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>
63 #define NR_STRIPES 256
64 #define STRIPE_SIZE PAGE_SIZE
65 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
66 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
67 #define IO_THRESHOLD 1
68 #define BYPASS_THRESHOLD 1
69 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
70 #define HASH_MASK (NR_HASH - 1)
72 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
74 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
75 * order without overlap. There may be several bio's per stripe+device, and
76 * a bio could span several devices.
77 * When walking this list for a particular stripe+device, we must never proceed
78 * beyond a bio that extends past this device, as the next bio might no longer
80 * This macro is used to determine the 'next' bio in the list, given the sector
81 * of the current stripe+device
83 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
85 * The following can be used to debug the driver
87 #define RAID5_PARANOIA 1
88 #if RAID5_PARANOIA && defined(CONFIG_SMP)
89 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
91 # define CHECK_DEVLOCK()
99 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
102 * We maintain a biased count of active stripes in the bottom 16 bits of
103 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
105 static inline int raid5_bi_phys_segments(struct bio *bio)
107 return bio->bi_phys_segments & 0xffff;
110 static inline int raid5_bi_hw_segments(struct bio *bio)
112 return (bio->bi_phys_segments >> 16) & 0xffff;
115 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
117 --bio->bi_phys_segments;
118 return raid5_bi_phys_segments(bio);
121 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
123 unsigned short val = raid5_bi_hw_segments(bio);
126 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
130 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
132 bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
135 /* Find first data disk in a raid6 stripe */
136 static inline int raid6_d0(struct stripe_head *sh)
139 /* ddf always start from first device */
141 /* md starts just after Q block */
142 if (sh->qd_idx == sh->disks - 1)
145 return sh->qd_idx + 1;
147 static inline int raid6_next_disk(int disk, int raid_disks)
150 return (disk < raid_disks) ? disk : 0;
153 /* When walking through the disks in a raid5, starting at raid6_d0,
154 * We need to map each disk to a 'slot', where the data disks are slot
155 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
156 * is raid_disks-1. This help does that mapping.
158 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
159 int *count, int syndrome_disks)
165 if (idx == sh->pd_idx)
166 return syndrome_disks;
167 if (idx == sh->qd_idx)
168 return syndrome_disks + 1;
174 static void return_io(struct bio *return_bi)
176 struct bio *bi = return_bi;
179 return_bi = bi->bi_next;
187 static void print_raid5_conf (raid5_conf_t *conf);
189 static int stripe_operations_active(struct stripe_head *sh)
191 return sh->check_state || sh->reconstruct_state ||
192 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
193 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
196 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
198 if (atomic_dec_and_test(&sh->count)) {
199 BUG_ON(!list_empty(&sh->lru));
200 BUG_ON(atomic_read(&conf->active_stripes)==0);
201 if (test_bit(STRIPE_HANDLE, &sh->state)) {
202 if (test_bit(STRIPE_DELAYED, &sh->state))
203 list_add_tail(&sh->lru, &conf->delayed_list);
204 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
205 sh->bm_seq - conf->seq_write > 0)
206 list_add_tail(&sh->lru, &conf->bitmap_list);
208 clear_bit(STRIPE_BIT_DELAY, &sh->state);
209 list_add_tail(&sh->lru, &conf->handle_list);
211 md_wakeup_thread(conf->mddev->thread);
213 BUG_ON(stripe_operations_active(sh));
214 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
215 atomic_dec(&conf->preread_active_stripes);
216 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
217 md_wakeup_thread(conf->mddev->thread);
219 atomic_dec(&conf->active_stripes);
220 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
221 list_add_tail(&sh->lru, &conf->inactive_list);
222 wake_up(&conf->wait_for_stripe);
223 if (conf->retry_read_aligned)
224 md_wakeup_thread(conf->mddev->thread);
230 static void release_stripe(struct stripe_head *sh)
232 raid5_conf_t *conf = sh->raid_conf;
235 spin_lock_irqsave(&conf->device_lock, flags);
236 __release_stripe(conf, sh);
237 spin_unlock_irqrestore(&conf->device_lock, flags);
240 static inline void remove_hash(struct stripe_head *sh)
242 pr_debug("remove_hash(), stripe %llu\n",
243 (unsigned long long)sh->sector);
245 hlist_del_init(&sh->hash);
248 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
250 struct hlist_head *hp = stripe_hash(conf, sh->sector);
252 pr_debug("insert_hash(), stripe %llu\n",
253 (unsigned long long)sh->sector);
256 hlist_add_head(&sh->hash, hp);
260 /* find an idle stripe, make sure it is unhashed, and return it. */
261 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
263 struct stripe_head *sh = NULL;
264 struct list_head *first;
267 if (list_empty(&conf->inactive_list))
269 first = conf->inactive_list.next;
270 sh = list_entry(first, struct stripe_head, lru);
271 list_del_init(first);
273 atomic_inc(&conf->active_stripes);
278 static void shrink_buffers(struct stripe_head *sh)
282 int num = sh->raid_conf->pool_size;
284 for (i = 0; i < num ; i++) {
288 sh->dev[i].page = NULL;
293 static int grow_buffers(struct stripe_head *sh)
296 int num = sh->raid_conf->pool_size;
298 for (i = 0; i < num; i++) {
301 if (!(page = alloc_page(GFP_KERNEL))) {
304 sh->dev[i].page = page;
309 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
310 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
311 struct stripe_head *sh);
313 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
315 raid5_conf_t *conf = sh->raid_conf;
318 BUG_ON(atomic_read(&sh->count) != 0);
319 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
320 BUG_ON(stripe_operations_active(sh));
323 pr_debug("init_stripe called, stripe %llu\n",
324 (unsigned long long)sh->sector);
328 sh->generation = conf->generation - previous;
329 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
331 stripe_set_idx(sector, conf, previous, sh);
335 for (i = sh->disks; i--; ) {
336 struct r5dev *dev = &sh->dev[i];
338 if (dev->toread || dev->read || dev->towrite || dev->written ||
339 test_bit(R5_LOCKED, &dev->flags)) {
340 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
341 (unsigned long long)sh->sector, i, dev->toread,
342 dev->read, dev->towrite, dev->written,
343 test_bit(R5_LOCKED, &dev->flags));
347 raid5_build_block(sh, i, previous);
349 insert_hash(conf, sh);
352 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
355 struct stripe_head *sh;
356 struct hlist_node *hn;
359 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
360 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
361 if (sh->sector == sector && sh->generation == generation)
363 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
368 * Need to check if array has failed when deciding whether to:
370 * - remove non-faulty devices
373 * This determination is simple when no reshape is happening.
374 * However if there is a reshape, we need to carefully check
375 * both the before and after sections.
376 * This is because some failed devices may only affect one
377 * of the two sections, and some non-in_sync devices may
378 * be insync in the section most affected by failed devices.
380 static int has_failed(raid5_conf_t *conf)
384 if (conf->mddev->reshape_position == MaxSector)
385 return conf->mddev->degraded > conf->max_degraded;
389 for (i = 0; i < conf->previous_raid_disks; i++) {
390 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
391 if (!rdev || test_bit(Faulty, &rdev->flags))
393 else if (test_bit(In_sync, &rdev->flags))
396 /* not in-sync or faulty.
397 * If the reshape increases the number of devices,
398 * this is being recovered by the reshape, so
399 * this 'previous' section is not in_sync.
400 * If the number of devices is being reduced however,
401 * the device can only be part of the array if
402 * we are reverting a reshape, so this section will
405 if (conf->raid_disks >= conf->previous_raid_disks)
409 if (degraded > conf->max_degraded)
413 for (i = 0; i < conf->raid_disks; i++) {
414 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
415 if (!rdev || test_bit(Faulty, &rdev->flags))
417 else if (test_bit(In_sync, &rdev->flags))
420 /* not in-sync or faulty.
421 * If reshape increases the number of devices, this
422 * section has already been recovered, else it
423 * almost certainly hasn't.
425 if (conf->raid_disks <= conf->previous_raid_disks)
429 if (degraded > conf->max_degraded)
434 static struct stripe_head *
435 get_active_stripe(raid5_conf_t *conf, sector_t sector,
436 int previous, int noblock, int noquiesce)
438 struct stripe_head *sh;
440 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
442 spin_lock_irq(&conf->device_lock);
445 wait_event_lock_irq(conf->wait_for_stripe,
446 conf->quiesce == 0 || noquiesce,
447 conf->device_lock, /* nothing */);
448 sh = __find_stripe(conf, sector, conf->generation - previous);
450 if (!conf->inactive_blocked)
451 sh = get_free_stripe(conf);
452 if (noblock && sh == NULL)
455 conf->inactive_blocked = 1;
456 wait_event_lock_irq(conf->wait_for_stripe,
457 !list_empty(&conf->inactive_list) &&
458 (atomic_read(&conf->active_stripes)
459 < (conf->max_nr_stripes *3/4)
460 || !conf->inactive_blocked),
463 conf->inactive_blocked = 0;
465 init_stripe(sh, sector, previous);
467 if (atomic_read(&sh->count)) {
468 BUG_ON(!list_empty(&sh->lru)
469 && !test_bit(STRIPE_EXPANDING, &sh->state));
471 if (!test_bit(STRIPE_HANDLE, &sh->state))
472 atomic_inc(&conf->active_stripes);
473 if (list_empty(&sh->lru) &&
474 !test_bit(STRIPE_EXPANDING, &sh->state))
476 list_del_init(&sh->lru);
479 } while (sh == NULL);
482 atomic_inc(&sh->count);
484 spin_unlock_irq(&conf->device_lock);
489 raid5_end_read_request(struct bio *bi, int error);
491 raid5_end_write_request(struct bio *bi, int error);
493 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
495 raid5_conf_t *conf = sh->raid_conf;
496 int i, disks = sh->disks;
500 for (i = disks; i--; ) {
504 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
505 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
509 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
514 bi = &sh->dev[i].req;
518 bi->bi_end_io = raid5_end_write_request;
520 bi->bi_end_io = raid5_end_read_request;
523 rdev = rcu_dereference(conf->disks[i].rdev);
524 if (rdev && test_bit(Faulty, &rdev->flags))
527 atomic_inc(&rdev->nr_pending);
531 if (s->syncing || s->expanding || s->expanded)
532 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
534 set_bit(STRIPE_IO_STARTED, &sh->state);
536 bi->bi_bdev = rdev->bdev;
537 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
538 __func__, (unsigned long long)sh->sector,
540 atomic_inc(&sh->count);
541 bi->bi_sector = sh->sector + rdev->data_offset;
542 bi->bi_flags = 1 << BIO_UPTODATE;
546 bi->bi_io_vec = &sh->dev[i].vec;
547 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
548 bi->bi_io_vec[0].bv_offset = 0;
549 bi->bi_size = STRIPE_SIZE;
552 test_bit(R5_ReWrite, &sh->dev[i].flags))
553 atomic_add(STRIPE_SECTORS,
554 &rdev->corrected_errors);
555 generic_make_request(bi);
558 set_bit(STRIPE_DEGRADED, &sh->state);
559 pr_debug("skip op %ld on disc %d for sector %llu\n",
560 bi->bi_rw, i, (unsigned long long)sh->sector);
561 clear_bit(R5_LOCKED, &sh->dev[i].flags);
562 set_bit(STRIPE_HANDLE, &sh->state);
567 static struct dma_async_tx_descriptor *
568 async_copy_data(int frombio, struct bio *bio, struct page *page,
569 sector_t sector, struct dma_async_tx_descriptor *tx)
572 struct page *bio_page;
575 struct async_submit_ctl submit;
576 enum async_tx_flags flags = 0;
578 if (bio->bi_sector >= sector)
579 page_offset = (signed)(bio->bi_sector - sector) * 512;
581 page_offset = (signed)(sector - bio->bi_sector) * -512;
584 flags |= ASYNC_TX_FENCE;
585 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
587 bio_for_each_segment(bvl, bio, i) {
588 int len = bvl->bv_len;
592 if (page_offset < 0) {
593 b_offset = -page_offset;
594 page_offset += b_offset;
598 if (len > 0 && page_offset + len > STRIPE_SIZE)
599 clen = STRIPE_SIZE - page_offset;
604 b_offset += bvl->bv_offset;
605 bio_page = bvl->bv_page;
607 tx = async_memcpy(page, bio_page, page_offset,
608 b_offset, clen, &submit);
610 tx = async_memcpy(bio_page, page, b_offset,
611 page_offset, clen, &submit);
613 /* chain the operations */
614 submit.depend_tx = tx;
616 if (clen < len) /* hit end of page */
624 static void ops_complete_biofill(void *stripe_head_ref)
626 struct stripe_head *sh = stripe_head_ref;
627 struct bio *return_bi = NULL;
628 raid5_conf_t *conf = sh->raid_conf;
631 pr_debug("%s: stripe %llu\n", __func__,
632 (unsigned long long)sh->sector);
634 /* clear completed biofills */
635 spin_lock_irq(&conf->device_lock);
636 for (i = sh->disks; i--; ) {
637 struct r5dev *dev = &sh->dev[i];
639 /* acknowledge completion of a biofill operation */
640 /* and check if we need to reply to a read request,
641 * new R5_Wantfill requests are held off until
642 * !STRIPE_BIOFILL_RUN
644 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
645 struct bio *rbi, *rbi2;
650 while (rbi && rbi->bi_sector <
651 dev->sector + STRIPE_SECTORS) {
652 rbi2 = r5_next_bio(rbi, dev->sector);
653 if (!raid5_dec_bi_phys_segments(rbi)) {
654 rbi->bi_next = return_bi;
661 spin_unlock_irq(&conf->device_lock);
662 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
664 return_io(return_bi);
666 set_bit(STRIPE_HANDLE, &sh->state);
670 static void ops_run_biofill(struct stripe_head *sh)
672 struct dma_async_tx_descriptor *tx = NULL;
673 raid5_conf_t *conf = sh->raid_conf;
674 struct async_submit_ctl submit;
677 pr_debug("%s: stripe %llu\n", __func__,
678 (unsigned long long)sh->sector);
680 for (i = sh->disks; i--; ) {
681 struct r5dev *dev = &sh->dev[i];
682 if (test_bit(R5_Wantfill, &dev->flags)) {
684 spin_lock_irq(&conf->device_lock);
685 dev->read = rbi = dev->toread;
687 spin_unlock_irq(&conf->device_lock);
688 while (rbi && rbi->bi_sector <
689 dev->sector + STRIPE_SECTORS) {
690 tx = async_copy_data(0, rbi, dev->page,
692 rbi = r5_next_bio(rbi, dev->sector);
697 atomic_inc(&sh->count);
698 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
699 async_trigger_callback(&submit);
702 static void mark_target_uptodate(struct stripe_head *sh, int target)
709 tgt = &sh->dev[target];
710 set_bit(R5_UPTODATE, &tgt->flags);
711 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
712 clear_bit(R5_Wantcompute, &tgt->flags);
715 static void ops_complete_compute(void *stripe_head_ref)
717 struct stripe_head *sh = stripe_head_ref;
719 pr_debug("%s: stripe %llu\n", __func__,
720 (unsigned long long)sh->sector);
722 /* mark the computed target(s) as uptodate */
723 mark_target_uptodate(sh, sh->ops.target);
724 mark_target_uptodate(sh, sh->ops.target2);
726 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
727 if (sh->check_state == check_state_compute_run)
728 sh->check_state = check_state_compute_result;
729 set_bit(STRIPE_HANDLE, &sh->state);
733 /* return a pointer to the address conversion region of the scribble buffer */
734 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
735 struct raid5_percpu *percpu)
737 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
740 static struct dma_async_tx_descriptor *
741 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
743 int disks = sh->disks;
744 struct page **xor_srcs = percpu->scribble;
745 int target = sh->ops.target;
746 struct r5dev *tgt = &sh->dev[target];
747 struct page *xor_dest = tgt->page;
749 struct dma_async_tx_descriptor *tx;
750 struct async_submit_ctl submit;
753 pr_debug("%s: stripe %llu block: %d\n",
754 __func__, (unsigned long long)sh->sector, target);
755 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
757 for (i = disks; i--; )
759 xor_srcs[count++] = sh->dev[i].page;
761 atomic_inc(&sh->count);
763 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
764 ops_complete_compute, sh, to_addr_conv(sh, percpu));
765 if (unlikely(count == 1))
766 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
768 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
773 /* set_syndrome_sources - populate source buffers for gen_syndrome
774 * @srcs - (struct page *) array of size sh->disks
775 * @sh - stripe_head to parse
777 * Populates srcs in proper layout order for the stripe and returns the
778 * 'count' of sources to be used in a call to async_gen_syndrome. The P
779 * destination buffer is recorded in srcs[count] and the Q destination
780 * is recorded in srcs[count+1]].
