2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/async.h>
51 #include <linux/seq_file.h>
52 #include <linux/cpu.h>
53 #include <linux/slab.h>
63 #define NR_STRIPES 256
64 #define STRIPE_SIZE PAGE_SIZE
65 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
66 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
67 #define IO_THRESHOLD 1
68 #define BYPASS_THRESHOLD 1
69 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
70 #define HASH_MASK (NR_HASH - 1)
72 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
74 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
75 * order without overlap. There may be several bio's per stripe+device, and
76 * a bio could span several devices.
77 * When walking this list for a particular stripe+device, we must never proceed
78 * beyond a bio that extends past this device, as the next bio might no longer
80 * This macro is used to determine the 'next' bio in the list, given the sector
81 * of the current stripe+device
83 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
85 * The following can be used to debug the driver
87 #define RAID5_PARANOIA 1
88 #if RAID5_PARANOIA && defined(CONFIG_SMP)
89 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
91 # define CHECK_DEVLOCK()
99 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
102 * We maintain a biased count of active stripes in the bottom 16 bits of
103 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
105 static inline int raid5_bi_phys_segments(struct bio *bio)
107 return bio->bi_phys_segments & 0xffff;
110 static inline int raid5_bi_hw_segments(struct bio *bio)
112 return (bio->bi_phys_segments >> 16) & 0xffff;
115 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
117 --bio->bi_phys_segments;
118 return raid5_bi_phys_segments(bio);
121 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
123 unsigned short val = raid5_bi_hw_segments(bio);
126 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
130 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
132 bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
135 /* Find first data disk in a raid6 stripe */
136 static inline int raid6_d0(struct stripe_head *sh)
139 /* ddf always start from first device */
141 /* md starts just after Q block */
142 if (sh->qd_idx == sh->disks - 1)
145 return sh->qd_idx + 1;
147 static inline int raid6_next_disk(int disk, int raid_disks)
150 return (disk < raid_disks) ? disk : 0;
153 /* When walking through the disks in a raid5, starting at raid6_d0,
154 * We need to map each disk to a 'slot', where the data disks are slot
155 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
156 * is raid_disks-1. This help does that mapping.
158 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
159 int *count, int syndrome_disks)
165 if (idx == sh->pd_idx)
166 return syndrome_disks;
167 if (idx == sh->qd_idx)
168 return syndrome_disks + 1;
174 static void return_io(struct bio *return_bi)
176 struct bio *bi = return_bi;
179 return_bi = bi->bi_next;
187 static void print_raid5_conf (raid5_conf_t *conf);
189 static int stripe_operations_active(struct stripe_head *sh)
191 return sh->check_state || sh->reconstruct_state ||
192 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
193 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
196 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
198 if (atomic_dec_and_test(&sh->count)) {
199 BUG_ON(!list_empty(&sh->lru));
200 BUG_ON(atomic_read(&conf->active_stripes)==0);
201 if (test_bit(STRIPE_HANDLE, &sh->state)) {
202 if (test_bit(STRIPE_DELAYED, &sh->state))
203 list_add_tail(&sh->lru, &conf->delayed_list);
204 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
205 sh->bm_seq - conf->seq_write > 0)
206 list_add_tail(&sh->lru, &conf->bitmap_list);
208 clear_bit(STRIPE_BIT_DELAY, &sh->state);
209 list_add_tail(&sh->lru, &conf->handle_list);
211 md_wakeup_thread(conf->mddev->thread);
213 BUG_ON(stripe_operations_active(sh));
214 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
215 atomic_dec(&conf->preread_active_stripes);
216 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
217 md_wakeup_thread(conf->mddev->thread);
219 atomic_dec(&conf->active_stripes);
220 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
221 list_add_tail(&sh->lru, &conf->inactive_list);
222 wake_up(&conf->wait_for_stripe);
223 if (conf->retry_read_aligned)
224 md_wakeup_thread(conf->mddev->thread);
230 static void release_stripe(struct stripe_head *sh)
232 raid5_conf_t *conf = sh->raid_conf;
235 spin_lock_irqsave(&conf->device_lock, flags);
236 __release_stripe(conf, sh);
237 spin_unlock_irqrestore(&conf->device_lock, flags);
240 static inline void remove_hash(struct stripe_head *sh)
242 pr_debug("remove_hash(), stripe %llu\n",
243 (unsigned long long)sh->sector);
245 hlist_del_init(&sh->hash);
248 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
250 struct hlist_head *hp = stripe_hash(conf, sh->sector);
252 pr_debug("insert_hash(), stripe %llu\n",
253 (unsigned long long)sh->sector);
256 hlist_add_head(&sh->hash, hp);
260 /* find an idle stripe, make sure it is unhashed, and return it. */
261 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
263 struct stripe_head *sh = NULL;
264 struct list_head *first;
267 if (list_empty(&conf->inactive_list))
269 first = conf->inactive_list.next;
270 sh = list_entry(first, struct stripe_head, lru);
271 list_del_init(first);
273 atomic_inc(&conf->active_stripes);
278 static void shrink_buffers(struct stripe_head *sh)
282 int num = sh->raid_conf->pool_size;
284 for (i = 0; i < num ; i++) {
288 sh->dev[i].page = NULL;
293 static int grow_buffers(struct stripe_head *sh)
296 int num = sh->raid_conf->pool_size;
298 for (i = 0; i < num; i++) {
301 if (!(page = alloc_page(GFP_KERNEL))) {
304 sh->dev[i].page = page;
309 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
310 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
311 struct stripe_head *sh);
313 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
315 raid5_conf_t *conf = sh->raid_conf;
318 BUG_ON(atomic_read(&sh->count) != 0);
319 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
320 BUG_ON(stripe_operations_active(sh));
323 pr_debug("init_stripe called, stripe %llu\n",
324 (unsigned long long)sh->sector);
328 sh->generation = conf->generation - previous;
329 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
331 stripe_set_idx(sector, conf, previous, sh);
335 for (i = sh->disks; i--; ) {
336 struct r5dev *dev = &sh->dev[i];
338 if (dev->toread || dev->read || dev->towrite || dev->written ||
339 test_bit(R5_LOCKED, &dev->flags)) {
340 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
341 (unsigned long long)sh->sector, i, dev->toread,
342 dev->read, dev->towrite, dev->written,
343 test_bit(R5_LOCKED, &dev->flags));
347 raid5_build_block(sh, i, previous);
349 insert_hash(conf, sh);
352 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
355 struct stripe_head *sh;
356 struct hlist_node *hn;
359 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
360 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
361 if (sh->sector == sector && sh->generation == generation)
363 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
368 * Need to check if array has failed when deciding whether to:
370 * - remove non-faulty devices
373 * This determination is simple when no reshape is happening.
374 * However if there is a reshape, we need to carefully check
375 * both the before and after sections.
376 * This is because some failed devices may only affect one
377 * of the two sections, and some non-in_sync devices may
378 * be insync in the section most affected by failed devices.
380 static int has_failed(raid5_conf_t *conf)
384 if (conf->mddev->reshape_position == MaxSector)
385 return conf->mddev->degraded > conf->max_degraded;
389 for (i = 0; i < conf->previous_raid_disks; i++) {
390 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
391 if (!rdev || test_bit(Faulty, &rdev->flags))
393 else if (test_bit(In_sync, &rdev->flags))
396 /* not in-sync or faulty.
397 * If the reshape increases the number of devices,
398 * this is being recovered by the reshape, so
399 * this 'previous' section is not in_sync.
400 * If the number of devices is being reduced however,
401 * the device can only be part of the array if
402 * we are reverting a reshape, so this section will
405 if (conf->raid_disks >= conf->previous_raid_disks)
409 if (degraded > conf->max_degraded)
413 for (i = 0; i < conf->raid_disks; i++) {
414 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
415 if (!rdev || test_bit(Faulty, &rdev->flags))
417 else if (test_bit(In_sync, &rdev->flags))
420 /* not in-sync or faulty.
421 * If reshape increases the number of devices, this
422 * section has already been recovered, else it
423 * almost certainly hasn't.
425 if (conf->raid_disks <= conf->previous_raid_disks)
429 if (degraded > conf->max_degraded)
434 static struct stripe_head *
435 get_active_stripe(raid5_conf_t *conf, sector_t sector,
436 int previous, int noblock, int noquiesce)
438 struct stripe_head *sh;
440 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
442 spin_lock_irq(&conf->device_lock);
445 wait_event_lock_irq(conf->wait_for_stripe,
446 conf->quiesce == 0 || noquiesce,
447 conf->device_lock, /* nothing */);
448 sh = __find_stripe(conf, sector, conf->generation - previous);
450 if (!conf->inactive_blocked)
451 sh = get_free_stripe(conf);
452 if (noblock && sh == NULL)
455 conf->inactive_blocked = 1;
456 wait_event_lock_irq(conf->wait_for_stripe,
457 !list_empty(&conf->inactive_list) &&
458 (atomic_read(&conf->active_stripes)
459 < (conf->max_nr_stripes *3/4)
460 || !conf->inactive_blocked),
463 conf->inactive_blocked = 0;
465 init_stripe(sh, sector, previous);
467 if (atomic_read(&sh->count)) {
468 BUG_ON(!list_empty(&sh->lru)
469 && !test_bit(STRIPE_EXPANDING, &sh->state));
471 if (!test_bit(STRIPE_HANDLE, &sh->state))
472 atomic_inc(&conf->active_stripes);
473 if (list_empty(&sh->lru) &&
474 !test_bit(STRIPE_EXPANDING, &sh->state))
476 list_del_init(&sh->lru);
479 } while (sh == NULL);
482 atomic_inc(&sh->count);
484 spin_unlock_irq(&conf->device_lock);
489 raid5_end_read_request(struct bio *bi, int error);
491 raid5_end_write_request(struct bio *bi, int error);
493 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
495 raid5_conf_t *conf = sh->raid_conf;
496 int i, disks = sh->disks;
500 for (i = disks; i--; ) {
504 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
505 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
509 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
514 bi = &sh->dev[i].req;
518 bi->bi_end_io = raid5_end_write_request;
520 bi->bi_end_io = raid5_end_read_request;
523 rdev = rcu_dereference(conf->disks[i].rdev);
524 if (rdev && test_bit(Faulty, &rdev->flags))
527 atomic_inc(&rdev->nr_pending);
531 if (s->syncing || s->expanding || s->expanded)
532 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
534 set_bit(STRIPE_IO_STARTED, &sh->state);
536 bi->bi_bdev = rdev->bdev;
537 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
538 __func__, (unsigned long long)sh->sector,
540 atomic_inc(&sh->count);
541 bi->bi_sector = sh->sector + rdev->data_offset;
542 bi->bi_flags = 1 << BIO_UPTODATE;
546 bi->bi_io_vec = &sh->dev[i].vec;
547 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
548 bi->bi_io_vec[0].bv_offset = 0;
549 bi->bi_size = STRIPE_SIZE;
552 test_bit(R5_ReWrite, &sh->dev[i].flags))
553 atomic_add(STRIPE_SECTORS,
554 &rdev->corrected_errors);
555 generic_make_request(bi);
558 set_bit(STRIPE_DEGRADED, &sh->state);
559 pr_debug("skip op %ld on disc %d for sector %llu\n",
560 bi->bi_rw, i, (unsigned long long)sh->sector);
561 clear_bit(R5_LOCKED, &sh->dev[i].flags);
562 set_bit(STRIPE_HANDLE, &sh->state);
567 static struct dma_async_tx_descriptor *
568 async_copy_data(int frombio, struct bio *bio, struct page *page,
569 sector_t sector, struct dma_async_tx_descriptor *tx)
572 struct page *bio_page;
575 struct async_submit_ctl submit;
576 enum async_tx_flags flags = 0;
578 if (bio->bi_sector >= sector)
579 page_offset = (signed)(bio->bi_sector - sector) * 512;
581 page_offset = (signed)(sector - bio->bi_sector) * -512;
584 flags |= ASYNC_TX_FENCE;
585 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
587 bio_for_each_segment(bvl, bio, i) {
588 int len = bvl->bv_len;
592 if (page_offset < 0) {
593 b_offset = -page_offset;
594 page_offset += b_offset;
598 if (len > 0 && page_offset + len > STRIPE_SIZE)
599 clen = STRIPE_SIZE - page_offset;
604 b_offset += bvl->bv_offset;
605 bio_page = bvl->bv_page;
607 tx = async_memcpy(page, bio_page, page_offset,
608 b_offset, clen, &submit);
610 tx = async_memcpy(bio_page, page, b_offset,
611 page_offset, clen, &submit);
613 /* chain the operations */
614 submit.depend_tx = tx;
616 if (clen < len) /* hit end of page */
624 static void ops_complete_biofill(void *stripe_head_ref)
626 struct stripe_head *sh = stripe_head_ref;
627 struct bio *return_bi = NULL;
628 raid5_conf_t *conf = sh->raid_conf;
631 pr_debug("%s: stripe %llu\n", __func__,
632 (unsigned long long)sh->sector);
634 /* clear completed biofills */
635 spin_lock_irq(&conf->device_lock);
636 for (i = sh->disks; i--; ) {
637 struct r5dev *dev = &sh->dev[i];
639 /* acknowledge completion of a biofill operation */
640 /* and check if we need to reply to a read request,
641 * new R5_Wantfill requests are held off until
642 * !STRIPE_BIOFILL_RUN
644 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
645 struct bio *rbi, *rbi2;
650 while (rbi && rbi->bi_sector <
651 dev->sector + STRIPE_SECTORS) {
652 rbi2 = r5_next_bio(rbi, dev->sector);
653 if (!raid5_dec_bi_phys_segments(rbi)) {
654 rbi->bi_next = return_bi;
661 spin_unlock_irq(&conf->device_lock);
662 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
664 return_io(return_bi);
666 set_bit(STRIPE_HANDLE, &sh->state);
670 static void ops_run_biofill(struct stripe_head *sh)
672 struct dma_async_tx_descriptor *tx = NULL;
673 raid5_conf_t *conf = sh->raid_conf;
674 struct async_submit_ctl submit;
677 pr_debug("%s: stripe %llu\n", __func__,
678 (unsigned long long)sh->sector);
680 for (i = sh->disks; i--; ) {
681 struct r5dev *dev = &sh->dev[i];
682 if (test_bit(R5_Wantfill, &dev->flags)) {
684 spin_lock_irq(&conf->device_lock);
685 dev->read = rbi = dev->toread;
687 spin_unlock_irq(&conf->device_lock);
688 while (rbi && rbi->bi_sector <
689 dev->sector + STRIPE_SECTORS) {
690 tx = async_copy_data(0, rbi, dev->page,
692 rbi = r5_next_bio(rbi, dev->sector);
697 atomic_inc(&sh->count);
698 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
699 async_trigger_callback(&submit);
702 static void mark_target_uptodate(struct stripe_head *sh, int target)
709 tgt = &sh->dev[target];
710 set_bit(R5_UPTODATE, &tgt->flags);
711 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
712 clear_bit(R5_Wantcompute, &tgt->flags);
715 static void ops_complete_compute(void *stripe_head_ref)
717 struct stripe_head *sh = stripe_head_ref;
719 pr_debug("%s: stripe %llu\n", __func__,
720 (unsigned long long)sh->sector);
722 /* mark the computed target(s) as uptodate */
723 mark_target_uptodate(sh, sh->ops.target);
724 mark_target_uptodate(sh, sh->ops.target2);
726 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
727 if (sh->check_state == check_state_compute_run)
728 sh->check_state = check_state_compute_result;
729 set_bit(STRIPE_HANDLE, &sh->state);
733 /* return a pointer to the address conversion region of the scribble buffer */
734 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
735 struct raid5_percpu *percpu)
737 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
740 static struct dma_async_tx_descriptor *
741 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
743 int disks = sh->disks;
744 struct page **xor_srcs = percpu->scribble;
745 int target = sh->ops.target;
746 struct r5dev *tgt = &sh->dev[target];
747 struct page *xor_dest = tgt->page;
749 struct dma_async_tx_descriptor *tx;
750 struct async_submit_ctl submit;
753 pr_debug("%s: stripe %llu block: %d\n",
754 __func__, (unsigned long long)sh->sector, target);
755 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
757 for (i = disks; i--; )
759 xor_srcs[count++] = sh->dev[i].page;
761 atomic_inc(&sh->count);
763 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
764 ops_complete_compute, sh, to_addr_conv(sh, percpu));
765 if (unlikely(count == 1))
766 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
768 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
773 /* set_syndrome_sources - populate source buffers for gen_syndrome
774 * @srcs - (struct page *) array of size sh->disks
775 * @sh - stripe_head to parse
777 * Populates srcs in proper layout order for the stripe and returns the
778 * 'count' of sources to be used in a call to async_gen_syndrome. The P
779 * destination buffer is recorded in srcs[count] and the Q destination
780 * is recorded in srcs[count+1]].
