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 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
204 list_add_tail(&sh->lru, &conf->delayed_list);
205 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
206 sh->bm_seq - conf->seq_write > 0)
207 list_add_tail(&sh->lru, &conf->bitmap_list);
209 clear_bit(STRIPE_DELAYED, &sh->state);
210 clear_bit(STRIPE_BIT_DELAY, &sh->state);
211 list_add_tail(&sh->lru, &conf->handle_list);
213 md_wakeup_thread(conf->mddev->thread);
215 BUG_ON(stripe_operations_active(sh));
216 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
217 atomic_dec(&conf->preread_active_stripes);
218 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
219 md_wakeup_thread(conf->mddev->thread);
221 atomic_dec(&conf->active_stripes);
222 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
223 list_add_tail(&sh->lru, &conf->inactive_list);
224 wake_up(&conf->wait_for_stripe);
225 if (conf->retry_read_aligned)
226 md_wakeup_thread(conf->mddev->thread);
232 static void release_stripe(struct stripe_head *sh)
234 raid5_conf_t *conf = sh->raid_conf;
237 spin_lock_irqsave(&conf->device_lock, flags);
238 __release_stripe(conf, sh);
239 spin_unlock_irqrestore(&conf->device_lock, flags);
242 static inline void remove_hash(struct stripe_head *sh)
244 pr_debug("remove_hash(), stripe %llu\n",
245 (unsigned long long)sh->sector);
247 hlist_del_init(&sh->hash);
250 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
252 struct hlist_head *hp = stripe_hash(conf, sh->sector);
254 pr_debug("insert_hash(), stripe %llu\n",
255 (unsigned long long)sh->sector);
258 hlist_add_head(&sh->hash, hp);
262 /* find an idle stripe, make sure it is unhashed, and return it. */
263 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
265 struct stripe_head *sh = NULL;
266 struct list_head *first;
269 if (list_empty(&conf->inactive_list))
271 first = conf->inactive_list.next;
272 sh = list_entry(first, struct stripe_head, lru);
273 list_del_init(first);
275 atomic_inc(&conf->active_stripes);
280 static void shrink_buffers(struct stripe_head *sh)
284 int num = sh->raid_conf->pool_size;
286 for (i = 0; i < num ; i++) {
290 sh->dev[i].page = NULL;
295 static int grow_buffers(struct stripe_head *sh)
298 int num = sh->raid_conf->pool_size;
300 for (i = 0; i < num; i++) {
303 if (!(page = alloc_page(GFP_KERNEL))) {
306 sh->dev[i].page = page;
311 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
312 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
313 struct stripe_head *sh);
315 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
317 raid5_conf_t *conf = sh->raid_conf;
320 BUG_ON(atomic_read(&sh->count) != 0);
321 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
322 BUG_ON(stripe_operations_active(sh));
325 pr_debug("init_stripe called, stripe %llu\n",
326 (unsigned long long)sh->sector);
330 sh->generation = conf->generation - previous;
331 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
333 stripe_set_idx(sector, conf, previous, sh);
337 for (i = sh->disks; i--; ) {
338 struct r5dev *dev = &sh->dev[i];
340 if (dev->toread || dev->read || dev->towrite || dev->written ||
341 test_bit(R5_LOCKED, &dev->flags)) {
342 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
343 (unsigned long long)sh->sector, i, dev->toread,
344 dev->read, dev->towrite, dev->written,
345 test_bit(R5_LOCKED, &dev->flags));
349 raid5_build_block(sh, i, previous);
351 insert_hash(conf, sh);
354 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
357 struct stripe_head *sh;
358 struct hlist_node *hn;
361 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
362 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
363 if (sh->sector == sector && sh->generation == generation)
365 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
370 * Need to check if array has failed when deciding whether to:
372 * - remove non-faulty devices
375 * This determination is simple when no reshape is happening.
376 * However if there is a reshape, we need to carefully check
377 * both the before and after sections.
378 * This is because some failed devices may only affect one
379 * of the two sections, and some non-in_sync devices may
380 * be insync in the section most affected by failed devices.
382 static int has_failed(raid5_conf_t *conf)
386 if (conf->mddev->reshape_position == MaxSector)
387 return conf->mddev->degraded > conf->max_degraded;
391 for (i = 0; i < conf->previous_raid_disks; i++) {
392 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
393 if (!rdev || test_bit(Faulty, &rdev->flags))
395 else if (test_bit(In_sync, &rdev->flags))
398 /* not in-sync or faulty.
399 * If the reshape increases the number of devices,
400 * this is being recovered by the reshape, so
401 * this 'previous' section is not in_sync.
402 * If the number of devices is being reduced however,
403 * the device can only be part of the array if
404 * we are reverting a reshape, so this section will
407 if (conf->raid_disks >= conf->previous_raid_disks)
411 if (degraded > conf->max_degraded)
415 for (i = 0; i < conf->raid_disks; i++) {
416 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
417 if (!rdev || test_bit(Faulty, &rdev->flags))
419 else if (test_bit(In_sync, &rdev->flags))
422 /* not in-sync or faulty.
423 * If reshape increases the number of devices, this
424 * section has already been recovered, else it
425 * almost certainly hasn't.
427 if (conf->raid_disks <= conf->previous_raid_disks)
431 if (degraded > conf->max_degraded)
436 static struct stripe_head *
437 get_active_stripe(raid5_conf_t *conf, sector_t sector,
438 int previous, int noblock, int noquiesce)
440 struct stripe_head *sh;
442 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
444 spin_lock_irq(&conf->device_lock);
447 wait_event_lock_irq(conf->wait_for_stripe,
448 conf->quiesce == 0 || noquiesce,
449 conf->device_lock, /* nothing */);
450 sh = __find_stripe(conf, sector, conf->generation - previous);
452 if (!conf->inactive_blocked)
453 sh = get_free_stripe(conf);
454 if (noblock && sh == NULL)
457 conf->inactive_blocked = 1;
458 wait_event_lock_irq(conf->wait_for_stripe,
459 !list_empty(&conf->inactive_list) &&
460 (atomic_read(&conf->active_stripes)
461 < (conf->max_nr_stripes *3/4)
462 || !conf->inactive_blocked),
465 conf->inactive_blocked = 0;
467 init_stripe(sh, sector, previous);
469 if (atomic_read(&sh->count)) {
470 BUG_ON(!list_empty(&sh->lru)
471 && !test_bit(STRIPE_EXPANDING, &sh->state));
473 if (!test_bit(STRIPE_HANDLE, &sh->state))
474 atomic_inc(&conf->active_stripes);
475 if (list_empty(&sh->lru) &&
476 !test_bit(STRIPE_EXPANDING, &sh->state))
478 list_del_init(&sh->lru);
481 } while (sh == NULL);
484 atomic_inc(&sh->count);
486 spin_unlock_irq(&conf->device_lock);
491 raid5_end_read_request(struct bio *bi, int error);
493 raid5_end_write_request(struct bio *bi, int error);
495 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
497 raid5_conf_t *conf = sh->raid_conf;
498 int i, disks = sh->disks;
502 for (i = disks; i--; ) {
506 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
507 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
511 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
516 bi = &sh->dev[i].req;
520 bi->bi_end_io = raid5_end_write_request;
522 bi->bi_end_io = raid5_end_read_request;
525 rdev = rcu_dereference(conf->disks[i].rdev);
526 if (rdev && test_bit(Faulty, &rdev->flags))
529 atomic_inc(&rdev->nr_pending);
533 if (s->syncing || s->expanding || s->expanded)
534 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
536 set_bit(STRIPE_IO_STARTED, &sh->state);
538 bi->bi_bdev = rdev->bdev;
539 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
540 __func__, (unsigned long long)sh->sector,
542 atomic_inc(&sh->count);
543 bi->bi_sector = sh->sector + rdev->data_offset;
544 bi->bi_flags = 1 << BIO_UPTODATE;
548 bi->bi_io_vec = &sh->dev[i].vec;
549 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
550 bi->bi_io_vec[0].bv_offset = 0;
551 bi->bi_size = STRIPE_SIZE;
554 test_bit(R5_ReWrite, &sh->dev[i].flags))
555 atomic_add(STRIPE_SECTORS,
556 &rdev->corrected_errors);
557 generic_make_request(bi);
560 set_bit(STRIPE_DEGRADED, &sh->state);
561 pr_debug("skip op %ld on disc %d for sector %llu\n",
562 bi->bi_rw, i, (unsigned long long)sh->sector);
563 clear_bit(R5_LOCKED, &sh->dev[i].flags);
564 set_bit(STRIPE_HANDLE, &sh->state);
569 static struct dma_async_tx_descriptor *
570 async_copy_data(int frombio, struct bio *bio, struct page *page,
571 sector_t sector, struct dma_async_tx_descriptor *tx)
574 struct page *bio_page;
577 struct async_submit_ctl submit;
578 enum async_tx_flags flags = 0;
580 if (bio->bi_sector >= sector)
581 page_offset = (signed)(bio->bi_sector - sector) * 512;
583 page_offset = (signed)(sector - bio->bi_sector) * -512;
586 flags |= ASYNC_TX_FENCE;
587 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
589 bio_for_each_segment(bvl, bio, i) {
590 int len = bvl->bv_len;
594 if (page_offset < 0) {
595 b_offset = -page_offset;
596 page_offset += b_offset;
600 if (len > 0 && page_offset + len > STRIPE_SIZE)
601 clen = STRIPE_SIZE - page_offset;
606 b_offset += bvl->bv_offset;
607 bio_page = bvl->bv_page;
609 tx = async_memcpy(page, bio_page, page_offset,
610 b_offset, clen, &submit);
612 tx = async_memcpy(bio_page, page, b_offset,
613 page_offset, clen, &submit);
615 /* chain the operations */
616 submit.depend_tx = tx;
618 if (clen < len) /* hit end of page */
626 static void ops_complete_biofill(void *stripe_head_ref)
628 struct stripe_head *sh = stripe_head_ref;
629 struct bio *return_bi = NULL;
630 raid5_conf_t *conf = sh->raid_conf;
633 pr_debug("%s: stripe %llu\n", __func__,
634 (unsigned long long)sh->sector);
636 /* clear completed biofills */
637 spin_lock_irq(&conf->device_lock);
638 for (i = sh->disks; i--; ) {
639 struct r5dev *dev = &sh->dev[i];
641 /* acknowledge completion of a biofill operation */
642 /* and check if we need to reply to a read request,
643 * new R5_Wantfill requests are held off until
644 * !STRIPE_BIOFILL_RUN
646 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
647 struct bio *rbi, *rbi2;
652 while (rbi && rbi->bi_sector <
653 dev->sector + STRIPE_SECTORS) {
654 rbi2 = r5_next_bio(rbi, dev->sector);
655 if (!raid5_dec_bi_phys_segments(rbi)) {
656 rbi->bi_next = return_bi;
663 spin_unlock_irq(&conf->device_lock);
664 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
666 return_io(return_bi);
668 set_bit(STRIPE_HANDLE, &sh->state);
672 static void ops_run_biofill(struct stripe_head *sh)
674 struct dma_async_tx_descriptor *tx = NULL;
675 raid5_conf_t *conf = sh->raid_conf;
676 struct async_submit_ctl submit;
679 pr_debug("%s: stripe %llu\n", __func__,
680 (unsigned long long)sh->sector);
682 for (i = sh->disks; i--; ) {
683 struct r5dev *dev = &sh->dev[i];
684 if (test_bit(R5_Wantfill, &dev->flags)) {
686 spin_lock_irq(&conf->device_lock);
687 dev->read = rbi = dev->toread;
689 spin_unlock_irq(&conf->device_lock);
690 while (rbi && rbi->bi_sector <
691 dev->sector + STRIPE_SECTORS) {
692 tx = async_copy_data(0, rbi, dev->page,
694 rbi = r5_next_bio(rbi, dev->sector);
699 atomic_inc(&sh->count);
700 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
701 async_trigger_callback(&submit);
704 static void mark_target_uptodate(struct stripe_head *sh, int target)
711 tgt = &sh->dev[target];
712 set_bit(R5_UPTODATE, &tgt->flags);
713 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
714 clear_bit(R5_Wantcompute, &tgt->flags);
717 static void ops_complete_compute(void *stripe_head_ref)
719 struct stripe_head *sh = stripe_head_ref;
721 pr_debug("%s: stripe %llu\n", __func__,
722 (unsigned long long)sh->sector);
724 /* mark the computed target(s) as uptodate */
725 mark_target_uptodate(sh, sh->ops.target);
726 mark_target_uptodate(sh, sh->ops.target2);
728 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
729 if (sh->check_state == check_state_compute_run)
730 sh->check_state = check_state_compute_result;
731 set_bit(STRIPE_HANDLE, &sh->state);
735 /* return a pointer to the address conversion region of the scribble buffer */
736 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
737 struct raid5_percpu *percpu)
739 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
742 static struct dma_async_tx_descriptor *
743 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
745 int disks = sh->disks;
746 struct page **xor_srcs = percpu->scribble;
747 int target = sh->ops.target;
748 struct r5dev *tgt = &sh->dev[target];
749 struct page *xor_dest = tgt->page;
751 struct dma_async_tx_descriptor *tx;
752 struct async_submit_ctl submit;
755 pr_debug("%s: stripe %llu block: %d\n",
756 __func__, (unsigned long long)sh->sector, target);
757 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
759 for (i = disks; i--; )
761 xor_srcs[count++] = sh->dev[i].page;
763 atomic_inc(&sh->count);
765 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
766 ops_complete_compute, sh, to_addr_conv(sh, percpu));
767 if (unlikely(count == 1))
768 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
770 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
775 /* set_syndrome_sources - populate source buffers for gen_syndrome
776 * @srcs - (struct page *) array of size sh->disks
777 * @sh - stripe_head to parse
779 * Populates srcs in proper layout order for the stripe and returns the
780 * 'count' of sources to be used in a call to async_gen_syndrome. The P
781 * destination buffer is recorded in srcs[count] and the Q destination
782 * is recorded in srcs[count+1]].
