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 = -1;
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_dirtying(raid5_conf_t *conf,
2472 struct stripe_head *sh,
2473 struct stripe_head_state *s,
2476 int rmw = 0, rcw = 0, i;
2477 if (conf->max_degraded == 2) {
2478 /* RAID6 requires 'rcw' in current implementation
2479 * Calculate the real rcw later - for now fake it
2480 * look like rcw is cheaper
2483 } else for (i = disks; i--; ) {
2484 /* would I have to read this buffer for read_modify_write */
2485 struct r5dev *dev = &sh->dev[i];
2486 if ((dev->towrite || i == sh->pd_idx) &&
2487 !test_bit(R5_LOCKED, &dev->flags) &&
2488 !(test_bit(R5_UPTODATE, &dev->flags) ||
2489 test_bit(R5_Wantcompute, &dev->flags))) {
2490 if (test_bit(R5_Insync, &dev->flags))
2493 rmw += 2*disks; /* cannot read it */
2495 /* Would I have to read this buffer for reconstruct_write */
2496 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2497 !test_bit(R5_LOCKED, &dev->flags) &&
2498 !(test_bit(R5_UPTODATE, &dev->flags) ||
2499 test_bit(R5_Wantcompute, &dev->flags))) {
2500 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2505 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2506 (unsigned long long)sh->sector, rmw, rcw);
2507 set_bit(STRIPE_HANDLE, &sh->state);
2508 if (rmw < rcw && rmw > 0)
2509 /* prefer read-modify-write, but need to get some data */
2510 for (i = disks; i--; ) {
2511 struct r5dev *dev = &sh->dev[i];
2512 if ((dev->towrite || i == sh->pd_idx) &&
2513 !test_bit(R5_LOCKED, &dev->flags) &&
2514 !(test_bit(R5_UPTODATE, &dev->flags) ||
2515 test_bit(R5_Wantcompute, &dev->flags)) &&
2516 test_bit(R5_Insync, &dev->flags)) {
2518 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2519 pr_debug("Read_old block "
2520 "%d for r-m-w\n", i);
2521 set_bit(R5_LOCKED, &dev->flags);
2522 set_bit(R5_Wantread, &dev->flags);
2525 set_bit(STRIPE_DELAYED, &sh->state);
2526 set_bit(STRIPE_HANDLE, &sh->state);
2530 if (rcw <= rmw && rcw > 0) {
2531 /* want reconstruct write, but need to get some data */
2533 for (i = disks; i--; ) {
2534 struct r5dev *dev = &sh->dev[i];
2535 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2536 i != sh->pd_idx && i != sh->qd_idx &&
2537 !test_bit(R5_LOCKED, &dev->flags) &&
2538 !(test_bit(R5_UPTODATE, &dev->flags) ||
2539 test_bit(R5_Wantcompute, &dev->flags))) {
2541 if (!test_bit(R5_Insync, &dev->flags))
2542 continue; /* it's a failed drive */
2544 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2545 pr_debug("Read_old block "
2546 "%d for Reconstruct\n", i);
2547 set_bit(R5_LOCKED, &dev->flags);
2548 set_bit(R5_Wantread, &dev->flags);
2551 set_bit(STRIPE_DELAYED, &sh->state);
2552 set_bit(STRIPE_HANDLE, &sh->state);
2557 /* now if nothing is locked, and if we have enough data,
2558 * we can start a write request
2560 /* since handle_stripe can be called at any time we need to handle the
2561 * case where a compute block operation has been submitted and then a
2562 * subsequent call wants to start a write request. raid_run_ops only
2563 * handles the case where compute block and reconstruct are requested
2564 * simultaneously. If this is not the case then new writes need to be
2565 * held off until the compute completes.
2567 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2568 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2569 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2570 schedule_reconstruction(sh, s, rcw == 0, 0);
2573 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2574 struct stripe_head_state *s, int disks)
2576 struct r5dev *dev = NULL;
2578 set_bit(STRIPE_HANDLE, &sh->state);
2580 switch (sh->check_state) {
2581 case check_state_idle:
2582 /* start a new check operation if there are no failures */
2583 if (s->failed == 0) {
2584 BUG_ON(s->uptodate != disks);
2585 sh->check_state = check_state_run;
2586 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2587 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2591 dev = &sh->dev[s->failed_num[0]];
2593 case check_state_compute_result:
2594 sh->check_state = check_state_idle;
2596 dev = &sh->dev[sh->pd_idx];
2598 /* check that a write has not made the stripe insync */
2599 if (test_bit(STRIPE_INSYNC, &sh->state))
2602 /* either failed parity check, or recovery is happening */
2603 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2604 BUG_ON(s->uptodate != disks);
2606 set_bit(R5_LOCKED, &dev->flags);
2608 set_bit(R5_Wantwrite, &dev->flags);
2610 clear_bit(STRIPE_DEGRADED, &sh->state);
2611 set_bit(STRIPE_INSYNC, &sh->state);
2613 case check_state_run:
2614 break; /* we will be called again upon completion */
2615 case check_state_check_result:
2616 sh->check_state = check_state_idle;
2618 /* if a failure occurred during the check operation, leave
2619 * STRIPE_INSYNC not set and let the stripe be handled again
2624 /* handle a successful check operation, if parity is correct
2625 * we are done. Otherwise update the mismatch count and repair
2626 * parity if !MD_RECOVERY_CHECK
2628 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2629 /* parity is correct (on disc,
2630 * not in buffer any more)
2632 set_bit(STRIPE_INSYNC, &sh->state);
2634 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2635 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2636 /* don't try to repair!! */
2637 set_bit(STRIPE_INSYNC, &sh->state);
2639 sh->check_state = check_state_compute_run;
2640 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2641 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2642 set_bit(R5_Wantcompute,
2643 &sh->dev[sh->pd_idx].flags);
2644 sh->ops.target = sh->pd_idx;
2645 sh->ops.target2 = -1;
2650 case check_state_compute_run:
2653 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2654 __func__, sh->check_state,
2655 (unsigned long long) sh->sector);
2661 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2662 struct stripe_head_state *s,
2665 int pd_idx = sh->pd_idx;
2666 int qd_idx = sh->qd_idx;
2669 set_bit(STRIPE_HANDLE, &sh->state);
2671 BUG_ON(s->failed > 2);
2673 /* Want to check and possibly repair P and Q.
2674 * However there could be one 'failed' device, in which
2675 * case we can only check one of them, possibly using the
2676 * other to generate missing data
2679 switch (sh->check_state) {
2680 case check_state_idle:
2681 /* start a new check operation if there are < 2 failures */
2682 if (s->failed == s->q_failed) {
2683 /* The only possible failed device holds Q, so it
2684 * makes sense to check P (If anything else were failed,
2685 * we would have used P to recreate it).
2687 sh->check_state = check_state_run;
2689 if (!s->q_failed && s->failed < 2) {
2690 /* Q is not failed, and we didn't use it to generate
2691 * anything, so it makes sense to check it
2693 if (sh->check_state == check_state_run)
2694 sh->check_state = check_state_run_pq;
2696 sh->check_state = check_state_run_q;
2699 /* discard potentially stale zero_sum_result */
2700 sh->ops.zero_sum_result = 0;
2702 if (sh->check_state == check_state_run) {
2703 /* async_xor_zero_sum destroys the contents of P */
2704 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2707 if (sh->check_state >= check_state_run &&
2708 sh->check_state <= check_state_run_pq) {
2709 /* async_syndrome_zero_sum preserves P and Q, so
2710 * no need to mark them !uptodate here
2712 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2716 /* we have 2-disk failure */
2717 BUG_ON(s->failed != 2);
2719 case check_state_compute_result:
2720 sh->check_state = check_state_idle;
2722 /* check that a write has not made the stripe insync */
2723 if (test_bit(STRIPE_INSYNC, &sh->state))
2726 /* now write out any block on a failed drive,
2727 * or P or Q if they were recomputed
2729 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2730 if (s->failed == 2) {
2731 dev = &sh->dev[s->failed_num[1]];
2733 set_bit(R5_LOCKED, &dev->flags);
2734 set_bit(R5_Wantwrite, &dev->flags);
2736 if (s->failed >= 1) {
2737 dev = &sh->dev[s->failed_num[0]];
2739 set_bit(R5_LOCKED, &dev->flags);
2740 set_bit(R5_Wantwrite, &dev->flags);
2742 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2743 dev = &sh->dev[pd_idx];
2745 set_bit(R5_LOCKED, &dev->flags);
2746 set_bit(R5_Wantwrite, &dev->flags);
2748 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2749 dev = &sh->dev[qd_idx];
2751 set_bit(R5_LOCKED, &dev->flags);
2752 set_bit(R5_Wantwrite, &dev->flags);
2754 clear_bit(STRIPE_DEGRADED, &sh->state);
2756 set_bit(STRIPE_INSYNC, &sh->state);
2758 case check_state_run:
2759 case check_state_run_q:
2760 case check_state_run_pq:
2761 break; /* we will be called again upon completion */
2762 case check_state_check_result:
2763 sh->check_state = check_state_idle;
2765 /* handle a successful check operation, if parity is correct
2766 * we are done. Otherwise update the mismatch count and repair
2767 * parity if !MD_RECOVERY_CHECK
2769 if (sh->ops.zero_sum_result == 0) {
2770 /* both parities are correct */
2772 set_bit(STRIPE_INSYNC, &sh->state);
2774 /* in contrast to the raid5 case we can validate
2775 * parity, but still have a failure to write
2778 sh->check_state = check_state_compute_result;
2779 /* Returning at this point means that we may go
2780 * off and bring p and/or q uptodate again so
2781 * we make sure to check zero_sum_result again
2782 * to verify if p or q need writeback
2786 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2787 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2788 /* don't try to repair!! */
2789 set_bit(STRIPE_INSYNC, &sh->state);
2791 int *target = &sh->ops.target;
2793 sh->ops.target = -1;
2794 sh->ops.target2 = -1;
2795 sh->check_state = check_state_compute_run;
2796 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2797 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2798 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2799 set_bit(R5_Wantcompute,
2800 &sh->dev[pd_idx].flags);
2802 target = &sh->ops.target2;
2805 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2806 set_bit(R5_Wantcompute,
2807 &sh->dev[qd_idx].flags);
2814 case check_state_compute_run:
2817 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2818 __func__, sh->check_state,
2819 (unsigned long long) sh->sector);
2824 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh)
2828 /* We have read all the blocks in this stripe and now we need to
2829 * copy some of them into a target stripe for expand.
