2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
65 #define NR_STRIPES 256
66 #define STRIPE_SIZE PAGE_SIZE
67 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
68 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
69 #define IO_THRESHOLD 1
70 #define BYPASS_THRESHOLD 1
71 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
72 #define HASH_MASK (NR_HASH - 1)
74 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
76 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
77 return &conf->stripe_hashtbl[hash];
80 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
81 * order without overlap. There may be several bio's per stripe+device, and
82 * a bio could span several devices.
83 * When walking this list for a particular stripe+device, we must never proceed
84 * beyond a bio that extends past this device, as the next bio might no longer
86 * This function is used to determine the 'next' bio in the list, given the sector
87 * of the current stripe+device
89 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
91 int sectors = bio->bi_size >> 9;
92 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
99 * We maintain a biased count of active stripes in the bottom 16 bits of
100 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
102 static inline int raid5_bi_processed_stripes(struct bio *bio)
104 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
105 return (atomic_read(segments) >> 16) & 0xffff;
108 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
110 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
111 return atomic_sub_return(1, segments) & 0xffff;
114 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
116 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
117 atomic_inc(segments);
120 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
123 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
127 old = atomic_read(segments);
128 new = (old & 0xffff) | (cnt << 16);
129 } while (atomic_cmpxchg(segments, old, new) != old);
132 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
134 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
135 atomic_set(segments, cnt);
138 /* Find first data disk in a raid6 stripe */
139 static inline int raid6_d0(struct stripe_head *sh)
142 /* ddf always start from first device */
144 /* md starts just after Q block */
145 if (sh->qd_idx == sh->disks - 1)
148 return sh->qd_idx + 1;
150 static inline int raid6_next_disk(int disk, int raid_disks)
153 return (disk < raid_disks) ? disk : 0;
156 /* When walking through the disks in a raid5, starting at raid6_d0,
157 * We need to map each disk to a 'slot', where the data disks are slot
158 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
159 * is raid_disks-1. This help does that mapping.
161 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
162 int *count, int syndrome_disks)
168 if (idx == sh->pd_idx)
169 return syndrome_disks;
170 if (idx == sh->qd_idx)
171 return syndrome_disks + 1;
177 static void return_io(struct bio *return_bi)
179 struct bio *bi = return_bi;
182 return_bi = bi->bi_next;
190 static void print_raid5_conf (struct r5conf *conf);
192 static int stripe_operations_active(struct stripe_head *sh)
194 return sh->check_state || sh->reconstruct_state ||
195 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
196 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
199 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
201 BUG_ON(!list_empty(&sh->lru));
202 BUG_ON(atomic_read(&conf->active_stripes)==0);
203 if (test_bit(STRIPE_HANDLE, &sh->state)) {
204 if (test_bit(STRIPE_DELAYED, &sh->state) &&
205 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
206 list_add_tail(&sh->lru, &conf->delayed_list);
207 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
208 sh->bm_seq - conf->seq_write > 0)
209 list_add_tail(&sh->lru, &conf->bitmap_list);
211 clear_bit(STRIPE_DELAYED, &sh->state);
212 clear_bit(STRIPE_BIT_DELAY, &sh->state);
213 list_add_tail(&sh->lru, &conf->handle_list);
215 md_wakeup_thread(conf->mddev->thread);
217 BUG_ON(stripe_operations_active(sh));
218 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
219 if (atomic_dec_return(&conf->preread_active_stripes)
221 md_wakeup_thread(conf->mddev->thread);
222 atomic_dec(&conf->active_stripes);
223 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
224 list_add_tail(&sh->lru, &conf->inactive_list);
225 wake_up(&conf->wait_for_stripe);
226 if (conf->retry_read_aligned)
227 md_wakeup_thread(conf->mddev->thread);
232 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
234 if (atomic_dec_and_test(&sh->count))
235 do_release_stripe(conf, sh);
238 static void release_stripe(struct stripe_head *sh)
240 struct r5conf *conf = sh->raid_conf;
243 local_irq_save(flags);
244 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
245 do_release_stripe(conf, sh);
246 spin_unlock(&conf->device_lock);
248 local_irq_restore(flags);
251 static inline void remove_hash(struct stripe_head *sh)
253 pr_debug("remove_hash(), stripe %llu\n",
254 (unsigned long long)sh->sector);
256 hlist_del_init(&sh->hash);
259 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
261 struct hlist_head *hp = stripe_hash(conf, sh->sector);
263 pr_debug("insert_hash(), stripe %llu\n",
264 (unsigned long long)sh->sector);
266 hlist_add_head(&sh->hash, hp);
270 /* find an idle stripe, make sure it is unhashed, and return it. */
271 static struct stripe_head *get_free_stripe(struct r5conf *conf)
273 struct stripe_head *sh = NULL;
274 struct list_head *first;
276 if (list_empty(&conf->inactive_list))
278 first = conf->inactive_list.next;
279 sh = list_entry(first, struct stripe_head, lru);
280 list_del_init(first);
282 atomic_inc(&conf->active_stripes);
287 static void shrink_buffers(struct stripe_head *sh)
291 int num = sh->raid_conf->pool_size;
293 for (i = 0; i < num ; i++) {
297 sh->dev[i].page = NULL;
302 static int grow_buffers(struct stripe_head *sh)
305 int num = sh->raid_conf->pool_size;
307 for (i = 0; i < num; i++) {
310 if (!(page = alloc_page(GFP_KERNEL))) {
313 sh->dev[i].page = page;
318 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
319 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
320 struct stripe_head *sh);
322 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
324 struct r5conf *conf = sh->raid_conf;
327 BUG_ON(atomic_read(&sh->count) != 0);
328 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
329 BUG_ON(stripe_operations_active(sh));
331 pr_debug("init_stripe called, stripe %llu\n",
332 (unsigned long long)sh->sector);
336 sh->generation = conf->generation - previous;
337 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
339 stripe_set_idx(sector, conf, previous, sh);
343 for (i = sh->disks; i--; ) {
344 struct r5dev *dev = &sh->dev[i];
346 if (dev->toread || dev->read || dev->towrite || dev->written ||
347 test_bit(R5_LOCKED, &dev->flags)) {
348 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
349 (unsigned long long)sh->sector, i, dev->toread,
350 dev->read, dev->towrite, dev->written,
351 test_bit(R5_LOCKED, &dev->flags));
355 raid5_build_block(sh, i, previous);
357 insert_hash(conf, sh);
360 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
363 struct stripe_head *sh;
364 struct hlist_node *hn;
366 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
367 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
368 if (sh->sector == sector && sh->generation == generation)
370 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
375 * Need to check if array has failed when deciding whether to:
377 * - remove non-faulty devices
380 * This determination is simple when no reshape is happening.
381 * However if there is a reshape, we need to carefully check
382 * both the before and after sections.
383 * This is because some failed devices may only affect one
384 * of the two sections, and some non-in_sync devices may
385 * be insync in the section most affected by failed devices.
387 static int calc_degraded(struct r5conf *conf)
389 int degraded, degraded2;
394 for (i = 0; i < conf->previous_raid_disks; i++) {
395 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
396 if (!rdev || test_bit(Faulty, &rdev->flags))
398 else if (test_bit(In_sync, &rdev->flags))
401 /* not in-sync or faulty.
402 * If the reshape increases the number of devices,
403 * this is being recovered by the reshape, so
404 * this 'previous' section is not in_sync.
405 * If the number of devices is being reduced however,
406 * the device can only be part of the array if
407 * we are reverting a reshape, so this section will
410 if (conf->raid_disks >= conf->previous_raid_disks)
414 if (conf->raid_disks == conf->previous_raid_disks)
418 for (i = 0; i < conf->raid_disks; i++) {
419 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
420 if (!rdev || test_bit(Faulty, &rdev->flags))
422 else if (test_bit(In_sync, &rdev->flags))
425 /* not in-sync or faulty.
426 * If reshape increases the number of devices, this
427 * section has already been recovered, else it
428 * almost certainly hasn't.
430 if (conf->raid_disks <= conf->previous_raid_disks)
434 if (degraded2 > degraded)
439 static int has_failed(struct r5conf *conf)
443 if (conf->mddev->reshape_position == MaxSector)
444 return conf->mddev->degraded > conf->max_degraded;
446 degraded = calc_degraded(conf);
447 if (degraded > conf->max_degraded)
452 static struct stripe_head *
453 get_active_stripe(struct r5conf *conf, sector_t sector,
454 int previous, int noblock, int noquiesce)
456 struct stripe_head *sh;
458 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
460 spin_lock_irq(&conf->device_lock);
463 wait_event_lock_irq(conf->wait_for_stripe,
464 conf->quiesce == 0 || noquiesce,
465 conf->device_lock, /* nothing */);
466 sh = __find_stripe(conf, sector, conf->generation - previous);
468 if (!conf->inactive_blocked)
469 sh = get_free_stripe(conf);
470 if (noblock && sh == NULL)
473 conf->inactive_blocked = 1;
474 wait_event_lock_irq(conf->wait_for_stripe,
475 !list_empty(&conf->inactive_list) &&
476 (atomic_read(&conf->active_stripes)
477 < (conf->max_nr_stripes *3/4)
478 || !conf->inactive_blocked),
481 conf->inactive_blocked = 0;
483 init_stripe(sh, sector, previous);
485 if (atomic_read(&sh->count)) {
486 BUG_ON(!list_empty(&sh->lru)
487 && !test_bit(STRIPE_EXPANDING, &sh->state));
489 if (!test_bit(STRIPE_HANDLE, &sh->state))
490 atomic_inc(&conf->active_stripes);
491 if (list_empty(&sh->lru) &&
492 !test_bit(STRIPE_EXPANDING, &sh->state))
494 list_del_init(&sh->lru);
497 } while (sh == NULL);
500 atomic_inc(&sh->count);
502 spin_unlock_irq(&conf->device_lock);
506 /* Determine if 'data_offset' or 'new_data_offset' should be used
507 * in this stripe_head.
509 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
511 sector_t progress = conf->reshape_progress;
512 /* Need a memory barrier to make sure we see the value
513 * of conf->generation, or ->data_offset that was set before
514 * reshape_progress was updated.
517 if (progress == MaxSector)
519 if (sh->generation == conf->generation - 1)
521 /* We are in a reshape, and this is a new-generation stripe,
522 * so use new_data_offset.
528 raid5_end_read_request(struct bio *bi, int error);
530 raid5_end_write_request(struct bio *bi, int error);
532 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
534 struct r5conf *conf = sh->raid_conf;
535 int i, disks = sh->disks;
539 for (i = disks; i--; ) {
541 int replace_only = 0;
542 struct bio *bi, *rbi;
543 struct md_rdev *rdev, *rrdev = NULL;
544 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
545 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
549 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
551 else if (test_and_clear_bit(R5_WantReplace,
552 &sh->dev[i].flags)) {
557 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
560 bi = &sh->dev[i].req;
561 rbi = &sh->dev[i].rreq; /* For writing to replacement */
566 bi->bi_end_io = raid5_end_write_request;
567 rbi->bi_end_io = raid5_end_write_request;
569 bi->bi_end_io = raid5_end_read_request;
572 rrdev = rcu_dereference(conf->disks[i].replacement);
573 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
574 rdev = rcu_dereference(conf->disks[i].rdev);
583 /* We raced and saw duplicates */
586 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
591 if (rdev && test_bit(Faulty, &rdev->flags))
594 atomic_inc(&rdev->nr_pending);
595 if (rrdev && test_bit(Faulty, &rrdev->flags))
598 atomic_inc(&rrdev->nr_pending);
601 /* We have already checked bad blocks for reads. Now
602 * need to check for writes. We never accept write errors
603 * on the replacement, so we don't to check rrdev.
605 while ((rw & WRITE) && rdev &&
606 test_bit(WriteErrorSeen, &rdev->flags)) {
609 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
610 &first_bad, &bad_sectors);
615 set_bit(BlockedBadBlocks, &rdev->flags);
616 if (!conf->mddev->external &&
617 conf->mddev->flags) {
618 /* It is very unlikely, but we might
619 * still need to write out the
620 * bad block log - better give it
622 md_check_recovery(conf->mddev);
625 * Because md_wait_for_blocked_rdev
626 * will dec nr_pending, we must
627 * increment it first.
629 atomic_inc(&rdev->nr_pending);
630 md_wait_for_blocked_rdev(rdev, conf->mddev);
632 /* Acknowledged bad block - skip the write */
633 rdev_dec_pending(rdev, conf->mddev);
639 if (s->syncing || s->expanding || s->expanded
641 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
643 set_bit(STRIPE_IO_STARTED, &sh->state);
645 bi->bi_bdev = rdev->bdev;
646 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
647 __func__, (unsigned long long)sh->sector,
649 atomic_inc(&sh->count);
650 if (use_new_offset(conf, sh))
651 bi->bi_sector = (sh->sector
652 + rdev->new_data_offset);
654 bi->bi_sector = (sh->sector
655 + rdev->data_offset);
656 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
657 bi->bi_rw |= REQ_FLUSH;
659 bi->bi_flags = 1 << BIO_UPTODATE;
661 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
662 bi->bi_io_vec[0].bv_offset = 0;
663 bi->bi_size = STRIPE_SIZE;
666 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
667 generic_make_request(bi);
670 if (s->syncing || s->expanding || s->expanded
672 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
674 set_bit(STRIPE_IO_STARTED, &sh->state);
676 rbi->bi_bdev = rrdev->bdev;
677 pr_debug("%s: for %llu schedule op %ld on "
678 "replacement disc %d\n",
679 __func__, (unsigned long long)sh->sector,
681 atomic_inc(&sh->count);
682 if (use_new_offset(conf, sh))
683 rbi->bi_sector = (sh->sector
684 + rrdev->new_data_offset);
686 rbi->bi_sector = (sh->sector
687 + rrdev->data_offset);
688 rbi->bi_flags = 1 << BIO_UPTODATE;
690 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
691 rbi->bi_io_vec[0].bv_offset = 0;
692 rbi->bi_size = STRIPE_SIZE;
694 generic_make_request(rbi);
696 if (!rdev && !rrdev) {
698 set_bit(STRIPE_DEGRADED, &sh->state);
699 pr_debug("skip op %ld on disc %d for sector %llu\n",
700 bi->bi_rw, i, (unsigned long long)sh->sector);
701 clear_bit(R5_LOCKED, &sh->dev[i].flags);
702 set_bit(STRIPE_HANDLE, &sh->state);
707 static struct dma_async_tx_descriptor *
708 async_copy_data(int frombio, struct bio *bio, struct page *page,
709 sector_t sector, struct dma_async_tx_descriptor *tx)
712 struct page *bio_page;
715 struct async_submit_ctl submit;
716 enum async_tx_flags flags = 0;
718 if (bio->bi_sector >= sector)
719 page_offset = (signed)(bio->bi_sector - sector) * 512;
721 page_offset = (signed)(sector - bio->bi_sector) * -512;
724 flags |= ASYNC_TX_FENCE;
725 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
727 bio_for_each_segment(bvl, bio, i) {
728 int len = bvl->bv_len;
732 if (page_offset < 0) {
733 b_offset = -page_offset;
734 page_offset += b_offset;
738 if (len > 0 && page_offset + len > STRIPE_SIZE)
739 clen = STRIPE_SIZE - page_offset;
744 b_offset += bvl->bv_offset;
745 bio_page = bvl->bv_page;
747 tx = async_memcpy(page, bio_page, page_offset,
748 b_offset, clen, &submit);
750 tx = async_memcpy(bio_page, page, b_offset,
751 page_offset, clen, &submit);
753 /* chain the operations */
754 submit.depend_tx = tx;
756 if (clen < len) /* hit end of page */
764 static void ops_complete_biofill(void *stripe_head_ref)
766 struct stripe_head *sh = stripe_head_ref;
767 struct bio *return_bi = NULL;
770 pr_debug("%s: stripe %llu\n", __func__,
771 (unsigned long long)sh->sector);
773 /* clear completed biofills */
774 for (i = sh->disks; i--; ) {
775 struct r5dev *dev = &sh->dev[i];
777 /* acknowledge completion of a biofill operation */
778 /* and check if we need to reply to a read request,
779 * new R5_Wantfill requests are held off until
780 * !STRIPE_BIOFILL_RUN
782 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
783 struct bio *rbi, *rbi2;
788 while (rbi && rbi->bi_sector <
789 dev->sector + STRIPE_SECTORS) {
790 rbi2 = r5_next_bio(rbi, dev->sector);
791 if (!raid5_dec_bi_active_stripes(rbi)) {
792 rbi->bi_next = return_bi;
799 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
801 return_io(return_bi);
803 set_bit(STRIPE_HANDLE, &sh->state);
807 static void ops_run_biofill(struct stripe_head *sh)
809 struct dma_async_tx_descriptor *tx = NULL;
810 struct async_submit_ctl submit;
813 pr_debug("%s: stripe %llu\n", __func__,
814 (unsigned long long)sh->sector);
816 for (i = sh->disks; i--; ) {
817 struct r5dev *dev = &sh->dev[i];
818 if (test_bit(R5_Wantfill, &dev->flags)) {
820 spin_lock_irq(&sh->stripe_lock);
821 dev->read = rbi = dev->toread;
823 spin_unlock_irq(&sh->stripe_lock);
824 while (rbi && rbi->bi_sector <
825 dev->sector + STRIPE_SECTORS) {
826 tx = async_copy_data(0, rbi, dev->page,
828 rbi = r5_next_bio(rbi, dev->sector);
833 atomic_inc(&sh->count);
834 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
835 async_trigger_callback(&submit);
838 static void mark_target_uptodate(struct stripe_head *sh, int target)
845 tgt = &sh->dev[target];
846 set_bit(R5_UPTODATE, &tgt->flags);
847 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
848 clear_bit(R5_Wantcompute, &tgt->flags);
851 static void ops_complete_compute(void *stripe_head_ref)
853 struct stripe_head *sh = stripe_head_ref;
855 pr_debug("%s: stripe %llu\n", __func__,
856 (unsigned long long)sh->sector);
858 /* mark the computed target(s) as uptodate */
859 mark_target_uptodate(sh, sh->ops.target);
860 mark_target_uptodate(sh, sh->ops.target2);
862 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
863 if (sh->check_state == check_state_compute_run)
864 sh->check_state = check_state_compute_result;
865 set_bit(STRIPE_HANDLE, &sh->state);
869 /* return a pointer to the address conversion region of the scribble buffer */
870 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
871 struct raid5_percpu *percpu)
873 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
876 static struct dma_async_tx_descriptor *
877 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
879 int disks = sh->disks;
880 struct page **xor_srcs = percpu->scribble;
881 int target = sh->ops.target;
882 struct r5dev *tgt = &sh->dev[target];
883 struct page *xor_dest = tgt->page;
885 struct dma_async_tx_descriptor *tx;
886 struct async_submit_ctl submit;
889 pr_debug("%s: stripe %llu block: %d\n",
890 __func__, (unsigned long long)sh->sector, target);
891 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
893 for (i = disks; i--; )
895 xor_srcs[count++] = sh->dev[i].page;
897 atomic_inc(&sh->count);
899 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
900 ops_complete_compute, sh, to_addr_conv(sh, percpu));
901 if (unlikely(count == 1))
902 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
904 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
909 /* set_syndrome_sources - populate source buffers for gen_syndrome
910 * @srcs - (struct page *) array of size sh->disks
911 * @sh - stripe_head to parse
913 * Populates srcs in proper layout order for the stripe and returns the
914 * 'count' of sources to be used in a call to async_gen_syndrome. The P
915 * destination buffer is recorded in srcs[count] and the Q destination
916 * is recorded in srcs[count+1]].
