct page pointer for each device in the array +2
1400 * 2/ room to convert each entry in (1) to its corresponding dma
1401 * (dma_map_page()) or page (page_address()) address.
1403 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1404 * calculate over all devices (not just the data blocks), using zeros in place
1405 * of the P and Q blocks.
1407 static size_t scribble_len(int num)
1411 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1416 static int resize_stripes(struct r5conf *conf, int newsize)
1418 /* Make all the stripes able to hold 'newsize' devices.
1419 * New slots in each stripe get 'page' set to a new page.
1421 * This happens in stages:
1422 * 1/ create a new kmem_cache and allocate the required number of
1424 * 2/ gather all the old stripe_heads and tranfer the pages across
1425 * to the new stripe_heads. This will have the side effect of
1426 * freezing the array as once all stripe_heads have been collected,
1427 * no IO will be possible. Old stripe heads are freed once their
1428 * pages have been transferred over, and the old kmem_cache is
1429 * freed when all stripes are done.
1430 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1431 * we simple return a failre status - no need to clean anything up.
1432 * 4/ allocate new pages for the new slots in the new stripe_heads.
1433 * If this fails, we don't bother trying the shrink the
1434 * stripe_heads down again, we just leave them as they are.
1435 * As each stripe_head is processed the new one is released into
1438 * Once step2 is started, we cannot afford to wait for a write,
1439 * so we use GFP_NOIO allocations.
1441 struct stripe_head *osh, *nsh;
1442 LIST_HEAD(newstripes);
1443 struct disk_info *ndisks;
1446 struct kmem_cache *sc;
1449 if (newsize <= conf->pool_size)
1450 return 0; /* never bother to shrink */
1452 err = md_allow_write(conf->mddev);
1457 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1458 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1463 for (i = conf->max_nr_stripes; i; i--) {
1464 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1468 nsh->raid_conf = conf;
1469 #ifdef CONFIG_MULTICORE_RAID456
1470 init_waitqueue_head(&nsh->ops.wait_for_ops);
1473 list_add(&nsh->lru, &newstripes);
1476 /* didn't get enough, give up */
1477 while (!list_empty(&newstripes)) {
1478 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1479 list_del(&nsh->lru);
1480 kmem_cache_free(sc, nsh);
1482 kmem_cache_destroy(sc);
1485 /* Step 2 - Must use GFP_NOIO now.
1486 * OK, we have enough stripes, start collecting inactive
1487 * stripes and copying them over
1489 list_for_each_entry(nsh, &newstripes, lru) {
1490 spin_lock_irq(&conf->device_lock);
1491 wait_event_lock_irq(conf->wait_for_stripe,
1492 !list_empty(&conf->inactive_list),
1495 osh = get_free_stripe(conf);
1496 spin_unlock_irq(&conf->device_lock);
1497 atomic_set(&nsh->count, 1);
1498 for(i=0; i<conf->pool_size; i++)
1499 nsh->dev[i].page = osh->dev[i].page;
1500 for( ; i<newsize; i++)
1501 nsh->dev[i].page = NULL;
1502 kmem_cache_free(conf->slab_cache, osh);
1504 kmem_cache_destroy(conf->slab_cache);
1507 * At this point, we are holding all the stripes so the array
1508 * is completely stalled, so now is a good time to resize
1509 * conf->disks and the scribble region
1511 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1513 for (i=0; i<conf->raid_disks; i++)
1514 ndisks[i] = conf->disks[i];
1516 conf->disks = ndisks;
1521 conf->scribble_len = scribble_len(newsize);
1522 for_each_present_cpu(cpu) {
1523 struct raid5_percpu *percpu;
1526 percpu = per_cpu_ptr(conf->percpu, cpu);
1527 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1530 kfree(percpu->scribble);
1531 percpu->scribble = scribble;
1539 /* Step 4, return new stripes to service */
1540 while(!list_empty(&newstripes)) {
1541 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1542 list_del_init(&nsh->lru);
1544 for (i=conf->raid_disks; i < newsize; i++)
1545 if (nsh->dev[i].page == NULL) {
1546 struct page *p = alloc_page(GFP_NOIO);
1547 nsh->dev[i].page = p;
1551 release_stripe(nsh);
1553 /* critical section pass, GFP_NOIO no longer needed */
1555 conf->slab_cache = sc;
1556 conf->active_name = 1-conf->active_name;
1558 conf->pool_size = newsize;
1562 static int drop_one_stripe(struct r5conf *conf)
1564 struct stripe_head *sh;
1566 spin_lock_irq(&conf->device_lock);
1567 sh = get_free_stripe(conf);
1568 spin_unlock_irq(&conf->device_lock);
1571 BUG_ON(atomic_read(&sh->count));
1573 kmem_cache_free(conf->slab_cache, sh);
1574 atomic_dec(&conf->active_stripes);
1578 static void shrink_stripes(struct r5conf *conf)
1580 while (drop_one_stripe(conf))
1583 if (conf->slab_cache)
1584 kmem_cache_destroy(conf->slab_cache);
1585 conf->slab_cache = NULL;
1588 static void raid5_end_read_request(struct bio * bi, int error)
1590 struct stripe_head *sh = bi->bi_private;
1591 struct r5conf *conf = sh->raid_conf;
1592 int disks = sh->disks, i;
1593 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1594 char b[BDEVNAME_SIZE];
1595 struct md_rdev *rdev;
1598 for (i=0 ; i<disks; i++)
1599 if (bi == &sh->dev[i].req)
1602 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1603 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1611 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1612 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1613 rdev = conf->disks[i].rdev;
1616 "md/raid:%s: read error corrected"
1617 " (%lu sectors at %llu on %s)\n",
1618 mdname(conf->mddev), STRIPE_SECTORS,
1619 (unsigned long long)(sh->sector
1620 + rdev->data_offset),
1621 bdevname(rdev->bdev, b));
1622 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1623 clear_bit(R5_ReadError, &sh->dev[i].flags);
1624 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1626 if (atomic_read(&conf->disks[i].rdev->read_errors))
1627 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1629 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1631 rdev = conf->disks[i].rdev;
1633 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1634 atomic_inc(&rdev->read_errors);
1635 if (conf->mddev->degraded >= conf->max_degraded)
1638 "md/raid:%s: read error not correctable "
1639 "(sector %llu on %s).\n",
1640 mdname(conf->mddev),
1641 (unsigned long long)(sh->sector
1642 + rdev->data_offset),
1644 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1648 "md/raid:%s: read error NOT corrected!! "
1649 "(sector %llu on %s).\n",
1650 mdname(conf->mddev),
1651 (unsigned long long)(sh->sector
1652 + rdev->data_offset),
1654 else if (atomic_read(&rdev->read_errors)
1655 > conf->max_nr_stripes)
1657 "md/raid:%s: Too many read errors, failing device %s.\n",
1658 mdname(conf->mddev), bdn);
1662 set_bit(R5_ReadError, &sh->dev[i].flags);
1664 clear_bit(R5_ReadError, &sh->dev[i].flags);
1665 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1666 md_error(conf->mddev, rdev);
1669 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1670 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1671 set_bit(STRIPE_HANDLE, &sh->state);
1675 static void raid5_end_write_request(struct bio *bi, int error)
1677 struct stripe_head *sh = bi->bi_private;
1678 struct r5conf *conf = sh->raid_conf;
1679 int disks = sh->disks, i;
1680 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1684 for (i=0 ; i<disks; i++)
1685 if (bi == &sh->dev[i].req)
1688 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1689 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1697 set_bit(STRIPE_DEGRADED, &sh->state);
1698 set_bit(WriteErrorSeen, &conf->disks[i].rdev->flags);
1699 set_bit(R5_WriteError, &sh->dev[i].flags);
1700 } else if (is_badblock(conf->disks[i].rdev, sh->sector, STRIPE_SECTORS,
1701 &first_bad, &bad_sectors))
1702 set_bit(R5_MadeGood, &sh->dev[i].flags);
1704 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1706 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1707 set_bit(STRIPE_HANDLE, &sh->state);
1712 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1714 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1716 struct r5dev *dev = &sh->dev[i];
1718 bio_init(&dev->req);
1719 dev->req.bi_io_vec = &dev->vec;
1721 dev->req.bi_max_vecs++;
1722 dev->vec.bv_page = dev->page;
1723 dev->vec.bv_len = STRIPE_SIZE;
1724 dev->vec.bv_offset = 0;
1726 dev->req.bi_sector = sh->sector;
1727 dev->req.bi_private = sh;
1730 dev->sector = compute_blocknr(sh, i, previous);
1733 static void error(struct mddev *mddev, struct md_rdev *rdev)
1735 char b[BDEVNAME_SIZE];
1736 struct r5conf *conf = mddev->private;
1737 pr_debug("raid456: error called\n");
1739 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1740 unsigned long flags;
1741 spin_lock_irqsave(&conf->device_lock, flags);
1743 spin_unlock_irqrestore(&conf->device_lock, flags);
1745 * if recovery was running, make sure it aborts.
1747 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1749 set_bit(Blocked, &rdev->flags);
1750 set_bit(Faulty, &rdev->flags);
1751 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1753 "md/raid:%s: Disk failure on %s, disabling device.\n"
1754 "md/raid:%s: Operation continuing on %d devices.\n",
1756 bdevname(rdev->bdev, b),
1758 conf->raid_disks - mddev->degraded);
1762 * Input: a 'big' sector number,
1763 * Output: index of the data and parity disk, and the sector # in them.
1765 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1766 int previous, int *dd_idx,
1767 struct stripe_head *sh)
1769 sector_t stripe, stripe2;
1770 sector_t chunk_number;
1771 unsigned int chunk_offset;
1774 sector_t new_sector;
1775 int algorithm = previous ? conf->prev_algo
1777 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1778 : conf->chunk_sectors;
1779 int raid_disks = previous ? conf->previous_raid_disks
1781 int data_disks = raid_disks - conf->max_degraded;
1783 /* First compute the information on this sector */
1786 * Compute the chunk number and the sector offset inside the chunk
1788 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1789 chunk_number = r_sector;
1792 * Compute the stripe number
1794 stripe = chunk_number;
1795 *dd_idx = sector_div(stripe, data_disks);
1798 * Select the parity disk based on the user selected algorithm.
1800 pd_idx = qd_idx = -1;
1801 switch(conf->level) {
1803 pd_idx = data_disks;
1806 switch (algorithm) {
1807 case ALGORITHM_LEFT_ASYMMETRIC:
1808 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1809 if (*dd_idx >= pd_idx)
1812 case ALGORITHM_RIGHT_ASYMMETRIC:
1813 pd_idx = sector_div(stripe2, raid_disks);
1814 if (*dd_idx >= pd_idx)
1817 case ALGORITHM_LEFT_SYMMETRIC:
1818 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1819 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1821 case ALGORITHM_RIGHT_SYMMETRIC:
1822 pd_idx = sector_div(stripe2, raid_disks);
1823 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1825 case ALGORITHM_PARITY_0:
1829 case ALGORITHM_PARITY_N:
1830 pd_idx = data_disks;
1838 switch (algorithm) {
1839 case ALGORITHM_LEFT_ASYMMETRIC:
1840 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1841 qd_idx = pd_idx + 1;
1842 if (pd_idx == raid_disks-1) {
1843 (*dd_idx)++; /* Q D D D P */
1845 } else if (*dd_idx >= pd_idx)
1846 (*dd_idx) += 2; /* D D P Q D */
1848 case ALGORITHM_RIGHT_ASYMMETRIC:
1849 pd_idx = sector_div(stripe2, raid_disks);
1850 qd_idx = pd_idx + 1;
1851 if (pd_idx == raid_disks-1) {
1852 (*dd_idx)++; /* Q D D D P */
1854 } else if (*dd_idx >= pd_idx)
1855 (*dd_idx) += 2; /* D D P Q D */
1857 case ALGORITHM_LEFT_SYMMETRIC:
1858 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1859 qd_idx = (pd_idx + 1) % raid_disks;
1860 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1862 case ALGORITHM_RIGHT_SYMMETRIC:
1863 pd_idx = sector_div(stripe2, raid_disks);
1864 qd_idx = (pd_idx + 1) % raid_disks;
1865 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1868 case ALGORITHM_PARITY_0:
1873 case ALGORITHM_PARITY_N:
1874 pd_idx = data_disks;
1875 qd_idx = data_disks + 1;
1878 case ALGORITHM_ROTATING_ZERO_RESTART:
1879 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1880 * of blocks for computing Q is different.
