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/async.h>
51 #include <linux/seq_file.h>
52 #include <linux/cpu.h>
53 #include <linux/slab.h>
54 #include <linux/ratelimit.h>
64 #define NR_STRIPES 256
65 #define STRIPE_SIZE PAGE_SIZE
66 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
67 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
68 #define IO_THRESHOLD 1
69 #define BYPASS_THRESHOLD 1
70 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
71 #define HASH_MASK (NR_HASH - 1)
73 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
75 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
76 * order without overlap. There may be several bio's per stripe+device, and
77 * a bio could span several devices.
78 * When walking this list for a particular stripe+device, we must never proceed
79 * beyond a bio that extends past this device, as the next bio might no longer
81 * This macro is used to determine the 'next' bio in the list, given the sector
82 * of the current stripe+device
84 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
86 * The following can be used to debug the driver
88 #define RAID5_PARANOIA 1
89 #if RAID5_PARANOIA && defined(CONFIG_SMP)
90 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
92 # define CHECK_DEVLOCK()
101 * We maintain a biased count of active stripes in the bottom 16 bits of
102 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
104 static inline int raid5_bi_phys_segments(struct bio *bio)
106 return bio->bi_phys_segments & 0xffff;
109 static inline int raid5_bi_hw_segments(struct bio *bio)
111 return (bio->bi_phys_segments >> 16) & 0xffff;
114 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
116 --bio->bi_phys_segments;
117 return raid5_bi_phys_segments(bio);
120 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
122 unsigned short val = raid5_bi_hw_segments(bio);
125 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
129 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
131 bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
134 /* Find first data disk in a raid6 stripe */
135 static inline int raid6_d0(struct stripe_head *sh)
138 /* ddf always start from first device */
140 /* md starts just after Q block */
141 if (sh->qd_idx == sh->disks - 1)
144 return sh->qd_idx + 1;
146 static inline int raid6_next_disk(int disk, int raid_disks)
149 return (disk < raid_disks) ? disk : 0;
152 /* When walking through the disks in a raid5, starting at raid6_d0,
153 * We need to map each disk to a 'slot', where the data disks are slot
154 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
155 * is raid_disks-1. This help does that mapping.
157 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
158 int *count, int syndrome_disks)
164 if (idx == sh->pd_idx)
165 return syndrome_disks;
166 if (idx == sh->qd_idx)
167 return syndrome_disks + 1;
173 static void return_io(struct bio *return_bi)
175 struct bio *bi = return_bi;
178 return_bi = bi->bi_next;
186 static void print_raid5_conf (raid5_conf_t *conf);
188 static int stripe_operations_active(struct stripe_head *sh)
190 return sh->check_state || sh->reconstruct_state ||
191 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
192 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
195 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
197 if (atomic_dec_and_test(&sh->count)) {
198 BUG_ON(!list_empty(&sh->lru));
199 BUG_ON(atomic_read(&conf->active_stripes)==0);
200 if (test_bit(STRIPE_HANDLE, &sh->state)) {
201 if (test_bit(STRIPE_DELAYED, &sh->state))
202 list_add_tail(&sh->lru, &conf->delayed_list);
203 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
204 sh->bm_seq - conf->seq_write > 0)
205 list_add_tail(&sh->lru, &conf->bitmap_list);
207 clear_bit(STRIPE_BIT_DELAY, &sh->state);
208 list_add_tail(&sh->lru, &conf->handle_list);
210 md_wakeup_thread(conf->mddev->thread);
212 BUG_ON(stripe_operations_active(sh));
213 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
214 atomic_dec(&conf->preread_active_stripes);
215 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
216 md_wakeup_thread(conf->mddev->thread);
218 atomic_dec(&conf->active_stripes);
219 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
220 list_add_tail(&sh->lru, &conf->inactive_list);
221 wake_up(&conf->wait_for_stripe);
222 if (conf->retry_read_aligned)
223 md_wakeup_thread(conf->mddev->thread);
229 static void release_stripe(struct stripe_head *sh)
231 raid5_conf_t *conf = sh->raid_conf;
234 spin_lock_irqsave(&conf->device_lock, flags);
235 __release_stripe(conf, sh);
236 spin_unlock_irqrestore(&conf->device_lock, flags);
239 static inline void remove_hash(struct stripe_head *sh)
241 pr_debug("remove_hash(), stripe %llu\n",
242 (unsigned long long)sh->sector);
244 hlist_del_init(&sh->hash);
247 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
249 struct hlist_head *hp = stripe_hash(conf, sh->sector);
251 pr_debug("insert_hash(), stripe %llu\n",
252 (unsigned long long)sh->sector);
255 hlist_add_head(&sh->hash, hp);
259 /* find an idle stripe, make sure it is unhashed, and return it. */
260 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
262 struct stripe_head *sh = NULL;
263 struct list_head *first;
266 if (list_empty(&conf->inactive_list))
268 first = conf->inactive_list.next;
269 sh = list_entry(first, struct stripe_head, lru);
270 list_del_init(first);
272 atomic_inc(&conf->active_stripes);
277 static void shrink_buffers(struct stripe_head *sh)
281 int num = sh->raid_conf->pool_size;
283 for (i = 0; i < num ; i++) {
287 sh->dev[i].page = NULL;
292 static int grow_buffers(struct stripe_head *sh)
295 int num = sh->raid_conf->pool_size;
297 for (i = 0; i < num; i++) {
300 if (!(page = alloc_page(GFP_KERNEL))) {
303 sh->dev[i].page = page;
308 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
309 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
310 struct stripe_head *sh);
312 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
314 raid5_conf_t *conf = sh->raid_conf;
317 BUG_ON(atomic_read(&sh->count) != 0);
318 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
319 BUG_ON(stripe_operations_active(sh));
322 pr_debug("init_stripe called, stripe %llu\n",
323 (unsigned long long)sh->sector);
327 sh->generation = conf->generation - previous;
328 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
330 stripe_set_idx(sector, conf, previous, sh);
334 for (i = sh->disks; i--; ) {
335 struct r5dev *dev = &sh->dev[i];
337 if (dev->toread || dev->read || dev->towrite || dev->written ||
338 test_bit(R5_LOCKED, &dev->flags)) {
339 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
340 (unsigned long long)sh->sector, i, dev->toread,
341 dev->read, dev->towrite, dev->written,
342 test_bit(R5_LOCKED, &dev->flags));
346 raid5_build_block(sh, i, previous);
348 insert_hash(conf, sh);
351 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
354 struct stripe_head *sh;
355 struct hlist_node *hn;
358 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
359 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
360 if (sh->sector == sector && sh->generation == generation)
362 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
367 * Need to check if array has failed when deciding whether to:
369 * - remove non-faulty devices
372 * This determination is simple when no reshape is happening.
373 * However if there is a reshape, we need to carefully check
374 * both the before and after sections.
375 * This is because some failed devices may only affect one
376 * of the two sections, and some non-in_sync devices may
377 * be insync in the section most affected by failed devices.
379 static int has_failed(raid5_conf_t *conf)
383 if (conf->mddev->reshape_position == MaxSector)
384 return conf->mddev->degraded > conf->max_degraded;
388 for (i = 0; i < conf->previous_raid_disks; i++) {
389 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
390 if (!rdev || test_bit(Faulty, &rdev->flags))
392 else if (test_bit(In_sync, &rdev->flags))
395 /* not in-sync or faulty.
396 * If the reshape increases the number of devices,
397 * this is being recovered by the reshape, so
398 * this 'previous' section is not in_sync.
399 * If the number of devices is being reduced however,
400 * the device can only be part of the array if
401 * we are reverting a reshape, so this section will
404 if (conf->raid_disks >= conf->previous_raid_disks)
408 if (degraded > conf->max_degraded)
412 for (i = 0; i < conf->raid_disks; i++) {
413 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
414 if (!rdev || test_bit(Faulty, &rdev->flags))
416 else if (test_bit(In_sync, &rdev->flags))
419 /* not in-sync or faulty.
420 * If reshape increases the number of devices, this
421 * section has already been recovered, else it
422 * almost certainly hasn't.
424 if (conf->raid_disks <= conf->previous_raid_disks)
428 if (degraded > conf->max_degraded)
433 static struct stripe_head *
434 get_active_stripe(raid5_conf_t *conf, sector_t sector,
435 int previous, int noblock, int noquiesce)
437 struct stripe_head *sh;
439 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
441 spin_lock_irq(&conf->device_lock);
444 wait_event_lock_irq(conf->wait_for_stripe,
445 conf->quiesce == 0 || noquiesce,
446 conf->device_lock, /* nothing */);
447 sh = __find_stripe(conf, sector, conf->generation - previous);
449 if (!conf->inactive_blocked)
450 sh = get_free_stripe(conf);
451 if (noblock && sh == NULL)
454 conf->inactive_blocked = 1;
455 wait_event_lock_irq(conf->wait_for_stripe,
456 !list_empty(&conf->inactive_list) &&
457 (atomic_read(&conf->active_stripes)
458 < (conf->max_nr_stripes *3/4)
459 || !conf->inactive_blocked),
462 conf->inactive_blocked = 0;
464 init_stripe(sh, sector, previous);
466 if (atomic_read(&sh->count)) {
467 BUG_ON(!list_empty(&sh->lru)
468 && !test_bit(STRIPE_EXPANDING, &sh->state));
470 if (!test_bit(STRIPE_HANDLE, &sh->state))
471 atomic_inc(&conf->active_stripes);
472 if (list_empty(&sh->lru) &&
473 !test_bit(STRIPE_EXPANDING, &sh->state))
475 list_del_init(&sh->lru);
478 } while (sh == NULL);
481 atomic_inc(&sh->count);
483 spin_unlock_irq(&conf->device_lock);
488 raid5_end_read_request(struct bio *bi, int error);
490 raid5_end_write_request(struct bio *bi, int error);
492 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
494 raid5_conf_t *conf = sh->raid_conf;
495 int i, disks = sh->disks;
499 for (i = disks; i--; ) {
503 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
504 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
508 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
513 bi = &sh->dev[i].req;
517 bi->bi_end_io = raid5_end_write_request;
519 bi->bi_end_io = raid5_end_read_request;
522 rdev = rcu_dereference(conf->disks[i].rdev);
523 if (rdev && test_bit(Faulty, &rdev->flags))
526 atomic_inc(&rdev->nr_pending);
530 if (s->syncing || s->expanding || s->expanded)
531 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
533 set_bit(STRIPE_IO_STARTED, &sh->state);
535 bi->bi_bdev = rdev->bdev;
536 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
537 __func__, (unsigned long long)sh->sector,
539 atomic_inc(&sh->count);
540 bi->bi_sector = sh->sector + rdev->data_offset;
541 bi->bi_flags = 1 << BIO_UPTODATE;
545 bi->bi_io_vec = &sh->dev[i].vec;
546 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
547 bi->bi_io_vec[0].bv_offset = 0;
548 bi->bi_size = STRIPE_SIZE;
551 test_bit(R5_ReWrite, &sh->dev[i].flags))
552 atomic_add(STRIPE_SECTORS,
553 &rdev->corrected_errors);
554 generic_make_request(bi);
557 set_bit(STRIPE_DEGRADED, &sh->state);
558 pr_debug("skip op %ld on disc %d for sector %llu\n",
559 bi->bi_rw, i, (unsigned long long)sh->sector);
560 clear_bit(R5_LOCKED, &sh->dev[i].flags);
561 set_bit(STRIPE_HANDLE, &sh->state);
566 static struct dma_async_tx_descriptor *
567 async_copy_data(int frombio, struct bio *bio, struct page *page,
568 sector_t sector, struct dma_async_tx_descriptor *tx)
571 struct page *bio_page;
574 struct async_submit_ctl submit;
575 enum async_tx_flags flags = 0;
577 if (bio->bi_sector >= sector)
578 page_offset = (signed)(bio->bi_sector - sector) * 512;
580 page_offset = (signed)(sector - bio->bi_sector) * -512;
583 flags |= ASYNC_TX_FENCE;
584 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
586 bio_for_each_segment(bvl, bio, i) {
587 int len = bvl->bv_len;
591 if (page_offset < 0) {
592 b_offset = -page_offset;
593 page_offset += b_offset;
597 if (len > 0 && page_offset + len > STRIPE_SIZE)
598 clen = STRIPE_SIZE - page_offset;
603 b_offset += bvl->bv_offset;
604 bio_page = bvl->bv_page;
606 tx = async_memcpy(page, bio_page, page_offset,
607 b_offset, clen, &submit);
609 tx = async_memcpy(bio_page, page, b_offset,
610 page_offset, clen, &submit);
612 /* chain the operations */
613 submit.depend_tx = tx;
615 if (clen < len) /* hit end of page */
623 static void ops_complete_biofill(void *stripe_head_ref)
625 struct stripe_head *sh = stripe_head_ref;
626 struct bio *return_bi = NULL;
627 raid5_conf_t *conf = sh->raid_conf;
630 pr_debug("%s: stripe %llu\n", __func__,
631 (unsigned long long)sh->sector);
633 /* clear completed biofills */
634 spin_lock_irq(&conf->device_lock);
635 for (i = sh->disks; i--; ) {
636 struct r5dev *dev = &sh->dev[i];
638 /* acknowledge completion of a biofill operation */
639 /* and check if we need to reply to a read request,
640 * new R5_Wantfill requests are held off until
641 * !STRIPE_BIOFILL_RUN
643 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
644 struct bio *rbi, *rbi2;
649 while (rbi && rbi->bi_sector <
650 dev->sector + STRIPE_SECTORS) {
651 rbi2 = r5_next_bio(rbi, dev->sector);
652 if (!raid5_dec_bi_phys_segments(rbi)) {
653 rbi->bi_next = return_bi;
660 spin_unlock_irq(&conf->device_lock);
661 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
663 return_io(return_bi);
665 set_bit(STRIPE_HANDLE, &sh->state);
669 static void ops_run_biofill(struct stripe_head *sh)
671 struct dma_async_tx_descriptor *tx = NULL;
672 raid5_conf_t *conf = sh->raid_conf;
673 struct async_submit_ctl submit;
676 pr_debug("%s: stripe %llu\n", __func__,
677 (unsigned long long)sh->sector);
679 for (i = sh->disks; i--; ) {
680 struct r5dev *dev = &sh->dev[i];
681 if (test_bit(R5_Wantfill, &dev->flags)) {
683 spin_lock_irq(&conf->device_lock);
684 dev->read = rbi = dev->toread;
686 spin_unlock_irq(&conf->device_lock);
687 while (rbi && rbi->bi_sector <
688 dev->sector + STRIPE_SECTORS) {
689 tx = async_copy_data(0, rbi, dev->page,
691 rbi = r5_next_bio(rbi, dev->sector);
696 atomic_inc(&sh->count);
697 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
698 async_trigger_callback(&submit);
701 static void mark_target_uptodate(struct stripe_head *sh, int target)
708 tgt = &sh->dev[target];
709 set_bit(R5_UPTODATE, &tgt->flags);
710 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
711 clear_bit(R5_Wantcompute, &tgt->flags);
714 static void ops_complete_compute(void *stripe_head_ref)
716 struct stripe_head *sh = stripe_head_ref;
718 pr_debug("%s: stripe %llu\n", __func__,
719 (unsigned long long)sh->sector);
721 /* mark the computed target(s) as uptodate */
722 mark_target_uptodate(sh, sh->ops.target);
723 mark_target_uptodate(sh, sh->ops.target2);
725 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
726 if (sh->check_state == check_state_compute_run)
727 sh->check_state = check_state_compute_result;
728 set_bit(STRIPE_HANDLE, &sh->state);
732 /* return a pointer to the address conversion region of the scribble buffer */
733 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
734 struct raid5_percpu *percpu)
736 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
739 static struct dma_async_tx_descriptor *
740 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
742 int disks = sh->disks;
743 struct page **xor_srcs = percpu->scribble;
744 int target = sh->ops.target;
745 struct r5dev *tgt = &sh->dev[target];
746 struct page *xor_dest = tgt->page;
748 struct dma_async_tx_descriptor *tx;
749 struct async_submit_ctl submit;
752 pr_debug("%s: stripe %llu block: %d\n",
753 __func__, (unsigned long long)sh->sector, target);
754 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
756 for (i = disks; i--; )
758 xor_srcs[count++] = sh->dev[i].page;
760 atomic_inc(&sh->count);
762 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
763 ops_complete_compute, sh, to_addr_conv(sh, percpu));
764 if (unlikely(count == 1))
765 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
767 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
772 /* set_syndrome_sources - populate source buffers for gen_syndrome
773 * @srcs - (struct page *) array of size sh->disks
774 * @sh - stripe_head to parse
776 * Populates srcs in proper layout order for the stripe and returns the
777 * 'count' of sources to be used in a call to async_gen_syndrome. The P
778 * destination buffer is recorded in srcs[count] and the Q destination
779 * is recorded in srcs[count+1]].