2831 struct dma_async_tx_descriptor *tx = NULL;
2832 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2833 for (i = 0; i < sh->disks; i++)
2834 if (i != sh->pd_idx && i != sh->qd_idx) {
2836 struct stripe_head *sh2;
2837 struct async_submit_ctl submit;
2839 sector_t bn = compute_blocknr(sh, i, 1);
2840 sector_t s = raid5_compute_sector(conf, bn, 0,
2842 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2844 /* so far only the early blocks of this stripe
2845 * have been requested. When later blocks
2846 * get requested, we will try again
2849 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2850 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2851 /* must have already done this block */
2852 release_stripe(sh2);
2856 /* place all the copies on one channel */
2857 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2858 tx = async_memcpy(sh2->dev[dd_idx].page,
2859 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2862 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2863 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2864 for (j = 0; j < conf->raid_disks; j++)
2865 if (j != sh2->pd_idx &&
2867 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2869 if (j == conf->raid_disks) {
2870 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2871 set_bit(STRIPE_HANDLE, &sh2->state);
2873 release_stripe(sh2);
2876 /* done submitting copies, wait for them to complete */
2879 dma_wait_for_async_tx(tx);
2885 * handle_stripe - do things to a stripe.
2887 * We lock the stripe and then examine the state of various bits
2888 * to see what needs to be done.
2890 * return some read request which now have data
2891 * return some write requests which are safely on disc
2892 * schedule a read on some buffers
2893 * schedule a write of some buffers
2894 * return confirmation of parity correctness
2896 * buffers are taken off read_list or write_list, and bh_cache buffers
2897 * get BH_Lock set before the stripe lock is released.
2901 static int handle_stripe5(struct stripe_head *sh, struct stripe_head_state *s)
2903 raid5_conf_t *conf = sh->raid_conf;
2904 int disks = sh->disks, i;
2907 /* Now to look around and see what can be done */
2909 spin_lock_irq(&conf->device_lock);
2910 for (i=disks; i--; ) {
2915 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2916 "written %p\n", i, dev->flags, dev->toread, dev->read,
2917 dev->towrite, dev->written);
2919 /* maybe we can request a biofill operation
2921 * new wantfill requests are only permitted while
2922 * ops_complete_biofill is guaranteed to be inactive
2924 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2925 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2926 set_bit(R5_Wantfill, &dev->flags);
2928 /* now count some things */
2929 if (test_bit(R5_LOCKED, &dev->flags))
2931 if (test_bit(R5_UPTODATE, &dev->flags))
2933 if (test_bit(R5_Wantcompute, &dev->flags))
2936 if (test_bit(R5_Wantfill, &dev->flags))
2938 else if (dev->toread)
2942 if (!test_bit(R5_OVERWRITE, &dev->flags))
2947 rdev = rcu_dereference(conf->disks[i].rdev);
2948 if (s->blocked_rdev == NULL &&
2949 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2950 s->blocked_rdev = rdev;
2951 atomic_inc(&rdev->nr_pending);
2953 clear_bit(R5_Insync, &dev->flags);
2956 else if (test_bit(In_sync, &rdev->flags))
2957 set_bit(R5_Insync, &dev->flags);
2959 /* could be in-sync depending on recovery/reshape status */
2960 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
2961 set_bit(R5_Insync, &dev->flags);
2963 if (!test_bit(R5_Insync, &dev->flags)) {
2964 /* The ReadError flag will just be confusing now */
2965 clear_bit(R5_ReadError, &dev->flags);
2966 clear_bit(R5_ReWrite, &dev->flags);
2968 if (test_bit(R5_ReadError, &dev->flags))
2969 clear_bit(R5_Insync, &dev->flags);
2970 if (!test_bit(R5_Insync, &dev->flags)) {
2972 s->failed_num[s->failed] = i;
2976 spin_unlock_irq(&conf->device_lock);
2982 static int handle_stripe6(struct stripe_head *sh, struct stripe_head_state *s)
2984 raid5_conf_t *conf = sh->raid_conf;
2985 int disks = sh->disks;
2989 /* Now to look around and see what can be done */
2992 spin_lock_irq(&conf->device_lock);
2993 for (i=disks; i--; ) {
2997 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2998 i, dev->flags, dev->toread, dev->towrite, dev->written);
2999 /* maybe we can reply to a read
3001 * new wantfill requests are only permitted while
3002 * ops_complete_biofill is guaranteed to be inactive
3004 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3005 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3006 set_bit(R5_Wantfill, &dev->flags);
3008 /* now count some things */
3009 if (test_bit(R5_LOCKED, &dev->flags))
3011 if (test_bit(R5_UPTODATE, &dev->flags))
3013 if (test_bit(R5_Wantcompute, &dev->flags)) {
3015 BUG_ON(s->compute > 2);
3018 if (test_bit(R5_Wantfill, &dev->flags)) {
3020 } else if (dev->toread)
3024 if (!test_bit(R5_OVERWRITE, &dev->flags))
3029 rdev = rcu_dereference(conf->disks[i].rdev);
3030 if (s->blocked_rdev == NULL &&
3031 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3032 s->blocked_rdev = rdev;
3033 atomic_inc(&rdev->nr_pending);
3035 clear_bit(R5_Insync, &dev->flags);
3038 else if (test_bit(In_sync, &rdev->flags))
3039 set_bit(R5_Insync, &dev->flags);
3041 /* in sync if before recovery_offset */
3042 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3043 set_bit(R5_Insync, &dev->flags);
3045 if (!test_bit(R5_Insync, &dev->flags)) {
3046 /* The ReadError flag will just be confusing now */
3047 clear_bit(R5_ReadError, &dev->flags);
3048 clear_bit(R5_ReWrite, &dev->flags);
3050 if (test_bit(R5_ReadError, &dev->flags))
3051 clear_bit(R5_Insync, &dev->flags);
3052 if (!test_bit(R5_Insync, &dev->flags)) {
3054 s->failed_num[s->failed] = i;
3058 spin_unlock_irq(&conf->device_lock);
3064 static void handle_stripe(struct stripe_head *sh)
3066 struct stripe_head_state s;
3068 raid5_conf_t *conf = sh->raid_conf;
3071 int disks = sh->disks;
3072 struct r5dev *pdev, *qdev;
3074 clear_bit(STRIPE_HANDLE, &sh->state);
3075 if (test_and_set_bit(STRIPE_ACTIVE, &sh->state)) {
3076 /* already being handled, ensure it gets handled
3077 * again when current action finishes */
3078 set_bit(STRIPE_HANDLE, &sh->state);
3082 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3083 set_bit(STRIPE_SYNCING, &sh->state);
3084 clear_bit(STRIPE_INSYNC, &sh->state);
3086 clear_bit(STRIPE_DELAYED, &sh->state);
3088 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3089 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3090 (unsigned long long)sh->sector, sh->state,
3091 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3092 sh->check_state, sh->reconstruct_state);
3093 memset(&s, 0, sizeof(s));
3095 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3096 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3097 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3098 s.failed_num[0] = -1;
3099 s.failed_num[1] = -1;
3101 if (conf->level == 6)
3102 done = handle_stripe6(sh, &s);
3104 done = handle_stripe5(sh, &s);
3109 if (unlikely(s.blocked_rdev)) {
3110 if (s.syncing || s.expanding || s.expanded ||
3111 s.to_write || s.written) {
3112 set_bit(STRIPE_HANDLE, &sh->state);
3115 /* There is nothing for the blocked_rdev to block */
3116 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3117 s.blocked_rdev = NULL;
3120 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3121 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3122 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3125 pr_debug("locked=%d uptodate=%d to_read=%d"
3126 " to_write=%d failed=%d failed_num=%d,%d\n",
3127 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3128 s.failed_num[0], s.failed_num[1]);
3129 /* check if the array has lost more than max_degraded devices and,
3130 * if so, some requests might need to be failed.
3132 if (s.failed > conf->max_degraded && s.to_read+s.to_write+s.written)
3133 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3134 if (s.failed > conf->max_degraded && s.syncing) {
3135 md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
3136 clear_bit(STRIPE_SYNCING, &sh->state);
3141 * might be able to return some write requests if the parity blocks
3142 * are safe, or on a failed drive
3144 pdev = &sh->dev[sh->pd_idx];
3145 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3146 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3147 qdev = &sh->dev[sh->qd_idx];
3148 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3149 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3153 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3154 && !test_bit(R5_LOCKED, &pdev->flags)
3155 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3156 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3157 && !test_bit(R5_LOCKED, &qdev->flags)
3158 && test_bit(R5_UPTODATE, &qdev->flags)))))
3159 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3161 /* Now we might consider reading some blocks, either to check/generate
3162 * parity, or to satisfy requests
3163 * or to load a block that is being partially written.
3165 if (s.to_read || s.non_overwrite
3166 || (conf->level == 6 && s.to_write && s.failed)
3167 || (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3168 handle_stripe_fill(sh, &s, disks);
3170 /* Now we check to see if any write operations have recently
3174 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3176 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3177 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3178 sh->reconstruct_state = reconstruct_state_idle;
3180 /* All the 'written' buffers and the parity block are ready to
3181 * be written back to disk
3183 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3184 BUG_ON(sh->qd_idx >= 0 &&
3185 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3186 for (i = disks; i--; ) {
3187 struct r5dev *dev = &sh->dev[i];
3188 if (test_bit(R5_LOCKED, &dev->flags) &&
3189 (i == sh->pd_idx || i == sh->qd_idx ||
3191 pr_debug("Writing block %d\n", i);
3192 set_bit(R5_Wantwrite, &dev->flags);
3195 if (!test_bit(R5_Insync, &dev->flags) ||
3196 ((i == sh->pd_idx || i == sh->qd_idx) &&
3198 set_bit(STRIPE_INSYNC, &sh->state);
3201 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3202 s.dec_preread_active = 1;
3205 /* Now to consider new write requests and what else, if anything
3206 * should be read. We do not handle new writes when:
3207 * 1/ A 'write' operation (copy+xor) is already in flight.
3208 * 2/ A 'check' operation is in flight, as it may clobber the parity
3211 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3212 handle_stripe_dirtying(conf, sh, &s, disks);
3214 /* maybe we need to check and possibly fix the parity for this stripe
3215 * Any reads will already have been scheduled, so we just see if enough
3216 * data is available. The parity check is held off while parity
3217 * dependent operations are in flight.
3219 if (sh->check_state ||
3220 (s.syncing && s.locked == 0 &&
3221 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3222 !test_bit(STRIPE_INSYNC, &sh->state))) {
3223 if (conf->level == 6)
3224 handle_parity_checks6(conf, sh, &s, disks);
3226 handle_parity_checks5(conf, sh, &s, disks);
3229 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3230 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3231 clear_bit(STRIPE_SYNCING, &sh->state);
3234 /* If the failed drives are just a ReadError, then we might need
3235 * to progress the repair/check process
3237 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3238 for (i = 0; i < s.failed; i++) {
3239 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3240 if (test_bit(R5_ReadError, &dev->flags)
3241 && !test_bit(R5_LOCKED, &dev->flags)
3242 && test_bit(R5_UPTODATE, &dev->flags)
3244 if (!test_bit(R5_ReWrite, &dev->flags)) {
3245 set_bit(R5_Wantwrite, &dev->flags);
3246 set_bit(R5_ReWrite, &dev->flags);
3247 set_bit(R5_LOCKED, &dev->flags);
3250 /* let's read it back */
3251 set_bit(R5_Wantread, &dev->flags);
3252 set_bit(R5_LOCKED, &dev->flags);
3259 /* Finish reconstruct operations initiated by the expansion process */
3260 if (sh->reconstruct_state == reconstruct_state_result) {
3261 struct stripe_head *sh_src
3262 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3263 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3264 /* sh cannot be written until sh_src has been read.