782 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
784 int disks = sh->disks;
785 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
786 int d0_idx = raid6_d0(sh);
790 for (i = 0; i < disks; i++)
796 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
798 srcs[slot] = sh->dev[i].page;
799 i = raid6_next_disk(i, disks);
800 } while (i != d0_idx);
802 return syndrome_disks;
805 static struct dma_async_tx_descriptor *
806 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
808 int disks = sh->disks;
809 struct page **blocks = percpu->scribble;
811 int qd_idx = sh->qd_idx;
812 struct dma_async_tx_descriptor *tx;
813 struct async_submit_ctl submit;
819 if (sh->ops.target < 0)
820 target = sh->ops.target2;
821 else if (sh->ops.target2 < 0)
822 target = sh->ops.target;
824 /* we should only have one valid target */
827 pr_debug("%s: stripe %llu block: %d\n",
828 __func__, (unsigned long long)sh->sector, target);
830 tgt = &sh->dev[target];
831 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
834 atomic_inc(&sh->count);
836 if (target == qd_idx) {
837 count = set_syndrome_sources(blocks, sh);
838 blocks[count] = NULL; /* regenerating p is not necessary */
839 BUG_ON(blocks[count+1] != dest); /* q should already be set */
840 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
841 ops_complete_compute, sh,
842 to_addr_conv(sh, percpu));
843 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
845 /* Compute any data- or p-drive using XOR */
847 for (i = disks; i-- ; ) {
848 if (i == target || i == qd_idx)
850 blocks[count++] = sh->dev[i].page;
853 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
854 NULL, ops_complete_compute, sh,
855 to_addr_conv(sh, percpu));
856 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
862 static struct dma_async_tx_descriptor *
863 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
865 int i, count, disks = sh->disks;
866 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
867 int d0_idx = raid6_d0(sh);
868 int faila = -1, failb = -1;
869 int target = sh->ops.target;
870 int target2 = sh->ops.target2;
871 struct r5dev *tgt = &sh->dev[target];
872 struct r5dev *tgt2 = &sh->dev[target2];
873 struct dma_async_tx_descriptor *tx;
874 struct page **blocks = percpu->scribble;
875 struct async_submit_ctl submit;
877 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
878 __func__, (unsigned long long)sh->sector, target, target2);
879 BUG_ON(target < 0 || target2 < 0);
880 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
881 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
883 /* we need to open-code set_syndrome_sources to handle the
884 * slot number conversion for 'faila' and 'failb'
886 for (i = 0; i < disks ; i++)
891 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
893 blocks[slot] = sh->dev[i].page;
899 i = raid6_next_disk(i, disks);
900 } while (i != d0_idx);
902 BUG_ON(faila == failb);
905 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
906 __func__, (unsigned long long)sh->sector, faila, failb);
908 atomic_inc(&sh->count);
910 if (failb == syndrome_disks+1) {
911 /* Q disk is one of the missing disks */
912 if (faila == syndrome_disks) {
913 /* Missing P+Q, just recompute */
914 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
915 ops_complete_compute, sh,
916 to_addr_conv(sh, percpu));
917 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
918 STRIPE_SIZE, &submit);
922 int qd_idx = sh->qd_idx;
924 /* Missing D+Q: recompute D from P, then recompute Q */
925 if (target == qd_idx)
926 data_target = target2;
928 data_target = target;
931 for (i = disks; i-- ; ) {
932 if (i == data_target || i == qd_idx)
934 blocks[count++] = sh->dev[i].page;
936 dest = sh->dev[data_target].page;
937 init_async_submit(&submit,
938 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
940 to_addr_conv(sh, percpu));
941 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
944 count = set_syndrome_sources(blocks, sh);
945 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
946 ops_complete_compute, sh,
947 to_addr_conv(sh, percpu));
948 return async_gen_syndrome(blocks, 0, count+2,
949 STRIPE_SIZE, &submit);
952 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
953 ops_complete_compute, sh,
954 to_addr_conv(sh, percpu));
955 if (failb == syndrome_disks) {
956 /* We're missing D+P. */
957 return async_raid6_datap_recov(syndrome_disks+2,
961 /* We're missing D+D. */
962 return async_raid6_2data_recov(syndrome_disks+2,
963 STRIPE_SIZE, faila, failb,
970 static void ops_complete_prexor(void *stripe_head_ref)
972 struct stripe_head *sh = stripe_head_ref;
974 pr_debug("%s: stripe %llu\n", __func__,
975 (unsigned long long)sh->sector);
978 static struct dma_async_tx_descriptor *
979 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
980 struct dma_async_tx_descriptor *tx)
982 int disks = sh->disks;
983 struct page **xor_srcs = percpu->scribble;
984 int count = 0, pd_idx = sh->pd_idx, i;
985 struct async_submit_ctl submit;
987 /* existing parity data subtracted */
988 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
990 pr_debug("%s: stripe %llu\n", __func__,
991 (unsigned long long)sh->sector);
993 for (i = disks; i--; ) {
994 struct r5dev *dev = &sh->dev[i];
995 /* Only process blocks that are known to be uptodate */
996 if (test_bit(R5_Wantdrain, &dev->flags))
997 xor_srcs[count++] = dev->page;
1000 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1001 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1002 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1007 static struct dma_async_tx_descriptor *
1008 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1010 int disks = sh->disks;
1013 pr_debug("%s: stripe %llu\n", __func__,
1014 (unsigned long long)sh->sector);
1016 for (i = disks; i--; ) {
1017 struct r5dev *dev = &sh->dev[i];
1020 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1023 spin_lock_irq(&sh->raid_conf->device_lock);
1024 chosen = dev->towrite;
1025 dev->towrite = NULL;
1026 BUG_ON(dev->written);
1027 wbi = dev->written = chosen;
1028 spin_unlock_irq(&sh->raid_conf->device_lock);
1030 while (wbi && wbi->bi_sector <
1031 dev->sector + STRIPE_SECTORS) {
1032 if (wbi->bi_rw & REQ_FUA)
1033 set_bit(R5_WantFUA, &dev->flags);
1034 tx = async_copy_data(1, wbi, dev->page,
1036 wbi = r5_next_bio(wbi, dev->sector);
1044 static void ops_complete_reconstruct(void *stripe_head_ref)
1046 struct stripe_head *sh = stripe_head_ref;
1047 int disks = sh->disks;
1048 int pd_idx = sh->pd_idx;
1049 int qd_idx = sh->qd_idx;
1053 pr_debug("%s: stripe %llu\n", __func__,
1054 (unsigned long long)sh->sector);
1056 for (i = disks; i--; )
1057 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1059 for (i = disks; i--; ) {
1060 struct r5dev *dev = &sh->dev[i];
1062 if (dev->written || i == pd_idx || i == qd_idx) {
1063 set_bit(R5_UPTODATE, &dev->flags);
1065 set_bit(R5_WantFUA, &dev->flags);
1069 if (sh->reconstruct_state == reconstruct_state_drain_run)
1070 sh->reconstruct_state = reconstruct_state_drain_result;
1071 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1072 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1074 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1075 sh->reconstruct_state = reconstruct_state_result;
1078 set_bit(STRIPE_HANDLE, &sh->state);
1083 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1084 struct dma_async_tx_descriptor *tx)
1086 int disks = sh->disks;
1087 struct page **xor_srcs = percpu->scribble;
1088 struct async_submit_ctl submit;
1089 int count = 0, pd_idx = sh->pd_idx, i;
1090 struct page *xor_dest;
1092 unsigned long flags;
1094 pr_debug("%s: stripe %llu\n", __func__,
1095 (unsigned long long)sh->sector);
1097 /* check if prexor is active which means only process blocks
1098 * that are part of a read-modify-write (written)
1100 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1102 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1103 for (i = disks; i--; ) {
1104 struct r5dev *dev = &sh->dev[i];
1106 xor_srcs[count++] = dev->page;
1109 xor_dest = sh->dev[pd_idx].page;
1110 for (i = disks; i--; ) {
1111 struct r5dev *dev = &sh->dev[i];
1113 xor_srcs[count++] = dev->page;
1117 /* 1/ if we prexor'd then the dest is reused as a source
1118 * 2/ if we did not prexor then we are redoing the parity
1119 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1120 * for the synchronous xor case
1122 flags = ASYNC_TX_ACK |
1123 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1125 atomic_inc(&sh->count);
1127 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1128 to_addr_conv(sh, percpu));
1129 if (unlikely(count == 1))
1130 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1132 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1136 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1137 struct dma_async_tx_descriptor *tx)
1139 struct async_submit_ctl submit;
1140 struct page **blocks = percpu->scribble;
1143 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1145 count = set_syndrome_sources(blocks, sh);
1147 atomic_inc(&sh->count);
1149 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1150 sh, to_addr_conv(sh, percpu));
1151 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1154 static void ops_complete_check(void *stripe_head_ref)
1156 struct stripe_head *sh = stripe_head_ref;
1158 pr_debug("%s: stripe %llu\n", __func__,
1159 (unsigned long long)sh->sector);
1161 sh->check_state = check_state_check_result;
1162 set_bit(STRIPE_HANDLE, &sh->state);
1166 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1168 int disks = sh->disks;
1169 int pd_idx = sh->pd_idx;
1170 int qd_idx = sh->qd_idx;
1171 struct page *xor_dest;
1172 struct page **xor_srcs = percpu->scribble;
1173 struct dma_async_tx_descriptor *tx;
1174 struct async_submit_ctl submit;
1178 pr_debug("%s: stripe %llu\n", __func__,
1179 (unsigned long long)sh->sector);
1182 xor_dest = sh->dev[pd_idx].page;
1183 xor_srcs[count++] = xor_dest;
1184 for (i = disks; i--; ) {
1185 if (i == pd_idx || i == qd_idx)
1187 xor_srcs[count++] = sh->dev[i].page;
1190 init_async_submit(&submit, 0, NULL, NULL, NULL,
1191 to_addr_conv(sh, percpu));
1192 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1193 &sh->ops.zero_sum_result, &submit);
1195 atomic_inc(&sh->count);
1196 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1197 tx = async_trigger_callback(&submit);
1200 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1202 struct page **srcs = percpu->scribble;
1203 struct async_submit_ctl submit;
1206 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1207 (unsigned long long)sh->sector, checkp);
1209 count = set_syndrome_sources(srcs, sh);
1213 atomic_inc(&sh->count);
1214 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1215 sh, to_addr_conv(sh, percpu));
1216 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1217 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1220 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1222 int overlap_clear = 0, i, disks = sh->disks;
1223 struct dma_async_tx_descriptor *tx = NULL;
1224 raid5_conf_t *conf = sh->raid_conf;
1225 int level = conf->level;
1226 struct raid5_percpu *percpu;
1230 percpu = per_cpu_ptr(conf->percpu, cpu);
1231 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1232 ops_run_biofill(sh);
1236 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1238 tx = ops_run_compute5(sh, percpu);
1240 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1241 tx = ops_run_compute6_1(sh, percpu);
1243 tx = ops_run_compute6_2(sh, percpu);
1245 /* terminate the chain if reconstruct is not set to be run */
1246 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1250 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1251 tx = ops_run_prexor(sh, percpu, tx);
1253 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1254 tx = ops_run_biodrain(sh, tx);
1258 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1260 ops_run_reconstruct5(sh, percpu, tx);
1262 ops_run_reconstruct6(sh, percpu, tx);
1265 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1266 if (sh->check_state == check_state_run)
1267 ops_run_check_p(sh, percpu);
1268 else if (sh->check_state == check_state_run_q)
1269 ops_run_check_pq(sh, percpu, 0);
1270 else if (sh->check_state == check_state_run_pq)
1271 ops_run_check_pq(sh, percpu, 1);
1277 for (i = disks; i--; ) {
1278 struct r5dev *dev = &sh->dev[i];
1279 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1280 wake_up(&sh->raid_conf->wait_for_overlap);
1285 #ifdef CONFIG_MULTICORE_RAID456
1286 static void async_run_ops(void *param, async_cookie_t cookie)
1288 struct stripe_head *sh = param;
1289 unsigned long ops_request = sh->ops.request;
1291 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1292 wake_up(&sh->ops.wait_for_ops);
1294 __raid_run_ops(sh, ops_request);
1298 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1300 /* since handle_stripe can be called outside of raid5d context
1301 * we need to ensure sh->ops.request is de-staged before another
1304 wait_event(sh->ops.wait_for_ops,
1305 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1306 sh->ops.request = ops_request;
1308 atomic_inc(&sh->count);
1309 async_schedule(async_run_ops, sh);
1312 #define raid_run_ops __raid_run_ops
1315 static int grow_one_stripe(raid5_conf_t *conf)
1317 struct stripe_head *sh;
1318 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1322 sh->raid_conf = conf;
1323 #ifdef CONFIG_MULTICORE_RAID456
1324 init_waitqueue_head(&sh->ops.wait_for_ops);
1327 if (grow_buffers(sh)) {
1329 kmem_cache_free(conf->slab_cache, sh);
1332 /* we just created an active stripe so... */
1333 atomic_set(&sh->count, 1);
1334 atomic_inc(&conf->active_stripes);
1335 INIT_LIST_HEAD(&sh->lru);
1340 static int grow_stripes(raid5_conf_t *conf, int num)
1342 struct kmem_cache *sc;
1343 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1345 if (conf->mddev->gendisk)
1346 sprintf(conf->cache_name[0],
1347 "raid%d-%s", conf->level, mdname(conf->mddev));
1349 sprintf(conf->cache_name[0],
1350 "raid%d-%p", conf->level, conf->mddev);
1351 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1353 conf->active_name = 0;
1354 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1355 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1359 conf->slab_cache = sc;
1360 conf->pool_size = devs;
1362 if (!grow_one_stripe(conf))
1368 * scribble_len - return the required size of the scribble region
1369 * @num - total number of disks in the array
1371 * The size must be enough to contain:
1372 * 1/ a struct page pointer for each device in the array +2
1373 * 2/ room to convert each entry in (1) to its corresponding dma
1374 * (dma_map_page()) or page (page_address()) address.
1376 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1377 * calculate over all devices (not just the data blocks), using zeros in place
1378 * of the P and Q blocks.
1380 static size_t scribble_len(int num)
1384 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1389 static int resize_stripes(raid5_conf_t *conf, int newsize)
1391 /* Make all the stripes able to hold 'newsize' devices.
1392 * New slots in each stripe get 'page' set to a new page.
1394 * This happens in stages:
1395 * 1/ create a new kmem_cache and allocate the required number of
1397 * 2/ gather all the old stripe_heads and tranfer the pages across
1398 * to the new stripe_heads. This will have the side effect of
1399 * freezing the array as once all stripe_heads have been collected,
1400 * no IO will be possible. Old stripe heads are freed once their
1401 * pages have been transferred over, and the old kmem_cache is
1402 * freed when all stripes are done.
1403 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1404 * we simple return a failre status - no need to clean anything up.
1405 * 4/ allocate new pages for the new slots in the new stripe_heads.
1406 * If this fails, we don't bother trying the shrink the
1407 * stripe_heads down again, we just leave them as they are.
1408 * As each stripe_head is processed the new one is released into
1411 * Once step2 is started, we cannot afford to wait for a write,
1412 * so we use GFP_NOIO allocations.
1414 struct stripe_head *osh, *nsh;
1415 LIST_HEAD(newstripes);
1416 struct disk_info *ndisks;
1419 struct kmem_cache *sc;
1422 if (newsize <= conf->pool_size)
1423 return 0; /* never bother to shrink */
1425 err = md_allow_write(conf->mddev);
1430 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1431 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1436 for (i = conf->max_nr_stripes; i; i--) {
1437 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1441 nsh->raid_conf = conf;
1442 #ifdef CONFIG_MULTICORE_RAID456
1443 init_waitqueue_head(&nsh->ops.wait_for_ops);
1446 list_add(&nsh->lru, &newstripes);
1449 /* didn't get enough, give up */
1450 while (!list_empty(&newstripes)) {
1451 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1452 list_del(&nsh->lru);
1453 kmem_cache_free(sc, nsh);
1455 kmem_cache_destroy(sc);
1458 /* Step 2 - Must use GFP_NOIO now.
1459 * OK, we have enough stripes, start collecting inactive
1460 * stripes and copying them over
1462 list_for_each_entry(nsh, &newstripes, lru) {
1463 spin_lock_irq(&conf->device_lock);
1464 wait_event_lock_irq(conf->wait_for_stripe,
1465 !list_empty(&conf->inactive_list),
1468 osh = get_free_stripe(conf);
1469 spin_unlock_irq(&conf->device_lock);
1470 atomic_set(&nsh->count, 1);
1471 for(i=0; i<conf->pool_size; i++)
1472 nsh->dev[i].page = osh->dev[i].page;
1473 for( ; i<newsize; i++)
1474 nsh->dev[i].page = NULL;
1475 kmem_cache_free(conf->slab_cache, osh);
1477 kmem_cache_destroy(conf->slab_cache);
1480 * At this point, we are holding all the stripes so the array
1481 * is completely stalled, so now is a good time to resize
1482 * conf->disks and the scribble region
1484 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1486 for (i=0; i<conf->raid_disks; i++)
1487 ndisks[i] = conf->disks[i];
1489 conf->disks = ndisks;
1494 conf->scribble_len = scribble_len(newsize);
1495 for_each_present_cpu(cpu) {
1496 struct raid5_percpu *percpu;
1499 percpu = per_cpu_ptr(conf->percpu, cpu);
1500 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1503 kfree(percpu->scribble);
1504 percpu->scribble = scribble;
1512 /* Step 4, return new stripes to service */
1513 while(!list_empty(&newstripes)) {
1514 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1515 list_del_init(&nsh->lru);
1517 for (i=conf->raid_disks; i < newsize; i++)
1518 if (nsh->dev[i].page == NULL) {
1519 struct page *p = alloc_page(GFP_NOIO);
1520 nsh->dev[i].page = p;
1524 release_stripe(nsh);
1526 /* critical section pass, GFP_NOIO no longer needed */
1528 conf->slab_cache = sc;
1529 conf->active_name = 1-conf->active_name;
1530 conf->pool_size = newsize;
1534 static int drop_one_stripe(raid5_conf_t *conf)
1536 struct stripe_head *sh;
1538 spin_lock_irq(&conf->device_lock);
1539 sh = get_free_stripe(conf);
1540 spin_unlock_irq(&conf->device_lock);
1543 BUG_ON(atomic_read(&sh->count));
1545 kmem_cache_free(conf->slab_cache, sh);
1546 atomic_dec(&conf->active_stripes);
1550 static void shrink_stripes(raid5_conf_t *conf)
1552 while (drop_one_stripe(conf))
1555 if (conf->slab_cache)
1556 kmem_cache_destroy(conf->slab_cache);
1557 conf->slab_cache = NULL;
1560 static void raid5_end_read_request(struct bio * bi, int error)
1562 struct stripe_head *sh = bi->bi_private;
1563 raid5_conf_t *conf = sh->raid_conf;
1564 int disks = sh->disks, i;
1565 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1566 char b[BDEVNAME_SIZE];
1570 for (i=0 ; i<disks; i++)
1571 if (bi == &sh->dev[i].req)
1574 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1575 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1583 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1584 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1585 rdev = conf->disks[i].rdev;
1586 printk_rl(KERN_INFO "md/raid:%s: read error corrected"
1587 " (%lu sectors at %llu on %s)\n",
1588 mdname(conf->mddev), STRIPE_SECTORS,
1589 (unsigned long long)(sh->sector
1590 + rdev->data_offset),
1591 bdevname(rdev->bdev, b));
1592 clear_bit(R5_ReadError, &sh->dev[i].flags);
1593 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1595 if (atomic_read(&conf->disks[i].rdev->read_errors))
1596 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1598 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1600 rdev = conf->disks[i].rdev;
1602 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1603 atomic_inc(&rdev->read_errors);
1604 if (conf->mddev->degraded >= conf->max_degraded)
1605 printk_rl(KERN_WARNING
1606 "md/raid:%s: read error not correctable "
1607 "(sector %llu on %s).\n",
1608 mdname(conf->mddev),
1609 (unsigned long long)(sh->sector
1610 + rdev->data_offset),
1612 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1614 printk_rl(KERN_WARNING
1615 "md/raid:%s: read error NOT corrected!! "
1616 "(sector %llu on %s).\n",
1617 mdname(conf->mddev),
1618 (unsigned long long)(sh->sector
1619 + rdev->data_offset),
1621 else if (atomic_read(&rdev->read_errors)
1622 > conf->max_nr_stripes)
1624 "md/raid:%s: Too many read errors, failing device %s.\n",
1625 mdname(conf->mddev), bdn);
1629 set_bit(R5_ReadError, &sh->dev[i].flags);
1631 clear_bit(R5_ReadError, &sh->dev[i].flags);
1632 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1633 md_error(conf->mddev, rdev);
1636 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1637 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1638 set_bit(STRIPE_HANDLE, &sh->state);
1642 static void raid5_end_write_request(struct bio *bi, int error)
1644 struct stripe_head *sh = bi->bi_private;
1645 raid5_conf_t *conf = sh->raid_conf;
1646 int disks = sh->disks, i;
1647 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1649 for (i=0 ; i<disks; i++)
1650 if (bi == &sh->dev[i].req)
1653 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1654 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1662 md_error(conf->mddev, conf->disks[i].rdev);
1664 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1666 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1667 set_bit(STRIPE_HANDLE, &sh->state);
1672 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1674 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1676 struct r5dev *dev = &sh->dev[i];
1678 bio_init(&dev->req);
1679 dev->req.bi_io_vec = &dev->vec;
1681 dev->req.bi_max_vecs++;
1682 dev->vec.bv_page = dev->page;
1683 dev->vec.bv_len = STRIPE_SIZE;
1684 dev->vec.bv_offset = 0;
1686 dev->req.bi_sector = sh->sector;
1687 dev->req.bi_private = sh;
1690 dev->sector = compute_blocknr(sh, i, previous);
1693 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1695 char b[BDEVNAME_SIZE];
1696 raid5_conf_t *conf = mddev->private;
1697 pr_debug("raid456: error called\n");
1699 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1700 unsigned long flags;
1701 spin_lock_irqsave(&conf->device_lock, flags);
1703 spin_unlock_irqrestore(&conf->device_lock, flags);
1705 * if recovery was running, make sure it aborts.