781 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
783 int disks = sh->disks;
784 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
785 int d0_idx = raid6_d0(sh);
789 for (i = 0; i < disks; i++)
795 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
797 srcs[slot] = sh->dev[i].page;
798 i = raid6_next_disk(i, disks);
799 } while (i != d0_idx);
801 return syndrome_disks;
804 static struct dma_async_tx_descriptor *
805 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
807 int disks = sh->disks;
808 struct page **blocks = percpu->scribble;
810 int qd_idx = sh->qd_idx;
811 struct dma_async_tx_descriptor *tx;
812 struct async_submit_ctl submit;
818 if (sh->ops.target < 0)
819 target = sh->ops.target2;
820 else if (sh->ops.target2 < 0)
821 target = sh->ops.target;
823 /* we should only have one valid target */
826 pr_debug("%s: stripe %llu block: %d\n",
827 __func__, (unsigned long long)sh->sector, target);
829 tgt = &sh->dev[target];
830 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
833 atomic_inc(&sh->count);
835 if (target == qd_idx) {
836 count = set_syndrome_sources(blocks, sh);
837 blocks[count] = NULL; /* regenerating p is not necessary */
838 BUG_ON(blocks[count+1] != dest); /* q should already be set */
839 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
840 ops_complete_compute, sh,
841 to_addr_conv(sh, percpu));
842 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
844 /* Compute any data- or p-drive using XOR */
846 for (i = disks; i-- ; ) {
847 if (i == target || i == qd_idx)
849 blocks[count++] = sh->dev[i].page;
852 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
853 NULL, ops_complete_compute, sh,
854 to_addr_conv(sh, percpu));
855 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
861 static struct dma_async_tx_descriptor *
862 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
864 int i, count, disks = sh->disks;
865 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
866 int d0_idx = raid6_d0(sh);
867 int faila = -1, failb = -1;
868 int target = sh->ops.target;
869 int target2 = sh->ops.target2;
870 struct r5dev *tgt = &sh->dev[target];
871 struct r5dev *tgt2 = &sh->dev[target2];
872 struct dma_async_tx_descriptor *tx;
873 struct page **blocks = percpu->scribble;
874 struct async_submit_ctl submit;
876 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
877 __func__, (unsigned long long)sh->sector, target, target2);
878 BUG_ON(target < 0 || target2 < 0);
879 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
880 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
882 /* we need to open-code set_syndrome_sources to handle the
883 * slot number conversion for 'faila' and 'failb'
885 for (i = 0; i < disks ; i++)
890 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
892 blocks[slot] = sh->dev[i].page;
898 i = raid6_next_disk(i, disks);
899 } while (i != d0_idx);
901 BUG_ON(faila == failb);
904 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
905 __func__, (unsigned long long)sh->sector, faila, failb);
907 atomic_inc(&sh->count);
909 if (failb == syndrome_disks+1) {
910 /* Q disk is one of the missing disks */
911 if (faila == syndrome_disks) {
912 /* Missing P+Q, just recompute */
913 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
914 ops_complete_compute, sh,
915 to_addr_conv(sh, percpu));
916 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
917 STRIPE_SIZE, &submit);
921 int qd_idx = sh->qd_idx;
923 /* Missing D+Q: recompute D from P, then recompute Q */
924 if (target == qd_idx)
925 data_target = target2;
927 data_target = target;
930 for (i = disks; i-- ; ) {
931 if (i == data_target || i == qd_idx)
933 blocks[count++] = sh->dev[i].page;
935 dest = sh->dev[data_target].page;
936 init_async_submit(&submit,
937 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
939 to_addr_conv(sh, percpu));
940 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
943 count = set_syndrome_sources(blocks, sh);
944 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
945 ops_complete_compute, sh,
946 to_addr_conv(sh, percpu));
947 return async_gen_syndrome(blocks, 0, count+2,
948 STRIPE_SIZE, &submit);
951 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
952 ops_complete_compute, sh,
953 to_addr_conv(sh, percpu));
954 if (failb == syndrome_disks) {
955 /* We're missing D+P. */
956 return async_raid6_datap_recov(syndrome_disks+2,
960 /* We're missing D+D. */
961 return async_raid6_2data_recov(syndrome_disks+2,
962 STRIPE_SIZE, faila, failb,
969 static void ops_complete_prexor(void *stripe_head_ref)
971 struct stripe_head *sh = stripe_head_ref;
973 pr_debug("%s: stripe %llu\n", __func__,
974 (unsigned long long)sh->sector);
977 static struct dma_async_tx_descriptor *
978 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
979 struct dma_async_tx_descriptor *tx)
981 int disks = sh->disks;
982 struct page **xor_srcs = percpu->scribble;
983 int count = 0, pd_idx = sh->pd_idx, i;
984 struct async_submit_ctl submit;
986 /* existing parity data subtracted */
987 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
989 pr_debug("%s: stripe %llu\n", __func__,
990 (unsigned long long)sh->sector);
992 for (i = disks; i--; ) {
993 struct r5dev *dev = &sh->dev[i];
994 /* Only process blocks that are known to be uptodate */
995 if (test_bit(R5_Wantdrain, &dev->flags))
996 xor_srcs[count++] = dev->page;
999 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1000 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1001 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1006 static struct dma_async_tx_descriptor *
1007 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1009 int disks = sh->disks;
1012 pr_debug("%s: stripe %llu\n", __func__,
1013 (unsigned long long)sh->sector);
1015 for (i = disks; i--; ) {
1016 struct r5dev *dev = &sh->dev[i];
1019 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1022 spin_lock_irq(&sh->raid_conf->device_lock);
1023 chosen = dev->towrite;
1024 dev->towrite = NULL;
1025 BUG_ON(dev->written);
1026 wbi = dev->written = chosen;
1027 spin_unlock_irq(&sh->raid_conf->device_lock);
1029 while (wbi && wbi->bi_sector <
1030 dev->sector + STRIPE_SECTORS) {
1031 if (wbi->bi_rw & REQ_FUA)
1032 set_bit(R5_WantFUA, &dev->flags);
1033 tx = async_copy_data(1, wbi, dev->page,
1035 wbi = r5_next_bio(wbi, dev->sector);
1043 static void ops_complete_reconstruct(void *stripe_head_ref)
1045 struct stripe_head *sh = stripe_head_ref;
1046 int disks = sh->disks;
1047 int pd_idx = sh->pd_idx;
1048 int qd_idx = sh->qd_idx;
1052 pr_debug("%s: stripe %llu\n", __func__,
1053 (unsigned long long)sh->sector);
1055 for (i = disks; i--; )
1056 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1058 for (i = disks; i--; ) {
1059 struct r5dev *dev = &sh->dev[i];
1061 if (dev->written || i == pd_idx || i == qd_idx) {
1062 set_bit(R5_UPTODATE, &dev->flags);
1064 set_bit(R5_WantFUA, &dev->flags);
1068 if (sh->reconstruct_state == reconstruct_state_drain_run)
1069 sh->reconstruct_state = reconstruct_state_drain_result;
1070 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1071 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1073 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1074 sh->reconstruct_state = reconstruct_state_result;
1077 set_bit(STRIPE_HANDLE, &sh->state);
1082 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1083 struct dma_async_tx_descriptor *tx)
1085 int disks = sh->disks;
1086 struct page **xor_srcs = percpu->scribble;
1087 struct async_submit_ctl submit;
1088 int count = 0, pd_idx = sh->pd_idx, i;
1089 struct page *xor_dest;
1091 unsigned long flags;
1093 pr_debug("%s: stripe %llu\n", __func__,
1094 (unsigned long long)sh->sector);
1096 /* check if prexor is active which means only process blocks
1097 * that are part of a read-modify-write (written)
1099 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1101 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1102 for (i = disks; i--; ) {
1103 struct r5dev *dev = &sh->dev[i];
1105 xor_srcs[count++] = dev->page;
1108 xor_dest = sh->dev[pd_idx].page;
1109 for (i = disks; i--; ) {
1110 struct r5dev *dev = &sh->dev[i];
1112 xor_srcs[count++] = dev->page;
1116 /* 1/ if we prexor'd then the dest is reused as a source
1117 * 2/ if we did not prexor then we are redoing the parity
1118 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1119 * for the synchronous xor case
1121 flags = ASYNC_TX_ACK |
1122 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1124 atomic_inc(&sh->count);
1126 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1127 to_addr_conv(sh, percpu));
1128 if (unlikely(count == 1))
1129 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1131 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1135 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1136 struct dma_async_tx_descriptor *tx)
1138 struct async_submit_ctl submit;
1139 struct page **blocks = percpu->scribble;
1142 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1144 count = set_syndrome_sources(blocks, sh);
1146 atomic_inc(&sh->count);
1148 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1149 sh, to_addr_conv(sh, percpu));
1150 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1153 static void ops_complete_check(void *stripe_head_ref)
1155 struct stripe_head *sh = stripe_head_ref;
1157 pr_debug("%s: stripe %llu\n", __func__,
1158 (unsigned long long)sh->sector);
1160 sh->check_state = check_state_check_result;
1161 set_bit(STRIPE_HANDLE, &sh->state);
1165 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1167 int disks = sh->disks;
1168 int pd_idx = sh->pd_idx;
1169 int qd_idx = sh->qd_idx;
1170 struct page *xor_dest;
1171 struct page **xor_srcs = percpu->scribble;
1172 struct dma_async_tx_descriptor *tx;
1173 struct async_submit_ctl submit;
1177 pr_debug("%s: stripe %llu\n", __func__,
1178 (unsigned long long)sh->sector);
1181 xor_dest = sh->dev[pd_idx].page;
1182 xor_srcs[count++] = xor_dest;
1183 for (i = disks; i--; ) {
1184 if (i == pd_idx || i == qd_idx)
1186 xor_srcs[count++] = sh->dev[i].page;
1189 init_async_submit(&submit, 0, NULL, NULL, NULL,
1190 to_addr_conv(sh, percpu));
1191 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1192 &sh->ops.zero_sum_result, &submit);
1194 atomic_inc(&sh->count);
1195 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1196 tx = async_trigger_callback(&submit);
1199 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1201 struct page **srcs = percpu->scribble;
1202 struct async_submit_ctl submit;
1205 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1206 (unsigned long long)sh->sector, checkp);
1208 count = set_syndrome_sources(srcs, sh);
1212 atomic_inc(&sh->count);
1213 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1214 sh, to_addr_conv(sh, percpu));
1215 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1216 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1219 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1221 int overlap_clear = 0, i, disks = sh->disks;
1222 struct dma_async_tx_descriptor *tx = NULL;
1223 raid5_conf_t *conf = sh->raid_conf;
1224 int level = conf->level;
1225 struct raid5_percpu *percpu;
1229 percpu = per_cpu_ptr(conf->percpu, cpu);
1230 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1231 ops_run_biofill(sh);
1235 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1237 tx = ops_run_compute5(sh, percpu);
1239 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1240 tx = ops_run_compute6_1(sh, percpu);
1242 tx = ops_run_compute6_2(sh, percpu);
1244 /* terminate the chain if reconstruct is not set to be run */
1245 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1249 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1250 tx = ops_run_prexor(sh, percpu, tx);
1252 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1253 tx = ops_run_biodrain(sh, tx);
1257 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1259 ops_run_reconstruct5(sh, percpu, tx);
1261 ops_run_reconstruct6(sh, percpu, tx);
1264 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1265 if (sh->check_state == check_state_run)
1266 ops_run_check_p(sh, percpu);
1267 else if (sh->check_state == check_state_run_q)
1268 ops_run_check_pq(sh, percpu, 0);
1269 else if (sh->check_state == check_state_run_pq)
1270 ops_run_check_pq(sh, percpu, 1);
1276 for (i = disks; i--; ) {
1277 struct r5dev *dev = &sh->dev[i];
1278 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1279 wake_up(&sh->raid_conf->wait_for_overlap);
1284 #ifdef CONFIG_MULTICORE_RAID456
1285 static void async_run_ops(void *param, async_cookie_t cookie)
1287 struct stripe_head *sh = param;
1288 unsigned long ops_request = sh->ops.request;
1290 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1291 wake_up(&sh->ops.wait_for_ops);
1293 __raid_run_ops(sh, ops_request);
1297 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1299 /* since handle_stripe can be called outside of raid5d context
1300 * we need to ensure sh->ops.request is de-staged before another
1303 wait_event(sh->ops.wait_for_ops,
1304 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1305 sh->ops.request = ops_request;
1307 atomic_inc(&sh->count);
1308 async_schedule(async_run_ops, sh);
1311 #define raid_run_ops __raid_run_ops
1314 static int grow_one_stripe(raid5_conf_t *conf)
1316 struct stripe_head *sh;
1317 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1321 sh->raid_conf = conf;
1322 #ifdef CONFIG_MULTICORE_RAID456
1323 init_waitqueue_head(&sh->ops.wait_for_ops);
1326 if (grow_buffers(sh)) {
1328 kmem_cache_free(conf->slab_cache, sh);
1331 /* we just created an active stripe so... */
1332 atomic_set(&sh->count, 1);
1333 atomic_inc(&conf->active_stripes);
1334 INIT_LIST_HEAD(&sh->lru);
1339 static int grow_stripes(raid5_conf_t *conf, int num)
1341 struct kmem_cache *sc;
1342 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1344 if (conf->mddev->gendisk)
1345 sprintf(conf->cache_name[0],
1346 "raid%d-%s", conf->level, mdname(conf->mddev));
1348 sprintf(conf->cache_name[0],
1349 "raid%d-%p", conf->level, conf->mddev);
1350 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1352 conf->active_name = 0;
1353 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1354 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1358 conf->slab_cache = sc;
1359 conf->pool_size = devs;
1361 if (!grow_one_stripe(conf))
1367 * scribble_len - return the required size of the scribble region
1368 * @num - total number of disks in the array
1370 * The size must be enough to contain:
1371 * 1/ a struct page pointer for each device in the array +2
1372 * 2/ room to convert each entry in (1) to its corresponding dma
1373 * (dma_map_page()) or page (page_address()) address.
1375 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1376 * calculate over all devices (not just the data blocks), using zeros in place
1377 * of the P and Q blocks.
1379 static size_t scribble_len(int num)
1383 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1388 static int resize_stripes(raid5_conf_t *conf, int newsize)
1390 /* Make all the stripes able to hold 'newsize' devices.
1391 * New slots in each stripe get 'page' set to a new page.
1393 * This happens in stages:
1394 * 1/ create a new kmem_cache and allocate the required number of
1396 * 2/ gather all the old stripe_heads and tranfer the pages across
1397 * to the new stripe_heads. This will have the side effect of
1398 * freezing the array as once all stripe_heads have been collected,
1399 * no IO will be possible. Old stripe heads are freed once their
1400 * pages have been transferred over, and the old kmem_cache is
1401 * freed when all stripes are done.
1402 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1403 * we simple return a failre status - no need to clean anything up.
1404 * 4/ allocate new pages for the new slots in the new stripe_heads.
1405 * If this fails, we don't bother trying the shrink the
1406 * stripe_heads down again, we just leave them as they are.
1407 * As each stripe_head is processed the new one is released into
1410 * Once step2 is started, we cannot afford to wait for a write,
1411 * so we use GFP_NOIO allocations.
1413 struct stripe_head *osh, *nsh;
1414 LIST_HEAD(newstripes);
1415 struct disk_info *ndisks;
1418 struct kmem_cache *sc;
1421 if (newsize <= conf->pool_size)
1422 return 0; /* never bother to shrink */
1424 err = md_allow_write(conf->mddev);
1429 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1430 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1435 for (i = conf->max_nr_stripes; i; i--) {
1436 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1440 nsh->raid_conf = conf;
1441 #ifdef CONFIG_MULTICORE_RAID456
1442 init_waitqueue_head(&nsh->ops.wait_for_ops);
1445 list_add(&nsh->lru, &newstripes);
1448 /* didn't get enough, give up */
1449 while (!list_empty(&newstripes)) {
1450 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1451 list_del(&nsh->lru);
1452 kmem_cache_free(sc, nsh);
1454 kmem_cache_destroy(sc);
1457 /* Step 2 - Must use GFP_NOIO now.