918 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
920 int disks = sh->disks;
921 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
922 int d0_idx = raid6_d0(sh);
926 for (i = 0; i < disks; i++)
932 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
934 srcs[slot] = sh->dev[i].page;
935 i = raid6_next_disk(i, disks);
936 } while (i != d0_idx);
938 return syndrome_disks;
941 static struct dma_async_tx_descriptor *
942 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
944 int disks = sh->disks;
945 struct page **blocks = percpu->scribble;
947 int qd_idx = sh->qd_idx;
948 struct dma_async_tx_descriptor *tx;
949 struct async_submit_ctl submit;
955 if (sh->ops.target < 0)
956 target = sh->ops.target2;
957 else if (sh->ops.target2 < 0)
958 target = sh->ops.target;
960 /* we should only have one valid target */
963 pr_debug("%s: stripe %llu block: %d\n",
964 __func__, (unsigned long long)sh->sector, target);
966 tgt = &sh->dev[target];
967 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
970 atomic_inc(&sh->count);
972 if (target == qd_idx) {
973 count = set_syndrome_sources(blocks, sh);
974 blocks[count] = NULL; /* regenerating p is not necessary */
975 BUG_ON(blocks[count+1] != dest); /* q should already be set */
976 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
977 ops_complete_compute, sh,
978 to_addr_conv(sh, percpu));
979 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
981 /* Compute any data- or p-drive using XOR */
983 for (i = disks; i-- ; ) {
984 if (i == target || i == qd_idx)
986 blocks[count++] = sh->dev[i].page;
989 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
990 NULL, ops_complete_compute, sh,
991 to_addr_conv(sh, percpu));
992 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
998 static struct dma_async_tx_descriptor *
999 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1001 int i, count, disks = sh->disks;
1002 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1003 int d0_idx = raid6_d0(sh);
1004 int faila = -1, failb = -1;
1005 int target = sh->ops.target;
1006 int target2 = sh->ops.target2;
1007 struct r5dev *tgt = &sh->dev[target];
1008 struct r5dev *tgt2 = &sh->dev[target2];
1009 struct dma_async_tx_descriptor *tx;
1010 struct page **blocks = percpu->scribble;
1011 struct async_submit_ctl submit;
1013 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1014 __func__, (unsigned long long)sh->sector, target, target2);
1015 BUG_ON(target < 0 || target2 < 0);
1016 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1017 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1019 /* we need to open-code set_syndrome_sources to handle the
1020 * slot number conversion for 'faila' and 'failb'
1022 for (i = 0; i < disks ; i++)
1027 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1029 blocks[slot] = sh->dev[i].page;
1035 i = raid6_next_disk(i, disks);
1036 } while (i != d0_idx);
1038 BUG_ON(faila == failb);
1041 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1042 __func__, (unsigned long long)sh->sector, faila, failb);
1044 atomic_inc(&sh->count);
1046 if (failb == syndrome_disks+1) {
1047 /* Q disk is one of the missing disks */
1048 if (faila == syndrome_disks) {
1049 /* Missing P+Q, just recompute */
1050 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1051 ops_complete_compute, sh,
1052 to_addr_conv(sh, percpu));
1053 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1054 STRIPE_SIZE, &submit);
1058 int qd_idx = sh->qd_idx;
1060 /* Missing D+Q: recompute D from P, then recompute Q */
1061 if (target == qd_idx)
1062 data_target = target2;
1064 data_target = target;
1067 for (i = disks; i-- ; ) {
1068 if (i == data_target || i == qd_idx)
1070 blocks[count++] = sh->dev[i].page;
1072 dest = sh->dev[data_target].page;
1073 init_async_submit(&submit,
1074 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1076 to_addr_conv(sh, percpu));
1077 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1080 count = set_syndrome_sources(blocks, sh);
1081 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1082 ops_complete_compute, sh,
1083 to_addr_conv(sh, percpu));
1084 return async_gen_syndrome(blocks, 0, count+2,
1085 STRIPE_SIZE, &submit);
1088 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1089 ops_complete_compute, sh,
1090 to_addr_conv(sh, percpu));
1091 if (failb == syndrome_disks) {
1092 /* We're missing D+P. */
1093 return async_raid6_datap_recov(syndrome_disks+2,
1097 /* We're missing D+D. */
1098 return async_raid6_2data_recov(syndrome_disks+2,
1099 STRIPE_SIZE, faila, failb,
1106 static void ops_complete_prexor(void *stripe_head_ref)
1108 struct stripe_head *sh = stripe_head_ref;
1110 pr_debug("%s: stripe %llu\n", __func__,
1111 (unsigned long long)sh->sector);
1114 static struct dma_async_tx_descriptor *
1115 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1116 struct dma_async_tx_descriptor *tx)
1118 int disks = sh->disks;
1119 struct page **xor_srcs = percpu->scribble;
1120 int count = 0, pd_idx = sh->pd_idx, i;
1121 struct async_submit_ctl submit;
1123 /* existing parity data subtracted */
1124 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1126 pr_debug("%s: stripe %llu\n", __func__,
1127 (unsigned long long)sh->sector);
1129 for (i = disks; i--; ) {
1130 struct r5dev *dev = &sh->dev[i];
1131 /* Only process blocks that are known to be uptodate */
1132 if (test_bit(R5_Wantdrain, &dev->flags))
1133 xor_srcs[count++] = dev->page;
1136 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1137 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1138 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1143 static struct dma_async_tx_descriptor *
1144 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1146 int disks = sh->disks;
1149 pr_debug("%s: stripe %llu\n", __func__,
1150 (unsigned long long)sh->sector);
1152 for (i = disks; i--; ) {
1153 struct r5dev *dev = &sh->dev[i];
1156 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1159 spin_lock_irq(&sh->stripe_lock);
1160 chosen = dev->towrite;
1161 dev->towrite = NULL;
1162 BUG_ON(dev->written);
1163 wbi = dev->written = chosen;
1164 spin_unlock_irq(&sh->stripe_lock);
1166 while (wbi && wbi->bi_sector <
1167 dev->sector + STRIPE_SECTORS) {
1168 if (wbi->bi_rw & REQ_FUA)
1169 set_bit(R5_WantFUA, &dev->flags);
1170 if (wbi->bi_rw & REQ_SYNC)
1171 set_bit(R5_SyncIO, &dev->flags);
1172 tx = async_copy_data(1, wbi, dev->page,
1174 wbi = r5_next_bio(wbi, dev->sector);
1182 static void ops_complete_reconstruct(void *stripe_head_ref)
1184 struct stripe_head *sh = stripe_head_ref;
1185 int disks = sh->disks;
1186 int pd_idx = sh->pd_idx;
1187 int qd_idx = sh->qd_idx;
1189 bool fua = false, sync = false;
1191 pr_debug("%s: stripe %llu\n", __func__,
1192 (unsigned long long)sh->sector);
1194 for (i = disks; i--; ) {
1195 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1196 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1199 for (i = disks; i--; ) {
1200 struct r5dev *dev = &sh->dev[i];
1202 if (dev->written || i == pd_idx || i == qd_idx) {
1203 set_bit(R5_UPTODATE, &dev->flags);
1205 set_bit(R5_WantFUA, &dev->flags);
1207 set_bit(R5_SyncIO, &dev->flags);
1211 if (sh->reconstruct_state == reconstruct_state_drain_run)
1212 sh->reconstruct_state = reconstruct_state_drain_result;
1213 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1214 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1216 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1217 sh->reconstruct_state = reconstruct_state_result;
1220 set_bit(STRIPE_HANDLE, &sh->state);
1225 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1226 struct dma_async_tx_descriptor *tx)
1228 int disks = sh->disks;
1229 struct page **xor_srcs = percpu->scribble;
1230 struct async_submit_ctl submit;
1231 int count = 0, pd_idx = sh->pd_idx, i;
1232 struct page *xor_dest;
1234 unsigned long flags;
1236 pr_debug("%s: stripe %llu\n", __func__,
1237 (unsigned long long)sh->sector);
1239 /* check if prexor is active which means only process blocks
1240 * that are part of a read-modify-write (written)
1242 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1244 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1245 for (i = disks; i--; ) {
1246 struct r5dev *dev = &sh->dev[i];
1248 xor_srcs[count++] = dev->page;
1251 xor_dest = sh->dev[pd_idx].page;
1252 for (i = disks; i--; ) {
1253 struct r5dev *dev = &sh->dev[i];
1255 xor_srcs[count++] = dev->page;
1259 /* 1/ if we prexor'd then the dest is reused as a source
1260 * 2/ if we did not prexor then we are redoing the parity
1261 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1262 * for the synchronous xor case
1264 flags = ASYNC_TX_ACK |
1265 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1267 atomic_inc(&sh->count);
1269 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1270 to_addr_conv(sh, percpu));
1271 if (unlikely(count == 1))
1272 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1274 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1278 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1279 struct dma_async_tx_descriptor *tx)
1281 struct async_submit_ctl submit;
1282 struct page **blocks = percpu->scribble;
1285 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1287 count = set_syndrome_sources(blocks, sh);
1289 atomic_inc(&sh->count);
1291 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1292 sh, to_addr_conv(sh, percpu));
1293 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1296 static void ops_complete_check(void *stripe_head_ref)
1298 struct stripe_head *sh = stripe_head_ref;
1300 pr_debug("%s: stripe %llu\n", __func__,
1301 (unsigned long long)sh->sector);
1303 sh->check_state = check_state_check_result;
1304 set_bit(STRIPE_HANDLE, &sh->state);
1308 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1310 int disks = sh->disks;
1311 int pd_idx = sh->pd_idx;
1312 int qd_idx = sh->qd_idx;
1313 struct page *xor_dest;
1314 struct page **xor_srcs = percpu->scribble;
1315 struct dma_async_tx_descriptor *tx;
1316 struct async_submit_ctl submit;
1320 pr_debug("%s: stripe %llu\n", __func__,
1321 (unsigned long long)sh->sector);
1324 xor_dest = sh->dev[pd_idx].page;
1325 xor_srcs[count++] = xor_dest;
1326 for (i = disks; i--; ) {
1327 if (i == pd_idx || i == qd_idx)
1329 xor_srcs[count++] = sh->dev[i].page;
1332 init_async_submit(&submit, 0, NULL, NULL, NULL,
1333 to_addr_conv(sh, percpu));
1334 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1335 &sh->ops.zero_sum_result, &submit);
1337 atomic_inc(&sh->count);
1338 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1339 tx = async_trigger_callback(&submit);
1342 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1344 struct page **srcs = percpu->scribble;
1345 struct async_submit_ctl submit;
1348 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1349 (unsigned long long)sh->sector, checkp);
1351 count = set_syndrome_sources(srcs, sh);
1355 atomic_inc(&sh->count);
1356 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1357 sh, to_addr_conv(sh, percpu));
1358 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1359 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1362 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1364 int overlap_clear = 0, i, disks = sh->disks;
1365 struct dma_async_tx_descriptor *tx = NULL;
1366 struct r5conf *conf = sh->raid_conf;
1367 int level = conf->level;
1368 struct raid5_percpu *percpu;
1372 percpu = per_cpu_ptr(conf->percpu, cpu);
1373 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1374 ops_run_biofill(sh);
1378 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1380 tx = ops_run_compute5(sh, percpu);
1382 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1383 tx = ops_run_compute6_1(sh, percpu);
1385 tx = ops_run_compute6_2(sh, percpu);
1387 /* terminate the chain if reconstruct is not set to be run */
1388 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1392 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1393 tx = ops_run_prexor(sh, percpu, tx);
1395 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1396 tx = ops_run_biodrain(sh, tx);
1400 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1402 ops_run_reconstruct5(sh, percpu, tx);
1404 ops_run_reconstruct6(sh, percpu, tx);
1407 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1408 if (sh->check_state == check_state_run)
1409 ops_run_check_p(sh, percpu);
1410 else if (sh->check_state == check_state_run_q)
1411 ops_run_check_pq(sh, percpu, 0);
1412 else if (sh->check_state == check_state_run_pq)
1413 ops_run_check_pq(sh, percpu, 1);
1419 for (i = disks; i--; ) {
1420 struct r5dev *dev = &sh->dev[i];
1421 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1422 wake_up(&sh->raid_conf->wait_for_overlap);
1427 #ifdef CONFIG_MULTICORE_RAID456
1428 static void async_run_ops(void *param, async_cookie_t cookie)
1430 struct stripe_head *sh = param;
1431 unsigned long ops_request = sh->ops.request;
1433 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1434 wake_up(&sh->ops.wait_for_ops);
1436 __raid_run_ops(sh, ops_request);
1440 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1442 /* since handle_stripe can be called outside of raid5d context
1443 * we need to ensure sh->ops.request is de-staged before another
1446 wait_event(sh->ops.wait_for_ops,
1447 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1448 sh->ops.request = ops_request;
1450 atomic_inc(&sh->count);
1451 async_schedule(async_run_ops, sh);
1454 #define raid_run_ops __raid_run_ops
1457 static int grow_one_stripe(struct r5conf *conf)
1459 struct stripe_head *sh;
1460 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1464 sh->raid_conf = conf;
1465 #ifdef CONFIG_MULTICORE_RAID456
1466 init_waitqueue_head(&sh->ops.wait_for_ops);
1469 spin_lock_init(&sh->stripe_lock);
1471 if (grow_buffers(sh)) {
1473 kmem_cache_free(conf->slab_cache, sh);
1476 /* we just created an active stripe so... */
1477 atomic_set(&sh->count, 1);
1478 atomic_inc(&conf->active_stripes);
1479 INIT_LIST_HEAD(&sh->lru);
1484 static int grow_stripes(struct r5conf *conf, int num)
1486 struct kmem_cache *sc;
1487 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1489 if (conf->mddev->gendisk)
1490 sprintf(conf->cache_name[0],
1491 "raid%d-%s", conf->level, mdname(conf->mddev));
1493 sprintf(conf->cache_name[0],
1494 "raid%d-%p", conf->level, conf->mddev);
1495 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1497 conf->active_name = 0;
1498 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1499 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1503 conf->slab_cache = sc;
1504 conf->pool_size = devs;
1506 if (!grow_one_stripe(conf))
1512 * scribble_len - return the required size of the scribble region
1513 * @num - total number of disks in the array
1515 * The size must be enough to contain:
1516 * 1/ a struct page pointer for each device in the array +2
1517 * 2/ room to convert each entry in (1) to its corresponding dma
1518 * (dma_map_page()) or page (page_address()) address.
1520 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1521 * calculate over all devices (not just the data blocks), using zeros in place
1522 * of the P and Q blocks.
1524 static size_t scribble_len(int num)
1528 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1533 static int resize_stripes(struct r5conf *conf, int newsize)
1535 /* Make all the stripes able to hold 'newsize' devices.
1536 * New slots in each stripe get 'page' set to a new page.
1538 * This happens in stages:
1539 * 1/ create a new kmem_cache and allocate the required number of
1541 * 2/ gather all the old stripe_heads and tranfer the pages across
1542 * to the new stripe_heads. This will have the side effect of
1543 * freezing the array as once all stripe_heads have been collected,
1544 * no IO will be possible. Old stripe heads are freed once their
1545 * pages have been transferred over, and the old kmem_cache is
1546 * freed when all stripes are done.
1547 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1548 * we simple return a failre status - no need to clean anything up.
1549 * 4/ allocate new pages for the new slots in the new stripe_heads.
1550 * If this fails, we don't bother trying the shrink the
1551 * stripe_heads down again, we just leave them as they are.
1552 * As each stripe_head is processed the new one is released into
1555 * Once step2 is started, we cannot afford to wait for a write,
1556 * so we use GFP_NOIO allocations.
1558 struct stripe_head *osh, *nsh;
1559 LIST_HEAD(newstripes);
1560 struct disk_info *ndisks;
1563 struct kmem_cache *sc;
1566 if (newsize <= conf->pool_size)
1567 return 0; /* never bother to shrink */
1569 err = md_allow_write(conf->mddev);
1574 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1575 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1580 for (i = conf->max_nr_stripes; i; i--) {
1581 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1585 nsh->raid_conf = conf;
1586 #ifdef CONFIG_MULTICORE_RAID456
1587 init_waitqueue_head(&nsh->ops.wait_for_ops);
1590 list_add(&nsh->lru, &newstripes);
1593 /* didn't get enough, give up */
1594 while (!list_empty(&newstripes)) {
1595 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1596 list_del(&nsh->lru);
1597 kmem_cache_free(sc, nsh);
1599 kmem_cache_destroy(sc);
1602 /* Step 2 - Must use GFP_NOIO now.
1603 * OK, we have enough stripes, start collecting inactive
1604 * stripes and copying them over
1606 list_for_each_entry(nsh, &newstripes, lru) {
1607 spin_lock_irq(&conf->device_lock);
1608 wait_event_lock_irq(conf->wait_for_stripe,
1609 !list_empty(&conf->inactive_list),
1612 osh = get_free_stripe(conf);
1613 spin_unlock_irq(&conf->device_lock);
1614 atomic_set(&nsh->count, 1);
1615 for(i=0; i<conf->pool_size; i++)
1616 nsh->dev[i].page = osh->dev[i].page;
1617 for( ; i<newsize; i++)
1618 nsh->dev[i].page = NULL;
1619 kmem_cache_free(conf->slab_cache, osh);
1621 kmem_cache_destroy(conf->slab_cache);
1624 * At this point, we are holding all the stripes so the array
1625 * is completely stalled, so now is a good time to resize
1626 * conf->disks and the scribble region
1628 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1630 for (i=0; i<conf->raid_disks; i++)
1631 ndisks[i] = conf->disks[i];
1633 conf->disks = ndisks;
1638 conf->scribble_len = scribble_len(newsize);
1639 for_each_present_cpu(cpu) {
1640 struct raid5_percpu *percpu;
1643 percpu = per_cpu_ptr(conf->percpu, cpu);
1644 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1647 kfree(percpu->scribble);
1648 percpu->scribble = scribble;
1656 /* Step 4, return new stripes to service */
1657 while(!list_empty(&newstripes)) {
1658 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1659 list_del_init(&nsh->lru);
1661 for (i=conf->raid_disks; i < newsize; i++)
1662 if (nsh->dev[i].page == NULL) {
1663 struct page *p = alloc_page(GFP_NOIO);
1664 nsh->dev[i].page = p;
1668 release_stripe(nsh);
1670 /* critical section pass, GFP_NOIO no longer needed */
1672 conf->slab_cache = sc;
1673 conf->active_name = 1-conf->active_name;
1674 conf->pool_size = newsize;
1678 static int drop_one_stripe(struct r5conf *conf)
1680 struct stripe_head *sh;
1682 spin_lock_irq(&conf->device_lock);
1683 sh = get_free_stripe(conf);
1684 spin_unlock_irq(&conf->device_lock);
1687 BUG_ON(atomic_read(&sh->count));
1689 kmem_cache_free(conf->slab_cache, sh);
1690 atomic_dec(&conf->active_stripes);
1694 static void shrink_stripes(struct r5conf *conf)
1696 while (drop_one_stripe(conf))
1699 if (conf->slab_cache)
1700 kmem_cache_destroy(conf->slab_cache);
1701 conf->slab_cache = NULL;
1704 static void raid5_end_read_request(struct bio * bi, int error)
1706 struct stripe_head *sh = bi->bi_private;
1707 struct r5conf *conf = sh->raid_conf;
1708 int disks = sh->disks, i;
1709 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1710 char b[BDEVNAME_SIZE];
1711 struct md_rdev *rdev = NULL;
1714 for (i=0 ; i<disks; i++)
1715 if (bi == &sh->dev[i].req)
1718 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1719 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1725 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1726 /* If replacement finished while this request was outstanding,
1727 * 'replacement' might be NULL already.