784 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
786 int disks = sh->disks;
787 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
788 int d0_idx = raid6_d0(sh);
792 for (i = 0; i < disks; i++)
798 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
800 srcs[slot] = sh->dev[i].page;
801 i = raid6_next_disk(i, disks);
802 } while (i != d0_idx);
804 return syndrome_disks;
807 static struct dma_async_tx_descriptor *
808 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
810 int disks = sh->disks;
811 struct page **blocks = percpu->scribble;
813 int qd_idx = sh->qd_idx;
814 struct dma_async_tx_descriptor *tx;
815 struct async_submit_ctl submit;
821 if (sh->ops.target < 0)
822 target = sh->ops.target2;
823 else if (sh->ops.target2 < 0)
824 target = sh->ops.target;
826 /* we should only have one valid target */
829 pr_debug("%s: stripe %llu block: %d\n",
830 __func__, (unsigned long long)sh->sector, target);
832 tgt = &sh->dev[target];
833 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
836 atomic_inc(&sh->count);
838 if (target == qd_idx) {
839 count = set_syndrome_sources(blocks, sh);
840 blocks[count] = NULL; /* regenerating p is not necessary */
841 BUG_ON(blocks[count+1] != dest); /* q should already be set */
842 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
843 ops_complete_compute, sh,
844 to_addr_conv(sh, percpu));
845 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
847 /* Compute any data- or p-drive using XOR */
849 for (i = disks; i-- ; ) {
850 if (i == target || i == qd_idx)
852 blocks[count++] = sh->dev[i].page;
855 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
856 NULL, ops_complete_compute, sh,
857 to_addr_conv(sh, percpu));
858 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
864 static struct dma_async_tx_descriptor *
865 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
867 int i, count, disks = sh->disks;
868 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
869 int d0_idx = raid6_d0(sh);
870 int faila = -1, failb = -1;
871 int target = sh->ops.target;
872 int target2 = sh->ops.target2;
873 struct r5dev *tgt = &sh->dev[target];
874 struct r5dev *tgt2 = &sh->dev[target2];
875 struct dma_async_tx_descriptor *tx;
876 struct page **blocks = percpu->scribble;
877 struct async_submit_ctl submit;
879 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
880 __func__, (unsigned long long)sh->sector, target, target2);
881 BUG_ON(target < 0 || target2 < 0);
882 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
883 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
885 /* we need to open-code set_syndrome_sources to handle the
886 * slot number conversion for 'faila' and 'failb'
888 for (i = 0; i < disks ; i++)
893 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
895 blocks[slot] = sh->dev[i].page;
901 i = raid6_next_disk(i, disks);
902 } while (i != d0_idx);
904 BUG_ON(faila == failb);
907 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
908 __func__, (unsigned long long)sh->sector, faila, failb);
910 atomic_inc(&sh->count);
912 if (failb == syndrome_disks+1) {
913 /* Q disk is one of the missing disks */
914 if (faila == syndrome_disks) {
915 /* Missing P+Q, just recompute */
916 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
917 ops_complete_compute, sh,
918 to_addr_conv(sh, percpu));
919 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
920 STRIPE_SIZE, &submit);
924 int qd_idx = sh->qd_idx;
926 /* Missing D+Q: recompute D from P, then recompute Q */
927 if (target == qd_idx)
928 data_target = target2;
930 data_target = target;
933 for (i = disks; i-- ; ) {
934 if (i == data_target || i == qd_idx)
936 blocks[count++] = sh->dev[i].page;
938 dest = sh->dev[data_target].page;
939 init_async_submit(&submit,
940 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
942 to_addr_conv(sh, percpu));
943 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
946 count = set_syndrome_sources(blocks, sh);
947 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
948 ops_complete_compute, sh,
949 to_addr_conv(sh, percpu));
950 return async_gen_syndrome(blocks, 0, count+2,
951 STRIPE_SIZE, &submit);
954 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
955 ops_complete_compute, sh,
956 to_addr_conv(sh, percpu));
957 if (failb == syndrome_disks) {
958 /* We're missing D+P. */
959 return async_raid6_datap_recov(syndrome_disks+2,
963 /* We're missing D+D. */
964 return async_raid6_2data_recov(syndrome_disks+2,
965 STRIPE_SIZE, faila, failb,
972 static void ops_complete_prexor(void *stripe_head_ref)
974 struct stripe_head *sh = stripe_head_ref;
976 pr_debug("%s: stripe %llu\n", __func__,
977 (unsigned long long)sh->sector);
980 static struct dma_async_tx_descriptor *
981 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
982 struct dma_async_tx_descriptor *tx)
984 int disks = sh->disks;
985 struct page **xor_srcs = percpu->scribble;
986 int count = 0, pd_idx = sh->pd_idx, i;
987 struct async_submit_ctl submit;
989 /* existing parity data subtracted */
990 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
992 pr_debug("%s: stripe %llu\n", __func__,
993 (unsigned long long)sh->sector);
995 for (i = disks; i--; ) {
996 struct r5dev *dev = &sh->dev[i];
997 /* Only process blocks that are known to be uptodate */
998 if (test_bit(R5_Wantdrain, &dev->flags))
999 xor_srcs[count++] = dev->page;
1002 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1003 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1004 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1009 static struct dma_async_tx_descriptor *
1010 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1012 int disks = sh->disks;
1015 pr_debug("%s: stripe %llu\n", __func__,
1016 (unsigned long long)sh->sector);
1018 for (i = disks; i--; ) {
1019 struct r5dev *dev = &sh->dev[i];
1022 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1025 spin_lock(&sh->lock);
1026 chosen = dev->towrite;
1027 dev->towrite = NULL;
1028 BUG_ON(dev->written);
1029 wbi = dev->written = chosen;
1030 spin_unlock(&sh->lock);
1032 while (wbi && wbi->bi_sector <
1033 dev->sector + STRIPE_SECTORS) {
1034 if (wbi->bi_rw & REQ_FUA)
1035 set_bit(R5_WantFUA, &dev->flags);
1036 tx = async_copy_data(1, wbi, dev->page,
1038 wbi = r5_next_bio(wbi, dev->sector);
1046 static void ops_complete_reconstruct(void *stripe_head_ref)
1048 struct stripe_head *sh = stripe_head_ref;
1049 int disks = sh->disks;
1050 int pd_idx = sh->pd_idx;
1051 int qd_idx = sh->qd_idx;
1055 pr_debug("%s: stripe %llu\n", __func__,
1056 (unsigned long long)sh->sector);
1058 for (i = disks; i--; )
1059 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1061 for (i = disks; i--; ) {
1062 struct r5dev *dev = &sh->dev[i];
1064 if (dev->written || i == pd_idx || i == qd_idx) {
1065 set_bit(R5_UPTODATE, &dev->flags);
1067 set_bit(R5_WantFUA, &dev->flags);
1071 if (sh->reconstruct_state == reconstruct_state_drain_run)
1072 sh->reconstruct_state = reconstruct_state_drain_result;
1073 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1074 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1076 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1077 sh->reconstruct_state = reconstruct_state_result;
1080 set_bit(STRIPE_HANDLE, &sh->state);
1085 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1086 struct dma_async_tx_descriptor *tx)
1088 int disks = sh->disks;
1089 struct page **xor_srcs = percpu->scribble;
1090 struct async_submit_ctl submit;
1091 int count = 0, pd_idx = sh->pd_idx, i;
1092 struct page *xor_dest;
1094 unsigned long flags;
1096 pr_debug("%s: stripe %llu\n", __func__,
1097 (unsigned long long)sh->sector);
1099 /* check if prexor is active which means only process blocks
1100 * that are part of a read-modify-write (written)
1102 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1104 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1105 for (i = disks; i--; ) {
1106 struct r5dev *dev = &sh->dev[i];
1108 xor_srcs[count++] = dev->page;
1111 xor_dest = sh->dev[pd_idx].page;
1112 for (i = disks; i--; ) {
1113 struct r5dev *dev = &sh->dev[i];
1115 xor_srcs[count++] = dev->page;
1119 /* 1/ if we prexor'd then the dest is reused as a source
1120 * 2/ if we did not prexor then we are redoing the parity
1121 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1122 * for the synchronous xor case
1124 flags = ASYNC_TX_ACK |
1125 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1127 atomic_inc(&sh->count);
1129 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1130 to_addr_conv(sh, percpu));
1131 if (unlikely(count == 1))
1132 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1134 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1138 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1139 struct dma_async_tx_descriptor *tx)
1141 struct async_submit_ctl submit;
1142 struct page **blocks = percpu->scribble;
1145 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1147 count = set_syndrome_sources(blocks, sh);
1149 atomic_inc(&sh->count);
1151 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1152 sh, to_addr_conv(sh, percpu));
1153 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1156 static void ops_complete_check(void *stripe_head_ref)
1158 struct stripe_head *sh = stripe_head_ref;
1160 pr_debug("%s: stripe %llu\n", __func__,
1161 (unsigned long long)sh->sector);
1163 sh->check_state = check_state_check_result;
1164 set_bit(STRIPE_HANDLE, &sh->state);
1168 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1170 int disks = sh->disks;
1171 int pd_idx = sh->pd_idx;
1172 int qd_idx = sh->qd_idx;
1173 struct page *xor_dest;
1174 struct page **xor_srcs = percpu->scribble;
1175 struct dma_async_tx_descriptor *tx;
1176 struct async_submit_ctl submit;
1180 pr_debug("%s: stripe %llu\n", __func__,
1181 (unsigned long long)sh->sector);
1184 xor_dest = sh->dev[pd_idx].page;
1185 xor_srcs[count++] = xor_dest;
1186 for (i = disks; i--; ) {
1187 if (i == pd_idx || i == qd_idx)
1189 xor_srcs[count++] = sh->dev[i].page;
1192 init_async_submit(&submit, 0, NULL, NULL, NULL,
1193 to_addr_conv(sh, percpu));
1194 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1195 &sh->ops.zero_sum_result, &submit);
1197 atomic_inc(&sh->count);
1198 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1199 tx = async_trigger_callback(&submit);
1202 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1204 struct page **srcs = percpu->scribble;
1205 struct async_submit_ctl submit;
1208 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1209 (unsigned long long)sh->sector, checkp);
1211 count = set_syndrome_sources(srcs, sh);
1215 atomic_inc(&sh->count);
1216 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1217 sh, to_addr_conv(sh, percpu));
1218 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1219 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1222 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1224 int overlap_clear = 0, i, disks = sh->disks;
1225 struct dma_async_tx_descriptor *tx = NULL;
1226 raid5_conf_t *conf = sh->raid_conf;
1227 int level = conf->level;
1228 struct raid5_percpu *percpu;
1232 percpu = per_cpu_ptr(conf->percpu, cpu);
1233 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1234 ops_run_biofill(sh);
1238 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1240 tx = ops_run_compute5(sh, percpu);
1242 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1243 tx = ops_run_compute6_1(sh, percpu);
1245 tx = ops_run_compute6_2(sh, percpu);
1247 /* terminate the chain if reconstruct is not set to be run */
1248 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1252 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1253 tx = ops_run_prexor(sh, percpu, tx);
1255 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1256 tx = ops_run_biodrain(sh, tx);
1260 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1262 ops_run_reconstruct5(sh, percpu, tx);
1264 ops_run_reconstruct6(sh, percpu, tx);
1267 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1268 if (sh->check_state == check_state_run)
1269 ops_run_check_p(sh, percpu);
1270 else if (sh->check_state == check_state_run_q)
1271 ops_run_check_pq(sh, percpu, 0);
1272 else if (sh->check_state == check_state_run_pq)
1273 ops_run_check_pq(sh, percpu, 1);
1279 for (i = disks; i--; ) {
1280 struct r5dev *dev = &sh->dev[i];
1281 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1282 wake_up(&sh->raid_conf->wait_for_overlap);
1287 #ifdef CONFIG_MULTICORE_RAID456
1288 static void async_run_ops(void *param, async_cookie_t cookie)
1290 struct stripe_head *sh = param;
1291 unsigned long ops_request = sh->ops.request;
1293 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1294 wake_up(&sh->ops.wait_for_ops);
1296 __raid_run_ops(sh, ops_request);
1300 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1302 /* since handle_stripe can be called outside of raid5d context
1303 * we need to ensure sh->ops.request is de-staged before another
1306 wait_event(sh->ops.wait_for_ops,
1307 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1308 sh->ops.request = ops_request;
1310 atomic_inc(&sh->count);
1311 async_schedule(async_run_ops, sh);
1314 #define raid_run_ops __raid_run_ops
1317 static int grow_one_stripe(raid5_conf_t *conf)
1319 struct stripe_head *sh;
1320 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1323 memset(sh, 0, sizeof(*sh) + (conf->pool_size-1)*sizeof(struct r5dev));
1324 sh->raid_conf = conf;
1325 spin_lock_init(&sh->lock);
1326 #ifdef CONFIG_MULTICORE_RAID456
1327 init_waitqueue_head(&sh->ops.wait_for_ops);
1330 if (grow_buffers(sh)) {
1332 kmem_cache_free(conf->slab_cache, sh);
1335 /* we just created an active stripe so... */
1336 atomic_set(&sh->count, 1);
1337 atomic_inc(&conf->active_stripes);
1338 INIT_LIST_HEAD(&sh->lru);
1343 static int grow_stripes(raid5_conf_t *conf, int num)
1345 struct kmem_cache *sc;
1346 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1348 if (conf->mddev->gendisk)
1349 sprintf(conf->cache_name[0],
1350 "raid%d-%s", conf->level, mdname(conf->mddev));
1352 sprintf(conf->cache_name[0],
1353 "raid%d-%p", conf->level, conf->mddev);
1354 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1356 conf->active_name = 0;
1357 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1358 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1362 conf->slab_cache = sc;
1363 conf->pool_size = devs;
1365 if (!grow_one_stripe(conf))
1371 * scribble_len - return the required size of the scribble region
1372 * @num - total number of disks in the array
1374 * The size must be enough to contain:
1375 * 1/ a struct page pointer for each device in the array +2
1376 * 2/ room to convert each entry in (1) to its corresponding dma
1377 * (dma_map_page()) or page (page_address()) address.
1379 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1380 * calculate over all devices (not just the data blocks), using zeros in place
1381 * of the P and Q blocks.
1383 static size_t scribble_len(int num)
1387 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1392 static int resize_stripes(raid5_conf_t *conf, int newsize)
1394 /* Make all the stripes able to hold 'newsize' devices.
1395 * New slots in each stripe get 'page' set to a new page.
1397 * This happens in stages:
1398 * 1/ create a new kmem_cache and allocate the required number of
1400 * 2/ gather all the old stripe_heads and tranfer the pages across
1401 * to the new stripe_heads. This will have the side effect of
1402 * freezing the array as once all stripe_heads have been collected,
1403 * no IO will be possible. Old stripe heads are freed once their
1404 * pages have been transferred over, and the old kmem_cache is
1405 * freed when all stripes are done.
1406 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1407 * we simple return a failre status - no need to clean anything up.
1408 * 4/ allocate new pages for the new slots in the new stripe_heads.
1409 * If this fails, we don't bother trying the shrink the
1410 * stripe_heads down again, we just leave them as they are.
1411 * As each stripe_head is processed the new one is released into
1414 * Once step2 is started, we cannot afford to wait for a write,
1415 * so we use GFP_NOIO allocations.
1417 struct stripe_head *osh, *nsh;
1418 LIST_HEAD(newstripes);
1419 struct disk_info *ndisks;
1422 struct kmem_cache *sc;
1425 if (newsize <= conf->pool_size)
1426 return 0; /* never bother to shrink */
1428 err = md_allow_write(conf->mddev);
1433 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1434 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1439 for (i = conf->max_nr_stripes; i; i--) {
1440 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1444 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1446 nsh->raid_conf = conf;
1447 spin_lock_init(&nsh->lock);
1448 #ifdef CONFIG_MULTICORE_RAID456
1449 init_waitqueue_head(&nsh->ops.wait_for_ops);
1452 list_add(&nsh->lru, &newstripes);
1455 /* didn't get enough, give up */
1456 while (!list_empty(&newstripes)) {
1457 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1458 list_del(&nsh->lru);
1459 kmem_cache_free(sc, nsh);
1461 kmem_cache_destroy(sc);
1464 /* Step 2 - Must use GFP_NOIO now.
1465 * OK, we have enough stripes, start collecting inactive
1466 * stripes and copying them over
1468 list_for_each_entry(nsh, &newstripes, lru) {
1469 spin_lock_irq(&conf->device_lock);
1470 wait_event_lock_irq(conf->wait_for_stripe,
1471 !list_empty(&conf->inactive_list),
1474 osh = get_free_stripe(conf);
1475 spin_unlock_irq(&conf->device_lock);
1476 atomic_set(&nsh->count, 1);
1477 for(i=0; i<conf->pool_size; i++)
1478 nsh->dev[i].page = osh->dev[i].page;
1479 for( ; i<newsize; i++)
1480 nsh->dev[i].page = NULL;
1481 kmem_cache_free(conf->slab_cache, osh);
1483 kmem_cache_destroy(conf->slab_cache);
1486 * At this point, we are holding all the stripes so the array
1487 * is completely stalled, so now is a good time to resize
1488 * conf->disks and the scribble region
1490 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1492 for (i=0; i<conf->raid_disks; i++)
1493 ndisks[i] = conf->disks[i];
1495 conf->disks = ndisks;
1500 conf->scribble_len = scribble_len(newsize);
1501 for_each_present_cpu(cpu) {
1502 struct raid5_percpu *percpu;
1505 percpu = per_cpu_ptr(conf->percpu, cpu);
1506 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1509 kfree(percpu->scribble);
1510 percpu->scribble = scribble;
1518 /* Step 4, return new stripes to service */
1519 while(!list_empty(&newstripes)) {
1520 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1521 list_del_init(&nsh->lru);
1523 for (i=conf->raid_disks; i < newsize; i++)
1524 if (nsh->dev[i].page == NULL) {
1525 struct page *p = alloc_page(GFP_NOIO);
1526 nsh->dev[i].page = p;
1530 release_stripe(nsh);
1532 /* critical section pass, GFP_NOIO no longer needed */
1534 conf->slab_cache = sc;
1535 conf->active_name = 1-conf->active_name;
1536 conf->pool_size = newsize;
1540 static int drop_one_stripe(raid5_conf_t *conf)
1542 struct stripe_head *sh;
1544 spin_lock_irq(&conf->device_lock);
1545 sh = get_free_stripe(conf);
1546 spin_unlock_irq(&conf->device_lock);
1549 BUG_ON(atomic_read(&sh->count));
1551 kmem_cache_free(conf->slab_cache, sh);
1552 atomic_dec(&conf->active_stripes);
1556 static void shrink_stripes(raid5_conf_t *conf)
1558 while (drop_one_stripe(conf))
1561 if (conf->slab_cache)
1562 kmem_cache_destroy(conf->slab_cache);
1563 conf->slab_cache = NULL;
1566 static void raid5_end_read_request(struct bio * bi, int error)
1568 struct stripe_head *sh = bi->bi_private;
1569 raid5_conf_t *conf = sh->raid_conf;
1570 int disks = sh->disks, i;
1571 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1572 char b[BDEVNAME_SIZE];
1576 for (i=0 ; i<disks; i++)
1577 if (bi == &sh->dev[i].req)
1580 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1581 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1589 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1590 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1591 rdev = conf->disks[i].rdev;
1592 printk_rl(KERN_INFO "md/raid:%s: read error corrected"
1593 " (%lu sectors at %llu on %s)\n",
1594 mdname(conf->mddev), STRIPE_SECTORS,
1595 (unsigned long long)(sh->sector
1596 + rdev->data_offset),
1597 bdevname(rdev->bdev, b));
1598 clear_bit(R5_ReadError, &sh->dev[i].flags);
1599 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1601 if (atomic_read(&conf->disks[i].rdev->read_errors))
1602 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1604 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1606 rdev = conf->disks[i].rdev;
1608 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1609 atomic_inc(&rdev->read_errors);
1610 if (conf->mddev->degraded >= conf->max_degraded)
1611 printk_rl(KERN_WARNING
1612 "md/raid:%s: read error not correctable "
1613 "(sector %llu on %s).\n",
1614 mdname(conf->mddev),
1615 (unsigned long long)(sh->sector
1616 + rdev->data_offset),
1618 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1620 printk_rl(KERN_WARNING
1621 "md/raid:%s: read error NOT corrected!! "
1622 "(sector %llu on %s).\n",
1623 mdname(conf->mddev),
1624 (unsigned long long)(sh->sector
1625 + rdev->data_offset),
1627 else if (atomic_read(&rdev->read_errors)
1628 > conf->max_nr_stripes)
1630 "md/raid:%s: Too many read errors, failing device %s.\n",
1631 mdname(conf->mddev), bdn);
1635 set_bit(R5_ReadError, &sh->dev[i].flags);
1637 clear_bit(R5_ReadError, &sh->dev[i].flags);
1638 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1639 md_error(conf->mddev, rdev);
1642 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1643 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1644 set_bit(STRIPE_HANDLE, &sh->state);
1648 static void raid5_end_write_request(struct bio *bi, int error)
1650 struct stripe_head *sh = bi->bi_private;
1651 raid5_conf_t *conf = sh->raid_conf;
1652 int disks = sh->disks, i;
1653 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1655 for (i=0 ; i<disks; i++)
1656 if (bi == &sh->dev[i].req)
1659 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1660 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1668 md_error(conf->mddev, conf->disks[i].rdev);
1670 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1672 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1673 set_bit(STRIPE_HANDLE, &sh->state);
1678 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1680 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1682 struct r5dev *dev = &sh->dev[i];
1684 bio_init(&dev->req);
1685 dev->req.bi_io_vec = &dev->vec;
1687 dev->req.bi_max_vecs++;
1688 dev->vec.bv_page = dev->page;
1689 dev->vec.bv_len = STRIPE_SIZE;
1690 dev->vec.bv_offset = 0;
1692 dev->req.bi_sector = sh->sector;
1693 dev->req.bi_private = sh;
1696 dev->sector = compute_blocknr(sh, i, previous);
1699 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1701 char b[BDEVNAME_SIZE];
1702 raid5_conf_t *conf = mddev->private;
1703 pr_debug("raid456: error called\n");
1705 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1706 unsigned long flags;
1707 spin_lock_irqsave(&conf->device_lock, flags);
1709 spin_unlock_irqrestore(&conf->device_lock, flags);
1711 * if recovery was running, make sure it aborts.
1713 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1715 set_bit(Faulty, &rdev->flags);
1716 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1718 "md/raid:%s: Disk failure on %s, disabling device.\n"
1719 "md/raid:%s: Operation continuing on %d devices.\n",
1721 bdevname(rdev->bdev, b),
1723 conf->raid_disks - mddev->degraded);
1727 * Input: a 'big' sector number,
1728 * Output: index of the data and parity disk, and the sector # in them.
1730 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1731 int previous, int *dd_idx,
1732 struct stripe_head *sh)
1734 sector_t stripe, stripe2;
1735 sector_t chunk_number;
1736 unsigned int chunk_offset;
1739 sector_t new_sector;
1740 int algorithm = previous ? conf->prev_algo
1742 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1743 : conf->chunk_sectors;
1744 int raid_disks = previous ? conf->previous_raid_disks
1746 int data_disks = raid_disks - conf->max_degraded;
1748 /* First compute the information on this sector */
1751 * Compute the chunk number and the sector offset inside the chunk
1753 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1754 chunk_number = r_sector;
1757 * Compute the stripe number
1759 stripe = chunk_number;
1760 *dd_idx = sector_div(stripe, data_disks);
1763 * Select the parity disk based on the user selected algorithm.