3265 * so arrange for sh to be delayed a little
3267 set_bit(STRIPE_DELAYED, &sh->state);
3268 set_bit(STRIPE_HANDLE, &sh->state);
3269 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3271 atomic_inc(&conf->preread_active_stripes);
3272 release_stripe(sh_src);
3276 release_stripe(sh_src);
3278 sh->reconstruct_state = reconstruct_state_idle;
3279 clear_bit(STRIPE_EXPANDING, &sh->state);
3280 for (i = conf->raid_disks; i--; ) {
3281 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3282 set_bit(R5_LOCKED, &sh->dev[i].flags);
3287 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3288 !sh->reconstruct_state) {
3289 /* Need to write out all blocks after computing parity */
3290 sh->disks = conf->raid_disks;
3291 stripe_set_idx(sh->sector, conf, 0, sh);
3292 schedule_reconstruction(sh, &s, 1, 1);
3293 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3294 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3295 atomic_dec(&conf->reshape_stripes);
3296 wake_up(&conf->wait_for_overlap);
3297 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3300 if (s.expanding && s.locked == 0 &&
3301 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3302 handle_stripe_expansion(conf, sh);
3305 /* wait for this device to become unblocked */
3306 if (unlikely(s.blocked_rdev))
3307 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3310 raid_run_ops(sh, s.ops_request);
3315 if (s.dec_preread_active) {
3316 /* We delay this until after ops_run_io so that if make_request
3317 * is waiting on a flush, it won't continue until the writes
3318 * have actually been submitted.
3320 atomic_dec(&conf->preread_active_stripes);
3321 if (atomic_read(&conf->preread_active_stripes) <
3323 md_wakeup_thread(conf->mddev->thread);
3326 return_io(s.return_bi);
3328 clear_bit(STRIPE_ACTIVE, &sh->state);
3331 static void raid5_activate_delayed(raid5_conf_t *conf)
3333 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3334 while (!list_empty(&conf->delayed_list)) {
3335 struct list_head *l = conf->delayed_list.next;
3336 struct stripe_head *sh;
3337 sh = list_entry(l, struct stripe_head, lru);
3339 clear_bit(STRIPE_DELAYED, &sh->state);
3340 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3341 atomic_inc(&conf->preread_active_stripes);
3342 list_add_tail(&sh->lru, &conf->hold_list);
3347 static void activate_bit_delay(raid5_conf_t *conf)
3349 /* device_lock is held */
3350 struct list_head head;
3351 list_add(&head, &conf->bitmap_list);
3352 list_del_init(&conf->bitmap_list);
3353 while (!list_empty(&head)) {
3354 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3355 list_del_init(&sh->lru);
3356 atomic_inc(&sh->count);
3357 __release_stripe(conf, sh);
3361 int md_raid5_congested(mddev_t *mddev, int bits)
3363 raid5_conf_t *conf = mddev->private;
3365 /* No difference between reads and writes. Just check
3366 * how busy the stripe_cache is
3369 if (conf->inactive_blocked)
3373 if (list_empty_careful(&conf->inactive_list))
3378 EXPORT_SYMBOL_GPL(md_raid5_congested);
3380 static int raid5_congested(void *data, int bits)
3382 mddev_t *mddev = data;
3384 return mddev_congested(mddev, bits) ||
3385 md_raid5_congested(mddev, bits);
3388 /* We want read requests to align with chunks where possible,
3389 * but write requests don't need to.
3391 static int raid5_mergeable_bvec(struct request_queue *q,
3392 struct bvec_merge_data *bvm,
3393 struct bio_vec *biovec)
3395 mddev_t *mddev = q->queuedata;
3396 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3398 unsigned int chunk_sectors = mddev->chunk_sectors;
3399 unsigned int bio_sectors = bvm->bi_size >> 9;
3401 if ((bvm->bi_rw & 1) == WRITE)
3402 return biovec->bv_len; /* always allow writes to be mergeable */
3404 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3405 chunk_sectors = mddev->new_chunk_sectors;
3406 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3407 if (max < 0) max = 0;
3408 if (max <= biovec->bv_len && bio_sectors == 0)
3409 return biovec->bv_len;
3415 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3417 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3418 unsigned int chunk_sectors = mddev->chunk_sectors;
3419 unsigned int bio_sectors = bio->bi_size >> 9;
3421 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3422 chunk_sectors = mddev->new_chunk_sectors;
3423 return chunk_sectors >=
3424 ((sector & (chunk_sectors - 1)) + bio_sectors);
3428 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3429 * later sampled by raid5d.
3431 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3433 unsigned long flags;
3435 spin_lock_irqsave(&conf->device_lock, flags);
3437 bi->bi_next = conf->retry_read_aligned_list;
3438 conf->retry_read_aligned_list = bi;
3440 spin_unlock_irqrestore(&conf->device_lock, flags);
3441 md_wakeup_thread(conf->mddev->thread);
3445 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3449 bi = conf->retry_read_aligned;
3451 conf->retry_read_aligned = NULL;
3454 bi = conf->retry_read_aligned_list;
3456 conf->retry_read_aligned_list = bi->bi_next;
3459 * this sets the active strip count to 1 and the processed
3460 * strip count to zero (upper 8 bits)
3462 bi->bi_phys_segments = 1; /* biased count of active stripes */
3470 * The "raid5_align_endio" should check if the read succeeded and if it
3471 * did, call bio_endio on the original bio (having bio_put the new bio
3473 * If the read failed..
3475 static void raid5_align_endio(struct bio *bi, int error)
3477 struct bio* raid_bi = bi->bi_private;
3480 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3485 rdev = (void*)raid_bi->bi_next;
3486 raid_bi->bi_next = NULL;
3487 mddev = rdev->mddev;
3488 conf = mddev->private;
3490 rdev_dec_pending(rdev, conf->mddev);
3492 if (!error && uptodate) {
3493 bio_endio(raid_bi, 0);
3494 if (atomic_dec_and_test(&conf->active_aligned_reads))
3495 wake_up(&conf->wait_for_stripe);
3500 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3502 add_bio_to_retry(raid_bi, conf);
3505 static int bio_fits_rdev(struct bio *bi)
3507 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3509 if ((bi->bi_size>>9) > queue_max_sectors(q))
3511 blk_recount_segments(q, bi);
3512 if (bi->bi_phys_segments > queue_max_segments(q))
3515 if (q->merge_bvec_fn)
3516 /* it's too hard to apply the merge_bvec_fn at this stage,
3525 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio)
3527 raid5_conf_t *conf = mddev->private;
3529 struct bio* align_bi;
3532 if (!in_chunk_boundary(mddev, raid_bio)) {
3533 pr_debug("chunk_aligned_read : non aligned\n");
3537 * use bio_clone_mddev to make a copy of the bio
3539 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3543 * set bi_end_io to a new function, and set bi_private to the
3546 align_bi->bi_end_io = raid5_align_endio;
3547 align_bi->bi_private = raid_bio;
3551 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3556 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3557 if (rdev && test_bit(In_sync, &rdev->flags)) {
3558 atomic_inc(&rdev->nr_pending);
3560 raid_bio->bi_next = (void*)rdev;
3561 align_bi->bi_bdev = rdev->bdev;
3562 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3563 align_bi->bi_sector += rdev->data_offset;
3565 if (!bio_fits_rdev(align_bi)) {
3566 /* too big in some way */
3568 rdev_dec_pending(rdev, mddev);
3572 spin_lock_irq(&conf->device_lock);
3573 wait_event_lock_irq(conf->wait_for_stripe,
3575 conf->device_lock, /* nothing */);
3576 atomic_inc(&conf->active_aligned_reads);
3577 spin_unlock_irq(&conf->device_lock);
3579 generic_make_request(align_bi);
3588 /* __get_priority_stripe - get the next stripe to process
3590 * Full stripe writes are allowed to pass preread active stripes up until
3591 * the bypass_threshold is exceeded. In general the bypass_count
3592 * increments when the handle_list is handled before the hold_list; however, it
3593 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3594 * stripe with in flight i/o. The bypass_count will be reset when the
3595 * head of the hold_list has changed, i.e. the head was promoted to the
3598 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3600 struct stripe_head *sh;
3602 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3604 list_empty(&conf->handle_list) ? "empty" : "busy",
3605 list_empty(&conf->hold_list) ? "empty" : "busy",
3606 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3608 if (!list_empty(&conf->handle_list)) {
3609 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3611 if (list_empty(&conf->hold_list))
3612 conf->bypass_count = 0;
3613 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3614 if (conf->hold_list.next == conf->last_hold)
3615 conf->bypass_count++;
3617 conf->last_hold = conf->hold_list.next;
3618 conf->bypass_count -= conf->bypass_threshold;
3619 if (conf->bypass_count < 0)
3620 conf->bypass_count = 0;
3623 } else if (!list_empty(&conf->hold_list) &&
3624 ((conf->bypass_threshold &&
3625 conf->bypass_count > conf->bypass_threshold) ||
3626 atomic_read(&conf->pending_full_writes) == 0)) {
3627 sh = list_entry(conf->hold_list.next,
3629 conf->bypass_count -= conf->bypass_threshold;
3630 if (conf->bypass_count < 0)
3631 conf->bypass_count = 0;
3635 list_del_init(&sh->lru);
3636 atomic_inc(&sh->count);
3637 BUG_ON(atomic_read(&sh->count) != 1);
3641 static int make_request(mddev_t *mddev, struct bio * bi)
3643 raid5_conf_t *conf = mddev->private;
3645 sector_t new_sector;
3646 sector_t logical_sector, last_sector;
3647 struct stripe_head *sh;
3648 const int rw = bio_data_dir(bi);
3652 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3653 md_flush_request(mddev, bi);
3657 md_write_start(mddev, bi);
3660 mddev->reshape_position == MaxSector &&
3661 chunk_aligned_read(mddev,bi))
3664 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3665 last_sector = bi->bi_sector + (bi->bi_size>>9);
3667 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3669 plugged = mddev_check_plugged(mddev);
3670 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3672 int disks, data_disks;
3677 disks = conf->raid_disks;
3678 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3679 if (unlikely(conf->reshape_progress != MaxSector)) {
3680 /* spinlock is needed as reshape_progress may be
3681 * 64bit on a 32bit platform, and so it might be
3682 * possible to see a half-updated value
3683 * Of course reshape_progress could change after
3684 * the lock is dropped, so once we get a reference
3685 * to the stripe that we think it is, we will have
3688 spin_lock_irq(&conf->device_lock);
3689 if (mddev->delta_disks < 0
3690 ? logical_sector < conf->reshape_progress
3691 : logical_sector >= conf->reshape_progress) {
3692 disks = conf->previous_raid_disks;
3695 if (mddev->delta_disks < 0
3696 ? logical_sector < conf->reshape_safe
3697 : logical_sector >= conf->reshape_safe) {
3698 spin_unlock_irq(&conf->device_lock);
3703 spin_unlock_irq(&conf->device_lock);
3705 data_disks = disks - conf->max_degraded;
3707 new_sector = raid5_compute_sector(conf, logical_sector,
3710 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3711 (unsigned long long)new_sector,
3712 (unsigned long long)logical_sector);
3714 sh = get_active_stripe(conf, new_sector, previous,
3715 (bi->bi_rw&RWA_MASK), 0);
3717 if (unlikely(previous)) {
3718 /* expansion might have moved on while waiting for a
3719 * stripe, so we must do the range check again.
