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->bm_write is the number of the last batch successfully written.
31 * conf->bm_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 bm_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/async_tx.h>
49 #include <linux/seq_file.h>
59 #define NR_STRIPES 256
60 #define STRIPE_SIZE PAGE_SIZE
61 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
62 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
63 #define IO_THRESHOLD 1
64 #define BYPASS_THRESHOLD 1
65 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
66 #define HASH_MASK (NR_HASH - 1)
68 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
70 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
71 * order without overlap. There may be several bio's per stripe+device, and
72 * a bio could span several devices.
73 * When walking this list for a particular stripe+device, we must never proceed
74 * beyond a bio that extends past this device, as the next bio might no longer
76 * This macro is used to determine the 'next' bio in the list, given the sector
77 * of the current stripe+device
79 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
81 * The following can be used to debug the driver
83 #define RAID5_PARANOIA 1
84 #if RAID5_PARANOIA && defined(CONFIG_SMP)
85 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
87 # define CHECK_DEVLOCK()
95 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
97 #if !RAID6_USE_EMPTY_ZERO_PAGE
98 /* In .bss so it's zeroed */
99 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
103 * We maintain a biased count of active stripes in the bottom 16 bits of
104 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
106 static inline int raid5_bi_phys_segments(struct bio *bio)
108 return bio->bi_phys_segments & 0xffff;
111 static inline int raid5_bi_hw_segments(struct bio *bio)
113 return (bio->bi_phys_segments >> 16) & 0xffff;
116 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
118 --bio->bi_phys_segments;
119 return raid5_bi_phys_segments(bio);
122 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
124 unsigned short val = raid5_bi_hw_segments(bio);
127 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
131 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
133 bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
136 /* Find first data disk in a raid6 stripe */
137 static inline int raid6_d0(struct stripe_head *sh)
140 /* ddf always start from first device */
142 /* md starts just after Q block */
143 if (sh->qd_idx == sh->disks - 1)
146 return sh->qd_idx + 1;
148 static inline int raid6_next_disk(int disk, int raid_disks)
151 return (disk < raid_disks) ? disk : 0;
154 /* When walking through the disks in a raid5, starting at raid6_d0,
155 * We need to map each disk to a 'slot', where the data disks are slot
156 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
157 * is raid_disks-1. This help does that mapping.
159 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
160 int *count, int syndrome_disks)
164 if (idx == sh->pd_idx)
165 return syndrome_disks;
166 if (idx == sh->qd_idx)
167 return syndrome_disks + 1;
172 static void return_io(struct bio *return_bi)
174 struct bio *bi = return_bi;
177 return_bi = bi->bi_next;
185 static void print_raid5_conf (raid5_conf_t *conf);
187 static int stripe_operations_active(struct stripe_head *sh)
189 return sh->check_state || sh->reconstruct_state ||
190 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
191 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
194 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
196 if (atomic_dec_and_test(&sh->count)) {
197 BUG_ON(!list_empty(&sh->lru));
198 BUG_ON(atomic_read(&conf->active_stripes)==0);
199 if (test_bit(STRIPE_HANDLE, &sh->state)) {
200 if (test_bit(STRIPE_DELAYED, &sh->state)) {
201 list_add_tail(&sh->lru, &conf->delayed_list);
202 blk_plug_device(conf->mddev->queue);
203 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
204 sh->bm_seq - conf->seq_write > 0) {
205 list_add_tail(&sh->lru, &conf->bitmap_list);
206 blk_plug_device(conf->mddev->queue);
208 clear_bit(STRIPE_BIT_DELAY, &sh->state);
209 list_add_tail(&sh->lru, &conf->handle_list);
211 md_wakeup_thread(conf->mddev->thread);
213 BUG_ON(stripe_operations_active(sh));
214 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
215 atomic_dec(&conf->preread_active_stripes);
216 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
217 md_wakeup_thread(conf->mddev->thread);
219 atomic_dec(&conf->active_stripes);
220 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
221 list_add_tail(&sh->lru, &conf->inactive_list);
222 wake_up(&conf->wait_for_stripe);
223 if (conf->retry_read_aligned)
224 md_wakeup_thread(conf->mddev->thread);
230 static void release_stripe(struct stripe_head *sh)
232 raid5_conf_t *conf = sh->raid_conf;
235 spin_lock_irqsave(&conf->device_lock, flags);
236 __release_stripe(conf, sh);
237 spin_unlock_irqrestore(&conf->device_lock, flags);
240 static inline void remove_hash(struct stripe_head *sh)
242 pr_debug("remove_hash(), stripe %llu\n",
243 (unsigned long long)sh->sector);
245 hlist_del_init(&sh->hash);
248 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
250 struct hlist_head *hp = stripe_hash(conf, sh->sector);
252 pr_debug("insert_hash(), stripe %llu\n",
253 (unsigned long long)sh->sector);
256 hlist_add_head(&sh->hash, hp);
260 /* find an idle stripe, make sure it is unhashed, and return it. */
261 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
263 struct stripe_head *sh = NULL;
264 struct list_head *first;
267 if (list_empty(&conf->inactive_list))
269 first = conf->inactive_list.next;
270 sh = list_entry(first, struct stripe_head, lru);
271 list_del_init(first);
273 atomic_inc(&conf->active_stripes);
278 static void shrink_buffers(struct stripe_head *sh, int num)
283 for (i=0; i<num ; i++) {
287 sh->dev[i].page = NULL;
292 static int grow_buffers(struct stripe_head *sh, int num)
296 for (i=0; i<num; i++) {
299 if (!(page = alloc_page(GFP_KERNEL))) {
302 sh->dev[i].page = page;
307 static void raid5_build_block(struct stripe_head *sh, int i);
308 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
309 struct stripe_head *sh);
311 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
313 raid5_conf_t *conf = sh->raid_conf;
316 BUG_ON(atomic_read(&sh->count) != 0);
317 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
318 BUG_ON(stripe_operations_active(sh));
321 pr_debug("init_stripe called, stripe %llu\n",
322 (unsigned long long)sh->sector);
326 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
328 stripe_set_idx(sector, conf, previous, sh);
332 for (i = sh->disks; i--; ) {
333 struct r5dev *dev = &sh->dev[i];
335 if (dev->toread || dev->read || dev->towrite || dev->written ||
336 test_bit(R5_LOCKED, &dev->flags)) {
337 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
338 (unsigned long long)sh->sector, i, dev->toread,
339 dev->read, dev->towrite, dev->written,
340 test_bit(R5_LOCKED, &dev->flags));
344 raid5_build_block(sh, i);
346 insert_hash(conf, sh);
349 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
351 struct stripe_head *sh;
352 struct hlist_node *hn;
355 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
356 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
357 if (sh->sector == sector && sh->disks == disks)
359 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
363 static void unplug_slaves(mddev_t *mddev);
364 static void raid5_unplug_device(struct request_queue *q);
366 static struct stripe_head *
367 get_active_stripe(raid5_conf_t *conf, sector_t sector,
368 int previous, int noblock)
370 struct stripe_head *sh;
371 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
373 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
375 spin_lock_irq(&conf->device_lock);
378 wait_event_lock_irq(conf->wait_for_stripe,
380 conf->device_lock, /* nothing */);
381 sh = __find_stripe(conf, sector, disks);
383 if (!conf->inactive_blocked)
384 sh = get_free_stripe(conf);
385 if (noblock && sh == NULL)
388 conf->inactive_blocked = 1;
389 wait_event_lock_irq(conf->wait_for_stripe,
390 !list_empty(&conf->inactive_list) &&
391 (atomic_read(&conf->active_stripes)
392 < (conf->max_nr_stripes *3/4)
393 || !conf->inactive_blocked),
395 raid5_unplug_device(conf->mddev->queue)
397 conf->inactive_blocked = 0;
399 init_stripe(sh, sector, previous);
401 if (atomic_read(&sh->count)) {
402 BUG_ON(!list_empty(&sh->lru));
404 if (!test_bit(STRIPE_HANDLE, &sh->state))
405 atomic_inc(&conf->active_stripes);
406 if (list_empty(&sh->lru) &&
407 !test_bit(STRIPE_EXPANDING, &sh->state))
409 list_del_init(&sh->lru);
412 } while (sh == NULL);
415 atomic_inc(&sh->count);
417 spin_unlock_irq(&conf->device_lock);
422 raid5_end_read_request(struct bio *bi, int error);
424 raid5_end_write_request(struct bio *bi, int error);
426 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
428 raid5_conf_t *conf = sh->raid_conf;
429 int i, disks = sh->disks;
433 for (i = disks; i--; ) {
437 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
439 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
444 bi = &sh->dev[i].req;
448 bi->bi_end_io = raid5_end_write_request;
450 bi->bi_end_io = raid5_end_read_request;
453 rdev = rcu_dereference(conf->disks[i].rdev);
454 if (rdev && test_bit(Faulty, &rdev->flags))
457 atomic_inc(&rdev->nr_pending);
461 if (s->syncing || s->expanding || s->expanded)
462 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
464 set_bit(STRIPE_IO_STARTED, &sh->state);
466 bi->bi_bdev = rdev->bdev;
467 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
468 __func__, (unsigned long long)sh->sector,
470 atomic_inc(&sh->count);
471 bi->bi_sector = sh->sector + rdev->data_offset;
472 bi->bi_flags = 1 << BIO_UPTODATE;
476 bi->bi_io_vec = &sh->dev[i].vec;
477 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
478 bi->bi_io_vec[0].bv_offset = 0;
479 bi->bi_size = STRIPE_SIZE;
482 test_bit(R5_ReWrite, &sh->dev[i].flags))
483 atomic_add(STRIPE_SECTORS,
484 &rdev->corrected_errors);
485 generic_make_request(bi);
488 set_bit(STRIPE_DEGRADED, &sh->state);
489 pr_debug("skip op %ld on disc %d for sector %llu\n",
490 bi->bi_rw, i, (unsigned long long)sh->sector);
491 clear_bit(R5_LOCKED, &sh->dev[i].flags);
492 set_bit(STRIPE_HANDLE, &sh->state);
497 static struct dma_async_tx_descriptor *
498 async_copy_data(int frombio, struct bio *bio, struct page *page,
499 sector_t sector, struct dma_async_tx_descriptor *tx)
502 struct page *bio_page;
506 if (bio->bi_sector >= sector)
507 page_offset = (signed)(bio->bi_sector - sector) * 512;
509 page_offset = (signed)(sector - bio->bi_sector) * -512;
510 bio_for_each_segment(bvl, bio, i) {
511 int len = bio_iovec_idx(bio, i)->bv_len;
515 if (page_offset < 0) {
516 b_offset = -page_offset;
517 page_offset += b_offset;
521 if (len > 0 && page_offset + len > STRIPE_SIZE)
522 clen = STRIPE_SIZE - page_offset;
527 b_offset += bio_iovec_idx(bio, i)->bv_offset;
528 bio_page = bio_iovec_idx(bio, i)->bv_page;
530 tx = async_memcpy(page, bio_page, page_offset,
535 tx = async_memcpy(bio_page, page, b_offset,
540 if (clen < len) /* hit end of page */
548 static void ops_complete_biofill(void *stripe_head_ref)
550 struct stripe_head *sh = stripe_head_ref;
551 struct bio *return_bi = NULL;
552 raid5_conf_t *conf = sh->raid_conf;
555 pr_debug("%s: stripe %llu\n", __func__,
556 (unsigned long long)sh->sector);
558 /* clear completed biofills */
559 spin_lock_irq(&conf->device_lock);
560 for (i = sh->disks; i--; ) {
561 struct r5dev *dev = &sh->dev[i];
563 /* acknowledge completion of a biofill operation */
564 /* and check if we need to reply to a read request,
565 * new R5_Wantfill requests are held off until
566 * !STRIPE_BIOFILL_RUN
568 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
569 struct bio *rbi, *rbi2;
574 while (rbi && rbi->bi_sector <
575 dev->sector + STRIPE_SECTORS) {
576 rbi2 = r5_next_bio(rbi, dev->sector);
577 if (!raid5_dec_bi_phys_segments(rbi)) {
578 rbi->bi_next = return_bi;
585 spin_unlock_irq(&conf->device_lock);
586 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
588 return_io(return_bi);
590 set_bit(STRIPE_HANDLE, &sh->state);
594 static void ops_run_biofill(struct stripe_head *sh)
596 struct dma_async_tx_descriptor *tx = NULL;
597 raid5_conf_t *conf = sh->raid_conf;
600 pr_debug("%s: stripe %llu\n", __func__,
601 (unsigned long long)sh->sector);
603 for (i = sh->disks; i--; ) {
604 struct r5dev *dev = &sh->dev[i];
605 if (test_bit(R5_Wantfill, &dev->flags)) {
607 spin_lock_irq(&conf->device_lock);
608 dev->read = rbi = dev->toread;
610 spin_unlock_irq(&conf->device_lock);
611 while (rbi && rbi->bi_sector <
612 dev->sector + STRIPE_SECTORS) {
613 tx = async_copy_data(0, rbi, dev->page,
615 rbi = r5_next_bio(rbi, dev->sector);
620 atomic_inc(&sh->count);
621 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
622 ops_complete_biofill, sh);
625 static void ops_complete_compute5(void *stripe_head_ref)
627 struct stripe_head *sh = stripe_head_ref;
628 int target = sh->ops.target;
629 struct r5dev *tgt = &sh->dev[target];
631 pr_debug("%s: stripe %llu\n", __func__,
632 (unsigned long long)sh->sector);
634 set_bit(R5_UPTODATE, &tgt->flags);
635 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
636 clear_bit(R5_Wantcompute, &tgt->flags);
637 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
638 if (sh->check_state == check_state_compute_run)
639 sh->check_state = check_state_compute_result;
640 set_bit(STRIPE_HANDLE, &sh->state);
644 static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
646 /* kernel stack size limits the total number of disks */
647 int disks = sh->disks;
648 struct page *xor_srcs[disks];
649 int target = sh->ops.target;
650 struct r5dev *tgt = &sh->dev[target];
651 struct page *xor_dest = tgt->page;
653 struct dma_async_tx_descriptor *tx;
656 pr_debug("%s: stripe %llu block: %d\n",
657 __func__, (unsigned long long)sh->sector, target);
658 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
660 for (i = disks; i--; )
662 xor_srcs[count++] = sh->dev[i].page;
664 atomic_inc(&sh->count);
666 if (unlikely(count == 1))
667 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
668 0, NULL, ops_complete_compute5, sh);
670 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
671 ASYNC_TX_XOR_ZERO_DST, NULL,
672 ops_complete_compute5, sh);
677 static void ops_complete_prexor(void *stripe_head_ref)
679 struct stripe_head *sh = stripe_head_ref;
681 pr_debug("%s: stripe %llu\n", __func__,
682 (unsigned long long)sh->sector);
685 static struct dma_async_tx_descriptor *
686 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
688 /* kernel stack size limits the total number of disks */
689 int disks = sh->disks;
690 struct page *xor_srcs[disks];
691 int count = 0, pd_idx = sh->pd_idx, i;
693 /* existing parity data subtracted */
694 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
696 pr_debug("%s: stripe %llu\n", __func__,
697 (unsigned long long)sh->sector);
699 for (i = disks; i--; ) {
700 struct r5dev *dev = &sh->dev[i];
701 /* Only process blocks that are known to be uptodate */
702 if (test_bit(R5_Wantdrain, &dev->flags))
703 xor_srcs[count++] = dev->page;
706 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
707 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
708 ops_complete_prexor, sh);
713 static struct dma_async_tx_descriptor *
714 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
716 int disks = sh->disks;
719 pr_debug("%s: stripe %llu\n", __func__,
720 (unsigned long long)sh->sector);
722 for (i = disks; i--; ) {
723 struct r5dev *dev = &sh->dev[i];
726 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
729 spin_lock(&sh->lock);
730 chosen = dev->towrite;
732 BUG_ON(dev->written);
733 wbi = dev->written = chosen;
734 spin_unlock(&sh->lock);
736 while (wbi && wbi->bi_sector <
737 dev->sector + STRIPE_SECTORS) {
738 tx = async_copy_data(1, wbi, dev->page,
740 wbi = r5_next_bio(wbi, dev->sector);
748 static void ops_complete_postxor(void *stripe_head_ref)
750 struct stripe_head *sh = stripe_head_ref;