1458 * OK, we have enough stripes, start collecting inactive
1459 * stripes and copying them over
1461 list_for_each_entry(nsh, &newstripes, lru) {
1462 spin_lock_irq(&conf->device_lock);
1463 wait_event_lock_irq(conf->wait_for_stripe,
1464 !list_empty(&conf->inactive_list),
1467 osh = get_free_stripe(conf);
1468 spin_unlock_irq(&conf->device_lock);
1469 atomic_set(&nsh->count, 1);
1470 for(i=0; i<conf->pool_size; i++)
1471 nsh->dev[i].page = osh->dev[i].page;
1472 for( ; i<newsize; i++)
1473 nsh->dev[i].page = NULL;
1474 kmem_cache_free(conf->slab_cache, osh);
1476 kmem_cache_destroy(conf->slab_cache);
1479 * At this point, we are holding all the stripes so the array
1480 * is completely stalled, so now is a good time to resize
1481 * conf->disks and the scribble region
1483 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1485 for (i=0; i<conf->raid_disks; i++)
1486 ndisks[i] = conf->disks[i];
1488 conf->disks = ndisks;
1493 conf->scribble_len = scribble_len(newsize);
1494 for_each_present_cpu(cpu) {
1495 struct raid5_percpu *percpu;
1498 percpu = per_cpu_ptr(conf->percpu, cpu);
1499 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1502 kfree(percpu->scribble);
1503 percpu->scribble = scribble;
1511 /* Step 4, return new stripes to service */
1512 while(!list_empty(&newstripes)) {
1513 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1514 list_del_init(&nsh->lru);
1516 for (i=conf->raid_disks; i < newsize; i++)
1517 if (nsh->dev[i].page == NULL) {
1518 struct page *p = alloc_page(GFP_NOIO);
1519 nsh->dev[i].page = p;
1523 release_stripe(nsh);
1525 /* critical section pass, GFP_NOIO no longer needed */
1527 conf->slab_cache = sc;
1528 conf->active_name = 1-conf->active_name;
1529 conf->pool_size = newsize;
1533 static int drop_one_stripe(raid5_conf_t *conf)
1535 struct stripe_head *sh;
1537 spin_lock_irq(&conf->device_lock);
1538 sh = get_free_stripe(conf);
1539 spin_unlock_irq(&conf->device_lock);
1542 BUG_ON(atomic_read(&sh->count));
1544 kmem_cache_free(conf->slab_cache, sh);
1545 atomic_dec(&conf->active_stripes);
1549 static void shrink_stripes(raid5_conf_t *conf)
1551 while (drop_one_stripe(conf))
1554 if (conf->slab_cache)
1555 kmem_cache_destroy(conf->slab_cache);
1556 conf->slab_cache = NULL;
1559 static void raid5_end_read_request(struct bio * bi, int error)
1561 struct stripe_head *sh = bi->bi_private;
1562 raid5_conf_t *conf = sh->raid_conf;
1563 int disks = sh->disks, i;
1564 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1565 char b[BDEVNAME_SIZE];
1569 for (i=0 ; i<disks; i++)
1570 if (bi == &sh->dev[i].req)
1573 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1574 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1582 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1583 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1584 rdev = conf->disks[i].rdev;
1587 "md/raid:%s: read error corrected"
1588 " (%lu sectors at %llu on %s)\n",
1589 mdname(conf->mddev), STRIPE_SECTORS,
1590 (unsigned long long)(sh->sector
1591 + rdev->data_offset),
1592 bdevname(rdev->bdev, b));
1593 clear_bit(R5_ReadError, &sh->dev[i].flags);
1594 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1596 if (atomic_read(&conf->disks[i].rdev->read_errors))
1597 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1599 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1601 rdev = conf->disks[i].rdev;
1603 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1604 atomic_inc(&rdev->read_errors);
1605 if (conf->mddev->degraded >= conf->max_degraded)
1608 "md/raid:%s: read error not correctable "
1609 "(sector %llu on %s).\n",
1610 mdname(conf->mddev),
1611 (unsigned long long)(sh->sector
1612 + rdev->data_offset),
1614 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1618 "md/raid:%s: read error NOT corrected!! "
1619 "(sector %llu on %s).\n",
1620 mdname(conf->mddev),
1621 (unsigned long long)(sh->sector
1622 + rdev->data_offset),
1624 else if (atomic_read(&rdev->read_errors)
1625 > conf->max_nr_stripes)
1627 "md/raid:%s: Too many read errors, failing device %s.\n",
1628 mdname(conf->mddev), bdn);
1632 set_bit(R5_ReadError, &sh->dev[i].flags);
1634 clear_bit(R5_ReadError, &sh->dev[i].flags);
1635 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1636 md_error(conf->mddev, rdev);
1639 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1640 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1641 set_bit(STRIPE_HANDLE, &sh->state);
1645 static void raid5_end_write_request(struct bio *bi, int error)
1647 struct stripe_head *sh = bi->bi_private;
1648 raid5_conf_t *conf = sh->raid_conf;
1649 int disks = sh->disks, i;
1650 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1652 for (i=0 ; i<disks; i++)
1653 if (bi == &sh->dev[i].req)
1656 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1657 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1665 md_error(conf->mddev, conf->disks[i].rdev);
1667 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1669 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1670 set_bit(STRIPE_HANDLE, &sh->state);
1675 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1677 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1679 struct r5dev *dev = &sh->dev[i];
1681 bio_init(&dev->req);
1682 dev->req.bi_io_vec = &dev->vec;
1684 dev->req.bi_max_vecs++;
1685 dev->vec.bv_page = dev->page;
1686 dev->vec.bv_len = STRIPE_SIZE;
1687 dev->vec.bv_offset = 0;
1689 dev->req.bi_sector = sh->sector;
1690 dev->req.bi_private = sh;
1693 dev->sector = compute_blocknr(sh, i, previous);
1696 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1698 char b[BDEVNAME_SIZE];
1699 raid5_conf_t *conf = mddev->private;
1700 pr_debug("raid456: error called\n");
1702 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1703 unsigned long flags;
1704 spin_lock_irqsave(&conf->device_lock, flags);
1706 spin_unlock_irqrestore(&conf->device_lock, flags);
1708 * if recovery was running, make sure it aborts.
1710 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1712 set_bit(Faulty, &rdev->flags);
1713 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1715 "md/raid:%s: Disk failure on %s, disabling device.\n"
1716 "md/raid:%s: Operation continuing on %d devices.\n",
1718 bdevname(rdev->bdev, b),
1720 conf->raid_disks - mddev->degraded);
1724 * Input: a 'big' sector number,
1725 * Output: index of the data and parity disk, and the sector # in them.
1727 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1728 int previous, int *dd_idx,
1729 struct stripe_head *sh)
1731 sector_t stripe, stripe2;
1732 sector_t chunk_number;
1733 unsigned int chunk_offset;
1736 sector_t new_sector;
1737 int algorithm = previous ? conf->prev_algo
1739 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1740 : conf->chunk_sectors;
1741 int raid_disks = previous ? conf->previous_raid_disks
1743 int data_disks = raid_disks - conf->max_degraded;
1745 /* First compute the information on this sector */
1748 * Compute the chunk number and the sector offset inside the chunk
1750 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1751 chunk_number = r_sector;
1754 * Compute the stripe number
1756 stripe = chunk_number;
1757 *dd_idx = sector_div(stripe, data_disks);
1760 * Select the parity disk based on the user selected algorithm.
1762 pd_idx = qd_idx = -1;
1763 switch(conf->level) {
1765 pd_idx = data_disks;
1768 switch (algorithm) {
1769 case ALGORITHM_LEFT_ASYMMETRIC:
1770 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1771 if (*dd_idx >= pd_idx)
1774 case ALGORITHM_RIGHT_ASYMMETRIC:
1775 pd_idx = sector_div(stripe2, raid_disks);
1776 if (*dd_idx >= pd_idx)
1779 case ALGORITHM_LEFT_SYMMETRIC:
1780 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1781 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1783 case ALGORITHM_RIGHT_SYMMETRIC:
1784 pd_idx = sector_div(stripe2, raid_disks);
1785 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1787 case ALGORITHM_PARITY_0:
1791 case ALGORITHM_PARITY_N:
1792 pd_idx = data_disks;
1800 switch (algorithm) {
1801 case ALGORITHM_LEFT_ASYMMETRIC:
1802 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1803 qd_idx = pd_idx + 1;
1804 if (pd_idx == raid_disks-1) {
1805 (*dd_idx)++; /* Q D D D P */
1807 } else if (*dd_idx >= pd_idx)
1808 (*dd_idx) += 2; /* D D P Q D */
1810 case ALGORITHM_RIGHT_ASYMMETRIC:
1811 pd_idx = sector_div(stripe2, raid_disks);
1812 qd_idx = pd_idx + 1;
1813 if (pd_idx == raid_disks-1) {
1814 (*dd_idx)++; /* Q D D D P */
1816 } else if (*dd_idx >= pd_idx)
1817 (*dd_idx) += 2; /* D D P Q D */
1819 case ALGORITHM_LEFT_SYMMETRIC:
1820 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1821 qd_idx = (pd_idx + 1) % raid_disks;
1822 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1824 case ALGORITHM_RIGHT_SYMMETRIC:
1825 pd_idx = sector_div(stripe2, raid_disks);
1826 qd_idx = (pd_idx + 1) % raid_disks;
1827 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1830 case ALGORITHM_PARITY_0:
1835 case ALGORITHM_PARITY_N:
1836 pd_idx = data_disks;
1837 qd_idx = data_disks + 1;
1840 case ALGORITHM_ROTATING_ZERO_RESTART:
1841 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1842 * of blocks for computing Q is different.
1844 pd_idx = sector_div(stripe2, raid_disks);
1845 qd_idx = pd_idx + 1;
1846 if (pd_idx == raid_disks-1) {
1847 (*dd_idx)++; /* Q D D D P */
1849 } else if (*dd_idx >= pd_idx)
1850 (*dd_idx) += 2; /* D D P Q D */
1854 case ALGORITHM_ROTATING_N_RESTART:
1855 /* Same a left_asymmetric, by first stripe is
1856 * D D D P Q rather than
1860 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1861 qd_idx = pd_idx + 1;
1862 if (pd_idx == raid_disks-1) {
1863 (*dd_idx)++; /* Q D D D P */
1865 } else if (*dd_idx >= pd_idx)
1866 (*dd_idx) += 2; /* D D P Q D */
1870 case ALGORITHM_ROTATING_N_CONTINUE:
1871 /* Same as left_symmetric but Q is before P */
1872 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1873 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1874 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1878 case ALGORITHM_LEFT_ASYMMETRIC_6:
1879 /* RAID5 left_asymmetric, with Q on last device */
1880 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1881 if (*dd_idx >= pd_idx)
1883 qd_idx = raid_disks - 1;
1886 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1887 pd_idx = sector_div(stripe2, raid_disks-1);
1888 if (*dd_idx >= pd_idx)
1890 qd_idx = raid_disks - 1;
1893 case ALGORITHM_LEFT_SYMMETRIC_6:
1894 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1895 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1896 qd_idx = raid_disks - 1;
1899 case ALGORITHM_RIGHT_SYMMETRIC_6:
1900 pd_idx = sector_div(stripe2, raid_disks-1);
1901 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1902 qd_idx = raid_disks - 1;
1905 case ALGORITHM_PARITY_0_6:
1908 qd_idx = raid_disks - 1;
1918 sh->pd_idx = pd_idx;
1919 sh->qd_idx = qd_idx;
1920 sh->ddf_layout = ddf_layout;
1923 * Finally, compute the new sector number
1925 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1930 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1932 raid5_conf_t *conf = sh->raid_conf;
1933 int raid_disks = sh->disks;
1934 int data_disks = raid_disks - conf->max_degraded;
1935 sector_t new_sector = sh->sector, check;
1936 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1937 : conf->chunk_sectors;
1938 int algorithm = previous ? conf->prev_algo
1942 sector_t chunk_number;
1943 int dummy1, dd_idx = i;
1945 struct stripe_head sh2;
1948 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1949 stripe = new_sector;
1951 if (i == sh->pd_idx)
1953 switch(conf->level) {
1956 switch (algorithm) {
1957 case ALGORITHM_LEFT_ASYMMETRIC:
1958 case ALGORITHM_RIGHT_ASYMMETRIC:
1962 case ALGORITHM_LEFT_SYMMETRIC:
1963 case ALGORITHM_RIGHT_SYMMETRIC:
1966 i -= (sh->pd_idx + 1);
1968 case ALGORITHM_PARITY_0:
1971 case ALGORITHM_PARITY_N:
1978 if (i == sh->qd_idx)
1979 return 0; /* It is the Q disk */
1980 switch (algorithm) {
1981 case ALGORITHM_LEFT_ASYMMETRIC:
1982 case ALGORITHM_RIGHT_ASYMMETRIC:
1983 case ALGORITHM_ROTATING_ZERO_RESTART:
1984 case ALGORITHM_ROTATING_N_RESTART:
1985 if (sh->pd_idx == raid_disks-1)
1986 i--; /* Q D D D P */
1987 else if (i > sh->pd_idx)
1988 i -= 2; /* D D P Q D */
1990 case ALGORITHM_LEFT_SYMMETRIC:
1991 case ALGORITHM_RIGHT_SYMMETRIC:
1992 if (sh->pd_idx == raid_disks-1)
1993 i--; /* Q D D D P */
1998 i -= (sh->pd_idx + 2);
2001 case ALGORITHM_PARITY_0:
2004 case ALGORITHM_PARITY_N:
2006 case ALGORITHM_ROTATING_N_CONTINUE:
2007 /* Like left_symmetric, but P is before Q */
2008 if (sh->pd_idx == 0)
2009 i--; /* P D D D Q */
2014 i -= (sh->pd_idx + 1);
2017 case ALGORITHM_LEFT_ASYMMETRIC_6:
2018 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2022 case ALGORITHM_LEFT_SYMMETRIC_6:
2023 case ALGORITHM_RIGHT_SYMMETRIC_6:
2025 i += data_disks + 1;
2026 i -= (sh->pd_idx + 1);
2028 case ALGORITHM_PARITY_0_6:
2037 chunk_number = stripe * data_disks + i;
2038 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2040 check = raid5_compute_sector(conf, r_sector,
2041 previous, &dummy1, &sh2);
2042 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2043 || sh2.qd_idx != sh->qd_idx) {
2044 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2045 mdname(conf->mddev));
2053 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2054 int rcw, int expand)
2056 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2057 raid5_conf_t *conf = sh->raid_conf;
2058 int level = conf->level;
2061 /* if we are not expanding this is a proper write request, and
2062 * there will be bios with new data to be drained into the
2066 sh->reconstruct_state = reconstruct_state_drain_run;
2067 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2069 sh->reconstruct_state = reconstruct_state_run;
2071 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2073 for (i = disks; i--; ) {
2074 struct r5dev *dev = &sh->dev[i];
2077 set_bit(R5_LOCKED, &dev->flags);
2078 set_bit(R5_Wantdrain, &dev->flags);
2080 clear_bit(R5_UPTODATE, &dev->flags);
2084 if (s->locked + conf->max_degraded == disks)
2085 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2086 atomic_inc(&conf->pending_full_writes);
2089 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2090 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2092 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2093 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2094 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2095 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2097 for (i = disks; i--; ) {
2098 struct r5dev *dev = &sh->dev[i];
2103 (test_bit(R5_UPTODATE, &dev->flags) ||
2104 test_bit(R5_Wantcompute, &dev->flags))) {
2105 set_bit(R5_Wantdrain, &dev->flags);
2106 set_bit(R5_LOCKED, &dev->flags);
2107 clear_bit(R5_UPTODATE, &dev->flags);
2113 /* keep the parity disk(s) locked while asynchronous operations
2116 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2117 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2121 int qd_idx = sh->qd_idx;
2122 struct r5dev *dev = &sh->dev[qd_idx];
2124 set_bit(R5_LOCKED, &dev->flags);
2125 clear_bit(R5_UPTODATE, &dev->flags);
2129 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2130 __func__, (unsigned long long)sh->sector,
2131 s->locked, s->ops_request);
2135 * Each stripe/dev can have one or more bion attached.
2136 * toread/towrite point to the first in a chain.
2137 * The bi_next chain must be in order.
2139 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2142 raid5_conf_t *conf = sh->raid_conf;
2145 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2146 (unsigned long long)bi->bi_sector,
2147 (unsigned long long)sh->sector);
2150 spin_lock_irq(&conf->device_lock);
2152 bip = &sh->dev[dd_idx].towrite;
2153 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2156 bip = &sh->dev[dd_idx].toread;
2157 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2158 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2160 bip = & (*bip)->bi_next;
2162 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2165 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2169 bi->bi_phys_segments++;
2172 /* check if page is covered */
2173 sector_t sector = sh->dev[dd_idx].sector;
2174 for (bi=sh->dev[dd_idx].towrite;
2175 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2176 bi && bi->bi_sector <= sector;
2177 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2178 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2179 sector = bi->bi_sector + (bi->bi_size>>9);
2181 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2182 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2184 spin_unlock_irq(&conf->device_lock);
2186 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2187 (unsigned long long)(*bip)->bi_sector,
2188 (unsigned long long)sh->sector, dd_idx);
2190 if (conf->mddev->bitmap && firstwrite) {
2191 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2193 sh->bm_seq = conf->seq_flush+1;
2194 set_bit(STRIPE_BIT_DELAY, &sh->state);
2199 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2200 spin_unlock_irq(&conf->device_lock);
2204 static void end_reshape(raid5_conf_t *conf);
2206 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2207 struct stripe_head *sh)
2209 int sectors_per_chunk =
2210 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2212 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2213 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2215 raid5_compute_sector(conf,
2216 stripe * (disks - conf->max_degraded)
2217 *sectors_per_chunk + chunk_offset,
2223 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2224 struct stripe_head_state *s, int disks,
2225 struct bio **return_bi)
2228 for (i = disks; i--; ) {
2232 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2235 rdev = rcu_dereference(conf->disks[i].rdev);
2236 if (rdev && test_bit(In_sync, &rdev->flags))
2237 /* multiple read failures in one stripe */
2238 md_error(conf->mddev, rdev);
2241 spin_lock_irq(&conf->device_lock);
2242 /* fail all writes first */
2243 bi = sh->dev[i].towrite;
2244 sh->dev[i].towrite = NULL;
2250 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2251 wake_up(&conf->wait_for_overlap);
2253 while (bi && bi->bi_sector <
2254 sh->dev[i].sector + STRIPE_SECTORS) {
2255 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2256 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2257 if (!raid5_dec_bi_phys_segments(bi)) {
2258 md_write_end(conf->mddev);
2259 bi->bi_next = *return_bi;
2264 /* and fail all 'written' */
2265 bi = sh->dev[i].written;
2266 sh->dev[i].written = NULL;
2267 if (bi) bitmap_end = 1;
2268 while (bi && bi->bi_sector <
2269 sh->dev[i].sector + STRIPE_SECTORS) {
2270 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2271 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2272 if (!raid5_dec_bi_phys_segments(bi)) {
2273 md_write_end(conf->mddev);
2274 bi->bi_next = *return_bi;
2280 /* fail any reads if this device is non-operational and
2281 * the data has not reached the cache yet.
2283 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2284 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2285 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2286 bi = sh->dev[i].toread;
2287 sh->dev[i].toread = NULL;
2288 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2289 wake_up(&conf->wait_for_overlap);
2290 if (bi) s->to_read--;
2291 while (bi && bi->bi_sector <
2292 sh->dev[i].sector + STRIPE_SECTORS) {
2293 struct bio *nextbi =
2294 r5_next_bio(bi, sh->dev[i].sector);
2295 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2296 if (!raid5_dec_bi_phys_segments(bi)) {
2297 bi->bi_next = *return_bi;
2303 spin_unlock_irq(&conf->device_lock);
2305 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2306 STRIPE_SECTORS, 0, 0);
2309 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2310 if (atomic_dec_and_test(&conf->pending_full_writes))
2311 md_wakeup_thread(conf->mddev->thread);
2314 /* fetch_block - checks the given member device to see if its data needs
2315 * to be read or computed to satisfy a request.