3720 * Expansion could still move past after this
3721 * test, but as we are holding a reference to
3722 * 'sh', we know that if that happens,
3723 * STRIPE_EXPANDING will get set and the expansion
3724 * won't proceed until we finish with the stripe.
3727 spin_lock_irq(&conf->device_lock);
3728 if (mddev->delta_disks < 0
3729 ? logical_sector >= conf->reshape_progress
3730 : logical_sector < conf->reshape_progress)
3731 /* mismatch, need to try again */
3733 spin_unlock_irq(&conf->device_lock);
3742 logical_sector >= mddev->suspend_lo &&
3743 logical_sector < mddev->suspend_hi) {
3745 /* As the suspend_* range is controlled by
3746 * userspace, we want an interruptible
3749 flush_signals(current);
3750 prepare_to_wait(&conf->wait_for_overlap,
3751 &w, TASK_INTERRUPTIBLE);
3752 if (logical_sector >= mddev->suspend_lo &&
3753 logical_sector < mddev->suspend_hi)
3758 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3759 !add_stripe_bio(sh, bi, dd_idx, rw)) {
3760 /* Stripe is busy expanding or
3761 * add failed due to overlap. Flush everything
3764 md_wakeup_thread(mddev->thread);
3769 finish_wait(&conf->wait_for_overlap, &w);
3770 set_bit(STRIPE_HANDLE, &sh->state);
3771 clear_bit(STRIPE_DELAYED, &sh->state);
3772 if ((bi->bi_rw & REQ_SYNC) &&
3773 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3774 atomic_inc(&conf->preread_active_stripes);
3777 /* cannot get stripe for read-ahead, just give-up */
3778 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3779 finish_wait(&conf->wait_for_overlap, &w);
3785 md_wakeup_thread(mddev->thread);
3787 spin_lock_irq(&conf->device_lock);
3788 remaining = raid5_dec_bi_phys_segments(bi);
3789 spin_unlock_irq(&conf->device_lock);
3790 if (remaining == 0) {
3793 md_write_end(mddev);
3801 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3803 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3805 /* reshaping is quite different to recovery/resync so it is
3806 * handled quite separately ... here.
3808 * On each call to sync_request, we gather one chunk worth of
3809 * destination stripes and flag them as expanding.
3810 * Then we find all the source stripes and request reads.
3811 * As the reads complete, handle_stripe will copy the data
3812 * into the destination stripe and release that stripe.
3814 raid5_conf_t *conf = mddev->private;
3815 struct stripe_head *sh;
3816 sector_t first_sector, last_sector;
3817 int raid_disks = conf->previous_raid_disks;
3818 int data_disks = raid_disks - conf->max_degraded;
3819 int new_data_disks = conf->raid_disks - conf->max_degraded;
3822 sector_t writepos, readpos, safepos;
3823 sector_t stripe_addr;
3824 int reshape_sectors;
3825 struct list_head stripes;
3827 if (sector_nr == 0) {
3828 /* If restarting in the middle, skip the initial sectors */
3829 if (mddev->delta_disks < 0 &&
3830 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3831 sector_nr = raid5_size(mddev, 0, 0)
3832 - conf->reshape_progress;
3833 } else if (mddev->delta_disks >= 0 &&
3834 conf->reshape_progress > 0)
3835 sector_nr = conf->reshape_progress;
3836 sector_div(sector_nr, new_data_disks);
3838 mddev->curr_resync_completed = sector_nr;
3839 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3845 /* We need to process a full chunk at a time.
3846 * If old and new chunk sizes differ, we need to process the
3849 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3850 reshape_sectors = mddev->new_chunk_sectors;
3852 reshape_sectors = mddev->chunk_sectors;
3854 /* we update the metadata when there is more than 3Meg
3855 * in the block range (that is rather arbitrary, should
3856 * probably be time based) or when the data about to be
3857 * copied would over-write the source of the data at
3858 * the front of the range.
3859 * i.e. one new_stripe along from reshape_progress new_maps
3860 * to after where reshape_safe old_maps to
3862 writepos = conf->reshape_progress;
3863 sector_div(writepos, new_data_disks);
3864 readpos = conf->reshape_progress;
3865 sector_div(readpos, data_disks);
3866 safepos = conf->reshape_safe;
3867 sector_div(safepos, data_disks);
3868 if (mddev->delta_disks < 0) {
3869 writepos -= min_t(sector_t, reshape_sectors, writepos);
3870 readpos += reshape_sectors;
3871 safepos += reshape_sectors;
3873 writepos += reshape_sectors;
3874 readpos -= min_t(sector_t, reshape_sectors, readpos);
3875 safepos -= min_t(sector_t, reshape_sectors, safepos);
3878 /* 'writepos' is the most advanced device address we might write.
3879 * 'readpos' is the least advanced device address we might read.
3880 * 'safepos' is the least address recorded in the metadata as having
3882 * If 'readpos' is behind 'writepos', then there is no way that we can
3883 * ensure safety in the face of a crash - that must be done by userspace
3884 * making a backup of the data. So in that case there is no particular
3885 * rush to update metadata.
3886 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3887 * update the metadata to advance 'safepos' to match 'readpos' so that
3888 * we can be safe in the event of a crash.
3889 * So we insist on updating metadata if safepos is behind writepos and
3890 * readpos is beyond writepos.
3891 * In any case, update the metadata every 10 seconds.
3892 * Maybe that number should be configurable, but I'm not sure it is
3893 * worth it.... maybe it could be a multiple of safemode_delay???
3895 if ((mddev->delta_disks < 0
3896 ? (safepos > writepos && readpos < writepos)
3897 : (safepos < writepos && readpos > writepos)) ||
3898 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3899 /* Cannot proceed until we've updated the superblock... */
3900 wait_event(conf->wait_for_overlap,
3901 atomic_read(&conf->reshape_stripes)==0);
3902 mddev->reshape_position = conf->reshape_progress;
3903 mddev->curr_resync_completed = sector_nr;
3904 conf->reshape_checkpoint = jiffies;
3905 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3906 md_wakeup_thread(mddev->thread);
3907 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3908 kthread_should_stop());
3909 spin_lock_irq(&conf->device_lock);
3910 conf->reshape_safe = mddev->reshape_position;
3911 spin_unlock_irq(&conf->device_lock);
3912 wake_up(&conf->wait_for_overlap);
3913 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3916 if (mddev->delta_disks < 0) {
3917 BUG_ON(conf->reshape_progress == 0);
3918 stripe_addr = writepos;
3919 BUG_ON((mddev->dev_sectors &
3920 ~((sector_t)reshape_sectors - 1))
3921 - reshape_sectors - stripe_addr
3924 BUG_ON(writepos != sector_nr + reshape_sectors);
3925 stripe_addr = sector_nr;
3927 INIT_LIST_HEAD(&stripes);
3928 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3930 int skipped_disk = 0;
3931 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
3932 set_bit(STRIPE_EXPANDING, &sh->state);
3933 atomic_inc(&conf->reshape_stripes);
3934 /* If any of this stripe is beyond the end of the old
3935 * array, then we need to zero those blocks
3937 for (j=sh->disks; j--;) {
3939 if (j == sh->pd_idx)
3941 if (conf->level == 6 &&
3944 s = compute_blocknr(sh, j, 0);
3945 if (s < raid5_size(mddev, 0, 0)) {
3949 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3950 set_bit(R5_Expanded, &sh->dev[j].flags);
3951 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3953 if (!skipped_disk) {
3954 set_bit(STRIPE_EXPAND_READY, &sh->state);
3955 set_bit(STRIPE_HANDLE, &sh->state);
3957 list_add(&sh->lru, &stripes);
3959 spin_lock_irq(&conf->device_lock);
3960 if (mddev->delta_disks < 0)
3961 conf->reshape_progress -= reshape_sectors * new_data_disks;
3963 conf->reshape_progress += reshape_sectors * new_data_disks;
3964 spin_unlock_irq(&conf->device_lock);
3965 /* Ok, those stripe are ready. We can start scheduling
3966 * reads on the source stripes.
3967 * The source stripes are determined by mapping the first and last
3968 * block on the destination stripes.
3971 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
3974 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
3975 * new_data_disks - 1),
3977 if (last_sector >= mddev->dev_sectors)
3978 last_sector = mddev->dev_sectors - 1;
3979 while (first_sector <= last_sector) {
3980 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
3981 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3982 set_bit(STRIPE_HANDLE, &sh->state);
3984 first_sector += STRIPE_SECTORS;
3986 /* Now that the sources are clearly marked, we can release
3987 * the destination stripes
3989 while (!list_empty(&stripes)) {
3990 sh = list_entry(stripes.next, struct stripe_head, lru);
3991 list_del_init(&sh->lru);
3994 /* If this takes us to the resync_max point where we have to pause,
3995 * then we need to write out the superblock.
3997 sector_nr += reshape_sectors;
3998 if ((sector_nr - mddev->curr_resync_completed) * 2
3999 >= mddev->resync_max - mddev->curr_resync_completed) {
4000 /* Cannot proceed until we've updated the superblock... */
4001 wait_event(conf->wait_for_overlap,
4002 atomic_read(&conf->reshape_stripes) == 0);
4003 mddev->reshape_position = conf->reshape_progress;
4004 mddev->curr_resync_completed = sector_nr;
4005 conf->reshape_checkpoint = jiffies;
4006 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4007 md_wakeup_thread(mddev->thread);
4008 wait_event(mddev->sb_wait,
4009 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4010 || kthread_should_stop());
4011 spin_lock_irq(&conf->device_lock);
4012 conf->reshape_safe = mddev->reshape_position;
4013 spin_unlock_irq(&conf->device_lock);
4014 wake_up(&conf->wait_for_overlap);
4015 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4017 return reshape_sectors;
4020 /* FIXME go_faster isn't used */
4021 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4023 raid5_conf_t *conf = mddev->private;
4024 struct stripe_head *sh;
4025 sector_t max_sector = mddev->dev_sectors;
4026 sector_t sync_blocks;
4027 int still_degraded = 0;
4030 if (sector_nr >= max_sector) {
4031 /* just being told to finish up .. nothing much to do */
4033 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4038 if (mddev->curr_resync < max_sector) /* aborted */
4039 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4041 else /* completed sync */
4043 bitmap_close_sync(mddev->bitmap);
4048 /* Allow raid5_quiesce to complete */
4049 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4051 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4052 return reshape_request(mddev, sector_nr, skipped);
4054 /* No need to check resync_max as we never do more than one
4055 * stripe, and as resync_max will always be on a chunk boundary,
4056 * if the check in md_do_sync didn't fire, there is no chance
4057 * of overstepping resync_max here
4060 /* if there is too many failed drives and we are trying
4061 * to resync, then assert that we are finished, because there is
4062 * nothing we can do.