1882 pd_idx = sector_div(stripe2, raid_disks);
1883 qd_idx = pd_idx + 1;
1884 if (pd_idx == raid_disks-1) {
1885 (*dd_idx)++; /* Q D D D P */
1887 } else if (*dd_idx >= pd_idx)
1888 (*dd_idx) += 2; /* D D P Q D */
1892 case ALGORITHM_ROTATING_N_RESTART:
1893 /* Same a left_asymmetric, by first stripe is
1894 * D D D P Q rather than
1898 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1899 qd_idx = pd_idx + 1;
1900 if (pd_idx == raid_disks-1) {
1901 (*dd_idx)++; /* Q D D D P */
1903 } else if (*dd_idx >= pd_idx)
1904 (*dd_idx) += 2; /* D D P Q D */
1908 case ALGORITHM_ROTATING_N_CONTINUE:
1909 /* Same as left_symmetric but Q is before P */
1910 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1911 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1912 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1916 case ALGORITHM_LEFT_ASYMMETRIC_6:
1917 /* RAID5 left_asymmetric, with Q on last device */
1918 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1919 if (*dd_idx >= pd_idx)
1921 qd_idx = raid_disks - 1;
1924 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1925 pd_idx = sector_div(stripe2, raid_disks-1);
1926 if (*dd_idx >= pd_idx)
1928 qd_idx = raid_disks - 1;
1931 case ALGORITHM_LEFT_SYMMETRIC_6:
1932 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1933 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1934 qd_idx = raid_disks - 1;
1937 case ALGORITHM_RIGHT_SYMMETRIC_6:
1938 pd_idx = sector_div(stripe2, raid_disks-1);
1939 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1940 qd_idx = raid_disks - 1;
1943 case ALGORITHM_PARITY_0_6:
1946 qd_idx = raid_disks - 1;
1956 sh->pd_idx = pd_idx;
1957 sh->qd_idx = qd_idx;
1958 sh->ddf_layout = ddf_layout;
1961 * Finally, compute the new sector number
1963 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1968 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1970 struct r5conf *conf = sh->raid_conf;
1971 int raid_disks = sh->disks;
1972 int data_disks = raid_disks - conf->max_degraded;
1973 sector_t new_sector = sh->sector, check;
1974 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1975 : conf->chunk_sectors;
1976 int algorithm = previous ? conf->prev_algo
1980 sector_t chunk_number;
1981 int dummy1, dd_idx = i;
1983 struct stripe_head sh2;
1986 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1987 stripe = new_sector;
1989 if (i == sh->pd_idx)
1991 switch(conf->level) {
1994 switch (algorithm) {
1995 case ALGORITHM_LEFT_ASYMMETRIC:
1996 case ALGORITHM_RIGHT_ASYMMETRIC:
2000 case ALGORITHM_LEFT_SYMMETRIC:
2001 case ALGORITHM_RIGHT_SYMMETRIC:
2004 i -= (sh->pd_idx + 1);
2006 case ALGORITHM_PARITY_0:
2009 case ALGORITHM_PARITY_N:
2016 if (i == sh->qd_idx)
2017 return 0; /* It is the Q disk */
2018 switch (algorithm) {
2019 case ALGORITHM_LEFT_ASYMMETRIC:
2020 case ALGORITHM_RIGHT_ASYMMETRIC:
2021 case ALGORITHM_ROTATING_ZERO_RESTART:
2022 case ALGORITHM_ROTATING_N_RESTART:
2023 if (sh->pd_idx == raid_disks-1)
2024 i--; /* Q D D D P */
2025 else if (i > sh->pd_idx)
2026 i -= 2; /* D D P Q D */
2028 case ALGORITHM_LEFT_SYMMETRIC:
2029 case ALGORITHM_RIGHT_SYMMETRIC:
2030 if (sh->pd_idx == raid_disks-1)
2031 i--; /* Q D D D P */
2036 i -= (sh->pd_idx + 2);
2039 case ALGORITHM_PARITY_0:
2042 case ALGORITHM_PARITY_N:
2044 case ALGORITHM_ROTATING_N_CONTINUE:
2045 /* Like left_symmetric, but P is before Q */
2046 if (sh->pd_idx == 0)
2047 i--; /* P D D D Q */
2052 i -= (sh->pd_idx + 1);
2055 case ALGORITHM_LEFT_ASYMMETRIC_6:
2056 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2060 case ALGORITHM_LEFT_SYMMETRIC_6:
2061 case ALGORITHM_RIGHT_SYMMETRIC_6:
2063 i += data_disks + 1;
2064 i -= (sh->pd_idx + 1);
2066 case ALGORITHM_PARITY_0_6:
2075 chunk_number = stripe * data_disks + i;
2076 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2078 check = raid5_compute_sector(conf, r_sector,
2079 previous, &dummy1, &sh2);
2080 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2081 || sh2.qd_idx != sh->qd_idx) {
2082 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2083 mdname(conf->mddev));
2091 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2092 int rcw, int expand)
2094 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2095 struct r5conf *conf = sh->raid_conf;
2096 int level = conf->level;
2099 /* if we are not expanding this is a proper write request, and
2100 * there will be bios with new data to be drained into the
2104 sh->reconstruct_state = reconstruct_state_drain_run;
2105 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2107 sh->reconstruct_state = reconstruct_state_run;
2109 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2111 for (i = disks; i--; ) {
2112 struct r5dev *dev = &sh->dev[i];
2115 set_bit(R5_LOCKED, &dev->flags);
2116 set_bit(R5_Wantdrain, &dev->flags);
2118 clear_bit(R5_UPTODATE, &dev->flags);
2122 if (s->locked + conf->max_degraded == disks)
2123 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2124 atomic_inc(&conf->pending_full_writes);
2127 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2128 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2130 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2131 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2132 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2133 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2135 for (i = disks; i--; ) {
2136 struct r5dev *dev = &sh->dev[i];
2141 (test_bit(R5_UPTODATE, &dev->flags) ||
2142 test_bit(R5_Wantcompute, &dev->flags))) {
2143 set_bit(R5_Wantdrain, &dev->flags);
2144 set_bit(R5_LOCKED, &dev->flags);
2145 clear_bit(R5_UPTODATE, &dev->flags);
2151 /* keep the parity disk(s) locked while asynchronous operations
2154 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2155 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2159 int qd_idx = sh->qd_idx;
2160 struct r5dev *dev = &sh->dev[qd_idx];
2162 set_bit(R5_LOCKED, &dev->flags);
2163 clear_bit(R5_UPTODATE, &dev->flags);
2167 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2168 __func__, (unsigned long long)sh->sector,
2169 s->locked, s->ops_request);
2173 * Each stripe/dev can have one or more bion attached.
2174 * toread/towrite point to the first in a chain.
2175 * The bi_next chain must be in order.
2177 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2180 struct r5conf *conf = sh->raid_conf;
2183 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2184 (unsigned long long)bi->bi_sector,
2185 (unsigned long long)sh->sector);
2188 spin_lock_irq(&conf->device_lock);
2190 bip = &sh->dev[dd_idx].towrite;
2191 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2194 bip = &sh->dev[dd_idx].toread;
2195 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2196 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2198 bip = & (*bip)->bi_next;
2200 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2203 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2207 bi->bi_phys_segments++;
2210 /* check if page is covered */
2211 sector_t sector = sh->dev[dd_idx].sector;
2212 for (bi=sh->dev[dd_idx].towrite;
2213 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2214 bi && bi->bi_sector <= sector;
2215 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2216 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2217 sector = bi->bi_sector + (bi->bi_size>>9);
2219 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2220 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2222 spin_unlock_irq(&conf->device_lock);
2224 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2225 (unsigned long long)(*bip)->bi_sector,
2226 (unsigned long long)sh->sector, dd_idx);
2228 if (conf->mddev->bitmap && firstwrite) {
2229 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2231 sh->bm_seq = conf->seq_flush+1;
2232 set_bit(STRIPE_BIT_DELAY, &sh->state);
2237 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2238 spin_unlock_irq(&conf->device_lock);
2242 static void end_reshape(struct r5conf *conf);
2244 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2245 struct stripe_head *sh)
2247 int sectors_per_chunk =
2248 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2250 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2251 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2253 raid5_compute_sector(conf,
2254 stripe * (disks - conf->max_degraded)
2255 *sectors_per_chunk + chunk_offset,
2261 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2262 struct stripe_head_state *s, int disks,
2263 struct bio **return_bi)
2266 for (i = disks; i--; ) {
2270 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2271 struct md_rdev *rdev;
2273 rdev = rcu_dereference(conf->disks[i].rdev);
2274 if (rdev && test_bit(In_sync, &rdev->flags))
2275 atomic_inc(&rdev->nr_pending);
2280 if (!rdev_set_badblocks(
2284 md_error(conf->mddev, rdev);
2285 rdev_dec_pending(rdev, conf->mddev);
2288 spin_lock_irq(&conf->device_lock);
2289 /* fail all writes first */
2290 bi = sh->dev[i].towrite;
2291 sh->dev[i].towrite = NULL;
2297 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2298 wake_up(&conf->wait_for_overlap);
2300 while (bi && bi->bi_sector <
2301 sh->dev[i].sector + STRIPE_SECTORS) {
2302 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2303 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2304 if (!raid5_dec_bi_phys_segments(bi)) {
2305 md_write_end(conf->mddev);
2306 bi->bi_next = *return_bi;
2311 /* and fail all 'written' */
2312 bi = sh->dev[i].written;
2313 sh->dev[i].written = NULL;
2314 if (bi) bitmap_end = 1;
2315 while (bi && bi->bi_sector <
2316 sh->dev[i].sector + STRIPE_SECTORS) {
2317 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2318 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2319 if (!raid5_dec_bi_phys_segments(bi)) {
2320 md_write_end(conf->mddev);
2321 bi->bi_next = *return_bi;
2327 /* fail any reads if this device is non-operational and
2328 * the data has not reached the cache yet.
2330 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2331 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2332 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2333 bi = sh->dev[i].toread;
2334 sh->dev[i].toread = NULL;
2335 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2336 wake_up(&conf->wait_for_overlap);
2337 if (bi) s->to_read--;
2338 while (bi && bi->bi_sector <
2339 sh->dev[i].sector + STRIPE_SECTORS) {
2340 struct bio *nextbi =
2341 r5_next_bio(bi, sh->dev[i].sector);
2342 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2343 if (!raid5_dec_bi_phys_segments(bi)) {
2344 bi->bi_next = *return_bi;
2350 spin_unlock_irq(&conf->device_lock);
2352 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2353 STRIPE_SECTORS, 0, 0);
2354 /* If we were in the middle of a write the parity block might
2355 * still be locked - so just clear all R5_LOCKED flags
2357 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2360 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2361 if (atomic_dec_and_test(&conf->pending_full_writes))
2362 md_wakeup_thread(conf->mddev->thread);
2366 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2367 struct stripe_head_state *s)
2372 md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
2373 clear_bit(STRIPE_SYNCING, &sh->state);
2375 /* There is nothing more to do for sync/check/repair.
2376 * For recover we need to record a bad block on all
2377 * non-sync devices, or abort the recovery
2379 if (!test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery))
2381 /* During recovery devices cannot be removed, so locking and
2382 * refcounting of rdevs is not needed
2384 for (i = 0; i < conf->raid_disks; i++) {
2385 struct md_rdev *rdev = conf->disks[i].rdev;
2387 || test_bit(Faulty, &rdev->flags)
2388 || test_bit(In_sync, &rdev->flags))
2390 if (!rdev_set_badblocks(rdev, sh->sector,
2395 conf->recovery_disabled = conf->mddev->recovery_disabled;
2396 set_bit(MD_RECOVERY_INTR, &conf->mddev->recovery);
2400 /* fetch_block - checks the given member device to see if its data needs
2401 * to be read or computed to satisfy a request.
2403 * Returns 1 when no more member devices need to be checked, otherwise returns
2404 * 0 to tell the loop in handle_stripe_fill to continue
2406 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2407 int disk_idx, int disks)
2409 struct r5dev *dev = &sh->dev[disk_idx];
2410 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2411 &sh->dev[s->failed_num[1]] };
2413 /* is the data in this block needed, and can we get it? */
2414 if (!test_bit(R5_LOCKED, &dev->flags) &&
2415 !test_bit(R5_UPTODATE, &dev->flags) &&
2417 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2418 s->syncing || s->expanding ||
2419 (s->failed >= 1 && fdev[0]->toread) ||
2420 (s->failed >= 2 && fdev[1]->toread) ||
2421 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2422 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2423 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2424 /* we would like to get this block, possibly by computing it,
2425 * otherwise read it if the backing disk is insync
2427 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2428 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2429 if ((s->uptodate == disks - 1) &&
2430 (s->failed && (disk_idx == s->failed_num[0] ||
2431 disk_idx == s->failed_num[1]))) {
2432 /* have disk failed, and we're requested to fetch it;
2435 pr_debug("Computing stripe %llu block %d\n",
2436 (unsigned long long)sh->sector, disk_idx);
2437 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2438 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2439 set_bit(R5_Wantcompute, &dev->flags);
2440 sh->ops.target = disk_idx;
2441 sh->ops.target2 = -1; /* no 2nd target */
2443 /* Careful: from this point on 'uptodate' is in the eye
2444 * of raid_run_ops which services 'compute' operations
2445 * before writes. R5_Wantcompute flags a block that will
2446 * be R5_UPTODATE by the time it is needed for a
2447 * subsequent operation.
2451 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2452 /* Computing 2-failure is *very* expensive; only
2453 * do it if failed >= 2
2456 for (other = disks; other--; ) {
2457 if (other == disk_idx)
2459 if (!test_bit(R5_UPTODATE,
2460 &sh->dev[other].flags))
2464 pr_debug("Computing stripe %llu blocks %d,%d\n",
2465 (unsigned long long)sh->sector,
2467 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2468 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2469 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2470 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2471 sh->ops.target = disk_idx;
2472 sh->ops.target2 = other;
2476 } else if (test_bit(R5_Insync, &dev->flags)) {
2477 set_bit(R5_LOCKED, &dev->flags);
2478 set_bit(R5_Wantread, &dev->flags);
2480 pr_debug("Reading block %d (sync=%d)\n",
2481 disk_idx, s->syncing);
2489 * handle_stripe_fill - read or compute data to satisfy pending requests.
2491 static void handle_stripe_fill(struct stripe_head *sh,
2492 struct stripe_head_state *s,
2497 /* look for blocks to read/compute, skip this if a compute
2498 * is already in flight, or if the stripe contents are in the
2499 * midst of changing due to a write
2501 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2502 !sh->reconstruct_state)
2503 for (i = disks; i--; )
2504 if (fetch_block(sh, s, i, disks))
2506 set_bit(STRIPE_HANDLE, &sh->state);
2510 /* handle_stripe_clean_event
2511 * any written block on an uptodate or failed drive can be returned.
2512 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2513 * never LOCKED, so we don't need to test 'failed' directly.