751 int disks = sh->disks, i, pd_idx = sh->pd_idx;
753 pr_debug("%s: stripe %llu\n", __func__,
754 (unsigned long long)sh->sector);
756 for (i = disks; i--; ) {
757 struct r5dev *dev = &sh->dev[i];
758 if (dev->written || i == pd_idx)
759 set_bit(R5_UPTODATE, &dev->flags);
762 if (sh->reconstruct_state == reconstruct_state_drain_run)
763 sh->reconstruct_state = reconstruct_state_drain_result;
764 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
765 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
767 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
768 sh->reconstruct_state = reconstruct_state_result;
771 set_bit(STRIPE_HANDLE, &sh->state);
776 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
778 /* kernel stack size limits the total number of disks */
779 int disks = sh->disks;
780 struct page *xor_srcs[disks];
782 int count = 0, pd_idx = sh->pd_idx, i;
783 struct page *xor_dest;
787 pr_debug("%s: stripe %llu\n", __func__,
788 (unsigned long long)sh->sector);
790 /* check if prexor is active which means only process blocks
791 * that are part of a read-modify-write (written)
793 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
795 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
796 for (i = disks; i--; ) {
797 struct r5dev *dev = &sh->dev[i];
799 xor_srcs[count++] = dev->page;
802 xor_dest = sh->dev[pd_idx].page;
803 for (i = disks; i--; ) {
804 struct r5dev *dev = &sh->dev[i];
806 xor_srcs[count++] = dev->page;
810 /* 1/ if we prexor'd then the dest is reused as a source
811 * 2/ if we did not prexor then we are redoing the parity
812 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
813 * for the synchronous xor case
815 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
816 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
818 atomic_inc(&sh->count);
820 if (unlikely(count == 1)) {
821 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
822 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
823 flags, tx, ops_complete_postxor, sh);
825 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
826 flags, tx, ops_complete_postxor, sh);
829 static void ops_complete_check(void *stripe_head_ref)
831 struct stripe_head *sh = stripe_head_ref;
833 pr_debug("%s: stripe %llu\n", __func__,
834 (unsigned long long)sh->sector);
836 sh->check_state = check_state_check_result;
837 set_bit(STRIPE_HANDLE, &sh->state);
841 static void ops_run_check(struct stripe_head *sh)
843 /* kernel stack size limits the total number of disks */
844 int disks = sh->disks;
845 struct page *xor_srcs[disks];
846 struct dma_async_tx_descriptor *tx;
848 int count = 0, pd_idx = sh->pd_idx, i;
849 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
851 pr_debug("%s: stripe %llu\n", __func__,
852 (unsigned long long)sh->sector);
854 for (i = disks; i--; ) {
855 struct r5dev *dev = &sh->dev[i];
857 xor_srcs[count++] = dev->page;
860 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
861 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
863 atomic_inc(&sh->count);
864 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
865 ops_complete_check, sh);
868 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
870 int overlap_clear = 0, i, disks = sh->disks;
871 struct dma_async_tx_descriptor *tx = NULL;
873 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
878 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
879 tx = ops_run_compute5(sh);
880 /* terminate the chain if postxor is not set to be run */
881 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
885 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
886 tx = ops_run_prexor(sh, tx);
888 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
889 tx = ops_run_biodrain(sh, tx);
893 if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
894 ops_run_postxor(sh, tx);
896 if (test_bit(STRIPE_OP_CHECK, &ops_request))
900 for (i = disks; i--; ) {
901 struct r5dev *dev = &sh->dev[i];
902 if (test_and_clear_bit(R5_Overlap, &dev->flags))
903 wake_up(&sh->raid_conf->wait_for_overlap);
907 static int grow_one_stripe(raid5_conf_t *conf)
909 struct stripe_head *sh;
910 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
913 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
914 sh->raid_conf = conf;
915 spin_lock_init(&sh->lock);
917 if (grow_buffers(sh, conf->raid_disks)) {
918 shrink_buffers(sh, conf->raid_disks);
919 kmem_cache_free(conf->slab_cache, sh);
922 sh->disks = conf->raid_disks;
923 /* we just created an active stripe so... */
924 atomic_set(&sh->count, 1);
925 atomic_inc(&conf->active_stripes);
926 INIT_LIST_HEAD(&sh->lru);
931 static int grow_stripes(raid5_conf_t *conf, int num)
933 struct kmem_cache *sc;
934 int devs = conf->raid_disks;
936 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
937 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
938 conf->active_name = 0;
939 sc = kmem_cache_create(conf->cache_name[conf->active_name],
940 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
944 conf->slab_cache = sc;
945 conf->pool_size = devs;
947 if (!grow_one_stripe(conf))
952 #ifdef CONFIG_MD_RAID5_RESHAPE
953 static int resize_stripes(raid5_conf_t *conf, int newsize)
955 /* Make all the stripes able to hold 'newsize' devices.
956 * New slots in each stripe get 'page' set to a new page.
958 * This happens in stages:
959 * 1/ create a new kmem_cache and allocate the required number of
961 * 2/ gather all the old stripe_heads and tranfer the pages across
962 * to the new stripe_heads. This will have the side effect of
963 * freezing the array as once all stripe_heads have been collected,
964 * no IO will be possible. Old stripe heads are freed once their
965 * pages have been transferred over, and the old kmem_cache is
966 * freed when all stripes are done.
967 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
968 * we simple return a failre status - no need to clean anything up.
969 * 4/ allocate new pages for the new slots in the new stripe_heads.
970 * If this fails, we don't bother trying the shrink the
971 * stripe_heads down again, we just leave them as they are.
972 * As each stripe_head is processed the new one is released into
975 * Once step2 is started, we cannot afford to wait for a write,
976 * so we use GFP_NOIO allocations.
978 struct stripe_head *osh, *nsh;
979 LIST_HEAD(newstripes);
980 struct disk_info *ndisks;
982 struct kmem_cache *sc;
985 if (newsize <= conf->pool_size)
986 return 0; /* never bother to shrink */
988 err = md_allow_write(conf->mddev);
993 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
994 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
999 for (i = conf->max_nr_stripes; i; i--) {
1000 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1004 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1006 nsh->raid_conf = conf;
1007 spin_lock_init(&nsh->lock);
1009 list_add(&nsh->lru, &newstripes);
1012 /* didn't get enough, give up */
1013 while (!list_empty(&newstripes)) {
1014 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1015 list_del(&nsh->lru);
1016 kmem_cache_free(sc, nsh);
1018 kmem_cache_destroy(sc);
1021 /* Step 2 - Must use GFP_NOIO now.
1022 * OK, we have enough stripes, start collecting inactive
1023 * stripes and copying them over
1025 list_for_each_entry(nsh, &newstripes, lru) {
1026 spin_lock_irq(&conf->device_lock);
1027 wait_event_lock_irq(conf->wait_for_stripe,
1028 !list_empty(&conf->inactive_list),
1030 unplug_slaves(conf->mddev)
1032 osh = get_free_stripe(conf);
1033 spin_unlock_irq(&conf->device_lock);
1034 atomic_set(&nsh->count, 1);
1035 for(i=0; i<conf->pool_size; i++)
1036 nsh->dev[i].page = osh->dev[i].page;
1037 for( ; i<newsize; i++)
1038 nsh->dev[i].page = NULL;
1039 kmem_cache_free(conf->slab_cache, osh);
1041 kmem_cache_destroy(conf->slab_cache);
1044 * At this point, we are holding all the stripes so the array
1045 * is completely stalled, so now is a good time to resize
1048 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1050 for (i=0; i<conf->raid_disks; i++)
1051 ndisks[i] = conf->disks[i];
1053 conf->disks = ndisks;
1057 /* Step 4, return new stripes to service */
1058 while(!list_empty(&newstripes)) {
1059 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1060 list_del_init(&nsh->lru);
1061 for (i=conf->raid_disks; i < newsize; i++)
1062 if (nsh->dev[i].page == NULL) {
1063 struct page *p = alloc_page(GFP_NOIO);
1064 nsh->dev[i].page = p;
1068 release_stripe(nsh);
1070 /* critical section pass, GFP_NOIO no longer needed */
1072 conf->slab_cache = sc;
1073 conf->active_name = 1-conf->active_name;
1074 conf->pool_size = newsize;
1079 static int drop_one_stripe(raid5_conf_t *conf)
1081 struct stripe_head *sh;
1083 spin_lock_irq(&conf->device_lock);
1084 sh = get_free_stripe(conf);
1085 spin_unlock_irq(&conf->device_lock);
1088 BUG_ON(atomic_read(&sh->count));
1089 shrink_buffers(sh, conf->pool_size);
1090 kmem_cache_free(conf->slab_cache, sh);
1091 atomic_dec(&conf->active_stripes);
1095 static void shrink_stripes(raid5_conf_t *conf)
1097 while (drop_one_stripe(conf))
1100 if (conf->slab_cache)
1101 kmem_cache_destroy(conf->slab_cache);
1102 conf->slab_cache = NULL;
1105 static void raid5_end_read_request(struct bio * bi, int error)
1107 struct stripe_head *sh = bi->bi_private;
1108 raid5_conf_t *conf = sh->raid_conf;
1109 int disks = sh->disks, i;
1110 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1111 char b[BDEVNAME_SIZE];
1115 for (i=0 ; i<disks; i++)
1116 if (bi == &sh->dev[i].req)
1119 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1120 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1128 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1129 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1130 rdev = conf->disks[i].rdev;
1131 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1132 " (%lu sectors at %llu on %s)\n",
1133 mdname(conf->mddev), STRIPE_SECTORS,
1134 (unsigned long long)(sh->sector
1135 + rdev->data_offset),
1136 bdevname(rdev->bdev, b));
1137 clear_bit(R5_ReadError, &sh->dev[i].flags);
1138 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1140 if (atomic_read(&conf->disks[i].rdev->read_errors))
1141 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1143 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1145 rdev = conf->disks[i].rdev;
1147 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1148 atomic_inc(&rdev->read_errors);
1149 if (conf->mddev->degraded)
1150 printk_rl(KERN_WARNING
1151 "raid5:%s: read error not correctable "
1152 "(sector %llu on %s).\n",
1153 mdname(conf->mddev),
1154 (unsigned long long)(sh->sector
1155 + rdev->data_offset),
1157 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1159 printk_rl(KERN_WARNING
1160 "raid5:%s: read error NOT corrected!! "
1161 "(sector %llu on %s).\n",
1162 mdname(conf->mddev),
1163 (unsigned long long)(sh->sector
1164 + rdev->data_offset),
1166 else if (atomic_read(&rdev->read_errors)
1167 > conf->max_nr_stripes)
1169 "raid5:%s: Too many read errors, failing device %s.\n",
1170 mdname(conf->mddev), bdn);
1174 set_bit(R5_ReadError, &sh->dev[i].flags);
1176 clear_bit(R5_ReadError, &sh->dev[i].flags);
1177 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1178 md_error(conf->mddev, rdev);
1181 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1182 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1183 set_bit(STRIPE_HANDLE, &sh->state);
1187 static void raid5_end_write_request(struct bio *bi, int error)
1189 struct stripe_head *sh = bi->bi_private;
1190 raid5_conf_t *conf = sh->raid_conf;
1191 int disks = sh->disks, i;
1192 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1194 for (i=0 ; i<disks; i++)
1195 if (bi == &sh->dev[i].req)
1198 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1199 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1207 md_error(conf->mddev, conf->disks[i].rdev);
1209 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1211 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1212 set_bit(STRIPE_HANDLE, &sh->state);
1217 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1219 static void raid5_build_block(struct stripe_head *sh, int i)
1221 struct r5dev *dev = &sh->dev[i];
1223 bio_init(&dev->req);
1224 dev->req.bi_io_vec = &dev->vec;
1226 dev->req.bi_max_vecs++;
1227 dev->vec.bv_page = dev->page;
1228 dev->vec.bv_len = STRIPE_SIZE;
1229 dev->vec.bv_offset = 0;
1231 dev->req.bi_sector = sh->sector;
1232 dev->req.bi_private = sh;
1235 dev->sector = compute_blocknr(sh, i);
1238 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1240 char b[BDEVNAME_SIZE];
1241 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1242 pr_debug("raid5: error called\n");
1244 if (!test_bit(Faulty, &rdev->flags)) {
1245 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1246 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1247 unsigned long flags;
1248 spin_lock_irqsave(&conf->device_lock, flags);
1250 spin_unlock_irqrestore(&conf->device_lock, flags);
1252 * if recovery was running, make sure it aborts.
1254 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1256 set_bit(Faulty, &rdev->flags);
1258 "raid5: Disk failure on %s, disabling device.\n"
1259 "raid5: Operation continuing on %d devices.\n",
1260 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1265 * Input: a 'big' sector number,
1266 * Output: index of the data and parity disk, and the sector # in them.
1268 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1269 int previous, int *dd_idx,
1270 struct stripe_head *sh)
1273 unsigned long chunk_number;
1274 unsigned int chunk_offset;
1277 sector_t new_sector;
1278 int sectors_per_chunk = conf->chunk_size >> 9;
1279 int raid_disks = previous ? conf->previous_raid_disks
1281 int data_disks = raid_disks - conf->max_degraded;
1283 /* First compute the information on this sector */
1286 * Compute the chunk number and the sector offset inside the chunk
1288 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1289 chunk_number = r_sector;
1290 BUG_ON(r_sector != chunk_number);
1293 * Compute the stripe number
1295 stripe = chunk_number / data_disks;
1298 * Compute the data disk and parity disk indexes inside the stripe
1300 *dd_idx = chunk_number % data_disks;
1303 * Select the parity disk based on the user selected algorithm.
1305 pd_idx = qd_idx = ~0;
1306 switch(conf->level) {
1308 pd_idx = data_disks;
1311 switch (conf->algorithm) {
1312 case ALGORITHM_LEFT_ASYMMETRIC:
1313 pd_idx = data_disks - stripe % raid_disks;
1314 if (*dd_idx >= pd_idx)
1317 case ALGORITHM_RIGHT_ASYMMETRIC:
1318 pd_idx = stripe % raid_disks;
1319 if (*dd_idx >= pd_idx)
1322 case ALGORITHM_LEFT_SYMMETRIC:
1323 pd_idx = data_disks - stripe % raid_disks;
1324 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1326 case ALGORITHM_RIGHT_SYMMETRIC:
1327 pd_idx = stripe % raid_disks;
1328 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1330 case ALGORITHM_PARITY_0:
1334 case ALGORITHM_PARITY_N:
1335 pd_idx = data_disks;
1338 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1345 switch (conf->algorithm) {
1346 case ALGORITHM_LEFT_ASYMMETRIC:
1347 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1348 qd_idx = pd_idx + 1;
1349 if (pd_idx == raid_disks-1) {
1350 (*dd_idx)++; /* Q D D D P */
1352 } else if (*dd_idx >= pd_idx)
1353 (*dd_idx) += 2; /* D D P Q D */
1355 case ALGORITHM_RIGHT_ASYMMETRIC:
1356 pd_idx = stripe % raid_disks;
1357 qd_idx = pd_idx + 1;
1358 if (pd_idx == raid_disks-1) {
1359 (*dd_idx)++; /* Q D D D P */
1361 } else if (*dd_idx >= pd_idx)
1362 (*dd_idx) += 2; /* D D P Q D */
1364 case ALGORITHM_LEFT_SYMMETRIC:
1365 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1366 qd_idx = (pd_idx + 1) % raid_disks;
1367 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1369 case ALGORITHM_RIGHT_SYMMETRIC:
1370 pd_idx = stripe % raid_disks;
1371 qd_idx = (pd_idx + 1) % raid_disks;
1372 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1375 case ALGORITHM_PARITY_0:
1380 case ALGORITHM_PARITY_N:
1381 pd_idx = data_disks;
1382 qd_idx = data_disks + 1;
1385 case ALGORITHM_ROTATING_ZERO_RESTART:
1386 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1387 * of blocks for computing Q is different.