4064 if (mddev->degraded >= conf->max_degraded &&
4065 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4066 sector_t rv = mddev->dev_sectors - sector_nr;
4070 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4071 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4072 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4073 /* we can skip this block, and probably more */
4074 sync_blocks /= STRIPE_SECTORS;
4076 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4080 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4082 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4084 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4085 /* make sure we don't swamp the stripe cache if someone else
4086 * is trying to get access
4088 schedule_timeout_uninterruptible(1);
4090 /* Need to check if array will still be degraded after recovery/resync
4091 * We don't need to check the 'failed' flag as when that gets set,
4094 for (i = 0; i < conf->raid_disks; i++)
4095 if (conf->disks[i].rdev == NULL)
4098 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4100 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4105 return STRIPE_SECTORS;
4108 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4110 /* We may not be able to submit a whole bio at once as there
4111 * may not be enough stripe_heads available.
4112 * We cannot pre-allocate enough stripe_heads as we may need
4113 * more than exist in the cache (if we allow ever large chunks).
4114 * So we do one stripe head at a time and record in
4115 * ->bi_hw_segments how many have been done.
4117 * We *know* that this entire raid_bio is in one chunk, so
4118 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4120 struct stripe_head *sh;
4122 sector_t sector, logical_sector, last_sector;
4127 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4128 sector = raid5_compute_sector(conf, logical_sector,
4130 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4132 for (; logical_sector < last_sector;
4133 logical_sector += STRIPE_SECTORS,
4134 sector += STRIPE_SECTORS,
4137 if (scnt < raid5_bi_hw_segments(raid_bio))
4138 /* already done this stripe */
4141 sh = get_active_stripe(conf, sector, 0, 1, 0);
4144 /* failed to get a stripe - must wait */
4145 raid5_set_bi_hw_segments(raid_bio, scnt);
4146 conf->retry_read_aligned = raid_bio;
4150 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4151 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4153 raid5_set_bi_hw_segments(raid_bio, scnt);
4154 conf->retry_read_aligned = raid_bio;
4162 spin_lock_irq(&conf->device_lock);
4163 remaining = raid5_dec_bi_phys_segments(raid_bio);
4164 spin_unlock_irq(&conf->device_lock);
4166 bio_endio(raid_bio, 0);
4167 if (atomic_dec_and_test(&conf->active_aligned_reads))
4168 wake_up(&conf->wait_for_stripe);
4174 * This is our raid5 kernel thread.
4176 * We scan the hash table for stripes which can be handled now.
4177 * During the scan, completed stripes are saved for us by the interrupt
4178 * handler, so that they will not have to wait for our next wakeup.
4180 static void raid5d(mddev_t *mddev)
4182 struct stripe_head *sh;
4183 raid5_conf_t *conf = mddev->private;
4185 struct blk_plug plug;
4187 pr_debug("+++ raid5d active\n");
4189 md_check_recovery(mddev);
4191 blk_start_plug(&plug);
4193 spin_lock_irq(&conf->device_lock);
4197 if (atomic_read(&mddev->plug_cnt) == 0 &&
4198 !list_empty(&conf->bitmap_list)) {
4199 /* Now is a good time to flush some bitmap updates */
4201 spin_unlock_irq(&conf->device_lock);
4202 bitmap_unplug(mddev->bitmap);
4203 spin_lock_irq(&conf->device_lock);
4204 conf->seq_write = conf->seq_flush;
4205 activate_bit_delay(conf);
4207 if (atomic_read(&mddev->plug_cnt) == 0)
4208 raid5_activate_delayed(conf);
4210 while ((bio = remove_bio_from_retry(conf))) {
4212 spin_unlock_irq(&conf->device_lock);
4213 ok = retry_aligned_read(conf, bio);
4214 spin_lock_irq(&conf->device_lock);
4220 sh = __get_priority_stripe(conf);
4224 spin_unlock_irq(&conf->device_lock);
4231 spin_lock_irq(&conf->device_lock);
4233 pr_debug("%d stripes handled\n", handled);
4235 spin_unlock_irq(&conf->device_lock);
4237 async_tx_issue_pending_all();
4238 blk_finish_plug(&plug);
4240 pr_debug("--- raid5d inactive\n");
4244 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4246 raid5_conf_t *conf = mddev->private;
4248 return sprintf(page, "%d\n", conf->max_nr_stripes);
4254 raid5_set_cache_size(mddev_t *mddev, int size)
4256 raid5_conf_t *conf = mddev->private;
4259 if (size <= 16 || size > 32768)
4261 while (size < conf->max_nr_stripes) {
4262 if (drop_one_stripe(conf))
4263 conf->max_nr_stripes--;
4267 err = md_allow_write(mddev);
4270 while (size > conf->max_nr_stripes) {
4271 if (grow_one_stripe(conf))
4272 conf->max_nr_stripes++;
4277 EXPORT_SYMBOL(raid5_set_cache_size);
4280 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4282 raid5_conf_t *conf = mddev->private;
4286 if (len >= PAGE_SIZE)
4291 if (strict_strtoul(page, 10, &new))
4293 err = raid5_set_cache_size(mddev, new);
4299 static struct md_sysfs_entry
4300 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4301 raid5_show_stripe_cache_size,
4302 raid5_store_stripe_cache_size);
4305 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4307 raid5_conf_t *conf = mddev->private;
4309 return sprintf(page, "%d\n", conf->bypass_threshold);
4315 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4317 raid5_conf_t *conf = mddev->private;
4319 if (len >= PAGE_SIZE)
4324 if (strict_strtoul(page, 10, &new))
4326 if (new > conf->max_nr_stripes)
4328 conf->bypass_threshold = new;
4332 static struct md_sysfs_entry
4333 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4335 raid5_show_preread_threshold,
4336 raid5_store_preread_threshold);
4339 stripe_cache_active_show(mddev_t *mddev, char *page)
4341 raid5_conf_t *conf = mddev->private;
4343 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4348 static struct md_sysfs_entry
4349 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4351 static struct attribute *raid5_attrs[] = {
4352 &raid5_stripecache_size.attr,
4353 &raid5_stripecache_active.attr,
4354 &raid5_preread_bypass_threshold.attr,
4357 static struct attribute_group raid5_attrs_group = {
4359 .attrs = raid5_attrs,
4363 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4365 raid5_conf_t *conf = mddev->private;
4368 sectors = mddev->dev_sectors;
4370 /* size is defined by the smallest of previous and new size */
4371 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4373 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4374 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4375 return sectors * (raid_disks - conf->max_degraded);
4378 static void raid5_free_percpu(raid5_conf_t *conf)
4380 struct raid5_percpu *percpu;
4387 for_each_possible_cpu(cpu) {
4388 percpu = per_cpu_ptr(conf->percpu, cpu);
4389 safe_put_page(percpu->spare_page);
4390 kfree(percpu->scribble);
4392 #ifdef CONFIG_HOTPLUG_CPU
4393 unregister_cpu_notifier(&conf->cpu_notify);
4397 free_percpu(conf->percpu);
4400 static void free_conf(raid5_conf_t *conf)
4402 shrink_stripes(conf);
4403 raid5_free_percpu(conf);
4405 kfree(conf->stripe_hashtbl);
4409 #ifdef CONFIG_HOTPLUG_CPU
4410 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4413 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4414 long cpu = (long)hcpu;
4415 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4418 case CPU_UP_PREPARE:
4419 case CPU_UP_PREPARE_FROZEN:
4420 if (conf->level == 6 && !percpu->spare_page)
4421 percpu->spare_page = alloc_page(GFP_KERNEL);
4422 if (!percpu->scribble)
4423 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4425 if (!percpu->scribble ||
4426 (conf->level == 6 && !percpu->spare_page)) {
4427 safe_put_page(percpu->spare_page);
4428 kfree(percpu->scribble);
4429 pr_err("%s: failed memory allocation for cpu%ld\n",
4431 return notifier_from_errno(-ENOMEM);
4435 case CPU_DEAD_FROZEN:
4436 safe_put_page(percpu->spare_page);
4437 kfree(percpu->scribble);
4438 percpu->spare_page = NULL;
4439 percpu->scribble = NULL;
4448 static int raid5_alloc_percpu(raid5_conf_t *conf)
4451 struct page *spare_page;
4452 struct raid5_percpu __percpu *allcpus;
4456 allcpus = alloc_percpu(struct raid5_percpu);
4459 conf->percpu = allcpus;
4463 for_each_present_cpu(cpu) {
4464 if (conf->level == 6) {
4465 spare_page = alloc_page(GFP_KERNEL);
4470 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4472 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4477 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4479 #ifdef CONFIG_HOTPLUG_CPU
4480 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4481 conf->cpu_notify.priority = 0;
4483 err = register_cpu_notifier(&conf->cpu_notify);
4490 static raid5_conf_t *setup_conf(mddev_t *mddev)
4493 int raid_disk, memory, max_disks;
4495 struct disk_info *disk;
4497 if (mddev->new_level != 5
4498 && mddev->new_level != 4
4499 && mddev->new_level != 6) {
4500 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4501 mdname(mddev), mddev->new_level);
4502 return ERR_PTR(-EIO);
4504 if ((mddev->new_level == 5
4505 && !algorithm_valid_raid5(mddev->new_layout)) ||
4506 (mddev->new_level == 6
4507 && !algorithm_valid_raid6(mddev->new_layout))) {
4508 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4509 mdname(mddev), mddev->new_layout);
4510 return ERR_PTR(-EIO);
4512 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4513 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4514 mdname(mddev), mddev->raid_disks);
4515 return ERR_PTR(-EINVAL);
4518 if (!mddev->new_chunk_sectors ||
4519 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4520 !is_power_of_2(mddev->new_chunk_sectors)) {
4521 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4522 mdname(mddev), mddev->new_chunk_sectors << 9);
4523 return ERR_PTR(-EINVAL);
4526 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4529 spin_lock_init(&conf->device_lock);
4530 init_waitqueue_head(&conf->wait_for_stripe);
4531 init_waitqueue_head(&conf->wait_for_overlap);
4532 INIT_LIST_HEAD(&conf->handle_list);
4533 INIT_LIST_HEAD(&conf->hold_list);
4534 INIT_LIST_HEAD(&conf->delayed_list);
4535 INIT_LIST_HEAD(&conf->bitmap_list);
4536 INIT_LIST_HEAD(&conf->inactive_list);
4537 atomic_set(&conf->active_stripes, 0);