1707 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1709 set_bit(Faulty, &rdev->flags);
1710 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1712 "md/raid:%s: Disk failure on %s, disabling device.\n"
1713 "md/raid:%s: Operation continuing on %d devices.\n",
1715 bdevname(rdev->bdev, b),
1717 conf->raid_disks - mddev->degraded);
1721 * Input: a 'big' sector number,
1722 * Output: index of the data and parity disk, and the sector # in them.
1724 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1725 int previous, int *dd_idx,
1726 struct stripe_head *sh)
1728 sector_t stripe, stripe2;
1729 sector_t chunk_number;
1730 unsigned int chunk_offset;
1733 sector_t new_sector;
1734 int algorithm = previous ? conf->prev_algo
1736 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1737 : conf->chunk_sectors;
1738 int raid_disks = previous ? conf->previous_raid_disks
1740 int data_disks = raid_disks - conf->max_degraded;
1742 /* First compute the information on this sector */
1745 * Compute the chunk number and the sector offset inside the chunk
1747 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1748 chunk_number = r_sector;
1751 * Compute the stripe number
1753 stripe = chunk_number;
1754 *dd_idx = sector_div(stripe, data_disks);
1757 * Select the parity disk based on the user selected algorithm.
1759 pd_idx = qd_idx = ~0;
1760 switch(conf->level) {
1762 pd_idx = data_disks;
1765 switch (algorithm) {
1766 case ALGORITHM_LEFT_ASYMMETRIC:
1767 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1768 if (*dd_idx >= pd_idx)
1771 case ALGORITHM_RIGHT_ASYMMETRIC:
1772 pd_idx = sector_div(stripe2, raid_disks);
1773 if (*dd_idx >= pd_idx)
1776 case ALGORITHM_LEFT_SYMMETRIC:
1777 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1778 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1780 case ALGORITHM_RIGHT_SYMMETRIC:
1781 pd_idx = sector_div(stripe2, raid_disks);
1782 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1784 case ALGORITHM_PARITY_0:
1788 case ALGORITHM_PARITY_N:
1789 pd_idx = data_disks;
1797 switch (algorithm) {
1798 case ALGORITHM_LEFT_ASYMMETRIC:
1799 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1800 qd_idx = pd_idx + 1;
1801 if (pd_idx == raid_disks-1) {
1802 (*dd_idx)++; /* Q D D D P */
1804 } else if (*dd_idx >= pd_idx)
1805 (*dd_idx) += 2; /* D D P Q D */
1807 case ALGORITHM_RIGHT_ASYMMETRIC:
1808 pd_idx = sector_div(stripe2, raid_disks);
1809 qd_idx = pd_idx + 1;
1810 if (pd_idx == raid_disks-1) {
1811 (*dd_idx)++; /* Q D D D P */
1813 } else if (*dd_idx >= pd_idx)
1814 (*dd_idx) += 2; /* D D P Q D */
1816 case ALGORITHM_LEFT_SYMMETRIC:
1817 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1818 qd_idx = (pd_idx + 1) % raid_disks;
1819 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1821 case ALGORITHM_RIGHT_SYMMETRIC:
1822 pd_idx = sector_div(stripe2, raid_disks);
1823 qd_idx = (pd_idx + 1) % raid_disks;
1824 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1827 case ALGORITHM_PARITY_0:
1832 case ALGORITHM_PARITY_N:
1833 pd_idx = data_disks;
1834 qd_idx = data_disks + 1;
1837 case ALGORITHM_ROTATING_ZERO_RESTART:
1838 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1839 * of blocks for computing Q is different.
1841 pd_idx = sector_div(stripe2, raid_disks);
1842 qd_idx = pd_idx + 1;
1843 if (pd_idx == raid_disks-1) {
1844 (*dd_idx)++; /* Q D D D P */
1846 } else if (*dd_idx >= pd_idx)
1847 (*dd_idx) += 2; /* D D P Q D */
1851 case ALGORITHM_ROTATING_N_RESTART:
1852 /* Same a left_asymmetric, by first stripe is
1853 * D D D P Q rather than
1857 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1858 qd_idx = pd_idx + 1;
1859 if (pd_idx == raid_disks-1) {
1860 (*dd_idx)++; /* Q D D D P */
1862 } else if (*dd_idx >= pd_idx)
1863 (*dd_idx) += 2; /* D D P Q D */
1867 case ALGORITHM_ROTATING_N_CONTINUE:
1868 /* Same as left_symmetric but Q is before P */
1869 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1870 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1871 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1875 case ALGORITHM_LEFT_ASYMMETRIC_6:
1876 /* RAID5 left_asymmetric, with Q on last device */
1877 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1878 if (*dd_idx >= pd_idx)
1880 qd_idx = raid_disks - 1;
1883 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1884 pd_idx = sector_div(stripe2, raid_disks-1);
1885 if (*dd_idx >= pd_idx)
1887 qd_idx = raid_disks - 1;
1890 case ALGORITHM_LEFT_SYMMETRIC_6:
1891 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1892 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1893 qd_idx = raid_disks - 1;
1896 case ALGORITHM_RIGHT_SYMMETRIC_6:
1897 pd_idx = sector_div(stripe2, raid_disks-1);
1898 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1899 qd_idx = raid_disks - 1;
1902 case ALGORITHM_PARITY_0_6:
1905 qd_idx = raid_disks - 1;
1915 sh->pd_idx = pd_idx;
1916 sh->qd_idx = qd_idx;
1917 sh->ddf_layout = ddf_layout;
1920 * Finally, compute the new sector number
1922 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1927 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1929 raid5_conf_t *conf = sh->raid_conf;
1930 int raid_disks = sh->disks;
1931 int data_disks = raid_disks - conf->max_degraded;
1932 sector_t new_sector = sh->sector, check;
1933 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1934 : conf->chunk_sectors;
1935 int algorithm = previous ? conf->prev_algo
1939 sector_t chunk_number;
1940 int dummy1, dd_idx = i;
1942 struct stripe_head sh2;
1945 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1946 stripe = new_sector;
1948 if (i == sh->pd_idx)
1950 switch(conf->level) {
1953 switch (algorithm) {
1954 case ALGORITHM_LEFT_ASYMMETRIC:
1955 case ALGORITHM_RIGHT_ASYMMETRIC:
1959 case ALGORITHM_LEFT_SYMMETRIC:
1960 case ALGORITHM_RIGHT_SYMMETRIC:
1963 i -= (sh->pd_idx + 1);
1965 case ALGORITHM_PARITY_0:
1968 case ALGORITHM_PARITY_N:
1975 if (i == sh->qd_idx)
1976 return 0; /* It is the Q disk */
1977 switch (algorithm) {
1978 case ALGORITHM_LEFT_ASYMMETRIC:
1979 case ALGORITHM_RIGHT_ASYMMETRIC:
1980 case ALGORITHM_ROTATING_ZERO_RESTART:
1981 case ALGORITHM_ROTATING_N_RESTART:
1982 if (sh->pd_idx == raid_disks-1)
1983 i--; /* Q D D D P */
1984 else if (i > sh->pd_idx)
1985 i -= 2; /* D D P Q D */
1987 case ALGORITHM_LEFT_SYMMETRIC:
1988 case ALGORITHM_RIGHT_SYMMETRIC:
1989 if (sh->pd_idx == raid_disks-1)
1990 i--; /* Q D D D P */
1995 i -= (sh->pd_idx + 2);
1998 case ALGORITHM_PARITY_0:
2001 case ALGORITHM_PARITY_N:
2003 case ALGORITHM_ROTATING_N_CONTINUE:
2004 /* Like left_symmetric, but P is before Q */
2005 if (sh->pd_idx == 0)
2006 i--; /* P D D D Q */
2011 i -= (sh->pd_idx + 1);
2014 case ALGORITHM_LEFT_ASYMMETRIC_6:
2015 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2019 case ALGORITHM_LEFT_SYMMETRIC_6:
2020 case ALGORITHM_RIGHT_SYMMETRIC_6:
2022 i += data_disks + 1;
2023 i -= (sh->pd_idx + 1);
2025 case ALGORITHM_PARITY_0_6:
2034 chunk_number = stripe * data_disks + i;
2035 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2037 check = raid5_compute_sector(conf, r_sector,
2038 previous, &dummy1, &sh2);
2039 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2040 || sh2.qd_idx != sh->qd_idx) {
2041 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2042 mdname(conf->mddev));
2050 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2051 int rcw, int expand)
2053 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2054 raid5_conf_t *conf = sh->raid_conf;
2055 int level = conf->level;
2058 /* if we are not expanding this is a proper write request, and
2059 * there will be bios with new data to be drained into the
2063 sh->reconstruct_state = reconstruct_state_drain_run;
2064 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2066 sh->reconstruct_state = reconstruct_state_run;
2068 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2070 for (i = disks; i--; ) {
2071 struct r5dev *dev = &sh->dev[i];
2074 set_bit(R5_LOCKED, &dev->flags);
2075 set_bit(R5_Wantdrain, &dev->flags);
2077 clear_bit(R5_UPTODATE, &dev->flags);
2081 if (s->locked + conf->max_degraded == disks)
2082 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2083 atomic_inc(&conf->pending_full_writes);
2086 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2087 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2089 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2090 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2091 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2092 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2094 for (i = disks; i--; ) {
2095 struct r5dev *dev = &sh->dev[i];
2100 (test_bit(R5_UPTODATE, &dev->flags) ||
2101 test_bit(R5_Wantcompute, &dev->flags))) {
2102 set_bit(R5_Wantdrain, &dev->flags);
2103 set_bit(R5_LOCKED, &dev->flags);
2104 clear_bit(R5_UPTODATE, &dev->flags);
2110 /* keep the parity disk(s) locked while asynchronous operations
2113 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2114 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2118 int qd_idx = sh->qd_idx;
2119 struct r5dev *dev = &sh->dev[qd_idx];
2121 set_bit(R5_LOCKED, &dev->flags);
2122 clear_bit(R5_UPTODATE, &dev->flags);
2126 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2127 __func__, (unsigned long long)sh->sector,
2128 s->locked, s->ops_request);
2132 * Each stripe/dev can have one or more bion attached.
2133 * toread/towrite point to the first in a chain.
2134 * The bi_next chain must be in order.
2136 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2139 raid5_conf_t *conf = sh->raid_conf;
2142 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2143 (unsigned long long)bi->bi_sector,
2144 (unsigned long long)sh->sector);
2147 spin_lock_irq(&conf->device_lock);
2149 bip = &sh->dev[dd_idx].towrite;
2150 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2153 bip = &sh->dev[dd_idx].toread;
2154 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2155 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2157 bip = & (*bip)->bi_next;
2159 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2162 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2166 bi->bi_phys_segments++;
2169 /* check if page is covered */
2170 sector_t sector = sh->dev[dd_idx].sector;
2171 for (bi=sh->dev[dd_idx].towrite;
2172 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2173 bi && bi->bi_sector <= sector;
2174 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2175 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2176 sector = bi->bi_sector + (bi->bi_size>>9);
2178 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2179 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2181 spin_unlock_irq(&conf->device_lock);
2183 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2184 (unsigned long long)(*bip)->bi_sector,
2185 (unsigned long long)sh->sector, dd_idx);
2187 if (conf->mddev->bitmap && firstwrite) {
2188 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2190 sh->bm_seq = conf->seq_flush+1;
2191 set_bit(STRIPE_BIT_DELAY, &sh->state);
2196 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2197 spin_unlock_irq(&conf->device_lock);
2201 static void end_reshape(raid5_conf_t *conf);
2203 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2204 struct stripe_head *sh)
2206 int sectors_per_chunk =
2207 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2209 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2210 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2212 raid5_compute_sector(conf,
2213 stripe * (disks - conf->max_degraded)
2214 *sectors_per_chunk + chunk_offset,
2220 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2221 struct stripe_head_state *s, int disks,
2222 struct bio **return_bi)
2225 for (i = disks; i--; ) {
2229 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2232 rdev = rcu_dereference(conf->disks[i].rdev);
2233 if (rdev && test_bit(In_sync, &rdev->flags))
2234 /* multiple read failures in one stripe */
2235 md_error(conf->mddev, rdev);
2238 spin_lock_irq(&conf->device_lock);
2239 /* fail all writes first */
2240 bi = sh->dev[i].towrite;
2241 sh->dev[i].towrite = NULL;
2247 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2248 wake_up(&conf->wait_for_overlap);
2250 while (bi && bi->bi_sector <
2251 sh->dev[i].sector + STRIPE_SECTORS) {
2252 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2253 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2254 if (!raid5_dec_bi_phys_segments(bi)) {
2255 md_write_end(conf->mddev);
2256 bi->bi_next = *return_bi;
2261 /* and fail all 'written' */
2262 bi = sh->dev[i].written;
2263 sh->dev[i].written = NULL;
2264 if (bi) bitmap_end = 1;
2265 while (bi && bi->bi_sector <
2266 sh->dev[i].sector + STRIPE_SECTORS) {
2267 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2268 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2269 if (!raid5_dec_bi_phys_segments(bi)) {
2270 md_write_end(conf->mddev);
2271 bi->bi_next = *return_bi;
2277 /* fail any reads if this device is non-operational and
2278 * the data has not reached the cache yet.
2280 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2281 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2282 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2283 bi = sh->dev[i].toread;
2284 sh->dev[i].toread = NULL;
2285 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2286 wake_up(&conf->wait_for_overlap);
2287 if (bi) s->to_read--;
2288 while (bi && bi->bi_sector <
2289 sh->dev[i].sector + STRIPE_SECTORS) {
2290 struct bio *nextbi =
2291 r5_next_bio(bi, sh->dev[i].sector);
2292 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2293 if (!raid5_dec_bi_phys_segments(bi)) {
2294 bi->bi_next = *return_bi;
2300 spin_unlock_irq(&conf->device_lock);
2302 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2303 STRIPE_SECTORS, 0, 0);
2306 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2307 if (atomic_dec_and_test(&conf->pending_full_writes))
2308 md_wakeup_thread(conf->mddev->thread);
2311 /* fetch_block5 - 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_fill5 to continue
2317 static int fetch_block5(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 *failed_dev = &sh->dev[s->failed_num[0]];
2323 /* is the data in this block needed, and can we get it? */
2324 if (!test_bit(R5_LOCKED, &dev->flags) &&
2325 !test_bit(R5_UPTODATE, &dev->flags) &&
2327 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2328 s->syncing || s->expanding ||
2330 (failed_dev->toread ||
2331 (failed_dev->towrite &&
2332 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2333 /* We would like to get this block, possibly by computing it,
2334 * otherwise read it if the backing disk is insync
2336 if ((s->uptodate == disks - 1) &&
2337 (s->failed && disk_idx == s->failed_num[0])) {
2338 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2339 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2340 set_bit(R5_Wantcompute, &dev->flags);
2341 sh->ops.target = disk_idx;
2342 sh->ops.target2 = -1;
2344 /* Careful: from this point on 'uptodate' is in the eye
2345 * of raid_run_ops which services 'compute' operations
2346 * before writes. R5_Wantcompute flags a block that will
2347 * be R5_UPTODATE by the time it is needed for a
2348 * subsequent operation.
2351 return 1; /* uptodate + compute == disks */
2352 } else if (test_bit(R5_Insync, &dev->flags)) {
2353 set_bit(R5_LOCKED, &dev->flags);
2354 set_bit(R5_Wantread, &dev->flags);
2356 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2365 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2367 static void handle_stripe_fill5(struct stripe_head *sh,
2368 struct stripe_head_state *s, int disks)
2372 /* look for blocks to read/compute, skip this if a compute
2373 * is already in flight, or if the stripe contents are in the
2374 * midst of changing due to a write
2376 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2377 !sh->reconstruct_state)
2378 for (i = disks; i--; )
2379 if (fetch_block5(sh, s, i, disks))
2381 set_bit(STRIPE_HANDLE, &sh->state);
2384 /* fetch_block6 - checks the given member device to see if its data needs
2385 * to be read or computed to satisfy a request.
2387 * Returns 1 when no more member devices need to be checked, otherwise returns
2388 * 0 to tell the loop in handle_stripe_fill6 to continue
2390 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2391 int disk_idx, int disks)
2393 struct r5dev *dev = &sh->dev[disk_idx];
2394 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2395 &sh->dev[s->failed_num[1]] };
2397 if (!test_bit(R5_LOCKED, &dev->flags) &&
2398 !test_bit(R5_UPTODATE, &dev->flags) &&
2400 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2401 s->syncing || s->expanding ||
2403 (fdev[0]->toread || s->to_write)) ||
2405 (fdev[1]->toread || s->to_write)))) {
2406 /* we would like to get this block, possibly by computing it,
2407 * otherwise read it if the backing disk is insync
2409 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2410 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2411 if ((s->uptodate == disks - 1) &&
2412 (s->failed && (disk_idx == s->failed_num[0] ||
2413 disk_idx == s->failed_num[1]))) {
2414 /* have disk failed, and we're requested to fetch it;
2417 pr_debug("Computing stripe %llu block %d\n",
2418 (unsigned long long)sh->sector, disk_idx);
2419 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2420 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2421 set_bit(R5_Wantcompute, &dev->flags);
2422 sh->ops.target = disk_idx;
2423 sh->ops.target2 = -1; /* no 2nd target */
2427 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2428 /* Computing 2-failure is *very* expensive; only
2429 * do it if failed >= 2
2432 for (other = disks; other--; ) {
2433 if (other == disk_idx)
2435 if (!test_bit(R5_UPTODATE,
2436 &sh->dev[other].flags))
2440 pr_debug("Computing stripe %llu blocks %d,%d\n",
2441 (unsigned long long)sh->sector,
2443 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2444 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2445 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2446 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2447 sh->ops.target = disk_idx;
2448 sh->ops.target2 = other;
2452 } else if (test_bit(R5_Insync, &dev->flags)) {
2453 set_bit(R5_LOCKED, &dev->flags);
2454 set_bit(R5_Wantread, &dev->flags);
2456 pr_debug("Reading block %d (sync=%d)\n",
2457 disk_idx, s->syncing);
2465 * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2467 static void handle_stripe_fill6(struct stripe_head *sh,
2468 struct stripe_head_state *s,
2473 /* look for blocks to read/compute, skip this if a compute
2474 * is already in flight, or if the stripe contents are in the
2475 * midst of changing due to a write
2477 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2478 !sh->reconstruct_state)
2479 for (i = disks; i--; )
2480 if (fetch_block6(sh, s, i, disks))
2482 set_bit(STRIPE_HANDLE, &sh->state);
2486 /* handle_stripe_clean_event
2487 * any written block on an uptodate or failed drive can be returned.