782 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
784 int disks = sh->disks;
785 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
786 int d0_idx = raid6_d0(sh);
790 for (i = 0; i < disks; i++)
796 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
798 srcs[slot] = sh->dev[i].page;
799 i = raid6_next_disk(i, disks);
800 } while (i != d0_idx);
802 return syndrome_disks;
805 static struct dma_async_tx_descriptor *
806 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
808 int disks = sh->disks;
809 struct page **blocks = percpu->scribble;
811 int qd_idx = sh->qd_idx;
812 struct dma_async_tx_descriptor *tx;
813 struct async_submit_ctl submit;
819 if (sh->ops.target < 0)
820 target = sh->ops.target2;
821 else if (sh->ops.target2 < 0)
822 target = sh->ops.target;
824 /* we should only have one valid target */
827 pr_debug("%s: stripe %llu block: %d\n",
828 __func__, (unsigned long long)sh->sector, target);
830 tgt = &sh->dev[target];
831 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
834 atomic_inc(&sh->count);
836 if (target == qd_idx) {
837 count = set_syndrome_sources(blocks, sh);
838 blocks[count] = NULL; /* regenerating p is not necessary */
839 BUG_ON(blocks[count+1] != dest); /* q should already be set */
840 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
841 ops_complete_compute, sh,
842 to_addr_conv(sh, percpu));
843 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
845 /* Compute any data- or p-drive using XOR */
847 for (i = disks; i-- ; ) {
848 if (i == target || i == qd_idx)
850 blocks[count++] = sh->dev[i].page;
853 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
854 NULL, ops_complete_compute, sh,
855 to_addr_conv(sh, percpu));
856 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
862 static struct dma_async_tx_descriptor *
863 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
865 int i, count, disks = sh->disks;
866 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
867 int d0_idx = raid6_d0(sh);
868 int faila = -1, failb = -1;
869 int target = sh->ops.target;
870 int target2 = sh->ops.target2;
871 struct r5dev *tgt = &sh->dev[target];
872 struct r5dev *tgt2 = &sh->dev[target2];
873 struct dma_async_tx_descriptor *tx;
874 struct page **blocks = percpu->scribble;
875 struct async_submit_ctl submit;
877 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
878 __func__, (unsigned long long)sh->sector, target, target2);
879 BUG_ON(target < 0 || target2 < 0);
880 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
881 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
883 /* we need to open-code set_syndrome_sources to handle the
884 * slot number conversion for 'faila' and 'failb'
886 for (i = 0; i < disks ; i++)
891 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
893 blocks[slot] = sh->dev[i].page;
899 i = raid6_next_disk(i, disks);
900 } while (i != d0_idx);
902 BUG_ON(faila == failb);
905 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
906 __func__, (unsigned long long)sh->sector, faila, failb);
908 atomic_inc(&sh->count);
910 if (failb == syndrome_disks+1) {
911 /* Q disk is one of the missing disks */
912 if (faila == syndrome_disks) {
913 /* Missing P+Q, just recompute */
914 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
915 ops_complete_compute, sh,
916 to_addr_conv(sh, percpu));
917 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
918 STRIPE_SIZE, &submit);
922 int qd_idx = sh->qd_idx;
924 /* Missing D+Q: recompute D from P, then recompute Q */
925 if (target == qd_idx)
926 data_target = target2;
928 data_target = target;
931 for (i = disks; i-- ; ) {
932 if (i == data_target || i == qd_idx)
934 blocks[count++] = sh->dev[i].page;
936 dest = sh->dev[data_target].page;
937 init_async_submit(&submit,
938 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
940 to_addr_conv(sh, percpu));
941 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
944 count = set_syndrome_sources(blocks, sh);
945 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
946 ops_complete_compute, sh,
947 to_addr_conv(sh, percpu));
948 return async_gen_syndrome(blocks, 0, count+2,
949 STRIPE_SIZE, &submit);
952 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
953 ops_complete_compute, sh,
954 to_addr_conv(sh, percpu));
955 if (failb == syndrome_disks) {
956 /* We're missing D+P. */
957 return async_raid6_datap_recov(syndrome_disks+2,
961 /* We're missing D+D. */
962 return async_raid6_2data_recov(syndrome_disks+2,
963 STRIPE_SIZE, faila, failb,
970 static void ops_complete_prexor(void *stripe_head_ref)
972 struct stripe_head *sh = stripe_head_ref;
974 pr_debug("%s: stripe %llu\n", __func__,
975 (unsigned long long)sh->sector);
978 static struct dma_async_tx_descriptor *
979 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
980 struct dma_async_tx_descriptor *tx)
982 int disks = sh->disks;
983 struct page **xor_srcs = percpu->scribble;
984 int count = 0, pd_idx = sh->pd_idx, i;
985 struct async_submit_ctl submit;
987 /* existing parity data subtracted */
988 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
990 pr_debug("%s: stripe %llu\n", __func__,
991 (unsigned long long)sh->sector);
993 for (i = disks; i--; ) {
994 struct r5dev *dev = &sh->dev[i];
995 /* Only process blocks that are known to be uptodate */
996 if (test_bit(R5_Wantdrain, &dev->flags))
997 xor_srcs[count++] = dev->page;
1000 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1001 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1002 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1007 static struct dma_async_tx_descriptor *
1008 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1010 int disks = sh->disks;
1013 pr_debug("%s: stripe %llu\n", __func__,
1014 (unsigned long long)sh->sector);
1016 for (i = disks; i--; ) {
1017 struct r5dev *dev = &sh->dev[i];
1020 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1023 spin_lock_irq(&sh->raid_conf->device_lock);
1024 chosen = dev->towrite;
1025 dev->towrite = NULL;
1026 BUG_ON(dev->written);
1027 wbi = dev->written = chosen;
1028 spin_unlock_irq(&sh->raid_conf->device_lock);
1030 while (wbi && wbi->bi_sector <
1031 dev->sector + STRIPE_SECTORS) {
1032 if (wbi->bi_rw & REQ_FUA)
1033 set_bit(R5_WantFUA, &dev->flags);
1034 tx = async_copy_data(1, wbi, dev->page,
1036 wbi = r5_next_bio(wbi, dev->sector);
1044 static void ops_complete_reconstruct(void *stripe_head_ref)
1046 struct stripe_head *sh = stripe_head_ref;
1047 int disks = sh->disks;
1048 int pd_idx = sh->pd_idx;
1049 int qd_idx = sh->qd_idx;
1053 pr_debug("%s: stripe %llu\n", __func__,
1054 (unsigned long long)sh->sector);
1056 for (i = disks; i--; )
1057 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1059 for (i = disks; i--; ) {
1060 struct r5dev *dev = &sh->dev[i];
1062 if (dev->written || i == pd_idx || i == qd_idx) {
1063 set_bit(R5_UPTODATE, &dev->flags);
1065 set_bit(R5_WantFUA, &dev->flags);
1069 if (sh->reconstruct_state == reconstruct_state_drain_run)
1070 sh->reconstruct_state = reconstruct_state_drain_result;
1071 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1072 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1074 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1075 sh->reconstruct_state = reconstruct_state_result;
1078 set_bit(STRIPE_HANDLE, &sh->state);
1083 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1084 struct dma_async_tx_descriptor *tx)
1086 int disks = sh->disks;
1087 struct page **xor_srcs = percpu->scribble;
1088 struct async_submit_ctl submit;
1089 int count = 0, pd_idx = sh->pd_idx, i;
1090 struct page *xor_dest;
1092 unsigned long flags;
1094 pr_debug("%s: stripe %llu\n", __func__,
1095 (unsigned long long)sh->sector);
1097 /* check if prexor is active which means only process blocks
1098 * that are part of a read-modify-write (written)
1100 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1102 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1103 for (i = disks; i--; ) {
1104 struct r5dev *dev = &sh->dev[i];
1106 xor_srcs[count++] = dev->page;
1109 xor_dest = sh->dev[pd_idx].page;
1110 for (i = disks; i--; ) {
1111 struct r5dev *dev = &sh->dev[i];
1113 xor_srcs[count++] = dev->page;
1117 /* 1/ if we prexor'd then the dest is reused as a source
1118 * 2/ if we did not prexor then we are redoing the parity
1119 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1120 * for the synchronous xor case
1122 flags = ASYNC_TX_ACK |
1123 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1125 atomic_inc(&sh->count);
1127 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1128 to_addr_conv(sh, percpu));
1129 if (unlikely(count == 1))
1130 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1132 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1136 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1137 struct dma_async_tx_descriptor *tx)
1139 struct async_submit_ctl submit;
1140 struct page **blocks = percpu->scribble;
1143 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1145 count = set_syndrome_sources(blocks, sh);
1147 atomic_inc(&sh->count);
1149 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1150 sh, to_addr_conv(sh, percpu));
1151 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1154 static void ops_complete_check(void *stripe_head_ref)
1156 struct stripe_head *sh = stripe_head_ref;
1158 pr_debug("%s: stripe %llu\n", __func__,
1159 (unsigned long long)sh->sector);
1161 sh->check_state = check_state_check_result;
1162 set_bit(STRIPE_HANDLE, &sh->state);
1166 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1168 int disks = sh->disks;
1169 int pd_idx = sh->pd_idx;
1170 int qd_idx = sh->qd_idx;
1171 struct page *xor_dest;
1172 struct page **xor_srcs = percpu->scribble;
1173 struct dma_async_tx_descriptor *tx;
1174 struct async_submit_ctl submit;
1178 pr_debug("%s: stripe %llu\n", __func__,
1179 (unsigned long long)sh->sector);
1182 xor_dest = sh->dev[pd_idx].page;
1183 xor_srcs[count++] = xor_dest;
1184 for (i = disks; i--; ) {
1185 if (i == pd_idx || i == qd_idx)
1187 xor_srcs[count++] = sh->dev[i].page;
1190 init_async_submit(&submit, 0, NULL, NULL, NULL,
1191 to_addr_conv(sh, percpu));
1192 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1193 &sh->ops.zero_sum_result, &submit);
1195 atomic_inc(&sh->count);
1196 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1197 tx = async_trigger_callback(&submit);
1200 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1202 struct page **srcs = percpu->scribble;
1203 struct async_submit_ctl submit;
1206 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1207 (unsigned long long)sh->sector, checkp);
1209 count = set_syndrome_sources(srcs, sh);
1213 atomic_inc(&sh->count);
1214 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1215 sh, to_addr_conv(sh, percpu));
1216 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1217 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1220 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1222 int overlap_clear = 0, i, disks = sh->disks;
1223 struct dma_async_tx_descriptor *tx = NULL;
1224 raid5_conf_t *conf = sh->raid_conf;
1225 int level = conf->level;
1226 struct raid5_percpu *percpu;
1230 percpu = per_cpu_ptr(conf->percpu, cpu);
1231 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1232 ops_run_biofill(sh);
1236 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1238 tx = ops_run_compute5(sh, percpu);
1240 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1241 tx = ops_run_compute6_1(sh, percpu);
1243 tx = ops_run_compute6_2(sh, percpu);
1245 /* terminate the chain if reconstruct is not set to be run */
1246 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1250 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1251 tx = ops_run_prexor(sh, percpu, tx);
1253 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1254 tx = ops_run_biodrain(sh, tx);
1258 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1260 ops_run_reconstruct5(sh, percpu, tx);
1262 ops_run_reconstruct6(sh, percpu, tx);
1265 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1266 if (sh->check_state == check_state_run)
1267 ops_run_check_p(sh, percpu);
1268 else if (sh->check_state == check_state_run_q)
1269 ops_run_check_pq(sh, percpu, 0);
1270 else if (sh->check_state == check_state_run_pq)
1271 ops_run_check_pq(sh, percpu, 1);
1277 for (i = disks; i--; ) {
1278 struct r5dev *dev = &sh->dev[i];
1279 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1280 wake_up(&sh->raid_conf->wait_for_overlap);
1285 #ifdef CONFIG_MULTICORE_RAID456
1286 static void async_run_ops(void *param, async_cookie_t cookie)
1288 struct stripe_head *sh = param;
1289 unsigned long ops_request = sh->ops.request;
1291 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1292 wake_up(&sh->ops.wait_for_ops);
1294 __raid_run_ops(sh, ops_request);
1298 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1300 /* since handle_stripe can be called outside of raid5d context
1301 * we need to ensure sh->ops.request is de-staged before another
1304 wait_event(sh->ops.wait_for_ops,
1305 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1306 sh->ops.request = ops_request;
1308 atomic_inc(&sh->count);
1309 async_schedule(async_run_ops, sh);
1312 #define raid_run_ops __raid_run_ops
1315 static int grow_one_stripe(raid5_conf_t *conf)
1317 struct stripe_head *sh;
1318 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1322 sh->raid_conf = conf;
1323 #ifdef CONFIG_MULTICORE_RAID456
1324 init_waitqueue_head(&sh->ops.wait_for_ops);
1327 if (grow_buffers(sh)) {
1329 kmem_cache_free(conf->slab_cache, sh);
1332 /* we just created an active stripe so... */
1333 atomic_set(&sh->count, 1);
1334 atomic_inc(&conf->active_stripes);
1335 INIT_LIST_HEAD(&sh->lru);
1340 static int grow_stripes(raid5_conf_t *conf, int num)
1342 struct kmem_cache *sc;
1343 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1345 if (conf->mddev->gendisk)
1346 sprintf(conf->cache_name[0],
1347 "raid%d-%s", conf->level, mdname(conf->mddev));
1349 sprintf(conf->cache_name[0],
1350 "raid%d-%p", conf->level, conf->mddev);
1351 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1353 conf->active_name = 0;
1354 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1355 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1359 conf->slab_cache = sc;
1360 conf->pool_size = devs;
1362 if (!grow_one_stripe(conf))
1368 * scribble_len - return the required size of the scribble region
1369 * @num - total number of disks in the array
1371 * The size must be enough to contain:
1372 * 1/ a struct page pointer for each device in the array +2
1373 * 2/ room to convert each entry in (1) to its corresponding dma
1374 * (dma_map_page()) or page (page_address()) address.
1376 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1377 * calculate over all devices (not just the data blocks), using zeros in place
1378 * of the P and Q blocks.
1380 static size_t scribble_len(int num)
1384 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1389 static int resize_stripes(raid5_conf_t *conf, int newsize)
1391 /* Make all the stripes able to hold 'newsize' devices.
1392 * New slots in each stripe get 'page' set to a new page.
1394 * This happens in stages:
1395 * 1/ create a new kmem_cache and allocate the required number of
1397 * 2/ gather all the old stripe_heads and tranfer the pages across
1398 * to the new stripe_heads. This will have the side effect of
1399 * freezing the array as once all stripe_heads have been collected,
1400 * no IO will be possible. Old stripe heads are freed once their
1401 * pages have been transferred over, and the old kmem_cache is
1402 * freed when all stripes are done.
1403 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1404 * we simple return a failre status - no need to clean anything up.
1405 * 4/ allocate new pages for the new slots in the new stripe_heads.
1406 * If this fails, we don't bother trying the shrink the
1407 * stripe_heads down again, we just leave them as they are.
1408 * As each stripe_head is processed the new one is released into
1411 * Once step2 is started, we cannot afford to wait for a write,
1412 * so we use GFP_NOIO allocations.
1414 struct stripe_head *osh, *nsh;
1415 LIST_HEAD(newstripes);
1416 struct disk_info *ndisks;
1419 struct kmem_cache *sc;
1422 if (newsize <= conf->pool_size)
1423 return 0; /* never bother to shrink */
1425 err = md_allow_write(conf->mddev);
1430 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1431 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1436 for (i = conf->max_nr_stripes; i; i--) {
1437 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1441 nsh->raid_conf = conf;
1442 #ifdef CONFIG_MULTICORE_RAID456
1443 init_waitqueue_head(&nsh->ops.wait_for_ops);
1446 list_add(&nsh->lru, &newstripes);
1449 /* didn't get enough, give up */
1450 while (!list_empty(&newstripes)) {
1451 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1452 list_del(&nsh->lru);
1453 kmem_cache_free(sc, nsh);
1455 kmem_cache_destroy(sc);
1458 /* Step 2 - Must use GFP_NOIO now.
1459 * OK, we have enough stripes, start collecting inactive
1460 * stripes and copying them over
1462 list_for_each_entry(nsh, &newstripes, lru) {
1463 spin_lock_irq(&conf->device_lock);
1464 wait_event_lock_irq(conf->wait_for_stripe,
1465 !list_empty(&conf->inactive_list),
1468 osh = get_free_stripe(conf);
1469 spin_unlock_irq(&conf->device_lock);
1470 atomic_set(&nsh->count, 1);
1471 for(i=0; i<conf->pool_size; i++)
1472 nsh->dev[i].page = osh->dev[i].page;
1473 for( ; i<newsize; i++)
1474 nsh->dev[i].page = NULL;
1475 kmem_cache_free(conf->slab_cache, osh);
1477 kmem_cache_destroy(conf->slab_cache);
1480 * At this point, we are holding all the stripes so the array
1481 * is completely stalled, so now is a good time to resize
1482 * conf->disks and the scribble region
1484 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1486 for (i=0; i<conf->raid_disks; i++)
1487 ndisks[i] = conf->disks[i];
1489 conf->disks = ndisks;
1494 conf->scribble_len = scribble_len(newsize);
1495 for_each_present_cpu(cpu) {
1496 struct raid5_percpu *percpu;
1499 percpu = per_cpu_ptr(conf->percpu, cpu);
1500 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1503 kfree(percpu->scribble);
1504 percpu->scribble = scribble;
1512 /* Step 4, return new stripes to service */
1513 while(!list_empty(&newstripes)) {
1514 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1515 list_del_init(&nsh->lru);
1517 for (i=conf->raid_disks; i < newsize; i++)
1518 if (nsh->dev[i].page == NULL) {
1519 struct page *p = alloc_page(GFP_NOIO);
1520 nsh->dev[i].page = p;
1524 release_stripe(nsh);
1526 /* critical section pass, GFP_NOIO no longer needed */
1528 conf->slab_cache = sc;
1529 conf->active_name = 1-conf->active_name;
1530 conf->pool_size = newsize;
1534 static int drop_one_stripe(raid5_conf_t *conf)
1536 struct stripe_head *sh;
1538 spin_lock_irq(&conf->device_lock);
1539 sh = get_free_stripe(conf);
1540 spin_unlock_irq(&conf->device_lock);
1543 BUG_ON(atomic_read(&sh->count));
1545 kmem_cache_free(conf->slab_cache, sh);
1546 atomic_dec(&conf->active_stripes);
1550 static void shrink_stripes(raid5_conf_t *conf)
1552 while (drop_one_stripe(conf))
1555 if (conf->slab_cache)
1556 kmem_cache_destroy(conf->slab_cache);
1557 conf->slab_cache = NULL;
1560 static void raid5_end_read_request(struct bio * bi, int error)
1562 struct stripe_head *sh = bi->bi_private;
1563 raid5_conf_t *conf = sh->raid_conf;
1564 int disks = sh->disks, i;
1565 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1566 char b[BDEVNAME_SIZE];
1570 for (i=0 ; i<disks; i++)
1571 if (bi == &sh->dev[i].req)
1574 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1575 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1583 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1584 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1585 rdev = conf->disks[i].rdev;
1586 printk_rl(KERN_INFO "md/raid:%s: read error corrected"
1587 " (%lu sectors at %llu on %s)\n",
1588 mdname(conf->mddev), STRIPE_SECTORS,
1589 (unsigned long long)(sh->sector
1590 + rdev->data_offset),
1591 bdevname(rdev->bdev, b));
1592 clear_bit(R5_ReadError, &sh->dev[i].flags);
1593 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1595 if (atomic_read(&conf->disks[i].rdev->read_errors))
1596 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1598 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1600 rdev = conf->disks[i].rdev;
1602 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1603 atomic_inc(&rdev->read_errors);
1604 if (conf->mddev->degraded >= conf->max_degraded)
1605 printk_rl(KERN_WARNING
1606 "md/raid:%s: read error not correctable "
1607 "(sector %llu on %s).\n",
1608 mdname(conf->mddev),
1609 (unsigned long long)(sh->sector
1610 + rdev->data_offset),
1612 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1614 printk_rl(KERN_WARNING
1615 "md/raid:%s: read error NOT corrected!! "
1616 "(sector %llu on %s).\n",
1617 mdname(conf->mddev),
1618 (unsigned long long)(sh->sector
1619 + rdev->data_offset),
1621 else if (atomic_read(&rdev->read_errors)
1622 > conf->max_nr_stripes)
1624 "md/raid:%s: Too many read errors, failing device %s.\n",
1625 mdname(conf->mddev), bdn);
1629 set_bit(R5_ReadError, &sh->dev[i].flags);
1631 clear_bit(R5_ReadError, &sh->dev[i].flags);
1632 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1633 md_error(conf->mddev, rdev);
1636 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1637 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1638 set_bit(STRIPE_HANDLE, &sh->state);
1642 static void raid5_end_write_request(struct bio *bi, int error)
1644 struct stripe_head *sh = bi->bi_private;
1645 raid5_conf_t *conf = sh->raid_conf;
1646 int disks = sh->disks, i;
1647 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1649 for (i=0 ; i<disks; i++)
1650 if (bi == &sh->dev[i].req)
1653 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1654 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1662 md_error(conf->mddev, conf->disks[i].rdev);
1664 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1666 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1667 set_bit(STRIPE_HANDLE, &sh->state);
1672 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1674 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1676 struct r5dev *dev = &sh->dev[i];
1678 bio_init(&dev->req);
1679 dev->req.bi_io_vec = &dev->vec;
1681 dev->req.bi_max_vecs++;
1682 dev->vec.bv_page = dev->page;
1683 dev->vec.bv_len = STRIPE_SIZE;
1684 dev->vec.bv_offset = 0;
1686 dev->req.bi_sector = sh->sector;
1687 dev->req.bi_private = sh;
1690 dev->sector = compute_blocknr(sh, i, previous);
1693 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1695 char b[BDEVNAME_SIZE];
1696 raid5_conf_t *conf = mddev->private;
1697 pr_debug("raid456: error called\n");
1699 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1700 unsigned long flags;
1701 spin_lock_irqsave(&conf->device_lock, flags);
1703 spin_unlock_irqrestore(&conf->device_lock, flags);
1705 * if recovery was running, make sure it aborts.