1765 pd_idx = qd_idx = ~0;
1766 switch(conf->level) {
1768 pd_idx = data_disks;
1771 switch (algorithm) {
1772 case ALGORITHM_LEFT_ASYMMETRIC:
1773 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1774 if (*dd_idx >= pd_idx)
1777 case ALGORITHM_RIGHT_ASYMMETRIC:
1778 pd_idx = sector_div(stripe2, raid_disks);
1779 if (*dd_idx >= pd_idx)
1782 case ALGORITHM_LEFT_SYMMETRIC:
1783 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1784 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1786 case ALGORITHM_RIGHT_SYMMETRIC:
1787 pd_idx = sector_div(stripe2, raid_disks);
1788 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1790 case ALGORITHM_PARITY_0:
1794 case ALGORITHM_PARITY_N:
1795 pd_idx = data_disks;
1803 switch (algorithm) {
1804 case ALGORITHM_LEFT_ASYMMETRIC:
1805 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1806 qd_idx = pd_idx + 1;
1807 if (pd_idx == raid_disks-1) {
1808 (*dd_idx)++; /* Q D D D P */
1810 } else if (*dd_idx >= pd_idx)
1811 (*dd_idx) += 2; /* D D P Q D */
1813 case ALGORITHM_RIGHT_ASYMMETRIC:
1814 pd_idx = sector_div(stripe2, raid_disks);
1815 qd_idx = pd_idx + 1;
1816 if (pd_idx == raid_disks-1) {
1817 (*dd_idx)++; /* Q D D D P */
1819 } else if (*dd_idx >= pd_idx)
1820 (*dd_idx) += 2; /* D D P Q D */
1822 case ALGORITHM_LEFT_SYMMETRIC:
1823 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1824 qd_idx = (pd_idx + 1) % raid_disks;
1825 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1827 case ALGORITHM_RIGHT_SYMMETRIC:
1828 pd_idx = sector_div(stripe2, raid_disks);
1829 qd_idx = (pd_idx + 1) % raid_disks;
1830 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1833 case ALGORITHM_PARITY_0:
1838 case ALGORITHM_PARITY_N:
1839 pd_idx = data_disks;
1840 qd_idx = data_disks + 1;
1843 case ALGORITHM_ROTATING_ZERO_RESTART:
1844 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1845 * of blocks for computing Q is different.
1847 pd_idx = sector_div(stripe2, raid_disks);
1848 qd_idx = pd_idx + 1;
1849 if (pd_idx == raid_disks-1) {
1850 (*dd_idx)++; /* Q D D D P */
1852 } else if (*dd_idx >= pd_idx)
1853 (*dd_idx) += 2; /* D D P Q D */
1857 case ALGORITHM_ROTATING_N_RESTART:
1858 /* Same a left_asymmetric, by first stripe is
1859 * D D D P Q rather than
1863 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1864 qd_idx = pd_idx + 1;
1865 if (pd_idx == raid_disks-1) {
1866 (*dd_idx)++; /* Q D D D P */
1868 } else if (*dd_idx >= pd_idx)
1869 (*dd_idx) += 2; /* D D P Q D */
1873 case ALGORITHM_ROTATING_N_CONTINUE:
1874 /* Same as left_symmetric but Q is before P */
1875 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1876 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1877 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1881 case ALGORITHM_LEFT_ASYMMETRIC_6:
1882 /* RAID5 left_asymmetric, with Q on last device */
1883 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1884 if (*dd_idx >= pd_idx)
1886 qd_idx = raid_disks - 1;
1889 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1890 pd_idx = sector_div(stripe2, raid_disks-1);
1891 if (*dd_idx >= pd_idx)
1893 qd_idx = raid_disks - 1;
1896 case ALGORITHM_LEFT_SYMMETRIC_6:
1897 pd_idx = data_disks - 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_RIGHT_SYMMETRIC_6:
1903 pd_idx = sector_div(stripe2, raid_disks-1);
1904 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1905 qd_idx = raid_disks - 1;
1908 case ALGORITHM_PARITY_0_6:
1911 qd_idx = raid_disks - 1;
1921 sh->pd_idx = pd_idx;
1922 sh->qd_idx = qd_idx;
1923 sh->ddf_layout = ddf_layout;
1926 * Finally, compute the new sector number
1928 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1933 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1935 raid5_conf_t *conf = sh->raid_conf;
1936 int raid_disks = sh->disks;
1937 int data_disks = raid_disks - conf->max_degraded;
1938 sector_t new_sector = sh->sector, check;
1939 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1940 : conf->chunk_sectors;
1941 int algorithm = previous ? conf->prev_algo
1945 sector_t chunk_number;
1946 int dummy1, dd_idx = i;
1948 struct stripe_head sh2;
1951 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1952 stripe = new_sector;
1954 if (i == sh->pd_idx)
1956 switch(conf->level) {
1959 switch (algorithm) {
1960 case ALGORITHM_LEFT_ASYMMETRIC:
1961 case ALGORITHM_RIGHT_ASYMMETRIC:
1965 case ALGORITHM_LEFT_SYMMETRIC:
1966 case ALGORITHM_RIGHT_SYMMETRIC:
1969 i -= (sh->pd_idx + 1);
1971 case ALGORITHM_PARITY_0:
1974 case ALGORITHM_PARITY_N:
1981 if (i == sh->qd_idx)
1982 return 0; /* It is the Q disk */
1983 switch (algorithm) {
1984 case ALGORITHM_LEFT_ASYMMETRIC:
1985 case ALGORITHM_RIGHT_ASYMMETRIC:
1986 case ALGORITHM_ROTATING_ZERO_RESTART:
1987 case ALGORITHM_ROTATING_N_RESTART:
1988 if (sh->pd_idx == raid_disks-1)
1989 i--; /* Q D D D P */
1990 else if (i > sh->pd_idx)
1991 i -= 2; /* D D P Q D */
1993 case ALGORITHM_LEFT_SYMMETRIC:
1994 case ALGORITHM_RIGHT_SYMMETRIC:
1995 if (sh->pd_idx == raid_disks-1)
1996 i--; /* Q D D D P */
2001 i -= (sh->pd_idx + 2);
2004 case ALGORITHM_PARITY_0:
2007 case ALGORITHM_PARITY_N:
2009 case ALGORITHM_ROTATING_N_CONTINUE:
2010 /* Like left_symmetric, but P is before Q */
2011 if (sh->pd_idx == 0)
2012 i--; /* P D D D Q */
2017 i -= (sh->pd_idx + 1);
2020 case ALGORITHM_LEFT_ASYMMETRIC_6:
2021 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2025 case ALGORITHM_LEFT_SYMMETRIC_6:
2026 case ALGORITHM_RIGHT_SYMMETRIC_6:
2028 i += data_disks + 1;
2029 i -= (sh->pd_idx + 1);
2031 case ALGORITHM_PARITY_0_6:
2040 chunk_number = stripe * data_disks + i;
2041 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2043 check = raid5_compute_sector(conf, r_sector,
2044 previous, &dummy1, &sh2);
2045 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2046 || sh2.qd_idx != sh->qd_idx) {
2047 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2048 mdname(conf->mddev));
2056 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2057 int rcw, int expand)
2059 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2060 raid5_conf_t *conf = sh->raid_conf;
2061 int level = conf->level;
2064 /* if we are not expanding this is a proper write request, and
2065 * there will be bios with new data to be drained into the
2069 sh->reconstruct_state = reconstruct_state_drain_run;
2070 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2072 sh->reconstruct_state = reconstruct_state_run;
2074 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2076 for (i = disks; i--; ) {
2077 struct r5dev *dev = &sh->dev[i];
2080 set_bit(R5_LOCKED, &dev->flags);
2081 set_bit(R5_Wantdrain, &dev->flags);
2083 clear_bit(R5_UPTODATE, &dev->flags);
2087 if (s->locked + conf->max_degraded == disks)
2088 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2089 atomic_inc(&conf->pending_full_writes);
2092 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2093 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2095 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2096 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2097 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2098 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2100 for (i = disks; i--; ) {
2101 struct r5dev *dev = &sh->dev[i];
2106 (test_bit(R5_UPTODATE, &dev->flags) ||
2107 test_bit(R5_Wantcompute, &dev->flags))) {
2108 set_bit(R5_Wantdrain, &dev->flags);
2109 set_bit(R5_LOCKED, &dev->flags);
2110 clear_bit(R5_UPTODATE, &dev->flags);
2116 /* keep the parity disk(s) locked while asynchronous operations
2119 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2120 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2124 int qd_idx = sh->qd_idx;
2125 struct r5dev *dev = &sh->dev[qd_idx];
2127 set_bit(R5_LOCKED, &dev->flags);
2128 clear_bit(R5_UPTODATE, &dev->flags);
2132 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2133 __func__, (unsigned long long)sh->sector,
2134 s->locked, s->ops_request);
2138 * Each stripe/dev can have one or more bion attached.
2139 * toread/towrite point to the first in a chain.
2140 * The bi_next chain must be in order.
2142 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2145 raid5_conf_t *conf = sh->raid_conf;
2148 pr_debug("adding bh b#%llu to stripe s#%llu\n",
2149 (unsigned long long)bi->bi_sector,
2150 (unsigned long long)sh->sector);
2153 spin_lock(&sh->lock);
2154 spin_lock_irq(&conf->device_lock);
2156 bip = &sh->dev[dd_idx].towrite;
2157 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2160 bip = &sh->dev[dd_idx].toread;
2161 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2162 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2164 bip = & (*bip)->bi_next;
2166 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2169 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2173 bi->bi_phys_segments++;
2174 spin_unlock_irq(&conf->device_lock);
2175 spin_unlock(&sh->lock);
2177 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2178 (unsigned long long)bi->bi_sector,
2179 (unsigned long long)sh->sector, dd_idx);
2181 if (conf->mddev->bitmap && firstwrite) {
2182 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2184 sh->bm_seq = conf->seq_flush+1;
2185 set_bit(STRIPE_BIT_DELAY, &sh->state);
2189 /* check if page is covered */
2190 sector_t sector = sh->dev[dd_idx].sector;
2191 for (bi=sh->dev[dd_idx].towrite;
2192 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2193 bi && bi->bi_sector <= sector;
2194 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2195 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2196 sector = bi->bi_sector + (bi->bi_size>>9);
2198 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2199 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2204 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2205 spin_unlock_irq(&conf->device_lock);
2206 spin_unlock(&sh->lock);
2210 static void end_reshape(raid5_conf_t *conf);
2212 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2213 struct stripe_head *sh)
2215 int sectors_per_chunk =
2216 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2218 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2219 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2221 raid5_compute_sector(conf,
2222 stripe * (disks - conf->max_degraded)
2223 *sectors_per_chunk + chunk_offset,
2229 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2230 struct stripe_head_state *s, int disks,
2231 struct bio **return_bi)
2234 for (i = disks; i--; ) {
2238 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2241 rdev = rcu_dereference(conf->disks[i].rdev);
2242 if (rdev && test_bit(In_sync, &rdev->flags))
2243 /* multiple read failures in one stripe */
2244 md_error(conf->mddev, rdev);
2247 spin_lock_irq(&conf->device_lock);
2248 /* fail all writes first */
2249 bi = sh->dev[i].towrite;
2250 sh->dev[i].towrite = NULL;
2256 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2257 wake_up(&conf->wait_for_overlap);
2259 while (bi && bi->bi_sector <
2260 sh->dev[i].sector + STRIPE_SECTORS) {
2261 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2262 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2263 if (!raid5_dec_bi_phys_segments(bi)) {
2264 md_write_end(conf->mddev);
2265 bi->bi_next = *return_bi;
2270 /* and fail all 'written' */
2271 bi = sh->dev[i].written;
2272 sh->dev[i].written = NULL;
2273 if (bi) bitmap_end = 1;
2274 while (bi && bi->bi_sector <
2275 sh->dev[i].sector + STRIPE_SECTORS) {
2276 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2277 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2278 if (!raid5_dec_bi_phys_segments(bi)) {
2279 md_write_end(conf->mddev);
2280 bi->bi_next = *return_bi;
2286 /* fail any reads if this device is non-operational and
2287 * the data has not reached the cache yet.
2289 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2290 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2291 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2292 bi = sh->dev[i].toread;
2293 sh->dev[i].toread = NULL;
2294 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2295 wake_up(&conf->wait_for_overlap);
2296 if (bi) s->to_read--;
2297 while (bi && bi->bi_sector <
2298 sh->dev[i].sector + STRIPE_SECTORS) {
2299 struct bio *nextbi =
2300 r5_next_bio(bi, sh->dev[i].sector);
2301 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2302 if (!raid5_dec_bi_phys_segments(bi)) {
2303 bi->bi_next = *return_bi;
2309 spin_unlock_irq(&conf->device_lock);
2311 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2312 STRIPE_SECTORS, 0, 0);
2315 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2316 if (atomic_dec_and_test(&conf->pending_full_writes))
2317 md_wakeup_thread(conf->mddev->thread);
2320 /* fetch_block5 - checks the given member device to see if its data needs
2321 * to be read or computed to satisfy a request.
2323 * Returns 1 when no more member devices need to be checked, otherwise returns
2324 * 0 to tell the loop in handle_stripe_fill5 to continue
2326 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2327 int disk_idx, int disks)
2329 struct r5dev *dev = &sh->dev[disk_idx];
2330 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2332 /* is the data in this block needed, and can we get it? */
2333 if (!test_bit(R5_LOCKED, &dev->flags) &&
2334 !test_bit(R5_UPTODATE, &dev->flags) &&
2336 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2337 s->syncing || s->expanding ||
2339 (failed_dev->toread ||
2340 (failed_dev->towrite &&
2341 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2342 /* We would like to get this block, possibly by computing it,
2343 * otherwise read it if the backing disk is insync
2345 if ((s->uptodate == disks - 1) &&
2346 (s->failed && disk_idx == s->failed_num)) {
2347 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2348 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2349 set_bit(R5_Wantcompute, &dev->flags);
2350 sh->ops.target = disk_idx;
2351 sh->ops.target2 = -1;
2353 /* Careful: from this point on 'uptodate' is in the eye
2354 * of raid_run_ops which services 'compute' operations
2355 * before writes. R5_Wantcompute flags a block that will
2356 * be R5_UPTODATE by the time it is needed for a
2357 * subsequent operation.
2360 return 1; /* uptodate + compute == disks */
2361 } else if (test_bit(R5_Insync, &dev->flags)) {
2362 set_bit(R5_LOCKED, &dev->flags);
2363 set_bit(R5_Wantread, &dev->flags);
2365 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2374 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2376 static void handle_stripe_fill5(struct stripe_head *sh,
2377 struct stripe_head_state *s, int disks)
2381 /* look for blocks to read/compute, skip this if a compute
2382 * is already in flight, or if the stripe contents are in the
2383 * midst of changing due to a write
2385 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2386 !sh->reconstruct_state)
2387 for (i = disks; i--; )
2388 if (fetch_block5(sh, s, i, disks))
2390 set_bit(STRIPE_HANDLE, &sh->state);
2393 /* fetch_block6 - checks the given member device to see if its data needs
2394 * to be read or computed to satisfy a request.
2396 * Returns 1 when no more member devices need to be checked, otherwise returns
2397 * 0 to tell the loop in handle_stripe_fill6 to continue
2399 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2400 struct r6_state *r6s, int disk_idx, int disks)
2402 struct r5dev *dev = &sh->dev[disk_idx];
2403 struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2404 &sh->dev[r6s->failed_num[1]] };
2406 if (!test_bit(R5_LOCKED, &dev->flags) &&
2407 !test_bit(R5_UPTODATE, &dev->flags) &&
2409 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2410 s->syncing || s->expanding ||
2412 (fdev[0]->toread || s->to_write)) ||
2414 (fdev[1]->toread || s->to_write)))) {
2415 /* we would like to get this block, possibly by computing it,
2416 * otherwise read it if the backing disk is insync
2418 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2419 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2420 if ((s->uptodate == disks - 1) &&
2421 (s->failed && (disk_idx == r6s->failed_num[0] ||
2422 disk_idx == r6s->failed_num[1]))) {
2423 /* have disk failed, and we're requested to fetch it;
2426 pr_debug("Computing stripe %llu block %d\n",
2427 (unsigned long long)sh->sector, disk_idx);
2428 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2429 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2430 set_bit(R5_Wantcompute, &dev->flags);
2431 sh->ops.target = disk_idx;
2432 sh->ops.target2 = -1; /* no 2nd target */
2436 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2437 /* Computing 2-failure is *very* expensive; only
2438 * do it if failed >= 2
2441 for (other = disks; other--; ) {
2442 if (other == disk_idx)
2444 if (!test_bit(R5_UPTODATE,
2445 &sh->dev[other].flags))
2449 pr_debug("Computing stripe %llu blocks %d,%d\n",
2450 (unsigned long long)sh->sector,
2452 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2453 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2454 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2455 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2456 sh->ops.target = disk_idx;
2457 sh->ops.target2 = other;
2461 } else if (test_bit(R5_Insync, &dev->flags)) {
2462 set_bit(R5_LOCKED, &dev->flags);
2463 set_bit(R5_Wantread, &dev->flags);
2465 pr_debug("Reading block %d (sync=%d)\n",
2466 disk_idx, s->syncing);
2474 * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2476 static void handle_stripe_fill6(struct stripe_head *sh,
2477 struct stripe_head_state *s, struct r6_state *r6s,
2482 /* look for blocks to read/compute, skip this if a compute
2483 * is already in flight, or if the stripe contents are in the
2484 * midst of changing due to a write
2486 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2487 !sh->reconstruct_state)
2488 for (i = disks; i--; )
2489 if (fetch_block6(sh, s, r6s, i, disks))
2491 set_bit(STRIPE_HANDLE, &sh->state);
2495 /* handle_stripe_clean_event
2496 * any written block on an uptodate or failed drive can be returned.
2497 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2498 * never LOCKED, so we don't need to test 'failed' directly.