2515 static void handle_stripe_clean_event(struct r5conf *conf,
2516 struct stripe_head *sh, int disks, struct bio **return_bi)
2521 for (i = disks; i--; )
2522 if (sh->dev[i].written) {
2524 if (!test_bit(R5_LOCKED, &dev->flags) &&
2525 test_bit(R5_UPTODATE, &dev->flags)) {
2526 /* We can return any write requests */
2527 struct bio *wbi, *wbi2;
2529 pr_debug("Return write for disc %d\n", i);
2530 spin_lock_irq(&conf->device_lock);
2532 dev->written = NULL;
2533 while (wbi && wbi->bi_sector <
2534 dev->sector + STRIPE_SECTORS) {
2535 wbi2 = r5_next_bio(wbi, dev->sector);
2536 if (!raid5_dec_bi_phys_segments(wbi)) {
2537 md_write_end(conf->mddev);
2538 wbi->bi_next = *return_bi;
2543 if (dev->towrite == NULL)
2545 spin_unlock_irq(&conf->device_lock);
2547 bitmap_endwrite(conf->mddev->bitmap,
2550 !test_bit(STRIPE_DEGRADED, &sh->state),
2555 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2556 if (atomic_dec_and_test(&conf->pending_full_writes))
2557 md_wakeup_thread(conf->mddev->thread);
2560 static void handle_stripe_dirtying(struct r5conf *conf,
2561 struct stripe_head *sh,
2562 struct stripe_head_state *s,
2565 int rmw = 0, rcw = 0, i;
2566 if (conf->max_degraded == 2) {
2567 /* RAID6 requires 'rcw' in current implementation
2568 * Calculate the real rcw later - for now fake it
2569 * look like rcw is cheaper
2572 } else for (i = disks; i--; ) {
2573 /* would I have to read this buffer for read_modify_write */
2574 struct r5dev *dev = &sh->dev[i];
2575 if ((dev->towrite || i == sh->pd_idx) &&
2576 !test_bit(R5_LOCKED, &dev->flags) &&
2577 !(test_bit(R5_UPTODATE, &dev->flags) ||
2578 test_bit(R5_Wantcompute, &dev->flags))) {
2579 if (test_bit(R5_Insync, &dev->flags))
2582 rmw += 2*disks; /* cannot read it */
2584 /* Would I have to read this buffer for reconstruct_write */
2585 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2586 !test_bit(R5_LOCKED, &dev->flags) &&
2587 !(test_bit(R5_UPTODATE, &dev->flags) ||
2588 test_bit(R5_Wantcompute, &dev->flags))) {
2589 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2594 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2595 (unsigned long long)sh->sector, rmw, rcw);
2596 set_bit(STRIPE_HANDLE, &sh->state);
2597 if (rmw < rcw && rmw > 0)
2598 /* prefer read-modify-write, but need to get some data */
2599 for (i = disks; i--; ) {
2600 struct r5dev *dev = &sh->dev[i];
2601 if ((dev->towrite || i == sh->pd_idx) &&
2602 !test_bit(R5_LOCKED, &dev->flags) &&
2603 !(test_bit(R5_UPTODATE, &dev->flags) ||
2604 test_bit(R5_Wantcompute, &dev->flags)) &&
2605 test_bit(R5_Insync, &dev->flags)) {
2607 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2608 pr_debug("Read_old block "
2609 "%d for r-m-w\n", i);
2610 set_bit(R5_LOCKED, &dev->flags);
2611 set_bit(R5_Wantread, &dev->flags);
2614 set_bit(STRIPE_DELAYED, &sh->state);
2615 set_bit(STRIPE_HANDLE, &sh->state);
2619 if (rcw <= rmw && rcw > 0) {
2620 /* want reconstruct write, but need to get some data */
2622 for (i = disks; i--; ) {
2623 struct r5dev *dev = &sh->dev[i];
2624 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2625 i != sh->pd_idx && i != sh->qd_idx &&
2626 !test_bit(R5_LOCKED, &dev->flags) &&
2627 !(test_bit(R5_UPTODATE, &dev->flags) ||
2628 test_bit(R5_Wantcompute, &dev->flags))) {
2630 if (!test_bit(R5_Insync, &dev->flags))
2631 continue; /* it's a failed drive */
2633 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2634 pr_debug("Read_old block "
2635 "%d for Reconstruct\n", i);
2636 set_bit(R5_LOCKED, &dev->flags);
2637 set_bit(R5_Wantread, &dev->flags);
2640 set_bit(STRIPE_DELAYED, &sh->state);
2641 set_bit(STRIPE_HANDLE, &sh->state);
2646 /* now if nothing is locked, and if we have enough data,
2647 * we can start a write request
2649 /* since handle_stripe can be called at any time we need to handle the
2650 * case where a compute block operation has been submitted and then a
2651 * subsequent call wants to start a write request. raid_run_ops only
2652 * handles the case where compute block and reconstruct are requested
2653 * simultaneously. If this is not the case then new writes need to be
2654 * held off until the compute completes.
2656 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2657 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2658 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2659 schedule_reconstruction(sh, s, rcw == 0, 0);
2662 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2663 struct stripe_head_state *s, int disks)
2665 struct r5dev *dev = NULL;
2667 set_bit(STRIPE_HANDLE, &sh->state);
2669 switch (sh->check_state) {
2670 case check_state_idle:
2671 /* start a new check operation if there are no failures */
2672 if (s->failed == 0) {
2673 BUG_ON(s->uptodate != disks);
2674 sh->check_state = check_state_run;
2675 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2676 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2680 dev = &sh->dev[s->failed_num[0]];
2682 case check_state_compute_result:
2683 sh->check_state = check_state_idle;
2685 dev = &sh->dev[sh->pd_idx];
2687 /* check that a write has not made the stripe insync */
2688 if (test_bit(STRIPE_INSYNC, &sh->state))
2691 /* either failed parity check, or recovery is happening */
2692 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2693 BUG_ON(s->uptodate != disks);
2695 set_bit(R5_LOCKED, &dev->flags);
2697 set_bit(R5_Wantwrite, &dev->flags);
2699 clear_bit(STRIPE_DEGRADED, &sh->state);
2700 set_bit(STRIPE_INSYNC, &sh->state);
2702 case check_state_run:
2703 break; /* we will be called again upon completion */
2704 case check_state_check_result:
2705 sh->check_state = check_state_idle;
2707 /* if a failure occurred during the check operation, leave
2708 * STRIPE_INSYNC not set and let the stripe be handled again
2713 /* handle a successful check operation, if parity is correct
2714 * we are done. Otherwise update the mismatch count and repair
2715 * parity if !MD_RECOVERY_CHECK
2717 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2718 /* parity is correct (on disc,
2719 * not in buffer any more)
2721 set_bit(STRIPE_INSYNC, &sh->state);
2723 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2724 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2725 /* don't try to repair!! */
2726 set_bit(STRIPE_INSYNC, &sh->state);
2728 sh->check_state = check_state_compute_run;
2729 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2730 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2731 set_bit(R5_Wantcompute,
2732 &sh->dev[sh->pd_idx].flags);
2733 sh->ops.target = sh->pd_idx;
2734 sh->ops.target2 = -1;
2739 case check_state_compute_run:
2742 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2743 __func__, sh->check_state,
2744 (unsigned long long) sh->sector);
2750 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2751 struct stripe_head_state *s,
2754 int pd_idx = sh->pd_idx;
2755 int qd_idx = sh->qd_idx;
2758 set_bit(STRIPE_HANDLE, &sh->state);
2760 BUG_ON(s->failed > 2);
2762 /* Want to check and possibly repair P and Q.
2763 * However there could be one 'failed' device, in which
2764 * case we can only check one of them, possibly using the
2765 * other to generate missing data
2768 switch (sh->check_state) {
2769 case check_state_idle:
2770 /* start a new check operation if there are < 2 failures */
2771 if (s->failed == s->q_failed) {
2772 /* The only possible failed device holds Q, so it
2773 * makes sense to check P (If anything else were failed,
2774 * we would have used P to recreate it).
2776 sh->check_state = check_state_run;
2778 if (!s->q_failed && s->failed < 2) {
2779 /* Q is not failed, and we didn't use it to generate
2780 * anything, so it makes sense to check it
2782 if (sh->check_state == check_state_run)
2783 sh->check_state = check_state_run_pq;
2785 sh->check_state = check_state_run_q;
2788 /* discard potentially stale zero_sum_result */
2789 sh->ops.zero_sum_result = 0;
2791 if (sh->check_state == check_state_run) {
2792 /* async_xor_zero_sum destroys the contents of P */
2793 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2796 if (sh->check_state >= check_state_run &&
2797 sh->check_state <= check_state_run_pq) {
2798 /* async_syndrome_zero_sum preserves P and Q, so
2799 * no need to mark them !uptodate here
2801 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2805 /* we have 2-disk failure */
2806 BUG_ON(s->failed != 2);
2808 case check_state_compute_result:
2809 sh->check_state = check_state_idle;
2811 /* check that a write has not made the stripe insync */
2812 if (test_bit(STRIPE_INSYNC, &sh->state))
2815 /* now write out any block on a failed drive,
2816 * or P or Q if they were recomputed
2818 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2819 if (s->failed == 2) {
2820 dev = &sh->dev[s->failed_num[1]];
2822 set_bit(R5_LOCKED, &dev->flags);
2823 set_bit(R5_Wantwrite, &dev->flags);
2825 if (s->failed >= 1) {
2826 dev = &sh->dev[s->failed_num[0]];
2828 set_bit(R5_LOCKED, &dev->flags);
2829 set_bit(R5_Wantwrite, &dev->flags);
2831 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2832 dev = &sh->dev[pd_idx];
2834 set_bit(R5_LOCKED, &dev->flags);
2835 set_bit(R5_Wantwrite, &dev->flags);
2837 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2838 dev = &sh->dev[qd_idx];
2840 set_bit(R5_LOCKED, &dev->flags);
2841 set_bit(R5_Wantwrite, &dev->flags);
2843 clear_bit(STRIPE_DEGRADED, &sh->state);
2845 set_bit(STRIPE_INSYNC, &sh->state);
2847 case check_state_run:
2848 case check_state_run_q:
2849 case check_state_run_pq:
2850 break; /* we will be called again upon completion */
2851 case check_state_check_result:
2852 sh->check_state = check_state_idle;
2854 /* handle a successful check operation, if parity is correct
2855 * we are done. Otherwise update the mismatch count and repair
2856 * parity if !MD_RECOVERY_CHECK
2858 if (sh->ops.zero_sum_result == 0) {
2859 /* both parities are correct */
2861 set_bit(STRIPE_INSYNC, &sh->state);
2863 /* in contrast to the raid5 case we can validate
2864 * parity, but still have a failure to write
2867 sh->check_state = check_state_compute_result;
2868 /* Returning at this point means that we may go
2869 * off and bring p and/or q uptodate again so
2870 * we make sure to check zero_sum_result again
2871 * to verify if p or q need writeback
2875 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2876 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2877 /* don't try to repair!! */
2878 set_bit(STRIPE_INSYNC, &sh->state);
2880 int *target = &sh->ops.target;
2882 sh->ops.target = -1;
2883 sh->ops.target2 = -1;
2884 sh->check_state = check_state_compute_run;
2885 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2886 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2887 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2888 set_bit(R5_Wantcompute,
2889 &sh->dev[pd_idx].flags);
2891 target = &sh->ops.target2;
2894 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2895 set_bit(R5_Wantcompute,
2896 &sh->dev[qd_idx].flags);
2903 case check_state_compute_run:
2906 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2907 __func__, sh->check_state,
2908 (unsigned long long) sh->sector);
2913 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
2917 /* We have read all the blocks in this stripe and now we need to
2918 * copy some of them into a target stripe for expand.
2920 struct dma_async_tx_descriptor *tx = NULL;
2921 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2922 for (i = 0; i < sh->disks; i++)
2923 if (i != sh->pd_idx && i != sh->qd_idx) {
2925 struct stripe_head *sh2;
2926 struct async_submit_ctl submit;
2928 sector_t bn = compute_blocknr(sh, i, 1);
2929 sector_t s = raid5_compute_sector(conf, bn, 0,
2931 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2933 /* so far only the early blocks of this stripe
2934 * have been requested. When later blocks
2935 * get requested, we will try again
2938 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2939 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2940 /* must have already done this block */
2941 release_stripe(sh2);
2945 /* place all the copies on one channel */
2946 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2947 tx = async_memcpy(sh2->dev[dd_idx].page,
2948 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2951 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2952 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2953 for (j = 0; j < conf->raid_disks; j++)
2954 if (j != sh2->pd_idx &&
2956 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2958 if (j == conf->raid_disks) {
2959 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2960 set_bit(STRIPE_HANDLE, &sh2->state);
2962 release_stripe(sh2);
2965 /* done submitting copies, wait for them to complete */
2968 dma_wait_for_async_tx(tx);
2974 * handle_stripe - do things to a stripe.
2976 * We lock the stripe and then examine the state of various bits
2977 * to see what needs to be done.
2979 * return some read request which now have data
2980 * return some write requests which are safely on disc
2981 * schedule a read on some buffers
2982 * schedule a write of some buffers
2983 * return confirmation of parity correctness
2985 * buffers are taken off read_list or write_list, and bh_cache buffers
2986 * get BH_Lock set before the stripe lock is released.
2990 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
2992 struct r5conf *conf = sh->raid_conf;
2993 int disks = sh->disks;
2997 memset(s, 0, sizeof(*s));
2999 s->syncing = test_bit(STRIPE_SYNCING, &sh->state);
3000 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3001 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3002 s->failed_num[0] = -1;
3003 s->failed_num[1] = -1;
3005 /* Now to look around and see what can be done */
3007 spin_lock_irq(&conf->device_lock);
3008 for (i=disks; i--; ) {
3009 struct md_rdev *rdev;
3016 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3017 i, dev->flags, dev->toread, dev->towrite, dev->written);
3018 /* maybe we can reply to a read
3020 * new wantfill requests are only permitted while
3021 * ops_complete_biofill is guaranteed to be inactive
3023 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3024 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3025 set_bit(R5_Wantfill, &dev->flags);
3027 /* now count some things */
3028 if (test_bit(R5_LOCKED, &dev->flags))
3030 if (test_bit(R5_UPTODATE, &dev->flags))
3032 if (test_bit(R5_Wantcompute, &dev->flags)) {
3034 BUG_ON(s->compute > 2);
3037 if (test_bit(R5_Wantfill, &dev->flags))
3039 else if (dev->toread)
3043 if (!test_bit(R5_OVERWRITE, &dev->flags))
3048 rdev = rcu_dereference(conf->disks[i].rdev);
3049 if (rdev && test_bit(Faulty, &rdev->flags))
3052 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3053 &first_bad, &bad_sectors);
3054 if (s->blocked_rdev == NULL
3055 && (test_bit(Blocked, &rdev->flags)
3058 set_bit(BlockedBadBlocks,
3060 s->blocked_rdev = rdev;
3061 atomic_inc(&rdev->nr_pending);
3064 clear_bit(R5_Insync, &dev->flags);
3068 /* also not in-sync */
3069 if (!test_bit(WriteErrorSeen, &rdev->flags)) {
3070 /* treat as in-sync, but with a read error
3071 * which we can now try to correct
3073 set_bit(R5_Insync, &dev->flags);
3074 set_bit(R5_ReadError, &dev->flags);
3076 } else if (test_bit(In_sync, &rdev->flags))
3077 set_bit(R5_Insync, &dev->flags);
3078 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3079 /* in sync if before recovery_offset */
3080 set_bit(R5_Insync, &dev->flags);
3081 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3082 test_bit(R5_Expanded, &dev->flags))
3083 /* If we've reshaped into here, we assume it is Insync.