1707 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1709 set_bit(Faulty, &rdev->flags);
1710 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1712 "md/raid:%s: Disk failure on %s, disabling device.\n"
1713 "md/raid:%s: Operation continuing on %d devices.\n",
1715 bdevname(rdev->bdev, b),
1717 conf->raid_disks - mddev->degraded);
1721 * Input: a 'big' sector number,
1722 * Output: index of the data and parity disk, and the sector # in them.
1724 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1725 int previous, int *dd_idx,
1726 struct stripe_head *sh)
1728 sector_t stripe, stripe2;
1729 sector_t chunk_number;
1730 unsigned int chunk_offset;
1733 sector_t new_sector;
1734 int algorithm = previous ? conf->prev_algo
1736 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1737 : conf->chunk_sectors;
1738 int raid_disks = previous ? conf->previous_raid_disks
1740 int data_disks = raid_disks - conf->max_degraded;
1742 /* First compute the information on this sector */
1745 * Compute the chunk number and the sector offset inside the chunk
1747 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1748 chunk_number = r_sector;
1751 * Compute the stripe number
1753 stripe = chunk_number;
1754 *dd_idx = sector_div(stripe, data_disks);
1757 * Select the parity disk based on the user selected algorithm.
1759 pd_idx = qd_idx = ~0;
1760 switch(conf->level) {
1762 pd_idx = data_disks;
1765 switch (algorithm) {
1766 case ALGORITHM_LEFT_ASYMMETRIC:
1767 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1768 if (*dd_idx >= pd_idx)
1771 case ALGORITHM_RIGHT_ASYMMETRIC:
1772 pd_idx = sector_div(stripe2, raid_disks);
1773 if (*dd_idx >= pd_idx)
1776 case ALGORITHM_LEFT_SYMMETRIC:
1777 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1778 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1780 case ALGORITHM_RIGHT_SYMMETRIC:
1781 pd_idx = sector_div(stripe2, raid_disks);
1782 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1784 case ALGORITHM_PARITY_0:
1788 case ALGORITHM_PARITY_N:
1789 pd_idx = data_disks;
1797 switch (algorithm) {
1798 case ALGORITHM_LEFT_ASYMMETRIC:
1799 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1800 qd_idx = pd_idx + 1;
1801 if (pd_idx == raid_disks-1) {
1802 (*dd_idx)++; /* Q D D D P */
1804 } else if (*dd_idx >= pd_idx)
1805 (*dd_idx) += 2; /* D D P Q D */
1807 case ALGORITHM_RIGHT_ASYMMETRIC:
1808 pd_idx = sector_div(stripe2, raid_disks);
1809 qd_idx = pd_idx + 1;
1810 if (pd_idx == raid_disks-1) {
1811 (*dd_idx)++; /* Q D D D P */
1813 } else if (*dd_idx >= pd_idx)
1814 (*dd_idx) += 2; /* D D P Q D */
1816 case ALGORITHM_LEFT_SYMMETRIC:
1817 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1818 qd_idx = (pd_idx + 1) % raid_disks;
1819 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1821 case ALGORITHM_RIGHT_SYMMETRIC:
1822 pd_idx = sector_div(stripe2, raid_disks);
1823 qd_idx = (pd_idx + 1) % raid_disks;
1824 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1827 case ALGORITHM_PARITY_0:
1832 case ALGORITHM_PARITY_N:
1833 pd_idx = data_disks;
1834 qd_idx = data_disks + 1;
1837 case ALGORITHM_ROTATING_ZERO_RESTART:
1838 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1839 * of blocks for computing Q is different.
1841 pd_idx = sector_div(stripe2, raid_disks);
1842 qd_idx = pd_idx + 1;
1843 if (pd_idx == raid_disks-1) {
1844 (*dd_idx)++; /* Q D D D P */
1846 } else if (*dd_idx >= pd_idx)
1847 (*dd_idx) += 2; /* D D P Q D */
1851 case ALGORITHM_ROTATING_N_RESTART:
1852 /* Same a left_asymmetric, by first stripe is
1853 * D D D P Q rather than
1857 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1858 qd_idx = pd_idx + 1;
1859 if (pd_idx == raid_disks-1) {
1860 (*dd_idx)++; /* Q D D D P */
1862 } else if (*dd_idx >= pd_idx)
1863 (*dd_idx) += 2; /* D D P Q D */
1867 case ALGORITHM_ROTATING_N_CONTINUE:
1868 /* Same as left_symmetric but Q is before P */
1869 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1870 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1871 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1875 case ALGORITHM_LEFT_ASYMMETRIC_6:
1876 /* RAID5 left_asymmetric, with Q on last device */
1877 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1878 if (*dd_idx >= pd_idx)
1880 qd_idx = raid_disks - 1;
1883 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1884 pd_idx = sector_div(stripe2, raid_disks-1);
1885 if (*dd_idx >= pd_idx)
1887 qd_idx = raid_disks - 1;
1890 case ALGORITHM_LEFT_SYMMETRIC_6:
1891 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1892 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1893 qd_idx = raid_disks - 1;
1896 case ALGORITHM_RIGHT_SYMMETRIC_6:
1897 pd_idx = sector_div(stripe2, raid_disks-1);
1898 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1899 qd_idx = raid_disks - 1;
1902 case ALGORITHM_PARITY_0_6:
1905 qd_idx = raid_disks - 1;
1915 sh->pd_idx = pd_idx;
1916 sh->qd_idx = qd_idx;
1917 sh->ddf_layout = ddf_layout;
1920 * Finally, compute the new sector number
1922 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1927 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1929 raid5_conf_t *conf = sh->raid_conf;
1930 int raid_disks = sh->disks;
1931 int data_disks = raid_disks - conf->max_degraded;
1932 sector_t new_sector = sh->sector, check;
1933 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1934 : conf->chunk_sectors;
1935 int algorithm = previous ? conf->prev_algo
1939 sector_t chunk_number;
1940 int dummy1, dd_idx = i;
1942 struct stripe_head sh2;
1945 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1946 stripe = new_sector;
1948 if (i == sh->pd_idx)
1950 switch(conf->level) {
1953 switch (algorithm) {
1954 case ALGORITHM_LEFT_ASYMMETRIC:
1955 case ALGORITHM_RIGHT_ASYMMETRIC:
1959 case ALGORITHM_LEFT_SYMMETRIC:
1960 case ALGORITHM_RIGHT_SYMMETRIC:
1963 i -= (sh->pd_idx + 1);
1965 case ALGORITHM_PARITY_0:
1968 case ALGORITHM_PARITY_N:
1975 if (i == sh->qd_idx)
1976 return 0; /* It is the Q disk */
1977 switch (algorithm) {
1978 case ALGORITHM_LEFT_ASYMMETRIC:
1979 case ALGORITHM_RIGHT_ASYMMETRIC:
1980 case ALGORITHM_ROTATING_ZERO_RESTART:
1981 case ALGORITHM_ROTATING_N_RESTART:
1982 if (sh->pd_idx == raid_disks-1)
1983 i--; /* Q D D D P */
1984 else if (i > sh->pd_idx)
1985 i -= 2; /* D D P Q D */
1987 case ALGORITHM_LEFT_SYMMETRIC:
1988 case ALGORITHM_RIGHT_SYMMETRIC:
1989 if (sh->pd_idx == raid_disks-1)
1990 i--; /* Q D D D P */
1995 i -= (sh->pd_idx + 2);
1998 case ALGORITHM_PARITY_0:
2001 case ALGORITHM_PARITY_N:
2003 case ALGORITHM_ROTATING_N_CONTINUE:
2004 /* Like left_symmetric, but P is before Q */
2005 if (sh->pd_idx == 0)
2006 i--; /* P D D D Q */
2011 i -= (sh->pd_idx + 1);
2014 case ALGORITHM_LEFT_ASYMMETRIC_6:
2015 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2019 case ALGORITHM_LEFT_SYMMETRIC_6:
2020 case ALGORITHM_RIGHT_SYMMETRIC_6:
2022 i += data_disks + 1;
2023 i -= (sh->pd_idx + 1);
2025 case ALGORITHM_PARITY_0_6:
2034 chunk_number = stripe * data_disks + i;
2035 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2037 check = raid5_compute_sector(conf, r_sector,
2038 previous, &dummy1, &sh2);
2039 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2040 || sh2.qd_idx != sh->qd_idx) {
2041 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2042 mdname(conf->mddev));
2050 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2051 int rcw, int expand)
2053 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2054 raid5_conf_t *conf = sh->raid_conf;
2055 int level = conf->level;
2058 /* if we are not expanding this is a proper write request, and
2059 * there will be bios with new data to be drained into the
2063 sh->reconstruct_state = reconstruct_state_drain_run;
2064 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2066 sh->reconstruct_state = reconstruct_state_run;
2068 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2070 for (i = disks; i--; ) {
2071 struct r5dev *dev = &sh->dev[i];
2074 set_bit(R5_LOCKED, &dev->flags);
2075 set_bit(R5_Wantdrain, &dev->flags);
2077 clear_bit(R5_UPTODATE, &dev->flags);
2081 if (s->locked + conf->max_degraded == disks)
2082 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2083 atomic_inc(&conf->pending_full_writes);
2086 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2087 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2089 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2090 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2091 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2092 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2094 for (i = disks; i--; ) {
2095 struct r5dev *dev = &sh->dev[i];
2100 (test_bit(R5_UPTODATE, &dev->flags) ||
2101 test_bit(R5_Wantcompute, &dev->flags))) {
2102 set_bit(R5_Wantdrain, &dev->flags);
2103 set_bit(R5_LOCKED, &dev->flags);
2104 clear_bit(R5_UPTODATE, &dev->flags);
2110 /* keep the parity disk(s) locked while asynchronous operations
2113 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2114 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2118 int qd_idx = sh->qd_idx;
2119 struct r5dev *dev = &sh->dev[qd_idx];
2121 set_bit(R5_LOCKED, &dev->flags);
2122 clear_bit(R5_UPTODATE, &dev->flags);
2126 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2127 __func__, (unsigned long long)sh->sector,
2128 s->locked, s->ops_request);
2132 * Each stripe/dev can have one or more bion attached.
2133 * toread/towrite point to the first in a chain.
2134 * The bi_next chain must be in order.
2136 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2139 raid5_conf_t *conf = sh->raid_conf;
2142 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2143 (unsigned long long)bi->bi_sector,
2144 (unsigned long long)sh->sector);
2147 spin_lock_irq(&conf->device_lock);
2149 bip = &sh->dev[dd_idx].towrite;
2150 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2153 bip = &sh->dev[dd_idx].toread;
2154 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2155 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2157 bip = & (*bip)->bi_next;
2159 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2162 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2166 bi->bi_phys_segments++;
2169 /* check if page is covered */
2170 sector_t sector = sh->dev[dd_idx].sector;
2171 for (bi=sh->dev[dd_idx].towrite;
2172 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2173 bi && bi->bi_sector <= sector;
2174 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2175 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2176 sector = bi->bi_sector + (bi->bi_size>>9);
2178 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2179 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2181 spin_unlock_irq(&conf->device_lock);
2183 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2184 (unsigned long long)(*bip)->bi_sector,
2185 (unsigned long long)sh->sector, dd_idx);
2187 if (conf->mddev->bitmap && firstwrite) {
2188 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2190 sh->bm_seq = conf->seq_flush+1;
2191 set_bit(STRIPE_BIT_DELAY, &sh->state);
2196 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2197 spin_unlock_irq(&conf->device_lock);
2201 static void end_reshape(raid5_conf_t *conf);
2203 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2204 struct stripe_head *sh)
2206 int sectors_per_chunk =
2207 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2209 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2210 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2212 raid5_compute_sector(conf,
2213 stripe * (disks - conf->max_degraded)
2214 *sectors_per_chunk + chunk_offset,
2220 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2221 struct stripe_head_state *s, int disks,
2222 struct bio **return_bi)
2225 for (i = disks; i--; ) {
2229 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2232 rdev = rcu_dereference(conf->disks[i].rdev);
2233 if (rdev && test_bit(In_sync, &rdev->flags))
2234 /* multiple read failures in one stripe */
2235 md_error(conf->mddev, rdev);
2238 spin_lock_irq(&conf->device_lock);
2239 /* fail all writes first */
2240 bi = sh->dev[i].towrite;
2241 sh->dev[i].towrite = NULL;
2247 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2248 wake_up(&conf->wait_for_overlap);
2250 while (bi && bi->bi_sector <
2251 sh->dev[i].sector + STRIPE_SECTORS) {
2252 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2253 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2254 if (!raid5_dec_bi_phys_segments(bi)) {
2255 md_write_end(conf->mddev);
2256 bi->bi_next = *return_bi;
2261 /* and fail all 'written' */
2262 bi = sh->dev[i].written;
2263 sh->dev[i].written = NULL;
2264 if (bi) bitmap_end = 1;
2265 while (bi && bi->bi_sector <
2266 sh->dev[i].sector + STRIPE_SECTORS) {
2267 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2268 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2269 if (!raid5_dec_bi_phys_segments(bi)) {
2270 md_write_end(conf->mddev);
2271 bi->bi_next = *return_bi;
2277 /* fail any reads if this device is non-operational and
2278 * the data has not reached the cache yet.
2280 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2281 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2282 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2283 bi = sh->dev[i].toread;
2284 sh->dev[i].toread = NULL;
2285 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2286 wake_up(&conf->wait_for_overlap);
2287 if (bi) s->to_read--;
2288 while (bi && bi->bi_sector <
2289 sh->dev[i].sector + STRIPE_SECTORS) {
2290 struct bio *nextbi =
2291 r5_next_bio(bi, sh->dev[i].sector);
2292 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2293 if (!raid5_dec_bi_phys_segments(bi)) {
2294 bi->bi_next = *return_bi;
2300 spin_unlock_irq(&conf->device_lock);
2302 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2303 STRIPE_SECTORS, 0, 0);
2306 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2307 if (atomic_dec_and_test(&conf->pending_full_writes))
2308 md_wakeup_thread(conf->mddev->thread);
2311 /* fetch_block - checks the given member device to see if its data needs
2312 * to be read or computed to satisfy a request.
2314 * Returns 1 when no more member devices need to be checked, otherwise returns
2315 * 0 to tell the loop in handle_stripe_fill to continue
2317 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2318 int disk_idx, int disks)
2320 struct r5dev *dev = &sh->dev[disk_idx];
2321 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2322 &sh->dev[s->failed_num[1]] };
2324 /* is the data in this block needed, and can we get it? */
2325 if (!test_bit(R5_LOCKED, &dev->flags) &&
2326 !test_bit(R5_UPTODATE, &dev->flags) &&
2328 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2329 s->syncing || s->expanding ||
2330 (s->failed >= 1 && fdev[0]->toread) ||
2331 (s->failed >= 2 && fdev[1]->toread) ||
2332 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2333 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2334 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2335 /* we would like to get this block, possibly by computing it,
2336 * otherwise read it if the backing disk is insync
2338 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2339 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2340 if ((s->uptodate == disks - 1) &&
2341 (s->failed && (disk_idx == s->failed_num[0] ||
2342 disk_idx == s->failed_num[1]))) {
2343 /* have disk failed, and we're requested to fetch it;
2346 pr_debug("Computing stripe %llu block %d\n",
2347 (unsigned long long)sh->sector, disk_idx);
2348 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2349 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2350 set_bit(R5_Wantcompute, &dev->flags);
2351 sh->ops.target = disk_idx;
2352 sh->ops.target2 = -1; /* no 2nd target */
2354 /* Careful: from this point on 'uptodate' is in the eye
2355 * of raid_run_ops which services 'compute' operations
2356 * before writes. R5_Wantcompute flags a block that will
2357 * be R5_UPTODATE by the time it is needed for a
2358 * subsequent operation.
2362 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2363 /* Computing 2-failure is *very* expensive; only
2364 * do it if failed >= 2
2367 for (other = disks; other--; ) {
2368 if (other == disk_idx)
2370 if (!test_bit(R5_UPTODATE,
2371 &sh->dev[other].flags))
2375 pr_debug("Computing stripe %llu blocks %d,%d\n",
2376 (unsigned long long)sh->sector,
2378 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2379 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2380 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2381 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2382 sh->ops.target = disk_idx;
2383 sh->ops.target2 = other;
2387 } else if (test_bit(R5_Insync, &dev->flags)) {
2388 set_bit(R5_LOCKED, &dev->flags);
2389 set_bit(R5_Wantread, &dev->flags);
2391 pr_debug("Reading block %d (sync=%d)\n",
2392 disk_idx, s->syncing);
2400 * handle_stripe_fill - read or compute data to satisfy pending requests.