1389 pd_idx = stripe % raid_disks;
1390 qd_idx = pd_idx + 1;
1391 if (pd_idx == raid_disks-1) {
1392 (*dd_idx)++; /* Q D D D P */
1394 } else if (*dd_idx >= pd_idx)
1395 (*dd_idx) += 2; /* D D P Q D */
1399 case ALGORITHM_ROTATING_N_RESTART:
1400 /* Same a left_asymmetric, by first stripe is
1401 * D D D P Q rather than
1404 pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1405 qd_idx = pd_idx + 1;
1406 if (pd_idx == raid_disks-1) {
1407 (*dd_idx)++; /* Q D D D P */
1409 } else if (*dd_idx >= pd_idx)
1410 (*dd_idx) += 2; /* D D P Q D */
1414 case ALGORITHM_ROTATING_N_CONTINUE:
1415 /* Same as left_symmetric but Q is before P */
1416 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1417 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1418 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1422 case ALGORITHM_LEFT_ASYMMETRIC_6:
1423 /* RAID5 left_asymmetric, with Q on last device */
1424 pd_idx = data_disks - stripe % (raid_disks-1);
1425 if (*dd_idx >= pd_idx)
1427 qd_idx = raid_disks - 1;
1430 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1431 pd_idx = stripe % (raid_disks-1);
1432 if (*dd_idx >= pd_idx)
1434 qd_idx = raid_disks - 1;
1437 case ALGORITHM_LEFT_SYMMETRIC_6:
1438 pd_idx = data_disks - stripe % (raid_disks-1);
1439 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1440 qd_idx = raid_disks - 1;
1443 case ALGORITHM_RIGHT_SYMMETRIC_6:
1444 pd_idx = stripe % (raid_disks-1);
1445 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1446 qd_idx = raid_disks - 1;
1449 case ALGORITHM_PARITY_0_6:
1452 qd_idx = raid_disks - 1;
1457 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1465 sh->pd_idx = pd_idx;
1466 sh->qd_idx = qd_idx;
1467 sh->ddf_layout = ddf_layout;
1470 * Finally, compute the new sector number
1472 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1477 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1479 raid5_conf_t *conf = sh->raid_conf;
1480 int raid_disks = sh->disks;
1481 int data_disks = raid_disks - conf->max_degraded;
1482 sector_t new_sector = sh->sector, check;
1483 int sectors_per_chunk = conf->chunk_size >> 9;
1486 int chunk_number, dummy1, dd_idx = i;
1488 struct stripe_head sh2;
1491 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1492 stripe = new_sector;
1493 BUG_ON(new_sector != stripe);
1495 if (i == sh->pd_idx)
1497 switch(conf->level) {
1500 switch (conf->algorithm) {
1501 case ALGORITHM_LEFT_ASYMMETRIC:
1502 case ALGORITHM_RIGHT_ASYMMETRIC:
1506 case ALGORITHM_LEFT_SYMMETRIC:
1507 case ALGORITHM_RIGHT_SYMMETRIC:
1510 i -= (sh->pd_idx + 1);
1512 case ALGORITHM_PARITY_0:
1515 case ALGORITHM_PARITY_N:
1518 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1524 if (i == sh->qd_idx)
1525 return 0; /* It is the Q disk */
1526 switch (conf->algorithm) {
1527 case ALGORITHM_LEFT_ASYMMETRIC:
1528 case ALGORITHM_RIGHT_ASYMMETRIC:
1529 case ALGORITHM_ROTATING_ZERO_RESTART:
1530 case ALGORITHM_ROTATING_N_RESTART:
1531 if (sh->pd_idx == raid_disks-1)
1532 i--; /* Q D D D P */
1533 else if (i > sh->pd_idx)
1534 i -= 2; /* D D P Q D */
1536 case ALGORITHM_LEFT_SYMMETRIC:
1537 case ALGORITHM_RIGHT_SYMMETRIC:
1538 if (sh->pd_idx == raid_disks-1)
1539 i--; /* Q D D D P */
1544 i -= (sh->pd_idx + 2);
1547 case ALGORITHM_PARITY_0:
1550 case ALGORITHM_PARITY_N:
1552 case ALGORITHM_ROTATING_N_CONTINUE:
1553 if (sh->pd_idx == 0)
1554 i--; /* P D D D Q */
1555 else if (i > sh->pd_idx)
1556 i -= 2; /* D D Q P D */
1558 case ALGORITHM_LEFT_ASYMMETRIC_6:
1559 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1563 case ALGORITHM_LEFT_SYMMETRIC_6:
1564 case ALGORITHM_RIGHT_SYMMETRIC_6:
1566 i += data_disks + 1;
1567 i -= (sh->pd_idx + 1);
1569 case ALGORITHM_PARITY_0_6:
1573 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1580 chunk_number = stripe * data_disks + i;
1581 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1583 check = raid5_compute_sector(conf, r_sector,
1584 (raid_disks != conf->raid_disks),
1586 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1587 || sh2.qd_idx != sh->qd_idx) {
1588 printk(KERN_ERR "compute_blocknr: map not correct\n");
1597 * Copy data between a page in the stripe cache, and one or more bion
1598 * The page could align with the middle of the bio, or there could be
1599 * several bion, each with several bio_vecs, which cover part of the page
1600 * Multiple bion are linked together on bi_next. There may be extras
1601 * at the end of this list. We ignore them.
1603 static void copy_data(int frombio, struct bio *bio,
1607 char *pa = page_address(page);
1608 struct bio_vec *bvl;
1612 if (bio->bi_sector >= sector)
1613 page_offset = (signed)(bio->bi_sector - sector) * 512;
1615 page_offset = (signed)(sector - bio->bi_sector) * -512;
1616 bio_for_each_segment(bvl, bio, i) {
1617 int len = bio_iovec_idx(bio,i)->bv_len;
1621 if (page_offset < 0) {
1622 b_offset = -page_offset;
1623 page_offset += b_offset;
1627 if (len > 0 && page_offset + len > STRIPE_SIZE)
1628 clen = STRIPE_SIZE - page_offset;
1632 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1634 memcpy(pa+page_offset, ba+b_offset, clen);
1636 memcpy(ba+b_offset, pa+page_offset, clen);
1637 __bio_kunmap_atomic(ba, KM_USER0);
1639 if (clen < len) /* hit end of page */
1645 #define check_xor() do { \
1646 if (count == MAX_XOR_BLOCKS) { \
1647 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1652 static void compute_parity6(struct stripe_head *sh, int method)
1654 raid5_conf_t *conf = sh->raid_conf;
1655 int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1656 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1658 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1659 void *ptrs[syndrome_disks+2];
1661 pd_idx = sh->pd_idx;
1662 qd_idx = sh->qd_idx;
1663 d0_idx = raid6_d0(sh);
1665 pr_debug("compute_parity, stripe %llu, method %d\n",
1666 (unsigned long long)sh->sector, method);
1669 case READ_MODIFY_WRITE:
1670 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1671 case RECONSTRUCT_WRITE:
1672 for (i= disks; i-- ;)
1673 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1674 chosen = sh->dev[i].towrite;
1675 sh->dev[i].towrite = NULL;
1677 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1678 wake_up(&conf->wait_for_overlap);
1680 BUG_ON(sh->dev[i].written);
1681 sh->dev[i].written = chosen;
1685 BUG(); /* Not implemented yet */
1688 for (i = disks; i--;)
1689 if (sh->dev[i].written) {
1690 sector_t sector = sh->dev[i].sector;
1691 struct bio *wbi = sh->dev[i].written;
1692 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1693 copy_data(1, wbi, sh->dev[i].page, sector);
1694 wbi = r5_next_bio(wbi, sector);
1697 set_bit(R5_LOCKED, &sh->dev[i].flags);
1698 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1701 /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1703 for (i = 0; i < disks; i++)
1704 ptrs[i] = (void *)raid6_empty_zero_page;
1709 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1711 ptrs[slot] = page_address(sh->dev[i].page);
1712 if (slot < syndrome_disks &&
1713 !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1714 printk(KERN_ERR "block %d/%d not uptodate "
1715 "on parity calc\n", i, count);
1719 i = raid6_next_disk(i, disks);
1720 } while (i != d0_idx);
1721 BUG_ON(count != syndrome_disks);
1723 raid6_call.gen_syndrome(syndrome_disks+2, STRIPE_SIZE, ptrs);
1726 case RECONSTRUCT_WRITE:
1727 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1728 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1729 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1730 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1733 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1734 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1740 /* Compute one missing block */
1741 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1743 int i, count, disks = sh->disks;
1744 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1745 int qd_idx = sh->qd_idx;
1747 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1748 (unsigned long long)sh->sector, dd_idx);
1750 if ( dd_idx == qd_idx ) {
1751 /* We're actually computing the Q drive */
1752 compute_parity6(sh, UPDATE_PARITY);
1754 dest = page_address(sh->dev[dd_idx].page);
1755 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1757 for (i = disks ; i--; ) {
1758 if (i == dd_idx || i == qd_idx)
1760 p = page_address(sh->dev[i].page);
1761 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1764 printk("compute_block() %d, stripe %llu, %d"
1765 " not present\n", dd_idx,
1766 (unsigned long long)sh->sector, i);
1771 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1772 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1773 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1777 /* Compute two missing blocks */
1778 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1780 int i, count, disks = sh->disks;
1781 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1782 int d0_idx = raid6_d0(sh);
1783 int faila = -1, failb = -1;
1784 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1785 void *ptrs[syndrome_disks+2];
1787 for (i = 0; i < disks ; i++)
1788 ptrs[i] = (void *)raid6_empty_zero_page;
1792 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1794 ptrs[slot] = page_address(sh->dev[i].page);
1800 i = raid6_next_disk(i, disks);
1801 } while (i != d0_idx);
1802 BUG_ON(count != syndrome_disks);
1804 BUG_ON(faila == failb);
1805 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1807 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1808 (unsigned long long)sh->sector, dd_idx1, dd_idx2,
1811 if (failb == syndrome_disks+1) {
1812 /* Q disk is one of the missing disks */
1813 if (faila == syndrome_disks) {
1814 /* Missing P+Q, just recompute */
1815 compute_parity6(sh, UPDATE_PARITY);
1818 /* We're missing D+Q; recompute D from P */
1819 compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
1822 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1827 /* We're missing D+P or D+D; */
1828 if (failb == syndrome_disks) {
1829 /* We're missing D+P. */
1830 raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE, faila, ptrs);
1832 /* We're missing D+D. */
1833 raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE, faila, failb,
1837 /* Both the above update both missing blocks */
1838 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1839 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1843 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1844 int rcw, int expand)
1846 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1849 /* if we are not expanding this is a proper write request, and
1850 * there will be bios with new data to be drained into the
1854 sh->reconstruct_state = reconstruct_state_drain_run;
1855 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1857 sh->reconstruct_state = reconstruct_state_run;
1859 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1861 for (i = disks; i--; ) {
1862 struct r5dev *dev = &sh->dev[i];
1865 set_bit(R5_LOCKED, &dev->flags);
1866 set_bit(R5_Wantdrain, &dev->flags);
1868 clear_bit(R5_UPTODATE, &dev->flags);
1872 if (s->locked + 1 == disks)
1873 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1874 atomic_inc(&sh->raid_conf->pending_full_writes);
1876 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1877 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1879 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1880 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1881 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1882 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1884 for (i = disks; i--; ) {
1885 struct r5dev *dev = &sh->dev[i];
1890 (test_bit(R5_UPTODATE, &dev->flags) ||
1891 test_bit(R5_Wantcompute, &dev->flags))) {
1892 set_bit(R5_Wantdrain, &dev->flags);
1893 set_bit(R5_LOCKED, &dev->flags);
1894 clear_bit(R5_UPTODATE, &dev->flags);
1900 /* keep the parity disk locked while asynchronous operations
1903 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1904 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1907 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1908 __func__, (unsigned long long)sh->sector,
1909 s->locked, s->ops_request);
1913 * Each stripe/dev can have one or more bion attached.
1914 * toread/towrite point to the first in a chain.
1915 * The bi_next chain must be in order.
1917 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1920 raid5_conf_t *conf = sh->raid_conf;
1923 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1924 (unsigned long long)bi->bi_sector,
1925 (unsigned long long)sh->sector);
1928 spin_lock(&sh->lock);
1929 spin_lock_irq(&conf->device_lock);
1931 bip = &sh->dev[dd_idx].towrite;
1932 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1935 bip = &sh->dev[dd_idx].toread;
1936 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1937 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1939 bip = & (*bip)->bi_next;
1941 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1944 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1948 bi->bi_phys_segments++;
1949 spin_unlock_irq(&conf->device_lock);
1950 spin_unlock(&sh->lock);
1952 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1953 (unsigned long long)bi->bi_sector,
1954 (unsigned long long)sh->sector, dd_idx);
1956 if (conf->mddev->bitmap && firstwrite) {
1957 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1959 sh->bm_seq = conf->seq_flush+1;
1960 set_bit(STRIPE_BIT_DELAY, &sh->state);
1964 /* check if page is covered */
1965 sector_t sector = sh->dev[dd_idx].sector;
1966 for (bi=sh->dev[dd_idx].towrite;
1967 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1968 bi && bi->bi_sector <= sector;
1969 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1970 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1971 sector = bi->bi_sector + (bi->bi_size>>9);
1973 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1974 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1979 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1980 spin_unlock_irq(&conf->device_lock);
1981 spin_unlock(&sh->lock);
1985 static void end_reshape(raid5_conf_t *conf);
1987 static int page_is_zero(struct page *p)
1989 char *a = page_address(p);
1990 return ((*(u32*)a) == 0 &&
1991 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1994 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
1995 struct stripe_head *sh)
1997 int sectors_per_chunk = conf->chunk_size >> 9;
1999 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2000 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2002 raid5_compute_sector(conf,
2003 stripe * (disks - conf->max_degraded)
2004 *sectors_per_chunk + chunk_offset,
2010 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2011 struct stripe_head_state *s, int disks,
2012 struct bio **return_bi)
2015 for (i = disks; i--; ) {
2019 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2022 rdev = rcu_dereference(conf->disks[i].rdev);
2023 if (rdev && test_bit(In_sync, &rdev->flags))
2024 /* multiple read failures in one stripe */
2025 md_error(conf->mddev, rdev);
2028 spin_lock_irq(&conf->device_lock);
2029 /* fail all writes first */
2030 bi = sh->dev[i].towrite;
2031 sh->dev[i].towrite = NULL;
2037 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2038 wake_up(&conf->wait_for_overlap);
2040 while (bi && bi->bi_sector <
2041 sh->dev[i].sector + STRIPE_SECTORS) {
2042 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2043 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2044 if (!raid5_dec_bi_phys_segments(bi)) {
2045 md_write_end(conf->mddev);
2046 bi->bi_next = *return_bi;
2051 /* and fail all 'written' */
2052 bi = sh->dev[i].written;
2053 sh->dev[i].written = NULL;
2054 if (bi) bitmap_end = 1;
2055 while (bi && bi->bi_sector <
2056 sh->dev[i].sector + STRIPE_SECTORS) {
2057 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2058 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2059 if (!raid5_dec_bi_phys_segments(bi)) {
2060 md_write_end(conf->mddev);
2061 bi->bi_next = *return_bi;
2067 /* fail any reads if this device is non-operational and
2068 * the data has not reached the cache yet.