3084 * We will shortly update recovery_offset to make
3087 set_bit(R5_Insync, &dev->flags);
3089 if (test_bit(R5_WriteError, &dev->flags)) {
3090 clear_bit(R5_Insync, &dev->flags);
3091 if (!test_bit(Faulty, &rdev->flags)) {
3092 s->handle_bad_blocks = 1;
3093 atomic_inc(&rdev->nr_pending);
3095 clear_bit(R5_WriteError, &dev->flags);
3097 if (test_bit(R5_MadeGood, &dev->flags)) {
3098 if (!test_bit(Faulty, &rdev->flags)) {
3099 s->handle_bad_blocks = 1;
3100 atomic_inc(&rdev->nr_pending);
3102 clear_bit(R5_MadeGood, &dev->flags);
3104 if (!test_bit(R5_Insync, &dev->flags)) {
3105 /* The ReadError flag will just be confusing now */
3106 clear_bit(R5_ReadError, &dev->flags);
3107 clear_bit(R5_ReWrite, &dev->flags);
3109 if (test_bit(R5_ReadError, &dev->flags))
3110 clear_bit(R5_Insync, &dev->flags);
3111 if (!test_bit(R5_Insync, &dev->flags)) {
3113 s->failed_num[s->failed] = i;
3117 spin_unlock_irq(&conf->device_lock);
3121 static void handle_stripe(struct stripe_head *sh)
3123 struct stripe_head_state s;
3124 struct r5conf *conf = sh->raid_conf;
3127 int disks = sh->disks;
3128 struct r5dev *pdev, *qdev;
3130 clear_bit(STRIPE_HANDLE, &sh->state);
3131 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3132 /* already being handled, ensure it gets handled
3133 * again when current action finishes */
3134 set_bit(STRIPE_HANDLE, &sh->state);
3138 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3139 set_bit(STRIPE_SYNCING, &sh->state);
3140 clear_bit(STRIPE_INSYNC, &sh->state);
3142 clear_bit(STRIPE_DELAYED, &sh->state);
3144 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3145 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3146 (unsigned long long)sh->sector, sh->state,
3147 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3148 sh->check_state, sh->reconstruct_state);
3150 analyse_stripe(sh, &s);
3152 if (s.handle_bad_blocks) {
3153 set_bit(STRIPE_HANDLE, &sh->state);
3157 if (unlikely(s.blocked_rdev)) {
3158 if (s.syncing || s.expanding || s.expanded ||
3159 s.to_write || s.written) {
3160 set_bit(STRIPE_HANDLE, &sh->state);
3163 /* There is nothing for the blocked_rdev to block */
3164 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3165 s.blocked_rdev = NULL;
3168 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3169 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3170 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3173 pr_debug("locked=%d uptodate=%d to_read=%d"
3174 " to_write=%d failed=%d failed_num=%d,%d\n",
3175 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3176 s.failed_num[0], s.failed_num[1]);
3177 /* check if the array has lost more than max_degraded devices and,
3178 * if so, some requests might need to be failed.
3180 if (s.failed > conf->max_degraded) {
3181 sh->check_state = 0;
3182 sh->reconstruct_state = 0;
3183 if (s.to_read+s.to_write+s.written)
3184 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3186 handle_failed_sync(conf, sh, &s);
3190 * might be able to return some write requests if the parity blocks
3191 * are safe, or on a failed drive
3193 pdev = &sh->dev[sh->pd_idx];
3194 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3195 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3196 qdev = &sh->dev[sh->qd_idx];
3197 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3198 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3202 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3203 && !test_bit(R5_LOCKED, &pdev->flags)
3204 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3205 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3206 && !test_bit(R5_LOCKED, &qdev->flags)
3207 && test_bit(R5_UPTODATE, &qdev->flags)))))
3208 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3210 /* Now we might consider reading some blocks, either to check/generate
3211 * parity, or to satisfy requests
3212 * or to load a block that is being partially written.
3214 if (s.to_read || s.non_overwrite
3215 || (conf->level == 6 && s.to_write && s.failed)
3216 || (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3217 handle_stripe_fill(sh, &s, disks);
3219 /* Now we check to see if any write operations have recently
3223 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3225 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3226 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3227 sh->reconstruct_state = reconstruct_state_idle;
3229 /* All the 'written' buffers and the parity block are ready to
3230 * be written back to disk
3232 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3233 BUG_ON(sh->qd_idx >= 0 &&
3234 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3235 for (i = disks; i--; ) {
3236 struct r5dev *dev = &sh->dev[i];
3237 if (test_bit(R5_LOCKED, &dev->flags) &&
3238 (i == sh->pd_idx || i == sh->qd_idx ||
3240 pr_debug("Writing block %d\n", i);
3241 set_bit(R5_Wantwrite, &dev->flags);
3246 if (!test_bit(R5_Insync, &dev->flags) ||
3247 ((i == sh->pd_idx || i == sh->qd_idx) &&
3249 set_bit(STRIPE_INSYNC, &sh->state);
3252 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3253 s.dec_preread_active = 1;
3256 /* Now to consider new write requests and what else, if anything
3257 * should be read. We do not handle new writes when:
3258 * 1/ A 'write' operation (copy+xor) is already in flight.
3259 * 2/ A 'check' operation is in flight, as it may clobber the parity
3262 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3263 handle_stripe_dirtying(conf, sh, &s, disks);
3265 /* maybe we need to check and possibly fix the parity for this stripe
3266 * Any reads will already have been scheduled, so we just see if enough
3267 * data is available. The parity check is held off while parity
3268 * dependent operations are in flight.
3270 if (sh->check_state ||
3271 (s.syncing && s.locked == 0 &&
3272 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3273 !test_bit(STRIPE_INSYNC, &sh->state))) {
3274 if (conf->level == 6)
3275 handle_parity_checks6(conf, sh, &s, disks);
3277 handle_parity_checks5(conf, sh, &s, disks);
3280 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3281 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3282 clear_bit(STRIPE_SYNCING, &sh->state);
3285 /* If the failed drives are just a ReadError, then we might need
3286 * to progress the repair/check process
3288 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3289 for (i = 0; i < s.failed; i++) {
3290 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3291 if (test_bit(R5_ReadError, &dev->flags)
3292 && !test_bit(R5_LOCKED, &dev->flags)
3293 && test_bit(R5_UPTODATE, &dev->flags)
3295 if (!test_bit(R5_ReWrite, &dev->flags)) {
3296 set_bit(R5_Wantwrite, &dev->flags);
3297 set_bit(R5_ReWrite, &dev->flags);
3298 set_bit(R5_LOCKED, &dev->flags);
3301 /* let's read it back */
3302 set_bit(R5_Wantread, &dev->flags);
3303 set_bit(R5_LOCKED, &dev->flags);
3310 /* Finish reconstruct operations initiated by the expansion process */
3311 if (sh->reconstruct_state == reconstruct_state_result) {
3312 struct stripe_head *sh_src
3313 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3314 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3315 /* sh cannot be written until sh_src has been read.
3316 * so arrange for sh to be delayed a little
3318 set_bit(STRIPE_DELAYED, &sh->state);
3319 set_bit(STRIPE_HANDLE, &sh->state);
3320 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3322 atomic_inc(&conf->preread_active_stripes);
3323 release_stripe(sh_src);
3327 release_stripe(sh_src);
3329 sh->reconstruct_state = reconstruct_state_idle;
3330 clear_bit(STRIPE_EXPANDING, &sh->state);
3331 for (i = conf->raid_disks; i--; ) {
3332 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3333 set_bit(R5_LOCKED, &sh->dev[i].flags);
3338 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3339 !sh->reconstruct_state) {
3340 /* Need to write out all blocks after computing parity */
3341 sh->disks = conf->raid_disks;
3342 stripe_set_idx(sh->sector, conf, 0, sh);
3343 schedule_reconstruction(sh, &s, 1, 1);
3344 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3345 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3346 atomic_dec(&conf->reshape_stripes);
3347 wake_up(&conf->wait_for_overlap);
3348 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3351 if (s.expanding && s.locked == 0 &&
3352 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3353 handle_stripe_expansion(conf, sh);
3356 /* wait for this device to become unblocked */
3357 if (conf->mddev->external && unlikely(s.blocked_rdev))
3358 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3360 if (s.handle_bad_blocks)
3361 for (i = disks; i--; ) {
3362 struct md_rdev *rdev;
3363 struct r5dev *dev = &sh->dev[i];
3364 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3365 /* We own a safe reference to the rdev */
3366 rdev = conf->disks[i].rdev;
3367 if (!rdev_set_badblocks(rdev, sh->sector,
3369 md_error(conf->mddev, rdev);
3370 rdev_dec_pending(rdev, conf->mddev);
3372 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3373 rdev = conf->disks[i].rdev;
3374 rdev_clear_badblocks(rdev, sh->sector,
3376 rdev_dec_pending(rdev, conf->mddev);
3381 raid_run_ops(sh, s.ops_request);
3385 if (s.dec_preread_active) {
3386 /* We delay this until after ops_run_io so that if make_request
3387 * is waiting on a flush, it won't continue until the writes
3388 * have actually been submitted.
3390 atomic_dec(&conf->preread_active_stripes);
3391 if (atomic_read(&conf->preread_active_stripes) <
3393 md_wakeup_thread(conf->mddev->thread);
3396 return_io(s.return_bi);
3398 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3401 static void raid5_activate_delayed(struct r5conf *conf)
3403 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3404 while (!list_empty(&conf->delayed_list)) {
3405 struct list_head *l = conf->delayed_list.next;
3406 struct stripe_head *sh;
3407 sh = list_entry(l, struct stripe_head, lru);
3409 clear_bit(STRIPE_DELAYED, &sh->state);
3410 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3411 atomic_inc(&conf->preread_active_stripes);
3412 list_add_tail(&sh->lru, &conf->hold_list);
3417 static void activate_bit_delay(struct r5conf *conf)
3419 /* device_lock is held */
3420 struct list_head head;
3421 list_add(&head, &conf->bitmap_list);
3422 list_del_init(&conf->bitmap_list);
3423 while (!list_empty(&head)) {
3424 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3425 list_del_init(&sh->lru);
3426 atomic_inc(&sh->count);
3427 __release_stripe(conf, sh);
3431 int md_raid5_congested(struct mddev *mddev, int bits)
3433 struct r5conf *conf = mddev->private;
3435 /* No difference between reads and writes. Just check
3436 * how busy the stripe_cache is
3439 if (conf->inactive_blocked)
3443 if (list_empty_careful(&conf->inactive_list))
3448 EXPORT_SYMBOL_GPL(md_raid5_congested);
3450 static int raid5_congested(void *data, int bits)
3452 struct mddev *mddev = data;
3454 return mddev_congested(mddev, bits) ||
3455 md_raid5_congested(mddev, bits);
3458 /* We want read requests to align with chunks where possible,
3459 * but write requests don't need to.
3461 static int raid5_mergeable_bvec(struct request_queue *q,
3462 struct bvec_merge_data *bvm,
3463 struct bio_vec *biovec)
3465 struct mddev *mddev = q->queuedata;
3466 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3468 unsigned int chunk_sectors = mddev->chunk_sectors;
3469 unsigned int bio_sectors = bvm->bi_size >> 9;
3471 if ((bvm->bi_rw & 1) == WRITE)
3472 return biovec->bv_len; /* always allow writes to be mergeable */
3474 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3475 chunk_sectors = mddev->new_chunk_sectors;
3476 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3477 if (max < 0) max = 0;
3478 if (max <= biovec->bv_len && bio_sectors == 0)
3479 return biovec->bv_len;
3485 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3487 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3488 unsigned int chunk_sectors = mddev->chunk_sectors;
3489 unsigned int bio_sectors = bio->bi_size >> 9;
3491 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3492 chunk_sectors = mddev->new_chunk_sectors;
3493 return chunk_sectors >=
3494 ((sector & (chunk_sectors - 1)) + bio_sectors);
3498 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3499 * later sampled by raid5d.
3501 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3503 unsigned long flags;
3505 spin_lock_irqsave(&conf->device_lock, flags);
3507 bi->bi_next = conf->retry_read_aligned_list;
3508 conf->retry_read_aligned_list = bi;
3510 spin_unlock_irqrestore(&conf->device_lock, flags);
3511 md_wakeup_thread(conf->mddev->thread);
3515 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3519 bi = conf->retry_read_aligned;
3521 conf->retry_read_aligned = NULL;
3524 bi = conf->retry_read_aligned_list;
3526 conf->retry_read_aligned_list = bi->bi_next;
3529 * this sets the active strip count to 1 and the processed
3530 * strip count to zero (upper 8 bits)
3532 bi->bi_phys_segments = 1; /* biased count of active stripes */
3540 * The "raid5_align_endio" should check if the read succeeded and if it
3541 * did, call bio_endio on the original bio (having bio_put the new bio
3543 * If the read failed..