2488 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2489 * never LOCKED, so we don't need to test 'failed' directly.
2491 static void handle_stripe_clean_event(raid5_conf_t *conf,
2492 struct stripe_head *sh, int disks, struct bio **return_bi)
2497 for (i = disks; i--; )
2498 if (sh->dev[i].written) {
2500 if (!test_bit(R5_LOCKED, &dev->flags) &&
2501 test_bit(R5_UPTODATE, &dev->flags)) {
2502 /* We can return any write requests */
2503 struct bio *wbi, *wbi2;
2505 pr_debug("Return write for disc %d\n", i);
2506 spin_lock_irq(&conf->device_lock);
2508 dev->written = NULL;
2509 while (wbi && wbi->bi_sector <
2510 dev->sector + STRIPE_SECTORS) {
2511 wbi2 = r5_next_bio(wbi, dev->sector);
2512 if (!raid5_dec_bi_phys_segments(wbi)) {
2513 md_write_end(conf->mddev);
2514 wbi->bi_next = *return_bi;
2519 if (dev->towrite == NULL)
2521 spin_unlock_irq(&conf->device_lock);
2523 bitmap_endwrite(conf->mddev->bitmap,
2526 !test_bit(STRIPE_DEGRADED, &sh->state),
2531 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2532 if (atomic_dec_and_test(&conf->pending_full_writes))
2533 md_wakeup_thread(conf->mddev->thread);
2536 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2537 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2539 int rmw = 0, rcw = 0, i;
2540 for (i = disks; i--; ) {
2541 /* would I have to read this buffer for read_modify_write */
2542 struct r5dev *dev = &sh->dev[i];
2543 if ((dev->towrite || i == sh->pd_idx) &&
2544 !test_bit(R5_LOCKED, &dev->flags) &&
2545 !(test_bit(R5_UPTODATE, &dev->flags) ||
2546 test_bit(R5_Wantcompute, &dev->flags))) {
2547 if (test_bit(R5_Insync, &dev->flags))
2550 rmw += 2*disks; /* cannot read it */
2552 /* Would I have to read this buffer for reconstruct_write */
2553 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2554 !test_bit(R5_LOCKED, &dev->flags) &&
2555 !(test_bit(R5_UPTODATE, &dev->flags) ||
2556 test_bit(R5_Wantcompute, &dev->flags))) {
2557 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2562 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2563 (unsigned long long)sh->sector, rmw, rcw);
2564 set_bit(STRIPE_HANDLE, &sh->state);
2565 if (rmw < rcw && rmw > 0)
2566 /* prefer read-modify-write, but need to get some data */
2567 for (i = disks; i--; ) {
2568 struct r5dev *dev = &sh->dev[i];
2569 if ((dev->towrite || i == sh->pd_idx) &&
2570 !test_bit(R5_LOCKED, &dev->flags) &&
2571 !(test_bit(R5_UPTODATE, &dev->flags) ||
2572 test_bit(R5_Wantcompute, &dev->flags)) &&
2573 test_bit(R5_Insync, &dev->flags)) {
2575 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2576 pr_debug("Read_old block "
2577 "%d for r-m-w\n", i);
2578 set_bit(R5_LOCKED, &dev->flags);
2579 set_bit(R5_Wantread, &dev->flags);
2582 set_bit(STRIPE_DELAYED, &sh->state);
2583 set_bit(STRIPE_HANDLE, &sh->state);
2587 if (rcw <= rmw && rcw > 0)
2588 /* want reconstruct write, but need to get some data */
2589 for (i = disks; i--; ) {
2590 struct r5dev *dev = &sh->dev[i];
2591 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2593 !test_bit(R5_LOCKED, &dev->flags) &&
2594 !(test_bit(R5_UPTODATE, &dev->flags) ||
2595 test_bit(R5_Wantcompute, &dev->flags)) &&
2596 test_bit(R5_Insync, &dev->flags)) {
2598 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2599 pr_debug("Read_old block "
2600 "%d for Reconstruct\n", i);
2601 set_bit(R5_LOCKED, &dev->flags);
2602 set_bit(R5_Wantread, &dev->flags);
2605 set_bit(STRIPE_DELAYED, &sh->state);
2606 set_bit(STRIPE_HANDLE, &sh->state);
2610 /* now if nothing is locked, and if we have enough data,
2611 * we can start a write request
2613 /* since handle_stripe can be called at any time we need to handle the
2614 * case where a compute block operation has been submitted and then a
2615 * subsequent call wants to start a write request. raid_run_ops only
2616 * handles the case where compute block and reconstruct are requested
2617 * simultaneously. If this is not the case then new writes need to be
2618 * held off until the compute completes.
2620 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2621 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2622 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2623 schedule_reconstruction(sh, s, rcw == 0, 0);
2626 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2627 struct stripe_head *sh, struct stripe_head_state *s,
2630 int rcw = 0, pd_idx = sh->pd_idx, i;
2631 int qd_idx = sh->qd_idx;
2633 set_bit(STRIPE_HANDLE, &sh->state);
2634 for (i = disks; i--; ) {
2635 struct r5dev *dev = &sh->dev[i];
2636 /* check if we haven't enough data */
2637 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2638 i != pd_idx && i != qd_idx &&
2639 !test_bit(R5_LOCKED, &dev->flags) &&
2640 !(test_bit(R5_UPTODATE, &dev->flags) ||
2641 test_bit(R5_Wantcompute, &dev->flags))) {
2643 if (!test_bit(R5_Insync, &dev->flags))
2644 continue; /* it's a failed drive */
2647 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2648 pr_debug("Read_old stripe %llu "
2649 "block %d for Reconstruct\n",
2650 (unsigned long long)sh->sector, i);
2651 set_bit(R5_LOCKED, &dev->flags);
2652 set_bit(R5_Wantread, &dev->flags);
2655 pr_debug("Request delayed stripe %llu "
2656 "block %d for Reconstruct\n",
2657 (unsigned long long)sh->sector, i);
2658 set_bit(STRIPE_DELAYED, &sh->state);
2659 set_bit(STRIPE_HANDLE, &sh->state);
2663 /* now if nothing is locked, and if we have enough data, we can start a
2666 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2667 s->locked == 0 && rcw == 0 &&
2668 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2669 schedule_reconstruction(sh, s, 1, 0);
2673 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2674 struct stripe_head_state *s, int disks)
2676 struct r5dev *dev = NULL;
2678 set_bit(STRIPE_HANDLE, &sh->state);
2680 switch (sh->check_state) {
2681 case check_state_idle:
2682 /* start a new check operation if there are no failures */
2683 if (s->failed == 0) {
2684 BUG_ON(s->uptodate != disks);
2685 sh->check_state = check_state_run;
2686 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2687 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2691 dev = &sh->dev[s->failed_num[0]];
2693 case check_state_compute_result:
2694 sh->check_state = check_state_idle;
2696 dev = &sh->dev[sh->pd_idx];
2698 /* check that a write has not made the stripe insync */
2699 if (test_bit(STRIPE_INSYNC, &sh->state))
2702 /* either failed parity check, or recovery is happening */
2703 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2704 BUG_ON(s->uptodate != disks);
2706 set_bit(R5_LOCKED, &dev->flags);
2708 set_bit(R5_Wantwrite, &dev->flags);
2710 clear_bit(STRIPE_DEGRADED, &sh->state);
2711 set_bit(STRIPE_INSYNC, &sh->state);
2713 case check_state_run:
2714 break; /* we will be called again upon completion */
2715 case check_state_check_result:
2716 sh->check_state = check_state_idle;
2718 /* if a failure occurred during the check operation, leave
2719 * STRIPE_INSYNC not set and let the stripe be handled again
2724 /* handle a successful check operation, if parity is correct
2725 * we are done. Otherwise update the mismatch count and repair
2726 * parity if !MD_RECOVERY_CHECK
2728 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2729 /* parity is correct (on disc,
2730 * not in buffer any more)
2732 set_bit(STRIPE_INSYNC, &sh->state);
2734 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2735 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2736 /* don't try to repair!! */
2737 set_bit(STRIPE_INSYNC, &sh->state);
2739 sh->check_state = check_state_compute_run;
2740 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2741 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2742 set_bit(R5_Wantcompute,
2743 &sh->dev[sh->pd_idx].flags);
2744 sh->ops.target = sh->pd_idx;
2745 sh->ops.target2 = -1;
2750 case check_state_compute_run:
2753 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2754 __func__, sh->check_state,
2755 (unsigned long long) sh->sector);
2761 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2762 struct stripe_head_state *s,
2765 int pd_idx = sh->pd_idx;
2766 int qd_idx = sh->qd_idx;
2769 set_bit(STRIPE_HANDLE, &sh->state);
2771 BUG_ON(s->failed > 2);
2773 /* Want to check and possibly repair P and Q.
2774 * However there could be one 'failed' device, in which
2775 * case we can only check one of them, possibly using the
2776 * other to generate missing data
2779 switch (sh->check_state) {
2780 case check_state_idle:
2781 /* start a new check operation if there are < 2 failures */
2782 if (s->failed == s->q_failed) {
2783 /* The only possible failed device holds Q, so it
2784 * makes sense to check P (If anything else were failed,
2785 * we would have used P to recreate it).
2787 sh->check_state = check_state_run;
2789 if (!s->q_failed && s->failed < 2) {
2790 /* Q is not failed, and we didn't use it to generate
2791 * anything, so it makes sense to check it
2793 if (sh->check_state == check_state_run)
2794 sh->check_state = check_state_run_pq;
2796 sh->check_state = check_state_run_q;
2799 /* discard potentially stale zero_sum_result */
2800 sh->ops.zero_sum_result = 0;
2802 if (sh->check_state == check_state_run) {
2803 /* async_xor_zero_sum destroys the contents of P */
2804 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2807 if (sh->check_state >= check_state_run &&
2808 sh->check_state <= check_state_run_pq) {
2809 /* async_syndrome_zero_sum preserves P and Q, so
2810 * no need to mark them !uptodate here
2812 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2816 /* we have 2-disk failure */
2817 BUG_ON(s->failed != 2);
2819 case check_state_compute_result:
2820 sh->check_state = check_state_idle;
2822 /* check that a write has not made the stripe insync */
2823 if (test_bit(STRIPE_INSYNC, &sh->state))
2826 /* now write out any block on a failed drive,
2827 * or P or Q if they were recomputed
2829 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2830 if (s->failed == 2) {
2831 dev = &sh->dev[s->failed_num[1]];
2833 set_bit(R5_LOCKED, &dev->flags);
2834 set_bit(R5_Wantwrite, &dev->flags);
2836 if (s->failed >= 1) {
2837 dev = &sh->dev[s->failed_num[0]];
2839 set_bit(R5_LOCKED, &dev->flags);
2840 set_bit(R5_Wantwrite, &dev->flags);
2842 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2843 dev = &sh->dev[pd_idx];
2845 set_bit(R5_LOCKED, &dev->flags);
2846 set_bit(R5_Wantwrite, &dev->flags);
2848 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2849 dev = &sh->dev[qd_idx];
2851 set_bit(R5_LOCKED, &dev->flags);
2852 set_bit(R5_Wantwrite, &dev->flags);
2854 clear_bit(STRIPE_DEGRADED, &sh->state);
2856 set_bit(STRIPE_INSYNC, &sh->state);
2858 case check_state_run:
2859 case check_state_run_q:
2860 case check_state_run_pq:
2861 break; /* we will be called again upon completion */
2862 case check_state_check_result:
2863 sh->check_state = check_state_idle;
2865 /* handle a successful check operation, if parity is correct
2866 * we are done. Otherwise update the mismatch count and repair
2867 * parity if !MD_RECOVERY_CHECK
2869 if (sh->ops.zero_sum_result == 0) {
2870 /* both parities are correct */
2872 set_bit(STRIPE_INSYNC, &sh->state);
2874 /* in contrast to the raid5 case we can validate
2875 * parity, but still have a failure to write
2878 sh->check_state = check_state_compute_result;
2879 /* Returning at this point means that we may go
2880 * off and bring p and/or q uptodate again so
2881 * we make sure to check zero_sum_result again
2882 * to verify if p or q need writeback
2886 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2887 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2888 /* don't try to repair!! */
2889 set_bit(STRIPE_INSYNC, &sh->state);
2891 int *target = &sh->ops.target;
2893 sh->ops.target = -1;
2894 sh->ops.target2 = -1;
2895 sh->check_state = check_state_compute_run;
2896 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2897 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2898 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2899 set_bit(R5_Wantcompute,
2900 &sh->dev[pd_idx].flags);
2902 target = &sh->ops.target2;
2905 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2906 set_bit(R5_Wantcompute,
2907 &sh->dev[qd_idx].flags);
2914 case check_state_compute_run:
2917 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2918 __func__, sh->check_state,
2919 (unsigned long long) sh->sector);
2924 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh)
2928 /* We have read all the blocks in this stripe and now we need to
2929 * copy some of them into a target stripe for expand.
2931 struct dma_async_tx_descriptor *tx = NULL;
2932 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2933 for (i = 0; i < sh->disks; i++)
2934 if (i != sh->pd_idx && i != sh->qd_idx) {
2936 struct stripe_head *sh2;
2937 struct async_submit_ctl submit;
2939 sector_t bn = compute_blocknr(sh, i, 1);
2940 sector_t s = raid5_compute_sector(conf, bn, 0,
2942 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2944 /* so far only the early blocks of this stripe
2945 * have been requested. When later blocks
2946 * get requested, we will try again
2949 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2950 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2951 /* must have already done this block */
2952 release_stripe(sh2);
2956 /* place all the copies on one channel */
2957 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2958 tx = async_memcpy(sh2->dev[dd_idx].page,
2959 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2962 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2963 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2964 for (j = 0; j < conf->raid_disks; j++)
2965 if (j != sh2->pd_idx &&
2967 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2969 if (j == conf->raid_disks) {
2970 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2971 set_bit(STRIPE_HANDLE, &sh2->state);
2973 release_stripe(sh2);
2976 /* done submitting copies, wait for them to complete */
2979 dma_wait_for_async_tx(tx);
2985 * handle_stripe - do things to a stripe.
2987 * We lock the stripe and then examine the state of various bits
2988 * to see what needs to be done.
2990 * return some read request which now have data
2991 * return some write requests which are safely on disc
2992 * schedule a read on some buffers
2993 * schedule a write of some buffers
2994 * return confirmation of parity correctness
2996 * buffers are taken off read_list or write_list, and bh_cache buffers
2997 * get BH_Lock set before the stripe lock is released.
3001 static int handle_stripe5(struct stripe_head *sh, struct stripe_head_state *s)
3003 raid5_conf_t *conf = sh->raid_conf;
3004 int disks = sh->disks, i;
3008 /* Now to look around and see what can be done */
3010 spin_lock_irq(&conf->device_lock);
3011 for (i=disks; i--; ) {
3016 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
3017 "written %p\n", i, dev->flags, dev->toread, dev->read,
3018 dev->towrite, dev->written);
3020 /* maybe we can request a biofill operation
3022 * new wantfill requests are only permitted while
3023 * ops_complete_biofill is guaranteed to be inactive
3025 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3026 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3027 set_bit(R5_Wantfill, &dev->flags);
3029 /* now count some things */
3030 if (test_bit(R5_LOCKED, &dev->flags))
3032 if (test_bit(R5_UPTODATE, &dev->flags))
3034 if (test_bit(R5_Wantcompute, &dev->flags))
3037 if (test_bit(R5_Wantfill, &dev->flags))
3039 else if (dev->toread)
3043 if (!test_bit(R5_OVERWRITE, &dev->flags))
3048 rdev = rcu_dereference(conf->disks[i].rdev);
3049 if (s->blocked_rdev == NULL &&
3050 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3051 s->blocked_rdev = rdev;
3052 atomic_inc(&rdev->nr_pending);
3054 clear_bit(R5_Insync, &dev->flags);
3057 else if (test_bit(In_sync, &rdev->flags))
3058 set_bit(R5_Insync, &dev->flags);
3060 /* could be in-sync depending on recovery/reshape status */
3061 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3062 set_bit(R5_Insync, &dev->flags);
3064 if (!test_bit(R5_Insync, &dev->flags)) {
3065 /* The ReadError flag will just be confusing now */
3066 clear_bit(R5_ReadError, &dev->flags);
3067 clear_bit(R5_ReWrite, &dev->flags);
3069 if (test_bit(R5_ReadError, &dev->flags))
3070 clear_bit(R5_Insync, &dev->flags);
3071 if (!test_bit(R5_Insync, &dev->flags)) {
3073 s->failed_num[0] = i;
3076 spin_unlock_irq(&conf->device_lock);
3079 if (unlikely(s->blocked_rdev)) {
3080 if (s->syncing || s->expanding || s->expanded ||
3081 s->to_write || s->written) {
3082 set_bit(STRIPE_HANDLE, &sh->state);
3085 /* There is nothing for the blocked_rdev to block */
3086 rdev_dec_pending(s->blocked_rdev, conf->mddev);
3087 s->blocked_rdev = NULL;
3090 if (s->to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3091 set_bit(STRIPE_OP_BIOFILL, &s->ops_request);
3092 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3095 pr_debug("locked=%d uptodate=%d to_read=%d"
3096 " to_write=%d failed=%d failed_num=%d\n",
3097 s->locked, s->uptodate, s->to_read, s->to_write,
3098 s->failed, s->failed_num[0]);
3099 /* check if the array has lost two devices and, if so, some requests might
3102 if (s->failed > 1 && s->to_read+s->to_write+s->written)
3103 handle_failed_stripe(conf, sh, s, disks, &s->return_bi);
3104 if (s->failed > 1 && s->syncing) {
3105 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3106 clear_bit(STRIPE_SYNCING, &sh->state);
3110 /* might be able to return some write requests if the parity block
3111 * is safe, or on a failed drive
3113 dev = &sh->dev[sh->pd_idx];
3115 ((test_bit(R5_Insync, &dev->flags) &&
3116 !test_bit(R5_LOCKED, &dev->flags) &&
3117 test_bit(R5_UPTODATE, &dev->flags)) ||
3118 (s->failed == 1 && s->failed_num[0] == sh->pd_idx)))
3119 handle_stripe_clean_event(conf, sh, disks, &s->return_bi);
3121 /* Now we might consider reading some blocks, either to check/generate
3122 * parity, or to satisfy requests
3123 * or to load a block that is being partially written.