4538 atomic_set(&conf->preread_active_stripes, 0);
4539 atomic_set(&conf->active_aligned_reads, 0);
4540 conf->bypass_threshold = BYPASS_THRESHOLD;
4542 conf->raid_disks = mddev->raid_disks;
4543 if (mddev->reshape_position == MaxSector)
4544 conf->previous_raid_disks = mddev->raid_disks;
4546 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4547 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4548 conf->scribble_len = scribble_len(max_disks);
4550 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4555 conf->mddev = mddev;
4557 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4560 conf->level = mddev->new_level;
4561 if (raid5_alloc_percpu(conf) != 0)
4564 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4566 list_for_each_entry(rdev, &mddev->disks, same_set) {
4567 raid_disk = rdev->raid_disk;
4568 if (raid_disk >= max_disks
4571 disk = conf->disks + raid_disk;
4575 if (test_bit(In_sync, &rdev->flags)) {
4576 char b[BDEVNAME_SIZE];
4577 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4579 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4580 } else if (rdev->saved_raid_disk != raid_disk)
4581 /* Cannot rely on bitmap to complete recovery */
4585 conf->chunk_sectors = mddev->new_chunk_sectors;
4586 conf->level = mddev->new_level;
4587 if (conf->level == 6)
4588 conf->max_degraded = 2;
4590 conf->max_degraded = 1;
4591 conf->algorithm = mddev->new_layout;
4592 conf->max_nr_stripes = NR_STRIPES;
4593 conf->reshape_progress = mddev->reshape_position;
4594 if (conf->reshape_progress != MaxSector) {
4595 conf->prev_chunk_sectors = mddev->chunk_sectors;
4596 conf->prev_algo = mddev->layout;
4599 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4600 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4601 if (grow_stripes(conf, conf->max_nr_stripes)) {
4603 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4604 mdname(mddev), memory);
4607 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4608 mdname(mddev), memory);
4610 conf->thread = md_register_thread(raid5d, mddev, NULL);
4611 if (!conf->thread) {
4613 "md/raid:%s: couldn't allocate thread.\n",
4623 return ERR_PTR(-EIO);
4625 return ERR_PTR(-ENOMEM);
4629 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4632 case ALGORITHM_PARITY_0:
4633 if (raid_disk < max_degraded)
4636 case ALGORITHM_PARITY_N:
4637 if (raid_disk >= raid_disks - max_degraded)
4640 case ALGORITHM_PARITY_0_6:
4641 if (raid_disk == 0 ||
4642 raid_disk == raid_disks - 1)
4645 case ALGORITHM_LEFT_ASYMMETRIC_6:
4646 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4647 case ALGORITHM_LEFT_SYMMETRIC_6:
4648 case ALGORITHM_RIGHT_SYMMETRIC_6:
4649 if (raid_disk == raid_disks - 1)
4655 static int run(mddev_t *mddev)
4658 int working_disks = 0;
4659 int dirty_parity_disks = 0;
4661 sector_t reshape_offset = 0;
4663 if (mddev->recovery_cp != MaxSector)
4664 printk(KERN_NOTICE "md/raid:%s: not clean"
4665 " -- starting background reconstruction\n",
4667 if (mddev->reshape_position != MaxSector) {
4668 /* Check that we can continue the reshape.
4669 * Currently only disks can change, it must
4670 * increase, and we must be past the point where
4671 * a stripe over-writes itself
4673 sector_t here_new, here_old;
4675 int max_degraded = (mddev->level == 6 ? 2 : 1);
4677 if (mddev->new_level != mddev->level) {
4678 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4679 "required - aborting.\n",
4683 old_disks = mddev->raid_disks - mddev->delta_disks;
4684 /* reshape_position must be on a new-stripe boundary, and one
4685 * further up in new geometry must map after here in old
4688 here_new = mddev->reshape_position;
4689 if (sector_div(here_new, mddev->new_chunk_sectors *
4690 (mddev->raid_disks - max_degraded))) {
4691 printk(KERN_ERR "md/raid:%s: reshape_position not "
4692 "on a stripe boundary\n", mdname(mddev));
4695 reshape_offset = here_new * mddev->new_chunk_sectors;
4696 /* here_new is the stripe we will write to */
4697 here_old = mddev->reshape_position;
4698 sector_div(here_old, mddev->chunk_sectors *
4699 (old_disks-max_degraded));
4700 /* here_old is the first stripe that we might need to read
4702 if (mddev->delta_disks == 0) {
4703 /* We cannot be sure it is safe to start an in-place
4704 * reshape. It is only safe if user-space if monitoring
4705 * and taking constant backups.
4706 * mdadm always starts a situation like this in
4707 * readonly mode so it can take control before
4708 * allowing any writes. So just check for that.
4710 if ((here_new * mddev->new_chunk_sectors !=
4711 here_old * mddev->chunk_sectors) ||
4713 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4714 " in read-only mode - aborting\n",
4718 } else if (mddev->delta_disks < 0
4719 ? (here_new * mddev->new_chunk_sectors <=
4720 here_old * mddev->chunk_sectors)
4721 : (here_new * mddev->new_chunk_sectors >=
4722 here_old * mddev->chunk_sectors)) {
4723 /* Reading from the same stripe as writing to - bad */
4724 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4725 "auto-recovery - aborting.\n",
4729 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4731 /* OK, we should be able to continue; */
4733 BUG_ON(mddev->level != mddev->new_level);
4734 BUG_ON(mddev->layout != mddev->new_layout);
4735 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4736 BUG_ON(mddev->delta_disks != 0);
4739 if (mddev->private == NULL)
4740 conf = setup_conf(mddev);
4742 conf = mddev->private;
4745 return PTR_ERR(conf);
4747 mddev->thread = conf->thread;
4748 conf->thread = NULL;
4749 mddev->private = conf;
4752 * 0 for a fully functional array, 1 or 2 for a degraded array.
4754 list_for_each_entry(rdev, &mddev->disks, same_set) {
4755 if (rdev->raid_disk < 0)
4757 if (test_bit(In_sync, &rdev->flags)) {
4761 /* This disc is not fully in-sync. However if it
4762 * just stored parity (beyond the recovery_offset),
4763 * when we don't need to be concerned about the
4764 * array being dirty.
4765 * When reshape goes 'backwards', we never have
4766 * partially completed devices, so we only need
4767 * to worry about reshape going forwards.
4769 /* Hack because v0.91 doesn't store recovery_offset properly. */
4770 if (mddev->major_version == 0 &&
4771 mddev->minor_version > 90)
4772 rdev->recovery_offset = reshape_offset;
4774 if (rdev->recovery_offset < reshape_offset) {
4775 /* We need to check old and new layout */
4776 if (!only_parity(rdev->raid_disk,
4779 conf->max_degraded))
4782 if (!only_parity(rdev->raid_disk,
4784 conf->previous_raid_disks,
4785 conf->max_degraded))
4787 dirty_parity_disks++;
4790 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4793 if (has_failed(conf)) {
4794 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4795 " (%d/%d failed)\n",
4796 mdname(mddev), mddev->degraded, conf->raid_disks);
4800 /* device size must be a multiple of chunk size */
4801 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4802 mddev->resync_max_sectors = mddev->dev_sectors;
4804 if (mddev->degraded > dirty_parity_disks &&
4805 mddev->recovery_cp != MaxSector) {
4806 if (mddev->ok_start_degraded)
4808 "md/raid:%s: starting dirty degraded array"
4809 " - data corruption possible.\n",
4813 "md/raid:%s: cannot start dirty degraded array.\n",
4819 if (mddev->degraded == 0)
4820 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4821 " devices, algorithm %d\n", mdname(mddev), conf->level,
4822 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4825 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4826 " out of %d devices, algorithm %d\n",
4827 mdname(mddev), conf->level,
4828 mddev->raid_disks - mddev->degraded,
4829 mddev->raid_disks, mddev->new_layout);
4831 print_raid5_conf(conf);
4833 if (conf->reshape_progress != MaxSector) {
4834 conf->reshape_safe = conf->reshape_progress;
4835 atomic_set(&conf->reshape_stripes, 0);
4836 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4837 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4838 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4839 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4840 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4845 /* Ok, everything is just fine now */
4846 if (mddev->to_remove == &raid5_attrs_group)
4847 mddev->to_remove = NULL;
4848 else if (mddev->kobj.sd &&
4849 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4851 "raid5: failed to create sysfs attributes for %s\n",
4853 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4857 /* read-ahead size must cover two whole stripes, which
4858 * is 2 * (datadisks) * chunksize where 'n' is the
4859 * number of raid devices
4861 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4862 int stripe = data_disks *
4863 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4864 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4865 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4867 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4869 mddev->queue->backing_dev_info.congested_data = mddev;
4870 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4872 chunk_size = mddev->chunk_sectors << 9;
4873 blk_queue_io_min(mddev->queue, chunk_size);
4874 blk_queue_io_opt(mddev->queue, chunk_size *
4875 (conf->raid_disks - conf->max_degraded));
4877 list_for_each_entry(rdev, &mddev->disks, same_set)
4878 disk_stack_limits(mddev->gendisk, rdev->bdev,
4879 rdev->data_offset << 9);
4884 md_unregister_thread(mddev->thread);
4885 mddev->thread = NULL;
4887 print_raid5_conf(conf);
4890 mddev->private = NULL;
4891 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
4895 static int stop(mddev_t *mddev)
4897 raid5_conf_t *conf = mddev->private;
4899 md_unregister_thread(mddev->thread);
4900 mddev->thread = NULL;
4902 mddev->queue->backing_dev_info.congested_fn = NULL;
4904 mddev->private = NULL;
4905 mddev->to_remove = &raid5_attrs_group;
4910 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4914 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4915 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4916 seq_printf(seq, "sh %llu, count %d.\n",
4917 (unsigned long long)sh->sector, atomic_read(&sh->count));
4918 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4919 for (i = 0; i < sh->disks; i++) {
4920 seq_printf(seq, "(cache%d: %p %ld) ",
4921 i, sh->dev[i].page, sh->dev[i].flags);