1728 * In that case it moved down to 'rdev'.
1729 * rdev is not removed until all requests are finished.
1731 rdev = conf->disks[i].replacement;
1733 rdev = conf->disks[i].rdev;
1735 if (use_new_offset(conf, sh))
1736 s = sh->sector + rdev->new_data_offset;
1738 s = sh->sector + rdev->data_offset;
1740 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1741 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1742 /* Note that this cannot happen on a
1743 * replacement device. We just fail those on
1748 "md/raid:%s: read error corrected"
1749 " (%lu sectors at %llu on %s)\n",
1750 mdname(conf->mddev), STRIPE_SECTORS,
1751 (unsigned long long)s,
1752 bdevname(rdev->bdev, b));
1753 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1754 clear_bit(R5_ReadError, &sh->dev[i].flags);
1755 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1756 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1757 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1759 if (atomic_read(&rdev->read_errors))
1760 atomic_set(&rdev->read_errors, 0);
1762 const char *bdn = bdevname(rdev->bdev, b);
1766 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1767 atomic_inc(&rdev->read_errors);
1768 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1771 "md/raid:%s: read error on replacement device "
1772 "(sector %llu on %s).\n",
1773 mdname(conf->mddev),
1774 (unsigned long long)s,
1776 else if (conf->mddev->degraded >= conf->max_degraded) {
1780 "md/raid:%s: read error not correctable "
1781 "(sector %llu on %s).\n",
1782 mdname(conf->mddev),
1783 (unsigned long long)s,
1785 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1790 "md/raid:%s: read error NOT corrected!! "
1791 "(sector %llu on %s).\n",
1792 mdname(conf->mddev),
1793 (unsigned long long)s,
1795 } else if (atomic_read(&rdev->read_errors)
1796 > conf->max_nr_stripes)
1798 "md/raid:%s: Too many read errors, failing device %s.\n",
1799 mdname(conf->mddev), bdn);
1803 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1804 set_bit(R5_ReadError, &sh->dev[i].flags);
1805 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1807 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1809 clear_bit(R5_ReadError, &sh->dev[i].flags);
1810 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1812 && test_bit(In_sync, &rdev->flags)
1813 && rdev_set_badblocks(
1814 rdev, sh->sector, STRIPE_SECTORS, 0)))
1815 md_error(conf->mddev, rdev);
1818 rdev_dec_pending(rdev, conf->mddev);
1819 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1820 set_bit(STRIPE_HANDLE, &sh->state);
1824 static void raid5_end_write_request(struct bio *bi, int error)
1826 struct stripe_head *sh = bi->bi_private;
1827 struct r5conf *conf = sh->raid_conf;
1828 int disks = sh->disks, i;
1829 struct md_rdev *uninitialized_var(rdev);
1830 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1833 int replacement = 0;
1835 for (i = 0 ; i < disks; i++) {
1836 if (bi == &sh->dev[i].req) {
1837 rdev = conf->disks[i].rdev;
1840 if (bi == &sh->dev[i].rreq) {
1841 rdev = conf->disks[i].replacement;
1845 /* rdev was removed and 'replacement'
1846 * replaced it. rdev is not removed
1847 * until all requests are finished.
1849 rdev = conf->disks[i].rdev;
1853 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1854 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1863 md_error(conf->mddev, rdev);
1864 else if (is_badblock(rdev, sh->sector,
1866 &first_bad, &bad_sectors))
1867 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1870 set_bit(WriteErrorSeen, &rdev->flags);
1871 set_bit(R5_WriteError, &sh->dev[i].flags);
1872 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1873 set_bit(MD_RECOVERY_NEEDED,
1874 &rdev->mddev->recovery);
1875 } else if (is_badblock(rdev, sh->sector,
1877 &first_bad, &bad_sectors))
1878 set_bit(R5_MadeGood, &sh->dev[i].flags);
1880 rdev_dec_pending(rdev, conf->mddev);
1882 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1883 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1884 set_bit(STRIPE_HANDLE, &sh->state);
1888 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1890 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1892 struct r5dev *dev = &sh->dev[i];
1894 bio_init(&dev->req);
1895 dev->req.bi_io_vec = &dev->vec;
1897 dev->req.bi_max_vecs++;
1898 dev->req.bi_private = sh;
1899 dev->vec.bv_page = dev->page;
1901 bio_init(&dev->rreq);
1902 dev->rreq.bi_io_vec = &dev->rvec;
1903 dev->rreq.bi_vcnt++;
1904 dev->rreq.bi_max_vecs++;
1905 dev->rreq.bi_private = sh;
1906 dev->rvec.bv_page = dev->page;
1909 dev->sector = compute_blocknr(sh, i, previous);
1912 static void error(struct mddev *mddev, struct md_rdev *rdev)
1914 char b[BDEVNAME_SIZE];
1915 struct r5conf *conf = mddev->private;
1916 unsigned long flags;
1917 pr_debug("raid456: error called\n");
1919 spin_lock_irqsave(&conf->device_lock, flags);
1920 clear_bit(In_sync, &rdev->flags);
1921 mddev->degraded = calc_degraded(conf);
1922 spin_unlock_irqrestore(&conf->device_lock, flags);
1923 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1925 set_bit(Blocked, &rdev->flags);
1926 set_bit(Faulty, &rdev->flags);
1927 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1929 "md/raid:%s: Disk failure on %s, disabling device.\n"
1930 "md/raid:%s: Operation continuing on %d devices.\n",
1932 bdevname(rdev->bdev, b),
1934 conf->raid_disks - mddev->degraded);
1938 * Input: a 'big' sector number,
1939 * Output: index of the data and parity disk, and the sector # in them.
1941 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1942 int previous, int *dd_idx,
1943 struct stripe_head *sh)
1945 sector_t stripe, stripe2;
1946 sector_t chunk_number;
1947 unsigned int chunk_offset;
1950 sector_t new_sector;
1951 int algorithm = previous ? conf->prev_algo
1953 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1954 : conf->chunk_sectors;
1955 int raid_disks = previous ? conf->previous_raid_disks
1957 int data_disks = raid_disks - conf->max_degraded;
1959 /* First compute the information on this sector */
1962 * Compute the chunk number and the sector offset inside the chunk
1964 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1965 chunk_number = r_sector;
1968 * Compute the stripe number
1970 stripe = chunk_number;
1971 *dd_idx = sector_div(stripe, data_disks);
1974 * Select the parity disk based on the user selected algorithm.
1976 pd_idx = qd_idx = -1;
1977 switch(conf->level) {
1979 pd_idx = data_disks;
1982 switch (algorithm) {
1983 case ALGORITHM_LEFT_ASYMMETRIC:
1984 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1985 if (*dd_idx >= pd_idx)
1988 case ALGORITHM_RIGHT_ASYMMETRIC:
1989 pd_idx = sector_div(stripe2, raid_disks);
1990 if (*dd_idx >= pd_idx)
1993 case ALGORITHM_LEFT_SYMMETRIC:
1994 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1995 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1997 case ALGORITHM_RIGHT_SYMMETRIC:
1998 pd_idx = sector_div(stripe2, raid_disks);
1999 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2001 case ALGORITHM_PARITY_0:
2005 case ALGORITHM_PARITY_N:
2006 pd_idx = data_disks;
2014 switch (algorithm) {
2015 case ALGORITHM_LEFT_ASYMMETRIC:
2016 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2017 qd_idx = pd_idx + 1;
2018 if (pd_idx == raid_disks-1) {
2019 (*dd_idx)++; /* Q D D D P */
2021 } else if (*dd_idx >= pd_idx)
2022 (*dd_idx) += 2; /* D D P Q D */
2024 case ALGORITHM_RIGHT_ASYMMETRIC:
2025 pd_idx = sector_div(stripe2, raid_disks);
2026 qd_idx = pd_idx + 1;
2027 if (pd_idx == raid_disks-1) {
2028 (*dd_idx)++; /* Q D D D P */
2030 } else if (*dd_idx >= pd_idx)
2031 (*dd_idx) += 2; /* D D P Q D */
2033 case ALGORITHM_LEFT_SYMMETRIC:
2034 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2035 qd_idx = (pd_idx + 1) % raid_disks;
2036 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2038 case ALGORITHM_RIGHT_SYMMETRIC:
2039 pd_idx = sector_div(stripe2, raid_disks);
2040 qd_idx = (pd_idx + 1) % raid_disks;
2041 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2044 case ALGORITHM_PARITY_0:
2049 case ALGORITHM_PARITY_N:
2050 pd_idx = data_disks;
2051 qd_idx = data_disks + 1;
2054 case ALGORITHM_ROTATING_ZERO_RESTART:
2055 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2056 * of blocks for computing Q is different.
2058 pd_idx = sector_div(stripe2, raid_disks);
2059 qd_idx = pd_idx + 1;
2060 if (pd_idx == raid_disks-1) {
2061 (*dd_idx)++; /* Q D D D P */
2063 } else if (*dd_idx >= pd_idx)
2064 (*dd_idx) += 2; /* D D P Q D */
2068 case ALGORITHM_ROTATING_N_RESTART:
2069 /* Same a left_asymmetric, by first stripe is
2070 * D D D P Q rather than
2074 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2075 qd_idx = pd_idx + 1;
2076 if (pd_idx == raid_disks-1) {
2077 (*dd_idx)++; /* Q D D D P */
2079 } else if (*dd_idx >= pd_idx)
2080 (*dd_idx) += 2; /* D D P Q D */
2084 case ALGORITHM_ROTATING_N_CONTINUE:
2085 /* Same as left_symmetric but Q is before P */
2086 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2087 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2088 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2092 case ALGORITHM_LEFT_ASYMMETRIC_6:
2093 /* RAID5 left_asymmetric, with Q on last device */
2094 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2095 if (*dd_idx >= pd_idx)
2097 qd_idx = raid_disks - 1;
2100 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2101 pd_idx = sector_div(stripe2, raid_disks-1);
2102 if (*dd_idx >= pd_idx)
2104 qd_idx = raid_disks - 1;
2107 case ALGORITHM_LEFT_SYMMETRIC_6:
2108 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2109 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2110 qd_idx = raid_disks - 1;
2113 case ALGORITHM_RIGHT_SYMMETRIC_6:
2114 pd_idx = sector_div(stripe2, raid_disks-1);
2115 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2116 qd_idx = raid_disks - 1;
2119 case ALGORITHM_PARITY_0_6:
2122 qd_idx = raid_disks - 1;
2132 sh->pd_idx = pd_idx;
2133 sh->qd_idx = qd_idx;
2134 sh->ddf_layout = ddf_layout;
2137 * Finally, compute the new sector number
2139 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2144 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2146 struct r5conf *conf = sh->raid_conf;
2147 int raid_disks = sh->disks;
2148 int data_disks = raid_disks - conf->max_degraded;
2149 sector_t new_sector = sh->sector, check;
2150 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2151 : conf->chunk_sectors;
2152 int algorithm = previous ? conf->prev_algo
2156 sector_t chunk_number;
2157 int dummy1, dd_idx = i;
2159 struct stripe_head sh2;
2162 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2163 stripe = new_sector;
2165 if (i == sh->pd_idx)
2167 switch(conf->level) {
2170 switch (algorithm) {
2171 case ALGORITHM_LEFT_ASYMMETRIC:
2172 case ALGORITHM_RIGHT_ASYMMETRIC:
2176 case ALGORITHM_LEFT_SYMMETRIC:
2177 case ALGORITHM_RIGHT_SYMMETRIC:
2180 i -= (sh->pd_idx + 1);
2182 case ALGORITHM_PARITY_0:
2185 case ALGORITHM_PARITY_N:
2192 if (i == sh->qd_idx)
2193 return 0; /* It is the Q disk */
2194 switch (algorithm) {
2195 case ALGORITHM_LEFT_ASYMMETRIC:
2196 case ALGORITHM_RIGHT_ASYMMETRIC:
2197 case ALGORITHM_ROTATING_ZERO_RESTART:
2198 case ALGORITHM_ROTATING_N_RESTART:
2199 if (sh->pd_idx == raid_disks-1)
2200 i--; /* Q D D D P */
2201 else if (i > sh->pd_idx)
2202 i -= 2; /* D D P Q D */
2204 case ALGORITHM_LEFT_SYMMETRIC:
2205 case ALGORITHM_RIGHT_SYMMETRIC:
2206 if (sh->pd_idx == raid_disks-1)
2207 i--; /* Q D D D P */
2212 i -= (sh->pd_idx + 2);
2215 case ALGORITHM_PARITY_0:
2218 case ALGORITHM_PARITY_N:
2220 case ALGORITHM_ROTATING_N_CONTINUE:
2221 /* Like left_symmetric, but P is before Q */
2222 if (sh->pd_idx == 0)
2223 i--; /* P D D D Q */
2228 i -= (sh->pd_idx + 1);
2231 case ALGORITHM_LEFT_ASYMMETRIC_6:
2232 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2236 case ALGORITHM_LEFT_SYMMETRIC_6:
2237 case ALGORITHM_RIGHT_SYMMETRIC_6:
2239 i += data_disks + 1;
2240 i -= (sh->pd_idx + 1);
2242 case ALGORITHM_PARITY_0_6:
2251 chunk_number = stripe * data_disks + i;
2252 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2254 check = raid5_compute_sector(conf, r_sector,
2255 previous, &dummy1, &sh2);
2256 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2257 || sh2.qd_idx != sh->qd_idx) {
2258 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2259 mdname(conf->mddev));
2267 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2268 int rcw, int expand)
2270 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2271 struct r5conf *conf = sh->raid_conf;
2272 int level = conf->level;
2275 /* if we are not expanding this is a proper write request, and
2276 * there will be bios with new data to be drained into the
2280 sh->reconstruct_state = reconstruct_state_drain_run;
2281 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2283 sh->reconstruct_state = reconstruct_state_run;
2285 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2287 for (i = disks; i--; ) {
2288 struct r5dev *dev = &sh->dev[i];
2291 set_bit(R5_LOCKED, &dev->flags);
2292 set_bit(R5_Wantdrain, &dev->flags);
2294 clear_bit(R5_UPTODATE, &dev->flags);
2298 if (s->locked + conf->max_degraded == disks)
2299 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2300 atomic_inc(&conf->pending_full_writes);
2303 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2304 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2306 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2307 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2308 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2309 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2311 for (i = disks; i--; ) {
2312 struct r5dev *dev = &sh->dev[i];
2317 (test_bit(R5_UPTODATE, &dev->flags) ||
2318 test_bit(R5_Wantcompute, &dev->flags))) {
2319 set_bit(R5_Wantdrain, &dev->flags);
2320 set_bit(R5_LOCKED, &dev->flags);
2321 clear_bit(R5_UPTODATE, &dev->flags);
2327 /* keep the parity disk(s) locked while asynchronous operations
2330 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2331 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2335 int qd_idx = sh->qd_idx;
2336 struct r5dev *dev = &sh->dev[qd_idx];
2338 set_bit(R5_LOCKED, &dev->flags);
2339 clear_bit(R5_UPTODATE, &dev->flags);
2343 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2344 __func__, (unsigned long long)sh->sector,
2345 s->locked, s->ops_request);
2349 * Each stripe/dev can have one or more bion attached.
2350 * toread/towrite point to the first in a chain.
2351 * The bi_next chain must be in order.
2353 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2356 struct r5conf *conf = sh->raid_conf;
2359 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2360 (unsigned long long)bi->bi_sector,
2361 (unsigned long long)sh->sector);
2364 * If several bio share a stripe. The bio bi_phys_segments acts as a
2365 * reference count to avoid race. The reference count should already be
2366 * increased before this function is called (for example, in
2367 * make_request()), so other bio sharing this stripe will not free the
2368 * stripe. If a stripe is owned by one stripe, the stripe lock will
2371 spin_lock_irq(&sh->stripe_lock);
2373 bip = &sh->dev[dd_idx].towrite;
2377 bip = &sh->dev[dd_idx].toread;
2378 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2379 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2381 bip = & (*bip)->bi_next;
2383 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2386 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2390 raid5_inc_bi_active_stripes(bi);
2393 /* check if page is covered */
2394 sector_t sector = sh->dev[dd_idx].sector;
2395 for (bi=sh->dev[dd_idx].towrite;
2396 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2397 bi && bi->bi_sector <= sector;
2398 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2399 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2400 sector = bi->bi_sector + (bi->bi_size>>9);
2402 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2403 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2405 spin_unlock_irq(&sh->stripe_lock);
2407 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2408 (unsigned long long)(*bip)->bi_sector,
2409 (unsigned long long)sh->sector, dd_idx);
2411 if (conf->mddev->bitmap && firstwrite) {
2412 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2414 sh->bm_seq = conf->seq_flush+1;
2415 set_bit(STRIPE_BIT_DELAY, &sh->state);
2420 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2421 spin_unlock_irq(&sh->stripe_lock);
2425 static void end_reshape(struct r5conf *conf);
2427 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2428 struct stripe_head *sh)
2430 int sectors_per_chunk =
2431 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2433 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2434 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2436 raid5_compute_sector(conf,
2437 stripe * (disks - conf->max_degraded)
2438 *sectors_per_chunk + chunk_offset,
2444 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2445 struct stripe_head_state *s, int disks,
2446 struct bio **return_bi)
2449 for (i = disks; i--; ) {
2453 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2454 struct md_rdev *rdev;
2456 rdev = rcu_dereference(conf->disks[i].rdev);
2457 if (rdev && test_bit(In_sync, &rdev->flags))
2458 atomic_inc(&rdev->nr_pending);
2463 if (!rdev_set_badblocks(
2467 md_error(conf->mddev, rdev);
2468 rdev_dec_pending(rdev, conf->mddev);
2471 spin_lock_irq(&sh->stripe_lock);
2472 /* fail all writes first */
2473 bi = sh->dev[i].towrite;
2474 sh->dev[i].towrite = NULL;
2475 spin_unlock_irq(&sh->stripe_lock);
2481 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2482 wake_up(&conf->wait_for_overlap);
2484 while (bi && bi->bi_sector <
2485 sh->dev[i].sector + STRIPE_SECTORS) {
2486 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2487 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2488 if (!raid5_dec_bi_active_stripes(bi)) {
2489 md_write_end(conf->mddev);
2490 bi->bi_next = *return_bi;
2496 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2497 STRIPE_SECTORS, 0, 0);
2499 /* and fail all 'written' */
2500 bi = sh->dev[i].written;
2501 sh->dev[i].written = NULL;
2502 if (bi) bitmap_end = 1;
2503 while (bi && bi->bi_sector <
2504 sh->dev[i].sector + STRIPE_SECTORS) {
2505 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2506 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2507 if (!raid5_dec_bi_active_stripes(bi)) {
2508 md_write_end(conf->mddev);
2509 bi->bi_next = *return_bi;
2515 /* fail any reads if this device is non-operational and
2516 * the data has not reached the cache yet.
2518 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2519 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2520 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2521 bi = sh->dev[i].toread;
2522 sh->dev[i].toread = NULL;
2523 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2524 wake_up(&conf->wait_for_overlap);
2525 if (bi) s->to_read--;
2526 while (bi && bi->bi_sector <
2527 sh->dev[i].sector + STRIPE_SECTORS) {
2528 struct bio *nextbi =
2529 r5_next_bio(bi, sh->dev[i].sector);
2530 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2531 if (!raid5_dec_bi_active_stripes(bi)) {
2532 bi->bi_next = *return_bi;
2539 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2540 STRIPE_SECTORS, 0, 0);
2541 /* If we were in the middle of a write the parity block might
2542 * still be locked - so just clear all R5_LOCKED flags
2544 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2547 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2548 if (atomic_dec_and_test(&conf->pending_full_writes))
2549 md_wakeup_thread(conf->mddev->thread);
2553 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2554 struct stripe_head_state *s)
2559 clear_bit(STRIPE_SYNCING, &sh->state);
2562 /* There is nothing more to do for sync/check/repair.