2402 static void handle_stripe_fill(struct stripe_head *sh,
2403 struct stripe_head_state *s,
2408 /* look for blocks to read/compute, skip this if a compute
2409 * is already in flight, or if the stripe contents are in the
2410 * midst of changing due to a write
2412 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2413 !sh->reconstruct_state)
2414 for (i = disks; i--; )
2415 if (fetch_block(sh, s, i, disks))
2417 set_bit(STRIPE_HANDLE, &sh->state);
2421 /* handle_stripe_clean_event
2422 * any written block on an uptodate or failed drive can be returned.
2423 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2424 * never LOCKED, so we don't need to test 'failed' directly.
2426 static void handle_stripe_clean_event(raid5_conf_t *conf,
2427 struct stripe_head *sh, int disks, struct bio **return_bi)
2432 for (i = disks; i--; )
2433 if (sh->dev[i].written) {
2435 if (!test_bit(R5_LOCKED, &dev->flags) &&
2436 test_bit(R5_UPTODATE, &dev->flags)) {
2437 /* We can return any write requests */
2438 struct bio *wbi, *wbi2;
2440 pr_debug("Return write for disc %d\n", i);
2441 spin_lock_irq(&conf->device_lock);
2443 dev->written = NULL;
2444 while (wbi && wbi->bi_sector <
2445 dev->sector + STRIPE_SECTORS) {
2446 wbi2 = r5_next_bio(wbi, dev->sector);
2447 if (!raid5_dec_bi_phys_segments(wbi)) {
2448 md_write_end(conf->mddev);
2449 wbi->bi_next = *return_bi;
2454 if (dev->towrite == NULL)
2456 spin_unlock_irq(&conf->device_lock);
2458 bitmap_endwrite(conf->mddev->bitmap,
2461 !test_bit(STRIPE_DEGRADED, &sh->state),
2466 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2467 if (atomic_dec_and_test(&conf->pending_full_writes))
2468 md_wakeup_thread(conf->mddev->thread);
2471 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2472 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2474 int rmw = 0, rcw = 0, i;
2475 for (i = disks; i--; ) {
2476 /* would I have to read this buffer for read_modify_write */
2477 struct r5dev *dev = &sh->dev[i];
2478 if ((dev->towrite || i == sh->pd_idx) &&
2479 !test_bit(R5_LOCKED, &dev->flags) &&
2480 !(test_bit(R5_UPTODATE, &dev->flags) ||
2481 test_bit(R5_Wantcompute, &dev->flags))) {
2482 if (test_bit(R5_Insync, &dev->flags))
2485 rmw += 2*disks; /* cannot read it */
2487 /* Would I have to read this buffer for reconstruct_write */
2488 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2489 !test_bit(R5_LOCKED, &dev->flags) &&
2490 !(test_bit(R5_UPTODATE, &dev->flags) ||
2491 test_bit(R5_Wantcompute, &dev->flags))) {
2492 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2497 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2498 (unsigned long long)sh->sector, rmw, rcw);
2499 set_bit(STRIPE_HANDLE, &sh->state);
2500 if (rmw < rcw && rmw > 0)
2501 /* prefer read-modify-write, but need to get some data */
2502 for (i = disks; i--; ) {
2503 struct r5dev *dev = &sh->dev[i];
2504 if ((dev->towrite || i == sh->pd_idx) &&
2505 !test_bit(R5_LOCKED, &dev->flags) &&
2506 !(test_bit(R5_UPTODATE, &dev->flags) ||
2507 test_bit(R5_Wantcompute, &dev->flags)) &&
2508 test_bit(R5_Insync, &dev->flags)) {
2510 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2511 pr_debug("Read_old block "
2512 "%d for r-m-w\n", i);
2513 set_bit(R5_LOCKED, &dev->flags);
2514 set_bit(R5_Wantread, &dev->flags);
2517 set_bit(STRIPE_DELAYED, &sh->state);
2518 set_bit(STRIPE_HANDLE, &sh->state);
2522 if (rcw <= rmw && rcw > 0)
2523 /* want reconstruct write, but need to get some data */
2524 for (i = disks; i--; ) {
2525 struct r5dev *dev = &sh->dev[i];
2526 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2528 !test_bit(R5_LOCKED, &dev->flags) &&
2529 !(test_bit(R5_UPTODATE, &dev->flags) ||
2530 test_bit(R5_Wantcompute, &dev->flags)) &&
2531 test_bit(R5_Insync, &dev->flags)) {
2533 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2534 pr_debug("Read_old block "
2535 "%d for Reconstruct\n", i);
2536 set_bit(R5_LOCKED, &dev->flags);
2537 set_bit(R5_Wantread, &dev->flags);
2540 set_bit(STRIPE_DELAYED, &sh->state);
2541 set_bit(STRIPE_HANDLE, &sh->state);
2545 /* now if nothing is locked, and if we have enough data,
2546 * we can start a write request
2548 /* since handle_stripe can be called at any time we need to handle the
2549 * case where a compute block operation has been submitted and then a
2550 * subsequent call wants to start a write request. raid_run_ops only
2551 * handles the case where compute block and reconstruct are requested
2552 * simultaneously. If this is not the case then new writes need to be
2553 * held off until the compute completes.
2555 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2556 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2557 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2558 schedule_reconstruction(sh, s, rcw == 0, 0);
2561 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2562 struct stripe_head *sh, struct stripe_head_state *s,
2565 int rcw = 0, pd_idx = sh->pd_idx, i;
2566 int qd_idx = sh->qd_idx;
2568 set_bit(STRIPE_HANDLE, &sh->state);
2569 for (i = disks; i--; ) {
2570 struct r5dev *dev = &sh->dev[i];
2571 /* check if we haven't enough data */
2572 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2573 i != pd_idx && i != qd_idx &&
2574 !test_bit(R5_LOCKED, &dev->flags) &&
2575 !(test_bit(R5_UPTODATE, &dev->flags) ||
2576 test_bit(R5_Wantcompute, &dev->flags))) {
2578 if (!test_bit(R5_Insync, &dev->flags))
2579 continue; /* it's a failed drive */
2582 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2583 pr_debug("Read_old stripe %llu "
2584 "block %d for Reconstruct\n",
2585 (unsigned long long)sh->sector, i);
2586 set_bit(R5_LOCKED, &dev->flags);
2587 set_bit(R5_Wantread, &dev->flags);
2590 pr_debug("Request delayed stripe %llu "
2591 "block %d for Reconstruct\n",
2592 (unsigned long long)sh->sector, i);
2593 set_bit(STRIPE_DELAYED, &sh->state);
2594 set_bit(STRIPE_HANDLE, &sh->state);
2598 /* now if nothing is locked, and if we have enough data, we can start a
2601 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2602 s->locked == 0 && rcw == 0 &&
2603 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2604 schedule_reconstruction(sh, s, 1, 0);
2608 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2609 struct stripe_head_state *s, int disks)
2611 struct r5dev *dev = NULL;
2613 set_bit(STRIPE_HANDLE, &sh->state);
2615 switch (sh->check_state) {
2616 case check_state_idle:
2617 /* start a new check operation if there are no failures */
2618 if (s->failed == 0) {
2619 BUG_ON(s->uptodate != disks);
2620 sh->check_state = check_state_run;
2621 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2622 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2626 dev = &sh->dev[s->failed_num[0]];
2628 case check_state_compute_result:
2629 sh->check_state = check_state_idle;
2631 dev = &sh->dev[sh->pd_idx];
2633 /* check that a write has not made the stripe insync */
2634 if (test_bit(STRIPE_INSYNC, &sh->state))
2637 /* either failed parity check, or recovery is happening */
2638 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2639 BUG_ON(s->uptodate != disks);
2641 set_bit(R5_LOCKED, &dev->flags);
2643 set_bit(R5_Wantwrite, &dev->flags);
2645 clear_bit(STRIPE_DEGRADED, &sh->state);
2646 set_bit(STRIPE_INSYNC, &sh->state);
2648 case check_state_run:
2649 break; /* we will be called again upon completion */
2650 case check_state_check_result:
2651 sh->check_state = check_state_idle;
2653 /* if a failure occurred during the check operation, leave
2654 * STRIPE_INSYNC not set and let the stripe be handled again
2659 /* handle a successful check operation, if parity is correct
2660 * we are done. Otherwise update the mismatch count and repair
2661 * parity if !MD_RECOVERY_CHECK
2663 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2664 /* parity is correct (on disc,
2665 * not in buffer any more)
2667 set_bit(STRIPE_INSYNC, &sh->state);
2669 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2670 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2671 /* don't try to repair!! */
2672 set_bit(STRIPE_INSYNC, &sh->state);
2674 sh->check_state = check_state_compute_run;
2675 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2676 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2677 set_bit(R5_Wantcompute,
2678 &sh->dev[sh->pd_idx].flags);
2679 sh->ops.target = sh->pd_idx;
2680 sh->ops.target2 = -1;
2685 case check_state_compute_run:
2688 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2689 __func__, sh->check_state,
2690 (unsigned long long) sh->sector);
2696 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2697 struct stripe_head_state *s,
2700 int pd_idx = sh->pd_idx;
2701 int qd_idx = sh->qd_idx;
2704 set_bit(STRIPE_HANDLE, &sh->state);
2706 BUG_ON(s->failed > 2);
2708 /* Want to check and possibly repair P and Q.
2709 * However there could be one 'failed' device, in which
2710 * case we can only check one of them, possibly using the
2711 * other to generate missing data
2714 switch (sh->check_state) {
2715 case check_state_idle:
2716 /* start a new check operation if there are < 2 failures */
2717 if (s->failed == s->q_failed) {
2718 /* The only possible failed device holds Q, so it
2719 * makes sense to check P (If anything else were failed,
2720 * we would have used P to recreate it).
2722 sh->check_state = check_state_run;
2724 if (!s->q_failed && s->failed < 2) {
2725 /* Q is not failed, and we didn't use it to generate
2726 * anything, so it makes sense to check it
2728 if (sh->check_state == check_state_run)
2729 sh->check_state = check_state_run_pq;
2731 sh->check_state = check_state_run_q;
2734 /* discard potentially stale zero_sum_result */
2735 sh->ops.zero_sum_result = 0;
2737 if (sh->check_state == check_state_run) {
2738 /* async_xor_zero_sum destroys the contents of P */
2739 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2742 if (sh->check_state >= check_state_run &&
2743 sh->check_state <= check_state_run_pq) {
2744 /* async_syndrome_zero_sum preserves P and Q, so
2745 * no need to mark them !uptodate here
2747 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2751 /* we have 2-disk failure */
2752 BUG_ON(s->failed != 2);
2754 case check_state_compute_result:
2755 sh->check_state = check_state_idle;
2757 /* check that a write has not made the stripe insync */
2758 if (test_bit(STRIPE_INSYNC, &sh->state))
2761 /* now write out any block on a failed drive,
2762 * or P or Q if they were recomputed
2764 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2765 if (s->failed == 2) {
2766 dev = &sh->dev[s->failed_num[1]];
2768 set_bit(R5_LOCKED, &dev->flags);
2769 set_bit(R5_Wantwrite, &dev->flags);
2771 if (s->failed >= 1) {
2772 dev = &sh->dev[s->failed_num[0]];
2774 set_bit(R5_LOCKED, &dev->flags);
2775 set_bit(R5_Wantwrite, &dev->flags);
2777 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2778 dev = &sh->dev[pd_idx];
2780 set_bit(R5_LOCKED, &dev->flags);
2781 set_bit(R5_Wantwrite, &dev->flags);
2783 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2784 dev = &sh->dev[qd_idx];
2786 set_bit(R5_LOCKED, &dev->flags);
2787 set_bit(R5_Wantwrite, &dev->flags);
2789 clear_bit(STRIPE_DEGRADED, &sh->state);
2791 set_bit(STRIPE_INSYNC, &sh->state);
2793 case check_state_run:
2794 case check_state_run_q:
2795 case check_state_run_pq:
2796 break; /* we will be called again upon completion */
2797 case check_state_check_result:
2798 sh->check_state = check_state_idle;
2800 /* handle a successful check operation, if parity is correct
2801 * we are done. Otherwise update the mismatch count and repair
2802 * parity if !MD_RECOVERY_CHECK
2804 if (sh->ops.zero_sum_result == 0) {
2805 /* both parities are correct */
2807 set_bit(STRIPE_INSYNC, &sh->state);
2809 /* in contrast to the raid5 case we can validate
2810 * parity, but still have a failure to write
2813 sh->check_state = check_state_compute_result;
2814 /* Returning at this point means that we may go
2815 * off and bring p and/or q uptodate again so
2816 * we make sure to check zero_sum_result again
2817 * to verify if p or q need writeback
2821 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2822 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2823 /* don't try to repair!! */
2824 set_bit(STRIPE_INSYNC, &sh->state);
2826 int *target = &sh->ops.target;
2828 sh->ops.target = -1;
2829 sh->ops.target2 = -1;
2830 sh->check_state = check_state_compute_run;
2831 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2832 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2833 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2834 set_bit(R5_Wantcompute,
2835 &sh->dev[pd_idx].flags);
2837 target = &sh->ops.target2;
2840 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2841 set_bit(R5_Wantcompute,
2842 &sh->dev[qd_idx].flags);
2849 case check_state_compute_run:
2852 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2853 __func__, sh->check_state,
2854 (unsigned long long) sh->sector);
2859 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh)
2863 /* We have read all the blocks in this stripe and now we need to
2864 * copy some of them into a target stripe for expand.
2866 struct dma_async_tx_descriptor *tx = NULL;
2867 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2868 for (i = 0; i < sh->disks; i++)
2869 if (i != sh->pd_idx && i != sh->qd_idx) {
2871 struct stripe_head *sh2;
2872 struct async_submit_ctl submit;
2874 sector_t bn = compute_blocknr(sh, i, 1);
2875 sector_t s = raid5_compute_sector(conf, bn, 0,
2877 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2879 /* so far only the early blocks of this stripe
2880 * have been requested. When later blocks
2881 * get requested, we will try again
2884 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2885 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2886 /* must have already done this block */
2887 release_stripe(sh2);
2891 /* place all the copies on one channel */
2892 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2893 tx = async_memcpy(sh2->dev[dd_idx].page,
2894 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2897 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2898 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2899 for (j = 0; j < conf->raid_disks; j++)
2900 if (j != sh2->pd_idx &&
2902 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2904 if (j == conf->raid_disks) {
2905 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2906 set_bit(STRIPE_HANDLE, &sh2->state);
2908 release_stripe(sh2);
2911 /* done submitting copies, wait for them to complete */
2914 dma_wait_for_async_tx(tx);
2920 * handle_stripe - do things to a stripe.
2922 * We lock the stripe and then examine the state of various bits
2923 * to see what needs to be done.
2925 * return some read request which now have data
2926 * return some write requests which are safely on disc
2927 * schedule a read on some buffers
2928 * schedule a write of some buffers
2929 * return confirmation of parity correctness
2931 * buffers are taken off read_list or write_list, and bh_cache buffers
2932 * get BH_Lock set before the stripe lock is released.
2936 static int handle_stripe5(struct stripe_head *sh, struct stripe_head_state *s)
2938 raid5_conf_t *conf = sh->raid_conf;
2939 int disks = sh->disks, i;
2943 /* Now to look around and see what can be done */
2945 spin_lock_irq(&conf->device_lock);
2946 for (i=disks; i--; ) {
2951 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2952 "written %p\n", i, dev->flags, dev->toread, dev->read,
2953 dev->towrite, dev->written);
2955 /* maybe we can request a biofill operation
2957 * new wantfill requests are only permitted while
2958 * ops_complete_biofill is guaranteed to be inactive
2960 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2961 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2962 set_bit(R5_Wantfill, &dev->flags);
2964 /* now count some things */
2965 if (test_bit(R5_LOCKED, &dev->flags))
2967 if (test_bit(R5_UPTODATE, &dev->flags))
2969 if (test_bit(R5_Wantcompute, &dev->flags))
2972 if (test_bit(R5_Wantfill, &dev->flags))
2974 else if (dev->toread)
2978 if (!test_bit(R5_OVERWRITE, &dev->flags))
2983 rdev = rcu_dereference(conf->disks[i].rdev);
2984 if (s->blocked_rdev == NULL &&
2985 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2986 s->blocked_rdev = rdev;
2987 atomic_inc(&rdev->nr_pending);
2989 clear_bit(R5_Insync, &dev->flags);
2992 else if (test_bit(In_sync, &rdev->flags))
2993 set_bit(R5_Insync, &dev->flags);
2995 /* could be in-sync depending on recovery/reshape status */
2996 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
2997 set_bit(R5_Insync, &dev->flags);
2999 if (!test_bit(R5_Insync, &dev->flags)) {
3000 /* The ReadError flag will just be confusing now */
3001 clear_bit(R5_ReadError, &dev->flags);
3002 clear_bit(R5_ReWrite, &dev->flags);
3004 if (test_bit(R5_ReadError, &dev->flags))
3005 clear_bit(R5_Insync, &dev->flags);
3006 if (!test_bit(R5_Insync, &dev->flags)) {
3008 s->failed_num[0] = i;
3011 spin_unlock_irq(&conf->device_lock);
3014 if (unlikely(s->blocked_rdev)) {
3015 if (s->syncing || s->expanding || s->expanded ||
3016 s->to_write || s->written) {
3017 set_bit(STRIPE_HANDLE, &sh->state);
3020 /* There is nothing for the blocked_rdev to block */
3021 rdev_dec_pending(s->blocked_rdev, conf->mddev);
3022 s->blocked_rdev = NULL;
3025 if (s->to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3026 set_bit(STRIPE_OP_BIOFILL, &s->ops_request);
3027 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3030 pr_debug("locked=%d uptodate=%d to_read=%d"
3031 " to_write=%d failed=%d failed_num=%d\n",
3032 s->locked, s->uptodate, s->to_read, s->to_write,
3033 s->failed, s->failed_num[0]);
3034 /* check if the array has lost two devices and, if so, some requests might
3037 if (s->failed > 1 && s->to_read+s->to_write+s->written)
3038 handle_failed_stripe(conf, sh, s, disks, &s->return_bi);
3039 if (s->failed > 1 && s->syncing) {
3040 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3041 clear_bit(STRIPE_SYNCING, &sh->state);
3045 /* might be able to return some write requests if the parity block
3046 * is safe, or on a failed drive
3048 dev = &sh->dev[sh->pd_idx];
3050 ((test_bit(R5_Insync, &dev->flags) &&
3051 !test_bit(R5_LOCKED, &dev->flags) &&
3052 test_bit(R5_UPTODATE, &dev->flags)) ||
3053 (s->failed == 1 && s->failed_num[0] == sh->pd_idx)))
3054 handle_stripe_clean_event(conf, sh, disks, &s->return_bi);
3056 /* Now we might consider reading some blocks, either to check/generate
3057 * parity, or to satisfy requests
3058 * or to load a block that is being partially written.