2317 * Returns 1 when no more member devices need to be checked, otherwise returns
2318 * 0 to tell the loop in handle_stripe_fill to continue
2320 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2321 int disk_idx, int disks)
2323 struct r5dev *dev = &sh->dev[disk_idx];
2324 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2325 &sh->dev[s->failed_num[1]] };
2327 /* is the data in this block needed, and can we get it? */
2328 if (!test_bit(R5_LOCKED, &dev->flags) &&
2329 !test_bit(R5_UPTODATE, &dev->flags) &&
2331 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2332 s->syncing || s->expanding ||
2333 (s->failed >= 1 && fdev[0]->toread) ||
2334 (s->failed >= 2 && fdev[1]->toread) ||
2335 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2336 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2337 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2338 /* we would like to get this block, possibly by computing it,
2339 * otherwise read it if the backing disk is insync
2341 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2342 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2343 if ((s->uptodate == disks - 1) &&
2344 (s->failed && (disk_idx == s->failed_num[0] ||
2345 disk_idx == s->failed_num[1]))) {
2346 /* have disk failed, and we're requested to fetch it;
2349 pr_debug("Computing stripe %llu block %d\n",
2350 (unsigned long long)sh->sector, disk_idx);
2351 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2352 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2353 set_bit(R5_Wantcompute, &dev->flags);
2354 sh->ops.target = disk_idx;
2355 sh->ops.target2 = -1; /* no 2nd target */
2357 /* Careful: from this point on 'uptodate' is in the eye
2358 * of raid_run_ops which services 'compute' operations
2359 * before writes. R5_Wantcompute flags a block that will
2360 * be R5_UPTODATE by the time it is needed for a
2361 * subsequent operation.
2365 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2366 /* Computing 2-failure is *very* expensive; only
2367 * do it if failed >= 2
2370 for (other = disks; other--; ) {
2371 if (other == disk_idx)
2373 if (!test_bit(R5_UPTODATE,
2374 &sh->dev[other].flags))
2378 pr_debug("Computing stripe %llu blocks %d,%d\n",
2379 (unsigned long long)sh->sector,
2381 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2382 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2383 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2384 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2385 sh->ops.target = disk_idx;
2386 sh->ops.target2 = other;
2390 } else if (test_bit(R5_Insync, &dev->flags)) {
2391 set_bit(R5_LOCKED, &dev->flags);
2392 set_bit(R5_Wantread, &dev->flags);
2394 pr_debug("Reading block %d (sync=%d)\n",
2395 disk_idx, s->syncing);
2403 * handle_stripe_fill - read or compute data to satisfy pending requests.
2405 static void handle_stripe_fill(struct stripe_head *sh,
2406 struct stripe_head_state *s,
2411 /* look for blocks to read/compute, skip this if a compute
2412 * is already in flight, or if the stripe contents are in the
2413 * midst of changing due to a write
2415 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2416 !sh->reconstruct_state)
2417 for (i = disks; i--; )
2418 if (fetch_block(sh, s, i, disks))
2420 set_bit(STRIPE_HANDLE, &sh->state);
2424 /* handle_stripe_clean_event
2425 * any written block on an uptodate or failed drive can be returned.
2426 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2427 * never LOCKED, so we don't need to test 'failed' directly.
2429 static void handle_stripe_clean_event(raid5_conf_t *conf,
2430 struct stripe_head *sh, int disks, struct bio **return_bi)
2435 for (i = disks; i--; )
2436 if (sh->dev[i].written) {
2438 if (!test_bit(R5_LOCKED, &dev->flags) &&
2439 test_bit(R5_UPTODATE, &dev->flags)) {
2440 /* We can return any write requests */
2441 struct bio *wbi, *wbi2;
2443 pr_debug("Return write for disc %d\n", i);
2444 spin_lock_irq(&conf->device_lock);
2446 dev->written = NULL;
2447 while (wbi && wbi->bi_sector <
2448 dev->sector + STRIPE_SECTORS) {
2449 wbi2 = r5_next_bio(wbi, dev->sector);
2450 if (!raid5_dec_bi_phys_segments(wbi)) {
2451 md_write_end(conf->mddev);
2452 wbi->bi_next = *return_bi;
2457 if (dev->towrite == NULL)
2459 spin_unlock_irq(&conf->device_lock);
2461 bitmap_endwrite(conf->mddev->bitmap,
2464 !test_bit(STRIPE_DEGRADED, &sh->state),
2469 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2470 if (atomic_dec_and_test(&conf->pending_full_writes))
2471 md_wakeup_thread(conf->mddev->thread);
2474 static void handle_stripe_dirtying(raid5_conf_t *conf,
2475 struct stripe_head *sh,
2476 struct stripe_head_state *s,
2479 int rmw = 0, rcw = 0, i;
2480 if (conf->max_degraded == 2) {
2481 /* RAID6 requires 'rcw' in current implementation
2482 * Calculate the real rcw later - for now fake it
2483 * look like rcw is cheaper
2486 } else for (i = disks; i--; ) {
2487 /* would I have to read this buffer for read_modify_write */
2488 struct r5dev *dev = &sh->dev[i];
2489 if ((dev->towrite || i == sh->pd_idx) &&
2490 !test_bit(R5_LOCKED, &dev->flags) &&
2491 !(test_bit(R5_UPTODATE, &dev->flags) ||
2492 test_bit(R5_Wantcompute, &dev->flags))) {
2493 if (test_bit(R5_Insync, &dev->flags))
2496 rmw += 2*disks; /* cannot read it */
2498 /* Would I have to read this buffer for reconstruct_write */
2499 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2500 !test_bit(R5_LOCKED, &dev->flags) &&
2501 !(test_bit(R5_UPTODATE, &dev->flags) ||
2502 test_bit(R5_Wantcompute, &dev->flags))) {
2503 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2508 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2509 (unsigned long long)sh->sector, rmw, rcw);
2510 set_bit(STRIPE_HANDLE, &sh->state);
2511 if (rmw < rcw && rmw > 0)
2512 /* prefer read-modify-write, but need to get some data */
2513 for (i = disks; i--; ) {
2514 struct r5dev *dev = &sh->dev[i];
2515 if ((dev->towrite || i == sh->pd_idx) &&
2516 !test_bit(R5_LOCKED, &dev->flags) &&
2517 !(test_bit(R5_UPTODATE, &dev->flags) ||
2518 test_bit(R5_Wantcompute, &dev->flags)) &&
2519 test_bit(R5_Insync, &dev->flags)) {
2521 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2522 pr_debug("Read_old block "
2523 "%d for r-m-w\n", i);
2524 set_bit(R5_LOCKED, &dev->flags);
2525 set_bit(R5_Wantread, &dev->flags);
2528 set_bit(STRIPE_DELAYED, &sh->state);
2529 set_bit(STRIPE_HANDLE, &sh->state);
2533 if (rcw <= rmw && rcw > 0) {
2534 /* want reconstruct write, but need to get some data */
2536 for (i = disks; i--; ) {
2537 struct r5dev *dev = &sh->dev[i];
2538 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2539 i != sh->pd_idx && i != sh->qd_idx &&
2540 !test_bit(R5_LOCKED, &dev->flags) &&
2541 !(test_bit(R5_UPTODATE, &dev->flags) ||
2542 test_bit(R5_Wantcompute, &dev->flags))) {
2544 if (!test_bit(R5_Insync, &dev->flags))
2545 continue; /* it's a failed drive */
2547 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2548 pr_debug("Read_old block "
2549 "%d for Reconstruct\n", i);
2550 set_bit(R5_LOCKED, &dev->flags);
2551 set_bit(R5_Wantread, &dev->flags);
2554 set_bit(STRIPE_DELAYED, &sh->state);
2555 set_bit(STRIPE_HANDLE, &sh->state);
2560 /* now if nothing is locked, and if we have enough data,
2561 * we can start a write request
2563 /* since handle_stripe can be called at any time we need to handle the
2564 * case where a compute block operation has been submitted and then a
2565 * subsequent call wants to start a write request. raid_run_ops only
2566 * handles the case where compute block and reconstruct are requested
2567 * simultaneously. If this is not the case then new writes need to be
2568 * held off until the compute completes.
2570 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2571 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2572 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2573 schedule_reconstruction(sh, s, rcw == 0, 0);
2576 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2577 struct stripe_head_state *s, int disks)
2579 struct r5dev *dev = NULL;
2581 set_bit(STRIPE_HANDLE, &sh->state);
2583 switch (sh->check_state) {
2584 case check_state_idle:
2585 /* start a new check operation if there are no failures */
2586 if (s->failed == 0) {
2587 BUG_ON(s->uptodate != disks);
2588 sh->check_state = check_state_run;
2589 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2590 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2594 dev = &sh->dev[s->failed_num[0]];
2596 case check_state_compute_result:
2597 sh->check_state = check_state_idle;
2599 dev = &sh->dev[sh->pd_idx];
2601 /* check that a write has not made the stripe insync */
2602 if (test_bit(STRIPE_INSYNC, &sh->state))
2605 /* either failed parity check, or recovery is happening */
2606 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2607 BUG_ON(s->uptodate != disks);
2609 set_bit(R5_LOCKED, &dev->flags);
2611 set_bit(R5_Wantwrite, &dev->flags);
2613 clear_bit(STRIPE_DEGRADED, &sh->state);
2614 set_bit(STRIPE_INSYNC, &sh->state);
2616 case check_state_run:
2617 break; /* we will be called again upon completion */
2618 case check_state_check_result:
2619 sh->check_state = check_state_idle;
2621 /* if a failure occurred during the check operation, leave
2622 * STRIPE_INSYNC not set and let the stripe be handled again
2627 /* handle a successful check operation, if parity is correct
2628 * we are done. Otherwise update the mismatch count and repair
2629 * parity if !MD_RECOVERY_CHECK
2631 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2632 /* parity is correct (on disc,
2633 * not in buffer any more)
2635 set_bit(STRIPE_INSYNC, &sh->state);
2637 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2638 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2639 /* don't try to repair!! */
2640 set_bit(STRIPE_INSYNC, &sh->state);
2642 sh->check_state = check_state_compute_run;
2643 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2644 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2645 set_bit(R5_Wantcompute,
2646 &sh->dev[sh->pd_idx].flags);
2647 sh->ops.target = sh->pd_idx;
2648 sh->ops.target2 = -1;
2653 case check_state_compute_run:
2656 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2657 __func__, sh->check_state,
2658 (unsigned long long) sh->sector);
2664 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2665 struct stripe_head_state *s,
2668 int pd_idx = sh->pd_idx;
2669 int qd_idx = sh->qd_idx;
2672 set_bit(STRIPE_HANDLE, &sh->state);
2674 BUG_ON(s->failed > 2);
2676 /* Want to check and possibly repair P and Q.
2677 * However there could be one 'failed' device, in which
2678 * case we can only check one of them, possibly using the
2679 * other to generate missing data
2682 switch (sh->check_state) {
2683 case check_state_idle:
2684 /* start a new check operation if there are < 2 failures */
2685 if (s->failed == s->q_failed) {
2686 /* The only possible failed device holds Q, so it
2687 * makes sense to check P (If anything else were failed,
2688 * we would have used P to recreate it).
2690 sh->check_state = check_state_run;
2692 if (!s->q_failed && s->failed < 2) {
2693 /* Q is not failed, and we didn't use it to generate
2694 * anything, so it makes sense to check it
2696 if (sh->check_state == check_state_run)
2697 sh->check_state = check_state_run_pq;
2699 sh->check_state = check_state_run_q;
2702 /* discard potentially stale zero_sum_result */
2703 sh->ops.zero_sum_result = 0;
2705 if (sh->check_state == check_state_run) {
2706 /* async_xor_zero_sum destroys the contents of P */
2707 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2710 if (sh->check_state >= check_state_run &&
2711 sh->check_state <= check_state_run_pq) {
2712 /* async_syndrome_zero_sum preserves P and Q, so
2713 * no need to mark them !uptodate here
2715 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2719 /* we have 2-disk failure */
2720 BUG_ON(s->failed != 2);
2722 case check_state_compute_result:
2723 sh->check_state = check_state_idle;
2725 /* check that a write has not made the stripe insync */
2726 if (test_bit(STRIPE_INSYNC, &sh->state))
2729 /* now write out any block on a failed drive,
2730 * or P or Q if they were recomputed
2732 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2733 if (s->failed == 2) {
2734 dev = &sh->dev[s->failed_num[1]];
2736 set_bit(R5_LOCKED, &dev->flags);
2737 set_bit(R5_Wantwrite, &dev->flags);
2739 if (s->failed >= 1) {
2740 dev = &sh->dev[s->failed_num[0]];
2742 set_bit(R5_LOCKED, &dev->flags);
2743 set_bit(R5_Wantwrite, &dev->flags);
2745 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2746 dev = &sh->dev[pd_idx];
2748 set_bit(R5_LOCKED, &dev->flags);
2749 set_bit(R5_Wantwrite, &dev->flags);
2751 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2752 dev = &sh->dev[qd_idx];
2754 set_bit(R5_LOCKED, &dev->flags);
2755 set_bit(R5_Wantwrite, &dev->flags);
2757 clear_bit(STRIPE_DEGRADED, &sh->state);
2759 set_bit(STRIPE_INSYNC, &sh->state);
2761 case check_state_run:
2762 case check_state_run_q:
2763 case check_state_run_pq:
2764 break; /* we will be called again upon completion */
2765 case check_state_check_result:
2766 sh->check_state = check_state_idle;
2768 /* handle a successful check operation, if parity is correct
2769 * we are done. Otherwise update the mismatch count and repair
2770 * parity if !MD_RECOVERY_CHECK
2772 if (sh->ops.zero_sum_result == 0) {
2773 /* both parities are correct */
2775 set_bit(STRIPE_INSYNC, &sh->state);
2777 /* in contrast to the raid5 case we can validate
2778 * parity, but still have a failure to write
2781 sh->check_state = check_state_compute_result;
2782 /* Returning at this point means that we may go
2783 * off and bring p and/or q uptodate again so
2784 * we make sure to check zero_sum_result again
2785 * to verify if p or q need writeback
2789 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2790 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2791 /* don't try to repair!! */
2792 set_bit(STRIPE_INSYNC, &sh->state);
2794 int *target = &sh->ops.target;
2796 sh->ops.target = -1;
2797 sh->ops.target2 = -1;
2798 sh->check_state = check_state_compute_run;
2799 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2800 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2801 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2802 set_bit(R5_Wantcompute,
2803 &sh->dev[pd_idx].flags);
2805 target = &sh->ops.target2;
2808 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2809 set_bit(R5_Wantcompute,
2810 &sh->dev[qd_idx].flags);
2817 case check_state_compute_run:
2820 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2821 __func__, sh->check_state,
2822 (unsigned long long) sh->sector);
2827 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh)
2831 /* We have read all the blocks in this stripe and now we need to
2832 * copy some of them into a target stripe for expand.
2834 struct dma_async_tx_descriptor *tx = NULL;
2835 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2836 for (i = 0; i < sh->disks; i++)
2837 if (i != sh->pd_idx && i != sh->qd_idx) {
2839 struct stripe_head *sh2;
2840 struct async_submit_ctl submit;
2842 sector_t bn = compute_blocknr(sh, i, 1);
2843 sector_t s = raid5_compute_sector(conf, bn, 0,
2845 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2847 /* so far only the early blocks of this stripe
2848 * have been requested. When later blocks
2849 * get requested, we will try again
2852 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2853 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2854 /* must have already done this block */
2855 release_stripe(sh2);
2859 /* place all the copies on one channel */
2860 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2861 tx = async_memcpy(sh2->dev[dd_idx].page,
2862 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2865 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2866 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2867 for (j = 0; j < conf->raid_disks; j++)
2868 if (j != sh2->pd_idx &&
2870 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2872 if (j == conf->raid_disks) {
2873 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2874 set_bit(STRIPE_HANDLE, &sh2->state);
2876 release_stripe(sh2);
2879 /* done submitting copies, wait for them to complete */
2882 dma_wait_for_async_tx(tx);
2888 * handle_stripe - do things to a stripe.
2890 * We lock the stripe and then examine the state of various bits
2891 * to see what needs to be done.
2893 * return some read request which now have data
2894 * return some write requests which are safely on disc
2895 * schedule a read on some buffers
2896 * schedule a write of some buffers
2897 * return confirmation of parity correctness
2899 * buffers are taken off read_list or write_list, and bh_cache buffers
2900 * get BH_Lock set before the stripe lock is released.