2563 * Don't even need to abort as that is handled elsewhere
2564 * if needed, and not always wanted e.g. if there is a known
2566 * For recover/replace we need to record a bad block on all
2567 * non-sync devices, or abort the recovery
2569 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2570 /* During recovery devices cannot be removed, so
2571 * locking and refcounting of rdevs is not needed
2573 for (i = 0; i < conf->raid_disks; i++) {
2574 struct md_rdev *rdev = conf->disks[i].rdev;
2576 && !test_bit(Faulty, &rdev->flags)
2577 && !test_bit(In_sync, &rdev->flags)
2578 && !rdev_set_badblocks(rdev, sh->sector,
2581 rdev = conf->disks[i].replacement;
2583 && !test_bit(Faulty, &rdev->flags)
2584 && !test_bit(In_sync, &rdev->flags)
2585 && !rdev_set_badblocks(rdev, sh->sector,
2590 conf->recovery_disabled =
2591 conf->mddev->recovery_disabled;
2593 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2596 static int want_replace(struct stripe_head *sh, int disk_idx)
2598 struct md_rdev *rdev;
2600 /* Doing recovery so rcu locking not required */
2601 rdev = sh->raid_conf->disks[disk_idx].replacement;
2603 && !test_bit(Faulty, &rdev->flags)
2604 && !test_bit(In_sync, &rdev->flags)
2605 && (rdev->recovery_offset <= sh->sector
2606 || rdev->mddev->recovery_cp <= sh->sector))
2612 /* fetch_block - checks the given member device to see if its data needs
2613 * to be read or computed to satisfy a request.
2615 * Returns 1 when no more member devices need to be checked, otherwise returns
2616 * 0 to tell the loop in handle_stripe_fill to continue
2618 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2619 int disk_idx, int disks)
2621 struct r5dev *dev = &sh->dev[disk_idx];
2622 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2623 &sh->dev[s->failed_num[1]] };
2625 /* is the data in this block needed, and can we get it? */
2626 if (!test_bit(R5_LOCKED, &dev->flags) &&
2627 !test_bit(R5_UPTODATE, &dev->flags) &&
2629 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2630 s->syncing || s->expanding ||
2631 (s->replacing && want_replace(sh, disk_idx)) ||
2632 (s->failed >= 1 && fdev[0]->toread) ||
2633 (s->failed >= 2 && fdev[1]->toread) ||
2634 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2635 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2636 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2637 /* we would like to get this block, possibly by computing it,
2638 * otherwise read it if the backing disk is insync
2640 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2641 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2642 if ((s->uptodate == disks - 1) &&
2643 (s->failed && (disk_idx == s->failed_num[0] ||
2644 disk_idx == s->failed_num[1]))) {
2645 /* have disk failed, and we're requested to fetch it;
2648 pr_debug("Computing stripe %llu block %d\n",
2649 (unsigned long long)sh->sector, disk_idx);
2650 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2651 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2652 set_bit(R5_Wantcompute, &dev->flags);
2653 sh->ops.target = disk_idx;
2654 sh->ops.target2 = -1; /* no 2nd target */
2656 /* Careful: from this point on 'uptodate' is in the eye
2657 * of raid_run_ops which services 'compute' operations
2658 * before writes. R5_Wantcompute flags a block that will
2659 * be R5_UPTODATE by the time it is needed for a
2660 * subsequent operation.
2664 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2665 /* Computing 2-failure is *very* expensive; only
2666 * do it if failed >= 2
2669 for (other = disks; other--; ) {
2670 if (other == disk_idx)
2672 if (!test_bit(R5_UPTODATE,
2673 &sh->dev[other].flags))
2677 pr_debug("Computing stripe %llu blocks %d,%d\n",
2678 (unsigned long long)sh->sector,
2680 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2681 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2682 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2683 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2684 sh->ops.target = disk_idx;
2685 sh->ops.target2 = other;
2689 } else if (test_bit(R5_Insync, &dev->flags)) {
2690 set_bit(R5_LOCKED, &dev->flags);
2691 set_bit(R5_Wantread, &dev->flags);
2693 pr_debug("Reading block %d (sync=%d)\n",
2694 disk_idx, s->syncing);
2702 * handle_stripe_fill - read or compute data to satisfy pending requests.
2704 static void handle_stripe_fill(struct stripe_head *sh,
2705 struct stripe_head_state *s,
2710 /* look for blocks to read/compute, skip this if a compute
2711 * is already in flight, or if the stripe contents are in the
2712 * midst of changing due to a write
2714 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2715 !sh->reconstruct_state)
2716 for (i = disks; i--; )
2717 if (fetch_block(sh, s, i, disks))
2719 set_bit(STRIPE_HANDLE, &sh->state);
2723 /* handle_stripe_clean_event
2724 * any written block on an uptodate or failed drive can be returned.
2725 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2726 * never LOCKED, so we don't need to test 'failed' directly.
2728 static void handle_stripe_clean_event(struct r5conf *conf,
2729 struct stripe_head *sh, int disks, struct bio **return_bi)
2734 for (i = disks; i--; )
2735 if (sh->dev[i].written) {
2737 if (!test_bit(R5_LOCKED, &dev->flags) &&
2738 test_bit(R5_UPTODATE, &dev->flags)) {
2739 /* We can return any write requests */
2740 struct bio *wbi, *wbi2;
2741 pr_debug("Return write for disc %d\n", i);
2743 dev->written = NULL;
2744 while (wbi && wbi->bi_sector <
2745 dev->sector + STRIPE_SECTORS) {
2746 wbi2 = r5_next_bio(wbi, dev->sector);
2747 if (!raid5_dec_bi_active_stripes(wbi)) {
2748 md_write_end(conf->mddev);
2749 wbi->bi_next = *return_bi;
2754 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2756 !test_bit(STRIPE_DEGRADED, &sh->state),
2761 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2762 if (atomic_dec_and_test(&conf->pending_full_writes))
2763 md_wakeup_thread(conf->mddev->thread);
2766 static void handle_stripe_dirtying(struct r5conf *conf,
2767 struct stripe_head *sh,
2768 struct stripe_head_state *s,
2771 int rmw = 0, rcw = 0, i;
2772 if (conf->max_degraded == 2) {
2773 /* RAID6 requires 'rcw' in current implementation
2774 * Calculate the real rcw later - for now fake it
2775 * look like rcw is cheaper
2778 } else for (i = disks; i--; ) {
2779 /* would I have to read this buffer for read_modify_write */
2780 struct r5dev *dev = &sh->dev[i];
2781 if ((dev->towrite || i == sh->pd_idx) &&
2782 !test_bit(R5_LOCKED, &dev->flags) &&
2783 !(test_bit(R5_UPTODATE, &dev->flags) ||
2784 test_bit(R5_Wantcompute, &dev->flags))) {
2785 if (test_bit(R5_Insync, &dev->flags))
2788 rmw += 2*disks; /* cannot read it */
2790 /* Would I have to read this buffer for reconstruct_write */
2791 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2792 !test_bit(R5_LOCKED, &dev->flags) &&
2793 !(test_bit(R5_UPTODATE, &dev->flags) ||
2794 test_bit(R5_Wantcompute, &dev->flags))) {
2795 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2800 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2801 (unsigned long long)sh->sector, rmw, rcw);
2802 set_bit(STRIPE_HANDLE, &sh->state);
2803 if (rmw < rcw && rmw > 0)
2804 /* prefer read-modify-write, but need to get some data */
2805 for (i = disks; i--; ) {
2806 struct r5dev *dev = &sh->dev[i];
2807 if ((dev->towrite || i == sh->pd_idx) &&
2808 !test_bit(R5_LOCKED, &dev->flags) &&
2809 !(test_bit(R5_UPTODATE, &dev->flags) ||
2810 test_bit(R5_Wantcompute, &dev->flags)) &&
2811 test_bit(R5_Insync, &dev->flags)) {
2813 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2814 pr_debug("Read_old block "
2815 "%d for r-m-w\n", i);
2816 set_bit(R5_LOCKED, &dev->flags);
2817 set_bit(R5_Wantread, &dev->flags);
2820 set_bit(STRIPE_DELAYED, &sh->state);
2821 set_bit(STRIPE_HANDLE, &sh->state);
2825 if (rcw <= rmw && rcw > 0) {
2826 /* want reconstruct write, but need to get some data */
2828 for (i = disks; i--; ) {
2829 struct r5dev *dev = &sh->dev[i];
2830 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2831 i != sh->pd_idx && i != sh->qd_idx &&
2832 !test_bit(R5_LOCKED, &dev->flags) &&
2833 !(test_bit(R5_UPTODATE, &dev->flags) ||
2834 test_bit(R5_Wantcompute, &dev->flags))) {
2836 if (!test_bit(R5_Insync, &dev->flags))
2837 continue; /* it's a failed drive */
2839 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2840 pr_debug("Read_old block "
2841 "%d for Reconstruct\n", i);
2842 set_bit(R5_LOCKED, &dev->flags);
2843 set_bit(R5_Wantread, &dev->flags);
2846 set_bit(STRIPE_DELAYED, &sh->state);
2847 set_bit(STRIPE_HANDLE, &sh->state);
2852 /* now if nothing is locked, and if we have enough data,
2853 * we can start a write request
2855 /* since handle_stripe can be called at any time we need to handle the
2856 * case where a compute block operation has been submitted and then a
2857 * subsequent call wants to start a write request. raid_run_ops only
2858 * handles the case where compute block and reconstruct are requested
2859 * simultaneously. If this is not the case then new writes need to be
2860 * held off until the compute completes.
2862 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2863 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2864 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2865 schedule_reconstruction(sh, s, rcw == 0, 0);
2868 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2869 struct stripe_head_state *s, int disks)
2871 struct r5dev *dev = NULL;
2873 set_bit(STRIPE_HANDLE, &sh->state);
2875 switch (sh->check_state) {
2876 case check_state_idle:
2877 /* start a new check operation if there are no failures */
2878 if (s->failed == 0) {
2879 BUG_ON(s->uptodate != disks);
2880 sh->check_state = check_state_run;
2881 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2882 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2886 dev = &sh->dev[s->failed_num[0]];
2888 case check_state_compute_result:
2889 sh->check_state = check_state_idle;
2891 dev = &sh->dev[sh->pd_idx];
2893 /* check that a write has not made the stripe insync */
2894 if (test_bit(STRIPE_INSYNC, &sh->state))
2897 /* either failed parity check, or recovery is happening */
2898 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2899 BUG_ON(s->uptodate != disks);
2901 set_bit(R5_LOCKED, &dev->flags);
2903 set_bit(R5_Wantwrite, &dev->flags);
2905 clear_bit(STRIPE_DEGRADED, &sh->state);
2906 set_bit(STRIPE_INSYNC, &sh->state);
2908 case check_state_run:
2909 break; /* we will be called again upon completion */
2910 case check_state_check_result:
2911 sh->check_state = check_state_idle;
2913 /* if a failure occurred during the check operation, leave
2914 * STRIPE_INSYNC not set and let the stripe be handled again
2919 /* handle a successful check operation, if parity is correct
2920 * we are done. Otherwise update the mismatch count and repair
2921 * parity if !MD_RECOVERY_CHECK
2923 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2924 /* parity is correct (on disc,
2925 * not in buffer any more)
2927 set_bit(STRIPE_INSYNC, &sh->state);
2929 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2930 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2931 /* don't try to repair!! */
2932 set_bit(STRIPE_INSYNC, &sh->state);
2934 sh->check_state = check_state_compute_run;
2935 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2936 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2937 set_bit(R5_Wantcompute,
2938 &sh->dev[sh->pd_idx].flags);
2939 sh->ops.target = sh->pd_idx;
2940 sh->ops.target2 = -1;
2945 case check_state_compute_run:
2948 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2949 __func__, sh->check_state,
2950 (unsigned long long) sh->sector);
2956 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2957 struct stripe_head_state *s,
2960 int pd_idx = sh->pd_idx;
2961 int qd_idx = sh->qd_idx;
2964 set_bit(STRIPE_HANDLE, &sh->state);
2966 BUG_ON(s->failed > 2);
2968 /* Want to check and possibly repair P and Q.
2969 * However there could be one 'failed' device, in which
2970 * case we can only check one of them, possibly using the
2971 * other to generate missing data
2974 switch (sh->check_state) {
2975 case check_state_idle:
2976 /* start a new check operation if there are < 2 failures */
2977 if (s->failed == s->q_failed) {
2978 /* The only possible failed device holds Q, so it
2979 * makes sense to check P (If anything else were failed,
2980 * we would have used P to recreate it).
2982 sh->check_state = check_state_run;
2984 if (!s->q_failed && s->failed < 2) {
2985 /* Q is not failed, and we didn't use it to generate
2986 * anything, so it makes sense to check it
2988 if (sh->check_state == check_state_run)
2989 sh->check_state = check_state_run_pq;
2991 sh->check_state = check_state_run_q;
2994 /* discard potentially stale zero_sum_result */
2995 sh->ops.zero_sum_result = 0;
2997 if (sh->check_state == check_state_run) {
2998 /* async_xor_zero_sum destroys the contents of P */
2999 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3002 if (sh->check_state >= check_state_run &&
3003 sh->check_state <= check_state_run_pq) {
3004 /* async_syndrome_zero_sum preserves P and Q, so
3005 * no need to mark them !uptodate here
3007 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3011 /* we have 2-disk failure */
3012 BUG_ON(s->failed != 2);
3014 case check_state_compute_result:
3015 sh->check_state = check_state_idle;
3017 /* check that a write has not made the stripe insync */
3018 if (test_bit(STRIPE_INSYNC, &sh->state))
3021 /* now write out any block on a failed drive,
3022 * or P or Q if they were recomputed
3024 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3025 if (s->failed == 2) {
3026 dev = &sh->dev[s->failed_num[1]];
3028 set_bit(R5_LOCKED, &dev->flags);
3029 set_bit(R5_Wantwrite, &dev->flags);
3031 if (s->failed >= 1) {
3032 dev = &sh->dev[s->failed_num[0]];
3034 set_bit(R5_LOCKED, &dev->flags);
3035 set_bit(R5_Wantwrite, &dev->flags);
3037 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3038 dev = &sh->dev[pd_idx];
3040 set_bit(R5_LOCKED, &dev->flags);
3041 set_bit(R5_Wantwrite, &dev->flags);
3043 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3044 dev = &sh->dev[qd_idx];
3046 set_bit(R5_LOCKED, &dev->flags);
3047 set_bit(R5_Wantwrite, &dev->flags);
3049 clear_bit(STRIPE_DEGRADED, &sh->state);
3051 set_bit(STRIPE_INSYNC, &sh->state);
3053 case check_state_run:
3054 case check_state_run_q:
3055 case check_state_run_pq:
3056 break; /* we will be called again upon completion */
3057 case check_state_check_result:
3058 sh->check_state = check_state_idle;
3060 /* handle a successful check operation, if parity is correct
3061 * we are done. Otherwise update the mismatch count and repair
3062 * parity if !MD_RECOVERY_CHECK
3064 if (sh->ops.zero_sum_result == 0) {
3065 /* both parities are correct */
3067 set_bit(STRIPE_INSYNC, &sh->state);
3069 /* in contrast to the raid5 case we can validate
3070 * parity, but still have a failure to write
3073 sh->check_state = check_state_compute_result;
3074 /* Returning at this point means that we may go
3075 * off and bring p and/or q uptodate again so
3076 * we make sure to check zero_sum_result again
3077 * to verify if p or q need writeback
3081 conf->mddev->resync_mismatches += STRIPE_SECTORS;
3082 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3083 /* don't try to repair!! */
3084 set_bit(STRIPE_INSYNC, &sh->state);
3086 int *target = &sh->ops.target;
3088 sh->ops.target = -1;
3089 sh->ops.target2 = -1;
3090 sh->check_state = check_state_compute_run;
3091 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3092 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3093 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3094 set_bit(R5_Wantcompute,
3095 &sh->dev[pd_idx].flags);
3097 target = &sh->ops.target2;
3100 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3101 set_bit(R5_Wantcompute,
3102 &sh->dev[qd_idx].flags);
3109 case check_state_compute_run:
3112 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3113 __func__, sh->check_state,
3114 (unsigned long long) sh->sector);
3119 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3123 /* We have read all the blocks in this stripe and now we need to
3124 * copy some of them into a target stripe for expand.
3126 struct dma_async_tx_descriptor *tx = NULL;
3127 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3128 for (i = 0; i < sh->disks; i++)
3129 if (i != sh->pd_idx && i != sh->qd_idx) {
3131 struct stripe_head *sh2;
3132 struct async_submit_ctl submit;
3134 sector_t bn = compute_blocknr(sh, i, 1);
3135 sector_t s = raid5_compute_sector(conf, bn, 0,
3137 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3139 /* so far only the early blocks of this stripe
3140 * have been requested. When later blocks
3141 * get requested, we will try again
3144 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3145 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3146 /* must have already done this block */
3147 release_stripe(sh2);
3151 /* place all the copies on one channel */
3152 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3153 tx = async_memcpy(sh2->dev[dd_idx].page,
3154 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3157 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3158 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3159 for (j = 0; j < conf->raid_disks; j++)
3160 if (j != sh2->pd_idx &&
3162 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3164 if (j == conf->raid_disks) {
3165 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3166 set_bit(STRIPE_HANDLE, &sh2->state);
3168 release_stripe(sh2);
3171 /* done submitting copies, wait for them to complete */
3174 dma_wait_for_async_tx(tx);
3179 * handle_stripe - do things to a stripe.
3181 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3182 * state of various bits to see what needs to be done.
3184 * return some read requests which now have data
3185 * return some write requests which are safely on storage
3186 * schedule a read on some buffers
3187 * schedule a write of some buffers
3188 * return confirmation of parity correctness
3192 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3194 struct r5conf *conf = sh->raid_conf;
3195 int disks = sh->disks;
3198 int do_recovery = 0;
3200 memset(s, 0, sizeof(*s));
3202 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3203 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3204 s->failed_num[0] = -1;
3205 s->failed_num[1] = -1;
3207 /* Now to look around and see what can be done */
3209 for (i=disks; i--; ) {
3210 struct md_rdev *rdev;
3217 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3219 dev->toread, dev->towrite, dev->written);
3220 /* maybe we can reply to a read
3222 * new wantfill requests are only permitted while
3223 * ops_complete_biofill is guaranteed to be inactive
3225 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3226 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3227 set_bit(R5_Wantfill, &dev->flags);
3229 /* now count some things */
3230 if (test_bit(R5_LOCKED, &dev->flags))
3232 if (test_bit(R5_UPTODATE, &dev->flags))
3234 if (test_bit(R5_Wantcompute, &dev->flags)) {
3236 BUG_ON(s->compute > 2);
3239 if (test_bit(R5_Wantfill, &dev->flags))
3241 else if (dev->toread)
3245 if (!test_bit(R5_OVERWRITE, &dev->flags))
3250 /* Prefer to use the replacement for reads, but only
3251 * if it is recovered enough and has no bad blocks.