3125 if (s->to_read || s->non_overwrite ||
3126 (s->syncing && (s->uptodate + s->compute < disks)) || s->expanding)
3127 handle_stripe_fill5(sh, s, disks);
3129 /* Now we check to see if any write operations have recently
3133 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3135 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3136 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3137 sh->reconstruct_state = reconstruct_state_idle;
3139 /* All the 'written' buffers and the parity block are ready to
3140 * be written back to disk
3142 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3143 for (i = disks; i--; ) {
3145 if (test_bit(R5_LOCKED, &dev->flags) &&
3146 (i == sh->pd_idx || dev->written)) {
3147 pr_debug("Writing block %d\n", i);
3148 set_bit(R5_Wantwrite, &dev->flags);
3151 if (!test_bit(R5_Insync, &dev->flags) ||
3152 (i == sh->pd_idx && s->failed == 0))
3153 set_bit(STRIPE_INSYNC, &sh->state);
3156 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3157 s->dec_preread_active = 1;
3160 /* Now to consider new write requests and what else, if anything
3161 * should be read. We do not handle new writes when:
3162 * 1/ A 'write' operation (copy+xor) is already in flight.
3163 * 2/ A 'check' operation is in flight, as it may clobber the parity
3166 if (s->to_write && !sh->reconstruct_state && !sh->check_state)
3167 handle_stripe_dirtying5(conf, sh, s, disks);
3169 /* maybe we need to check and possibly fix the parity for this stripe
3170 * Any reads will already have been scheduled, so we just see if enough
3171 * data is available. The parity check is held off while parity
3172 * dependent operations are in flight.
3174 if (sh->check_state ||
3175 (s->syncing && s->locked == 0 &&
3176 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3177 !test_bit(STRIPE_INSYNC, &sh->state)))
3178 handle_parity_checks5(conf, sh, s, disks);
3182 static int handle_stripe6(struct stripe_head *sh, struct stripe_head_state *s)
3184 raid5_conf_t *conf = sh->raid_conf;
3185 int disks = sh->disks;
3186 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3187 struct r5dev *dev, *pdev, *qdev;
3189 /* Now to look around and see what can be done */
3192 spin_lock_irq(&conf->device_lock);
3193 for (i=disks; i--; ) {
3197 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3198 i, dev->flags, dev->toread, dev->towrite, dev->written);
3199 /* maybe we can reply to a read
3201 * new wantfill requests are only permitted while
3202 * ops_complete_biofill is guaranteed to be inactive
3204 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3205 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3206 set_bit(R5_Wantfill, &dev->flags);
3208 /* now count some things */
3209 if (test_bit(R5_LOCKED, &dev->flags))
3211 if (test_bit(R5_UPTODATE, &dev->flags))
3213 if (test_bit(R5_Wantcompute, &dev->flags)) {
3215 BUG_ON(s->compute > 2);
3218 if (test_bit(R5_Wantfill, &dev->flags)) {
3220 } else if (dev->toread)
3224 if (!test_bit(R5_OVERWRITE, &dev->flags))
3229 rdev = rcu_dereference(conf->disks[i].rdev);
3230 if (s->blocked_rdev == NULL &&
3231 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3232 s->blocked_rdev = rdev;
3233 atomic_inc(&rdev->nr_pending);
3235 clear_bit(R5_Insync, &dev->flags);
3238 else if (test_bit(In_sync, &rdev->flags))
3239 set_bit(R5_Insync, &dev->flags);
3241 /* in sync if before recovery_offset */
3242 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3243 set_bit(R5_Insync, &dev->flags);
3245 if (!test_bit(R5_Insync, &dev->flags)) {
3246 /* The ReadError flag will just be confusing now */
3247 clear_bit(R5_ReadError, &dev->flags);
3248 clear_bit(R5_ReWrite, &dev->flags);
3250 if (test_bit(R5_ReadError, &dev->flags))
3251 clear_bit(R5_Insync, &dev->flags);
3252 if (!test_bit(R5_Insync, &dev->flags)) {
3254 s->failed_num[s->failed] = i;
3258 spin_unlock_irq(&conf->device_lock);
3261 if (unlikely(s->blocked_rdev)) {
3262 if (s->syncing || s->expanding || s->expanded ||
3263 s->to_write || s->written) {
3264 set_bit(STRIPE_HANDLE, &sh->state);
3267 /* There is nothing for the blocked_rdev to block */
3268 rdev_dec_pending(s->blocked_rdev, conf->mddev);
3269 s->blocked_rdev = NULL;
3272 if (s->to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3273 set_bit(STRIPE_OP_BIOFILL, &s->ops_request);
3274 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3277 pr_debug("locked=%d uptodate=%d to_read=%d"
3278 " to_write=%d failed=%d failed_num=%d,%d\n",
3279 s->locked, s->uptodate, s->to_read, s->to_write, s->failed,
3280 s->failed_num[0], s->failed_num[1]);
3281 /* check if the array has lost >2 devices and, if so, some requests
3282 * might need to be failed
3284 if (s->failed > 2 && s->to_read+s->to_write+s->written)
3285 handle_failed_stripe(conf, sh, s, disks, &s->return_bi);
3286 if (s->failed > 2 && s->syncing) {
3287 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3288 clear_bit(STRIPE_SYNCING, &sh->state);
3293 * might be able to return some write requests if the parity blocks
3294 * are safe, or on a failed drive
3296 pdev = &sh->dev[pd_idx];
3297 s->p_failed = (s->failed >= 1 && s->failed_num[0] == pd_idx)
3298 || (s->failed >= 2 && s->failed_num[1] == pd_idx);
3299 qdev = &sh->dev[qd_idx];
3300 s->q_failed = (s->failed >= 1 && s->failed_num[0] == qd_idx)
3301 || (s->failed >= 2 && s->failed_num[1] == qd_idx);
3304 (s->p_failed || ((test_bit(R5_Insync, &pdev->flags)
3305 && !test_bit(R5_LOCKED, &pdev->flags)
3306 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3307 (s->q_failed || ((test_bit(R5_Insync, &qdev->flags)
3308 && !test_bit(R5_LOCKED, &qdev->flags)
3309 && test_bit(R5_UPTODATE, &qdev->flags)))))
3310 handle_stripe_clean_event(conf, sh, disks, &s->return_bi);
3312 /* Now we might consider reading some blocks, either to check/generate
3313 * parity, or to satisfy requests
3314 * or to load a block that is being partially written.
3316 if (s->to_read || s->non_overwrite || (s->to_write && s->failed) ||
3317 (s->syncing && (s->uptodate + s->compute < disks)) || s->expanding)
3318 handle_stripe_fill6(sh, s, disks);
3320 /* Now we check to see if any write operations have recently
3323 if (sh->reconstruct_state == reconstruct_state_drain_result) {
3325 sh->reconstruct_state = reconstruct_state_idle;
3326 /* All the 'written' buffers and the parity blocks are ready to
3327 * be written back to disk
3329 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3330 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3331 for (i = disks; i--; ) {
3333 if (test_bit(R5_LOCKED, &dev->flags) &&
3334 (i == sh->pd_idx || i == qd_idx ||
3336 pr_debug("Writing block %d\n", i);
3337 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3338 set_bit(R5_Wantwrite, &dev->flags);
3339 if (!test_bit(R5_Insync, &dev->flags) ||
3340 ((i == sh->pd_idx || i == qd_idx) &&
3342 set_bit(STRIPE_INSYNC, &sh->state);
3345 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3346 s->dec_preread_active = 1;
3349 /* Now to consider new write requests and what else, if anything
3350 * should be read. We do not handle new writes when:
3351 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3352 * 2/ A 'check' operation is in flight, as it may clobber the parity
3355 if (s->to_write && !sh->reconstruct_state && !sh->check_state)
3356 handle_stripe_dirtying6(conf, sh, s, disks);
3358 /* maybe we need to check and possibly fix the parity for this stripe
3359 * Any reads will already have been scheduled, so we just see if enough
3360 * data is available. The parity check is held off while parity
3361 * dependent operations are in flight.
3363 if (sh->check_state ||
3364 (s->syncing && s->locked == 0 &&
3365 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3366 !test_bit(STRIPE_INSYNC, &sh->state)))
3367 handle_parity_checks6(conf, sh, s, disks);
3371 static void handle_stripe(struct stripe_head *sh)
3373 struct stripe_head_state s;
3376 raid5_conf_t *conf = sh->raid_conf;
3378 clear_bit(STRIPE_HANDLE, &sh->state);
3379 if (test_and_set_bit(STRIPE_ACTIVE, &sh->state)) {
3380 /* already being handled, ensure it gets handled
3381 * again when current action finishes */
3382 set_bit(STRIPE_HANDLE, &sh->state);
3386 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3387 set_bit(STRIPE_SYNCING, &sh->state);
3388 clear_bit(STRIPE_INSYNC, &sh->state);
3390 clear_bit(STRIPE_DELAYED, &sh->state);
3392 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3393 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3394 (unsigned long long)sh->sector, sh->state,
3395 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3396 sh->check_state, sh->reconstruct_state);
3397 memset(&s, 0, sizeof(s));
3399 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3400 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3401 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3403 if (conf->level == 6)
3404 done = handle_stripe6(sh, &s);
3406 done = handle_stripe5(sh, &s);
3412 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3413 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3414 clear_bit(STRIPE_SYNCING, &sh->state);
3417 /* If the failed drives are just a ReadError, then we might need
3418 * to progress the repair/check process
3420 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3421 for (i = 0; i < s.failed; i++) {
3422 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3423 if (test_bit(R5_ReadError, &dev->flags)
3424 && !test_bit(R5_LOCKED, &dev->flags)
3425 && test_bit(R5_UPTODATE, &dev->flags)
3427 if (!test_bit(R5_ReWrite, &dev->flags)) {
3428 set_bit(R5_Wantwrite, &dev->flags);
3429 set_bit(R5_ReWrite, &dev->flags);
3430 set_bit(R5_LOCKED, &dev->flags);
3433 /* let's read it back */
3434 set_bit(R5_Wantread, &dev->flags);
3435 set_bit(R5_LOCKED, &dev->flags);
3442 /* Finish reconstruct operations initiated by the expansion process */
3443 if (sh->reconstruct_state == reconstruct_state_result) {
3444 struct stripe_head *sh_src
3445 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3446 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3447 /* sh cannot be written until sh_src has been read.
3448 * so arrange for sh to be delayed a little
3450 set_bit(STRIPE_DELAYED, &sh->state);
3451 set_bit(STRIPE_HANDLE, &sh->state);
3452 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3454 atomic_inc(&conf->preread_active_stripes);
3455 release_stripe(sh_src);
3459 release_stripe(sh_src);
3461 sh->reconstruct_state = reconstruct_state_idle;
3462 clear_bit(STRIPE_EXPANDING, &sh->state);
3463 for (i = conf->raid_disks; i--; ) {
3464 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3465 set_bit(R5_LOCKED, &sh->dev[i].flags);
3470 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3471 !sh->reconstruct_state) {
3472 /* Need to write out all blocks after computing parity */
3473 sh->disks = conf->raid_disks;
3474 stripe_set_idx(sh->sector, conf, 0, sh);
3475 schedule_reconstruction(sh, &s, 1, 1);
3476 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3477 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3478 atomic_dec(&conf->reshape_stripes);
3479 wake_up(&conf->wait_for_overlap);
3480 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3483 if (s.expanding && s.locked == 0 &&
3484 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3485 handle_stripe_expansion(conf, sh);
3488 /* wait for this device to become unblocked */
3489 if (unlikely(s.blocked_rdev))
3490 md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3493 raid_run_ops(sh, s.ops_request);
3498 if (s.dec_preread_active) {
3499 /* We delay this until after ops_run_io so that if make_request
3500 * is waiting on a flush, it won't continue until the writes
3501 * have actually been submitted.
3503 atomic_dec(&conf->preread_active_stripes);
3504 if (atomic_read(&conf->preread_active_stripes) <
3506 md_wakeup_thread(conf->mddev->thread);
3509 return_io(s.return_bi);
3511 clear_bit(STRIPE_ACTIVE, &sh->state);
3514 static void raid5_activate_delayed(raid5_conf_t *conf)
3516 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3517 while (!list_empty(&conf->delayed_list)) {
3518 struct list_head *l = conf->delayed_list.next;
3519 struct stripe_head *sh;
3520 sh = list_entry(l, struct stripe_head, lru);
3522 clear_bit(STRIPE_DELAYED, &sh->state);
3523 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3524 atomic_inc(&conf->preread_active_stripes);
3525 list_add_tail(&sh->lru, &conf->hold_list);
3530 static void activate_bit_delay(raid5_conf_t *conf)
3532 /* device_lock is held */
3533 struct list_head head;
3534 list_add(&head, &conf->bitmap_list);
3535 list_del_init(&conf->bitmap_list);
3536 while (!list_empty(&head)) {
3537 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3538 list_del_init(&sh->lru);
3539 atomic_inc(&sh->count);
3540 __release_stripe(conf, sh);
3544 int md_raid5_congested(mddev_t *mddev, int bits)
3546 raid5_conf_t *conf = mddev->private;
3548 /* No difference between reads and writes. Just check
3549 * how busy the stripe_cache is
3552 if (conf->inactive_blocked)
3556 if (list_empty_careful(&conf->inactive_list))
3561 EXPORT_SYMBOL_GPL(md_raid5_congested);
3563 static int raid5_congested(void *data, int bits)
3565 mddev_t *mddev = data;
3567 return mddev_congested(mddev, bits) ||
3568 md_raid5_congested(mddev, bits);
3571 /* We want read requests to align with chunks where possible,
3572 * but write requests don't need to.
3574 static int raid5_mergeable_bvec(struct request_queue *q,
3575 struct bvec_merge_data *bvm,
3576 struct bio_vec *biovec)
3578 mddev_t *mddev = q->queuedata;
3579 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3581 unsigned int chunk_sectors = mddev->chunk_sectors;
3582 unsigned int bio_sectors = bvm->bi_size >> 9;
3584 if ((bvm->bi_rw & 1) == WRITE)
3585 return biovec->bv_len; /* always allow writes to be mergeable */
3587 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3588 chunk_sectors = mddev->new_chunk_sectors;
3589 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3590 if (max < 0) max = 0;
3591 if (max <= biovec->bv_len && bio_sectors == 0)
3592 return biovec->bv_len;
3598 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3600 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3601 unsigned int chunk_sectors = mddev->chunk_sectors;
3602 unsigned int bio_sectors = bio->bi_size >> 9;
3604 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3605 chunk_sectors = mddev->new_chunk_sectors;
3606 return chunk_sectors >=
3607 ((sector & (chunk_sectors - 1)) + bio_sectors);
3611 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3612 * later sampled by raid5d.
3614 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3616 unsigned long flags;
3618 spin_lock_irqsave(&conf->device_lock, flags);
3620 bi->bi_next = conf->retry_read_aligned_list;
3621 conf->retry_read_aligned_list = bi;
3623 spin_unlock_irqrestore(&conf->device_lock, flags);
3624 md_wakeup_thread(conf->mddev->thread);
3628 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3632 bi = conf->retry_read_aligned;
3634 conf->retry_read_aligned = NULL;
3637 bi = conf->retry_read_aligned_list;
3639 conf->retry_read_aligned_list = bi->bi_next;
3642 * this sets the active strip count to 1 and the processed
3643 * strip count to zero (upper 8 bits)
3645 bi->bi_phys_segments = 1; /* biased count of active stripes */
3653 * The "raid5_align_endio" should check if the read succeeded and if it
3654 * did, call bio_endio on the original bio (having bio_put the new bio
3656 * If the read failed..