2904 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
2906 raid5_conf_t *conf = sh->raid_conf;
2907 int disks = sh->disks;
2911 memset(s, 0, sizeof(*s));
2913 s->syncing = test_bit(STRIPE_SYNCING, &sh->state);
2914 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2915 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2916 s->failed_num[0] = -1;
2917 s->failed_num[1] = -1;
2919 /* Now to look around and see what can be done */
2921 spin_lock_irq(&conf->device_lock);
2922 for (i=disks; i--; ) {
2927 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2928 i, dev->flags, dev->toread, dev->towrite, dev->written);
2929 /* maybe we can reply to a read
2931 * new wantfill requests are only permitted while
2932 * ops_complete_biofill is guaranteed to be inactive
2934 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2935 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2936 set_bit(R5_Wantfill, &dev->flags);
2938 /* now count some things */
2939 if (test_bit(R5_LOCKED, &dev->flags))
2941 if (test_bit(R5_UPTODATE, &dev->flags))
2943 if (test_bit(R5_Wantcompute, &dev->flags)) {
2945 BUG_ON(s->compute > 2);
2948 if (test_bit(R5_Wantfill, &dev->flags))
2950 else if (dev->toread)
2954 if (!test_bit(R5_OVERWRITE, &dev->flags))
2959 rdev = rcu_dereference(conf->disks[i].rdev);
2960 if (s->blocked_rdev == NULL &&
2961 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2962 s->blocked_rdev = rdev;
2963 atomic_inc(&rdev->nr_pending);
2965 clear_bit(R5_Insync, &dev->flags);
2968 else if (test_bit(In_sync, &rdev->flags))
2969 set_bit(R5_Insync, &dev->flags);
2971 /* in sync if before recovery_offset */
2972 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
2973 set_bit(R5_Insync, &dev->flags);
2975 if (!test_bit(R5_Insync, &dev->flags)) {
2976 /* The ReadError flag will just be confusing now */
2977 clear_bit(R5_ReadError, &dev->flags);
2978 clear_bit(R5_ReWrite, &dev->flags);
2980 if (test_bit(R5_ReadError, &dev->flags))
2981 clear_bit(R5_Insync, &dev->flags);
2982 if (!test_bit(R5_Insync, &dev->flags)) {
2984 s->failed_num[s->failed] = i;
2988 spin_unlock_irq(&conf->device_lock);
2992 static void handle_stripe(struct stripe_head *sh)
2994 struct stripe_head_state s;
2995 raid5_conf_t *conf = sh->raid_conf;
2998 int disks = sh->disks;
2999 struct r5dev *pdev, *qdev;
3001 clear_bit(STRIPE_HANDLE, &sh->state);
3002 if (test_and_set_bit(STRIPE_ACTIVE, &sh->state)) {
3003 /* already being handled, ensure it gets handled
3004 * again when current action finishes */
3005 set_bit(STRIPE_HANDLE, &sh->state);
3009 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3010 set_bit(STRIPE_SYNCING, &sh->state);
3011 clear_bit(STRIPE_INSYNC, &sh->state);
3013 clear_bit(STRIPE_DELAYED, &sh->state);
3015 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3016 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3017 (unsigned long long)sh->sector, sh->state,
3018 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3019 sh->check_state, sh->reconstruct_state);
3021 analyse_stripe(sh, &s);
3023 if (unlikely(s.blocked_rdev)) {
3024 if (s.syncing || s.expanding || s.expanded ||
3025 s.to_write || s.written) {
3026 set_bit(STRIPE_HANDLE, &sh->state);
3029 /* There is nothing for the blocked_rdev to block */
3030 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3031 s.blocked_rdev = NULL;
3034 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3035 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3036 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3039 pr_debug("locked=%d uptodate=%d to_read=%d"
3040 " to_write=%d failed=%d failed_num=%d,%d\n",
3041 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3042 s.failed_num[0], s.failed_num[1]);
3043 /* check if the array has lost more than max_degraded devices and,
3044 * if so, some requests might need to be failed.
3046 if (s.failed > conf->max_degraded && s.to_read+s.to_write+s.written)
3047 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3048 if (s.failed > conf->max_degraded && s.syncing) {
3049 md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
3050 clear_bit(STRIPE_SYNCING, &sh->state);
3055 * might be able to return some write requests if the parity blocks
3056 * are safe, or on a failed drive
3058 pdev = &sh->dev[sh->pd_idx];
3059 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3060 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3061 qdev = &sh->dev[sh->qd_idx];
3062 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3063 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3067 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3068 && !test_bit(R5_LOCKED, &pdev->flags)
3069 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3070 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3071 && !test_bit(R5_LOCKED, &qdev->flags)
3072 && test_bit(R5_UPTODATE, &qdev->flags)))))
3073 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3075 /* Now we might consider reading some blocks, either to check/generate
3076 * parity, or to satisfy requests
3077 * or to load a block that is being partially written.
3079 if (s.to_read || s.non_overwrite
3080 || (conf->level == 6 && s.to_write && s.failed)
3081 || (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3082 handle_stripe_fill(sh, &s, disks);
3084 /* Now we check to see if any write operations have recently
3088 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3090 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3091 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3092 sh->reconstruct_state = reconstruct_state_idle;
3094 /* All the 'written' buffers and the parity block are ready to
3095 * be written back to disk
3097 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3098 BUG_ON(sh->qd_idx >= 0 &&
3099 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3100 for (i = disks; i--; ) {
3101 struct r5dev *dev = &sh->dev[i];
3102 if (test_bit(R5_LOCKED, &dev->flags) &&
3103 (i == sh->pd_idx || i == sh->qd_idx ||
3105 pr_debug("Writing block %d\n", i);
3106 set_bit(R5_Wantwrite, &dev->flags);
3109 if (!test_bit(R5_Insync, &dev->flags) ||
3110 ((i == sh->pd_idx || i == sh->qd_idx) &&
3112 set_bit(STRIPE_INSYNC, &sh->state);
3115 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3116 s.dec_preread_active = 1;
3119 /* Now to consider new write requests and what else, if anything
3120 * should be read. We do not handle new writes when:
3121 * 1/ A 'write' operation (copy+xor) is already in flight.
3122 * 2/ A 'check' operation is in flight, as it may clobber the parity
3125 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3126 handle_stripe_dirtying(conf, sh, &s, disks);
3128 /* maybe we need to check and possibly fix the parity for this stripe
3129 * Any reads will already have been scheduled, so we just see if enough
3130 * data is available. The parity check is held off while parity
3131 * dependent operations are in flight.
3133 if (sh->check_state ||
3134 (s.syncing && s.locked == 0 &&
3135 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3136 !test_bit(STRIPE_INSYNC, &sh->state))) {
3137 if (conf->level == 6)
3138 handle_parity_checks6(conf, sh, &s, disks);
3140 handle_parity_checks5(conf, sh, &s, disks);
3143 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3144 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3145 clear_bit(STRIPE_SYNCING, &sh->state);
3148 /* If the failed drives are just a ReadError, then we might need
3149 * to progress the repair/check process
3151 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3152 for (i = 0; i < s.failed; i++) {
3153 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3154 if (test_bit(R5_ReadError, &dev->flags)
3155 && !test_bit(R5_LOCKED, &dev->flags)
3156 && test_bit(R5_UPTODATE, &dev->flags)
3158 if (!test_bit(R5_ReWrite, &dev->flags)) {
3159 set_bit(R5_Wantwrite, &dev->flags);
3160 set_bit(R5_ReWrite, &dev->flags);
3161 set_bit(R5_LOCKED, &dev->flags);
3164 /* let's read it back */
3165 set_bit(R5_Wantread, &dev->flags);
3166 set_bit(R5_LOCKED, &dev->flags);
3173 /* Finish reconstruct operations initiated by the expansion process */
3174 if (sh->reconstruct_state == reconstruct_state_result) {
3175 struct stripe_head *sh_src
3176 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3177 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3178 /* sh cannot be written until sh_src has been read.
3179 * so arrange for sh to be delayed a little
3181 set_bit(STRIPE_DELAYED, &sh->state);
3182 set_bit(STRIPE_HANDLE, &sh->state);
3183 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3185 atomic_inc(&conf->preread_active_stripes);
3186 release_stripe(sh_src);
3190 release_stripe(sh_src);
3192 sh->reconstruct_state = reconstruct_state_idle;
3193 clear_bit(STRIPE_EXPANDING, &sh->state);
3194 for (i = conf->raid_disks; i--; ) {
3195 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3196 set_bit(R5_LOCKED, &sh->dev[i].flags);
3201 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3202 !sh->reconstruct_state) {
3203 /* Need to write out all blocks after computing parity */
3204 sh->disks = conf->raid_disks;
3205 stripe_set_idx(sh->sector, conf, 0, sh);
3206 schedule_reconstruction(sh, &s, 1, 1);
3207 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3208 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3209 atomic_dec(&conf->reshape_stripes);
3210 wake_up(&conf->wait_for_overlap);
3211 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3214 if (s.expanding && s.locked == 0 &&
3215 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3216 handle_stripe_expansion(conf, sh);
3219 /* wait for this device to become unblocked */
3220 if (unlikely(s.blocked_rdev))
3221 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3224 raid_run_ops(sh, s.ops_request);
3229 if (s.dec_preread_active) {
3230 /* We delay this until after ops_run_io so that if make_request
3231 * is waiting on a flush, it won't continue until the writes
3232 * have actually been submitted.
3234 atomic_dec(&conf->preread_active_stripes);
3235 if (atomic_read(&conf->preread_active_stripes) <
3237 md_wakeup_thread(conf->mddev->thread);
3240 return_io(s.return_bi);
3242 clear_bit(STRIPE_ACTIVE, &sh->state);
3245 static void raid5_activate_delayed(raid5_conf_t *conf)
3247 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3248 while (!list_empty(&conf->delayed_list)) {
3249 struct list_head *l = conf->delayed_list.next;
3250 struct stripe_head *sh;
3251 sh = list_entry(l, struct stripe_head, lru);
3253 clear_bit(STRIPE_DELAYED, &sh->state);
3254 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3255 atomic_inc(&conf->preread_active_stripes);
3256 list_add_tail(&sh->lru, &conf->hold_list);
3261 static void activate_bit_delay(raid5_conf_t *conf)
3263 /* device_lock is held */
3264 struct list_head head;
3265 list_add(&head, &conf->bitmap_list);
3266 list_del_init(&conf->bitmap_list);
3267 while (!list_empty(&head)) {
3268 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3269 list_del_init(&sh->lru);
3270 atomic_inc(&sh->count);
3271 __release_stripe(conf, sh);
3275 int md_raid5_congested(mddev_t *mddev, int bits)
3277 raid5_conf_t *conf = mddev->private;
3279 /* No difference between reads and writes. Just check
3280 * how busy the stripe_cache is
3283 if (conf->inactive_blocked)
3287 if (list_empty_careful(&conf->inactive_list))
3292 EXPORT_SYMBOL_GPL(md_raid5_congested);
3294 static int raid5_congested(void *data, int bits)
3296 mddev_t *mddev = data;
3298 return mddev_congested(mddev, bits) ||
3299 md_raid5_congested(mddev, bits);
3302 /* We want read requests to align with chunks where possible,
3303 * but write requests don't need to.
3305 static int raid5_mergeable_bvec(struct request_queue *q,
3306 struct bvec_merge_data *bvm,
3307 struct bio_vec *biovec)
3309 mddev_t *mddev = q->queuedata;
3310 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3312 unsigned int chunk_sectors = mddev->chunk_sectors;
3313 unsigned int bio_sectors = bvm->bi_size >> 9;
3315 if ((bvm->bi_rw & 1) == WRITE)
3316 return biovec->bv_len; /* always allow writes to be mergeable */
3318 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3319 chunk_sectors = mddev->new_chunk_sectors;
3320 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3321 if (max < 0) max = 0;
3322 if (max <= biovec->bv_len && bio_sectors == 0)
3323 return biovec->bv_len;
3329 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3331 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3332 unsigned int chunk_sectors = mddev->chunk_sectors;
3333 unsigned int bio_sectors = bio->bi_size >> 9;
3335 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3336 chunk_sectors = mddev->new_chunk_sectors;
3337 return chunk_sectors >=
3338 ((sector & (chunk_sectors - 1)) + bio_sectors);
3342 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3343 * later sampled by raid5d.
3345 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3347 unsigned long flags;
3349 spin_lock_irqsave(&conf->device_lock, flags);
3351 bi->bi_next = conf->retry_read_aligned_list;
3352 conf->retry_read_aligned_list = bi;
3354 spin_unlock_irqrestore(&conf->device_lock, flags);
3355 md_wakeup_thread(conf->mddev->thread);
3359 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3363 bi = conf->retry_read_aligned;
3365 conf->retry_read_aligned = NULL;
3368 bi = conf->retry_read_aligned_list;
3370 conf->retry_read_aligned_list = bi->bi_next;
3373 * this sets the active strip count to 1 and the processed
3374 * strip count to zero (upper 8 bits)
3376 bi->bi_phys_segments = 1; /* biased count of active stripes */
3384 * The "raid5_align_endio" should check if the read succeeded and if it
3385 * did, call bio_endio on the original bio (having bio_put the new bio
3387 * If the read failed..
3389 static void raid5_align_endio(struct bio *bi, int error)
3391 struct bio* raid_bi = bi->bi_private;
3394 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3399 rdev = (void*)raid_bi->bi_next;
3400 raid_bi->bi_next = NULL;
3401 mddev = rdev->mddev;
3402 conf = mddev->private;
3404 rdev_dec_pending(rdev, conf->mddev);
3406 if (!error && uptodate) {
3407 bio_endio(raid_bi, 0);
3408 if (atomic_dec_and_test(&conf->active_aligned_reads))
3409 wake_up(&conf->wait_for_stripe);
3414 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3416 add_bio_to_retry(raid_bi, conf);
3419 static int bio_fits_rdev(struct bio *bi)
3421 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3423 if ((bi->bi_size>>9) > queue_max_sectors(q))
3425 blk_recount_segments(q, bi);
3426 if (bi->bi_phys_segments > queue_max_segments(q))
3429 if (q->merge_bvec_fn)
3430 /* it's too hard to apply the merge_bvec_fn at this stage,
3439 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio)
3441 raid5_conf_t *conf = mddev->private;
3443 struct bio* align_bi;
3446 if (!in_chunk_boundary(mddev, raid_bio)) {
3447 pr_debug("chunk_aligned_read : non aligned\n");
3451 * use bio_clone_mddev to make a copy of the bio
3453 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3457 * set bi_end_io to a new function, and set bi_private to the
3460 align_bi->bi_end_io = raid5_align_endio;
3461 align_bi->bi_private = raid_bio;
3465 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3470 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3471 if (rdev && test_bit(In_sync, &rdev->flags)) {
3472 atomic_inc(&rdev->nr_pending);
3474 raid_bio->bi_next = (void*)rdev;
3475 align_bi->bi_bdev = rdev->bdev;
3476 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3477 align_bi->bi_sector += rdev->data_offset;
3479 if (!bio_fits_rdev(align_bi)) {
3480 /* too big in some way */
3482 rdev_dec_pending(rdev, mddev);
3486 spin_lock_irq(&conf->device_lock);
3487 wait_event_lock_irq(conf->wait_for_stripe,
3489 conf->device_lock, /* nothing */);
3490 atomic_inc(&conf->active_aligned_reads);
3491 spin_unlock_irq(&conf->device_lock);
3493 generic_make_request(align_bi);
3502 /* __get_priority_stripe - get the next stripe to process
3504 * Full stripe writes are allowed to pass preread active stripes up until
3505 * the bypass_threshold is exceeded. In general the bypass_count
3506 * increments when the handle_list is handled before the hold_list; however, it
3507 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3508 * stripe with in flight i/o. The bypass_count will be reset when the
3509 * head of the hold_list has changed, i.e. the head was promoted to the
3512 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3514 struct stripe_head *sh;
3516 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3518 list_empty(&conf->handle_list) ? "empty" : "busy",
3519 list_empty(&conf->hold_list) ? "empty" : "busy",
3520 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3522 if (!list_empty(&conf->handle_list)) {
3523 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3525 if (list_empty(&conf->hold_list))
3526 conf->bypass_count = 0;
3527 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3528 if (conf->hold_list.next == conf->last_hold)
3529 conf->bypass_count++;
3531 conf->last_hold = conf->hold_list.next;
3532 conf->bypass_count -= conf->bypass_threshold;
3533 if (conf->bypass_count < 0)
3534 conf->bypass_count = 0;
3537 } else if (!list_empty(&conf->hold_list) &&
3538 ((conf->bypass_threshold &&
3539 conf->bypass_count > conf->bypass_threshold) ||
3540 atomic_read(&conf->pending_full_writes) == 0)) {
3541 sh = list_entry(conf->hold_list.next,
3543 conf->bypass_count -= conf->bypass_threshold;
3544 if (conf->bypass_count < 0)
3545 conf->bypass_count = 0;
3549 list_del_init(&sh->lru);
3550 atomic_inc(&sh->count);
3551 BUG_ON(atomic_read(&sh->count) != 1);
3555 static int make_request(mddev_t *mddev, struct bio * bi)
3557 raid5_conf_t *conf = mddev->private;
3559 sector_t new_sector;
3560 sector_t logical_sector, last_sector;
3561 struct stripe_head *sh;
3562 const int rw = bio_data_dir(bi);
3566 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3567 md_flush_request(mddev, bi);
3571 md_write_start(mddev, bi);
3574 mddev->reshape_position == MaxSector &&
3575 chunk_aligned_read(mddev,bi))
3578 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3579 last_sector = bi->bi_sector + (bi->bi_size>>9);
3581 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3583 plugged = mddev_check_plugged(mddev);
3584 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3586 int disks, data_disks;
3591 disks = conf->raid_disks;
3592 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3593 if (unlikely(conf->reshape_progress != MaxSector)) {
3594 /* spinlock is needed as reshape_progress may be
3595 * 64bit on a 32bit platform, and so it might be
3596 * possible to see a half-updated value
3597 * Of course reshape_progress could change after
3598 * the lock is dropped, so once we get a reference
3599 * to the stripe that we think it is, we will have
3602 spin_lock_irq(&conf->device_lock);
3603 if (mddev->delta_disks < 0
3604 ? logical_sector < conf->reshape_progress
3605 : logical_sector >= conf->reshape_progress) {
3606 disks = conf->previous_raid_disks;
3609 if (mddev->delta_disks < 0
3610 ? logical_sector < conf->reshape_safe
3611 : logical_sector >= conf->reshape_safe) {
3612 spin_unlock_irq(&conf->device_lock);
3617 spin_unlock_irq(&conf->device_lock);
3619 data_disks = disks - conf->max_degraded;
3621 new_sector = raid5_compute_sector(conf, logical_sector,
3624 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3625 (unsigned long long)new_sector,
3626 (unsigned long long)logical_sector);
3628 sh = get_active_stripe(conf, new_sector, previous,
3629 (bi->bi_rw&RWA_MASK), 0);
3631 if (unlikely(previous)) {
3632 /* expansion might have moved on while waiting for a
3633 * stripe, so we must do the range check again.