2070 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2071 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2072 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2073 bi = sh->dev[i].toread;
2074 sh->dev[i].toread = NULL;
2075 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2076 wake_up(&conf->wait_for_overlap);
2077 if (bi) s->to_read--;
2078 while (bi && bi->bi_sector <
2079 sh->dev[i].sector + STRIPE_SECTORS) {
2080 struct bio *nextbi =
2081 r5_next_bio(bi, sh->dev[i].sector);
2082 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2083 if (!raid5_dec_bi_phys_segments(bi)) {
2084 bi->bi_next = *return_bi;
2090 spin_unlock_irq(&conf->device_lock);
2092 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2093 STRIPE_SECTORS, 0, 0);
2096 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2097 if (atomic_dec_and_test(&conf->pending_full_writes))
2098 md_wakeup_thread(conf->mddev->thread);
2101 /* fetch_block5 - checks the given member device to see if its data needs
2102 * to be read or computed to satisfy a request.
2104 * Returns 1 when no more member devices need to be checked, otherwise returns
2105 * 0 to tell the loop in handle_stripe_fill5 to continue
2107 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2108 int disk_idx, int disks)
2110 struct r5dev *dev = &sh->dev[disk_idx];
2111 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2113 /* is the data in this block needed, and can we get it? */
2114 if (!test_bit(R5_LOCKED, &dev->flags) &&
2115 !test_bit(R5_UPTODATE, &dev->flags) &&
2117 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2118 s->syncing || s->expanding ||
2120 (failed_dev->toread ||
2121 (failed_dev->towrite &&
2122 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2123 /* We would like to get this block, possibly by computing it,
2124 * otherwise read it if the backing disk is insync
2126 if ((s->uptodate == disks - 1) &&
2127 (s->failed && disk_idx == s->failed_num)) {
2128 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2129 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2130 set_bit(R5_Wantcompute, &dev->flags);
2131 sh->ops.target = disk_idx;
2133 /* Careful: from this point on 'uptodate' is in the eye
2134 * of raid5_run_ops which services 'compute' operations
2135 * before writes. R5_Wantcompute flags a block that will
2136 * be R5_UPTODATE by the time it is needed for a
2137 * subsequent operation.
2140 return 1; /* uptodate + compute == disks */
2141 } else if (test_bit(R5_Insync, &dev->flags)) {
2142 set_bit(R5_LOCKED, &dev->flags);
2143 set_bit(R5_Wantread, &dev->flags);
2145 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2154 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2156 static void handle_stripe_fill5(struct stripe_head *sh,
2157 struct stripe_head_state *s, int disks)
2161 /* look for blocks to read/compute, skip this if a compute
2162 * is already in flight, or if the stripe contents are in the
2163 * midst of changing due to a write
2165 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2166 !sh->reconstruct_state)
2167 for (i = disks; i--; )
2168 if (fetch_block5(sh, s, i, disks))
2170 set_bit(STRIPE_HANDLE, &sh->state);
2173 static void handle_stripe_fill6(struct stripe_head *sh,
2174 struct stripe_head_state *s, struct r6_state *r6s,
2178 for (i = disks; i--; ) {
2179 struct r5dev *dev = &sh->dev[i];
2180 if (!test_bit(R5_LOCKED, &dev->flags) &&
2181 !test_bit(R5_UPTODATE, &dev->flags) &&
2182 (dev->toread || (dev->towrite &&
2183 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2184 s->syncing || s->expanding ||
2186 (sh->dev[r6s->failed_num[0]].toread ||
2189 (sh->dev[r6s->failed_num[1]].toread ||
2191 /* we would like to get this block, possibly
2192 * by computing it, but we might not be able to
2194 if ((s->uptodate == disks - 1) &&
2195 (s->failed && (i == r6s->failed_num[0] ||
2196 i == r6s->failed_num[1]))) {
2197 pr_debug("Computing stripe %llu block %d\n",
2198 (unsigned long long)sh->sector, i);
2199 compute_block_1(sh, i, 0);
2201 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2202 /* Computing 2-failure is *very* expensive; only
2203 * do it if failed >= 2
2206 for (other = disks; other--; ) {
2209 if (!test_bit(R5_UPTODATE,
2210 &sh->dev[other].flags))
2214 pr_debug("Computing stripe %llu blocks %d,%d\n",
2215 (unsigned long long)sh->sector,
2217 compute_block_2(sh, i, other);
2219 } else if (test_bit(R5_Insync, &dev->flags)) {
2220 set_bit(R5_LOCKED, &dev->flags);
2221 set_bit(R5_Wantread, &dev->flags);
2223 pr_debug("Reading block %d (sync=%d)\n",
2228 set_bit(STRIPE_HANDLE, &sh->state);
2232 /* handle_stripe_clean_event
2233 * any written block on an uptodate or failed drive can be returned.
2234 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2235 * never LOCKED, so we don't need to test 'failed' directly.
2237 static void handle_stripe_clean_event(raid5_conf_t *conf,
2238 struct stripe_head *sh, int disks, struct bio **return_bi)
2243 for (i = disks; i--; )
2244 if (sh->dev[i].written) {
2246 if (!test_bit(R5_LOCKED, &dev->flags) &&
2247 test_bit(R5_UPTODATE, &dev->flags)) {
2248 /* We can return any write requests */
2249 struct bio *wbi, *wbi2;
2251 pr_debug("Return write for disc %d\n", i);
2252 spin_lock_irq(&conf->device_lock);
2254 dev->written = NULL;
2255 while (wbi && wbi->bi_sector <
2256 dev->sector + STRIPE_SECTORS) {
2257 wbi2 = r5_next_bio(wbi, dev->sector);
2258 if (!raid5_dec_bi_phys_segments(wbi)) {
2259 md_write_end(conf->mddev);
2260 wbi->bi_next = *return_bi;
2265 if (dev->towrite == NULL)
2267 spin_unlock_irq(&conf->device_lock);
2269 bitmap_endwrite(conf->mddev->bitmap,
2272 !test_bit(STRIPE_DEGRADED, &sh->state),
2277 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2278 if (atomic_dec_and_test(&conf->pending_full_writes))
2279 md_wakeup_thread(conf->mddev->thread);
2282 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2283 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2285 int rmw = 0, rcw = 0, i;
2286 for (i = disks; i--; ) {
2287 /* would I have to read this buffer for read_modify_write */
2288 struct r5dev *dev = &sh->dev[i];
2289 if ((dev->towrite || i == sh->pd_idx) &&
2290 !test_bit(R5_LOCKED, &dev->flags) &&
2291 !(test_bit(R5_UPTODATE, &dev->flags) ||
2292 test_bit(R5_Wantcompute, &dev->flags))) {
2293 if (test_bit(R5_Insync, &dev->flags))
2296 rmw += 2*disks; /* cannot read it */
2298 /* Would I have to read this buffer for reconstruct_write */
2299 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2300 !test_bit(R5_LOCKED, &dev->flags) &&
2301 !(test_bit(R5_UPTODATE, &dev->flags) ||
2302 test_bit(R5_Wantcompute, &dev->flags))) {
2303 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2308 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2309 (unsigned long long)sh->sector, rmw, rcw);
2310 set_bit(STRIPE_HANDLE, &sh->state);
2311 if (rmw < rcw && rmw > 0)
2312 /* prefer read-modify-write, but need to get some data */
2313 for (i = disks; i--; ) {
2314 struct r5dev *dev = &sh->dev[i];
2315 if ((dev->towrite || i == sh->pd_idx) &&
2316 !test_bit(R5_LOCKED, &dev->flags) &&
2317 !(test_bit(R5_UPTODATE, &dev->flags) ||
2318 test_bit(R5_Wantcompute, &dev->flags)) &&
2319 test_bit(R5_Insync, &dev->flags)) {
2321 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2322 pr_debug("Read_old block "
2323 "%d for r-m-w\n", i);
2324 set_bit(R5_LOCKED, &dev->flags);
2325 set_bit(R5_Wantread, &dev->flags);
2328 set_bit(STRIPE_DELAYED, &sh->state);
2329 set_bit(STRIPE_HANDLE, &sh->state);
2333 if (rcw <= rmw && rcw > 0)
2334 /* want reconstruct write, but need to get some data */
2335 for (i = disks; i--; ) {
2336 struct r5dev *dev = &sh->dev[i];
2337 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2339 !test_bit(R5_LOCKED, &dev->flags) &&
2340 !(test_bit(R5_UPTODATE, &dev->flags) ||
2341 test_bit(R5_Wantcompute, &dev->flags)) &&
2342 test_bit(R5_Insync, &dev->flags)) {
2344 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2345 pr_debug("Read_old block "
2346 "%d for Reconstruct\n", i);
2347 set_bit(R5_LOCKED, &dev->flags);
2348 set_bit(R5_Wantread, &dev->flags);
2351 set_bit(STRIPE_DELAYED, &sh->state);
2352 set_bit(STRIPE_HANDLE, &sh->state);
2356 /* now if nothing is locked, and if we have enough data,
2357 * we can start a write request
2359 /* since handle_stripe can be called at any time we need to handle the
2360 * case where a compute block operation has been submitted and then a
2361 * subsequent call wants to start a write request. raid5_run_ops only
2362 * handles the case where compute block and postxor are requested
2363 * simultaneously. If this is not the case then new writes need to be
2364 * held off until the compute completes.
2366 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2367 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2368 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2369 schedule_reconstruction5(sh, s, rcw == 0, 0);
2372 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2373 struct stripe_head *sh, struct stripe_head_state *s,
2374 struct r6_state *r6s, int disks)
2376 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2377 int qd_idx = r6s->qd_idx;
2378 for (i = disks; i--; ) {
2379 struct r5dev *dev = &sh->dev[i];
2380 /* Would I have to read this buffer for reconstruct_write */
2381 if (!test_bit(R5_OVERWRITE, &dev->flags)
2382 && i != pd_idx && i != qd_idx
2383 && (!test_bit(R5_LOCKED, &dev->flags)
2385 !test_bit(R5_UPTODATE, &dev->flags)) {
2386 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2388 pr_debug("raid6: must_compute: "
2389 "disk %d flags=%#lx\n", i, dev->flags);
2394 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2395 (unsigned long long)sh->sector, rcw, must_compute);
2396 set_bit(STRIPE_HANDLE, &sh->state);
2399 /* want reconstruct write, but need to get some data */
2400 for (i = disks; i--; ) {
2401 struct r5dev *dev = &sh->dev[i];
2402 if (!test_bit(R5_OVERWRITE, &dev->flags)
2403 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2404 && !test_bit(R5_LOCKED, &dev->flags) &&
2405 !test_bit(R5_UPTODATE, &dev->flags) &&
2406 test_bit(R5_Insync, &dev->flags)) {
2408 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2409 pr_debug("Read_old stripe %llu "
2410 "block %d for Reconstruct\n",
2411 (unsigned long long)sh->sector, i);
2412 set_bit(R5_LOCKED, &dev->flags);
2413 set_bit(R5_Wantread, &dev->flags);
2416 pr_debug("Request delayed stripe %llu "
2417 "block %d for Reconstruct\n",
2418 (unsigned long long)sh->sector, i);
2419 set_bit(STRIPE_DELAYED, &sh->state);
2420 set_bit(STRIPE_HANDLE, &sh->state);
2424 /* now if nothing is locked, and if we have enough data, we can start a
2427 if (s->locked == 0 && rcw == 0 &&
2428 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2429 if (must_compute > 0) {
2430 /* We have failed blocks and need to compute them */
2431 switch (s->failed) {
2435 compute_block_1(sh, r6s->failed_num[0], 0);
2438 compute_block_2(sh, r6s->failed_num[0],
2439 r6s->failed_num[1]);
2441 default: /* This request should have been failed? */
2446 pr_debug("Computing parity for stripe %llu\n",
2447 (unsigned long long)sh->sector);
2448 compute_parity6(sh, RECONSTRUCT_WRITE);
2449 /* now every locked buffer is ready to be written */
2450 for (i = disks; i--; )
2451 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2452 pr_debug("Writing stripe %llu block %d\n",
2453 (unsigned long long)sh->sector, i);
2455 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2457 if (s->locked == disks)
2458 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2459 atomic_inc(&conf->pending_full_writes);
2460 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2461 set_bit(STRIPE_INSYNC, &sh->state);
2463 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2464 atomic_dec(&conf->preread_active_stripes);
2465 if (atomic_read(&conf->preread_active_stripes) <
2467 md_wakeup_thread(conf->mddev->thread);
2472 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2473 struct stripe_head_state *s, int disks)
2475 struct r5dev *dev = NULL;
2477 set_bit(STRIPE_HANDLE, &sh->state);
2479 switch (sh->check_state) {
2480 case check_state_idle:
2481 /* start a new check operation if there are no failures */
2482 if (s->failed == 0) {
2483 BUG_ON(s->uptodate != disks);
2484 sh->check_state = check_state_run;
2485 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2486 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2490 dev = &sh->dev[s->failed_num];
2492 case check_state_compute_result:
2493 sh->check_state = check_state_idle;
2495 dev = &sh->dev[sh->pd_idx];
2497 /* check that a write has not made the stripe insync */
2498 if (test_bit(STRIPE_INSYNC, &sh->state))
2501 /* either failed parity check, or recovery is happening */
2502 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2503 BUG_ON(s->uptodate != disks);
2505 set_bit(R5_LOCKED, &dev->flags);
2507 set_bit(R5_Wantwrite, &dev->flags);
2509 clear_bit(STRIPE_DEGRADED, &sh->state);
2510 set_bit(STRIPE_INSYNC, &sh->state);
2512 case check_state_run:
2513 break; /* we will be called again upon completion */
2514 case check_state_check_result:
2515 sh->check_state = check_state_idle;
2517 /* if a failure occurred during the check operation, leave
2518 * STRIPE_INSYNC not set and let the stripe be handled again
2523 /* handle a successful check operation, if parity is correct
2524 * we are done. Otherwise update the mismatch count and repair
2525 * parity if !MD_RECOVERY_CHECK
2527 if (sh->ops.zero_sum_result == 0)
2528 /* parity is correct (on disc,
2529 * not in buffer any more)
2531 set_bit(STRIPE_INSYNC, &sh->state);
2533 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2534 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2535 /* don't try to repair!! */
2536 set_bit(STRIPE_INSYNC, &sh->state);
2538 sh->check_state = check_state_compute_run;
2539 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2540 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2541 set_bit(R5_Wantcompute,
2542 &sh->dev[sh->pd_idx].flags);
2543 sh->ops.target = sh->pd_idx;
2548 case check_state_compute_run:
2551 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2552 __func__, sh->check_state,
2553 (unsigned long long) sh->sector);
2559 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2560 struct stripe_head_state *s,
2561 struct r6_state *r6s, struct page *tmp_page,
2564 int update_p = 0, update_q = 0;
2566 int pd_idx = sh->pd_idx;
2567 int qd_idx = r6s->qd_idx;
2569 set_bit(STRIPE_HANDLE, &sh->state);
2571 BUG_ON(s->failed > 2);
2572 BUG_ON(s->uptodate < disks);
2573 /* Want to check and possibly repair P and Q.
2574 * However there could be one 'failed' device, in which
2575 * case we can only check one of them, possibly using the
2576 * other to generate missing data
2579 /* If !tmp_page, we cannot do the calculations,
2580 * but as we have set STRIPE_HANDLE, we will soon be called
2581 * by stripe_handle with a tmp_page - just wait until then.