3253 rdev = rcu_dereference(conf->disks[i].replacement);
3254 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3255 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3256 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3257 &first_bad, &bad_sectors))
3258 set_bit(R5_ReadRepl, &dev->flags);
3261 set_bit(R5_NeedReplace, &dev->flags);
3262 rdev = rcu_dereference(conf->disks[i].rdev);
3263 clear_bit(R5_ReadRepl, &dev->flags);
3265 if (rdev && test_bit(Faulty, &rdev->flags))
3268 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3269 &first_bad, &bad_sectors);
3270 if (s->blocked_rdev == NULL
3271 && (test_bit(Blocked, &rdev->flags)
3274 set_bit(BlockedBadBlocks,
3276 s->blocked_rdev = rdev;
3277 atomic_inc(&rdev->nr_pending);
3280 clear_bit(R5_Insync, &dev->flags);
3284 /* also not in-sync */
3285 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3286 test_bit(R5_UPTODATE, &dev->flags)) {
3287 /* treat as in-sync, but with a read error
3288 * which we can now try to correct
3290 set_bit(R5_Insync, &dev->flags);
3291 set_bit(R5_ReadError, &dev->flags);
3293 } else if (test_bit(In_sync, &rdev->flags))
3294 set_bit(R5_Insync, &dev->flags);
3295 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3296 /* in sync if before recovery_offset */
3297 set_bit(R5_Insync, &dev->flags);
3298 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3299 test_bit(R5_Expanded, &dev->flags))
3300 /* If we've reshaped into here, we assume it is Insync.
3301 * We will shortly update recovery_offset to make
3304 set_bit(R5_Insync, &dev->flags);
3306 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3307 /* This flag does not apply to '.replacement'
3308 * only to .rdev, so make sure to check that*/
3309 struct md_rdev *rdev2 = rcu_dereference(
3310 conf->disks[i].rdev);
3312 clear_bit(R5_Insync, &dev->flags);
3313 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3314 s->handle_bad_blocks = 1;
3315 atomic_inc(&rdev2->nr_pending);
3317 clear_bit(R5_WriteError, &dev->flags);
3319 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3320 /* This flag does not apply to '.replacement'
3321 * only to .rdev, so make sure to check that*/
3322 struct md_rdev *rdev2 = rcu_dereference(
3323 conf->disks[i].rdev);
3324 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3325 s->handle_bad_blocks = 1;
3326 atomic_inc(&rdev2->nr_pending);
3328 clear_bit(R5_MadeGood, &dev->flags);
3330 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3331 struct md_rdev *rdev2 = rcu_dereference(
3332 conf->disks[i].replacement);
3333 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3334 s->handle_bad_blocks = 1;
3335 atomic_inc(&rdev2->nr_pending);
3337 clear_bit(R5_MadeGoodRepl, &dev->flags);
3339 if (!test_bit(R5_Insync, &dev->flags)) {
3340 /* The ReadError flag will just be confusing now */
3341 clear_bit(R5_ReadError, &dev->flags);
3342 clear_bit(R5_ReWrite, &dev->flags);
3344 if (test_bit(R5_ReadError, &dev->flags))
3345 clear_bit(R5_Insync, &dev->flags);
3346 if (!test_bit(R5_Insync, &dev->flags)) {
3348 s->failed_num[s->failed] = i;
3350 if (rdev && !test_bit(Faulty, &rdev->flags))
3354 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3355 /* If there is a failed device being replaced,
3356 * we must be recovering.
3357 * else if we are after recovery_cp, we must be syncing
3358 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3359 * else we can only be replacing
3360 * sync and recovery both need to read all devices, and so
3361 * use the same flag.
3364 sh->sector >= conf->mddev->recovery_cp ||
3365 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3373 static void handle_stripe(struct stripe_head *sh)
3375 struct stripe_head_state s;
3376 struct r5conf *conf = sh->raid_conf;
3379 int disks = sh->disks;
3380 struct r5dev *pdev, *qdev;
3382 clear_bit(STRIPE_HANDLE, &sh->state);
3383 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3384 /* already being handled, ensure it gets handled
3385 * again when current action finishes */
3386 set_bit(STRIPE_HANDLE, &sh->state);
3390 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3391 set_bit(STRIPE_SYNCING, &sh->state);
3392 clear_bit(STRIPE_INSYNC, &sh->state);
3394 clear_bit(STRIPE_DELAYED, &sh->state);
3396 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3397 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3398 (unsigned long long)sh->sector, sh->state,
3399 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3400 sh->check_state, sh->reconstruct_state);
3402 analyse_stripe(sh, &s);
3404 if (s.handle_bad_blocks) {
3405 set_bit(STRIPE_HANDLE, &sh->state);
3409 if (unlikely(s.blocked_rdev)) {
3410 if (s.syncing || s.expanding || s.expanded ||
3411 s.replacing || s.to_write || s.written) {
3412 set_bit(STRIPE_HANDLE, &sh->state);
3415 /* There is nothing for the blocked_rdev to block */
3416 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3417 s.blocked_rdev = NULL;
3420 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3421 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3422 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3425 pr_debug("locked=%d uptodate=%d to_read=%d"
3426 " to_write=%d failed=%d failed_num=%d,%d\n",
3427 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3428 s.failed_num[0], s.failed_num[1]);
3429 /* check if the array has lost more than max_degraded devices and,
3430 * if so, some requests might need to be failed.
3432 if (s.failed > conf->max_degraded) {
3433 sh->check_state = 0;
3434 sh->reconstruct_state = 0;
3435 if (s.to_read+s.to_write+s.written)
3436 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3437 if (s.syncing + s.replacing)
3438 handle_failed_sync(conf, sh, &s);
3442 * might be able to return some write requests if the parity blocks
3443 * are safe, or on a failed drive
3445 pdev = &sh->dev[sh->pd_idx];
3446 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3447 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3448 qdev = &sh->dev[sh->qd_idx];
3449 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3450 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3454 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3455 && !test_bit(R5_LOCKED, &pdev->flags)
3456 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3457 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3458 && !test_bit(R5_LOCKED, &qdev->flags)
3459 && test_bit(R5_UPTODATE, &qdev->flags)))))
3460 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3462 /* Now we might consider reading some blocks, either to check/generate
3463 * parity, or to satisfy requests
3464 * or to load a block that is being partially written.
3466 if (s.to_read || s.non_overwrite
3467 || (conf->level == 6 && s.to_write && s.failed)
3468 || (s.syncing && (s.uptodate + s.compute < disks))
3471 handle_stripe_fill(sh, &s, disks);
3473 /* Now we check to see if any write operations have recently
3477 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3479 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3480 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3481 sh->reconstruct_state = reconstruct_state_idle;
3483 /* All the 'written' buffers and the parity block are ready to
3484 * be written back to disk
3486 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3487 BUG_ON(sh->qd_idx >= 0 &&
3488 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3489 for (i = disks; i--; ) {
3490 struct r5dev *dev = &sh->dev[i];
3491 if (test_bit(R5_LOCKED, &dev->flags) &&
3492 (i == sh->pd_idx || i == sh->qd_idx ||
3494 pr_debug("Writing block %d\n", i);
3495 set_bit(R5_Wantwrite, &dev->flags);
3498 if (!test_bit(R5_Insync, &dev->flags) ||
3499 ((i == sh->pd_idx || i == sh->qd_idx) &&
3501 set_bit(STRIPE_INSYNC, &sh->state);
3504 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3505 s.dec_preread_active = 1;
3508 /* Now to consider new write requests and what else, if anything
3509 * should be read. We do not handle new writes when:
3510 * 1/ A 'write' operation (copy+xor) is already in flight.
3511 * 2/ A 'check' operation is in flight, as it may clobber the parity
3514 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3515 handle_stripe_dirtying(conf, sh, &s, disks);
3517 /* maybe we need to check and possibly fix the parity for this stripe
3518 * Any reads will already have been scheduled, so we just see if enough
3519 * data is available. The parity check is held off while parity
3520 * dependent operations are in flight.
3522 if (sh->check_state ||
3523 (s.syncing && s.locked == 0 &&
3524 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3525 !test_bit(STRIPE_INSYNC, &sh->state))) {
3526 if (conf->level == 6)
3527 handle_parity_checks6(conf, sh, &s, disks);
3529 handle_parity_checks5(conf, sh, &s, disks);
3532 if (s.replacing && s.locked == 0
3533 && !test_bit(STRIPE_INSYNC, &sh->state)) {
3534 /* Write out to replacement devices where possible */
3535 for (i = 0; i < conf->raid_disks; i++)
3536 if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3537 test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3538 set_bit(R5_WantReplace, &sh->dev[i].flags);
3539 set_bit(R5_LOCKED, &sh->dev[i].flags);
3542 set_bit(STRIPE_INSYNC, &sh->state);
3544 if ((s.syncing || s.replacing) && s.locked == 0 &&
3545 test_bit(STRIPE_INSYNC, &sh->state)) {
3546 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3547 clear_bit(STRIPE_SYNCING, &sh->state);
3550 /* If the failed drives are just a ReadError, then we might need
3551 * to progress the repair/check process
3553 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3554 for (i = 0; i < s.failed; i++) {
3555 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3556 if (test_bit(R5_ReadError, &dev->flags)
3557 && !test_bit(R5_LOCKED, &dev->flags)
3558 && test_bit(R5_UPTODATE, &dev->flags)
3560 if (!test_bit(R5_ReWrite, &dev->flags)) {
3561 set_bit(R5_Wantwrite, &dev->flags);
3562 set_bit(R5_ReWrite, &dev->flags);
3563 set_bit(R5_LOCKED, &dev->flags);
3566 /* let's read it back */
3567 set_bit(R5_Wantread, &dev->flags);
3568 set_bit(R5_LOCKED, &dev->flags);
3575 /* Finish reconstruct operations initiated by the expansion process */
3576 if (sh->reconstruct_state == reconstruct_state_result) {
3577 struct stripe_head *sh_src
3578 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3579 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3580 /* sh cannot be written until sh_src has been read.
3581 * so arrange for sh to be delayed a little
3583 set_bit(STRIPE_DELAYED, &sh->state);
3584 set_bit(STRIPE_HANDLE, &sh->state);
3585 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3587 atomic_inc(&conf->preread_active_stripes);
3588 release_stripe(sh_src);
3592 release_stripe(sh_src);
3594 sh->reconstruct_state = reconstruct_state_idle;
3595 clear_bit(STRIPE_EXPANDING, &sh->state);
3596 for (i = conf->raid_disks; i--; ) {
3597 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3598 set_bit(R5_LOCKED, &sh->dev[i].flags);
3603 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3604 !sh->reconstruct_state) {
3605 /* Need to write out all blocks after computing parity */
3606 sh->disks = conf->raid_disks;
3607 stripe_set_idx(sh->sector, conf, 0, sh);
3608 schedule_reconstruction(sh, &s, 1, 1);
3609 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3610 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3611 atomic_dec(&conf->reshape_stripes);
3612 wake_up(&conf->wait_for_overlap);
3613 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3616 if (s.expanding && s.locked == 0 &&
3617 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3618 handle_stripe_expansion(conf, sh);
3621 /* wait for this device to become unblocked */
3622 if (unlikely(s.blocked_rdev)) {
3623 if (conf->mddev->external)
3624 md_wait_for_blocked_rdev(s.blocked_rdev,
3627 /* Internal metadata will immediately
3628 * be written by raid5d, so we don't
3629 * need to wait here.
3631 rdev_dec_pending(s.blocked_rdev,
3635 if (s.handle_bad_blocks)
3636 for (i = disks; i--; ) {
3637 struct md_rdev *rdev;
3638 struct r5dev *dev = &sh->dev[i];
3639 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3640 /* We own a safe reference to the rdev */
3641 rdev = conf->disks[i].rdev;
3642 if (!rdev_set_badblocks(rdev, sh->sector,
3644 md_error(conf->mddev, rdev);
3645 rdev_dec_pending(rdev, conf->mddev);
3647 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3648 rdev = conf->disks[i].rdev;
3649 rdev_clear_badblocks(rdev, sh->sector,
3651 rdev_dec_pending(rdev, conf->mddev);
3653 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3654 rdev = conf->disks[i].replacement;
3656 /* rdev have been moved down */
3657 rdev = conf->disks[i].rdev;
3658 rdev_clear_badblocks(rdev, sh->sector,
3660 rdev_dec_pending(rdev, conf->mddev);
3665 raid_run_ops(sh, s.ops_request);
3669 if (s.dec_preread_active) {
3670 /* We delay this until after ops_run_io so that if make_request
3671 * is waiting on a flush, it won't continue until the writes
3672 * have actually been submitted.
3674 atomic_dec(&conf->preread_active_stripes);
3675 if (atomic_read(&conf->preread_active_stripes) <
3677 md_wakeup_thread(conf->mddev->thread);
3680 return_io(s.return_bi);
3682 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3685 static void raid5_activate_delayed(struct r5conf *conf)
3687 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3688 while (!list_empty(&conf->delayed_list)) {
3689 struct list_head *l = conf->delayed_list.next;
3690 struct stripe_head *sh;
3691 sh = list_entry(l, struct stripe_head, lru);
3693 clear_bit(STRIPE_DELAYED, &sh->state);
3694 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3695 atomic_inc(&conf->preread_active_stripes);
3696 list_add_tail(&sh->lru, &conf->hold_list);
3701 static void activate_bit_delay(struct r5conf *conf)
3703 /* device_lock is held */
3704 struct list_head head;
3705 list_add(&head, &conf->bitmap_list);
3706 list_del_init(&conf->bitmap_list);
3707 while (!list_empty(&head)) {
3708 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3709 list_del_init(&sh->lru);
3710 atomic_inc(&sh->count);
3711 __release_stripe(conf, sh);
3715 int md_raid5_congested(struct mddev *mddev, int bits)
3717 struct r5conf *conf = mddev->private;
3719 /* No difference between reads and writes. Just check
3720 * how busy the stripe_cache is
3723 if (conf->inactive_blocked)
3727 if (list_empty_careful(&conf->inactive_list))
3732 EXPORT_SYMBOL_GPL(md_raid5_congested);
3734 static int raid5_congested(void *data, int bits)
3736 struct mddev *mddev = data;
3738 return mddev_congested(mddev, bits) ||
3739 md_raid5_congested(mddev, bits);
3742 /* We want read requests to align with chunks where possible,
3743 * but write requests don't need to.
3745 static int raid5_mergeable_bvec(struct request_queue *q,
3746 struct bvec_merge_data *bvm,
3747 struct bio_vec *biovec)
3749 struct mddev *mddev = q->queuedata;
3750 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3752 unsigned int chunk_sectors = mddev->chunk_sectors;
3753 unsigned int bio_sectors = bvm->bi_size >> 9;
3755 if ((bvm->bi_rw & 1) == WRITE)
3756 return biovec->bv_len; /* always allow writes to be mergeable */
3758 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3759 chunk_sectors = mddev->new_chunk_sectors;
3760 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3761 if (max < 0) max = 0;
3762 if (max <= biovec->bv_len && bio_sectors == 0)
3763 return biovec->bv_len;
3769 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3771 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3772 unsigned int chunk_sectors = mddev->chunk_sectors;
3773 unsigned int bio_sectors = bio->bi_size >> 9;
3775 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3776 chunk_sectors = mddev->new_chunk_sectors;
3777 return chunk_sectors >=
3778 ((sector & (chunk_sectors - 1)) + bio_sectors);
3782 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3783 * later sampled by raid5d.
3785 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3787 unsigned long flags;
3789 spin_lock_irqsave(&conf->device_lock, flags);
3791 bi->bi_next = conf->retry_read_aligned_list;
3792 conf->retry_read_aligned_list = bi;
3794 spin_unlock_irqrestore(&conf->device_lock, flags);
3795 md_wakeup_thread(conf->mddev->thread);
3799 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3803 bi = conf->retry_read_aligned;
3805 conf->retry_read_aligned = NULL;
3808 bi = conf->retry_read_aligned_list;
3810 conf->retry_read_aligned_list = bi->bi_next;
3813 * this sets the active strip count to 1 and the processed
3814 * strip count to zero (upper 8 bits)
3816 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
3824 * The "raid5_align_endio" should check if the read succeeded and if it
3825 * did, call bio_endio on the original bio (having bio_put the new bio
3827 * If the read failed..
3829 static void raid5_align_endio(struct bio *bi, int error)
3831 struct bio* raid_bi = bi->bi_private;
3832 struct mddev *mddev;
3833 struct r5conf *conf;
3834 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3835 struct md_rdev *rdev;
3839 rdev = (void*)raid_bi->bi_next;
3840 raid_bi->bi_next = NULL;
3841 mddev = rdev->mddev;
3842 conf = mddev->private;
3844 rdev_dec_pending(rdev, conf->mddev);
3846 if (!error && uptodate) {
3847 bio_endio(raid_bi, 0);
3848 if (atomic_dec_and_test(&conf->active_aligned_reads))
3849 wake_up(&conf->wait_for_stripe);
3854 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3856 add_bio_to_retry(raid_bi, conf);
3859 static int bio_fits_rdev(struct bio *bi)
3861 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3863 if ((bi->bi_size>>9) > queue_max_sectors(q))
3865 blk_recount_segments(q, bi);
3866 if (bi->bi_phys_segments > queue_max_segments(q))
3869 if (q->merge_bvec_fn)
3870 /* it's too hard to apply the merge_bvec_fn at this stage,
3879 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3881 struct r5conf *conf = mddev->private;
3883 struct bio* align_bi;
3884 struct md_rdev *rdev;
3885 sector_t end_sector;
3887 if (!in_chunk_boundary(mddev, raid_bio)) {
3888 pr_debug("chunk_aligned_read : non aligned\n");
3892 * use bio_clone_mddev to make a copy of the bio
3894 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3898 * set bi_end_io to a new function, and set bi_private to the
3901 align_bi->bi_end_io = raid5_align_endio;
3902 align_bi->bi_private = raid_bio;
3906 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3910 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3912 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3913 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3914 rdev->recovery_offset < end_sector) {
3915 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3917 (test_bit(Faulty, &rdev->flags) ||
3918 !(test_bit(In_sync, &rdev->flags) ||
3919 rdev->recovery_offset >= end_sector)))
3926 atomic_inc(&rdev->nr_pending);
3928 raid_bio->bi_next = (void*)rdev;
3929 align_bi->bi_bdev = rdev->bdev;
3930 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3932 if (!bio_fits_rdev(align_bi) ||
3933 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3934 &first_bad, &bad_sectors)) {
3935 /* too big in some way, or has a known bad block */
3937 rdev_dec_pending(rdev, mddev);
3941 /* No reshape active, so we can trust rdev->data_offset */
3942 align_bi->bi_sector += rdev->data_offset;
3944 spin_lock_irq(&conf->device_lock);
3945 wait_event_lock_irq(conf->wait_for_stripe,
3947 conf->device_lock, /* nothing */);
3948 atomic_inc(&conf->active_aligned_reads);
3949 spin_unlock_irq(&conf->device_lock);
3951 generic_make_request(align_bi);
3960 /* __get_priority_stripe - get the next stripe to process
3962 * Full stripe writes are allowed to pass preread active stripes up until
3963 * the bypass_threshold is exceeded. In general the bypass_count
3964 * increments when the handle_list is handled before the hold_list; however, it
3965 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3966 * stripe with in flight i/o. The bypass_count will be reset when the
3967 * head of the hold_list has changed, i.e. the head was promoted to the
3970 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3972 struct stripe_head *sh;
3974 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3976 list_empty(&conf->handle_list) ? "empty" : "busy",
3977 list_empty(&conf->hold_list) ? "empty" : "busy",
3978 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3980 if (!list_empty(&conf->handle_list)) {
3981 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3983 if (list_empty(&conf->hold_list))
3984 conf->bypass_count = 0;
3985 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3986 if (conf->hold_list.next == conf->last_hold)
3987 conf->bypass_count++;
3989 conf->last_hold = conf->hold_list.next;
3990 conf->bypass_count -= conf->bypass_threshold;
3991 if (conf->bypass_count < 0)
3992 conf->bypass_count = 0;
3995 } else if (!list_empty(&conf->hold_list) &&
3996 ((conf->bypass_threshold &&
3997 conf->bypass_count > conf->bypass_threshold) ||
3998 atomic_read(&conf->pending_full_writes) == 0)) {
3999 sh = list_entry(conf->hold_list.next,
4001 conf->bypass_count -= conf->bypass_threshold;
4002 if (conf->bypass_count < 0)
4003 conf->bypass_count = 0;
4007 list_del_init(&sh->lru);
4008 atomic_inc(&sh->count);
4009 BUG_ON(atomic_read(&sh->count) != 1);
4013 static void make_request(struct mddev *mddev, struct bio * bi)
4015 struct r5conf *conf = mddev->private;
4017 sector_t new_sector;
4018 sector_t logical_sector, last_sector;
4019 struct stripe_head *sh;
4020 const int rw = bio_data_dir(bi);
4023 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4024 md_flush_request(mddev, bi);
4028 md_write_start(mddev, bi);
4031 mddev->reshape_position == MaxSector &&
4032 chunk_aligned_read(mddev,bi))
4035 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4036 last_sector = bi->bi_sector + (bi->bi_size>>9);
4038 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4040 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4046 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4047 if (unlikely(conf->reshape_progress != MaxSector)) {
4048 /* spinlock is needed as reshape_progress may be
4049 * 64bit on a 32bit platform, and so it might be
4050 * possible to see a half-updated value
4051 * Of course reshape_progress could change after
4052 * the lock is dropped, so once we get a reference
4053 * to the stripe that we think it is, we will have
4056 spin_lock_irq(&conf->device_lock);
4057 if (mddev->reshape_backwards
4058 ? logical_sector < conf->reshape_progress
4059 : logical_sector >= conf->reshape_progress) {
4062 if (mddev->reshape_backwards
4063 ? logical_sector < conf->reshape_safe
4064 : logical_sector >= conf->reshape_safe) {
4065 spin_unlock_irq(&conf->device_lock);
4070 spin_unlock_irq(&conf->device_lock);
4073 new_sector = raid5_compute_sector(conf, logical_sector,
4076 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4077 (unsigned long long)new_sector,
4078 (unsigned long long)logical_sector);
4080 sh = get_active_stripe(conf, new_sector, previous,
4081 (bi->bi_rw&RWA_MASK), 0);
4083 if (unlikely(previous)) {
4084 /* expansion might have moved on while waiting for a
4085 * stripe, so we must do the range check again.