3545 static void raid5_align_endio(struct bio *bi, int error)
3547 struct bio* raid_bi = bi->bi_private;
3548 struct mddev *mddev;
3549 struct r5conf *conf;
3550 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3551 struct md_rdev *rdev;
3555 rdev = (void*)raid_bi->bi_next;
3556 raid_bi->bi_next = NULL;
3557 mddev = rdev->mddev;
3558 conf = mddev->private;
3560 rdev_dec_pending(rdev, conf->mddev);
3562 if (!error && uptodate) {
3563 bio_endio(raid_bi, 0);
3564 if (atomic_dec_and_test(&conf->active_aligned_reads))
3565 wake_up(&conf->wait_for_stripe);
3570 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3572 add_bio_to_retry(raid_bi, conf);
3575 static int bio_fits_rdev(struct bio *bi)
3577 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3579 if ((bi->bi_size>>9) > queue_max_sectors(q))
3581 blk_recount_segments(q, bi);
3582 if (bi->bi_phys_segments > queue_max_segments(q))
3585 if (q->merge_bvec_fn)
3586 /* it's too hard to apply the merge_bvec_fn at this stage,
3595 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3597 struct r5conf *conf = mddev->private;
3599 struct bio* align_bi;
3600 struct md_rdev *rdev;
3602 if (!in_chunk_boundary(mddev, raid_bio)) {
3603 pr_debug("chunk_aligned_read : non aligned\n");
3607 * use bio_clone_mddev to make a copy of the bio
3609 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3613 * set bi_end_io to a new function, and set bi_private to the
3616 align_bi->bi_end_io = raid5_align_endio;
3617 align_bi->bi_private = raid_bio;
3621 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3626 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3627 if (rdev && test_bit(In_sync, &rdev->flags)) {
3631 atomic_inc(&rdev->nr_pending);
3633 raid_bio->bi_next = (void*)rdev;
3634 align_bi->bi_bdev = rdev->bdev;
3635 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3637 if (!bio_fits_rdev(align_bi) ||
3638 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3639 &first_bad, &bad_sectors)) {
3640 /* too big in some way, or has a known bad block */
3642 rdev_dec_pending(rdev, mddev);
3646 /* No reshape active, so we can trust rdev->data_offset */
3647 align_bi->bi_sector += rdev->data_offset;
3649 spin_lock_irq(&conf->device_lock);
3650 wait_event_lock_irq(conf->wait_for_stripe,
3652 conf->device_lock, /* nothing */);
3653 atomic_inc(&conf->active_aligned_reads);
3654 spin_unlock_irq(&conf->device_lock);
3656 generic_make_request(align_bi);
3665 /* __get_priority_stripe - get the next stripe to process
3667 * Full stripe writes are allowed to pass preread active stripes up until
3668 * the bypass_threshold is exceeded. In general the bypass_count
3669 * increments when the handle_list is handled before the hold_list; however, it
3670 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3671 * stripe with in flight i/o. The bypass_count will be reset when the
3672 * head of the hold_list has changed, i.e. the head was promoted to the
3675 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3677 struct stripe_head *sh;
3679 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3681 list_empty(&conf->handle_list) ? "empty" : "busy",
3682 list_empty(&conf->hold_list) ? "empty" : "busy",
3683 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3685 if (!list_empty(&conf->handle_list)) {
3686 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3688 if (list_empty(&conf->hold_list))
3689 conf->bypass_count = 0;
3690 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3691 if (conf->hold_list.next == conf->last_hold)
3692 conf->bypass_count++;
3694 conf->last_hold = conf->hold_list.next;
3695 conf->bypass_count -= conf->bypass_threshold;
3696 if (conf->bypass_count < 0)
3697 conf->bypass_count = 0;
3700 } else if (!list_empty(&conf->hold_list) &&
3701 ((conf->bypass_threshold &&
3702 conf->bypass_count > conf->bypass_threshold) ||
3703 atomic_read(&conf->pending_full_writes) == 0)) {
3704 sh = list_entry(conf->hold_list.next,
3706 conf->bypass_count -= conf->bypass_threshold;
3707 if (conf->bypass_count < 0)
3708 conf->bypass_count = 0;
3712 list_del_init(&sh->lru);
3713 atomic_inc(&sh->count);
3714 BUG_ON(atomic_read(&sh->count) != 1);
3718 static void make_request(struct mddev *mddev, struct bio * bi)
3720 struct r5conf *conf = mddev->private;
3722 sector_t new_sector;
3723 sector_t logical_sector, last_sector;
3724 struct stripe_head *sh;
3725 const int rw = bio_data_dir(bi);
3729 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3730 md_flush_request(mddev, bi);
3734 md_write_start(mddev, bi);
3737 mddev->reshape_position == MaxSector &&
3738 chunk_aligned_read(mddev,bi))
3741 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3742 last_sector = bi->bi_sector + (bi->bi_size>>9);
3744 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3746 plugged = mddev_check_plugged(mddev);
3747 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3749 int disks, data_disks;
3754 disks = conf->raid_disks;
3755 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3756 if (unlikely(conf->reshape_progress != MaxSector)) {
3757 /* spinlock is needed as reshape_progress may be
3758 * 64bit on a 32bit platform, and so it might be
3759 * possible to see a half-updated value
3760 * Of course reshape_progress could change after
3761 * the lock is dropped, so once we get a reference
3762 * to the stripe that we think it is, we will have
3765 spin_lock_irq(&conf->device_lock);
3766 if (mddev->delta_disks < 0
3767 ? logical_sector < conf->reshape_progress
3768 : logical_sector >= conf->reshape_progress) {
3769 disks = conf->previous_raid_disks;
3772 if (mddev->delta_disks < 0
3773 ? logical_sector < conf->reshape_safe
3774 : logical_sector >= conf->reshape_safe) {
3775 spin_unlock_irq(&conf->device_lock);
3780 spin_unlock_irq(&conf->device_lock);
3782 data_disks = disks - conf->max_degraded;
3784 new_sector = raid5_compute_sector(conf, logical_sector,
3787 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3788 (unsigned long long)new_sector,
3789 (unsigned long long)logical_sector);
3791 sh = get_active_stripe(conf, new_sector, previous,
3792 (bi->bi_rw&RWA_MASK), 0);
3794 if (unlikely(previous)) {
3795 /* expansion might have moved on while waiting for a
3796 * stripe, so we must do the range check again.
3797 * Expansion could still move past after this
3798 * test, but as we are holding a reference to
3799 * 'sh', we know that if that happens,
3800 * STRIPE_EXPANDING will get set and the expansion
3801 * won't proceed until we finish with the stripe.
3804 spin_lock_irq(&conf->device_lock);
3805 if (mddev->delta_disks < 0
3806 ? logical_sector >= conf->reshape_progress
3807 : logical_sector < conf->reshape_progress)
3808 /* mismatch, need to try again */
3810 spin_unlock_irq(&conf->device_lock);
3819 logical_sector >= mddev->suspend_lo &&
3820 logical_sector < mddev->suspend_hi) {
3822 /* As the suspend_* range is controlled by
3823 * userspace, we want an interruptible
3826 flush_signals(current);
3827 prepare_to_wait(&conf->wait_for_overlap,
3828 &w, TASK_INTERRUPTIBLE);
3829 if (logical_sector >= mddev->suspend_lo &&
3830 logical_sector < mddev->suspend_hi)
3835 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3836 !add_stripe_bio(sh, bi, dd_idx, rw)) {
3837 /* Stripe is busy expanding or
3838 * add failed due to overlap. Flush everything
3841 md_wakeup_thread(mddev->thread);
3846 finish_wait(&conf->wait_for_overlap, &w);
3847 set_bit(STRIPE_HANDLE, &sh->state);
3848 clear_bit(STRIPE_DELAYED, &sh->state);
3849 if ((bi->bi_rw & REQ_SYNC) &&
3850 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3851 atomic_inc(&conf->preread_active_stripes);
3854 /* cannot get stripe for read-ahead, just give-up */
3855 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3856 finish_wait(&conf->wait_for_overlap, &w);
3862 md_wakeup_thread(mddev->thread);
3864 spin_lock_irq(&conf->device_lock);
3865 remaining = raid5_dec_bi_phys_segments(bi);
3866 spin_unlock_irq(&conf->device_lock);
3867 if (remaining == 0) {
3870 md_write_end(mddev);
3876 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
3878 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
3880 /* reshaping is quite different to recovery/resync so it is
3881 * handled quite separately ... here.
3883 * On each call to sync_request, we gather one chunk worth of
3884 * destination stripes and flag them as expanding.
3885 * Then we find all the source stripes and request reads.
3886 * As the reads complete, handle_stripe will copy the data
3887 * into the destination stripe and release that stripe.
3889 struct r5conf *conf = mddev->private;
3890 struct stripe_head *sh;
3891 sector_t first_sector, last_sector;
3892 int raid_disks = conf->previous_raid_disks;
3893 int data_disks = raid_disks - conf->max_degraded;
3894 int new_data_disks = conf->raid_disks - conf->max_degraded;
3897 sector_t writepos, readpos, safepos;
3898 sector_t stripe_addr;
3899 int reshape_sectors;
3900 struct list_head stripes;
3902 if (sector_nr == 0) {
3903 /* If restarting in the middle, skip the initial sectors */
3904 if (mddev->delta_disks < 0 &&
3905 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3906 sector_nr = raid5_size(mddev, 0, 0)
3907 - conf->reshape_progress;
3908 } else if (mddev->delta_disks >= 0 &&
3909 conf->reshape_progress > 0)
3910 sector_nr = conf->reshape_progress;
3911 sector_div(sector_nr, new_data_disks);
3913 mddev->curr_resync_completed = sector_nr;
3914 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3920 /* We need to process a full chunk at a time.
3921 * If old and new chunk sizes differ, we need to process the
3924 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3925 reshape_sectors = mddev->new_chunk_sectors;
3927 reshape_sectors = mddev->chunk_sectors;
3929 /* we update the metadata when there is more than 3Meg
3930 * in the block range (that is rather arbitrary, should
3931 * probably be time based) or when the data about to be
3932 * copied would over-write the source of the data at
3933 * the front of the range.
3934 * i.e. one new_stripe along from reshape_progress new_maps
3935 * to after where reshape_safe old_maps to
3937 writepos = conf->reshape_progress;
3938 sector_div(writepos, new_data_disks);
3939 readpos = conf->reshape_progress;
3940 sector_div(readpos, data_disks);
3941 safepos = conf->reshape_safe;
3942 sector_div(safepos, data_disks);
3943 if (mddev->delta_disks < 0) {
3944 writepos -= min_t(sector_t, reshape_sectors, writepos);
3945 readpos += reshape_sectors;
3946 safepos += reshape_sectors;
3948 writepos += reshape_sectors;
3949 readpos -= min_t(sector_t, reshape_sectors, readpos);
3950 safepos -= min_t(sector_t, reshape_sectors, safepos);
3953 /* 'writepos' is the most advanced device address we might write.
3954 * 'readpos' is the least advanced device address we might read.
3955 * 'safepos' is the least address recorded in the metadata as having
3957 * If 'readpos' is behind 'writepos', then there is no way that we can
3958 * ensure safety in the face of a crash - that must be done by userspace
3959 * making a backup of the data. So in that case there is no particular
3960 * rush to update metadata.
3961 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3962 * update the metadata to advance 'safepos' to match 'readpos' so that
3963 * we can be safe in the event of a crash.
3964 * So we insist on updating metadata if safepos is behind writepos and
3965 * readpos is beyond writepos.
3966 * In any case, update the metadata every 10 seconds.
3967 * Maybe that number should be configurable, but I'm not sure it is
3968 * worth it.... maybe it could be a multiple of safemode_delay???
3970 if ((mddev->delta_disks < 0
3971 ? (safepos > writepos && readpos < writepos)
3972 : (safepos < writepos && readpos > writepos)) ||
3973 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3974 /* Cannot proceed until we've updated the superblock... */
3975 wait_event(conf->wait_for_overlap,
3976 atomic_read(&conf->reshape_stripes)==0);
3977 mddev->reshape_position = conf->reshape_progress;
3978 mddev->curr_resync_completed = sector_nr;
3979 conf->reshape_checkpoint = jiffies;
3980 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3981 md_wakeup_thread(mddev->thread);
3982 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3983 kthread_should_stop());
3984 spin_lock_irq(&conf->device_lock);
3985 conf->reshape_safe = mddev->reshape_position;
3986 spin_unlock_irq(&conf->device_lock);
3987 wake_up(&conf->wait_for_overlap);
3988 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3991 if (mddev->delta_disks < 0) {
3992 BUG_ON(conf->reshape_progress == 0);
3993 stripe_addr = writepos;
3994 BUG_ON((mddev->dev_sectors &
3995 ~((sector_t)reshape_sectors - 1))
3996 - reshape_sectors - stripe_addr
3999 BUG_ON(writepos != sector_nr + reshape_sectors);
4000 stripe_addr = sector_nr;
4002 INIT_LIST_HEAD(&stripes);
4003 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4005 int skipped_disk = 0;
4006 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4007 set_bit(STRIPE_EXPANDING, &sh->state);
4008 atomic_inc(&conf->reshape_stripes);
4009 /* If any of this stripe is beyond the end of the old
4010 * array, then we need to zero those blocks
4012 for (j=sh->disks; j--;) {
4014 if (j == sh->pd_idx)
4016 if (conf->level == 6 &&
4019 s = compute_blocknr(sh, j, 0);
4020 if (s < raid5_size(mddev, 0, 0)) {
4024 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4025 set_bit(R5_Expanded, &sh->dev[j].flags);
4026 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4028 if (!skipped_disk) {
4029 set_bit(STRIPE_EXPAND_READY, &sh->state);
4030 set_bit(STRIPE_HANDLE, &sh->state);
4032 list_add(&sh->lru, &stripes);
4034 spin_lock_irq(&conf->device_lock);
4035 if (mddev->delta_disks < 0)
4036 conf->reshape_progress -= reshape_sectors * new_data_disks;
4038 conf->reshape_progress += reshape_sectors * new_data_disks;
4039 spin_unlock_irq(&conf->device_lock);
4040 /* Ok, those stripe are ready. We can start scheduling
4041 * reads on the source stripes.
4042 * The source stripes are determined by mapping the first and last
4043 * block on the destination stripes.
4046 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4049 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4050 * new_data_disks - 1),
4052 if (last_sector >= mddev->dev_sectors)
4053 last_sector = mddev->dev_sectors - 1;
4054 while (first_sector <= last_sector) {
4055 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4056 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4057 set_bit(STRIPE_HANDLE, &sh->state);
4059 first_sector += STRIPE_SECTORS;
4061 /* Now that the sources are clearly marked, we can release
4062 * the destination stripes
4064 while (!list_empty(&stripes)) {
4065 sh = list_entry(stripes.next, struct stripe_head, lru);
4066 list_del_init(&sh->lru);
4069 /* If this takes us to the resync_max point where we have to pause,
4070 * then we need to write out the superblock.