4923 seq_printf(seq, "\n");
4926 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4928 struct stripe_head *sh;
4929 struct hlist_node *hn;
4932 spin_lock_irq(&conf->device_lock);
4933 for (i = 0; i < NR_HASH; i++) {
4934 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4935 if (sh->raid_conf != conf)
4940 spin_unlock_irq(&conf->device_lock);
4944 static void status(struct seq_file *seq, mddev_t *mddev)
4946 raid5_conf_t *conf = mddev->private;
4949 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4950 mddev->chunk_sectors / 2, mddev->layout);
4951 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4952 for (i = 0; i < conf->raid_disks; i++)
4953 seq_printf (seq, "%s",
4954 conf->disks[i].rdev &&
4955 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4956 seq_printf (seq, "]");
4958 seq_printf (seq, "\n");
4959 printall(seq, conf);
4963 static void print_raid5_conf (raid5_conf_t *conf)
4966 struct disk_info *tmp;
4968 printk(KERN_DEBUG "RAID conf printout:\n");
4970 printk("(conf==NULL)\n");
4973 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
4975 conf->raid_disks - conf->mddev->degraded);
4977 for (i = 0; i < conf->raid_disks; i++) {
4978 char b[BDEVNAME_SIZE];
4979 tmp = conf->disks + i;
4981 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
4982 i, !test_bit(Faulty, &tmp->rdev->flags),
4983 bdevname(tmp->rdev->bdev, b));
4987 static int raid5_spare_active(mddev_t *mddev)
4990 raid5_conf_t *conf = mddev->private;
4991 struct disk_info *tmp;
4993 unsigned long flags;
4995 for (i = 0; i < conf->raid_disks; i++) {
4996 tmp = conf->disks + i;
4998 && tmp->rdev->recovery_offset == MaxSector
4999 && !test_bit(Faulty, &tmp->rdev->flags)
5000 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5002 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5005 spin_lock_irqsave(&conf->device_lock, flags);
5006 mddev->degraded -= count;
5007 spin_unlock_irqrestore(&conf->device_lock, flags);
5008 print_raid5_conf(conf);
5012 static int raid5_remove_disk(mddev_t *mddev, int number)
5014 raid5_conf_t *conf = mddev->private;
5017 struct disk_info *p = conf->disks + number;
5019 print_raid5_conf(conf);
5022 if (number >= conf->raid_disks &&
5023 conf->reshape_progress == MaxSector)
5024 clear_bit(In_sync, &rdev->flags);
5026 if (test_bit(In_sync, &rdev->flags) ||
5027 atomic_read(&rdev->nr_pending)) {
5031 /* Only remove non-faulty devices if recovery
5034 if (!test_bit(Faulty, &rdev->flags) &&
5035 !has_failed(conf) &&
5036 number < conf->raid_disks) {
5042 if (atomic_read(&rdev->nr_pending)) {
5043 /* lost the race, try later */
5050 print_raid5_conf(conf);
5054 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5056 raid5_conf_t *conf = mddev->private;
5059 struct disk_info *p;
5061 int last = conf->raid_disks - 1;
5063 if (has_failed(conf))
5064 /* no point adding a device */
5067 if (rdev->raid_disk >= 0)
5068 first = last = rdev->raid_disk;
5071 * find the disk ... but prefer rdev->saved_raid_disk
5074 if (rdev->saved_raid_disk >= 0 &&
5075 rdev->saved_raid_disk >= first &&
5076 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5077 disk = rdev->saved_raid_disk;
5080 for ( ; disk <= last ; disk++)
5081 if ((p=conf->disks + disk)->rdev == NULL) {
5082 clear_bit(In_sync, &rdev->flags);
5083 rdev->raid_disk = disk;
5085 if (rdev->saved_raid_disk != disk)
5087 rcu_assign_pointer(p->rdev, rdev);
5090 print_raid5_conf(conf);
5094 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5096 /* no resync is happening, and there is enough space
5097 * on all devices, so we can resize.
5098 * We need to make sure resync covers any new space.
5099 * If the array is shrinking we should possibly wait until
5100 * any io in the removed space completes, but it hardly seems
5103 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5104 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5105 mddev->raid_disks));
5106 if (mddev->array_sectors >
5107 raid5_size(mddev, sectors, mddev->raid_disks))
5109 set_capacity(mddev->gendisk, mddev->array_sectors);
5110 revalidate_disk(mddev->gendisk);
5111 if (sectors > mddev->dev_sectors &&
5112 mddev->recovery_cp > mddev->dev_sectors) {
5113 mddev->recovery_cp = mddev->dev_sectors;
5114 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5116 mddev->dev_sectors = sectors;
5117 mddev->resync_max_sectors = sectors;
5121 static int check_stripe_cache(mddev_t *mddev)
5123 /* Can only proceed if there are plenty of stripe_heads.
5124 * We need a minimum of one full stripe,, and for sensible progress
5125 * it is best to have about 4 times that.
5126 * If we require 4 times, then the default 256 4K stripe_heads will
5127 * allow for chunk sizes up to 256K, which is probably OK.
5128 * If the chunk size is greater, user-space should request more
5129 * stripe_heads first.
5131 raid5_conf_t *conf = mddev->private;
5132 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5133 > conf->max_nr_stripes ||
5134 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5135 > conf->max_nr_stripes) {
5136 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5138 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5145 static int check_reshape(mddev_t *mddev)
5147 raid5_conf_t *conf = mddev->private;
5149 if (mddev->delta_disks == 0 &&
5150 mddev->new_layout == mddev->layout &&
5151 mddev->new_chunk_sectors == mddev->chunk_sectors)
5152 return 0; /* nothing to do */
5154 /* Cannot grow a bitmap yet */
5156 if (has_failed(conf))
5158 if (mddev->delta_disks < 0) {
5159 /* We might be able to shrink, but the devices must
5160 * be made bigger first.
5161 * For raid6, 4 is the minimum size.
5162 * Otherwise 2 is the minimum
5165 if (mddev->level == 6)
5167 if (mddev->raid_disks + mddev->delta_disks < min)
5171 if (!check_stripe_cache(mddev))
5174 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5177 static int raid5_start_reshape(mddev_t *mddev)
5179 raid5_conf_t *conf = mddev->private;
5182 unsigned long flags;
5184 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5187 if (!check_stripe_cache(mddev))
5190 list_for_each_entry(rdev, &mddev->disks, same_set)
5191 if (!test_bit(In_sync, &rdev->flags)
5192 && !test_bit(Faulty, &rdev->flags))
5195 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5196 /* Not enough devices even to make a degraded array
5201 /* Refuse to reduce size of the array. Any reductions in
5202 * array size must be through explicit setting of array_size
5205 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5206 < mddev->array_sectors) {
5207 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5208 "before number of disks\n", mdname(mddev));
5212 atomic_set(&conf->reshape_stripes, 0);
5213 spin_lock_irq(&conf->device_lock);
5214 conf->previous_raid_disks = conf->raid_disks;
5215 conf->raid_disks += mddev->delta_disks;
5216 conf->prev_chunk_sectors = conf->chunk_sectors;
5217 conf->chunk_sectors = mddev->new_chunk_sectors;
5218 conf->prev_algo = conf->algorithm;
5219 conf->algorithm = mddev->new_layout;
5220 if (mddev->delta_disks < 0)
5221 conf->reshape_progress = raid5_size(mddev, 0, 0);
5223 conf->reshape_progress = 0;
5224 conf->reshape_safe = conf->reshape_progress;
5226 spin_unlock_irq(&conf->device_lock);
5228 /* Add some new drives, as many as will fit.
5229 * We know there are enough to make the newly sized array work.
5230 * Don't add devices if we are reducing the number of
5231 * devices in the array. This is because it is not possible
5232 * to correctly record the "partially reconstructed" state of
5233 * such devices during the reshape and confusion could result.
5235 if (mddev->delta_disks >= 0) {
5236 int added_devices = 0;
5237 list_for_each_entry(rdev, &mddev->disks, same_set)
5238 if (rdev->raid_disk < 0 &&
5239 !test_bit(Faulty, &rdev->flags)) {
5240 if (raid5_add_disk(mddev, rdev) == 0) {
5243 >= conf->previous_raid_disks) {
5244 set_bit(In_sync, &rdev->flags);
5247 rdev->recovery_offset = 0;
5248 sprintf(nm, "rd%d", rdev->raid_disk);
5249 if (sysfs_create_link(&mddev->kobj,
5251 /* Failure here is OK */;
5253 } else if (rdev->raid_disk >= conf->previous_raid_disks
5254 && !test_bit(Faulty, &rdev->flags)) {
5255 /* This is a spare that was manually added */
5256 set_bit(In_sync, &rdev->flags);
5260 /* When a reshape changes the number of devices,
5261 * ->degraded is measured against the larger of the
5262 * pre and post number of devices.
5264 spin_lock_irqsave(&conf->device_lock, flags);
5265 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5267 spin_unlock_irqrestore(&conf->device_lock, flags);
5269 mddev->raid_disks = conf->raid_disks;
5270 mddev->reshape_position = conf->reshape_progress;
5271 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5273 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5274 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5275 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5276 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5277 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5279 if (!mddev->sync_thread) {
5280 mddev->recovery = 0;
5281 spin_lock_irq(&conf->device_lock);
5282 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5283 conf->reshape_progress = MaxSector;
5284 spin_unlock_irq(&conf->device_lock);
5287 conf->reshape_checkpoint = jiffies;
5288 md_wakeup_thread(mddev->sync_thread);
5289 md_new_event(mddev);
5293 /* This is called from the reshape thread and should make any
5294 * changes needed in 'conf'
5296 static void end_reshape(raid5_conf_t *conf)
5299 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5301 spin_lock_irq(&conf->device_lock);
5302 conf->previous_raid_disks = conf->raid_disks;
5303 conf->reshape_progress = MaxSector;
5304 spin_unlock_irq(&conf->device_lock);
5305 wake_up(&conf->wait_for_overlap);
5307 /* read-ahead size must cover two whole stripes, which is
5308 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5310 if (conf->mddev->queue) {
5311 int data_disks = conf->raid_disks - conf->max_degraded;
5312 int stripe = data_disks * ((conf->chunk_sectors << 9)
5314 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5315 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5320 /* This is called from the raid5d thread with mddev_lock held.
5321 * It makes config changes to the device.