3060 if (s->to_read || s->non_overwrite ||
3061 (s->syncing && (s->uptodate + s->compute < disks)) || s->expanding)
3062 handle_stripe_fill(sh, s, disks);
3064 /* Now we check to see if any write operations have recently
3068 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3070 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3071 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3072 sh->reconstruct_state = reconstruct_state_idle;
3074 /* All the 'written' buffers and the parity block are ready to
3075 * be written back to disk
3077 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3078 for (i = disks; i--; ) {
3080 if (test_bit(R5_LOCKED, &dev->flags) &&
3081 (i == sh->pd_idx || dev->written)) {
3082 pr_debug("Writing block %d\n", i);
3083 set_bit(R5_Wantwrite, &dev->flags);
3086 if (!test_bit(R5_Insync, &dev->flags) ||
3087 (i == sh->pd_idx && s->failed == 0))
3088 set_bit(STRIPE_INSYNC, &sh->state);
3091 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3092 s->dec_preread_active = 1;
3095 /* Now to consider new write requests and what else, if anything
3096 * should be read. We do not handle new writes when:
3097 * 1/ A 'write' operation (copy+xor) is already in flight.
3098 * 2/ A 'check' operation is in flight, as it may clobber the parity
3101 if (s->to_write && !sh->reconstruct_state && !sh->check_state)
3102 handle_stripe_dirtying5(conf, sh, s, disks);
3104 /* maybe we need to check and possibly fix the parity for this stripe
3105 * Any reads will already have been scheduled, so we just see if enough
3106 * data is available. The parity check is held off while parity
3107 * dependent operations are in flight.
3109 if (sh->check_state ||
3110 (s->syncing && s->locked == 0 &&
3111 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3112 !test_bit(STRIPE_INSYNC, &sh->state)))
3113 handle_parity_checks5(conf, sh, s, disks);
3117 static int handle_stripe6(struct stripe_head *sh, struct stripe_head_state *s)
3119 raid5_conf_t *conf = sh->raid_conf;
3120 int disks = sh->disks;
3121 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3122 struct r5dev *dev, *pdev, *qdev;
3124 /* Now to look around and see what can be done */
3127 spin_lock_irq(&conf->device_lock);
3128 for (i=disks; i--; ) {
3132 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3133 i, dev->flags, dev->toread, dev->towrite, dev->written);
3134 /* maybe we can reply to a read
3136 * new wantfill requests are only permitted while
3137 * ops_complete_biofill is guaranteed to be inactive
3139 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3140 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3141 set_bit(R5_Wantfill, &dev->flags);
3143 /* now count some things */
3144 if (test_bit(R5_LOCKED, &dev->flags))
3146 if (test_bit(R5_UPTODATE, &dev->flags))
3148 if (test_bit(R5_Wantcompute, &dev->flags)) {
3150 BUG_ON(s->compute > 2);
3153 if (test_bit(R5_Wantfill, &dev->flags)) {
3155 } else if (dev->toread)
3159 if (!test_bit(R5_OVERWRITE, &dev->flags))
3164 rdev = rcu_dereference(conf->disks[i].rdev);
3165 if (s->blocked_rdev == NULL &&
3166 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3167 s->blocked_rdev = rdev;
3168 atomic_inc(&rdev->nr_pending);
3170 clear_bit(R5_Insync, &dev->flags);
3173 else if (test_bit(In_sync, &rdev->flags))
3174 set_bit(R5_Insync, &dev->flags);
3176 /* in sync if before recovery_offset */
3177 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3178 set_bit(R5_Insync, &dev->flags);
3180 if (!test_bit(R5_Insync, &dev->flags)) {
3181 /* The ReadError flag will just be confusing now */
3182 clear_bit(R5_ReadError, &dev->flags);
3183 clear_bit(R5_ReWrite, &dev->flags);
3185 if (test_bit(R5_ReadError, &dev->flags))
3186 clear_bit(R5_Insync, &dev->flags);
3187 if (!test_bit(R5_Insync, &dev->flags)) {
3189 s->failed_num[s->failed] = i;
3193 spin_unlock_irq(&conf->device_lock);
3196 if (unlikely(s->blocked_rdev)) {
3197 if (s->syncing || s->expanding || s->expanded ||
3198 s->to_write || s->written) {
3199 set_bit(STRIPE_HANDLE, &sh->state);
3202 /* There is nothing for the blocked_rdev to block */
3203 rdev_dec_pending(s->blocked_rdev, conf->mddev);
3204 s->blocked_rdev = NULL;
3207 if (s->to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3208 set_bit(STRIPE_OP_BIOFILL, &s->ops_request);
3209 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3212 pr_debug("locked=%d uptodate=%d to_read=%d"
3213 " to_write=%d failed=%d failed_num=%d,%d\n",
3214 s->locked, s->uptodate, s->to_read, s->to_write, s->failed,
3215 s->failed_num[0], s->failed_num[1]);
3216 /* check if the array has lost >2 devices and, if so, some requests
3217 * might need to be failed
3219 if (s->failed > 2 && s->to_read+s->to_write+s->written)
3220 handle_failed_stripe(conf, sh, s, disks, &s->return_bi);
3221 if (s->failed > 2 && s->syncing) {
3222 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3223 clear_bit(STRIPE_SYNCING, &sh->state);
3228 * might be able to return some write requests if the parity blocks
3229 * are safe, or on a failed drive
3231 pdev = &sh->dev[pd_idx];
3232 s->p_failed = (s->failed >= 1 && s->failed_num[0] == pd_idx)
3233 || (s->failed >= 2 && s->failed_num[1] == pd_idx);
3234 qdev = &sh->dev[qd_idx];
3235 s->q_failed = (s->failed >= 1 && s->failed_num[0] == qd_idx)
3236 || (s->failed >= 2 && s->failed_num[1] == qd_idx);
3239 (s->p_failed || ((test_bit(R5_Insync, &pdev->flags)
3240 && !test_bit(R5_LOCKED, &pdev->flags)
3241 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3242 (s->q_failed || ((test_bit(R5_Insync, &qdev->flags)
3243 && !test_bit(R5_LOCKED, &qdev->flags)
3244 && test_bit(R5_UPTODATE, &qdev->flags)))))
3245 handle_stripe_clean_event(conf, sh, disks, &s->return_bi);
3247 /* Now we might consider reading some blocks, either to check/generate
3248 * parity, or to satisfy requests
3249 * or to load a block that is being partially written.
3251 if (s->to_read || s->non_overwrite || (s->to_write && s->failed) ||
3252 (s->syncing && (s->uptodate + s->compute < disks)) || s->expanding)
3253 handle_stripe_fill(sh, s, disks);
3255 /* Now we check to see if any write operations have recently
3258 if (sh->reconstruct_state == reconstruct_state_drain_result) {
3260 sh->reconstruct_state = reconstruct_state_idle;
3261 /* All the 'written' buffers and the parity blocks are ready to
3262 * be written back to disk
3264 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3265 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3266 for (i = disks; i--; ) {
3268 if (test_bit(R5_LOCKED, &dev->flags) &&
3269 (i == sh->pd_idx || i == qd_idx ||
3271 pr_debug("Writing block %d\n", i);
3272 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3273 set_bit(R5_Wantwrite, &dev->flags);
3274 if (!test_bit(R5_Insync, &dev->flags) ||
3275 ((i == sh->pd_idx || i == qd_idx) &&
3277 set_bit(STRIPE_INSYNC, &sh->state);
3280 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3281 s->dec_preread_active = 1;
3284 /* Now to consider new write requests and what else, if anything
3285 * should be read. We do not handle new writes when:
3286 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3287 * 2/ A 'check' operation is in flight, as it may clobber the parity
3290 if (s->to_write && !sh->reconstruct_state && !sh->check_state)
3291 handle_stripe_dirtying6(conf, sh, s, disks);
3293 /* maybe we need to check and possibly fix the parity for this stripe
3294 * Any reads will already have been scheduled, so we just see if enough
3295 * data is available. The parity check is held off while parity
3296 * dependent operations are in flight.
3298 if (sh->check_state ||
3299 (s->syncing && s->locked == 0 &&
3300 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3301 !test_bit(STRIPE_INSYNC, &sh->state)))
3302 handle_parity_checks6(conf, sh, s, disks);
3306 static void handle_stripe(struct stripe_head *sh)
3308 struct stripe_head_state s;
3311 raid5_conf_t *conf = sh->raid_conf;
3313 clear_bit(STRIPE_HANDLE, &sh->state);
3314 if (test_and_set_bit(STRIPE_ACTIVE, &sh->state)) {
3315 /* already being handled, ensure it gets handled
3316 * again when current action finishes */
3317 set_bit(STRIPE_HANDLE, &sh->state);
3321 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3322 set_bit(STRIPE_SYNCING, &sh->state);
3323 clear_bit(STRIPE_INSYNC, &sh->state);
3325 clear_bit(STRIPE_DELAYED, &sh->state);
3327 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3328 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3329 (unsigned long long)sh->sector, sh->state,
3330 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3331 sh->check_state, sh->reconstruct_state);
3332 memset(&s, 0, sizeof(s));
3334 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3335 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3336 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3337 s.failed_num[0] = -1;
3338 s.failed_num[1] = -1;
3340 if (conf->level == 6)
3341 done = handle_stripe6(sh, &s);
3343 done = handle_stripe5(sh, &s);
3349 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3350 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3351 clear_bit(STRIPE_SYNCING, &sh->state);
3354 /* If the failed drives are just a ReadError, then we might need
3355 * to progress the repair/check process
3357 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3358 for (i = 0; i < s.failed; i++) {
3359 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3360 if (test_bit(R5_ReadError, &dev->flags)
3361 && !test_bit(R5_LOCKED, &dev->flags)
3362 && test_bit(R5_UPTODATE, &dev->flags)
3364 if (!test_bit(R5_ReWrite, &dev->flags)) {
3365 set_bit(R5_Wantwrite, &dev->flags);
3366 set_bit(R5_ReWrite, &dev->flags);
3367 set_bit(R5_LOCKED, &dev->flags);
3370 /* let's read it back */
3371 set_bit(R5_Wantread, &dev->flags);
3372 set_bit(R5_LOCKED, &dev->flags);
3379 /* Finish reconstruct operations initiated by the expansion process */
3380 if (sh->reconstruct_state == reconstruct_state_result) {
3381 struct stripe_head *sh_src
3382 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3383 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3384 /* sh cannot be written until sh_src has been read.
3385 * so arrange for sh to be delayed a little
3387 set_bit(STRIPE_DELAYED, &sh->state);
3388 set_bit(STRIPE_HANDLE, &sh->state);
3389 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3391 atomic_inc(&conf->preread_active_stripes);
3392 release_stripe(sh_src);
3396 release_stripe(sh_src);
3398 sh->reconstruct_state = reconstruct_state_idle;
3399 clear_bit(STRIPE_EXPANDING, &sh->state);
3400 for (i = conf->raid_disks; i--; ) {
3401 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3402 set_bit(R5_LOCKED, &sh->dev[i].flags);
3407 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3408 !sh->reconstruct_state) {
3409 /* Need to write out all blocks after computing parity */
3410 sh->disks = conf->raid_disks;
3411 stripe_set_idx(sh->sector, conf, 0, sh);
3412 schedule_reconstruction(sh, &s, 1, 1);
3413 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3414 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3415 atomic_dec(&conf->reshape_stripes);
3416 wake_up(&conf->wait_for_overlap);
3417 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3420 if (s.expanding && s.locked == 0 &&
3421 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3422 handle_stripe_expansion(conf, sh);
3425 /* wait for this device to become unblocked */
3426 if (unlikely(s.blocked_rdev))
3427 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3430 raid_run_ops(sh, s.ops_request);
3435 if (s.dec_preread_active) {
3436 /* We delay this until after ops_run_io so that if make_request
3437 * is waiting on a flush, it won't continue until the writes
3438 * have actually been submitted.
3440 atomic_dec(&conf->preread_active_stripes);
3441 if (atomic_read(&conf->preread_active_stripes) <
3443 md_wakeup_thread(conf->mddev->thread);
3446 return_io(s.return_bi);
3448 clear_bit(STRIPE_ACTIVE, &sh->state);
3451 static void raid5_activate_delayed(raid5_conf_t *conf)
3453 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3454 while (!list_empty(&conf->delayed_list)) {
3455 struct list_head *l = conf->delayed_list.next;
3456 struct stripe_head *sh;
3457 sh = list_entry(l, struct stripe_head, lru);
3459 clear_bit(STRIPE_DELAYED, &sh->state);
3460 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3461 atomic_inc(&conf->preread_active_stripes);
3462 list_add_tail(&sh->lru, &conf->hold_list);
3467 static void activate_bit_delay(raid5_conf_t *conf)
3469 /* device_lock is held */
3470 struct list_head head;
3471 list_add(&head, &conf->bitmap_list);
3472 list_del_init(&conf->bitmap_list);
3473 while (!list_empty(&head)) {
3474 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3475 list_del_init(&sh->lru);
3476 atomic_inc(&sh->count);
3477 __release_stripe(conf, sh);
3481 int md_raid5_congested(mddev_t *mddev, int bits)
3483 raid5_conf_t *conf = mddev->private;
3485 /* No difference between reads and writes. Just check
3486 * how busy the stripe_cache is
3489 if (conf->inactive_blocked)
3493 if (list_empty_careful(&conf->inactive_list))
3498 EXPORT_SYMBOL_GPL(md_raid5_congested);
3500 static int raid5_congested(void *data, int bits)
3502 mddev_t *mddev = data;
3504 return mddev_congested(mddev, bits) ||
3505 md_raid5_congested(mddev, bits);
3508 /* We want read requests to align with chunks where possible,
3509 * but write requests don't need to.
3511 static int raid5_mergeable_bvec(struct request_queue *q,
3512 struct bvec_merge_data *bvm,
3513 struct bio_vec *biovec)
3515 mddev_t *mddev = q->queuedata;
3516 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3518 unsigned int chunk_sectors = mddev->chunk_sectors;
3519 unsigned int bio_sectors = bvm->bi_size >> 9;
3521 if ((bvm->bi_rw & 1) == WRITE)
3522 return biovec->bv_len; /* always allow writes to be mergeable */
3524 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3525 chunk_sectors = mddev->new_chunk_sectors;
3526 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3527 if (max < 0) max = 0;
3528 if (max <= biovec->bv_len && bio_sectors == 0)
3529 return biovec->bv_len;
3535 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3537 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3538 unsigned int chunk_sectors = mddev->chunk_sectors;
3539 unsigned int bio_sectors = bio->bi_size >> 9;
3541 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3542 chunk_sectors = mddev->new_chunk_sectors;
3543 return chunk_sectors >=
3544 ((sector & (chunk_sectors - 1)) + bio_sectors);
3548 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3549 * later sampled by raid5d.
3551 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3553 unsigned long flags;
3555 spin_lock_irqsave(&conf->device_lock, flags);
3557 bi->bi_next = conf->retry_read_aligned_list;
3558 conf->retry_read_aligned_list = bi;
3560 spin_unlock_irqrestore(&conf->device_lock, flags);
3561 md_wakeup_thread(conf->mddev->thread);
3565 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3569 bi = conf->retry_read_aligned;
3571 conf->retry_read_aligned = NULL;
3574 bi = conf->retry_read_aligned_list;
3576 conf->retry_read_aligned_list = bi->bi_next;
3579 * this sets the active strip count to 1 and the processed
3580 * strip count to zero (upper 8 bits)
3582 bi->bi_phys_segments = 1; /* biased count of active stripes */
3590 * The "raid5_align_endio" should check if the read succeeded and if it
3591 * did, call bio_endio on the original bio (having bio_put the new bio
3593 * If the read failed..