2584 if (s->failed == r6s->q_failed) {
2585 /* The only possible failed device holds 'Q', so it
2586 * makes sense to check P (If anything else were failed,
2587 * we would have used P to recreate it).
2589 compute_block_1(sh, pd_idx, 1);
2590 if (!page_is_zero(sh->dev[pd_idx].page)) {
2591 compute_block_1(sh, pd_idx, 0);
2595 if (!r6s->q_failed && s->failed < 2) {
2596 /* q is not failed, and we didn't use it to generate
2597 * anything, so it makes sense to check it
2599 memcpy(page_address(tmp_page),
2600 page_address(sh->dev[qd_idx].page),
2602 compute_parity6(sh, UPDATE_PARITY);
2603 if (memcmp(page_address(tmp_page),
2604 page_address(sh->dev[qd_idx].page),
2605 STRIPE_SIZE) != 0) {
2606 clear_bit(STRIPE_INSYNC, &sh->state);
2610 if (update_p || update_q) {
2611 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2612 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2613 /* don't try to repair!! */
2614 update_p = update_q = 0;
2617 /* now write out any block on a failed drive,
2618 * or P or Q if they need it
2621 if (s->failed == 2) {
2622 dev = &sh->dev[r6s->failed_num[1]];
2624 set_bit(R5_LOCKED, &dev->flags);
2625 set_bit(R5_Wantwrite, &dev->flags);
2627 if (s->failed >= 1) {
2628 dev = &sh->dev[r6s->failed_num[0]];
2630 set_bit(R5_LOCKED, &dev->flags);
2631 set_bit(R5_Wantwrite, &dev->flags);
2635 dev = &sh->dev[pd_idx];
2637 set_bit(R5_LOCKED, &dev->flags);
2638 set_bit(R5_Wantwrite, &dev->flags);
2641 dev = &sh->dev[qd_idx];
2643 set_bit(R5_LOCKED, &dev->flags);
2644 set_bit(R5_Wantwrite, &dev->flags);
2646 clear_bit(STRIPE_DEGRADED, &sh->state);
2648 set_bit(STRIPE_INSYNC, &sh->state);
2652 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2653 struct r6_state *r6s)
2657 /* We have read all the blocks in this stripe and now we need to
2658 * copy some of them into a target stripe for expand.
2660 struct dma_async_tx_descriptor *tx = NULL;
2661 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2662 for (i = 0; i < sh->disks; i++)
2663 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2665 struct stripe_head *sh2;
2667 sector_t bn = compute_blocknr(sh, i);
2668 sector_t s = raid5_compute_sector(conf, bn, 0,
2670 sh2 = get_active_stripe(conf, s, 0, 1);
2672 /* so far only the early blocks of this stripe
2673 * have been requested. When later blocks
2674 * get requested, we will try again
2677 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2678 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2679 /* must have already done this block */
2680 release_stripe(sh2);
2684 /* place all the copies on one channel */
2685 tx = async_memcpy(sh2->dev[dd_idx].page,
2686 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2687 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2689 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2690 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2691 for (j = 0; j < conf->raid_disks; j++)
2692 if (j != sh2->pd_idx &&
2693 (!r6s || j != sh2->qd_idx) &&
2694 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2696 if (j == conf->raid_disks) {
2697 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2698 set_bit(STRIPE_HANDLE, &sh2->state);
2700 release_stripe(sh2);
2703 /* done submitting copies, wait for them to complete */
2706 dma_wait_for_async_tx(tx);
2712 * handle_stripe - do things to a stripe.
2714 * We lock the stripe and then examine the state of various bits
2715 * to see what needs to be done.
2717 * return some read request which now have data
2718 * return some write requests which are safely on disc
2719 * schedule a read on some buffers
2720 * schedule a write of some buffers
2721 * return confirmation of parity correctness
2723 * buffers are taken off read_list or write_list, and bh_cache buffers
2724 * get BH_Lock set before the stripe lock is released.
2728 static bool handle_stripe5(struct stripe_head *sh)
2730 raid5_conf_t *conf = sh->raid_conf;
2731 int disks = sh->disks, i;
2732 struct bio *return_bi = NULL;
2733 struct stripe_head_state s;
2735 mdk_rdev_t *blocked_rdev = NULL;
2738 memset(&s, 0, sizeof(s));
2739 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2740 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2741 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2742 sh->reconstruct_state);
2744 spin_lock(&sh->lock);
2745 clear_bit(STRIPE_HANDLE, &sh->state);
2746 clear_bit(STRIPE_DELAYED, &sh->state);
2748 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2749 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2750 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2752 /* Now to look around and see what can be done */
2754 for (i=disks; i--; ) {
2756 struct r5dev *dev = &sh->dev[i];
2757 clear_bit(R5_Insync, &dev->flags);
2759 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2760 "written %p\n", i, dev->flags, dev->toread, dev->read,
2761 dev->towrite, dev->written);
2763 /* maybe we can request a biofill operation
2765 * new wantfill requests are only permitted while
2766 * ops_complete_biofill is guaranteed to be inactive
2768 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2769 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2770 set_bit(R5_Wantfill, &dev->flags);
2772 /* now count some things */
2773 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2774 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2775 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2777 if (test_bit(R5_Wantfill, &dev->flags))
2779 else if (dev->toread)
2783 if (!test_bit(R5_OVERWRITE, &dev->flags))
2788 rdev = rcu_dereference(conf->disks[i].rdev);
2789 if (blocked_rdev == NULL &&
2790 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2791 blocked_rdev = rdev;
2792 atomic_inc(&rdev->nr_pending);
2794 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2795 /* The ReadError flag will just be confusing now */
2796 clear_bit(R5_ReadError, &dev->flags);
2797 clear_bit(R5_ReWrite, &dev->flags);
2799 if (!rdev || !test_bit(In_sync, &rdev->flags)
2800 || test_bit(R5_ReadError, &dev->flags)) {
2804 set_bit(R5_Insync, &dev->flags);
2808 if (unlikely(blocked_rdev)) {
2809 if (s.syncing || s.expanding || s.expanded ||
2810 s.to_write || s.written) {
2811 set_bit(STRIPE_HANDLE, &sh->state);
2814 /* There is nothing for the blocked_rdev to block */
2815 rdev_dec_pending(blocked_rdev, conf->mddev);
2816 blocked_rdev = NULL;
2819 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2820 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2821 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2824 pr_debug("locked=%d uptodate=%d to_read=%d"
2825 " to_write=%d failed=%d failed_num=%d\n",
2826 s.locked, s.uptodate, s.to_read, s.to_write,
2827 s.failed, s.failed_num);
2828 /* check if the array has lost two devices and, if so, some requests might
2831 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2832 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2833 if (s.failed > 1 && s.syncing) {
2834 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2835 clear_bit(STRIPE_SYNCING, &sh->state);
2839 /* might be able to return some write requests if the parity block
2840 * is safe, or on a failed drive
2842 dev = &sh->dev[sh->pd_idx];
2844 ((test_bit(R5_Insync, &dev->flags) &&
2845 !test_bit(R5_LOCKED, &dev->flags) &&
2846 test_bit(R5_UPTODATE, &dev->flags)) ||
2847 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2848 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2850 /* Now we might consider reading some blocks, either to check/generate
2851 * parity, or to satisfy requests
2852 * or to load a block that is being partially written.
2854 if (s.to_read || s.non_overwrite ||
2855 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2856 handle_stripe_fill5(sh, &s, disks);
2858 /* Now we check to see if any write operations have recently
2862 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2864 if (sh->reconstruct_state == reconstruct_state_drain_result ||
2865 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2866 sh->reconstruct_state = reconstruct_state_idle;
2868 /* All the 'written' buffers and the parity block are ready to
2869 * be written back to disk
2871 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2872 for (i = disks; i--; ) {
2874 if (test_bit(R5_LOCKED, &dev->flags) &&
2875 (i == sh->pd_idx || dev->written)) {
2876 pr_debug("Writing block %d\n", i);
2877 set_bit(R5_Wantwrite, &dev->flags);
2880 if (!test_bit(R5_Insync, &dev->flags) ||
2881 (i == sh->pd_idx && s.failed == 0))
2882 set_bit(STRIPE_INSYNC, &sh->state);
2885 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2886 atomic_dec(&conf->preread_active_stripes);
2887 if (atomic_read(&conf->preread_active_stripes) <
2889 md_wakeup_thread(conf->mddev->thread);
2893 /* Now to consider new write requests and what else, if anything
2894 * should be read. We do not handle new writes when:
2895 * 1/ A 'write' operation (copy+xor) is already in flight.
2896 * 2/ A 'check' operation is in flight, as it may clobber the parity
2899 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2900 handle_stripe_dirtying5(conf, sh, &s, disks);
2902 /* maybe we need to check and possibly fix the parity for this stripe
2903 * Any reads will already have been scheduled, so we just see if enough
2904 * data is available. The parity check is held off while parity
2905 * dependent operations are in flight.
2907 if (sh->check_state ||
2908 (s.syncing && s.locked == 0 &&
2909 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2910 !test_bit(STRIPE_INSYNC, &sh->state)))
2911 handle_parity_checks5(conf, sh, &s, disks);
2913 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2914 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2915 clear_bit(STRIPE_SYNCING, &sh->state);
2918 /* If the failed drive is just a ReadError, then we might need to progress
2919 * the repair/check process
2921 if (s.failed == 1 && !conf->mddev->ro &&
2922 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2923 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2924 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2926 dev = &sh->dev[s.failed_num];
2927 if (!test_bit(R5_ReWrite, &dev->flags)) {
2928 set_bit(R5_Wantwrite, &dev->flags);
2929 set_bit(R5_ReWrite, &dev->flags);
2930 set_bit(R5_LOCKED, &dev->flags);
2933 /* let's read it back */
2934 set_bit(R5_Wantread, &dev->flags);
2935 set_bit(R5_LOCKED, &dev->flags);
2940 /* Finish reconstruct operations initiated by the expansion process */
2941 if (sh->reconstruct_state == reconstruct_state_result) {
2942 sh->reconstruct_state = reconstruct_state_idle;
2943 clear_bit(STRIPE_EXPANDING, &sh->state);
2944 for (i = conf->raid_disks; i--; ) {
2945 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2946 set_bit(R5_LOCKED, &sh->dev[i].flags);
2951 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2952 !sh->reconstruct_state) {
2953 /* Need to write out all blocks after computing parity */
2954 sh->disks = conf->raid_disks;
2955 stripe_set_idx(sh->sector, conf, 0, sh);
2956 schedule_reconstruction5(sh, &s, 1, 1);
2957 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
2958 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2959 atomic_dec(&conf->reshape_stripes);
2960 wake_up(&conf->wait_for_overlap);
2961 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2964 if (s.expanding && s.locked == 0 &&
2965 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2966 handle_stripe_expansion(conf, sh, NULL);
2969 spin_unlock(&sh->lock);
2971 /* wait for this device to become unblocked */
2972 if (unlikely(blocked_rdev))
2973 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2976 raid5_run_ops(sh, s.ops_request);
2980 return_io(return_bi);
2982 return blocked_rdev == NULL;
2985 static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2987 raid5_conf_t *conf = sh->raid_conf;
2988 int disks = sh->disks;
2989 struct bio *return_bi = NULL;
2990 int i, pd_idx = sh->pd_idx;
2991 struct stripe_head_state s;
2992 struct r6_state r6s;
2993 struct r5dev *dev, *pdev, *qdev;
2994 mdk_rdev_t *blocked_rdev = NULL;
2996 r6s.qd_idx = sh->qd_idx;
2997 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2998 "pd_idx=%d, qd_idx=%d\n",
2999 (unsigned long long)sh->sector, sh->state,
3000 atomic_read(&sh->count), pd_idx, r6s.qd_idx);
3001 memset(&s, 0, sizeof(s));
3003 spin_lock(&sh->lock);
3004 clear_bit(STRIPE_HANDLE, &sh->state);
3005 clear_bit(STRIPE_DELAYED, &sh->state);
3007 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3008 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3009 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3010 /* Now to look around and see what can be done */
3013 for (i=disks; i--; ) {
3016 clear_bit(R5_Insync, &dev->flags);
3018 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3019 i, dev->flags, dev->toread, dev->towrite, dev->written);
3020 /* maybe we can reply to a read */
3021 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3022 struct bio *rbi, *rbi2;
3023 pr_debug("Return read for disc %d\n", i);
3024 spin_lock_irq(&conf->device_lock);
3027 if (test_and_clear_bit(R5_Overlap, &dev->flags))
3028 wake_up(&conf->wait_for_overlap);
3029 spin_unlock_irq(&conf->device_lock);
3030 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3031 copy_data(0, rbi, dev->page, dev->sector);
3032 rbi2 = r5_next_bio(rbi, dev->sector);
3033 spin_lock_irq(&conf->device_lock);
3034 if (!raid5_dec_bi_phys_segments(rbi)) {
3035 rbi->bi_next = return_bi;
3038 spin_unlock_irq(&conf->device_lock);
3043 /* now count some things */
3044 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3045 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3052 if (!test_bit(R5_OVERWRITE, &dev->flags))
3057 rdev = rcu_dereference(conf->disks[i].rdev);
3058 if (blocked_rdev == NULL &&
3059 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3060 blocked_rdev = rdev;
3061 atomic_inc(&rdev->nr_pending);
3063 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3064 /* The ReadError flag will just be confusing now */
3065 clear_bit(R5_ReadError, &dev->flags);
3066 clear_bit(R5_ReWrite, &dev->flags);
3068 if (!rdev || !test_bit(In_sync, &rdev->flags)
3069 || test_bit(R5_ReadError, &dev->flags)) {
3071 r6s.failed_num[s.failed] = i;
3074 set_bit(R5_Insync, &dev->flags);
3078 if (unlikely(blocked_rdev)) {
3079 if (s.syncing || s.expanding || s.expanded ||
3080 s.to_write || s.written) {
3081 set_bit(STRIPE_HANDLE, &sh->state);
3084 /* There is nothing for the blocked_rdev to block */
3085 rdev_dec_pending(blocked_rdev, conf->mddev);
3086 blocked_rdev = NULL;
3089 pr_debug("locked=%d uptodate=%d to_read=%d"
3090 " to_write=%d failed=%d failed_num=%d,%d\n",
3091 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3092 r6s.failed_num[0], r6s.failed_num[1]);
3093 /* check if the array has lost >2 devices and, if so, some requests
3094 * might need to be failed
3096 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3097 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3098 if (s.failed > 2 && s.syncing) {
3099 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3100 clear_bit(STRIPE_SYNCING, &sh->state);
3105 * might be able to return some write requests if the parity blocks
3106 * are safe, or on a failed drive
3108 pdev = &sh->dev[pd_idx];
3109 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3110 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3111 qdev = &sh->dev[r6s.qd_idx];
3112 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3113 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3116 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3117 && !test_bit(R5_LOCKED, &pdev->flags)
3118 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3119 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3120 && !test_bit(R5_LOCKED, &qdev->flags)
3121 && test_bit(R5_UPTODATE, &qdev->flags)))))
3122 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3124 /* Now we might consider reading some blocks, either to check/generate
3125 * parity, or to satisfy requests
3126 * or to load a block that is being partially written.