3634 * Expansion could still move past after this
3635 * test, but as we are holding a reference to
3636 * 'sh', we know that if that happens,
3637 * STRIPE_EXPANDING will get set and the expansion
3638 * won't proceed until we finish with the stripe.
3641 spin_lock_irq(&conf->device_lock);
3642 if (mddev->delta_disks < 0
3643 ? logical_sector >= conf->reshape_progress
3644 : logical_sector < conf->reshape_progress)
3645 /* mismatch, need to try again */
3647 spin_unlock_irq(&conf->device_lock);
3656 logical_sector >= mddev->suspend_lo &&
3657 logical_sector < mddev->suspend_hi) {
3659 /* As the suspend_* range is controlled by
3660 * userspace, we want an interruptible
3663 flush_signals(current);
3664 prepare_to_wait(&conf->wait_for_overlap,
3665 &w, TASK_INTERRUPTIBLE);
3666 if (logical_sector >= mddev->suspend_lo &&
3667 logical_sector < mddev->suspend_hi)
3672 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3673 !add_stripe_bio(sh, bi, dd_idx, rw)) {
3674 /* Stripe is busy expanding or
3675 * add failed due to overlap. Flush everything
3678 md_wakeup_thread(mddev->thread);
3683 finish_wait(&conf->wait_for_overlap, &w);
3684 set_bit(STRIPE_HANDLE, &sh->state);
3685 clear_bit(STRIPE_DELAYED, &sh->state);
3686 if ((bi->bi_rw & REQ_SYNC) &&
3687 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3688 atomic_inc(&conf->preread_active_stripes);
3691 /* cannot get stripe for read-ahead, just give-up */
3692 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3693 finish_wait(&conf->wait_for_overlap, &w);
3699 md_wakeup_thread(mddev->thread);
3701 spin_lock_irq(&conf->device_lock);
3702 remaining = raid5_dec_bi_phys_segments(bi);
3703 spin_unlock_irq(&conf->device_lock);
3704 if (remaining == 0) {
3707 md_write_end(mddev);
3715 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3717 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3719 /* reshaping is quite different to recovery/resync so it is
3720 * handled quite separately ... here.
3722 * On each call to sync_request, we gather one chunk worth of
3723 * destination stripes and flag them as expanding.
3724 * Then we find all the source stripes and request reads.
3725 * As the reads complete, handle_stripe will copy the data
3726 * into the destination stripe and release that stripe.
3728 raid5_conf_t *conf = mddev->private;
3729 struct stripe_head *sh;
3730 sector_t first_sector, last_sector;
3731 int raid_disks = conf->previous_raid_disks;
3732 int data_disks = raid_disks - conf->max_degraded;
3733 int new_data_disks = conf->raid_disks - conf->max_degraded;
3736 sector_t writepos, readpos, safepos;
3737 sector_t stripe_addr;
3738 int reshape_sectors;
3739 struct list_head stripes;
3741 if (sector_nr == 0) {
3742 /* If restarting in the middle, skip the initial sectors */
3743 if (mddev->delta_disks < 0 &&
3744 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3745 sector_nr = raid5_size(mddev, 0, 0)
3746 - conf->reshape_progress;
3747 } else if (mddev->delta_disks >= 0 &&
3748 conf->reshape_progress > 0)
3749 sector_nr = conf->reshape_progress;
3750 sector_div(sector_nr, new_data_disks);
3752 mddev->curr_resync_completed = sector_nr;
3753 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3759 /* We need to process a full chunk at a time.
3760 * If old and new chunk sizes differ, we need to process the
3763 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3764 reshape_sectors = mddev->new_chunk_sectors;
3766 reshape_sectors = mddev->chunk_sectors;
3768 /* we update the metadata when there is more than 3Meg
3769 * in the block range (that is rather arbitrary, should
3770 * probably be time based) or when the data about to be
3771 * copied would over-write the source of the data at
3772 * the front of the range.
3773 * i.e. one new_stripe along from reshape_progress new_maps
3774 * to after where reshape_safe old_maps to
3776 writepos = conf->reshape_progress;
3777 sector_div(writepos, new_data_disks);
3778 readpos = conf->reshape_progress;
3779 sector_div(readpos, data_disks);
3780 safepos = conf->reshape_safe;
3781 sector_div(safepos, data_disks);
3782 if (mddev->delta_disks < 0) {
3783 writepos -= min_t(sector_t, reshape_sectors, writepos);
3784 readpos += reshape_sectors;
3785 safepos += reshape_sectors;
3787 writepos += reshape_sectors;
3788 readpos -= min_t(sector_t, reshape_sectors, readpos);
3789 safepos -= min_t(sector_t, reshape_sectors, safepos);
3792 /* 'writepos' is the most advanced device address we might write.
3793 * 'readpos' is the least advanced device address we might read.
3794 * 'safepos' is the least address recorded in the metadata as having
3796 * If 'readpos' is behind 'writepos', then there is no way that we can
3797 * ensure safety in the face of a crash - that must be done by userspace
3798 * making a backup of the data. So in that case there is no particular
3799 * rush to update metadata.
3800 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3801 * update the metadata to advance 'safepos' to match 'readpos' so that
3802 * we can be safe in the event of a crash.
3803 * So we insist on updating metadata if safepos is behind writepos and
3804 * readpos is beyond writepos.
3805 * In any case, update the metadata every 10 seconds.
3806 * Maybe that number should be configurable, but I'm not sure it is
3807 * worth it.... maybe it could be a multiple of safemode_delay???
3809 if ((mddev->delta_disks < 0
3810 ? (safepos > writepos && readpos < writepos)
3811 : (safepos < writepos && readpos > writepos)) ||
3812 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3813 /* Cannot proceed until we've updated the superblock... */
3814 wait_event(conf->wait_for_overlap,
3815 atomic_read(&conf->reshape_stripes)==0);
3816 mddev->reshape_position = conf->reshape_progress;
3817 mddev->curr_resync_completed = sector_nr;
3818 conf->reshape_checkpoint = jiffies;
3819 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3820 md_wakeup_thread(mddev->thread);
3821 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3822 kthread_should_stop());
3823 spin_lock_irq(&conf->device_lock);
3824 conf->reshape_safe = mddev->reshape_position;
3825 spin_unlock_irq(&conf->device_lock);
3826 wake_up(&conf->wait_for_overlap);
3827 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3830 if (mddev->delta_disks < 0) {
3831 BUG_ON(conf->reshape_progress == 0);
3832 stripe_addr = writepos;
3833 BUG_ON((mddev->dev_sectors &
3834 ~((sector_t)reshape_sectors - 1))
3835 - reshape_sectors - stripe_addr
3838 BUG_ON(writepos != sector_nr + reshape_sectors);
3839 stripe_addr = sector_nr;
3841 INIT_LIST_HEAD(&stripes);
3842 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3844 int skipped_disk = 0;
3845 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
3846 set_bit(STRIPE_EXPANDING, &sh->state);
3847 atomic_inc(&conf->reshape_stripes);
3848 /* If any of this stripe is beyond the end of the old
3849 * array, then we need to zero those blocks
3851 for (j=sh->disks; j--;) {
3853 if (j == sh->pd_idx)
3855 if (conf->level == 6 &&
3858 s = compute_blocknr(sh, j, 0);
3859 if (s < raid5_size(mddev, 0, 0)) {
3863 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3864 set_bit(R5_Expanded, &sh->dev[j].flags);
3865 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3867 if (!skipped_disk) {
3868 set_bit(STRIPE_EXPAND_READY, &sh->state);
3869 set_bit(STRIPE_HANDLE, &sh->state);
3871 list_add(&sh->lru, &stripes);
3873 spin_lock_irq(&conf->device_lock);
3874 if (mddev->delta_disks < 0)
3875 conf->reshape_progress -= reshape_sectors * new_data_disks;
3877 conf->reshape_progress += reshape_sectors * new_data_disks;
3878 spin_unlock_irq(&conf->device_lock);
3879 /* Ok, those stripe are ready. We can start scheduling
3880 * reads on the source stripes.
3881 * The source stripes are determined by mapping the first and last
3882 * block on the destination stripes.
3885 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
3888 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
3889 * new_data_disks - 1),
3891 if (last_sector >= mddev->dev_sectors)
3892 last_sector = mddev->dev_sectors - 1;
3893 while (first_sector <= last_sector) {
3894 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
3895 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3896 set_bit(STRIPE_HANDLE, &sh->state);
3898 first_sector += STRIPE_SECTORS;
3900 /* Now that the sources are clearly marked, we can release
3901 * the destination stripes
3903 while (!list_empty(&stripes)) {
3904 sh = list_entry(stripes.next, struct stripe_head, lru);
3905 list_del_init(&sh->lru);
3908 /* If this takes us to the resync_max point where we have to pause,
3909 * then we need to write out the superblock.
3911 sector_nr += reshape_sectors;
3912 if ((sector_nr - mddev->curr_resync_completed) * 2
3913 >= mddev->resync_max - mddev->curr_resync_completed) {
3914 /* Cannot proceed until we've updated the superblock... */
3915 wait_event(conf->wait_for_overlap,
3916 atomic_read(&conf->reshape_stripes) == 0);
3917 mddev->reshape_position = conf->reshape_progress;
3918 mddev->curr_resync_completed = sector_nr;
3919 conf->reshape_checkpoint = jiffies;
3920 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3921 md_wakeup_thread(mddev->thread);
3922 wait_event(mddev->sb_wait,
3923 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3924 || kthread_should_stop());
3925 spin_lock_irq(&conf->device_lock);
3926 conf->reshape_safe = mddev->reshape_position;
3927 spin_unlock_irq(&conf->device_lock);
3928 wake_up(&conf->wait_for_overlap);
3929 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3931 return reshape_sectors;
3934 /* FIXME go_faster isn't used */
3935 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3937 raid5_conf_t *conf = mddev->private;
3938 struct stripe_head *sh;
3939 sector_t max_sector = mddev->dev_sectors;
3940 sector_t sync_blocks;
3941 int still_degraded = 0;
3944 if (sector_nr >= max_sector) {
3945 /* just being told to finish up .. nothing much to do */
3947 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3952 if (mddev->curr_resync < max_sector) /* aborted */
3953 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3955 else /* completed sync */
3957 bitmap_close_sync(mddev->bitmap);
3962 /* Allow raid5_quiesce to complete */
3963 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
3965 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3966 return reshape_request(mddev, sector_nr, skipped);
3968 /* No need to check resync_max as we never do more than one
3969 * stripe, and as resync_max will always be on a chunk boundary,
3970 * if the check in md_do_sync didn't fire, there is no chance
3971 * of overstepping resync_max here
3974 /* if there is too many failed drives and we are trying
3975 * to resync, then assert that we are finished, because there is
3976 * nothing we can do.
3978 if (mddev->degraded >= conf->max_degraded &&
3979 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3980 sector_t rv = mddev->dev_sectors - sector_nr;
3984 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3985 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3986 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3987 /* we can skip this block, and probably more */
3988 sync_blocks /= STRIPE_SECTORS;
3990 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3994 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3996 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
3998 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
3999 /* make sure we don't swamp the stripe cache if someone else
4000 * is trying to get access
4002 schedule_timeout_uninterruptible(1);
4004 /* Need to check if array will still be degraded after recovery/resync
4005 * We don't need to check the 'failed' flag as when that gets set,
4008 for (i = 0; i < conf->raid_disks; i++)
4009 if (conf->disks[i].rdev == NULL)
4012 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4014 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4019 return STRIPE_SECTORS;
4022 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4024 /* We may not be able to submit a whole bio at once as there
4025 * may not be enough stripe_heads available.
4026 * We cannot pre-allocate enough stripe_heads as we may need
4027 * more than exist in the cache (if we allow ever large chunks).
4028 * So we do one stripe head at a time and record in
4029 * ->bi_hw_segments how many have been done.
4031 * We *know* that this entire raid_bio is in one chunk, so
4032 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4034 struct stripe_head *sh;
4036 sector_t sector, logical_sector, last_sector;
4041 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4042 sector = raid5_compute_sector(conf, logical_sector,
4044 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4046 for (; logical_sector < last_sector;
4047 logical_sector += STRIPE_SECTORS,
4048 sector += STRIPE_SECTORS,
4051 if (scnt < raid5_bi_hw_segments(raid_bio))
4052 /* already done this stripe */
4055 sh = get_active_stripe(conf, sector, 0, 1, 0);
4058 /* failed to get a stripe - must wait */
4059 raid5_set_bi_hw_segments(raid_bio, scnt);
4060 conf->retry_read_aligned = raid_bio;
4064 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4065 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4067 raid5_set_bi_hw_segments(raid_bio, scnt);
4068 conf->retry_read_aligned = raid_bio;
4076 spin_lock_irq(&conf->device_lock);
4077 remaining = raid5_dec_bi_phys_segments(raid_bio);
4078 spin_unlock_irq(&conf->device_lock);
4080 bio_endio(raid_bio, 0);
4081 if (atomic_dec_and_test(&conf->active_aligned_reads))
4082 wake_up(&conf->wait_for_stripe);
4088 * This is our raid5 kernel thread.
4090 * We scan the hash table for stripes which can be handled now.
4091 * During the scan, completed stripes are saved for us by the interrupt
4092 * handler, so that they will not have to wait for our next wakeup.