2500 static void handle_stripe_clean_event(raid5_conf_t *conf,
2501 struct stripe_head *sh, int disks, struct bio **return_bi)
2506 for (i = disks; i--; )
2507 if (sh->dev[i].written) {
2509 if (!test_bit(R5_LOCKED, &dev->flags) &&
2510 test_bit(R5_UPTODATE, &dev->flags)) {
2511 /* We can return any write requests */
2512 struct bio *wbi, *wbi2;
2514 pr_debug("Return write for disc %d\n", i);
2515 spin_lock_irq(&conf->device_lock);
2517 dev->written = NULL;
2518 while (wbi && wbi->bi_sector <
2519 dev->sector + STRIPE_SECTORS) {
2520 wbi2 = r5_next_bio(wbi, dev->sector);
2521 if (!raid5_dec_bi_phys_segments(wbi)) {
2522 md_write_end(conf->mddev);
2523 wbi->bi_next = *return_bi;
2528 if (dev->towrite == NULL)
2530 spin_unlock_irq(&conf->device_lock);
2532 bitmap_endwrite(conf->mddev->bitmap,
2535 !test_bit(STRIPE_DEGRADED, &sh->state),
2540 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2541 if (atomic_dec_and_test(&conf->pending_full_writes))
2542 md_wakeup_thread(conf->mddev->thread);
2545 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2546 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2548 int rmw = 0, rcw = 0, i;
2549 for (i = disks; i--; ) {
2550 /* would I have to read this buffer for read_modify_write */
2551 struct r5dev *dev = &sh->dev[i];
2552 if ((dev->towrite || i == sh->pd_idx) &&
2553 !test_bit(R5_LOCKED, &dev->flags) &&
2554 !(test_bit(R5_UPTODATE, &dev->flags) ||
2555 test_bit(R5_Wantcompute, &dev->flags))) {
2556 if (test_bit(R5_Insync, &dev->flags))
2559 rmw += 2*disks; /* cannot read it */
2561 /* Would I have to read this buffer for reconstruct_write */
2562 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2563 !test_bit(R5_LOCKED, &dev->flags) &&
2564 !(test_bit(R5_UPTODATE, &dev->flags) ||
2565 test_bit(R5_Wantcompute, &dev->flags))) {
2566 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2571 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2572 (unsigned long long)sh->sector, rmw, rcw);
2573 set_bit(STRIPE_HANDLE, &sh->state);
2574 if (rmw < rcw && rmw > 0)
2575 /* prefer read-modify-write, but need to get some data */
2576 for (i = disks; i--; ) {
2577 struct r5dev *dev = &sh->dev[i];
2578 if ((dev->towrite || i == sh->pd_idx) &&
2579 !test_bit(R5_LOCKED, &dev->flags) &&
2580 !(test_bit(R5_UPTODATE, &dev->flags) ||
2581 test_bit(R5_Wantcompute, &dev->flags)) &&
2582 test_bit(R5_Insync, &dev->flags)) {
2584 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2585 pr_debug("Read_old block "
2586 "%d for r-m-w\n", i);
2587 set_bit(R5_LOCKED, &dev->flags);
2588 set_bit(R5_Wantread, &dev->flags);
2591 set_bit(STRIPE_DELAYED, &sh->state);
2592 set_bit(STRIPE_HANDLE, &sh->state);
2596 if (rcw <= rmw && rcw > 0)
2597 /* want reconstruct write, but need to get some data */
2598 for (i = disks; i--; ) {
2599 struct r5dev *dev = &sh->dev[i];
2600 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2602 !test_bit(R5_LOCKED, &dev->flags) &&
2603 !(test_bit(R5_UPTODATE, &dev->flags) ||
2604 test_bit(R5_Wantcompute, &dev->flags)) &&
2605 test_bit(R5_Insync, &dev->flags)) {
2607 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2608 pr_debug("Read_old block "
2609 "%d for Reconstruct\n", i);
2610 set_bit(R5_LOCKED, &dev->flags);
2611 set_bit(R5_Wantread, &dev->flags);
2614 set_bit(STRIPE_DELAYED, &sh->state);
2615 set_bit(STRIPE_HANDLE, &sh->state);
2619 /* now if nothing is locked, and if we have enough data,
2620 * we can start a write request
2622 /* since handle_stripe can be called at any time we need to handle the
2623 * case where a compute block operation has been submitted and then a
2624 * subsequent call wants to start a write request. raid_run_ops only
2625 * handles the case where compute block and reconstruct are requested
2626 * simultaneously. If this is not the case then new writes need to be
2627 * held off until the compute completes.
2629 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2630 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2631 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2632 schedule_reconstruction(sh, s, rcw == 0, 0);
2635 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2636 struct stripe_head *sh, struct stripe_head_state *s,
2637 struct r6_state *r6s, int disks)
2639 int rcw = 0, pd_idx = sh->pd_idx, i;
2640 int qd_idx = sh->qd_idx;
2642 set_bit(STRIPE_HANDLE, &sh->state);
2643 for (i = disks; i--; ) {
2644 struct r5dev *dev = &sh->dev[i];
2645 /* check if we haven't enough data */
2646 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2647 i != pd_idx && i != qd_idx &&
2648 !test_bit(R5_LOCKED, &dev->flags) &&
2649 !(test_bit(R5_UPTODATE, &dev->flags) ||
2650 test_bit(R5_Wantcompute, &dev->flags))) {
2652 if (!test_bit(R5_Insync, &dev->flags))
2653 continue; /* it's a failed drive */
2656 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2657 pr_debug("Read_old stripe %llu "
2658 "block %d for Reconstruct\n",
2659 (unsigned long long)sh->sector, i);
2660 set_bit(R5_LOCKED, &dev->flags);
2661 set_bit(R5_Wantread, &dev->flags);
2664 pr_debug("Request delayed stripe %llu "
2665 "block %d for Reconstruct\n",
2666 (unsigned long long)sh->sector, i);
2667 set_bit(STRIPE_DELAYED, &sh->state);
2668 set_bit(STRIPE_HANDLE, &sh->state);
2672 /* now if nothing is locked, and if we have enough data, we can start a
2675 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2676 s->locked == 0 && rcw == 0 &&
2677 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2678 schedule_reconstruction(sh, s, 1, 0);
2682 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2683 struct stripe_head_state *s, int disks)
2685 struct r5dev *dev = NULL;
2687 set_bit(STRIPE_HANDLE, &sh->state);
2689 switch (sh->check_state) {
2690 case check_state_idle:
2691 /* start a new check operation if there are no failures */
2692 if (s->failed == 0) {
2693 BUG_ON(s->uptodate != disks);
2694 sh->check_state = check_state_run;
2695 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2696 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2700 dev = &sh->dev[s->failed_num];
2702 case check_state_compute_result:
2703 sh->check_state = check_state_idle;
2705 dev = &sh->dev[sh->pd_idx];
2707 /* check that a write has not made the stripe insync */
2708 if (test_bit(STRIPE_INSYNC, &sh->state))
2711 /* either failed parity check, or recovery is happening */
2712 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2713 BUG_ON(s->uptodate != disks);
2715 set_bit(R5_LOCKED, &dev->flags);
2717 set_bit(R5_Wantwrite, &dev->flags);
2719 clear_bit(STRIPE_DEGRADED, &sh->state);
2720 set_bit(STRIPE_INSYNC, &sh->state);
2722 case check_state_run:
2723 break; /* we will be called again upon completion */
2724 case check_state_check_result:
2725 sh->check_state = check_state_idle;
2727 /* if a failure occurred during the check operation, leave
2728 * STRIPE_INSYNC not set and let the stripe be handled again
2733 /* handle a successful check operation, if parity is correct
2734 * we are done. Otherwise update the mismatch count and repair
2735 * parity if !MD_RECOVERY_CHECK
2737 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2738 /* parity is correct (on disc,
2739 * not in buffer any more)
2741 set_bit(STRIPE_INSYNC, &sh->state);
2743 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2744 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2745 /* don't try to repair!! */
2746 set_bit(STRIPE_INSYNC, &sh->state);
2748 sh->check_state = check_state_compute_run;
2749 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2750 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2751 set_bit(R5_Wantcompute,
2752 &sh->dev[sh->pd_idx].flags);
2753 sh->ops.target = sh->pd_idx;
2754 sh->ops.target2 = -1;
2759 case check_state_compute_run:
2762 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2763 __func__, sh->check_state,
2764 (unsigned long long) sh->sector);
2770 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2771 struct stripe_head_state *s,
2772 struct r6_state *r6s, int disks)
2774 int pd_idx = sh->pd_idx;
2775 int qd_idx = sh->qd_idx;
2778 set_bit(STRIPE_HANDLE, &sh->state);
2780 BUG_ON(s->failed > 2);
2782 /* Want to check and possibly repair P and Q.
2783 * However there could be one 'failed' device, in which
2784 * case we can only check one of them, possibly using the
2785 * other to generate missing data
2788 switch (sh->check_state) {
2789 case check_state_idle:
2790 /* start a new check operation if there are < 2 failures */
2791 if (s->failed == r6s->q_failed) {
2792 /* The only possible failed device holds Q, so it
2793 * makes sense to check P (If anything else were failed,
2794 * we would have used P to recreate it).
2796 sh->check_state = check_state_run;
2798 if (!r6s->q_failed && s->failed < 2) {
2799 /* Q is not failed, and we didn't use it to generate
2800 * anything, so it makes sense to check it
2802 if (sh->check_state == check_state_run)
2803 sh->check_state = check_state_run_pq;
2805 sh->check_state = check_state_run_q;
2808 /* discard potentially stale zero_sum_result */
2809 sh->ops.zero_sum_result = 0;
2811 if (sh->check_state == check_state_run) {
2812 /* async_xor_zero_sum destroys the contents of P */
2813 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2816 if (sh->check_state >= check_state_run &&
2817 sh->check_state <= check_state_run_pq) {
2818 /* async_syndrome_zero_sum preserves P and Q, so
2819 * no need to mark them !uptodate here
2821 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2825 /* we have 2-disk failure */
2826 BUG_ON(s->failed != 2);
2828 case check_state_compute_result:
2829 sh->check_state = check_state_idle;
2831 /* check that a write has not made the stripe insync */
2832 if (test_bit(STRIPE_INSYNC, &sh->state))
2835 /* now write out any block on a failed drive,
2836 * or P or Q if they were recomputed
2838 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2839 if (s->failed == 2) {
2840 dev = &sh->dev[r6s->failed_num[1]];
2842 set_bit(R5_LOCKED, &dev->flags);
2843 set_bit(R5_Wantwrite, &dev->flags);
2845 if (s->failed >= 1) {
2846 dev = &sh->dev[r6s->failed_num[0]];
2848 set_bit(R5_LOCKED, &dev->flags);
2849 set_bit(R5_Wantwrite, &dev->flags);
2851 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2852 dev = &sh->dev[pd_idx];
2854 set_bit(R5_LOCKED, &dev->flags);
2855 set_bit(R5_Wantwrite, &dev->flags);
2857 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2858 dev = &sh->dev[qd_idx];
2860 set_bit(R5_LOCKED, &dev->flags);
2861 set_bit(R5_Wantwrite, &dev->flags);
2863 clear_bit(STRIPE_DEGRADED, &sh->state);
2865 set_bit(STRIPE_INSYNC, &sh->state);
2867 case check_state_run:
2868 case check_state_run_q:
2869 case check_state_run_pq:
2870 break; /* we will be called again upon completion */
2871 case check_state_check_result:
2872 sh->check_state = check_state_idle;
2874 /* handle a successful check operation, if parity is correct
2875 * we are done. Otherwise update the mismatch count and repair
2876 * parity if !MD_RECOVERY_CHECK
2878 if (sh->ops.zero_sum_result == 0) {
2879 /* both parities are correct */
2881 set_bit(STRIPE_INSYNC, &sh->state);
2883 /* in contrast to the raid5 case we can validate
2884 * parity, but still have a failure to write
2887 sh->check_state = check_state_compute_result;
2888 /* Returning at this point means that we may go
2889 * off and bring p and/or q uptodate again so
2890 * we make sure to check zero_sum_result again
2891 * to verify if p or q need writeback
2895 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2896 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2897 /* don't try to repair!! */
2898 set_bit(STRIPE_INSYNC, &sh->state);
2900 int *target = &sh->ops.target;
2902 sh->ops.target = -1;
2903 sh->ops.target2 = -1;
2904 sh->check_state = check_state_compute_run;
2905 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2906 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2907 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2908 set_bit(R5_Wantcompute,
2909 &sh->dev[pd_idx].flags);
2911 target = &sh->ops.target2;
2914 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2915 set_bit(R5_Wantcompute,
2916 &sh->dev[qd_idx].flags);
2923 case check_state_compute_run:
2926 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2927 __func__, sh->check_state,
2928 (unsigned long long) sh->sector);
2933 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2934 struct r6_state *r6s)
2938 /* We have read all the blocks in this stripe and now we need to
2939 * copy some of them into a target stripe for expand.
2941 struct dma_async_tx_descriptor *tx = NULL;
2942 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2943 for (i = 0; i < sh->disks; i++)
2944 if (i != sh->pd_idx && i != sh->qd_idx) {
2946 struct stripe_head *sh2;
2947 struct async_submit_ctl submit;
2949 sector_t bn = compute_blocknr(sh, i, 1);
2950 sector_t s = raid5_compute_sector(conf, bn, 0,
2952 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2954 /* so far only the early blocks of this stripe
2955 * have been requested. When later blocks
2956 * get requested, we will try again
2959 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2960 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2961 /* must have already done this block */
2962 release_stripe(sh2);
2966 /* place all the copies on one channel */
2967 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2968 tx = async_memcpy(sh2->dev[dd_idx].page,
2969 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2972 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2973 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2974 for (j = 0; j < conf->raid_disks; j++)
2975 if (j != sh2->pd_idx &&
2976 (!r6s || j != sh2->qd_idx) &&
2977 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2979 if (j == conf->raid_disks) {
2980 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2981 set_bit(STRIPE_HANDLE, &sh2->state);
2983 release_stripe(sh2);
2986 /* done submitting copies, wait for them to complete */
2989 dma_wait_for_async_tx(tx);
2995 * handle_stripe - do things to a stripe.
2997 * We lock the stripe and then examine the state of various bits
2998 * to see what needs to be done.
3000 * return some read request which now have data
3001 * return some write requests which are safely on disc
3002 * schedule a read on some buffers
3003 * schedule a write of some buffers
3004 * return confirmation of parity correctness
3006 * buffers are taken off read_list or write_list, and bh_cache buffers
3007 * get BH_Lock set before the stripe lock is released.
3011 static void handle_stripe5(struct stripe_head *sh)
3013 raid5_conf_t *conf = sh->raid_conf;
3014 int disks = sh->disks, i;
3015 struct bio *return_bi = NULL;
3016 struct stripe_head_state s;
3018 mdk_rdev_t *blocked_rdev = NULL;
3020 int dec_preread_active = 0;
3022 memset(&s, 0, sizeof(s));
3023 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
3024 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
3025 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
3026 sh->reconstruct_state);
3028 spin_lock(&sh->lock);
3029 clear_bit(STRIPE_HANDLE, &sh->state);
3030 clear_bit(STRIPE_DELAYED, &sh->state);
3032 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3033 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3034 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3036 /* Now to look around and see what can be done */
3038 for (i=disks; i--; ) {
3043 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
3044 "written %p\n", i, dev->flags, dev->toread, dev->read,
3045 dev->towrite, dev->written);
3047 /* maybe we can request a biofill operation
3049 * new wantfill requests are only permitted while
3050 * ops_complete_biofill is guaranteed to be inactive
3052 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3053 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3054 set_bit(R5_Wantfill, &dev->flags);
3056 /* now count some things */
3057 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3058 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3059 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
3061 if (test_bit(R5_Wantfill, &dev->flags))
3063 else if (dev->toread)
3067 if (!test_bit(R5_OVERWRITE, &dev->flags))
3072 rdev = rcu_dereference(conf->disks[i].rdev);
3073 if (blocked_rdev == NULL &&
3074 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3075 blocked_rdev = rdev;
3076 atomic_inc(&rdev->nr_pending);
3078 clear_bit(R5_Insync, &dev->flags);
3081 else if (test_bit(In_sync, &rdev->flags))
3082 set_bit(R5_Insync, &dev->flags);
3083 else if (!test_bit(Faulty, &rdev->flags)) {
3084 /* could be in-sync depending on recovery/reshape status */
3085 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3086 set_bit(R5_Insync, &dev->flags);
3088 if (!test_bit(R5_Insync, &dev->flags)) {
3089 /* The ReadError flag will just be confusing now */
3090 clear_bit(R5_ReadError, &dev->flags);
3091 clear_bit(R5_ReWrite, &dev->flags);
3093 if (test_bit(R5_ReadError, &dev->flags))
3094 clear_bit(R5_Insync, &dev->flags);
3095 if (!test_bit(R5_Insync, &dev->flags)) {
3102 if (unlikely(blocked_rdev)) {
3103 if (s.syncing || s.expanding || s.expanded ||
3104 s.to_write || s.written) {
3105 set_bit(STRIPE_HANDLE, &sh->state);
3108 /* There is nothing for the blocked_rdev to block */
3109 rdev_dec_pending(blocked_rdev, conf->mddev);
3110 blocked_rdev = NULL;
3113 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3114 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3115 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3118 pr_debug("locked=%d uptodate=%d to_read=%d"
3119 " to_write=%d failed=%d failed_num=%d\n",
3120 s.locked, s.uptodate, s.to_read, s.to_write,
3121 s.failed, s.failed_num);
3122 /* check if the array has lost two devices and, if so, some requests might
3126 sh->check_state = 0;
3127 sh->reconstruct_state = 0;
3128 if (s.to_read+s.to_write+s.written)
3129 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3131 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3132 clear_bit(STRIPE_SYNCING, &sh->state);
3137 /* might be able to return some write requests if the parity block
3138 * is safe, or on a failed drive
3140 dev = &sh->dev[sh->pd_idx];
3142 ((test_bit(R5_Insync, &dev->flags) &&
3143 !test_bit(R5_LOCKED, &dev->flags) &&
3144 test_bit(R5_UPTODATE, &dev->flags)) ||
3145 (s.failed == 1 && s.failed_num == sh->pd_idx)))
3146 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3148 /* Now we might consider reading some blocks, either to check/generate
3149 * parity, or to satisfy requests
3150 * or to load a block that is being partially written.
3152 if (s.to_read || s.non_overwrite ||
3153 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3154 handle_stripe_fill5(sh, &s, disks);
3156 /* Now we check to see if any write operations have recently
3160 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3162 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3163 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3164 sh->reconstruct_state = reconstruct_state_idle;
3166 /* All the 'written' buffers and the parity block are ready to
3167 * be written back to disk
3169 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3170 for (i = disks; i--; ) {
3172 if (test_bit(R5_LOCKED, &dev->flags) &&
3173 (i == sh->pd_idx || dev->written)) {
3174 pr_debug("Writing block %d\n", i);
3175 set_bit(R5_Wantwrite, &dev->flags);
3178 if (!test_bit(R5_Insync, &dev->flags) ||
3179 (i == sh->pd_idx && s.failed == 0))
3180 set_bit(STRIPE_INSYNC, &sh->state);
3183 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3184 dec_preread_active = 1;
3187 /* Now to consider new write requests and what else, if anything
3188 * should be read. We do not handle new writes when:
3189 * 1/ A 'write' operation (copy+xor) is already in flight.