5323 static void raid5_finish_reshape(mddev_t *mddev)
5325 raid5_conf_t *conf = mddev->private;
5327 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5329 if (mddev->delta_disks > 0) {
5330 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5331 set_capacity(mddev->gendisk, mddev->array_sectors);
5332 revalidate_disk(mddev->gendisk);
5335 mddev->degraded = conf->raid_disks;
5336 for (d = 0; d < conf->raid_disks ; d++)
5337 if (conf->disks[d].rdev &&
5339 &conf->disks[d].rdev->flags))
5341 for (d = conf->raid_disks ;
5342 d < conf->raid_disks - mddev->delta_disks;
5344 mdk_rdev_t *rdev = conf->disks[d].rdev;
5345 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5347 sprintf(nm, "rd%d", rdev->raid_disk);
5348 sysfs_remove_link(&mddev->kobj, nm);
5349 rdev->raid_disk = -1;
5353 mddev->layout = conf->algorithm;
5354 mddev->chunk_sectors = conf->chunk_sectors;
5355 mddev->reshape_position = MaxSector;
5356 mddev->delta_disks = 0;
5360 static void raid5_quiesce(mddev_t *mddev, int state)
5362 raid5_conf_t *conf = mddev->private;
5365 case 2: /* resume for a suspend */
5366 wake_up(&conf->wait_for_overlap);
5369 case 1: /* stop all writes */
5370 spin_lock_irq(&conf->device_lock);
5371 /* '2' tells resync/reshape to pause so that all
5372 * active stripes can drain
5375 wait_event_lock_irq(conf->wait_for_stripe,
5376 atomic_read(&conf->active_stripes) == 0 &&
5377 atomic_read(&conf->active_aligned_reads) == 0,
5378 conf->device_lock, /* nothing */);
5380 spin_unlock_irq(&conf->device_lock);
5381 /* allow reshape to continue */
5382 wake_up(&conf->wait_for_overlap);
5385 case 0: /* re-enable writes */
5386 spin_lock_irq(&conf->device_lock);
5388 wake_up(&conf->wait_for_stripe);
5389 wake_up(&conf->wait_for_overlap);
5390 spin_unlock_irq(&conf->device_lock);
5396 static void *raid45_takeover_raid0(mddev_t *mddev, int level)
5398 struct raid0_private_data *raid0_priv = mddev->private;
5401 /* for raid0 takeover only one zone is supported */
5402 if (raid0_priv->nr_strip_zones > 1) {
5403 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5405 return ERR_PTR(-EINVAL);
5408 sectors = raid0_priv->strip_zone[0].zone_end;
5409 sector_div(sectors, raid0_priv->strip_zone[0].nb_dev);
5410 mddev->dev_sectors = sectors;
5411 mddev->new_level = level;
5412 mddev->new_layout = ALGORITHM_PARITY_N;
5413 mddev->new_chunk_sectors = mddev->chunk_sectors;
5414 mddev->raid_disks += 1;
5415 mddev->delta_disks = 1;
5416 /* make sure it will be not marked as dirty */
5417 mddev->recovery_cp = MaxSector;
5419 return setup_conf(mddev);
5423 static void *raid5_takeover_raid1(mddev_t *mddev)
5427 if (mddev->raid_disks != 2 ||
5428 mddev->degraded > 1)
5429 return ERR_PTR(-EINVAL);
5431 /* Should check if there are write-behind devices? */
5433 chunksect = 64*2; /* 64K by default */
5435 /* The array must be an exact multiple of chunksize */
5436 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5439 if ((chunksect<<9) < STRIPE_SIZE)
5440 /* array size does not allow a suitable chunk size */
5441 return ERR_PTR(-EINVAL);
5443 mddev->new_level = 5;
5444 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5445 mddev->new_chunk_sectors = chunksect;
5447 return setup_conf(mddev);
5450 static void *raid5_takeover_raid6(mddev_t *mddev)
5454 switch (mddev->layout) {
5455 case ALGORITHM_LEFT_ASYMMETRIC_6:
5456 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5458 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5459 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5461 case ALGORITHM_LEFT_SYMMETRIC_6:
5462 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5464 case ALGORITHM_RIGHT_SYMMETRIC_6:
5465 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5467 case ALGORITHM_PARITY_0_6:
5468 new_layout = ALGORITHM_PARITY_0;
5470 case ALGORITHM_PARITY_N:
5471 new_layout = ALGORITHM_PARITY_N;
5474 return ERR_PTR(-EINVAL);
5476 mddev->new_level = 5;
5477 mddev->new_layout = new_layout;
5478 mddev->delta_disks = -1;
5479 mddev->raid_disks -= 1;
5480 return setup_conf(mddev);
5484 static int raid5_check_reshape(mddev_t *mddev)
5486 /* For a 2-drive array, the layout and chunk size can be changed
5487 * immediately as not restriping is needed.
5488 * For larger arrays we record the new value - after validation
5489 * to be used by a reshape pass.
5491 raid5_conf_t *conf = mddev->private;
5492 int new_chunk = mddev->new_chunk_sectors;
5494 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5496 if (new_chunk > 0) {
5497 if (!is_power_of_2(new_chunk))
5499 if (new_chunk < (PAGE_SIZE>>9))
5501 if (mddev->array_sectors & (new_chunk-1))
5502 /* not factor of array size */
5506 /* They look valid */
5508 if (mddev->raid_disks == 2) {
5509 /* can make the change immediately */
5510 if (mddev->new_layout >= 0) {
5511 conf->algorithm = mddev->new_layout;
5512 mddev->layout = mddev->new_layout;
5514 if (new_chunk > 0) {
5515 conf->chunk_sectors = new_chunk ;
5516 mddev->chunk_sectors = new_chunk;
5518 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5519 md_wakeup_thread(mddev->thread);
5521 return check_reshape(mddev);
5524 static int raid6_check_reshape(mddev_t *mddev)
5526 int new_chunk = mddev->new_chunk_sectors;
5528 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5530 if (new_chunk > 0) {
5531 if (!is_power_of_2(new_chunk))
5533 if (new_chunk < (PAGE_SIZE >> 9))
5535 if (mddev->array_sectors & (new_chunk-1))
5536 /* not factor of array size */
5540 /* They look valid */
5541 return check_reshape(mddev);
5544 static void *raid5_takeover(mddev_t *mddev)
5546 /* raid5 can take over:
5547 * raid0 - if there is only one strip zone - make it a raid4 layout
5548 * raid1 - if there are two drives. We need to know the chunk size
5549 * raid4 - trivial - just use a raid4 layout.
5550 * raid6 - Providing it is a *_6 layout
5552 if (mddev->level == 0)
5553 return raid45_takeover_raid0(mddev, 5);
5554 if (mddev->level == 1)
5555 return raid5_takeover_raid1(mddev);
5556 if (mddev->level == 4) {
5557 mddev->new_layout = ALGORITHM_PARITY_N;
5558 mddev->new_level = 5;
5559 return setup_conf(mddev);
5561 if (mddev->level == 6)
5562 return raid5_takeover_raid6(mddev);
5564 return ERR_PTR(-EINVAL);
5567 static void *raid4_takeover(mddev_t *mddev)
5569 /* raid4 can take over:
5570 * raid0 - if there is only one strip zone
5571 * raid5 - if layout is right
5573 if (mddev->level == 0)
5574 return raid45_takeover_raid0(mddev, 4);
5575 if (mddev->level == 5 &&
5576 mddev->layout == ALGORITHM_PARITY_N) {
5577 mddev->new_layout = 0;
5578 mddev->new_level = 4;
5579 return setup_conf(mddev);
5581 return ERR_PTR(-EINVAL);
5584 static struct mdk_personality raid5_personality;
5586 static void *raid6_takeover(mddev_t *mddev)
5588 /* Currently can only take over a raid5. We map the
5589 * personality to an equivalent raid6 personality
5590 * with the Q block at the end.
5594 if (mddev->pers != &raid5_personality)
5595 return ERR_PTR(-EINVAL);
5596 if (mddev->degraded > 1)
5597 return ERR_PTR(-EINVAL);
5598 if (mddev->raid_disks > 253)
5599 return ERR_PTR(-EINVAL);
5600 if (mddev->raid_disks < 3)
5601 return ERR_PTR(-EINVAL);
5603 switch (mddev->layout) {
5604 case ALGORITHM_LEFT_ASYMMETRIC:
5605 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5607 case ALGORITHM_RIGHT_ASYMMETRIC:
5608 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5610 case ALGORITHM_LEFT_SYMMETRIC:
5611 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5613 case ALGORITHM_RIGHT_SYMMETRIC:
5614 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5616 case ALGORITHM_PARITY_0:
5617 new_layout = ALGORITHM_PARITY_0_6;
5619 case ALGORITHM_PARITY_N:
5620 new_layout = ALGORITHM_PARITY_N;
5623 return ERR_PTR(-EINVAL);
5625 mddev->new_level = 6;
5626 mddev->new_layout = new_layout;
5627 mddev->delta_disks = 1;
5628 mddev->raid_disks += 1;
5629 return setup_conf(mddev);
5633 static struct mdk_personality raid6_personality =
5637 .owner = THIS_MODULE,
5638 .make_request = make_request,
5642 .error_handler = error,
5643 .hot_add_disk = raid5_add_disk,
5644 .hot_remove_disk= raid5_remove_disk,
5645 .spare_active = raid5_spare_active,
5646 .sync_request = sync_request,
5647 .resize = raid5_resize,
5649 .check_reshape = raid6_check_reshape,
5650 .start_reshape = raid5_start_reshape,
5651 .finish_reshape = raid5_finish_reshape,
5652 .quiesce = raid5_quiesce,
5653 .takeover = raid6_takeover,
5655 static struct mdk_personality raid5_personality =
5659 .owner = THIS_MODULE,
5660 .make_request = make_request,
5664 .error_handler = error,
5665 .hot_add_disk = raid5_add_disk,
5666 .hot_remove_disk= raid5_remove_disk,
5667 .spare_active = raid5_spare_active,
5668 .sync_request = sync_request,
5669 .resize = raid5_resize,
5671 .check_reshape = raid5_check_reshape,
5672 .start_reshape = raid5_start_reshape,
5673 .finish_reshape = raid5_finish_reshape,
5674 .quiesce = raid5_quiesce,
5675 .takeover = raid5_takeover,
5678 static struct mdk_personality raid4_personality =
5682 .owner = THIS_MODULE,
5683 .make_request = make_request,
5687 .error_handler = error,
5688 .hot_add_disk = raid5_add_disk,
5689 .hot_remove_disk= raid5_remove_disk,
5690 .spare_active = raid5_spare_active,
5691 .sync_request = sync_request,
5692 .resize = raid5_resize,
5694 .check_reshape = raid5_check_reshape,
5695 .start_reshape = raid5_start_reshape,
5696 .finish_reshape = raid5_finish_reshape,
5697 .quiesce = raid5_quiesce,
5698 .takeover = raid4_takeover,
5701 static int __init raid5_init(void)
5703 register_md_personality(&raid6_personality);
5704 register_md_personality(&raid5_personality);
5705 register_md_personality(&raid4_personality);
5709 static void raid5_exit(void)
5711 unregister_md_personality(&raid6_personality);
5712 unregister_md_personality(&raid5_personality);
5713 unregister_md_personality(&raid4_personality);
5716 module_init(raid5_init);
5717 module_exit(raid5_exit);
5718 MODULE_LICENSE("GPL");
5719 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5720 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5721 MODULE_ALIAS("md-raid5");
5722 MODULE_ALIAS("md-raid4");
5723 MODULE_ALIAS("md-level-5");
5724 MODULE_ALIAS("md-level-4");
5725 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5726 MODULE_ALIAS("md-raid6");
5727 MODULE_ALIAS("md-level-6");
5729 /* This used to be two separate modules, they were: */
5730 MODULE_ALIAS("raid5");
5731 MODULE_ALIAS("raid6");