3128 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3129 (s.syncing && (s.uptodate < disks)) || s.expanding)
3130 handle_stripe_fill6(sh, &s, &r6s, disks);
3132 /* now to consider writing and what else, if anything should be read */
3134 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3136 /* maybe we need to check and possibly fix the parity for this stripe
3137 * Any reads will already have been scheduled, so we just see if enough
3140 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3141 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3143 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3144 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3145 clear_bit(STRIPE_SYNCING, &sh->state);
3148 /* If the failed drives are just a ReadError, then we might need
3149 * to progress the repair/check process
3151 if (s.failed <= 2 && !conf->mddev->ro)
3152 for (i = 0; i < s.failed; i++) {
3153 dev = &sh->dev[r6s.failed_num[i]];
3154 if (test_bit(R5_ReadError, &dev->flags)
3155 && !test_bit(R5_LOCKED, &dev->flags)
3156 && test_bit(R5_UPTODATE, &dev->flags)
3158 if (!test_bit(R5_ReWrite, &dev->flags)) {
3159 set_bit(R5_Wantwrite, &dev->flags);
3160 set_bit(R5_ReWrite, &dev->flags);
3161 set_bit(R5_LOCKED, &dev->flags);
3163 /* let's read it back */
3164 set_bit(R5_Wantread, &dev->flags);
3165 set_bit(R5_LOCKED, &dev->flags);
3170 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3171 /* Need to write out all blocks after computing P&Q */
3172 sh->disks = conf->raid_disks;
3173 stripe_set_idx(sh->sector, conf, 0, sh);
3174 compute_parity6(sh, RECONSTRUCT_WRITE);
3175 for (i = conf->raid_disks ; i-- ; ) {
3176 set_bit(R5_LOCKED, &sh->dev[i].flags);
3178 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3180 clear_bit(STRIPE_EXPANDING, &sh->state);
3181 } else if (s.expanded) {
3182 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3183 atomic_dec(&conf->reshape_stripes);
3184 wake_up(&conf->wait_for_overlap);
3185 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3188 if (s.expanding && s.locked == 0 &&
3189 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3190 handle_stripe_expansion(conf, sh, &r6s);
3193 spin_unlock(&sh->lock);
3195 /* wait for this device to become unblocked */
3196 if (unlikely(blocked_rdev))
3197 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3201 return_io(return_bi);
3203 return blocked_rdev == NULL;
3206 /* returns true if the stripe was handled */
3207 static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3209 if (sh->raid_conf->level == 6)
3210 return handle_stripe6(sh, tmp_page);
3212 return handle_stripe5(sh);
3217 static void raid5_activate_delayed(raid5_conf_t *conf)
3219 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3220 while (!list_empty(&conf->delayed_list)) {
3221 struct list_head *l = conf->delayed_list.next;
3222 struct stripe_head *sh;
3223 sh = list_entry(l, struct stripe_head, lru);
3225 clear_bit(STRIPE_DELAYED, &sh->state);
3226 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3227 atomic_inc(&conf->preread_active_stripes);
3228 list_add_tail(&sh->lru, &conf->hold_list);
3231 blk_plug_device(conf->mddev->queue);
3234 static void activate_bit_delay(raid5_conf_t *conf)
3236 /* device_lock is held */
3237 struct list_head head;
3238 list_add(&head, &conf->bitmap_list);
3239 list_del_init(&conf->bitmap_list);
3240 while (!list_empty(&head)) {
3241 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3242 list_del_init(&sh->lru);
3243 atomic_inc(&sh->count);
3244 __release_stripe(conf, sh);
3248 static void unplug_slaves(mddev_t *mddev)
3250 raid5_conf_t *conf = mddev_to_conf(mddev);
3254 for (i=0; i<mddev->raid_disks; i++) {
3255 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3256 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3257 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3259 atomic_inc(&rdev->nr_pending);
3262 blk_unplug(r_queue);
3264 rdev_dec_pending(rdev, mddev);
3271 static void raid5_unplug_device(struct request_queue *q)
3273 mddev_t *mddev = q->queuedata;
3274 raid5_conf_t *conf = mddev_to_conf(mddev);
3275 unsigned long flags;
3277 spin_lock_irqsave(&conf->device_lock, flags);
3279 if (blk_remove_plug(q)) {
3281 raid5_activate_delayed(conf);
3283 md_wakeup_thread(mddev->thread);
3285 spin_unlock_irqrestore(&conf->device_lock, flags);
3287 unplug_slaves(mddev);
3290 static int raid5_congested(void *data, int bits)
3292 mddev_t *mddev = data;
3293 raid5_conf_t *conf = mddev_to_conf(mddev);
3295 /* No difference between reads and writes. Just check
3296 * how busy the stripe_cache is
3298 if (conf->inactive_blocked)
3302 if (list_empty_careful(&conf->inactive_list))
3308 /* We want read requests to align with chunks where possible,
3309 * but write requests don't need to.
3311 static int raid5_mergeable_bvec(struct request_queue *q,
3312 struct bvec_merge_data *bvm,
3313 struct bio_vec *biovec)
3315 mddev_t *mddev = q->queuedata;
3316 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3318 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3319 unsigned int bio_sectors = bvm->bi_size >> 9;
3321 if ((bvm->bi_rw & 1) == WRITE)
3322 return biovec->bv_len; /* always allow writes to be mergeable */
3324 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3325 if (max < 0) max = 0;
3326 if (max <= biovec->bv_len && bio_sectors == 0)
3327 return biovec->bv_len;
3333 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3335 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3336 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3337 unsigned int bio_sectors = bio->bi_size >> 9;
3339 return chunk_sectors >=
3340 ((sector & (chunk_sectors - 1)) + bio_sectors);
3344 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3345 * later sampled by raid5d.
3347 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3349 unsigned long flags;
3351 spin_lock_irqsave(&conf->device_lock, flags);
3353 bi->bi_next = conf->retry_read_aligned_list;
3354 conf->retry_read_aligned_list = bi;
3356 spin_unlock_irqrestore(&conf->device_lock, flags);
3357 md_wakeup_thread(conf->mddev->thread);
3361 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3365 bi = conf->retry_read_aligned;
3367 conf->retry_read_aligned = NULL;
3370 bi = conf->retry_read_aligned_list;
3372 conf->retry_read_aligned_list = bi->bi_next;
3375 * this sets the active strip count to 1 and the processed
3376 * strip count to zero (upper 8 bits)
3378 bi->bi_phys_segments = 1; /* biased count of active stripes */
3386 * The "raid5_align_endio" should check if the read succeeded and if it
3387 * did, call bio_endio on the original bio (having bio_put the new bio
3389 * If the read failed..
3391 static void raid5_align_endio(struct bio *bi, int error)
3393 struct bio* raid_bi = bi->bi_private;
3396 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3401 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3402 conf = mddev_to_conf(mddev);
3403 rdev = (void*)raid_bi->bi_next;
3404 raid_bi->bi_next = NULL;
3406 rdev_dec_pending(rdev, conf->mddev);
3408 if (!error && uptodate) {
3409 bio_endio(raid_bi, 0);
3410 if (atomic_dec_and_test(&conf->active_aligned_reads))
3411 wake_up(&conf->wait_for_stripe);
3416 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3418 add_bio_to_retry(raid_bi, conf);
3421 static int bio_fits_rdev(struct bio *bi)
3423 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3425 if ((bi->bi_size>>9) > q->max_sectors)
3427 blk_recount_segments(q, bi);
3428 if (bi->bi_phys_segments > q->max_phys_segments)
3431 if (q->merge_bvec_fn)
3432 /* it's too hard to apply the merge_bvec_fn at this stage,
3441 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3443 mddev_t *mddev = q->queuedata;
3444 raid5_conf_t *conf = mddev_to_conf(mddev);
3445 unsigned int dd_idx;
3446 struct bio* align_bi;
3449 if (!in_chunk_boundary(mddev, raid_bio)) {
3450 pr_debug("chunk_aligned_read : non aligned\n");
3454 * use bio_clone to make a copy of the bio
3456 align_bi = bio_clone(raid_bio, GFP_NOIO);
3460 * set bi_end_io to a new function, and set bi_private to the
3463 align_bi->bi_end_io = raid5_align_endio;
3464 align_bi->bi_private = raid_bio;
3468 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3473 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3474 if (rdev && test_bit(In_sync, &rdev->flags)) {
3475 atomic_inc(&rdev->nr_pending);
3477 raid_bio->bi_next = (void*)rdev;
3478 align_bi->bi_bdev = rdev->bdev;
3479 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3480 align_bi->bi_sector += rdev->data_offset;
3482 if (!bio_fits_rdev(align_bi)) {
3483 /* too big in some way */
3485 rdev_dec_pending(rdev, mddev);
3489 spin_lock_irq(&conf->device_lock);
3490 wait_event_lock_irq(conf->wait_for_stripe,
3492 conf->device_lock, /* nothing */);
3493 atomic_inc(&conf->active_aligned_reads);
3494 spin_unlock_irq(&conf->device_lock);
3496 generic_make_request(align_bi);
3505 /* __get_priority_stripe - get the next stripe to process
3507 * Full stripe writes are allowed to pass preread active stripes up until
3508 * the bypass_threshold is exceeded. In general the bypass_count
3509 * increments when the handle_list is handled before the hold_list; however, it
3510 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3511 * stripe with in flight i/o. The bypass_count will be reset when the
3512 * head of the hold_list has changed, i.e. the head was promoted to the
3515 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3517 struct stripe_head *sh;
3519 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3521 list_empty(&conf->handle_list) ? "empty" : "busy",
3522 list_empty(&conf->hold_list) ? "empty" : "busy",
3523 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3525 if (!list_empty(&conf->handle_list)) {
3526 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3528 if (list_empty(&conf->hold_list))
3529 conf->bypass_count = 0;
3530 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3531 if (conf->hold_list.next == conf->last_hold)
3532 conf->bypass_count++;
3534 conf->last_hold = conf->hold_list.next;
3535 conf->bypass_count -= conf->bypass_threshold;
3536 if (conf->bypass_count < 0)
3537 conf->bypass_count = 0;
3540 } else if (!list_empty(&conf->hold_list) &&
3541 ((conf->bypass_threshold &&
3542 conf->bypass_count > conf->bypass_threshold) ||
3543 atomic_read(&conf->pending_full_writes) == 0)) {
3544 sh = list_entry(conf->hold_list.next,
3546 conf->bypass_count -= conf->bypass_threshold;
3547 if (conf->bypass_count < 0)
3548 conf->bypass_count = 0;
3552 list_del_init(&sh->lru);
3553 atomic_inc(&sh->count);
3554 BUG_ON(atomic_read(&sh->count) != 1);
3558 static int make_request(struct request_queue *q, struct bio * bi)
3560 mddev_t *mddev = q->queuedata;
3561 raid5_conf_t *conf = mddev_to_conf(mddev);
3563 sector_t new_sector;
3564 sector_t logical_sector, last_sector;
3565 struct stripe_head *sh;
3566 const int rw = bio_data_dir(bi);
3569 if (unlikely(bio_barrier(bi))) {
3570 bio_endio(bi, -EOPNOTSUPP);
3574 md_write_start(mddev, bi);
3576 cpu = part_stat_lock();
3577 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3578 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3583 mddev->reshape_position == MaxSector &&
3584 chunk_aligned_read(q,bi))
3587 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3588 last_sector = bi->bi_sector + (bi->bi_size>>9);
3590 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3592 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3594 int disks, data_disks;
3599 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3600 if (likely(conf->expand_progress == MaxSector))
3601 disks = conf->raid_disks;
3603 /* spinlock is needed as expand_progress may be
3604 * 64bit on a 32bit platform, and so it might be
3605 * possible to see a half-updated value
3606 * Ofcourse expand_progress could change after
3607 * the lock is dropped, so once we get a reference
3608 * to the stripe that we think it is, we will have
3611 spin_lock_irq(&conf->device_lock);
3612 disks = conf->raid_disks;
3613 if (logical_sector >= conf->expand_progress) {
3614 disks = conf->previous_raid_disks;
3617 if (logical_sector >= conf->expand_lo) {
3618 spin_unlock_irq(&conf->device_lock);
3623 spin_unlock_irq(&conf->device_lock);
3625 data_disks = disks - conf->max_degraded;
3627 new_sector = raid5_compute_sector(conf, logical_sector,
3630 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3631 (unsigned long long)new_sector,
3632 (unsigned long long)logical_sector);
3634 sh = get_active_stripe(conf, new_sector, previous,
3635 (bi->bi_rw&RWA_MASK));
3637 if (unlikely(conf->expand_progress != MaxSector)) {
3638 /* expansion might have moved on while waiting for a
3639 * stripe, so we must do the range check again.
3640 * Expansion could still move past after this
3641 * test, but as we are holding a reference to
3642 * 'sh', we know that if that happens,
3643 * STRIPE_EXPANDING will get set and the expansion
3644 * won't proceed until we finish with the stripe.
3647 spin_lock_irq(&conf->device_lock);
3648 if (logical_sector < conf->expand_progress &&
3649 disks == conf->previous_raid_disks)
3650 /* mismatch, need to try again */
3652 spin_unlock_irq(&conf->device_lock);
3658 /* FIXME what if we get a false positive because these
3659 * are being updated.
3661 if (logical_sector >= mddev->suspend_lo &&
3662 logical_sector < mddev->suspend_hi) {
3668 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3669 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3670 /* Stripe is busy expanding or
3671 * add failed due to overlap. Flush everything
3674 raid5_unplug_device(mddev->queue);
3679 finish_wait(&conf->wait_for_overlap, &w);
3680 set_bit(STRIPE_HANDLE, &sh->state);
3681 clear_bit(STRIPE_DELAYED, &sh->state);
3684 /* cannot get stripe for read-ahead, just give-up */
3685 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3686 finish_wait(&conf->wait_for_overlap, &w);
3691 spin_lock_irq(&conf->device_lock);
3692 remaining = raid5_dec_bi_phys_segments(bi);
3693 spin_unlock_irq(&conf->device_lock);
3694 if (remaining == 0) {
3697 md_write_end(mddev);
3704 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3706 /* reshaping is quite different to recovery/resync so it is
3707 * handled quite separately ... here.
3709 * On each call to sync_request, we gather one chunk worth of
3710 * destination stripes and flag them as expanding.
3711 * Then we find all the source stripes and request reads.
3712 * As the reads complete, handle_stripe will copy the data
3713 * into the destination stripe and release that stripe.
3715 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3716 struct stripe_head *sh;
3717 sector_t first_sector, last_sector;
3718 int raid_disks = conf->previous_raid_disks;
3719 int data_disks = raid_disks - conf->max_degraded;
3720 int new_data_disks = conf->raid_disks - conf->max_degraded;
3723 sector_t writepos, safepos, gap;
3725 if (sector_nr == 0 &&
3726 conf->expand_progress != 0) {
3727 /* restarting in the middle, skip the initial sectors */
3728 sector_nr = conf->expand_progress;
3729 sector_div(sector_nr, new_data_disks);
3734 /* we update the metadata when there is more than 3Meg
3735 * in the block range (that is rather arbitrary, should
3736 * probably be time based) or when the data about to be
3737 * copied would over-write the source of the data at
3738 * the front of the range.