4072 sector_nr += reshape_sectors;
4073 if ((sector_nr - mddev->curr_resync_completed) * 2
4074 >= mddev->resync_max - mddev->curr_resync_completed) {
4075 /* Cannot proceed until we've updated the superblock... */
4076 wait_event(conf->wait_for_overlap,
4077 atomic_read(&conf->reshape_stripes) == 0);
4078 mddev->reshape_position = conf->reshape_progress;
4079 mddev->curr_resync_completed = sector_nr;
4080 conf->reshape_checkpoint = jiffies;
4081 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4082 md_wakeup_thread(mddev->thread);
4083 wait_event(mddev->sb_wait,
4084 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4085 || kthread_should_stop());
4086 spin_lock_irq(&conf->device_lock);
4087 conf->reshape_safe = mddev->reshape_position;
4088 spin_unlock_irq(&conf->device_lock);
4089 wake_up(&conf->wait_for_overlap);
4090 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4092 return reshape_sectors;
4095 /* FIXME go_faster isn't used */
4096 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4098 struct r5conf *conf = mddev->private;
4099 struct stripe_head *sh;
4100 sector_t max_sector = mddev->dev_sectors;
4101 sector_t sync_blocks;
4102 int still_degraded = 0;
4105 if (sector_nr >= max_sector) {
4106 /* just being told to finish up .. nothing much to do */
4108 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4113 if (mddev->curr_resync < max_sector) /* aborted */
4114 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4116 else /* completed sync */
4118 bitmap_close_sync(mddev->bitmap);
4123 /* Allow raid5_quiesce to complete */
4124 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4126 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4127 return reshape_request(mddev, sector_nr, skipped);
4129 /* No need to check resync_max as we never do more than one
4130 * stripe, and as resync_max will always be on a chunk boundary,
4131 * if the check in md_do_sync didn't fire, there is no chance
4132 * of overstepping resync_max here
4135 /* if there is too many failed drives and we are trying
4136 * to resync, then assert that we are finished, because there is
4137 * nothing we can do.
4139 if (mddev->degraded >= conf->max_degraded &&
4140 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4141 sector_t rv = mddev->dev_sectors - sector_nr;
4145 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4146 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4147 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4148 /* we can skip this block, and probably more */
4149 sync_blocks /= STRIPE_SECTORS;
4151 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4155 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4157 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4159 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4160 /* make sure we don't swamp the stripe cache if someone else
4161 * is trying to get access
4163 schedule_timeout_uninterruptible(1);
4165 /* Need to check if array will still be degraded after recovery/resync
4166 * We don't need to check the 'failed' flag as when that gets set,
4169 for (i = 0; i < conf->raid_disks; i++)
4170 if (conf->disks[i].rdev == NULL)
4173 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4175 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4180 return STRIPE_SECTORS;
4183 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4185 /* We may not be able to submit a whole bio at once as there
4186 * may not be enough stripe_heads available.
4187 * We cannot pre-allocate enough stripe_heads as we may need
4188 * more than exist in the cache (if we allow ever large chunks).
4189 * So we do one stripe head at a time and record in
4190 * ->bi_hw_segments how many have been done.
4192 * We *know* that this entire raid_bio is in one chunk, so
4193 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4195 struct stripe_head *sh;
4197 sector_t sector, logical_sector, last_sector;
4202 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4203 sector = raid5_compute_sector(conf, logical_sector,
4205 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4207 for (; logical_sector < last_sector;
4208 logical_sector += STRIPE_SECTORS,
4209 sector += STRIPE_SECTORS,
4212 if (scnt < raid5_bi_hw_segments(raid_bio))
4213 /* already done this stripe */
4216 sh = get_active_stripe(conf, sector, 0, 1, 0);
4219 /* failed to get a stripe - must wait */
4220 raid5_set_bi_hw_segments(raid_bio, scnt);
4221 conf->retry_read_aligned = raid_bio;
4225 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4226 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4228 raid5_set_bi_hw_segments(raid_bio, scnt);
4229 conf->retry_read_aligned = raid_bio;
4237 spin_lock_irq(&conf->device_lock);
4238 remaining = raid5_dec_bi_phys_segments(raid_bio);
4239 spin_unlock_irq(&conf->device_lock);
4241 bio_endio(raid_bio, 0);
4242 if (atomic_dec_and_test(&conf->active_aligned_reads))
4243 wake_up(&conf->wait_for_stripe);
4249 * This is our raid5 kernel thread.
4251 * We scan the hash table for stripes which can be handled now.
4252 * During the scan, completed stripes are saved for us by the interrupt
4253 * handler, so that they will not have to wait for our next wakeup.
4255 static void raid5d(struct mddev *mddev)
4257 struct stripe_head *sh;
4258 struct r5conf *conf = mddev->private;
4260 struct blk_plug plug;
4262 pr_debug("+++ raid5d active\n");
4264 md_check_recovery(mddev);
4266 blk_start_plug(&plug);
4268 spin_lock_irq(&conf->device_lock);
4272 if (atomic_read(&mddev->plug_cnt) == 0 &&
4273 !list_empty(&conf->bitmap_list)) {
4274 /* Now is a good time to flush some bitmap updates */
4276 spin_unlock_irq(&conf->device_lock);
4277 bitmap_unplug(mddev->bitmap);
4278 spin_lock_irq(&conf->device_lock);
4279 conf->seq_write = conf->seq_flush;
4280 activate_bit_delay(conf);
4282 if (atomic_read(&mddev->plug_cnt) == 0)
4283 raid5_activate_delayed(conf);
4285 while ((bio = remove_bio_from_retry(conf))) {
4287 spin_unlock_irq(&conf->device_lock);
4288 ok = retry_aligned_read(conf, bio);
4289 spin_lock_irq(&conf->device_lock);
4295 sh = __get_priority_stripe(conf);
4299 spin_unlock_irq(&conf->device_lock);
4306 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4307 md_check_recovery(mddev);
4309 spin_lock_irq(&conf->device_lock);
4311 pr_debug("%d stripes handled\n", handled);
4313 spin_unlock_irq(&conf->device_lock);
4315 async_tx_issue_pending_all();
4316 blk_finish_plug(&plug);
4318 pr_debug("--- raid5d inactive\n");
4322 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4324 struct r5conf *conf = mddev->private;
4326 return sprintf(page, "%d\n", conf->max_nr_stripes);
4332 raid5_set_cache_size(struct mddev *mddev, int size)
4334 struct r5conf *conf = mddev->private;
4337 if (size <= 16 || size > 32768)
4339 while (size < conf->max_nr_stripes) {
4340 if (drop_one_stripe(conf))
4341 conf->max_nr_stripes--;
4345 err = md_allow_write(mddev);
4348 while (size > conf->max_nr_stripes) {
4349 if (grow_one_stripe(conf))
4350 conf->max_nr_stripes++;
4355 EXPORT_SYMBOL(raid5_set_cache_size);
4358 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4360 struct r5conf *conf = mddev->private;
4364 if (len >= PAGE_SIZE)
4369 if (strict_strtoul(page, 10, &new))
4371 err = raid5_set_cache_size(mddev, new);
4377 static struct md_sysfs_entry
4378 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4379 raid5_show_stripe_cache_size,
4380 raid5_store_stripe_cache_size);
4383 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4385 struct r5conf *conf = mddev->private;
4387 return sprintf(page, "%d\n", conf->bypass_threshold);
4393 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4395 struct r5conf *conf = mddev->private;
4397 if (len >= PAGE_SIZE)
4402 if (strict_strtoul(page, 10, &new))
4404 if (new > conf->max_nr_stripes)
4406 conf->bypass_threshold = new;
4410 static struct md_sysfs_entry
4411 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4413 raid5_show_preread_threshold,
4414 raid5_store_preread_threshold);
4417 stripe_cache_active_show(struct mddev *mddev, char *page)
4419 struct r5conf *conf = mddev->private;
4421 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4426 static struct md_sysfs_entry
4427 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4429 static struct attribute *raid5_attrs[] = {
4430 &raid5_stripecache_size.attr,
4431 &raid5_stripecache_active.attr,
4432 &raid5_preread_bypass_threshold.attr,
4435 static struct attribute_group raid5_attrs_group = {
4437 .attrs = raid5_attrs,
4441 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4443 struct r5conf *conf = mddev->private;
4446 sectors = mddev->dev_sectors;
4448 /* size is defined by the smallest of previous and new size */
4449 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4451 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4452 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4453 return sectors * (raid_disks - conf->max_degraded);
4456 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
4458 safe_put_page(percpu->spare_page);
4459 kfree(percpu->scribble);
4460 percpu->spare_page = NULL;
4461 percpu->scribble = NULL;
4464 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
4466 if (conf->level == 6 && !percpu->spare_page)
4467 percpu->spare_page = alloc_page(GFP_KERNEL);
4468 if (!percpu->scribble)
4469 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4471 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
4472 free_scratch_buffer(conf, percpu);
4479 static void raid5_free_percpu(struct r5conf *conf)
4486 #ifdef CONFIG_HOTPLUG_CPU
4487 unregister_cpu_notifier(&conf->cpu_notify);
4491 for_each_possible_cpu(cpu)
4492 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
4495 free_percpu(conf->percpu);
4498 static void free_conf(struct r5conf *conf)
4500 shrink_stripes(conf);
4501 raid5_free_percpu(conf);
4503 kfree(conf->stripe_hashtbl);
4507 #ifdef CONFIG_HOTPLUG_CPU
4508 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4511 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4512 long cpu = (long)hcpu;
4513 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4516 case CPU_UP_PREPARE:
4517 case CPU_UP_PREPARE_FROZEN:
4518 if (alloc_scratch_buffer(conf, percpu)) {
4519 pr_err("%s: failed memory allocation for cpu%ld\n",
4521 return notifier_from_errno(-ENOMEM);
4525 case CPU_DEAD_FROZEN:
4526 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
4535 static int raid5_alloc_percpu(struct r5conf *conf)
4540 conf->percpu = alloc_percpu(struct raid5_percpu);
4544 #ifdef CONFIG_HOTPLUG_CPU
4545 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4546 conf->cpu_notify.priority = 0;
4547 err = register_cpu_notifier(&conf->cpu_notify);
4553 for_each_present_cpu(cpu) {
4554 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
4556 pr_err("%s: failed memory allocation for cpu%ld\n",
4566 static struct r5conf *setup_conf(struct mddev *mddev)
4568 struct r5conf *conf;
4569 int raid_disk, memory, max_disks;
4570 struct md_rdev *rdev;
4571 struct disk_info *disk;
4573 if (mddev->new_level != 5
4574 && mddev->new_level != 4
4575 && mddev->new_level != 6) {
4576 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4577 mdname(mddev), mddev->new_level);
4578 return ERR_PTR(-EIO);
4580 if ((mddev->new_level == 5
4581 && !algorithm_valid_raid5(mddev->new_layout)) ||
4582 (mddev->new_level == 6
4583 && !algorithm_valid_raid6(mddev->new_layout))) {
4584 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4585 mdname(mddev), mddev->new_layout);
4586 return ERR_PTR(-EIO);
4588 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4589 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4590 mdname(mddev), mddev->raid_disks);
4591 return ERR_PTR(-EINVAL);
4594 if (!mddev->new_chunk_sectors ||
4595 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4596 !is_power_of_2(mddev->new_chunk_sectors)) {
4597 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4598 mdname(mddev), mddev->new_chunk_sectors << 9);
4599 return ERR_PTR(-EINVAL);
4602 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4605 spin_lock_init(&conf->device_lock);
4606 init_waitqueue_head(&conf->wait_for_stripe);
4607 init_waitqueue_head(&conf->wait_for_overlap);
4608 INIT_LIST_HEAD(&conf->handle_list);
4609 INIT_LIST_HEAD(&conf->hold_list);
4610 INIT_LIST_HEAD(&conf->delayed_list);
4611 INIT_LIST_HEAD(&conf->bitmap_list);
4612 INIT_LIST_HEAD(&conf->inactive_list);
4613 atomic_set(&conf->active_stripes, 0);
4614 atomic_set(&conf->preread_active_stripes, 0);
4615 atomic_set(&conf->active_aligned_reads, 0);
4616 conf->bypass_threshold = BYPASS_THRESHOLD;
4617 conf->recovery_disabled = mddev->recovery_disabled - 1;
4619 conf->raid_disks = mddev->raid_disks;
4620 if (mddev->reshape_position == MaxSector)
4621 conf->previous_raid_disks = mddev->raid_disks;
4623 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4624 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4625 conf->scribble_len = scribble_len(max_disks);
4627 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4632 conf->mddev = mddev;
4634 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4637 conf->level = mddev->new_level;
4638 if (raid5_alloc_percpu(conf) != 0)
4641 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4643 list_for_each_entry(rdev, &mddev->disks, same_set) {
4644 raid_disk = rdev->raid_disk;
4645 if (raid_disk >= max_disks
4648 disk = conf->disks + raid_disk;
4652 if (test_bit(In_sync, &rdev->flags)) {
4653 char b[BDEVNAME_SIZE];
4654 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4656 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4657 } else if (rdev->saved_raid_disk != raid_disk)
4658 /* Cannot rely on bitmap to complete recovery */
4662 conf->chunk_sectors = mddev->new_chunk_sectors;
4663 conf->level = mddev->new_level;
4664 if (conf->level == 6)
4665 conf->max_degraded = 2;
4667 conf->max_degraded = 1;
4668 conf->algorithm = mddev->new_layout;
4669 conf->max_nr_stripes = NR_STRIPES;
4670 conf->reshape_progress = mddev->reshape_position;
4671 if (conf->reshape_progress != MaxSector) {
4672 conf->prev_chunk_sectors = mddev->chunk_sectors;
4673 conf->prev_algo = mddev->layout;
4676 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4677 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4678 if (grow_stripes(conf, conf->max_nr_stripes)) {
4680 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4681 mdname(mddev), memory);
4684 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4685 mdname(mddev), memory);
4687 conf->thread = md_register_thread(raid5d, mddev, NULL);
4688 if (!conf->thread) {
4690 "md/raid:%s: couldn't allocate thread.\n",
4700 return ERR_PTR(-EIO);
4702 return ERR_PTR(-ENOMEM);
4706 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4709 case ALGORITHM_PARITY_0:
4710 if (raid_disk < max_degraded)
4713 case ALGORITHM_PARITY_N:
4714 if (raid_disk >= raid_disks - max_degraded)
4717 case ALGORITHM_PARITY_0_6:
4718 if (raid_disk == 0 ||
4719 raid_disk == raid_disks - 1)
4722 case ALGORITHM_LEFT_ASYMMETRIC_6:
4723 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4724 case ALGORITHM_LEFT_SYMMETRIC_6:
4725 case ALGORITHM_RIGHT_SYMMETRIC_6:
4726 if (raid_disk == raid_disks - 1)
4732 static int run(struct mddev *mddev)
4734 struct r5conf *conf;
4735 int working_disks = 0;
4736 int dirty_parity_disks = 0;
4737 struct md_rdev *rdev;
4738 sector_t reshape_offset = 0;
4740 if (mddev->recovery_cp != MaxSector)
4741 printk(KERN_NOTICE "md/raid:%s: not clean"
4742 " -- starting background reconstruction\n",
4744 if (mddev->reshape_position != MaxSector) {
4745 /* Check that we can continue the reshape.