3190 * 2/ A 'check' operation is in flight, as it may clobber the parity
3193 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3194 handle_stripe_dirtying5(conf, sh, &s, disks);
3196 /* maybe we need to check and possibly fix the parity for this stripe
3197 * Any reads will already have been scheduled, so we just see if enough
3198 * data is available. The parity check is held off while parity
3199 * dependent operations are in flight.
3201 if (sh->check_state ||
3202 (s.syncing && s.locked == 0 &&
3203 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3204 !test_bit(STRIPE_INSYNC, &sh->state)))
3205 handle_parity_checks5(conf, sh, &s, disks);
3207 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3208 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3209 clear_bit(STRIPE_SYNCING, &sh->state);
3212 /* If the failed drive is just a ReadError, then we might need to progress
3213 * the repair/check process
3215 if (s.failed == 1 && !conf->mddev->ro &&
3216 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3217 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3218 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3220 dev = &sh->dev[s.failed_num];
3221 if (!test_bit(R5_ReWrite, &dev->flags)) {
3222 set_bit(R5_Wantwrite, &dev->flags);
3223 set_bit(R5_ReWrite, &dev->flags);
3224 set_bit(R5_LOCKED, &dev->flags);
3227 /* let's read it back */
3228 set_bit(R5_Wantread, &dev->flags);
3229 set_bit(R5_LOCKED, &dev->flags);
3234 /* Finish reconstruct operations initiated by the expansion process */
3235 if (sh->reconstruct_state == reconstruct_state_result) {
3236 struct stripe_head *sh2
3237 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3238 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3239 /* sh cannot be written until sh2 has been read.
3240 * so arrange for sh to be delayed a little
3242 set_bit(STRIPE_DELAYED, &sh->state);
3243 set_bit(STRIPE_HANDLE, &sh->state);
3244 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3246 atomic_inc(&conf->preread_active_stripes);
3247 release_stripe(sh2);
3251 release_stripe(sh2);
3253 sh->reconstruct_state = reconstruct_state_idle;
3254 clear_bit(STRIPE_EXPANDING, &sh->state);
3255 for (i = conf->raid_disks; i--; ) {
3256 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3257 set_bit(R5_LOCKED, &sh->dev[i].flags);
3262 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3263 !sh->reconstruct_state) {
3264 /* Need to write out all blocks after computing parity */
3265 sh->disks = conf->raid_disks;
3266 stripe_set_idx(sh->sector, conf, 0, sh);
3267 schedule_reconstruction(sh, &s, 1, 1);
3268 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3269 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3270 atomic_dec(&conf->reshape_stripes);
3271 wake_up(&conf->wait_for_overlap);
3272 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3275 if (s.expanding && s.locked == 0 &&
3276 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3277 handle_stripe_expansion(conf, sh, NULL);
3280 spin_unlock(&sh->lock);
3282 /* wait for this device to become unblocked */
3283 if (unlikely(blocked_rdev))
3284 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3287 raid_run_ops(sh, s.ops_request);
3291 if (dec_preread_active) {
3292 /* We delay this until after ops_run_io so that if make_request
3293 * is waiting on a flush, it won't continue until the writes
3294 * have actually been submitted.
3296 atomic_dec(&conf->preread_active_stripes);
3297 if (atomic_read(&conf->preread_active_stripes) <
3299 md_wakeup_thread(conf->mddev->thread);
3301 return_io(return_bi);
3304 static void handle_stripe6(struct stripe_head *sh)
3306 raid5_conf_t *conf = sh->raid_conf;
3307 int disks = sh->disks;
3308 struct bio *return_bi = NULL;
3309 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3310 struct stripe_head_state s;
3311 struct r6_state r6s;
3312 struct r5dev *dev, *pdev, *qdev;
3313 mdk_rdev_t *blocked_rdev = NULL;
3314 int dec_preread_active = 0;
3316 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3317 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3318 (unsigned long long)sh->sector, sh->state,
3319 atomic_read(&sh->count), pd_idx, qd_idx,
3320 sh->check_state, sh->reconstruct_state);
3321 memset(&s, 0, sizeof(s));
3323 spin_lock(&sh->lock);
3324 clear_bit(STRIPE_HANDLE, &sh->state);
3325 clear_bit(STRIPE_DELAYED, &sh->state);
3327 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3328 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3329 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3330 /* Now to look around and see what can be done */
3333 for (i=disks; i--; ) {
3337 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3338 i, dev->flags, dev->toread, dev->towrite, dev->written);
3339 /* maybe we can reply to a read
3341 * new wantfill requests are only permitted while
3342 * ops_complete_biofill is guaranteed to be inactive
3344 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3345 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3346 set_bit(R5_Wantfill, &dev->flags);
3348 /* now count some things */
3349 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3350 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3351 if (test_bit(R5_Wantcompute, &dev->flags)) {
3353 BUG_ON(s.compute > 2);
3356 if (test_bit(R5_Wantfill, &dev->flags)) {
3358 } else if (dev->toread)
3362 if (!test_bit(R5_OVERWRITE, &dev->flags))
3367 rdev = rcu_dereference(conf->disks[i].rdev);
3368 if (blocked_rdev == NULL &&
3369 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3370 blocked_rdev = rdev;
3371 atomic_inc(&rdev->nr_pending);
3373 clear_bit(R5_Insync, &dev->flags);
3376 else if (test_bit(In_sync, &rdev->flags))
3377 set_bit(R5_Insync, &dev->flags);
3378 else if (!test_bit(Faulty, &rdev->flags)) {
3379 /* in sync if before recovery_offset */
3380 if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3381 set_bit(R5_Insync, &dev->flags);
3383 if (!test_bit(R5_Insync, &dev->flags)) {
3384 /* The ReadError flag will just be confusing now */
3385 clear_bit(R5_ReadError, &dev->flags);
3386 clear_bit(R5_ReWrite, &dev->flags);
3388 if (test_bit(R5_ReadError, &dev->flags))
3389 clear_bit(R5_Insync, &dev->flags);
3390 if (!test_bit(R5_Insync, &dev->flags)) {
3392 r6s.failed_num[s.failed] = i;
3398 if (unlikely(blocked_rdev)) {
3399 if (s.syncing || s.expanding || s.expanded ||
3400 s.to_write || s.written) {
3401 set_bit(STRIPE_HANDLE, &sh->state);
3404 /* There is nothing for the blocked_rdev to block */
3405 rdev_dec_pending(blocked_rdev, conf->mddev);
3406 blocked_rdev = NULL;
3409 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3410 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3411 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3414 pr_debug("locked=%d uptodate=%d to_read=%d"
3415 " to_write=%d failed=%d failed_num=%d,%d\n",
3416 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3417 r6s.failed_num[0], r6s.failed_num[1]);
3418 /* check if the array has lost >2 devices and, if so, some requests
3419 * might need to be failed
3422 sh->check_state = 0;
3423 sh->reconstruct_state = 0;
3424 if (s.to_read+s.to_write+s.written)
3425 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3427 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3428 clear_bit(STRIPE_SYNCING, &sh->state);
3434 * might be able to return some write requests if the parity blocks
3435 * are safe, or on a failed drive
3437 pdev = &sh->dev[pd_idx];
3438 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3439 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3440 qdev = &sh->dev[qd_idx];
3441 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3442 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3445 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3446 && !test_bit(R5_LOCKED, &pdev->flags)
3447 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3448 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3449 && !test_bit(R5_LOCKED, &qdev->flags)
3450 && test_bit(R5_UPTODATE, &qdev->flags)))))
3451 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3453 /* Now we might consider reading some blocks, either to check/generate
3454 * parity, or to satisfy requests
3455 * or to load a block that is being partially written.
3457 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3458 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3459 handle_stripe_fill6(sh, &s, &r6s, disks);
3461 /* Now we check to see if any write operations have recently
3464 if (sh->reconstruct_state == reconstruct_state_drain_result) {
3466 sh->reconstruct_state = reconstruct_state_idle;
3467 /* All the 'written' buffers and the parity blocks are ready to
3468 * be written back to disk
3470 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3471 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3472 for (i = disks; i--; ) {
3474 if (test_bit(R5_LOCKED, &dev->flags) &&
3475 (i == sh->pd_idx || i == qd_idx ||
3477 pr_debug("Writing block %d\n", i);
3478 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3479 set_bit(R5_Wantwrite, &dev->flags);
3480 if (!test_bit(R5_Insync, &dev->flags) ||
3481 ((i == sh->pd_idx || i == qd_idx) &&
3483 set_bit(STRIPE_INSYNC, &sh->state);
3486 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3487 dec_preread_active = 1;
3490 /* Now to consider new write requests and what else, if anything
3491 * should be read. We do not handle new writes when:
3492 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3493 * 2/ A 'check' operation is in flight, as it may clobber the parity
3496 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3497 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3499 /* maybe we need to check and possibly fix the parity for this stripe
3500 * Any reads will already have been scheduled, so we just see if enough
3501 * data is available. The parity check is held off while parity
3502 * dependent operations are in flight.
3504 if (sh->check_state ||
3505 (s.syncing && s.locked == 0 &&
3506 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3507 !test_bit(STRIPE_INSYNC, &sh->state)))
3508 handle_parity_checks6(conf, sh, &s, &r6s, disks);
3510 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3511 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3512 clear_bit(STRIPE_SYNCING, &sh->state);
3515 /* If the failed drives are just a ReadError, then we might need
3516 * to progress the repair/check process
3518 if (s.failed <= 2 && !conf->mddev->ro)
3519 for (i = 0; i < s.failed; i++) {
3520 dev = &sh->dev[r6s.failed_num[i]];
3521 if (test_bit(R5_ReadError, &dev->flags)
3522 && !test_bit(R5_LOCKED, &dev->flags)
3523 && test_bit(R5_UPTODATE, &dev->flags)
3525 if (!test_bit(R5_ReWrite, &dev->flags)) {
3526 set_bit(R5_Wantwrite, &dev->flags);
3527 set_bit(R5_ReWrite, &dev->flags);
3528 set_bit(R5_LOCKED, &dev->flags);
3531 /* let's read it back */
3532 set_bit(R5_Wantread, &dev->flags);
3533 set_bit(R5_LOCKED, &dev->flags);
3539 /* Finish reconstruct operations initiated by the expansion process */
3540 if (sh->reconstruct_state == reconstruct_state_result) {
3541 sh->reconstruct_state = reconstruct_state_idle;
3542 clear_bit(STRIPE_EXPANDING, &sh->state);
3543 for (i = conf->raid_disks; i--; ) {
3544 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3545 set_bit(R5_LOCKED, &sh->dev[i].flags);
3550 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3551 !sh->reconstruct_state) {
3552 struct stripe_head *sh2
3553 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3554 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3555 /* sh cannot be written until sh2 has been read.
3556 * so arrange for sh to be delayed a little
3558 set_bit(STRIPE_DELAYED, &sh->state);
3559 set_bit(STRIPE_HANDLE, &sh->state);
3560 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3562 atomic_inc(&conf->preread_active_stripes);
3563 release_stripe(sh2);
3567 release_stripe(sh2);
3569 /* Need to write out all blocks after computing P&Q */
3570 sh->disks = conf->raid_disks;
3571 stripe_set_idx(sh->sector, conf, 0, sh);
3572 schedule_reconstruction(sh, &s, 1, 1);
3573 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3574 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3575 atomic_dec(&conf->reshape_stripes);
3576 wake_up(&conf->wait_for_overlap);
3577 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3580 if (s.expanding && s.locked == 0 &&
3581 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3582 handle_stripe_expansion(conf, sh, &r6s);
3585 spin_unlock(&sh->lock);
3587 /* wait for this device to become unblocked */
3588 if (unlikely(blocked_rdev))
3589 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3592 raid_run_ops(sh, s.ops_request);
3597 if (dec_preread_active) {
3598 /* We delay this until after ops_run_io so that if make_request
3599 * is waiting on a flush, it won't continue until the writes
3600 * have actually been submitted.
3602 atomic_dec(&conf->preread_active_stripes);
3603 if (atomic_read(&conf->preread_active_stripes) <
3605 md_wakeup_thread(conf->mddev->thread);
3608 return_io(return_bi);
3611 static void handle_stripe(struct stripe_head *sh)
3613 if (sh->raid_conf->level == 6)
3619 static void raid5_activate_delayed(raid5_conf_t *conf)
3621 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3622 while (!list_empty(&conf->delayed_list)) {
3623 struct list_head *l = conf->delayed_list.next;
3624 struct stripe_head *sh;
3625 sh = list_entry(l, struct stripe_head, lru);
3627 clear_bit(STRIPE_DELAYED, &sh->state);
3628 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3629 atomic_inc(&conf->preread_active_stripes);
3630 list_add_tail(&sh->lru, &conf->hold_list);
3635 static void activate_bit_delay(raid5_conf_t *conf)
3637 /* device_lock is held */
3638 struct list_head head;
3639 list_add(&head, &conf->bitmap_list);
3640 list_del_init(&conf->bitmap_list);
3641 while (!list_empty(&head)) {
3642 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3643 list_del_init(&sh->lru);
3644 atomic_inc(&sh->count);
3645 __release_stripe(conf, sh);
3649 int md_raid5_congested(mddev_t *mddev, int bits)
3651 raid5_conf_t *conf = mddev->private;
3653 /* No difference between reads and writes. Just check
3654 * how busy the stripe_cache is
3657 if (conf->inactive_blocked)
3661 if (list_empty_careful(&conf->inactive_list))
3666 EXPORT_SYMBOL_GPL(md_raid5_congested);
3668 static int raid5_congested(void *data, int bits)
3670 mddev_t *mddev = data;
3672 return mddev_congested(mddev, bits) ||
3673 md_raid5_congested(mddev, bits);
3676 /* We want read requests to align with chunks where possible,
3677 * but write requests don't need to.
3679 static int raid5_mergeable_bvec(struct request_queue *q,
3680 struct bvec_merge_data *bvm,
3681 struct bio_vec *biovec)
3683 mddev_t *mddev = q->queuedata;
3684 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3686 unsigned int chunk_sectors = mddev->chunk_sectors;
3687 unsigned int bio_sectors = bvm->bi_size >> 9;
3689 if ((bvm->bi_rw & 1) == WRITE)
3690 return biovec->bv_len; /* always allow writes to be mergeable */
3692 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3693 chunk_sectors = mddev->new_chunk_sectors;
3694 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3695 if (max < 0) max = 0;
3696 if (max <= biovec->bv_len && bio_sectors == 0)
3697 return biovec->bv_len;
3703 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3705 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3706 unsigned int chunk_sectors = mddev->chunk_sectors;
3707 unsigned int bio_sectors = bio->bi_size >> 9;
3709 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3710 chunk_sectors = mddev->new_chunk_sectors;
3711 return chunk_sectors >=
3712 ((sector & (chunk_sectors - 1)) + bio_sectors);
3716 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3717 * later sampled by raid5d.
3719 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3721 unsigned long flags;
3723 spin_lock_irqsave(&conf->device_lock, flags);
3725 bi->bi_next = conf->retry_read_aligned_list;
3726 conf->retry_read_aligned_list = bi;
3728 spin_unlock_irqrestore(&conf->device_lock, flags);
3729 md_wakeup_thread(conf->mddev->thread);
3733 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3737 bi = conf->retry_read_aligned;
3739 conf->retry_read_aligned = NULL;
3742 bi = conf->retry_read_aligned_list;
3744 conf->retry_read_aligned_list = bi->bi_next;
3747 * this sets the active strip count to 1 and the processed
3748 * strip count to zero (upper 8 bits)
3750 bi->bi_phys_segments = 1; /* biased count of active stripes */
3758 * The "raid5_align_endio" should check if the read succeeded and if it
3759 * did, call bio_endio on the original bio (having bio_put the new bio
3761 * If the read failed..
3763 static void raid5_align_endio(struct bio *bi, int error)
3765 struct bio* raid_bi = bi->bi_private;
3768 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3773 rdev = (void*)raid_bi->bi_next;
3774 raid_bi->bi_next = NULL;
3775 mddev = rdev->mddev;
3776 conf = mddev->private;
3778 rdev_dec_pending(rdev, conf->mddev);
3780 if (!error && uptodate) {
3781 bio_endio(raid_bi, 0);
3782 if (atomic_dec_and_test(&conf->active_aligned_reads))
3783 wake_up(&conf->wait_for_stripe);
3788 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3790 add_bio_to_retry(raid_bi, conf);
3793 static int bio_fits_rdev(struct bio *bi)
3795 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3797 if ((bi->bi_size>>9) > queue_max_sectors(q))
3799 blk_recount_segments(q, bi);
3800 if (bi->bi_phys_segments > queue_max_segments(q))
3803 if (q->merge_bvec_fn)
3804 /* it's too hard to apply the merge_bvec_fn at this stage,
3813 static int chunk_aligned_read(mddev_t *mddev, struct bio * raid_bio)
3815 raid5_conf_t *conf = mddev->private;
3817 struct bio* align_bi;
3820 if (!in_chunk_boundary(mddev, raid_bio)) {
3821 pr_debug("chunk_aligned_read : non aligned\n");
3825 * use bio_clone_mddev to make a copy of the bio
3827 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3831 * set bi_end_io to a new function, and set bi_private to the
3834 align_bi->bi_end_io = raid5_align_endio;
3835 align_bi->bi_private = raid_bio;
3839 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3844 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3845 if (rdev && test_bit(In_sync, &rdev->flags)) {
3846 atomic_inc(&rdev->nr_pending);
3848 raid_bio->bi_next = (void*)rdev;
3849 align_bi->bi_bdev = rdev->bdev;
3850 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3852 if (!bio_fits_rdev(align_bi)) {
3853 /* too big in some way */
3855 rdev_dec_pending(rdev, mddev);
3859 /* No reshape active, so we can trust rdev->data_offset */
3860 align_bi->bi_sector += rdev->data_offset;
3862 spin_lock_irq(&conf->device_lock);
3863 wait_event_lock_irq(conf->wait_for_stripe,
3865 conf->device_lock, /* nothing */);
3866 atomic_inc(&conf->active_aligned_reads);
3867 spin_unlock_irq(&conf->device_lock);
3869 generic_make_request(align_bi);
3878 /* __get_priority_stripe - get the next stripe to process
3880 * Full stripe writes are allowed to pass preread active stripes up until
3881 * the bypass_threshold is exceeded. In general the bypass_count
3882 * increments when the handle_list is handled before the hold_list; however, it
3883 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3884 * stripe with in flight i/o. The bypass_count will be reset when the
3885 * head of the hold_list has changed, i.e. the head was promoted to the
3888 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3890 struct stripe_head *sh;
3892 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3894 list_empty(&conf->handle_list) ? "empty" : "busy",
3895 list_empty(&conf->hold_list) ? "empty" : "busy",
3896 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3898 if (!list_empty(&conf->handle_list)) {
3899 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3901 if (list_empty(&conf->hold_list))
3902 conf->bypass_count = 0;
3903 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3904 if (conf->hold_list.next == conf->last_hold)
3905 conf->bypass_count++;
3907 conf->last_hold = conf->hold_list.next;
3908 conf->bypass_count -= conf->bypass_threshold;
3909 if (conf->bypass_count < 0)
3910 conf->bypass_count = 0;
3913 } else if (!list_empty(&conf->hold_list) &&
3914 ((conf->bypass_threshold &&
3915 conf->bypass_count > conf->bypass_threshold) ||
3916 atomic_read(&conf->pending_full_writes) == 0)) {
3917 sh = list_entry(conf->hold_list.next,
3919 conf->bypass_count -= conf->bypass_threshold;
3920 if (conf->bypass_count < 0)
3921 conf->bypass_count = 0;
3925 list_del_init(&sh->lru);
3926 atomic_inc(&sh->count);
3927 BUG_ON(atomic_read(&sh->count) != 1);
3931 static int make_request(mddev_t *mddev, struct bio * bi)
3933 raid5_conf_t *conf = mddev->private;
3935 sector_t new_sector;
3936 sector_t logical_sector, last_sector;
3937 struct stripe_head *sh;
3938 const int rw = bio_data_dir(bi);
3942 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3943 md_flush_request(mddev, bi);
3947 md_write_start(mddev, bi);
3950 mddev->reshape_position == MaxSector &&
3951 chunk_aligned_read(mddev,bi))
3954 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3955 last_sector = bi->bi_sector + (bi->bi_size>>9);
3957 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3959 plugged = mddev_check_plugged(mddev);
3960 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3962 int disks, data_disks;
3967 disks = conf->raid_disks;
3968 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3969 if (unlikely(conf->reshape_progress != MaxSector)) {
3970 /* spinlock is needed as reshape_progress may be
3971 * 64bit on a 32bit platform, and so it might be
3972 * possible to see a half-updated value
3973 * Of course reshape_progress could change after
3974 * the lock is dropped, so once we get a reference
3975 * to the stripe that we think it is, we will have
3978 spin_lock_irq(&conf->device_lock);
3979 if (mddev->delta_disks < 0
3980 ? logical_sector < conf->reshape_progress
3981 : logical_sector >= conf->reshape_progress) {
3982 disks = conf->previous_raid_disks;
3985 if (mddev->delta_disks < 0
3986 ? logical_sector < conf->reshape_safe
3987 : logical_sector >= conf->reshape_safe) {
3988 spin_unlock_irq(&conf->device_lock);
3993 spin_unlock_irq(&conf->device_lock);
3995 data_disks = disks - conf->max_degraded;
3997 new_sector = raid5_compute_sector(conf, logical_sector,
4000 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4001 (unsigned long long)new_sector,
4002 (unsigned long long)logical_sector);
4004 sh = get_active_stripe(conf, new_sector, previous,
4005 (bi->bi_rw&RWA_MASK), 0);
4007 if (unlikely(previous)) {
4008 /* expansion might have moved on while waiting for a
4009 * stripe, so we must do the range check again.