3658 static void raid5_align_endio(struct bio *bi, int error)
3660 struct bio* raid_bi = bi->bi_private;
3663 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3668 rdev = (void*)raid_bi->bi_next;
3669 raid_bi->bi_next = NULL;
3670 mddev = rdev->mddev;
3671 conf = mddev->private;
3673 rdev_dec_pending(rdev, conf->mddev);
3675 if (!error && uptodate) {
3676 bio_endio(raid_bi, 0);
3677 if (atomic_dec_and_test(&conf->active_aligned_reads))
3678 wake_up(&conf->wait_for_stripe);
3683 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3685 add_bio_to_retry(raid_bi, conf);
3688 static int bio_fits_rdev(struct bio *bi)
3690 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3692 if ((bi->bi_size>>9) > queue_max_sectors(q))
3694 blk_recount_segments(q, bi);
3695 if (bi->bi_phys_segments > queue_max_segments(q))
3698 if (q->merge_bvec_fn)
3699 /* it's too hard to apply the merge_bvec_fn at this stage,
3708 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio)
3710 raid5_conf_t *conf = mddev->private;
3712 struct bio* align_bi;
3715 if (!in_chunk_boundary(mddev, raid_bio)) {
3716 pr_debug("chunk_aligned_read : non aligned\n");
3720 * use bio_clone_mddev to make a copy of the bio
3722 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3726 * set bi_end_io to a new function, and set bi_private to the
3729 align_bi->bi_end_io = raid5_align_endio;
3730 align_bi->bi_private = raid_bio;
3734 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3739 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3740 if (rdev && test_bit(In_sync, &rdev->flags)) {
3741 atomic_inc(&rdev->nr_pending);
3743 raid_bio->bi_next = (void*)rdev;
3744 align_bi->bi_bdev = rdev->bdev;
3745 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3746 align_bi->bi_sector += rdev->data_offset;
3748 if (!bio_fits_rdev(align_bi)) {
3749 /* too big in some way */
3751 rdev_dec_pending(rdev, mddev);
3755 spin_lock_irq(&conf->device_lock);
3756 wait_event_lock_irq(conf->wait_for_stripe,
3758 conf->device_lock, /* nothing */);
3759 atomic_inc(&conf->active_aligned_reads);
3760 spin_unlock_irq(&conf->device_lock);
3762 generic_make_request(align_bi);
3771 /* __get_priority_stripe - get the next stripe to process
3773 * Full stripe writes are allowed to pass preread active stripes up until
3774 * the bypass_threshold is exceeded. In general the bypass_count
3775 * increments when the handle_list is handled before the hold_list; however, it
3776 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3777 * stripe with in flight i/o. The bypass_count will be reset when the
3778 * head of the hold_list has changed, i.e. the head was promoted to the
3781 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3783 struct stripe_head *sh;
3785 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3787 list_empty(&conf->handle_list) ? "empty" : "busy",
3788 list_empty(&conf->hold_list) ? "empty" : "busy",
3789 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3791 if (!list_empty(&conf->handle_list)) {
3792 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3794 if (list_empty(&conf->hold_list))
3795 conf->bypass_count = 0;
3796 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3797 if (conf->hold_list.next == conf->last_hold)
3798 conf->bypass_count++;
3800 conf->last_hold = conf->hold_list.next;
3801 conf->bypass_count -= conf->bypass_threshold;
3802 if (conf->bypass_count < 0)
3803 conf->bypass_count = 0;
3806 } else if (!list_empty(&conf->hold_list) &&
3807 ((conf->bypass_threshold &&
3808 conf->bypass_count > conf->bypass_threshold) ||
3809 atomic_read(&conf->pending_full_writes) == 0)) {
3810 sh = list_entry(conf->hold_list.next,
3812 conf->bypass_count -= conf->bypass_threshold;
3813 if (conf->bypass_count < 0)
3814 conf->bypass_count = 0;
3818 list_del_init(&sh->lru);
3819 atomic_inc(&sh->count);
3820 BUG_ON(atomic_read(&sh->count) != 1);
3824 static int make_request(mddev_t *mddev, struct bio * bi)
3826 raid5_conf_t *conf = mddev->private;
3828 sector_t new_sector;
3829 sector_t logical_sector, last_sector;
3830 struct stripe_head *sh;
3831 const int rw = bio_data_dir(bi);
3835 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3836 md_flush_request(mddev, bi);
3840 md_write_start(mddev, bi);
3843 mddev->reshape_position == MaxSector &&
3844 chunk_aligned_read(mddev,bi))
3847 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3848 last_sector = bi->bi_sector + (bi->bi_size>>9);
3850 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3852 plugged = mddev_check_plugged(mddev);
3853 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3855 int disks, data_disks;
3860 disks = conf->raid_disks;
3861 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3862 if (unlikely(conf->reshape_progress != MaxSector)) {
3863 /* spinlock is needed as reshape_progress may be
3864 * 64bit on a 32bit platform, and so it might be
3865 * possible to see a half-updated value
3866 * Of course reshape_progress could change after
3867 * the lock is dropped, so once we get a reference
3868 * to the stripe that we think it is, we will have
3871 spin_lock_irq(&conf->device_lock);
3872 if (mddev->delta_disks < 0
3873 ? logical_sector < conf->reshape_progress
3874 : logical_sector >= conf->reshape_progress) {
3875 disks = conf->previous_raid_disks;
3878 if (mddev->delta_disks < 0
3879 ? logical_sector < conf->reshape_safe
3880 : logical_sector >= conf->reshape_safe) {
3881 spin_unlock_irq(&conf->device_lock);
3886 spin_unlock_irq(&conf->device_lock);
3888 data_disks = disks - conf->max_degraded;
3890 new_sector = raid5_compute_sector(conf, logical_sector,
3893 pr_debug("raid456: make_request, sector %llu logical %llu\n",
3894 (unsigned long long)new_sector,
3895 (unsigned long long)logical_sector);
3897 sh = get_active_stripe(conf, new_sector, previous,
3898 (bi->bi_rw&RWA_MASK), 0);
3900 if (unlikely(previous)) {
3901 /* expansion might have moved on while waiting for a
3902 * stripe, so we must do the range check again.
3903 * Expansion could still move past after this
3904 * test, but as we are holding a reference to
3905 * 'sh', we know that if that happens,
3906 * STRIPE_EXPANDING will get set and the expansion
3907 * won't proceed until we finish with the stripe.
3910 spin_lock_irq(&conf->device_lock);
3911 if (mddev->delta_disks < 0
3912 ? logical_sector >= conf->reshape_progress
3913 : logical_sector < conf->reshape_progress)
3914 /* mismatch, need to try again */
3916 spin_unlock_irq(&conf->device_lock);
3925 logical_sector >= mddev->suspend_lo &&
3926 logical_sector < mddev->suspend_hi) {
3928 /* As the suspend_* range is controlled by
3929 * userspace, we want an interruptible
3932 flush_signals(current);
3933 prepare_to_wait(&conf->wait_for_overlap,
3934 &w, TASK_INTERRUPTIBLE);
3935 if (logical_sector >= mddev->suspend_lo &&
3936 logical_sector < mddev->suspend_hi)
3941 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3942 !add_stripe_bio(sh, bi, dd_idx, rw)) {
3943 /* Stripe is busy expanding or
3944 * add failed due to overlap. Flush everything
3947 md_wakeup_thread(mddev->thread);
3952 finish_wait(&conf->wait_for_overlap, &w);
3953 set_bit(STRIPE_HANDLE, &sh->state);
3954 clear_bit(STRIPE_DELAYED, &sh->state);
3955 if ((bi->bi_rw & REQ_SYNC) &&
3956 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3957 atomic_inc(&conf->preread_active_stripes);
3960 /* cannot get stripe for read-ahead, just give-up */
3961 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3962 finish_wait(&conf->wait_for_overlap, &w);
3968 md_wakeup_thread(mddev->thread);
3970 spin_lock_irq(&conf->device_lock);
3971 remaining = raid5_dec_bi_phys_segments(bi);
3972 spin_unlock_irq(&conf->device_lock);
3973 if (remaining == 0) {
3976 md_write_end(mddev);
3984 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3986 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3988 /* reshaping is quite different to recovery/resync so it is
3989 * handled quite separately ... here.
3991 * On each call to sync_request, we gather one chunk worth of
3992 * destination stripes and flag them as expanding.
3993 * Then we find all the source stripes and request reads.
3994 * As the reads complete, handle_stripe will copy the data
3995 * into the destination stripe and release that stripe.
3997 raid5_conf_t *conf = mddev->private;
3998 struct stripe_head *sh;
3999 sector_t first_sector, last_sector;
4000 int raid_disks = conf->previous_raid_disks;
4001 int data_disks = raid_disks - conf->max_degraded;
4002 int new_data_disks = conf->raid_disks - conf->max_degraded;
4005 sector_t writepos, readpos, safepos;
4006 sector_t stripe_addr;
4007 int reshape_sectors;
4008 struct list_head stripes;
4010 if (sector_nr == 0) {
4011 /* If restarting in the middle, skip the initial sectors */
4012 if (mddev->delta_disks < 0 &&
4013 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4014 sector_nr = raid5_size(mddev, 0, 0)
4015 - conf->reshape_progress;
4016 } else if (mddev->delta_disks >= 0 &&
4017 conf->reshape_progress > 0)
4018 sector_nr = conf->reshape_progress;
4019 sector_div(sector_nr, new_data_disks);
4021 mddev->curr_resync_completed = sector_nr;
4022 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4028 /* We need to process a full chunk at a time.
4029 * If old and new chunk sizes differ, we need to process the
4032 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4033 reshape_sectors = mddev->new_chunk_sectors;
4035 reshape_sectors = mddev->chunk_sectors;
4037 /* we update the metadata when there is more than 3Meg
4038 * in the block range (that is rather arbitrary, should
4039 * probably be time based) or when the data about to be
4040 * copied would over-write the source of the data at
4041 * the front of the range.
4042 * i.e. one new_stripe along from reshape_progress new_maps
4043 * to after where reshape_safe old_maps to
4045 writepos = conf->reshape_progress;
4046 sector_div(writepos, new_data_disks);
4047 readpos = conf->reshape_progress;
4048 sector_div(readpos, data_disks);
4049 safepos = conf->reshape_safe;
4050 sector_div(safepos, data_disks);
4051 if (mddev->delta_disks < 0) {
4052 writepos -= min_t(sector_t, reshape_sectors, writepos);
4053 readpos += reshape_sectors;
4054 safepos += reshape_sectors;
4056 writepos += reshape_sectors;
4057 readpos -= min_t(sector_t, reshape_sectors, readpos);
4058 safepos -= min_t(sector_t, reshape_sectors, safepos);
4061 /* 'writepos' is the most advanced device address we might write.
4062 * 'readpos' is the least advanced device address we might read.
4063 * 'safepos' is the least address recorded in the metadata as having
4065 * If 'readpos' is behind 'writepos', then there is no way that we can
4066 * ensure safety in the face of a crash - that must be done by userspace
4067 * making a backup of the data. So in that case there is no particular
4068 * rush to update metadata.
4069 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4070 * update the metadata to advance 'safepos' to match 'readpos' so that
4071 * we can be safe in the event of a crash.
4072 * So we insist on updating metadata if safepos is behind writepos and
4073 * readpos is beyond writepos.
4074 * In any case, update the metadata every 10 seconds.
4075 * Maybe that number should be configurable, but I'm not sure it is
4076 * worth it.... maybe it could be a multiple of safemode_delay???
4078 if ((mddev->delta_disks < 0
4079 ? (safepos > writepos && readpos < writepos)
4080 : (safepos < writepos && readpos > writepos)) ||
4081 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4082 /* Cannot proceed until we've updated the superblock... */
4083 wait_event(conf->wait_for_overlap,
4084 atomic_read(&conf->reshape_stripes)==0);
4085 mddev->reshape_position = conf->reshape_progress;
4086 mddev->curr_resync_completed = sector_nr;
4087 conf->reshape_checkpoint = jiffies;
4088 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4089 md_wakeup_thread(mddev->thread);
4090 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4091 kthread_should_stop());
4092 spin_lock_irq(&conf->device_lock);
4093 conf->reshape_safe = mddev->reshape_position;
4094 spin_unlock_irq(&conf->device_lock);
4095 wake_up(&conf->wait_for_overlap);
4096 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4099 if (mddev->delta_disks < 0) {
4100 BUG_ON(conf->reshape_progress == 0);
4101 stripe_addr = writepos;
4102 BUG_ON((mddev->dev_sectors &
4103 ~((sector_t)reshape_sectors - 1))
4104 - reshape_sectors - stripe_addr
4107 BUG_ON(writepos != sector_nr + reshape_sectors);
4108 stripe_addr = sector_nr;
4110 INIT_LIST_HEAD(&stripes);
4111 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4113 int skipped_disk = 0;
4114 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4115 set_bit(STRIPE_EXPANDING, &sh->state);
4116 atomic_inc(&conf->reshape_stripes);
4117 /* If any of this stripe is beyond the end of the old
4118 * array, then we need to zero those blocks
4120 for (j=sh->disks; j--;) {
4122 if (j == sh->pd_idx)
4124 if (conf->level == 6 &&
4127 s = compute_blocknr(sh, j, 0);
4128 if (s < raid5_size(mddev, 0, 0)) {
4132 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4133 set_bit(R5_Expanded, &sh->dev[j].flags);
4134 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4136 if (!skipped_disk) {
4137 set_bit(STRIPE_EXPAND_READY, &sh->state);
4138 set_bit(STRIPE_HANDLE, &sh->state);
4140 list_add(&sh->lru, &stripes);
4142 spin_lock_irq(&conf->device_lock);
4143 if (mddev->delta_disks < 0)
4144 conf->reshape_progress -= reshape_sectors * new_data_disks;
4146 conf->reshape_progress += reshape_sectors * new_data_disks;
4147 spin_unlock_irq(&conf->device_lock);
4148 /* Ok, those stripe are ready. We can start scheduling
4149 * reads on the source stripes.
4150 * The source stripes are determined by mapping the first and last
4151 * block on the destination stripes.
4154 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4157 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4158 * new_data_disks - 1),
4160 if (last_sector >= mddev->dev_sectors)
4161 last_sector = mddev->dev_sectors - 1;
4162 while (first_sector <= last_sector) {
4163 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4164 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4165 set_bit(STRIPE_HANDLE, &sh->state);
4167 first_sector += STRIPE_SECTORS;
4169 /* Now that the sources are clearly marked, we can release
4170 * the destination stripes
4172 while (!list_empty(&stripes)) {
4173 sh = list_entry(stripes.next, struct stripe_head, lru);
4174 list_del_init(&sh->lru);
4177 /* If this takes us to the resync_max point where we have to pause,
4178 * then we need to write out the superblock.
4180 sector_nr += reshape_sectors;
4181 if ((sector_nr - mddev->curr_resync_completed) * 2
4182 >= mddev->resync_max - mddev->curr_resync_completed) {
4183 /* Cannot proceed until we've updated the superblock... */
4184 wait_event(conf->wait_for_overlap,
4185 atomic_read(&conf->reshape_stripes) == 0);
4186 mddev->reshape_position = conf->reshape_progress;
4187 mddev->curr_resync_completed = sector_nr;
4188 conf->reshape_checkpoint = jiffies;
4189 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4190 md_wakeup_thread(mddev->thread);
4191 wait_event(mddev->sb_wait,
4192 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4193 || kthread_should_stop());
4194 spin_lock_irq(&conf->device_lock);
4195 conf->reshape_safe = mddev->reshape_position;
4196 spin_unlock_irq(&conf->device_lock);
4197 wake_up(&conf->wait_for_overlap);
4198 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4200 return reshape_sectors;
4203 /* FIXME go_faster isn't used */
4204 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4206 raid5_conf_t *conf = mddev->private;
4207 struct stripe_head *sh;
4208 sector_t max_sector = mddev->dev_sectors;
4209 sector_t sync_blocks;
4210 int still_degraded = 0;
4213 if (sector_nr >= max_sector) {
4214 /* just being told to finish up .. nothing much to do */
4216 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4221 if (mddev->curr_resync < max_sector) /* aborted */
4222 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4224 else /* completed sync */
4226 bitmap_close_sync(mddev->bitmap);
4231 /* Allow raid5_quiesce to complete */
4232 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4234 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4235 return reshape_request(mddev, sector_nr, skipped);
4237 /* No need to check resync_max as we never do more than one
4238 * stripe, and as resync_max will always be on a chunk boundary,
4239 * if the check in md_do_sync didn't fire, there is no chance
4240 * of overstepping resync_max here
4243 /* if there is too many failed drives and we are trying
4244 * to resync, then assert that we are finished, because there is
4245 * nothing we can do.
4247 if (mddev->degraded >= conf->max_degraded &&
4248 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4249 sector_t rv = mddev->dev_sectors - sector_nr;
4253 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4254 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4255 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4256 /* we can skip this block, and probably more */
4257 sync_blocks /= STRIPE_SECTORS;
4259 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4263 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4265 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4267 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4268 /* make sure we don't swamp the stripe cache if someone else
4269 * is trying to get access
4271 schedule_timeout_uninterruptible(1);
4273 /* Need to check if array will still be degraded after recovery/resync
4274 * We don't need to check the 'failed' flag as when that gets set,
4277 for (i = 0; i < conf->raid_disks; i++)
4278 if (conf->disks[i].rdev == NULL)
4281 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4283 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4288 return STRIPE_SECTORS;
4291 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4293 /* We may not be able to submit a whole bio at once as there
4294 * may not be enough stripe_heads available.
4295 * We cannot pre-allocate enough stripe_heads as we may need
4296 * more than exist in the cache (if we allow ever large chunks).
4297 * So we do one stripe head at a time and record in
4298 * ->bi_hw_segments how many have been done.
4300 * We *know* that this entire raid_bio is in one chunk, so
4301 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4303 struct stripe_head *sh;
4305 sector_t sector, logical_sector, last_sector;
4310 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4311 sector = raid5_compute_sector(conf, logical_sector,
4313 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4315 for (; logical_sector < last_sector;
4316 logical_sector += STRIPE_SECTORS,
4317 sector += STRIPE_SECTORS,
4320 if (scnt < raid5_bi_hw_segments(raid_bio))
4321 /* already done this stripe */
4324 sh = get_active_stripe(conf, sector, 0, 1, 0);
4327 /* failed to get a stripe - must wait */
4328 raid5_set_bi_hw_segments(raid_bio, scnt);
4329 conf->retry_read_aligned = raid_bio;
4333 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4334 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4336 raid5_set_bi_hw_segments(raid_bio, scnt);
4337 conf->retry_read_aligned = raid_bio;
4345 spin_lock_irq(&conf->device_lock);
4346 remaining = raid5_dec_bi_phys_segments(raid_bio);
4347 spin_unlock_irq(&conf->device_lock);
4349 bio_endio(raid_bio, 0);
4350 if (atomic_dec_and_test(&conf->active_aligned_reads))
4351 wake_up(&conf->wait_for_stripe);
4357 * This is our raid5 kernel thread.
4359 * We scan the hash table for stripes which can be handled now.
4360 * During the scan, completed stripes are saved for us by the interrupt
4361 * handler, so that they will not have to wait for our next wakeup.