4746 * Currently only disks can change, it must
4747 * increase, and we must be past the point where
4748 * a stripe over-writes itself
4750 sector_t here_new, here_old;
4752 int max_degraded = (mddev->level == 6 ? 2 : 1);
4754 if (mddev->new_level != mddev->level) {
4755 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4756 "required - aborting.\n",
4760 old_disks = mddev->raid_disks - mddev->delta_disks;
4761 /* reshape_position must be on a new-stripe boundary, and one
4762 * further up in new geometry must map after here in old
4765 here_new = mddev->reshape_position;
4766 if (sector_div(here_new, mddev->new_chunk_sectors *
4767 (mddev->raid_disks - max_degraded))) {
4768 printk(KERN_ERR "md/raid:%s: reshape_position not "
4769 "on a stripe boundary\n", mdname(mddev));
4772 reshape_offset = here_new * mddev->new_chunk_sectors;
4773 /* here_new is the stripe we will write to */
4774 here_old = mddev->reshape_position;
4775 sector_div(here_old, mddev->chunk_sectors *
4776 (old_disks-max_degraded));
4777 /* here_old is the first stripe that we might need to read
4779 if (mddev->delta_disks == 0) {
4780 /* We cannot be sure it is safe to start an in-place
4781 * reshape. It is only safe if user-space if monitoring
4782 * and taking constant backups.
4783 * mdadm always starts a situation like this in
4784 * readonly mode so it can take control before
4785 * allowing any writes. So just check for that.
4787 if ((here_new * mddev->new_chunk_sectors !=
4788 here_old * mddev->chunk_sectors) ||
4790 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4791 " in read-only mode - aborting\n",
4795 } else if (mddev->delta_disks < 0
4796 ? (here_new * mddev->new_chunk_sectors <=
4797 here_old * mddev->chunk_sectors)
4798 : (here_new * mddev->new_chunk_sectors >=
4799 here_old * mddev->chunk_sectors)) {
4800 /* Reading from the same stripe as writing to - bad */
4801 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4802 "auto-recovery - aborting.\n",
4806 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4808 /* OK, we should be able to continue; */
4810 BUG_ON(mddev->level != mddev->new_level);
4811 BUG_ON(mddev->layout != mddev->new_layout);
4812 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4813 BUG_ON(mddev->delta_disks != 0);
4816 if (mddev->private == NULL)
4817 conf = setup_conf(mddev);
4819 conf = mddev->private;
4822 return PTR_ERR(conf);
4824 mddev->thread = conf->thread;
4825 conf->thread = NULL;
4826 mddev->private = conf;
4829 * 0 for a fully functional array, 1 or 2 for a degraded array.
4831 list_for_each_entry(rdev, &mddev->disks, same_set) {
4832 if (rdev->raid_disk < 0)
4834 if (test_bit(In_sync, &rdev->flags)) {
4838 /* This disc is not fully in-sync. However if it
4839 * just stored parity (beyond the recovery_offset),
4840 * when we don't need to be concerned about the
4841 * array being dirty.
4842 * When reshape goes 'backwards', we never have
4843 * partially completed devices, so we only need
4844 * to worry about reshape going forwards.
4846 /* Hack because v0.91 doesn't store recovery_offset properly. */
4847 if (mddev->major_version == 0 &&
4848 mddev->minor_version > 90)
4849 rdev->recovery_offset = reshape_offset;
4851 if (rdev->recovery_offset < reshape_offset) {
4852 /* We need to check old and new layout */
4853 if (!only_parity(rdev->raid_disk,
4856 conf->max_degraded))
4859 if (!only_parity(rdev->raid_disk,
4861 conf->previous_raid_disks,
4862 conf->max_degraded))
4864 dirty_parity_disks++;
4867 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4870 if (has_failed(conf)) {
4871 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4872 " (%d/%d failed)\n",
4873 mdname(mddev), mddev->degraded, conf->raid_disks);
4877 /* device size must be a multiple of chunk size */
4878 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4879 mddev->resync_max_sectors = mddev->dev_sectors;
4881 if (mddev->degraded > dirty_parity_disks &&
4882 mddev->recovery_cp != MaxSector) {
4883 if (mddev->ok_start_degraded)
4885 "md/raid:%s: starting dirty degraded array"
4886 " - data corruption possible.\n",
4890 "md/raid:%s: cannot start dirty degraded array.\n",
4896 if (mddev->degraded == 0)
4897 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4898 " devices, algorithm %d\n", mdname(mddev), conf->level,
4899 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4902 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4903 " out of %d devices, algorithm %d\n",
4904 mdname(mddev), conf->level,
4905 mddev->raid_disks - mddev->degraded,
4906 mddev->raid_disks, mddev->new_layout);
4908 print_raid5_conf(conf);
4910 if (conf->reshape_progress != MaxSector) {
4911 conf->reshape_safe = conf->reshape_progress;
4912 atomic_set(&conf->reshape_stripes, 0);
4913 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4914 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4915 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4916 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4917 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4922 /* Ok, everything is just fine now */
4923 if (mddev->to_remove == &raid5_attrs_group)
4924 mddev->to_remove = NULL;
4925 else if (mddev->kobj.sd &&
4926 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4928 "raid5: failed to create sysfs attributes for %s\n",
4930 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4934 /* read-ahead size must cover two whole stripes, which
4935 * is 2 * (datadisks) * chunksize where 'n' is the
4936 * number of raid devices
4938 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4939 int stripe = data_disks *
4940 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4941 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4942 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4944 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4946 mddev->queue->backing_dev_info.congested_data = mddev;
4947 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4949 chunk_size = mddev->chunk_sectors << 9;
4950 blk_queue_io_min(mddev->queue, chunk_size);
4951 blk_queue_io_opt(mddev->queue, chunk_size *
4952 (conf->raid_disks - conf->max_degraded));
4954 list_for_each_entry(rdev, &mddev->disks, same_set)
4955 disk_stack_limits(mddev->gendisk, rdev->bdev,
4956 rdev->data_offset << 9);
4961 md_unregister_thread(&mddev->thread);
4962 print_raid5_conf(conf);
4964 mddev->private = NULL;
4965 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
4969 static int stop(struct mddev *mddev)
4971 struct r5conf *conf = mddev->private;
4973 md_unregister_thread(&mddev->thread);
4975 mddev->queue->backing_dev_info.congested_fn = NULL;
4977 mddev->private = NULL;
4978 mddev->to_remove = &raid5_attrs_group;
4982 static void status(struct seq_file *seq, struct mddev *mddev)
4984 struct r5conf *conf = mddev->private;
4987 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4988 mddev->chunk_sectors / 2, mddev->layout);
4989 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4990 for (i = 0; i < conf->raid_disks; i++)
4991 seq_printf (seq, "%s",
4992 conf->disks[i].rdev &&
4993 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4994 seq_printf (seq, "]");
4997 static void print_raid5_conf (struct r5conf *conf)
5000 struct disk_info *tmp;
5002 printk(KERN_DEBUG "RAID conf printout:\n");
5004 printk("(conf==NULL)\n");
5007 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5009 conf->raid_disks - conf->mddev->degraded);
5011 for (i = 0; i < conf->raid_disks; i++) {
5012 char b[BDEVNAME_SIZE];
5013 tmp = conf->disks + i;
5015 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5016 i, !test_bit(Faulty, &tmp->rdev->flags),
5017 bdevname(tmp->rdev->bdev, b));
5021 static int raid5_spare_active(struct mddev *mddev)
5024 struct r5conf *conf = mddev->private;
5025 struct disk_info *tmp;
5027 unsigned long flags;
5029 for (i = 0; i < conf->raid_disks; i++) {
5030 tmp = conf->disks + i;
5032 && tmp->rdev->recovery_offset == MaxSector
5033 && !test_bit(Faulty, &tmp->rdev->flags)
5034 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5036 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5039 spin_lock_irqsave(&conf->device_lock, flags);
5040 mddev->degraded -= count;
5041 spin_unlock_irqrestore(&conf->device_lock, flags);
5042 print_raid5_conf(conf);
5046 static int raid5_remove_disk(struct mddev *mddev, int number)
5048 struct r5conf *conf = mddev->private;
5050 struct md_rdev *rdev;
5051 struct disk_info *p = conf->disks + number;
5053 print_raid5_conf(conf);
5056 if (number >= conf->raid_disks &&
5057 conf->reshape_progress == MaxSector)
5058 clear_bit(In_sync, &rdev->flags);
5060 if (test_bit(In_sync, &rdev->flags) ||
5061 atomic_read(&rdev->nr_pending)) {
5065 /* Only remove non-faulty devices if recovery
5068 if (!test_bit(Faulty, &rdev->flags) &&
5069 mddev->recovery_disabled != conf->recovery_disabled &&
5070 !has_failed(conf) &&
5071 number < conf->raid_disks) {
5077 if (atomic_read(&rdev->nr_pending)) {
5078 /* lost the race, try later */
5085 print_raid5_conf(conf);
5089 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5091 struct r5conf *conf = mddev->private;
5094 struct disk_info *p;
5096 int last = conf->raid_disks - 1;
5098 if (mddev->recovery_disabled == conf->recovery_disabled)
5101 if (has_failed(conf))
5102 /* no point adding a device */
5105 if (rdev->raid_disk >= 0)
5106 first = last = rdev->raid_disk;
5109 * find the disk ... but prefer rdev->saved_raid_disk
5112 if (rdev->saved_raid_disk >= 0 &&
5113 rdev->saved_raid_disk >= first &&
5114 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5115 disk = rdev->saved_raid_disk;
5118 for ( ; disk <= last ; disk++)
5119 if ((p=conf->disks + disk)->rdev == NULL) {
5120 clear_bit(In_sync, &rdev->flags);
5121 rdev->raid_disk = disk;
5123 if (rdev->saved_raid_disk != disk)
5125 rcu_assign_pointer(p->rdev, rdev);
5128 print_raid5_conf(conf);
5132 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5134 /* no resync is happening, and there is enough space
5135 * on all devices, so we can resize.
5136 * We need to make sure resync covers any new space.
5137 * If the array is shrinking we should possibly wait until
5138 * any io in the removed space completes, but it hardly seems
5141 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5142 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5143 mddev->raid_disks));
5144 if (mddev->array_sectors >
5145 raid5_size(mddev, sectors, mddev->raid_disks))
5147 set_capacity(mddev->gendisk, mddev->array_sectors);
5148 revalidate_disk(mddev->gendisk);
5149 if (sectors > mddev->dev_sectors &&
5150 mddev->recovery_cp > mddev->dev_sectors) {
5151 mddev->recovery_cp = mddev->dev_sectors;
5152 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5154 mddev->dev_sectors = sectors;
5155 mddev->resync_max_sectors = sectors;
5159 static int check_stripe_cache(struct mddev *mddev)
5161 /* Can only proceed if there are plenty of stripe_heads.
5162 * We need a minimum of one full stripe,, and for sensible progress
5163 * it is best to have about 4 times that.
5164 * If we require 4 times, then the default 256 4K stripe_heads will
5165 * allow for chunk sizes up to 256K, which is probably OK.
5166 * If the chunk size is greater, user-space should request more
5167 * stripe_heads first.
5169 struct r5conf *conf = mddev->private;
5170 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5171 > conf->max_nr_stripes ||
5172 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5173 > conf->max_nr_stripes) {
5174 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5176 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5183 static int check_reshape(struct mddev *mddev)
5185 struct r5conf *conf = mddev->private;
5187 if (mddev->delta_disks == 0 &&
5188 mddev->new_layout == mddev->layout &&
5189 mddev->new_chunk_sectors == mddev->chunk_sectors)
5190 return 0; /* nothing to do */
5192 /* Cannot grow a bitmap yet */
5194 if (has_failed(conf))
5196 if (mddev->delta_disks < 0) {
5197 /* We might be able to shrink, but the devices must
5198 * be made bigger first.
5199 * For raid6, 4 is the minimum size.
5200 * Otherwise 2 is the minimum
5203 if (mddev->level == 6)
5205 if (mddev->raid_disks + mddev->delta_disks < min)
5209 if (!check_stripe_cache(mddev))
5212 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5215 static int raid5_start_reshape(struct mddev *mddev)
5217 struct r5conf *conf = mddev->private;
5218 struct md_rdev *rdev;
5220 unsigned long flags;
5222 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5225 if (!check_stripe_cache(mddev))
5228 list_for_each_entry(rdev, &mddev->disks, same_set)
5229 if (!test_bit(In_sync, &rdev->flags)
5230 && !test_bit(Faulty, &rdev->flags))
5233 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5234 /* Not enough devices even to make a degraded array
5239 /* Refuse to reduce size of the array. Any reductions in
5240 * array size must be through explicit setting of array_size
5243 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5244 < mddev->array_sectors) {
5245 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5246 "before number of disks\n", mdname(mddev));
5250 atomic_set(&conf->reshape_stripes, 0);
5251 spin_lock_irq(&conf->device_lock);
5252 conf->previous_raid_disks = conf->raid_disks;
5253 conf->raid_disks += mddev->delta_disks;
5254 conf->prev_chunk_sectors = conf->chunk_sectors;
5255 conf->chunk_sectors = mddev->new_chunk_sectors;
5256 conf->prev_algo = conf->algorithm;
5257 conf->algorithm = mddev->new_layout;
5258 if (mddev->delta_disks < 0)
5259 conf->reshape_progress = raid5_size(mddev, 0, 0);
5261 conf->reshape_progress = 0;
5262 conf->reshape_safe = conf->reshape_progress;
5264 spin_unlock_irq(&conf->device_lock);
5266 /* Add some new drives, as many as will fit.
5267 * We know there are enough to make the newly sized array work.