4086 * Expansion could still move past after this
4087 * test, but as we are holding a reference to
4088 * 'sh', we know that if that happens,
4089 * STRIPE_EXPANDING will get set and the expansion
4090 * won't proceed until we finish with the stripe.
4093 spin_lock_irq(&conf->device_lock);
4094 if (mddev->reshape_backwards
4095 ? logical_sector >= conf->reshape_progress
4096 : logical_sector < conf->reshape_progress)
4097 /* mismatch, need to try again */
4099 spin_unlock_irq(&conf->device_lock);
4108 logical_sector >= mddev->suspend_lo &&
4109 logical_sector < mddev->suspend_hi) {
4111 /* As the suspend_* range is controlled by
4112 * userspace, we want an interruptible
4115 flush_signals(current);
4116 prepare_to_wait(&conf->wait_for_overlap,
4117 &w, TASK_INTERRUPTIBLE);
4118 if (logical_sector >= mddev->suspend_lo &&
4119 logical_sector < mddev->suspend_hi)
4124 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4125 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4126 /* Stripe is busy expanding or
4127 * add failed due to overlap. Flush everything
4130 md_wakeup_thread(mddev->thread);
4135 finish_wait(&conf->wait_for_overlap, &w);
4136 set_bit(STRIPE_HANDLE, &sh->state);
4137 clear_bit(STRIPE_DELAYED, &sh->state);
4138 if ((bi->bi_rw & REQ_NOIDLE) &&
4139 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4140 atomic_inc(&conf->preread_active_stripes);
4141 mddev_check_plugged(mddev);
4144 /* cannot get stripe for read-ahead, just give-up */
4145 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4146 finish_wait(&conf->wait_for_overlap, &w);
4151 remaining = raid5_dec_bi_active_stripes(bi);
4152 if (remaining == 0) {
4155 md_write_end(mddev);
4161 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4163 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4165 /* reshaping is quite different to recovery/resync so it is
4166 * handled quite separately ... here.
4168 * On each call to sync_request, we gather one chunk worth of
4169 * destination stripes and flag them as expanding.
4170 * Then we find all the source stripes and request reads.
4171 * As the reads complete, handle_stripe will copy the data
4172 * into the destination stripe and release that stripe.
4174 struct r5conf *conf = mddev->private;
4175 struct stripe_head *sh;
4176 sector_t first_sector, last_sector;
4177 int raid_disks = conf->previous_raid_disks;
4178 int data_disks = raid_disks - conf->max_degraded;
4179 int new_data_disks = conf->raid_disks - conf->max_degraded;
4182 sector_t writepos, readpos, safepos;
4183 sector_t stripe_addr;
4184 int reshape_sectors;
4185 struct list_head stripes;
4187 if (sector_nr == 0) {
4188 /* If restarting in the middle, skip the initial sectors */
4189 if (mddev->reshape_backwards &&
4190 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4191 sector_nr = raid5_size(mddev, 0, 0)
4192 - conf->reshape_progress;
4193 } else if (!mddev->reshape_backwards &&
4194 conf->reshape_progress > 0)
4195 sector_nr = conf->reshape_progress;
4196 sector_div(sector_nr, new_data_disks);
4198 mddev->curr_resync_completed = sector_nr;
4199 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4205 /* We need to process a full chunk at a time.
4206 * If old and new chunk sizes differ, we need to process the
4209 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4210 reshape_sectors = mddev->new_chunk_sectors;
4212 reshape_sectors = mddev->chunk_sectors;
4214 /* We update the metadata at least every 10 seconds, or when
4215 * the data about to be copied would over-write the source of
4216 * the data at the front of the range. i.e. one new_stripe
4217 * along from reshape_progress new_maps to after where
4218 * reshape_safe old_maps to
4220 writepos = conf->reshape_progress;
4221 sector_div(writepos, new_data_disks);
4222 readpos = conf->reshape_progress;
4223 sector_div(readpos, data_disks);
4224 safepos = conf->reshape_safe;
4225 sector_div(safepos, data_disks);
4226 if (mddev->reshape_backwards) {
4227 writepos -= min_t(sector_t, reshape_sectors, writepos);
4228 readpos += reshape_sectors;
4229 safepos += reshape_sectors;
4231 writepos += reshape_sectors;
4232 readpos -= min_t(sector_t, reshape_sectors, readpos);
4233 safepos -= min_t(sector_t, reshape_sectors, safepos);
4236 /* Having calculated the 'writepos' possibly use it
4237 * to set 'stripe_addr' which is where we will write to.
4239 if (mddev->reshape_backwards) {
4240 BUG_ON(conf->reshape_progress == 0);
4241 stripe_addr = writepos;
4242 BUG_ON((mddev->dev_sectors &
4243 ~((sector_t)reshape_sectors - 1))
4244 - reshape_sectors - stripe_addr
4247 BUG_ON(writepos != sector_nr + reshape_sectors);
4248 stripe_addr = sector_nr;
4251 /* 'writepos' is the most advanced device address we might write.
4252 * 'readpos' is the least advanced device address we might read.
4253 * 'safepos' is the least address recorded in the metadata as having
4255 * If there is a min_offset_diff, these are adjusted either by
4256 * increasing the safepos/readpos if diff is negative, or
4257 * increasing writepos if diff is positive.
4258 * If 'readpos' is then behind 'writepos', there is no way that we can
4259 * ensure safety in the face of a crash - that must be done by userspace
4260 * making a backup of the data. So in that case there is no particular
4261 * rush to update metadata.
4262 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4263 * update the metadata to advance 'safepos' to match 'readpos' so that
4264 * we can be safe in the event of a crash.
4265 * So we insist on updating metadata if safepos is behind writepos and
4266 * readpos is beyond writepos.
4267 * In any case, update the metadata every 10 seconds.
4268 * Maybe that number should be configurable, but I'm not sure it is
4269 * worth it.... maybe it could be a multiple of safemode_delay???
4271 if (conf->min_offset_diff < 0) {
4272 safepos += -conf->min_offset_diff;
4273 readpos += -conf->min_offset_diff;
4275 writepos += conf->min_offset_diff;
4277 if ((mddev->reshape_backwards
4278 ? (safepos > writepos && readpos < writepos)
4279 : (safepos < writepos && readpos > writepos)) ||
4280 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4281 /* Cannot proceed until we've updated the superblock... */
4282 wait_event(conf->wait_for_overlap,
4283 atomic_read(&conf->reshape_stripes)==0);
4284 mddev->reshape_position = conf->reshape_progress;
4285 mddev->curr_resync_completed = sector_nr;
4286 conf->reshape_checkpoint = jiffies;
4287 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4288 md_wakeup_thread(mddev->thread);
4289 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4290 kthread_should_stop());
4291 spin_lock_irq(&conf->device_lock);
4292 conf->reshape_safe = mddev->reshape_position;
4293 spin_unlock_irq(&conf->device_lock);
4294 wake_up(&conf->wait_for_overlap);
4295 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4298 INIT_LIST_HEAD(&stripes);
4299 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4301 int skipped_disk = 0;
4302 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4303 set_bit(STRIPE_EXPANDING, &sh->state);
4304 atomic_inc(&conf->reshape_stripes);
4305 /* If any of this stripe is beyond the end of the old
4306 * array, then we need to zero those blocks
4308 for (j=sh->disks; j--;) {
4310 if (j == sh->pd_idx)
4312 if (conf->level == 6 &&
4315 s = compute_blocknr(sh, j, 0);
4316 if (s < raid5_size(mddev, 0, 0)) {
4320 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4321 set_bit(R5_Expanded, &sh->dev[j].flags);
4322 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4324 if (!skipped_disk) {
4325 set_bit(STRIPE_EXPAND_READY, &sh->state);
4326 set_bit(STRIPE_HANDLE, &sh->state);
4328 list_add(&sh->lru, &stripes);
4330 spin_lock_irq(&conf->device_lock);
4331 if (mddev->reshape_backwards)
4332 conf->reshape_progress -= reshape_sectors * new_data_disks;
4334 conf->reshape_progress += reshape_sectors * new_data_disks;
4335 spin_unlock_irq(&conf->device_lock);
4336 /* Ok, those stripe are ready. We can start scheduling
4337 * reads on the source stripes.
4338 * The source stripes are determined by mapping the first and last
4339 * block on the destination stripes.
4342 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4345 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4346 * new_data_disks - 1),
4348 if (last_sector >= mddev->dev_sectors)
4349 last_sector = mddev->dev_sectors - 1;
4350 while (first_sector <= last_sector) {
4351 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4352 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4353 set_bit(STRIPE_HANDLE, &sh->state);
4355 first_sector += STRIPE_SECTORS;
4357 /* Now that the sources are clearly marked, we can release
4358 * the destination stripes
4360 while (!list_empty(&stripes)) {
4361 sh = list_entry(stripes.next, struct stripe_head, lru);
4362 list_del_init(&sh->lru);
4365 /* If this takes us to the resync_max point where we have to pause,
4366 * then we need to write out the superblock.
4368 sector_nr += reshape_sectors;
4369 if ((sector_nr - mddev->curr_resync_completed) * 2
4370 >= mddev->resync_max - mddev->curr_resync_completed) {
4371 /* Cannot proceed until we've updated the superblock... */
4372 wait_event(conf->wait_for_overlap,
4373 atomic_read(&conf->reshape_stripes) == 0);
4374 mddev->reshape_position = conf->reshape_progress;
4375 mddev->curr_resync_completed = sector_nr;
4376 conf->reshape_checkpoint = jiffies;
4377 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4378 md_wakeup_thread(mddev->thread);
4379 wait_event(mddev->sb_wait,
4380 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4381 || kthread_should_stop());
4382 spin_lock_irq(&conf->device_lock);
4383 conf->reshape_safe = mddev->reshape_position;
4384 spin_unlock_irq(&conf->device_lock);
4385 wake_up(&conf->wait_for_overlap);
4386 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4388 return reshape_sectors;
4391 /* FIXME go_faster isn't used */
4392 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4394 struct r5conf *conf = mddev->private;
4395 struct stripe_head *sh;
4396 sector_t max_sector = mddev->dev_sectors;
4397 sector_t sync_blocks;
4398 int still_degraded = 0;
4401 if (sector_nr >= max_sector) {
4402 /* just being told to finish up .. nothing much to do */
4404 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4409 if (mddev->curr_resync < max_sector) /* aborted */
4410 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4412 else /* completed sync */
4414 bitmap_close_sync(mddev->bitmap);
4419 /* Allow raid5_quiesce to complete */
4420 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4422 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4423 return reshape_request(mddev, sector_nr, skipped);
4425 /* No need to check resync_max as we never do more than one
4426 * stripe, and as resync_max will always be on a chunk boundary,
4427 * if the check in md_do_sync didn't fire, there is no chance
4428 * of overstepping resync_max here
4431 /* if there is too many failed drives and we are trying
4432 * to resync, then assert that we are finished, because there is
4433 * nothing we can do.
4435 if (mddev->degraded >= conf->max_degraded &&
4436 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4437 sector_t rv = mddev->dev_sectors - sector_nr;
4441 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4442 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4443 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4444 /* we can skip this block, and probably more */
4445 sync_blocks /= STRIPE_SECTORS;
4447 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4450 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4452 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4454 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4455 /* make sure we don't swamp the stripe cache if someone else
4456 * is trying to get access
4458 schedule_timeout_uninterruptible(1);
4460 /* Need to check if array will still be degraded after recovery/resync
4461 * We don't need to check the 'failed' flag as when that gets set,
4464 for (i = 0; i < conf->raid_disks; i++)
4465 if (conf->disks[i].rdev == NULL)
4468 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4470 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4475 return STRIPE_SECTORS;
4478 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4480 /* We may not be able to submit a whole bio at once as there
4481 * may not be enough stripe_heads available.
4482 * We cannot pre-allocate enough stripe_heads as we may need
4483 * more than exist in the cache (if we allow ever large chunks).
4484 * So we do one stripe head at a time and record in
4485 * ->bi_hw_segments how many have been done.
4487 * We *know* that this entire raid_bio is in one chunk, so
4488 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4490 struct stripe_head *sh;
4492 sector_t sector, logical_sector, last_sector;
4497 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4498 sector = raid5_compute_sector(conf, logical_sector,
4500 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4502 for (; logical_sector < last_sector;
4503 logical_sector += STRIPE_SECTORS,
4504 sector += STRIPE_SECTORS,
4507 if (scnt < raid5_bi_processed_stripes(raid_bio))
4508 /* already done this stripe */
4511 sh = get_active_stripe(conf, sector, 0, 1, 0);
4514 /* failed to get a stripe - must wait */
4515 raid5_set_bi_processed_stripes(raid_bio, scnt);
4516 conf->retry_read_aligned = raid_bio;
4520 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4522 raid5_set_bi_processed_stripes(raid_bio, scnt);
4523 conf->retry_read_aligned = raid_bio;
4527 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4532 remaining = raid5_dec_bi_active_stripes(raid_bio);
4534 bio_endio(raid_bio, 0);
4535 if (atomic_dec_and_test(&conf->active_aligned_reads))
4536 wake_up(&conf->wait_for_stripe);
4542 * This is our raid5 kernel thread.
4544 * We scan the hash table for stripes which can be handled now.
4545 * During the scan, completed stripes are saved for us by the interrupt
4546 * handler, so that they will not have to wait for our next wakeup.