4363 static void raid5d(mddev_t *mddev)
4365 struct stripe_head *sh;
4366 raid5_conf_t *conf = mddev->private;
4368 struct blk_plug plug;
4370 pr_debug("+++ raid5d active\n");
4372 md_check_recovery(mddev);
4374 blk_start_plug(&plug);
4376 spin_lock_irq(&conf->device_lock);
4380 if (atomic_read(&mddev->plug_cnt) == 0 &&
4381 !list_empty(&conf->bitmap_list)) {
4382 /* Now is a good time to flush some bitmap updates */
4384 spin_unlock_irq(&conf->device_lock);
4385 bitmap_unplug(mddev->bitmap);
4386 spin_lock_irq(&conf->device_lock);
4387 conf->seq_write = conf->seq_flush;
4388 activate_bit_delay(conf);
4390 if (atomic_read(&mddev->plug_cnt) == 0)
4391 raid5_activate_delayed(conf);
4393 while ((bio = remove_bio_from_retry(conf))) {
4395 spin_unlock_irq(&conf->device_lock);
4396 ok = retry_aligned_read(conf, bio);
4397 spin_lock_irq(&conf->device_lock);
4403 sh = __get_priority_stripe(conf);
4407 spin_unlock_irq(&conf->device_lock);
4414 spin_lock_irq(&conf->device_lock);
4416 pr_debug("%d stripes handled\n", handled);
4418 spin_unlock_irq(&conf->device_lock);
4420 async_tx_issue_pending_all();
4421 blk_finish_plug(&plug);
4423 pr_debug("--- raid5d inactive\n");
4427 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4429 raid5_conf_t *conf = mddev->private;
4431 return sprintf(page, "%d\n", conf->max_nr_stripes);
4437 raid5_set_cache_size(mddev_t *mddev, int size)
4439 raid5_conf_t *conf = mddev->private;
4442 if (size <= 16 || size > 32768)
4444 while (size < conf->max_nr_stripes) {
4445 if (drop_one_stripe(conf))
4446 conf->max_nr_stripes--;
4450 err = md_allow_write(mddev);
4453 while (size > conf->max_nr_stripes) {
4454 if (grow_one_stripe(conf))
4455 conf->max_nr_stripes++;
4460 EXPORT_SYMBOL(raid5_set_cache_size);
4463 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4465 raid5_conf_t *conf = mddev->private;
4469 if (len >= PAGE_SIZE)
4474 if (strict_strtoul(page, 10, &new))
4476 err = raid5_set_cache_size(mddev, new);
4482 static struct md_sysfs_entry
4483 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4484 raid5_show_stripe_cache_size,
4485 raid5_store_stripe_cache_size);
4488 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4490 raid5_conf_t *conf = mddev->private;
4492 return sprintf(page, "%d\n", conf->bypass_threshold);
4498 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4500 raid5_conf_t *conf = mddev->private;
4502 if (len >= PAGE_SIZE)
4507 if (strict_strtoul(page, 10, &new))
4509 if (new > conf->max_nr_stripes)
4511 conf->bypass_threshold = new;
4515 static struct md_sysfs_entry
4516 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4518 raid5_show_preread_threshold,
4519 raid5_store_preread_threshold);
4522 stripe_cache_active_show(mddev_t *mddev, char *page)
4524 raid5_conf_t *conf = mddev->private;
4526 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4531 static struct md_sysfs_entry
4532 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4534 static struct attribute *raid5_attrs[] = {
4535 &raid5_stripecache_size.attr,
4536 &raid5_stripecache_active.attr,
4537 &raid5_preread_bypass_threshold.attr,
4540 static struct attribute_group raid5_attrs_group = {
4542 .attrs = raid5_attrs,
4546 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4548 raid5_conf_t *conf = mddev->private;
4551 sectors = mddev->dev_sectors;
4553 /* size is defined by the smallest of previous and new size */
4554 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4556 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4557 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4558 return sectors * (raid_disks - conf->max_degraded);
4561 static void raid5_free_percpu(raid5_conf_t *conf)
4563 struct raid5_percpu *percpu;
4570 for_each_possible_cpu(cpu) {
4571 percpu = per_cpu_ptr(conf->percpu, cpu);
4572 safe_put_page(percpu->spare_page);
4573 kfree(percpu->scribble);
4575 #ifdef CONFIG_HOTPLUG_CPU
4576 unregister_cpu_notifier(&conf->cpu_notify);
4580 free_percpu(conf->percpu);
4583 static void free_conf(raid5_conf_t *conf)
4585 shrink_stripes(conf);
4586 raid5_free_percpu(conf);
4588 kfree(conf->stripe_hashtbl);
4592 #ifdef CONFIG_HOTPLUG_CPU
4593 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4596 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4597 long cpu = (long)hcpu;
4598 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4601 case CPU_UP_PREPARE:
4602 case CPU_UP_PREPARE_FROZEN:
4603 if (conf->level == 6 && !percpu->spare_page)
4604 percpu->spare_page = alloc_page(GFP_KERNEL);
4605 if (!percpu->scribble)
4606 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4608 if (!percpu->scribble ||
4609 (conf->level == 6 && !percpu->spare_page)) {
4610 safe_put_page(percpu->spare_page);
4611 kfree(percpu->scribble);
4612 pr_err("%s: failed memory allocation for cpu%ld\n",
4614 return notifier_from_errno(-ENOMEM);
4618 case CPU_DEAD_FROZEN:
4619 safe_put_page(percpu->spare_page);
4620 kfree(percpu->scribble);
4621 percpu->spare_page = NULL;
4622 percpu->scribble = NULL;
4631 static int raid5_alloc_percpu(raid5_conf_t *conf)
4634 struct page *spare_page;
4635 struct raid5_percpu __percpu *allcpus;
4639 allcpus = alloc_percpu(struct raid5_percpu);
4642 conf->percpu = allcpus;
4646 for_each_present_cpu(cpu) {
4647 if (conf->level == 6) {
4648 spare_page = alloc_page(GFP_KERNEL);
4653 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4655 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4660 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4662 #ifdef CONFIG_HOTPLUG_CPU
4663 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4664 conf->cpu_notify.priority = 0;
4666 err = register_cpu_notifier(&conf->cpu_notify);
4673 static raid5_conf_t *setup_conf(mddev_t *mddev)
4676 int raid_disk, memory, max_disks;
4678 struct disk_info *disk;
4680 if (mddev->new_level != 5
4681 && mddev->new_level != 4
4682 && mddev->new_level != 6) {
4683 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4684 mdname(mddev), mddev->new_level);
4685 return ERR_PTR(-EIO);
4687 if ((mddev->new_level == 5
4688 && !algorithm_valid_raid5(mddev->new_layout)) ||
4689 (mddev->new_level == 6
4690 && !algorithm_valid_raid6(mddev->new_layout))) {
4691 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4692 mdname(mddev), mddev->new_layout);
4693 return ERR_PTR(-EIO);
4695 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4696 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4697 mdname(mddev), mddev->raid_disks);
4698 return ERR_PTR(-EINVAL);
4701 if (!mddev->new_chunk_sectors ||
4702 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4703 !is_power_of_2(mddev->new_chunk_sectors)) {
4704 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4705 mdname(mddev), mddev->new_chunk_sectors << 9);
4706 return ERR_PTR(-EINVAL);
4709 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4712 spin_lock_init(&conf->device_lock);
4713 init_waitqueue_head(&conf->wait_for_stripe);
4714 init_waitqueue_head(&conf->wait_for_overlap);
4715 INIT_LIST_HEAD(&conf->handle_list);
4716 INIT_LIST_HEAD(&conf->hold_list);
4717 INIT_LIST_HEAD(&conf->delayed_list);
4718 INIT_LIST_HEAD(&conf->bitmap_list);
4719 INIT_LIST_HEAD(&conf->inactive_list);
4720 atomic_set(&conf->active_stripes, 0);
4721 atomic_set(&conf->preread_active_stripes, 0);
4722 atomic_set(&conf->active_aligned_reads, 0);
4723 conf->bypass_threshold = BYPASS_THRESHOLD;
4725 conf->raid_disks = mddev->raid_disks;
4726 if (mddev->reshape_position == MaxSector)
4727 conf->previous_raid_disks = mddev->raid_disks;
4729 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4730 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4731 conf->scribble_len = scribble_len(max_disks);
4733 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4738 conf->mddev = mddev;
4740 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4743 conf->level = mddev->new_level;
4744 if (raid5_alloc_percpu(conf) != 0)
4747 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4749 list_for_each_entry(rdev, &mddev->disks, same_set) {
4750 raid_disk = rdev->raid_disk;
4751 if (raid_disk >= max_disks
4754 disk = conf->disks + raid_disk;
4758 if (test_bit(In_sync, &rdev->flags)) {
4759 char b[BDEVNAME_SIZE];
4760 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4762 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4763 } else if (rdev->saved_raid_disk != raid_disk)
4764 /* Cannot rely on bitmap to complete recovery */
4768 conf->chunk_sectors = mddev->new_chunk_sectors;
4769 conf->level = mddev->new_level;
4770 if (conf->level == 6)
4771 conf->max_degraded = 2;
4773 conf->max_degraded = 1;
4774 conf->algorithm = mddev->new_layout;
4775 conf->max_nr_stripes = NR_STRIPES;
4776 conf->reshape_progress = mddev->reshape_position;
4777 if (conf->reshape_progress != MaxSector) {
4778 conf->prev_chunk_sectors = mddev->chunk_sectors;
4779 conf->prev_algo = mddev->layout;
4782 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4783 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4784 if (grow_stripes(conf, conf->max_nr_stripes)) {
4786 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4787 mdname(mddev), memory);
4790 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4791 mdname(mddev), memory);
4793 conf->thread = md_register_thread(raid5d, mddev, NULL);
4794 if (!conf->thread) {
4796 "md/raid:%s: couldn't allocate thread.\n",
4806 return ERR_PTR(-EIO);
4808 return ERR_PTR(-ENOMEM);
4812 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4815 case ALGORITHM_PARITY_0:
4816 if (raid_disk < max_degraded)
4819 case ALGORITHM_PARITY_N:
4820 if (raid_disk >= raid_disks - max_degraded)
4823 case ALGORITHM_PARITY_0_6:
4824 if (raid_disk == 0 ||
4825 raid_disk == raid_disks - 1)
4828 case ALGORITHM_LEFT_ASYMMETRIC_6:
4829 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4830 case ALGORITHM_LEFT_SYMMETRIC_6:
4831 case ALGORITHM_RIGHT_SYMMETRIC_6:
4832 if (raid_disk == raid_disks - 1)
4838 static int run(mddev_t *mddev)
4841 int working_disks = 0;
4842 int dirty_parity_disks = 0;
4844 sector_t reshape_offset = 0;
4846 if (mddev->recovery_cp != MaxSector)
4847 printk(KERN_NOTICE "md/raid:%s: not clean"
4848 " -- starting background reconstruction\n",
4850 if (mddev->reshape_position != MaxSector) {
4851 /* Check that we can continue the reshape.
4852 * Currently only disks can change, it must
4853 * increase, and we must be past the point where
4854 * a stripe over-writes itself
4856 sector_t here_new, here_old;
4858 int max_degraded = (mddev->level == 6 ? 2 : 1);
4860 if (mddev->new_level != mddev->level) {
4861 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4862 "required - aborting.\n",
4866 old_disks = mddev->raid_disks - mddev->delta_disks;
4867 /* reshape_position must be on a new-stripe boundary, and one
4868 * further up in new geometry must map after here in old
4871 here_new = mddev->reshape_position;
4872 if (sector_div(here_new, mddev->new_chunk_sectors *
4873 (mddev->raid_disks - max_degraded))) {
4874 printk(KERN_ERR "md/raid:%s: reshape_position not "
4875 "on a stripe boundary\n", mdname(mddev));
4878 reshape_offset = here_new * mddev->new_chunk_sectors;
4879 /* here_new is the stripe we will write to */
4880 here_old = mddev->reshape_position;
4881 sector_div(here_old, mddev->chunk_sectors *
4882 (old_disks-max_degraded));
4883 /* here_old is the first stripe that we might need to read
4885 if (mddev->delta_disks == 0) {
4886 /* We cannot be sure it is safe to start an in-place
4887 * reshape. It is only safe if user-space if monitoring
4888 * and taking constant backups.
4889 * mdadm always starts a situation like this in
4890 * readonly mode so it can take control before
4891 * allowing any writes. So just check for that.
4893 if ((here_new * mddev->new_chunk_sectors !=
4894 here_old * mddev->chunk_sectors) ||
4896 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4897 " in read-only mode - aborting\n",
4901 } else if (mddev->delta_disks < 0
4902 ? (here_new * mddev->new_chunk_sectors <=
4903 here_old * mddev->chunk_sectors)
4904 : (here_new * mddev->new_chunk_sectors >=
4905 here_old * mddev->chunk_sectors)) {
4906 /* Reading from the same stripe as writing to - bad */
4907 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4908 "auto-recovery - aborting.\n",
4912 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4914 /* OK, we should be able to continue; */
4916 BUG_ON(mddev->level != mddev->new_level);
4917 BUG_ON(mddev->layout != mddev->new_layout);
4918 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4919 BUG_ON(mddev->delta_disks != 0);
4922 if (mddev->private == NULL)
4923 conf = setup_conf(mddev);
4925 conf = mddev->private;
4928 return PTR_ERR(conf);
4930 mddev->thread = conf->thread;
4931 conf->thread = NULL;
4932 mddev->private = conf;
4935 * 0 for a fully functional array, 1 or 2 for a degraded array.
4937 list_for_each_entry(rdev, &mddev->disks, same_set) {
4938 if (rdev->raid_disk < 0)
4940 if (test_bit(In_sync, &rdev->flags)) {
4944 /* This disc is not fully in-sync. However if it
4945 * just stored parity (beyond the recovery_offset),
4946 * when we don't need to be concerned about the
4947 * array being dirty.
4948 * When reshape goes 'backwards', we never have
4949 * partially completed devices, so we only need
4950 * to worry about reshape going forwards.
4952 /* Hack because v0.91 doesn't store recovery_offset properly. */
4953 if (mddev->major_version == 0 &&
4954 mddev->minor_version > 90)
4955 rdev->recovery_offset = reshape_offset;
4957 if (rdev->recovery_offset < reshape_offset) {
4958 /* We need to check old and new layout */
4959 if (!only_parity(rdev->raid_disk,
4962 conf->max_degraded))
4965 if (!only_parity(rdev->raid_disk,
4967 conf->previous_raid_disks,
4968 conf->max_degraded))
4970 dirty_parity_disks++;
4973 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4976 if (has_failed(conf)) {
4977 printk(KERN_ERR "md/raid:%s: not enough operational devices"
4978 " (%d/%d failed)\n",
4979 mdname(mddev), mddev->degraded, conf->raid_disks);
4983 /* device size must be a multiple of chunk size */
4984 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4985 mddev->resync_max_sectors = mddev->dev_sectors;
4987 if (mddev->degraded > dirty_parity_disks &&
4988 mddev->recovery_cp != MaxSector) {
4989 if (mddev->ok_start_degraded)
4991 "md/raid:%s: starting dirty degraded array"
4992 " - data corruption possible.\n",
4996 "md/raid:%s: cannot start dirty degraded array.\n",
5002 if (mddev->degraded == 0)
5003 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5004 " devices, algorithm %d\n", mdname(mddev), conf->level,
5005 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5008 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5009 " out of %d devices, algorithm %d\n",
5010 mdname(mddev), conf->level,
5011 mddev->raid_disks - mddev->degraded,
5012 mddev->raid_disks, mddev->new_layout);
5014 print_raid5_conf(conf);
5016 if (conf->reshape_progress != MaxSector) {
5017 conf->reshape_safe = conf->reshape_progress;
5018 atomic_set(&conf->reshape_stripes, 0);
5019 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5020 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5021 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5022 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5023 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5028 /* Ok, everything is just fine now */
5029 if (mddev->to_remove == &raid5_attrs_group)
5030 mddev->to_remove = NULL;
5031 else if (mddev->kobj.sd &&
5032 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5034 "raid5: failed to create sysfs attributes for %s\n",
5036 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5040 /* read-ahead size must cover two whole stripes, which
5041 * is 2 * (datadisks) * chunksize where 'n' is the
5042 * number of raid devices
5044 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5045 int stripe = data_disks *
5046 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5047 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5048 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5050 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5052 mddev->queue->backing_dev_info.congested_data = mddev;
5053 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5055 chunk_size = mddev->chunk_sectors << 9;
5056 blk_queue_io_min(mddev->queue, chunk_size);
5057 blk_queue_io_opt(mddev->queue, chunk_size *
5058 (conf->raid_disks - conf->max_degraded));
5060 list_for_each_entry(rdev, &mddev->disks, same_set)
5061 disk_stack_limits(mddev->gendisk, rdev->bdev,
5062 rdev->data_offset << 9);
5067 md_unregister_thread(mddev->thread);
5068 mddev->thread = NULL;
5070 print_raid5_conf(conf);
5073 mddev->private = NULL;
5074 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5078 static int stop(mddev_t *mddev)
5080 raid5_conf_t *conf = mddev->private;
5082 md_unregister_thread(mddev->thread);
5083 mddev->thread = NULL;
5085 mddev->queue->backing_dev_info.congested_fn = NULL;
5087 mddev->private = NULL;
5088 mddev->to_remove = &raid5_attrs_group;
5093 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5097 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5098 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5099 seq_printf(seq, "sh %llu, count %d.\n",
5100 (unsigned long long)sh->sector, atomic_read(&sh->count));
5101 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5102 for (i = 0; i < sh->disks; i++) {
5103 seq_printf(seq, "(cache%d: %p %ld) ",
5104 i, sh->dev[i].page, sh->dev[i].flags);
5106 seq_printf(seq, "\n");
5109 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5111 struct stripe_head *sh;
5112 struct hlist_node *hn;
5115 spin_lock_irq(&conf->device_lock);
5116 for (i = 0; i < NR_HASH; i++) {
5117 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5118 if (sh->raid_conf != conf)
5123 spin_unlock_irq(&conf->device_lock);
5127 static void status(struct seq_file *seq, mddev_t *mddev)
5129 raid5_conf_t *conf = mddev->private;
5132 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5133 mddev->chunk_sectors / 2, mddev->layout);
5134 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5135 for (i = 0; i < conf->raid_disks; i++)
5136 seq_printf (seq, "%s",
5137 conf->disks[i].rdev &&
5138 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5139 seq_printf (seq, "]");
5141 seq_printf (seq, "\n");
5142 printall(seq, conf);
5146 static void print_raid5_conf (raid5_conf_t *conf)
5149 struct disk_info *tmp;
5151 printk(KERN_DEBUG "RAID conf printout:\n");
5153 printk("(conf==NULL)\n");
5156 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5158 conf->raid_disks - conf->mddev->degraded);
5160 for (i = 0; i < conf->raid_disks; i++) {
5161 char b[BDEVNAME_SIZE];
5162 tmp = conf->disks + i;
5164 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5165 i, !test_bit(Faulty, &tmp->rdev->flags),
5166 bdevname(tmp->rdev->bdev, b));
5170 static int raid5_spare_active(mddev_t *mddev)
5173 raid5_conf_t *conf = mddev->private;
5174 struct disk_info *tmp;
5176 unsigned long flags;
5178 for (i = 0; i < conf->raid_disks; i++) {
5179 tmp = conf->disks + i;
5181 && tmp->rdev->recovery_offset == MaxSector
5182 && !test_bit(Faulty, &tmp->rdev->flags)
5183 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5185 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5188 spin_lock_irqsave(&conf->device_lock, flags);
5189 mddev->degraded -= count;
5190 spin_unlock_irqrestore(&conf->device_lock, flags);
5191 print_raid5_conf(conf);
5195 static int raid5_remove_disk(mddev_t *mddev, int number)
5197 raid5_conf_t *conf = mddev->private;
5200 struct disk_info *p = conf->disks + number;
5202 print_raid5_conf(conf);
5205 if (number >= conf->raid_disks &&
5206 conf->reshape_progress == MaxSector)
5207 clear_bit(In_sync, &rdev->flags);
5209 if (test_bit(In_sync, &rdev->flags) ||
5210 atomic_read(&rdev->nr_pending)) {
5214 /* Only remove non-faulty devices if recovery
5217 if (!test_bit(Faulty, &rdev->flags) &&
5218 !has_failed(conf) &&
5219 number < conf->raid_disks) {
5225 if (atomic_read(&rdev->nr_pending)) {
5226 /* lost the race, try later */
5233 print_raid5_conf(conf);
5237 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5239 raid5_conf_t *conf = mddev->private;
5242 struct disk_info *p;
5244 int last = conf->raid_disks - 1;
5246 if (has_failed(conf))
5247 /* no point adding a device */
5250 if (rdev->raid_disk >= 0)
5251 first = last = rdev->raid_disk;
5254 * find the disk ... but prefer rdev->saved_raid_disk
5257 if (rdev->saved_raid_disk >= 0 &&
5258 rdev->saved_raid_disk >= first &&
5259 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5260 disk = rdev->saved_raid_disk;
5263 for ( ; disk <= last ; disk++)
5264 if ((p=conf->disks + disk)->rdev == NULL) {
5265 clear_bit(In_sync, &rdev->flags);
5266 rdev->raid_disk = disk;
5268 if (rdev->saved_raid_disk != disk)
5270 rcu_assign_pointer(p->rdev, rdev);
5273 print_raid5_conf(conf);
5277 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5279 /* no resync is happening, and there is enough space
5280 * on all devices, so we can resize.