4010 * Expansion could still move past after this
4011 * test, but as we are holding a reference to
4012 * 'sh', we know that if that happens,
4013 * STRIPE_EXPANDING will get set and the expansion
4014 * won't proceed until we finish with the stripe.
4017 spin_lock_irq(&conf->device_lock);
4018 if (mddev->delta_disks < 0
4019 ? logical_sector >= conf->reshape_progress
4020 : logical_sector < conf->reshape_progress)
4021 /* mismatch, need to try again */
4023 spin_unlock_irq(&conf->device_lock);
4031 if (bio_data_dir(bi) == WRITE &&
4032 logical_sector >= mddev->suspend_lo &&
4033 logical_sector < mddev->suspend_hi) {
4035 /* As the suspend_* range is controlled by
4036 * userspace, we want an interruptible
4039 flush_signals(current);
4040 prepare_to_wait(&conf->wait_for_overlap,
4041 &w, TASK_INTERRUPTIBLE);
4042 if (logical_sector >= mddev->suspend_lo &&
4043 logical_sector < mddev->suspend_hi)
4048 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4049 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
4050 /* Stripe is busy expanding or
4051 * add failed due to overlap. Flush everything
4054 md_wakeup_thread(mddev->thread);
4059 finish_wait(&conf->wait_for_overlap, &w);
4060 set_bit(STRIPE_HANDLE, &sh->state);
4061 clear_bit(STRIPE_DELAYED, &sh->state);
4062 if ((bi->bi_rw & REQ_SYNC) &&
4063 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4064 atomic_inc(&conf->preread_active_stripes);
4067 /* cannot get stripe for read-ahead, just give-up */
4068 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4069 finish_wait(&conf->wait_for_overlap, &w);
4075 md_wakeup_thread(mddev->thread);
4077 spin_lock_irq(&conf->device_lock);
4078 remaining = raid5_dec_bi_phys_segments(bi);
4079 spin_unlock_irq(&conf->device_lock);
4080 if (remaining == 0) {
4083 md_write_end(mddev);
4091 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
4093 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
4095 /* reshaping is quite different to recovery/resync so it is
4096 * handled quite separately ... here.
4098 * On each call to sync_request, we gather one chunk worth of
4099 * destination stripes and flag them as expanding.
4100 * Then we find all the source stripes and request reads.
4101 * As the reads complete, handle_stripe will copy the data
4102 * into the destination stripe and release that stripe.
4104 raid5_conf_t *conf = mddev->private;
4105 struct stripe_head *sh;
4106 sector_t first_sector, last_sector;
4107 int raid_disks = conf->previous_raid_disks;
4108 int data_disks = raid_disks - conf->max_degraded;
4109 int new_data_disks = conf->raid_disks - conf->max_degraded;
4112 sector_t writepos, readpos, safepos;
4113 sector_t stripe_addr;
4114 int reshape_sectors;
4115 struct list_head stripes;
4117 if (sector_nr == 0) {
4118 /* If restarting in the middle, skip the initial sectors */
4119 if (mddev->delta_disks < 0 &&
4120 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4121 sector_nr = raid5_size(mddev, 0, 0)
4122 - conf->reshape_progress;
4123 } else if (mddev->delta_disks >= 0 &&
4124 conf->reshape_progress > 0)
4125 sector_nr = conf->reshape_progress;
4126 sector_div(sector_nr, new_data_disks);
4128 mddev->curr_resync_completed = sector_nr;
4129 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4135 /* We need to process a full chunk at a time.
4136 * If old and new chunk sizes differ, we need to process the
4139 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4140 reshape_sectors = mddev->new_chunk_sectors;
4142 reshape_sectors = mddev->chunk_sectors;
4144 /* we update the metadata when there is more than 3Meg
4145 * in the block range (that is rather arbitrary, should
4146 * probably be time based) or when the data about to be
4147 * copied would over-write the source of the data at
4148 * the front of the range.
4149 * i.e. one new_stripe along from reshape_progress new_maps
4150 * to after where reshape_safe old_maps to
4152 writepos = conf->reshape_progress;
4153 sector_div(writepos, new_data_disks);
4154 readpos = conf->reshape_progress;
4155 sector_div(readpos, data_disks);
4156 safepos = conf->reshape_safe;
4157 sector_div(safepos, data_disks);
4158 if (mddev->delta_disks < 0) {
4159 writepos -= min_t(sector_t, reshape_sectors, writepos);
4160 readpos += reshape_sectors;
4161 safepos += reshape_sectors;
4163 writepos += reshape_sectors;
4164 readpos -= min_t(sector_t, reshape_sectors, readpos);
4165 safepos -= min_t(sector_t, reshape_sectors, safepos);
4168 /* 'writepos' is the most advanced device address we might write.
4169 * 'readpos' is the least advanced device address we might read.
4170 * 'safepos' is the least address recorded in the metadata as having
4172 * If 'readpos' is behind 'writepos', then there is no way that we can
4173 * ensure safety in the face of a crash - that must be done by userspace
4174 * making a backup of the data. So in that case there is no particular
4175 * rush to update metadata.
4176 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4177 * update the metadata to advance 'safepos' to match 'readpos' so that
4178 * we can be safe in the event of a crash.
4179 * So we insist on updating metadata if safepos is behind writepos and
4180 * readpos is beyond writepos.
4181 * In any case, update the metadata every 10 seconds.
4182 * Maybe that number should be configurable, but I'm not sure it is
4183 * worth it.... maybe it could be a multiple of safemode_delay???
4185 if ((mddev->delta_disks < 0
4186 ? (safepos > writepos && readpos < writepos)
4187 : (safepos < writepos && readpos > writepos)) ||
4188 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4189 /* Cannot proceed until we've updated the superblock... */
4190 wait_event(conf->wait_for_overlap,
4191 atomic_read(&conf->reshape_stripes)==0);
4192 mddev->reshape_position = conf->reshape_progress;
4193 mddev->curr_resync_completed = sector_nr;
4194 conf->reshape_checkpoint = jiffies;
4195 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4196 md_wakeup_thread(mddev->thread);
4197 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4198 kthread_should_stop());
4199 spin_lock_irq(&conf->device_lock);
4200 conf->reshape_safe = mddev->reshape_position;
4201 spin_unlock_irq(&conf->device_lock);
4202 wake_up(&conf->wait_for_overlap);
4203 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4206 if (mddev->delta_disks < 0) {
4207 BUG_ON(conf->reshape_progress == 0);
4208 stripe_addr = writepos;
4209 BUG_ON((mddev->dev_sectors &
4210 ~((sector_t)reshape_sectors - 1))
4211 - reshape_sectors - stripe_addr
4214 BUG_ON(writepos != sector_nr + reshape_sectors);
4215 stripe_addr = sector_nr;
4217 INIT_LIST_HEAD(&stripes);
4218 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4220 int skipped_disk = 0;
4221 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4222 set_bit(STRIPE_EXPANDING, &sh->state);
4223 atomic_inc(&conf->reshape_stripes);
4224 /* If any of this stripe is beyond the end of the old
4225 * array, then we need to zero those blocks
4227 for (j=sh->disks; j--;) {
4229 if (j == sh->pd_idx)
4231 if (conf->level == 6 &&
4234 s = compute_blocknr(sh, j, 0);
4235 if (s < raid5_size(mddev, 0, 0)) {
4239 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4240 set_bit(R5_Expanded, &sh->dev[j].flags);
4241 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4243 if (!skipped_disk) {
4244 set_bit(STRIPE_EXPAND_READY, &sh->state);
4245 set_bit(STRIPE_HANDLE, &sh->state);
4247 list_add(&sh->lru, &stripes);
4249 spin_lock_irq(&conf->device_lock);
4250 if (mddev->delta_disks < 0)
4251 conf->reshape_progress -= reshape_sectors * new_data_disks;
4253 conf->reshape_progress += reshape_sectors * new_data_disks;
4254 spin_unlock_irq(&conf->device_lock);
4255 /* Ok, those stripe are ready. We can start scheduling
4256 * reads on the source stripes.
4257 * The source stripes are determined by mapping the first and last
4258 * block on the destination stripes.
4261 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4264 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4265 * new_data_disks - 1),
4267 if (last_sector >= mddev->dev_sectors)
4268 last_sector = mddev->dev_sectors - 1;
4269 while (first_sector <= last_sector) {
4270 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4271 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4272 set_bit(STRIPE_HANDLE, &sh->state);
4274 first_sector += STRIPE_SECTORS;
4276 /* Now that the sources are clearly marked, we can release
4277 * the destination stripes
4279 while (!list_empty(&stripes)) {
4280 sh = list_entry(stripes.next, struct stripe_head, lru);
4281 list_del_init(&sh->lru);
4284 /* If this takes us to the resync_max point where we have to pause,
4285 * then we need to write out the superblock.
4287 sector_nr += reshape_sectors;
4288 if ((sector_nr - mddev->curr_resync_completed) * 2
4289 >= mddev->resync_max - mddev->curr_resync_completed) {
4290 /* Cannot proceed until we've updated the superblock... */
4291 wait_event(conf->wait_for_overlap,
4292 atomic_read(&conf->reshape_stripes) == 0);
4293 mddev->reshape_position = conf->reshape_progress;
4294 mddev->curr_resync_completed = sector_nr;
4295 conf->reshape_checkpoint = jiffies;
4296 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4297 md_wakeup_thread(mddev->thread);
4298 wait_event(mddev->sb_wait,
4299 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4300 || kthread_should_stop());
4301 spin_lock_irq(&conf->device_lock);
4302 conf->reshape_safe = mddev->reshape_position;
4303 spin_unlock_irq(&conf->device_lock);
4304 wake_up(&conf->wait_for_overlap);
4305 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4307 return reshape_sectors;
4310 /* FIXME go_faster isn't used */
4311 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4313 raid5_conf_t *conf = mddev->private;
4314 struct stripe_head *sh;
4315 sector_t max_sector = mddev->dev_sectors;
4316 sector_t sync_blocks;
4317 int still_degraded = 0;
4320 if (sector_nr >= max_sector) {
4321 /* just being told to finish up .. nothing much to do */
4323 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4328 if (mddev->curr_resync < max_sector) /* aborted */
4329 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4331 else /* completed sync */
4333 bitmap_close_sync(mddev->bitmap);
4338 /* Allow raid5_quiesce to complete */
4339 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4341 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4342 return reshape_request(mddev, sector_nr, skipped);
4344 /* No need to check resync_max as we never do more than one
4345 * stripe, and as resync_max will always be on a chunk boundary,
4346 * if the check in md_do_sync didn't fire, there is no chance
4347 * of overstepping resync_max here
4350 /* if there is too many failed drives and we are trying
4351 * to resync, then assert that we are finished, because there is
4352 * nothing we can do.
4354 if (mddev->degraded >= conf->max_degraded &&
4355 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4356 sector_t rv = mddev->dev_sectors - sector_nr;
4360 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4361 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4362 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4363 /* we can skip this block, and probably more */
4364 sync_blocks /= STRIPE_SECTORS;
4366 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4370 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4372 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4374 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4375 /* make sure we don't swamp the stripe cache if someone else
4376 * is trying to get access
4378 schedule_timeout_uninterruptible(1);
4380 /* Need to check if array will still be degraded after recovery/resync
4381 * We don't need to check the 'failed' flag as when that gets set,
4384 for (i = 0; i < conf->raid_disks; i++)
4385 if (conf->disks[i].rdev == NULL)
4388 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4390 spin_lock(&sh->lock);
4391 set_bit(STRIPE_SYNCING, &sh->state);
4392 clear_bit(STRIPE_INSYNC, &sh->state);
4393 spin_unlock(&sh->lock);
4398 return STRIPE_SECTORS;
4401 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4403 /* We may not be able to submit a whole bio at once as there
4404 * may not be enough stripe_heads available.
4405 * We cannot pre-allocate enough stripe_heads as we may need
4406 * more than exist in the cache (if we allow ever large chunks).
4407 * So we do one stripe head at a time and record in
4408 * ->bi_hw_segments how many have been done.
4410 * We *know* that this entire raid_bio is in one chunk, so
4411 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4413 struct stripe_head *sh;
4415 sector_t sector, logical_sector, last_sector;
4420 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4421 sector = raid5_compute_sector(conf, logical_sector,
4423 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4425 for (; logical_sector < last_sector;
4426 logical_sector += STRIPE_SECTORS,
4427 sector += STRIPE_SECTORS,
4430 if (scnt < raid5_bi_hw_segments(raid_bio))
4431 /* already done this stripe */
4434 sh = get_active_stripe(conf, sector, 0, 1, 0);
4437 /* failed to get a stripe - must wait */
4438 raid5_set_bi_hw_segments(raid_bio, scnt);
4439 conf->retry_read_aligned = raid_bio;
4443 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4444 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4446 raid5_set_bi_hw_segments(raid_bio, scnt);
4447 conf->retry_read_aligned = raid_bio;
4455 spin_lock_irq(&conf->device_lock);
4456 remaining = raid5_dec_bi_phys_segments(raid_bio);
4457 spin_unlock_irq(&conf->device_lock);
4459 bio_endio(raid_bio, 0);
4460 if (atomic_dec_and_test(&conf->active_aligned_reads))
4461 wake_up(&conf->wait_for_stripe);
4467 * This is our raid5 kernel thread.
4469 * We scan the hash table for stripes which can be handled now.
4470 * During the scan, completed stripes are saved for us by the interrupt
4471 * handler, so that they will not have to wait for our next wakeup.