4548 static void raid5d(struct mddev *mddev)
4550 struct stripe_head *sh;
4551 struct r5conf *conf = mddev->private;
4553 struct blk_plug plug;
4555 pr_debug("+++ raid5d active\n");
4557 md_check_recovery(mddev);
4559 blk_start_plug(&plug);
4561 spin_lock_irq(&conf->device_lock);
4566 !list_empty(&conf->bitmap_list)) {
4567 /* Now is a good time to flush some bitmap updates */
4569 spin_unlock_irq(&conf->device_lock);
4570 bitmap_unplug(mddev->bitmap);
4571 spin_lock_irq(&conf->device_lock);
4572 conf->seq_write = conf->seq_flush;
4573 activate_bit_delay(conf);
4575 raid5_activate_delayed(conf);
4577 while ((bio = remove_bio_from_retry(conf))) {
4579 spin_unlock_irq(&conf->device_lock);
4580 ok = retry_aligned_read(conf, bio);
4581 spin_lock_irq(&conf->device_lock);
4587 sh = __get_priority_stripe(conf);
4591 spin_unlock_irq(&conf->device_lock);
4598 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4599 md_check_recovery(mddev);
4601 spin_lock_irq(&conf->device_lock);
4603 pr_debug("%d stripes handled\n", handled);
4605 spin_unlock_irq(&conf->device_lock);
4607 async_tx_issue_pending_all();
4608 blk_finish_plug(&plug);
4610 pr_debug("--- raid5d inactive\n");
4614 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4616 struct r5conf *conf = mddev->private;
4618 return sprintf(page, "%d\n", conf->max_nr_stripes);
4624 raid5_set_cache_size(struct mddev *mddev, int size)
4626 struct r5conf *conf = mddev->private;
4629 if (size <= 16 || size > 32768)
4631 while (size < conf->max_nr_stripes) {
4632 if (drop_one_stripe(conf))
4633 conf->max_nr_stripes--;
4637 err = md_allow_write(mddev);
4640 while (size > conf->max_nr_stripes) {
4641 if (grow_one_stripe(conf))
4642 conf->max_nr_stripes++;
4647 EXPORT_SYMBOL(raid5_set_cache_size);
4650 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4652 struct r5conf *conf = mddev->private;
4656 if (len >= PAGE_SIZE)
4661 if (strict_strtoul(page, 10, &new))
4663 err = raid5_set_cache_size(mddev, new);
4669 static struct md_sysfs_entry
4670 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4671 raid5_show_stripe_cache_size,
4672 raid5_store_stripe_cache_size);
4675 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4677 struct r5conf *conf = mddev->private;
4679 return sprintf(page, "%d\n", conf->bypass_threshold);
4685 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4687 struct r5conf *conf = mddev->private;
4689 if (len >= PAGE_SIZE)
4694 if (strict_strtoul(page, 10, &new))
4696 if (new > conf->max_nr_stripes)
4698 conf->bypass_threshold = new;
4702 static struct md_sysfs_entry
4703 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4705 raid5_show_preread_threshold,
4706 raid5_store_preread_threshold);
4709 stripe_cache_active_show(struct mddev *mddev, char *page)
4711 struct r5conf *conf = mddev->private;
4713 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4718 static struct md_sysfs_entry
4719 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4721 static struct attribute *raid5_attrs[] = {
4722 &raid5_stripecache_size.attr,
4723 &raid5_stripecache_active.attr,
4724 &raid5_preread_bypass_threshold.attr,
4727 static struct attribute_group raid5_attrs_group = {
4729 .attrs = raid5_attrs,
4733 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4735 struct r5conf *conf = mddev->private;
4738 sectors = mddev->dev_sectors;
4740 /* size is defined by the smallest of previous and new size */
4741 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4743 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4744 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4745 return sectors * (raid_disks - conf->max_degraded);
4748 static void raid5_free_percpu(struct r5conf *conf)
4750 struct raid5_percpu *percpu;
4757 for_each_possible_cpu(cpu) {
4758 percpu = per_cpu_ptr(conf->percpu, cpu);
4759 safe_put_page(percpu->spare_page);
4760 kfree(percpu->scribble);
4762 #ifdef CONFIG_HOTPLUG_CPU
4763 unregister_cpu_notifier(&conf->cpu_notify);
4767 free_percpu(conf->percpu);
4770 static void free_conf(struct r5conf *conf)
4772 shrink_stripes(conf);
4773 raid5_free_percpu(conf);
4775 kfree(conf->stripe_hashtbl);
4779 #ifdef CONFIG_HOTPLUG_CPU
4780 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4783 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4784 long cpu = (long)hcpu;
4785 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4788 case CPU_UP_PREPARE:
4789 case CPU_UP_PREPARE_FROZEN:
4790 if (conf->level == 6 && !percpu->spare_page)
4791 percpu->spare_page = alloc_page(GFP_KERNEL);
4792 if (!percpu->scribble)
4793 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4795 if (!percpu->scribble ||
4796 (conf->level == 6 && !percpu->spare_page)) {
4797 safe_put_page(percpu->spare_page);
4798 kfree(percpu->scribble);
4799 pr_err("%s: failed memory allocation for cpu%ld\n",
4801 return notifier_from_errno(-ENOMEM);
4805 case CPU_DEAD_FROZEN:
4806 safe_put_page(percpu->spare_page);
4807 kfree(percpu->scribble);
4808 percpu->spare_page = NULL;
4809 percpu->scribble = NULL;
4818 static int raid5_alloc_percpu(struct r5conf *conf)
4821 struct page *spare_page;
4822 struct raid5_percpu __percpu *allcpus;
4826 allcpus = alloc_percpu(struct raid5_percpu);
4829 conf->percpu = allcpus;
4833 for_each_present_cpu(cpu) {
4834 if (conf->level == 6) {
4835 spare_page = alloc_page(GFP_KERNEL);
4840 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4842 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4847 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4849 #ifdef CONFIG_HOTPLUG_CPU
4850 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4851 conf->cpu_notify.priority = 0;
4853 err = register_cpu_notifier(&conf->cpu_notify);
4860 static struct r5conf *setup_conf(struct mddev *mddev)
4862 struct r5conf *conf;
4863 int raid_disk, memory, max_disks;
4864 struct md_rdev *rdev;
4865 struct disk_info *disk;
4868 if (mddev->new_level != 5
4869 && mddev->new_level != 4
4870 && mddev->new_level != 6) {
4871 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4872 mdname(mddev), mddev->new_level);
4873 return ERR_PTR(-EIO);
4875 if ((mddev->new_level == 5
4876 && !algorithm_valid_raid5(mddev->new_layout)) ||
4877 (mddev->new_level == 6
4878 && !algorithm_valid_raid6(mddev->new_layout))) {
4879 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4880 mdname(mddev), mddev->new_layout);
4881 return ERR_PTR(-EIO);
4883 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4884 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4885 mdname(mddev), mddev->raid_disks);
4886 return ERR_PTR(-EINVAL);
4889 if (!mddev->new_chunk_sectors ||
4890 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4891 !is_power_of_2(mddev->new_chunk_sectors)) {
4892 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4893 mdname(mddev), mddev->new_chunk_sectors << 9);
4894 return ERR_PTR(-EINVAL);
4897 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4900 spin_lock_init(&conf->device_lock);
4901 init_waitqueue_head(&conf->wait_for_stripe);
4902 init_waitqueue_head(&conf->wait_for_overlap);
4903 INIT_LIST_HEAD(&conf->handle_list);
4904 INIT_LIST_HEAD(&conf->hold_list);
4905 INIT_LIST_HEAD(&conf->delayed_list);
4906 INIT_LIST_HEAD(&conf->bitmap_list);
4907 INIT_LIST_HEAD(&conf->inactive_list);
4908 atomic_set(&conf->active_stripes, 0);
4909 atomic_set(&conf->preread_active_stripes, 0);
4910 atomic_set(&conf->active_aligned_reads, 0);
4911 conf->bypass_threshold = BYPASS_THRESHOLD;
4912 conf->recovery_disabled = mddev->recovery_disabled - 1;
4914 conf->raid_disks = mddev->raid_disks;
4915 if (mddev->reshape_position == MaxSector)
4916 conf->previous_raid_disks = mddev->raid_disks;
4918 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4919 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4920 conf->scribble_len = scribble_len(max_disks);
4922 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4927 conf->mddev = mddev;
4929 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4932 conf->level = mddev->new_level;
4933 if (raid5_alloc_percpu(conf) != 0)
4936 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4938 rdev_for_each(rdev, mddev) {
4939 raid_disk = rdev->raid_disk;
4940 if (raid_disk >= max_disks
4943 disk = conf->disks + raid_disk;
4945 if (test_bit(Replacement, &rdev->flags)) {
4946 if (disk->replacement)
4948 disk->replacement = rdev;
4955 if (test_bit(In_sync, &rdev->flags)) {
4956 char b[BDEVNAME_SIZE];
4957 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4959 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4960 } else if (rdev->saved_raid_disk != raid_disk)
4961 /* Cannot rely on bitmap to complete recovery */
4965 conf->chunk_sectors = mddev->new_chunk_sectors;
4966 conf->level = mddev->new_level;
4967 if (conf->level == 6)
4968 conf->max_degraded = 2;
4970 conf->max_degraded = 1;
4971 conf->algorithm = mddev->new_layout;
4972 conf->max_nr_stripes = NR_STRIPES;
4973 conf->reshape_progress = mddev->reshape_position;
4974 if (conf->reshape_progress != MaxSector) {
4975 conf->prev_chunk_sectors = mddev->chunk_sectors;
4976 conf->prev_algo = mddev->layout;
4979 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4980 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4981 if (grow_stripes(conf, conf->max_nr_stripes)) {
4983 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4984 mdname(mddev), memory);
4987 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4988 mdname(mddev), memory);
4990 sprintf(pers_name, "raid%d", mddev->new_level);
4991 conf->thread = md_register_thread(raid5d, mddev, pers_name);
4992 if (!conf->thread) {
4994 "md/raid:%s: couldn't allocate thread.\n",
5004 return ERR_PTR(-EIO);
5006 return ERR_PTR(-ENOMEM);
5010 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5013 case ALGORITHM_PARITY_0:
5014 if (raid_disk < max_degraded)
5017 case ALGORITHM_PARITY_N:
5018 if (raid_disk >= raid_disks - max_degraded)
5021 case ALGORITHM_PARITY_0_6:
5022 if (raid_disk == 0 ||
5023 raid_disk == raid_disks - 1)
5026 case ALGORITHM_LEFT_ASYMMETRIC_6:
5027 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5028 case ALGORITHM_LEFT_SYMMETRIC_6:
5029 case ALGORITHM_RIGHT_SYMMETRIC_6:
5030 if (raid_disk == raid_disks - 1)
5036 static int run(struct mddev *mddev)
5038 struct r5conf *conf;
5039 int working_disks = 0;
5040 int dirty_parity_disks = 0;
5041 struct md_rdev *rdev;
5042 sector_t reshape_offset = 0;
5044 long long min_offset_diff = 0;
5047 if (mddev->recovery_cp != MaxSector)
5048 printk(KERN_NOTICE "md/raid:%s: not clean"
5049 " -- starting background reconstruction\n",
5052 rdev_for_each(rdev, mddev) {
5054 if (rdev->raid_disk < 0)
5056 diff = (rdev->new_data_offset - rdev->data_offset);
5058 min_offset_diff = diff;
5060 } else if (mddev->reshape_backwards &&
5061 diff < min_offset_diff)
5062 min_offset_diff = diff;
5063 else if (!mddev->reshape_backwards &&
5064 diff > min_offset_diff)
5065 min_offset_diff = diff;
5068 if (mddev->reshape_position != MaxSector) {
5069 /* Check that we can continue the reshape.
5070 * Difficulties arise if the stripe we would write to
5071 * next is at or after the stripe we would read from next.
5072 * For a reshape that changes the number of devices, this
5073 * is only possible for a very short time, and mdadm makes
5074 * sure that time appears to have past before assembling
5075 * the array. So we fail if that time hasn't passed.
5076 * For a reshape that keeps the number of devices the same
5077 * mdadm must be monitoring the reshape can keeping the
5078 * critical areas read-only and backed up. It will start
5079 * the array in read-only mode, so we check for that.
5081 sector_t here_new, here_old;
5083 int max_degraded = (mddev->level == 6 ? 2 : 1);
5085 if (mddev->new_level != mddev->level) {
5086 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5087 "required - aborting.\n",
5091 old_disks = mddev->raid_disks - mddev->delta_disks;
5092 /* reshape_position must be on a new-stripe boundary, and one
5093 * further up in new geometry must map after here in old
5096 here_new = mddev->reshape_position;
5097 if (sector_div(here_new, mddev->new_chunk_sectors *
5098 (mddev->raid_disks - max_degraded))) {
5099 printk(KERN_ERR "md/raid:%s: reshape_position not "
5100 "on a stripe boundary\n", mdname(mddev));
5103 reshape_offset = here_new * mddev->new_chunk_sectors;
5104 /* here_new is the stripe we will write to */
5105 here_old = mddev->reshape_position;
5106 sector_div(here_old, mddev->chunk_sectors *
5107 (old_disks-max_degraded));
5108 /* here_old is the first stripe that we might need to read
5110 if (mddev->delta_disks == 0) {
5111 if ((here_new * mddev->new_chunk_sectors !=
5112 here_old * mddev->chunk_sectors)) {
5113 printk(KERN_ERR "md/raid:%s: reshape position is"
5114 " confused - aborting\n", mdname(mddev));
5117 /* We cannot be sure it is safe to start an in-place
5118 * reshape. It is only safe if user-space is monitoring
5119 * and taking constant backups.
5120 * mdadm always starts a situation like this in
5121 * readonly mode so it can take control before
5122 * allowing any writes. So just check for that.
5124 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5125 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5126 /* not really in-place - so OK */;
5127 else if (mddev->ro == 0) {
5128 printk(KERN_ERR "md/raid:%s: in-place reshape "
5129 "must be started in read-only mode "
5134 } else if (mddev->reshape_backwards
5135 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5136 here_old * mddev->chunk_sectors)
5137 : (here_new * mddev->new_chunk_sectors >=
5138 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5139 /* Reading from the same stripe as writing to - bad */
5140 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5141 "auto-recovery - aborting.\n",
5145 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5147 /* OK, we should be able to continue; */
5149 BUG_ON(mddev->level != mddev->new_level);
5150 BUG_ON(mddev->layout != mddev->new_layout);
5151 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5152 BUG_ON(mddev->delta_disks != 0);
5155 if (mddev->private == NULL)
5156 conf = setup_conf(mddev);
5158 conf = mddev->private;
5161 return PTR_ERR(conf);
5163 conf->min_offset_diff = min_offset_diff;
5164 mddev->thread = conf->thread;
5165 conf->thread = NULL;
5166 mddev->private = conf;
5168 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5170 rdev = conf->disks[i].rdev;
5171 if (!rdev && conf->disks[i].replacement) {
5172 /* The replacement is all we have yet */
5173 rdev = conf->disks[i].replacement;
5174 conf->disks[i].replacement = NULL;
5175 clear_bit(Replacement, &rdev->flags);
5176 conf->disks[i].rdev = rdev;
5180 if (conf->disks[i].replacement &&
5181 conf->reshape_progress != MaxSector) {
5182 /* replacements and reshape simply do not mix. */
5183 printk(KERN_ERR "md: cannot handle concurrent "
5184 "replacement and reshape.\n");
5187 if (test_bit(In_sync, &rdev->flags)) {
5191 /* This disc is not fully in-sync. However if it
5192 * just stored parity (beyond the recovery_offset),
5193 * when we don't need to be concerned about the
5194 * array being dirty.
5195 * When reshape goes 'backwards', we never have
5196 * partially completed devices, so we only need
5197 * to worry about reshape going forwards.
5199 /* Hack because v0.91 doesn't store recovery_offset properly. */
5200 if (mddev->major_version == 0 &&
5201 mddev->minor_version > 90)
5202 rdev->recovery_offset = reshape_offset;
5204 if (rdev->recovery_offset < reshape_offset) {
5205 /* We need to check old and new layout */
5206 if (!only_parity(rdev->raid_disk,
5209 conf->max_degraded))
5212 if (!only_parity(rdev->raid_disk,
5214 conf->previous_raid_disks,
5215 conf->max_degraded))
5217 dirty_parity_disks++;
5221 * 0 for a fully functional array, 1 or 2 for a degraded array.
5223 mddev->degraded = calc_degraded(conf);
5225 if (has_failed(conf)) {
5226 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5227 " (%d/%d failed)\n",
5228 mdname(mddev), mddev->degraded, conf->raid_disks);
5232 /* device size must be a multiple of chunk size */
5233 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5234 mddev->resync_max_sectors = mddev->dev_sectors;
5236 if (mddev->degraded > dirty_parity_disks &&
5237 mddev->recovery_cp != MaxSector) {
5238 if (mddev->ok_start_degraded)
5240 "md/raid:%s: starting dirty degraded array"
5241 " - data corruption possible.\n",
5245 "md/raid:%s: cannot start dirty degraded array.\n",
5251 if (mddev->degraded == 0)
5252 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5253 " devices, algorithm %d\n", mdname(mddev), conf->level,
5254 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5257 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5258 " out of %d devices, algorithm %d\n",
5259 mdname(mddev), conf->level,
5260 mddev->raid_disks - mddev->degraded,
5261 mddev->raid_disks, mddev->new_layout);
5263 print_raid5_conf(conf);
5265 if (conf->reshape_progress != MaxSector) {
5266 conf->reshape_safe = conf->reshape_progress;
5267 atomic_set(&conf->reshape_stripes, 0);
5268 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5269 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5270 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5271 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5272 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5277 /* Ok, everything is just fine now */
5278 if (mddev->to_remove == &raid5_attrs_group)
5279 mddev->to_remove = NULL;
5280 else if (mddev->kobj.sd &&
5281 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5283 "raid5: failed to create sysfs attributes for %s\n",
5285 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5289 /* read-ahead size must cover two whole stripes, which
5290 * is 2 * (datadisks) * chunksize where 'n' is the
5291 * number of raid devices
5293 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5294 int stripe = data_disks *
5295 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5296 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5297 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5299 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5301 mddev->queue->backing_dev_info.congested_data = mddev;
5302 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5304 chunk_size = mddev->chunk_sectors << 9;
5305 blk_queue_io_min(mddev->queue, chunk_size);
5306 blk_queue_io_opt(mddev->queue, chunk_size *
5307 (conf->raid_disks - conf->max_degraded));
5309 rdev_for_each(rdev, mddev) {
5310 disk_stack_limits(mddev->gendisk, rdev->bdev,
5311 rdev->data_offset << 9);
5312 disk_stack_limits(mddev->gendisk, rdev->bdev,
5313 rdev->new_data_offset << 9);
5319 md_unregister_thread(&mddev->thread);
5320 print_raid5_conf(conf);
5322 mddev->private = NULL;
5323 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5327 static int stop(struct mddev *mddev)
5329 struct r5conf *conf = mddev->private;
5331 md_unregister_thread(&mddev->thread);
5333 mddev->queue->backing_dev_info.congested_fn = NULL;
5335 mddev->private = NULL;
5336 mddev->to_remove = &raid5_attrs_group;
5340 static void status(struct seq_file *seq, struct mddev *mddev)
5342 struct r5conf *conf = mddev->private;
5345 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5346 mddev->chunk_sectors / 2, mddev->layout);
5347 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5348 for (i = 0; i < conf->raid_disks; i++)
5349 seq_printf (seq, "%s",
5350 conf->disks[i].rdev &&
5351 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5352 seq_printf (seq, "]");
5355 static void print_raid5_conf (struct r5conf *conf)
5358 struct disk_info *tmp;
5360 printk(KERN_DEBUG "RAID conf printout:\n");
5362 printk("(conf==NULL)\n");
5365 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5367 conf->raid_disks - conf->mddev->degraded);
5369 for (i = 0; i < conf->raid_disks; i++) {
5370 char b[BDEVNAME_SIZE];
5371 tmp = conf->disks + i;
5373 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5374 i, !test_bit(Faulty, &tmp->rdev->flags),
5375 bdevname(tmp->rdev->bdev, b));
5379 static int raid5_spare_active(struct mddev *mddev)
5382 struct r5conf *conf = mddev->private;
5383 struct disk_info *tmp;
5385 unsigned long flags;
5387 for (i = 0; i < conf->raid_disks; i++) {
5388 tmp = conf->disks + i;
5389 if (tmp->replacement
5390 && tmp->replacement->recovery_offset == MaxSector
5391 && !test_bit(Faulty, &tmp->replacement->flags)
5392 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5393 /* Replacement has just become active. */
5395 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5398 /* Replaced device not technically faulty,
5399 * but we need to be sure it gets removed
5400 * and never re-added.
5402 set_bit(Faulty, &tmp->rdev->flags);
5403 sysfs_notify_dirent_safe(
5404 tmp->rdev->sysfs_state);
5406 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5407 } else if (tmp->rdev
5408 && tmp->rdev->recovery_offset == MaxSector
5409 && !test_bit(Faulty, &tmp->rdev->flags)
5410 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5412 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5415 spin_lock_irqsave(&conf->device_lock, flags);
5416 mddev->degraded = calc_degraded(conf);
5417 spin_unlock_irqrestore(&conf->device_lock, flags);
5418 print_raid5_conf(conf);
5422 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5424 struct r5conf *conf = mddev->private;
5426 int number = rdev->raid_disk;
5427 struct md_rdev **rdevp;
5428 struct disk_info *p = conf->disks + number;
5430 print_raid5_conf(conf);
5431 if (rdev == p->rdev)
5433 else if (rdev == p->replacement)
5434 rdevp = &p->replacement;
5438 if (number >= conf->raid_disks &&
5439 conf->reshape_progress == MaxSector)
5440 clear_bit(In_sync, &rdev->flags);
5442 if (test_bit(In_sync, &rdev->flags) ||
5443 atomic_read(&rdev->nr_pending)) {
5447 /* Only remove non-faulty devices if recovery
5450 if (!test_bit(Faulty, &rdev->flags) &&
5451 mddev->recovery_disabled != conf->recovery_disabled &&
5452 !has_failed(conf) &&
5453 (!p->replacement || p->replacement == rdev) &&
5454 number < conf->raid_disks) {
5460 if (atomic_read(&rdev->nr_pending)) {
5461 /* lost the race, try later */
5464 } else if (p->replacement) {
5465 /* We must have just cleared 'rdev' */
5466 p->rdev = p->replacement;
5467 clear_bit(Replacement, &p->replacement->flags);
5468 smp_mb(); /* Make sure other CPUs may see both as identical
5469 * but will never see neither - if they are careful
5471 p->replacement = NULL;
5472 clear_bit(WantReplacement, &rdev->flags);
5474 /* We might have just removed the Replacement as faulty-
5475 * clear the bit just in case
5477 clear_bit(WantReplacement, &rdev->flags);
5480 print_raid5_conf(conf);
5484 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5486 struct r5conf *conf = mddev->private;
5489 struct disk_info *p;
5491 int last = conf->raid_disks - 1;
5493 if (mddev->recovery_disabled == conf->recovery_disabled)
5496 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5497 /* no point adding a device */
5500 if (rdev->raid_disk >= 0)
5501 first = last = rdev->raid_disk;
5504 * find the disk ... but prefer rdev->saved_raid_disk
5507 if (rdev->saved_raid_disk >= 0 &&
5508 rdev->saved_raid_disk >= first &&
5509 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5510 first = rdev->saved_raid_disk;
5512 for (disk = first; disk <= last; disk++) {
5513 p = conf->disks + disk;
5514 if (p->rdev == NULL) {
5515 clear_bit(In_sync, &rdev->flags);
5516 rdev->raid_disk = disk;
5518 if (rdev->saved_raid_disk != disk)
5520 rcu_assign_pointer(p->rdev, rdev);
5524 for (disk = first; disk <= last; disk++) {
5525 p = conf->disks + disk;
5526 if (test_bit(WantReplacement, &p->rdev->flags) &&
5527 p->replacement == NULL) {
5528 clear_bit(In_sync, &rdev->flags);
5529 set_bit(Replacement, &rdev->flags);
5530 rdev->raid_disk = disk;
5533 rcu_assign_pointer(p->replacement, rdev);
5538 print_raid5_conf(conf);
5542 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5544 /* no resync is happening, and there is enough space
5545 * on all devices, so we can resize.