5281 * We need to make sure resync covers any new space.
5282 * If the array is shrinking we should possibly wait until
5283 * any io in the removed space completes, but it hardly seems
5286 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5287 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5288 mddev->raid_disks));
5289 if (mddev->array_sectors >
5290 raid5_size(mddev, sectors, mddev->raid_disks))
5292 set_capacity(mddev->gendisk, mddev->array_sectors);
5293 revalidate_disk(mddev->gendisk);
5294 if (sectors > mddev->dev_sectors &&
5295 mddev->recovery_cp > mddev->dev_sectors) {
5296 mddev->recovery_cp = mddev->dev_sectors;
5297 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5299 mddev->dev_sectors = sectors;
5300 mddev->resync_max_sectors = sectors;
5304 static int check_stripe_cache(mddev_t *mddev)
5306 /* Can only proceed if there are plenty of stripe_heads.
5307 * We need a minimum of one full stripe,, and for sensible progress
5308 * it is best to have about 4 times that.
5309 * If we require 4 times, then the default 256 4K stripe_heads will
5310 * allow for chunk sizes up to 256K, which is probably OK.
5311 * If the chunk size is greater, user-space should request more
5312 * stripe_heads first.
5314 raid5_conf_t *conf = mddev->private;
5315 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5316 > conf->max_nr_stripes ||
5317 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5318 > conf->max_nr_stripes) {
5319 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5321 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5328 static int check_reshape(mddev_t *mddev)
5330 raid5_conf_t *conf = mddev->private;
5332 if (mddev->delta_disks == 0 &&
5333 mddev->new_layout == mddev->layout &&
5334 mddev->new_chunk_sectors == mddev->chunk_sectors)
5335 return 0; /* nothing to do */
5337 /* Cannot grow a bitmap yet */
5339 if (has_failed(conf))
5341 if (mddev->delta_disks < 0) {
5342 /* We might be able to shrink, but the devices must
5343 * be made bigger first.
5344 * For raid6, 4 is the minimum size.
5345 * Otherwise 2 is the minimum
5348 if (mddev->level == 6)
5350 if (mddev->raid_disks + mddev->delta_disks < min)
5354 if (!check_stripe_cache(mddev))
5357 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5360 static int raid5_start_reshape(mddev_t *mddev)
5362 raid5_conf_t *conf = mddev->private;
5365 unsigned long flags;
5367 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5370 if (!check_stripe_cache(mddev))
5373 list_for_each_entry(rdev, &mddev->disks, same_set)
5374 if (!test_bit(In_sync, &rdev->flags)
5375 && !test_bit(Faulty, &rdev->flags))
5378 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5379 /* Not enough devices even to make a degraded array
5384 /* Refuse to reduce size of the array. Any reductions in
5385 * array size must be through explicit setting of array_size
5388 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5389 < mddev->array_sectors) {
5390 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5391 "before number of disks\n", mdname(mddev));
5395 atomic_set(&conf->reshape_stripes, 0);
5396 spin_lock_irq(&conf->device_lock);
5397 conf->previous_raid_disks = conf->raid_disks;
5398 conf->raid_disks += mddev->delta_disks;
5399 conf->prev_chunk_sectors = conf->chunk_sectors;
5400 conf->chunk_sectors = mddev->new_chunk_sectors;
5401 conf->prev_algo = conf->algorithm;
5402 conf->algorithm = mddev->new_layout;
5403 if (mddev->delta_disks < 0)
5404 conf->reshape_progress = raid5_size(mddev, 0, 0);
5406 conf->reshape_progress = 0;
5407 conf->reshape_safe = conf->reshape_progress;
5409 spin_unlock_irq(&conf->device_lock);
5411 /* Add some new drives, as many as will fit.
5412 * We know there are enough to make the newly sized array work.
5413 * Don't add devices if we are reducing the number of
5414 * devices in the array. This is because it is not possible
5415 * to correctly record the "partially reconstructed" state of
5416 * such devices during the reshape and confusion could result.
5418 if (mddev->delta_disks >= 0) {
5419 int added_devices = 0;
5420 list_for_each_entry(rdev, &mddev->disks, same_set)
5421 if (rdev->raid_disk < 0 &&
5422 !test_bit(Faulty, &rdev->flags)) {
5423 if (raid5_add_disk(mddev, rdev) == 0) {
5426 >= conf->previous_raid_disks) {
5427 set_bit(In_sync, &rdev->flags);
5430 rdev->recovery_offset = 0;
5431 sprintf(nm, "rd%d", rdev->raid_disk);
5432 if (sysfs_create_link(&mddev->kobj,
5434 /* Failure here is OK */;
5436 } else if (rdev->raid_disk >= conf->previous_raid_disks
5437 && !test_bit(Faulty, &rdev->flags)) {
5438 /* This is a spare that was manually added */
5439 set_bit(In_sync, &rdev->flags);
5443 /* When a reshape changes the number of devices,
5444 * ->degraded is measured against the larger of the
5445 * pre and post number of devices.
5447 spin_lock_irqsave(&conf->device_lock, flags);
5448 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5450 spin_unlock_irqrestore(&conf->device_lock, flags);
5452 mddev->raid_disks = conf->raid_disks;
5453 mddev->reshape_position = conf->reshape_progress;
5454 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5456 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5457 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5458 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5459 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5460 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5462 if (!mddev->sync_thread) {
5463 mddev->recovery = 0;
5464 spin_lock_irq(&conf->device_lock);
5465 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5466 conf->reshape_progress = MaxSector;
5467 spin_unlock_irq(&conf->device_lock);
5470 conf->reshape_checkpoint = jiffies;
5471 md_wakeup_thread(mddev->sync_thread);
5472 md_new_event(mddev);
5476 /* This is called from the reshape thread and should make any
5477 * changes needed in 'conf'
5479 static void end_reshape(raid5_conf_t *conf)
5482 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5484 spin_lock_irq(&conf->device_lock);
5485 conf->previous_raid_disks = conf->raid_disks;
5486 conf->reshape_progress = MaxSector;
5487 spin_unlock_irq(&conf->device_lock);
5488 wake_up(&conf->wait_for_overlap);
5490 /* read-ahead size must cover two whole stripes, which is
5491 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5493 if (conf->mddev->queue) {
5494 int data_disks = conf->raid_disks - conf->max_degraded;
5495 int stripe = data_disks * ((conf->chunk_sectors << 9)
5497 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5498 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5503 /* This is called from the raid5d thread with mddev_lock held.
5504 * It makes config changes to the device.
5506 static void raid5_finish_reshape(mddev_t *mddev)
5508 raid5_conf_t *conf = mddev->private;
5510 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5512 if (mddev->delta_disks > 0) {
5513 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5514 set_capacity(mddev->gendisk, mddev->array_sectors);
5515 revalidate_disk(mddev->gendisk);
5518 mddev->degraded = conf->raid_disks;
5519 for (d = 0; d < conf->raid_disks ; d++)
5520 if (conf->disks[d].rdev &&
5522 &conf->disks[d].rdev->flags))
5524 for (d = conf->raid_disks ;
5525 d < conf->raid_disks - mddev->delta_disks;
5527 mdk_rdev_t *rdev = conf->disks[d].rdev;
5528 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5530 sprintf(nm, "rd%d", rdev->raid_disk);
5531 sysfs_remove_link(&mddev->kobj, nm);
5532 rdev->raid_disk = -1;
5536 mddev->layout = conf->algorithm;
5537 mddev->chunk_sectors = conf->chunk_sectors;
5538 mddev->reshape_position = MaxSector;
5539 mddev->delta_disks = 0;
5543 static void raid5_quiesce(mddev_t *mddev, int state)
5545 raid5_conf_t *conf = mddev->private;
5548 case 2: /* resume for a suspend */
5549 wake_up(&conf->wait_for_overlap);
5552 case 1: /* stop all writes */
5553 spin_lock_irq(&conf->device_lock);
5554 /* '2' tells resync/reshape to pause so that all
5555 * active stripes can drain
5558 wait_event_lock_irq(conf->wait_for_stripe,
5559 atomic_read(&conf->active_stripes) == 0 &&
5560 atomic_read(&conf->active_aligned_reads) == 0,
5561 conf->device_lock, /* nothing */);
5563 spin_unlock_irq(&conf->device_lock);
5564 /* allow reshape to continue */
5565 wake_up(&conf->wait_for_overlap);
5568 case 0: /* re-enable writes */
5569 spin_lock_irq(&conf->device_lock);
5571 wake_up(&conf->wait_for_stripe);
5572 wake_up(&conf->wait_for_overlap);
5573 spin_unlock_irq(&conf->device_lock);
5579 static void *raid45_takeover_raid0(mddev_t *mddev, int level)
5581 struct raid0_private_data *raid0_priv = mddev->private;
5584 /* for raid0 takeover only one zone is supported */
5585 if (raid0_priv->nr_strip_zones > 1) {
5586 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5588 return ERR_PTR(-EINVAL);
5591 sectors = raid0_priv->strip_zone[0].zone_end;
5592 sector_div(sectors, raid0_priv->strip_zone[0].nb_dev);
5593 mddev->dev_sectors = sectors;
5594 mddev->new_level = level;
5595 mddev->new_layout = ALGORITHM_PARITY_N;
5596 mddev->new_chunk_sectors = mddev->chunk_sectors;
5597 mddev->raid_disks += 1;
5598 mddev->delta_disks = 1;
5599 /* make sure it will be not marked as dirty */
5600 mddev->recovery_cp = MaxSector;
5602 return setup_conf(mddev);
5606 static void *raid5_takeover_raid1(mddev_t *mddev)
5610 if (mddev->raid_disks != 2 ||
5611 mddev->degraded > 1)
5612 return ERR_PTR(-EINVAL);
5614 /* Should check if there are write-behind devices? */
5616 chunksect = 64*2; /* 64K by default */
5618 /* The array must be an exact multiple of chunksize */
5619 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5622 if ((chunksect<<9) < STRIPE_SIZE)
5623 /* array size does not allow a suitable chunk size */
5624 return ERR_PTR(-EINVAL);
5626 mddev->new_level = 5;
5627 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5628 mddev->new_chunk_sectors = chunksect;
5630 return setup_conf(mddev);
5633 static void *raid5_takeover_raid6(mddev_t *mddev)
5637 switch (mddev->layout) {
5638 case ALGORITHM_LEFT_ASYMMETRIC_6:
5639 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5641 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5642 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5644 case ALGORITHM_LEFT_SYMMETRIC_6:
5645 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5647 case ALGORITHM_RIGHT_SYMMETRIC_6:
5648 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5650 case ALGORITHM_PARITY_0_6:
5651 new_layout = ALGORITHM_PARITY_0;
5653 case ALGORITHM_PARITY_N:
5654 new_layout = ALGORITHM_PARITY_N;
5657 return ERR_PTR(-EINVAL);
5659 mddev->new_level = 5;
5660 mddev->new_layout = new_layout;
5661 mddev->delta_disks = -1;
5662 mddev->raid_disks -= 1;
5663 return setup_conf(mddev);
5667 static int raid5_check_reshape(mddev_t *mddev)
5669 /* For a 2-drive array, the layout and chunk size can be changed
5670 * immediately as not restriping is needed.
5671 * For larger arrays we record the new value - after validation
5672 * to be used by a reshape pass.
5674 raid5_conf_t *conf = mddev->private;
5675 int new_chunk = mddev->new_chunk_sectors;
5677 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5679 if (new_chunk > 0) {
5680 if (!is_power_of_2(new_chunk))
5682 if (new_chunk < (PAGE_SIZE>>9))
5684 if (mddev->array_sectors & (new_chunk-1))
5685 /* not factor of array size */
5689 /* They look valid */
5691 if (mddev->raid_disks == 2) {
5692 /* can make the change immediately */
5693 if (mddev->new_layout >= 0) {
5694 conf->algorithm = mddev->new_layout;
5695 mddev->layout = mddev->new_layout;
5697 if (new_chunk > 0) {
5698 conf->chunk_sectors = new_chunk ;
5699 mddev->chunk_sectors = new_chunk;
5701 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5702 md_wakeup_thread(mddev->thread);
5704 return check_reshape(mddev);
5707 static int raid6_check_reshape(mddev_t *mddev)
5709 int new_chunk = mddev->new_chunk_sectors;
5711 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5713 if (new_chunk > 0) {
5714 if (!is_power_of_2(new_chunk))
5716 if (new_chunk < (PAGE_SIZE >> 9))
5718 if (mddev->array_sectors & (new_chunk-1))
5719 /* not factor of array size */
5723 /* They look valid */
5724 return check_reshape(mddev);
5727 static void *raid5_takeover(mddev_t *mddev)
5729 /* raid5 can take over:
5730 * raid0 - if there is only one strip zone - make it a raid4 layout
5731 * raid1 - if there are two drives. We need to know the chunk size
5732 * raid4 - trivial - just use a raid4 layout.
5733 * raid6 - Providing it is a *_6 layout
5735 if (mddev->level == 0)
5736 return raid45_takeover_raid0(mddev, 5);
5737 if (mddev->level == 1)
5738 return raid5_takeover_raid1(mddev);
5739 if (mddev->level == 4) {
5740 mddev->new_layout = ALGORITHM_PARITY_N;
5741 mddev->new_level = 5;
5742 return setup_conf(mddev);
5744 if (mddev->level == 6)
5745 return raid5_takeover_raid6(mddev);
5747 return ERR_PTR(-EINVAL);
5750 static void *raid4_takeover(mddev_t *mddev)
5752 /* raid4 can take over:
5753 * raid0 - if there is only one strip zone
5754 * raid5 - if layout is right
5756 if (mddev->level == 0)
5757 return raid45_takeover_raid0(mddev, 4);
5758 if (mddev->level == 5 &&
5759 mddev->layout == ALGORITHM_PARITY_N) {
5760 mddev->new_layout = 0;
5761 mddev->new_level = 4;
5762 return setup_conf(mddev);
5764 return ERR_PTR(-EINVAL);
5767 static struct mdk_personality raid5_personality;
5769 static void *raid6_takeover(mddev_t *mddev)
5771 /* Currently can only take over a raid5. We map the
5772 * personality to an equivalent raid6 personality
5773 * with the Q block at the end.
5777 if (mddev->pers != &raid5_personality)
5778 return ERR_PTR(-EINVAL);
5779 if (mddev->degraded > 1)
5780 return ERR_PTR(-EINVAL);
5781 if (mddev->raid_disks > 253)
5782 return ERR_PTR(-EINVAL);
5783 if (mddev->raid_disks < 3)
5784 return ERR_PTR(-EINVAL);
5786 switch (mddev->layout) {
5787 case ALGORITHM_LEFT_ASYMMETRIC:
5788 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5790 case ALGORITHM_RIGHT_ASYMMETRIC:
5791 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5793 case ALGORITHM_LEFT_SYMMETRIC:
5794 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5796 case ALGORITHM_RIGHT_SYMMETRIC:
5797 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5799 case ALGORITHM_PARITY_0:
5800 new_layout = ALGORITHM_PARITY_0_6;
5802 case ALGORITHM_PARITY_N:
5803 new_layout = ALGORITHM_PARITY_N;
5806 return ERR_PTR(-EINVAL);
5808 mddev->new_level = 6;
5809 mddev->new_layout = new_layout;
5810 mddev->delta_disks = 1;
5811 mddev->raid_disks += 1;
5812 return setup_conf(mddev);
5816 static struct mdk_personality raid6_personality =
5820 .owner = THIS_MODULE,
5821 .make_request = make_request,
5825 .error_handler = error,
5826 .hot_add_disk = raid5_add_disk,
5827 .hot_remove_disk= raid5_remove_disk,
5828 .spare_active = raid5_spare_active,
5829 .sync_request = sync_request,
5830 .resize = raid5_resize,
5832 .check_reshape = raid6_check_reshape,
5833 .start_reshape = raid5_start_reshape,
5834 .finish_reshape = raid5_finish_reshape,
5835 .quiesce = raid5_quiesce,
5836 .takeover = raid6_takeover,
5838 static struct mdk_personality raid5_personality =
5842 .owner = THIS_MODULE,
5843 .make_request = make_request,
5847 .error_handler = error,
5848 .hot_add_disk = raid5_add_disk,
5849 .hot_remove_disk= raid5_remove_disk,
5850 .spare_active = raid5_spare_active,
5851 .sync_request = sync_request,
5852 .resize = raid5_resize,
5854 .check_reshape = raid5_check_reshape,
5855 .start_reshape = raid5_start_reshape,
5856 .finish_reshape = raid5_finish_reshape,
5857 .quiesce = raid5_quiesce,
5858 .takeover = raid5_takeover,
5861 static struct mdk_personality raid4_personality =
5865 .owner = THIS_MODULE,
5866 .make_request = make_request,
5870 .error_handler = error,
5871 .hot_add_disk = raid5_add_disk,
5872 .hot_remove_disk= raid5_remove_disk,
5873 .spare_active = raid5_spare_active,
5874 .sync_request = sync_request,
5875 .resize = raid5_resize,
5877 .check_reshape = raid5_check_reshape,
5878 .start_reshape = raid5_start_reshape,
5879 .finish_reshape = raid5_finish_reshape,
5880 .quiesce = raid5_quiesce,
5881 .takeover = raid4_takeover,
5884 static int __init raid5_init(void)
5886 register_md_personality(&raid6_personality);
5887 register_md_personality(&raid5_personality);
5888 register_md_personality(&raid4_personality);
5892 static void raid5_exit(void)
5894 unregister_md_personality(&raid6_personality);
5895 unregister_md_personality(&raid5_personality);
5896 unregister_md_personality(&raid4_personality);
5899 module_init(raid5_init);
5900 module_exit(raid5_exit);
5901 MODULE_LICENSE("GPL");
5902 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5903 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5904 MODULE_ALIAS("md-raid5");
5905 MODULE_ALIAS("md-raid4");
5906 MODULE_ALIAS("md-level-5");
5907 MODULE_ALIAS("md-level-4");
5908 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5909 MODULE_ALIAS("md-raid6");
5910 MODULE_ALIAS("md-level-6");
5912 /* This used to be two separate modules, they were: */
5913 MODULE_ALIAS("raid5");
5914 MODULE_ALIAS("raid6");