4473 static void raid5d(mddev_t *mddev)
4475 struct stripe_head *sh;
4476 raid5_conf_t *conf = mddev->private;
4478 struct blk_plug plug;
4480 pr_debug("+++ raid5d active\n");
4482 md_check_recovery(mddev);
4484 blk_start_plug(&plug);
4486 spin_lock_irq(&conf->device_lock);
4490 if (atomic_read(&mddev->plug_cnt) == 0 &&
4491 !list_empty(&conf->bitmap_list)) {
4492 /* Now is a good time to flush some bitmap updates */
4494 spin_unlock_irq(&conf->device_lock);
4495 bitmap_unplug(mddev->bitmap);
4496 spin_lock_irq(&conf->device_lock);
4497 conf->seq_write = conf->seq_flush;
4498 activate_bit_delay(conf);
4500 if (atomic_read(&mddev->plug_cnt) == 0)
4501 raid5_activate_delayed(conf);
4503 while ((bio = remove_bio_from_retry(conf))) {
4505 spin_unlock_irq(&conf->device_lock);
4506 ok = retry_aligned_read(conf, bio);
4507 spin_lock_irq(&conf->device_lock);
4513 sh = __get_priority_stripe(conf);
4517 spin_unlock_irq(&conf->device_lock);
4524 spin_lock_irq(&conf->device_lock);
4526 pr_debug("%d stripes handled\n", handled);
4528 spin_unlock_irq(&conf->device_lock);
4530 async_tx_issue_pending_all();
4531 blk_finish_plug(&plug);
4533 pr_debug("--- raid5d inactive\n");
4537 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4539 raid5_conf_t *conf = mddev->private;
4541 return sprintf(page, "%d\n", conf->max_nr_stripes);
4547 raid5_set_cache_size(mddev_t *mddev, int size)
4549 raid5_conf_t *conf = mddev->private;
4552 if (size <= 16 || size > 32768)
4554 while (size < conf->max_nr_stripes) {
4555 if (drop_one_stripe(conf))
4556 conf->max_nr_stripes--;
4560 err = md_allow_write(mddev);
4563 while (size > conf->max_nr_stripes) {
4564 if (grow_one_stripe(conf))
4565 conf->max_nr_stripes++;
4570 EXPORT_SYMBOL(raid5_set_cache_size);
4573 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4575 raid5_conf_t *conf = mddev->private;
4579 if (len >= PAGE_SIZE)
4584 if (strict_strtoul(page, 10, &new))
4586 err = raid5_set_cache_size(mddev, new);
4592 static struct md_sysfs_entry
4593 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4594 raid5_show_stripe_cache_size,
4595 raid5_store_stripe_cache_size);
4598 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4600 raid5_conf_t *conf = mddev->private;
4602 return sprintf(page, "%d\n", conf->bypass_threshold);
4608 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4610 raid5_conf_t *conf = mddev->private;
4612 if (len >= PAGE_SIZE)
4617 if (strict_strtoul(page, 10, &new))
4619 if (new > conf->max_nr_stripes)
4621 conf->bypass_threshold = new;
4625 static struct md_sysfs_entry
4626 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4628 raid5_show_preread_threshold,
4629 raid5_store_preread_threshold);
4632 stripe_cache_active_show(mddev_t *mddev, char *page)
4634 raid5_conf_t *conf = mddev->private;
4636 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4641 static struct md_sysfs_entry
4642 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4644 static struct attribute *raid5_attrs[] = {
4645 &raid5_stripecache_size.attr,
4646 &raid5_stripecache_active.attr,
4647 &raid5_preread_bypass_threshold.attr,
4650 static struct attribute_group raid5_attrs_group = {
4652 .attrs = raid5_attrs,
4656 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4658 raid5_conf_t *conf = mddev->private;
4661 sectors = mddev->dev_sectors;
4663 /* size is defined by the smallest of previous and new size */
4664 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4666 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4667 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4668 return sectors * (raid_disks - conf->max_degraded);
4671 static void raid5_free_percpu(raid5_conf_t *conf)
4673 struct raid5_percpu *percpu;
4680 for_each_possible_cpu(cpu) {
4681 percpu = per_cpu_ptr(conf->percpu, cpu);
4682 safe_put_page(percpu->spare_page);
4683 kfree(percpu->scribble);
4685 #ifdef CONFIG_HOTPLUG_CPU
4686 unregister_cpu_notifier(&conf->cpu_notify);
4690 free_percpu(conf->percpu);
4693 static void free_conf(raid5_conf_t *conf)
4695 shrink_stripes(conf);
4696 raid5_free_percpu(conf);
4698 kfree(conf->stripe_hashtbl);
4702 #ifdef CONFIG_HOTPLUG_CPU
4703 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4706 raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4707 long cpu = (long)hcpu;
4708 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4711 case CPU_UP_PREPARE:
4712 case CPU_UP_PREPARE_FROZEN:
4713 if (conf->level == 6 && !percpu->spare_page)
4714 percpu->spare_page = alloc_page(GFP_KERNEL);
4715 if (!percpu->scribble)
4716 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4718 if (!percpu->scribble ||
4719 (conf->level == 6 && !percpu->spare_page)) {
4720 safe_put_page(percpu->spare_page);
4721 kfree(percpu->scribble);
4722 pr_err("%s: failed memory allocation for cpu%ld\n",
4724 return notifier_from_errno(-ENOMEM);
4728 case CPU_DEAD_FROZEN:
4729 safe_put_page(percpu->spare_page);
4730 kfree(percpu->scribble);
4731 percpu->spare_page = NULL;
4732 percpu->scribble = NULL;
4741 static int raid5_alloc_percpu(raid5_conf_t *conf)
4744 struct page *spare_page;
4745 struct raid5_percpu __percpu *allcpus;
4749 allcpus = alloc_percpu(struct raid5_percpu);
4752 conf->percpu = allcpus;
4756 for_each_present_cpu(cpu) {
4757 if (conf->level == 6) {
4758 spare_page = alloc_page(GFP_KERNEL);
4763 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4765 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4770 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4772 #ifdef CONFIG_HOTPLUG_CPU
4773 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4774 conf->cpu_notify.priority = 0;
4776 err = register_cpu_notifier(&conf->cpu_notify);
4783 static raid5_conf_t *setup_conf(mddev_t *mddev)
4786 int raid_disk, memory, max_disks;
4788 struct disk_info *disk;
4790 if (mddev->new_level != 5
4791 && mddev->new_level != 4
4792 && mddev->new_level != 6) {
4793 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4794 mdname(mddev), mddev->new_level);
4795 return ERR_PTR(-EIO);
4797 if ((mddev->new_level == 5
4798 && !algorithm_valid_raid5(mddev->new_layout)) ||
4799 (mddev->new_level == 6
4800 && !algorithm_valid_raid6(mddev->new_layout))) {
4801 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4802 mdname(mddev), mddev->new_layout);
4803 return ERR_PTR(-EIO);
4805 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4806 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4807 mdname(mddev), mddev->raid_disks);
4808 return ERR_PTR(-EINVAL);
4811 if (!mddev->new_chunk_sectors ||
4812 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4813 !is_power_of_2(mddev->new_chunk_sectors)) {
4814 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4815 mdname(mddev), mddev->new_chunk_sectors << 9);
4816 return ERR_PTR(-EINVAL);
4819 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4822 spin_lock_init(&conf->device_lock);
4823 init_waitqueue_head(&conf->wait_for_stripe);
4824 init_waitqueue_head(&conf->wait_for_overlap);
4825 INIT_LIST_HEAD(&conf->handle_list);
4826 INIT_LIST_HEAD(&conf->hold_list);
4827 INIT_LIST_HEAD(&conf->delayed_list);
4828 INIT_LIST_HEAD(&conf->bitmap_list);
4829 INIT_LIST_HEAD(&conf->inactive_list);
4830 atomic_set(&conf->active_stripes, 0);
4831 atomic_set(&conf->preread_active_stripes, 0);
4832 atomic_set(&conf->active_aligned_reads, 0);
4833 conf->bypass_threshold = BYPASS_THRESHOLD;
4835 conf->raid_disks = mddev->raid_disks;
4836 if (mddev->reshape_position == MaxSector)
4837 conf->previous_raid_disks = mddev->raid_disks;
4839 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4840 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4841 conf->scribble_len = scribble_len(max_disks);
4843 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4848 conf->mddev = mddev;
4850 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4853 conf->level = mddev->new_level;
4854 if (raid5_alloc_percpu(conf) != 0)
4857 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4859 list_for_each_entry(rdev, &mddev->disks, same_set) {
4860 raid_disk = rdev->raid_disk;
4861 if (raid_disk >= max_disks
4864 disk = conf->disks + raid_disk;
4868 if (test_bit(In_sync, &rdev->flags)) {
4869 char b[BDEVNAME_SIZE];
4870 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4872 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4873 } else if (rdev->saved_raid_disk != raid_disk)
4874 /* Cannot rely on bitmap to complete recovery */
4878 conf->chunk_sectors = mddev->new_chunk_sectors;
4879 conf->level = mddev->new_level;
4880 if (conf->level == 6)
4881 conf->max_degraded = 2;
4883 conf->max_degraded = 1;
4884 conf->algorithm = mddev->new_layout;
4885 conf->max_nr_stripes = NR_STRIPES;
4886 conf->reshape_progress = mddev->reshape_position;
4887 if (conf->reshape_progress != MaxSector) {
4888 conf->prev_chunk_sectors = mddev->chunk_sectors;
4889 conf->prev_algo = mddev->layout;
4892 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4893 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4894 if (grow_stripes(conf, conf->max_nr_stripes)) {
4896 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4897 mdname(mddev), memory);
4900 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4901 mdname(mddev), memory);
4903 conf->thread = md_register_thread(raid5d, mddev, NULL);
4904 if (!conf->thread) {
4906 "md/raid:%s: couldn't allocate thread.\n",
4916 return ERR_PTR(-EIO);
4918 return ERR_PTR(-ENOMEM);
4922 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4925 case ALGORITHM_PARITY_0:
4926 if (raid_disk < max_degraded)
4929 case ALGORITHM_PARITY_N:
4930 if (raid_disk >= raid_disks - max_degraded)
4933 case ALGORITHM_PARITY_0_6:
4934 if (raid_disk == 0 ||
4935 raid_disk == raid_disks - 1)
4938 case ALGORITHM_LEFT_ASYMMETRIC_6:
4939 case ALGORITHM_RIGHT_ASYMMETRIC_6:
4940 case ALGORITHM_LEFT_SYMMETRIC_6:
4941 case ALGORITHM_RIGHT_SYMMETRIC_6:
4942 if (raid_disk == raid_disks - 1)
4948 static int run(mddev_t *mddev)
4951 int working_disks = 0;
4952 int dirty_parity_disks = 0;
4954 sector_t reshape_offset = 0;
4956 if (mddev->recovery_cp != MaxSector)
4957 printk(KERN_NOTICE "md/raid:%s: not clean"
4958 " -- starting background reconstruction\n",
4960 if (mddev->reshape_position != MaxSector) {
4961 /* Check that we can continue the reshape.
4962 * Currently only disks can change, it must
4963 * increase, and we must be past the point where
4964 * a stripe over-writes itself
4966 sector_t here_new, here_old;
4968 int max_degraded = (mddev->level == 6 ? 2 : 1);
4970 if (mddev->new_level != mddev->level) {
4971 printk(KERN_ERR "md/raid:%s: unsupported reshape "
4972 "required - aborting.\n",
4976 old_disks = mddev->raid_disks - mddev->delta_disks;
4977 /* reshape_position must be on a new-stripe boundary, and one
4978 * further up in new geometry must map after here in old
4981 here_new = mddev->reshape_position;
4982 if (sector_div(here_new, mddev->new_chunk_sectors *
4983 (mddev->raid_disks - max_degraded))) {
4984 printk(KERN_ERR "md/raid:%s: reshape_position not "
4985 "on a stripe boundary\n", mdname(mddev));
4988 reshape_offset = here_new * mddev->new_chunk_sectors;
4989 /* here_new is the stripe we will write to */
4990 here_old = mddev->reshape_position;
4991 sector_div(here_old, mddev->chunk_sectors *
4992 (old_disks-max_degraded));
4993 /* here_old is the first stripe that we might need to read
4995 if (mddev->delta_disks == 0) {
4996 /* We cannot be sure it is safe to start an in-place
4997 * reshape. It is only safe if user-space if monitoring
4998 * and taking constant backups.
4999 * mdadm always starts a situation like this in
5000 * readonly mode so it can take control before
5001 * allowing any writes. So just check for that.
5003 if ((here_new * mddev->new_chunk_sectors !=
5004 here_old * mddev->chunk_sectors) ||
5006 printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
5007 " in read-only mode - aborting\n",
5011 } else if (mddev->delta_disks < 0
5012 ? (here_new * mddev->new_chunk_sectors <=
5013 here_old * mddev->chunk_sectors)
5014 : (here_new * mddev->new_chunk_sectors >=
5015 here_old * mddev->chunk_sectors)) {
5016 /* Reading from the same stripe as writing to - bad */
5017 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5018 "auto-recovery - aborting.\n",
5022 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5024 /* OK, we should be able to continue; */
5026 BUG_ON(mddev->level != mddev->new_level);
5027 BUG_ON(mddev->layout != mddev->new_layout);
5028 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5029 BUG_ON(mddev->delta_disks != 0);
5032 if (mddev->private == NULL)
5033 conf = setup_conf(mddev);
5035 conf = mddev->private;
5038 return PTR_ERR(conf);
5040 mddev->thread = conf->thread;
5041 conf->thread = NULL;
5042 mddev->private = conf;
5045 * 0 for a fully functional array, 1 or 2 for a degraded array.
5047 list_for_each_entry(rdev, &mddev->disks, same_set) {
5048 if (rdev->raid_disk < 0)
5050 if (test_bit(In_sync, &rdev->flags)) {
5054 /* This disc is not fully in-sync. However if it
5055 * just stored parity (beyond the recovery_offset),
5056 * when we don't need to be concerned about the
5057 * array being dirty.
5058 * When reshape goes 'backwards', we never have
5059 * partially completed devices, so we only need
5060 * to worry about reshape going forwards.
5062 /* Hack because v0.91 doesn't store recovery_offset properly. */
5063 if (mddev->major_version == 0 &&
5064 mddev->minor_version > 90)
5065 rdev->recovery_offset = reshape_offset;
5067 if (rdev->recovery_offset < reshape_offset) {
5068 /* We need to check old and new layout */
5069 if (!only_parity(rdev->raid_disk,
5072 conf->max_degraded))
5075 if (!only_parity(rdev->raid_disk,
5077 conf->previous_raid_disks,
5078 conf->max_degraded))
5080 dirty_parity_disks++;
5083 mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
5086 if (has_failed(conf)) {
5087 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5088 " (%d/%d failed)\n",
5089 mdname(mddev), mddev->degraded, conf->raid_disks);
5093 /* device size must be a multiple of chunk size */
5094 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5095 mddev->resync_max_sectors = mddev->dev_sectors;
5097 if (mddev->degraded > dirty_parity_disks &&
5098 mddev->recovery_cp != MaxSector) {
5099 if (mddev->ok_start_degraded)
5101 "md/raid:%s: starting dirty degraded array"
5102 " - data corruption possible.\n",
5106 "md/raid:%s: cannot start dirty degraded array.\n",
5112 if (mddev->degraded == 0)
5113 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5114 " devices, algorithm %d\n", mdname(mddev), conf->level,
5115 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5118 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5119 " out of %d devices, algorithm %d\n",
5120 mdname(mddev), conf->level,
5121 mddev->raid_disks - mddev->degraded,
5122 mddev->raid_disks, mddev->new_layout);
5124 print_raid5_conf(conf);
5126 if (conf->reshape_progress != MaxSector) {
5127 conf->reshape_safe = conf->reshape_progress;
5128 atomic_set(&conf->reshape_stripes, 0);
5129 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5130 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5131 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5132 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5133 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5138 /* Ok, everything is just fine now */
5139 if (mddev->to_remove == &raid5_attrs_group)
5140 mddev->to_remove = NULL;
5141 else if (mddev->kobj.sd &&
5142 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5144 "raid5: failed to create sysfs attributes for %s\n",
5146 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5150 /* read-ahead size must cover two whole stripes, which
5151 * is 2 * (datadisks) * chunksize where 'n' is the
5152 * number of raid devices
5154 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5155 int stripe = data_disks *
5156 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5157 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5158 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5160 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5162 mddev->queue->backing_dev_info.congested_data = mddev;
5163 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5165 chunk_size = mddev->chunk_sectors << 9;
5166 blk_queue_io_min(mddev->queue, chunk_size);
5167 blk_queue_io_opt(mddev->queue, chunk_size *
5168 (conf->raid_disks - conf->max_degraded));
5170 list_for_each_entry(rdev, &mddev->disks, same_set)
5171 disk_stack_limits(mddev->gendisk, rdev->bdev,
5172 rdev->data_offset << 9);
5177 md_unregister_thread(&mddev->thread);
5179 print_raid5_conf(conf);
5182 mddev->private = NULL;
5183 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5187 static int stop(mddev_t *mddev)
5189 raid5_conf_t *conf = mddev->private;
5191 md_unregister_thread(&mddev->thread);
5193 mddev->queue->backing_dev_info.congested_fn = NULL;
5195 mddev->private = NULL;
5196 mddev->to_remove = &raid5_attrs_group;
5201 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5205 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5206 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5207 seq_printf(seq, "sh %llu, count %d.\n",
5208 (unsigned long long)sh->sector, atomic_read(&sh->count));
5209 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5210 for (i = 0; i < sh->disks; i++) {
5211 seq_printf(seq, "(cache%d: %p %ld) ",
5212 i, sh->dev[i].page, sh->dev[i].flags);
5214 seq_printf(seq, "\n");
5217 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5219 struct stripe_head *sh;
5220 struct hlist_node *hn;
5223 spin_lock_irq(&conf->device_lock);
5224 for (i = 0; i < NR_HASH; i++) {
5225 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5226 if (sh->raid_conf != conf)
5231 spin_unlock_irq(&conf->device_lock);
5235 static void status(struct seq_file *seq, mddev_t *mddev)
5237 raid5_conf_t *conf = mddev->private;
5240 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5241 mddev->chunk_sectors / 2, mddev->layout);
5242 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5243 for (i = 0; i < conf->raid_disks; i++)
5244 seq_printf (seq, "%s",
5245 conf->disks[i].rdev &&
5246 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5247 seq_printf (seq, "]");
5249 seq_printf (seq, "\n");
5250 printall(seq, conf);
5254 static void print_raid5_conf (raid5_conf_t *conf)
5257 struct disk_info *tmp;
5259 printk(KERN_DEBUG "RAID conf printout:\n");
5261 printk("(conf==NULL)\n");
5264 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5266 conf->raid_disks - conf->mddev->degraded);
5268 for (i = 0; i < conf->raid_disks; i++) {
5269 char b[BDEVNAME_SIZE];
5270 tmp = conf->disks + i;
5272 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5273 i, !test_bit(Faulty, &tmp->rdev->flags),
5274 bdevname(tmp->rdev->bdev, b));
5278 static int raid5_spare_active(mddev_t *mddev)
5281 raid5_conf_t *conf = mddev->private;
5282 struct disk_info *tmp;
5284 unsigned long flags;
5286 for (i = 0; i < conf->raid_disks; i++) {
5287 tmp = conf->disks + i;
5289 && tmp->rdev->recovery_offset == MaxSector
5290 && !test_bit(Faulty, &tmp->rdev->flags)
5291 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5293 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5296 spin_lock_irqsave(&conf->device_lock, flags);
5297 mddev->degraded -= count;
5298 spin_unlock_irqrestore(&conf->device_lock, flags);
5299 print_raid5_conf(conf);
5303 static int raid5_remove_disk(mddev_t *mddev, int number)
5305 raid5_conf_t *conf = mddev->private;
5308 struct disk_info *p = conf->disks + number;
5310 print_raid5_conf(conf);
5313 if (number >= conf->raid_disks &&
5314 conf->reshape_progress == MaxSector)
5315 clear_bit(In_sync, &rdev->flags);
5317 if (test_bit(In_sync, &rdev->flags) ||
5318 atomic_read(&rdev->nr_pending)) {
5322 /* Only remove non-faulty devices if recovery
5325 if (!test_bit(Faulty, &rdev->flags) &&
5326 !has_failed(conf) &&
5327 number < conf->raid_disks) {
5333 if (atomic_read(&rdev->nr_pending)) {
5334 /* lost the race, try later */
5341 print_raid5_conf(conf);
5345 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5347 raid5_conf_t *conf = mddev->private;
5350 struct disk_info *p;
5352 int last = conf->raid_disks - 1;
5354 if (has_failed(conf))
5355 /* no point adding a device */
5358 if (rdev->raid_disk >= 0)
5359 first = last = rdev->raid_disk;
5362 * find the disk ... but prefer rdev->saved_raid_disk
5365 if (rdev->saved_raid_disk >= 0 &&
5366 rdev->saved_raid_disk >= first &&
5367 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5368 disk = rdev->saved_raid_disk;
5371 for ( ; disk <= last ; disk++)
5372 if ((p=conf->disks + disk)->rdev == NULL) {
5373 clear_bit(In_sync, &rdev->flags);
5374 rdev->raid_disk = disk;
5376 if (rdev->saved_raid_disk != disk)
5378 rcu_assign_pointer(p->rdev, rdev);
5381 print_raid5_conf(conf);
5385 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5387 /* no resync is happening, and there is enough space
5388 * on all devices, so we can resize.