4094 static void raid5d(mddev_t *mddev)
4096 struct stripe_head *sh;
4097 raid5_conf_t *conf = mddev->private;
4099 struct blk_plug plug;
4101 pr_debug("+++ raid5d active\n");
4103 md_check_recovery(mddev);
4105 blk_start_plug(&plug);
4107 spin_lock_irq(&conf->device_lock);
4111 if (atomic_read(&mddev->plug_cnt) == 0 &&
4112 !list_empty(&conf->bitmap_list)) {
4113 /* Now is a good time to flush some bitmap updates */
4115 spin_unlock_irq(&conf->device_lock);
4116 bitmap_unplug(mddev->bitmap);
4117 spin_lock_irq(&conf->device_lock);
4118 conf->seq_write = conf->seq_flush;
4119 activate_bit_delay(conf);
4121 if (atomic_read(&mddev->plug_cnt) == 0)
4122 raid5_activate_delayed(conf);
4124 while ((bio = remove_bio_from_retry(conf))) {
4126 spin_unlock_irq(&conf->device_lock);
4127 ok = retry_aligned_read(conf, bio);
4128 spin_lock_irq(&conf->device_lock);
4134 sh = __get_priority_stripe(conf);
4138 spin_unlock_irq(&conf->device_lock);
4145 spin_lock_irq(&conf->device_lock);
4147 pr_debug("%d stripes handled\n", handled);
4149 spin_unlock_irq(&conf->device_lock);
4151 async_tx_issue_pending_all();
4152 blk_finish_plug(&plug);
4154 pr_debug("--- raid5d inactive\n");
4158 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4160 raid5_conf_t *conf = mddev->private;
4162 return sprintf(page, "%d\n", conf->max_nr_stripes);
4168 raid5_set_cache_size(mddev_t *mddev, int size)
4170 raid5_conf_t *conf = mddev->private;
4173 if (size <= 16 || size > 32768)
4175 while (size < conf->max_nr_stripes) {
4176 if (drop_one_stripe(conf))
4177 conf->max_nr_stripes--;
4181 err = md_allow_write(mddev);
4184 while (size > conf->max_nr_stripes) {
4185 if (grow_one_stripe(conf))
4186 conf->max_nr_stripes++;
4191 EXPORT_SYMBOL(raid5_set_cache_size);
4194 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4196 raid5_conf_t *conf = mddev->private;
4200 if (len >= PAGE_SIZE)
4205 if (strict_strtoul(page, 10, &new))
4207 err = raid5_set_cache_size(mddev, new);
4213 static struct md_sysfs_entry
4214 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4215 raid5_show_stripe_cache_size,
4216 raid5_store_stripe_cache_size);
4219 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4221 raid5_conf_t *conf = mddev->private;
4223 return sprintf(page, "%d\n", conf->bypass_threshold);
4229 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4231 raid5_conf_t *conf = mddev->private;
4233 if (len >= PAGE_SIZE)
4238 if (strict_strtoul(page, 10, &new))
4240 if (new > conf->max_nr_stripes)
4242 conf->bypass_threshold = new;
4246 static struct md_sysfs_entry
4247 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4249 raid5_show_preread_threshold,
4250 raid5_store_preread_threshold);
4253 stripe_cache_active_show(mddev_t *mddev, char *page)
4255 raid5_conf_t *conf = mddev->private;
4257 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4262 static struct md_sysfs_entry
4263 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4265 static struct attribute *raid5_attrs[] = {
4266 &raid5_stripecache_size.attr,
4267 &raid5_stripecache_active.attr,
4268 &raid5_preread_bypass_threshold.attr,
4271 static struct attribute_group raid5_attrs_group = {
4273 .attrs = raid5_attrs,
4277 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4279 raid5_conf_t *conf = mddev->private;
4282 sectors = mddev->dev_sectors;
4284 /* size is defined by the smallest of previous and new size */
4285 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4287 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4288 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4289 return sectors * (raid_disks - conf->max_degraded);
4292 static void raid5_free_percpu(raid5_conf_t *conf)
4294 struct raid5_percpu *percpu;
4301 for_each_possible_cpu(cpu) {
4302 percpu = per_cpu_ptr(conf->percpu, cpu);
4303 safe_put_page(percpu->spare_page);
4304 kfree(percpu->scribble);
4306 #ifdef CONFIG_HOTPLUG_CPU
4307 unregister_cpu_notifier(&conf->cpu_notify);
4311 free_percpu(conf->percpu);
4314 static void free_conf(raid5_conf_t *conf)
4316 shrink_stripes(conf);
4317 raid5_free_percpu(conf);
4319 kfree(conf->stripe_hashtbl);
4323 #ifdef CONFIG_HOTPLUG_CPU
4324 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4327 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4328 long cpu = (long)hcpu;
4329 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4332 case CPU_UP_PREPARE:
4333 case CPU_UP_PREPARE_FROZEN:
4334 if (conf->level == 6 && !percpu->spare_page)
4335 percpu->spare_page = alloc_page(GFP_KERNEL);
4336 if (!percpu->scribble)
4337 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4339 if (!percpu->scribble ||
4340 (conf->level == 6 && !percpu->spare_page)) {
4341 safe_put_page(percpu->spare_page);
4342 kfree(percpu->scribble);
4343 pr_err("%s: failed memory allocation for cpu%ld\n",
4345 return notifier_from_errno(-ENOMEM);
4349 case CPU_DEAD_FROZEN:
4350 safe_put_page(percpu->spare_page);
4351 kfree(percpu->scribble);
4352 percpu->spare_page = NULL;
4353 percpu->scribble = NULL;
4362 static int raid5_alloc_percpu(raid5_conf_t *conf)
4365 struct page *spare_page;
4366 struct raid5_percpu __percpu *allcpus;
4370 allcpus = alloc_percpu(struct raid5_percpu);
4373 conf->percpu = allcpus;
4377 for_each_present_cpu(cpu) {
4378 if (conf->level == 6) {
4379 spare_page = alloc_page(GFP_KERNEL);
4384 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4386 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4391 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4393 #ifdef CONFIG_HOTPLUG_CPU
4394 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4395 conf->cpu_notify.priority = 0;
4397 err = register_cpu_notifier(&conf->cpu_notify);
4404 static raid5_conf_t *setup_conf(mddev_t *mddev)
4407 int raid_disk, memory, max_disks;
4409 struct disk_info *disk;
4411 if (mddev->new_level != 5
4412 && mddev->new_level != 4
4413 && mddev->new_level != 6) {
4414 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4415 mdname(mddev), mddev->new_level);
4416 return ERR_PTR(-EIO);
4418 if ((mddev->new_level == 5
4419 && !algorithm_valid_raid5(mddev->new_layout)) ||
4420 (mddev->new_level == 6
4421 && !algorithm_valid_raid6(mddev->new_layout))) {
4422 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4423 mdname(mddev), mddev->new_layout);
4424 return ERR_PTR(-EIO);
4426 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4427 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4428 mdname(mddev), mddev->raid_disks);
4429 return ERR_PTR(-EINVAL);
4432 if (!mddev->new_chunk_sectors ||
4433 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4434 !is_power_of_2(mddev->new_chunk_sectors)) {
4435 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4436 mdname(mddev), mddev->new_chunk_sectors << 9);
4437 return ERR_PTR(-EINVAL);
4440 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4443 spin_lock_init(&conf->device_lock);
4444 init_waitqueue_head(&conf->wait_for_stripe);
4445 init_waitqueue_head(&conf->wait_for_overlap);
4446 INIT_LIST_HEAD(&conf->handle_list);
4447 INIT_LIST_HEAD(&conf->hold_list);
4448 INIT_LIST_HEAD(&conf->delayed_list);
4449 INIT_LIST_HEAD(&conf->bitmap_list);
4450 INIT_LIST_HEAD(&conf->inactive_list);
4451 atomic_set(&conf->active_stripes, 0);
4452 atomic_set(&conf->preread_active_stripes, 0);
4453 atomic_set(&conf->active_aligned_reads, 0);
4454 conf->bypass_threshold = BYPASS_THRESHOLD;
4456 conf->raid_disks = mddev->raid_disks;
4457 if (mddev->reshape_position == MaxSector)
4458 conf->previous_raid_disks = mddev->raid_disks;
4460 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4461 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4462 conf->scribble_len = scribble_len(max_disks);
4464 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4469 conf->mddev = mddev;
4471 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4474 conf->level = mddev->new_level;
4475 if (raid5_alloc_percpu(conf) != 0)
4478 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4480 list_for_each_entry(rdev, &mddev->disks, same_set) {
4481 raid_disk = rdev->raid_disk;
4482 if (raid_disk >= max_disks
4485 disk = conf->disks + raid_disk;
4489 if (test_bit(In_sync, &rdev->flags)) {
4490 char b[BDEVNAME_SIZE];
4491 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4493 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4494 } else if (rdev->saved_raid_disk != raid_disk)
4495 /* Cannot rely on bitmap to complete recovery */
4499 conf->chunk_sectors = mddev->new_chunk_sectors;
4500 conf->level = mddev->new_level;
4501 if (conf->level == 6)
4502 conf->max_degraded = 2;
4504 conf->max_degraded = 1;
4505 conf->algorithm = mddev->new_layout;
4506 conf->max_nr_stripes = NR_STRIPES;
4507 conf->reshape_progress = mddev->reshape_position;
4508 if (conf->reshape_progress != MaxSector) {
4509 conf->prev_chunk_sectors = mddev->chunk_sectors;
4510 conf->prev_algo = mddev->layout;
4513 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4514 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4515 if (grow_stripes(conf, conf->max_nr_stripes)) {
4517 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4518 mdname(mddev), memory);
4521 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4522 mdname(mddev), memory);
4524 conf->thread = md_register_thread(raid5d, mddev, NULL);
4525 if (!conf->thread) {
4527 "md/raid:%s: couldn't allocate thread.\n",
4537 return ERR_PTR(-EIO);
4539 return ERR_PTR(-ENOMEM);
4543 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4546 case ALGORITHM_PARITY_0:
4547 if (raid_disk < max_degraded)
4550 case ALGORITHM_PARITY_N:
4551 if (raid_disk >= raid_disks - max_degraded)
4554 case ALGORITHM_PARITY_0_6:
4555 if (raid_disk == 0 ||
4556 raid_disk == raid_disks - 1)
4559 case ALGORITHM_LEFT_ASYMMETRIC_6:
4560 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4561 case ALGORITHM_LEFT_SYMMETRIC_6:
4562 case ALGORITHM_RIGHT_SYMMETRIC_6:
4563 if (raid_disk == raid_disks - 1)
4569 static int run(mddev_t *mddev)
4572 int working_disks = 0;
4573 int dirty_parity_disks = 0;
4575 sector_t reshape_offset = 0;
4577 if (mddev->recovery_cp != MaxSector)
4578 printk(KERN_NOTICE "md/raid:%s: not clean"
4579 " -- starting background reconstruction\n",
4581 if (mddev->reshape_position != MaxSector) {
4582 /* Check that we can continue the reshape.
4583 * Currently only disks can change, it must
4584 * increase, and we must be past the point where
4585 * a stripe over-writes itself
4587 sector_t here_new, here_old;
4589 int max_degraded = (mddev->level == 6 ? 2 : 1);
4591 if (mddev->new_level != mddev->level) {
4592 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4593 "required - aborting.\n",
4597 old_disks = mddev->raid_disks - mddev->delta_disks;
4598 /* reshape_position must be on a new-stripe boundary, and one
4599 * further up in new geometry must map after here in old
4602 here_new = mddev->reshape_position;
4603 if (sector_div(here_new, mddev->new_chunk_sectors *
4604 (mddev->raid_disks - max_degraded))) {
4605 printk(KERN_ERR "md/raid:%s: reshape_position not "
4606 "on a stripe boundary\n", mdname(mddev));
4609 reshape_offset = here_new * mddev->new_chunk_sectors;
4610 /* here_new is the stripe we will write to */
4611 here_old = mddev->reshape_position;
4612 sector_div(here_old, mddev->chunk_sectors *
4613 (old_disks-max_degraded));
4614 /* here_old is the first stripe that we might need to read
4616 if (mddev->delta_disks == 0) {
4617 /* We cannot be sure it is safe to start an in-place
4618 * reshape. It is only safe if user-space if monitoring
4619 * and taking constant backups.
4620 * mdadm always starts a situation like this in
4621 * readonly mode so it can take control before
4622 * allowing any writes. So just check for that.
4624 if ((here_new * mddev->new_chunk_sectors !=
4625 here_old * mddev->chunk_sectors) ||
4627 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4628 " in read-only mode - aborting\n",
4632 } else if (mddev->delta_disks < 0
4633 ? (here_new * mddev->new_chunk_sectors <=
4634 here_old * mddev->chunk_sectors)
4635 : (here_new * mddev->new_chunk_sectors >=
4636 here_old * mddev->chunk_sectors)) {
4637 /* Reading from the same stripe as writing to - bad */
4638 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4639 "auto-recovery - aborting.\n",
4643 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4645 /* OK, we should be able to continue; */
4647 BUG_ON(mddev->level != mddev->new_level);
4648 BUG_ON(mddev->layout != mddev->new_layout);
4649 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4650 BUG_ON(mddev->delta_disks != 0);
4653 if (mddev->private == NULL)
4654 conf = setup_conf(mddev);
4656 conf = mddev->private;
4659 return PTR_ERR(conf);
4661 mddev->thread = conf->thread;
4662 conf->thread = NULL;
4663 mddev->private = conf;
4666 * 0 for a fully functional array, 1 or 2 for a degraded array.
4668 list_for_each_entry(rdev, &mddev->disks, same_set) {
4669 if (rdev->raid_disk < 0)
4671 if (test_bit(In_sync, &rdev->flags)) {
4675 /* This disc is not fully in-sync. However if it
4676 * just stored parity (beyond the recovery_offset),
4677 * when we don't need to be concerned about the
4678 * array being dirty.
4679 * When reshape goes 'backwards', we never have
4680 * partially completed devices, so we only need
4681 * to worry about reshape going forwards.
4683 /* Hack because v0.91 doesn't store recovery_offset properly. */
4684 if (mddev->major_version == 0 &&
4685 mddev->minor_version > 90)
4686 rdev->recovery_offset = reshape_offset;
4688 if (rdev->recovery_offset < reshape_offset) {
4689 /* We need to check old and new layout */
4690 if (!only_parity(rdev->raid_disk,
4693 conf->max_degraded))
4696 if (!only_parity(rdev->raid_disk,
4698 conf->previous_raid_disks,
4699 conf->max_degraded))
4701 dirty_parity_disks++;
4704 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4707 if (has_failed(conf)) {
4708 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4709 " (%d/%d failed)\n",
4710 mdname(mddev), mddev->degraded, conf->raid_disks);
4714 /* device size must be a multiple of chunk size */
4715 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4716 mddev->resync_max_sectors = mddev->dev_sectors;
4718 if (mddev->degraded > dirty_parity_disks &&
4719 mddev->recovery_cp != MaxSector) {
4720 if (mddev->ok_start_degraded)
4722 "md/raid:%s: starting dirty degraded array"
4723 " - data corruption possible.\n",
4727 "md/raid:%s: cannot start dirty degraded array.\n",
4733 if (mddev->degraded == 0)
4734 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4735 " devices, algorithm %d\n", mdname(mddev), conf->level,
4736 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4739 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4740 " out of %d devices, algorithm %d\n",
4741 mdname(mddev), conf->level,
4742 mddev->raid_disks - mddev->degraded,
4743 mddev->raid_disks, mddev->new_layout);
4745 print_raid5_conf(conf);
4747 if (conf->reshape_progress != MaxSector) {
4748 conf->reshape_safe = conf->reshape_progress;
4749 atomic_set(&conf->reshape_stripes, 0);
4750 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4751 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4752 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4753 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4754 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4759 /* Ok, everything is just fine now */
4760 if (mddev->to_remove == &raid5_attrs_group)
4761 mddev->to_remove = NULL;
4762 else if (mddev->kobj.sd &&
4763 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4765 "raid5: failed to create sysfs attributes for %s\n",
4767 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4771 /* read-ahead size must cover two whole stripes, which
4772 * is 2 * (datadisks) * chunksize where 'n' is the
4773 * number of raid devices
4775 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4776 int stripe = data_disks *
4777 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4778 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4779 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4781 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4783 mddev->queue->backing_dev_info.congested_data = mddev;
4784 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4786 chunk_size = mddev->chunk_sectors << 9;
4787 blk_queue_io_min(mddev->queue, chunk_size);
4788 blk_queue_io_opt(mddev->queue, chunk_size *
4789 (conf->raid_disks - conf->max_degraded));
4791 list_for_each_entry(rdev, &mddev->disks, same_set)
4792 disk_stack_limits(mddev->gendisk, rdev->bdev,
4793 rdev->data_offset << 9);
4798 md_unregister_thread(mddev->thread);
4799 mddev->thread = NULL;
4801 print_raid5_conf(conf);
4804 mddev->private = NULL;
4805 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
4809 static int stop(mddev_t *mddev)
4811 raid5_conf_t *conf = mddev->private;
4813 md_unregister_thread(mddev->thread);
4814 mddev->thread = NULL;
4816 mddev->queue->backing_dev_info.congested_fn = NULL;
4818 mddev->private = NULL;
4819 mddev->to_remove = &raid5_attrs_group;
4824 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4828 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4829 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4830 seq_printf(seq, "sh %llu, count %d.\n",
4831 (unsigned long long)sh->sector, atomic_read(&sh->count));
4832 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4833 for (i = 0; i < sh->disks; i++) {
4834 seq_printf(seq, "(cache%d: %p %ld) ",
4835 i, sh->dev[i].page, sh->dev[i].flags);
4837 seq_printf(seq, "\n");
4840 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4842 struct stripe_head *sh;
4843 struct hlist_node *hn;
4846 spin_lock_irq(&conf->device_lock);
4847 for (i = 0; i < NR_HASH; i++) {
4848 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4849 if (sh->raid_conf != conf)
4854 spin_unlock_irq(&conf->device_lock);
4858 static void status(struct seq_file *seq, mddev_t *mddev)
4860 raid5_conf_t *conf = mddev->private;
4863 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4864 mddev->chunk_sectors / 2, mddev->layout);
4865 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4866 for (i = 0; i < conf->raid_disks; i++)
4867 seq_printf (seq, "%s",
4868 conf->disks[i].rdev &&
4869 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4870 seq_printf (seq, "]");
4872 seq_printf (seq, "\n");
4873 printall(seq, conf);
4877 static void print_raid5_conf (raid5_conf_t *conf)
4880 struct disk_info *tmp;
4882 printk(KERN_DEBUG "RAID conf printout:\n");
4884 printk("(conf==NULL)\n");
4887 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
4889 conf->raid_disks - conf->mddev->degraded);
4891 for (i = 0; i < conf->raid_disks; i++) {
4892 char b[BDEVNAME_SIZE];
4893 tmp = conf->disks + i;
4895 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
4896 i, !test_bit(Faulty, &tmp->rdev->flags),
4897 bdevname(tmp->rdev->bdev, b));
4901 static int raid5_spare_active(mddev_t *mddev)
4904 raid5_conf_t *conf = mddev->private;
4905 struct disk_info *tmp;
4907 unsigned long flags;
4909 for (i = 0; i < conf->raid_disks; i++) {
4910 tmp = conf->disks + i;
4912 && tmp->rdev->recovery_offset == MaxSector
4913 && !test_bit(Faulty, &tmp->rdev->flags)
4914 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4916 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
4919 spin_lock_irqsave(&conf->device_lock, flags);
4920 mddev->degraded -= count;
4921 spin_unlock_irqrestore(&conf->device_lock, flags);
4922 print_raid5_conf(conf);
4926 static int raid5_remove_disk(mddev_t *mddev, int number)
4928 raid5_conf_t *conf = mddev->private;
4931 struct disk_info *p = conf->disks + number;
4933 print_raid5_conf(conf);
4936 if (number >= conf->raid_disks &&
4937 conf->reshape_progress == MaxSector)
4938 clear_bit(In_sync, &rdev->flags);
4940 if (test_bit(In_sync, &rdev->flags) ||
4941 atomic_read(&rdev->nr_pending)) {
4945 /* Only remove non-faulty devices if recovery
4948 if (!test_bit(Faulty, &rdev->flags) &&
4949 !has_failed(conf) &&
4950 number < conf->raid_disks) {
4956 if (atomic_read(&rdev->nr_pending)) {
4957 /* lost the race, try later */
4964 print_raid5_conf(conf);
4968 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4970 raid5_conf_t *conf = mddev->private;
4973 struct disk_info *p;
4975 int last = conf->raid_disks - 1;
4977 if (has_failed(conf))
4978 /* no point adding a device */
4981 if (rdev->raid_disk >= 0)
4982 first = last = rdev->raid_disk;
4985 * find the disk ... but prefer rdev->saved_raid_disk
4988 if (rdev->saved_raid_disk >= 0 &&
4989 rdev->saved_raid_disk >= first &&
4990 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4991 disk = rdev->saved_raid_disk;
4994 for ( ; disk <= last ; disk++)
4995 if ((p=conf->disks + disk)->rdev == NULL) {
4996 clear_bit(In_sync, &rdev->flags);
4997 rdev->raid_disk = disk;
4999 if (rdev->saved_raid_disk != disk)
5001 rcu_assign_pointer(p->rdev, rdev);
5004 print_raid5_conf(conf);
5008 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5010 /* no resync is happening, and there is enough space
5011 * on all devices, so we can resize.
5012 * We need to make sure resync covers any new space.