3739 * i.e. one new_stripe forward from expand_progress new_maps
3740 * to after where expand_lo old_maps to
3742 writepos = conf->expand_progress +
3743 conf->chunk_size/512*(new_data_disks);
3744 sector_div(writepos, new_data_disks);
3745 safepos = conf->expand_lo;
3746 sector_div(safepos, data_disks);
3747 gap = conf->expand_progress - conf->expand_lo;
3749 if (writepos >= safepos ||
3750 gap > (new_data_disks)*3000*2 /*3Meg*/) {
3751 /* Cannot proceed until we've updated the superblock... */
3752 wait_event(conf->wait_for_overlap,
3753 atomic_read(&conf->reshape_stripes)==0);
3754 mddev->reshape_position = conf->expand_progress;
3755 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3756 md_wakeup_thread(mddev->thread);
3757 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3758 kthread_should_stop());
3759 spin_lock_irq(&conf->device_lock);
3760 conf->expand_lo = mddev->reshape_position;
3761 spin_unlock_irq(&conf->device_lock);
3762 wake_up(&conf->wait_for_overlap);
3765 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3768 sh = get_active_stripe(conf, sector_nr+i, 0, 0);
3769 set_bit(STRIPE_EXPANDING, &sh->state);
3770 atomic_inc(&conf->reshape_stripes);
3771 /* If any of this stripe is beyond the end of the old
3772 * array, then we need to zero those blocks
3774 for (j=sh->disks; j--;) {
3776 if (j == sh->pd_idx)
3778 if (conf->level == 6 &&
3781 s = compute_blocknr(sh, j);
3782 if (s < mddev->array_sectors) {
3786 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3787 set_bit(R5_Expanded, &sh->dev[j].flags);
3788 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3791 set_bit(STRIPE_EXPAND_READY, &sh->state);
3792 set_bit(STRIPE_HANDLE, &sh->state);
3796 spin_lock_irq(&conf->device_lock);
3797 conf->expand_progress = (sector_nr + i) * new_data_disks;
3798 spin_unlock_irq(&conf->device_lock);
3799 /* Ok, those stripe are ready. We can start scheduling
3800 * reads on the source stripes.
3801 * The source stripes are determined by mapping the first and last
3802 * block on the destination stripes.
3805 raid5_compute_sector(conf, sector_nr*(new_data_disks),
3808 raid5_compute_sector(conf, ((sector_nr+conf->chunk_size/512)
3809 *(new_data_disks) - 1),
3811 if (last_sector >= mddev->dev_sectors)
3812 last_sector = mddev->dev_sectors - 1;
3813 while (first_sector <= last_sector) {
3814 sh = get_active_stripe(conf, first_sector, 1, 0);
3815 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3816 set_bit(STRIPE_HANDLE, &sh->state);
3818 first_sector += STRIPE_SECTORS;
3820 /* If this takes us to the resync_max point where we have to pause,
3821 * then we need to write out the superblock.
3823 sector_nr += conf->chunk_size>>9;
3824 if (sector_nr >= mddev->resync_max) {
3825 /* Cannot proceed until we've updated the superblock... */
3826 wait_event(conf->wait_for_overlap,
3827 atomic_read(&conf->reshape_stripes) == 0);
3828 mddev->reshape_position = conf->expand_progress;
3829 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3830 md_wakeup_thread(mddev->thread);
3831 wait_event(mddev->sb_wait,
3832 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3833 || kthread_should_stop());
3834 spin_lock_irq(&conf->device_lock);
3835 conf->expand_lo = mddev->reshape_position;
3836 spin_unlock_irq(&conf->device_lock);
3837 wake_up(&conf->wait_for_overlap);
3839 return conf->chunk_size>>9;
3842 /* FIXME go_faster isn't used */
3843 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3845 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3846 struct stripe_head *sh;
3847 sector_t max_sector = mddev->dev_sectors;
3849 int still_degraded = 0;
3852 if (sector_nr >= max_sector) {
3853 /* just being told to finish up .. nothing much to do */
3854 unplug_slaves(mddev);
3855 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3860 if (mddev->curr_resync < max_sector) /* aborted */
3861 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3863 else /* completed sync */
3865 bitmap_close_sync(mddev->bitmap);
3870 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3871 return reshape_request(mddev, sector_nr, skipped);
3873 /* No need to check resync_max as we never do more than one
3874 * stripe, and as resync_max will always be on a chunk boundary,
3875 * if the check in md_do_sync didn't fire, there is no chance
3876 * of overstepping resync_max here
3879 /* if there is too many failed drives and we are trying
3880 * to resync, then assert that we are finished, because there is
3881 * nothing we can do.
3883 if (mddev->degraded >= conf->max_degraded &&
3884 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3885 sector_t rv = mddev->dev_sectors - sector_nr;
3889 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3890 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3891 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3892 /* we can skip this block, and probably more */
3893 sync_blocks /= STRIPE_SECTORS;
3895 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3899 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3901 sh = get_active_stripe(conf, sector_nr, 0, 1);
3903 sh = get_active_stripe(conf, sector_nr, 0, 0);
3904 /* make sure we don't swamp the stripe cache if someone else
3905 * is trying to get access
3907 schedule_timeout_uninterruptible(1);
3909 /* Need to check if array will still be degraded after recovery/resync
3910 * We don't need to check the 'failed' flag as when that gets set,
3913 for (i=0; i<mddev->raid_disks; i++)
3914 if (conf->disks[i].rdev == NULL)
3917 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3919 spin_lock(&sh->lock);
3920 set_bit(STRIPE_SYNCING, &sh->state);
3921 clear_bit(STRIPE_INSYNC, &sh->state);
3922 spin_unlock(&sh->lock);
3924 /* wait for any blocked device to be handled */
3925 while(unlikely(!handle_stripe(sh, NULL)))
3929 return STRIPE_SECTORS;
3932 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3934 /* We may not be able to submit a whole bio at once as there
3935 * may not be enough stripe_heads available.
3936 * We cannot pre-allocate enough stripe_heads as we may need
3937 * more than exist in the cache (if we allow ever large chunks).
3938 * So we do one stripe head at a time and record in
3939 * ->bi_hw_segments how many have been done.
3941 * We *know* that this entire raid_bio is in one chunk, so
3942 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3944 struct stripe_head *sh;
3946 sector_t sector, logical_sector, last_sector;
3951 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3952 sector = raid5_compute_sector(conf, logical_sector,
3954 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3956 for (; logical_sector < last_sector;
3957 logical_sector += STRIPE_SECTORS,
3958 sector += STRIPE_SECTORS,
3961 if (scnt < raid5_bi_hw_segments(raid_bio))
3962 /* already done this stripe */
3965 sh = get_active_stripe(conf, sector, 0, 1);
3968 /* failed to get a stripe - must wait */
3969 raid5_set_bi_hw_segments(raid_bio, scnt);
3970 conf->retry_read_aligned = raid_bio;
3974 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3975 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3977 raid5_set_bi_hw_segments(raid_bio, scnt);
3978 conf->retry_read_aligned = raid_bio;
3982 handle_stripe(sh, NULL);
3986 spin_lock_irq(&conf->device_lock);
3987 remaining = raid5_dec_bi_phys_segments(raid_bio);
3988 spin_unlock_irq(&conf->device_lock);
3990 bio_endio(raid_bio, 0);
3991 if (atomic_dec_and_test(&conf->active_aligned_reads))
3992 wake_up(&conf->wait_for_stripe);
3999 * This is our raid5 kernel thread.
4001 * We scan the hash table for stripes which can be handled now.
4002 * During the scan, completed stripes are saved for us by the interrupt
4003 * handler, so that they will not have to wait for our next wakeup.
4005 static void raid5d(mddev_t *mddev)
4007 struct stripe_head *sh;
4008 raid5_conf_t *conf = mddev_to_conf(mddev);
4011 pr_debug("+++ raid5d active\n");
4013 md_check_recovery(mddev);
4016 spin_lock_irq(&conf->device_lock);
4020 if (conf->seq_flush != conf->seq_write) {
4021 int seq = conf->seq_flush;
4022 spin_unlock_irq(&conf->device_lock);
4023 bitmap_unplug(mddev->bitmap);
4024 spin_lock_irq(&conf->device_lock);
4025 conf->seq_write = seq;
4026 activate_bit_delay(conf);
4029 while ((bio = remove_bio_from_retry(conf))) {
4031 spin_unlock_irq(&conf->device_lock);
4032 ok = retry_aligned_read(conf, bio);
4033 spin_lock_irq(&conf->device_lock);
4039 sh = __get_priority_stripe(conf);
4043 spin_unlock_irq(&conf->device_lock);
4046 handle_stripe(sh, conf->spare_page);
4049 spin_lock_irq(&conf->device_lock);
4051 pr_debug("%d stripes handled\n", handled);
4053 spin_unlock_irq(&conf->device_lock);
4055 async_tx_issue_pending_all();
4056 unplug_slaves(mddev);
4058 pr_debug("--- raid5d inactive\n");
4062 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4064 raid5_conf_t *conf = mddev_to_conf(mddev);
4066 return sprintf(page, "%d\n", conf->max_nr_stripes);
4072 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4074 raid5_conf_t *conf = mddev_to_conf(mddev);
4078 if (len >= PAGE_SIZE)
4083 if (strict_strtoul(page, 10, &new))
4085 if (new <= 16 || new > 32768)
4087 while (new < conf->max_nr_stripes) {
4088 if (drop_one_stripe(conf))
4089 conf->max_nr_stripes--;
4093 err = md_allow_write(mddev);
4096 while (new > conf->max_nr_stripes) {
4097 if (grow_one_stripe(conf))
4098 conf->max_nr_stripes++;
4104 static struct md_sysfs_entry
4105 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4106 raid5_show_stripe_cache_size,
4107 raid5_store_stripe_cache_size);
4110 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4112 raid5_conf_t *conf = mddev_to_conf(mddev);
4114 return sprintf(page, "%d\n", conf->bypass_threshold);
4120 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4122 raid5_conf_t *conf = mddev_to_conf(mddev);
4124 if (len >= PAGE_SIZE)
4129 if (strict_strtoul(page, 10, &new))
4131 if (new > conf->max_nr_stripes)
4133 conf->bypass_threshold = new;
4137 static struct md_sysfs_entry
4138 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4140 raid5_show_preread_threshold,
4141 raid5_store_preread_threshold);
4144 stripe_cache_active_show(mddev_t *mddev, char *page)
4146 raid5_conf_t *conf = mddev_to_conf(mddev);
4148 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4153 static struct md_sysfs_entry
4154 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4156 static struct attribute *raid5_attrs[] = {
4157 &raid5_stripecache_size.attr,
4158 &raid5_stripecache_active.attr,
4159 &raid5_preread_bypass_threshold.attr,
4162 static struct attribute_group raid5_attrs_group = {
4164 .attrs = raid5_attrs,
4167 static raid5_conf_t *setup_conf(mddev_t *mddev)
4170 int raid_disk, memory;
4172 struct disk_info *disk;
4174 if (mddev->new_level != 5
4175 && mddev->new_level != 4
4176 && mddev->new_level != 6) {
4177 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4178 mdname(mddev), mddev->new_level);
4179 return ERR_PTR(-EIO);
4181 if ((mddev->new_level == 5
4182 && !algorithm_valid_raid5(mddev->new_layout)) ||
4183 (mddev->new_level == 6
4184 && !algorithm_valid_raid6(mddev->new_layout))) {
4185 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4186 mdname(mddev), mddev->new_layout);
4187 return ERR_PTR(-EIO);
4189 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4190 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4191 mdname(mddev), mddev->raid_disks);
4192 return ERR_PTR(-EINVAL);
4195 if (!mddev->new_chunk || mddev->new_chunk % PAGE_SIZE) {
4196 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4197 mddev->new_chunk, mdname(mddev));
4198 return ERR_PTR(-EINVAL);
4201 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4205 conf->raid_disks = mddev->raid_disks;
4206 if (mddev->reshape_position == MaxSector)
4207 conf->previous_raid_disks = mddev->raid_disks;
4209 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4211 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4216 conf->mddev = mddev;
4218 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4221 if (mddev->new_level == 6) {
4222 conf->spare_page = alloc_page(GFP_KERNEL);
4223 if (!conf->spare_page)
4226 spin_lock_init(&conf->device_lock);
4227 init_waitqueue_head(&conf->wait_for_stripe);
4228 init_waitqueue_head(&conf->wait_for_overlap);
4229 INIT_LIST_HEAD(&conf->handle_list);
4230 INIT_LIST_HEAD(&conf->hold_list);
4231 INIT_LIST_HEAD(&conf->delayed_list);
4232 INIT_LIST_HEAD(&conf->bitmap_list);
4233 INIT_LIST_HEAD(&conf->inactive_list);
4234 atomic_set(&conf->active_stripes, 0);
4235 atomic_set(&conf->preread_active_stripes, 0);
4236 atomic_set(&conf->active_aligned_reads, 0);
4237 conf->bypass_threshold = BYPASS_THRESHOLD;
4239 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4241 list_for_each_entry(rdev, &mddev->disks, same_set) {
4242 raid_disk = rdev->raid_disk;
4243 if (raid_disk >= conf->raid_disks
4246 disk = conf->disks + raid_disk;
4250 if (test_bit(In_sync, &rdev->flags)) {
4251 char b[BDEVNAME_SIZE];
4252 printk(KERN_INFO "raid5: device %s operational as raid"
4253 " disk %d\n", bdevname(rdev->bdev,b),
4256 /* Cannot rely on bitmap to complete recovery */
4260 conf->chunk_size = mddev->new_chunk;
4261 conf->level = mddev->new_level;
4262 if (conf->level == 6)
4263 conf->max_degraded = 2;
4265 conf->max_degraded = 1;
4266 conf->algorithm = mddev->new_layout;
4267 conf->max_nr_stripes = NR_STRIPES;
4268 conf->expand_progress = mddev->reshape_position;
4270 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4271 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4272 if (grow_stripes(conf, conf->max_nr_stripes)) {
4274 "raid5: couldn't allocate %dkB for buffers\n", memory);
4277 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4278 memory, mdname(mddev));
4280 conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4281 if (!conf->thread) {
4283 "raid5: couldn't allocate thread for %s\n",
4292 shrink_stripes(conf);
4293 safe_put_page(conf->spare_page);
4295 kfree(conf->stripe_hashtbl);
4297 return ERR_PTR(-EIO);
4299 return ERR_PTR(-ENOMEM);
4302 static int run(mddev_t *mddev)
4305 int working_disks = 0;
4308 if (mddev->reshape_position != MaxSector) {
4309 /* Check that we can continue the reshape.
4310 * Currently only disks can change, it must
4311 * increase, and we must be past the point where
4312 * a stripe over-writes itself
4314 sector_t here_new, here_old;
4316 int max_degraded = (mddev->level == 5 ? 1 : 2);
4318 if (mddev->new_level != mddev->level ||
4319 mddev->new_layout != mddev->layout ||
4320 mddev->new_chunk != mddev->chunk_size) {
4321 printk(KERN_ERR "raid5: %s: unsupported reshape "
4322 "required - aborting.\n",
4326 if (mddev->delta_disks <= 0) {
4327 printk(KERN_ERR "raid5: %s: unsupported reshape "
4328 "(reduce disks) required - aborting.\n",
4332 old_disks = mddev->raid_disks - mddev->delta_disks;
4333 /* reshape_position must be on a new-stripe boundary, and one
4334 * further up in new geometry must map after here in old
4337 here_new = mddev->reshape_position;
4338 if (sector_div(here_new, (mddev->chunk_size>>9)*
4339 (mddev->raid_disks - max_degraded))) {
4340 printk(KERN_ERR "raid5: reshape_position not "
4341 "on a stripe boundary\n");
4344 /* here_new is the stripe we will write to */
4345 here_old = mddev->reshape_position;
4346 sector_div(here_old, (mddev->chunk_size>>9)*
4347 (old_disks-max_degraded));
4348 /* here_old is the first stripe that we might need to read
4350 if (here_new >= here_old) {
4351 /* Reading from the same stripe as writing to - bad */
4352 printk(KERN_ERR "raid5: reshape_position too early for "
4353 "auto-recovery - aborting.\n");
4356 printk(KERN_INFO "raid5: reshape will continue\n");
4357 /* OK, we should be able to continue; */
4359 BUG_ON(mddev->level != mddev->new_level);
4360 BUG_ON(mddev->layout != mddev->new_layout);
4361 BUG_ON(mddev->chunk_size != mddev->new_chunk);
4362 BUG_ON(mddev->delta_disks != 0);
4364 conf = setup_conf(mddev);
4369 return PTR_ERR(conf);
4371 mddev->thread = conf->thread;
4372 conf->thread = NULL;
4373 mddev->private = conf;
4376 * 0 for a fully functional array, 1 or 2 for a degraded array.