3595 static void raid5_align_endio(struct bio *bi, int error)
3597 struct bio* raid_bi = bi->bi_private;
3600 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3605 rdev = (void*)raid_bi->bi_next;
3606 raid_bi->bi_next = NULL;
3607 mddev = rdev->mddev;
3608 conf = mddev->private;
3610 rdev_dec_pending(rdev, conf->mddev);
3612 if (!error && uptodate) {
3613 bio_endio(raid_bi, 0);
3614 if (atomic_dec_and_test(&conf->active_aligned_reads))
3615 wake_up(&conf->wait_for_stripe);
3620 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3622 add_bio_to_retry(raid_bi, conf);
3625 static int bio_fits_rdev(struct bio *bi)
3627 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3629 if ((bi->bi_size>>9) > queue_max_sectors(q))
3631 blk_recount_segments(q, bi);
3632 if (bi->bi_phys_segments > queue_max_segments(q))
3635 if (q->merge_bvec_fn)
3636 /* it's too hard to apply the merge_bvec_fn at this stage,
3645 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio)
3647 raid5_conf_t *conf = mddev->private;
3649 struct bio* align_bi;
3652 if (!in_chunk_boundary(mddev, raid_bio)) {
3653 pr_debug("chunk_aligned_read : non aligned\n");
3657 * use bio_clone_mddev to make a copy of the bio
3659 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3663 * set bi_end_io to a new function, and set bi_private to the
3666 align_bi->bi_end_io = raid5_align_endio;
3667 align_bi->bi_private = raid_bio;
3671 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3676 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3677 if (rdev && test_bit(In_sync, &rdev->flags)) {
3678 atomic_inc(&rdev->nr_pending);
3680 raid_bio->bi_next = (void*)rdev;
3681 align_bi->bi_bdev = rdev->bdev;
3682 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3683 align_bi->bi_sector += rdev->data_offset;
3685 if (!bio_fits_rdev(align_bi)) {
3686 /* too big in some way */
3688 rdev_dec_pending(rdev, mddev);
3692 spin_lock_irq(&conf->device_lock);
3693 wait_event_lock_irq(conf->wait_for_stripe,
3695 conf->device_lock, /* nothing */);
3696 atomic_inc(&conf->active_aligned_reads);
3697 spin_unlock_irq(&conf->device_lock);
3699 generic_make_request(align_bi);
3708 /* __get_priority_stripe - get the next stripe to process
3710 * Full stripe writes are allowed to pass preread active stripes up until
3711 * the bypass_threshold is exceeded. In general the bypass_count
3712 * increments when the handle_list is handled before the hold_list; however, it
3713 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3714 * stripe with in flight i/o. The bypass_count will be reset when the
3715 * head of the hold_list has changed, i.e. the head was promoted to the
3718 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3720 struct stripe_head *sh;
3722 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3724 list_empty(&conf->handle_list) ? "empty" : "busy",
3725 list_empty(&conf->hold_list) ? "empty" : "busy",
3726 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3728 if (!list_empty(&conf->handle_list)) {
3729 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3731 if (list_empty(&conf->hold_list))
3732 conf->bypass_count = 0;
3733 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3734 if (conf->hold_list.next == conf->last_hold)
3735 conf->bypass_count++;
3737 conf->last_hold = conf->hold_list.next;
3738 conf->bypass_count -= conf->bypass_threshold;
3739 if (conf->bypass_count < 0)
3740 conf->bypass_count = 0;
3743 } else if (!list_empty(&conf->hold_list) &&
3744 ((conf->bypass_threshold &&
3745 conf->bypass_count > conf->bypass_threshold) ||
3746 atomic_read(&conf->pending_full_writes) == 0)) {
3747 sh = list_entry(conf->hold_list.next,
3749 conf->bypass_count -= conf->bypass_threshold;
3750 if (conf->bypass_count < 0)
3751 conf->bypass_count = 0;
3755 list_del_init(&sh->lru);
3756 atomic_inc(&sh->count);
3757 BUG_ON(atomic_read(&sh->count) != 1);
3761 static int make_request(mddev_t *mddev, struct bio * bi)
3763 raid5_conf_t *conf = mddev->private;
3765 sector_t new_sector;
3766 sector_t logical_sector, last_sector;
3767 struct stripe_head *sh;
3768 const int rw = bio_data_dir(bi);
3772 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3773 md_flush_request(mddev, bi);
3777 md_write_start(mddev, bi);
3780 mddev->reshape_position == MaxSector &&
3781 chunk_aligned_read(mddev,bi))
3784 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3785 last_sector = bi->bi_sector + (bi->bi_size>>9);
3787 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3789 plugged = mddev_check_plugged(mddev);
3790 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3792 int disks, data_disks;
3797 disks = conf->raid_disks;
3798 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3799 if (unlikely(conf->reshape_progress != MaxSector)) {
3800 /* spinlock is needed as reshape_progress may be
3801 * 64bit on a 32bit platform, and so it might be
3802 * possible to see a half-updated value
3803 * Of course reshape_progress could change after
3804 * the lock is dropped, so once we get a reference
3805 * to the stripe that we think it is, we will have
3808 spin_lock_irq(&conf->device_lock);
3809 if (mddev->delta_disks < 0
3810 ? logical_sector < conf->reshape_progress
3811 : logical_sector >= conf->reshape_progress) {
3812 disks = conf->previous_raid_disks;
3815 if (mddev->delta_disks < 0
3816 ? logical_sector < conf->reshape_safe
3817 : logical_sector >= conf->reshape_safe) {
3818 spin_unlock_irq(&conf->device_lock);
3823 spin_unlock_irq(&conf->device_lock);
3825 data_disks = disks - conf->max_degraded;
3827 new_sector = raid5_compute_sector(conf, logical_sector,
3830 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3831 (unsigned long long)new_sector,
3832 (unsigned long long)logical_sector);
3834 sh = get_active_stripe(conf, new_sector, previous,
3835 (bi->bi_rw&RWA_MASK), 0);
3837 if (unlikely(previous)) {
3838 /* expansion might have moved on while waiting for a
3839 * stripe, so we must do the range check again.
3840 * Expansion could still move past after this
3841 * test, but as we are holding a reference to
3842 * 'sh', we know that if that happens,
3843 * STRIPE_EXPANDING will get set and the expansion
3844 * won't proceed until we finish with the stripe.
3847 spin_lock_irq(&conf->device_lock);
3848 if (mddev->delta_disks < 0
3849 ? logical_sector >= conf->reshape_progress
3850 : logical_sector < conf->reshape_progress)
3851 /* mismatch, need to try again */
3853 spin_unlock_irq(&conf->device_lock);
3862 logical_sector >= mddev->suspend_lo &&
3863 logical_sector < mddev->suspend_hi) {
3865 /* As the suspend_* range is controlled by
3866 * userspace, we want an interruptible
3869 flush_signals(current);
3870 prepare_to_wait(&conf->wait_for_overlap,
3871 &w, TASK_INTERRUPTIBLE);
3872 if (logical_sector >= mddev->suspend_lo &&
3873 logical_sector < mddev->suspend_hi)
3878 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3879 !add_stripe_bio(sh, bi, dd_idx, rw)) {
3880 /* Stripe is busy expanding or
3881 * add failed due to overlap. Flush everything
3884 md_wakeup_thread(mddev->thread);
3889 finish_wait(&conf->wait_for_overlap, &w);
3890 set_bit(STRIPE_HANDLE, &sh->state);
3891 clear_bit(STRIPE_DELAYED, &sh->state);
3892 if ((bi->bi_rw & REQ_SYNC) &&
3893 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3894 atomic_inc(&conf->preread_active_stripes);
3897 /* cannot get stripe for read-ahead, just give-up */
3898 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3899 finish_wait(&conf->wait_for_overlap, &w);
3905 md_wakeup_thread(mddev->thread);
3907 spin_lock_irq(&conf->device_lock);
3908 remaining = raid5_dec_bi_phys_segments(bi);
3909 spin_unlock_irq(&conf->device_lock);
3910 if (remaining == 0) {
3913 md_write_end(mddev);
3921 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3923 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3925 /* reshaping is quite different to recovery/resync so it is
3926 * handled quite separately ... here.
3928 * On each call to sync_request, we gather one chunk worth of
3929 * destination stripes and flag them as expanding.
3930 * Then we find all the source stripes and request reads.
3931 * As the reads complete, handle_stripe will copy the data
3932 * into the destination stripe and release that stripe.
3934 raid5_conf_t *conf = mddev->private;
3935 struct stripe_head *sh;
3936 sector_t first_sector, last_sector;
3937 int raid_disks = conf->previous_raid_disks;
3938 int data_disks = raid_disks - conf->max_degraded;
3939 int new_data_disks = conf->raid_disks - conf->max_degraded;
3942 sector_t writepos, readpos, safepos;
3943 sector_t stripe_addr;
3944 int reshape_sectors;
3945 struct list_head stripes;
3947 if (sector_nr == 0) {
3948 /* If restarting in the middle, skip the initial sectors */
3949 if (mddev->delta_disks < 0 &&
3950 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3951 sector_nr = raid5_size(mddev, 0, 0)
3952 - conf->reshape_progress;
3953 } else if (mddev->delta_disks >= 0 &&
3954 conf->reshape_progress > 0)
3955 sector_nr = conf->reshape_progress;
3956 sector_div(sector_nr, new_data_disks);
3958 mddev->curr_resync_completed = sector_nr;
3959 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3965 /* We need to process a full chunk at a time.
3966 * If old and new chunk sizes differ, we need to process the
3969 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3970 reshape_sectors = mddev->new_chunk_sectors;
3972 reshape_sectors = mddev->chunk_sectors;
3974 /* we update the metadata when there is more than 3Meg
3975 * in the block range (that is rather arbitrary, should
3976 * probably be time based) or when the data about to be
3977 * copied would over-write the source of the data at
3978 * the front of the range.
3979 * i.e. one new_stripe along from reshape_progress new_maps
3980 * to after where reshape_safe old_maps to
3982 writepos = conf->reshape_progress;
3983 sector_div(writepos, new_data_disks);
3984 readpos = conf->reshape_progress;
3985 sector_div(readpos, data_disks);
3986 safepos = conf->reshape_safe;
3987 sector_div(safepos, data_disks);
3988 if (mddev->delta_disks < 0) {
3989 writepos -= min_t(sector_t, reshape_sectors, writepos);
3990 readpos += reshape_sectors;
3991 safepos += reshape_sectors;
3993 writepos += reshape_sectors;
3994 readpos -= min_t(sector_t, reshape_sectors, readpos);
3995 safepos -= min_t(sector_t, reshape_sectors, safepos);
3998 /* 'writepos' is the most advanced device address we might write.
3999 * 'readpos' is the least advanced device address we might read.
4000 * 'safepos' is the least address recorded in the metadata as having
4002 * If 'readpos' is behind 'writepos', then there is no way that we can
4003 * ensure safety in the face of a crash - that must be done by userspace
4004 * making a backup of the data. So in that case there is no particular
4005 * rush to update metadata.
4006 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4007 * update the metadata to advance 'safepos' to match 'readpos' so that
4008 * we can be safe in the event of a crash.
4009 * So we insist on updating metadata if safepos is behind writepos and
4010 * readpos is beyond writepos.
4011 * In any case, update the metadata every 10 seconds.
4012 * Maybe that number should be configurable, but I'm not sure it is
4013 * worth it.... maybe it could be a multiple of safemode_delay???
4015 if ((mddev->delta_disks < 0
4016 ? (safepos > writepos && readpos < writepos)
4017 : (safepos < writepos && readpos > writepos)) ||
4018 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4019 /* Cannot proceed until we've updated the superblock... */
4020 wait_event(conf->wait_for_overlap,
4021 atomic_read(&conf->reshape_stripes)==0);
4022 mddev->reshape_position = conf->reshape_progress;
4023 mddev->curr_resync_completed = sector_nr;
4024 conf->reshape_checkpoint = jiffies;
4025 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4026 md_wakeup_thread(mddev->thread);
4027 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4028 kthread_should_stop());
4029 spin_lock_irq(&conf->device_lock);
4030 conf->reshape_safe = mddev->reshape_position;
4031 spin_unlock_irq(&conf->device_lock);
4032 wake_up(&conf->wait_for_overlap);
4033 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4036 if (mddev->delta_disks < 0) {
4037 BUG_ON(conf->reshape_progress == 0);
4038 stripe_addr = writepos;
4039 BUG_ON((mddev->dev_sectors &
4040 ~((sector_t)reshape_sectors - 1))
4041 - reshape_sectors - stripe_addr
4044 BUG_ON(writepos != sector_nr + reshape_sectors);
4045 stripe_addr = sector_nr;
4047 INIT_LIST_HEAD(&stripes);
4048 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4050 int skipped_disk = 0;
4051 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4052 set_bit(STRIPE_EXPANDING, &sh->state);
4053 atomic_inc(&conf->reshape_stripes);
4054 /* If any of this stripe is beyond the end of the old
4055 * array, then we need to zero those blocks
4057 for (j=sh->disks; j--;) {
4059 if (j == sh->pd_idx)
4061 if (conf->level == 6 &&
4064 s = compute_blocknr(sh, j, 0);
4065 if (s < raid5_size(mddev, 0, 0)) {
4069 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4070 set_bit(R5_Expanded, &sh->dev[j].flags);
4071 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4073 if (!skipped_disk) {
4074 set_bit(STRIPE_EXPAND_READY, &sh->state);
4075 set_bit(STRIPE_HANDLE, &sh->state);
4077 list_add(&sh->lru, &stripes);
4079 spin_lock_irq(&conf->device_lock);
4080 if (mddev->delta_disks < 0)
4081 conf->reshape_progress -= reshape_sectors * new_data_disks;
4083 conf->reshape_progress += reshape_sectors * new_data_disks;
4084 spin_unlock_irq(&conf->device_lock);
4085 /* Ok, those stripe are ready. We can start scheduling
4086 * reads on the source stripes.
4087 * The source stripes are determined by mapping the first and last
4088 * block on the destination stripes.
4091 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4094 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4095 * new_data_disks - 1),
4097 if (last_sector >= mddev->dev_sectors)
4098 last_sector = mddev->dev_sectors - 1;
4099 while (first_sector <= last_sector) {
4100 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4101 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4102 set_bit(STRIPE_HANDLE, &sh->state);
4104 first_sector += STRIPE_SECTORS;
4106 /* Now that the sources are clearly marked, we can release
4107 * the destination stripes
4109 while (!list_empty(&stripes)) {
4110 sh = list_entry(stripes.next, struct stripe_head, lru);
4111 list_del_init(&sh->lru);
4114 /* If this takes us to the resync_max point where we have to pause,
4115 * then we need to write out the superblock.
4117 sector_nr += reshape_sectors;
4118 if ((sector_nr - mddev->curr_resync_completed) * 2
4119 >= mddev->resync_max - mddev->curr_resync_completed) {
4120 /* Cannot proceed until we've updated the superblock... */
4121 wait_event(conf->wait_for_overlap,
4122 atomic_read(&conf->reshape_stripes) == 0);
4123 mddev->reshape_position = conf->reshape_progress;
4124 mddev->curr_resync_completed = sector_nr;
4125 conf->reshape_checkpoint = jiffies;
4126 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4127 md_wakeup_thread(mddev->thread);
4128 wait_event(mddev->sb_wait,
4129 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4130 || kthread_should_stop());
4131 spin_lock_irq(&conf->device_lock);
4132 conf->reshape_safe = mddev->reshape_position;
4133 spin_unlock_irq(&conf->device_lock);
4134 wake_up(&conf->wait_for_overlap);
4135 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4137 return reshape_sectors;
4140 /* FIXME go_faster isn't used */
4141 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4143 raid5_conf_t *conf = mddev->private;
4144 struct stripe_head *sh;
4145 sector_t max_sector = mddev->dev_sectors;
4146 sector_t sync_blocks;
4147 int still_degraded = 0;
4150 if (sector_nr >= max_sector) {
4151 /* just being told to finish up .. nothing much to do */
4153 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4158 if (mddev->curr_resync < max_sector) /* aborted */
4159 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4161 else /* completed sync */
4163 bitmap_close_sync(mddev->bitmap);
4168 /* Allow raid5_quiesce to complete */
4169 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4171 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4172 return reshape_request(mddev, sector_nr, skipped);
4174 /* No need to check resync_max as we never do more than one
4175 * stripe, and as resync_max will always be on a chunk boundary,
4176 * if the check in md_do_sync didn't fire, there is no chance
4177 * of overstepping resync_max here
4180 /* if there is too many failed drives and we are trying
4181 * to resync, then assert that we are finished, because there is
4182 * nothing we can do.
4184 if (mddev->degraded >= conf->max_degraded &&
4185 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4186 sector_t rv = mddev->dev_sectors - sector_nr;
4190 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4191 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4192 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4193 /* we can skip this block, and probably more */
4194 sync_blocks /= STRIPE_SECTORS;
4196 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4200 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4202 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4204 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4205 /* make sure we don't swamp the stripe cache if someone else
4206 * is trying to get access
4208 schedule_timeout_uninterruptible(1);
4210 /* Need to check if array will still be degraded after recovery/resync
4211 * We don't need to check the 'failed' flag as when that gets set,
4214 for (i = 0; i < conf->raid_disks; i++)
4215 if (conf->disks[i].rdev == NULL)
4218 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4220 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4225 return STRIPE_SECTORS;
4228 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4230 /* We may not be able to submit a whole bio at once as there
4231 * may not be enough stripe_heads available.
4232 * We cannot pre-allocate enough stripe_heads as we may need
4233 * more than exist in the cache (if we allow ever large chunks).
4234 * So we do one stripe head at a time and record in
4235 * ->bi_hw_segments how many have been done.
4237 * We *know* that this entire raid_bio is in one chunk, so
4238 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4240 struct stripe_head *sh;
4242 sector_t sector, logical_sector, last_sector;
4247 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4248 sector = raid5_compute_sector(conf, logical_sector,
4250 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4252 for (; logical_sector < last_sector;
4253 logical_sector += STRIPE_SECTORS,
4254 sector += STRIPE_SECTORS,
4257 if (scnt < raid5_bi_hw_segments(raid_bio))
4258 /* already done this stripe */
4261 sh = get_active_stripe(conf, sector, 0, 1, 0);
4264 /* failed to get a stripe - must wait */
4265 raid5_set_bi_hw_segments(raid_bio, scnt);
4266 conf->retry_read_aligned = raid_bio;
4270 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4271 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4273 raid5_set_bi_hw_segments(raid_bio, scnt);
4274 conf->retry_read_aligned = raid_bio;
4282 spin_lock_irq(&conf->device_lock);
4283 remaining = raid5_dec_bi_phys_segments(raid_bio);
4284 spin_unlock_irq(&conf->device_lock);
4286 bio_endio(raid_bio, 0);
4287 if (atomic_dec_and_test(&conf->active_aligned_reads))
4288 wake_up(&conf->wait_for_stripe);
4294 * This is our raid5 kernel thread.
4296 * We scan the hash table for stripes which can be handled now.
4297 * During the scan, completed stripes are saved for us by the interrupt
4298 * handler, so that they will not have to wait for our next wakeup.