5546 * We need to make sure resync covers any new space.
5547 * If the array is shrinking we should possibly wait until
5548 * any io in the removed space completes, but it hardly seems
5552 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5553 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5554 if (mddev->external_size &&
5555 mddev->array_sectors > newsize)
5557 if (mddev->bitmap) {
5558 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5562 md_set_array_sectors(mddev, newsize);
5563 set_capacity(mddev->gendisk, mddev->array_sectors);
5564 revalidate_disk(mddev->gendisk);
5565 if (sectors > mddev->dev_sectors &&
5566 mddev->recovery_cp > mddev->dev_sectors) {
5567 mddev->recovery_cp = mddev->dev_sectors;
5568 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5570 mddev->dev_sectors = sectors;
5571 mddev->resync_max_sectors = sectors;
5575 static int check_stripe_cache(struct mddev *mddev)
5577 /* Can only proceed if there are plenty of stripe_heads.
5578 * We need a minimum of one full stripe,, and for sensible progress
5579 * it is best to have about 4 times that.
5580 * If we require 4 times, then the default 256 4K stripe_heads will
5581 * allow for chunk sizes up to 256K, which is probably OK.
5582 * If the chunk size is greater, user-space should request more
5583 * stripe_heads first.
5585 struct r5conf *conf = mddev->private;
5586 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5587 > conf->max_nr_stripes ||
5588 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5589 > conf->max_nr_stripes) {
5590 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5592 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5599 static int check_reshape(struct mddev *mddev)
5601 struct r5conf *conf = mddev->private;
5603 if (mddev->delta_disks == 0 &&
5604 mddev->new_layout == mddev->layout &&
5605 mddev->new_chunk_sectors == mddev->chunk_sectors)
5606 return 0; /* nothing to do */
5607 if (has_failed(conf))
5609 if (mddev->delta_disks < 0) {
5610 /* We might be able to shrink, but the devices must
5611 * be made bigger first.
5612 * For raid6, 4 is the minimum size.
5613 * Otherwise 2 is the minimum
5616 if (mddev->level == 6)
5618 if (mddev->raid_disks + mddev->delta_disks < min)
5622 if (!check_stripe_cache(mddev))
5625 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5628 static int raid5_start_reshape(struct mddev *mddev)
5630 struct r5conf *conf = mddev->private;
5631 struct md_rdev *rdev;
5633 unsigned long flags;
5635 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5638 if (!check_stripe_cache(mddev))
5641 if (has_failed(conf))
5644 rdev_for_each(rdev, mddev) {
5645 if (!test_bit(In_sync, &rdev->flags)
5646 && !test_bit(Faulty, &rdev->flags))
5650 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5651 /* Not enough devices even to make a degraded array
5656 /* Refuse to reduce size of the array. Any reductions in
5657 * array size must be through explicit setting of array_size
5660 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5661 < mddev->array_sectors) {
5662 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5663 "before number of disks\n", mdname(mddev));
5667 atomic_set(&conf->reshape_stripes, 0);
5668 spin_lock_irq(&conf->device_lock);
5669 conf->previous_raid_disks = conf->raid_disks;
5670 conf->raid_disks += mddev->delta_disks;
5671 conf->prev_chunk_sectors = conf->chunk_sectors;
5672 conf->chunk_sectors = mddev->new_chunk_sectors;
5673 conf->prev_algo = conf->algorithm;
5674 conf->algorithm = mddev->new_layout;
5676 /* Code that selects data_offset needs to see the generation update
5677 * if reshape_progress has been set - so a memory barrier needed.
5680 if (mddev->reshape_backwards)
5681 conf->reshape_progress = raid5_size(mddev, 0, 0);
5683 conf->reshape_progress = 0;
5684 conf->reshape_safe = conf->reshape_progress;
5685 spin_unlock_irq(&conf->device_lock);
5687 /* Add some new drives, as many as will fit.
5688 * We know there are enough to make the newly sized array work.
5689 * Don't add devices if we are reducing the number of
5690 * devices in the array. This is because it is not possible
5691 * to correctly record the "partially reconstructed" state of
5692 * such devices during the reshape and confusion could result.
5694 if (mddev->delta_disks >= 0) {
5695 rdev_for_each(rdev, mddev)
5696 if (rdev->raid_disk < 0 &&
5697 !test_bit(Faulty, &rdev->flags)) {
5698 if (raid5_add_disk(mddev, rdev) == 0) {
5700 >= conf->previous_raid_disks)
5701 set_bit(In_sync, &rdev->flags);
5703 rdev->recovery_offset = 0;
5705 if (sysfs_link_rdev(mddev, rdev))
5706 /* Failure here is OK */;
5708 } else if (rdev->raid_disk >= conf->previous_raid_disks
5709 && !test_bit(Faulty, &rdev->flags)) {
5710 /* This is a spare that was manually added */
5711 set_bit(In_sync, &rdev->flags);
5714 /* When a reshape changes the number of devices,
5715 * ->degraded is measured against the larger of the
5716 * pre and post number of devices.
5718 spin_lock_irqsave(&conf->device_lock, flags);
5719 mddev->degraded = calc_degraded(conf);
5720 spin_unlock_irqrestore(&conf->device_lock, flags);
5722 mddev->raid_disks = conf->raid_disks;
5723 mddev->reshape_position = conf->reshape_progress;
5724 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5726 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5727 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5728 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5729 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5730 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5732 if (!mddev->sync_thread) {
5733 mddev->recovery = 0;
5734 spin_lock_irq(&conf->device_lock);
5735 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5736 rdev_for_each(rdev, mddev)
5737 rdev->new_data_offset = rdev->data_offset;
5739 conf->reshape_progress = MaxSector;
5740 mddev->reshape_position = MaxSector;
5741 spin_unlock_irq(&conf->device_lock);
5744 conf->reshape_checkpoint = jiffies;
5745 md_wakeup_thread(mddev->sync_thread);
5746 md_new_event(mddev);
5750 /* This is called from the reshape thread and should make any
5751 * changes needed in 'conf'
5753 static void end_reshape(struct r5conf *conf)
5756 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5757 struct md_rdev *rdev;
5759 spin_lock_irq(&conf->device_lock);
5760 conf->previous_raid_disks = conf->raid_disks;
5761 rdev_for_each(rdev, conf->mddev)
5762 rdev->data_offset = rdev->new_data_offset;
5764 conf->reshape_progress = MaxSector;
5765 spin_unlock_irq(&conf->device_lock);
5766 wake_up(&conf->wait_for_overlap);
5768 /* read-ahead size must cover two whole stripes, which is
5769 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5771 if (conf->mddev->queue) {
5772 int data_disks = conf->raid_disks - conf->max_degraded;
5773 int stripe = data_disks * ((conf->chunk_sectors << 9)
5775 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5776 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5781 /* This is called from the raid5d thread with mddev_lock held.
5782 * It makes config changes to the device.
5784 static void raid5_finish_reshape(struct mddev *mddev)
5786 struct r5conf *conf = mddev->private;
5788 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5790 if (mddev->delta_disks > 0) {
5791 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5792 set_capacity(mddev->gendisk, mddev->array_sectors);
5793 revalidate_disk(mddev->gendisk);
5796 spin_lock_irq(&conf->device_lock);
5797 mddev->degraded = calc_degraded(conf);
5798 spin_unlock_irq(&conf->device_lock);
5799 for (d = conf->raid_disks ;
5800 d < conf->raid_disks - mddev->delta_disks;
5802 struct md_rdev *rdev = conf->disks[d].rdev;
5804 clear_bit(In_sync, &rdev->flags);
5805 rdev = conf->disks[d].replacement;
5807 clear_bit(In_sync, &rdev->flags);
5810 mddev->layout = conf->algorithm;
5811 mddev->chunk_sectors = conf->chunk_sectors;
5812 mddev->reshape_position = MaxSector;
5813 mddev->delta_disks = 0;
5814 mddev->reshape_backwards = 0;
5818 static void raid5_quiesce(struct mddev *mddev, int state)
5820 struct r5conf *conf = mddev->private;
5823 case 2: /* resume for a suspend */
5824 wake_up(&conf->wait_for_overlap);
5827 case 1: /* stop all writes */
5828 spin_lock_irq(&conf->device_lock);
5829 /* '2' tells resync/reshape to pause so that all
5830 * active stripes can drain
5833 wait_event_lock_irq(conf->wait_for_stripe,
5834 atomic_read(&conf->active_stripes) == 0 &&
5835 atomic_read(&conf->active_aligned_reads) == 0,
5836 conf->device_lock, /* nothing */);
5838 spin_unlock_irq(&conf->device_lock);
5839 /* allow reshape to continue */
5840 wake_up(&conf->wait_for_overlap);
5843 case 0: /* re-enable writes */
5844 spin_lock_irq(&conf->device_lock);
5846 wake_up(&conf->wait_for_stripe);
5847 wake_up(&conf->wait_for_overlap);
5848 spin_unlock_irq(&conf->device_lock);
5854 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5856 struct r0conf *raid0_conf = mddev->private;
5859 /* for raid0 takeover only one zone is supported */
5860 if (raid0_conf->nr_strip_zones > 1) {
5861 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5863 return ERR_PTR(-EINVAL);
5866 sectors = raid0_conf->strip_zone[0].zone_end;
5867 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5868 mddev->dev_sectors = sectors;
5869 mddev->new_level = level;
5870 mddev->new_layout = ALGORITHM_PARITY_N;
5871 mddev->new_chunk_sectors = mddev->chunk_sectors;
5872 mddev->raid_disks += 1;
5873 mddev->delta_disks = 1;
5874 /* make sure it will be not marked as dirty */
5875 mddev->recovery_cp = MaxSector;
5877 return setup_conf(mddev);
5881 static void *raid5_takeover_raid1(struct mddev *mddev)
5885 if (mddev->raid_disks != 2 ||
5886 mddev->degraded > 1)
5887 return ERR_PTR(-EINVAL);
5889 /* Should check if there are write-behind devices? */
5891 chunksect = 64*2; /* 64K by default */
5893 /* The array must be an exact multiple of chunksize */
5894 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5897 if ((chunksect<<9) < STRIPE_SIZE)
5898 /* array size does not allow a suitable chunk size */
5899 return ERR_PTR(-EINVAL);
5901 mddev->new_level = 5;
5902 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5903 mddev->new_chunk_sectors = chunksect;
5905 return setup_conf(mddev);
5908 static void *raid5_takeover_raid6(struct mddev *mddev)
5912 switch (mddev->layout) {
5913 case ALGORITHM_LEFT_ASYMMETRIC_6:
5914 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5916 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5917 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5919 case ALGORITHM_LEFT_SYMMETRIC_6:
5920 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5922 case ALGORITHM_RIGHT_SYMMETRIC_6:
5923 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5925 case ALGORITHM_PARITY_0_6:
5926 new_layout = ALGORITHM_PARITY_0;
5928 case ALGORITHM_PARITY_N:
5929 new_layout = ALGORITHM_PARITY_N;
5932 return ERR_PTR(-EINVAL);
5934 mddev->new_level = 5;
5935 mddev->new_layout = new_layout;
5936 mddev->delta_disks = -1;
5937 mddev->raid_disks -= 1;
5938 return setup_conf(mddev);
5942 static int raid5_check_reshape(struct mddev *mddev)
5944 /* For a 2-drive array, the layout and chunk size can be changed
5945 * immediately as not restriping is needed.
5946 * For larger arrays we record the new value - after validation
5947 * to be used by a reshape pass.
5949 struct r5conf *conf = mddev->private;
5950 int new_chunk = mddev->new_chunk_sectors;
5952 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5954 if (new_chunk > 0) {
5955 if (!is_power_of_2(new_chunk))
5957 if (new_chunk < (PAGE_SIZE>>9))
5959 if (mddev->array_sectors & (new_chunk-1))
5960 /* not factor of array size */
5964 /* They look valid */
5966 if (mddev->raid_disks == 2) {
5967 /* can make the change immediately */
5968 if (mddev->new_layout >= 0) {
5969 conf->algorithm = mddev->new_layout;
5970 mddev->layout = mddev->new_layout;
5972 if (new_chunk > 0) {
5973 conf->chunk_sectors = new_chunk ;
5974 mddev->chunk_sectors = new_chunk;
5976 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5977 md_wakeup_thread(mddev->thread);
5979 return check_reshape(mddev);
5982 static int raid6_check_reshape(struct mddev *mddev)
5984 int new_chunk = mddev->new_chunk_sectors;
5986 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5988 if (new_chunk > 0) {
5989 if (!is_power_of_2(new_chunk))
5991 if (new_chunk < (PAGE_SIZE >> 9))
5993 if (mddev->array_sectors & (new_chunk-1))
5994 /* not factor of array size */
5998 /* They look valid */
5999 return check_reshape(mddev);
6002 static void *raid5_takeover(struct mddev *mddev)
6004 /* raid5 can take over:
6005 * raid0 - if there is only one strip zone - make it a raid4 layout
6006 * raid1 - if there are two drives. We need to know the chunk size
6007 * raid4 - trivial - just use a raid4 layout.
6008 * raid6 - Providing it is a *_6 layout
6010 if (mddev->level == 0)
6011 return raid45_takeover_raid0(mddev, 5);
6012 if (mddev->level == 1)
6013 return raid5_takeover_raid1(mddev);
6014 if (mddev->level == 4) {
6015 mddev->new_layout = ALGORITHM_PARITY_N;
6016 mddev->new_level = 5;
6017 return setup_conf(mddev);
6019 if (mddev->level == 6)
6020 return raid5_takeover_raid6(mddev);
6022 return ERR_PTR(-EINVAL);
6025 static void *raid4_takeover(struct mddev *mddev)
6027 /* raid4 can take over:
6028 * raid0 - if there is only one strip zone
6029 * raid5 - if layout is right
6031 if (mddev->level == 0)
6032 return raid45_takeover_raid0(mddev, 4);
6033 if (mddev->level == 5 &&
6034 mddev->layout == ALGORITHM_PARITY_N) {
6035 mddev->new_layout = 0;
6036 mddev->new_level = 4;
6037 return setup_conf(mddev);
6039 return ERR_PTR(-EINVAL);
6042 static struct md_personality raid5_personality;
6044 static void *raid6_takeover(struct mddev *mddev)
6046 /* Currently can only take over a raid5. We map the
6047 * personality to an equivalent raid6 personality
6048 * with the Q block at the end.
6052 if (mddev->pers != &raid5_personality)
6053 return ERR_PTR(-EINVAL);
6054 if (mddev->degraded > 1)
6055 return ERR_PTR(-EINVAL);
6056 if (mddev->raid_disks > 253)
6057 return ERR_PTR(-EINVAL);
6058 if (mddev->raid_disks < 3)
6059 return ERR_PTR(-EINVAL);
6061 switch (mddev->layout) {
6062 case ALGORITHM_LEFT_ASYMMETRIC:
6063 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6065 case ALGORITHM_RIGHT_ASYMMETRIC:
6066 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6068 case ALGORITHM_LEFT_SYMMETRIC:
6069 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6071 case ALGORITHM_RIGHT_SYMMETRIC:
6072 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6074 case ALGORITHM_PARITY_0:
6075 new_layout = ALGORITHM_PARITY_0_6;
6077 case ALGORITHM_PARITY_N:
6078 new_layout = ALGORITHM_PARITY_N;
6081 return ERR_PTR(-EINVAL);
6083 mddev->new_level = 6;
6084 mddev->new_layout = new_layout;
6085 mddev->delta_disks = 1;
6086 mddev->raid_disks += 1;
6087 return setup_conf(mddev);
6091 static struct md_personality raid6_personality =
6095 .owner = THIS_MODULE,
6096 .make_request = make_request,
6100 .error_handler = error,
6101 .hot_add_disk = raid5_add_disk,
6102 .hot_remove_disk= raid5_remove_disk,
6103 .spare_active = raid5_spare_active,
6104 .sync_request = sync_request,
6105 .resize = raid5_resize,
6107 .check_reshape = raid6_check_reshape,
6108 .start_reshape = raid5_start_reshape,
6109 .finish_reshape = raid5_finish_reshape,
6110 .quiesce = raid5_quiesce,
6111 .takeover = raid6_takeover,
6113 static struct md_personality raid5_personality =
6117 .owner = THIS_MODULE,
6118 .make_request = make_request,
6122 .error_handler = error,
6123 .hot_add_disk = raid5_add_disk,
6124 .hot_remove_disk= raid5_remove_disk,
6125 .spare_active = raid5_spare_active,
6126 .sync_request = sync_request,
6127 .resize = raid5_resize,
6129 .check_reshape = raid5_check_reshape,
6130 .start_reshape = raid5_start_reshape,
6131 .finish_reshape = raid5_finish_reshape,
6132 .quiesce = raid5_quiesce,
6133 .takeover = raid5_takeover,
6136 static struct md_personality raid4_personality =
6140 .owner = THIS_MODULE,
6141 .make_request = make_request,
6145 .error_handler = error,
6146 .hot_add_disk = raid5_add_disk,
6147 .hot_remove_disk= raid5_remove_disk,
6148 .spare_active = raid5_spare_active,
6149 .sync_request = sync_request,
6150 .resize = raid5_resize,
6152 .check_reshape = raid5_check_reshape,
6153 .start_reshape = raid5_start_reshape,
6154 .finish_reshape = raid5_finish_reshape,
6155 .quiesce = raid5_quiesce,
6156 .takeover = raid4_takeover,
6159 static int __init raid5_init(void)
6161 register_md_personality(&raid6_personality);
6162 register_md_personality(&raid5_personality);
6163 register_md_personality(&raid4_personality);
6167 static void raid5_exit(void)
6169 unregister_md_personality(&raid6_personality);
6170 unregister_md_personality(&raid5_personality);
6171 unregister_md_personality(&raid4_personality);
6174 module_init(raid5_init);
6175 module_exit(raid5_exit);
6176 MODULE_LICENSE("GPL");
6177 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6178 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6179 MODULE_ALIAS("md-raid5");
6180 MODULE_ALIAS("md-raid4");
6181 MODULE_ALIAS("md-level-5");
6182 MODULE_ALIAS("md-level-4");
6183 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6184 MODULE_ALIAS("md-raid6");
6185 MODULE_ALIAS("md-level-6");
6187 /* This used to be two separate modules, they were: */
6188 MODULE_ALIAS("raid5");
6189 MODULE_ALIAS("raid6");