5013 * If the array is shrinking we should possibly wait until
5014 * any io in the removed space completes, but it hardly seems
5017 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5018 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5019 mddev->raid_disks));
5020 if (mddev->array_sectors >
5021 raid5_size(mddev, sectors, mddev->raid_disks))
5023 set_capacity(mddev->gendisk, mddev->array_sectors);
5024 revalidate_disk(mddev->gendisk);
5025 if (sectors > mddev->dev_sectors &&
5026 mddev->recovery_cp > mddev->dev_sectors) {
5027 mddev->recovery_cp = mddev->dev_sectors;
5028 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5030 mddev->dev_sectors = sectors;
5031 mddev->resync_max_sectors = sectors;
5035 static int check_stripe_cache(mddev_t *mddev)
5037 /* Can only proceed if there are plenty of stripe_heads.
5038 * We need a minimum of one full stripe,, and for sensible progress
5039 * it is best to have about 4 times that.
5040 * If we require 4 times, then the default 256 4K stripe_heads will
5041 * allow for chunk sizes up to 256K, which is probably OK.
5042 * If the chunk size is greater, user-space should request more
5043 * stripe_heads first.
5045 raid5_conf_t *conf = mddev->private;
5046 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5047 > conf->max_nr_stripes ||
5048 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5049 > conf->max_nr_stripes) {
5050 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5052 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5059 static int check_reshape(mddev_t *mddev)
5061 raid5_conf_t *conf = mddev->private;
5063 if (mddev->delta_disks == 0 &&
5064 mddev->new_layout == mddev->layout &&
5065 mddev->new_chunk_sectors == mddev->chunk_sectors)
5066 return 0; /* nothing to do */
5068 /* Cannot grow a bitmap yet */
5070 if (has_failed(conf))
5072 if (mddev->delta_disks < 0) {
5073 /* We might be able to shrink, but the devices must
5074 * be made bigger first.
5075 * For raid6, 4 is the minimum size.
5076 * Otherwise 2 is the minimum
5079 if (mddev->level == 6)
5081 if (mddev->raid_disks + mddev->delta_disks < min)
5085 if (!check_stripe_cache(mddev))
5088 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5091 static int raid5_start_reshape(mddev_t *mddev)
5093 raid5_conf_t *conf = mddev->private;
5096 unsigned long flags;
5098 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5101 if (!check_stripe_cache(mddev))
5104 list_for_each_entry(rdev, &mddev->disks, same_set)
5105 if (!test_bit(In_sync, &rdev->flags)
5106 && !test_bit(Faulty, &rdev->flags))
5109 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5110 /* Not enough devices even to make a degraded array
5115 /* Refuse to reduce size of the array. Any reductions in
5116 * array size must be through explicit setting of array_size
5119 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5120 < mddev->array_sectors) {
5121 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5122 "before number of disks\n", mdname(mddev));
5126 atomic_set(&conf->reshape_stripes, 0);
5127 spin_lock_irq(&conf->device_lock);
5128 conf->previous_raid_disks = conf->raid_disks;
5129 conf->raid_disks += mddev->delta_disks;
5130 conf->prev_chunk_sectors = conf->chunk_sectors;
5131 conf->chunk_sectors = mddev->new_chunk_sectors;
5132 conf->prev_algo = conf->algorithm;
5133 conf->algorithm = mddev->new_layout;
5134 if (mddev->delta_disks < 0)
5135 conf->reshape_progress = raid5_size(mddev, 0, 0);
5137 conf->reshape_progress = 0;
5138 conf->reshape_safe = conf->reshape_progress;
5140 spin_unlock_irq(&conf->device_lock);
5142 /* Add some new drives, as many as will fit.
5143 * We know there are enough to make the newly sized array work.
5144 * Don't add devices if we are reducing the number of
5145 * devices in the array. This is because it is not possible
5146 * to correctly record the "partially reconstructed" state of
5147 * such devices during the reshape and confusion could result.
5149 if (mddev->delta_disks >= 0) {
5150 int added_devices = 0;
5151 list_for_each_entry(rdev, &mddev->disks, same_set)
5152 if (rdev->raid_disk < 0 &&
5153 !test_bit(Faulty, &rdev->flags)) {
5154 if (raid5_add_disk(mddev, rdev) == 0) {
5156 >= conf->previous_raid_disks) {
5157 set_bit(In_sync, &rdev->flags);
5160 rdev->recovery_offset = 0;
5162 if (sysfs_link_rdev(mddev, rdev))
5163 /* Failure here is OK */;
5165 } else if (rdev->raid_disk >= conf->previous_raid_disks
5166 && !test_bit(Faulty, &rdev->flags)) {
5167 /* This is a spare that was manually added */
5168 set_bit(In_sync, &rdev->flags);
5172 /* When a reshape changes the number of devices,
5173 * ->degraded is measured against the larger of the
5174 * pre and post number of devices.
5176 spin_lock_irqsave(&conf->device_lock, flags);
5177 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5179 spin_unlock_irqrestore(&conf->device_lock, flags);
5181 mddev->raid_disks = conf->raid_disks;
5182 mddev->reshape_position = conf->reshape_progress;
5183 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5185 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5186 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5187 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5188 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5189 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5191 if (!mddev->sync_thread) {
5192 mddev->recovery = 0;
5193 spin_lock_irq(&conf->device_lock);
5194 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5195 conf->reshape_progress = MaxSector;
5196 spin_unlock_irq(&conf->device_lock);
5199 conf->reshape_checkpoint = jiffies;
5200 md_wakeup_thread(mddev->sync_thread);
5201 md_new_event(mddev);
5205 /* This is called from the reshape thread and should make any
5206 * changes needed in 'conf'
5208 static void end_reshape(raid5_conf_t *conf)
5211 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5213 spin_lock_irq(&conf->device_lock);
5214 conf->previous_raid_disks = conf->raid_disks;
5215 conf->reshape_progress = MaxSector;
5216 spin_unlock_irq(&conf->device_lock);
5217 wake_up(&conf->wait_for_overlap);
5219 /* read-ahead size must cover two whole stripes, which is
5220 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5222 if (conf->mddev->queue) {
5223 int data_disks = conf->raid_disks - conf->max_degraded;
5224 int stripe = data_disks * ((conf->chunk_sectors << 9)
5226 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5227 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5232 /* This is called from the raid5d thread with mddev_lock held.
5233 * It makes config changes to the device.
5235 static void raid5_finish_reshape(mddev_t *mddev)
5237 raid5_conf_t *conf = mddev->private;
5239 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5241 if (mddev->delta_disks > 0) {
5242 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5243 set_capacity(mddev->gendisk, mddev->array_sectors);
5244 revalidate_disk(mddev->gendisk);
5247 mddev->degraded = conf->raid_disks;
5248 for (d = 0; d < conf->raid_disks ; d++)
5249 if (conf->disks[d].rdev &&
5251 &conf->disks[d].rdev->flags))
5253 for (d = conf->raid_disks ;
5254 d < conf->raid_disks - mddev->delta_disks;
5256 mdk_rdev_t *rdev = conf->disks[d].rdev;
5257 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5258 sysfs_unlink_rdev(mddev, rdev);
5259 rdev->raid_disk = -1;
5263 mddev->layout = conf->algorithm;
5264 mddev->chunk_sectors = conf->chunk_sectors;
5265 mddev->reshape_position = MaxSector;
5266 mddev->delta_disks = 0;
5270 static void raid5_quiesce(mddev_t *mddev, int state)
5272 raid5_conf_t *conf = mddev->private;
5275 case 2: /* resume for a suspend */
5276 wake_up(&conf->wait_for_overlap);
5279 case 1: /* stop all writes */
5280 spin_lock_irq(&conf->device_lock);
5281 /* '2' tells resync/reshape to pause so that all
5282 * active stripes can drain
5285 wait_event_lock_irq(conf->wait_for_stripe,
5286 atomic_read(&conf->active_stripes) == 0 &&
5287 atomic_read(&conf->active_aligned_reads) == 0,
5288 conf->device_lock, /* nothing */);
5290 spin_unlock_irq(&conf->device_lock);
5291 /* allow reshape to continue */
5292 wake_up(&conf->wait_for_overlap);
5295 case 0: /* re-enable writes */
5296 spin_lock_irq(&conf->device_lock);
5298 wake_up(&conf->wait_for_stripe);
5299 wake_up(&conf->wait_for_overlap);
5300 spin_unlock_irq(&conf->device_lock);
5306 static void *raid45_takeover_raid0(mddev_t *mddev, int level)
5308 struct raid0_private_data *raid0_priv = mddev->private;
5311 /* for raid0 takeover only one zone is supported */
5312 if (raid0_priv->nr_strip_zones > 1) {
5313 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5315 return ERR_PTR(-EINVAL);
5318 sectors = raid0_priv->strip_zone[0].zone_end;
5319 sector_div(sectors, raid0_priv->strip_zone[0].nb_dev);
5320 mddev->dev_sectors = sectors;
5321 mddev->new_level = level;
5322 mddev->new_layout = ALGORITHM_PARITY_N;
5323 mddev->new_chunk_sectors = mddev->chunk_sectors;
5324 mddev->raid_disks += 1;
5325 mddev->delta_disks = 1;
5326 /* make sure it will be not marked as dirty */
5327 mddev->recovery_cp = MaxSector;
5329 return setup_conf(mddev);
5333 static void *raid5_takeover_raid1(mddev_t *mddev)
5337 if (mddev->raid_disks != 2 ||
5338 mddev->degraded > 1)
5339 return ERR_PTR(-EINVAL);
5341 /* Should check if there are write-behind devices? */
5343 chunksect = 64*2; /* 64K by default */
5345 /* The array must be an exact multiple of chunksize */
5346 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5349 if ((chunksect<<9) < STRIPE_SIZE)
5350 /* array size does not allow a suitable chunk size */
5351 return ERR_PTR(-EINVAL);
5353 mddev->new_level = 5;
5354 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5355 mddev->new_chunk_sectors = chunksect;
5357 return setup_conf(mddev);
5360 static void *raid5_takeover_raid6(mddev_t *mddev)
5364 switch (mddev->layout) {
5365 case ALGORITHM_LEFT_ASYMMETRIC_6:
5366 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5368 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5369 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5371 case ALGORITHM_LEFT_SYMMETRIC_6:
5372 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5374 case ALGORITHM_RIGHT_SYMMETRIC_6:
5375 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5377 case ALGORITHM_PARITY_0_6:
5378 new_layout = ALGORITHM_PARITY_0;
5380 case ALGORITHM_PARITY_N:
5381 new_layout = ALGORITHM_PARITY_N;
5384 return ERR_PTR(-EINVAL);
5386 mddev->new_level = 5;
5387 mddev->new_layout = new_layout;
5388 mddev->delta_disks = -1;
5389 mddev->raid_disks -= 1;
5390 return setup_conf(mddev);
5394 static int raid5_check_reshape(mddev_t *mddev)
5396 /* For a 2-drive array, the layout and chunk size can be changed
5397 * immediately as not restriping is needed.
5398 * For larger arrays we record the new value - after validation
5399 * to be used by a reshape pass.
5401 raid5_conf_t *conf = mddev->private;
5402 int new_chunk = mddev->new_chunk_sectors;
5404 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5406 if (new_chunk > 0) {
5407 if (!is_power_of_2(new_chunk))
5409 if (new_chunk < (PAGE_SIZE>>9))
5411 if (mddev->array_sectors & (new_chunk-1))
5412 /* not factor of array size */
5416 /* They look valid */
5418 if (mddev->raid_disks == 2) {
5419 /* can make the change immediately */
5420 if (mddev->new_layout >= 0) {
5421 conf->algorithm = mddev->new_layout;
5422 mddev->layout = mddev->new_layout;
5424 if (new_chunk > 0) {
5425 conf->chunk_sectors = new_chunk ;
5426 mddev->chunk_sectors = new_chunk;
5428 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5429 md_wakeup_thread(mddev->thread);
5431 return check_reshape(mddev);
5434 static int raid6_check_reshape(mddev_t *mddev)
5436 int new_chunk = mddev->new_chunk_sectors;
5438 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5440 if (new_chunk > 0) {
5441 if (!is_power_of_2(new_chunk))
5443 if (new_chunk < (PAGE_SIZE >> 9))
5445 if (mddev->array_sectors & (new_chunk-1))
5446 /* not factor of array size */
5450 /* They look valid */
5451 return check_reshape(mddev);
5454 static void *raid5_takeover(mddev_t *mddev)
5456 /* raid5 can take over:
5457 * raid0 - if there is only one strip zone - make it a raid4 layout
5458 * raid1 - if there are two drives. We need to know the chunk size
5459 * raid4 - trivial - just use a raid4 layout.
5460 * raid6 - Providing it is a *_6 layout
5462 if (mddev->level == 0)
5463 return raid45_takeover_raid0(mddev, 5);
5464 if (mddev->level == 1)
5465 return raid5_takeover_raid1(mddev);
5466 if (mddev->level == 4) {
5467 mddev->new_layout = ALGORITHM_PARITY_N;
5468 mddev->new_level = 5;
5469 return setup_conf(mddev);
5471 if (mddev->level == 6)
5472 return raid5_takeover_raid6(mddev);
5474 return ERR_PTR(-EINVAL);
5477 static void *raid4_takeover(mddev_t *mddev)
5479 /* raid4 can take over:
5480 * raid0 - if there is only one strip zone
5481 * raid5 - if layout is right
5483 if (mddev->level == 0)
5484 return raid45_takeover_raid0(mddev, 4);
5485 if (mddev->level == 5 &&
5486 mddev->layout == ALGORITHM_PARITY_N) {
5487 mddev->new_layout = 0;
5488 mddev->new_level = 4;
5489 return setup_conf(mddev);
5491 return ERR_PTR(-EINVAL);
5494 static struct mdk_personality raid5_personality;
5496 static void *raid6_takeover(mddev_t *mddev)
5498 /* Currently can only take over a raid5. We map the
5499 * personality to an equivalent raid6 personality
5500 * with the Q block at the end.
5504 if (mddev->pers != &raid5_personality)
5505 return ERR_PTR(-EINVAL);
5506 if (mddev->degraded > 1)
5507 return ERR_PTR(-EINVAL);
5508 if (mddev->raid_disks > 253)
5509 return ERR_PTR(-EINVAL);
5510 if (mddev->raid_disks < 3)
5511 return ERR_PTR(-EINVAL);
5513 switch (mddev->layout) {
5514 case ALGORITHM_LEFT_ASYMMETRIC:
5515 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5517 case ALGORITHM_RIGHT_ASYMMETRIC:
5518 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5520 case ALGORITHM_LEFT_SYMMETRIC:
5521 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5523 case ALGORITHM_RIGHT_SYMMETRIC:
5524 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5526 case ALGORITHM_PARITY_0:
5527 new_layout = ALGORITHM_PARITY_0_6;
5529 case ALGORITHM_PARITY_N:
5530 new_layout = ALGORITHM_PARITY_N;
5533 return ERR_PTR(-EINVAL);
5535 mddev->new_level = 6;
5536 mddev->new_layout = new_layout;
5537 mddev->delta_disks = 1;
5538 mddev->raid_disks += 1;
5539 return setup_conf(mddev);
5543 static struct mdk_personality raid6_personality =
5547 .owner = THIS_MODULE,
5548 .make_request = make_request,
5552 .error_handler = error,
5553 .hot_add_disk = raid5_add_disk,
5554 .hot_remove_disk= raid5_remove_disk,
5555 .spare_active = raid5_spare_active,
5556 .sync_request = sync_request,
5557 .resize = raid5_resize,
5559 .check_reshape = raid6_check_reshape,
5560 .start_reshape = raid5_start_reshape,
5561 .finish_reshape = raid5_finish_reshape,
5562 .quiesce = raid5_quiesce,
5563 .takeover = raid6_takeover,
5565 static struct mdk_personality raid5_personality =
5569 .owner = THIS_MODULE,
5570 .make_request = make_request,
5574 .error_handler = error,
5575 .hot_add_disk = raid5_add_disk,
5576 .hot_remove_disk= raid5_remove_disk,
5577 .spare_active = raid5_spare_active,
5578 .sync_request = sync_request,
5579 .resize = raid5_resize,
5581 .check_reshape = raid5_check_reshape,
5582 .start_reshape = raid5_start_reshape,
5583 .finish_reshape = raid5_finish_reshape,
5584 .quiesce = raid5_quiesce,
5585 .takeover = raid5_takeover,
5588 static struct mdk_personality raid4_personality =
5592 .owner = THIS_MODULE,
5593 .make_request = make_request,
5597 .error_handler = error,
5598 .hot_add_disk = raid5_add_disk,
5599 .hot_remove_disk= raid5_remove_disk,
5600 .spare_active = raid5_spare_active,
5601 .sync_request = sync_request,
5602 .resize = raid5_resize,
5604 .check_reshape = raid5_check_reshape,
5605 .start_reshape = raid5_start_reshape,
5606 .finish_reshape = raid5_finish_reshape,
5607 .quiesce = raid5_quiesce,
5608 .takeover = raid4_takeover,
5611 static int __init raid5_init(void)
5613 register_md_personality(&raid6_personality);
5614 register_md_personality(&raid5_personality);
5615 register_md_personality(&raid4_personality);
5619 static void raid5_exit(void)
5621 unregister_md_personality(&raid6_personality);
5622 unregister_md_personality(&raid5_personality);
5623 unregister_md_personality(&raid4_personality);
5626 module_init(raid5_init);
5627 module_exit(raid5_exit);
5628 MODULE_LICENSE("GPL");
5629 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5630 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5631 MODULE_ALIAS("md-raid5");
5632 MODULE_ALIAS("md-raid4");
5633 MODULE_ALIAS("md-level-5");
5634 MODULE_ALIAS("md-level-4");
5635 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5636 MODULE_ALIAS("md-raid6");
5637 MODULE_ALIAS("md-level-6");
5639 /* This used to be two separate modules, they were: */
5640 MODULE_ALIAS("raid5");
5641 MODULE_ALIAS("raid6");