4300 static void raid5d(mddev_t *mddev)
4302 struct stripe_head *sh;
4303 raid5_conf_t *conf = mddev->private;
4305 struct blk_plug plug;
4307 pr_debug("+++ raid5d active\n");
4309 md_check_recovery(mddev);
4311 blk_start_plug(&plug);
4313 spin_lock_irq(&conf->device_lock);
4317 if (atomic_read(&mddev->plug_cnt) == 0 &&
4318 !list_empty(&conf->bitmap_list)) {
4319 /* Now is a good time to flush some bitmap updates */
4321 spin_unlock_irq(&conf->device_lock);
4322 bitmap_unplug(mddev->bitmap);
4323 spin_lock_irq(&conf->device_lock);
4324 conf->seq_write = conf->seq_flush;
4325 activate_bit_delay(conf);
4327 if (atomic_read(&mddev->plug_cnt) == 0)
4328 raid5_activate_delayed(conf);
4330 while ((bio = remove_bio_from_retry(conf))) {
4332 spin_unlock_irq(&conf->device_lock);
4333 ok = retry_aligned_read(conf, bio);
4334 spin_lock_irq(&conf->device_lock);
4340 sh = __get_priority_stripe(conf);
4344 spin_unlock_irq(&conf->device_lock);
4351 spin_lock_irq(&conf->device_lock);
4353 pr_debug("%d stripes handled\n", handled);
4355 spin_unlock_irq(&conf->device_lock);
4357 async_tx_issue_pending_all();
4358 blk_finish_plug(&plug);
4360 pr_debug("--- raid5d inactive\n");
4364 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4366 raid5_conf_t *conf = mddev->private;
4368 return sprintf(page, "%d\n", conf->max_nr_stripes);
4374 raid5_set_cache_size(mddev_t *mddev, int size)
4376 raid5_conf_t *conf = mddev->private;
4379 if (size <= 16 || size > 32768)
4381 while (size < conf->max_nr_stripes) {
4382 if (drop_one_stripe(conf))
4383 conf->max_nr_stripes--;
4387 err = md_allow_write(mddev);
4390 while (size > conf->max_nr_stripes) {
4391 if (grow_one_stripe(conf))
4392 conf->max_nr_stripes++;
4397 EXPORT_SYMBOL(raid5_set_cache_size);
4400 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4402 raid5_conf_t *conf = mddev->private;
4406 if (len >= PAGE_SIZE)
4411 if (strict_strtoul(page, 10, &new))
4413 err = raid5_set_cache_size(mddev, new);
4419 static struct md_sysfs_entry
4420 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4421 raid5_show_stripe_cache_size,
4422 raid5_store_stripe_cache_size);
4425 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4427 raid5_conf_t *conf = mddev->private;
4429 return sprintf(page, "%d\n", conf->bypass_threshold);
4435 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4437 raid5_conf_t *conf = mddev->private;
4439 if (len >= PAGE_SIZE)
4444 if (strict_strtoul(page, 10, &new))
4446 if (new > conf->max_nr_stripes)
4448 conf->bypass_threshold = new;
4452 static struct md_sysfs_entry
4453 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4455 raid5_show_preread_threshold,
4456 raid5_store_preread_threshold);
4459 stripe_cache_active_show(mddev_t *mddev, char *page)
4461 raid5_conf_t *conf = mddev->private;
4463 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4468 static struct md_sysfs_entry
4469 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4471 static struct attribute *raid5_attrs[] = {
4472 &raid5_stripecache_size.attr,
4473 &raid5_stripecache_active.attr,
4474 &raid5_preread_bypass_threshold.attr,
4477 static struct attribute_group raid5_attrs_group = {
4479 .attrs = raid5_attrs,
4483 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4485 raid5_conf_t *conf = mddev->private;
4488 sectors = mddev->dev_sectors;
4490 /* size is defined by the smallest of previous and new size */
4491 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4493 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4494 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4495 return sectors * (raid_disks - conf->max_degraded);
4498 static void raid5_free_percpu(raid5_conf_t *conf)
4500 struct raid5_percpu *percpu;
4507 for_each_possible_cpu(cpu) {
4508 percpu = per_cpu_ptr(conf->percpu, cpu);
4509 safe_put_page(percpu->spare_page);
4510 kfree(percpu->scribble);
4512 #ifdef CONFIG_HOTPLUG_CPU
4513 unregister_cpu_notifier(&conf->cpu_notify);
4517 free_percpu(conf->percpu);
4520 static void free_conf(raid5_conf_t *conf)
4522 shrink_stripes(conf);
4523 raid5_free_percpu(conf);
4525 kfree(conf->stripe_hashtbl);
4529 #ifdef CONFIG_HOTPLUG_CPU
4530 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4533 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4534 long cpu = (long)hcpu;
4535 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4538 case CPU_UP_PREPARE:
4539 case CPU_UP_PREPARE_FROZEN:
4540 if (conf->level == 6 && !percpu->spare_page)
4541 percpu->spare_page = alloc_page(GFP_KERNEL);
4542 if (!percpu->scribble)
4543 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4545 if (!percpu->scribble ||
4546 (conf->level == 6 && !percpu->spare_page)) {
4547 safe_put_page(percpu->spare_page);
4548 kfree(percpu->scribble);
4549 pr_err("%s: failed memory allocation for cpu%ld\n",
4551 return notifier_from_errno(-ENOMEM);
4555 case CPU_DEAD_FROZEN:
4556 safe_put_page(percpu->spare_page);
4557 kfree(percpu->scribble);
4558 percpu->spare_page = NULL;
4559 percpu->scribble = NULL;
4568 static int raid5_alloc_percpu(raid5_conf_t *conf)
4571 struct page *spare_page;
4572 struct raid5_percpu __percpu *allcpus;
4576 allcpus = alloc_percpu(struct raid5_percpu);
4579 conf->percpu = allcpus;
4583 for_each_present_cpu(cpu) {
4584 if (conf->level == 6) {
4585 spare_page = alloc_page(GFP_KERNEL);
4590 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4592 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4597 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4599 #ifdef CONFIG_HOTPLUG_CPU
4600 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4601 conf->cpu_notify.priority = 0;
4603 err = register_cpu_notifier(&conf->cpu_notify);
4610 static raid5_conf_t *setup_conf(mddev_t *mddev)
4613 int raid_disk, memory, max_disks;
4615 struct disk_info *disk;
4617 if (mddev->new_level != 5
4618 && mddev->new_level != 4
4619 && mddev->new_level != 6) {
4620 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4621 mdname(mddev), mddev->new_level);
4622 return ERR_PTR(-EIO);
4624 if ((mddev->new_level == 5
4625 && !algorithm_valid_raid5(mddev->new_layout)) ||
4626 (mddev->new_level == 6
4627 && !algorithm_valid_raid6(mddev->new_layout))) {
4628 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4629 mdname(mddev), mddev->new_layout);
4630 return ERR_PTR(-EIO);
4632 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4633 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4634 mdname(mddev), mddev->raid_disks);
4635 return ERR_PTR(-EINVAL);
4638 if (!mddev->new_chunk_sectors ||
4639 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4640 !is_power_of_2(mddev->new_chunk_sectors)) {
4641 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4642 mdname(mddev), mddev->new_chunk_sectors << 9);
4643 return ERR_PTR(-EINVAL);
4646 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4649 spin_lock_init(&conf->device_lock);
4650 init_waitqueue_head(&conf->wait_for_stripe);
4651 init_waitqueue_head(&conf->wait_for_overlap);
4652 INIT_LIST_HEAD(&conf->handle_list);
4653 INIT_LIST_HEAD(&conf->hold_list);
4654 INIT_LIST_HEAD(&conf->delayed_list);
4655 INIT_LIST_HEAD(&conf->bitmap_list);
4656 INIT_LIST_HEAD(&conf->inactive_list);
4657 atomic_set(&conf->active_stripes, 0);
4658 atomic_set(&conf->preread_active_stripes, 0);
4659 atomic_set(&conf->active_aligned_reads, 0);
4660 conf->bypass_threshold = BYPASS_THRESHOLD;
4662 conf->raid_disks = mddev->raid_disks;
4663 if (mddev->reshape_position == MaxSector)
4664 conf->previous_raid_disks = mddev->raid_disks;
4666 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4667 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4668 conf->scribble_len = scribble_len(max_disks);
4670 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4675 conf->mddev = mddev;
4677 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4680 conf->level = mddev->new_level;
4681 if (raid5_alloc_percpu(conf) != 0)
4684 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4686 list_for_each_entry(rdev, &mddev->disks, same_set) {
4687 raid_disk = rdev->raid_disk;
4688 if (raid_disk >= max_disks
4691 disk = conf->disks + raid_disk;
4695 if (test_bit(In_sync, &rdev->flags)) {
4696 char b[BDEVNAME_SIZE];
4697 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4699 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4700 } else if (rdev->saved_raid_disk != raid_disk)
4701 /* Cannot rely on bitmap to complete recovery */
4705 conf->chunk_sectors = mddev->new_chunk_sectors;
4706 conf->level = mddev->new_level;
4707 if (conf->level == 6)
4708 conf->max_degraded = 2;
4710 conf->max_degraded = 1;
4711 conf->algorithm = mddev->new_layout;
4712 conf->max_nr_stripes = NR_STRIPES;
4713 conf->reshape_progress = mddev->reshape_position;
4714 if (conf->reshape_progress != MaxSector) {
4715 conf->prev_chunk_sectors = mddev->chunk_sectors;
4716 conf->prev_algo = mddev->layout;
4719 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4720 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4721 if (grow_stripes(conf, conf->max_nr_stripes)) {
4723 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4724 mdname(mddev), memory);
4727 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4728 mdname(mddev), memory);
4730 conf->thread = md_register_thread(raid5d, mddev, NULL);
4731 if (!conf->thread) {
4733 "md/raid:%s: couldn't allocate thread.\n",
4743 return ERR_PTR(-EIO);
4745 return ERR_PTR(-ENOMEM);
4749 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4752 case ALGORITHM_PARITY_0:
4753 if (raid_disk < max_degraded)
4756 case ALGORITHM_PARITY_N:
4757 if (raid_disk >= raid_disks - max_degraded)
4760 case ALGORITHM_PARITY_0_6:
4761 if (raid_disk == 0 ||
4762 raid_disk == raid_disks - 1)
4765 case ALGORITHM_LEFT_ASYMMETRIC_6:
4766 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4767 case ALGORITHM_LEFT_SYMMETRIC_6:
4768 case ALGORITHM_RIGHT_SYMMETRIC_6:
4769 if (raid_disk == raid_disks - 1)
4775 static int run(mddev_t *mddev)
4778 int working_disks = 0;
4779 int dirty_parity_disks = 0;
4781 sector_t reshape_offset = 0;
4783 if (mddev->recovery_cp != MaxSector)
4784 printk(KERN_NOTICE "md/raid:%s: not clean"
4785 " -- starting background reconstruction\n",
4787 if (mddev->reshape_position != MaxSector) {
4788 /* Check that we can continue the reshape.
4789 * Currently only disks can change, it must
4790 * increase, and we must be past the point where
4791 * a stripe over-writes itself
4793 sector_t here_new, here_old;
4795 int max_degraded = (mddev->level == 6 ? 2 : 1);
4797 if (mddev->new_level != mddev->level) {
4798 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4799 "required - aborting.\n",
4803 old_disks = mddev->raid_disks - mddev->delta_disks;
4804 /* reshape_position must be on a new-stripe boundary, and one
4805 * further up in new geometry must map after here in old
4808 here_new = mddev->reshape_position;
4809 if (sector_div(here_new, mddev->new_chunk_sectors *
4810 (mddev->raid_disks - max_degraded))) {
4811 printk(KERN_ERR "md/raid:%s: reshape_position not "
4812 "on a stripe boundary\n", mdname(mddev));
4815 reshape_offset = here_new * mddev->new_chunk_sectors;
4816 /* here_new is the stripe we will write to */
4817 here_old = mddev->reshape_position;
4818 sector_div(here_old, mddev->chunk_sectors *
4819 (old_disks-max_degraded));
4820 /* here_old is the first stripe that we might need to read
4822 if (mddev->delta_disks == 0) {
4823 /* We cannot be sure it is safe to start an in-place
4824 * reshape. It is only safe if user-space if monitoring
4825 * and taking constant backups.
4826 * mdadm always starts a situation like this in
4827 * readonly mode so it can take control before
4828 * allowing any writes. So just check for that.
4830 if ((here_new * mddev->new_chunk_sectors !=
4831 here_old * mddev->chunk_sectors) ||
4833 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4834 " in read-only mode - aborting\n",
4838 } else if (mddev->delta_disks < 0
4839 ? (here_new * mddev->new_chunk_sectors <=
4840 here_old * mddev->chunk_sectors)
4841 : (here_new * mddev->new_chunk_sectors >=
4842 here_old * mddev->chunk_sectors)) {
4843 /* Reading from the same stripe as writing to - bad */
4844 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4845 "auto-recovery - aborting.\n",
4849 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4851 /* OK, we should be able to continue; */
4853 BUG_ON(mddev->level != mddev->new_level);
4854 BUG_ON(mddev->layout != mddev->new_layout);
4855 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4856 BUG_ON(mddev->delta_disks != 0);
4859 if (mddev->private == NULL)
4860 conf = setup_conf(mddev);
4862 conf = mddev->private;
4865 return PTR_ERR(conf);
4867 mddev->thread = conf->thread;
4868 conf->thread = NULL;
4869 mddev->private = conf;
4872 * 0 for a fully functional array, 1 or 2 for a degraded array.
4874 list_for_each_entry(rdev, &mddev->disks, same_set) {
4875 if (rdev->raid_disk < 0)
4877 if (test_bit(In_sync, &rdev->flags)) {
4881 /* This disc is not fully in-sync. However if it
4882 * just stored parity (beyond the recovery_offset),
4883 * when we don't need to be concerned about the
4884 * array being dirty.
4885 * When reshape goes 'backwards', we never have
4886 * partially completed devices, so we only need
4887 * to worry about reshape going forwards.
4889 /* Hack because v0.91 doesn't store recovery_offset properly. */
4890 if (mddev->major_version == 0 &&
4891 mddev->minor_version > 90)
4892 rdev->recovery_offset = reshape_offset;
4894 if (rdev->recovery_offset < reshape_offset) {
4895 /* We need to check old and new layout */
4896 if (!only_parity(rdev->raid_disk,
4899 conf->max_degraded))
4902 if (!only_parity(rdev->raid_disk,
4904 conf->previous_raid_disks,
4905 conf->max_degraded))
4907 dirty_parity_disks++;
4910 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4913 if (has_failed(conf)) {
4914 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4915 " (%d/%d failed)\n",
4916 mdname(mddev), mddev->degraded, conf->raid_disks);
4920 /* device size must be a multiple of chunk size */
4921 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4922 mddev->resync_max_sectors = mddev->dev_sectors;
4924 if (mddev->degraded > dirty_parity_disks &&
4925 mddev->recovery_cp != MaxSector) {
4926 if (mddev->ok_start_degraded)
4928 "md/raid:%s: starting dirty degraded array"
4929 " - data corruption possible.\n",
4933 "md/raid:%s: cannot start dirty degraded array.\n",
4939 if (mddev->degraded == 0)
4940 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4941 " devices, algorithm %d\n", mdname(mddev), conf->level,
4942 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4945 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4946 " out of %d devices, algorithm %d\n",
4947 mdname(mddev), conf->level,
4948 mddev->raid_disks - mddev->degraded,
4949 mddev->raid_disks, mddev->new_layout);
4951 print_raid5_conf(conf);
4953 if (conf->reshape_progress != MaxSector) {
4954 conf->reshape_safe = conf->reshape_progress;
4955 atomic_set(&conf->reshape_stripes, 0);
4956 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4957 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4958 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4959 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4960 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4965 /* Ok, everything is just fine now */
4966 if (mddev->to_remove == &raid5_attrs_group)
4967 mddev->to_remove = NULL;
4968 else if (mddev->kobj.sd &&
4969 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4971 "raid5: failed to create sysfs attributes for %s\n",
4973 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4977 /* read-ahead size must cover two whole stripes, which
4978 * is 2 * (datadisks) * chunksize where 'n' is the
4979 * number of raid devices
4981 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4982 int stripe = data_disks *
4983 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4984 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4985 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4987 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4989 mddev->queue->backing_dev_info.congested_data = mddev;
4990 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4992 chunk_size = mddev->chunk_sectors << 9;
4993 blk_queue_io_min(mddev->queue, chunk_size);
4994 blk_queue_io_opt(mddev->queue, chunk_size *
4995 (conf->raid_disks - conf->max_degraded));
4997 list_for_each_entry(rdev, &mddev->disks, same_set)
4998 disk_stack_limits(mddev->gendisk, rdev->bdev,
4999 rdev->data_offset << 9);
5004 md_unregister_thread(mddev->thread);
5005 mddev->thread = NULL;
5007 print_raid5_conf(conf);
5010 mddev->private = NULL;
5011 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5015 static int stop(mddev_t *mddev)
5017 raid5_conf_t *conf = mddev->private;
5019 md_unregister_thread(mddev->thread);
5020 mddev->thread = NULL;
5022 mddev->queue->backing_dev_info.congested_fn = NULL;
5024 mddev->private = NULL;
5025 mddev->to_remove = &raid5_attrs_group;
5030 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5034 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5035 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5036 seq_printf(seq, "sh %llu, count %d.\n",
5037 (unsigned long long)sh->sector, atomic_read(&sh->count));
5038 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5039 for (i = 0; i < sh->disks; i++) {
5040 seq_printf(seq, "(cache%d: %p %ld) ",
5041 i, sh->dev[i].page, sh->dev[i].flags);
5043 seq_printf(seq, "\n");
5046 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5048 struct stripe_head *sh;
5049 struct hlist_node *hn;
5052 spin_lock_irq(&conf->device_lock);
5053 for (i = 0; i < NR_HASH; i++) {
5054 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5055 if (sh->raid_conf != conf)
5060 spin_unlock_irq(&conf->device_lock);
5064 static void status(struct seq_file *seq, mddev_t *mddev)
5066 raid5_conf_t *conf = mddev->private;
5069 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5070 mddev->chunk_sectors / 2, mddev->layout);
5071 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5072 for (i = 0; i < conf->raid_disks; i++)
5073 seq_printf (seq, "%s",
5074 conf->disks[i].rdev &&
5075 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5076 seq_printf (seq, "]");
5078 seq_printf (seq, "\n");
5079 printall(seq, conf);
5083 static void print_raid5_conf (raid5_conf_t *conf)
5086 struct disk_info *tmp;
5088 printk(KERN_DEBUG "RAID conf printout:\n");
5090 printk("(conf==NULL)\n");
5093 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5095 conf->raid_disks - conf->mddev->degraded);
5097 for (i = 0; i < conf->raid_disks; i++) {
5098 char b[BDEVNAME_SIZE];
5099 tmp = conf->disks + i;
5101 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5102 i, !test_bit(Faulty, &tmp->rdev->flags),
5103 bdevname(tmp->rdev->bdev, b));
5107 static int raid5_spare_active(mddev_t *mddev)
5110 raid5_conf_t *conf = mddev->private;
5111 struct disk_info *tmp;
5113 unsigned long flags;
5115 for (i = 0; i < conf->raid_disks; i++) {
5116 tmp = conf->disks + i;
5118 && tmp->rdev->recovery_offset == MaxSector
5119 && !test_bit(Faulty, &tmp->rdev->flags)
5120 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5122 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5125 spin_lock_irqsave(&conf->device_lock, flags);
5126 mddev->degraded -= count;
5127 spin_unlock_irqrestore(&conf->device_lock, flags);
5128 print_raid5_conf(conf);
5132 static int raid5_remove_disk(mddev_t *mddev, int number)
5134 raid5_conf_t *conf = mddev->private;
5137 struct disk_info *p = conf->disks + number;
5139 print_raid5_conf(conf);
5142 if (number >= conf->raid_disks &&
5143 conf->reshape_progress == MaxSector)
5144 clear_bit(In_sync, &rdev->flags);
5146 if (test_bit(In_sync, &rdev->flags) ||
5147 atomic_read(&rdev->nr_pending)) {
5151 /* Only remove non-faulty devices if recovery
5154 if (!test_bit(Faulty, &rdev->flags) &&
5155 !has_failed(conf) &&
5156 number < conf->raid_disks) {
5162 if (atomic_read(&rdev->nr_pending)) {
5163 /* lost the race, try later */
5170 print_raid5_conf(conf);
5174 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5176 raid5_conf_t *conf = mddev->private;
5179 struct disk_info *p;
5181 int last = conf->raid_disks - 1;
5183 if (has_failed(conf))
5184 /* no point adding a device */
5187 if (rdev->raid_disk >= 0)
5188 first = last = rdev->raid_disk;
5191 * find the disk ... but prefer rdev->saved_raid_disk
5194 if (rdev->saved_raid_disk >= 0 &&
5195 rdev->saved_raid_disk >= first &&
5196 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5197 disk = rdev->saved_raid_disk;
5200 for ( ; disk <= last ; disk++)
5201 if ((p=conf->disks + disk)->rdev == NULL) {
5202 clear_bit(In_sync, &rdev->flags);
5203 rdev->raid_disk = disk;
5205 if (rdev->saved_raid_disk != disk)
5207 rcu_assign_pointer(p->rdev, rdev);
5210 print_raid5_conf(conf);
5214 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5216 /* no resync is happening, and there is enough space
5217 * on all devices, so we can resize.