5389 * We need to make sure resync covers any new space.
5390 * If the array is shrinking we should possibly wait until
5391 * any io in the removed space completes, but it hardly seems
5394 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5395 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5396 mddev->raid_disks));
5397 if (mddev->array_sectors >
5398 raid5_size(mddev, sectors, mddev->raid_disks))
5400 set_capacity(mddev->gendisk, mddev->array_sectors);
5401 revalidate_disk(mddev->gendisk);
5402 if (sectors > mddev->dev_sectors &&
5403 mddev->recovery_cp > mddev->dev_sectors) {
5404 mddev->recovery_cp = mddev->dev_sectors;
5405 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5407 mddev->dev_sectors = sectors;
5408 mddev->resync_max_sectors = sectors;
5412 static int check_stripe_cache(mddev_t *mddev)
5414 /* Can only proceed if there are plenty of stripe_heads.
5415 * We need a minimum of one full stripe,, and for sensible progress
5416 * it is best to have about 4 times that.
5417 * If we require 4 times, then the default 256 4K stripe_heads will
5418 * allow for chunk sizes up to 256K, which is probably OK.
5419 * If the chunk size is greater, user-space should request more
5420 * stripe_heads first.
5422 raid5_conf_t *conf = mddev->private;
5423 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5424 > conf->max_nr_stripes ||
5425 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5426 > conf->max_nr_stripes) {
5427 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5429 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5436 static int check_reshape(mddev_t *mddev)
5438 raid5_conf_t *conf = mddev->private;
5440 if (mddev->delta_disks == 0 &&
5441 mddev->new_layout == mddev->layout &&
5442 mddev->new_chunk_sectors == mddev->chunk_sectors)
5443 return 0; /* nothing to do */
5445 /* Cannot grow a bitmap yet */
5447 if (has_failed(conf))
5449 if (mddev->delta_disks < 0) {
5450 /* We might be able to shrink, but the devices must
5451 * be made bigger first.
5452 * For raid6, 4 is the minimum size.
5453 * Otherwise 2 is the minimum
5456 if (mddev->level == 6)
5458 if (mddev->raid_disks + mddev->delta_disks < min)
5462 if (!check_stripe_cache(mddev))
5465 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5468 static int raid5_start_reshape(mddev_t *mddev)
5470 raid5_conf_t *conf = mddev->private;
5473 unsigned long flags;
5475 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5478 if (!check_stripe_cache(mddev))
5481 list_for_each_entry(rdev, &mddev->disks, same_set)
5482 if (!test_bit(In_sync, &rdev->flags)
5483 && !test_bit(Faulty, &rdev->flags))
5486 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5487 /* Not enough devices even to make a degraded array
5492 /* Refuse to reduce size of the array. Any reductions in
5493 * array size must be through explicit setting of array_size
5496 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5497 < mddev->array_sectors) {
5498 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5499 "before number of disks\n", mdname(mddev));
5503 atomic_set(&conf->reshape_stripes, 0);
5504 spin_lock_irq(&conf->device_lock);
5505 conf->previous_raid_disks = conf->raid_disks;
5506 conf->raid_disks += mddev->delta_disks;
5507 conf->prev_chunk_sectors = conf->chunk_sectors;
5508 conf->chunk_sectors = mddev->new_chunk_sectors;
5509 conf->prev_algo = conf->algorithm;
5510 conf->algorithm = mddev->new_layout;
5511 if (mddev->delta_disks < 0)
5512 conf->reshape_progress = raid5_size(mddev, 0, 0);
5514 conf->reshape_progress = 0;
5515 conf->reshape_safe = conf->reshape_progress;
5517 spin_unlock_irq(&conf->device_lock);
5519 /* Add some new drives, as many as will fit.
5520 * We know there are enough to make the newly sized array work.
5521 * Don't add devices if we are reducing the number of
5522 * devices in the array. This is because it is not possible
5523 * to correctly record the "partially reconstructed" state of
5524 * such devices during the reshape and confusion could result.
5526 if (mddev->delta_disks >= 0) {
5527 int added_devices = 0;
5528 list_for_each_entry(rdev, &mddev->disks, same_set)
5529 if (rdev->raid_disk < 0 &&
5530 !test_bit(Faulty, &rdev->flags)) {
5531 if (raid5_add_disk(mddev, rdev) == 0) {
5534 >= conf->previous_raid_disks) {
5535 set_bit(In_sync, &rdev->flags);
5538 rdev->recovery_offset = 0;
5539 sprintf(nm, "rd%d", rdev->raid_disk);
5540 if (sysfs_create_link(&mddev->kobj,
5542 /* Failure here is OK */;
5544 } else if (rdev->raid_disk >= conf->previous_raid_disks
5545 && !test_bit(Faulty, &rdev->flags)) {
5546 /* This is a spare that was manually added */
5547 set_bit(In_sync, &rdev->flags);
5551 /* When a reshape changes the number of devices,
5552 * ->degraded is measured against the larger of the
5553 * pre and post number of devices.
5555 spin_lock_irqsave(&conf->device_lock, flags);
5556 mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5558 spin_unlock_irqrestore(&conf->device_lock, flags);
5560 mddev->raid_disks = conf->raid_disks;
5561 mddev->reshape_position = conf->reshape_progress;
5562 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5564 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5565 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5566 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5567 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5568 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5570 if (!mddev->sync_thread) {
5571 mddev->recovery = 0;
5572 spin_lock_irq(&conf->device_lock);
5573 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5574 conf->reshape_progress = MaxSector;
5575 spin_unlock_irq(&conf->device_lock);
5578 conf->reshape_checkpoint = jiffies;
5579 md_wakeup_thread(mddev->sync_thread);
5580 md_new_event(mddev);
5584 /* This is called from the reshape thread and should make any
5585 * changes needed in 'conf'
5587 static void end_reshape(raid5_conf_t *conf)
5590 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5592 spin_lock_irq(&conf->device_lock);
5593 conf->previous_raid_disks = conf->raid_disks;
5594 conf->reshape_progress = MaxSector;
5595 spin_unlock_irq(&conf->device_lock);
5596 wake_up(&conf->wait_for_overlap);
5598 /* read-ahead size must cover two whole stripes, which is
5599 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5601 if (conf->mddev->queue) {
5602 int data_disks = conf->raid_disks - conf->max_degraded;
5603 int stripe = data_disks * ((conf->chunk_sectors << 9)
5605 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5606 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5611 /* This is called from the raid5d thread with mddev_lock held.
5612 * It makes config changes to the device.
5614 static void raid5_finish_reshape(mddev_t *mddev)
5616 raid5_conf_t *conf = mddev->private;
5618 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5620 if (mddev->delta_disks > 0) {
5621 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5622 set_capacity(mddev->gendisk, mddev->array_sectors);
5623 revalidate_disk(mddev->gendisk);
5626 mddev->degraded = conf->raid_disks;
5627 for (d = 0; d < conf->raid_disks ; d++)
5628 if (conf->disks[d].rdev &&
5630 &conf->disks[d].rdev->flags))
5632 for (d = conf->raid_disks ;
5633 d < conf->raid_disks - mddev->delta_disks;
5635 mdk_rdev_t *rdev = conf->disks[d].rdev;
5636 if (rdev && raid5_remove_disk(mddev, d) == 0) {
5638 sprintf(nm, "rd%d", rdev->raid_disk);
5639 sysfs_remove_link(&mddev->kobj, nm);
5640 rdev->raid_disk = -1;
5644 mddev->layout = conf->algorithm;
5645 mddev->chunk_sectors = conf->chunk_sectors;
5646 mddev->reshape_position = MaxSector;
5647 mddev->delta_disks = 0;
5651 static void raid5_quiesce(mddev_t *mddev, int state)
5653 raid5_conf_t *conf = mddev->private;
5656 case 2: /* resume for a suspend */
5657 wake_up(&conf->wait_for_overlap);
5660 case 1: /* stop all writes */
5661 spin_lock_irq(&conf->device_lock);
5662 /* '2' tells resync/reshape to pause so that all
5663 * active stripes can drain
5666 wait_event_lock_irq(conf->wait_for_stripe,
5667 atomic_read(&conf->active_stripes) == 0 &&
5668 atomic_read(&conf->active_aligned_reads) == 0,
5669 conf->device_lock, /* nothing */);
5671 spin_unlock_irq(&conf->device_lock);
5672 /* allow reshape to continue */
5673 wake_up(&conf->wait_for_overlap);
5676 case 0: /* re-enable writes */
5677 spin_lock_irq(&conf->device_lock);
5679 wake_up(&conf->wait_for_stripe);
5680 wake_up(&conf->wait_for_overlap);
5681 spin_unlock_irq(&conf->device_lock);
5687 static void *raid45_takeover_raid0(mddev_t *mddev, int level)
5689 struct raid0_private_data *raid0_priv = mddev->private;
5692 /* for raid0 takeover only one zone is supported */
5693 if (raid0_priv->nr_strip_zones > 1) {
5694 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5696 return ERR_PTR(-EINVAL);
5699 sectors = raid0_priv->strip_zone[0].zone_end;
5700 sector_div(sectors, raid0_priv->strip_zone[0].nb_dev);
5701 mddev->dev_sectors = sectors;
5702 mddev->new_level = level;
5703 mddev->new_layout = ALGORITHM_PARITY_N;
5704 mddev->new_chunk_sectors = mddev->chunk_sectors;
5705 mddev->raid_disks += 1;
5706 mddev->delta_disks = 1;
5707 /* make sure it will be not marked as dirty */
5708 mddev->recovery_cp = MaxSector;
5710 return setup_conf(mddev);
5714 static void *raid5_takeover_raid1(mddev_t *mddev)
5718 if (mddev->raid_disks != 2 ||
5719 mddev->degraded > 1)
5720 return ERR_PTR(-EINVAL);
5722 /* Should check if there are write-behind devices? */
5724 chunksect = 64*2; /* 64K by default */
5726 /* The array must be an exact multiple of chunksize */
5727 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5730 if ((chunksect<<9) < STRIPE_SIZE)
5731 /* array size does not allow a suitable chunk size */
5732 return ERR_PTR(-EINVAL);
5734 mddev->new_level = 5;
5735 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5736 mddev->new_chunk_sectors = chunksect;
5738 return setup_conf(mddev);
5741 static void *raid5_takeover_raid6(mddev_t *mddev)
5745 switch (mddev->layout) {
5746 case ALGORITHM_LEFT_ASYMMETRIC_6:
5747 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5749 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5750 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5752 case ALGORITHM_LEFT_SYMMETRIC_6:
5753 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5755 case ALGORITHM_RIGHT_SYMMETRIC_6:
5756 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5758 case ALGORITHM_PARITY_0_6:
5759 new_layout = ALGORITHM_PARITY_0;
5761 case ALGORITHM_PARITY_N:
5762 new_layout = ALGORITHM_PARITY_N;
5765 return ERR_PTR(-EINVAL);
5767 mddev->new_level = 5;
5768 mddev->new_layout = new_layout;
5769 mddev->delta_disks = -1;
5770 mddev->raid_disks -= 1;
5771 return setup_conf(mddev);
5775 static int raid5_check_reshape(mddev_t *mddev)
5777 /* For a 2-drive array, the layout and chunk size can be changed
5778 * immediately as not restriping is needed.
5779 * For larger arrays we record the new value - after validation
5780 * to be used by a reshape pass.
5782 raid5_conf_t *conf = mddev->private;
5783 int new_chunk = mddev->new_chunk_sectors;
5785 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5787 if (new_chunk > 0) {
5788 if (!is_power_of_2(new_chunk))
5790 if (new_chunk < (PAGE_SIZE>>9))
5792 if (mddev->array_sectors & (new_chunk-1))
5793 /* not factor of array size */
5797 /* They look valid */
5799 if (mddev->raid_disks == 2) {
5800 /* can make the change immediately */
5801 if (mddev->new_layout >= 0) {
5802 conf->algorithm = mddev->new_layout;
5803 mddev->layout = mddev->new_layout;
5805 if (new_chunk > 0) {
5806 conf->chunk_sectors = new_chunk ;
5807 mddev->chunk_sectors = new_chunk;
5809 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5810 md_wakeup_thread(mddev->thread);
5812 return check_reshape(mddev);
5815 static int raid6_check_reshape(mddev_t *mddev)
5817 int new_chunk = mddev->new_chunk_sectors;
5819 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5821 if (new_chunk > 0) {
5822 if (!is_power_of_2(new_chunk))
5824 if (new_chunk < (PAGE_SIZE >> 9))
5826 if (mddev->array_sectors & (new_chunk-1))
5827 /* not factor of array size */
5831 /* They look valid */
5832 return check_reshape(mddev);
5835 static void *raid5_takeover(mddev_t *mddev)
5837 /* raid5 can take over:
5838 * raid0 - if there is only one strip zone - make it a raid4 layout
5839 * raid1 - if there are two drives. We need to know the chunk size
5840 * raid4 - trivial - just use a raid4 layout.
5841 * raid6 - Providing it is a *_6 layout
5843 if (mddev->level == 0)
5844 return raid45_takeover_raid0(mddev, 5);
5845 if (mddev->level == 1)
5846 return raid5_takeover_raid1(mddev);
5847 if (mddev->level == 4) {
5848 mddev->new_layout = ALGORITHM_PARITY_N;
5849 mddev->new_level = 5;
5850 return setup_conf(mddev);
5852 if (mddev->level == 6)
5853 return raid5_takeover_raid6(mddev);
5855 return ERR_PTR(-EINVAL);
5858 static void *raid4_takeover(mddev_t *mddev)
5860 /* raid4 can take over:
5861 * raid0 - if there is only one strip zone
5862 * raid5 - if layout is right
5864 if (mddev->level == 0)
5865 return raid45_takeover_raid0(mddev, 4);
5866 if (mddev->level == 5 &&
5867 mddev->layout == ALGORITHM_PARITY_N) {
5868 mddev->new_layout = 0;
5869 mddev->new_level = 4;
5870 return setup_conf(mddev);
5872 return ERR_PTR(-EINVAL);
5875 static struct mdk_personality raid5_personality;
5877 static void *raid6_takeover(mddev_t *mddev)
5879 /* Currently can only take over a raid5. We map the
5880 * personality to an equivalent raid6 personality
5881 * with the Q block at the end.
5885 if (mddev->pers != &raid5_personality)
5886 return ERR_PTR(-EINVAL);
5887 if (mddev->degraded > 1)
5888 return ERR_PTR(-EINVAL);
5889 if (mddev->raid_disks > 253)
5890 return ERR_PTR(-EINVAL);
5891 if (mddev->raid_disks < 3)
5892 return ERR_PTR(-EINVAL);
5894 switch (mddev->layout) {
5895 case ALGORITHM_LEFT_ASYMMETRIC:
5896 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5898 case ALGORITHM_RIGHT_ASYMMETRIC:
5899 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5901 case ALGORITHM_LEFT_SYMMETRIC:
5902 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5904 case ALGORITHM_RIGHT_SYMMETRIC:
5905 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5907 case ALGORITHM_PARITY_0:
5908 new_layout = ALGORITHM_PARITY_0_6;
5910 case ALGORITHM_PARITY_N:
5911 new_layout = ALGORITHM_PARITY_N;
5914 return ERR_PTR(-EINVAL);
5916 mddev->new_level = 6;
5917 mddev->new_layout = new_layout;
5918 mddev->delta_disks = 1;
5919 mddev->raid_disks += 1;
5920 return setup_conf(mddev);
5924 static struct mdk_personality raid6_personality =
5928 .owner = THIS_MODULE,
5929 .make_request = make_request,
5933 .error_handler = error,
5934 .hot_add_disk = raid5_add_disk,
5935 .hot_remove_disk= raid5_remove_disk,
5936 .spare_active = raid5_spare_active,
5937 .sync_request = sync_request,
5938 .resize = raid5_resize,
5940 .check_reshape = raid6_check_reshape,
5941 .start_reshape = raid5_start_reshape,
5942 .finish_reshape = raid5_finish_reshape,
5943 .quiesce = raid5_quiesce,
5944 .takeover = raid6_takeover,
5946 static struct mdk_personality raid5_personality =
5950 .owner = THIS_MODULE,
5951 .make_request = make_request,
5955 .error_handler = error,
5956 .hot_add_disk = raid5_add_disk,
5957 .hot_remove_disk= raid5_remove_disk,
5958 .spare_active = raid5_spare_active,
5959 .sync_request = sync_request,
5960 .resize = raid5_resize,
5962 .check_reshape = raid5_check_reshape,
5963 .start_reshape = raid5_start_reshape,
5964 .finish_reshape = raid5_finish_reshape,
5965 .quiesce = raid5_quiesce,
5966 .takeover = raid5_takeover,
5969 static struct mdk_personality raid4_personality =
5973 .owner = THIS_MODULE,
5974 .make_request = make_request,
5978 .error_handler = error,
5979 .hot_add_disk = raid5_add_disk,
5980 .hot_remove_disk= raid5_remove_disk,
5981 .spare_active = raid5_spare_active,
5982 .sync_request = sync_request,
5983 .resize = raid5_resize,
5985 .check_reshape = raid5_check_reshape,
5986 .start_reshape = raid5_start_reshape,
5987 .finish_reshape = raid5_finish_reshape,
5988 .quiesce = raid5_quiesce,
5989 .takeover = raid4_takeover,
5992 static int __init raid5_init(void)
5994 register_md_personality(&raid6_personality);
5995 register_md_personality(&raid5_personality);
5996 register_md_personality(&raid4_personality);
6000 static void raid5_exit(void)
6002 unregister_md_personality(&raid6_personality);
6003 unregister_md_personality(&raid5_personality);
6004 unregister_md_personality(&raid4_personality);
6007 module_init(raid5_init);
6008 module_exit(raid5_exit);
6009 MODULE_LICENSE("GPL");
6010 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6011 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6012 MODULE_ALIAS("md-raid5");
6013 MODULE_ALIAS("md-raid4");
6014 MODULE_ALIAS("md-level-5");
6015 MODULE_ALIAS("md-level-4");
6016 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6017 MODULE_ALIAS("md-raid6");
6018 MODULE_ALIAS("md-level-6");
6020 /* This used to be two separate modules, they were: */
6021 MODULE_ALIAS("raid5");
6022 MODULE_ALIAS("raid6");