5268 * Don't add devices if we are reducing the number of
5269 * devices in the array. This is because it is not possible
5270 * to correctly record the "partially reconstructed" state of
5271 * such devices during the reshape and confusion could result.
5273 if (mddev->delta_disks >= 0) {
5274 int added_devices = 0;
5275 list_for_each_entry(rdev, &mddev->disks, same_set)
5276 if (rdev->raid_disk < 0 &&
5277 !test_bit(Faulty, &rdev->flags)) {
5278 if (raid5_add_disk(mddev, rdev) == 0) {
5280 >= conf->previous_raid_disks) {
5281 set_bit(In_sync, &rdev->flags);
5284 rdev->recovery_offset = 0;
5286 if (sysfs_link_rdev(mddev, rdev))
5287 /* Failure here is OK */;
5289 } else if (rdev->raid_disk >= conf->previous_raid_disks
5290 && !test_bit(Faulty, &rdev->flags)) {
5291 /* This is a spare that was manually added */
5292 set_bit(In_sync, &rdev->flags);
5296 /* When a reshape changes the number of devices,
5297 * ->degraded is measured against the larger of the
5298 * pre and post number of devices.
5300 spin_lock_irqsave(&conf->device_lock, flags);
5301 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5303 spin_unlock_irqrestore(&conf->device_lock, flags);
5305 mddev->raid_disks = conf->raid_disks;
5306 mddev->reshape_position = conf->reshape_progress;
5307 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5309 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5310 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5311 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5312 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5313 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5315 if (!mddev->sync_thread) {
5316 mddev->recovery = 0;
5317 spin_lock_irq(&conf->device_lock);
5318 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5319 conf->reshape_progress = MaxSector;
5320 spin_unlock_irq(&conf->device_lock);
5323 conf->reshape_checkpoint = jiffies;
5324 md_wakeup_thread(mddev->sync_thread);
5325 md_new_event(mddev);
5329 /* This is called from the reshape thread and should make any
5330 * changes needed in 'conf'
5332 static void end_reshape(struct r5conf *conf)
5335 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5337 spin_lock_irq(&conf->device_lock);
5338 conf->previous_raid_disks = conf->raid_disks;
5339 conf->reshape_progress = MaxSector;
5340 spin_unlock_irq(&conf->device_lock);
5341 wake_up(&conf->wait_for_overlap);
5343 /* read-ahead size must cover two whole stripes, which is
5344 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5346 if (conf->mddev->queue) {
5347 int data_disks = conf->raid_disks - conf->max_degraded;
5348 int stripe = data_disks * ((conf->chunk_sectors << 9)
5350 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5351 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5356 /* This is called from the raid5d thread with mddev_lock held.
5357 * It makes config changes to the device.
5359 static void raid5_finish_reshape(struct mddev *mddev)
5361 struct r5conf *conf = mddev->private;
5363 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5365 if (mddev->delta_disks > 0) {
5366 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5367 set_capacity(mddev->gendisk, mddev->array_sectors);
5368 revalidate_disk(mddev->gendisk);
5371 mddev->degraded = conf->raid_disks;
5372 for (d = 0; d < conf->raid_disks ; d++)
5373 if (conf->disks[d].rdev &&
5375 &conf->disks[d].rdev->flags))
5377 for (d = conf->raid_disks ;
5378 d < conf->raid_disks - mddev->delta_disks;
5380 struct md_rdev *rdev = conf->disks[d].rdev;
5381 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5382 sysfs_unlink_rdev(mddev, rdev);
5383 rdev->raid_disk = -1;
5387 mddev->layout = conf->algorithm;
5388 mddev->chunk_sectors = conf->chunk_sectors;
5389 mddev->reshape_position = MaxSector;
5390 mddev->delta_disks = 0;
5394 static void raid5_quiesce(struct mddev *mddev, int state)
5396 struct r5conf *conf = mddev->private;
5399 case 2: /* resume for a suspend */
5400 wake_up(&conf->wait_for_overlap);
5403 case 1: /* stop all writes */
5404 spin_lock_irq(&conf->device_lock);
5405 /* '2' tells resync/reshape to pause so that all
5406 * active stripes can drain
5409 wait_event_lock_irq(conf->wait_for_stripe,
5410 atomic_read(&conf->active_stripes) == 0 &&
5411 atomic_read(&conf->active_aligned_reads) == 0,
5412 conf->device_lock, /* nothing */);
5414 spin_unlock_irq(&conf->device_lock);
5415 /* allow reshape to continue */
5416 wake_up(&conf->wait_for_overlap);
5419 case 0: /* re-enable writes */
5420 spin_lock_irq(&conf->device_lock);
5422 wake_up(&conf->wait_for_stripe);
5423 wake_up(&conf->wait_for_overlap);
5424 spin_unlock_irq(&conf->device_lock);
5430 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5432 struct r0conf *raid0_conf = mddev->private;
5435 /* for raid0 takeover only one zone is supported */
5436 if (raid0_conf->nr_strip_zones > 1) {
5437 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5439 return ERR_PTR(-EINVAL);
5442 sectors = raid0_conf->strip_zone[0].zone_end;
5443 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5444 mddev->dev_sectors = sectors;
5445 mddev->new_level = level;
5446 mddev->new_layout = ALGORITHM_PARITY_N;
5447 mddev->new_chunk_sectors = mddev->chunk_sectors;
5448 mddev->raid_disks += 1;
5449 mddev->delta_disks = 1;
5450 /* make sure it will be not marked as dirty */
5451 mddev->recovery_cp = MaxSector;
5453 return setup_conf(mddev);
5457 static void *raid5_takeover_raid1(struct mddev *mddev)
5461 if (mddev->raid_disks != 2 ||
5462 mddev->degraded > 1)
5463 return ERR_PTR(-EINVAL);
5465 /* Should check if there are write-behind devices? */
5467 chunksect = 64*2; /* 64K by default */
5469 /* The array must be an exact multiple of chunksize */
5470 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5473 if ((chunksect<<9) < STRIPE_SIZE)
5474 /* array size does not allow a suitable chunk size */
5475 return ERR_PTR(-EINVAL);
5477 mddev->new_level = 5;
5478 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5479 mddev->new_chunk_sectors = chunksect;
5481 return setup_conf(mddev);
5484 static void *raid5_takeover_raid6(struct mddev *mddev)
5488 switch (mddev->layout) {
5489 case ALGORITHM_LEFT_ASYMMETRIC_6:
5490 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5492 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5493 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5495 case ALGORITHM_LEFT_SYMMETRIC_6:
5496 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5498 case ALGORITHM_RIGHT_SYMMETRIC_6:
5499 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5501 case ALGORITHM_PARITY_0_6:
5502 new_layout = ALGORITHM_PARITY_0;
5504 case ALGORITHM_PARITY_N:
5505 new_layout = ALGORITHM_PARITY_N;
5508 return ERR_PTR(-EINVAL);
5510 mddev->new_level = 5;
5511 mddev->new_layout = new_layout;
5512 mddev->delta_disks = -1;
5513 mddev->raid_disks -= 1;
5514 return setup_conf(mddev);
5518 static int raid5_check_reshape(struct mddev *mddev)
5520 /* For a 2-drive array, the layout and chunk size can be changed
5521 * immediately as not restriping is needed.
5522 * For larger arrays we record the new value - after validation
5523 * to be used by a reshape pass.
5525 struct r5conf *conf = mddev->private;
5526 int new_chunk = mddev->new_chunk_sectors;
5528 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5530 if (new_chunk > 0) {
5531 if (!is_power_of_2(new_chunk))
5533 if (new_chunk < (PAGE_SIZE>>9))
5535 if (mddev->array_sectors & (new_chunk-1))
5536 /* not factor of array size */
5540 /* They look valid */
5542 if (mddev->raid_disks == 2) {
5543 /* can make the change immediately */
5544 if (mddev->new_layout >= 0) {
5545 conf->algorithm = mddev->new_layout;
5546 mddev->layout = mddev->new_layout;
5548 if (new_chunk > 0) {
5549 conf->chunk_sectors = new_chunk ;
5550 mddev->chunk_sectors = new_chunk;
5552 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5553 md_wakeup_thread(mddev->thread);
5555 return check_reshape(mddev);
5558 static int raid6_check_reshape(struct mddev *mddev)
5560 int new_chunk = mddev->new_chunk_sectors;
5562 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5564 if (new_chunk > 0) {
5565 if (!is_power_of_2(new_chunk))
5567 if (new_chunk < (PAGE_SIZE >> 9))
5569 if (mddev->array_sectors & (new_chunk-1))
5570 /* not factor of array size */
5574 /* They look valid */
5575 return check_reshape(mddev);
5578 static void *raid5_takeover(struct mddev *mddev)
5580 /* raid5 can take over:
5581 * raid0 - if there is only one strip zone - make it a raid4 layout
5582 * raid1 - if there are two drives. We need to know the chunk size
5583 * raid4 - trivial - just use a raid4 layout.
5584 * raid6 - Providing it is a *_6 layout
5586 if (mddev->level == 0)
5587 return raid45_takeover_raid0(mddev, 5);
5588 if (mddev->level == 1)
5589 return raid5_takeover_raid1(mddev);
5590 if (mddev->level == 4) {
5591 mddev->new_layout = ALGORITHM_PARITY_N;
5592 mddev->new_level = 5;
5593 return setup_conf(mddev);
5595 if (mddev->level == 6)
5596 return raid5_takeover_raid6(mddev);
5598 return ERR_PTR(-EINVAL);
5601 static void *raid4_takeover(struct mddev *mddev)
5603 /* raid4 can take over:
5604 * raid0 - if there is only one strip zone
5605 * raid5 - if layout is right
5607 if (mddev->level == 0)
5608 return raid45_takeover_raid0(mddev, 4);
5609 if (mddev->level == 5 &&
5610 mddev->layout == ALGORITHM_PARITY_N) {
5611 mddev->new_layout = 0;
5612 mddev->new_level = 4;
5613 return setup_conf(mddev);
5615 return ERR_PTR(-EINVAL);
5618 static struct md_personality raid5_personality;
5620 static void *raid6_takeover(struct mddev *mddev)
5622 /* Currently can only take over a raid5. We map the
5623 * personality to an equivalent raid6 personality
5624 * with the Q block at the end.
5628 if (mddev->pers != &raid5_personality)
5629 return ERR_PTR(-EINVAL);
5630 if (mddev->degraded > 1)
5631 return ERR_PTR(-EINVAL);
5632 if (mddev->raid_disks > 253)
5633 return ERR_PTR(-EINVAL);
5634 if (mddev->raid_disks < 3)
5635 return ERR_PTR(-EINVAL);
5637 switch (mddev->layout) {
5638 case ALGORITHM_LEFT_ASYMMETRIC:
5639 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5641 case ALGORITHM_RIGHT_ASYMMETRIC:
5642 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5644 case ALGORITHM_LEFT_SYMMETRIC:
5645 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5647 case ALGORITHM_RIGHT_SYMMETRIC:
5648 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5650 case ALGORITHM_PARITY_0:
5651 new_layout = ALGORITHM_PARITY_0_6;
5653 case ALGORITHM_PARITY_N:
5654 new_layout = ALGORITHM_PARITY_N;
5657 return ERR_PTR(-EINVAL);
5659 mddev->new_level = 6;
5660 mddev->new_layout = new_layout;
5661 mddev->delta_disks = 1;
5662 mddev->raid_disks += 1;
5663 return setup_conf(mddev);
5667 static struct md_personality raid6_personality =
5671 .owner = THIS_MODULE,
5672 .make_request = make_request,
5676 .error_handler = error,
5677 .hot_add_disk = raid5_add_disk,
5678 .hot_remove_disk= raid5_remove_disk,
5679 .spare_active = raid5_spare_active,
5680 .sync_request = sync_request,
5681 .resize = raid5_resize,
5683 .check_reshape = raid6_check_reshape,
5684 .start_reshape = raid5_start_reshape,
5685 .finish_reshape = raid5_finish_reshape,
5686 .quiesce = raid5_quiesce,
5687 .takeover = raid6_takeover,
5689 static struct md_personality raid5_personality =
5693 .owner = THIS_MODULE,
5694 .make_request = make_request,
5698 .error_handler = error,
5699 .hot_add_disk = raid5_add_disk,
5700 .hot_remove_disk= raid5_remove_disk,
5701 .spare_active = raid5_spare_active,
5702 .sync_request = sync_request,
5703 .resize = raid5_resize,
5705 .check_reshape = raid5_check_reshape,
5706 .start_reshape = raid5_start_reshape,
5707 .finish_reshape = raid5_finish_reshape,
5708 .quiesce = raid5_quiesce,
5709 .takeover = raid5_takeover,
5712 static struct md_personality raid4_personality =
5716 .owner = THIS_MODULE,
5717 .make_request = make_request,
5721 .error_handler = error,
5722 .hot_add_disk = raid5_add_disk,
5723 .hot_remove_disk= raid5_remove_disk,
5724 .spare_active = raid5_spare_active,
5725 .sync_request = sync_request,
5726 .resize = raid5_resize,
5728 .check_reshape = raid5_check_reshape,
5729 .start_reshape = raid5_start_reshape,
5730 .finish_reshape = raid5_finish_reshape,
5731 .quiesce = raid5_quiesce,
5732 .takeover = raid4_takeover,
5735 static int __init raid5_init(void)
5737 register_md_personality(&raid6_personality);
5738 register_md_personality(&raid5_personality);
5739 register_md_personality(&raid4_personality);
5743 static void raid5_exit(void)
5745 unregister_md_personality(&raid6_personality);
5746 unregister_md_personality(&raid5_personality);
5747 unregister_md_personality(&raid4_personality);
5750 module_init(raid5_init);
5751 module_exit(raid5_exit);
5752 MODULE_LICENSE("GPL");
5753 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5754 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5755 MODULE_ALIAS("md-raid5");
5756 MODULE_ALIAS("md-raid4");
5757 MODULE_ALIAS("md-level-5");
5758 MODULE_ALIAS("md-level-4");
5759 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5760 MODULE_ALIAS("md-raid6");
5761 MODULE_ALIAS("md-level-6");
5763 /* This used to be two separate modules, they were: */
5764 MODULE_ALIAS("raid5");
5765 MODULE_ALIAS("raid6");