5218 * We need to make sure resync covers any new space.
5219 * If the array is shrinking we should possibly wait until
5220 * any io in the removed space completes, but it hardly seems
5223 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5224 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5225 mddev->raid_disks));
5226 if (mddev->array_sectors >
5227 raid5_size(mddev, sectors, mddev->raid_disks))
5229 set_capacity(mddev->gendisk, mddev->array_sectors);
5230 revalidate_disk(mddev->gendisk);
5231 if (sectors > mddev->dev_sectors &&
5232 mddev->recovery_cp > mddev->dev_sectors) {
5233 mddev->recovery_cp = mddev->dev_sectors;
5234 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5236 mddev->dev_sectors = sectors;
5237 mddev->resync_max_sectors = sectors;
5241 static int check_stripe_cache(mddev_t *mddev)
5243 /* Can only proceed if there are plenty of stripe_heads.
5244 * We need a minimum of one full stripe,, and for sensible progress
5245 * it is best to have about 4 times that.
5246 * If we require 4 times, then the default 256 4K stripe_heads will
5247 * allow for chunk sizes up to 256K, which is probably OK.
5248 * If the chunk size is greater, user-space should request more
5249 * stripe_heads first.
5251 raid5_conf_t *conf = mddev->private;
5252 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5253 > conf->max_nr_stripes ||
5254 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5255 > conf->max_nr_stripes) {
5256 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5258 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5265 static int check_reshape(mddev_t *mddev)
5267 raid5_conf_t *conf = mddev->private;
5269 if (mddev->delta_disks == 0 &&
5270 mddev->new_layout == mddev->layout &&
5271 mddev->new_chunk_sectors == mddev->chunk_sectors)
5272 return 0; /* nothing to do */
5274 /* Cannot grow a bitmap yet */
5276 if (has_failed(conf))
5278 if (mddev->delta_disks < 0) {
5279 /* We might be able to shrink, but the devices must
5280 * be made bigger first.
5281 * For raid6, 4 is the minimum size.
5282 * Otherwise 2 is the minimum
5285 if (mddev->level == 6)
5287 if (mddev->raid_disks + mddev->delta_disks < min)
5291 if (!check_stripe_cache(mddev))
5294 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5297 static int raid5_start_reshape(mddev_t *mddev)
5299 raid5_conf_t *conf = mddev->private;
5302 unsigned long flags;
5304 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5307 if (!check_stripe_cache(mddev))
5310 list_for_each_entry(rdev, &mddev->disks, same_set)
5311 if (!test_bit(In_sync, &rdev->flags)
5312 && !test_bit(Faulty, &rdev->flags))
5315 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5316 /* Not enough devices even to make a degraded array
5321 /* Refuse to reduce size of the array. Any reductions in
5322 * array size must be through explicit setting of array_size
5325 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5326 < mddev->array_sectors) {
5327 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5328 "before number of disks\n", mdname(mddev));
5332 atomic_set(&conf->reshape_stripes, 0);
5333 spin_lock_irq(&conf->device_lock);
5334 conf->previous_raid_disks = conf->raid_disks;
5335 conf->raid_disks += mddev->delta_disks;
5336 conf->prev_chunk_sectors = conf->chunk_sectors;
5337 conf->chunk_sectors = mddev->new_chunk_sectors;
5338 conf->prev_algo = conf->algorithm;
5339 conf->algorithm = mddev->new_layout;
5340 if (mddev->delta_disks < 0)
5341 conf->reshape_progress = raid5_size(mddev, 0, 0);
5343 conf->reshape_progress = 0;
5344 conf->reshape_safe = conf->reshape_progress;
5346 spin_unlock_irq(&conf->device_lock);
5348 /* Add some new drives, as many as will fit.
5349 * We know there are enough to make the newly sized array work.
5350 * Don't add devices if we are reducing the number of
5351 * devices in the array. This is because it is not possible
5352 * to correctly record the "partially reconstructed" state of
5353 * such devices during the reshape and confusion could result.
5355 if (mddev->delta_disks >= 0) {
5356 int added_devices = 0;
5357 list_for_each_entry(rdev, &mddev->disks, same_set)
5358 if (rdev->raid_disk < 0 &&
5359 !test_bit(Faulty, &rdev->flags)) {
5360 if (raid5_add_disk(mddev, rdev) == 0) {
5363 >= conf->previous_raid_disks) {
5364 set_bit(In_sync, &rdev->flags);
5367 rdev->recovery_offset = 0;
5368 sprintf(nm, "rd%d", rdev->raid_disk);
5369 if (sysfs_create_link(&mddev->kobj,
5371 /* Failure here is OK */;
5373 } else if (rdev->raid_disk >= conf->previous_raid_disks
5374 && !test_bit(Faulty, &rdev->flags)) {
5375 /* This is a spare that was manually added */
5376 set_bit(In_sync, &rdev->flags);
5380 /* When a reshape changes the number of devices,
5381 * ->degraded is measured against the larger of the
5382 * pre and post number of devices.
5384 spin_lock_irqsave(&conf->device_lock, flags);
5385 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5387 spin_unlock_irqrestore(&conf->device_lock, flags);
5389 mddev->raid_disks = conf->raid_disks;
5390 mddev->reshape_position = conf->reshape_progress;
5391 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5393 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5394 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5395 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5396 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5397 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5399 if (!mddev->sync_thread) {
5400 mddev->recovery = 0;
5401 spin_lock_irq(&conf->device_lock);
5402 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5403 conf->reshape_progress = MaxSector;
5404 spin_unlock_irq(&conf->device_lock);
5407 conf->reshape_checkpoint = jiffies;
5408 md_wakeup_thread(mddev->sync_thread);
5409 md_new_event(mddev);
5413 /* This is called from the reshape thread and should make any
5414 * changes needed in 'conf'
5416 static void end_reshape(raid5_conf_t *conf)
5419 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5421 spin_lock_irq(&conf->device_lock);
5422 conf->previous_raid_disks = conf->raid_disks;
5423 conf->reshape_progress = MaxSector;
5424 spin_unlock_irq(&conf->device_lock);
5425 wake_up(&conf->wait_for_overlap);
5427 /* read-ahead size must cover two whole stripes, which is
5428 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5430 if (conf->mddev->queue) {
5431 int data_disks = conf->raid_disks - conf->max_degraded;
5432 int stripe = data_disks * ((conf->chunk_sectors << 9)
5434 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5435 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5440 /* This is called from the raid5d thread with mddev_lock held.
5441 * It makes config changes to the device.
5443 static void raid5_finish_reshape(mddev_t *mddev)
5445 raid5_conf_t *conf = mddev->private;
5447 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5449 if (mddev->delta_disks > 0) {
5450 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5451 set_capacity(mddev->gendisk, mddev->array_sectors);
5452 revalidate_disk(mddev->gendisk);
5455 mddev->degraded = conf->raid_disks;
5456 for (d = 0; d < conf->raid_disks ; d++)
5457 if (conf->disks[d].rdev &&
5459 &conf->disks[d].rdev->flags))
5461 for (d = conf->raid_disks ;
5462 d < conf->raid_disks - mddev->delta_disks;
5464 mdk_rdev_t *rdev = conf->disks[d].rdev;
5465 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5467 sprintf(nm, "rd%d", rdev->raid_disk);
5468 sysfs_remove_link(&mddev->kobj, nm);
5469 rdev->raid_disk = -1;
5473 mddev->layout = conf->algorithm;
5474 mddev->chunk_sectors = conf->chunk_sectors;
5475 mddev->reshape_position = MaxSector;
5476 mddev->delta_disks = 0;
5480 static void raid5_quiesce(mddev_t *mddev, int state)
5482 raid5_conf_t *conf = mddev->private;
5485 case 2: /* resume for a suspend */
5486 wake_up(&conf->wait_for_overlap);
5489 case 1: /* stop all writes */
5490 spin_lock_irq(&conf->device_lock);
5491 /* '2' tells resync/reshape to pause so that all
5492 * active stripes can drain
5495 wait_event_lock_irq(conf->wait_for_stripe,
5496 atomic_read(&conf->active_stripes) == 0 &&
5497 atomic_read(&conf->active_aligned_reads) == 0,
5498 conf->device_lock, /* nothing */);
5500 spin_unlock_irq(&conf->device_lock);
5501 /* allow reshape to continue */
5502 wake_up(&conf->wait_for_overlap);
5505 case 0: /* re-enable writes */
5506 spin_lock_irq(&conf->device_lock);
5508 wake_up(&conf->wait_for_stripe);
5509 wake_up(&conf->wait_for_overlap);
5510 spin_unlock_irq(&conf->device_lock);
5516 static void *raid45_takeover_raid0(mddev_t *mddev, int level)
5518 struct raid0_private_data *raid0_priv = mddev->private;
5521 /* for raid0 takeover only one zone is supported */
5522 if (raid0_priv->nr_strip_zones > 1) {
5523 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5525 return ERR_PTR(-EINVAL);
5528 sectors = raid0_priv->strip_zone[0].zone_end;
5529 sector_div(sectors, raid0_priv->strip_zone[0].nb_dev);
5530 mddev->dev_sectors = sectors;
5531 mddev->new_level = level;
5532 mddev->new_layout = ALGORITHM_PARITY_N;
5533 mddev->new_chunk_sectors = mddev->chunk_sectors;
5534 mddev->raid_disks += 1;
5535 mddev->delta_disks = 1;
5536 /* make sure it will be not marked as dirty */
5537 mddev->recovery_cp = MaxSector;
5539 return setup_conf(mddev);
5543 static void *raid5_takeover_raid1(mddev_t *mddev)
5547 if (mddev->raid_disks != 2 ||
5548 mddev->degraded > 1)
5549 return ERR_PTR(-EINVAL);
5551 /* Should check if there are write-behind devices? */
5553 chunksect = 64*2; /* 64K by default */
5555 /* The array must be an exact multiple of chunksize */
5556 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5559 if ((chunksect<<9) < STRIPE_SIZE)
5560 /* array size does not allow a suitable chunk size */
5561 return ERR_PTR(-EINVAL);
5563 mddev->new_level = 5;
5564 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5565 mddev->new_chunk_sectors = chunksect;
5567 return setup_conf(mddev);
5570 static void *raid5_takeover_raid6(mddev_t *mddev)
5574 switch (mddev->layout) {
5575 case ALGORITHM_LEFT_ASYMMETRIC_6:
5576 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5578 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5579 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5581 case ALGORITHM_LEFT_SYMMETRIC_6:
5582 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5584 case ALGORITHM_RIGHT_SYMMETRIC_6:
5585 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5587 case ALGORITHM_PARITY_0_6:
5588 new_layout = ALGORITHM_PARITY_0;
5590 case ALGORITHM_PARITY_N:
5591 new_layout = ALGORITHM_PARITY_N;
5594 return ERR_PTR(-EINVAL);
5596 mddev->new_level = 5;
5597 mddev->new_layout = new_layout;
5598 mddev->delta_disks = -1;
5599 mddev->raid_disks -= 1;
5600 return setup_conf(mddev);
5604 static int raid5_check_reshape(mddev_t *mddev)
5606 /* For a 2-drive array, the layout and chunk size can be changed
5607 * immediately as not restriping is needed.
5608 * For larger arrays we record the new value - after validation
5609 * to be used by a reshape pass.
5611 raid5_conf_t *conf = mddev->private;
5612 int new_chunk = mddev->new_chunk_sectors;
5614 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5616 if (new_chunk > 0) {
5617 if (!is_power_of_2(new_chunk))
5619 if (new_chunk < (PAGE_SIZE>>9))
5621 if (mddev->array_sectors & (new_chunk-1))
5622 /* not factor of array size */
5626 /* They look valid */
5628 if (mddev->raid_disks == 2) {
5629 /* can make the change immediately */
5630 if (mddev->new_layout >= 0) {
5631 conf->algorithm = mddev->new_layout;
5632 mddev->layout = mddev->new_layout;
5634 if (new_chunk > 0) {
5635 conf->chunk_sectors = new_chunk ;
5636 mddev->chunk_sectors = new_chunk;
5638 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5639 md_wakeup_thread(mddev->thread);
5641 return check_reshape(mddev);
5644 static int raid6_check_reshape(mddev_t *mddev)
5646 int new_chunk = mddev->new_chunk_sectors;
5648 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5650 if (new_chunk > 0) {
5651 if (!is_power_of_2(new_chunk))
5653 if (new_chunk < (PAGE_SIZE >> 9))
5655 if (mddev->array_sectors & (new_chunk-1))
5656 /* not factor of array size */
5660 /* They look valid */
5661 return check_reshape(mddev);
5664 static void *raid5_takeover(mddev_t *mddev)
5666 /* raid5 can take over:
5667 * raid0 - if there is only one strip zone - make it a raid4 layout
5668 * raid1 - if there are two drives. We need to know the chunk size
5669 * raid4 - trivial - just use a raid4 layout.
5670 * raid6 - Providing it is a *_6 layout
5672 if (mddev->level == 0)
5673 return raid45_takeover_raid0(mddev, 5);
5674 if (mddev->level == 1)
5675 return raid5_takeover_raid1(mddev);
5676 if (mddev->level == 4) {
5677 mddev->new_layout = ALGORITHM_PARITY_N;
5678 mddev->new_level = 5;
5679 return setup_conf(mddev);
5681 if (mddev->level == 6)
5682 return raid5_takeover_raid6(mddev);
5684 return ERR_PTR(-EINVAL);
5687 static void *raid4_takeover(mddev_t *mddev)
5689 /* raid4 can take over:
5690 * raid0 - if there is only one strip zone
5691 * raid5 - if layout is right
5693 if (mddev->level == 0)
5694 return raid45_takeover_raid0(mddev, 4);
5695 if (mddev->level == 5 &&
5696 mddev->layout == ALGORITHM_PARITY_N) {
5697 mddev->new_layout = 0;
5698 mddev->new_level = 4;
5699 return setup_conf(mddev);
5701 return ERR_PTR(-EINVAL);
5704 static struct mdk_personality raid5_personality;
5706 static void *raid6_takeover(mddev_t *mddev)
5708 /* Currently can only take over a raid5. We map the
5709 * personality to an equivalent raid6 personality
5710 * with the Q block at the end.
5714 if (mddev->pers != &raid5_personality)
5715 return ERR_PTR(-EINVAL);
5716 if (mddev->degraded > 1)
5717 return ERR_PTR(-EINVAL);
5718 if (mddev->raid_disks > 253)
5719 return ERR_PTR(-EINVAL);
5720 if (mddev->raid_disks < 3)
5721 return ERR_PTR(-EINVAL);
5723 switch (mddev->layout) {
5724 case ALGORITHM_LEFT_ASYMMETRIC:
5725 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5727 case ALGORITHM_RIGHT_ASYMMETRIC:
5728 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5730 case ALGORITHM_LEFT_SYMMETRIC:
5731 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5733 case ALGORITHM_RIGHT_SYMMETRIC:
5734 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5736 case ALGORITHM_PARITY_0:
5737 new_layout = ALGORITHM_PARITY_0_6;
5739 case ALGORITHM_PARITY_N:
5740 new_layout = ALGORITHM_PARITY_N;
5743 return ERR_PTR(-EINVAL);
5745 mddev->new_level = 6;
5746 mddev->new_layout = new_layout;
5747 mddev->delta_disks = 1;
5748 mddev->raid_disks += 1;
5749 return setup_conf(mddev);
5753 static struct mdk_personality raid6_personality =
5757 .owner = THIS_MODULE,
5758 .make_request = make_request,
5762 .error_handler = error,
5763 .hot_add_disk = raid5_add_disk,
5764 .hot_remove_disk= raid5_remove_disk,
5765 .spare_active = raid5_spare_active,
5766 .sync_request = sync_request,
5767 .resize = raid5_resize,
5769 .check_reshape = raid6_check_reshape,
5770 .start_reshape = raid5_start_reshape,
5771 .finish_reshape = raid5_finish_reshape,
5772 .quiesce = raid5_quiesce,
5773 .takeover = raid6_takeover,
5775 static struct mdk_personality raid5_personality =
5779 .owner = THIS_MODULE,
5780 .make_request = make_request,
5784 .error_handler = error,
5785 .hot_add_disk = raid5_add_disk,
5786 .hot_remove_disk= raid5_remove_disk,
5787 .spare_active = raid5_spare_active,
5788 .sync_request = sync_request,
5789 .resize = raid5_resize,
5791 .check_reshape = raid5_check_reshape,
5792 .start_reshape = raid5_start_reshape,
5793 .finish_reshape = raid5_finish_reshape,
5794 .quiesce = raid5_quiesce,
5795 .takeover = raid5_takeover,
5798 static struct mdk_personality raid4_personality =
5802 .owner = THIS_MODULE,
5803 .make_request = make_request,
5807 .error_handler = error,
5808 .hot_add_disk = raid5_add_disk,
5809 .hot_remove_disk= raid5_remove_disk,
5810 .spare_active = raid5_spare_active,
5811 .sync_request = sync_request,
5812 .resize = raid5_resize,
5814 .check_reshape = raid5_check_reshape,
5815 .start_reshape = raid5_start_reshape,
5816 .finish_reshape = raid5_finish_reshape,
5817 .quiesce = raid5_quiesce,
5818 .takeover = raid4_takeover,
5821 static int __init raid5_init(void)
5823 register_md_personality(&raid6_personality);
5824 register_md_personality(&raid5_personality);
5825 register_md_personality(&raid4_personality);
5829 static void raid5_exit(void)
5831 unregister_md_personality(&raid6_personality);
5832 unregister_md_personality(&raid5_personality);
5833 unregister_md_personality(&raid4_personality);
5836 module_init(raid5_init);
5837 module_exit(raid5_exit);
5838 MODULE_LICENSE("GPL");
5839 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5840 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5841 MODULE_ALIAS("md-raid5");
5842 MODULE_ALIAS("md-raid4");
5843 MODULE_ALIAS("md-level-5");
5844 MODULE_ALIAS("md-level-4");
5845 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5846 MODULE_ALIAS("md-raid6");
5847 MODULE_ALIAS("md-level-6");
5849 /* This used to be two separate modules, they were: */
5850 MODULE_ALIAS("raid5");
5851 MODULE_ALIAS("raid6");