4378 list_for_each_entry(rdev, &mddev->disks, same_set)
4379 if (rdev->raid_disk >= 0 &&
4380 test_bit(In_sync, &rdev->flags))
4383 mddev->degraded = conf->raid_disks - working_disks;
4385 if (mddev->degraded > conf->max_degraded) {
4386 printk(KERN_ERR "raid5: not enough operational devices for %s"
4387 " (%d/%d failed)\n",
4388 mdname(mddev), mddev->degraded, conf->raid_disks);
4392 /* device size must be a multiple of chunk size */
4393 mddev->dev_sectors &= ~(mddev->chunk_size / 512 - 1);
4394 mddev->resync_max_sectors = mddev->dev_sectors;
4396 if (mddev->degraded > 0 &&
4397 mddev->recovery_cp != MaxSector) {
4398 if (mddev->ok_start_degraded)
4400 "raid5: starting dirty degraded array: %s"
4401 "- data corruption possible.\n",
4405 "raid5: cannot start dirty degraded array for %s\n",
4411 if (mddev->degraded == 0)
4412 printk("raid5: raid level %d set %s active with %d out of %d"
4413 " devices, algorithm %d\n", conf->level, mdname(mddev),
4414 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4417 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4418 " out of %d devices, algorithm %d\n", conf->level,
4419 mdname(mddev), mddev->raid_disks - mddev->degraded,
4420 mddev->raid_disks, conf->algorithm);
4422 print_raid5_conf(conf);
4424 if (conf->expand_progress != MaxSector) {
4425 printk("...ok start reshape thread\n");
4426 conf->expand_lo = conf->expand_progress;
4427 atomic_set(&conf->reshape_stripes, 0);
4428 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4429 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4430 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4431 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4432 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4436 /* read-ahead size must cover two whole stripes, which is
4437 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4440 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4441 int stripe = data_disks *
4442 (mddev->chunk_size / PAGE_SIZE);
4443 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4444 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4447 /* Ok, everything is just fine now */
4448 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4450 "raid5: failed to create sysfs attributes for %s\n",
4453 mddev->queue->queue_lock = &conf->device_lock;
4455 mddev->queue->unplug_fn = raid5_unplug_device;
4456 mddev->queue->backing_dev_info.congested_data = mddev;
4457 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4459 mddev->array_sectors = mddev->dev_sectors *
4460 (conf->previous_raid_disks - conf->max_degraded);
4462 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4466 md_unregister_thread(mddev->thread);
4467 mddev->thread = NULL;
4469 shrink_stripes(conf);
4470 print_raid5_conf(conf);
4471 safe_put_page(conf->spare_page);
4473 kfree(conf->stripe_hashtbl);
4476 mddev->private = NULL;
4477 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4483 static int stop(mddev_t *mddev)
4485 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4487 md_unregister_thread(mddev->thread);
4488 mddev->thread = NULL;
4489 shrink_stripes(conf);
4490 kfree(conf->stripe_hashtbl);
4491 mddev->queue->backing_dev_info.congested_fn = NULL;
4492 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4493 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4496 mddev->private = NULL;
4501 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4505 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4506 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4507 seq_printf(seq, "sh %llu, count %d.\n",
4508 (unsigned long long)sh->sector, atomic_read(&sh->count));
4509 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4510 for (i = 0; i < sh->disks; i++) {
4511 seq_printf(seq, "(cache%d: %p %ld) ",
4512 i, sh->dev[i].page, sh->dev[i].flags);
4514 seq_printf(seq, "\n");
4517 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4519 struct stripe_head *sh;
4520 struct hlist_node *hn;
4523 spin_lock_irq(&conf->device_lock);
4524 for (i = 0; i < NR_HASH; i++) {
4525 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4526 if (sh->raid_conf != conf)
4531 spin_unlock_irq(&conf->device_lock);
4535 static void status(struct seq_file *seq, mddev_t *mddev)
4537 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4540 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4541 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4542 for (i = 0; i < conf->raid_disks; i++)
4543 seq_printf (seq, "%s",
4544 conf->disks[i].rdev &&
4545 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4546 seq_printf (seq, "]");
4548 seq_printf (seq, "\n");
4549 printall(seq, conf);
4553 static void print_raid5_conf (raid5_conf_t *conf)
4556 struct disk_info *tmp;
4558 printk("RAID5 conf printout:\n");
4560 printk("(conf==NULL)\n");
4563 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4564 conf->raid_disks - conf->mddev->degraded);
4566 for (i = 0; i < conf->raid_disks; i++) {
4567 char b[BDEVNAME_SIZE];
4568 tmp = conf->disks + i;
4570 printk(" disk %d, o:%d, dev:%s\n",
4571 i, !test_bit(Faulty, &tmp->rdev->flags),
4572 bdevname(tmp->rdev->bdev,b));
4576 static int raid5_spare_active(mddev_t *mddev)
4579 raid5_conf_t *conf = mddev->private;
4580 struct disk_info *tmp;
4582 for (i = 0; i < conf->raid_disks; i++) {
4583 tmp = conf->disks + i;
4585 && !test_bit(Faulty, &tmp->rdev->flags)
4586 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4587 unsigned long flags;
4588 spin_lock_irqsave(&conf->device_lock, flags);
4590 spin_unlock_irqrestore(&conf->device_lock, flags);
4593 print_raid5_conf(conf);
4597 static int raid5_remove_disk(mddev_t *mddev, int number)
4599 raid5_conf_t *conf = mddev->private;
4602 struct disk_info *p = conf->disks + number;
4604 print_raid5_conf(conf);
4607 if (test_bit(In_sync, &rdev->flags) ||
4608 atomic_read(&rdev->nr_pending)) {
4612 /* Only remove non-faulty devices if recovery
4615 if (!test_bit(Faulty, &rdev->flags) &&
4616 mddev->degraded <= conf->max_degraded) {
4622 if (atomic_read(&rdev->nr_pending)) {
4623 /* lost the race, try later */
4630 print_raid5_conf(conf);
4634 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4636 raid5_conf_t *conf = mddev->private;
4639 struct disk_info *p;
4641 int last = conf->raid_disks - 1;
4643 if (mddev->degraded > conf->max_degraded)
4644 /* no point adding a device */
4647 if (rdev->raid_disk >= 0)
4648 first = last = rdev->raid_disk;
4651 * find the disk ... but prefer rdev->saved_raid_disk
4654 if (rdev->saved_raid_disk >= 0 &&
4655 rdev->saved_raid_disk >= first &&
4656 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4657 disk = rdev->saved_raid_disk;
4660 for ( ; disk <= last ; disk++)
4661 if ((p=conf->disks + disk)->rdev == NULL) {
4662 clear_bit(In_sync, &rdev->flags);
4663 rdev->raid_disk = disk;
4665 if (rdev->saved_raid_disk != disk)
4667 rcu_assign_pointer(p->rdev, rdev);
4670 print_raid5_conf(conf);
4674 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4676 /* no resync is happening, and there is enough space
4677 * on all devices, so we can resize.
4678 * We need to make sure resync covers any new space.
4679 * If the array is shrinking we should possibly wait until
4680 * any io in the removed space completes, but it hardly seems
4683 raid5_conf_t *conf = mddev_to_conf(mddev);
4685 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4686 mddev->array_sectors = sectors * (mddev->raid_disks
4687 - conf->max_degraded);
4688 set_capacity(mddev->gendisk, mddev->array_sectors);
4690 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
4691 mddev->recovery_cp = mddev->dev_sectors;
4692 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4694 mddev->dev_sectors = sectors;
4695 mddev->resync_max_sectors = sectors;
4699 #ifdef CONFIG_MD_RAID5_RESHAPE
4700 static int raid5_check_reshape(mddev_t *mddev)
4702 raid5_conf_t *conf = mddev_to_conf(mddev);
4705 if (mddev->delta_disks < 0 ||
4706 mddev->new_level != mddev->level)
4707 return -EINVAL; /* Cannot shrink array or change level yet */
4708 if (mddev->delta_disks == 0)
4709 return 0; /* nothing to do */
4711 /* Cannot grow a bitmap yet */
4714 /* Can only proceed if there are plenty of stripe_heads.
4715 * We need a minimum of one full stripe,, and for sensible progress
4716 * it is best to have about 4 times that.
4717 * If we require 4 times, then the default 256 4K stripe_heads will
4718 * allow for chunk sizes up to 256K, which is probably OK.
4719 * If the chunk size is greater, user-space should request more
4720 * stripe_heads first.
4722 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4723 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4724 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4725 (mddev->chunk_size / STRIPE_SIZE)*4);
4729 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4733 if (mddev->degraded > conf->max_degraded)
4735 /* looks like we might be able to manage this */
4739 static int raid5_start_reshape(mddev_t *mddev)
4741 raid5_conf_t *conf = mddev_to_conf(mddev);
4744 int added_devices = 0;
4745 unsigned long flags;
4747 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4750 list_for_each_entry(rdev, &mddev->disks, same_set)
4751 if (rdev->raid_disk < 0 &&
4752 !test_bit(Faulty, &rdev->flags))
4755 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4756 /* Not enough devices even to make a degraded array
4761 atomic_set(&conf->reshape_stripes, 0);
4762 spin_lock_irq(&conf->device_lock);
4763 conf->previous_raid_disks = conf->raid_disks;
4764 conf->raid_disks += mddev->delta_disks;
4765 conf->expand_progress = 0;
4766 conf->expand_lo = 0;
4767 spin_unlock_irq(&conf->device_lock);
4769 /* Add some new drives, as many as will fit.
4770 * We know there are enough to make the newly sized array work.
4772 list_for_each_entry(rdev, &mddev->disks, same_set)
4773 if (rdev->raid_disk < 0 &&
4774 !test_bit(Faulty, &rdev->flags)) {
4775 if (raid5_add_disk(mddev, rdev) == 0) {
4777 set_bit(In_sync, &rdev->flags);
4779 rdev->recovery_offset = 0;
4780 sprintf(nm, "rd%d", rdev->raid_disk);
4781 if (sysfs_create_link(&mddev->kobj,
4784 "raid5: failed to create "
4785 " link %s for %s\n",
4791 spin_lock_irqsave(&conf->device_lock, flags);
4792 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4793 spin_unlock_irqrestore(&conf->device_lock, flags);
4794 mddev->raid_disks = conf->raid_disks;
4795 mddev->reshape_position = 0;
4796 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4798 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4799 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4800 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4801 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4802 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4804 if (!mddev->sync_thread) {
4805 mddev->recovery = 0;
4806 spin_lock_irq(&conf->device_lock);
4807 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4808 conf->expand_progress = MaxSector;
4809 spin_unlock_irq(&conf->device_lock);
4812 md_wakeup_thread(mddev->sync_thread);
4813 md_new_event(mddev);
4818 static void end_reshape(raid5_conf_t *conf)
4820 struct block_device *bdev;
4822 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4823 conf->mddev->array_sectors = conf->mddev->dev_sectors *
4824 (conf->raid_disks - conf->max_degraded);
4825 set_capacity(conf->mddev->gendisk, conf->mddev->array_sectors);
4826 conf->mddev->changed = 1;
4828 bdev = bdget_disk(conf->mddev->gendisk, 0);
4830 mutex_lock(&bdev->bd_inode->i_mutex);
4831 i_size_write(bdev->bd_inode,
4832 (loff_t)conf->mddev->array_sectors << 9);
4833 mutex_unlock(&bdev->bd_inode->i_mutex);
4836 spin_lock_irq(&conf->device_lock);
4837 conf->expand_progress = MaxSector;
4838 spin_unlock_irq(&conf->device_lock);
4839 conf->mddev->reshape_position = MaxSector;
4841 /* read-ahead size must cover two whole stripes, which is
4842 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4845 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4846 int stripe = data_disks *
4847 (conf->mddev->chunk_size / PAGE_SIZE);
4848 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4849 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4854 static void raid5_quiesce(mddev_t *mddev, int state)
4856 raid5_conf_t *conf = mddev_to_conf(mddev);
4859 case 2: /* resume for a suspend */
4860 wake_up(&conf->wait_for_overlap);
4863 case 1: /* stop all writes */
4864 spin_lock_irq(&conf->device_lock);
4866 wait_event_lock_irq(conf->wait_for_stripe,
4867 atomic_read(&conf->active_stripes) == 0 &&
4868 atomic_read(&conf->active_aligned_reads) == 0,
4869 conf->device_lock, /* nothing */);
4870 spin_unlock_irq(&conf->device_lock);
4873 case 0: /* re-enable writes */
4874 spin_lock_irq(&conf->device_lock);
4876 wake_up(&conf->wait_for_stripe);
4877 wake_up(&conf->wait_for_overlap);
4878 spin_unlock_irq(&conf->device_lock);
4883 static struct mdk_personality raid6_personality =
4887 .owner = THIS_MODULE,
4888 .make_request = make_request,
4892 .error_handler = error,
4893 .hot_add_disk = raid5_add_disk,
4894 .hot_remove_disk= raid5_remove_disk,
4895 .spare_active = raid5_spare_active,
4896 .sync_request = sync_request,
4897 .resize = raid5_resize,
4898 #ifdef CONFIG_MD_RAID5_RESHAPE
4899 .check_reshape = raid5_check_reshape,
4900 .start_reshape = raid5_start_reshape,
4902 .quiesce = raid5_quiesce,
4904 static struct mdk_personality raid5_personality =
4908 .owner = THIS_MODULE,
4909 .make_request = make_request,
4913 .error_handler = error,
4914 .hot_add_disk = raid5_add_disk,
4915 .hot_remove_disk= raid5_remove_disk,
4916 .spare_active = raid5_spare_active,
4917 .sync_request = sync_request,
4918 .resize = raid5_resize,
4919 #ifdef CONFIG_MD_RAID5_RESHAPE
4920 .check_reshape = raid5_check_reshape,
4921 .start_reshape = raid5_start_reshape,
4923 .quiesce = raid5_quiesce,
4926 static struct mdk_personality raid4_personality =
4930 .owner = THIS_MODULE,
4931 .make_request = make_request,
4935 .error_handler = error,
4936 .hot_add_disk = raid5_add_disk,
4937 .hot_remove_disk= raid5_remove_disk,
4938 .spare_active = raid5_spare_active,
4939 .sync_request = sync_request,
4940 .resize = raid5_resize,
4941 #ifdef CONFIG_MD_RAID5_RESHAPE
4942 .check_reshape = raid5_check_reshape,
4943 .start_reshape = raid5_start_reshape,
4945 .quiesce = raid5_quiesce,
4948 static int __init raid5_init(void)
4952 e = raid6_select_algo();
4955 register_md_personality(&raid6_personality);
4956 register_md_personality(&raid5_personality);
4957 register_md_personality(&raid4_personality);
4961 static void raid5_exit(void)
4963 unregister_md_personality(&raid6_personality);
4964 unregister_md_personality(&raid5_personality);
4965 unregister_md_personality(&raid4_personality);
4968 module_init(raid5_init);
4969 module_exit(raid5_exit);
4970 MODULE_LICENSE("GPL");
4971 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4972 MODULE_ALIAS("md-raid5");
4973 MODULE_ALIAS("md-raid4");
4974 MODULE_ALIAS("md-level-5");
4975 MODULE_ALIAS("md-level-4");
4976 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4977 MODULE_ALIAS("md-raid6");
4978 MODULE_ALIAS("md-level-6");
4980 /* This used to be two separate modules, they were: */
4981 MODULE_ALIAS("raid5");
4982 MODULE_ALIAS("raid6");